]>
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 | |
f9ab4918 | 44 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 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 | |
aec89dc5 | 191 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 192 | { |
7996a8b5 | 193 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
194 | if (force_atomic) |
195 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
196 | q->mq_freeze_depth--; |
197 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
198 | if (!q->mq_freeze_depth) { | |
bdd63160 | 199 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 200 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 201 | } |
7996a8b5 | 202 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 203 | } |
aec89dc5 CH |
204 | |
205 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
206 | { | |
207 | __blk_mq_unfreeze_queue(q, false); | |
208 | } | |
b4c6a028 | 209 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 210 | |
852ec809 BVA |
211 | /* |
212 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
213 | * mpt3sas driver such that this function can be removed. | |
214 | */ | |
215 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
216 | { | |
8814ce8a | 217 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
218 | } |
219 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
220 | ||
6a83e74d | 221 | /** |
69e07c4a | 222 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
223 | * @q: request queue. |
224 | * | |
225 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
226 | * callback function is invoked. Once this function is returned, we make |
227 | * sure no dispatch can happen until the queue is unquiesced via | |
228 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
229 | */ |
230 | void blk_mq_quiesce_queue(struct request_queue *q) | |
231 | { | |
232 | struct blk_mq_hw_ctx *hctx; | |
233 | unsigned int i; | |
234 | bool rcu = false; | |
235 | ||
1d9e9bc6 | 236 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 237 | |
6a83e74d BVA |
238 | queue_for_each_hw_ctx(q, hctx, i) { |
239 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 240 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
241 | else |
242 | rcu = true; | |
243 | } | |
244 | if (rcu) | |
245 | synchronize_rcu(); | |
246 | } | |
247 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
248 | ||
e4e73913 ML |
249 | /* |
250 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
251 | * @q: request queue. | |
252 | * | |
253 | * This function recovers queue into the state before quiescing | |
254 | * which is done by blk_mq_quiesce_queue. | |
255 | */ | |
256 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
257 | { | |
8814ce8a | 258 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 259 | |
1d9e9bc6 ML |
260 | /* dispatch requests which are inserted during quiescing */ |
261 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
262 | } |
263 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
264 | ||
aed3ea94 JA |
265 | void blk_mq_wake_waiters(struct request_queue *q) |
266 | { | |
267 | struct blk_mq_hw_ctx *hctx; | |
268 | unsigned int i; | |
269 | ||
270 | queue_for_each_hw_ctx(q, hctx, i) | |
271 | if (blk_mq_hw_queue_mapped(hctx)) | |
272 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
273 | } | |
274 | ||
fe1f4526 | 275 | /* |
9a91b05b HT |
276 | * Only need start/end time stamping if we have iostat or |
277 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
278 | */ |
279 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
280 | { | |
9a91b05b | 281 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
282 | } |
283 | ||
e4cdf1a1 | 284 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 285 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 286 | { |
e4cdf1a1 CH |
287 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
288 | struct request *rq = tags->static_rqs[tag]; | |
c3a148d2 | 289 | |
42fdc5e4 | 290 | if (data->q->elevator) { |
76647368 | 291 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
292 | rq->internal_tag = tag; |
293 | } else { | |
e4cdf1a1 | 294 | rq->tag = tag; |
76647368 | 295 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
296 | } |
297 | ||
af76e555 | 298 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
299 | rq->q = data->q; |
300 | rq->mq_ctx = data->ctx; | |
ea4f995e | 301 | rq->mq_hctx = data->hctx; |
568f2700 | 302 | rq->rq_flags = 0; |
7ea4d8a4 | 303 | rq->cmd_flags = data->cmd_flags; |
0854bcdc BVA |
304 | if (data->flags & BLK_MQ_REQ_PM) |
305 | rq->rq_flags |= RQF_PM; | |
e4cdf1a1 | 306 | if (blk_queue_io_stat(data->q)) |
e8064021 | 307 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 308 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
309 | INIT_HLIST_NODE(&rq->hash); |
310 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
311 | rq->rq_disk = NULL; |
312 | rq->part = NULL; | |
6f816b4b TH |
313 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
314 | rq->alloc_time_ns = alloc_time_ns; | |
315 | #endif | |
fe1f4526 JA |
316 | if (blk_mq_need_time_stamp(rq)) |
317 | rq->start_time_ns = ktime_get_ns(); | |
318 | else | |
319 | rq->start_time_ns = 0; | |
544ccc8d | 320 | rq->io_start_time_ns = 0; |
3d244306 | 321 | rq->stats_sectors = 0; |
af76e555 CH |
322 | rq->nr_phys_segments = 0; |
323 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
324 | rq->nr_integrity_segments = 0; | |
325 | #endif | |
a892c8d5 | 326 | blk_crypto_rq_set_defaults(rq); |
af76e555 | 327 | /* tag was already set */ |
079076b3 | 328 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 329 | |
f6be4fb4 JA |
330 | rq->timeout = 0; |
331 | ||
af76e555 CH |
332 | rq->end_io = NULL; |
333 | rq->end_io_data = NULL; | |
af76e555 | 334 | |
7ea4d8a4 | 335 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; |
12f5b931 | 336 | refcount_set(&rq->ref, 1); |
7ea4d8a4 CH |
337 | |
338 | if (!op_is_flush(data->cmd_flags)) { | |
339 | struct elevator_queue *e = data->q->elevator; | |
340 | ||
341 | rq->elv.icq = NULL; | |
342 | if (e && e->type->ops.prepare_request) { | |
343 | if (e->type->icq_cache) | |
344 | blk_mq_sched_assign_ioc(rq); | |
345 | ||
346 | e->type->ops.prepare_request(rq); | |
347 | rq->rq_flags |= RQF_ELVPRIV; | |
348 | } | |
349 | } | |
350 | ||
351 | data->hctx->queued++; | |
e4cdf1a1 | 352 | return rq; |
5dee8577 CH |
353 | } |
354 | ||
e6e7abff | 355 | static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data) |
d2c0d383 | 356 | { |
e6e7abff | 357 | struct request_queue *q = data->q; |
d2c0d383 | 358 | struct elevator_queue *e = q->elevator; |
6f816b4b | 359 | u64 alloc_time_ns = 0; |
600c3b0c | 360 | unsigned int tag; |
d2c0d383 | 361 | |
6f816b4b TH |
362 | /* alloc_time includes depth and tag waits */ |
363 | if (blk_queue_rq_alloc_time(q)) | |
364 | alloc_time_ns = ktime_get_ns(); | |
365 | ||
f9afca4d | 366 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 367 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
368 | |
369 | if (e) { | |
d2c0d383 | 370 | /* |
8d663f34 | 371 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
372 | * dispatch list. Don't include reserved tags in the |
373 | * limiting, as it isn't useful. | |
d2c0d383 | 374 | */ |
f9afca4d | 375 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 376 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 377 | e->type->ops.limit_depth && |
17a51199 | 378 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 379 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
380 | } |
381 | ||
bf0beec0 | 382 | retry: |
600c3b0c CH |
383 | data->ctx = blk_mq_get_ctx(q); |
384 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 385 | if (!e) |
600c3b0c CH |
386 | blk_mq_tag_busy(data->hctx); |
387 | ||
bf0beec0 ML |
388 | /* |
389 | * Waiting allocations only fail because of an inactive hctx. In that | |
390 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
391 | * should have migrated us to an online CPU by now. | |
392 | */ | |
e4cdf1a1 | 393 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
394 | if (tag == BLK_MQ_NO_TAG) { |
395 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
396 | return NULL; | |
397 | ||
398 | /* | |
399 | * Give up the CPU and sleep for a random short time to ensure | |
400 | * that thread using a realtime scheduling class are migrated | |
70f15a4f | 401 | * off the CPU, and thus off the hctx that is going away. |
bf0beec0 ML |
402 | */ |
403 | msleep(3); | |
404 | goto retry; | |
405 | } | |
7ea4d8a4 | 406 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
407 | } |
408 | ||
cd6ce148 | 409 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 410 | blk_mq_req_flags_t flags) |
320ae51f | 411 | { |
e6e7abff CH |
412 | struct blk_mq_alloc_data data = { |
413 | .q = q, | |
414 | .flags = flags, | |
415 | .cmd_flags = op, | |
416 | }; | |
bd166ef1 | 417 | struct request *rq; |
a492f075 | 418 | int ret; |
320ae51f | 419 | |
3a0a5299 | 420 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
421 | if (ret) |
422 | return ERR_PTR(ret); | |
320ae51f | 423 | |
e6e7abff | 424 | rq = __blk_mq_alloc_request(&data); |
bd166ef1 | 425 | if (!rq) |
a5ea5811 | 426 | goto out_queue_exit; |
0c4de0f3 CH |
427 | rq->__data_len = 0; |
428 | rq->__sector = (sector_t) -1; | |
429 | rq->bio = rq->biotail = NULL; | |
320ae51f | 430 | return rq; |
a5ea5811 CH |
431 | out_queue_exit: |
432 | blk_queue_exit(q); | |
433 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 434 | } |
4bb659b1 | 435 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 436 | |
cd6ce148 | 437 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 438 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 439 | { |
e6e7abff CH |
440 | struct blk_mq_alloc_data data = { |
441 | .q = q, | |
442 | .flags = flags, | |
443 | .cmd_flags = op, | |
444 | }; | |
600c3b0c | 445 | u64 alloc_time_ns = 0; |
6d2809d5 | 446 | unsigned int cpu; |
600c3b0c | 447 | unsigned int tag; |
1f5bd336 ML |
448 | int ret; |
449 | ||
600c3b0c CH |
450 | /* alloc_time includes depth and tag waits */ |
451 | if (blk_queue_rq_alloc_time(q)) | |
452 | alloc_time_ns = ktime_get_ns(); | |
453 | ||
1f5bd336 ML |
454 | /* |
455 | * If the tag allocator sleeps we could get an allocation for a | |
456 | * different hardware context. No need to complicate the low level | |
457 | * allocator for this for the rare use case of a command tied to | |
458 | * a specific queue. | |
459 | */ | |
600c3b0c | 460 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
461 | return ERR_PTR(-EINVAL); |
462 | ||
463 | if (hctx_idx >= q->nr_hw_queues) | |
464 | return ERR_PTR(-EIO); | |
465 | ||
3a0a5299 | 466 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
467 | if (ret) |
468 | return ERR_PTR(ret); | |
469 | ||
c8712c6a CH |
470 | /* |
471 | * Check if the hardware context is actually mapped to anything. | |
472 | * If not tell the caller that it should skip this queue. | |
473 | */ | |
a5ea5811 | 474 | ret = -EXDEV; |
e6e7abff CH |
475 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
476 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 477 | goto out_queue_exit; |
e6e7abff CH |
478 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
479 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 480 | |
42fdc5e4 | 481 | if (!q->elevator) |
600c3b0c CH |
482 | blk_mq_tag_busy(data.hctx); |
483 | ||
a5ea5811 | 484 | ret = -EWOULDBLOCK; |
600c3b0c CH |
485 | tag = blk_mq_get_tag(&data); |
486 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 487 | goto out_queue_exit; |
600c3b0c CH |
488 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
489 | ||
a5ea5811 CH |
490 | out_queue_exit: |
491 | blk_queue_exit(q); | |
492 | return ERR_PTR(ret); | |
1f5bd336 ML |
493 | } |
494 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
495 | ||
12f5b931 KB |
496 | static void __blk_mq_free_request(struct request *rq) |
497 | { | |
498 | struct request_queue *q = rq->q; | |
499 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 500 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
501 | const int sched_tag = rq->internal_tag; |
502 | ||
a892c8d5 | 503 | blk_crypto_free_request(rq); |
986d413b | 504 | blk_pm_mark_last_busy(rq); |
ea4f995e | 505 | rq->mq_hctx = NULL; |
76647368 | 506 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 507 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 508 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 509 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
510 | blk_mq_sched_restart(hctx); |
511 | blk_queue_exit(q); | |
512 | } | |
513 | ||
6af54051 | 514 | void blk_mq_free_request(struct request *rq) |
320ae51f | 515 | { |
320ae51f | 516 | struct request_queue *q = rq->q; |
6af54051 CH |
517 | struct elevator_queue *e = q->elevator; |
518 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 519 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 520 | |
5bbf4e5a | 521 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
522 | if (e && e->type->ops.finish_request) |
523 | e->type->ops.finish_request(rq); | |
6af54051 CH |
524 | if (rq->elv.icq) { |
525 | put_io_context(rq->elv.icq->ioc); | |
526 | rq->elv.icq = NULL; | |
527 | } | |
528 | } | |
320ae51f | 529 | |
6af54051 | 530 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 531 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 532 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 533 | |
7beb2f84 | 534 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 535 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 536 | |
a7905043 | 537 | rq_qos_done(q, rq); |
0d2602ca | 538 | |
12f5b931 KB |
539 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
540 | if (refcount_dec_and_test(&rq->ref)) | |
541 | __blk_mq_free_request(rq); | |
320ae51f | 542 | } |
1a3b595a | 543 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 544 | |
2a842aca | 545 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 546 | { |
fe1f4526 JA |
547 | u64 now = 0; |
548 | ||
549 | if (blk_mq_need_time_stamp(rq)) | |
550 | now = ktime_get_ns(); | |
522a7775 | 551 | |
4bc6339a OS |
552 | if (rq->rq_flags & RQF_STATS) { |
553 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 554 | blk_stat_add(rq, now); |
4bc6339a OS |
555 | } |
556 | ||
87890092 | 557 | blk_mq_sched_completed_request(rq, now); |
ed88660a | 558 | |
522a7775 | 559 | blk_account_io_done(rq, now); |
0d11e6ac | 560 | |
91b63639 | 561 | if (rq->end_io) { |
a7905043 | 562 | rq_qos_done(rq->q, rq); |
320ae51f | 563 | rq->end_io(rq, error); |
91b63639 | 564 | } else { |
320ae51f | 565 | blk_mq_free_request(rq); |
91b63639 | 566 | } |
320ae51f | 567 | } |
c8a446ad | 568 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 569 | |
2a842aca | 570 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
571 | { |
572 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
573 | BUG(); | |
c8a446ad | 574 | __blk_mq_end_request(rq, error); |
63151a44 | 575 | } |
c8a446ad | 576 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 577 | |
f9ab4918 | 578 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 579 | { |
f9ab4918 SAS |
580 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
581 | struct request *rq, *next; | |
c3077b5d | 582 | |
f9ab4918 | 583 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 584 | rq->q->mq_ops->complete(rq); |
320ae51f | 585 | } |
320ae51f | 586 | |
f9ab4918 | 587 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 588 | { |
f9ab4918 | 589 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
590 | } |
591 | ||
c3077b5d CH |
592 | static int blk_softirq_cpu_dead(unsigned int cpu) |
593 | { | |
f9ab4918 | 594 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
595 | return 0; |
596 | } | |
597 | ||
40d09b53 | 598 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 599 | { |
f9ab4918 | 600 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
601 | } |
602 | ||
96339526 CH |
603 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
604 | { | |
605 | int cpu = raw_smp_processor_id(); | |
606 | ||
607 | if (!IS_ENABLED(CONFIG_SMP) || | |
608 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
609 | return false; | |
71425189 SAS |
610 | /* |
611 | * With force threaded interrupts enabled, raising softirq from an SMP | |
612 | * function call will always result in waking the ksoftirqd thread. | |
613 | * This is probably worse than completing the request on a different | |
614 | * cache domain. | |
615 | */ | |
91cc470e | 616 | if (force_irqthreads()) |
71425189 | 617 | return false; |
96339526 CH |
618 | |
619 | /* same CPU or cache domain? Complete locally */ | |
620 | if (cpu == rq->mq_ctx->cpu || | |
621 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
622 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
623 | return false; | |
624 | ||
625 | /* don't try to IPI to an offline CPU */ | |
626 | return cpu_online(rq->mq_ctx->cpu); | |
627 | } | |
628 | ||
f9ab4918 SAS |
629 | static void blk_mq_complete_send_ipi(struct request *rq) |
630 | { | |
631 | struct llist_head *list; | |
632 | unsigned int cpu; | |
633 | ||
634 | cpu = rq->mq_ctx->cpu; | |
635 | list = &per_cpu(blk_cpu_done, cpu); | |
636 | if (llist_add(&rq->ipi_list, list)) { | |
637 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
638 | smp_call_function_single_async(cpu, &rq->csd); | |
639 | } | |
640 | } | |
641 | ||
642 | static void blk_mq_raise_softirq(struct request *rq) | |
643 | { | |
644 | struct llist_head *list; | |
645 | ||
646 | preempt_disable(); | |
647 | list = this_cpu_ptr(&blk_cpu_done); | |
648 | if (llist_add(&rq->ipi_list, list)) | |
649 | raise_softirq(BLOCK_SOFTIRQ); | |
650 | preempt_enable(); | |
651 | } | |
652 | ||
40d09b53 | 653 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 654 | { |
af78ff7c | 655 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 656 | |
4ab32bf3 JA |
657 | /* |
658 | * For a polled request, always complete locallly, it's pointless | |
659 | * to redirect the completion. | |
660 | */ | |
40d09b53 CH |
661 | if (rq->cmd_flags & REQ_HIPRI) |
662 | return false; | |
