]>
Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
f75782e4 | 13 | #include <linux/kmemleak.h> |
320ae51f JA |
14 | #include <linux/mm.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/smp.h> | |
19 | #include <linux/llist.h> | |
20 | #include <linux/list_sort.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/cache.h> | |
23 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
a892c8d5 | 29 | #include <linux/blk-crypto.h> |
320ae51f JA |
30 | |
31 | #include <trace/events/block.h> | |
32 | ||
33 | #include <linux/blk-mq.h> | |
54d4e6ab | 34 | #include <linux/t10-pi.h> |
320ae51f JA |
35 | #include "blk.h" |
36 | #include "blk-mq.h" | |
9c1051aa | 37 | #include "blk-mq-debugfs.h" |
320ae51f | 38 | #include "blk-mq-tag.h" |
986d413b | 39 | #include "blk-pm.h" |
cf43e6be | 40 | #include "blk-stat.h" |
bd166ef1 | 41 | #include "blk-mq-sched.h" |
c1c80384 | 42 | #include "blk-rq-qos.h" |
320ae51f | 43 | |
c3077b5d CH |
44 | static DEFINE_PER_CPU(struct list_head, blk_cpu_done); |
45 | ||
34dbad5d OS |
46 | static void blk_mq_poll_stats_start(struct request_queue *q); |
47 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
48 | ||
720b8ccc SB |
49 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
50 | { | |
3d244306 | 51 | int ddir, sectors, bucket; |
720b8ccc | 52 | |
99c749a4 | 53 | ddir = rq_data_dir(rq); |
3d244306 | 54 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 55 | |
3d244306 | 56 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
57 | |
58 | if (bucket < 0) | |
59 | return -1; | |
60 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
61 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
62 | ||
63 | return bucket; | |
64 | } | |
65 | ||
320ae51f | 66 | /* |
85fae294 YY |
67 | * Check if any of the ctx, dispatch list or elevator |
68 | * have pending work in this hardware queue. | |
320ae51f | 69 | */ |
79f720a7 | 70 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 71 | { |
79f720a7 JA |
72 | return !list_empty_careful(&hctx->dispatch) || |
73 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 74 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
75 | } |
76 | ||
320ae51f JA |
77 | /* |
78 | * Mark this ctx as having pending work in this hardware queue | |
79 | */ | |
80 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
f31967f0 JA |
83 | const int bit = ctx->index_hw[hctx->type]; |
84 | ||
85 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
86 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
87 | } |
88 | ||
89 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
90 | struct blk_mq_ctx *ctx) | |
91 | { | |
f31967f0 JA |
92 | const int bit = ctx->index_hw[hctx->type]; |
93 | ||
94 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
95 | } |
96 | ||
f299b7c7 | 97 | struct mq_inflight { |
8446fe92 | 98 | struct block_device *part; |
a2e80f6f | 99 | unsigned int inflight[2]; |
f299b7c7 JA |
100 | }; |
101 | ||
7baa8572 | 102 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
103 | struct request *rq, void *priv, |
104 | bool reserved) | |
105 | { | |
106 | struct mq_inflight *mi = priv; | |
107 | ||
b0d97557 JX |
108 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
109 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 110 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
111 | |
112 | return true; | |
f299b7c7 JA |
113 | } |
114 | ||
8446fe92 CH |
115 | unsigned int blk_mq_in_flight(struct request_queue *q, |
116 | struct block_device *part) | |
f299b7c7 | 117 | { |
a2e80f6f | 118 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 119 | |
f299b7c7 | 120 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 121 | |
a2e80f6f | 122 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
123 | } |
124 | ||
8446fe92 CH |
125 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
126 | unsigned int inflight[2]) | |
bf0ddaba | 127 | { |
a2e80f6f | 128 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 129 | |
bb4e6b14 | 130 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
131 | inflight[0] = mi.inflight[0]; |
132 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
133 | } |
134 | ||
1671d522 | 135 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 136 | { |
7996a8b5 BL |
137 | mutex_lock(&q->mq_freeze_lock); |
138 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 139 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 140 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 141 | if (queue_is_mq(q)) |
055f6e18 | 142 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
143 | } else { |
144 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 145 | } |
f3af020b | 146 | } |
1671d522 | 147 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 148 | |
6bae363e | 149 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 150 | { |
3ef28e83 | 151 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 152 | } |
6bae363e | 153 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 154 | |
f91328c4 KB |
155 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
156 | unsigned long timeout) | |
157 | { | |
158 | return wait_event_timeout(q->mq_freeze_wq, | |
159 | percpu_ref_is_zero(&q->q_usage_counter), | |
160 | timeout); | |
161 | } | |
162 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 163 | |
f3af020b TH |
164 | /* |
165 | * Guarantee no request is in use, so we can change any data structure of | |
166 | * the queue afterward. | |
167 | */ | |
3ef28e83 | 168 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 169 | { |
3ef28e83 DW |
170 | /* |
171 | * In the !blk_mq case we are only calling this to kill the | |
172 | * q_usage_counter, otherwise this increases the freeze depth | |
173 | * and waits for it to return to zero. For this reason there is | |
174 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
175 | * exported to drivers as the only user for unfreeze is blk_mq. | |
176 | */ | |
1671d522 | 177 | blk_freeze_queue_start(q); |
f3af020b TH |
178 | blk_mq_freeze_queue_wait(q); |
179 | } | |
3ef28e83 DW |
180 | |
181 | void blk_mq_freeze_queue(struct request_queue *q) | |
182 | { | |
183 | /* | |
184 | * ...just an alias to keep freeze and unfreeze actions balanced | |
185 | * in the blk_mq_* namespace | |
186 | */ | |
187 | blk_freeze_queue(q); | |
188 | } | |
c761d96b | 189 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 190 | |
b4c6a028 | 191 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 192 | { |
7996a8b5 BL |
193 | mutex_lock(&q->mq_freeze_lock); |
194 | q->mq_freeze_depth--; | |
195 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
196 | if (!q->mq_freeze_depth) { | |
bdd63160 | 197 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 198 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 199 | } |
7996a8b5 | 200 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 201 | } |
b4c6a028 | 202 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 203 | |
852ec809 BVA |
204 | /* |
205 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
206 | * mpt3sas driver such that this function can be removed. | |
207 | */ | |
208 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
209 | { | |
8814ce8a | 210 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
211 | } |
212 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
213 | ||
6a83e74d | 214 | /** |
69e07c4a | 215 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
216 | * @q: request queue. |
217 | * | |
218 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
219 | * callback function is invoked. Once this function is returned, we make |
220 | * sure no dispatch can happen until the queue is unquiesced via | |
221 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
222 | */ |
223 | void blk_mq_quiesce_queue(struct request_queue *q) | |
224 | { | |
225 | struct blk_mq_hw_ctx *hctx; | |
226 | unsigned int i; | |
227 | bool rcu = false; | |
228 | ||
1d9e9bc6 | 229 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 230 | |
6a83e74d BVA |
231 | queue_for_each_hw_ctx(q, hctx, i) { |
232 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 233 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
234 | else |
235 | rcu = true; | |
236 | } | |
237 | if (rcu) | |
238 | synchronize_rcu(); | |
239 | } | |
240 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
241 | ||
e4e73913 ML |
242 | /* |
243 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
244 | * @q: request queue. | |
245 | * | |
246 | * This function recovers queue into the state before quiescing | |
247 | * which is done by blk_mq_quiesce_queue. | |
248 | */ | |
249 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
250 | { | |
8814ce8a | 251 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 252 | |
1d9e9bc6 ML |
253 | /* dispatch requests which are inserted during quiescing */ |
254 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
255 | } |
256 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
257 | ||
aed3ea94 JA |
258 | void blk_mq_wake_waiters(struct request_queue *q) |
259 | { | |
260 | struct blk_mq_hw_ctx *hctx; | |
261 | unsigned int i; | |
262 | ||
263 | queue_for_each_hw_ctx(q, hctx, i) | |
264 | if (blk_mq_hw_queue_mapped(hctx)) | |
265 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
266 | } | |
267 | ||
fe1f4526 | 268 | /* |
9a91b05b HT |
269 | * Only need start/end time stamping if we have iostat or |
270 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
271 | */ |
272 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
273 | { | |
9a91b05b | 274 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
275 | } |
276 | ||
e4cdf1a1 | 277 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 278 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 279 | { |
e4cdf1a1 CH |
280 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
281 | struct request *rq = tags->static_rqs[tag]; | |
c3a148d2 | 282 | |
42fdc5e4 | 283 | if (data->q->elevator) { |
76647368 | 284 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
285 | rq->internal_tag = tag; |
286 | } else { | |
e4cdf1a1 | 287 | rq->tag = tag; |
76647368 | 288 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
289 | } |
290 | ||
af76e555 | 291 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
292 | rq->q = data->q; |
293 | rq->mq_ctx = data->ctx; | |
ea4f995e | 294 | rq->mq_hctx = data->hctx; |
568f2700 | 295 | rq->rq_flags = 0; |
7ea4d8a4 | 296 | rq->cmd_flags = data->cmd_flags; |
0854bcdc BVA |
297 | if (data->flags & BLK_MQ_REQ_PM) |
298 | rq->rq_flags |= RQF_PM; | |
e4cdf1a1 | 299 | if (blk_queue_io_stat(data->q)) |
e8064021 | 300 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 301 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
302 | INIT_HLIST_NODE(&rq->hash); |
303 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
304 | rq->rq_disk = NULL; |
305 | rq->part = NULL; | |
6f816b4b TH |
306 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
307 | rq->alloc_time_ns = alloc_time_ns; | |
308 | #endif | |
fe1f4526 JA |
309 | if (blk_mq_need_time_stamp(rq)) |
310 | rq->start_time_ns = ktime_get_ns(); | |
311 | else | |
312 | rq->start_time_ns = 0; | |
544ccc8d | 313 | rq->io_start_time_ns = 0; |
3d244306 | 314 | rq->stats_sectors = 0; |
af76e555 CH |
315 | rq->nr_phys_segments = 0; |
316 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
317 | rq->nr_integrity_segments = 0; | |
318 | #endif | |
a892c8d5 | 319 | blk_crypto_rq_set_defaults(rq); |
af76e555 | 320 | /* tag was already set */ |
079076b3 | 321 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 322 | |
f6be4fb4 JA |
323 | rq->timeout = 0; |
324 | ||
af76e555 CH |
325 | rq->end_io = NULL; |
326 | rq->end_io_data = NULL; | |
af76e555 | 327 | |
7ea4d8a4 | 328 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; |
12f5b931 | 329 | refcount_set(&rq->ref, 1); |
7ea4d8a4 CH |
330 | |
331 | if (!op_is_flush(data->cmd_flags)) { | |
332 | struct elevator_queue *e = data->q->elevator; | |
333 | ||
334 | rq->elv.icq = NULL; | |
335 | if (e && e->type->ops.prepare_request) { | |
336 | if (e->type->icq_cache) | |
337 | blk_mq_sched_assign_ioc(rq); | |
338 | ||
339 | e->type->ops.prepare_request(rq); | |
340 | rq->rq_flags |= RQF_ELVPRIV; | |
341 | } | |
342 | } | |
343 | ||
344 | data->hctx->queued++; | |
e4cdf1a1 | 345 | return rq; |
5dee8577 CH |
346 | } |
347 | ||
e6e7abff | 348 | static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data) |
d2c0d383 | 349 | { |
e6e7abff | 350 | struct request_queue *q = data->q; |
d2c0d383 | 351 | struct elevator_queue *e = q->elevator; |
6f816b4b | 352 | u64 alloc_time_ns = 0; |
600c3b0c | 353 | unsigned int tag; |
d2c0d383 | 354 | |
6f816b4b TH |
355 | /* alloc_time includes depth and tag waits */ |
356 | if (blk_queue_rq_alloc_time(q)) | |
357 | alloc_time_ns = ktime_get_ns(); | |
358 | ||
f9afca4d | 359 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 360 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
361 | |
362 | if (e) { | |
d2c0d383 CH |
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 | */ |
f9afca4d JA |
368 | if (!op_is_flush(data->cmd_flags) && |
369 | e->type->ops.limit_depth && | |
17a51199 | 370 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 371 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
372 | } |
373 | ||
bf0beec0 | 374 | retry: |
600c3b0c CH |
375 | data->ctx = blk_mq_get_ctx(q); |
376 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 377 | if (!e) |
600c3b0c CH |
378 | blk_mq_tag_busy(data->hctx); |
379 | ||
bf0beec0 ML |
380 | /* |
381 | * Waiting allocations only fail because of an inactive hctx. In that | |
382 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
383 | * should have migrated us to an online CPU by now. | |
384 | */ | |
e4cdf1a1 | 385 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
386 | if (tag == BLK_MQ_NO_TAG) { |
387 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
388 | return NULL; | |
389 | ||
390 | /* | |
391 | * Give up the CPU and sleep for a random short time to ensure | |
392 | * that thread using a realtime scheduling class are migrated | |
70f15a4f | 393 | * off the CPU, and thus off the hctx that is going away. |
bf0beec0 ML |
394 | */ |
395 | msleep(3); | |
396 | goto retry; | |
397 | } | |
7ea4d8a4 | 398 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
399 | } |
400 | ||
cd6ce148 | 401 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 402 | blk_mq_req_flags_t flags) |
320ae51f | 403 | { |
e6e7abff CH |
404 | struct blk_mq_alloc_data data = { |
405 | .q = q, | |
406 | .flags = flags, | |
407 | .cmd_flags = op, | |
408 | }; | |
bd166ef1 | 409 | struct request *rq; |
a492f075 | 410 | int ret; |
320ae51f | 411 | |
3a0a5299 | 412 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
413 | if (ret) |
414 | return ERR_PTR(ret); | |
320ae51f | 415 | |
e6e7abff | 416 | rq = __blk_mq_alloc_request(&data); |
bd166ef1 | 417 | if (!rq) |
a5ea5811 | 418 | goto out_queue_exit; |
0c4de0f3 CH |
419 | rq->__data_len = 0; |
420 | rq->__sector = (sector_t) -1; | |
421 | rq->bio = rq->biotail = NULL; | |
320ae51f | 422 | return rq; |
a5ea5811 CH |
423 | out_queue_exit: |
424 | blk_queue_exit(q); | |
425 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 426 | } |
4bb659b1 | 427 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 428 | |
cd6ce148 | 429 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 430 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 431 | { |
e6e7abff CH |
432 | struct blk_mq_alloc_data data = { |
433 | .q = q, | |
434 | .flags = flags, | |
435 | .cmd_flags = op, | |
436 | }; | |
600c3b0c | 437 | u64 alloc_time_ns = 0; |
6d2809d5 | 438 | unsigned int cpu; |
600c3b0c | 439 | unsigned int tag; |
1f5bd336 ML |
440 | int ret; |
441 | ||
600c3b0c CH |
442 | /* alloc_time includes depth and tag waits */ |
443 | if (blk_queue_rq_alloc_time(q)) | |
444 | alloc_time_ns = ktime_get_ns(); | |
445 | ||
1f5bd336 ML |
446 | /* |
447 | * If the tag allocator sleeps we could get an allocation for a | |
448 | * different hardware context. No need to complicate the low level | |
449 | * allocator for this for the rare use case of a command tied to | |
450 | * a specific queue. | |
451 | */ | |
600c3b0c | 452 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
453 | return ERR_PTR(-EINVAL); |
454 | ||
455 | if (hctx_idx >= q->nr_hw_queues) | |
456 | return ERR_PTR(-EIO); | |
457 | ||
3a0a5299 | 458 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
459 | if (ret) |
460 | return ERR_PTR(ret); | |
461 | ||
c8712c6a CH |
462 | /* |
463 | * Check if the hardware context is actually mapped to anything. | |
464 | * If not tell the caller that it should skip this queue. | |
465 | */ | |
a5ea5811 | 466 | ret = -EXDEV; |
e6e7abff CH |
467 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
468 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 469 | goto out_queue_exit; |
e6e7abff CH |
470 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
471 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 472 | |
42fdc5e4 | 473 | if (!q->elevator) |
600c3b0c CH |
474 | blk_mq_tag_busy(data.hctx); |
475 | ||
a5ea5811 | 476 | ret = -EWOULDBLOCK; |
600c3b0c CH |
477 | tag = blk_mq_get_tag(&data); |
478 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 479 | goto out_queue_exit; |
600c3b0c CH |
480 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
481 | ||
a5ea5811 CH |
482 | out_queue_exit: |
483 | blk_queue_exit(q); | |
484 | return ERR_PTR(ret); | |
1f5bd336 ML |
485 | } |
486 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
487 | ||
12f5b931 KB |
488 | static void __blk_mq_free_request(struct request *rq) |
489 | { | |
490 | struct request_queue *q = rq->q; | |
491 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 492 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
493 | const int sched_tag = rq->internal_tag; |
494 | ||
a892c8d5 | 495 | blk_crypto_free_request(rq); |
986d413b | 496 | blk_pm_mark_last_busy(rq); |
ea4f995e | 497 | rq->mq_hctx = NULL; |
76647368 | 498 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 499 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 500 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 501 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
