]>
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 | |
ea435e1b | 41 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie); |
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
720b8ccc SB |
45 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
46 | { | |
47 | int ddir, bytes, bucket; | |
48 | ||
99c749a4 | 49 | ddir = rq_data_dir(rq); |
720b8ccc SB |
50 | bytes = blk_rq_bytes(rq); |
51 | ||
52 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
53 | ||
54 | if (bucket < 0) | |
55 | return -1; | |
56 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
57 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
58 | ||
59 | return bucket; | |
60 | } | |
61 | ||
320ae51f JA |
62 | /* |
63 | * Check if any of the ctx's have pending work in this hardware queue | |
64 | */ | |
79f720a7 | 65 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 66 | { |
79f720a7 JA |
67 | return !list_empty_careful(&hctx->dispatch) || |
68 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 69 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
70 | } |
71 | ||
320ae51f JA |
72 | /* |
73 | * Mark this ctx as having pending work in this hardware queue | |
74 | */ | |
75 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
76 | struct blk_mq_ctx *ctx) | |
77 | { | |
88459642 OS |
78 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
79 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
80 | } |
81 | ||
82 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
83 | struct blk_mq_ctx *ctx) | |
84 | { | |
88459642 | 85 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
86 | } |
87 | ||
f299b7c7 JA |
88 | struct mq_inflight { |
89 | struct hd_struct *part; | |
90 | unsigned int *inflight; | |
91 | }; | |
92 | ||
93 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
94 | struct request *rq, void *priv, | |
95 | bool reserved) | |
96 | { | |
97 | struct mq_inflight *mi = priv; | |
98 | ||
6131837b OS |
99 | /* |
100 | * index[0] counts the specific partition that was asked for. index[1] | |
101 | * counts the ones that are active on the whole device, so increment | |
102 | * that if mi->part is indeed a partition, and not a whole device. | |
103 | */ | |
104 | if (rq->part == mi->part) | |
105 | mi->inflight[0]++; | |
106 | if (mi->part->partno) | |
107 | mi->inflight[1]++; | |
f299b7c7 JA |
108 | } |
109 | ||
110 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
111 | unsigned int inflight[2]) | |
112 | { | |
113 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
114 | ||
b8d62b3a | 115 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
116 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
117 | } | |
118 | ||
bf0ddaba OS |
119 | static void blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
120 | struct request *rq, void *priv, | |
121 | bool reserved) | |
122 | { | |
123 | struct mq_inflight *mi = priv; | |
124 | ||
125 | if (rq->part == mi->part) | |
126 | mi->inflight[rq_data_dir(rq)]++; | |
127 | } | |
128 | ||
129 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
130 | unsigned int inflight[2]) | |
131 | { | |
132 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
133 | ||
134 | inflight[0] = inflight[1] = 0; | |
135 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
136 | } | |
137 | ||
1671d522 | 138 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 139 | { |
4ecd4fef | 140 | int freeze_depth; |
cddd5d17 | 141 | |
4ecd4fef CH |
142 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
143 | if (freeze_depth == 1) { | |
3ef28e83 | 144 | percpu_ref_kill(&q->q_usage_counter); |
055f6e18 ML |
145 | if (q->mq_ops) |
146 | blk_mq_run_hw_queues(q, false); | |
cddd5d17 | 147 | } |
f3af020b | 148 | } |
1671d522 | 149 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 150 | |
6bae363e | 151 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 152 | { |
3ef28e83 | 153 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 154 | } |
6bae363e | 155 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 156 | |
f91328c4 KB |
157 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
158 | unsigned long timeout) | |
159 | { | |
160 | return wait_event_timeout(q->mq_freeze_wq, | |
161 | percpu_ref_is_zero(&q->q_usage_counter), | |
162 | timeout); | |
163 | } | |
164 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 165 | |
f3af020b TH |
166 | /* |
167 | * Guarantee no request is in use, so we can change any data structure of | |
168 | * the queue afterward. | |
169 | */ | |
3ef28e83 | 170 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 171 | { |
3ef28e83 DW |
172 | /* |
173 | * In the !blk_mq case we are only calling this to kill the | |
174 | * q_usage_counter, otherwise this increases the freeze depth | |
175 | * and waits for it to return to zero. For this reason there is | |
176 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
177 | * exported to drivers as the only user for unfreeze is blk_mq. | |
178 | */ | |
1671d522 | 179 | blk_freeze_queue_start(q); |
454be724 ML |
180 | if (!q->mq_ops) |
181 | blk_drain_queue(q); | |
f3af020b TH |
182 | blk_mq_freeze_queue_wait(q); |
183 | } | |
3ef28e83 DW |
184 | |
185 | void blk_mq_freeze_queue(struct request_queue *q) | |
186 | { | |
187 | /* | |
188 | * ...just an alias to keep freeze and unfreeze actions balanced | |
189 | * in the blk_mq_* namespace | |
190 | */ | |
191 | blk_freeze_queue(q); | |
192 | } | |
c761d96b | 193 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 194 | |
b4c6a028 | 195 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 196 | { |
4ecd4fef | 197 | int freeze_depth; |
320ae51f | 198 | |
4ecd4fef CH |
199 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
200 | WARN_ON_ONCE(freeze_depth < 0); | |
201 | if (!freeze_depth) { | |
bdd63160 | 202 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 203 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 204 | } |
320ae51f | 205 | } |
b4c6a028 | 206 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 207 | |
852ec809 BVA |
208 | /* |
209 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
210 | * mpt3sas driver such that this function can be removed. | |
211 | */ | |
212 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
213 | { | |
8814ce8a | 214 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
215 | } |
216 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
217 | ||
6a83e74d | 218 | /** |
69e07c4a | 219 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
220 | * @q: request queue. |
221 | * | |
222 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
223 | * callback function is invoked. Once this function is returned, we make |
224 | * sure no dispatch can happen until the queue is unquiesced via | |
225 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
226 | */ |
227 | void blk_mq_quiesce_queue(struct request_queue *q) | |
228 | { | |
229 | struct blk_mq_hw_ctx *hctx; | |
230 | unsigned int i; | |
231 | bool rcu = false; | |
232 | ||
1d9e9bc6 | 233 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 234 | |
6a83e74d BVA |
235 | queue_for_each_hw_ctx(q, hctx, i) { |
236 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 237 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
238 | else |
239 | rcu = true; | |
240 | } | |
241 | if (rcu) | |
242 | synchronize_rcu(); | |
243 | } | |
244 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
245 | ||
e4e73913 ML |
246 | /* |
247 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
248 | * @q: request queue. | |
249 | * | |
250 | * This function recovers queue into the state before quiescing | |
251 | * which is done by blk_mq_quiesce_queue. | |
252 | */ | |
253 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
254 | { | |
8814ce8a | 255 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 256 | |
1d9e9bc6 ML |
257 | /* dispatch requests which are inserted during quiescing */ |
258 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
259 | } |
260 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
261 | ||
aed3ea94 JA |
262 | void blk_mq_wake_waiters(struct request_queue *q) |
263 | { | |
264 | struct blk_mq_hw_ctx *hctx; | |
265 | unsigned int i; | |
266 | ||
267 | queue_for_each_hw_ctx(q, hctx, i) | |
268 | if (blk_mq_hw_queue_mapped(hctx)) | |
269 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
270 | } | |
271 | ||
320ae51f JA |
272 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
273 | { | |
274 | return blk_mq_has_free_tags(hctx->tags); | |
275 | } | |
276 | EXPORT_SYMBOL(blk_mq_can_queue); | |
277 | ||
e4cdf1a1 CH |
278 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
279 | unsigned int tag, unsigned int op) | |
320ae51f | 280 | { |
e4cdf1a1 CH |
281 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
282 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 283 | req_flags_t rq_flags = 0; |
c3a148d2 | 284 | |
e4cdf1a1 CH |
285 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
286 | rq->tag = -1; | |
287 | rq->internal_tag = tag; | |
288 | } else { | |
d263ed99 | 289 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 290 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
291 | atomic_inc(&data->hctx->nr_active); |
292 | } | |
293 | rq->tag = tag; | |
294 | rq->internal_tag = -1; | |
295 | data->hctx->tags->rqs[rq->tag] = rq; | |
296 | } | |
297 | ||
af76e555 | 298 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
299 | rq->q = data->q; |
300 | rq->mq_ctx = data->ctx; | |
bf9ae8c5 | 301 | rq->rq_flags = rq_flags; |
7c3fb70f | 302 | rq->cpu = -1; |
ef295ecf | 303 | rq->cmd_flags = op; |
1b6d65a0 BVA |
304 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
305 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 306 | if (blk_queue_io_stat(data->q)) |
e8064021 | 307 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 308 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
309 | INIT_HLIST_NODE(&rq->hash); |
310 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
311 | rq->rq_disk = NULL; |
312 | rq->part = NULL; | |
522a7775 | 313 | rq->start_time_ns = ktime_get_ns(); |
544ccc8d | 314 | rq->io_start_time_ns = 0; |
af76e555 CH |
315 | rq->nr_phys_segments = 0; |
316 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
317 | rq->nr_integrity_segments = 0; | |
318 | #endif | |
af76e555 CH |
319 | rq->special = NULL; |
320 | /* tag was already set */ | |
af76e555 | 321 | rq->extra_len = 0; |
e14575b3 | 322 | rq->__deadline = 0; |
af76e555 | 323 | |
af76e555 | 324 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
325 | rq->timeout = 0; |
326 | ||
af76e555 CH |
327 | rq->end_io = NULL; |
328 | rq->end_io_data = NULL; | |
329 | rq->next_rq = NULL; | |
330 | ||
7c3fb70f JA |
331 | #ifdef CONFIG_BLK_CGROUP |
332 | rq->rl = NULL; | |
7c3fb70f JA |
333 | #endif |
334 | ||
e4cdf1a1 | 335 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 336 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 337 | return rq; |
5dee8577 CH |
338 | } |
339 | ||
d2c0d383 CH |
340 | static struct request *blk_mq_get_request(struct request_queue *q, |
341 | struct bio *bio, unsigned int op, | |
342 | struct blk_mq_alloc_data *data) | |
343 | { | |
344 | struct elevator_queue *e = q->elevator; | |
345 | struct request *rq; | |
e4cdf1a1 | 346 | unsigned int tag; |
21e768b4 | 347 | bool put_ctx_on_error = false; |
d2c0d383 CH |
348 | |
349 | blk_queue_enter_live(q); | |
350 | data->q = q; | |
21e768b4 BVA |
351 | if (likely(!data->ctx)) { |
352 | data->ctx = blk_mq_get_ctx(q); | |
353 | put_ctx_on_error = true; | |
354 | } | |
d2c0d383 CH |
355 | if (likely(!data->hctx)) |
356 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
357 | if (op & REQ_NOWAIT) |
358 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
359 | |
360 | if (e) { | |
361 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
362 | ||
363 | /* | |
364 | * Flush requests are special and go directly to the | |
17a51199 JA |
365 | * dispatch list. Don't include reserved tags in the |
366 | * limiting, as it isn't useful. | |
d2c0d383 | 367 | */ |
17a51199 JA |
368 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth && |
369 | !(data->flags & BLK_MQ_REQ_RESERVED)) | |
5bbf4e5a | 370 | e->type->ops.mq.limit_depth(op, data); |
d263ed99 JW |
371 | } else { |
372 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
373 | } |
374 | ||
e4cdf1a1 CH |
375 | tag = blk_mq_get_tag(data); |
376 | if (tag == BLK_MQ_TAG_FAIL) { | |
21e768b4 BVA |
377 | if (put_ctx_on_error) { |
378 | blk_mq_put_ctx(data->ctx); | |
1ad43c00 ML |
379 | data->ctx = NULL; |
380 | } | |
037cebb8 CH |
381 | blk_queue_exit(q); |
382 | return NULL; | |
d2c0d383 CH |
383 | } |
384 | ||
e4cdf1a1 | 385 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
386 | if (!op_is_flush(op)) { |
387 | rq->elv.icq = NULL; | |
5bbf4e5a | 388 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
389 | if (e->type->icq_cache && rq_ioc(bio)) |
390 | blk_mq_sched_assign_ioc(rq, bio); | |
391 | ||
5bbf4e5a CH |
392 | e->type->ops.mq.prepare_request(rq, bio); |
393 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 394 | } |
037cebb8 CH |
395 | } |
396 | data->hctx->queued++; | |
397 | return rq; | |
d2c0d383 CH |
398 | } |
399 | ||
cd6ce148 | 400 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 401 | blk_mq_req_flags_t flags) |
320ae51f | 402 | { |
5a797e00 | 403 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 404 | struct request *rq; |
a492f075 | 405 | int ret; |
320ae51f | 406 | |
3a0a5299 | 407 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
408 | if (ret) |
409 | return ERR_PTR(ret); | |
320ae51f | 410 | |
cd6ce148 | 411 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 412 | blk_queue_exit(q); |
841bac2c | 413 | |
bd166ef1 | 414 | if (!rq) |
a492f075 | 415 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 416 | |
1ad43c00 | 417 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 418 | |
0c4de0f3 CH |
419 | rq->__data_len = 0; |
420 | rq->__sector = (sector_t) -1; | |
421 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
422 | return rq; |
423 | } | |
4bb659b1 | 424 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 425 | |
cd6ce148 | 426 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 427 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 428 | { |
6d2809d5 | 429 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 430 | struct request *rq; |
6d2809d5 | 431 | unsigned int cpu; |
1f5bd336 ML |
432 | int ret; |
433 | ||
434 | /* | |
435 | * If the tag allocator sleeps we could get an allocation for a | |
436 | * different hardware context. No need to complicate the low level | |
437 | * allocator for this for the rare use case of a command tied to | |