38535201 | 663 | |
96339526 | 664 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
665 | blk_mq_complete_send_ipi(rq); |
666 | return true; | |
3d6efbf6 | 667 | } |
40d09b53 | 668 | |
f9ab4918 SAS |
669 | if (rq->q->nr_hw_queues == 1) { |
670 | blk_mq_raise_softirq(rq); | |
671 | return true; | |
672 | } | |
673 | return false; | |
40d09b53 CH |
674 | } |
675 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
676 | ||
677 | /** | |
678 | * blk_mq_complete_request - end I/O on a request | |
679 | * @rq: the request being processed | |
680 | * | |
681 | * Description: | |
682 | * Complete a request by scheduling the ->complete_rq operation. | |
683 | **/ | |
684 | void blk_mq_complete_request(struct request *rq) | |
685 | { | |
686 | if (!blk_mq_complete_request_remote(rq)) | |
687 | rq->q->mq_ops->complete(rq); | |
320ae51f | 688 | } |
15f73f5b | 689 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 690 | |
04ced159 | 691 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 692 | __releases(hctx->srcu) |
04ced159 JA |
693 | { |
694 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
695 | rcu_read_unlock(); | |
696 | else | |
05707b64 | 697 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
698 | } |
699 | ||
700 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 701 | __acquires(hctx->srcu) |
04ced159 | 702 | { |
08b5a6e2 JA |
703 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
704 | /* shut up gcc false positive */ | |
705 | *srcu_idx = 0; | |
04ced159 | 706 | rcu_read_lock(); |
08b5a6e2 | 707 | } else |
05707b64 | 708 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
709 | } |
710 | ||
105663f7 AA |
711 | /** |
712 | * blk_mq_start_request - Start processing a request | |
713 | * @rq: Pointer to request to be started | |
714 | * | |
715 | * Function used by device drivers to notify the block layer that a request | |
716 | * is going to be processed now, so blk layer can do proper initializations | |
717 | * such as starting the timeout timer. | |
718 | */ | |
e2490073 | 719 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
720 | { |
721 | struct request_queue *q = rq->q; | |
722 | ||
a54895fa | 723 | trace_block_rq_issue(rq); |
320ae51f | 724 | |
cf43e6be | 725 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 726 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 727 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 728 | rq->rq_flags |= RQF_STATS; |
a7905043 | 729 | rq_qos_issue(q, rq); |
cf43e6be JA |
730 | } |
731 | ||
1d9bd516 | 732 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 733 | |
1d9bd516 | 734 | blk_add_timer(rq); |
12f5b931 | 735 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 736 | |
54d4e6ab MG |
737 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
738 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
739 | q->integrity.profile->prepare_fn(rq); | |
740 | #endif | |
320ae51f | 741 | } |
e2490073 | 742 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 743 | |
ed0791b2 | 744 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
745 | { |
746 | struct request_queue *q = rq->q; | |
747 | ||
923218f6 ML |
748 | blk_mq_put_driver_tag(rq); |
749 | ||
a54895fa | 750 | trace_block_rq_requeue(rq); |
a7905043 | 751 | rq_qos_requeue(q, rq); |
49f5baa5 | 752 | |
12f5b931 KB |
753 | if (blk_mq_request_started(rq)) { |
754 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 755 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 756 | } |
320ae51f JA |
757 | } |
758 | ||
2b053aca | 759 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 760 | { |
ed0791b2 | 761 | __blk_mq_requeue_request(rq); |
ed0791b2 | 762 | |
105976f5 ML |
763 | /* this request will be re-inserted to io scheduler queue */ |
764 | blk_mq_sched_requeue_request(rq); | |
765 | ||
2b053aca | 766 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
767 | } |
768 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
769 | ||
6fca6a61 CH |
770 | static void blk_mq_requeue_work(struct work_struct *work) |
771 | { | |
772 | struct request_queue *q = | |
2849450a | 773 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
774 | LIST_HEAD(rq_list); |
775 | struct request *rq, *next; | |
6fca6a61 | 776 | |
18e9781d | 777 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 778 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 779 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
780 | |
781 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 782 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
783 | continue; |
784 | ||
e8064021 | 785 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 786 | list_del_init(&rq->queuelist); |
aef1897c JW |
787 | /* |
788 | * If RQF_DONTPREP, rq has contained some driver specific | |
789 | * data, so insert it to hctx dispatch list to avoid any | |
790 | * merge. | |
791 | */ | |
792 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 793 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
794 | else |
795 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
796 | } |
797 | ||
798 | while (!list_empty(&rq_list)) { | |
799 | rq = list_entry(rq_list.next, struct request, queuelist); | |
800 | list_del_init(&rq->queuelist); | |
9e97d295 | 801 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
802 | } |
803 | ||
52d7f1b5 | 804 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
805 | } |
806 | ||
2b053aca BVA |
807 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
808 | bool kick_requeue_list) | |
6fca6a61 CH |
809 | { |
810 | struct request_queue *q = rq->q; | |
811 | unsigned long flags; | |
812 | ||
813 | /* | |
814 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 815 | * request head insertion from the workqueue. |
6fca6a61 | 816 | */ |
e8064021 | 817 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
818 | |
819 | spin_lock_irqsave(&q->requeue_lock, flags); | |
820 | if (at_head) { | |
e8064021 | 821 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
822 | list_add(&rq->queuelist, &q->requeue_list); |
823 | } else { | |
824 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
825 | } | |
826 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
827 | |
828 | if (kick_requeue_list) | |
829 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 830 | } |
6fca6a61 CH |
831 | |
832 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
833 | { | |
ae943d20 | 834 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
835 | } |
836 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
837 | ||
2849450a MS |
838 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
839 | unsigned long msecs) | |
840 | { | |
d4acf365 BVA |
841 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
842 | msecs_to_jiffies(msecs)); | |
2849450a MS |
843 | } |
844 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
845 | ||
0e62f51f JA |
846 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
847 | { | |
88c7b2b7 JA |
848 | if (tag < tags->nr_tags) { |
849 | prefetch(tags->rqs[tag]); | |
4ee86bab | 850 | return tags->rqs[tag]; |
88c7b2b7 | 851 | } |
4ee86bab HR |
852 | |
853 | return NULL; | |
24d2f903 CH |
854 | } |
855 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
856 | ||
3c94d83c JA |
857 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
858 | void *priv, bool reserved) | |
ae879912 JA |
859 | { |
860 | /* | |
05a4fed6 | 861 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 862 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 863 | */ |
05a4fed6 | 864 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
865 | bool *busy = priv; |
866 | ||
867 | *busy = true; | |
868 | return false; | |
869 | } | |
870 | ||
871 | return true; | |
872 | } | |
873 | ||
3c94d83c | 874 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
875 | { |
876 | bool busy = false; | |
877 | ||
3c94d83c | 878 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
879 | return busy; |
880 | } | |
3c94d83c | 881 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 882 | |
358f70da | 883 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 884 | { |
da661267 | 885 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
886 | if (req->q->mq_ops->timeout) { |
887 | enum blk_eh_timer_return ret; | |
888 | ||
889 | ret = req->q->mq_ops->timeout(req, reserved); | |
890 | if (ret == BLK_EH_DONE) | |
891 | return; | |
892 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 893 | } |
d1210d5a CH |
894 | |
895 | blk_add_timer(req); | |
87ee7b11 | 896 | } |
5b3f25fc | 897 | |
12f5b931 | 898 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 899 | { |
12f5b931 | 900 | unsigned long deadline; |
87ee7b11 | 901 | |
12f5b931 KB |
902 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
903 | return false; | |
da661267 CH |
904 | if (rq->rq_flags & RQF_TIMED_OUT) |
905 | return false; | |
a7af0af3 | 906 | |
079076b3 | 907 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
908 | if (time_after_eq(jiffies, deadline)) |
909 | return true; | |
a7af0af3 | 910 | |
12f5b931 KB |
911 | if (*next == 0) |
912 | *next = deadline; | |
913 | else if (time_after(*next, deadline)) | |
914 | *next = deadline; | |
915 | return false; | |
87ee7b11 JA |
916 | } |
917 | ||
2e315dc0 ML |
918 | void blk_mq_put_rq_ref(struct request *rq) |
919 | { | |
a9ed27a7 | 920 | if (is_flush_rq(rq)) |
2e315dc0 ML |
921 | rq->end_io(rq, 0); |
922 | else if (refcount_dec_and_test(&rq->ref)) | |
923 | __blk_mq_free_request(rq); | |
924 | } | |
925 | ||
7baa8572 | 926 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
927 | struct request *rq, void *priv, bool reserved) |
928 | { | |
12f5b931 KB |
929 | unsigned long *next = priv; |
930 | ||
931 | /* | |
c797b40c ML |
932 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
933 | * be reallocated underneath the timeout handler's processing, then | |
934 | * the expire check is reliable. If the request is not expired, then | |
935 | * it was completed and reallocated as a new request after returning | |
936 | * from blk_mq_check_expired(). | |
1d9bd516 | 937 | */ |
12f5b931 | 938 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 939 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 940 | return true; |
1d9bd516 TH |
941 | } |
942 | ||
287922eb | 943 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 944 | { |
287922eb CH |
945 | struct request_queue *q = |
946 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 947 | unsigned long next = 0; |
1d9bd516 | 948 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 949 | int i; |
320ae51f | 950 | |
71f79fb3 GKB |
951 | /* A deadlock might occur if a request is stuck requiring a |
952 | * timeout at the same time a queue freeze is waiting | |
953 | * completion, since the timeout code would not be able to | |
954 | * acquire the queue reference here. | |
955 | * | |
956 | * That's why we don't use blk_queue_enter here; instead, we use | |
957 | * percpu_ref_tryget directly, because we need to be able to | |
958 | * obtain a reference even in the short window between the queue | |
959 | * starting to freeze, by dropping the first reference in | |
1671d522 | 960 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
961 | * consumed, marked by the instant q_usage_counter reaches |
962 | * zero. | |
963 | */ | |
964 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
965 | return; |
966 | ||
12f5b931 | 967 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 968 | |
12f5b931 KB |
969 | if (next != 0) { |
970 | mod_timer(&q->timeout, next); | |
0d2602ca | 971 | } else { |
fcd36c36 BVA |
972 | /* |
973 | * Request timeouts are handled as a forward rolling timer. If | |
974 | * we end up here it means that no requests are pending and | |
975 | * also that no request has been pending for a while. Mark | |
976 | * each hctx as idle. | |
977 | */ | |
f054b56c ML |
978 | queue_for_each_hw_ctx(q, hctx, i) { |
979 | /* the hctx may be unmapped, so check it here */ | |
980 | if (blk_mq_hw_queue_mapped(hctx)) | |
981 | blk_mq_tag_idle(hctx); | |
982 | } | |
0d2602ca | 983 | } |
287922eb | 984 | blk_queue_exit(q); |
320ae51f JA |
985 | } |
986 | ||
88459642 OS |
987 | struct flush_busy_ctx_data { |
988 | struct blk_mq_hw_ctx *hctx; | |
989 | struct list_head *list; | |
990 | }; | |
991 | ||
992 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
993 | { | |
994 | struct flush_busy_ctx_data *flush_data = data; | |
995 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
996 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 997 | enum hctx_type type = hctx->type; |
88459642 | 998 | |
88459642 | 999 | spin_lock(&ctx->lock); |
c16d6b5a | 1000 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1001 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1002 | spin_unlock(&ctx->lock); |
1003 | return true; | |
1004 | } | |
1005 | ||
1429d7c9 JA |
1006 | /* |
1007 | * Process software queues that have been marked busy, splicing them | |
1008 | * to the for-dispatch | |
1009 | */ | |
2c3ad667 | 1010 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1011 | { |
88459642 OS |
1012 | struct flush_busy_ctx_data data = { |
1013 | .hctx = hctx, | |
1014 | .list = list, | |
1015 | }; | |
1429d7c9 | 1016 | |
88459642 | 1017 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1018 | } |
2c3ad667 | 1019 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1020 | |
b347689f ML |
1021 | struct dispatch_rq_data { |
1022 | struct blk_mq_hw_ctx *hctx; | |
1023 | struct request *rq; | |
1024 | }; | |
1025 | ||
1026 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1027 | void *data) | |
1028 | { | |
1029 | struct dispatch_rq_data *dispatch_data = data; | |
1030 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1031 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1032 | enum hctx_type type = hctx->type; |
b347689f ML |
1033 | |
1034 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1035 | if (!list_empty(&ctx->rq_lists[type])) { |
1036 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1037 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1038 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1039 | sbitmap_clear_bit(sb, bitnr); |
1040 | } | |
1041 | spin_unlock(&ctx->lock); | |
1042 | ||
1043 | return !dispatch_data->rq; | |
1044 | } | |
1045 | ||
1046 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1047 | struct blk_mq_ctx *start) | |
1048 | { | |
f31967f0 | 1049 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1050 | struct dispatch_rq_data data = { |
1051 | .hctx = hctx, | |
1052 | .rq = NULL, | |
1053 | }; | |
1054 | ||
1055 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1056 | dispatch_rq_from_ctx, &data); | |
1057 | ||
1058 | return data.rq; | |
1059 | } | |
1060 | ||
703fd1c0 JA |
1061 | static inline unsigned int queued_to_index(unsigned int queued) |
1062 | { | |
1063 | if (!queued) | |
1064 | return 0; | |
1429d7c9 | 1065 | |
703fd1c0 | 1066 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1067 | } |
1068 | ||
570e9b73 ML |
1069 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1070 | { | |
222a5ae0 | 1071 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1072 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1073 | int tag; |
1074 | ||
568f2700 ML |
1075 | blk_mq_tag_busy(rq->mq_hctx); |
1076 | ||
570e9b73 | 1077 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
222a5ae0 | 1078 | bt = rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1079 | tag_offset = 0; |
28500850 ML |
1080 | } else { |
1081 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1082 | return false; | |
570e9b73 ML |
1083 | } |
1084 | ||
570e9b73 ML |
1085 | tag = __sbitmap_queue_get(bt); |
1086 | if (tag == BLK_MQ_NO_TAG) | |
1087 | return false; | |
1088 | ||
1089 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1090 | return true; |
1091 | } | |
1092 | ||
61347154 | 1093 | bool blk_mq_get_driver_tag(struct request *rq) |
570e9b73 | 1094 | { |
568f2700 ML |
1095 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1096 | ||
1097 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1098 | return false; | |
1099 | ||
51db1c37 | 1100 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1101 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1102 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1103 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1104 | } |
1105 | hctx->tags->rqs[rq->tag] = rq; | |
1106 | return true; | |
570e9b73 ML |
1107 | } |
1108 | ||
eb619fdb JA |
1109 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1110 | int flags, void *key) | |
da55f2cc OS |
1111 | { |
1112 | struct blk_mq_hw_ctx *hctx; | |
1113 | ||
1114 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1115 | ||
5815839b | 1116 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1117 | if (!list_empty(&wait->entry)) { |
1118 | struct sbitmap_queue *sbq; | |
1119 | ||
1120 | list_del_init(&wait->entry); | |
222a5ae0 | 1121 | sbq = hctx->tags->bitmap_tags; |
e8618575 JA |
1122 | atomic_dec(&sbq->ws_active); |
1123 | } | |
5815839b ML |
1124 | spin_unlock(&hctx->dispatch_wait_lock); |
1125 | ||
da55f2cc OS |
1126 | blk_mq_run_hw_queue(hctx, true); |
1127 | return 1; | |
1128 | } | |
1129 | ||
f906a6a0 JA |
1130 | /* |
1131 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1132 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1133 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1134 | * marking us as waiting. |
1135 | */ | |
2278d69f | 1136 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1137 | struct request *rq) |
da55f2cc | 1138 | { |
222a5ae0 | 1139 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; |
5815839b | 1140 | struct wait_queue_head *wq; |
f906a6a0 JA |
1141 | wait_queue_entry_t *wait; |
1142 | bool ret; | |
da55f2cc | 1143 | |
51db1c37 | 1144 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1145 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1146 | |
c27d53fb BVA |
1147 | /* |
1148 | * It's possible that a tag was freed in the window between the | |
1149 | * allocation failure and adding the hardware queue to the wait | |
