Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
f75782e4 | 13 | #include <linux/kmemleak.h> |
320ae51f JA |
14 | #include <linux/mm.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/smp.h> | |
19 | #include <linux/llist.h> | |
20 | #include <linux/list_sort.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/cache.h> | |
23 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
a892c8d5 | 29 | #include <linux/blk-crypto.h> |
320ae51f JA |
30 | |
31 | #include <trace/events/block.h> | |
32 | ||
33 | #include <linux/blk-mq.h> | |
54d4e6ab | 34 | #include <linux/t10-pi.h> |
320ae51f JA |
35 | #include "blk.h" |
36 | #include "blk-mq.h" | |
9c1051aa | 37 | #include "blk-mq-debugfs.h" |
320ae51f | 38 | #include "blk-mq-tag.h" |
986d413b | 39 | #include "blk-pm.h" |
cf43e6be | 40 | #include "blk-stat.h" |
bd166ef1 | 41 | #include "blk-mq-sched.h" |
c1c80384 | 42 | #include "blk-rq-qos.h" |
320ae51f | 43 | |
c3077b5d CH |
44 | static DEFINE_PER_CPU(struct list_head, blk_cpu_done); |
45 | ||
34dbad5d OS |
46 | static void blk_mq_poll_stats_start(struct request_queue *q); |
47 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
48 | ||
720b8ccc SB |
49 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
50 | { | |
3d244306 | 51 | int ddir, sectors, bucket; |
720b8ccc | 52 | |
99c749a4 | 53 | ddir = rq_data_dir(rq); |
3d244306 | 54 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 55 | |
3d244306 | 56 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
57 | |
58 | if (bucket < 0) | |
59 | return -1; | |
60 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
61 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
62 | ||
63 | return bucket; | |
64 | } | |
65 | ||
320ae51f | 66 | /* |
85fae294 YY |
67 | * Check if any of the ctx, dispatch list or elevator |
68 | * have pending work in this hardware queue. | |
320ae51f | 69 | */ |
79f720a7 | 70 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 71 | { |
79f720a7 JA |
72 | return !list_empty_careful(&hctx->dispatch) || |
73 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 74 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
75 | } |
76 | ||
320ae51f JA |
77 | /* |
78 | * Mark this ctx as having pending work in this hardware queue | |
79 | */ | |
80 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
f31967f0 JA |
83 | const int bit = ctx->index_hw[hctx->type]; |
84 | ||
85 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
86 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
87 | } |
88 | ||
89 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
90 | struct blk_mq_ctx *ctx) | |
91 | { | |
f31967f0 JA |
92 | const int bit = ctx->index_hw[hctx->type]; |
93 | ||
94 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
95 | } |
96 | ||
f299b7c7 | 97 | struct mq_inflight { |
8446fe92 | 98 | struct block_device *part; |
a2e80f6f | 99 | unsigned int inflight[2]; |
f299b7c7 JA |
100 | }; |
101 | ||
7baa8572 | 102 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
103 | struct request *rq, void *priv, |
104 | bool reserved) | |
105 | { | |
106 | struct mq_inflight *mi = priv; | |
107 | ||
b0d97557 JX |
108 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
109 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 110 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
111 | |
112 | return true; | |
f299b7c7 JA |
113 | } |
114 | ||
8446fe92 CH |
115 | unsigned int blk_mq_in_flight(struct request_queue *q, |
116 | struct block_device *part) | |
f299b7c7 | 117 | { |
a2e80f6f | 118 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 119 | |
f299b7c7 | 120 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 121 | |
a2e80f6f | 122 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
123 | } |
124 | ||
8446fe92 CH |
125 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
126 | unsigned int inflight[2]) | |
bf0ddaba | 127 | { |
a2e80f6f | 128 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 129 | |
bb4e6b14 | 130 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
131 | inflight[0] = mi.inflight[0]; |
132 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
133 | } |
134 | ||
1671d522 | 135 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 136 | { |
7996a8b5 BL |
137 | mutex_lock(&q->mq_freeze_lock); |
138 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 139 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 140 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 141 | if (queue_is_mq(q)) |
055f6e18 | 142 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
143 | } else { |
144 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 145 | } |
f3af020b | 146 | } |
1671d522 | 147 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 148 | |
6bae363e | 149 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 150 | { |
3ef28e83 | 151 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 152 | } |
6bae363e | 153 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 154 | |
f91328c4 KB |
155 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
156 | unsigned long timeout) | |
157 | { | |
158 | return wait_event_timeout(q->mq_freeze_wq, | |
159 | percpu_ref_is_zero(&q->q_usage_counter), | |
160 | timeout); | |
161 | } | |
162 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 163 | |
f3af020b TH |
164 | /* |
165 | * Guarantee no request is in use, so we can change any data structure of | |
166 | * the queue afterward. | |
167 | */ | |
3ef28e83 | 168 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 169 | { |
3ef28e83 DW |
170 | /* |
171 | * In the !blk_mq case we are only calling this to kill the | |
172 | * q_usage_counter, otherwise this increases the freeze depth | |
173 | * and waits for it to return to zero. For this reason there is | |
174 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
175 | * exported to drivers as the only user for unfreeze is blk_mq. | |
176 | */ | |
1671d522 | 177 | blk_freeze_queue_start(q); |
f3af020b TH |
178 | blk_mq_freeze_queue_wait(q); |
179 | } | |
3ef28e83 DW |
180 | |
181 | void blk_mq_freeze_queue(struct request_queue *q) | |
182 | { | |
183 | /* | |
184 | * ...just an alias to keep freeze and unfreeze actions balanced | |
185 | * in the blk_mq_* namespace | |
186 | */ | |
187 | blk_freeze_queue(q); | |
188 | } | |
c761d96b | 189 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 190 | |
b4c6a028 | 191 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 192 | { |
7996a8b5 BL |
193 | mutex_lock(&q->mq_freeze_lock); |
194 | q->mq_freeze_depth--; | |
195 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
196 | if (!q->mq_freeze_depth) { | |
bdd63160 | 197 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 198 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 199 | } |
7996a8b5 | 200 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 201 | } |
b4c6a028 | 202 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 203 | |
852ec809 BVA |
204 | /* |
205 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
206 | * mpt3sas driver such that this function can be removed. | |
207 | */ | |
208 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
209 | { | |
8814ce8a | 210 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
211 | } |
212 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
213 | ||
6a83e74d | 214 | /** |
69e07c4a | 215 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
216 | * @q: request queue. |
217 | * | |
218 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
219 | * callback function is invoked. Once this function is returned, we make |
220 | * sure no dispatch can happen until the queue is unquiesced via | |
221 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
222 | */ |
223 | void blk_mq_quiesce_queue(struct request_queue *q) | |
224 | { | |
225 | struct blk_mq_hw_ctx *hctx; | |
226 | unsigned int i; | |
227 | bool rcu = false; | |
228 | ||
1d9e9bc6 | 229 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 230 | |
6a83e74d BVA |
231 | queue_for_each_hw_ctx(q, hctx, i) { |
232 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 233 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
234 | else |
235 | rcu = true; | |
236 | } | |
237 | if (rcu) | |
238 | synchronize_rcu(); | |
239 | } | |
240 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
241 | ||
e4e73913 ML |
242 | /* |
243 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
244 | * @q: request queue. | |
245 | * | |
246 | * This function recovers queue into the state before quiescing | |
247 | * which is done by blk_mq_quiesce_queue. | |
248 | */ | |
249 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
250 | { | |
8814ce8a | 251 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 252 | |
1d9e9bc6 ML |
253 | /* dispatch requests which are inserted during quiescing */ |
254 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
255 | } |
256 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
257 | ||
aed3ea94 JA |
258 | void blk_mq_wake_waiters(struct request_queue *q) |
259 | { | |
260 | struct blk_mq_hw_ctx *hctx; | |
261 | unsigned int i; | |
262 | ||
263 | queue_for_each_hw_ctx(q, hctx, i) | |
264 | if (blk_mq_hw_queue_mapped(hctx)) | |
265 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
266 | } | |
267 | ||
fe1f4526 | 268 | /* |
9a91b05b HT |
269 | * Only need start/end time stamping if we have iostat or |
270 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
271 | */ |
272 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
273 | { | |
9a91b05b | 274 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
275 | } |
276 | ||
e4cdf1a1 | 277 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 278 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 279 | { |
e4cdf1a1 CH |
280 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
281 | struct request *rq = tags->static_rqs[tag]; | |
c3a148d2 | 282 | |
42fdc5e4 | 283 | if (data->q->elevator) { |
76647368 | 284 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
285 | rq->internal_tag = tag; |
286 | } else { | |
e4cdf1a1 | 287 | rq->tag = tag; |
76647368 | 288 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
289 | } |
290 | ||
af76e555 | 291 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
292 | rq->q = data->q; |
293 | rq->mq_ctx = data->ctx; | |
ea4f995e | 294 | rq->mq_hctx = data->hctx; |
568f2700 | 295 | rq->rq_flags = 0; |
7ea4d8a4 | 296 | rq->cmd_flags = data->cmd_flags; |
1b6d65a0 BVA |
297 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
298 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 299 | if (blk_queue_io_stat(data->q)) |
e8064021 | 300 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 301 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
302 | INIT_HLIST_NODE(&rq->hash); |
303 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
304 | rq->rq_disk = NULL; |
305 | rq->part = NULL; | |
6f816b4b TH |
306 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
307 | rq->alloc_time_ns = alloc_time_ns; | |
308 | #endif | |
fe1f4526 JA |
309 | if (blk_mq_need_time_stamp(rq)) |
310 | rq->start_time_ns = ktime_get_ns(); | |
311 | else | |
312 | rq->start_time_ns = 0; | |
544ccc8d | 313 | rq->io_start_time_ns = 0; |
3d244306 | 314 | rq->stats_sectors = 0; |