438 | * a specific queue. | |
439 | */ | |
440 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
441 | return ERR_PTR(-EINVAL); | |
442 | ||
443 | if (hctx_idx >= q->nr_hw_queues) | |
444 | return ERR_PTR(-EIO); | |
445 | ||
3a0a5299 | 446 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
447 | if (ret) |
448 | return ERR_PTR(ret); | |
449 | ||
c8712c6a CH |
450 | /* |
451 | * Check if the hardware context is actually mapped to anything. | |
452 | * If not tell the caller that it should skip this queue. | |
453 | */ | |
6d2809d5 OS |
454 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
455 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
456 | blk_queue_exit(q); | |
457 | return ERR_PTR(-EXDEV); | |
c8712c6a | 458 | } |
20e4d813 | 459 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 460 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 461 | |
cd6ce148 | 462 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 463 | blk_queue_exit(q); |
c8712c6a | 464 | |
6d2809d5 OS |
465 | if (!rq) |
466 | return ERR_PTR(-EWOULDBLOCK); | |
467 | ||
468 | return rq; | |
1f5bd336 ML |
469 | } |
470 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
471 | ||
12f5b931 KB |
472 | static void __blk_mq_free_request(struct request *rq) |
473 | { | |
474 | struct request_queue *q = rq->q; | |
475 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
476 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
477 | const int sched_tag = rq->internal_tag; | |
478 | ||
986d413b | 479 | blk_pm_mark_last_busy(rq); |
12f5b931 KB |
480 | if (rq->tag != -1) |
481 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
482 | if (sched_tag != -1) | |
483 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
484 | blk_mq_sched_restart(hctx); | |
485 | blk_queue_exit(q); | |
486 | } | |
487 | ||
6af54051 | 488 | void blk_mq_free_request(struct request *rq) |
320ae51f | 489 | { |
320ae51f | 490 | struct request_queue *q = rq->q; |
6af54051 CH |
491 | struct elevator_queue *e = q->elevator; |
492 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
493 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
6af54051 | 494 | |
5bbf4e5a | 495 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
496 | if (e && e->type->ops.mq.finish_request) |
497 | e->type->ops.mq.finish_request(rq); | |
498 | if (rq->elv.icq) { | |
499 | put_io_context(rq->elv.icq->ioc); | |
500 | rq->elv.icq = NULL; | |
501 | } | |
502 | } | |
320ae51f | 503 | |
6af54051 | 504 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 505 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 506 | atomic_dec(&hctx->nr_active); |
87760e5e | 507 | |
7beb2f84 JA |
508 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
509 | laptop_io_completion(q->backing_dev_info); | |
510 | ||
a7905043 | 511 | rq_qos_done(q, rq); |
0d2602ca | 512 | |
85acb3ba SL |
513 | if (blk_rq_rl(rq)) |
514 | blk_put_rl(blk_rq_rl(rq)); | |
515 | ||
12f5b931 KB |
516 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
517 | if (refcount_dec_and_test(&rq->ref)) | |
518 | __blk_mq_free_request(rq); | |
320ae51f | 519 | } |
1a3b595a | 520 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 521 | |
2a842aca | 522 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 523 | { |
522a7775 OS |
524 | u64 now = ktime_get_ns(); |
525 | ||
4bc6339a OS |
526 | if (rq->rq_flags & RQF_STATS) { |
527 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 528 | blk_stat_add(rq, now); |
4bc6339a OS |
529 | } |
530 | ||
ed88660a OS |
531 | if (rq->internal_tag != -1) |
532 | blk_mq_sched_completed_request(rq, now); | |
533 | ||
522a7775 | 534 | blk_account_io_done(rq, now); |
0d11e6ac | 535 | |
91b63639 | 536 | if (rq->end_io) { |
a7905043 | 537 | rq_qos_done(rq->q, rq); |
320ae51f | 538 | rq->end_io(rq, error); |
91b63639 CH |
539 | } else { |
540 | if (unlikely(blk_bidi_rq(rq))) | |
541 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 542 | blk_mq_free_request(rq); |
91b63639 | 543 | } |
320ae51f | 544 | } |
c8a446ad | 545 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 546 | |
2a842aca | 547 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
548 | { |
549 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
550 | BUG(); | |
c8a446ad | 551 | __blk_mq_end_request(rq, error); |
63151a44 | 552 | } |
c8a446ad | 553 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 554 | |
30a91cb4 | 555 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 556 | { |
3d6efbf6 | 557 | struct request *rq = data; |
320ae51f | 558 | |
30a91cb4 | 559 | rq->q->softirq_done_fn(rq); |
320ae51f | 560 | } |
320ae51f | 561 | |
453f8341 | 562 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
563 | { |
564 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 565 | bool shared = false; |
320ae51f JA |
566 | int cpu; |
567 | ||
0fc09f92 | 568 | if (!blk_mq_mark_complete(rq)) |
12f5b931 | 569 | return; |
453f8341 | 570 | |
36e76539 ML |
571 | /* |
572 | * Most of single queue controllers, there is only one irq vector | |
573 | * for handling IO completion, and the only irq's affinity is set | |
574 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
575 | * irq is handled on one specific CPU. | |
576 | * | |
577 | * So complete IO reqeust in softirq context in case of single queue | |
578 | * for not degrading IO performance by irqsoff latency. | |
579 | */ | |
580 | if (rq->q->nr_hw_queues == 1) { | |
581 | __blk_complete_request(rq); | |
582 | return; | |
583 | } | |
584 | ||
38535201 | 585 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
586 | rq->q->softirq_done_fn(rq); |
587 | return; | |
588 | } | |
320ae51f JA |
589 | |
590 | cpu = get_cpu(); | |
38535201 CH |
591 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
592 | shared = cpus_share_cache(cpu, ctx->cpu); | |
593 | ||
594 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 595 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
596 | rq->csd.info = rq; |
597 | rq->csd.flags = 0; | |
c46fff2a | 598 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 599 | } else { |
30a91cb4 | 600 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 601 | } |
320ae51f JA |
602 | put_cpu(); |
603 | } | |
30a91cb4 | 604 | |
04ced159 | 605 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 606 | __releases(hctx->srcu) |
04ced159 JA |
607 | { |
608 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
609 | rcu_read_unlock(); | |
610 | else | |
05707b64 | 611 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
612 | } |
613 | ||
614 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 615 | __acquires(hctx->srcu) |
04ced159 | 616 | { |
08b5a6e2 JA |
617 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
618 | /* shut up gcc false positive */ | |
619 | *srcu_idx = 0; | |
04ced159 | 620 | rcu_read_lock(); |
08b5a6e2 | 621 | } else |
05707b64 | 622 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
623 | } |
624 | ||
30a91cb4 CH |
625 | /** |
626 | * blk_mq_complete_request - end I/O on a request | |
627 | * @rq: the request being processed | |
628 | * | |
629 | * Description: | |
630 | * Ends all I/O on a request. It does not handle partial completions. | |
631 | * The actual completion happens out-of-order, through a IPI handler. | |
632 | **/ | |
08e0029a | 633 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 634 | { |
12f5b931 | 635 | if (unlikely(blk_should_fake_timeout(rq->q))) |
30a91cb4 | 636 | return; |
12f5b931 | 637 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
638 | } |
639 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 640 | |
973c0191 KB |
641 | int blk_mq_request_started(struct request *rq) |
642 | { | |
5a61c363 | 643 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
644 | } |
645 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
646 | ||
e2490073 | 647 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
648 | { |
649 | struct request_queue *q = rq->q; | |
650 | ||
bd166ef1 JA |
651 | blk_mq_sched_started_request(rq); |
652 | ||
320ae51f JA |
653 | trace_block_rq_issue(q, rq); |
654 | ||
cf43e6be | 655 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d OS |
656 | rq->io_start_time_ns = ktime_get_ns(); |
657 | #ifdef CONFIG_BLK_DEV_THROTTLING_LOW | |
658 | rq->throtl_size = blk_rq_sectors(rq); | |
659 | #endif | |
cf43e6be | 660 | rq->rq_flags |= RQF_STATS; |
a7905043 | 661 | rq_qos_issue(q, rq); |
cf43e6be JA |
662 | } |
663 | ||
1d9bd516 | 664 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 665 | |
1d9bd516 | 666 | blk_add_timer(rq); |
12f5b931 | 667 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
668 | |
669 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
670 | /* | |
671 | * Make sure space for the drain appears. We know we can do | |
672 | * this because max_hw_segments has been adjusted to be one | |
673 | * fewer than the device can handle. | |
674 | */ | |
675 | rq->nr_phys_segments++; | |
676 | } | |
320ae51f | 677 | } |
e2490073 | 678 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 679 | |
ed0791b2 | 680 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
681 | { |
682 | struct request_queue *q = rq->q; | |
683 | ||
923218f6 ML |
684 | blk_mq_put_driver_tag(rq); |
685 | ||
320ae51f | 686 | trace_block_rq_requeue(q, rq); |
a7905043 | 687 | rq_qos_requeue(q, rq); |
49f5baa5 | 688 | |
12f5b931 KB |
689 | if (blk_mq_request_started(rq)) { |
690 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 691 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
692 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
693 | rq->nr_phys_segments--; | |
694 | } | |
320ae51f JA |
695 | } |
696 | ||
2b053aca | 697 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 698 | { |
ed0791b2 | 699 | __blk_mq_requeue_request(rq); |
ed0791b2 | 700 | |
105976f5 ML |
701 | /* this request will be re-inserted to io scheduler queue */ |
702 | blk_mq_sched_requeue_request(rq); | |
703 | ||
ed0791b2 | 704 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 705 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
706 | } |
707 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
708 | ||
6fca6a61 CH |
709 | static void blk_mq_requeue_work(struct work_struct *work) |
710 | { | |
711 | struct request_queue *q = | |
2849450a | 712 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
713 | LIST_HEAD(rq_list); |
714 | struct request *rq, *next; | |
6fca6a61 | 715 | |
18e9781d | 716 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 717 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 718 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
719 | |
720 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 721 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
722 | continue; |
723 | ||
e8064021 | 724 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 725 | list_del_init(&rq->queuelist); |
9e97d295 | 726 | blk_mq_sched_insert_request(rq, true, false, false); |
6fca6a61 CH |
727 | } |
728 | ||
729 | while (!list_empty(&rq_list)) { | |
730 | rq = list_entry(rq_list.next, struct request, queuelist); | |
731 | list_del_init(&rq->queuelist); | |
9e97d295 | 732 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
733 | } |
734 | ||
52d7f1b5 | 735 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
736 | } |
737 | ||
2b053aca BVA |
738 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
739 | bool kick_requeue_list) | |
6fca6a61 CH |
740 | { |
741 | struct request_queue *q = rq->q; | |
742 | unsigned long flags; | |
743 | ||
744 | /* | |
745 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 746 | * request head insertion from the workqueue. |
6fca6a61 | 747 | */ |
e8064021 | 748 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
749 | |
750 | spin_lock_irqsave(&q->requeue_lock, flags); | |
751 | if (at_head) { | |
e8064021 | 752 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
753 | list_add(&rq->queuelist, &q->requeue_list); |
754 | } else { | |
755 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
756 | } | |
757 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
758 | |
759 | if (kick_requeue_list) | |
760 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
761 | } |
762 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
763 | ||
764 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
765 | { | |
ae943d20 | 766 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
767 | } |
768 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
769 | ||
2849450a MS |
770 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
771 | unsigned long msecs) | |
772 | { | |
d4acf365 BVA |
773 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
774 | msecs_to_jiffies(msecs)); | |
2849450a MS |
775 | } |
776 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
777 | ||
0e62f51f JA |
778 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
779 | { | |
88c7b2b7 JA |
780 | if (tag < tags->nr_tags) { |
781 | prefetch(tags->rqs[tag]); | |
4ee86bab | 782 | return tags->rqs[tag]; |
88c7b2b7 | 783 | } |
4ee86bab HR |
784 | |
785 | return NULL; | |
24d2f903 CH |
786 | } |
787 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
788 | ||
358f70da | 789 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 790 | { |
da661267 | 791 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
792 | if (req->q->mq_ops->timeout) { |
793 | enum blk_eh_timer_return ret; | |
794 | ||
795 | ret = req->q->mq_ops->timeout(req, reserved); | |
796 | if (ret == BLK_EH_DONE) | |
797 | return; | |
798 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 799 | } |
d1210d5a CH |
800 | |
801 | blk_add_timer(req); | |
87ee7b11 | 802 | } |
5b3f25fc | 803 | |
12f5b931 | 804 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 805 | { |
12f5b931 | 806 | unsigned long deadline; |
87ee7b11 | 807 | |
12f5b931 KB |
808 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
809 | return false; | |
da661267 CH |
810 | if (rq->rq_flags & RQF_TIMED_OUT) |
811 | return false; | |
a7af0af3 | 812 | |
12f5b931 KB |
813 | deadline = blk_rq_deadline(rq); |
814 | if (time_after_eq(jiffies, deadline)) | |
815 | return true; | |
a7af0af3 | 816 | |
12f5b931 KB |
817 | if (*next == 0) |
818 | *next = deadline; | |
819 | else if (time_after(*next, deadline)) | |
820 | *next = deadline; | |
821 | return false; | |
87ee7b11 JA |
822 | } |
823 | ||
12f5b931 | 824 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
825 | struct request *rq, void *priv, bool reserved) |
826 | { | |
12f5b931 KB |
827 | unsigned long *next = priv; |
828 | ||
829 | /* | |
830 | * Just do a quick check if it is expired before locking the request in | |
831 | * so we're not unnecessarilly synchronizing across CPUs. | |
832 | */ | |
833 | if (!blk_mq_req_expired(rq, next)) | |