1150 | * queue. | |
1151 | * | |
1152 | * Don't clear RESTART here, someone else could have set it. | |
1153 | * At most this will cost an extra queue run. | |
1154 | */ | |
8ab6bb9e | 1155 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1156 | } |
eb619fdb | 1157 | |
2278d69f | 1158 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1159 | if (!list_empty_careful(&wait->entry)) |
1160 | return false; | |
1161 | ||
e8618575 | 1162 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1163 | |
1164 | spin_lock_irq(&wq->lock); | |
1165 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1166 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1167 | spin_unlock(&hctx->dispatch_wait_lock); |
1168 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1169 | return false; |
eb619fdb JA |
1170 | } |
1171 | ||
e8618575 | 1172 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1173 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1174 | __add_wait_queue(wq, wait); | |
c27d53fb | 1175 | |
da55f2cc | 1176 | /* |
eb619fdb JA |
1177 | * It's possible that a tag was freed in the window between the |
1178 | * allocation failure and adding the hardware queue to the wait | |
1179 | * queue. | |
da55f2cc | 1180 | */ |
8ab6bb9e | 1181 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1182 | if (!ret) { |
5815839b ML |
1183 | spin_unlock(&hctx->dispatch_wait_lock); |
1184 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1185 | return false; |
eb619fdb | 1186 | } |
c27d53fb BVA |
1187 | |
1188 | /* | |
1189 | * We got a tag, remove ourselves from the wait queue to ensure | |
1190 | * someone else gets the wakeup. | |
1191 | */ | |
c27d53fb | 1192 | list_del_init(&wait->entry); |
e8618575 | 1193 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1194 | spin_unlock(&hctx->dispatch_wait_lock); |
1195 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1196 | |
1197 | return true; | |
da55f2cc OS |
1198 | } |
1199 | ||
6e768717 ML |
1200 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1201 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1202 | /* | |
1203 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1204 | * - EWMA is one simple way to compute running average value | |
1205 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1206 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1207 | * factor doesn't matter because EWMA decreases exponentially | |
1208 | */ | |
1209 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1210 | { | |
1211 | unsigned int ewma; | |
1212 | ||
6e768717 ML |
1213 | ewma = hctx->dispatch_busy; |
1214 | ||
1215 | if (!ewma && !busy) | |
1216 | return; | |
1217 | ||
1218 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1219 | if (busy) | |
1220 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1221 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1222 | ||
1223 | hctx->dispatch_busy = ewma; | |
1224 | } | |
1225 | ||
86ff7c2a ML |
1226 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1227 | ||
c92a4103 JT |
1228 | static void blk_mq_handle_dev_resource(struct request *rq, |
1229 | struct list_head *list) | |
1230 | { | |
1231 | struct request *next = | |
1232 | list_first_entry_or_null(list, struct request, queuelist); | |
1233 | ||
1234 | /* | |
1235 | * If an I/O scheduler has been configured and we got a driver tag for | |
1236 | * the next request already, free it. | |
1237 | */ | |
1238 | if (next) | |
1239 | blk_mq_put_driver_tag(next); | |
1240 | ||
1241 | list_add(&rq->queuelist, list); | |
1242 | __blk_mq_requeue_request(rq); | |
1243 | } | |
1244 | ||
0512a75b KB |
1245 | static void blk_mq_handle_zone_resource(struct request *rq, |
1246 | struct list_head *zone_list) | |
1247 | { | |
1248 | /* | |
1249 | * If we end up here it is because we cannot dispatch a request to a | |
1250 | * specific zone due to LLD level zone-write locking or other zone | |
1251 | * related resource not being available. In this case, set the request | |
1252 | * aside in zone_list for retrying it later. | |
1253 | */ | |
1254 | list_add(&rq->queuelist, zone_list); | |
1255 | __blk_mq_requeue_request(rq); | |
1256 | } | |
1257 | ||
75383524 ML |
1258 | enum prep_dispatch { |
1259 | PREP_DISPATCH_OK, | |
1260 | PREP_DISPATCH_NO_TAG, | |
1261 | PREP_DISPATCH_NO_BUDGET, | |
1262 | }; | |
1263 | ||
1264 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1265 | bool need_budget) | |
1266 | { | |
1267 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1268 | int budget_token = -1; |
75383524 | 1269 | |
2a5a24aa ML |
1270 | if (need_budget) { |
1271 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1272 | if (budget_token < 0) { | |
1273 | blk_mq_put_driver_tag(rq); | |
1274 | return PREP_DISPATCH_NO_BUDGET; | |
1275 | } | |
1276 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1277 | } |
1278 | ||
1279 | if (!blk_mq_get_driver_tag(rq)) { | |
1280 | /* | |
1281 | * The initial allocation attempt failed, so we need to | |
1282 | * rerun the hardware queue when a tag is freed. The | |
1283 | * waitqueue takes care of that. If the queue is run | |
1284 | * before we add this entry back on the dispatch list, | |
1285 | * we'll re-run it below. | |
1286 | */ | |
1287 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1288 | /* |
1289 | * All budgets not got from this function will be put | |
1290 | * together during handling partial dispatch | |
1291 | */ | |
1292 | if (need_budget) | |
2a5a24aa | 1293 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1294 | return PREP_DISPATCH_NO_TAG; |
1295 | } | |
1296 | } | |
1297 | ||
1298 | return PREP_DISPATCH_OK; | |
1299 | } | |
1300 | ||
1fd40b5e ML |
1301 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1302 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1303 | struct list_head *list) |
1fd40b5e | 1304 | { |
2a5a24aa | 1305 | struct request *rq; |
1fd40b5e | 1306 | |
2a5a24aa ML |
1307 | list_for_each_entry(rq, list, queuelist) { |
1308 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1309 | |
2a5a24aa ML |
1310 | if (budget_token >= 0) |
1311 | blk_mq_put_dispatch_budget(q, budget_token); | |
1312 | } | |
1fd40b5e ML |
1313 | } |
1314 | ||
1f57f8d4 JA |
1315 | /* |
1316 | * Returns true if we did some work AND can potentially do more. | |
1317 | */ | |
445874e8 | 1318 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1319 | unsigned int nr_budgets) |
320ae51f | 1320 | { |
75383524 | 1321 | enum prep_dispatch prep; |
445874e8 | 1322 | struct request_queue *q = hctx->queue; |
6d6f167c | 1323 | struct request *rq, *nxt; |
fc17b653 | 1324 | int errors, queued; |
86ff7c2a | 1325 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1326 | LIST_HEAD(zone_list); |
9586e67b | 1327 | bool needs_resource = false; |
320ae51f | 1328 | |
81380ca1 OS |
1329 | if (list_empty(list)) |
1330 | return false; | |
1331 | ||
320ae51f JA |
1332 | /* |
1333 | * Now process all the entries, sending them to the driver. | |
1334 | */ | |
93efe981 | 1335 | errors = queued = 0; |
81380ca1 | 1336 | do { |
74c45052 | 1337 | struct blk_mq_queue_data bd; |
320ae51f | 1338 | |
f04c3df3 | 1339 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1340 | |
445874e8 | 1341 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1342 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1343 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1344 | break; |
de148297 | 1345 | |
320ae51f | 1346 | list_del_init(&rq->queuelist); |
320ae51f | 1347 | |
74c45052 | 1348 | bd.rq = rq; |
113285b4 JA |
1349 | |
1350 | /* | |
1351 | * Flag last if we have no more requests, or if we have more | |
1352 | * but can't assign a driver tag to it. | |
1353 | */ | |
1354 | if (list_empty(list)) | |
1355 | bd.last = true; | |
1356 | else { | |
113285b4 | 1357 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1358 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1359 | } |
74c45052 | 1360 | |
1fd40b5e ML |
1361 | /* |
1362 | * once the request is queued to lld, no need to cover the | |
1363 | * budget any more | |
1364 | */ | |
1365 | if (nr_budgets) | |
1366 | nr_budgets--; | |
74c45052 | 1367 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1368 | switch (ret) { |
1369 | case BLK_STS_OK: | |
1370 | queued++; | |
320ae51f | 1371 | break; |
7bf13729 | 1372 | case BLK_STS_RESOURCE: |
9586e67b NA |
1373 | needs_resource = true; |
1374 | fallthrough; | |
7bf13729 ML |
1375 | case BLK_STS_DEV_RESOURCE: |
1376 | blk_mq_handle_dev_resource(rq, list); | |
1377 | goto out; | |
1378 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1379 | /* |
1380 | * Move the request to zone_list and keep going through | |
1381 | * the dispatch list to find more requests the drive can | |
1382 | * accept. | |
1383 | */ | |
1384 | blk_mq_handle_zone_resource(rq, &zone_list); | |
9586e67b | 1385 | needs_resource = true; |
7bf13729 ML |
1386 | break; |
1387 | default: | |
93efe981 | 1388 | errors++; |
e21ee5a6 | 1389 | blk_mq_end_request(rq, ret); |
320ae51f | 1390 | } |
81380ca1 | 1391 | } while (!list_empty(list)); |
7bf13729 | 1392 | out: |
0512a75b KB |
1393 | if (!list_empty(&zone_list)) |
1394 | list_splice_tail_init(&zone_list, list); | |
1395 | ||
703fd1c0 | 1396 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1397 | |
632bfb63 | 1398 | /* If we didn't flush the entire list, we could have told the driver |
1399 | * there was more coming, but that turned out to be a lie. | |
1400 | */ | |
1401 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1402 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1403 | /* |
1404 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1405 | * that is where we will continue on next queue run. | |
1406 | */ | |
f04c3df3 | 1407 | if (!list_empty(list)) { |
86ff7c2a | 1408 | bool needs_restart; |
75383524 ML |
1409 | /* For non-shared tags, the RESTART check will suffice */ |
1410 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1411 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
86ff7c2a | 1412 | |
2a5a24aa ML |
1413 | if (nr_budgets) |
1414 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1415 | |
320ae51f | 1416 | spin_lock(&hctx->lock); |
01e99aec | 1417 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1418 | spin_unlock(&hctx->lock); |
f04c3df3 | 1419 | |
d7d8535f ML |
1420 | /* |
1421 | * Order adding requests to hctx->dispatch and checking | |
1422 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1423 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1424 | * miss the new added requests to hctx->dispatch, meantime | |
1425 | * SCHED_RESTART is observed here. | |
1426 | */ | |
1427 | smp_mb(); | |
1428 | ||
9ba52e58 | 1429 | /* |
710c785f BVA |
1430 | * If SCHED_RESTART was set by the caller of this function and |
1431 | * it is no longer set that means that it was cleared by another | |
1432 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1433 | * |
eb619fdb JA |
1434 | * If 'no_tag' is set, that means that we failed getting |
1435 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1436 | * waitqueue is no longer active, ensure that we run the queue | |
1437 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1438 | * |
710c785f BVA |
1439 | * If no I/O scheduler has been configured it is possible that |
1440 | * the hardware queue got stopped and restarted before requests | |
1441 | * were pushed back onto the dispatch list. Rerun the queue to | |
1442 | * avoid starvation. Notes: | |
1443 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1444 | * been stopped before rerunning a queue. | |
1445 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1446 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1447 | * and dm-rq. |
86ff7c2a ML |
1448 | * |
1449 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1450 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 | 1451 | * that could otherwise occur if the queue is idle. We'll do |
9586e67b NA |
1452 | * similar if we couldn't get budget or couldn't lock a zone |
1453 | * and SCHED_RESTART is set. | |
bd166ef1 | 1454 | */ |
86ff7c2a | 1455 | needs_restart = blk_mq_sched_needs_restart(hctx); |
9586e67b NA |
1456 | if (prep == PREP_DISPATCH_NO_BUDGET) |
1457 | needs_resource = true; | |
86ff7c2a | 1458 | if (!needs_restart || |
eb619fdb | 1459 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1460 | blk_mq_run_hw_queue(hctx, true); |
9586e67b | 1461 | else if (needs_restart && needs_resource) |
86ff7c2a | 1462 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1463 | |
6e768717 | 1464 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1465 | return false; |
6e768717 ML |
1466 | } else |
1467 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1468 | |
93efe981 | 1469 | return (queued + errors) != 0; |
f04c3df3 JA |
1470 | } |
1471 | ||
105663f7 AA |
1472 | /** |
1473 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1474 | * @hctx: Pointer to the hardware queue to run. | |
1475 | * | |
1476 | * Send pending requests to the hardware. | |
1477 | */ | |
6a83e74d BVA |
1478 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1479 | { | |
1480 | int srcu_idx; | |
1481 | ||
b7a71e66 JA |
1482 | /* |
1483 | * We can't run the queue inline with ints disabled. Ensure that | |
1484 | * we catch bad users of this early. | |
1485 | */ | |
1486 | WARN_ON_ONCE(in_interrupt()); | |
1487 | ||
04ced159 | 1488 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1489 | |
04ced159 JA |
1490 | hctx_lock(hctx, &srcu_idx); |
1491 | blk_mq_sched_dispatch_requests(hctx); | |
1492 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1493 | } |
1494 | ||
f82ddf19 ML |
1495 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1496 | { | |
1497 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1498 | ||
1499 | if (cpu >= nr_cpu_ids) | |
1500 | cpu = cpumask_first(hctx->cpumask); | |
1501 | return cpu; | |
1502 | } | |
1503 | ||
506e931f JA |
1504 | /* |
1505 | * It'd be great if the workqueue API had a way to pass | |
1506 | * in a mask and had some smarts for more clever placement. | |
1507 | * For now we just round-robin here, switching for every | |
1508 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1509 | */ | |
1510 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1511 | { | |
7bed4595 | 1512 | bool tried = false; |
476f8c98 | 1513 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1514 | |
b657d7e6 CH |
1515 | if (hctx->queue->nr_hw_queues == 1) |
1516 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1517 | |
1518 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1519 | select_cpu: |
476f8c98 | 1520 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1521 | cpu_online_mask); |
506e931f | 1522 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1523 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1524 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1525 | } | |
1526 | ||
7bed4595 ML |
1527 | /* |
1528 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1529 | * and it should only happen in the path of handling CPU DEAD. | |
1530 | */ | |
476f8c98 | 1531 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1532 | if (!tried) { |
1533 | tried = true; | |
1534 | goto select_cpu; | |
1535 | } | |
1536 | ||
1537 | /* | |
1538 | * Make sure to re-select CPU next time once after CPUs | |
1539 | * in hctx->cpumask become online again. | |
1540 | */ | |
476f8c98 | 1541 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1542 | hctx->next_cpu_batch = 1; |
1543 | return WORK_CPU_UNBOUND; | |
1544 | } | |
476f8c98 ML |
1545 | |
1546 | hctx->next_cpu = next_cpu; | |
1547 | return next_cpu; | |
506e931f JA |
1548 | } |
1549 | ||
105663f7 AA |
1550 | /** |
1551 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1552 | * @hctx: Pointer to the hardware queue to run. | |
1553 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1554 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1555 | * |
1556 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1557 | * with a delay of @msecs. | |
1558 | */ | |
7587a5ae BVA |
1559 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1560 | unsigned long msecs) | |
320ae51f | 1561 | { |
5435c023 | 1562 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1563 | return; |
1564 | ||
1b792f2f | 1565 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1566 | int cpu = get_cpu(); |
1567 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1568 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1569 | put_cpu(); |
398205b8 PB |
1570 | return; |
1571 | } | |
e4043dcf | 1572 | |
2a90d4aa | 1573 | put_cpu(); |
e4043dcf | 1574 | } |
398205b8 | 1575 | |
ae943d20 BVA |
1576 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1577 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1578 | } |
1579 | ||
105663f7 AA |
1580 | /** |
1581 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1582 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1583 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1584 | * |
1585 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1586 | */ | |
7587a5ae BVA |
1587 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1588 | { | |
1589 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1590 | } | |
1591 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1592 | ||
105663f7 AA |
1593 | /** |
1594 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1595 | * @hctx: Pointer to the hardware queue to run. | |
1596 | * @async: If we want to run the queue asynchronously. | |
1597 | * | |
1598 | * Check if the request queue is not in a quiesced state and if there are | |
1599 | * pending requests to be sent. If this is true, run the queue to send requests | |
1600 | * to hardware. | |
1601 | */ | |
626fb735 | 1602 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1603 | { |
24f5a90f ML |
1604 | int srcu_idx; |
1605 | bool need_run; | |
1606 | ||
1607 | /* | |
1608 | * When queue is quiesced, we may be switching io scheduler, or | |
1609 | * updating nr_hw_queues, or other things, and we can't run queue | |
1610 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1611 | * | |