af76e555 CH |
315 | rq->nr_phys_segments = 0; |
316 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
317 | rq->nr_integrity_segments = 0; | |
318 | #endif | |
a892c8d5 | 319 | blk_crypto_rq_set_defaults(rq); |
af76e555 | 320 | /* tag was already set */ |
079076b3 | 321 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 322 | |
f6be4fb4 JA |
323 | rq->timeout = 0; |
324 | ||
af76e555 CH |
325 | rq->end_io = NULL; |
326 | rq->end_io_data = NULL; | |
af76e555 | 327 | |
7ea4d8a4 | 328 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; |
12f5b931 | 329 | refcount_set(&rq->ref, 1); |
7ea4d8a4 CH |
330 | |
331 | if (!op_is_flush(data->cmd_flags)) { | |
332 | struct elevator_queue *e = data->q->elevator; | |
333 | ||
334 | rq->elv.icq = NULL; | |
335 | if (e && e->type->ops.prepare_request) { | |
336 | if (e->type->icq_cache) | |
337 | blk_mq_sched_assign_ioc(rq); | |
338 | ||
339 | e->type->ops.prepare_request(rq); | |
340 | rq->rq_flags |= RQF_ELVPRIV; | |
341 | } | |
342 | } | |
343 | ||
344 | data->hctx->queued++; | |
e4cdf1a1 | 345 | return rq; |
5dee8577 CH |
346 | } |
347 | ||
e6e7abff | 348 | static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data) |
d2c0d383 | 349 | { |
e6e7abff | 350 | struct request_queue *q = data->q; |
d2c0d383 | 351 | struct elevator_queue *e = q->elevator; |
6f816b4b | 352 | u64 alloc_time_ns = 0; |
600c3b0c | 353 | unsigned int tag; |
d2c0d383 | 354 | |
6f816b4b TH |
355 | /* alloc_time includes depth and tag waits */ |
356 | if (blk_queue_rq_alloc_time(q)) | |
357 | alloc_time_ns = ktime_get_ns(); | |
358 | ||
f9afca4d | 359 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 360 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
361 | |
362 | if (e) { | |
d2c0d383 CH |
363 | /* |
364 | * Flush requests are special and go directly to the | |
17a51199 JA |
365 | * dispatch list. Don't include reserved tags in the |
366 | * limiting, as it isn't useful. | |
d2c0d383 | 367 | */ |
f9afca4d JA |
368 | if (!op_is_flush(data->cmd_flags) && |
369 | e->type->ops.limit_depth && | |
17a51199 | 370 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 371 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
372 | } |
373 | ||
bf0beec0 | 374 | retry: |
600c3b0c CH |
375 | data->ctx = blk_mq_get_ctx(q); |
376 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 377 | if (!e) |
600c3b0c CH |
378 | blk_mq_tag_busy(data->hctx); |
379 | ||
bf0beec0 ML |
380 | /* |
381 | * Waiting allocations only fail because of an inactive hctx. In that | |
382 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
383 | * should have migrated us to an online CPU by now. | |
384 | */ | |
e4cdf1a1 | 385 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
386 | if (tag == BLK_MQ_NO_TAG) { |
387 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
388 | return NULL; | |
389 | ||
390 | /* | |
391 | * Give up the CPU and sleep for a random short time to ensure | |
392 | * that thread using a realtime scheduling class are migrated | |
70f15a4f | 393 | * off the CPU, and thus off the hctx that is going away. |
bf0beec0 ML |
394 | */ |
395 | msleep(3); | |
396 | goto retry; | |
397 | } | |
7ea4d8a4 | 398 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
399 | } |
400 | ||
cd6ce148 | 401 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 402 | blk_mq_req_flags_t flags) |
320ae51f | 403 | { |
e6e7abff CH |
404 | struct blk_mq_alloc_data data = { |
405 | .q = q, | |
406 | .flags = flags, | |
407 | .cmd_flags = op, | |
408 | }; | |
bd166ef1 | 409 | struct request *rq; |
a492f075 | 410 | int ret; |
320ae51f | 411 | |
3a0a5299 | 412 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
413 | if (ret) |
414 | return ERR_PTR(ret); | |
320ae51f | 415 | |
e6e7abff | 416 | rq = __blk_mq_alloc_request(&data); |
bd166ef1 | 417 | if (!rq) |
a5ea5811 | 418 | goto out_queue_exit; |
0c4de0f3 CH |
419 | rq->__data_len = 0; |
420 | rq->__sector = (sector_t) -1; | |
421 | rq->bio = rq->biotail = NULL; | |
320ae51f | 422 | return rq; |
a5ea5811 CH |
423 | out_queue_exit: |
424 | blk_queue_exit(q); | |
425 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 426 | } |
4bb659b1 | 427 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 428 | |
cd6ce148 | 429 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 430 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 431 | { |
e6e7abff CH |
432 | struct blk_mq_alloc_data data = { |
433 | .q = q, | |
434 | .flags = flags, | |
435 | .cmd_flags = op, | |
436 | }; | |
600c3b0c | 437 | u64 alloc_time_ns = 0; |
6d2809d5 | 438 | unsigned int cpu; |
600c3b0c | 439 | unsigned int tag; |
1f5bd336 ML |
440 | int ret; |
441 | ||
600c3b0c CH |
442 | /* alloc_time includes depth and tag waits */ |
443 | if (blk_queue_rq_alloc_time(q)) | |
444 | alloc_time_ns = ktime_get_ns(); | |
445 | ||
1f5bd336 ML |
446 | /* |
447 | * If the tag allocator sleeps we could get an allocation for a | |
448 | * different hardware context. No need to complicate the low level | |
449 | * allocator for this for the rare use case of a command tied to | |
450 | * a specific queue. | |
451 | */ | |
600c3b0c | 452 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
453 | return ERR_PTR(-EINVAL); |
454 | ||
455 | if (hctx_idx >= q->nr_hw_queues) | |
456 | return ERR_PTR(-EIO); | |
457 | ||
3a0a5299 | 458 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
459 | if (ret) |
460 | return ERR_PTR(ret); | |
461 | ||
c8712c6a CH |
462 | /* |
463 | * Check if the hardware context is actually mapped to anything. | |
464 | * If not tell the caller that it should skip this queue. | |
465 | */ | |
a5ea5811 | 466 | ret = -EXDEV; |
e6e7abff CH |
467 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
468 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 469 | goto out_queue_exit; |
e6e7abff CH |
470 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
471 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 472 | |
42fdc5e4 | 473 | if (!q->elevator) |
600c3b0c CH |
474 | blk_mq_tag_busy(data.hctx); |
475 | ||
a5ea5811 | 476 | ret = -EWOULDBLOCK; |
600c3b0c CH |
477 | tag = blk_mq_get_tag(&data); |
478 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 479 | goto out_queue_exit; |
600c3b0c CH |
480 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
481 | ||
a5ea5811 CH |
482 | out_queue_exit: |
483 | blk_queue_exit(q); | |
484 | return ERR_PTR(ret); | |
1f5bd336 ML |
485 | } |
486 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
487 | ||
12f5b931 KB |
488 | static void __blk_mq_free_request(struct request *rq) |
489 | { | |
490 | struct request_queue *q = rq->q; | |
491 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 492 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
493 | const int sched_tag = rq->internal_tag; |
494 | ||
a892c8d5 | 495 | blk_crypto_free_request(rq); |
986d413b | 496 | blk_pm_mark_last_busy(rq); |
ea4f995e | 497 | rq->mq_hctx = NULL; |
76647368 | 498 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 499 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 500 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 501 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
502 | blk_mq_sched_restart(hctx); |
503 | blk_queue_exit(q); | |
504 | } | |
505 | ||
6af54051 | 506 | void blk_mq_free_request(struct request *rq) |
320ae51f | 507 | { |
320ae51f | 508 | struct request_queue *q = rq->q; |
6af54051 CH |
509 | struct elevator_queue *e = q->elevator; |
510 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 511 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 512 | |
5bbf4e5a | 513 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
514 | if (e && e->type->ops.finish_request) |
515 | e->type->ops.finish_request(rq); | |
6af54051 CH |
516 | if (rq->elv.icq) { |
517 | put_io_context(rq->elv.icq->ioc); | |
518 | rq->elv.icq = NULL; | |
519 | } | |
520 | } | |
320ae51f | 521 | |
6af54051 | 522 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 523 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 524 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 525 | |
7beb2f84 JA |
526 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
527 | laptop_io_completion(q->backing_dev_info); | |
528 | ||
a7905043 | 529 | rq_qos_done(q, rq); |
0d2602ca | 530 | |
12f5b931 KB |
531 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
532 | if (refcount_dec_and_test(&rq->ref)) | |
533 | __blk_mq_free_request(rq); | |
320ae51f | 534 | } |
1a3b595a | 535 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 536 | |
2a842aca | 537 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 538 | { |
fe1f4526 JA |
539 | u64 now = 0; |
540 | ||
541 | if (blk_mq_need_time_stamp(rq)) | |
542 | now = ktime_get_ns(); | |
522a7775 | 543 | |
4bc6339a OS |
544 | if (rq->rq_flags & RQF_STATS) { |
545 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 546 | blk_stat_add(rq, now); |
4bc6339a OS |
547 | } |
548 | ||
87890092 | 549 | blk_mq_sched_completed_request(rq, now); |
ed88660a | 550 | |
522a7775 | 551 | blk_account_io_done(rq, now); |
0d11e6ac | 552 | |
91b63639 | 553 | if (rq->end_io) { |
a7905043 | 554 | rq_qos_done(rq->q, rq); |
320ae51f | 555 | rq->end_io(rq, error); |
91b63639 | 556 | } else { |
320ae51f | 557 | blk_mq_free_request(rq); |
91b63639 | 558 | } |
320ae51f | 559 | } |
c8a446ad | 560 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 561 | |
2a842aca | 562 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
563 | { |
564 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
565 | BUG(); | |
c8a446ad | 566 | __blk_mq_end_request(rq, error); |
63151a44 | 567 | } |
c8a446ad | 568 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 569 | |
c3077b5d CH |
570 | /* |
571 | * Softirq action handler - move entries to local list and loop over them | |
572 | * while passing them to the queue registered handler. | |
573 | */ | |
574 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) | |
320ae51f | 575 | { |
c3077b5d | 576 | struct list_head *cpu_list, local_list; |
320ae51f | 577 | |
c3077b5d CH |
578 | local_irq_disable(); |
579 | cpu_list = this_cpu_ptr(&blk_cpu_done); | |
580 | list_replace_init(cpu_list, &local_list); | |
581 | local_irq_enable(); | |
582 | ||
583 | while (!list_empty(&local_list)) { | |
584 | struct request *rq; | |
585 | ||
586 | rq = list_entry(local_list.next, struct request, ipi_list); | |
587 | list_del_init(&rq->ipi_list); | |
588 | rq->q->mq_ops->complete(rq); | |
589 | } | |
320ae51f | 590 | } |
320ae51f | 591 | |
115243f5 | 592 | static void blk_mq_trigger_softirq(struct request *rq) |
320ae51f | 593 | { |
d391a7a3 CH |
594 | struct list_head *list; |
595 | unsigned long flags; | |
c3077b5d | 596 | |
d391a7a3 CH |
597 | local_irq_save(flags); |
598 | list = this_cpu_ptr(&blk_cpu_done); | |
c3077b5d CH |
599 | list_add_tail(&rq->ipi_list, list); |
600 | ||
115243f5 CH |
601 | /* |
602 | * If the list only contains our just added request, signal a raise of | |
603 | * the softirq. If there are already entries there, someone already | |
604 | * raised the irq but it hasn't run yet. | |
605 | */ | |
c3077b5d CH |
606 | if (list->next == &rq->ipi_list) |
607 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