834 | return; | |
835 | ||
836 | /* | |
837 | * We have reason to believe the request may be expired. Take a | |
838 | * reference on the request to lock this request lifetime into its | |
839 | * currently allocated context to prevent it from being reallocated in | |
840 | * the event the completion by-passes this timeout handler. | |
841 | * | |
842 | * If the reference was already released, then the driver beat the | |
843 | * timeout handler to posting a natural completion. | |
844 | */ | |
845 | if (!refcount_inc_not_zero(&rq->ref)) | |
846 | return; | |
847 | ||
1d9bd516 | 848 | /* |
12f5b931 KB |
849 | * The request is now locked and cannot be reallocated underneath the |
850 | * timeout handler's processing. Re-verify this exact request is truly | |
851 | * expired; if it is not expired, then the request was completed and | |
852 | * reallocated as a new request. | |
1d9bd516 | 853 | */ |
12f5b931 | 854 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 855 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
856 | if (refcount_dec_and_test(&rq->ref)) |
857 | __blk_mq_free_request(rq); | |
1d9bd516 TH |
858 | } |
859 | ||
287922eb | 860 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 861 | { |
287922eb CH |
862 | struct request_queue *q = |
863 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 864 | unsigned long next = 0; |
1d9bd516 | 865 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 866 | int i; |
320ae51f | 867 | |
71f79fb3 GKB |
868 | /* A deadlock might occur if a request is stuck requiring a |
869 | * timeout at the same time a queue freeze is waiting | |
870 | * completion, since the timeout code would not be able to | |
871 | * acquire the queue reference here. | |
872 | * | |
873 | * That's why we don't use blk_queue_enter here; instead, we use | |
874 | * percpu_ref_tryget directly, because we need to be able to | |
875 | * obtain a reference even in the short window between the queue | |
876 | * starting to freeze, by dropping the first reference in | |
1671d522 | 877 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
878 | * consumed, marked by the instant q_usage_counter reaches |
879 | * zero. | |
880 | */ | |
881 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
882 | return; |
883 | ||
12f5b931 | 884 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 885 | |
12f5b931 KB |
886 | if (next != 0) { |
887 | mod_timer(&q->timeout, next); | |
0d2602ca | 888 | } else { |
fcd36c36 BVA |
889 | /* |
890 | * Request timeouts are handled as a forward rolling timer. If | |
891 | * we end up here it means that no requests are pending and | |
892 | * also that no request has been pending for a while. Mark | |
893 | * each hctx as idle. | |
894 | */ | |
f054b56c ML |
895 | queue_for_each_hw_ctx(q, hctx, i) { |
896 | /* the hctx may be unmapped, so check it here */ | |
897 | if (blk_mq_hw_queue_mapped(hctx)) | |
898 | blk_mq_tag_idle(hctx); | |
899 | } | |
0d2602ca | 900 | } |
287922eb | 901 | blk_queue_exit(q); |
320ae51f JA |
902 | } |
903 | ||
88459642 OS |
904 | struct flush_busy_ctx_data { |
905 | struct blk_mq_hw_ctx *hctx; | |
906 | struct list_head *list; | |
907 | }; | |
908 | ||
909 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
910 | { | |
911 | struct flush_busy_ctx_data *flush_data = data; | |
912 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
913 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
914 | ||
88459642 OS |
915 | spin_lock(&ctx->lock); |
916 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
e9a99a63 | 917 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
918 | spin_unlock(&ctx->lock); |
919 | return true; | |
920 | } | |
921 | ||
1429d7c9 JA |
922 | /* |
923 | * Process software queues that have been marked busy, splicing them | |
924 | * to the for-dispatch | |
925 | */ | |
2c3ad667 | 926 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 927 | { |
88459642 OS |
928 | struct flush_busy_ctx_data data = { |
929 | .hctx = hctx, | |
930 | .list = list, | |
931 | }; | |
1429d7c9 | 932 | |
88459642 | 933 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 934 | } |
2c3ad667 | 935 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 936 | |
b347689f ML |
937 | struct dispatch_rq_data { |
938 | struct blk_mq_hw_ctx *hctx; | |
939 | struct request *rq; | |
940 | }; | |
941 | ||
942 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
943 | void *data) | |
944 | { | |
945 | struct dispatch_rq_data *dispatch_data = data; | |
946 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
947 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
948 | ||
949 | spin_lock(&ctx->lock); | |
b4f6f38d | 950 | if (!list_empty(&ctx->rq_list)) { |
b347689f ML |
951 | dispatch_data->rq = list_entry_rq(ctx->rq_list.next); |
952 | list_del_init(&dispatch_data->rq->queuelist); | |
953 | if (list_empty(&ctx->rq_list)) | |
954 | sbitmap_clear_bit(sb, bitnr); | |
955 | } | |
956 | spin_unlock(&ctx->lock); | |
957 | ||
958 | return !dispatch_data->rq; | |
959 | } | |
960 | ||
961 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
962 | struct blk_mq_ctx *start) | |
963 | { | |
964 | unsigned off = start ? start->index_hw : 0; | |
965 | struct dispatch_rq_data data = { | |
966 | .hctx = hctx, | |
967 | .rq = NULL, | |
968 | }; | |
969 | ||
970 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
971 | dispatch_rq_from_ctx, &data); | |
972 | ||
973 | return data.rq; | |
974 | } | |
975 | ||
703fd1c0 JA |
976 | static inline unsigned int queued_to_index(unsigned int queued) |
977 | { | |
978 | if (!queued) | |
979 | return 0; | |
1429d7c9 | 980 | |
703fd1c0 | 981 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
982 | } |
983 | ||
8ab6bb9e | 984 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
985 | { |
986 | struct blk_mq_alloc_data data = { | |
987 | .q = rq->q, | |
bd166ef1 | 988 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
8ab6bb9e | 989 | .flags = BLK_MQ_REQ_NOWAIT, |
bd166ef1 | 990 | }; |
d263ed99 | 991 | bool shared; |
5feeacdd | 992 | |
81380ca1 OS |
993 | if (rq->tag != -1) |
994 | goto done; | |
bd166ef1 | 995 | |
415b806d SG |
996 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
997 | data.flags |= BLK_MQ_REQ_RESERVED; | |
998 | ||
d263ed99 | 999 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1000 | rq->tag = blk_mq_get_tag(&data); |
1001 | if (rq->tag >= 0) { | |
d263ed99 | 1002 | if (shared) { |
200e86b3 JA |
1003 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1004 | atomic_inc(&data.hctx->nr_active); | |
1005 | } | |
bd166ef1 | 1006 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1007 | } |
1008 | ||
81380ca1 | 1009 | done: |
81380ca1 | 1010 | return rq->tag != -1; |
bd166ef1 JA |
1011 | } |
1012 | ||
eb619fdb JA |
1013 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1014 | int flags, void *key) | |
da55f2cc OS |
1015 | { |
1016 | struct blk_mq_hw_ctx *hctx; | |
1017 | ||
1018 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1019 | ||
5815839b | 1020 | spin_lock(&hctx->dispatch_wait_lock); |
eb619fdb | 1021 | list_del_init(&wait->entry); |
5815839b ML |
1022 | spin_unlock(&hctx->dispatch_wait_lock); |
1023 | ||
da55f2cc OS |
1024 | blk_mq_run_hw_queue(hctx, true); |
1025 | return 1; | |
1026 | } | |
1027 | ||
f906a6a0 JA |
1028 | /* |
1029 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1030 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1031 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1032 | * marking us as waiting. |
1033 | */ | |
2278d69f | 1034 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1035 | struct request *rq) |
da55f2cc | 1036 | { |
5815839b | 1037 | struct wait_queue_head *wq; |
f906a6a0 JA |
1038 | wait_queue_entry_t *wait; |
1039 | bool ret; | |
da55f2cc | 1040 | |
2278d69f ML |
1041 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
1042 | if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
1043 | set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state); | |
f906a6a0 | 1044 | |
c27d53fb BVA |
1045 | /* |
1046 | * It's possible that a tag was freed in the window between the | |
1047 | * allocation failure and adding the hardware queue to the wait | |
1048 | * queue. | |
1049 | * | |
1050 | * Don't clear RESTART here, someone else could have set it. | |
1051 | * At most this will cost an extra queue run. | |
1052 | */ | |
8ab6bb9e | 1053 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1054 | } |
eb619fdb | 1055 | |
2278d69f | 1056 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1057 | if (!list_empty_careful(&wait->entry)) |
1058 | return false; | |
1059 | ||
5815839b ML |
1060 | wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait; |
1061 | ||
1062 | spin_lock_irq(&wq->lock); | |
1063 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1064 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1065 | spin_unlock(&hctx->dispatch_wait_lock); |
1066 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1067 | return false; |
eb619fdb JA |
1068 | } |
1069 | ||
5815839b ML |
1070 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1071 | __add_wait_queue(wq, wait); | |
c27d53fb | 1072 | |
da55f2cc | 1073 | /* |
eb619fdb JA |
1074 | * It's possible that a tag was freed in the window between the |
1075 | * allocation failure and adding the hardware queue to the wait | |
1076 | * queue. | |
da55f2cc | 1077 | */ |
8ab6bb9e | 1078 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1079 | if (!ret) { |
5815839b ML |
1080 | spin_unlock(&hctx->dispatch_wait_lock); |
1081 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1082 | return false; |
eb619fdb | 1083 | } |
c27d53fb BVA |
1084 | |
1085 | /* | |
1086 | * We got a tag, remove ourselves from the wait queue to ensure | |
1087 | * someone else gets the wakeup. | |
1088 | */ | |
c27d53fb | 1089 | list_del_init(&wait->entry); |
5815839b ML |
1090 | spin_unlock(&hctx->dispatch_wait_lock); |
1091 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1092 | |
1093 | return true; | |
da55f2cc OS |
1094 | } |
1095 | ||
6e768717 ML |
1096 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1097 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1098 | /* | |
1099 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1100 | * - EWMA is one simple way to compute running average value | |
1101 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1102 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1103 | * factor doesn't matter because EWMA decreases exponentially | |
1104 | */ | |
1105 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1106 | { | |
1107 | unsigned int ewma; | |
1108 | ||
1109 | if (hctx->queue->elevator) | |
1110 | return; | |
1111 | ||
1112 | ewma = hctx->dispatch_busy; | |
1113 | ||
1114 | if (!ewma && !busy) | |
1115 | return; | |
1116 | ||
1117 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1118 | if (busy) | |
1119 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1120 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1121 | ||
1122 | hctx->dispatch_busy = ewma; | |
1123 | } | |
1124 | ||
86ff7c2a ML |
1125 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1126 | ||
1f57f8d4 JA |
1127 | /* |
1128 | * Returns true if we did some work AND can potentially do more. | |
1129 | */ | |
de148297 | 1130 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1131 | bool got_budget) |
320ae51f | 1132 | { |
81380ca1 | 1133 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1134 | struct request *rq, *nxt; |
eb619fdb | 1135 | bool no_tag = false; |
fc17b653 | 1136 | int errors, queued; |
86ff7c2a | 1137 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1138 | |
81380ca1 OS |
1139 | if (list_empty(list)) |
1140 | return false; | |
1141 | ||
de148297 ML |
1142 | WARN_ON(!list_is_singular(list) && got_budget); |
1143 | ||
320ae51f JA |
1144 | /* |
1145 | * Now process all the entries, sending them to the driver. | |
1146 | */ | |
93efe981 | 1147 | errors = queued = 0; |
81380ca1 | 1148 | do { |
74c45052 | 1149 | struct blk_mq_queue_data bd; |
320ae51f | 1150 | |
f04c3df3 | 1151 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b ML |
1152 | |
1153 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
1154 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) | |
1155 | break; | |
1156 | ||
8ab6bb9e | 1157 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1158 | /* |
da55f2cc | 1159 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1160 | * rerun the hardware queue when a tag is freed. The |
1161 | * waitqueue takes care of that. If the queue is run | |
1162 | * before we add this entry back on the dispatch list, | |
1163 | * we'll re-run it below. | |
3c782d67 | 1164 | */ |
2278d69f | 1165 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1166 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1167 | /* |
1168 | * For non-shared tags, the RESTART check | |
1169 | * will suffice. | |
1170 | */ | |
1171 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1172 | no_tag = true; | |
de148297 ML |
1173 | break; |
1174 | } | |
1175 | } | |
1176 | ||
320ae51f | 1177 | list_del_init(&rq->queuelist); |
320ae51f | 1178 | |
74c45052 | 1179 | bd.rq = rq; |
113285b4 JA |
1180 | |
1181 | /* | |
1182 | * Flag last if we have no more requests, or if we have more | |
1183 | * but can't assign a driver tag to it. | |
1184 | */ | |
1185 | if (list_empty(list)) | |
1186 | bd.last = true; | |
1187 | else { | |
113285b4 | 1188 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1189 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1190 | } |
74c45052 JA |
1191 | |
1192 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1193 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1194 | /* |
1195 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1196 | * driver tag for the next request already, free it |
1197 | * again. | |
6d6f167c JW |
1198 | */ |
1199 | if (!list_empty(list)) { | |
1200 | nxt = list_first_entry(list, struct request, queuelist); | |
1201 | blk_mq_put_driver_tag(nxt); | |
1202 | } | |
f04c3df3 | 1203 | list_add(&rq->queuelist, list); |
ed0791b2 | 1204 | __blk_mq_requeue_request(rq); |
320ae51f | 1205 | break; |
fc17b653 CH |
1206 | } |
1207 | ||
1208 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1209 | errors++; |
2a842aca | 1210 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1211 | continue; |
320ae51f JA |
1212 | } |
1213 | ||
fc17b653 | 1214 | queued++; |
81380ca1 | 1215 | } while (!list_empty(list)); |
320ae51f | 1216 | |
703fd1c0 | 1217 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1218 | |
1219 | /* | |
1220 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1221 | * that is where we will continue on next queue run. | |