502 | blk_mq_sched_restart(hctx); |
503 | blk_queue_exit(q); | |
504 | } | |
505 | ||
6af54051 | 506 | void blk_mq_free_request(struct request *rq) |
320ae51f | 507 | { |
320ae51f | 508 | struct request_queue *q = rq->q; |
6af54051 CH |
509 | struct elevator_queue *e = q->elevator; |
510 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 511 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 512 | |
5bbf4e5a | 513 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
514 | if (e && e->type->ops.finish_request) |
515 | e->type->ops.finish_request(rq); | |
6af54051 CH |
516 | if (rq->elv.icq) { |
517 | put_io_context(rq->elv.icq->ioc); | |
518 | rq->elv.icq = NULL; | |
519 | } | |
520 | } | |
320ae51f | 521 | |
6af54051 | 522 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 523 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 524 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 525 | |
7beb2f84 JA |
526 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
527 | laptop_io_completion(q->backing_dev_info); | |
528 | ||
a7905043 | 529 | rq_qos_done(q, rq); |
0d2602ca | 530 | |
12f5b931 KB |
531 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
532 | if (refcount_dec_and_test(&rq->ref)) | |
533 | __blk_mq_free_request(rq); | |
320ae51f | 534 | } |
1a3b595a | 535 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 536 | |
2a842aca | 537 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 538 | { |
fe1f4526 JA |
539 | u64 now = 0; |
540 | ||
541 | if (blk_mq_need_time_stamp(rq)) | |
542 | now = ktime_get_ns(); | |
522a7775 | 543 | |
4bc6339a OS |
544 | if (rq->rq_flags & RQF_STATS) { |
545 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 546 | blk_stat_add(rq, now); |
4bc6339a OS |
547 | } |
548 | ||
87890092 | 549 | blk_mq_sched_completed_request(rq, now); |
ed88660a | 550 | |
522a7775 | 551 | blk_account_io_done(rq, now); |
0d11e6ac | 552 | |
91b63639 | 553 | if (rq->end_io) { |
a7905043 | 554 | rq_qos_done(rq->q, rq); |
320ae51f | 555 | rq->end_io(rq, error); |
91b63639 | 556 | } else { |
320ae51f | 557 | blk_mq_free_request(rq); |
91b63639 | 558 | } |
320ae51f | 559 | } |
c8a446ad | 560 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 561 | |
2a842aca | 562 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
563 | { |
564 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
565 | BUG(); | |
c8a446ad | 566 | __blk_mq_end_request(rq, error); |
63151a44 | 567 | } |
c8a446ad | 568 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 569 | |
c3077b5d CH |
570 | /* |
571 | * Softirq action handler - move entries to local list and loop over them | |
572 | * while passing them to the queue registered handler. | |
573 | */ | |
574 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) | |
320ae51f | 575 | { |
c3077b5d | 576 | struct list_head *cpu_list, local_list; |
320ae51f | 577 | |
c3077b5d CH |
578 | local_irq_disable(); |
579 | cpu_list = this_cpu_ptr(&blk_cpu_done); | |
580 | list_replace_init(cpu_list, &local_list); | |
581 | local_irq_enable(); | |
582 | ||
583 | while (!list_empty(&local_list)) { | |
584 | struct request *rq; | |
585 | ||
586 | rq = list_entry(local_list.next, struct request, ipi_list); | |
587 | list_del_init(&rq->ipi_list); | |
588 | rq->q->mq_ops->complete(rq); | |
589 | } | |
320ae51f | 590 | } |
320ae51f | 591 | |
115243f5 | 592 | static void blk_mq_trigger_softirq(struct request *rq) |
320ae51f | 593 | { |
d391a7a3 CH |
594 | struct list_head *list; |
595 | unsigned long flags; | |
c3077b5d | 596 | |
d391a7a3 CH |
597 | local_irq_save(flags); |
598 | list = this_cpu_ptr(&blk_cpu_done); | |
c3077b5d CH |
599 | list_add_tail(&rq->ipi_list, list); |
600 | ||
115243f5 CH |
601 | /* |
602 | * If the list only contains our just added request, signal a raise of | |
603 | * the softirq. If there are already entries there, someone already | |
604 | * raised the irq but it hasn't run yet. | |
605 | */ | |
c3077b5d CH |
606 | if (list->next == &rq->ipi_list) |
607 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
d391a7a3 | 608 | local_irq_restore(flags); |
115243f5 CH |
609 | } |
610 | ||
c3077b5d CH |
611 | static int blk_softirq_cpu_dead(unsigned int cpu) |
612 | { | |
613 | /* | |
614 | * If a CPU goes away, splice its entries to the current CPU | |
615 | * and trigger a run of the softirq | |
616 | */ | |
617 | local_irq_disable(); | |
618 | list_splice_init(&per_cpu(blk_cpu_done, cpu), | |
619 | this_cpu_ptr(&blk_cpu_done)); | |
620 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
621 | local_irq_enable(); | |
622 | ||
623 | return 0; | |
624 | } | |
625 | ||
40d09b53 CH |
626 | |
627 | static void __blk_mq_complete_request_remote(void *data) | |
c3077b5d | 628 | { |
40d09b53 | 629 | struct request *rq = data; |
320ae51f | 630 | |
36e76539 | 631 | /* |
d391a7a3 CH |
632 | * For most of single queue controllers, there is only one irq vector |
633 | * for handling I/O completion, and the only irq's affinity is set | |
634 | * to all possible CPUs. On most of ARCHs, this affinity means the irq | |
635 | * is handled on one specific CPU. | |
36e76539 | 636 | * |
d391a7a3 CH |
637 | * So complete I/O requests in softirq context in case of single queue |
638 | * devices to avoid degrading I/O performance due to irqsoff latency. | |
36e76539 | 639 | */ |
d391a7a3 CH |
640 | if (rq->q->nr_hw_queues == 1) |
641 | blk_mq_trigger_softirq(rq); | |
642 | else | |
643 | rq->q->mq_ops->complete(rq); | |
c3077b5d CH |
644 | } |
645 | ||
96339526 CH |
646 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
647 | { | |
648 | int cpu = raw_smp_processor_id(); | |
649 | ||
650 | if (!IS_ENABLED(CONFIG_SMP) || | |
651 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
652 | return false; | |
71425189 SAS |
653 | /* |
654 | * With force threaded interrupts enabled, raising softirq from an SMP | |
655 | * function call will always result in waking the ksoftirqd thread. | |
656 | * This is probably worse than completing the request on a different | |
657 | * cache domain. | |
658 | */ | |
659 | if (force_irqthreads) | |
660 | return false; | |
96339526 CH |
661 | |
662 | /* same CPU or cache domain? Complete locally */ | |
663 | if (cpu == rq->mq_ctx->cpu || | |
664 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
665 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
666 | return false; | |
667 | ||
668 | /* don't try to IPI to an offline CPU */ | |
669 | return cpu_online(rq->mq_ctx->cpu); | |
670 | } | |
671 | ||
40d09b53 | 672 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 673 | { |
af78ff7c | 674 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 675 | |
4ab32bf3 JA |
676 | /* |
677 | * For a polled request, always complete locallly, it's pointless | |
678 | * to redirect the completion. | |
679 | */ | |
40d09b53 CH |
680 | if (rq->cmd_flags & REQ_HIPRI) |
681 | return false; | |
38535201 | 682 | |
96339526 | 683 | if (blk_mq_complete_need_ipi(rq)) { |
545b8c8d | 684 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); |
96339526 | 685 | smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd); |
3d6efbf6 | 686 | } else { |
40d09b53 CH |
687 | if (rq->q->nr_hw_queues > 1) |
688 | return false; | |
689 | blk_mq_trigger_softirq(rq); | |
3d6efbf6 | 690 | } |
40d09b53 CH |
691 | |
692 | return true; | |
693 | } | |
694 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
695 | ||
696 | /** | |
697 | * blk_mq_complete_request - end I/O on a request | |
698 | * @rq: the request being processed | |
699 | * | |
700 | * Description: | |
701 | * Complete a request by scheduling the ->complete_rq operation. | |
702 | **/ | |
703 | void blk_mq_complete_request(struct request *rq) | |
704 | { | |
705 | if (!blk_mq_complete_request_remote(rq)) | |
706 | rq->q->mq_ops->complete(rq); | |
320ae51f | 707 | } |
15f73f5b | 708 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 709 | |
04ced159 | 710 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 711 | __releases(hctx->srcu) |
04ced159 JA |
712 | { |
713 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
714 | rcu_read_unlock(); | |
715 | else | |
05707b64 | 716 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
717 | } |
718 | ||
719 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 720 | __acquires(hctx->srcu) |
04ced159 | 721 | { |
08b5a6e2 JA |
722 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
723 | /* shut up gcc false positive */ | |
724 | *srcu_idx = 0; | |
04ced159 | 725 | rcu_read_lock(); |
08b5a6e2 | 726 | } else |
05707b64 | 727 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
728 | } |
729 | ||
105663f7 AA |
730 | /** |
731 | * blk_mq_start_request - Start processing a request | |
732 | * @rq: Pointer to request to be started | |
733 | * | |
734 | * Function used by device drivers to notify the block layer that a request | |
735 | * is going to be processed now, so blk layer can do proper initializations | |
736 | * such as starting the timeout timer. | |
737 | */ | |
e2490073 | 738 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
739 | { |
740 | struct request_queue *q = rq->q; | |
741 | ||
a54895fa | 742 | trace_block_rq_issue(rq); |
320ae51f | 743 | |
cf43e6be | 744 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 745 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 746 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 747 | rq->rq_flags |= RQF_STATS; |
a7905043 | 748 | rq_qos_issue(q, rq); |
cf43e6be JA |
749 | } |
750 | ||
1d9bd516 | 751 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 752 | |
1d9bd516 | 753 | blk_add_timer(rq); |
12f5b931 | 754 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 755 | |
54d4e6ab MG |
756 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
757 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
758 | q->integrity.profile->prepare_fn(rq); | |
759 | #endif | |
320ae51f | 760 | } |
e2490073 | 761 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 762 | |
ed0791b2 | 763 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
764 | { |
765 | struct request_queue *q = rq->q; | |
766 | ||
923218f6 ML |
767 | blk_mq_put_driver_tag(rq); |
768 | ||
a54895fa | 769 | trace_block_rq_requeue(rq); |
a7905043 | 770 | rq_qos_requeue(q, rq); |
49f5baa5 | 771 | |
12f5b931 KB |
772 | if (blk_mq_request_started(rq)) { |
773 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 774 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 775 | } |
320ae51f JA |
776 | } |
777 | ||
2b053aca | 778 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 779 | { |
ed0791b2 | 780 | __blk_mq_requeue_request(rq); |
ed0791b2 | 781 | |
105976f5 ML |
782 | /* this request will be re-inserted to io scheduler queue */ |
783 | blk_mq_sched_requeue_request(rq); | |
784 | ||
7d692330 | 785 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 786 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
787 | } |
788 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
789 | ||
6fca6a61 CH |
790 | static void blk_mq_requeue_work(struct work_struct *work) |
791 | { | |
792 | struct request_queue *q = | |
2849450a | 793 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
794 | LIST_HEAD(rq_list); |
795 | struct request *rq, *next; | |
6fca6a61 | 796 | |
18e9781d | 797 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 798 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 799 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
800 | |
801 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 802 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
803 | continue; |
804 | ||
e8064021 | 805 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 806 | list_del_init(&rq->queuelist); |
aef1897c JW |
807 | /* |
808 | * If RQF_DONTPREP, rq has contained some driver specific | |
809 | * data, so insert it to hctx dispatch list to avoid any | |
810 | * merge. | |
811 | */ | |
812 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 813 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
814 | else |
815 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
816 | } |
817 | ||
818 | while (!list_empty(&rq_list)) { | |
819 | rq = list_entry(rq_list.next, struct request, queuelist); | |
820 | list_del_init(&rq->queuelist); | |
9e97d295 | 821 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
822 | } |
823 | ||
52d7f1b5 | 824 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
825 | } |
826 | ||
2b053aca BVA |
827 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
828 | bool kick_requeue_list) | |
6fca6a61 CH |
829 | { |
830 | struct request_queue *q = rq->q; | |
831 | unsigned long flags; | |
832 | ||
833 | /* | |
834 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 835 | * request head insertion from the workqueue. |
6fca6a61 | 836 | */ |
e8064021 | 837 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
838 | |
839 | spin_lock_irqsave(&q->requeue_lock, flags); | |
840 | if (at_head) { | |
e8064021 | 841 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
842 | list_add(&rq->queuelist, &q->requeue_list); |
843 | } else { | |
844 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
845 | } | |
846 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
847 | |
848 | if (kick_requeue_list) | |
849 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 850 | } |
6fca6a61 CH |
851 | |
852 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
853 | { | |
ae943d20 | 854 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
855 | } |
856 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
857 | ||
2849450a MS |
858 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
859 | unsigned long msecs) | |
860 | { | |
d4acf365 BVA |
861 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
862 | msecs_to_jiffies(msecs)); | |
2849450a MS |
863 | } |
864 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
865 | ||
0e62f51f JA |
866 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
867 | { | |
88c7b2b7 JA |
868 | if (tag < tags->nr_tags) { |
869 | prefetch(tags->rqs[tag]); | |
4ee86bab | 870 | return tags->rqs[tag]; |
88c7b2b7 | 871 | } |
4ee86bab HR |
872 | |
873 | return NULL; | |
24d2f903 CH |
874 | } |
875 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
876 | ||
3c94d83c JA |
877 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
878 | void *priv, bool reserved) | |
ae879912 JA |
879 | { |
880 | /* | |
05a4fed6 | 881 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 882 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 883 | */ |
05a4fed6 | 884 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
885 | bool *busy = priv; |
886 | ||
887 | *busy = true; | |
888 | return false; | |
889 | } | |
890 | ||
891 | return true; | |
892 | } | |
893 | ||
3c94d83c | 894 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
895 | { |
896 | bool busy = false; | |
897 | ||
3c94d83c | 898 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
899 | return busy; |
900 | } | |
3c94d83c | 901 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 902 | |
358f70da | 903 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 904 | { |
da661267 | 905 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
906 | if (req->q->mq_ops->timeout) { |
907 | enum blk_eh_timer_return ret; | |
908 | ||
909 | ret = req->q->mq_ops->timeout(req, reserved); | |
910 | if (ret == BLK_EH_DONE) | |
911 | return; | |
912 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 913 | } |
d1210d5a CH |
914 | |
915 | blk_add_timer(req); | |
87ee7b11 | 916 | } |
5b3f25fc | 917 | |
12f5b931 | 918 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 919 | { |
12f5b931 | 920 | unsigned long deadline; |
87ee7b11 | 921 | |
12f5b931 KB |
922 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
923 | return false; | |
da661267 CH |
924 | if (rq->rq_flags & RQF_TIMED_OUT) |
925 | return false; | |
a7af0af3 | 926 | |
079076b3 | 927 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
928 | if (time_after_eq(jiffies, deadline)) |
929 | return true; | |
a7af0af3 | 930 | |
12f5b931 KB |
931 | if (*next == 0) |
932 | *next = deadline; | |
933 | else if (time_after(*next, deadline)) | |
934 | *next = deadline; | |
935 | return false; | |
87ee7b11 JA |
936 | } |
937 | ||
7baa8572 | 938 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
939 | struct request *rq, void *priv, bool reserved) |
940 | { | |
12f5b931 KB |
941 | unsigned long *next = priv; |
942 | ||
943 | /* | |
944 | * Just do a quick check if it is expired before locking the request in | |
945 | * so we're not unnecessarilly synchronizing across CPUs. | |
946 | */ | |
947 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 948 | return true; |
12f5b931 KB |
949 | |
950 | /* | |
951 | * We have reason to believe the request may be expired. Take a | |
952 | * reference on the request to lock this request lifetime into its | |
953 | * currently allocated context to prevent it from being reallocated in | |
954 | * the event the completion by-passes this timeout handler. | |
955 | * | |
956 | * If the reference was already released, then the driver beat the | |
957 | * timeout handler to posting a natural completion. | |
958 | */ | |
959 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 960 | return true; |
12f5b931 | 961 | |
1d9bd516 | 962 | /* |
12f5b931 KB |
963 | * The request is now locked and cannot be reallocated underneath the |
964 | * timeout handler's processing. Re-verify this exact request is truly | |
965 | * expired; if it is not expired, then the request was completed and | |
966 | * reallocated as a new request. | |
1d9bd516 | 967 | */ |
12f5b931 | 968 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 969 | blk_mq_rq_timed_out(rq, reserved); |
8d699663 YY |
970 | |
971 | if (is_flush_rq(rq, hctx)) | |
972 | rq->end_io(rq, 0); | |
973 | else if (refcount_dec_and_test(&rq->ref)) | |
12f5b931 | 974 | __blk_mq_free_request(rq); |
7baa8572 JA |
975 | |
976 | return true; | |
1d9bd516 TH |
977 | } |
978 | ||
287922eb | 979 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 980 | { |
287922eb CH |
981 | struct request_queue *q = |