d391a7a3 | 608 | local_irq_restore(flags); |
115243f5 CH |
609 | } |
610 | ||
c3077b5d CH |
611 | static int blk_softirq_cpu_dead(unsigned int cpu) |
612 | { | |
613 | /* | |
614 | * If a CPU goes away, splice its entries to the current CPU | |
615 | * and trigger a run of the softirq | |
616 | */ | |
617 | local_irq_disable(); | |
618 | list_splice_init(&per_cpu(blk_cpu_done, cpu), | |
619 | this_cpu_ptr(&blk_cpu_done)); | |
620 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
621 | local_irq_enable(); | |
622 | ||
623 | return 0; | |
624 | } | |
625 | ||
40d09b53 CH |
626 | |
627 | static void __blk_mq_complete_request_remote(void *data) | |
c3077b5d | 628 | { |
40d09b53 | 629 | struct request *rq = data; |
320ae51f | 630 | |
36e76539 | 631 | /* |
d391a7a3 CH |
632 | * For most of single queue controllers, there is only one irq vector |
633 | * for handling I/O completion, and the only irq's affinity is set | |
634 | * to all possible CPUs. On most of ARCHs, this affinity means the irq | |
635 | * is handled on one specific CPU. | |
36e76539 | 636 | * |
d391a7a3 CH |
637 | * So complete I/O requests in softirq context in case of single queue |
638 | * devices to avoid degrading I/O performance due to irqsoff latency. | |
36e76539 | 639 | */ |
d391a7a3 CH |
640 | if (rq->q->nr_hw_queues == 1) |
641 | blk_mq_trigger_softirq(rq); | |
642 | else | |
643 | rq->q->mq_ops->complete(rq); | |
c3077b5d CH |
644 | } |
645 | ||
96339526 CH |
646 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
647 | { | |
648 | int cpu = raw_smp_processor_id(); | |
649 | ||
650 | if (!IS_ENABLED(CONFIG_SMP) || | |
651 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
652 | return false; | |
653 | ||
654 | /* same CPU or cache domain? Complete locally */ | |
655 | if (cpu == rq->mq_ctx->cpu || | |
656 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
657 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
658 | return false; | |
659 | ||
660 | /* don't try to IPI to an offline CPU */ | |
661 | return cpu_online(rq->mq_ctx->cpu); | |
662 | } | |
663 | ||
40d09b53 | 664 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 665 | { |
af78ff7c | 666 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 667 | |
4ab32bf3 JA |
668 | /* |
669 | * For a polled request, always complete locallly, it's pointless | |
670 | * to redirect the completion. | |
671 | */ | |
40d09b53 CH |
672 | if (rq->cmd_flags & REQ_HIPRI) |
673 | return false; | |
38535201 | 674 | |
96339526 | 675 | if (blk_mq_complete_need_ipi(rq)) { |
30a91cb4 | 676 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
677 | rq->csd.info = rq; |
678 | rq->csd.flags = 0; | |
96339526 | 679 | smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd); |
3d6efbf6 | 680 | } else { |
40d09b53 CH |
681 | if (rq->q->nr_hw_queues > 1) |
682 | return false; | |
683 | blk_mq_trigger_softirq(rq); | |
3d6efbf6 | 684 | } |
40d09b53 CH |
685 | |
686 | return true; | |
687 | } | |
688 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
689 | ||
690 | /** | |
691 | * blk_mq_complete_request - end I/O on a request | |
692 | * @rq: the request being processed | |
693 | * | |
694 | * Description: | |
695 | * Complete a request by scheduling the ->complete_rq operation. | |
696 | **/ | |
697 | void blk_mq_complete_request(struct request *rq) | |
698 | { | |
699 | if (!blk_mq_complete_request_remote(rq)) | |
700 | rq->q->mq_ops->complete(rq); | |
320ae51f | 701 | } |
15f73f5b | 702 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 703 | |
04ced159 | 704 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 705 | __releases(hctx->srcu) |
04ced159 JA |
706 | { |
707 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
708 | rcu_read_unlock(); | |
709 | else | |
05707b64 | 710 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
711 | } |
712 | ||
713 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 714 | __acquires(hctx->srcu) |
04ced159 | 715 | { |
08b5a6e2 JA |
716 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
717 | /* shut up gcc false positive */ | |
718 | *srcu_idx = 0; | |
04ced159 | 719 | rcu_read_lock(); |
08b5a6e2 | 720 | } else |
05707b64 | 721 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
722 | } |
723 | ||
105663f7 AA |
724 | /** |
725 | * blk_mq_start_request - Start processing a request | |
726 | * @rq: Pointer to request to be started | |
727 | * | |
728 | * Function used by device drivers to notify the block layer that a request | |
729 | * is going to be processed now, so blk layer can do proper initializations | |
730 | * such as starting the timeout timer. | |
731 | */ | |
e2490073 | 732 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
733 | { |
734 | struct request_queue *q = rq->q; | |
735 | ||
736 | trace_block_rq_issue(q, rq); | |
737 | ||
cf43e6be | 738 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 739 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 740 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 741 | rq->rq_flags |= RQF_STATS; |
a7905043 | 742 | rq_qos_issue(q, rq); |
cf43e6be JA |
743 | } |
744 | ||
1d9bd516 | 745 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 746 | |
1d9bd516 | 747 | blk_add_timer(rq); |
12f5b931 | 748 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 749 | |
54d4e6ab MG |
750 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
751 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
752 | q->integrity.profile->prepare_fn(rq); | |
753 | #endif | |
320ae51f | 754 | } |
e2490073 | 755 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 756 | |
ed0791b2 | 757 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
758 | { |
759 | struct request_queue *q = rq->q; | |
760 | ||
923218f6 ML |
761 | blk_mq_put_driver_tag(rq); |
762 | ||
320ae51f | 763 | trace_block_rq_requeue(q, rq); |
a7905043 | 764 | rq_qos_requeue(q, rq); |
49f5baa5 | 765 | |
12f5b931 KB |
766 | if (blk_mq_request_started(rq)) { |
767 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 768 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 769 | } |
320ae51f JA |
770 | } |
771 | ||
2b053aca | 772 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 773 | { |
ed0791b2 | 774 | __blk_mq_requeue_request(rq); |
ed0791b2 | 775 | |
105976f5 ML |
776 | /* this request will be re-inserted to io scheduler queue */ |
777 | blk_mq_sched_requeue_request(rq); | |
778 | ||
7d692330 | 779 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 780 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
781 | } |
782 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
783 | ||
6fca6a61 CH |
784 | static void blk_mq_requeue_work(struct work_struct *work) |
785 | { | |
786 | struct request_queue *q = | |
2849450a | 787 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
788 | LIST_HEAD(rq_list); |
789 | struct request *rq, *next; | |
6fca6a61 | 790 | |
18e9781d | 791 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 792 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 793 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
794 | |
795 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 796 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
797 | continue; |
798 | ||
e8064021 | 799 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 800 | list_del_init(&rq->queuelist); |
aef1897c JW |
801 | /* |
802 | * If RQF_DONTPREP, rq has contained some driver specific | |
803 | * data, so insert it to hctx dispatch list to avoid any | |
804 | * merge. | |
805 | */ | |
806 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 807 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
808 | else |
809 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
810 | } |
811 | ||
812 | while (!list_empty(&rq_list)) { | |
813 | rq = list_entry(rq_list.next, struct request, queuelist); | |
814 | list_del_init(&rq->queuelist); | |
9e97d295 | 815 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
816 | } |
817 | ||
52d7f1b5 | 818 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
819 | } |
820 | ||
2b053aca BVA |
821 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
822 | bool kick_requeue_list) | |
6fca6a61 CH |
823 | { |
824 | struct request_queue *q = rq->q; | |
825 | unsigned long flags; | |
826 | ||
827 | /* | |
828 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 829 | * request head insertion from the workqueue. |
6fca6a61 | 830 | */ |
e8064021 | 831 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
832 | |
833 | spin_lock_irqsave(&q->requeue_lock, flags); | |
834 | if (at_head) { | |
e8064021 | 835 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
836 | list_add(&rq->queuelist, &q->requeue_list); |
837 | } else { | |
838 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
839 | } | |
840 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
841 | |
842 | if (kick_requeue_list) | |
843 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 844 | } |
6fca6a61 CH |
845 | |
846 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
847 | { | |
ae943d20 | 848 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
849 | } |
850 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
851 | ||
2849450a MS |
852 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
853 | unsigned long msecs) | |
854 | { | |
d4acf365 BVA |
855 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
856 | msecs_to_jiffies(msecs)); | |
2849450a MS |
857 | } |
858 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
859 | ||
0e62f51f JA |
860 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
861 | { | |
88c7b2b7 JA |
862 | if (tag < tags->nr_tags) { |
863 | prefetch(tags->rqs[tag]); | |
4ee86bab | 864 | return tags->rqs[tag]; |
88c7b2b7 | 865 | } |
4ee86bab HR |
866 | |
867 | return NULL; | |
24d2f903 CH |
868 | } |
869 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
870 | ||
3c94d83c JA |
871 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
872 | void *priv, bool reserved) | |
ae879912 JA |
873 | { |
874 | /* | |
05a4fed6 | 875 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 876 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 877 | */ |
05a4fed6 | 878 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
879 | bool *busy = priv; |
880 | ||
881 | *busy = true; | |
882 | return false; | |
883 | } | |
884 | ||
885 | return true; | |
886 | } | |
887 | ||
3c94d83c | 888 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
889 | { |
890 | bool busy = false; | |
891 | ||
3c94d83c | 892 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
893 | return busy; |
894 | } | |
3c94d83c | 895 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 896 | |
358f70da | 897 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 898 | { |
da661267 | 899 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
900 | if (req->q->mq_ops->timeout) { |
901 | enum blk_eh_timer_return ret; | |
902 | ||
903 | ret = req->q->mq_ops->timeout(req, reserved); | |
904 | if (ret == BLK_EH_DONE) | |
905 | return; | |
906 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 907 | } |
d1210d5a CH |
908 | |
909 | blk_add_timer(req); | |
87ee7b11 | 910 | } |
5b3f25fc | 911 | |
12f5b931 | 912 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 913 | { |