1222 | */ | |
f04c3df3 | 1223 | if (!list_empty(list)) { |
86ff7c2a ML |
1224 | bool needs_restart; |
1225 | ||
320ae51f | 1226 | spin_lock(&hctx->lock); |
c13660a0 | 1227 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1228 | spin_unlock(&hctx->lock); |
f04c3df3 | 1229 | |
9ba52e58 | 1230 | /* |
710c785f BVA |
1231 | * If SCHED_RESTART was set by the caller of this function and |
1232 | * it is no longer set that means that it was cleared by another | |
1233 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1234 | * |
eb619fdb JA |
1235 | * If 'no_tag' is set, that means that we failed getting |
1236 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1237 | * waitqueue is no longer active, ensure that we run the queue | |
1238 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1239 | * |
710c785f BVA |
1240 | * If no I/O scheduler has been configured it is possible that |
1241 | * the hardware queue got stopped and restarted before requests | |
1242 | * were pushed back onto the dispatch list. Rerun the queue to | |
1243 | * avoid starvation. Notes: | |
1244 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1245 | * been stopped before rerunning a queue. | |
1246 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1247 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1248 | * and dm-rq. |
86ff7c2a ML |
1249 | * |
1250 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1251 | * bit is set, run queue after a delay to avoid IO stalls | |
1252 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1253 | */ |
86ff7c2a ML |
1254 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1255 | if (!needs_restart || | |
eb619fdb | 1256 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1257 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1258 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1259 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1260 | |
6e768717 | 1261 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1262 | return false; |
6e768717 ML |
1263 | } else |
1264 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1265 | |
1f57f8d4 JA |
1266 | /* |
1267 | * If the host/device is unable to accept more work, inform the | |
1268 | * caller of that. | |
1269 | */ | |
1270 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1271 | return false; | |
1272 | ||
93efe981 | 1273 | return (queued + errors) != 0; |
f04c3df3 JA |
1274 | } |
1275 | ||
6a83e74d BVA |
1276 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1277 | { | |
1278 | int srcu_idx; | |
1279 | ||
b7a71e66 JA |
1280 | /* |
1281 | * We should be running this queue from one of the CPUs that | |
1282 | * are mapped to it. | |
7df938fb ML |
1283 | * |
1284 | * There are at least two related races now between setting | |
1285 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1286 | * __blk_mq_run_hw_queue(): | |
1287 | * | |
1288 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1289 | * but later it becomes online, then this warning is harmless | |
1290 | * at all | |
1291 | * | |
1292 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1293 | * but later it becomes offline, then the warning can't be | |
1294 | * triggered, and we depend on blk-mq timeout handler to | |
1295 | * handle dispatched requests to this hctx | |
b7a71e66 | 1296 | */ |
7df938fb ML |
1297 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1298 | cpu_online(hctx->next_cpu)) { | |
1299 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1300 | raw_smp_processor_id(), | |
1301 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1302 | dump_stack(); | |
1303 | } | |
6a83e74d | 1304 | |
b7a71e66 JA |
1305 | /* |
1306 | * We can't run the queue inline with ints disabled. Ensure that | |
1307 | * we catch bad users of this early. | |
1308 | */ | |
1309 | WARN_ON_ONCE(in_interrupt()); | |
1310 | ||
04ced159 | 1311 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1312 | |
04ced159 JA |
1313 | hctx_lock(hctx, &srcu_idx); |
1314 | blk_mq_sched_dispatch_requests(hctx); | |
1315 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1316 | } |
1317 | ||
f82ddf19 ML |
1318 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1319 | { | |
1320 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1321 | ||
1322 | if (cpu >= nr_cpu_ids) | |
1323 | cpu = cpumask_first(hctx->cpumask); | |
1324 | return cpu; | |
1325 | } | |
1326 | ||
506e931f JA |
1327 | /* |
1328 | * It'd be great if the workqueue API had a way to pass | |
1329 | * in a mask and had some smarts for more clever placement. | |
1330 | * For now we just round-robin here, switching for every | |
1331 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1332 | */ | |
1333 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1334 | { | |
7bed4595 | 1335 | bool tried = false; |
476f8c98 | 1336 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1337 | |
b657d7e6 CH |
1338 | if (hctx->queue->nr_hw_queues == 1) |
1339 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1340 | |
1341 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1342 | select_cpu: |
476f8c98 | 1343 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1344 | cpu_online_mask); |
506e931f | 1345 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1346 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1347 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1348 | } | |
1349 | ||
7bed4595 ML |
1350 | /* |
1351 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1352 | * and it should only happen in the path of handling CPU DEAD. | |
1353 | */ | |
476f8c98 | 1354 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1355 | if (!tried) { |
1356 | tried = true; | |
1357 | goto select_cpu; | |
1358 | } | |
1359 | ||
1360 | /* | |
1361 | * Make sure to re-select CPU next time once after CPUs | |
1362 | * in hctx->cpumask become online again. | |
1363 | */ | |
476f8c98 | 1364 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1365 | hctx->next_cpu_batch = 1; |
1366 | return WORK_CPU_UNBOUND; | |
1367 | } | |
476f8c98 ML |
1368 | |
1369 | hctx->next_cpu = next_cpu; | |
1370 | return next_cpu; | |
506e931f JA |
1371 | } |
1372 | ||
7587a5ae BVA |
1373 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1374 | unsigned long msecs) | |
320ae51f | 1375 | { |
5435c023 | 1376 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1377 | return; |
1378 | ||
1b792f2f | 1379 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1380 | int cpu = get_cpu(); |
1381 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1382 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1383 | put_cpu(); |
398205b8 PB |
1384 | return; |
1385 | } | |
e4043dcf | 1386 | |
2a90d4aa | 1387 | put_cpu(); |
e4043dcf | 1388 | } |
398205b8 | 1389 | |
ae943d20 BVA |
1390 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1391 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1392 | } |
1393 | ||
1394 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1395 | { | |
1396 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1397 | } | |
1398 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1399 | ||
79f720a7 | 1400 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1401 | { |
24f5a90f ML |
1402 | int srcu_idx; |
1403 | bool need_run; | |
1404 | ||
1405 | /* | |
1406 | * When queue is quiesced, we may be switching io scheduler, or | |
1407 | * updating nr_hw_queues, or other things, and we can't run queue | |
1408 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1409 | * | |
1410 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1411 | * quiesced. | |
1412 | */ | |
04ced159 JA |
1413 | hctx_lock(hctx, &srcu_idx); |
1414 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1415 | blk_mq_hctx_has_pending(hctx); | |
1416 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1417 | |
1418 | if (need_run) { | |
79f720a7 JA |
1419 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1420 | return true; | |
1421 | } | |
1422 | ||
1423 | return false; | |
320ae51f | 1424 | } |
5b727272 | 1425 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1426 | |
b94ec296 | 1427 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1428 | { |
1429 | struct blk_mq_hw_ctx *hctx; | |
1430 | int i; | |
1431 | ||
1432 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1433 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1434 | continue; |
1435 | ||
b94ec296 | 1436 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1437 | } |
1438 | } | |
b94ec296 | 1439 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1440 | |
fd001443 BVA |
1441 | /** |
1442 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1443 | * @q: request queue. | |
1444 | * | |
1445 | * The caller is responsible for serializing this function against | |
1446 | * blk_mq_{start,stop}_hw_queue(). | |
1447 | */ | |
1448 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1449 | { | |
1450 | struct blk_mq_hw_ctx *hctx; | |
1451 | int i; | |
1452 | ||
1453 | queue_for_each_hw_ctx(q, hctx, i) | |
1454 | if (blk_mq_hctx_stopped(hctx)) | |
1455 | return true; | |
1456 | ||
1457 | return false; | |
1458 | } | |
1459 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1460 | ||
39a70c76 ML |
1461 | /* |
1462 | * This function is often used for pausing .queue_rq() by driver when | |
1463 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1464 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1465 | * |
1466 | * We do not guarantee that dispatch can be drained or blocked | |
1467 | * after blk_mq_stop_hw_queue() returns. Please use | |
1468 | * blk_mq_quiesce_queue() for that requirement. | |
1469 | */ | |
2719aa21 JA |
1470 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1471 | { | |
641a9ed6 | 1472 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1473 | |
641a9ed6 | 1474 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1475 | } |
641a9ed6 | 1476 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1477 | |
39a70c76 ML |
1478 | /* |
1479 | * This function is often used for pausing .queue_rq() by driver when | |
1480 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1481 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1482 | * |
1483 | * We do not guarantee that dispatch can be drained or blocked | |
1484 | * after blk_mq_stop_hw_queues() returns. Please use | |
1485 | * blk_mq_quiesce_queue() for that requirement. | |
1486 | */ | |
2719aa21 JA |
1487 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1488 | { | |
641a9ed6 ML |
1489 | struct blk_mq_hw_ctx *hctx; |
1490 | int i; | |
1491 | ||
1492 | queue_for_each_hw_ctx(q, hctx, i) | |
1493 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1494 | } |
1495 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1496 | ||
320ae51f JA |
1497 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1498 | { | |
1499 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1500 | |
0ffbce80 | 1501 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1502 | } |
1503 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1504 | ||
2f268556 CH |
1505 | void blk_mq_start_hw_queues(struct request_queue *q) |
1506 | { | |
1507 | struct blk_mq_hw_ctx *hctx; | |
1508 | int i; | |
1509 | ||
1510 | queue_for_each_hw_ctx(q, hctx, i) | |
1511 | blk_mq_start_hw_queue(hctx); | |
1512 | } | |
1513 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1514 | ||
ae911c5e JA |
1515 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1516 | { | |
1517 | if (!blk_mq_hctx_stopped(hctx)) | |
1518 | return; | |
1519 | ||
1520 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1521 | blk_mq_run_hw_queue(hctx, async); | |
1522 | } | |
1523 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1524 | ||
1b4a3258 | 1525 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1526 | { |
1527 | struct blk_mq_hw_ctx *hctx; | |
1528 | int i; | |
1529 | ||
ae911c5e JA |
1530 | queue_for_each_hw_ctx(q, hctx, i) |
1531 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1532 | } |
1533 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1534 | ||
70f4db63 | 1535 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1536 | { |
1537 | struct blk_mq_hw_ctx *hctx; | |
1538 | ||
9f993737 | 1539 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1540 | |
21c6e939 | 1541 | /* |
15fe8a90 | 1542 | * If we are stopped, don't run the queue. |
21c6e939 | 1543 | */ |
15fe8a90 | 1544 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1545 | return; |
7587a5ae BVA |
1546 | |
1547 | __blk_mq_run_hw_queue(hctx); | |
1548 | } | |
1549 | ||
cfd0c552 | 1550 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1551 | struct request *rq, |
1552 | bool at_head) | |
320ae51f | 1553 | { |
e57690fe JA |
1554 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1555 | ||
7b607814 BVA |
1556 | lockdep_assert_held(&ctx->lock); |
1557 | ||
01b983c9 JA |
1558 | trace_block_rq_insert(hctx->queue, rq); |
1559 | ||
72a0a36e CH |
1560 | if (at_head) |
1561 | list_add(&rq->queuelist, &ctx->rq_list); | |
1562 | else | |
1563 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1564 | } |
4bb659b1 | 1565 | |
2c3ad667 JA |
1566 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1567 | bool at_head) | |
cfd0c552 ML |
1568 | { |
1569 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1570 | ||
7b607814 BVA |
1571 | lockdep_assert_held(&ctx->lock); |
1572 | ||
e57690fe | 1573 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1574 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1575 | } |
1576 | ||
157f377b JA |
1577 | /* |
1578 | * Should only be used carefully, when the caller knows we want to | |
1579 | * bypass a potential IO scheduler on the target device. | |
1580 | */ | |
b0850297 | 1581 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b JA |
1582 | { |
1583 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1584 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1585 | ||
1586 | spin_lock(&hctx->lock); | |
1587 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1588 | spin_unlock(&hctx->lock); | |
1589 | ||
b0850297 ML |
1590 | if (run_queue) |
1591 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1592 | } |
1593 | ||
bd166ef1 JA |
1594 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1595 | struct list_head *list) | |
320ae51f JA |
1596 | |
1597 | { | |
3f0cedc7 ML |
1598 | struct request *rq; |
1599 | ||
320ae51f JA |
1600 | /* |
1601 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1602 | * offline now | |
1603 | */ | |
3f0cedc7 | 1604 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1605 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1606 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1607 | } |
3f0cedc7 ML |
1608 | |