1612 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1613 | * quiesced. | |
1614 | */ | |
04ced159 JA |
1615 | hctx_lock(hctx, &srcu_idx); |
1616 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1617 | blk_mq_hctx_has_pending(hctx); | |
1618 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1619 | |
626fb735 | 1620 | if (need_run) |
79f720a7 | 1621 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1622 | } |
5b727272 | 1623 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1624 | |
b6e68ee8 JK |
1625 | /* |
1626 | * Is the request queue handled by an IO scheduler that does not respect | |
1627 | * hardware queues when dispatching? | |
1628 | */ | |
1629 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
1630 | { | |
1631 | struct elevator_queue *e = q->elevator; | |
1632 | ||
1633 | if (e && e->type->ops.dispatch_request && | |
1634 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
1635 | return true; | |
1636 | return false; | |
1637 | } | |
1638 | ||
1639 | /* | |
1640 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
1641 | * scheduler. | |
1642 | */ | |
1643 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
1644 | { | |
1645 | struct blk_mq_hw_ctx *hctx; | |
1646 | ||
1647 | /* | |
1648 | * If the IO scheduler does not respect hardware queues when | |
1649 | * dispatching, we just don't bother with multiple HW queues and | |
1650 | * dispatch from hctx for the current CPU since running multiple queues | |
1651 | * just causes lock contention inside the scheduler and pointless cache | |
1652 | * bouncing. | |
1653 | */ | |
1654 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
1655 | raw_smp_processor_id()); | |
1656 | if (!blk_mq_hctx_stopped(hctx)) | |
1657 | return hctx; | |
1658 | return NULL; | |
1659 | } | |
1660 | ||
105663f7 | 1661 | /** |
24f7bb88 | 1662 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1663 | * @q: Pointer to the request queue to run. |
1664 | * @async: If we want to run the queue asynchronously. | |
1665 | */ | |
b94ec296 | 1666 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 1667 | { |
b6e68ee8 | 1668 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
1669 | int i; |
1670 | ||
b6e68ee8 JK |
1671 | sq_hctx = NULL; |
1672 | if (blk_mq_has_sqsched(q)) | |
1673 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 1674 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 1675 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 1676 | continue; |
b6e68ee8 JK |
1677 | /* |
1678 | * Dispatch from this hctx either if there's no hctx preferred | |
1679 | * by IO scheduler or if it has requests that bypass the | |
1680 | * scheduler. | |
1681 | */ | |
1682 | if (!sq_hctx || sq_hctx == hctx || | |
1683 | !list_empty_careful(&hctx->dispatch)) | |
1684 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
1685 | } |
1686 | } | |
b94ec296 | 1687 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1688 | |
b9151e7b DA |
1689 | /** |
1690 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1691 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1692 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1693 | */ |
1694 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1695 | { | |
b6e68ee8 | 1696 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
1697 | int i; |
1698 | ||
b6e68ee8 JK |
1699 | sq_hctx = NULL; |
1700 | if (blk_mq_has_sqsched(q)) | |
1701 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
1702 | queue_for_each_hw_ctx(q, hctx, i) { |
1703 | if (blk_mq_hctx_stopped(hctx)) | |
1704 | continue; | |
b6e68ee8 JK |
1705 | /* |
1706 | * Dispatch from this hctx either if there's no hctx preferred | |
1707 | * by IO scheduler or if it has requests that bypass the | |
1708 | * scheduler. | |
1709 | */ | |
1710 | if (!sq_hctx || sq_hctx == hctx || | |
1711 | !list_empty_careful(&hctx->dispatch)) | |
1712 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
1713 | } |
1714 | } | |
1715 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1716 | ||
fd001443 BVA |
1717 | /** |
1718 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1719 | * @q: request queue. | |
1720 | * | |
1721 | * The caller is responsible for serializing this function against | |
1722 | * blk_mq_{start,stop}_hw_queue(). | |
1723 | */ | |
1724 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1725 | { | |
1726 | struct blk_mq_hw_ctx *hctx; | |
1727 | int i; | |
1728 | ||
1729 | queue_for_each_hw_ctx(q, hctx, i) | |
1730 | if (blk_mq_hctx_stopped(hctx)) | |
1731 | return true; | |
1732 | ||
1733 | return false; | |
1734 | } | |
1735 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1736 | ||
39a70c76 ML |
1737 | /* |
1738 | * This function is often used for pausing .queue_rq() by driver when | |
1739 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1740 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1741 | * |
1742 | * We do not guarantee that dispatch can be drained or blocked | |
1743 | * after blk_mq_stop_hw_queue() returns. Please use | |
1744 | * blk_mq_quiesce_queue() for that requirement. | |
1745 | */ | |
2719aa21 JA |
1746 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1747 | { | |
641a9ed6 | 1748 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1749 | |
641a9ed6 | 1750 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1751 | } |
641a9ed6 | 1752 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1753 | |
39a70c76 ML |
1754 | /* |
1755 | * This function is often used for pausing .queue_rq() by driver when | |
1756 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1757 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1758 | * |
1759 | * We do not guarantee that dispatch can be drained or blocked | |
1760 | * after blk_mq_stop_hw_queues() returns. Please use | |
1761 | * blk_mq_quiesce_queue() for that requirement. | |
1762 | */ | |
2719aa21 JA |
1763 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1764 | { | |
641a9ed6 ML |
1765 | struct blk_mq_hw_ctx *hctx; |
1766 | int i; | |
1767 | ||
1768 | queue_for_each_hw_ctx(q, hctx, i) | |
1769 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1770 | } |
1771 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1772 | ||
320ae51f JA |
1773 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1774 | { | |
1775 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1776 | |
0ffbce80 | 1777 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1778 | } |
1779 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1780 | ||
2f268556 CH |
1781 | void blk_mq_start_hw_queues(struct request_queue *q) |
1782 | { | |
1783 | struct blk_mq_hw_ctx *hctx; | |
1784 | int i; | |
1785 | ||
1786 | queue_for_each_hw_ctx(q, hctx, i) | |
1787 | blk_mq_start_hw_queue(hctx); | |
1788 | } | |
1789 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1790 | ||
ae911c5e JA |
1791 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1792 | { | |
1793 | if (!blk_mq_hctx_stopped(hctx)) | |
1794 | return; | |
1795 | ||
1796 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1797 | blk_mq_run_hw_queue(hctx, async); | |
1798 | } | |
1799 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1800 | ||
1b4a3258 | 1801 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1802 | { |
1803 | struct blk_mq_hw_ctx *hctx; | |
1804 | int i; | |
1805 | ||
ae911c5e JA |
1806 | queue_for_each_hw_ctx(q, hctx, i) |
1807 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1808 | } |
1809 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1810 | ||
70f4db63 | 1811 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1812 | { |
1813 | struct blk_mq_hw_ctx *hctx; | |
1814 | ||
9f993737 | 1815 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1816 | |
21c6e939 | 1817 | /* |
15fe8a90 | 1818 | * If we are stopped, don't run the queue. |
21c6e939 | 1819 | */ |
0841031a | 1820 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 1821 | return; |
7587a5ae BVA |
1822 | |
1823 | __blk_mq_run_hw_queue(hctx); | |
1824 | } | |
1825 | ||
cfd0c552 | 1826 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1827 | struct request *rq, |
1828 | bool at_head) | |
320ae51f | 1829 | { |
e57690fe | 1830 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1831 | enum hctx_type type = hctx->type; |
e57690fe | 1832 | |
7b607814 BVA |
1833 | lockdep_assert_held(&ctx->lock); |
1834 | ||
a54895fa | 1835 | trace_block_rq_insert(rq); |
01b983c9 | 1836 | |
72a0a36e | 1837 | if (at_head) |
c16d6b5a | 1838 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1839 | else |
c16d6b5a | 1840 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1841 | } |
4bb659b1 | 1842 | |
2c3ad667 JA |
1843 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1844 | bool at_head) | |
cfd0c552 ML |
1845 | { |
1846 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1847 | ||
7b607814 BVA |
1848 | lockdep_assert_held(&ctx->lock); |
1849 | ||
e57690fe | 1850 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1851 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1852 | } |
1853 | ||
105663f7 AA |
1854 | /** |
1855 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1856 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 1857 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
1858 | * @run_queue: If we should run the hardware queue after inserting the request. |
1859 | * | |
157f377b JA |
1860 | * Should only be used carefully, when the caller knows we want to |
1861 | * bypass a potential IO scheduler on the target device. | |
1862 | */ | |
01e99aec ML |
1863 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1864 | bool run_queue) | |
157f377b | 1865 | { |
ea4f995e | 1866 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1867 | |
1868 | spin_lock(&hctx->lock); | |
01e99aec ML |
1869 | if (at_head) |
1870 | list_add(&rq->queuelist, &hctx->dispatch); | |
1871 | else | |
1872 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1873 | spin_unlock(&hctx->lock); |
1874 | ||
b0850297 ML |
1875 | if (run_queue) |
1876 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1877 | } |
1878 | ||
bd166ef1 JA |
1879 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1880 | struct list_head *list) | |
320ae51f JA |
1881 | |
1882 | { | |
3f0cedc7 | 1883 | struct request *rq; |
c16d6b5a | 1884 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1885 | |
320ae51f JA |
1886 | /* |
1887 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1888 | * offline now | |
1889 | */ | |
3f0cedc7 | 1890 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1891 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 1892 | trace_block_rq_insert(rq); |
320ae51f | 1893 | } |
3f0cedc7 ML |
1894 | |
1895 | spin_lock(&ctx->lock); | |
c16d6b5a | 1896 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1897 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1898 | spin_unlock(&ctx->lock); |
320ae51f JA |
1899 | } |
1900 | ||
4f0f586b ST |
1901 | static int plug_rq_cmp(void *priv, const struct list_head *a, |
1902 | const struct list_head *b) | |
320ae51f JA |
1903 | { |
1904 | struct request *rqa = container_of(a, struct request, queuelist); | |
1905 | struct request *rqb = container_of(b, struct request, queuelist); | |
1906 | ||
7d30a621 PB |
1907 | if (rqa->mq_ctx != rqb->mq_ctx) |
1908 | return rqa->mq_ctx > rqb->mq_ctx; | |
1909 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1910 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1911 | |
1912 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1913 | } |
1914 | ||
1915 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1916 | { | |
320ae51f | 1917 | LIST_HEAD(list); |
320ae51f | 1918 | |
95ed0c5b PB |
1919 | if (list_empty(&plug->mq_list)) |
1920 | return; | |
320ae51f JA |
1921 | list_splice_init(&plug->mq_list, &list); |
1922 | ||
ce5b009c JA |
1923 | if (plug->rq_count > 2 && plug->multiple_queues) |
1924 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1925 | |
bcc816df DZ |
1926 | plug->rq_count = 0; |
1927 | ||
95ed0c5b PB |
1928 | do { |
1929 | struct list_head rq_list; | |
1930 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1931 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1932 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1933 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1934 | unsigned int depth = 1; | |
1935 | ||
1936 | list_for_each_continue(pos, &list) { | |
1937 | rq = list_entry_rq(pos); | |
1938 | BUG_ON(!rq->q); | |
1939 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1940 | break; | |
1941 | depth++; | |
320ae51f JA |
1942 | } |
1943 | ||
95ed0c5b PB |
1944 | list_cut_before(&rq_list, &list, pos); |
1945 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1946 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1947 | from_schedule); |
95ed0c5b | 1948 | } while(!list_empty(&list)); |
320ae51f JA |
1949 | } |
1950 | ||
14ccb66b CH |
1951 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1952 | unsigned int nr_segs) | |
320ae51f | 1953 | { |
93f221ae EB |
1954 | int err; |
1955 | ||
f924cdde CH |
1956 | if (bio->bi_opf & REQ_RAHEAD) |
1957 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1958 | ||
1959 | rq->__sector = bio->bi_iter.bi_sector; | |
1960 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1961 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1962 | |
1963 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
1964 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1965 | WARN_ON_ONCE(err); | |
4b570521 | 1966 | |
b5af37ab | 1967 | blk_account_io_start(rq); |
320ae51f JA |
1968 | } |
1969 | ||
0f95549c MS |
1970 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1971 | struct request *rq, | |
be94f058 | 1972 | blk_qc_t *cookie, bool last) |
f984df1f | 1973 | { |
f984df1f | 1974 | struct request_queue *q = rq->q; |
f984df1f SL |
1975 | struct blk_mq_queue_data bd = { |
1976 | .rq = rq, | |
be94f058 | 1977 | .last = last, |
f984df1f | 1978 | }; |
bd166ef1 | 1979 | blk_qc_t new_cookie; |
f06345ad | 1980 | blk_status_t ret; |
0f95549c MS |
1981 | |
1982 | new_cookie = request_to_qc_t(hctx, rq); | |
1983 | ||
1984 | /* | |
1985 | * For OK queue, we are done. For error, caller may kill it. | |
1986 | * Any other error (busy), just add it to our list as we | |
1987 | * previously would have done. | |
1988 | */ | |
1989 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1990 | switch (ret) { | |
1991 | case BLK_STS_OK: | |
6ce3dd6e | 1992 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1993 | *cookie = new_cookie; |
1994 | break; | |
1995 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1996 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1997 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1998 | __blk_mq_requeue_request(rq); |
1999 | break; | |
2000 | default: | |
6ce3dd6e | 2001 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2002 | *cookie = BLK_QC_T_NONE; |
2003 | break; | |
2004 | } | |
2005 | ||
2006 | return ret; | |
2007 | } | |
2008 | ||
fd9c40f6 | 2009 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2010 | struct request *rq, |
396eaf21 | 2011 | blk_qc_t *cookie, |
fd9c40f6 | 2012 | bool bypass_insert, bool last) |
0f95549c MS |
2013 | { |
2014 | struct request_queue *q = rq->q; | |
d964f04a | 2015 | bool run_queue = true; |
2a5a24aa | 2016 | int budget_token; |
d964f04a | 2017 | |
23d4ee19 | 2018 | /* |
fd9c40f6 | 2019 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2020 | * |
fd9c40f6 BVA |
2021 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2022 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2023 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2024 | */ |
fd9c40f6 | 2025 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2026 | run_queue = false; |
fd9c40f6 BVA |
2027 | bypass_insert = false; |
2028 | goto insert; | |
d964f04a | 2029 | } |
f984df1f | 2030 | |
fd9c40f6 BVA |
2031 | if (q->elevator && !bypass_insert) |
2032 | goto insert; | |
2253efc8 | 2033 | |
2a5a24aa ML |
2034 | budget_token = blk_mq_get_dispatch_budget(q); |
2035 | if (budget_token < 0) | |
fd9c40f6 | 2036 | goto insert; |
bd166ef1 | 2037 | |
2a5a24aa ML |
2038 | blk_mq_set_rq_budget_token(rq, budget_token); |
2039 | ||
8ab6bb9e | 2040 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2041 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2042 | goto insert; |
88022d72 | 2043 | } |
de148297 | 2044 | |
fd9c40f6 BVA |
2045 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2046 | insert: | |
2047 | if (bypass_insert) | |
2048 | return BLK_STS_RESOURCE; | |
2049 | ||
db03f88f ML |
2050 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2051 | ||
fd9c40f6 BVA |
2052 | return BLK_STS_OK; |
2053 | } | |
2054 | ||
105663f7 AA |
2055 | /** |
2056 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2057 | * @hctx: Pointer of the associated hardware queue. | |
2058 | * @rq: Pointer to request to be sent. | |
2059 | * @cookie: Request queue cookie. | |
2060 | * | |
2061 | * If the device has enough resources to accept a new request now, send the | |
2062 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2063 | * we can try send it another time in the future. Requests inserted at this | |
2064 | * queue have higher priority. | |
2065 | */ | |
fd9c40f6 BVA |
2066 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2067 | struct request *rq, blk_qc_t *cookie) | |
2068 | { | |
2069 | blk_status_t ret; | |
2070 | int srcu_idx; | |
2071 | ||
2072 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2073 | ||
2074 | hctx_lock(hctx, &srcu_idx); | |
2075 | ||
2076 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2077 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2078 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2079 | else if (ret != BLK_STS_OK) |
2080 | blk_mq_end_request(rq, ret); | |
2081 | ||
2082 | hctx_unlock(hctx, srcu_idx); | |
2083 | } | |
2084 | ||
2085 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2086 | { | |
2087 | blk_status_t ret; | |
2088 | int srcu_idx; | |
2089 | blk_qc_t unused_cookie; | |
2090 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2091 | ||