12f5b931 | 914 | unsigned long deadline; |
87ee7b11 | 915 | |
12f5b931 KB |
916 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
917 | return false; | |
da661267 CH |
918 | if (rq->rq_flags & RQF_TIMED_OUT) |
919 | return false; | |
a7af0af3 | 920 | |
079076b3 | 921 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
922 | if (time_after_eq(jiffies, deadline)) |
923 | return true; | |
a7af0af3 | 924 | |
12f5b931 KB |
925 | if (*next == 0) |
926 | *next = deadline; | |
927 | else if (time_after(*next, deadline)) | |
928 | *next = deadline; | |
929 | return false; | |
87ee7b11 JA |
930 | } |
931 | ||
7baa8572 | 932 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
933 | struct request *rq, void *priv, bool reserved) |
934 | { | |
12f5b931 KB |
935 | unsigned long *next = priv; |
936 | ||
937 | /* | |
938 | * Just do a quick check if it is expired before locking the request in | |
939 | * so we're not unnecessarilly synchronizing across CPUs. | |
940 | */ | |
941 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 942 | return true; |
12f5b931 KB |
943 | |
944 | /* | |
945 | * We have reason to believe the request may be expired. Take a | |
946 | * reference on the request to lock this request lifetime into its | |
947 | * currently allocated context to prevent it from being reallocated in | |
948 | * the event the completion by-passes this timeout handler. | |
949 | * | |
950 | * If the reference was already released, then the driver beat the | |
951 | * timeout handler to posting a natural completion. | |
952 | */ | |
953 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 954 | return true; |
12f5b931 | 955 | |
1d9bd516 | 956 | /* |
12f5b931 KB |
957 | * The request is now locked and cannot be reallocated underneath the |
958 | * timeout handler's processing. Re-verify this exact request is truly | |
959 | * expired; if it is not expired, then the request was completed and | |
960 | * reallocated as a new request. | |
1d9bd516 | 961 | */ |
12f5b931 | 962 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 963 | blk_mq_rq_timed_out(rq, reserved); |
8d699663 YY |
964 | |
965 | if (is_flush_rq(rq, hctx)) | |
966 | rq->end_io(rq, 0); | |
967 | else if (refcount_dec_and_test(&rq->ref)) | |
12f5b931 | 968 | __blk_mq_free_request(rq); |
7baa8572 JA |
969 | |
970 | return true; | |
1d9bd516 TH |
971 | } |
972 | ||
287922eb | 973 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 974 | { |
287922eb CH |
975 | struct request_queue *q = |
976 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 977 | unsigned long next = 0; |
1d9bd516 | 978 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 979 | int i; |
320ae51f | 980 | |
71f79fb3 GKB |
981 | /* A deadlock might occur if a request is stuck requiring a |
982 | * timeout at the same time a queue freeze is waiting | |
983 | * completion, since the timeout code would not be able to | |
984 | * acquire the queue reference here. | |
985 | * | |
986 | * That's why we don't use blk_queue_enter here; instead, we use | |
987 | * percpu_ref_tryget directly, because we need to be able to | |
988 | * obtain a reference even in the short window between the queue | |
989 | * starting to freeze, by dropping the first reference in | |
1671d522 | 990 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
991 | * consumed, marked by the instant q_usage_counter reaches |
992 | * zero. | |
993 | */ | |
994 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
995 | return; |
996 | ||
12f5b931 | 997 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 998 | |
12f5b931 KB |
999 | if (next != 0) { |
1000 | mod_timer(&q->timeout, next); | |
0d2602ca | 1001 | } else { |
fcd36c36 BVA |
1002 | /* |
1003 | * Request timeouts are handled as a forward rolling timer. If | |
1004 | * we end up here it means that no requests are pending and | |
1005 | * also that no request has been pending for a while. Mark | |
1006 | * each hctx as idle. | |
1007 | */ | |
f054b56c ML |
1008 | queue_for_each_hw_ctx(q, hctx, i) { |
1009 | /* the hctx may be unmapped, so check it here */ | |
1010 | if (blk_mq_hw_queue_mapped(hctx)) | |
1011 | blk_mq_tag_idle(hctx); | |
1012 | } | |
0d2602ca | 1013 | } |
287922eb | 1014 | blk_queue_exit(q); |
320ae51f JA |
1015 | } |
1016 | ||
88459642 OS |
1017 | struct flush_busy_ctx_data { |
1018 | struct blk_mq_hw_ctx *hctx; | |
1019 | struct list_head *list; | |
1020 | }; | |
1021 | ||
1022 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1023 | { | |
1024 | struct flush_busy_ctx_data *flush_data = data; | |
1025 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1026 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1027 | enum hctx_type type = hctx->type; |
88459642 | 1028 | |
88459642 | 1029 | spin_lock(&ctx->lock); |
c16d6b5a | 1030 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1031 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1032 | spin_unlock(&ctx->lock); |
1033 | return true; | |
1034 | } | |
1035 | ||
1429d7c9 JA |
1036 | /* |
1037 | * Process software queues that have been marked busy, splicing them | |
1038 | * to the for-dispatch | |
1039 | */ | |
2c3ad667 | 1040 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1041 | { |
88459642 OS |
1042 | struct flush_busy_ctx_data data = { |
1043 | .hctx = hctx, | |
1044 | .list = list, | |
1045 | }; | |
1429d7c9 | 1046 | |
88459642 | 1047 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1048 | } |
2c3ad667 | 1049 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1050 | |
b347689f ML |
1051 | struct dispatch_rq_data { |
1052 | struct blk_mq_hw_ctx *hctx; | |
1053 | struct request *rq; | |
1054 | }; | |
1055 | ||
1056 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1057 | void *data) | |
1058 | { | |
1059 | struct dispatch_rq_data *dispatch_data = data; | |
1060 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1061 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1062 | enum hctx_type type = hctx->type; |
b347689f ML |
1063 | |
1064 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1065 | if (!list_empty(&ctx->rq_lists[type])) { |
1066 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1067 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1068 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1069 | sbitmap_clear_bit(sb, bitnr); |
1070 | } | |
1071 | spin_unlock(&ctx->lock); | |
1072 | ||
1073 | return !dispatch_data->rq; | |
1074 | } | |
1075 | ||
1076 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1077 | struct blk_mq_ctx *start) | |
1078 | { | |
f31967f0 | 1079 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1080 | struct dispatch_rq_data data = { |
1081 | .hctx = hctx, | |
1082 | .rq = NULL, | |
1083 | }; | |
1084 | ||
1085 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1086 | dispatch_rq_from_ctx, &data); | |
1087 | ||
1088 | return data.rq; | |
1089 | } | |
1090 | ||
703fd1c0 JA |
1091 | static inline unsigned int queued_to_index(unsigned int queued) |
1092 | { | |
1093 | if (!queued) | |
1094 | return 0; | |
1429d7c9 | 1095 | |
703fd1c0 | 1096 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1097 | } |
1098 | ||
570e9b73 ML |
1099 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1100 | { | |
222a5ae0 | 1101 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1102 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1103 | int tag; |
1104 | ||
568f2700 ML |
1105 | blk_mq_tag_busy(rq->mq_hctx); |
1106 | ||
570e9b73 | 1107 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
222a5ae0 | 1108 | bt = rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1109 | tag_offset = 0; |
28500850 ML |
1110 | } else { |
1111 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1112 | return false; | |
570e9b73 ML |
1113 | } |
1114 | ||
570e9b73 ML |
1115 | tag = __sbitmap_queue_get(bt); |
1116 | if (tag == BLK_MQ_NO_TAG) | |
1117 | return false; | |
1118 | ||
1119 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1120 | return true; |
1121 | } | |
1122 | ||
1123 | static bool blk_mq_get_driver_tag(struct request *rq) | |
1124 | { | |
568f2700 ML |
1125 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1126 | ||
1127 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1128 | return false; | |
1129 | ||
51db1c37 | 1130 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1131 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1132 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1133 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1134 | } |
1135 | hctx->tags->rqs[rq->tag] = rq; | |
1136 | return true; | |
570e9b73 ML |
1137 | } |
1138 | ||
eb619fdb JA |
1139 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1140 | int flags, void *key) | |
da55f2cc OS |
1141 | { |
1142 | struct blk_mq_hw_ctx *hctx; | |
1143 | ||
1144 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1145 | ||
5815839b | 1146 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1147 | if (!list_empty(&wait->entry)) { |
1148 | struct sbitmap_queue *sbq; | |
1149 | ||
1150 | list_del_init(&wait->entry); | |
222a5ae0 | 1151 | sbq = hctx->tags->bitmap_tags; |
e8618575 JA |
1152 | atomic_dec(&sbq->ws_active); |
1153 | } | |
5815839b ML |
1154 | spin_unlock(&hctx->dispatch_wait_lock); |
1155 | ||
da55f2cc OS |
1156 | blk_mq_run_hw_queue(hctx, true); |
1157 | return 1; | |
1158 | } | |
1159 | ||
f906a6a0 JA |
1160 | /* |
1161 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1162 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1163 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1164 | * marking us as waiting. |
1165 | */ | |
2278d69f | 1166 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1167 | struct request *rq) |
da55f2cc | 1168 | { |
222a5ae0 | 1169 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; |
5815839b | 1170 | struct wait_queue_head *wq; |
f906a6a0 JA |
1171 | wait_queue_entry_t *wait; |
1172 | bool ret; | |
da55f2cc | 1173 | |
51db1c37 | 1174 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1175 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1176 | |
c27d53fb BVA |
1177 | /* |
1178 | * It's possible that a tag was freed in the window between the | |
1179 | * allocation failure and adding the hardware queue to the wait | |
1180 | * queue. | |
1181 | * | |
1182 | * Don't clear RESTART here, someone else could have set it. | |
1183 | * At most this will cost an extra queue run. | |
1184 | */ | |
8ab6bb9e | 1185 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1186 | } |
eb619fdb | 1187 | |
2278d69f | 1188 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1189 | if (!list_empty_careful(&wait->entry)) |
1190 | return false; | |
1191 | ||
e8618575 | 1192 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1193 | |
1194 | spin_lock_irq(&wq->lock); | |
1195 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1196 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1197 | spin_unlock(&hctx->dispatch_wait_lock); |
1198 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1199 | return false; |
eb619fdb JA |
1200 | } |
1201 | ||
e8618575 | 1202 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1203 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1204 | __add_wait_queue(wq, wait); | |