1609 | spin_lock(&ctx->lock); | |
1610 | list_splice_tail_init(list, &ctx->rq_list); | |
cfd0c552 | 1611 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1612 | spin_unlock(&ctx->lock); |
320ae51f JA |
1613 | } |
1614 | ||
1615 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1616 | { | |
1617 | struct request *rqa = container_of(a, struct request, queuelist); | |
1618 | struct request *rqb = container_of(b, struct request, queuelist); | |
1619 | ||
1620 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1621 | (rqa->mq_ctx == rqb->mq_ctx && | |
1622 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1623 | } | |
1624 | ||
1625 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1626 | { | |
1627 | struct blk_mq_ctx *this_ctx; | |
1628 | struct request_queue *this_q; | |
1629 | struct request *rq; | |
1630 | LIST_HEAD(list); | |
1631 | LIST_HEAD(ctx_list); | |
1632 | unsigned int depth; | |
1633 | ||
1634 | list_splice_init(&plug->mq_list, &list); | |
1635 | ||
1636 | list_sort(NULL, &list, plug_ctx_cmp); | |
1637 | ||
1638 | this_q = NULL; | |
1639 | this_ctx = NULL; | |
1640 | depth = 0; | |
1641 | ||
1642 | while (!list_empty(&list)) { | |
1643 | rq = list_entry_rq(list.next); | |
1644 | list_del_init(&rq->queuelist); | |
1645 | BUG_ON(!rq->q); | |
1646 | if (rq->mq_ctx != this_ctx) { | |
1647 | if (this_ctx) { | |
587562d0 | 1648 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1649 | blk_mq_sched_insert_requests(this_q, this_ctx, |
1650 | &ctx_list, | |
1651 | from_schedule); | |
320ae51f JA |
1652 | } |
1653 | ||
1654 | this_ctx = rq->mq_ctx; | |
1655 | this_q = rq->q; | |
1656 | depth = 0; | |
1657 | } | |
1658 | ||
1659 | depth++; | |
1660 | list_add_tail(&rq->queuelist, &ctx_list); | |
1661 | } | |
1662 | ||
1663 | /* | |
1664 | * If 'this_ctx' is set, we know we have entries to complete | |
1665 | * on 'ctx_list'. Do those. | |
1666 | */ | |
1667 | if (this_ctx) { | |
587562d0 | 1668 | trace_block_unplug(this_q, depth, !from_schedule); |
bd166ef1 JA |
1669 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, |
1670 | from_schedule); | |
320ae51f JA |
1671 | } |
1672 | } | |
1673 | ||
1674 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1675 | { | |
da8d7f07 | 1676 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1677 | |
85acb3ba SL |
1678 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1679 | ||
6e85eaf3 | 1680 | blk_account_io_start(rq, true); |
320ae51f JA |
1681 | } |
1682 | ||
fd2d3326 JA |
1683 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1684 | { | |
bd166ef1 JA |
1685 | if (rq->tag != -1) |
1686 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1687 | ||
1688 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1689 | } |
1690 | ||
0f95549c MS |
1691 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1692 | struct request *rq, | |
1693 | blk_qc_t *cookie) | |
f984df1f | 1694 | { |
f984df1f | 1695 | struct request_queue *q = rq->q; |
f984df1f SL |
1696 | struct blk_mq_queue_data bd = { |
1697 | .rq = rq, | |
d945a365 | 1698 | .last = true, |
f984df1f | 1699 | }; |
bd166ef1 | 1700 | blk_qc_t new_cookie; |
f06345ad | 1701 | blk_status_t ret; |
0f95549c MS |
1702 | |
1703 | new_cookie = request_to_qc_t(hctx, rq); | |
1704 | ||
1705 | /* | |
1706 | * For OK queue, we are done. For error, caller may kill it. | |
1707 | * Any other error (busy), just add it to our list as we | |
1708 | * previously would have done. | |
1709 | */ | |
1710 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1711 | switch (ret) { | |
1712 | case BLK_STS_OK: | |
6ce3dd6e | 1713 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1714 | *cookie = new_cookie; |
1715 | break; | |
1716 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1717 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1718 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1719 | __blk_mq_requeue_request(rq); |
1720 | break; | |
1721 | default: | |
6ce3dd6e | 1722 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1723 | *cookie = BLK_QC_T_NONE; |
1724 | break; | |
1725 | } | |
1726 | ||
1727 | return ret; | |
1728 | } | |
1729 | ||
0f95549c MS |
1730 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1731 | struct request *rq, | |
396eaf21 ML |
1732 | blk_qc_t *cookie, |
1733 | bool bypass_insert) | |
0f95549c MS |
1734 | { |
1735 | struct request_queue *q = rq->q; | |
d964f04a ML |
1736 | bool run_queue = true; |
1737 | ||
23d4ee19 ML |
1738 | /* |
1739 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
1740 | * | |
1741 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
c77ff7fd | 1742 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, |
23d4ee19 ML |
1743 | * and avoid driver to try to dispatch again. |
1744 | */ | |
f4560ffe | 1745 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1746 | run_queue = false; |
23d4ee19 | 1747 | bypass_insert = false; |
d964f04a ML |
1748 | goto insert; |
1749 | } | |
f984df1f | 1750 | |
396eaf21 | 1751 | if (q->elevator && !bypass_insert) |
2253efc8 BVA |
1752 | goto insert; |
1753 | ||
0bca799b | 1754 | if (!blk_mq_get_dispatch_budget(hctx)) |
bd166ef1 JA |
1755 | goto insert; |
1756 | ||
8ab6bb9e | 1757 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1758 | blk_mq_put_dispatch_budget(hctx); |
de148297 | 1759 | goto insert; |
88022d72 | 1760 | } |
de148297 | 1761 | |
0f95549c | 1762 | return __blk_mq_issue_directly(hctx, rq, cookie); |
2253efc8 | 1763 | insert: |
396eaf21 ML |
1764 | if (bypass_insert) |
1765 | return BLK_STS_RESOURCE; | |
0f95549c | 1766 | |
23d4ee19 | 1767 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
0f95549c | 1768 | return BLK_STS_OK; |
f984df1f SL |
1769 | } |
1770 | ||
5eb6126e CH |
1771 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1772 | struct request *rq, blk_qc_t *cookie) | |
1773 | { | |
0f95549c | 1774 | blk_status_t ret; |
04ced159 | 1775 | int srcu_idx; |
bf4907c0 | 1776 | |
04ced159 | 1777 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1778 | |
04ced159 | 1779 | hctx_lock(hctx, &srcu_idx); |
0f95549c | 1780 | |
396eaf21 | 1781 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
86ff7c2a | 1782 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
23d4ee19 | 1783 | blk_mq_sched_insert_request(rq, false, true, false); |
0f95549c MS |
1784 | else if (ret != BLK_STS_OK) |
1785 | blk_mq_end_request(rq, ret); | |
1786 | ||
04ced159 | 1787 | hctx_unlock(hctx, srcu_idx); |
5eb6126e CH |
1788 | } |
1789 | ||
c77ff7fd | 1790 | blk_status_t blk_mq_request_issue_directly(struct request *rq) |
396eaf21 ML |
1791 | { |
1792 | blk_status_t ret; | |
1793 | int srcu_idx; | |
1794 | blk_qc_t unused_cookie; | |
1795 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1796 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1797 | ||
1798 | hctx_lock(hctx, &srcu_idx); | |
1799 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true); | |
1800 | hctx_unlock(hctx, srcu_idx); | |
1801 | ||
1802 | return ret; | |
5eb6126e CH |
1803 | } |
1804 | ||
6ce3dd6e ML |
1805 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1806 | struct list_head *list) | |
1807 | { | |
1808 | while (!list_empty(list)) { | |
1809 | blk_status_t ret; | |
1810 | struct request *rq = list_first_entry(list, struct request, | |
1811 | queuelist); | |
1812 | ||
1813 | list_del_init(&rq->queuelist); | |
1814 | ret = blk_mq_request_issue_directly(rq); | |
1815 | if (ret != BLK_STS_OK) { | |
8824f622 ML |
1816 | if (ret == BLK_STS_RESOURCE || |
1817 | ret == BLK_STS_DEV_RESOURCE) { | |
1818 | list_add(&rq->queuelist, list); | |
1819 | break; | |
1820 | } | |
1821 | blk_mq_end_request(rq, ret); | |
6ce3dd6e ML |
1822 | } |
1823 | } | |
1824 | } | |
1825 | ||
dece1635 | 1826 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1827 | { |
ef295ecf | 1828 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1829 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1830 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1831 | struct request *rq; |
5eb6126e | 1832 | unsigned int request_count = 0; |
f984df1f | 1833 | struct blk_plug *plug; |
5b3f341f | 1834 | struct request *same_queue_rq = NULL; |
7b371636 | 1835 | blk_qc_t cookie; |
07068d5b JA |
1836 | |
1837 | blk_queue_bounce(q, &bio); | |
1838 | ||
af67c31f | 1839 | blk_queue_split(q, &bio); |
f36ea50c | 1840 | |
e23947bd | 1841 | if (!bio_integrity_prep(bio)) |
dece1635 | 1842 | return BLK_QC_T_NONE; |
07068d5b | 1843 | |
87c279e6 OS |
1844 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1845 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1846 | return BLK_QC_T_NONE; | |
f984df1f | 1847 | |
bd166ef1 JA |
1848 | if (blk_mq_sched_bio_merge(q, bio)) |
1849 | return BLK_QC_T_NONE; | |
1850 | ||
c1c80384 | 1851 | rq_qos_throttle(q, bio, NULL); |
87760e5e | 1852 | |
bd166ef1 JA |
1853 | trace_block_getrq(q, bio, bio->bi_opf); |
1854 | ||
d2c0d383 | 1855 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e | 1856 | if (unlikely(!rq)) { |
c1c80384 | 1857 | rq_qos_cleanup(q, bio); |
03a07c92 GR |
1858 | if (bio->bi_opf & REQ_NOWAIT) |
1859 | bio_wouldblock_error(bio); | |
dece1635 | 1860 | return BLK_QC_T_NONE; |
87760e5e JA |
1861 | } |
1862 | ||
c1c80384 | 1863 | rq_qos_track(q, rq, bio); |
07068d5b | 1864 | |
fd2d3326 | 1865 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1866 | |
f984df1f | 1867 | plug = current->plug; |
07068d5b | 1868 | if (unlikely(is_flush_fua)) { |
f984df1f | 1869 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1870 | blk_mq_bio_to_request(rq, bio); |
923218f6 ML |
1871 | |
1872 | /* bypass scheduler for flush rq */ | |
1873 | blk_insert_flush(rq); | |
1874 | blk_mq_run_hw_queue(data.hctx, true); | |
a4d907b6 | 1875 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1876 | struct request *last = NULL; |
1877 | ||
b00c53e8 | 1878 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1879 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1880 | |
1881 | /* | |
1882 | * @request_count may become stale because of schedule | |
1883 | * out, so check the list again. | |
1884 | */ | |
1885 | if (list_empty(&plug->mq_list)) | |
1886 | request_count = 0; | |
254d259d CH |
1887 | else if (blk_queue_nomerges(q)) |
1888 | request_count = blk_plug_queued_count(q); | |
1889 | ||
676d0607 | 1890 | if (!request_count) |
e6c4438b | 1891 | trace_block_plug(q); |
600271d9 SL |
1892 | else |
1893 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1894 | |
600271d9 SL |
1895 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1896 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1897 | blk_flush_plug_list(plug, false); |
1898 | trace_block_plug(q); | |
320ae51f | 1899 | } |
b094f89c | 1900 | |
e6c4438b | 1901 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1902 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1903 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1904 | |
07068d5b | 1905 | /* |
6a83e74d | 1906 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1907 | * Otherwise the existing request in the plug list will be |
1908 | * issued. So the plug list will have one request at most | |
2299722c CH |
1909 | * The plug list might get flushed before this. If that happens, |
1910 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1911 | */ |
2299722c CH |
1912 | if (list_empty(&plug->mq_list)) |
1913 | same_queue_rq = NULL; | |
1914 | if (same_queue_rq) | |
1915 | list_del_init(&same_queue_rq->queuelist); | |
1916 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1917 | ||
bf4907c0 JA |
1918 | blk_mq_put_ctx(data.ctx); |
1919 | ||
dad7a3be ML |
1920 | if (same_queue_rq) { |
1921 | data.hctx = blk_mq_map_queue(q, | |
1922 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1923 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1924 | &cookie); | |
dad7a3be | 1925 | } |
6ce3dd6e ML |
1926 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
1927 | !data.hctx->dispatch_busy)) { | |
bf4907c0 | 1928 | blk_mq_put_ctx(data.ctx); |
2299722c | 1929 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1930 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 1931 | } else { |
b00c53e8 | 1932 | blk_mq_put_ctx(data.ctx); |
ab42f35d | 1933 | blk_mq_bio_to_request(rq, bio); |
8fa9f556 | 1934 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 1935 | } |
320ae51f | 1936 | |
7b371636 | 1937 | return cookie; |
320ae51f JA |
1938 | } |
1939 | ||
cc71a6f4 JA |
1940 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1941 | unsigned int hctx_idx) | |
95363efd | 1942 | { |
e9b267d9 | 1943 | struct page *page; |
320ae51f | 1944 | |
24d2f903 | 1945 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1946 | int i; |
320ae51f | 1947 | |
24d2f903 | 1948 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1949 | struct request *rq = tags->static_rqs[i]; |
1950 | ||
1951 | if (!rq) | |
e9b267d9 | 1952 | continue; |
d6296d39 | 1953 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1954 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1955 | } |
320ae51f | 1956 | } |
320ae51f | 1957 | |
24d2f903 CH |
1958 | while (!list_empty(&tags->page_list)) { |
1959 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1960 | list_del_init(&page->lru); |
f75782e4 CM |
1961 | /* |
1962 | * Remove kmemleak object previously allocated in | |
1963 | * blk_mq_init_rq_map(). | |
1964 | */ | |
1965 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1966 | __free_pages(page, page->private); |
1967 | } | |
cc71a6f4 | 1968 | } |
320ae51f | 1969 | |
cc71a6f4 JA |
1970 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1971 | { | |
24d2f903 | 1972 | kfree(tags->rqs); |
cc71a6f4 | 1973 | tags->rqs = NULL; |
2af8cbe3 JA |
1974 | kfree(tags->static_rqs); |
1975 | tags->static_rqs = NULL; | |
320ae51f | 1976 | |
24d2f903 | 1977 | blk_mq_free_tags(tags); |
320ae51f JA |
1978 | } |
1979 | ||
cc71a6f4 JA |
1980 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1981 | unsigned int hctx_idx, | |
1982 | unsigned int nr_tags, | |
1983 | unsigned int reserved_tags) | |
320ae51f | 1984 | { |
24d2f903 | 1985 | struct blk_mq_tags *tags; |
59f082e4 | 1986 | int node; |
320ae51f | 1987 | |
59f082e4 SL |
1988 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1989 | if (node == NUMA_NO_NODE) | |
1990 | node = set->numa_node; | |
1991 | ||
1992 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1993 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1994 | if (!tags) |
1995 | return NULL; | |
320ae51f | 1996 | |
590b5b7d | 1997 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 1998 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1999 | node); |
24d2f903 CH |
2000 | if (!tags->rqs) { |
2001 | blk_mq_free_tags(tags); | |
2002 | return NULL; | |
2003 | } | |
320ae51f | 2004 | |