2092 | hctx_lock(hctx, &srcu_idx); | |
2093 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2094 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2095 | |
2096 | return ret; | |
5eb6126e CH |
2097 | } |
2098 | ||
6ce3dd6e ML |
2099 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2100 | struct list_head *list) | |
2101 | { | |
536167d4 | 2102 | int queued = 0; |
632bfb63 | 2103 | int errors = 0; |
536167d4 | 2104 | |
6ce3dd6e | 2105 | while (!list_empty(list)) { |
fd9c40f6 | 2106 | blk_status_t ret; |
6ce3dd6e ML |
2107 | struct request *rq = list_first_entry(list, struct request, |
2108 | queuelist); | |
2109 | ||
2110 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2111 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2112 | if (ret != BLK_STS_OK) { | |
2113 | if (ret == BLK_STS_RESOURCE || | |
2114 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2115 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2116 | list_empty(list)); |
fd9c40f6 BVA |
2117 | break; |
2118 | } | |
2119 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2120 | errors++; |
536167d4 KB |
2121 | } else |
2122 | queued++; | |
6ce3dd6e | 2123 | } |
d666ba98 JA |
2124 | |
2125 | /* | |
2126 | * If we didn't flush the entire list, we could have told | |
2127 | * the driver there was more coming, but that turned out to | |
2128 | * be a lie. | |
2129 | */ | |
632bfb63 | 2130 | if ((!list_empty(list) || errors) && |
2131 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2132 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2133 | } |
2134 | ||
ce5b009c JA |
2135 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2136 | { | |
2137 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2138 | plug->rq_count++; | |
2139 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2140 | struct request *tmp; | |
2141 | ||
2142 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2143 | queuelist); | |
2144 | if (tmp->q != rq->q) | |
2145 | plug->multiple_queues = true; | |
2146 | } | |
2147 | } | |
2148 | ||
7f2a6a69 | 2149 | /* |
316c6c88 | 2150 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
7f2a6a69 SL |
2151 | * queues. This is important for md arrays to benefit from merging |
2152 | * requests. | |
2153 | */ | |
2154 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2155 | { | |
2156 | if (plug->multiple_queues) | |
316c6c88 | 2157 | return BLK_MAX_REQUEST_COUNT * 2; |
7f2a6a69 SL |
2158 | return BLK_MAX_REQUEST_COUNT; |
2159 | } | |
2160 | ||
105663f7 | 2161 | /** |
c62b37d9 | 2162 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2163 | * @bio: Bio pointer. |
2164 | * | |
2165 | * Builds up a request structure from @q and @bio and send to the device. The | |
2166 | * request may not be queued directly to hardware if: | |
2167 | * * This request can be merged with another one | |
2168 | * * We want to place request at plug queue for possible future merging | |
2169 | * * There is an IO scheduler active at this queue | |
2170 | * | |
2171 | * It will not queue the request if there is an error with the bio, or at the | |
2172 | * request creation. | |
2173 | * | |
2174 | * Returns: Request queue cookie. | |
2175 | */ | |
c62b37d9 | 2176 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2177 | { |
309dca30 | 2178 | struct request_queue *q = bio->bi_bdev->bd_disk->queue; |
ef295ecf | 2179 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2180 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6e7abff CH |
2181 | struct blk_mq_alloc_data data = { |
2182 | .q = q, | |
2183 | }; | |
07068d5b | 2184 | struct request *rq; |
f984df1f | 2185 | struct blk_plug *plug; |
5b3f341f | 2186 | struct request *same_queue_rq = NULL; |
14ccb66b | 2187 | unsigned int nr_segs; |
7b371636 | 2188 | blk_qc_t cookie; |
a892c8d5 | 2189 | blk_status_t ret; |
cc29e1bf | 2190 | bool hipri; |
07068d5b JA |
2191 | |
2192 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2193 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2194 | |
e23947bd | 2195 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2196 | goto queue_exit; |
07068d5b | 2197 | |
87c279e6 | 2198 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2199 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2200 | goto queue_exit; |
f984df1f | 2201 | |
14ccb66b | 2202 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2203 | goto queue_exit; |
bd166ef1 | 2204 | |
d5337560 | 2205 | rq_qos_throttle(q, bio); |
87760e5e | 2206 | |
cc29e1bf JX |
2207 | hipri = bio->bi_opf & REQ_HIPRI; |
2208 | ||
7809167d | 2209 | data.cmd_flags = bio->bi_opf; |
e6e7abff | 2210 | rq = __blk_mq_alloc_request(&data); |
87760e5e | 2211 | if (unlikely(!rq)) { |
c1c80384 | 2212 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2213 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2214 | bio_wouldblock_error(bio); |
ac7c5675 | 2215 | goto queue_exit; |
87760e5e JA |
2216 | } |
2217 | ||
e8a676d6 | 2218 | trace_block_getrq(bio); |
d6f1dda2 | 2219 | |
c1c80384 | 2220 | rq_qos_track(q, rq, bio); |
07068d5b | 2221 | |
fd2d3326 | 2222 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2223 | |
970d168d BVA |
2224 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2225 | ||
a892c8d5 ST |
2226 | ret = blk_crypto_init_request(rq); |
2227 | if (ret != BLK_STS_OK) { | |
2228 | bio->bi_status = ret; | |
2229 | bio_endio(bio); | |
2230 | blk_mq_free_request(rq); | |
2231 | return BLK_QC_T_NONE; | |
2232 | } | |
2233 | ||
b49773e7 | 2234 | plug = blk_mq_plug(q, bio); |
07068d5b | 2235 | if (unlikely(is_flush_fua)) { |
105663f7 | 2236 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2237 | blk_insert_flush(rq); |
2238 | blk_mq_run_hw_queue(data.hctx, true); | |
03f26d8f ML |
2239 | } else if (plug && (q->nr_hw_queues == 1 || |
2240 | blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) || | |
2241 | q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2242 | /* |
2243 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2244 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2245 | * |
2246 | * Use normal plugging if this disk is slow HDD, as sequential | |
2247 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2248 | */ |
5f0ed774 | 2249 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2250 | struct request *last = NULL; |
2251 | ||
676d0607 | 2252 | if (!request_count) |
e6c4438b | 2253 | trace_block_plug(q); |
600271d9 SL |
2254 | else |
2255 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2256 | |
7f2a6a69 | 2257 | if (request_count >= blk_plug_max_rq_count(plug) || (last && |
600271d9 | 2258 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
e6c4438b JM |
2259 | blk_flush_plug_list(plug, false); |
2260 | trace_block_plug(q); | |
320ae51f | 2261 | } |
b094f89c | 2262 | |
ce5b009c | 2263 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2264 | } else if (q->elevator) { |
105663f7 | 2265 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2266 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2267 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2268 | /* |
6a83e74d | 2269 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2270 | * Otherwise the existing request in the plug list will be |
2271 | * issued. So the plug list will have one request at most | |
2299722c CH |
2272 | * The plug list might get flushed before this. If that happens, |
2273 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2274 | */ |
2299722c CH |
2275 | if (list_empty(&plug->mq_list)) |
2276 | same_queue_rq = NULL; | |
4711b573 | 2277 | if (same_queue_rq) { |
2299722c | 2278 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2279 | plug->rq_count--; |
2280 | } | |
ce5b009c | 2281 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2282 | trace_block_plug(q); |
2299722c | 2283 | |
dad7a3be | 2284 | if (same_queue_rq) { |
ea4f995e | 2285 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2286 | trace_block_unplug(q, 1, true); |
2299722c | 2287 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2288 | &cookie); |
dad7a3be | 2289 | } |
a12de1d4 ML |
2290 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2291 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2292 | /* |
2293 | * There is no scheduler and we can try to send directly | |
2294 | * to the hardware. | |
2295 | */ | |
fd9c40f6 | 2296 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2297 | } else { |
105663f7 | 2298 | /* Default case. */ |
8fa9f556 | 2299 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2300 | } |
320ae51f | 2301 | |
cc29e1bf JX |
2302 | if (!hipri) |
2303 | return BLK_QC_T_NONE; | |
7b371636 | 2304 | return cookie; |
ac7c5675 CH |
2305 | queue_exit: |
2306 | blk_queue_exit(q); | |
2307 | return BLK_QC_T_NONE; | |
320ae51f JA |
2308 | } |
2309 | ||
bd63141d ML |
2310 | static size_t order_to_size(unsigned int order) |
2311 | { | |
2312 | return (size_t)PAGE_SIZE << order; | |
2313 | } | |
2314 | ||
2315 | /* called before freeing request pool in @tags */ | |
2316 | static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set, | |
2317 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
2318 | { | |
2319 | struct blk_mq_tags *drv_tags = set->tags[hctx_idx]; | |
2320 | struct page *page; | |
2321 | unsigned long flags; | |
2322 | ||
2323 | list_for_each_entry(page, &tags->page_list, lru) { | |
2324 | unsigned long start = (unsigned long)page_address(page); | |
2325 | unsigned long end = start + order_to_size(page->private); | |
2326 | int i; | |
2327 | ||
2328 | for (i = 0; i < set->queue_depth; i++) { | |
2329 | struct request *rq = drv_tags->rqs[i]; | |
2330 | unsigned long rq_addr = (unsigned long)rq; | |
2331 | ||
2332 | if (rq_addr >= start && rq_addr < end) { | |
2333 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
2334 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
2335 | } | |
2336 | } | |
2337 | } | |
2338 | ||
2339 | /* | |
2340 | * Wait until all pending iteration is done. | |
2341 | * | |
2342 | * Request reference is cleared and it is guaranteed to be observed | |
2343 | * after the ->lock is released. | |
2344 | */ | |
2345 | spin_lock_irqsave(&drv_tags->lock, flags); | |
2346 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
2347 | } | |
2348 | ||
cc71a6f4 JA |
2349 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2350 | unsigned int hctx_idx) | |
95363efd | 2351 | { |
e9b267d9 | 2352 | struct page *page; |
320ae51f | 2353 | |
24d2f903 | 2354 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2355 | int i; |
320ae51f | 2356 | |
24d2f903 | 2357 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2358 | struct request *rq = tags->static_rqs[i]; |
2359 | ||
2360 | if (!rq) | |
e9b267d9 | 2361 | continue; |
d6296d39 | 2362 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2363 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2364 | } |
320ae51f | 2365 | } |
320ae51f | 2366 | |
bd63141d ML |
2367 | blk_mq_clear_rq_mapping(set, tags, hctx_idx); |
2368 | ||
24d2f903 CH |
2369 | while (!list_empty(&tags->page_list)) { |
2370 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2371 | list_del_init(&page->lru); |
f75782e4 CM |
2372 | /* |
2373 | * Remove kmemleak object previously allocated in | |
273938bf | 2374 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2375 | */ |
2376 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2377 | __free_pages(page, page->private); |
2378 | } | |
cc71a6f4 | 2379 | } |
320ae51f | 2380 | |
1c0706a7 | 2381 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) |
cc71a6f4 | 2382 | { |
24d2f903 | 2383 | kfree(tags->rqs); |
cc71a6f4 | 2384 | tags->rqs = NULL; |
2af8cbe3 JA |
2385 | kfree(tags->static_rqs); |
2386 | tags->static_rqs = NULL; | |
320ae51f | 2387 | |
1c0706a7 | 2388 | blk_mq_free_tags(tags, flags); |
320ae51f JA |
2389 | } |
2390 | ||
cc71a6f4 JA |
2391 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2392 | unsigned int hctx_idx, | |
2393 | unsigned int nr_tags, | |
1c0706a7 JG |
2394 | unsigned int reserved_tags, |
2395 | unsigned int flags) | |
320ae51f | 2396 | { |
24d2f903 | 2397 | struct blk_mq_tags *tags; |
59f082e4 | 2398 | int node; |
320ae51f | 2399 | |
7d76f856 | 2400 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2401 | if (node == NUMA_NO_NODE) |
2402 | node = set->numa_node; | |
2403 | ||
1c0706a7 | 2404 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); |
24d2f903 CH |
2405 | if (!tags) |
2406 | return NULL; | |
320ae51f | 2407 | |
590b5b7d | 2408 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2409 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2410 | node); |
24d2f903 | 2411 | if (!tags->rqs) { |
1c0706a7 | 2412 | blk_mq_free_tags(tags, flags); |
24d2f903 CH |
2413 | return NULL; |
2414 | } | |
320ae51f | 2415 | |
590b5b7d KC |
2416 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2417 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2418 | node); | |
2af8cbe3 JA |
2419 | if (!tags->static_rqs) { |
2420 | kfree(tags->rqs); | |
1c0706a7 | 2421 | blk_mq_free_tags(tags, flags); |
2af8cbe3 JA |
2422 | return NULL; |
2423 | } | |
2424 | ||
cc71a6f4 JA |
2425 | return tags; |
2426 | } | |
2427 | ||
1d9bd516 TH |
2428 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2429 | unsigned int hctx_idx, int node) | |
2430 | { | |
2431 | int ret; | |
2432 | ||
2433 | if (set->ops->init_request) { | |
2434 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2435 | if (ret) | |
2436 | return ret; | |
2437 | } | |
2438 | ||
12f5b931 | 2439 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2440 | return 0; |
2441 | } | |
2442 | ||
cc71a6f4 JA |
2443 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2444 | unsigned int hctx_idx, unsigned int depth) | |
2445 | { | |
2446 | unsigned int i, j, entries_per_page, max_order = 4; | |
2447 | size_t rq_size, left; | |
59f082e4 SL |
2448 | int node; |
2449 | ||
7d76f856 | 2450 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2451 | if (node == NUMA_NO_NODE) |
2452 | node = set->numa_node; | |
cc71a6f4 JA |
2453 | |
2454 | INIT_LIST_HEAD(&tags->page_list); | |
2455 | ||
320ae51f JA |
2456 | /* |
2457 | * rq_size is the size of the request plus driver payload, rounded | |
2458 | * to the cacheline size | |
2459 | */ | |
24d2f903 | 2460 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2461 | cache_line_size()); |
cc71a6f4 | 2462 | left = rq_size * depth; |
320ae51f | 2463 | |
cc71a6f4 | 2464 | for (i = 0; i < depth; ) { |
320ae51f JA |
2465 | int this_order = max_order; |
2466 | struct page *page; | |
2467 | int to_do; | |
2468 | void *p; | |
2469 | ||
b3a834b1 | 2470 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2471 | this_order--; |
2472 | ||
2473 | do { | |
59f082e4 | 2474 | page = alloc_pages_node(node, |
36e1f3d1 | 2475 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2476 | this_order); |
320ae51f JA |
2477 | if (page) |
2478 | break; | |
2479 | if (!this_order--) | |
2480 | break; | |
2481 | if (order_to_size(this_order) < rq_size) | |
2482 | break; | |
2483 | } while (1); | |
2484 | ||
2485 | if (!page) | |
24d2f903 | 2486 | goto fail; |
320ae51f JA |
2487 | |
2488 | page->private = this_order; | |
24d2f903 | 2489 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2490 | |
2491 | p = page_address(page); | |
f75782e4 CM |
2492 | /* |
2493 | * Allow kmemleak to scan these pages as they contain pointers | |
2494 | * to additional allocations like via ops->init_request(). | |
2495 | */ | |
36e1f3d1 | 2496 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2497 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2498 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2499 | left -= to_do * rq_size; |
2500 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2501 | struct request *rq = p; |
2502 | ||
2503 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2504 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2505 | tags->static_rqs[i] = NULL; | |
2506 | goto fail; | |
e9b267d9 CH |
2507 | } |
2508 | ||
320ae51f JA |
2509 | p += rq_size; |
2510 | i++; | |
2511 | } | |
2512 | } | |
cc71a6f4 | 2513 | return 0; |
320ae51f | 2514 | |
24d2f903 | 2515 | fail: |
cc71a6f4 JA |
2516 | blk_mq_free_rqs(set, tags, hctx_idx); |
2517 | return -ENOMEM; | |
320ae51f JA |
2518 | } |
2519 | ||
bf0beec0 ML |
2520 | struct rq_iter_data { |
2521 | struct blk_mq_hw_ctx *hctx; | |
2522 | bool has_rq; | |
2523 | }; | |
2524 | ||
2525 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2526 | { | |
2527 | struct rq_iter_data *iter_data = data; | |
2528 | ||
2529 | if (rq->mq_hctx != iter_data->hctx) | |
2530 | return true; | |
2531 | iter_data->has_rq = true; | |
2532 | return false; | |
2533 | } | |
2534 | ||
2535 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2536 | { | |
2537 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2538 | hctx->sched_tags : hctx->tags; | |
2539 | struct rq_iter_data data = { | |
2540 | .hctx = hctx, | |
2541 | }; | |
2542 | ||
2543 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2544 | return data.has_rq; | |
2545 | } | |
2546 | ||
2547 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2548 | struct blk_mq_hw_ctx *hctx) | |
2549 | { | |
2550 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2551 | return false; | |
2552 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2553 | return false; | |
2554 | return true; | |
2555 | } | |
2556 | ||
2557 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2558 | { | |
2559 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2560 | struct blk_mq_hw_ctx, cpuhp_online); | |
2561 | ||
2562 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2563 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2564 | return 0; | |
2565 | ||
2566 | /* | |
2567 | * Prevent new request from being allocated on the current hctx. | |
2568 | * | |
2569 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2570 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2571 | * seen once we return from the tag allocator. | |
2572 | */ | |
2573 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2574 | smp_mb__after_atomic(); | |
2575 | ||
2576 | /* | |