c27d53fb | 1205 | |
da55f2cc | 1206 | /* |
eb619fdb JA |
1207 | * It's possible that a tag was freed in the window between the |
1208 | * allocation failure and adding the hardware queue to the wait | |
1209 | * queue. | |
da55f2cc | 1210 | */ |
8ab6bb9e | 1211 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1212 | if (!ret) { |
5815839b ML |
1213 | spin_unlock(&hctx->dispatch_wait_lock); |
1214 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1215 | return false; |
eb619fdb | 1216 | } |
c27d53fb BVA |
1217 | |
1218 | /* | |
1219 | * We got a tag, remove ourselves from the wait queue to ensure | |
1220 | * someone else gets the wakeup. | |
1221 | */ | |
c27d53fb | 1222 | list_del_init(&wait->entry); |
e8618575 | 1223 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1224 | spin_unlock(&hctx->dispatch_wait_lock); |
1225 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1226 | |
1227 | return true; | |
da55f2cc OS |
1228 | } |
1229 | ||
6e768717 ML |
1230 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1231 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1232 | /* | |
1233 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1234 | * - EWMA is one simple way to compute running average value | |
1235 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1236 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1237 | * factor doesn't matter because EWMA decreases exponentially | |
1238 | */ | |
1239 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1240 | { | |
1241 | unsigned int ewma; | |
1242 | ||
1243 | if (hctx->queue->elevator) | |
1244 | return; | |
1245 | ||
1246 | ewma = hctx->dispatch_busy; | |
1247 | ||
1248 | if (!ewma && !busy) | |
1249 | return; | |
1250 | ||
1251 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1252 | if (busy) | |
1253 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1254 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1255 | ||
1256 | hctx->dispatch_busy = ewma; | |
1257 | } | |
1258 | ||
86ff7c2a ML |
1259 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1260 | ||
c92a4103 JT |
1261 | static void blk_mq_handle_dev_resource(struct request *rq, |
1262 | struct list_head *list) | |
1263 | { | |
1264 | struct request *next = | |
1265 | list_first_entry_or_null(list, struct request, queuelist); | |
1266 | ||
1267 | /* | |
1268 | * If an I/O scheduler has been configured and we got a driver tag for | |
1269 | * the next request already, free it. | |
1270 | */ | |
1271 | if (next) | |
1272 | blk_mq_put_driver_tag(next); | |
1273 | ||
1274 | list_add(&rq->queuelist, list); | |
1275 | __blk_mq_requeue_request(rq); | |
1276 | } | |
1277 | ||
0512a75b KB |
1278 | static void blk_mq_handle_zone_resource(struct request *rq, |
1279 | struct list_head *zone_list) | |
1280 | { | |
1281 | /* | |
1282 | * If we end up here it is because we cannot dispatch a request to a | |
1283 | * specific zone due to LLD level zone-write locking or other zone | |
1284 | * related resource not being available. In this case, set the request | |
1285 | * aside in zone_list for retrying it later. | |
1286 | */ | |
1287 | list_add(&rq->queuelist, zone_list); | |
1288 | __blk_mq_requeue_request(rq); | |
1289 | } | |
1290 | ||
75383524 ML |
1291 | enum prep_dispatch { |
1292 | PREP_DISPATCH_OK, | |
1293 | PREP_DISPATCH_NO_TAG, | |
1294 | PREP_DISPATCH_NO_BUDGET, | |
1295 | }; | |
1296 | ||
1297 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1298 | bool need_budget) | |
1299 | { | |
1300 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1301 | ||
1302 | if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) { | |
1303 | blk_mq_put_driver_tag(rq); | |
1304 | return PREP_DISPATCH_NO_BUDGET; | |
1305 | } | |
1306 | ||
1307 | if (!blk_mq_get_driver_tag(rq)) { | |
1308 | /* | |
1309 | * The initial allocation attempt failed, so we need to | |
1310 | * rerun the hardware queue when a tag is freed. The | |
1311 | * waitqueue takes care of that. If the queue is run | |
1312 | * before we add this entry back on the dispatch list, | |
1313 | * we'll re-run it below. | |
1314 | */ | |
1315 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1316 | /* |
1317 | * All budgets not got from this function will be put | |
1318 | * together during handling partial dispatch | |
1319 | */ | |
1320 | if (need_budget) | |
1321 | blk_mq_put_dispatch_budget(rq->q); | |
75383524 ML |
1322 | return PREP_DISPATCH_NO_TAG; |
1323 | } | |
1324 | } | |
1325 | ||
1326 | return PREP_DISPATCH_OK; | |
1327 | } | |
1328 | ||
1fd40b5e ML |
1329 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1330 | static void blk_mq_release_budgets(struct request_queue *q, | |
1331 | unsigned int nr_budgets) | |
1332 | { | |
1333 | int i; | |
1334 | ||
1335 | for (i = 0; i < nr_budgets; i++) | |
1336 | blk_mq_put_dispatch_budget(q); | |
1337 | } | |
1338 | ||
1f57f8d4 JA |
1339 | /* |
1340 | * Returns true if we did some work AND can potentially do more. | |
1341 | */ | |
445874e8 | 1342 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1343 | unsigned int nr_budgets) |
320ae51f | 1344 | { |
75383524 | 1345 | enum prep_dispatch prep; |
445874e8 | 1346 | struct request_queue *q = hctx->queue; |
6d6f167c | 1347 | struct request *rq, *nxt; |
fc17b653 | 1348 | int errors, queued; |
86ff7c2a | 1349 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1350 | LIST_HEAD(zone_list); |
320ae51f | 1351 | |
81380ca1 OS |
1352 | if (list_empty(list)) |
1353 | return false; | |
1354 | ||
320ae51f JA |
1355 | /* |
1356 | * Now process all the entries, sending them to the driver. | |
1357 | */ | |
93efe981 | 1358 | errors = queued = 0; |
81380ca1 | 1359 | do { |
74c45052 | 1360 | struct blk_mq_queue_data bd; |
320ae51f | 1361 | |
f04c3df3 | 1362 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1363 | |
445874e8 | 1364 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1365 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1366 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1367 | break; |
de148297 | 1368 | |
320ae51f | 1369 | list_del_init(&rq->queuelist); |
320ae51f | 1370 | |
74c45052 | 1371 | bd.rq = rq; |
113285b4 JA |
1372 | |
1373 | /* | |
1374 | * Flag last if we have no more requests, or if we have more | |
1375 | * but can't assign a driver tag to it. | |
1376 | */ | |
1377 | if (list_empty(list)) | |
1378 | bd.last = true; | |
1379 | else { | |
113285b4 | 1380 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1381 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1382 | } |
74c45052 | 1383 | |
1fd40b5e ML |
1384 | /* |
1385 | * once the request is queued to lld, no need to cover the | |
1386 | * budget any more | |
1387 | */ | |
1388 | if (nr_budgets) | |
1389 | nr_budgets--; | |
74c45052 | 1390 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1391 | switch (ret) { |
1392 | case BLK_STS_OK: | |
1393 | queued++; | |
320ae51f | 1394 | break; |
7bf13729 ML |
1395 | case BLK_STS_RESOURCE: |
1396 | case BLK_STS_DEV_RESOURCE: | |
1397 | blk_mq_handle_dev_resource(rq, list); | |
1398 | goto out; | |
1399 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1400 | /* |
1401 | * Move the request to zone_list and keep going through | |
1402 | * the dispatch list to find more requests the drive can | |
1403 | * accept. | |
1404 | */ | |
1405 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1406 | break; |
1407 | default: | |
93efe981 | 1408 | errors++; |
2a842aca | 1409 | blk_mq_end_request(rq, BLK_STS_IOERR); |
320ae51f | 1410 | } |
81380ca1 | 1411 | } while (!list_empty(list)); |
7bf13729 | 1412 | out: |
0512a75b KB |
1413 | if (!list_empty(&zone_list)) |
1414 | list_splice_tail_init(&zone_list, list); | |
1415 | ||
703fd1c0 | 1416 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1417 | |
632bfb63 | 1418 | /* If we didn't flush the entire list, we could have told the driver |
1419 | * there was more coming, but that turned out to be a lie. | |
1420 | */ | |
1421 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1422 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1423 | /* |
1424 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1425 | * that is where we will continue on next queue run. | |
1426 | */ | |
f04c3df3 | 1427 | if (!list_empty(list)) { |
86ff7c2a | 1428 | bool needs_restart; |
75383524 ML |
1429 | /* For non-shared tags, the RESTART check will suffice */ |
1430 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1431 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1432 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1433 | |
1fd40b5e | 1434 | blk_mq_release_budgets(q, nr_budgets); |
86ff7c2a | 1435 | |
320ae51f | 1436 | spin_lock(&hctx->lock); |
01e99aec | 1437 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1438 | spin_unlock(&hctx->lock); |
f04c3df3 | 1439 | |
d7d8535f ML |
1440 | /* |
1441 | * Order adding requests to hctx->dispatch and checking | |
1442 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1443 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1444 | * miss the new added requests to hctx->dispatch, meantime | |
1445 | * SCHED_RESTART is observed here. | |
1446 | */ | |
1447 | smp_mb(); | |
1448 | ||
9ba52e58 | 1449 | /* |
710c785f BVA |
1450 | * If SCHED_RESTART was set by the caller of this function and |
1451 | * it is no longer set that means that it was cleared by another | |
1452 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1453 | * |
eb619fdb JA |
1454 | * If 'no_tag' is set, that means that we failed getting |
1455 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1456 | * waitqueue is no longer active, ensure that we run the queue | |
1457 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1458 | * |
710c785f BVA |
1459 | * If no I/O scheduler has been configured it is possible that |
1460 | * the hardware queue got stopped and restarted before requests | |
1461 | * were pushed back onto the dispatch list. Rerun the queue to | |
1462 | * avoid starvation. Notes: | |
1463 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1464 | * been stopped before rerunning a queue. | |
1465 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1466 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1467 | * and dm-rq. |
86ff7c2a ML |
1468 | * |
1469 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1470 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1471 | * that could otherwise occur if the queue is idle. We'll do |
1472 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1473 | */ |
86ff7c2a ML |
1474 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1475 | if (!needs_restart || | |
eb619fdb | 1476 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1477 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1478 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1479 | no_budget_avail)) | |
86ff7c2a | 1480 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1481 | |
6e768717 | 1482 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1483 | return false; |
6e768717 ML |
1484 | } else |
1485 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1486 | |
93efe981 | 1487 | return (queued + errors) != 0; |
f04c3df3 JA |
1488 | } |
1489 | ||
105663f7 AA |
1490 | /** |