590b5b7d KC |
2005 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2006 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2007 | node); | |
2af8cbe3 JA |
2008 | if (!tags->static_rqs) { |
2009 | kfree(tags->rqs); | |
2010 | blk_mq_free_tags(tags); | |
2011 | return NULL; | |
2012 | } | |
2013 | ||
cc71a6f4 JA |
2014 | return tags; |
2015 | } | |
2016 | ||
2017 | static size_t order_to_size(unsigned int order) | |
2018 | { | |
2019 | return (size_t)PAGE_SIZE << order; | |
2020 | } | |
2021 | ||
1d9bd516 TH |
2022 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2023 | unsigned int hctx_idx, int node) | |
2024 | { | |
2025 | int ret; | |
2026 | ||
2027 | if (set->ops->init_request) { | |
2028 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2029 | if (ret) | |
2030 | return ret; | |
2031 | } | |
2032 | ||
12f5b931 | 2033 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2034 | return 0; |
2035 | } | |
2036 | ||
cc71a6f4 JA |
2037 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2038 | unsigned int hctx_idx, unsigned int depth) | |
2039 | { | |
2040 | unsigned int i, j, entries_per_page, max_order = 4; | |
2041 | size_t rq_size, left; | |
59f082e4 SL |
2042 | int node; |
2043 | ||
2044 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
2045 | if (node == NUMA_NO_NODE) | |
2046 | node = set->numa_node; | |
cc71a6f4 JA |
2047 | |
2048 | INIT_LIST_HEAD(&tags->page_list); | |
2049 | ||
320ae51f JA |
2050 | /* |
2051 | * rq_size is the size of the request plus driver payload, rounded | |
2052 | * to the cacheline size | |
2053 | */ | |
24d2f903 | 2054 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2055 | cache_line_size()); |
cc71a6f4 | 2056 | left = rq_size * depth; |
320ae51f | 2057 | |
cc71a6f4 | 2058 | for (i = 0; i < depth; ) { |
320ae51f JA |
2059 | int this_order = max_order; |
2060 | struct page *page; | |
2061 | int to_do; | |
2062 | void *p; | |
2063 | ||
b3a834b1 | 2064 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2065 | this_order--; |
2066 | ||
2067 | do { | |
59f082e4 | 2068 | page = alloc_pages_node(node, |
36e1f3d1 | 2069 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2070 | this_order); |
320ae51f JA |
2071 | if (page) |
2072 | break; | |
2073 | if (!this_order--) | |
2074 | break; | |
2075 | if (order_to_size(this_order) < rq_size) | |
2076 | break; | |
2077 | } while (1); | |
2078 | ||
2079 | if (!page) | |
24d2f903 | 2080 | goto fail; |
320ae51f JA |
2081 | |
2082 | page->private = this_order; | |
24d2f903 | 2083 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2084 | |
2085 | p = page_address(page); | |
f75782e4 CM |
2086 | /* |
2087 | * Allow kmemleak to scan these pages as they contain pointers | |
2088 | * to additional allocations like via ops->init_request(). | |
2089 | */ | |
36e1f3d1 | 2090 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2091 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2092 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2093 | left -= to_do * rq_size; |
2094 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2095 | struct request *rq = p; |
2096 | ||
2097 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2098 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2099 | tags->static_rqs[i] = NULL; | |
2100 | goto fail; | |
e9b267d9 CH |
2101 | } |
2102 | ||
320ae51f JA |
2103 | p += rq_size; |
2104 | i++; | |
2105 | } | |
2106 | } | |
cc71a6f4 | 2107 | return 0; |
320ae51f | 2108 | |
24d2f903 | 2109 | fail: |
cc71a6f4 JA |
2110 | blk_mq_free_rqs(set, tags, hctx_idx); |
2111 | return -ENOMEM; | |
320ae51f JA |
2112 | } |
2113 | ||
e57690fe JA |
2114 | /* |
2115 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2116 | * software queue to the hw queue dispatch list, and ensure that it | |
2117 | * gets run. | |
2118 | */ | |
9467f859 | 2119 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2120 | { |
9467f859 | 2121 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2122 | struct blk_mq_ctx *ctx; |
2123 | LIST_HEAD(tmp); | |
2124 | ||
9467f859 | 2125 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2126 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
2127 | |
2128 | spin_lock(&ctx->lock); | |
2129 | if (!list_empty(&ctx->rq_list)) { | |
2130 | list_splice_init(&ctx->rq_list, &tmp); | |
2131 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2132 | } | |
2133 | spin_unlock(&ctx->lock); | |
2134 | ||
2135 | if (list_empty(&tmp)) | |
9467f859 | 2136 | return 0; |
484b4061 | 2137 | |
e57690fe JA |
2138 | spin_lock(&hctx->lock); |
2139 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2140 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2141 | |
2142 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2143 | return 0; |
484b4061 JA |
2144 | } |
2145 | ||
9467f859 | 2146 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2147 | { |
9467f859 TG |
2148 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2149 | &hctx->cpuhp_dead); | |
484b4061 JA |
2150 | } |
2151 | ||
c3b4afca | 2152 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2153 | static void blk_mq_exit_hctx(struct request_queue *q, |
2154 | struct blk_mq_tag_set *set, | |
2155 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2156 | { | |
8ab0b7dc ML |
2157 | if (blk_mq_hw_queue_mapped(hctx)) |
2158 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2159 | |
f70ced09 | 2160 | if (set->ops->exit_request) |
d6296d39 | 2161 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2162 | |
08e98fc6 ML |
2163 | if (set->ops->exit_hctx) |
2164 | set->ops->exit_hctx(hctx, hctx_idx); | |
2165 | ||
6a83e74d | 2166 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2167 | cleanup_srcu_struct(hctx->srcu); |
6a83e74d | 2168 | |
9467f859 | 2169 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2170 | blk_free_flush_queue(hctx->fq); |
88459642 | 2171 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2172 | } |
2173 | ||
624dbe47 ML |
2174 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2175 | struct blk_mq_tag_set *set, int nr_queue) | |
2176 | { | |
2177 | struct blk_mq_hw_ctx *hctx; | |
2178 | unsigned int i; | |
2179 | ||
2180 | queue_for_each_hw_ctx(q, hctx, i) { | |
2181 | if (i == nr_queue) | |
2182 | break; | |
477e19de | 2183 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2184 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2185 | } |
624dbe47 ML |
2186 | } |
2187 | ||
08e98fc6 ML |
2188 | static int blk_mq_init_hctx(struct request_queue *q, |
2189 | struct blk_mq_tag_set *set, | |
2190 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2191 | { |
08e98fc6 ML |
2192 | int node; |
2193 | ||
2194 | node = hctx->numa_node; | |
2195 | if (node == NUMA_NO_NODE) | |
2196 | node = hctx->numa_node = set->numa_node; | |
2197 | ||
9f993737 | 2198 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2199 | spin_lock_init(&hctx->lock); |
2200 | INIT_LIST_HEAD(&hctx->dispatch); | |
2201 | hctx->queue = q; | |
2404e607 | 2202 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2203 | |
9467f859 | 2204 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2205 | |
2206 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2207 | |
2208 | /* | |
08e98fc6 ML |
2209 | * Allocate space for all possible cpus to avoid allocation at |
2210 | * runtime | |
320ae51f | 2211 | */ |
d904bfa7 | 2212 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
5b202853 | 2213 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node); |
08e98fc6 ML |
2214 | if (!hctx->ctxs) |
2215 | goto unregister_cpu_notifier; | |
320ae51f | 2216 | |
5b202853 JW |
2217 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
2218 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node)) | |
08e98fc6 | 2219 | goto free_ctxs; |
320ae51f | 2220 | |
08e98fc6 | 2221 | hctx->nr_ctx = 0; |
320ae51f | 2222 | |
5815839b | 2223 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2224 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2225 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2226 | ||
08e98fc6 ML |
2227 | if (set->ops->init_hctx && |
2228 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2229 | goto free_bitmap; | |
320ae51f | 2230 | |
5b202853 JW |
2231 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
2232 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
f70ced09 | 2233 | if (!hctx->fq) |
d48ece20 | 2234 | goto exit_hctx; |
320ae51f | 2235 | |
1d9bd516 | 2236 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node)) |
f70ced09 | 2237 | goto free_fq; |
320ae51f | 2238 | |
6a83e74d | 2239 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2240 | init_srcu_struct(hctx->srcu); |
6a83e74d | 2241 | |
08e98fc6 | 2242 | return 0; |
320ae51f | 2243 | |
f70ced09 ML |
2244 | free_fq: |
2245 | kfree(hctx->fq); | |
2246 | exit_hctx: | |
2247 | if (set->ops->exit_hctx) | |
2248 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2249 | free_bitmap: |
88459642 | 2250 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2251 | free_ctxs: |
2252 | kfree(hctx->ctxs); | |
2253 | unregister_cpu_notifier: | |
9467f859 | 2254 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2255 | return -1; |
2256 | } | |
320ae51f | 2257 | |
320ae51f JA |
2258 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2259 | unsigned int nr_hw_queues) | |
2260 | { | |
2261 | unsigned int i; | |
2262 | ||
2263 | for_each_possible_cpu(i) { | |
2264 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2265 | struct blk_mq_hw_ctx *hctx; | |
2266 | ||
320ae51f JA |
2267 | __ctx->cpu = i; |
2268 | spin_lock_init(&__ctx->lock); | |
2269 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2270 | __ctx->queue = q; | |
2271 | ||
320ae51f JA |
2272 | /* |
2273 | * Set local node, IFF we have more than one hw queue. If | |
2274 | * not, we remain on the home node of the device | |
2275 | */ | |
20e4d813 | 2276 | hctx = blk_mq_map_queue(q, i); |
320ae51f | 2277 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) |
bffed457 | 2278 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2279 | } |
2280 | } | |
2281 | ||
cc71a6f4 JA |
2282 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2283 | { | |
2284 | int ret = 0; | |
2285 | ||
2286 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2287 | set->queue_depth, set->reserved_tags); | |
2288 | if (!set->tags[hctx_idx]) | |
2289 | return false; | |
2290 | ||
2291 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2292 | set->queue_depth); | |
2293 | if (!ret) | |
2294 | return true; | |
2295 | ||
2296 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2297 | set->tags[hctx_idx] = NULL; | |
2298 | return false; | |
2299 | } | |
2300 | ||
2301 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2302 | unsigned int hctx_idx) | |
2303 | { | |
bd166ef1 JA |
2304 | if (set->tags[hctx_idx]) { |
2305 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2306 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2307 | set->tags[hctx_idx] = NULL; | |
2308 | } | |
cc71a6f4 JA |
2309 | } |
2310 | ||
4b855ad3 | 2311 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2312 | { |
4412efec | 2313 | unsigned int i, hctx_idx; |
320ae51f JA |
2314 | struct blk_mq_hw_ctx *hctx; |
2315 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2316 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2317 | |
60de074b AM |
2318 | /* |
2319 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2320 | */ | |
2321 | mutex_lock(&q->sysfs_lock); | |
2322 | ||
320ae51f | 2323 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2324 | cpumask_clear(hctx->cpumask); |
320ae51f | 2325 | hctx->nr_ctx = 0; |
d416c92c | 2326 | hctx->dispatch_from = NULL; |
320ae51f JA |
2327 | } |
2328 | ||
2329 | /* | |
4b855ad3 | 2330 | * Map software to hardware queues. |
4412efec ML |
2331 | * |
2332 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2333 | */ |
20e4d813 | 2334 | for_each_possible_cpu(i) { |
4412efec ML |
2335 | hctx_idx = q->mq_map[i]; |
2336 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2337 | if (!set->tags[hctx_idx] && | |
2338 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2339 | /* | |
2340 | * If tags initialization fail for some hctx, | |
2341 | * that hctx won't be brought online. In this | |
2342 | * case, remap the current ctx to hctx[0] which | |
2343 | * is guaranteed to always have tags allocated | |
2344 | */ | |
2345 | q->mq_map[i] = 0; | |
2346 | } | |
2347 | ||
897bb0c7 | 2348 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2349 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2350 | |
e4043dcf | 2351 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2352 | ctx->index_hw = hctx->nr_ctx; |
2353 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2354 | } | |
506e931f | 2355 | |
60de074b AM |
2356 | mutex_unlock(&q->sysfs_lock); |
2357 | ||
506e931f | 2358 | queue_for_each_hw_ctx(q, hctx, i) { |
4412efec ML |
2359 | /* |
2360 | * If no software queues are mapped to this hardware queue, | |
2361 | * disable it and free the request entries. | |
2362 | */ | |
2363 | if (!hctx->nr_ctx) { | |
2364 | /* Never unmap queue 0. We need it as a | |
2365 | * fallback in case of a new remap fails | |
2366 | * allocation | |
2367 | */ | |
2368 | if (i && set->tags[i]) | |
2369 | blk_mq_free_map_and_requests(set, i); | |
2370 | ||
2371 | hctx->tags = NULL; | |
2372 | continue; | |
2373 | } | |
484b4061 | 2374 | |
2a34c087 ML |
2375 | hctx->tags = set->tags[i]; |
2376 | WARN_ON(!hctx->tags); | |
2377 | ||
889fa31f CY |
2378 | /* |
2379 | * Set the map size to the number of mapped software queues. | |
2380 | * This is more accurate and more efficient than looping | |
2381 | * over all possibly mapped software queues. | |
2382 | */ | |
88459642 | 2383 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2384 | |
484b4061 JA |
2385 | /* |
2386 | * Initialize batch roundrobin counts | |
2387 | */ | |
f82ddf19 | 2388 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2389 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2390 | } | |
320ae51f JA |
2391 | } |
2392 | ||
8e8320c9 JA |
2393 | /* |
2394 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2395 | * the queue isn't live yet. | |
2396 | */ | |
2404e607 | 2397 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2398 | { |
2399 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2400 | int i; |
2401 | ||
2404e607 | 2402 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2403 | if (shared) |
2404e607 | 2404 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2405 | else |
2404e607 JM |
2406 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2407 | } | |
2408 | } | |
2409 | ||
8e8320c9 JA |
2410 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2411 | bool shared) | |