2577 | * Try to grab a reference to the queue and wait for any outstanding | |
2578 | * requests. If we could not grab a reference the queue has been | |
2579 | * frozen and there are no requests. | |
2580 | */ | |
2581 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2582 | while (blk_mq_hctx_has_requests(hctx)) | |
2583 | msleep(5); | |
2584 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2585 | } | |
2586 | ||
2587 | return 0; | |
2588 | } | |
2589 | ||
2590 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2591 | { | |
2592 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2593 | struct blk_mq_hw_ctx, cpuhp_online); | |
2594 | ||
2595 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2596 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2597 | return 0; | |
2598 | } | |
2599 | ||
e57690fe JA |
2600 | /* |
2601 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2602 | * software queue to the hw queue dispatch list, and ensure that it | |
2603 | * gets run. | |
2604 | */ | |
9467f859 | 2605 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2606 | { |
9467f859 | 2607 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2608 | struct blk_mq_ctx *ctx; |
2609 | LIST_HEAD(tmp); | |
c16d6b5a | 2610 | enum hctx_type type; |
484b4061 | 2611 | |
9467f859 | 2612 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2613 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2614 | return 0; | |
2615 | ||
e57690fe | 2616 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2617 | type = hctx->type; |
484b4061 JA |
2618 | |
2619 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2620 | if (!list_empty(&ctx->rq_lists[type])) { |
2621 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2622 | blk_mq_hctx_clear_pending(hctx, ctx); |
2623 | } | |
2624 | spin_unlock(&ctx->lock); | |
2625 | ||
2626 | if (list_empty(&tmp)) | |
9467f859 | 2627 | return 0; |
484b4061 | 2628 | |
e57690fe JA |
2629 | spin_lock(&hctx->lock); |
2630 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2631 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2632 | |
2633 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2634 | return 0; |
484b4061 JA |
2635 | } |
2636 | ||
9467f859 | 2637 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2638 | { |
bf0beec0 ML |
2639 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2640 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2641 | &hctx->cpuhp_online); | |
9467f859 TG |
2642 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2643 | &hctx->cpuhp_dead); | |
484b4061 JA |
2644 | } |
2645 | ||
364b6181 ML |
2646 | /* |
2647 | * Before freeing hw queue, clearing the flush request reference in | |
2648 | * tags->rqs[] for avoiding potential UAF. | |
2649 | */ | |
2650 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
2651 | unsigned int queue_depth, struct request *flush_rq) | |
2652 | { | |
2653 | int i; | |
2654 | unsigned long flags; | |
2655 | ||
2656 | /* The hw queue may not be mapped yet */ | |
2657 | if (!tags) | |
2658 | return; | |
2659 | ||
2660 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
2661 | ||
2662 | for (i = 0; i < queue_depth; i++) | |
2663 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
2664 | ||
2665 | /* | |
2666 | * Wait until all pending iteration is done. | |
2667 | * | |
2668 | * Request reference is cleared and it is guaranteed to be observed | |
2669 | * after the ->lock is released. | |
2670 | */ | |
2671 | spin_lock_irqsave(&tags->lock, flags); | |
2672 | spin_unlock_irqrestore(&tags->lock, flags); | |
2673 | } | |
2674 | ||
c3b4afca | 2675 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2676 | static void blk_mq_exit_hctx(struct request_queue *q, |
2677 | struct blk_mq_tag_set *set, | |
2678 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2679 | { | |
364b6181 ML |
2680 | struct request *flush_rq = hctx->fq->flush_rq; |
2681 | ||
8ab0b7dc ML |
2682 | if (blk_mq_hw_queue_mapped(hctx)) |
2683 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2684 | |
364b6181 ML |
2685 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
2686 | set->queue_depth, flush_rq); | |
f70ced09 | 2687 | if (set->ops->exit_request) |
364b6181 | 2688 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 2689 | |
08e98fc6 ML |
2690 | if (set->ops->exit_hctx) |
2691 | set->ops->exit_hctx(hctx, hctx_idx); | |
2692 | ||
9467f859 | 2693 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2694 | |
2695 | spin_lock(&q->unused_hctx_lock); | |
2696 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2697 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2698 | } |
2699 | ||
624dbe47 ML |
2700 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2701 | struct blk_mq_tag_set *set, int nr_queue) | |
2702 | { | |
2703 | struct blk_mq_hw_ctx *hctx; | |
2704 | unsigned int i; | |
2705 | ||
2706 | queue_for_each_hw_ctx(q, hctx, i) { | |
2707 | if (i == nr_queue) | |
2708 | break; | |
477e19de | 2709 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2710 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2711 | } |
624dbe47 ML |
2712 | } |
2713 | ||
7c6c5b7c ML |
2714 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2715 | { | |
2716 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2717 | ||
2718 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2719 | __alignof__(struct blk_mq_hw_ctx)) != | |
2720 | sizeof(struct blk_mq_hw_ctx)); | |
2721 | ||
2722 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2723 | hw_ctx_size += sizeof(struct srcu_struct); | |
2724 | ||
2725 | return hw_ctx_size; | |
2726 | } | |
2727 | ||
08e98fc6 ML |
2728 | static int blk_mq_init_hctx(struct request_queue *q, |
2729 | struct blk_mq_tag_set *set, | |
2730 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2731 | { |
7c6c5b7c ML |
2732 | hctx->queue_num = hctx_idx; |
2733 | ||
bf0beec0 ML |
2734 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2735 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2736 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2737 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2738 | ||
2739 | hctx->tags = set->tags[hctx_idx]; | |
2740 | ||
2741 | if (set->ops->init_hctx && | |
2742 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2743 | goto unregister_cpu_notifier; | |
08e98fc6 | 2744 | |
7c6c5b7c ML |
2745 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2746 | hctx->numa_node)) | |
2747 | goto exit_hctx; | |
2748 | return 0; | |
2749 | ||
2750 | exit_hctx: | |
2751 | if (set->ops->exit_hctx) | |
2752 | set->ops->exit_hctx(hctx, hctx_idx); | |
2753 | unregister_cpu_notifier: | |
2754 | blk_mq_remove_cpuhp(hctx); | |
2755 | return -1; | |
2756 | } | |
2757 | ||
2758 | static struct blk_mq_hw_ctx * | |
2759 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2760 | int node) | |
2761 | { | |
2762 | struct blk_mq_hw_ctx *hctx; | |
2763 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2764 | ||
2765 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2766 | if (!hctx) | |
2767 | goto fail_alloc_hctx; | |
2768 | ||
2769 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2770 | goto free_hctx; | |
2771 | ||
2772 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2773 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2774 | node = set->numa_node; |
2775 | hctx->numa_node = node; | |
08e98fc6 | 2776 | |
9f993737 | 2777 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2778 | spin_lock_init(&hctx->lock); |
2779 | INIT_LIST_HEAD(&hctx->dispatch); | |
2780 | hctx->queue = q; | |
51db1c37 | 2781 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2782 | |
2f8f1336 ML |
2783 | INIT_LIST_HEAD(&hctx->hctx_list); |
2784 | ||
320ae51f | 2785 | /* |
08e98fc6 ML |
2786 | * Allocate space for all possible cpus to avoid allocation at |
2787 | * runtime | |
320ae51f | 2788 | */ |
d904bfa7 | 2789 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2790 | gfp, node); |
08e98fc6 | 2791 | if (!hctx->ctxs) |
7c6c5b7c | 2792 | goto free_cpumask; |
320ae51f | 2793 | |
5b202853 | 2794 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 2795 | gfp, node, false, false)) |
08e98fc6 | 2796 | goto free_ctxs; |
08e98fc6 | 2797 | hctx->nr_ctx = 0; |
320ae51f | 2798 | |
5815839b | 2799 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2800 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2801 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2802 | ||
754a1572 | 2803 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2804 | if (!hctx->fq) |
7c6c5b7c | 2805 | goto free_bitmap; |
320ae51f | 2806 | |
6a83e74d | 2807 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2808 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2809 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2810 | |
7c6c5b7c | 2811 | return hctx; |
320ae51f | 2812 | |
08e98fc6 | 2813 | free_bitmap: |
88459642 | 2814 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2815 | free_ctxs: |
2816 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2817 | free_cpumask: |
2818 | free_cpumask_var(hctx->cpumask); | |
2819 | free_hctx: | |
2820 | kfree(hctx); | |
2821 | fail_alloc_hctx: | |
2822 | return NULL; | |
08e98fc6 | 2823 | } |
320ae51f | 2824 | |
320ae51f JA |
2825 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2826 | unsigned int nr_hw_queues) | |
2827 | { | |
b3c661b1 JA |
2828 | struct blk_mq_tag_set *set = q->tag_set; |
2829 | unsigned int i, j; | |
320ae51f JA |
2830 | |
2831 | for_each_possible_cpu(i) { | |
2832 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2833 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2834 | int k; |
320ae51f | 2835 | |
320ae51f JA |
2836 | __ctx->cpu = i; |
2837 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2838 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2839 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2840 | ||
320ae51f JA |
2841 | __ctx->queue = q; |
2842 | ||
320ae51f JA |
2843 | /* |
2844 | * Set local node, IFF we have more than one hw queue. If | |
2845 | * not, we remain on the home node of the device | |
2846 | */ | |
b3c661b1 JA |
2847 | for (j = 0; j < set->nr_maps; j++) { |
2848 | hctx = blk_mq_map_queue_type(q, j, i); | |
2849 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2850 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2851 | } |
320ae51f JA |
2852 | } |
2853 | } | |
2854 | ||
03b63b02 WZ |
2855 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, |
2856 | int hctx_idx) | |
cc71a6f4 | 2857 | { |
1c0706a7 | 2858 | unsigned int flags = set->flags; |
cc71a6f4 JA |
2859 | int ret = 0; |
2860 | ||
2861 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1c0706a7 | 2862 | set->queue_depth, set->reserved_tags, flags); |
cc71a6f4 JA |
2863 | if (!set->tags[hctx_idx]) |
2864 | return false; | |
2865 | ||
2866 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2867 | set->queue_depth); | |
2868 | if (!ret) | |
2869 | return true; | |
2870 | ||
1c0706a7 | 2871 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
cc71a6f4 JA |
2872 | set->tags[hctx_idx] = NULL; |
2873 | return false; | |
2874 | } | |
2875 | ||
2876 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2877 | unsigned int hctx_idx) | |
2878 | { | |
1c0706a7 JG |
2879 | unsigned int flags = set->flags; |
2880 | ||
4e6db0f2 | 2881 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 | 2882 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
1c0706a7 | 2883 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
bd166ef1 JA |
2884 | set->tags[hctx_idx] = NULL; |
2885 | } | |
cc71a6f4 JA |
2886 | } |
2887 | ||
4b855ad3 | 2888 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2889 | { |
b3c661b1 | 2890 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2891 | struct blk_mq_hw_ctx *hctx; |
2892 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2893 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2894 | |
2895 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2896 | cpumask_clear(hctx->cpumask); |
320ae51f | 2897 | hctx->nr_ctx = 0; |
d416c92c | 2898 | hctx->dispatch_from = NULL; |
320ae51f JA |
2899 | } |
2900 | ||
2901 | /* | |
4b855ad3 | 2902 | * Map software to hardware queues. |
4412efec ML |
2903 | * |
2904 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2905 | */ |
20e4d813 | 2906 | for_each_possible_cpu(i) { |
4412efec | 2907 | |
897bb0c7 | 2908 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2909 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2910 | if (!set->map[j].nr_queues) { |
2911 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2912 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2913 | continue; |
bb94aea1 | 2914 | } |
fd689871 ML |
2915 | hctx_idx = set->map[j].mq_map[i]; |
2916 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2917 | if (!set->tags[hctx_idx] && | |
03b63b02 | 2918 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { |
fd689871 ML |
2919 | /* |
2920 | * If tags initialization fail for some hctx, | |
2921 | * that hctx won't be brought online. In this | |
2922 | * case, remap the current ctx to hctx[0] which | |
2923 | * is guaranteed to always have tags allocated | |
2924 | */ | |
2925 | set->map[j].mq_map[i] = 0; | |
2926 | } | |
e5edd5f2 | 2927 | |
b3c661b1 | 2928 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2929 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2930 | /* |
2931 | * If the CPU is already set in the mask, then we've | |
2932 | * mapped this one already. This can happen if | |
2933 | * devices share queues across queue maps. | |
2934 | */ | |
2935 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2936 | continue; | |
2937 | ||
2938 | cpumask_set_cpu(i, hctx->cpumask); | |
2939 | hctx->type = j; | |
2940 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2941 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2942 | ||
2943 | /* | |
2944 | * If the nr_ctx type overflows, we have exceeded the | |
2945 | * amount of sw queues we can support. | |
2946 | */ | |
2947 | BUG_ON(!hctx->nr_ctx); | |
2948 | } | |
bb94aea1 JW |
2949 | |
2950 | for (; j < HCTX_MAX_TYPES; j++) | |
2951 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2952 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2953 | } |
506e931f JA |
2954 | |
2955 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2956 | /* |
2957 | * If no software queues are mapped to this hardware queue, | |
2958 | * disable it and free the request entries. | |
2959 | */ | |
2960 | if (!hctx->nr_ctx) { | |
2961 | /* Never unmap queue 0. We need it as a | |
2962 | * fallback in case of a new remap fails | |
2963 | * allocation | |
2964 | */ | |
2965 | if (i && set->tags[i]) | |
2966 | blk_mq_free_map_and_requests(set, i); | |
2967 | ||
2968 | hctx->tags = NULL; | |
2969 | continue; | |
2970 | } | |
484b4061 | 2971 | |
2a34c087 ML |
2972 | hctx->tags = set->tags[i]; |
2973 | WARN_ON(!hctx->tags); | |
2974 | ||
889fa31f CY |
2975 | /* |
2976 | * Set the map size to the number of mapped software queues. | |
2977 | * This is more accurate and more efficient than looping | |
2978 | * over all possibly mapped software queues. | |
2979 | */ | |
88459642 | 2980 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2981 | |
484b4061 JA |
2982 | /* |
2983 | * Initialize batch roundrobin counts | |
2984 | */ | |
f82ddf19 | 2985 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2986 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2987 | } | |
320ae51f JA |
2988 | } |
2989 | ||
8e8320c9 JA |
2990 | /* |
2991 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2992 | * the queue isn't live yet. | |
2993 | */ | |
2404e607 | 2994 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2995 | { |
2996 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2997 | int i; |
2998 | ||
2404e607 | 2999 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3000 | if (shared) { |
51db1c37 | 3001 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3002 | } else { |
3003 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3004 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3005 | } |
2404e607 JM |
3006 | } |
3007 | } | |
3008 | ||
655ac300 HR |
3009 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3010 | bool shared) | |
2404e607 JM |
3011 | { |
3012 | struct request_queue *q; | |
0d2602ca | 3013 | |
705cda97 BVA |
3014 | lockdep_assert_held(&set->tag_list_lock); |
3015 | ||
0d2602ca JA |
3016 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3017 | blk_mq_freeze_queue(q); | |
2404e607 | 3018 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3019 | blk_mq_unfreeze_queue(q); |
3020 | } | |
3021 | } | |
3022 | ||
3023 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3024 | { | |
3025 | struct blk_mq_tag_set *set = q->tag_set; | |
3026 | ||
0d2602ca | 3027 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3028 | list_del(&q->tag_set_list); |
2404e607 JM |
3029 | if (list_is_singular(&set->tag_list)) { |
3030 | /* just transitioned to unshared */ | |
51db1c37 | 3031 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3032 | /* update existing queue */ |
655ac300 | 3033 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3034 | } |
0d2602ca | 3035 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3036 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3037 | } |
3038 | ||
3039 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3040 | struct request_queue *q) | |
3041 | { | |
0d2602ca | 3042 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3043 | |
ff821d27 JA |
3044 | /* |
3045 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3046 | */ | |
3047 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3048 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3049 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3050 | /* update existing queue */ |
655ac300 | 3051 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3052 | } |
51db1c37 | 3053 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3054 | queue_set_hctx_shared(q, true); |
08c875cb | 3055 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3056 | |
0d2602ca JA |