982 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 983 | unsigned long next = 0; |
1d9bd516 | 984 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 985 | int i; |
320ae51f | 986 | |
71f79fb3 GKB |
987 | /* A deadlock might occur if a request is stuck requiring a |
988 | * timeout at the same time a queue freeze is waiting | |
989 | * completion, since the timeout code would not be able to | |
990 | * acquire the queue reference here. | |
991 | * | |
992 | * That's why we don't use blk_queue_enter here; instead, we use | |
993 | * percpu_ref_tryget directly, because we need to be able to | |
994 | * obtain a reference even in the short window between the queue | |
995 | * starting to freeze, by dropping the first reference in | |
1671d522 | 996 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
997 | * consumed, marked by the instant q_usage_counter reaches |
998 | * zero. | |
999 | */ | |
1000 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1001 | return; |
1002 | ||
12f5b931 | 1003 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1004 | |
12f5b931 KB |
1005 | if (next != 0) { |
1006 | mod_timer(&q->timeout, next); | |
0d2602ca | 1007 | } else { |
fcd36c36 BVA |
1008 | /* |
1009 | * Request timeouts are handled as a forward rolling timer. If | |
1010 | * we end up here it means that no requests are pending and | |
1011 | * also that no request has been pending for a while. Mark | |
1012 | * each hctx as idle. | |
1013 | */ | |
f054b56c ML |
1014 | queue_for_each_hw_ctx(q, hctx, i) { |
1015 | /* the hctx may be unmapped, so check it here */ | |
1016 | if (blk_mq_hw_queue_mapped(hctx)) | |
1017 | blk_mq_tag_idle(hctx); | |
1018 | } | |
0d2602ca | 1019 | } |
287922eb | 1020 | blk_queue_exit(q); |
320ae51f JA |
1021 | } |
1022 | ||
88459642 OS |
1023 | struct flush_busy_ctx_data { |
1024 | struct blk_mq_hw_ctx *hctx; | |
1025 | struct list_head *list; | |
1026 | }; | |
1027 | ||
1028 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1029 | { | |
1030 | struct flush_busy_ctx_data *flush_data = data; | |
1031 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1032 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1033 | enum hctx_type type = hctx->type; |
88459642 | 1034 | |
88459642 | 1035 | spin_lock(&ctx->lock); |
c16d6b5a | 1036 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1037 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1038 | spin_unlock(&ctx->lock); |
1039 | return true; | |
1040 | } | |
1041 | ||
1429d7c9 JA |
1042 | /* |
1043 | * Process software queues that have been marked busy, splicing them | |
1044 | * to the for-dispatch | |
1045 | */ | |
2c3ad667 | 1046 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1047 | { |
88459642 OS |
1048 | struct flush_busy_ctx_data data = { |
1049 | .hctx = hctx, | |
1050 | .list = list, | |
1051 | }; | |
1429d7c9 | 1052 | |
88459642 | 1053 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1054 | } |
2c3ad667 | 1055 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1056 | |
b347689f ML |
1057 | struct dispatch_rq_data { |
1058 | struct blk_mq_hw_ctx *hctx; | |
1059 | struct request *rq; | |
1060 | }; | |
1061 | ||
1062 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1063 | void *data) | |
1064 | { | |
1065 | struct dispatch_rq_data *dispatch_data = data; | |
1066 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1067 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1068 | enum hctx_type type = hctx->type; |
b347689f ML |
1069 | |
1070 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1071 | if (!list_empty(&ctx->rq_lists[type])) { |
1072 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1073 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1074 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1075 | sbitmap_clear_bit(sb, bitnr); |
1076 | } | |
1077 | spin_unlock(&ctx->lock); | |
1078 | ||
1079 | return !dispatch_data->rq; | |
1080 | } | |
1081 | ||
1082 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1083 | struct blk_mq_ctx *start) | |
1084 | { | |
f31967f0 | 1085 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1086 | struct dispatch_rq_data data = { |
1087 | .hctx = hctx, | |
1088 | .rq = NULL, | |
1089 | }; | |
1090 | ||
1091 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1092 | dispatch_rq_from_ctx, &data); | |
1093 | ||
1094 | return data.rq; | |
1095 | } | |
1096 | ||
703fd1c0 JA |
1097 | static inline unsigned int queued_to_index(unsigned int queued) |
1098 | { | |
1099 | if (!queued) | |
1100 | return 0; | |
1429d7c9 | 1101 | |
703fd1c0 | 1102 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1103 | } |
1104 | ||
570e9b73 ML |
1105 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1106 | { | |
222a5ae0 | 1107 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1108 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1109 | int tag; |
1110 | ||
568f2700 ML |
1111 | blk_mq_tag_busy(rq->mq_hctx); |
1112 | ||
570e9b73 | 1113 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
222a5ae0 | 1114 | bt = rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1115 | tag_offset = 0; |
28500850 ML |
1116 | } else { |
1117 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1118 | return false; | |
570e9b73 ML |
1119 | } |
1120 | ||
570e9b73 ML |
1121 | tag = __sbitmap_queue_get(bt); |
1122 | if (tag == BLK_MQ_NO_TAG) | |
1123 | return false; | |
1124 | ||
1125 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1126 | return true; |
1127 | } | |
1128 | ||
1129 | static bool blk_mq_get_driver_tag(struct request *rq) | |
1130 | { | |
568f2700 ML |
1131 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1132 | ||
1133 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1134 | return false; | |
1135 | ||
51db1c37 | 1136 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1137 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1138 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1139 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1140 | } |
1141 | hctx->tags->rqs[rq->tag] = rq; | |
1142 | return true; | |
570e9b73 ML |
1143 | } |
1144 | ||
eb619fdb JA |
1145 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1146 | int flags, void *key) | |
da55f2cc OS |
1147 | { |
1148 | struct blk_mq_hw_ctx *hctx; | |
1149 | ||
1150 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1151 | ||
5815839b | 1152 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1153 | if (!list_empty(&wait->entry)) { |
1154 | struct sbitmap_queue *sbq; | |
1155 | ||
1156 | list_del_init(&wait->entry); | |
222a5ae0 | 1157 | sbq = hctx->tags->bitmap_tags; |
e8618575 JA |
1158 | atomic_dec(&sbq->ws_active); |
1159 | } | |
5815839b ML |
1160 | spin_unlock(&hctx->dispatch_wait_lock); |
1161 | ||
da55f2cc OS |
1162 | blk_mq_run_hw_queue(hctx, true); |
1163 | return 1; | |
1164 | } | |
1165 | ||
f906a6a0 JA |
1166 | /* |
1167 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1168 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1169 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1170 | * marking us as waiting. |
1171 | */ | |
2278d69f | 1172 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1173 | struct request *rq) |
da55f2cc | 1174 | { |
222a5ae0 | 1175 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; |
5815839b | 1176 | struct wait_queue_head *wq; |
f906a6a0 JA |
1177 | wait_queue_entry_t *wait; |
1178 | bool ret; | |
da55f2cc | 1179 | |
51db1c37 | 1180 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1181 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1182 | |
c27d53fb BVA |
1183 | /* |
1184 | * It's possible that a tag was freed in the window between the | |
1185 | * allocation failure and adding the hardware queue to the wait | |
1186 | * queue. | |
1187 | * | |
1188 | * Don't clear RESTART here, someone else could have set it. | |
1189 | * At most this will cost an extra queue run. | |
1190 | */ | |
8ab6bb9e | 1191 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1192 | } |
eb619fdb | 1193 | |
2278d69f | 1194 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1195 | if (!list_empty_careful(&wait->entry)) |
1196 | return false; | |
1197 | ||
e8618575 | 1198 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1199 | |
1200 | spin_lock_irq(&wq->lock); | |
1201 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1202 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1203 | spin_unlock(&hctx->dispatch_wait_lock); |
1204 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1205 | return false; |
eb619fdb JA |
1206 | } |
1207 | ||
e8618575 | 1208 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1209 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1210 | __add_wait_queue(wq, wait); | |
c27d53fb | 1211 | |
da55f2cc | 1212 | /* |
eb619fdb JA |
1213 | * It's possible that a tag was freed in the window between the |
1214 | * allocation failure and adding the hardware queue to the wait | |
1215 | * queue. | |
da55f2cc | 1216 | */ |
8ab6bb9e | 1217 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1218 | if (!ret) { |
5815839b ML |
1219 | spin_unlock(&hctx->dispatch_wait_lock); |
1220 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1221 | return false; |
eb619fdb | 1222 | } |
c27d53fb BVA |
1223 | |
1224 | /* | |
1225 | * We got a tag, remove ourselves from the wait queue to ensure | |
1226 | * someone else gets the wakeup. | |
1227 | */ | |
c27d53fb | 1228 | list_del_init(&wait->entry); |
e8618575 | 1229 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1230 | spin_unlock(&hctx->dispatch_wait_lock); |
1231 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1232 | |
1233 | return true; | |
da55f2cc OS |
1234 | } |
1235 | ||
6e768717 ML |
1236 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1237 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1238 | /* | |
1239 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1240 | * - EWMA is one simple way to compute running average value | |
1241 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1242 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1243 | * factor doesn't matter because EWMA decreases exponentially | |
1244 | */ | |
1245 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1246 | { | |
1247 | unsigned int ewma; | |
1248 | ||
1249 | if (hctx->queue->elevator) | |
1250 | return; | |
1251 | ||
1252 | ewma = hctx->dispatch_busy; | |
1253 | ||
1254 | if (!ewma && !busy) | |
1255 | return; | |
1256 | ||
1257 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1258 | if (busy) | |
1259 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1260 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1261 | ||
1262 | hctx->dispatch_busy = ewma; | |
1263 | } | |
1264 | ||
86ff7c2a ML |
1265 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1266 | ||
c92a4103 JT |
1267 | static void blk_mq_handle_dev_resource(struct request *rq, |
1268 | struct list_head *list) | |
1269 | { | |
1270 | struct request *next = | |
1271 | list_first_entry_or_null(list, struct request, queuelist); | |
1272 | ||
1273 | /* | |
1274 | * If an I/O scheduler has been configured and we got a driver tag for | |
1275 | * the next request already, free it. | |
1276 | */ | |
1277 | if (next) | |
1278 | blk_mq_put_driver_tag(next); | |
1279 | ||
1280 | list_add(&rq->queuelist, list); | |
1281 | __blk_mq_requeue_request(rq); | |
1282 | } | |
1283 | ||
0512a75b KB |
1284 | static void blk_mq_handle_zone_resource(struct request *rq, |
1285 | struct list_head *zone_list) | |
1286 | { | |
1287 | /* | |
1288 | * If we end up here it is because we cannot dispatch a request to a | |
1289 | * specific zone due to LLD level zone-write locking or other zone | |
1290 | * related resource not being available. In this case, set the request | |
1291 | * aside in zone_list for retrying it later. | |
1292 | */ | |
1293 | list_add(&rq->queuelist, zone_list); | |
1294 | __blk_mq_requeue_request(rq); | |
1295 | } | |
1296 | ||
75383524 ML |
1297 | enum prep_dispatch { |
1298 | PREP_DISPATCH_OK, | |
1299 | PREP_DISPATCH_NO_TAG, | |
1300 | PREP_DISPATCH_NO_BUDGET, | |
1301 | }; | |
1302 | ||
1303 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1304 | bool need_budget) | |
1305 | { | |
1306 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1307 | ||
1308 | if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) { | |
1309 | blk_mq_put_driver_tag(rq); | |
1310 | return PREP_DISPATCH_NO_BUDGET; | |
1311 | } | |
1312 | ||
1313 | if (!blk_mq_get_driver_tag(rq)) { | |
1314 | /* | |
1315 | * The initial allocation attempt failed, so we need to | |
1316 | * rerun the hardware queue when a tag is freed. The | |
1317 | * waitqueue takes care of that. If the queue is run | |
1318 | * before we add this entry back on the dispatch list, | |
1319 | * we'll re-run it below. | |
1320 | */ | |
1321 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1322 | /* |
1323 | * All budgets not got from this function will be put | |
1324 | * together during handling partial dispatch | |
1325 | */ | |
1326 | if (need_budget) | |
1327 | blk_mq_put_dispatch_budget(rq->q); | |
75383524 ML |
1328 | return PREP_DISPATCH_NO_TAG; |
1329 | } | |
1330 | } | |
1331 | ||
1332 | return PREP_DISPATCH_OK; | |
1333 | } | |
1334 | ||
1fd40b5e ML |
1335 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1336 | static void blk_mq_release_budgets(struct request_queue *q, | |
1337 | unsigned int nr_budgets) | |
1338 | { | |
1339 | int i; | |
1340 | ||
1341 | for (i = 0; i < nr_budgets; i++) | |
1342 | blk_mq_put_dispatch_budget(q); | |
1343 | } | |
1344 | ||
1f57f8d4 JA |
1345 | /* |
1346 | * Returns true if we did some work AND can potentially do more. | |
1347 | */ | |
445874e8 | 1348 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1349 | unsigned int nr_budgets) |
320ae51f | 1350 | { |
75383524 | 1351 | enum prep_dispatch prep; |
445874e8 | 1352 | struct request_queue *q = hctx->queue; |
6d6f167c | 1353 | struct request *rq, *nxt; |
fc17b653 | 1354 | int errors, queued; |
86ff7c2a | 1355 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1356 | LIST_HEAD(zone_list); |
320ae51f | 1357 | |
81380ca1 OS |
1358 | if (list_empty(list)) |
1359 | return false; | |
1360 | ||
320ae51f JA |
1361 | /* |
1362 | * Now process all the entries, sending them to the driver. | |
1363 | */ | |
93efe981 | 1364 | errors = queued = 0; |
81380ca1 | 1365 | do { |
74c45052 | 1366 | struct blk_mq_queue_data bd; |
320ae51f | 1367 | |
f04c3df3 | 1368 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1369 | |
445874e8 | 1370 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1371 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1372 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1373 | break; |
de148297 | 1374 | |
320ae51f | 1375 | list_del_init(&rq->queuelist); |
320ae51f | 1376 | |
74c45052 | 1377 | bd.rq = rq; |
113285b4 JA |
1378 | |
1379 | /* | |
1380 | * Flag last if we have no more requests, or if we have more | |
1381 | * but can't assign a driver tag to it. | |
1382 | */ | |
1383 | if (list_empty(list)) | |
1384 | bd.last = true; | |
1385 | else { | |
113285b4 | 1386 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1387 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1388 | } |
74c45052 | 1389 | |
1fd40b5e ML |
1390 | /* |
1391 | * once the request is queued to lld, no need to cover the | |
1392 | * budget any more | |
1393 | */ | |
1394 | if (nr_budgets) | |
1395 | nr_budgets--; | |
74c45052 | 1396 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1397 | switch (ret) { |
1398 | case BLK_STS_OK: | |
1399 | queued++; | |
320ae51f | 1400 | break; |
7bf13729 ML |
1401 | case BLK_STS_RESOURCE: |
1402 | case BLK_STS_DEV_RESOURCE: | |
1403 | blk_mq_handle_dev_resource(rq, list); | |
1404 | goto out; | |
1405 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1406 | /* |
1407 | * Move the request to zone_list and keep going through | |
1408 | * the dispatch list to find more requests the drive can | |
1409 | * accept. | |
1410 | */ | |
1411 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1412 | break; |
1413 | default: | |
93efe981 | 1414 | errors++; |
e21ee5a6 | 1415 | blk_mq_end_request(rq, ret); |
320ae51f | 1416 | } |
81380ca1 | 1417 | } while (!list_empty(list)); |
7bf13729 | 1418 | out: |
0512a75b KB |
1419 | if (!list_empty(&zone_list)) |
1420 | list_splice_tail_init(&zone_list, list); | |
1421 | ||
703fd1c0 | 1422 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1423 | |
632bfb63 | 1424 | /* If we didn't flush the entire list, we could have told the driver |
1425 | * there was more coming, but that turned out to be a lie. | |
1426 | */ | |
1427 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1428 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1429 | /* |
1430 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1431 | * that is where we will continue on next queue run. | |
1432 | */ | |
f04c3df3 | 1433 | if (!list_empty(list)) { |
86ff7c2a | 1434 | bool needs_restart; |
75383524 ML |
1435 | /* For non-shared tags, the RESTART check will suffice */ |
1436 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1437 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1438 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1439 | |
1fd40b5e | 1440 | blk_mq_release_budgets(q, nr_budgets); |
86ff7c2a | 1441 | |
320ae51f | 1442 | spin_lock(&hctx->lock); |
01e99aec | 1443 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1444 | spin_unlock(&hctx->lock); |
f04c3df3 | 1445 | |
d7d8535f ML |
1446 | /* |
1447 | * Order adding requests to hctx->dispatch and checking | |
1448 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1449 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1450 | * miss the new added requests to hctx->dispatch, meantime | |
1451 | * SCHED_RESTART is observed here. | |
1452 | */ | |
1453 | smp_mb(); | |
1454 | ||
9ba52e58 | 1455 | /* |
710c785f BVA |
1456 | * If SCHED_RESTART was set by the caller of this function and |
1457 | * it is no longer set that means that it was cleared by another | |
1458 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1459 | * |
eb619fdb JA |
1460 | * If 'no_tag' is set, that means that we failed getting |