2404e607 JM |
2412 | { |
2413 | struct request_queue *q; | |
0d2602ca | 2414 | |
705cda97 BVA |
2415 | lockdep_assert_held(&set->tag_list_lock); |
2416 | ||
0d2602ca JA |
2417 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2418 | blk_mq_freeze_queue(q); | |
2404e607 | 2419 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2420 | blk_mq_unfreeze_queue(q); |
2421 | } | |
2422 | } | |
2423 | ||
2424 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2425 | { | |
2426 | struct blk_mq_tag_set *set = q->tag_set; | |
2427 | ||
0d2602ca | 2428 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2429 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2430 | if (list_is_singular(&set->tag_list)) { |
2431 | /* just transitioned to unshared */ | |
2432 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2433 | /* update existing queue */ | |
2434 | blk_mq_update_tag_set_depth(set, false); | |
2435 | } | |
0d2602ca | 2436 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2437 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2438 | } |
2439 | ||
2440 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2441 | struct request_queue *q) | |
2442 | { | |
2443 | q->tag_set = set; | |
2444 | ||
2445 | mutex_lock(&set->tag_list_lock); | |
2404e607 | 2446 | |
ff821d27 JA |
2447 | /* |
2448 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2449 | */ | |
2450 | if (!list_empty(&set->tag_list) && | |
2451 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2452 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2453 | /* update existing queue */ | |
2454 | blk_mq_update_tag_set_depth(set, true); | |
2455 | } | |
2456 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2457 | queue_set_hctx_shared(q, true); | |
705cda97 | 2458 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2459 | |
0d2602ca JA |
2460 | mutex_unlock(&set->tag_list_lock); |
2461 | } | |
2462 | ||
e09aae7e ML |
2463 | /* |
2464 | * It is the actual release handler for mq, but we do it from | |
2465 | * request queue's release handler for avoiding use-after-free | |
2466 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2467 | * but we can't group ctx/kctx kobj without it. | |
2468 | */ | |
2469 | void blk_mq_release(struct request_queue *q) | |
2470 | { | |
2471 | struct blk_mq_hw_ctx *hctx; | |
2472 | unsigned int i; | |
2473 | ||
2474 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2475 | queue_for_each_hw_ctx(q, hctx, i) { |
2476 | if (!hctx) | |
2477 | continue; | |
6c8b232e | 2478 | kobject_put(&hctx->kobj); |
c3b4afca | 2479 | } |
e09aae7e | 2480 | |
a723bab3 AM |
2481 | q->mq_map = NULL; |
2482 | ||
e09aae7e ML |
2483 | kfree(q->queue_hw_ctx); |
2484 | ||
7ea5fe31 ML |
2485 | /* |
2486 | * release .mq_kobj and sw queue's kobject now because | |
2487 | * both share lifetime with request queue. | |
2488 | */ | |
2489 | blk_mq_sysfs_deinit(q); | |
2490 | ||
e09aae7e ML |
2491 | free_percpu(q->queue_ctx); |
2492 | } | |
2493 | ||
24d2f903 | 2494 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2495 | { |
2496 | struct request_queue *uninit_q, *q; | |
2497 | ||
5ee0524b | 2498 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL); |
b62c21b7 MS |
2499 | if (!uninit_q) |
2500 | return ERR_PTR(-ENOMEM); | |
2501 | ||
2502 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2503 | if (IS_ERR(q)) | |
2504 | blk_cleanup_queue(uninit_q); | |
2505 | ||
2506 | return q; | |
2507 | } | |
2508 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2509 | ||
07319678 BVA |
2510 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2511 | { | |
2512 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2513 | ||
05707b64 | 2514 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), |
07319678 BVA |
2515 | __alignof__(struct blk_mq_hw_ctx)) != |
2516 | sizeof(struct blk_mq_hw_ctx)); | |
2517 | ||
2518 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2519 | hw_ctx_size += sizeof(struct srcu_struct); | |
2520 | ||
2521 | return hw_ctx_size; | |
2522 | } | |
2523 | ||
34d11ffa JW |
2524 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2525 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2526 | int hctx_idx, int node) | |
2527 | { | |
2528 | struct blk_mq_hw_ctx *hctx; | |
2529 | ||
2530 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), | |
2531 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2532 | node); | |
2533 | if (!hctx) | |
2534 | return NULL; | |
2535 | ||
2536 | if (!zalloc_cpumask_var_node(&hctx->cpumask, | |
2537 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2538 | node)) { | |
2539 | kfree(hctx); | |
2540 | return NULL; | |
2541 | } | |
2542 | ||
2543 | atomic_set(&hctx->nr_active, 0); | |
2544 | hctx->numa_node = node; | |
2545 | hctx->queue_num = hctx_idx; | |
2546 | ||
2547 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) { | |
2548 | free_cpumask_var(hctx->cpumask); | |
2549 | kfree(hctx); | |
2550 | return NULL; | |
2551 | } | |
2552 | blk_mq_hctx_kobj_init(hctx); | |
2553 | ||
2554 | return hctx; | |
2555 | } | |
2556 | ||
868f2f0b KB |
2557 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2558 | struct request_queue *q) | |
320ae51f | 2559 | { |
868f2f0b KB |
2560 | int i, j; |
2561 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2562 | |
fb350e0a ML |
2563 | /* protect against switching io scheduler */ |
2564 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2565 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2566 | int node; |
34d11ffa | 2567 | struct blk_mq_hw_ctx *hctx; |
868f2f0b KB |
2568 | |
2569 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
34d11ffa JW |
2570 | /* |
2571 | * If the hw queue has been mapped to another numa node, | |
2572 | * we need to realloc the hctx. If allocation fails, fallback | |
2573 | * to use the previous one. | |
2574 | */ | |
2575 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2576 | continue; | |
868f2f0b | 2577 | |
34d11ffa JW |
2578 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2579 | if (hctx) { | |
2580 | if (hctxs[i]) { | |
2581 | blk_mq_exit_hctx(q, set, hctxs[i], i); | |
2582 | kobject_put(&hctxs[i]->kobj); | |
2583 | } | |
2584 | hctxs[i] = hctx; | |
2585 | } else { | |
2586 | if (hctxs[i]) | |
2587 | pr_warn("Allocate new hctx on node %d fails,\ | |
2588 | fallback to previous one on node %d\n", | |
2589 | node, hctxs[i]->numa_node); | |
2590 | else | |
2591 | break; | |
868f2f0b | 2592 | } |
320ae51f | 2593 | } |
34d11ffa | 2594 | |
868f2f0b KB |
2595 | for (j = i; j < q->nr_hw_queues; j++) { |
2596 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2597 | ||
2598 | if (hctx) { | |
cc71a6f4 JA |
2599 | if (hctx->tags) |
2600 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2601 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2602 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2603 | hctxs[j] = NULL; |
2604 | ||
2605 | } | |
2606 | } | |
2607 | q->nr_hw_queues = i; | |
fb350e0a | 2608 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2609 | } |
2610 | ||
2611 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2612 | struct request_queue *q) | |
2613 | { | |
66841672 ML |
2614 | /* mark the queue as mq asap */ |
2615 | q->mq_ops = set->ops; | |
2616 | ||
34dbad5d | 2617 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2618 | blk_mq_poll_stats_bkt, |
2619 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2620 | if (!q->poll_cb) |
2621 | goto err_exit; | |
2622 | ||
868f2f0b KB |
2623 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2624 | if (!q->queue_ctx) | |
c7de5726 | 2625 | goto err_exit; |
868f2f0b | 2626 | |
737f98cf ML |
2627 | /* init q->mq_kobj and sw queues' kobjects */ |
2628 | blk_mq_sysfs_init(q); | |
2629 | ||
590b5b7d | 2630 | q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)), |
868f2f0b KB |
2631 | GFP_KERNEL, set->numa_node); |
2632 | if (!q->queue_hw_ctx) | |
2633 | goto err_percpu; | |
2634 | ||
bdd17e75 | 2635 | q->mq_map = set->mq_map; |
868f2f0b KB |
2636 | |
2637 | blk_mq_realloc_hw_ctxs(set, q); | |
2638 | if (!q->nr_hw_queues) | |
2639 | goto err_hctxs; | |
320ae51f | 2640 | |
287922eb | 2641 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2642 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2643 | |
2644 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2645 | |
94eddfbe | 2646 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2647 | |
05f1dd53 | 2648 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
f78bac2c | 2649 | queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q); |
05f1dd53 | 2650 | |
1be036e9 CH |
2651 | q->sg_reserved_size = INT_MAX; |
2652 | ||
2849450a | 2653 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2654 | INIT_LIST_HEAD(&q->requeue_list); |
2655 | spin_lock_init(&q->requeue_lock); | |
2656 | ||
254d259d | 2657 | blk_queue_make_request(q, blk_mq_make_request); |
ea435e1b CH |
2658 | if (q->mq_ops->poll) |
2659 | q->poll_fn = blk_mq_poll; | |
07068d5b | 2660 | |
eba71768 JA |
2661 | /* |
2662 | * Do this after blk_queue_make_request() overrides it... | |
2663 | */ | |
2664 | q->nr_requests = set->queue_depth; | |
2665 | ||
64f1c21e JA |
2666 | /* |
2667 | * Default to classic polling | |
2668 | */ | |
2669 | q->poll_nsec = -1; | |
2670 | ||
24d2f903 CH |
2671 | if (set->ops->complete) |
2672 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2673 | |
24d2f903 | 2674 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2675 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2676 | blk_mq_map_swqueue(q); |
4593fdbe | 2677 | |
d3484991 JA |
2678 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2679 | int ret; | |
2680 | ||
131d08e1 | 2681 | ret = elevator_init_mq(q); |
d3484991 JA |
2682 | if (ret) |
2683 | return ERR_PTR(ret); | |
2684 | } | |
2685 | ||
320ae51f | 2686 | return q; |
18741986 | 2687 | |
320ae51f | 2688 | err_hctxs: |
868f2f0b | 2689 | kfree(q->queue_hw_ctx); |
320ae51f | 2690 | err_percpu: |
868f2f0b | 2691 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2692 | err_exit: |
2693 | q->mq_ops = NULL; | |
320ae51f JA |
2694 | return ERR_PTR(-ENOMEM); |
2695 | } | |
b62c21b7 | 2696 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2697 | |
2698 | void blk_mq_free_queue(struct request_queue *q) | |
2699 | { | |
624dbe47 | 2700 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2701 | |
0d2602ca | 2702 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2703 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2704 | } |
320ae51f | 2705 | |
a5164405 JA |
2706 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2707 | { | |
2708 | int i; | |
2709 | ||
cc71a6f4 JA |
2710 | for (i = 0; i < set->nr_hw_queues; i++) |
2711 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2712 | goto out_unwind; |
a5164405 JA |
2713 | |
2714 | return 0; | |
2715 | ||
2716 | out_unwind: | |
2717 | while (--i >= 0) | |
cc71a6f4 | 2718 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2719 | |
a5164405 JA |
2720 | return -ENOMEM; |
2721 | } | |
2722 | ||
2723 | /* | |
2724 | * Allocate the request maps associated with this tag_set. Note that this | |
2725 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2726 | * will be updated to reflect the allocated depth. | |
2727 | */ | |
2728 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2729 | { | |
2730 | unsigned int depth; | |
2731 | int err; | |
2732 | ||
2733 | depth = set->queue_depth; | |
2734 | do { | |
2735 | err = __blk_mq_alloc_rq_maps(set); | |
2736 | if (!err) | |
2737 | break; | |
2738 | ||
2739 | set->queue_depth >>= 1; | |
2740 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2741 | err = -ENOMEM; | |
2742 | break; | |
2743 | } | |
2744 | } while (set->queue_depth); | |
2745 | ||
2746 | if (!set->queue_depth || err) { | |
2747 | pr_err("blk-mq: failed to allocate request map\n"); | |
2748 | return -ENOMEM; | |
2749 | } | |
2750 | ||
2751 | if (depth != set->queue_depth) | |
2752 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2753 | depth, set->queue_depth); | |
2754 | ||
2755 | return 0; | |
2756 | } | |
2757 | ||
ebe8bddb OS |
2758 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2759 | { | |
7d4901a9 | 2760 | if (set->ops->map_queues) { |
7d4901a9 ML |
2761 | /* |
2762 | * transport .map_queues is usually done in the following | |
2763 | * way: | |
2764 | * | |
2765 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
2766 | * mask = get_cpu_mask(queue) | |
2767 | * for_each_cpu(cpu, mask) | |
2768 | * set->mq_map[cpu] = queue; | |
2769 | * } | |
2770 | * | |
2771 | * When we need to remap, the table has to be cleared for | |
2772 | * killing stale mapping since one CPU may not be mapped | |
2773 | * to any hw queue. | |
2774 | */ | |
0da73d00 | 2775 | blk_mq_clear_mq_map(set); |
7d4901a9 | 2776 | |
ebe8bddb | 2777 | return set->ops->map_queues(set); |
7d4901a9 | 2778 | } else |
ebe8bddb OS |
2779 | return blk_mq_map_queues(set); |
2780 | } | |
2781 | ||
a4391c64 JA |
2782 | /* |
2783 | * Alloc a tag set to be associated with one or more request queues. | |
2784 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 2785 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
2786 | * value will be stored in set->queue_depth. |
2787 | */ | |
24d2f903 CH |
2788 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2789 | { | |
da695ba2 CH |
2790 | int ret; |
2791 | ||
205fb5f5 BVA |
2792 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2793 | ||
24d2f903 CH |
2794 | if (!set->nr_hw_queues) |
2795 | return -EINVAL; | |
a4391c64 | 2796 | if (!set->queue_depth) |
24d2f903 CH |
2797 | return -EINVAL; |
2798 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2799 | return -EINVAL; | |
2800 | ||
7d7e0f90 | 2801 | if (!set->ops->queue_rq) |
24d2f903 CH |
2802 | return -EINVAL; |
2803 | ||
de148297 ML |
2804 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2805 | return -EINVAL; | |
2806 | ||
a4391c64 JA |
2807 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2808 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2809 | BLK_MQ_MAX_DEPTH); | |
2810 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2811 | } | |
24d2f903 | 2812 | |
6637fadf SL |
2813 | /* |
2814 | * If a crashdump is active, then we are potentially in a very | |
2815 | * memory constrained environment. Limit us to 1 queue and | |
2816 | * 64 tags to prevent using too much memory. | |
2817 | */ | |
2818 | if (is_kdump_kernel()) { | |
2819 | set->nr_hw_queues = 1; | |
2820 | set->queue_depth = min(64U, set->queue_depth); | |
2821 | } | |
868f2f0b KB |
2822 | /* |
2823 | * There is no use for more h/w queues than cpus. | |
2824 | */ | |
2825 | if (set->nr_hw_queues > nr_cpu_ids) | |