1491 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1492 | * @hctx: Pointer to the hardware queue to run. | |
1493 | * | |
1494 | * Send pending requests to the hardware. | |
1495 | */ | |
6a83e74d BVA |
1496 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1497 | { | |
1498 | int srcu_idx; | |
1499 | ||
b7a71e66 JA |
1500 | /* |
1501 | * We should be running this queue from one of the CPUs that | |
1502 | * are mapped to it. | |
7df938fb ML |
1503 | * |
1504 | * There are at least two related races now between setting | |
1505 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1506 | * __blk_mq_run_hw_queue(): | |
1507 | * | |
1508 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1509 | * but later it becomes online, then this warning is harmless | |
1510 | * at all | |
1511 | * | |
1512 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1513 | * but later it becomes offline, then the warning can't be | |
1514 | * triggered, and we depend on blk-mq timeout handler to | |
1515 | * handle dispatched requests to this hctx | |
b7a71e66 | 1516 | */ |
7df938fb ML |
1517 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1518 | cpu_online(hctx->next_cpu)) { | |
1519 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1520 | raw_smp_processor_id(), | |
1521 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1522 | dump_stack(); | |
1523 | } | |
6a83e74d | 1524 | |
b7a71e66 JA |
1525 | /* |
1526 | * We can't run the queue inline with ints disabled. Ensure that | |
1527 | * we catch bad users of this early. | |
1528 | */ | |
1529 | WARN_ON_ONCE(in_interrupt()); | |
1530 | ||
04ced159 | 1531 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1532 | |
04ced159 JA |
1533 | hctx_lock(hctx, &srcu_idx); |
1534 | blk_mq_sched_dispatch_requests(hctx); | |
1535 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1536 | } |
1537 | ||
f82ddf19 ML |
1538 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1539 | { | |
1540 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1541 | ||
1542 | if (cpu >= nr_cpu_ids) | |
1543 | cpu = cpumask_first(hctx->cpumask); | |
1544 | return cpu; | |
1545 | } | |
1546 | ||
506e931f JA |
1547 | /* |
1548 | * It'd be great if the workqueue API had a way to pass | |
1549 | * in a mask and had some smarts for more clever placement. | |
1550 | * For now we just round-robin here, switching for every | |
1551 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1552 | */ | |
1553 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1554 | { | |
7bed4595 | 1555 | bool tried = false; |
476f8c98 | 1556 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1557 | |
b657d7e6 CH |
1558 | if (hctx->queue->nr_hw_queues == 1) |
1559 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1560 | |
1561 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1562 | select_cpu: |
476f8c98 | 1563 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1564 | cpu_online_mask); |
506e931f | 1565 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1566 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1567 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1568 | } | |
1569 | ||
7bed4595 ML |
1570 | /* |
1571 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1572 | * and it should only happen in the path of handling CPU DEAD. | |
1573 | */ | |
476f8c98 | 1574 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1575 | if (!tried) { |
1576 | tried = true; | |
1577 | goto select_cpu; | |
1578 | } | |
1579 | ||
1580 | /* | |
1581 | * Make sure to re-select CPU next time once after CPUs | |
1582 | * in hctx->cpumask become online again. | |
1583 | */ | |
476f8c98 | 1584 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1585 | hctx->next_cpu_batch = 1; |
1586 | return WORK_CPU_UNBOUND; | |
1587 | } | |
476f8c98 ML |
1588 | |
1589 | hctx->next_cpu = next_cpu; | |
1590 | return next_cpu; | |
506e931f JA |
1591 | } |
1592 | ||
105663f7 AA |
1593 | /** |
1594 | * __blk_mq_delay_run_hw_queue - Run (or schedule 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 | * @msecs: Microseconds of delay to wait before running the queue. | |
1598 | * | |
1599 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1600 | * with a delay of @msecs. | |
1601 | */ | |
7587a5ae BVA |
1602 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1603 | unsigned long msecs) | |
320ae51f | 1604 | { |
5435c023 | 1605 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1606 | return; |
1607 | ||
1b792f2f | 1608 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1609 | int cpu = get_cpu(); |
1610 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1611 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1612 | put_cpu(); |
398205b8 PB |
1613 | return; |
1614 | } | |
e4043dcf | 1615 | |
2a90d4aa | 1616 | put_cpu(); |
e4043dcf | 1617 | } |
398205b8 | 1618 | |
ae943d20 BVA |
1619 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1620 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1621 | } |
1622 | ||
105663f7 AA |
1623 | /** |
1624 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1625 | * @hctx: Pointer to the hardware queue to run. | |
1626 | * @msecs: Microseconds of delay to wait before running the queue. | |
1627 | * | |
1628 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1629 | */ | |
7587a5ae BVA |
1630 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1631 | { | |
1632 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1633 | } | |
1634 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1635 | ||
105663f7 AA |
1636 | /** |
1637 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1638 | * @hctx: Pointer to the hardware queue to run. | |
1639 | * @async: If we want to run the queue asynchronously. | |
1640 | * | |
1641 | * Check if the request queue is not in a quiesced state and if there are | |
1642 | * pending requests to be sent. If this is true, run the queue to send requests | |
1643 | * to hardware. | |
1644 | */ | |
626fb735 | 1645 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1646 | { |
24f5a90f ML |
1647 | int srcu_idx; |
1648 | bool need_run; | |
1649 | ||
1650 | /* | |
1651 | * When queue is quiesced, we may be switching io scheduler, or | |
1652 | * updating nr_hw_queues, or other things, and we can't run queue | |
1653 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1654 | * | |
1655 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1656 | * quiesced. | |
1657 | */ | |
04ced159 JA |
1658 | hctx_lock(hctx, &srcu_idx); |
1659 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1660 | blk_mq_hctx_has_pending(hctx); | |
1661 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1662 | |
626fb735 | 1663 | if (need_run) |
79f720a7 | 1664 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1665 | } |
5b727272 | 1666 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1667 | |
105663f7 | 1668 | /** |
24f7bb88 | 1669 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1670 | * @q: Pointer to the request queue to run. |
1671 | * @async: If we want to run the queue asynchronously. | |
1672 | */ | |
b94ec296 | 1673 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1674 | { |
1675 | struct blk_mq_hw_ctx *hctx; | |
1676 | int i; | |
1677 | ||
1678 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1679 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1680 | continue; |
1681 | ||
b94ec296 | 1682 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1683 | } |
1684 | } | |
b94ec296 | 1685 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1686 | |
b9151e7b DA |
1687 | /** |
1688 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1689 | * @q: Pointer to the request queue to run. | |
1690 | * @msecs: Microseconds of delay to wait before running the queues. | |
1691 | */ | |
1692 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1693 | { | |
1694 | struct blk_mq_hw_ctx *hctx; | |
1695 | int i; | |
1696 | ||
1697 | queue_for_each_hw_ctx(q, hctx, i) { | |
1698 | if (blk_mq_hctx_stopped(hctx)) | |
1699 | continue; | |
1700 | ||
1701 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
1702 | } | |
1703 | } | |
1704 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1705 | ||
fd001443 BVA |
1706 | /** |
1707 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1708 | * @q: request queue. | |
1709 | * | |
1710 | * The caller is responsible for serializing this function against | |
1711 | * blk_mq_{start,stop}_hw_queue(). | |
1712 | */ | |
1713 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1714 | { | |
1715 | struct blk_mq_hw_ctx *hctx; | |
1716 | int i; | |
1717 | ||
1718 | queue_for_each_hw_ctx(q, hctx, i) | |
1719 | if (blk_mq_hctx_stopped(hctx)) | |
1720 | return true; | |
1721 | ||
1722 | return false; | |
1723 | } | |
1724 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1725 | ||
39a70c76 ML |
1726 | /* |
1727 | * This function is often used for pausing .queue_rq() by driver when | |
1728 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1729 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1730 | * |
1731 | * We do not guarantee that dispatch can be drained or blocked | |
1732 | * after blk_mq_stop_hw_queue() returns. Please use | |
1733 | * blk_mq_quiesce_queue() for that requirement. | |
1734 | */ | |
2719aa21 JA |
1735 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1736 | { | |
641a9ed6 | 1737 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1738 | |
641a9ed6 | 1739 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1740 | } |
641a9ed6 | 1741 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1742 | |
39a70c76 ML |
1743 | /* |
1744 | * This function is often used for pausing .queue_rq() by driver when | |
1745 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1746 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1747 | * |
1748 | * We do not guarantee that dispatch can be drained or blocked | |
1749 | * after blk_mq_stop_hw_queues() returns. Please use | |
1750 | * blk_mq_quiesce_queue() for that requirement. | |
1751 | */ | |
2719aa21 JA |
1752 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1753 | { | |
641a9ed6 ML |
1754 | struct blk_mq_hw_ctx *hctx; |
1755 | int i; | |
1756 | ||
1757 | queue_for_each_hw_ctx(q, hctx, i) | |
1758 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1759 | } |
1760 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1761 | ||
320ae51f JA |
1762 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1763 | { | |
1764 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1765 | |
0ffbce80 | 1766 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1767 | } |
1768 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1769 | ||
2f268556 CH |
1770 | void blk_mq_start_hw_queues(struct request_queue *q) |
1771 | { | |
1772 | struct blk_mq_hw_ctx *hctx; | |
1773 | int i; | |
1774 | ||
1775 | queue_for_each_hw_ctx(q, hctx, i) | |
1776 | blk_mq_start_hw_queue(hctx); | |
1777 | } | |
1778 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1779 | ||
ae911c5e JA |
1780 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1781 | { | |
1782 | if (!blk_mq_hctx_stopped(hctx)) | |
1783 | return; | |
1784 | ||
1785 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1786 | blk_mq_run_hw_queue(hctx, async); | |
1787 | } | |