1461 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1462 | * waitqueue is no longer active, ensure that we run the queue | |
1463 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1464 | * |
710c785f BVA |
1465 | * If no I/O scheduler has been configured it is possible that |
1466 | * the hardware queue got stopped and restarted before requests | |
1467 | * were pushed back onto the dispatch list. Rerun the queue to | |
1468 | * avoid starvation. Notes: | |
1469 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1470 | * been stopped before rerunning a queue. | |
1471 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1472 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1473 | * and dm-rq. |
86ff7c2a ML |
1474 | * |
1475 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1476 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1477 | * that could otherwise occur if the queue is idle. We'll do |
1478 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1479 | */ |
86ff7c2a ML |
1480 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1481 | if (!needs_restart || | |
eb619fdb | 1482 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1483 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1484 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1485 | no_budget_avail)) | |
86ff7c2a | 1486 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1487 | |
6e768717 | 1488 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1489 | return false; |
6e768717 ML |
1490 | } else |
1491 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1492 | |
93efe981 | 1493 | return (queued + errors) != 0; |
f04c3df3 JA |
1494 | } |
1495 | ||
105663f7 AA |
1496 | /** |
1497 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1498 | * @hctx: Pointer to the hardware queue to run. | |
1499 | * | |
1500 | * Send pending requests to the hardware. | |
1501 | */ | |
6a83e74d BVA |
1502 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1503 | { | |
1504 | int srcu_idx; | |
1505 | ||
b7a71e66 JA |
1506 | /* |
1507 | * We can't run the queue inline with ints disabled. Ensure that | |
1508 | * we catch bad users of this early. | |
1509 | */ | |
1510 | WARN_ON_ONCE(in_interrupt()); | |
1511 | ||
04ced159 | 1512 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1513 | |
04ced159 JA |
1514 | hctx_lock(hctx, &srcu_idx); |
1515 | blk_mq_sched_dispatch_requests(hctx); | |
1516 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1517 | } |
1518 | ||
f82ddf19 ML |
1519 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1520 | { | |
1521 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1522 | ||
1523 | if (cpu >= nr_cpu_ids) | |
1524 | cpu = cpumask_first(hctx->cpumask); | |
1525 | return cpu; | |
1526 | } | |
1527 | ||
506e931f JA |
1528 | /* |
1529 | * It'd be great if the workqueue API had a way to pass | |
1530 | * in a mask and had some smarts for more clever placement. | |
1531 | * For now we just round-robin here, switching for every | |
1532 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1533 | */ | |
1534 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1535 | { | |
7bed4595 | 1536 | bool tried = false; |
476f8c98 | 1537 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1538 | |
b657d7e6 CH |
1539 | if (hctx->queue->nr_hw_queues == 1) |
1540 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1541 | |
1542 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1543 | select_cpu: |
476f8c98 | 1544 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1545 | cpu_online_mask); |
506e931f | 1546 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1547 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1548 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1549 | } | |
1550 | ||
7bed4595 ML |
1551 | /* |
1552 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1553 | * and it should only happen in the path of handling CPU DEAD. | |
1554 | */ | |
476f8c98 | 1555 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1556 | if (!tried) { |
1557 | tried = true; | |
1558 | goto select_cpu; | |
1559 | } | |
1560 | ||
1561 | /* | |
1562 | * Make sure to re-select CPU next time once after CPUs | |
1563 | * in hctx->cpumask become online again. | |
1564 | */ | |
476f8c98 | 1565 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1566 | hctx->next_cpu_batch = 1; |
1567 | return WORK_CPU_UNBOUND; | |
1568 | } | |
476f8c98 ML |
1569 | |
1570 | hctx->next_cpu = next_cpu; | |
1571 | return next_cpu; | |
506e931f JA |
1572 | } |
1573 | ||
105663f7 AA |
1574 | /** |
1575 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1576 | * @hctx: Pointer to the hardware queue to run. | |
1577 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1578 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1579 | * |
1580 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1581 | * with a delay of @msecs. | |
1582 | */ | |
7587a5ae BVA |
1583 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1584 | unsigned long msecs) | |
320ae51f | 1585 | { |
5435c023 | 1586 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1587 | return; |
1588 | ||
1b792f2f | 1589 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1590 | int cpu = get_cpu(); |
1591 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1592 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1593 | put_cpu(); |
398205b8 PB |
1594 | return; |
1595 | } | |
e4043dcf | 1596 | |
2a90d4aa | 1597 | put_cpu(); |
e4043dcf | 1598 | } |
398205b8 | 1599 | |
ae943d20 BVA |
1600 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1601 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1602 | } |
1603 | ||
105663f7 AA |
1604 | /** |
1605 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1606 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1607 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1608 | * |
1609 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1610 | */ | |
7587a5ae BVA |
1611 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1612 | { | |
1613 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1614 | } | |
1615 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1616 | ||
105663f7 AA |
1617 | /** |
1618 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1619 | * @hctx: Pointer to the hardware queue to run. | |
1620 | * @async: If we want to run the queue asynchronously. | |
1621 | * | |
1622 | * Check if the request queue is not in a quiesced state and if there are | |
1623 | * pending requests to be sent. If this is true, run the queue to send requests | |
1624 | * to hardware. | |
1625 | */ | |
626fb735 | 1626 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1627 | { |
24f5a90f ML |
1628 | int srcu_idx; |
1629 | bool need_run; | |
1630 | ||
1631 | /* | |
1632 | * When queue is quiesced, we may be switching io scheduler, or | |
1633 | * updating nr_hw_queues, or other things, and we can't run queue | |
1634 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1635 | * | |
1636 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1637 | * quiesced. | |
1638 | */ | |
04ced159 JA |
1639 | hctx_lock(hctx, &srcu_idx); |
1640 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1641 | blk_mq_hctx_has_pending(hctx); | |
1642 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1643 | |
626fb735 | 1644 | if (need_run) |
79f720a7 | 1645 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1646 | } |
5b727272 | 1647 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1648 | |
105663f7 | 1649 | /** |
24f7bb88 | 1650 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1651 | * @q: Pointer to the request queue to run. |
1652 | * @async: If we want to run the queue asynchronously. | |
1653 | */ | |
b94ec296 | 1654 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1655 | { |
1656 | struct blk_mq_hw_ctx *hctx; | |
1657 | int i; | |
1658 | ||
1659 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1660 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1661 | continue; |
1662 | ||
b94ec296 | 1663 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1664 | } |
1665 | } | |
b94ec296 | 1666 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1667 | |
b9151e7b DA |
1668 | /** |
1669 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1670 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1671 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1672 | */ |
1673 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1674 | { | |
1675 | struct blk_mq_hw_ctx *hctx; | |
1676 | int i; | |
1677 | ||
1678 | queue_for_each_hw_ctx(q, hctx, i) { | |
1679 | if (blk_mq_hctx_stopped(hctx)) | |
1680 | continue; | |
1681 | ||
1682 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
1683 | } | |
1684 | } | |
1685 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1686 | ||
fd001443 BVA |
1687 | /** |
1688 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1689 | * @q: request queue. | |
1690 | * | |
1691 | * The caller is responsible for serializing this function against | |
1692 | * blk_mq_{start,stop}_hw_queue(). | |
1693 | */ | |
1694 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1695 | { | |
1696 | struct blk_mq_hw_ctx *hctx; | |
1697 | int i; | |
1698 | ||
1699 | queue_for_each_hw_ctx(q, hctx, i) | |
1700 | if (blk_mq_hctx_stopped(hctx)) | |
1701 | return true; | |
1702 | ||
1703 | return false; | |
1704 | } | |
1705 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1706 | ||
39a70c76 ML |
1707 | /* |
1708 | * This function is often used for pausing .queue_rq() by driver when | |
1709 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1710 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1711 | * |
1712 | * We do not guarantee that dispatch can be drained or blocked | |
1713 | * after blk_mq_stop_hw_queue() returns. Please use | |
1714 | * blk_mq_quiesce_queue() for that requirement. | |
1715 | */ | |
2719aa21 JA |
1716 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1717 | { | |
641a9ed6 | 1718 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1719 | |
641a9ed6 | 1720 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1721 | } |
641a9ed6 | 1722 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1723 | |
39a70c76 ML |
1724 | /* |
1725 | * This function is often used for pausing .queue_rq() by driver when | |
1726 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1727 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1728 | * |
1729 | * We do not guarantee that dispatch can be drained or blocked | |
1730 | * after blk_mq_stop_hw_queues() returns. Please use | |
1731 | * blk_mq_quiesce_queue() for that requirement. | |
1732 | */ | |
2719aa21 JA |
1733 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1734 | { | |
641a9ed6 ML |
1735 | struct blk_mq_hw_ctx *hctx; |
1736 | int i; | |
1737 | ||
1738 | queue_for_each_hw_ctx(q, hctx, i) | |
1739 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1740 | } |
1741 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1742 | ||
320ae51f JA |
1743 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1744 | { | |
1745 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1746 | |
0ffbce80 | 1747 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1748 | } |
1749 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1750 | ||
2f268556 CH |
1751 | void blk_mq_start_hw_queues(struct request_queue *q) |
1752 | { | |
1753 | struct blk_mq_hw_ctx *hctx; | |
1754 | int i; | |
1755 | ||
1756 | queue_for_each_hw_ctx(q, hctx, i) | |
1757 | blk_mq_start_hw_queue(hctx); | |
1758 | } | |
1759 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1760 | ||
ae911c5e JA |
1761 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1762 | { | |
1763 | if (!blk_mq_hctx_stopped(hctx)) | |
1764 | return; | |
1765 | ||
1766 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1767 | blk_mq_run_hw_queue(hctx, async); | |
1768 | } | |
1769 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1770 | ||
1b4a3258 | 1771 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1772 | { |
1773 | struct blk_mq_hw_ctx *hctx; | |
1774 | int i; | |
1775 | ||
ae911c5e JA |
1776 | queue_for_each_hw_ctx(q, hctx, i) |
1777 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1778 | } |
1779 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1780 | ||
70f4db63 | 1781 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1782 | { |
1783 | struct blk_mq_hw_ctx *hctx; | |
1784 | ||
9f993737 | 1785 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1786 | |
21c6e939 | 1787 | /* |
15fe8a90 | 1788 | * If we are stopped, don't run the queue. |
21c6e939 | 1789 | */ |
0841031a | 1790 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 1791 | return; |
7587a5ae BVA |
1792 | |
1793 | __blk_mq_run_hw_queue(hctx); | |
1794 | } | |
1795 | ||
cfd0c552 | 1796 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1797 | struct request *rq, |
1798 | bool at_head) | |
320ae51f | 1799 | { |
e57690fe | 1800 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1801 | enum hctx_type type = hctx->type; |
e57690fe | 1802 | |
7b607814 BVA |
1803 | lockdep_assert_held(&ctx->lock); |
1804 | ||
a54895fa | 1805 | trace_block_rq_insert(rq); |
01b983c9 | 1806 | |
72a0a36e | 1807 | if (at_head) |
c16d6b5a | 1808 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1809 | else |
c16d6b5a | 1810 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1811 | } |
4bb659b1 | 1812 | |
2c3ad667 JA |
1813 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1814 | bool at_head) | |
cfd0c552 ML |
1815 | { |
1816 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1817 | ||
7b607814 BVA |
1818 | lockdep_assert_held(&ctx->lock); |
1819 | ||
e57690fe | 1820 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1821 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1822 | } |
1823 | ||
105663f7 AA |
1824 | /** |
1825 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1826 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 1827 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
1828 | * @run_queue: If we should run the hardware queue after inserting the request. |
1829 | * | |
157f377b JA |
1830 | * Should only be used carefully, when the caller knows we want to |
1831 | * bypass a potential IO scheduler on the target device. | |
1832 | */ | |
01e99aec ML |
1833 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1834 | bool run_queue) | |
157f377b | 1835 | { |
ea4f995e | 1836 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1837 | |
1838 | spin_lock(&hctx->lock); | |
01e99aec ML |
1839 | if (at_head) |
1840 | list_add(&rq->queuelist, &hctx->dispatch); | |
1841 | else | |
1842 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1843 | spin_unlock(&hctx->lock); |
1844 | ||
b0850297 ML |
1845 | if (run_queue) |
1846 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1847 | } |
1848 | ||
bd166ef1 JA |
1849 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1850 | struct list_head *list) | |
320ae51f JA |
1851 | |
1852 | { | |
3f0cedc7 | 1853 | struct request *rq; |
c16d6b5a | 1854 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1855 | |
320ae51f JA |
1856 | /* |
1857 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1858 | * offline now | |
1859 | */ | |
3f0cedc7 | 1860 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1861 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 1862 | trace_block_rq_insert(rq); |
320ae51f | 1863 | } |
3f0cedc7 ML |
1864 | |
1865 | spin_lock(&ctx->lock); | |
c16d6b5a | 1866 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1867 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1868 | spin_unlock(&ctx->lock); |
320ae51f JA |
1869 | } |
1870 | ||
3110fc79 | 1871 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1872 | { |
1873 | struct request *rqa = container_of(a, struct request, queuelist); | |
1874 | struct request *rqb = container_of(b, struct request, queuelist); | |
1875 | ||
7d30a621 PB |
1876 | if (rqa->mq_ctx != rqb->mq_ctx) |
1877 | return rqa->mq_ctx > rqb->mq_ctx; | |
1878 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1879 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1880 | |
1881 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1882 | } |
1883 | ||
1884 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1885 | { | |
320ae51f | 1886 | LIST_HEAD(list); |
320ae51f | 1887 | |
95ed0c5b PB |
1888 | if (list_empty(&plug->mq_list)) |
1889 | return; | |
320ae51f JA |
1890 | list_splice_init(&plug->mq_list, &list); |
1891 | ||
ce5b009c JA |
1892 | if (plug->rq_count > 2 && plug->multiple_queues) |
1893 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1894 | |
bcc816df DZ |
1895 | plug->rq_count = 0; |
1896 | ||
95ed0c5b PB |
1897 | do { |
1898 | struct list_head rq_list; | |
1899 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1900 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1901 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1902 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1903 | unsigned int depth = 1; | |
1904 | ||
1905 | list_for_each_continue(pos, &list) { | |
1906 | rq = list_entry_rq(pos); | |
1907 | BUG_ON(!rq->q); | |
1908 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1909 | break; | |
1910 | depth++; | |
320ae51f JA |
1911 | } |
1912 | ||
95ed0c5b PB |
1913 | list_cut_before(&rq_list, &list, pos); |
1914 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1915 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1916 | from_schedule); |
95ed0c5b | 1917 | } while(!list_empty(&list)); |
320ae51f JA |
1918 | } |
1919 | ||
14ccb66b CH |
1920 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1921 | unsigned int nr_segs) | |
320ae51f | 1922 | { |
93f221ae EB |
1923 | int err; |
1924 | ||
f924cdde CH |
1925 | if (bio->bi_opf & REQ_RAHEAD) |
1926 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1927 | ||
1928 | rq->__sector = bio->bi_iter.bi_sector; | |
1929 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1930 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1931 | |
1932 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
1933 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1934 | WARN_ON_ONCE(err); | |
4b570521 | 1935 | |