3057 | mutex_unlock(&set->tag_list_lock); |
3058 | } | |
3059 | ||
1db4909e ML |
3060 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3061 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3062 | { | |
3063 | struct blk_mq_ctxs *ctxs; | |
3064 | int cpu; | |
3065 | ||
3066 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3067 | if (!ctxs) | |
3068 | return -ENOMEM; | |
3069 | ||
3070 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3071 | if (!ctxs->queue_ctx) | |
3072 | goto fail; | |
3073 | ||
3074 | for_each_possible_cpu(cpu) { | |
3075 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3076 | ctx->ctxs = ctxs; | |
3077 | } | |
3078 | ||
3079 | q->mq_kobj = &ctxs->kobj; | |
3080 | q->queue_ctx = ctxs->queue_ctx; | |
3081 | ||
3082 | return 0; | |
3083 | fail: | |
3084 | kfree(ctxs); | |
3085 | return -ENOMEM; | |
3086 | } | |
3087 | ||
e09aae7e ML |
3088 | /* |
3089 | * It is the actual release handler for mq, but we do it from | |
3090 | * request queue's release handler for avoiding use-after-free | |
3091 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3092 | * but we can't group ctx/kctx kobj without it. | |
3093 | */ | |
3094 | void blk_mq_release(struct request_queue *q) | |
3095 | { | |
2f8f1336 ML |
3096 | struct blk_mq_hw_ctx *hctx, *next; |
3097 | int i; | |
e09aae7e | 3098 | |
2f8f1336 ML |
3099 | queue_for_each_hw_ctx(q, hctx, i) |
3100 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3101 | ||
3102 | /* all hctx are in .unused_hctx_list now */ | |
3103 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3104 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3105 | kobject_put(&hctx->kobj); |
c3b4afca | 3106 | } |
e09aae7e ML |
3107 | |
3108 | kfree(q->queue_hw_ctx); | |
3109 | ||
7ea5fe31 ML |
3110 | /* |
3111 | * release .mq_kobj and sw queue's kobject now because | |
3112 | * both share lifetime with request queue. | |
3113 | */ | |
3114 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3115 | } |
3116 | ||
5ec780a6 | 3117 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3118 | void *queuedata) |
b62c21b7 | 3119 | { |
26a9750a CH |
3120 | struct request_queue *q; |
3121 | int ret; | |
b62c21b7 | 3122 | |
26a9750a CH |
3123 | q = blk_alloc_queue(set->numa_node); |
3124 | if (!q) | |
b62c21b7 | 3125 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3126 | q->queuedata = queuedata; |
3127 | ret = blk_mq_init_allocated_queue(set, q); | |
3128 | if (ret) { | |
3129 | blk_cleanup_queue(q); | |
3130 | return ERR_PTR(ret); | |
3131 | } | |
b62c21b7 MS |
3132 | return q; |
3133 | } | |
2f227bb9 CH |
3134 | |
3135 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3136 | { | |
3137 | return blk_mq_init_queue_data(set, NULL); | |
3138 | } | |
b62c21b7 MS |
3139 | EXPORT_SYMBOL(blk_mq_init_queue); |
3140 | ||
4dcc4874 CH |
3141 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3142 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3143 | { |
3144 | struct request_queue *q; | |
b461dfc4 | 3145 | struct gendisk *disk; |
9316a9ed | 3146 | |
b461dfc4 CH |
3147 | q = blk_mq_init_queue_data(set, queuedata); |
3148 | if (IS_ERR(q)) | |
3149 | return ERR_CAST(q); | |
9316a9ed | 3150 | |
4a1fa41d | 3151 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3152 | if (!disk) { |
3153 | blk_cleanup_queue(q); | |
3154 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3155 | } |
b461dfc4 | 3156 | return disk; |
9316a9ed | 3157 | } |
b461dfc4 | 3158 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3159 | |
34d11ffa JW |
3160 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3161 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3162 | int hctx_idx, int node) | |
3163 | { | |
2f8f1336 | 3164 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3165 | |
2f8f1336 ML |
3166 | /* reuse dead hctx first */ |
3167 | spin_lock(&q->unused_hctx_lock); | |
3168 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3169 | if (tmp->numa_node == node) { | |
3170 | hctx = tmp; | |
3171 | break; | |
3172 | } | |
3173 | } | |
3174 | if (hctx) | |
3175 | list_del_init(&hctx->hctx_list); | |
3176 | spin_unlock(&q->unused_hctx_lock); | |
3177 | ||
3178 | if (!hctx) | |
3179 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3180 | if (!hctx) |
7c6c5b7c | 3181 | goto fail; |
34d11ffa | 3182 | |
7c6c5b7c ML |
3183 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3184 | goto free_hctx; | |
34d11ffa JW |
3185 | |
3186 | return hctx; | |
7c6c5b7c ML |
3187 | |
3188 | free_hctx: | |
3189 | kobject_put(&hctx->kobj); | |
3190 | fail: | |
3191 | return NULL; | |
34d11ffa JW |
3192 | } |
3193 | ||
868f2f0b KB |
3194 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3195 | struct request_queue *q) | |
320ae51f | 3196 | { |
e01ad46d | 3197 | int i, j, end; |
868f2f0b | 3198 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3199 | |
ac0d6b92 BVA |
3200 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3201 | struct blk_mq_hw_ctx **new_hctxs; | |
3202 | ||
3203 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3204 | sizeof(*new_hctxs), GFP_KERNEL, | |
3205 | set->numa_node); | |
3206 | if (!new_hctxs) | |
3207 | return; | |
3208 | if (hctxs) | |
3209 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3210 | sizeof(*hctxs)); | |
3211 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3212 | kfree(hctxs); |
3213 | hctxs = new_hctxs; | |
3214 | } | |
3215 | ||
fb350e0a ML |
3216 | /* protect against switching io scheduler */ |
3217 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3218 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3219 | int node; |
34d11ffa | 3220 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3221 | |
7d76f856 | 3222 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3223 | /* |
3224 | * If the hw queue has been mapped to another numa node, | |
3225 | * we need to realloc the hctx. If allocation fails, fallback | |
3226 | * to use the previous one. | |
3227 | */ | |
3228 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3229 | continue; | |
868f2f0b | 3230 | |
34d11ffa JW |
3231 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3232 | if (hctx) { | |
2f8f1336 | 3233 | if (hctxs[i]) |
34d11ffa | 3234 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3235 | hctxs[i] = hctx; |
3236 | } else { | |
3237 | if (hctxs[i]) | |
3238 | pr_warn("Allocate new hctx on node %d fails,\ | |
3239 | fallback to previous one on node %d\n", | |
3240 | node, hctxs[i]->numa_node); | |
3241 | else | |
3242 | break; | |
868f2f0b | 3243 | } |
320ae51f | 3244 | } |
e01ad46d JW |
3245 | /* |
3246 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3247 | * hctxs and keep the previous q->nr_hw_queues. | |
3248 | */ | |
3249 | if (i != set->nr_hw_queues) { | |
3250 | j = q->nr_hw_queues; | |
3251 | end = i; | |
3252 | } else { | |
3253 | j = i; | |
3254 | end = q->nr_hw_queues; | |
3255 | q->nr_hw_queues = set->nr_hw_queues; | |
3256 | } | |
34d11ffa | 3257 | |
e01ad46d | 3258 | for (; j < end; j++) { |
868f2f0b KB |
3259 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3260 | ||
3261 | if (hctx) { | |
cc71a6f4 JA |
3262 | if (hctx->tags) |
3263 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 3264 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3265 | hctxs[j] = NULL; |
868f2f0b KB |
3266 | } |
3267 | } | |
fb350e0a | 3268 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3269 | } |
3270 | ||
26a9750a CH |
3271 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
3272 | struct request_queue *q) | |
868f2f0b | 3273 | { |
66841672 ML |
3274 | /* mark the queue as mq asap */ |
3275 | q->mq_ops = set->ops; | |
3276 | ||
34dbad5d | 3277 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3278 | blk_mq_poll_stats_bkt, |
3279 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3280 | if (!q->poll_cb) |
3281 | goto err_exit; | |
3282 | ||
1db4909e | 3283 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3284 | goto err_poll; |
868f2f0b | 3285 | |
737f98cf ML |
3286 | /* init q->mq_kobj and sw queues' kobjects */ |
3287 | blk_mq_sysfs_init(q); | |
3288 | ||
2f8f1336 ML |
3289 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3290 | spin_lock_init(&q->unused_hctx_lock); | |
3291 | ||
868f2f0b KB |
3292 | blk_mq_realloc_hw_ctxs(set, q); |
3293 | if (!q->nr_hw_queues) | |
3294 | goto err_hctxs; | |
320ae51f | 3295 | |
287922eb | 3296 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3297 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3298 | |
a8908939 | 3299 | q->tag_set = set; |
320ae51f | 3300 | |
94eddfbe | 3301 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3302 | if (set->nr_maps > HCTX_TYPE_POLL && |
3303 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3304 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3305 | |
2849450a | 3306 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3307 | INIT_LIST_HEAD(&q->requeue_list); |
3308 | spin_lock_init(&q->requeue_lock); | |
3309 | ||
eba71768 JA |
3310 | q->nr_requests = set->queue_depth; |
3311 | ||
64f1c21e JA |
3312 | /* |
3313 | * Default to classic polling | |
3314 | */ | |
29ece8b4 | 3315 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3316 | |
24d2f903 | 3317 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3318 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3319 | blk_mq_map_swqueue(q); |
26a9750a | 3320 | return 0; |
18741986 | 3321 | |
320ae51f | 3322 | err_hctxs: |
868f2f0b | 3323 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3324 | q->nr_hw_queues = 0; |
1db4909e | 3325 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3326 | err_poll: |
3327 | blk_stat_free_callback(q->poll_cb); | |
3328 | q->poll_cb = NULL; | |
c7de5726 ML |
3329 | err_exit: |
3330 | q->mq_ops = NULL; | |
26a9750a | 3331 | return -ENOMEM; |
320ae51f | 3332 | } |
b62c21b7 | 3333 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3334 | |
c7e2d94b ML |
3335 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3336 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3337 | { |
630ef623 | 3338 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3339 | |
630ef623 | 3340 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 3341 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
3342 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
3343 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 3344 | } |
320ae51f | 3345 | |
a5164405 JA |
3346 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3347 | { | |
3348 | int i; | |
3349 | ||
8229cca8 | 3350 | for (i = 0; i < set->nr_hw_queues; i++) { |
03b63b02 | 3351 | if (!__blk_mq_alloc_map_and_request(set, i)) |
a5164405 | 3352 | goto out_unwind; |
8229cca8 XT |
3353 | cond_resched(); |
3354 | } | |
a5164405 JA |
3355 | |
3356 | return 0; | |
3357 | ||
3358 | out_unwind: | |
3359 | while (--i >= 0) | |
2e194422 | 3360 | blk_mq_free_map_and_requests(set, i); |
a5164405 | 3361 | |
a5164405 JA |
3362 | return -ENOMEM; |
3363 | } | |
3364 | ||
3365 | /* | |
3366 | * Allocate the request maps associated with this tag_set. Note that this | |
3367 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3368 | * will be updated to reflect the allocated depth. | |
3369 | */ | |
79fab528 | 3370 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) |
a5164405 JA |
3371 | { |
3372 | unsigned int depth; | |
3373 | int err; | |
3374 | ||
3375 | depth = set->queue_depth; | |
3376 | do { | |
3377 | err = __blk_mq_alloc_rq_maps(set); | |
3378 | if (!err) | |
3379 | break; | |
3380 | ||
3381 | set->queue_depth >>= 1; | |
3382 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3383 | err = -ENOMEM; | |
3384 | break; | |
3385 | } | |
3386 | } while (set->queue_depth); | |
3387 | ||
3388 | if (!set->queue_depth || err) { | |
3389 | pr_err("blk-mq: failed to allocate request map\n"); | |
3390 | return -ENOMEM; | |
3391 | } | |
3392 | ||
3393 | if (depth != set->queue_depth) | |
3394 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3395 | depth, set->queue_depth); | |
3396 | ||
3397 | return 0; | |
3398 | } | |
3399 | ||
ebe8bddb OS |
3400 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3401 | { | |
6e66b493 BVA |
3402 | /* |
3403 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3404 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3405 | * number of hardware queues. | |
3406 | */ | |
3407 | if (set->nr_maps == 1) | |
3408 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3409 | ||
59388702 | 3410 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3411 | int i; |
3412 | ||
7d4901a9 ML |
3413 | /* |
3414 | * transport .map_queues is usually done in the following | |
3415 | * way: | |
3416 | * | |
3417 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3418 | * mask = get_cpu_mask(queue) | |
3419 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3420 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3421 | * } |
3422 | * | |
3423 | * When we need to remap, the table has to be cleared for | |
3424 | * killing stale mapping since one CPU may not be mapped | |
3425 | * to any hw queue. | |
3426 | */ | |
b3c661b1 JA |
3427 | for (i = 0; i < set->nr_maps; i++) |
3428 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3429 | |
ebe8bddb | 3430 | return set->ops->map_queues(set); |
b3c661b1 JA |
3431 | } else { |
3432 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3433 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3434 | } |
ebe8bddb OS |
3435 | } |
3436 | ||
f7e76dbc BVA |
3437 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3438 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3439 | { | |
3440 | struct blk_mq_tags **new_tags; | |
3441 | ||
3442 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3443 | return 0; | |
3444 | ||
3445 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3446 | GFP_KERNEL, set->numa_node); | |
3447 | if (!new_tags) | |
3448 | return -ENOMEM; | |
3449 | ||
3450 | if (set->tags) | |
3451 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3452 | sizeof(*set->tags)); | |
3453 | kfree(set->tags); | |
3454 | set->tags = new_tags; | |
3455 | set->nr_hw_queues = new_nr_hw_queues; | |
3456 | ||
3457 | return 0; | |
3458 | } | |
3459 | ||
91cdf265 MI |
3460 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3461 | int new_nr_hw_queues) | |
3462 | { | |
3463 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3464 | } | |
3465 | ||
a4391c64 JA |
3466 | /* |
3467 | * Alloc a tag set to be associated with one or more request queues. | |
3468 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3469 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3470 | * value will be stored in set->queue_depth. |
3471 | */ | |
24d2f903 CH |
3472 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3473 | { | |
b3c661b1 | 3474 | int i, ret; |
da695ba2 | 3475 | |
205fb5f5 BVA |
3476 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3477 | ||
24d2f903 CH |
3478 | if (!set->nr_hw_queues) |
3479 | return -EINVAL; | |
a4391c64 | 3480 | if (!set->queue_depth) |
24d2f903 CH |
3481 | return -EINVAL; |
3482 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3483 | return -EINVAL; | |
3484 | ||
7d7e0f90 | 3485 | if (!set->ops->queue_rq) |
24d2f903 CH |
3486 | return -EINVAL; |
3487 | ||
de148297 ML |
3488 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3489 | return -EINVAL; | |
3490 | ||
a4391c64 JA |
3491 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3492 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3493 | BLK_MQ_MAX_DEPTH); | |
3494 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3495 | } | |
24d2f903 | 3496 | |
b3c661b1 JA |
3497 | if (!set->nr_maps) |
3498 | set->nr_maps = 1; | |
3499 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3500 | return -EINVAL; | |
3501 | ||
6637fadf SL |
3502 | /* |
3503 | * If a crashdump is active, then we are potentially in a very | |
3504 | * memory constrained environment. Limit us to 1 queue and | |
3505 | * 64 tags to prevent using too much memory. | |
3506 | */ | |
3507 | if (is_kdump_kernel()) { | |
3508 | set->nr_hw_queues = 1; | |
59388702 | 3509 | set->nr_maps = 1; |
6637fadf SL |
3510 | set->queue_depth = min(64U, set->queue_depth); |
3511 | } | |
868f2f0b | 3512 | /* |
392546ae JA |
3513 | * There is no use for more h/w queues than cpus if we just have |
3514 | * a single map | |
868f2f0b | 3515 | */ |
392546ae | 3516 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3517 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3518 | |
91cdf265 | 3519 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3520 | return -ENOMEM; |
24d2f903 | 3521 | |
da695ba2 | 3522 | ret = -ENOMEM; |
b3c661b1 JA |
3523 | for (i = 0; i < set->nr_maps; i++) { |
3524 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3525 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3526 | GFP_KERNEL, set->numa_node); |
3527 | if (!set->map[i].mq_map) | |
3528 | goto out_free_mq_map; | |
59388702 | 3529 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3530 | } |
bdd17e75 | 3531 | |
ebe8bddb | 3532 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3533 | if (ret) |
3534 | goto out_free_mq_map; | |
3535 | ||
79fab528 | 3536 | ret = blk_mq_alloc_map_and_requests(set); |
da695ba2 | 3537 | if (ret) |
bdd17e75 | 3538 | goto out_free_mq_map; |
24d2f903 | 3539 | |
32bc15af | 3540 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
f1b49fdc JG |
3541 | atomic_set(&set->active_queues_shared_sbitmap, 0); |
3542 | ||
56b68085 | 3543 | if (blk_mq_init_shared_sbitmap(set)) { |
32bc15af JG |
3544 | ret = -ENOMEM; |
3545 | goto out_free_mq_rq_maps; | |
3546 | } | |
3547 | } | |
3548 | ||
0d2602ca JA |
3549 | mutex_init(&set->tag_list_lock); |
3550 | INIT_LIST_HEAD(&set->tag_list); | |
3551 | ||
24d2f903 | 3552 | return 0; |
bdd17e75 | 3553 | |
32bc15af JG |
3554 | out_free_mq_rq_maps: |
3555 | for (i = 0; i < set->nr_hw_queues; i++) | |
3556 | blk_mq_free_map_and_requests(set, i); | |
bdd17e75 | 3557 | out_free_mq_map: |
b3c661b1 JA |
3558 | for (i = 0; i < set->nr_maps; i++) { |
3559 | kfree(set->map[i].mq_map); | |
3560 | set->map[i].mq_map = NULL; | |