1788 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1789 | ||
1b4a3258 | 1790 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1791 | { |
1792 | struct blk_mq_hw_ctx *hctx; | |
1793 | int i; | |
1794 | ||
ae911c5e JA |
1795 | queue_for_each_hw_ctx(q, hctx, i) |
1796 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1797 | } |
1798 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1799 | ||
70f4db63 | 1800 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1801 | { |
1802 | struct blk_mq_hw_ctx *hctx; | |
1803 | ||
9f993737 | 1804 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1805 | |
21c6e939 | 1806 | /* |
15fe8a90 | 1807 | * If we are stopped, don't run the queue. |
21c6e939 | 1808 | */ |
0841031a | 1809 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 1810 | return; |
7587a5ae BVA |
1811 | |
1812 | __blk_mq_run_hw_queue(hctx); | |
1813 | } | |
1814 | ||
cfd0c552 | 1815 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1816 | struct request *rq, |
1817 | bool at_head) | |
320ae51f | 1818 | { |
e57690fe | 1819 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1820 | enum hctx_type type = hctx->type; |
e57690fe | 1821 | |
7b607814 BVA |
1822 | lockdep_assert_held(&ctx->lock); |
1823 | ||
01b983c9 JA |
1824 | trace_block_rq_insert(hctx->queue, rq); |
1825 | ||
72a0a36e | 1826 | if (at_head) |
c16d6b5a | 1827 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1828 | else |
c16d6b5a | 1829 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1830 | } |
4bb659b1 | 1831 | |
2c3ad667 JA |
1832 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1833 | bool at_head) | |
cfd0c552 ML |
1834 | { |
1835 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1836 | ||
7b607814 BVA |
1837 | lockdep_assert_held(&ctx->lock); |
1838 | ||
e57690fe | 1839 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1840 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1841 | } |
1842 | ||
105663f7 AA |
1843 | /** |
1844 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1845 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 1846 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
1847 | * @run_queue: If we should run the hardware queue after inserting the request. |
1848 | * | |
157f377b JA |
1849 | * Should only be used carefully, when the caller knows we want to |
1850 | * bypass a potential IO scheduler on the target device. | |
1851 | */ | |
01e99aec ML |
1852 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1853 | bool run_queue) | |
157f377b | 1854 | { |
ea4f995e | 1855 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1856 | |
1857 | spin_lock(&hctx->lock); | |
01e99aec ML |
1858 | if (at_head) |
1859 | list_add(&rq->queuelist, &hctx->dispatch); | |
1860 | else | |
1861 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1862 | spin_unlock(&hctx->lock); |
1863 | ||
b0850297 ML |
1864 | if (run_queue) |
1865 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1866 | } |
1867 | ||
bd166ef1 JA |
1868 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1869 | struct list_head *list) | |
320ae51f JA |
1870 | |
1871 | { | |
3f0cedc7 | 1872 | struct request *rq; |
c16d6b5a | 1873 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1874 | |
320ae51f JA |
1875 | /* |
1876 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1877 | * offline now | |
1878 | */ | |
3f0cedc7 | 1879 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1880 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1881 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1882 | } |
3f0cedc7 ML |
1883 | |
1884 | spin_lock(&ctx->lock); | |
c16d6b5a | 1885 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1886 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1887 | spin_unlock(&ctx->lock); |
320ae51f JA |
1888 | } |
1889 | ||
3110fc79 | 1890 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1891 | { |
1892 | struct request *rqa = container_of(a, struct request, queuelist); | |
1893 | struct request *rqb = container_of(b, struct request, queuelist); | |
1894 | ||
7d30a621 PB |
1895 | if (rqa->mq_ctx != rqb->mq_ctx) |
1896 | return rqa->mq_ctx > rqb->mq_ctx; | |
1897 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1898 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1899 | |
1900 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1901 | } |
1902 | ||
1903 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1904 | { | |
320ae51f | 1905 | LIST_HEAD(list); |
320ae51f | 1906 | |
95ed0c5b PB |
1907 | if (list_empty(&plug->mq_list)) |
1908 | return; | |
320ae51f JA |
1909 | list_splice_init(&plug->mq_list, &list); |
1910 | ||
ce5b009c JA |
1911 | if (plug->rq_count > 2 && plug->multiple_queues) |
1912 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1913 | |
bcc816df DZ |
1914 | plug->rq_count = 0; |
1915 | ||
95ed0c5b PB |
1916 | do { |
1917 | struct list_head rq_list; | |
1918 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1919 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1920 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1921 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1922 | unsigned int depth = 1; | |
1923 | ||
1924 | list_for_each_continue(pos, &list) { | |
1925 | rq = list_entry_rq(pos); | |
1926 | BUG_ON(!rq->q); | |
1927 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1928 | break; | |
1929 | depth++; | |
320ae51f JA |
1930 | } |
1931 | ||
95ed0c5b PB |
1932 | list_cut_before(&rq_list, &list, pos); |
1933 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1934 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1935 | from_schedule); |
95ed0c5b | 1936 | } while(!list_empty(&list)); |
320ae51f JA |
1937 | } |
1938 | ||
14ccb66b CH |
1939 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1940 | unsigned int nr_segs) | |
320ae51f | 1941 | { |
93f221ae EB |
1942 | int err; |
1943 | ||
f924cdde CH |
1944 | if (bio->bi_opf & REQ_RAHEAD) |
1945 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1946 | ||
1947 | rq->__sector = bio->bi_iter.bi_sector; | |
1948 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1949 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1950 | |
1951 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
1952 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1953 | WARN_ON_ONCE(err); | |
4b570521 | 1954 | |
b5af37ab | 1955 | blk_account_io_start(rq); |
320ae51f JA |
1956 | } |
1957 | ||
0f95549c MS |
1958 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1959 | struct request *rq, | |
be94f058 | 1960 | blk_qc_t *cookie, bool last) |
f984df1f | 1961 | { |
f984df1f | 1962 | struct request_queue *q = rq->q; |
f984df1f SL |
1963 | struct blk_mq_queue_data bd = { |
1964 | .rq = rq, | |
be94f058 | 1965 | .last = last, |
f984df1f | 1966 | }; |
bd166ef1 | 1967 | blk_qc_t new_cookie; |
f06345ad | 1968 | blk_status_t ret; |
0f95549c MS |
1969 | |
1970 | new_cookie = request_to_qc_t(hctx, rq); | |
1971 | ||
1972 | /* | |
1973 | * For OK queue, we are done. For error, caller may kill it. | |
1974 | * Any other error (busy), just add it to our list as we | |
1975 | * previously would have done. | |
1976 | */ | |
1977 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1978 | switch (ret) { | |
1979 | case BLK_STS_OK: | |
6ce3dd6e | 1980 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1981 | *cookie = new_cookie; |
1982 | break; | |
1983 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1984 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1985 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1986 | __blk_mq_requeue_request(rq); |
1987 | break; | |
1988 | default: | |
6ce3dd6e | 1989 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1990 | *cookie = BLK_QC_T_NONE; |
1991 | break; | |
1992 | } | |
1993 | ||
1994 | return ret; | |
1995 | } | |
1996 | ||
fd9c40f6 | 1997 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1998 | struct request *rq, |
396eaf21 | 1999 | blk_qc_t *cookie, |
fd9c40f6 | 2000 | bool bypass_insert, bool last) |
0f95549c MS |
2001 | { |
2002 | struct request_queue *q = rq->q; | |
d964f04a ML |
2003 | bool run_queue = true; |
2004 | ||
23d4ee19 | 2005 | /* |
fd9c40f6 | 2006 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2007 | * |
fd9c40f6 BVA |
2008 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2009 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2010 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2011 | */ |
fd9c40f6 | 2012 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2013 | run_queue = false; |
fd9c40f6 BVA |
2014 | bypass_insert = false; |
2015 | goto insert; | |
d964f04a | 2016 | } |
f984df1f | 2017 | |
fd9c40f6 BVA |
2018 | if (q->elevator && !bypass_insert) |
2019 | goto insert; | |
2253efc8 | 2020 | |
65c76369 | 2021 | if (!blk_mq_get_dispatch_budget(q)) |
fd9c40f6 | 2022 | goto insert; |
bd166ef1 | 2023 | |
8ab6bb9e | 2024 | if (!blk_mq_get_driver_tag(rq)) { |
65c76369 | 2025 | blk_mq_put_dispatch_budget(q); |
fd9c40f6 | 2026 | goto insert; |
88022d72 | 2027 | } |
de148297 | 2028 | |
fd9c40f6 BVA |
2029 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2030 | insert: | |
2031 | if (bypass_insert) | |
2032 | return BLK_STS_RESOURCE; | |
2033 | ||
db03f88f ML |
2034 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2035 | ||
fd9c40f6 BVA |
2036 | return BLK_STS_OK; |
2037 | } | |
2038 | ||
105663f7 AA |
2039 | /** |
2040 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2041 | * @hctx: Pointer of the associated hardware queue. | |
2042 | * @rq: Pointer to request to be sent. | |
2043 | * @cookie: Request queue cookie. | |
2044 | * | |
2045 | * If the device has enough resources to accept a new request now, send the | |
2046 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2047 | * we can try send it another time in the future. Requests inserted at this | |
2048 | * queue have higher priority. | |
2049 | */ | |
fd9c40f6 BVA |
2050 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2051 | struct request *rq, blk_qc_t *cookie) | |
2052 | { | |
2053 | blk_status_t ret; | |
2054 | int srcu_idx; | |
2055 | ||
2056 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2057 | ||
2058 | hctx_lock(hctx, &srcu_idx); | |
2059 | ||
2060 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2061 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2062 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2063 | else if (ret != BLK_STS_OK) |
2064 | blk_mq_end_request(rq, ret); | |
2065 | ||
2066 | hctx_unlock(hctx, srcu_idx); | |
2067 | } | |
2068 | ||
2069 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2070 | { | |
2071 | blk_status_t ret; | |
2072 | int srcu_idx; | |
2073 | blk_qc_t unused_cookie; | |
2074 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2075 | ||
2076 | hctx_lock(hctx, &srcu_idx); | |
2077 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2078 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2079 | |
2080 | return ret; | |
5eb6126e CH |
2081 | } |
2082 | ||
6ce3dd6e ML |
2083 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2084 | struct list_head *list) | |
2085 | { | |
536167d4 | 2086 | int queued = 0; |
632bfb63 | 2087 | int errors = 0; |
536167d4 | 2088 | |
6ce3dd6e | 2089 | while (!list_empty(list)) { |