b5af37ab | 1936 | blk_account_io_start(rq); |
320ae51f JA |
1937 | } |
1938 | ||
0f95549c MS |
1939 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1940 | struct request *rq, | |
be94f058 | 1941 | blk_qc_t *cookie, bool last) |
f984df1f | 1942 | { |
f984df1f | 1943 | struct request_queue *q = rq->q; |
f984df1f SL |
1944 | struct blk_mq_queue_data bd = { |
1945 | .rq = rq, | |
be94f058 | 1946 | .last = last, |
f984df1f | 1947 | }; |
bd166ef1 | 1948 | blk_qc_t new_cookie; |
f06345ad | 1949 | blk_status_t ret; |
0f95549c MS |
1950 | |
1951 | new_cookie = request_to_qc_t(hctx, rq); | |
1952 | ||
1953 | /* | |
1954 | * For OK queue, we are done. For error, caller may kill it. | |
1955 | * Any other error (busy), just add it to our list as we | |
1956 | * previously would have done. | |
1957 | */ | |
1958 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1959 | switch (ret) { | |
1960 | case BLK_STS_OK: | |
6ce3dd6e | 1961 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1962 | *cookie = new_cookie; |
1963 | break; | |
1964 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1965 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1966 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1967 | __blk_mq_requeue_request(rq); |
1968 | break; | |
1969 | default: | |
6ce3dd6e | 1970 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1971 | *cookie = BLK_QC_T_NONE; |
1972 | break; | |
1973 | } | |
1974 | ||
1975 | return ret; | |
1976 | } | |
1977 | ||
fd9c40f6 | 1978 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1979 | struct request *rq, |
396eaf21 | 1980 | blk_qc_t *cookie, |
fd9c40f6 | 1981 | bool bypass_insert, bool last) |
0f95549c MS |
1982 | { |
1983 | struct request_queue *q = rq->q; | |
d964f04a ML |
1984 | bool run_queue = true; |
1985 | ||
23d4ee19 | 1986 | /* |
fd9c40f6 | 1987 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1988 | * |
fd9c40f6 BVA |
1989 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1990 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1991 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1992 | */ |
fd9c40f6 | 1993 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1994 | run_queue = false; |
fd9c40f6 BVA |
1995 | bypass_insert = false; |
1996 | goto insert; | |
d964f04a | 1997 | } |
f984df1f | 1998 | |
fd9c40f6 BVA |
1999 | if (q->elevator && !bypass_insert) |
2000 | goto insert; | |
2253efc8 | 2001 | |
65c76369 | 2002 | if (!blk_mq_get_dispatch_budget(q)) |
fd9c40f6 | 2003 | goto insert; |
bd166ef1 | 2004 | |
8ab6bb9e | 2005 | if (!blk_mq_get_driver_tag(rq)) { |
65c76369 | 2006 | blk_mq_put_dispatch_budget(q); |
fd9c40f6 | 2007 | goto insert; |
88022d72 | 2008 | } |
de148297 | 2009 | |
fd9c40f6 BVA |
2010 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2011 | insert: | |
2012 | if (bypass_insert) | |
2013 | return BLK_STS_RESOURCE; | |
2014 | ||
db03f88f ML |
2015 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2016 | ||
fd9c40f6 BVA |
2017 | return BLK_STS_OK; |
2018 | } | |
2019 | ||
105663f7 AA |
2020 | /** |
2021 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2022 | * @hctx: Pointer of the associated hardware queue. | |
2023 | * @rq: Pointer to request to be sent. | |
2024 | * @cookie: Request queue cookie. | |
2025 | * | |
2026 | * If the device has enough resources to accept a new request now, send the | |
2027 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2028 | * we can try send it another time in the future. Requests inserted at this | |
2029 | * queue have higher priority. | |
2030 | */ | |
fd9c40f6 BVA |
2031 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2032 | struct request *rq, blk_qc_t *cookie) | |
2033 | { | |
2034 | blk_status_t ret; | |
2035 | int srcu_idx; | |
2036 | ||
2037 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2038 | ||
2039 | hctx_lock(hctx, &srcu_idx); | |
2040 | ||
2041 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2042 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2043 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2044 | else if (ret != BLK_STS_OK) |
2045 | blk_mq_end_request(rq, ret); | |
2046 | ||
2047 | hctx_unlock(hctx, srcu_idx); | |
2048 | } | |
2049 | ||
2050 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2051 | { | |
2052 | blk_status_t ret; | |
2053 | int srcu_idx; | |
2054 | blk_qc_t unused_cookie; | |
2055 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2056 | ||
2057 | hctx_lock(hctx, &srcu_idx); | |
2058 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2059 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2060 | |
2061 | return ret; | |
5eb6126e CH |
2062 | } |
2063 | ||
6ce3dd6e ML |
2064 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2065 | struct list_head *list) | |
2066 | { | |
536167d4 | 2067 | int queued = 0; |
632bfb63 | 2068 | int errors = 0; |
536167d4 | 2069 | |
6ce3dd6e | 2070 | while (!list_empty(list)) { |
fd9c40f6 | 2071 | blk_status_t ret; |
6ce3dd6e ML |
2072 | struct request *rq = list_first_entry(list, struct request, |
2073 | queuelist); | |
2074 | ||
2075 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2076 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2077 | if (ret != BLK_STS_OK) { | |
2078 | if (ret == BLK_STS_RESOURCE || | |
2079 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2080 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2081 | list_empty(list)); |
fd9c40f6 BVA |
2082 | break; |
2083 | } | |
2084 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2085 | errors++; |
536167d4 KB |
2086 | } else |
2087 | queued++; | |
6ce3dd6e | 2088 | } |
d666ba98 JA |
2089 | |
2090 | /* | |
2091 | * If we didn't flush the entire list, we could have told | |
2092 | * the driver there was more coming, but that turned out to | |
2093 | * be a lie. | |
2094 | */ | |
632bfb63 | 2095 | if ((!list_empty(list) || errors) && |
2096 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2097 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2098 | } |
2099 | ||
ce5b009c JA |
2100 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2101 | { | |
2102 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2103 | plug->rq_count++; | |
2104 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2105 | struct request *tmp; | |
2106 | ||
2107 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2108 | queuelist); | |
2109 | if (tmp->q != rq->q) | |
2110 | plug->multiple_queues = true; | |
2111 | } | |
2112 | } | |
2113 | ||
105663f7 | 2114 | /** |
c62b37d9 | 2115 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2116 | * @bio: Bio pointer. |
2117 | * | |
2118 | * Builds up a request structure from @q and @bio and send to the device. The | |
2119 | * request may not be queued directly to hardware if: | |
2120 | * * This request can be merged with another one | |
2121 | * * We want to place request at plug queue for possible future merging | |
2122 | * * There is an IO scheduler active at this queue | |
2123 | * | |
2124 | * It will not queue the request if there is an error with the bio, or at the | |
2125 | * request creation. | |
2126 | * | |
2127 | * Returns: Request queue cookie. | |
2128 | */ | |
c62b37d9 | 2129 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2130 | { |
c62b37d9 | 2131 | struct request_queue *q = bio->bi_disk->queue; |
ef295ecf | 2132 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2133 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6e7abff CH |
2134 | struct blk_mq_alloc_data data = { |
2135 | .q = q, | |
2136 | }; | |
07068d5b | 2137 | struct request *rq; |
f984df1f | 2138 | struct blk_plug *plug; |
5b3f341f | 2139 | struct request *same_queue_rq = NULL; |
14ccb66b | 2140 | unsigned int nr_segs; |
7b371636 | 2141 | blk_qc_t cookie; |
a892c8d5 | 2142 | blk_status_t ret; |
cc29e1bf | 2143 | bool hipri; |
07068d5b JA |
2144 | |
2145 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2146 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2147 | |
e23947bd | 2148 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2149 | goto queue_exit; |
07068d5b | 2150 | |
87c279e6 | 2151 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2152 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2153 | goto queue_exit; |
f984df1f | 2154 | |
14ccb66b | 2155 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2156 | goto queue_exit; |
bd166ef1 | 2157 | |
d5337560 | 2158 | rq_qos_throttle(q, bio); |
87760e5e | 2159 | |
cc29e1bf JX |
2160 | hipri = bio->bi_opf & REQ_HIPRI; |
2161 | ||
7809167d | 2162 | data.cmd_flags = bio->bi_opf; |
e6e7abff | 2163 | rq = __blk_mq_alloc_request(&data); |
87760e5e | 2164 | if (unlikely(!rq)) { |
c1c80384 | 2165 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2166 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2167 | bio_wouldblock_error(bio); |
ac7c5675 | 2168 | goto queue_exit; |
87760e5e JA |
2169 | } |
2170 | ||
e8a676d6 | 2171 | trace_block_getrq(bio); |
d6f1dda2 | 2172 | |
c1c80384 | 2173 | rq_qos_track(q, rq, bio); |
07068d5b | 2174 | |
fd2d3326 | 2175 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2176 | |
970d168d BVA |
2177 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2178 | ||
a892c8d5 ST |
2179 | ret = blk_crypto_init_request(rq); |
2180 | if (ret != BLK_STS_OK) { | |
2181 | bio->bi_status = ret; | |
2182 | bio_endio(bio); | |
2183 | blk_mq_free_request(rq); | |
2184 | return BLK_QC_T_NONE; | |
2185 | } | |
2186 | ||
b49773e7 | 2187 | plug = blk_mq_plug(q, bio); |
07068d5b | 2188 | if (unlikely(is_flush_fua)) { |
105663f7 | 2189 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2190 | blk_insert_flush(rq); |
2191 | blk_mq_run_hw_queue(data.hctx, true); | |
3154df26 ML |
2192 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs || |
2193 | !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2194 | /* |
2195 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2196 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2197 | * |
2198 | * Use normal plugging if this disk is slow HDD, as sequential | |
2199 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2200 | */ |
5f0ed774 | 2201 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2202 | struct request *last = NULL; |
2203 | ||
676d0607 | 2204 | if (!request_count) |
e6c4438b | 2205 | trace_block_plug(q); |
600271d9 SL |
2206 | else |
2207 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2208 | |
600271d9 SL |
2209 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
2210 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
2211 | blk_flush_plug_list(plug, false); |
2212 | trace_block_plug(q); | |
320ae51f | 2213 | } |
b094f89c | 2214 | |
ce5b009c | 2215 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2216 | } else if (q->elevator) { |
105663f7 | 2217 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2218 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2219 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2220 | /* |
6a83e74d | 2221 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2222 | * Otherwise the existing request in the plug list will be |
2223 | * issued. So the plug list will have one request at most | |
2299722c CH |
2224 | * The plug list might get flushed before this. If that happens, |
2225 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2226 | */ |
2299722c CH |
2227 | if (list_empty(&plug->mq_list)) |
2228 | same_queue_rq = NULL; | |
4711b573 | 2229 | if (same_queue_rq) { |
2299722c | 2230 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2231 | plug->rq_count--; |
2232 | } | |
ce5b009c | 2233 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2234 | trace_block_plug(q); |
2299722c | 2235 | |
dad7a3be | 2236 | if (same_queue_rq) { |
ea4f995e | 2237 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2238 | trace_block_unplug(q, 1, true); |
2299722c | 2239 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2240 | &cookie); |
dad7a3be | 2241 | } |
a12de1d4 ML |
2242 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2243 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2244 | /* |
2245 | * There is no scheduler and we can try to send directly | |
2246 | * to the hardware. | |
2247 | */ | |
fd9c40f6 | 2248 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2249 | } else { |
105663f7 | 2250 | /* Default case. */ |
8fa9f556 | 2251 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2252 | } |
320ae51f | 2253 | |
cc29e1bf JX |
2254 | if (!hipri) |
2255 | return BLK_QC_T_NONE; | |
7b371636 | 2256 | return cookie; |
ac7c5675 CH |
2257 | queue_exit: |
2258 | blk_queue_exit(q); | |
2259 | return BLK_QC_T_NONE; | |
320ae51f JA |
2260 | } |
2261 | ||
cc71a6f4 JA |
2262 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2263 | unsigned int hctx_idx) | |
95363efd | 2264 | { |
e9b267d9 | 2265 | struct page *page; |
320ae51f | 2266 | |
24d2f903 | 2267 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2268 | int i; |
320ae51f | 2269 | |
24d2f903 | 2270 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2271 | struct request *rq = tags->static_rqs[i]; |
2272 | ||
2273 | if (!rq) | |
e9b267d9 | 2274 | continue; |
d6296d39 | 2275 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2276 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2277 | } |
320ae51f | 2278 | } |
320ae51f | 2279 | |
24d2f903 CH |
2280 | while (!list_empty(&tags->page_list)) { |
2281 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2282 | list_del_init(&page->lru); |
f75782e4 CM |
2283 | /* |
2284 | * Remove kmemleak object previously allocated in | |
273938bf | 2285 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2286 | */ |
2287 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2288 | __free_pages(page, page->private); |
2289 | } | |
cc71a6f4 | 2290 | } |
320ae51f | 2291 | |
1c0706a7 | 2292 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) |
cc71a6f4 | 2293 | { |
24d2f903 | 2294 | kfree(tags->rqs); |
cc71a6f4 | 2295 | tags->rqs = NULL; |
2af8cbe3 JA |
2296 | kfree(tags->static_rqs); |
2297 | tags->static_rqs = NULL; | |
320ae51f | 2298 | |
1c0706a7 | 2299 | blk_mq_free_tags(tags, flags); |
320ae51f JA |
2300 | } |
2301 | ||
cc71a6f4 JA |
2302 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2303 | unsigned int hctx_idx, | |
2304 | unsigned int nr_tags, | |
1c0706a7 JG |
2305 | unsigned int reserved_tags, |
2306 | unsigned int flags) | |
320ae51f | 2307 | { |
24d2f903 | 2308 | struct blk_mq_tags *tags; |
59f082e4 | 2309 | int node; |
320ae51f | 2310 | |
7d76f856 | 2311 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2312 | if (node == NUMA_NO_NODE) |
2313 | node = set->numa_node; | |
2314 | ||
1c0706a7 | 2315 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); |
24d2f903 CH |
2316 | if (!tags) |
2317 | return NULL; | |
320ae51f | 2318 | |
590b5b7d | 2319 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2320 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2321 | node); |
24d2f903 | 2322 | if (!tags->rqs) { |
1c0706a7 | 2323 | blk_mq_free_tags(tags, flags); |
24d2f903 CH |
2324 | return NULL; |
2325 | } | |
320ae51f | 2326 | |
590b5b7d KC |
2327 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2328 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2329 | node); | |
2af8cbe3 JA |
2330 | if (!tags->static_rqs) { |
2331 | kfree(tags->rqs); | |
1c0706a7 | 2332 | blk_mq_free_tags(tags, flags); |
2af8cbe3 JA |
2333 | return NULL; |
2334 | } | |
2335 | ||
cc71a6f4 JA |
2336 | return tags; |
2337 | } | |
2338 | ||
2339 | static size_t order_to_size(unsigned int order) | |
2340 | { | |
2341 | return (size_t)PAGE_SIZE << order; | |
2342 | } | |
2343 | ||
1d9bd516 TH |
2344 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2345 | unsigned int hctx_idx, int node) | |
2346 | { | |
2347 | int ret; | |
2348 | ||
2349 | if (set->ops->init_request) { | |
2350 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2351 | if (ret) | |
2352 | return ret; | |
2353 | } | |
2354 | ||
12f5b931 | 2355 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2356 | return 0; |
2357 | } | |
2358 | ||
cc71a6f4 JA |
2359 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2360 | unsigned int hctx_idx, unsigned int depth) | |
2361 | { | |
2362 | unsigned int i, j, entries_per_page, max_order = 4; | |
2363 | size_t rq_size, left; | |
59f082e4 SL |
2364 | int node; |
2365 | ||
7d76f856 | 2366 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2367 | if (node == NUMA_NO_NODE) |
2368 | node = set->numa_node; | |
cc71a6f4 JA |
2369 | |
2370 | INIT_LIST_HEAD(&tags->page_list); | |
2371 | ||
320ae51f JA |
2372 | /* |
2373 | * rq_size is the size of the request plus driver payload, rounded | |
2374 | * to the cacheline size | |
2375 | */ | |
24d2f903 | 2376 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2377 | cache_line_size()); |
cc71a6f4 | 2378 | left = rq_size * depth; |
320ae51f | 2379 | |
cc71a6f4 | 2380 | for (i = 0; i < depth; ) { |
320ae51f JA |
2381 | int this_order = max_order; |
2382 | struct page *page; | |
2383 | int to_do; | |
2384 | void *p; | |
2385 | ||
b3a834b1 | 2386 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2387 | this_order--; |
2388 | ||
2389 | do { | |
59f082e4 | 2390 | page = alloc_pages_node(node, |
36e1f3d1 | 2391 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2392 | this_order); |
320ae51f JA |
2393 | if (page) |
2394 | break; | |
2395 | if (!this_order--) | |
2396 | break; | |
2397 | if (order_to_size(this_order) < rq_size) | |
2398 | break; | |
2399 | } while (1); | |
2400 | ||
2401 | if (!page) | |
24d2f903 | 2402 | goto fail; |
320ae51f JA |
2403 | |
2404 | page->private = this_order; | |
24d2f903 | 2405 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2406 | |
2407 | p = page_address(page); | |
f75782e4 CM |
2408 | /* |
2409 | * Allow kmemleak to scan these pages as they contain pointers | |
2410 | * to additional allocations like via ops->init_request(). | |
2411 | */ | |
36e1f3d1 | 2412 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2413 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2414 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2415 | left -= to_do * rq_size; |
2416 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2417 | struct request *rq = p; |