3561 | } | |
5676e7b6 RE |
3562 | kfree(set->tags); |
3563 | set->tags = NULL; | |
da695ba2 | 3564 | return ret; |
24d2f903 CH |
3565 | } |
3566 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3567 | ||
cdb14e0f CH |
3568 | /* allocate and initialize a tagset for a simple single-queue device */ |
3569 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
3570 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
3571 | unsigned int set_flags) | |
3572 | { | |
3573 | memset(set, 0, sizeof(*set)); | |
3574 | set->ops = ops; | |
3575 | set->nr_hw_queues = 1; | |
3576 | set->nr_maps = 1; | |
3577 | set->queue_depth = queue_depth; | |
3578 | set->numa_node = NUMA_NO_NODE; | |
3579 | set->flags = set_flags; | |
3580 | return blk_mq_alloc_tag_set(set); | |
3581 | } | |
3582 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
3583 | ||
24d2f903 CH |
3584 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
3585 | { | |
b3c661b1 | 3586 | int i, j; |
24d2f903 | 3587 | |
f7e76dbc | 3588 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3589 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3590 | |
32bc15af JG |
3591 | if (blk_mq_is_sbitmap_shared(set->flags)) |
3592 | blk_mq_exit_shared_sbitmap(set); | |
3593 | ||
b3c661b1 JA |
3594 | for (j = 0; j < set->nr_maps; j++) { |
3595 | kfree(set->map[j].mq_map); | |
3596 | set->map[j].mq_map = NULL; | |
3597 | } | |
bdd17e75 | 3598 | |
981bd189 | 3599 | kfree(set->tags); |
5676e7b6 | 3600 | set->tags = NULL; |
24d2f903 CH |
3601 | } |
3602 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3603 | ||
e3a2b3f9 JA |
3604 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3605 | { | |
3606 | struct blk_mq_tag_set *set = q->tag_set; | |
3607 | struct blk_mq_hw_ctx *hctx; | |
3608 | int i, ret; | |
3609 | ||
bd166ef1 | 3610 | if (!set) |
e3a2b3f9 JA |
3611 | return -EINVAL; |
3612 | ||
e5fa8140 AZ |
3613 | if (q->nr_requests == nr) |
3614 | return 0; | |
3615 | ||
70f36b60 | 3616 | blk_mq_freeze_queue(q); |
24f5a90f | 3617 | blk_mq_quiesce_queue(q); |
70f36b60 | 3618 | |
e3a2b3f9 JA |
3619 | ret = 0; |
3620 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3621 | if (!hctx->tags) |
3622 | continue; | |
bd166ef1 JA |
3623 | /* |
3624 | * If we're using an MQ scheduler, just update the scheduler | |
3625 | * queue depth. This is similar to what the old code would do. | |
3626 | */ | |
70f36b60 | 3627 | if (!hctx->sched_tags) { |
c2e82a23 | 3628 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 | 3629 | false); |
32bc15af JG |
3630 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) |
3631 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
70f36b60 JA |
3632 | } else { |
3633 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3634 | nr, true); | |
d97e594c JG |
3635 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
3636 | hctx->sched_tags->bitmap_tags = | |
3637 | &q->sched_bitmap_tags; | |
3638 | hctx->sched_tags->breserved_tags = | |
3639 | &q->sched_breserved_tags; | |
3640 | } | |
70f36b60 | 3641 | } |
e3a2b3f9 JA |
3642 | if (ret) |
3643 | break; | |
77f1e0a5 JA |
3644 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3645 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 3646 | } |
d97e594c | 3647 | if (!ret) { |
e3a2b3f9 | 3648 | q->nr_requests = nr; |
d97e594c JG |
3649 | if (q->elevator && blk_mq_is_sbitmap_shared(set->flags)) |
3650 | sbitmap_queue_resize(&q->sched_bitmap_tags, | |
3651 | nr - set->reserved_tags); | |
3652 | } | |
e3a2b3f9 | 3653 | |
24f5a90f | 3654 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3655 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3656 | |
e3a2b3f9 JA |
3657 | return ret; |
3658 | } | |
3659 | ||
d48ece20 JW |
3660 | /* |
3661 | * request_queue and elevator_type pair. | |
3662 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3663 | * the elevator_type associated with a request_queue. | |
3664 | */ | |
3665 | struct blk_mq_qe_pair { | |
3666 | struct list_head node; | |
3667 | struct request_queue *q; | |
3668 | struct elevator_type *type; | |
3669 | }; | |
3670 | ||
3671 | /* | |
3672 | * Cache the elevator_type in qe pair list and switch the | |
3673 | * io scheduler to 'none' | |
3674 | */ | |
3675 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3676 | struct request_queue *q) | |
3677 | { | |
3678 | struct blk_mq_qe_pair *qe; | |
3679 | ||
3680 | if (!q->elevator) | |
3681 | return true; | |
3682 | ||
3683 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3684 | if (!qe) | |
3685 | return false; | |
3686 | ||
3687 | INIT_LIST_HEAD(&qe->node); | |
3688 | qe->q = q; | |
3689 | qe->type = q->elevator->type; | |
3690 | list_add(&qe->node, head); | |
3691 | ||
3692 | mutex_lock(&q->sysfs_lock); | |
3693 | /* | |
3694 | * After elevator_switch_mq, the previous elevator_queue will be | |
3695 | * released by elevator_release. The reference of the io scheduler | |
3696 | * module get by elevator_get will also be put. So we need to get | |
3697 | * a reference of the io scheduler module here to prevent it to be | |
3698 | * removed. | |
3699 | */ | |
3700 | __module_get(qe->type->elevator_owner); | |
3701 | elevator_switch_mq(q, NULL); | |
3702 | mutex_unlock(&q->sysfs_lock); | |
3703 | ||
3704 | return true; | |
3705 | } | |
3706 | ||
3707 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3708 | struct request_queue *q) | |
3709 | { | |
3710 | struct blk_mq_qe_pair *qe; | |
3711 | struct elevator_type *t = NULL; | |
3712 | ||
3713 | list_for_each_entry(qe, head, node) | |
3714 | if (qe->q == q) { | |
3715 | t = qe->type; | |
3716 | break; | |
3717 | } | |
3718 | ||
3719 | if (!t) | |
3720 | return; | |
3721 | ||
3722 | list_del(&qe->node); | |
3723 | kfree(qe); | |
3724 | ||
3725 | mutex_lock(&q->sysfs_lock); | |
3726 | elevator_switch_mq(q, t); | |
3727 | mutex_unlock(&q->sysfs_lock); | |
3728 | } | |
3729 | ||
e4dc2b32 KB |
3730 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3731 | int nr_hw_queues) | |
868f2f0b KB |
3732 | { |
3733 | struct request_queue *q; | |
d48ece20 | 3734 | LIST_HEAD(head); |
e01ad46d | 3735 | int prev_nr_hw_queues; |
868f2f0b | 3736 | |
705cda97 BVA |
3737 | lockdep_assert_held(&set->tag_list_lock); |
3738 | ||
392546ae | 3739 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3740 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3741 | if (nr_hw_queues < 1) |
3742 | return; | |
3743 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3744 | return; |
3745 | ||
3746 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3747 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3748 | /* |
3749 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3750 | * with the previous scheduler. We will switch back once we are done | |
3751 | * updating the new sw to hw queue mappings. | |
3752 | */ | |
3753 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3754 | if (!blk_mq_elv_switch_none(&head, q)) | |
3755 | goto switch_back; | |
868f2f0b | 3756 | |
477e19de JW |
3757 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3758 | blk_mq_debugfs_unregister_hctxs(q); | |
3759 | blk_mq_sysfs_unregister(q); | |
3760 | } | |
3761 | ||
a2584e43 | 3762 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3763 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3764 | 0) | |
3765 | goto reregister; | |
3766 | ||
868f2f0b | 3767 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3768 | fallback: |
aa880ad6 | 3769 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3770 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3771 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3772 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3773 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3774 | nr_hw_queues, prev_nr_hw_queues); | |
3775 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3776 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3777 | goto fallback; |
3778 | } | |
477e19de JW |
3779 | blk_mq_map_swqueue(q); |
3780 | } | |
3781 | ||
f7e76dbc | 3782 | reregister: |
477e19de JW |
3783 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3784 | blk_mq_sysfs_register(q); | |
3785 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3786 | } |
3787 | ||
d48ece20 JW |
3788 | switch_back: |
3789 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3790 | blk_mq_elv_switch_back(&head, q); | |
3791 | ||
868f2f0b KB |
3792 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3793 | blk_mq_unfreeze_queue(q); | |
3794 | } | |
e4dc2b32 KB |
3795 | |
3796 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3797 | { | |
3798 | mutex_lock(&set->tag_list_lock); | |
3799 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3800 | mutex_unlock(&set->tag_list_lock); | |
3801 | } | |
868f2f0b KB |
3802 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3803 | ||
34dbad5d OS |
3804 | /* Enable polling stats and return whether they were already enabled. */ |
3805 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3806 | { | |
3807 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3808 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3809 | return true; |
3810 | blk_stat_add_callback(q, q->poll_cb); | |
3811 | return false; | |
3812 | } | |
3813 | ||
3814 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3815 | { | |
3816 | /* | |
3817 | * We don't arm the callback if polling stats are not enabled or the | |
3818 | * callback is already active. | |
3819 | */ | |
3820 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3821 | blk_stat_is_active(q->poll_cb)) | |
3822 | return; | |
3823 | ||
3824 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3825 | } | |
3826 | ||
3827 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3828 | { | |
3829 | struct request_queue *q = cb->data; | |
720b8ccc | 3830 | int bucket; |
34dbad5d | 3831 | |
720b8ccc SB |
3832 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3833 | if (cb->stat[bucket].nr_samples) | |
3834 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3835 | } | |
34dbad5d OS |
3836 | } |
3837 | ||
64f1c21e | 3838 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3839 | struct request *rq) |
3840 | { | |
64f1c21e | 3841 | unsigned long ret = 0; |
720b8ccc | 3842 | int bucket; |
64f1c21e JA |
3843 | |
3844 | /* | |
3845 | * If stats collection isn't on, don't sleep but turn it on for | |
3846 | * future users | |
3847 | */ | |
34dbad5d | 3848 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3849 | return 0; |
3850 | ||
64f1c21e JA |
3851 | /* |
3852 | * As an optimistic guess, use half of the mean service time | |
3853 | * for this type of request. We can (and should) make this smarter. | |
3854 | * For instance, if the completion latencies are tight, we can | |
3855 | * get closer than just half the mean. This is especially | |
3856 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3857 | * than ~10 usec. We do use the stats for the relevant IO size |
3858 | * if available which does lead to better estimates. | |
64f1c21e | 3859 | */ |
720b8ccc SB |
3860 | bucket = blk_mq_poll_stats_bkt(rq); |
3861 | if (bucket < 0) | |
3862 | return ret; | |
3863 | ||
3864 | if (q->poll_stat[bucket].nr_samples) | |
3865 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3866 | |
3867 | return ret; | |
3868 | } | |
3869 | ||
06426adf JA |
3870 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3871 | struct request *rq) | |
3872 | { | |
3873 | struct hrtimer_sleeper hs; | |
3874 | enum hrtimer_mode mode; | |
64f1c21e | 3875 | unsigned int nsecs; |
06426adf JA |
3876 | ktime_t kt; |
3877 | ||
76a86f9d | 3878 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3879 | return false; |
3880 | ||
3881 | /* | |
1052b8ac | 3882 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3883 | * |
64f1c21e JA |
3884 | * 0: use half of prev avg |
3885 | * >0: use this specific value | |
3886 | */ | |
1052b8ac | 3887 | if (q->poll_nsec > 0) |
64f1c21e JA |
3888 | nsecs = q->poll_nsec; |
3889 | else | |
cae740a0 | 3890 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3891 | |
3892 | if (!nsecs) | |
06426adf JA |
3893 | return false; |
3894 | ||
76a86f9d | 3895 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3896 | |
3897 | /* | |
3898 | * This will be replaced with the stats tracking code, using | |
3899 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3900 | */ | |
8b0e1953 | 3901 | kt = nsecs; |
06426adf JA |
3902 | |
3903 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3904 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3905 | hrtimer_set_expires(&hs.timer, kt); |
3906 | ||
06426adf | 3907 | do { |
5a61c363 | 3908 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3909 | break; |
3910 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3911 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3912 | if (hs.task) |
3913 | io_schedule(); | |
3914 | hrtimer_cancel(&hs.timer); | |
3915 | mode = HRTIMER_MODE_ABS; | |
3916 | } while (hs.task && !signal_pending(current)); | |
3917 | ||
3918 | __set_current_state(TASK_RUNNING); | |
3919 | destroy_hrtimer_on_stack(&hs.timer); | |
3920 | return true; | |
3921 | } | |
3922 | ||
1052b8ac JA |
3923 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3924 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3925 | { |
1052b8ac JA |
3926 | struct request *rq; |
3927 | ||
29ece8b4 | 3928 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3929 | return false; |
3930 | ||
3931 | if (!blk_qc_t_is_internal(cookie)) | |
3932 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3933 | else { | |
3934 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3935 | /* | |
3936 | * With scheduling, if the request has completed, we'll | |
3937 | * get a NULL return here, as we clear the sched tag when | |
3938 | * that happens. The request still remains valid, like always, | |
3939 | * so we should be safe with just the NULL check. | |
3940 | */ | |
3941 | if (!rq) | |
3942 | return false; | |
3943 | } | |
3944 | ||
cae740a0 | 3945 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3946 | } |
3947 | ||
529262d5 CH |
3948 | /** |
3949 | * blk_poll - poll for IO completions | |
3950 | * @q: the queue | |
3951 | * @cookie: cookie passed back at IO submission time | |
3952 | * @spin: whether to spin for completions | |
3953 | * | |
3954 | * Description: | |
3955 | * Poll for completions on the passed in queue. Returns number of | |
3956 | * completed entries found. If @spin is true, then blk_poll will continue | |
3957 | * looping until at least one completion is found, unless the task is | |
3958 | * otherwise marked running (or we need to reschedule). | |
3959 | */ | |
3960 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3961 | { |
3962 | struct blk_mq_hw_ctx *hctx; | |
2f064a59 | 3963 | unsigned int state; |
bbd7bb70 | 3964 | |
529262d5 CH |
3965 | if (!blk_qc_t_valid(cookie) || |
3966 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3967 | return 0; |
3968 | ||
529262d5 CH |
3969 | if (current->plug) |
3970 | blk_flush_plug_list(current->plug, false); | |
3971 | ||
1052b8ac JA |
3972 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3973 | ||
06426adf JA |
3974 | /* |
3975 | * If we sleep, have the caller restart the poll loop to reset | |
3976 | * the state. Like for the other success return cases, the | |
3977 | * caller is responsible for checking if the IO completed. If | |
3978 | * the IO isn't complete, we'll get called again and will go | |
f6f371f7 PB |
3979 | * straight to the busy poll loop. If specified not to spin, |
3980 | * we also should not sleep. | |
06426adf | 3981 | */ |
f6f371f7 | 3982 | if (spin && blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3983 | return 1; |
06426adf | 3984 | |
bbd7bb70 JA |
3985 | hctx->poll_considered++; |
3986 | ||
d6c23bb3 | 3987 | state = get_current_state(); |
aa61bec3 | 3988 | do { |
bbd7bb70 JA |
3989 | int ret; |
3990 | ||
3991 | hctx->poll_invoked++; | |
3992 | ||
9743139c | 3993 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3994 | if (ret > 0) { |
3995 | hctx->poll_success++; | |
849a3700 | 3996 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3997 | return ret; |
bbd7bb70 JA |
3998 | } |
3999 | ||
4000 | if (signal_pending_state(state, current)) | |
849a3700 | 4001 | __set_current_state(TASK_RUNNING); |
bbd7bb70 | 4002 | |
b03fbd4f | 4003 | if (task_is_running(current)) |
85f4d4b6 | 4004 | return 1; |
0a1b8b87 | 4005 | if (ret < 0 || !spin) |
bbd7bb70 JA |
4006 | break; |
4007 | cpu_relax(); | |
aa61bec3 | 4008 | } while (!need_resched()); |
bbd7bb70 | 4009 | |
67b4110f | 4010 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4011 | return 0; |
bbd7bb70 | 4012 | } |
529262d5 | 4013 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 4014 | |
9cf2bab6 JA |
4015 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4016 | { | |
4017 | return rq->mq_ctx->cpu; | |
4018 | } | |
4019 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4020 | ||
9a2b5cd4 ML |
4021 | void blk_mq_cancel_work_sync(struct request_queue *q) |
4022 | { | |
4023 | if (queue_is_mq(q)) { | |
4024 | struct blk_mq_hw_ctx *hctx; | |
4025 | int i; | |
4026 | ||
4027 | cancel_delayed_work_sync(&q->requeue_work); | |
4028 | ||
4029 | queue_for_each_hw_ctx(q, hctx, i) | |
4030 | cancel_delayed_work_sync(&hctx->run_work); | |
4031 | } | |
4032 | } | |
4033 | ||
320ae51f JA |
4034 | static int __init blk_mq_init(void) |
4035 | { | |
c3077b5d CH |
4036 | int i; |
4037 | ||
4038 | for_each_possible_cpu(i) | |
f9ab4918 | 4039 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4040 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4041 | ||
4042 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4043 | "block/softirq:dead", NULL, | |
4044 | blk_softirq_cpu_dead); | |
9467f859 TG |
4045 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4046 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4047 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4048 | blk_mq_hctx_notify_online, | |
4049 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4050 | return 0; |
4051 | } | |
4052 | subsys_initcall(blk_mq_init); |