2826 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2827 | |
590b5b7d | 2828 | set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2829 | GFP_KERNEL, set->numa_node); |
2830 | if (!set->tags) | |
a5164405 | 2831 | return -ENOMEM; |
24d2f903 | 2832 | |
da695ba2 | 2833 | ret = -ENOMEM; |
590b5b7d KC |
2834 | set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map), |
2835 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2836 | if (!set->mq_map) |
2837 | goto out_free_tags; | |
2838 | ||
ebe8bddb | 2839 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2840 | if (ret) |
2841 | goto out_free_mq_map; | |
2842 | ||
2843 | ret = blk_mq_alloc_rq_maps(set); | |
2844 | if (ret) | |
bdd17e75 | 2845 | goto out_free_mq_map; |
24d2f903 | 2846 | |
0d2602ca JA |
2847 | mutex_init(&set->tag_list_lock); |
2848 | INIT_LIST_HEAD(&set->tag_list); | |
2849 | ||
24d2f903 | 2850 | return 0; |
bdd17e75 CH |
2851 | |
2852 | out_free_mq_map: | |
2853 | kfree(set->mq_map); | |
2854 | set->mq_map = NULL; | |
2855 | out_free_tags: | |
5676e7b6 RE |
2856 | kfree(set->tags); |
2857 | set->tags = NULL; | |
da695ba2 | 2858 | return ret; |
24d2f903 CH |
2859 | } |
2860 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2861 | ||
2862 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2863 | { | |
2864 | int i; | |
2865 | ||
cc71a6f4 JA |
2866 | for (i = 0; i < nr_cpu_ids; i++) |
2867 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2868 | |
bdd17e75 CH |
2869 | kfree(set->mq_map); |
2870 | set->mq_map = NULL; | |
2871 | ||
981bd189 | 2872 | kfree(set->tags); |
5676e7b6 | 2873 | set->tags = NULL; |
24d2f903 CH |
2874 | } |
2875 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2876 | ||
e3a2b3f9 JA |
2877 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2878 | { | |
2879 | struct blk_mq_tag_set *set = q->tag_set; | |
2880 | struct blk_mq_hw_ctx *hctx; | |
2881 | int i, ret; | |
2882 | ||
bd166ef1 | 2883 | if (!set) |
e3a2b3f9 JA |
2884 | return -EINVAL; |
2885 | ||
70f36b60 | 2886 | blk_mq_freeze_queue(q); |
24f5a90f | 2887 | blk_mq_quiesce_queue(q); |
70f36b60 | 2888 | |
e3a2b3f9 JA |
2889 | ret = 0; |
2890 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2891 | if (!hctx->tags) |
2892 | continue; | |
bd166ef1 JA |
2893 | /* |
2894 | * If we're using an MQ scheduler, just update the scheduler | |
2895 | * queue depth. This is similar to what the old code would do. | |
2896 | */ | |
70f36b60 | 2897 | if (!hctx->sched_tags) { |
c2e82a23 | 2898 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
2899 | false); |
2900 | } else { | |
2901 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2902 | nr, true); | |
2903 | } | |
e3a2b3f9 JA |
2904 | if (ret) |
2905 | break; | |
2906 | } | |
2907 | ||
2908 | if (!ret) | |
2909 | q->nr_requests = nr; | |
2910 | ||
24f5a90f | 2911 | blk_mq_unquiesce_queue(q); |
70f36b60 | 2912 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2913 | |
e3a2b3f9 JA |
2914 | return ret; |
2915 | } | |
2916 | ||
d48ece20 JW |
2917 | /* |
2918 | * request_queue and elevator_type pair. | |
2919 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
2920 | * the elevator_type associated with a request_queue. | |
2921 | */ | |
2922 | struct blk_mq_qe_pair { | |
2923 | struct list_head node; | |
2924 | struct request_queue *q; | |
2925 | struct elevator_type *type; | |
2926 | }; | |
2927 | ||
2928 | /* | |
2929 | * Cache the elevator_type in qe pair list and switch the | |
2930 | * io scheduler to 'none' | |
2931 | */ | |
2932 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
2933 | struct request_queue *q) | |
2934 | { | |
2935 | struct blk_mq_qe_pair *qe; | |
2936 | ||
2937 | if (!q->elevator) | |
2938 | return true; | |
2939 | ||
2940 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
2941 | if (!qe) | |
2942 | return false; | |
2943 | ||
2944 | INIT_LIST_HEAD(&qe->node); | |
2945 | qe->q = q; | |
2946 | qe->type = q->elevator->type; | |
2947 | list_add(&qe->node, head); | |
2948 | ||
2949 | mutex_lock(&q->sysfs_lock); | |
2950 | /* | |
2951 | * After elevator_switch_mq, the previous elevator_queue will be | |
2952 | * released by elevator_release. The reference of the io scheduler | |
2953 | * module get by elevator_get will also be put. So we need to get | |
2954 | * a reference of the io scheduler module here to prevent it to be | |
2955 | * removed. | |
2956 | */ | |
2957 | __module_get(qe->type->elevator_owner); | |
2958 | elevator_switch_mq(q, NULL); | |
2959 | mutex_unlock(&q->sysfs_lock); | |
2960 | ||
2961 | return true; | |
2962 | } | |
2963 | ||
2964 | static void blk_mq_elv_switch_back(struct list_head *head, | |
2965 | struct request_queue *q) | |
2966 | { | |
2967 | struct blk_mq_qe_pair *qe; | |
2968 | struct elevator_type *t = NULL; | |
2969 | ||
2970 | list_for_each_entry(qe, head, node) | |
2971 | if (qe->q == q) { | |
2972 | t = qe->type; | |
2973 | break; | |
2974 | } | |
2975 | ||
2976 | if (!t) | |
2977 | return; | |
2978 | ||
2979 | list_del(&qe->node); | |
2980 | kfree(qe); | |
2981 | ||
2982 | mutex_lock(&q->sysfs_lock); | |
2983 | elevator_switch_mq(q, t); | |
2984 | mutex_unlock(&q->sysfs_lock); | |
2985 | } | |
2986 | ||
e4dc2b32 KB |
2987 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2988 | int nr_hw_queues) | |
868f2f0b KB |
2989 | { |
2990 | struct request_queue *q; | |
d48ece20 | 2991 | LIST_HEAD(head); |
868f2f0b | 2992 | |
705cda97 BVA |
2993 | lockdep_assert_held(&set->tag_list_lock); |
2994 | ||
868f2f0b KB |
2995 | if (nr_hw_queues > nr_cpu_ids) |
2996 | nr_hw_queues = nr_cpu_ids; | |
2997 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2998 | return; | |
2999 | ||
3000 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3001 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3002 | /* |
3003 | * Sync with blk_mq_queue_tag_busy_iter. | |
3004 | */ | |
3005 | synchronize_rcu(); | |
d48ece20 JW |
3006 | /* |
3007 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3008 | * with the previous scheduler. We will switch back once we are done | |
3009 | * updating the new sw to hw queue mappings. | |
3010 | */ | |
3011 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3012 | if (!blk_mq_elv_switch_none(&head, q)) | |
3013 | goto switch_back; | |
868f2f0b | 3014 | |
477e19de JW |
3015 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3016 | blk_mq_debugfs_unregister_hctxs(q); | |
3017 | blk_mq_sysfs_unregister(q); | |
3018 | } | |
3019 | ||
868f2f0b | 3020 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3021 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3022 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3023 | blk_mq_realloc_hw_ctxs(set, q); | |
477e19de JW |
3024 | blk_mq_map_swqueue(q); |
3025 | } | |
3026 | ||
3027 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3028 | blk_mq_sysfs_register(q); | |
3029 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3030 | } |
3031 | ||
d48ece20 JW |
3032 | switch_back: |
3033 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3034 | blk_mq_elv_switch_back(&head, q); | |
3035 | ||
868f2f0b KB |
3036 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3037 | blk_mq_unfreeze_queue(q); | |
3038 | } | |
e4dc2b32 KB |
3039 | |
3040 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3041 | { | |
3042 | mutex_lock(&set->tag_list_lock); | |
3043 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3044 | mutex_unlock(&set->tag_list_lock); | |
3045 | } | |
868f2f0b KB |
3046 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3047 | ||
34dbad5d OS |
3048 | /* Enable polling stats and return whether they were already enabled. */ |
3049 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3050 | { | |
3051 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3052 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3053 | return true; |
3054 | blk_stat_add_callback(q, q->poll_cb); | |
3055 | return false; | |
3056 | } | |
3057 | ||
3058 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3059 | { | |
3060 | /* | |
3061 | * We don't arm the callback if polling stats are not enabled or the | |
3062 | * callback is already active. | |
3063 | */ | |
3064 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3065 | blk_stat_is_active(q->poll_cb)) | |
3066 | return; | |
3067 | ||
3068 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3069 | } | |
3070 | ||
3071 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3072 | { | |
3073 | struct request_queue *q = cb->data; | |
720b8ccc | 3074 | int bucket; |
34dbad5d | 3075 | |
720b8ccc SB |
3076 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3077 | if (cb->stat[bucket].nr_samples) | |
3078 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3079 | } | |
34dbad5d OS |
3080 | } |
3081 | ||
64f1c21e JA |
3082 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3083 | struct blk_mq_hw_ctx *hctx, | |
3084 | struct request *rq) | |
3085 | { | |
64f1c21e | 3086 | unsigned long ret = 0; |
720b8ccc | 3087 | int bucket; |
64f1c21e JA |
3088 | |
3089 | /* | |
3090 | * If stats collection isn't on, don't sleep but turn it on for | |
3091 | * future users | |
3092 | */ | |
34dbad5d | 3093 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3094 | return 0; |
3095 | ||
64f1c21e JA |
3096 | /* |
3097 | * As an optimistic guess, use half of the mean service time | |
3098 | * for this type of request. We can (and should) make this smarter. | |
3099 | * For instance, if the completion latencies are tight, we can | |
3100 | * get closer than just half the mean. This is especially | |
3101 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3102 | * than ~10 usec. We do use the stats for the relevant IO size |
3103 | * if available which does lead to better estimates. | |
64f1c21e | 3104 | */ |
720b8ccc SB |
3105 | bucket = blk_mq_poll_stats_bkt(rq); |
3106 | if (bucket < 0) | |
3107 | return ret; | |
3108 | ||
3109 | if (q->poll_stat[bucket].nr_samples) | |
3110 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3111 | |
3112 | return ret; | |
3113 | } | |
3114 | ||
06426adf | 3115 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3116 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3117 | struct request *rq) |
3118 | { | |
3119 | struct hrtimer_sleeper hs; | |
3120 | enum hrtimer_mode mode; | |
64f1c21e | 3121 | unsigned int nsecs; |
06426adf JA |
3122 | ktime_t kt; |
3123 | ||
76a86f9d | 3124 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3125 | return false; |
3126 | ||
3127 | /* | |
3128 | * poll_nsec can be: | |
3129 | * | |
3130 | * -1: don't ever hybrid sleep | |
3131 | * 0: use half of prev avg | |
3132 | * >0: use this specific value | |
3133 | */ | |
3134 | if (q->poll_nsec == -1) | |
3135 | return false; | |
3136 | else if (q->poll_nsec > 0) | |
3137 | nsecs = q->poll_nsec; | |
3138 | else | |
3139 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3140 | ||
3141 | if (!nsecs) | |
06426adf JA |
3142 | return false; |
3143 | ||
76a86f9d | 3144 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3145 | |
3146 | /* | |
3147 | * This will be replaced with the stats tracking code, using | |
3148 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3149 | */ | |
8b0e1953 | 3150 | kt = nsecs; |
06426adf JA |
3151 | |
3152 | mode = HRTIMER_MODE_REL; | |
3153 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3154 | hrtimer_set_expires(&hs.timer, kt); | |
3155 | ||
3156 | hrtimer_init_sleeper(&hs, current); | |
3157 | do { | |
5a61c363 | 3158 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3159 | break; |
3160 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3161 | hrtimer_start_expires(&hs.timer, mode); | |
3162 | if (hs.task) | |
3163 | io_schedule(); | |
3164 | hrtimer_cancel(&hs.timer); | |
3165 | mode = HRTIMER_MODE_ABS; | |
3166 | } while (hs.task && !signal_pending(current)); | |
3167 | ||
3168 | __set_current_state(TASK_RUNNING); | |
3169 | destroy_hrtimer_on_stack(&hs.timer); | |
3170 | return true; | |
3171 | } | |
3172 | ||
bbd7bb70 JA |
3173 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
3174 | { | |
3175 | struct request_queue *q = hctx->queue; | |
3176 | long state; | |
3177 | ||
06426adf JA |
3178 | /* |
3179 | * If we sleep, have the caller restart the poll loop to reset | |
3180 | * the state. Like for the other success return cases, the | |
3181 | * caller is responsible for checking if the IO completed. If | |
3182 | * the IO isn't complete, we'll get called again and will go | |
3183 | * straight to the busy poll loop. | |
3184 | */ | |
64f1c21e | 3185 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
3186 | return true; |
3187 | ||
bbd7bb70 JA |
3188 | hctx->poll_considered++; |
3189 | ||
3190 | state = current->state; | |
3191 | while (!need_resched()) { | |
3192 | int ret; | |
3193 | ||
3194 | hctx->poll_invoked++; | |
3195 | ||
3196 | ret = q->mq_ops->poll(hctx, rq->tag); | |
3197 | if (ret > 0) { | |
3198 | hctx->poll_success++; | |
3199 | set_current_state(TASK_RUNNING); | |
3200 | return true; | |
3201 | } | |
3202 | ||
3203 | if (signal_pending_state(state, current)) | |
3204 | set_current_state(TASK_RUNNING); | |
3205 | ||
3206 | if (current->state == TASK_RUNNING) | |
3207 | return true; | |
3208 | if (ret < 0) | |
3209 | break; | |
3210 | cpu_relax(); | |
3211 | } | |
3212 | ||
67b4110f | 3213 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3214 | return false; |
3215 | } | |
3216 | ||
ea435e1b | 3217 | static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) |
bbd7bb70 JA |
3218 | { |
3219 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3220 | struct request *rq; |
3221 | ||
ea435e1b | 3222 | if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) |
bbd7bb70 JA |
3223 | return false; |
3224 | ||
bbd7bb70 | 3225 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
bd166ef1 JA |
3226 | if (!blk_qc_t_is_internal(cookie)) |
3227 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 3228 | else { |
bd166ef1 | 3229 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
3230 | /* |
3231 | * With scheduling, if the request has completed, we'll | |
3232 | * get a NULL return here, as we clear the sched tag when | |
3233 | * that happens. The request still remains valid, like always, | |
3234 | * so we should be safe with just the NULL check. | |
3235 | */ | |
3236 | if (!rq) | |
3237 | return false; | |
3238 | } | |
bbd7bb70 JA |
3239 | |
3240 | return __blk_mq_poll(hctx, rq); | |
3241 | } | |
bbd7bb70 | 3242 | |
320ae51f JA |
3243 | static int __init blk_mq_init(void) |
3244 | { | |
9467f859 TG |
3245 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3246 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3247 | return 0; |
3248 | } | |
3249 | subsys_initcall(blk_mq_init); |