2418 | ||
2419 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2420 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2421 | tags->static_rqs[i] = NULL; | |
2422 | goto fail; | |
e9b267d9 CH |
2423 | } |
2424 | ||
320ae51f JA |
2425 | p += rq_size; |
2426 | i++; | |
2427 | } | |
2428 | } | |
cc71a6f4 | 2429 | return 0; |
320ae51f | 2430 | |
24d2f903 | 2431 | fail: |
cc71a6f4 JA |
2432 | blk_mq_free_rqs(set, tags, hctx_idx); |
2433 | return -ENOMEM; | |
320ae51f JA |
2434 | } |
2435 | ||
bf0beec0 ML |
2436 | struct rq_iter_data { |
2437 | struct blk_mq_hw_ctx *hctx; | |
2438 | bool has_rq; | |
2439 | }; | |
2440 | ||
2441 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2442 | { | |
2443 | struct rq_iter_data *iter_data = data; | |
2444 | ||
2445 | if (rq->mq_hctx != iter_data->hctx) | |
2446 | return true; | |
2447 | iter_data->has_rq = true; | |
2448 | return false; | |
2449 | } | |
2450 | ||
2451 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2452 | { | |
2453 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2454 | hctx->sched_tags : hctx->tags; | |
2455 | struct rq_iter_data data = { | |
2456 | .hctx = hctx, | |
2457 | }; | |
2458 | ||
2459 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2460 | return data.has_rq; | |
2461 | } | |
2462 | ||
2463 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2464 | struct blk_mq_hw_ctx *hctx) | |
2465 | { | |
2466 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2467 | return false; | |
2468 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2469 | return false; | |
2470 | return true; | |
2471 | } | |
2472 | ||
2473 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2474 | { | |
2475 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2476 | struct blk_mq_hw_ctx, cpuhp_online); | |
2477 | ||
2478 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2479 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2480 | return 0; | |
2481 | ||
2482 | /* | |
2483 | * Prevent new request from being allocated on the current hctx. | |
2484 | * | |
2485 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2486 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2487 | * seen once we return from the tag allocator. | |
2488 | */ | |
2489 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2490 | smp_mb__after_atomic(); | |
2491 | ||
2492 | /* | |
2493 | * Try to grab a reference to the queue and wait for any outstanding | |
2494 | * requests. If we could not grab a reference the queue has been | |
2495 | * frozen and there are no requests. | |
2496 | */ | |
2497 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2498 | while (blk_mq_hctx_has_requests(hctx)) | |
2499 | msleep(5); | |
2500 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2501 | } | |
2502 | ||
2503 | return 0; | |
2504 | } | |
2505 | ||
2506 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2507 | { | |
2508 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2509 | struct blk_mq_hw_ctx, cpuhp_online); | |
2510 | ||
2511 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2512 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2513 | return 0; | |
2514 | } | |
2515 | ||
e57690fe JA |
2516 | /* |
2517 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2518 | * software queue to the hw queue dispatch list, and ensure that it | |
2519 | * gets run. | |
2520 | */ | |
9467f859 | 2521 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2522 | { |
9467f859 | 2523 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2524 | struct blk_mq_ctx *ctx; |
2525 | LIST_HEAD(tmp); | |
c16d6b5a | 2526 | enum hctx_type type; |
484b4061 | 2527 | |
9467f859 | 2528 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2529 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2530 | return 0; | |
2531 | ||
e57690fe | 2532 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2533 | type = hctx->type; |
484b4061 JA |
2534 | |
2535 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2536 | if (!list_empty(&ctx->rq_lists[type])) { |
2537 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2538 | blk_mq_hctx_clear_pending(hctx, ctx); |
2539 | } | |
2540 | spin_unlock(&ctx->lock); | |
2541 | ||
2542 | if (list_empty(&tmp)) | |
9467f859 | 2543 | return 0; |
484b4061 | 2544 | |
e57690fe JA |
2545 | spin_lock(&hctx->lock); |
2546 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2547 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2548 | |
2549 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2550 | return 0; |
484b4061 JA |
2551 | } |
2552 | ||
9467f859 | 2553 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2554 | { |
bf0beec0 ML |
2555 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2556 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2557 | &hctx->cpuhp_online); | |
9467f859 TG |
2558 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2559 | &hctx->cpuhp_dead); | |
484b4061 JA |
2560 | } |
2561 | ||
c3b4afca | 2562 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2563 | static void blk_mq_exit_hctx(struct request_queue *q, |
2564 | struct blk_mq_tag_set *set, | |
2565 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2566 | { | |
8ab0b7dc ML |
2567 | if (blk_mq_hw_queue_mapped(hctx)) |
2568 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2569 | |
f70ced09 | 2570 | if (set->ops->exit_request) |
d6296d39 | 2571 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2572 | |
08e98fc6 ML |
2573 | if (set->ops->exit_hctx) |
2574 | set->ops->exit_hctx(hctx, hctx_idx); | |
2575 | ||
9467f859 | 2576 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2577 | |
2578 | spin_lock(&q->unused_hctx_lock); | |
2579 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2580 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2581 | } |
2582 | ||
624dbe47 ML |
2583 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2584 | struct blk_mq_tag_set *set, int nr_queue) | |
2585 | { | |
2586 | struct blk_mq_hw_ctx *hctx; | |
2587 | unsigned int i; | |
2588 | ||
2589 | queue_for_each_hw_ctx(q, hctx, i) { | |
2590 | if (i == nr_queue) | |
2591 | break; | |
477e19de | 2592 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2593 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2594 | } |
624dbe47 ML |
2595 | } |
2596 | ||
7c6c5b7c ML |
2597 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2598 | { | |
2599 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2600 | ||
2601 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2602 | __alignof__(struct blk_mq_hw_ctx)) != | |
2603 | sizeof(struct blk_mq_hw_ctx)); | |
2604 | ||
2605 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2606 | hw_ctx_size += sizeof(struct srcu_struct); | |
2607 | ||
2608 | return hw_ctx_size; | |
2609 | } | |
2610 | ||
08e98fc6 ML |
2611 | static int blk_mq_init_hctx(struct request_queue *q, |
2612 | struct blk_mq_tag_set *set, | |
2613 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2614 | { |
7c6c5b7c ML |
2615 | hctx->queue_num = hctx_idx; |
2616 | ||
bf0beec0 ML |
2617 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2618 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2619 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2620 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2621 | ||
2622 | hctx->tags = set->tags[hctx_idx]; | |
2623 | ||
2624 | if (set->ops->init_hctx && | |
2625 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2626 | goto unregister_cpu_notifier; | |
08e98fc6 | 2627 | |
7c6c5b7c ML |
2628 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2629 | hctx->numa_node)) | |
2630 | goto exit_hctx; | |
2631 | return 0; | |
2632 | ||
2633 | exit_hctx: | |
2634 | if (set->ops->exit_hctx) | |
2635 | set->ops->exit_hctx(hctx, hctx_idx); | |
2636 | unregister_cpu_notifier: | |
2637 | blk_mq_remove_cpuhp(hctx); | |
2638 | return -1; | |
2639 | } | |
2640 | ||
2641 | static struct blk_mq_hw_ctx * | |
2642 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2643 | int node) | |
2644 | { | |
2645 | struct blk_mq_hw_ctx *hctx; | |
2646 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2647 | ||
2648 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2649 | if (!hctx) | |
2650 | goto fail_alloc_hctx; | |
2651 | ||
2652 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2653 | goto free_hctx; | |
2654 | ||
2655 | atomic_set(&hctx->nr_active, 0); | |
b445547e | 2656 | atomic_set(&hctx->elevator_queued, 0); |
08e98fc6 | 2657 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2658 | node = set->numa_node; |
2659 | hctx->numa_node = node; | |
08e98fc6 | 2660 | |
9f993737 | 2661 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2662 | spin_lock_init(&hctx->lock); |
2663 | INIT_LIST_HEAD(&hctx->dispatch); | |
2664 | hctx->queue = q; | |
51db1c37 | 2665 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2666 | |
2f8f1336 ML |
2667 | INIT_LIST_HEAD(&hctx->hctx_list); |
2668 | ||
320ae51f | 2669 | /* |
08e98fc6 ML |
2670 | * Allocate space for all possible cpus to avoid allocation at |
2671 | * runtime | |
320ae51f | 2672 | */ |
d904bfa7 | 2673 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2674 | gfp, node); |
08e98fc6 | 2675 | if (!hctx->ctxs) |
7c6c5b7c | 2676 | goto free_cpumask; |
320ae51f | 2677 | |
5b202853 | 2678 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2679 | gfp, node)) |
08e98fc6 | 2680 | goto free_ctxs; |
08e98fc6 | 2681 | hctx->nr_ctx = 0; |
320ae51f | 2682 | |
5815839b | 2683 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2684 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2685 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2686 | ||
754a1572 | 2687 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2688 | if (!hctx->fq) |
7c6c5b7c | 2689 | goto free_bitmap; |
320ae51f | 2690 | |
6a83e74d | 2691 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2692 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2693 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2694 | |
7c6c5b7c | 2695 | return hctx; |
320ae51f | 2696 | |
08e98fc6 | 2697 | free_bitmap: |
88459642 | 2698 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2699 | free_ctxs: |
2700 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2701 | free_cpumask: |
2702 | free_cpumask_var(hctx->cpumask); | |
2703 | free_hctx: | |
2704 | kfree(hctx); | |
2705 | fail_alloc_hctx: | |
2706 | return NULL; | |
08e98fc6 | 2707 | } |
320ae51f | 2708 | |
320ae51f JA |
2709 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2710 | unsigned int nr_hw_queues) | |
2711 | { | |
b3c661b1 JA |
2712 | struct blk_mq_tag_set *set = q->tag_set; |
2713 | unsigned int i, j; | |
320ae51f JA |
2714 | |
2715 | for_each_possible_cpu(i) { | |
2716 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2717 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2718 | int k; |
320ae51f | 2719 | |
320ae51f JA |
2720 | __ctx->cpu = i; |
2721 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2722 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2723 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2724 | ||
320ae51f JA |
2725 | __ctx->queue = q; |
2726 | ||
320ae51f JA |
2727 | /* |
2728 | * Set local node, IFF we have more than one hw queue. If | |
2729 | * not, we remain on the home node of the device | |
2730 | */ | |
b3c661b1 JA |
2731 | for (j = 0; j < set->nr_maps; j++) { |
2732 | hctx = blk_mq_map_queue_type(q, j, i); | |
2733 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2734 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2735 | } |
320ae51f JA |
2736 | } |
2737 | } | |
2738 | ||
03b63b02 WZ |
2739 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, |
2740 | int hctx_idx) | |
cc71a6f4 | 2741 | { |
1c0706a7 | 2742 | unsigned int flags = set->flags; |
cc71a6f4 JA |
2743 | int ret = 0; |
2744 | ||
2745 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1c0706a7 | 2746 | set->queue_depth, set->reserved_tags, flags); |
cc71a6f4 JA |
2747 | if (!set->tags[hctx_idx]) |
2748 | return false; | |
2749 | ||
2750 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2751 | set->queue_depth); | |
2752 | if (!ret) | |
2753 | return true; | |
2754 | ||
1c0706a7 | 2755 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
cc71a6f4 JA |
2756 | set->tags[hctx_idx] = NULL; |
2757 | return false; | |
2758 | } | |
2759 | ||
2760 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2761 | unsigned int hctx_idx) | |
2762 | { | |
1c0706a7 JG |
2763 | unsigned int flags = set->flags; |
2764 | ||
4e6db0f2 | 2765 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 | 2766 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
1c0706a7 | 2767 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
bd166ef1 JA |
2768 | set->tags[hctx_idx] = NULL; |
2769 | } | |
cc71a6f4 JA |
2770 | } |
2771 | ||
4b855ad3 | 2772 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2773 | { |
b3c661b1 | 2774 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2775 | struct blk_mq_hw_ctx *hctx; |
2776 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2777 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2778 | |
2779 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2780 | cpumask_clear(hctx->cpumask); |
320ae51f | 2781 | hctx->nr_ctx = 0; |
d416c92c | 2782 | hctx->dispatch_from = NULL; |
320ae51f JA |
2783 | } |
2784 | ||
2785 | /* | |
4b855ad3 | 2786 | * Map software to hardware queues. |
4412efec ML |
2787 | * |
2788 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2789 | */ |
20e4d813 | 2790 | for_each_possible_cpu(i) { |
4412efec | 2791 | |
897bb0c7 | 2792 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2793 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2794 | if (!set->map[j].nr_queues) { |
2795 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2796 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2797 | continue; |
bb94aea1 | 2798 | } |
fd689871 ML |
2799 | hctx_idx = set->map[j].mq_map[i]; |
2800 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2801 | if (!set->tags[hctx_idx] && | |
03b63b02 | 2802 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { |
fd689871 ML |
2803 | /* |
2804 | * If tags initialization fail for some hctx, | |
2805 | * that hctx won't be brought online. In this | |
2806 | * case, remap the current ctx to hctx[0] which | |
2807 | * is guaranteed to always have tags allocated | |
2808 | */ | |
2809 | set->map[j].mq_map[i] = 0; | |
2810 | } | |
e5edd5f2 | 2811 | |
b3c661b1 | 2812 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2813 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2814 | /* |
2815 | * If the CPU is already set in the mask, then we've | |
2816 | * mapped this one already. This can happen if | |
2817 | * devices share queues across queue maps. | |
2818 | */ | |
2819 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2820 | continue; | |
2821 | ||
2822 | cpumask_set_cpu(i, hctx->cpumask); | |
2823 | hctx->type = j; | |
2824 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2825 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2826 | ||
2827 | /* | |
2828 | * If the nr_ctx type overflows, we have exceeded the | |
2829 | * amount of sw queues we can support. | |
2830 | */ | |
2831 | BUG_ON(!hctx->nr_ctx); | |
2832 | } | |
bb94aea1 JW |
2833 | |
2834 | for (; j < HCTX_MAX_TYPES; j++) | |
2835 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2836 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2837 | } |
506e931f JA |
2838 | |
2839 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2840 | /* |
2841 | * If no software queues are mapped to this hardware queue, | |
2842 | * disable it and free the request entries. | |
2843 | */ | |
2844 | if (!hctx->nr_ctx) { | |
2845 | /* Never unmap queue 0. We need it as a | |
2846 | * fallback in case of a new remap fails | |
2847 | * allocation | |
2848 | */ | |
2849 | if (i && set->tags[i]) | |
2850 | blk_mq_free_map_and_requests(set, i); | |
2851 | ||
2852 | hctx->tags = NULL; | |
2853 | continue; | |
2854 | } | |
484b4061 | 2855 | |
2a34c087 ML |
2856 | hctx->tags = set->tags[i]; |
2857 | WARN_ON(!hctx->tags); | |
2858 | ||
889fa31f CY |
2859 | /* |
2860 | * Set the map size to the number of mapped software queues. | |
2861 | * This is more accurate and more efficient than looping | |
2862 | * over all possibly mapped software queues. | |
2863 | */ | |
88459642 | 2864 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2865 | |
484b4061 JA |
2866 | /* |
2867 | * Initialize batch roundrobin counts | |
2868 | */ | |
f82ddf19 | 2869 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2870 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2871 | } | |
320ae51f JA |
2872 | } |
2873 | ||
8e8320c9 JA |
2874 | /* |
2875 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2876 | * the queue isn't live yet. | |
2877 | */ | |
2404e607 | 2878 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2879 | { |
2880 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2881 | int i; |
2882 | ||
2404e607 | 2883 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2884 | if (shared) |
51db1c37 | 2885 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
97889f9a | 2886 | else |
51db1c37 | 2887 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 JM |
2888 | } |
2889 | } | |
2890 | ||
655ac300 HR |
2891 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
2892 | bool shared) | |
2404e607 JM |
2893 | { |
2894 | struct request_queue *q; | |
0d2602ca | 2895 | |
705cda97 BVA |
2896 | lockdep_assert_held(&set->tag_list_lock); |
2897 | ||
0d2602ca JA |
2898 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2899 | blk_mq_freeze_queue(q); | |
2404e607 | 2900 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2901 | blk_mq_unfreeze_queue(q); |
2902 | } | |
2903 | } | |
2904 | ||
2905 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2906 | { | |
2907 | struct blk_mq_tag_set *set = q->tag_set; | |
2908 | ||
0d2602ca | 2909 | mutex_lock(&set->tag_list_lock); |
08c875cb | 2910 | list_del(&q->tag_set_list); |
2404e607 JM |
2911 | if (list_is_singular(&set->tag_list)) { |
2912 | /* just transitioned to unshared */ | |
51db1c37 | 2913 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 2914 | /* update existing queue */ |
655ac300 | 2915 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 2916 | } |
0d2602ca | 2917 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2918 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2919 | } |
2920 | ||
2921 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2922 | struct request_queue *q) | |
2923 | { | |
0d2602ca | 2924 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2925 | |
ff821d27 JA |
2926 | /* |
2927 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2928 | */ | |
2929 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
2930 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
2931 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 2932 | /* update existing queue */ |
655ac300 | 2933 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 2934 | } |
51db1c37 | 2935 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 2936 | queue_set_hctx_shared(q, true); |
08c875cb | 2937 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2938 | |
0d2602ca JA |
2939 | mutex_unlock(&set->tag_list_lock); |
2940 | } | |
2941 | ||
1db4909e ML |
2942 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2943 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2944 | { | |
2945 | struct blk_mq_ctxs *ctxs; | |
2946 | int cpu; | |
2947 | ||
2948 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2949 | if (!ctxs) | |
2950 | return -ENOMEM; | |
2951 | ||
2952 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2953 | if (!ctxs->queue_ctx) | |
2954 | goto fail; | |
2955 | ||
2956 | for_each_possible_cpu(cpu) { | |
2957 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2958 | ctx->ctxs = ctxs; | |
2959 | } | |
2960 | ||
2961 | q->mq_kobj = &ctxs->kobj; | |
2962 | q->queue_ctx = ctxs->queue_ctx; | |
2963 | ||
2964 | return 0; | |
2965 | fail: | |
2966 | kfree(ctxs); | |
2967 | return -ENOMEM; | |
2968 | } | |
2969 | ||
e09aae7e ML |
2970 | /* |
2971 | * It is the actual release handler for mq, but we do it from | |
2972 | * request queue's release handler for avoiding use-after-free | |
2973 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2974 | * but we can't group ctx/kctx kobj without it. | |
2975 | */ | |
2976 | void blk_mq_release(struct request_queue *q) | |
2977 | { | |
2f8f1336 ML |
2978 | struct blk_mq_hw_ctx *hctx, *next; |
2979 | int i; | |
e09aae7e | 2980 | |
2f8f1336 ML |
2981 | queue_for_each_hw_ctx(q, hctx, i) |
2982 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2983 | ||
2984 | /* all hctx are in .unused_hctx_list now */ | |
2985 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2986 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 2987 | kobject_put(&hctx->kobj); |
c3b4afca | 2988 | } |
e09aae7e ML |
2989 | |
2990 | kfree(q->queue_hw_ctx); | |
2991 | ||
7ea5fe31 ML |
2992 | /* |
2993 | * release .mq_kobj and sw queue's kobject now because | |
2994 | * both share lifetime with request queue. | |
2995 | */ | |
2996 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2997 | } |
2998 | ||
2f227bb9 CH |
2999 | struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
3000 | void *queuedata) | |
b62c21b7 MS |
3001 | { |
3002 | struct request_queue *uninit_q, *q; | |
3003 | ||
c62b37d9 | 3004 | uninit_q = blk_alloc_queue(set->numa_node); |
b62c21b7 MS |
3005 | if (!uninit_q) |
3006 | return ERR_PTR(-ENOMEM); | |
2f227bb9 | 3007 | uninit_q->queuedata = queuedata; |
b62c21b7 | 3008 | |
737eb78e DLM |
3009 | /* |
3010 | * Initialize the queue without an elevator. device_add_disk() will do | |
3011 | * the initialization. | |
3012 | */ | |
3013 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
b62c21b7 MS |
3014 | if (IS_ERR(q)) |
3015 | blk_cleanup_queue(uninit_q); | |
3016 | ||
3017 | return q; | |
3018 | } | |
2f227bb9 CH |
3019 | EXPORT_SYMBOL_GPL(blk_mq_init_queue_data); |
3020 | ||
3021 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3022 | { | |
3023 | return blk_mq_init_queue_data(set, NULL); | |
3024 | } | |
b62c21b7 MS |
3025 | EXPORT_SYMBOL(blk_mq_init_queue); |
3026 | ||
9316a9ed JA |
3027 | /* |
3028 | * Helper for setting up a queue with mq ops, given queue depth, and | |
3029 | * the passed in mq ops flags. | |
3030 | */ | |
3031 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
3032 | const struct blk_mq_ops *ops, | |
3033 | unsigned int queue_depth, | |
3034 | unsigned int set_flags) | |
3035 | { | |
3036 | struct request_queue *q; | |
3037 | int ret; | |
3038 | ||
3039 | memset(set, 0, sizeof(*set)); | |
3040 | set->ops = ops; | |
3041 | set->nr_hw_queues = 1; | |
b3c661b1 | 3042 | set->nr_maps = 1; |
9316a9ed JA |
3043 | set->queue_depth = queue_depth; |
3044 | set->numa_node = NUMA_NO_NODE; | |
3045 | set->flags = set_flags; | |
3046 | ||
3047 | ret = blk_mq_alloc_tag_set(set); | |
3048 | if (ret) | |
3049 | return ERR_PTR(ret); | |
3050 | ||
3051 | q = blk_mq_init_queue(set); | |
3052 | if (IS_ERR(q)) { | |
3053 | blk_mq_free_tag_set(set); | |
3054 | return q; | |
3055 | } | |
3056 | ||
3057 | return q; | |
3058 | } | |
3059 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
3060 | ||
34d11ffa JW |
3061 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3062 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3063 | int hctx_idx, int node) | |
3064 | { | |
2f8f1336 | 3065 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3066 | |
2f8f1336 ML |
3067 | /* reuse dead hctx first */ |
3068 | spin_lock(&q->unused_hctx_lock); | |
3069 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3070 | if (tmp->numa_node == node) { | |
3071 | hctx = tmp; | |
3072 | break; | |
3073 | } | |
3074 | } | |
3075 | if (hctx) | |
3076 | list_del_init(&hctx->hctx_list); | |
3077 | spin_unlock(&q->unused_hctx_lock); | |
3078 | ||
3079 | if (!hctx) | |
3080 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3081 | if (!hctx) |
7c6c5b7c | 3082 | goto fail; |
34d11ffa | 3083 | |
7c6c5b7c ML |
3084 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3085 | goto free_hctx; | |
34d11ffa JW |
3086 | |
3087 | return hctx; | |
7c6c5b7c ML |
3088 | |
3089 | free_hctx: | |
3090 | kobject_put(&hctx->kobj); | |
3091 | fail: | |
3092 | return NULL; | |
34d11ffa JW |
3093 | } |
3094 | ||
868f2f0b KB |
3095 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3096 | struct request_queue *q) | |
320ae51f | 3097 | { |
e01ad46d | 3098 | int i, j, end; |
868f2f0b | 3099 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3100 | |
ac0d6b92 BVA |
3101 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3102 | struct blk_mq_hw_ctx **new_hctxs; | |
3103 | ||
3104 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3105 | sizeof(*new_hctxs), GFP_KERNEL, | |
3106 | set->numa_node); | |
3107 | if (!new_hctxs) | |
3108 | return; | |
3109 | if (hctxs) | |
3110 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3111 | sizeof(*hctxs)); | |
3112 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3113 | kfree(hctxs); |
3114 | hctxs = new_hctxs; | |
3115 | } | |
3116 | ||
fb350e0a ML |
3117 | /* protect against switching io scheduler */ |
3118 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3119 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3120 | int node; |
34d11ffa | 3121 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3122 | |
7d76f856 | 3123 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3124 | /* |
3125 | * If the hw queue has been mapped to another numa node, | |
3126 | * we need to realloc the hctx. If allocation fails, fallback | |
3127 | * to use the previous one. | |
3128 | */ | |
3129 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3130 | continue; | |
868f2f0b | 3131 | |
34d11ffa JW |
3132 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3133 | if (hctx) { | |
2f8f1336 | 3134 | if (hctxs[i]) |
34d11ffa | 3135 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3136 | hctxs[i] = hctx; |
3137 | } else { | |
3138 | if (hctxs[i]) | |
3139 | pr_warn("Allocate new hctx on node %d fails,\ | |
3140 | fallback to previous one on node %d\n", | |
3141 | node, hctxs[i]->numa_node); | |
3142 | else | |
3143 | break; | |
868f2f0b | 3144 | } |
320ae51f | 3145 | } |
e01ad46d JW |
3146 | /* |
3147 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3148 | * hctxs and keep the previous q->nr_hw_queues. | |
3149 | */ | |
3150 | if (i != set->nr_hw_queues) { | |
3151 | j = q->nr_hw_queues; | |
3152 | end = i; | |
3153 | } else { | |
3154 | j = i; | |
3155 | end = q->nr_hw_queues; | |
3156 | q->nr_hw_queues = set->nr_hw_queues; | |
3157 | } | |
34d11ffa | 3158 | |
e01ad46d | 3159 | for (; j < end; j++) { |
868f2f0b KB |
3160 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3161 | ||
3162 | if (hctx) { | |
cc71a6f4 JA |
3163 | if (hctx->tags) |
3164 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 3165 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3166 | hctxs[j] = NULL; |
868f2f0b KB |
3167 | } |
3168 | } | |
fb350e0a | 3169 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3170 | } |
3171 | ||
3172 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
737eb78e DLM |
3173 | struct request_queue *q, |
3174 | bool elevator_init) | |
868f2f0b | 3175 | { |
66841672 ML |
3176 | /* mark the queue as mq asap */ |
3177 | q->mq_ops = set->ops; | |
3178 | ||
34dbad5d | 3179 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3180 | blk_mq_poll_stats_bkt, |
3181 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3182 | if (!q->poll_cb) |
3183 | goto err_exit; | |
3184 | ||
1db4909e | 3185 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3186 | goto err_poll; |
868f2f0b | 3187 | |
737f98cf ML |
3188 | /* init q->mq_kobj and sw queues' kobjects */ |
3189 | blk_mq_sysfs_init(q); | |
3190 | ||
2f8f1336 ML |
3191 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3192 | spin_lock_init(&q->unused_hctx_lock); | |
3193 | ||
868f2f0b KB |
3194 | blk_mq_realloc_hw_ctxs(set, q); |
3195 | if (!q->nr_hw_queues) | |
3196 | goto err_hctxs; | |
320ae51f | 3197 | |
287922eb | 3198 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3199 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3200 | |
a8908939 | 3201 | q->tag_set = set; |
320ae51f | 3202 | |
94eddfbe | 3203 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3204 | if (set->nr_maps > HCTX_TYPE_POLL && |
3205 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3206 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3207 | |
1be036e9 CH |
3208 | q->sg_reserved_size = INT_MAX; |
3209 | ||
2849450a | 3210 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3211 | INIT_LIST_HEAD(&q->requeue_list); |
3212 | spin_lock_init(&q->requeue_lock); | |
3213 | ||
eba71768 JA |
3214 | q->nr_requests = set->queue_depth; |
3215 | ||
64f1c21e JA |
3216 | /* |
3217 | * Default to classic polling | |
3218 | */ | |
29ece8b4 | 3219 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3220 | |
24d2f903 | 3221 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3222 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3223 | blk_mq_map_swqueue(q); |
4593fdbe | 3224 | |
737eb78e DLM |
3225 | if (elevator_init) |
3226 | elevator_init_mq(q); | |
d3484991 | 3227 | |
320ae51f | 3228 | return q; |
18741986 | 3229 | |
320ae51f | 3230 | err_hctxs: |
868f2f0b | 3231 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3232 | q->nr_hw_queues = 0; |
1db4909e | 3233 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3234 | err_poll: |
3235 | blk_stat_free_callback(q->poll_cb); | |
3236 | q->poll_cb = NULL; | |
c7de5726 ML |
3237 | err_exit: |
3238 | q->mq_ops = NULL; | |
320ae51f JA |
3239 | return ERR_PTR(-ENOMEM); |
3240 | } | |
b62c21b7 | 3241 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3242 | |
c7e2d94b ML |
3243 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3244 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3245 | { |
624dbe47 | 3246 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3247 | |
0d2602ca | 3248 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 3249 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 3250 | } |
320ae51f | 3251 | |
a5164405 JA |
3252 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3253 | { | |
3254 | int i; | |
3255 | ||
8229cca8 | 3256 | for (i = 0; i < set->nr_hw_queues; i++) { |
03b63b02 | 3257 | if (!__blk_mq_alloc_map_and_request(set, i)) |
a5164405 | 3258 | goto out_unwind; |
8229cca8 XT |
3259 | cond_resched(); |
3260 | } | |
a5164405 JA |
3261 | |
3262 | return 0; | |
3263 | ||
3264 | out_unwind: | |
3265 | while (--i >= 0) | |
2e194422 | 3266 | blk_mq_free_map_and_requests(set, i); |
a5164405 | 3267 | |
a5164405 JA |
3268 | return -ENOMEM; |
3269 | } | |
3270 | ||
3271 | /* | |
3272 | * Allocate the request maps associated with this tag_set. Note that this | |
3273 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3274 | * will be updated to reflect the allocated depth. | |
3275 | */ | |
79fab528 | 3276 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) |
a5164405 JA |
3277 | { |
3278 | unsigned int depth; | |
3279 | int err; | |
3280 | ||
3281 | depth = set->queue_depth; | |
3282 | do { | |
3283 | err = __blk_mq_alloc_rq_maps(set); | |
3284 | if (!err) | |
3285 | break; | |
3286 | ||
3287 | set->queue_depth >>= 1; | |
3288 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3289 | err = -ENOMEM; | |
3290 | break; | |
3291 | } | |
3292 | } while (set->queue_depth); | |
3293 | ||
3294 | if (!set->queue_depth || err) { | |
3295 | pr_err("blk-mq: failed to allocate request map\n"); | |
3296 | return -ENOMEM; | |
3297 | } | |
3298 | ||
3299 | if (depth != set->queue_depth) | |
3300 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3301 | depth, set->queue_depth); | |
3302 | ||
3303 | return 0; | |
3304 | } | |
3305 | ||
ebe8bddb OS |
3306 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3307 | { | |
6e66b493 BVA |
3308 | /* |
3309 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3310 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3311 | * number of hardware queues. | |
3312 | */ | |
3313 | if (set->nr_maps == 1) | |
3314 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3315 | ||
59388702 | 3316 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3317 | int i; |
3318 | ||
7d4901a9 ML |
3319 | /* |
3320 | * transport .map_queues is usually done in the following | |
3321 | * way: | |
3322 | * | |
3323 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3324 | * mask = get_cpu_mask(queue) | |
3325 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3326 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3327 | * } |
3328 | * | |
3329 | * When we need to remap, the table has to be cleared for | |
3330 | * killing stale mapping since one CPU may not be mapped | |
3331 | * to any hw queue. | |
3332 | */ | |
b3c661b1 JA |
3333 | for (i = 0; i < set->nr_maps; i++) |
3334 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3335 | |
ebe8bddb | 3336 | return set->ops->map_queues(set); |
b3c661b1 JA |
3337 | } else { |
3338 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3339 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3340 | } |
ebe8bddb OS |
3341 | } |
3342 | ||
f7e76dbc BVA |
3343 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3344 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3345 | { | |
3346 | struct blk_mq_tags **new_tags; | |
3347 | ||
3348 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3349 | return 0; | |
3350 | ||
3351 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3352 | GFP_KERNEL, set->numa_node); | |
3353 | if (!new_tags) | |
3354 | return -ENOMEM; | |
3355 | ||
3356 | if (set->tags) | |
3357 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3358 | sizeof(*set->tags)); | |
3359 | kfree(set->tags); | |
3360 | set->tags = new_tags; | |
3361 | set->nr_hw_queues = new_nr_hw_queues; | |
3362 | ||
3363 | return 0; | |
3364 | } | |
3365 | ||
91cdf265 MI |
3366 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3367 | int new_nr_hw_queues) | |
3368 | { | |
3369 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3370 | } | |
3371 | ||
a4391c64 JA |
3372 | /* |
3373 | * Alloc a tag set to be associated with one or more request queues. | |
3374 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3375 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3376 | * value will be stored in set->queue_depth. |
3377 | */ | |
24d2f903 CH |
3378 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3379 | { | |
b3c661b1 | 3380 | int i, ret; |
da695ba2 | 3381 | |
205fb5f5 BVA |
3382 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3383 | ||
24d2f903 CH |
3384 | if (!set->nr_hw_queues) |
3385 | return -EINVAL; | |
a4391c64 | 3386 | if (!set->queue_depth) |
24d2f903 CH |
3387 | return -EINVAL; |
3388 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3389 | return -EINVAL; | |
3390 | ||
7d7e0f90 | 3391 | if (!set->ops->queue_rq) |
24d2f903 CH |
3392 | return -EINVAL; |
3393 | ||
de148297 ML |
3394 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3395 | return -EINVAL; | |
3396 | ||
a4391c64 JA |
3397 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3398 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3399 | BLK_MQ_MAX_DEPTH); | |
3400 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3401 | } | |
24d2f903 | 3402 | |
b3c661b1 JA |
3403 | if (!set->nr_maps) |
3404 | set->nr_maps = 1; | |
3405 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3406 | return -EINVAL; | |
3407 | ||
6637fadf SL |
3408 | /* |
3409 | * If a crashdump is active, then we are potentially in a very | |
3410 | * memory constrained environment. Limit us to 1 queue and | |
3411 | * 64 tags to prevent using too much memory. | |
3412 | */ | |
3413 | if (is_kdump_kernel()) { | |
3414 | set->nr_hw_queues = 1; | |
59388702 | 3415 | set->nr_maps = 1; |
6637fadf SL |
3416 | set->queue_depth = min(64U, set->queue_depth); |
3417 | } | |
868f2f0b | 3418 | /* |
392546ae JA |
3419 | * There is no use for more h/w queues than cpus if we just have |
3420 | * a single map | |
868f2f0b | 3421 | */ |
392546ae | 3422 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3423 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3424 | |
91cdf265 | 3425 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3426 | return -ENOMEM; |
24d2f903 | 3427 | |
da695ba2 | 3428 | ret = -ENOMEM; |
b3c661b1 JA |
3429 | for (i = 0; i < set->nr_maps; i++) { |
3430 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3431 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3432 | GFP_KERNEL, set->numa_node); |
3433 | if (!set->map[i].mq_map) | |
3434 | goto out_free_mq_map; | |
59388702 | 3435 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3436 | } |
bdd17e75 | 3437 | |
ebe8bddb | 3438 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3439 | if (ret) |
3440 | goto out_free_mq_map; | |
3441 | ||
79fab528 | 3442 | ret = blk_mq_alloc_map_and_requests(set); |
da695ba2 | 3443 | if (ret) |
bdd17e75 | 3444 | goto out_free_mq_map; |
24d2f903 | 3445 | |
32bc15af | 3446 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
f1b49fdc JG |
3447 | atomic_set(&set->active_queues_shared_sbitmap, 0); |
3448 | ||
32bc15af JG |
3449 | if (blk_mq_init_shared_sbitmap(set, set->flags)) { |
3450 | ret = -ENOMEM; | |
3451 | goto out_free_mq_rq_maps; | |
3452 | } | |
3453 | } | |
3454 | ||
0d2602ca JA |
3455 | mutex_init(&set->tag_list_lock); |
3456 | INIT_LIST_HEAD(&set->tag_list); | |
3457 | ||
24d2f903 | 3458 | return 0; |
bdd17e75 | 3459 | |
32bc15af JG |
3460 | out_free_mq_rq_maps: |
3461 | for (i = 0; i < set->nr_hw_queues; i++) | |
3462 | blk_mq_free_map_and_requests(set, i); | |
bdd17e75 | 3463 | out_free_mq_map: |
b3c661b1 JA |
3464 | for (i = 0; i < set->nr_maps; i++) { |
3465 | kfree(set->map[i].mq_map); | |
3466 | set->map[i].mq_map = NULL; | |
3467 | } | |
5676e7b6 RE |
3468 | kfree(set->tags); |
3469 | set->tags = NULL; | |
da695ba2 | 3470 | return ret; |
24d2f903 CH |
3471 | } |
3472 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3473 | ||
3474 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3475 | { | |
b3c661b1 | 3476 | int i, j; |
24d2f903 | 3477 | |
f7e76dbc | 3478 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3479 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3480 | |
32bc15af JG |
3481 | if (blk_mq_is_sbitmap_shared(set->flags)) |
3482 | blk_mq_exit_shared_sbitmap(set); | |
3483 | ||
b3c661b1 JA |
3484 | for (j = 0; j < set->nr_maps; j++) { |
3485 | kfree(set->map[j].mq_map); | |
3486 | set->map[j].mq_map = NULL; | |
3487 | } | |
bdd17e75 | 3488 | |
981bd189 | 3489 | kfree(set->tags); |
5676e7b6 | 3490 | set->tags = NULL; |
24d2f903 CH |
3491 | } |
3492 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3493 | ||
e3a2b3f9 JA |
3494 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3495 | { | |
3496 | struct blk_mq_tag_set *set = q->tag_set; | |
3497 | struct blk_mq_hw_ctx *hctx; | |
3498 | int i, ret; | |
3499 | ||
bd166ef1 | 3500 | if (!set) |
e3a2b3f9 JA |
3501 | return -EINVAL; |
3502 | ||
e5fa8140 AZ |
3503 | if (q->nr_requests == nr) |
3504 | return 0; | |
3505 | ||
70f36b60 | 3506 | blk_mq_freeze_queue(q); |
24f5a90f | 3507 | blk_mq_quiesce_queue(q); |
70f36b60 | 3508 | |
e3a2b3f9 JA |
3509 | ret = 0; |
3510 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3511 | if (!hctx->tags) |
3512 | continue; | |
bd166ef1 JA |
3513 | /* |
3514 | * If we're using an MQ scheduler, just update the scheduler | |
3515 | * queue depth. This is similar to what the old code would do. | |
3516 | */ | |
70f36b60 | 3517 | if (!hctx->sched_tags) { |
c2e82a23 | 3518 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 | 3519 | false); |
32bc15af JG |
3520 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) |
3521 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
70f36b60 JA |
3522 | } else { |
3523 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3524 | nr, true); | |
3525 | } | |
e3a2b3f9 JA |
3526 | if (ret) |
3527 | break; | |
77f1e0a5 JA |
3528 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3529 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3530 | } |
3531 | ||
3532 | if (!ret) | |
3533 | q->nr_requests = nr; | |
3534 | ||
24f5a90f | 3535 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3536 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3537 | |
e3a2b3f9 JA |
3538 | return ret; |
3539 | } | |
3540 | ||
d48ece20 JW |
3541 | /* |
3542 | * request_queue and elevator_type pair. | |
3543 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3544 | * the elevator_type associated with a request_queue. | |
3545 | */ | |
3546 | struct blk_mq_qe_pair { | |
3547 | struct list_head node; | |
3548 | struct request_queue *q; | |
3549 | struct elevator_type *type; | |
3550 | }; | |
3551 | ||
3552 | /* | |
3553 | * Cache the elevator_type in qe pair list and switch the | |
3554 | * io scheduler to 'none' | |
3555 | */ | |
3556 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3557 | struct request_queue *q) | |
3558 | { | |
3559 | struct blk_mq_qe_pair *qe; | |
3560 | ||
3561 | if (!q->elevator) | |
3562 | return true; | |
3563 | ||
3564 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3565 | if (!qe) | |
3566 | return false; | |
3567 | ||
3568 | INIT_LIST_HEAD(&qe->node); | |
3569 | qe->q = q; | |
3570 | qe->type = q->elevator->type; | |
3571 | list_add(&qe->node, head); | |
3572 | ||
3573 | mutex_lock(&q->sysfs_lock); | |
3574 | /* | |
3575 | * After elevator_switch_mq, the previous elevator_queue will be | |
3576 | * released by elevator_release. The reference of the io scheduler | |
3577 | * module get by elevator_get will also be put. So we need to get | |
3578 | * a reference of the io scheduler module here to prevent it to be | |
3579 | * removed. | |
3580 | */ | |
3581 | __module_get(qe->type->elevator_owner); | |
3582 | elevator_switch_mq(q, NULL); | |
3583 | mutex_unlock(&q->sysfs_lock); | |
3584 | ||
3585 | return true; | |
3586 | } | |
3587 | ||
3588 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3589 | struct request_queue *q) | |
3590 | { | |
3591 | struct blk_mq_qe_pair *qe; | |
3592 | struct elevator_type *t = NULL; | |
3593 | ||
3594 | list_for_each_entry(qe, head, node) | |
3595 | if (qe->q == q) { | |
3596 | t = qe->type; | |
3597 | break; | |
3598 | } | |
3599 | ||
3600 | if (!t) | |
3601 | return; | |
3602 | ||
3603 | list_del(&qe->node); | |
3604 | kfree(qe); | |
3605 | ||
3606 | mutex_lock(&q->sysfs_lock); | |
3607 | elevator_switch_mq(q, t); | |
3608 | mutex_unlock(&q->sysfs_lock); | |
3609 | } | |
3610 | ||
e4dc2b32 KB |
3611 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3612 | int nr_hw_queues) | |
868f2f0b KB |
3613 | { |
3614 | struct request_queue *q; | |
d48ece20 | 3615 | LIST_HEAD(head); |
e01ad46d | 3616 | int prev_nr_hw_queues; |
868f2f0b | 3617 | |
705cda97 BVA |
3618 | lockdep_assert_held(&set->tag_list_lock); |
3619 | ||
392546ae | 3620 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3621 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3622 | if (nr_hw_queues < 1) |
3623 | return; | |
3624 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3625 | return; |
3626 | ||
3627 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3628 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3629 | /* |
3630 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3631 | * with the previous scheduler. We will switch back once we are done | |
3632 | * updating the new sw to hw queue mappings. | |
3633 | */ | |
3634 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3635 | if (!blk_mq_elv_switch_none(&head, q)) | |
3636 | goto switch_back; | |
868f2f0b | 3637 | |
477e19de JW |
3638 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3639 | blk_mq_debugfs_unregister_hctxs(q); | |
3640 | blk_mq_sysfs_unregister(q); | |
3641 | } | |
3642 | ||
a2584e43 | 3643 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3644 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3645 | 0) | |
3646 | goto reregister; | |
3647 | ||
868f2f0b | 3648 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3649 | fallback: |
aa880ad6 | 3650 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3651 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3652 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3653 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3654 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3655 | nr_hw_queues, prev_nr_hw_queues); | |
3656 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3657 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3658 | goto fallback; |
3659 | } | |
477e19de JW |
3660 | blk_mq_map_swqueue(q); |
3661 | } | |
3662 | ||
f7e76dbc | 3663 | reregister: |
477e19de JW |
3664 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3665 | blk_mq_sysfs_register(q); | |
3666 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3667 | } |
3668 | ||
d48ece20 JW |
3669 | switch_back: |
3670 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3671 | blk_mq_elv_switch_back(&head, q); | |
3672 | ||
868f2f0b KB |
3673 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3674 | blk_mq_unfreeze_queue(q); | |
3675 | } | |
e4dc2b32 KB |
3676 | |
3677 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3678 | { | |
3679 | mutex_lock(&set->tag_list_lock); | |
3680 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3681 | mutex_unlock(&set->tag_list_lock); | |
3682 | } | |
868f2f0b KB |
3683 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3684 | ||
34dbad5d OS |
3685 | /* Enable polling stats and return whether they were already enabled. */ |
3686 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3687 | { | |
3688 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3689 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3690 | return true; |
3691 | blk_stat_add_callback(q, q->poll_cb); | |
3692 | return false; | |
3693 | } | |
3694 | ||
3695 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3696 | { | |
3697 | /* | |
3698 | * We don't arm the callback if polling stats are not enabled or the | |
3699 | * callback is already active. | |
3700 | */ | |
3701 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3702 | blk_stat_is_active(q->poll_cb)) | |
3703 | return; | |
3704 | ||
3705 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3706 | } | |
3707 | ||
3708 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3709 | { | |
3710 | struct request_queue *q = cb->data; | |
720b8ccc | 3711 | int bucket; |
34dbad5d | 3712 | |
720b8ccc SB |
3713 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3714 | if (cb->stat[bucket].nr_samples) | |
3715 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3716 | } | |
34dbad5d OS |
3717 | } |
3718 | ||
64f1c21e | 3719 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3720 | struct request *rq) |
3721 | { | |
64f1c21e | 3722 | unsigned long ret = 0; |
720b8ccc | 3723 | int bucket; |
64f1c21e JA |
3724 | |
3725 | /* | |
3726 | * If stats collection isn't on, don't sleep but turn it on for | |
3727 | * future users | |
3728 | */ | |
34dbad5d | 3729 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3730 | return 0; |
3731 | ||
64f1c21e JA |
3732 | /* |
3733 | * As an optimistic guess, use half of the mean service time | |
3734 | * for this type of request. We can (and should) make this smarter. | |
3735 | * For instance, if the completion latencies are tight, we can | |
3736 | * get closer than just half the mean. This is especially | |
3737 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3738 | * than ~10 usec. We do use the stats for the relevant IO size |
3739 | * if available which does lead to better estimates. | |
64f1c21e | 3740 | */ |
720b8ccc SB |
3741 | bucket = blk_mq_poll_stats_bkt(rq); |
3742 | if (bucket < 0) | |
3743 | return ret; | |
3744 | ||
3745 | if (q->poll_stat[bucket].nr_samples) | |
3746 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3747 | |
3748 | return ret; | |
3749 | } | |
3750 | ||
06426adf JA |
3751 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3752 | struct request *rq) | |
3753 | { | |
3754 | struct hrtimer_sleeper hs; | |
3755 | enum hrtimer_mode mode; | |
64f1c21e | 3756 | unsigned int nsecs; |
06426adf JA |
3757 | ktime_t kt; |
3758 | ||
76a86f9d | 3759 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3760 | return false; |
3761 | ||
3762 | /* | |
1052b8ac | 3763 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3764 | * |
64f1c21e JA |
3765 | * 0: use half of prev avg |
3766 | * >0: use this specific value | |
3767 | */ | |
1052b8ac | 3768 | if (q->poll_nsec > 0) |
64f1c21e JA |
3769 | nsecs = q->poll_nsec; |
3770 | else | |
cae740a0 | 3771 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3772 | |
3773 | if (!nsecs) | |
06426adf JA |
3774 | return false; |
3775 | ||
76a86f9d | 3776 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3777 | |
3778 | /* | |
3779 | * This will be replaced with the stats tracking code, using | |
3780 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3781 | */ | |
8b0e1953 | 3782 | kt = nsecs; |
06426adf JA |
3783 | |
3784 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3785 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3786 | hrtimer_set_expires(&hs.timer, kt); |
3787 | ||
06426adf | 3788 | do { |
5a61c363 | 3789 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3790 | break; |
3791 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3792 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3793 | if (hs.task) |
3794 | io_schedule(); | |
3795 | hrtimer_cancel(&hs.timer); | |
3796 | mode = HRTIMER_MODE_ABS; | |
3797 | } while (hs.task && !signal_pending(current)); | |
3798 | ||
3799 | __set_current_state(TASK_RUNNING); | |
3800 | destroy_hrtimer_on_stack(&hs.timer); | |
3801 | return true; | |
3802 | } | |
3803 | ||
1052b8ac JA |
3804 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3805 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3806 | { |
1052b8ac JA |
3807 | struct request *rq; |
3808 | ||
29ece8b4 | 3809 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3810 | return false; |
3811 | ||
3812 | if (!blk_qc_t_is_internal(cookie)) | |
3813 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3814 | else { | |
3815 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3816 | /* | |
3817 | * With scheduling, if the request has completed, we'll | |
3818 | * get a NULL return here, as we clear the sched tag when | |
3819 | * that happens. The request still remains valid, like always, | |
3820 | * so we should be safe with just the NULL check. | |
3821 | */ | |
3822 | if (!rq) | |
3823 | return false; | |
3824 | } | |
3825 | ||
cae740a0 | 3826 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3827 | } |
3828 | ||
529262d5 CH |
3829 | /** |
3830 | * blk_poll - poll for IO completions | |
3831 | * @q: the queue | |
3832 | * @cookie: cookie passed back at IO submission time | |
3833 | * @spin: whether to spin for completions | |
3834 | * | |
3835 | * Description: | |
3836 | * Poll for completions on the passed in queue. Returns number of | |
3837 | * completed entries found. If @spin is true, then blk_poll will continue | |
3838 | * looping until at least one completion is found, unless the task is | |
3839 | * otherwise marked running (or we need to reschedule). | |
3840 | */ | |
3841 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3842 | { |
3843 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3844 | long state; |
3845 | ||
529262d5 CH |
3846 | if (!blk_qc_t_valid(cookie) || |
3847 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3848 | return 0; |
3849 | ||
529262d5 CH |
3850 | if (current->plug) |
3851 | blk_flush_plug_list(current->plug, false); | |
3852 | ||
1052b8ac JA |
3853 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3854 | ||
06426adf JA |
3855 | /* |
3856 | * If we sleep, have the caller restart the poll loop to reset | |
3857 | * the state. Like for the other success return cases, the | |
3858 | * caller is responsible for checking if the IO completed. If | |
3859 | * the IO isn't complete, we'll get called again and will go | |
f6f371f7 PB |
3860 | * straight to the busy poll loop. If specified not to spin, |
3861 | * we also should not sleep. | |
06426adf | 3862 | */ |
f6f371f7 | 3863 | if (spin && blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3864 | return 1; |
06426adf | 3865 | |
bbd7bb70 JA |
3866 | hctx->poll_considered++; |
3867 | ||
3868 | state = current->state; | |
aa61bec3 | 3869 | do { |
bbd7bb70 JA |
3870 | int ret; |
3871 | ||
3872 | hctx->poll_invoked++; | |
3873 | ||
9743139c | 3874 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3875 | if (ret > 0) { |
3876 | hctx->poll_success++; | |
849a3700 | 3877 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3878 | return ret; |
bbd7bb70 JA |
3879 | } |
3880 | ||
3881 | if (signal_pending_state(state, current)) | |
849a3700 | 3882 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3883 | |
3884 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3885 | return 1; |
0a1b8b87 | 3886 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3887 | break; |
3888 | cpu_relax(); | |
aa61bec3 | 3889 | } while (!need_resched()); |
bbd7bb70 | 3890 | |
67b4110f | 3891 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3892 | return 0; |
bbd7bb70 | 3893 | } |
529262d5 | 3894 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3895 | |
9cf2bab6 JA |
3896 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3897 | { | |
3898 | return rq->mq_ctx->cpu; | |
3899 | } | |
3900 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3901 | ||
320ae51f JA |
3902 | static int __init blk_mq_init(void) |
3903 | { | |
c3077b5d CH |
3904 | int i; |
3905 | ||
3906 | for_each_possible_cpu(i) | |
3907 | INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i)); | |
3908 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); | |
3909 | ||
3910 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
3911 | "block/softirq:dead", NULL, | |
3912 | blk_softirq_cpu_dead); | |
9467f859 TG |
3913 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3914 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
3915 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
3916 | blk_mq_hctx_notify_online, | |
3917 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
3918 | return 0; |
3919 | } | |
3920 | subsys_initcall(blk_mq_init); |