fd9c40f6 | 2090 | blk_status_t ret; |
6ce3dd6e ML |
2091 | struct request *rq = list_first_entry(list, struct request, |
2092 | queuelist); | |
2093 | ||
2094 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2095 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2096 | if (ret != BLK_STS_OK) { | |
2097 | if (ret == BLK_STS_RESOURCE || | |
2098 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2099 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2100 | list_empty(list)); |
fd9c40f6 BVA |
2101 | break; |
2102 | } | |
2103 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2104 | errors++; |
536167d4 KB |
2105 | } else |
2106 | queued++; | |
6ce3dd6e | 2107 | } |
d666ba98 JA |
2108 | |
2109 | /* | |
2110 | * If we didn't flush the entire list, we could have told | |
2111 | * the driver there was more coming, but that turned out to | |
2112 | * be a lie. | |
2113 | */ | |
632bfb63 | 2114 | if ((!list_empty(list) || errors) && |
2115 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2116 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2117 | } |
2118 | ||
ce5b009c JA |
2119 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2120 | { | |
2121 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2122 | plug->rq_count++; | |
2123 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2124 | struct request *tmp; | |
2125 | ||
2126 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2127 | queuelist); | |
2128 | if (tmp->q != rq->q) | |
2129 | plug->multiple_queues = true; | |
2130 | } | |
2131 | } | |
2132 | ||
105663f7 | 2133 | /** |
c62b37d9 | 2134 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2135 | * @bio: Bio pointer. |
2136 | * | |
2137 | * Builds up a request structure from @q and @bio and send to the device. The | |
2138 | * request may not be queued directly to hardware if: | |
2139 | * * This request can be merged with another one | |
2140 | * * We want to place request at plug queue for possible future merging | |
2141 | * * There is an IO scheduler active at this queue | |
2142 | * | |
2143 | * It will not queue the request if there is an error with the bio, or at the | |
2144 | * request creation. | |
2145 | * | |
2146 | * Returns: Request queue cookie. | |
2147 | */ | |
c62b37d9 | 2148 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2149 | { |
c62b37d9 | 2150 | struct request_queue *q = bio->bi_disk->queue; |
ef295ecf | 2151 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2152 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6e7abff CH |
2153 | struct blk_mq_alloc_data data = { |
2154 | .q = q, | |
2155 | }; | |
07068d5b | 2156 | struct request *rq; |
f984df1f | 2157 | struct blk_plug *plug; |
5b3f341f | 2158 | struct request *same_queue_rq = NULL; |
14ccb66b | 2159 | unsigned int nr_segs; |
7b371636 | 2160 | blk_qc_t cookie; |
a892c8d5 | 2161 | blk_status_t ret; |
07068d5b JA |
2162 | |
2163 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2164 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2165 | |
e23947bd | 2166 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2167 | goto queue_exit; |
07068d5b | 2168 | |
87c279e6 | 2169 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2170 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2171 | goto queue_exit; |
f984df1f | 2172 | |
14ccb66b | 2173 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2174 | goto queue_exit; |
bd166ef1 | 2175 | |
d5337560 | 2176 | rq_qos_throttle(q, bio); |
87760e5e | 2177 | |
7809167d | 2178 | data.cmd_flags = bio->bi_opf; |
e6e7abff | 2179 | rq = __blk_mq_alloc_request(&data); |
87760e5e | 2180 | if (unlikely(!rq)) { |
c1c80384 | 2181 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2182 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2183 | bio_wouldblock_error(bio); |
ac7c5675 | 2184 | goto queue_exit; |
87760e5e JA |
2185 | } |
2186 | ||
d6f1dda2 XW |
2187 | trace_block_getrq(q, bio, bio->bi_opf); |
2188 | ||
c1c80384 | 2189 | rq_qos_track(q, rq, bio); |
07068d5b | 2190 | |
fd2d3326 | 2191 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2192 | |
970d168d BVA |
2193 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2194 | ||
a892c8d5 ST |
2195 | ret = blk_crypto_init_request(rq); |
2196 | if (ret != BLK_STS_OK) { | |
2197 | bio->bi_status = ret; | |
2198 | bio_endio(bio); | |
2199 | blk_mq_free_request(rq); | |
2200 | return BLK_QC_T_NONE; | |
2201 | } | |
2202 | ||
b49773e7 | 2203 | plug = blk_mq_plug(q, bio); |
07068d5b | 2204 | if (unlikely(is_flush_fua)) { |
105663f7 | 2205 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2206 | blk_insert_flush(rq); |
2207 | blk_mq_run_hw_queue(data.hctx, true); | |
3154df26 ML |
2208 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs || |
2209 | !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2210 | /* |
2211 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2212 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2213 | * |
2214 | * Use normal plugging if this disk is slow HDD, as sequential | |
2215 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2216 | */ |
5f0ed774 | 2217 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2218 | struct request *last = NULL; |
2219 | ||
676d0607 | 2220 | if (!request_count) |
e6c4438b | 2221 | trace_block_plug(q); |
600271d9 SL |
2222 | else |
2223 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2224 | |
600271d9 SL |
2225 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
2226 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
2227 | blk_flush_plug_list(plug, false); |
2228 | trace_block_plug(q); | |
320ae51f | 2229 | } |
b094f89c | 2230 | |
ce5b009c | 2231 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2232 | } else if (q->elevator) { |
105663f7 | 2233 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2234 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2235 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2236 | /* |
6a83e74d | 2237 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2238 | * Otherwise the existing request in the plug list will be |
2239 | * issued. So the plug list will have one request at most | |
2299722c CH |
2240 | * The plug list might get flushed before this. If that happens, |
2241 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2242 | */ |
2299722c CH |
2243 | if (list_empty(&plug->mq_list)) |
2244 | same_queue_rq = NULL; | |
4711b573 | 2245 | if (same_queue_rq) { |
2299722c | 2246 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2247 | plug->rq_count--; |
2248 | } | |
ce5b009c | 2249 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2250 | trace_block_plug(q); |
2299722c | 2251 | |
dad7a3be | 2252 | if (same_queue_rq) { |
ea4f995e | 2253 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2254 | trace_block_unplug(q, 1, true); |
2299722c | 2255 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2256 | &cookie); |
dad7a3be | 2257 | } |
a12de1d4 ML |
2258 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2259 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2260 | /* |
2261 | * There is no scheduler and we can try to send directly | |
2262 | * to the hardware. | |
2263 | */ | |
fd9c40f6 | 2264 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2265 | } else { |
105663f7 | 2266 | /* Default case. */ |
8fa9f556 | 2267 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2268 | } |
320ae51f | 2269 | |
7b371636 | 2270 | return cookie; |
ac7c5675 CH |
2271 | queue_exit: |
2272 | blk_queue_exit(q); | |
2273 | return BLK_QC_T_NONE; | |
320ae51f JA |
2274 | } |
2275 | ||
cc71a6f4 JA |
2276 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2277 | unsigned int hctx_idx) | |
95363efd | 2278 | { |
e9b267d9 | 2279 | struct page *page; |
320ae51f | 2280 | |
24d2f903 | 2281 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2282 | int i; |
320ae51f | 2283 | |
24d2f903 | 2284 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2285 | struct request *rq = tags->static_rqs[i]; |
2286 | ||
2287 | if (!rq) | |
e9b267d9 | 2288 | continue; |
d6296d39 | 2289 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2290 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2291 | } |
320ae51f | 2292 | } |
320ae51f | 2293 | |
24d2f903 CH |
2294 | while (!list_empty(&tags->page_list)) { |
2295 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2296 | list_del_init(&page->lru); |
f75782e4 CM |
2297 | /* |
2298 | * Remove kmemleak object previously allocated in | |
273938bf | 2299 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2300 | */ |
2301 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2302 | __free_pages(page, page->private); |
2303 | } | |
cc71a6f4 | 2304 | } |
320ae51f | 2305 | |
1c0706a7 | 2306 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) |
cc71a6f4 | 2307 | { |
24d2f903 | 2308 | kfree(tags->rqs); |
cc71a6f4 | 2309 | tags->rqs = NULL; |
2af8cbe3 JA |
2310 | kfree(tags->static_rqs); |
2311 | tags->static_rqs = NULL; | |
320ae51f | 2312 | |
1c0706a7 | 2313 | blk_mq_free_tags(tags, flags); |
320ae51f JA |
2314 | } |
2315 | ||
cc71a6f4 JA |
2316 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2317 | unsigned int hctx_idx, | |
2318 | unsigned int nr_tags, | |
1c0706a7 JG |
2319 | unsigned int reserved_tags, |
2320 | unsigned int flags) | |
320ae51f | 2321 | { |
24d2f903 | 2322 | struct blk_mq_tags *tags; |
59f082e4 | 2323 | int node; |
320ae51f | 2324 | |
7d76f856 | 2325 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2326 | if (node == NUMA_NO_NODE) |
2327 | node = set->numa_node; | |
2328 | ||
1c0706a7 | 2329 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); |
24d2f903 CH |
2330 | if (!tags) |
2331 | return NULL; | |
320ae51f | 2332 | |
590b5b7d | 2333 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2334 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2335 | node); |
24d2f903 | 2336 | if (!tags->rqs) { |
1c0706a7 | 2337 | blk_mq_free_tags(tags, flags); |
24d2f903 CH |
2338 | return NULL; |
2339 | } | |
320ae51f | 2340 | |
590b5b7d KC |
2341 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2342 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2343 | node); | |
2af8cbe3 JA |
2344 | if (!tags->static_rqs) { |
2345 | kfree(tags->rqs); | |
1c0706a7 | 2346 | blk_mq_free_tags(tags, flags); |
2af8cbe3 JA |
2347 | return NULL; |
2348 | } | |
2349 | ||
cc71a6f4 JA |
2350 | return tags; |
2351 | } | |
2352 | ||
2353 | static size_t order_to_size(unsigned int order) | |
2354 | { | |
2355 | return (size_t)PAGE_SIZE << order; | |
2356 | } | |
2357 | ||
1d9bd516 TH |
2358 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2359 | unsigned int hctx_idx, int node) | |
2360 | { | |
2361 | int ret; | |
2362 | ||
2363 | if (set->ops->init_request) { | |
2364 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2365 | if (ret) | |
2366 | return ret; | |
2367 | } | |
2368 | ||
12f5b931 | 2369 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2370 | return 0; |
2371 | } | |
2372 | ||
cc71a6f4 JA |
2373 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2374 | unsigned int hctx_idx, unsigned int depth) | |
2375 | { | |
2376 | unsigned int i, j, entries_per_page, max_order = 4; | |
2377 | size_t rq_size, left; | |
59f082e4 SL |
2378 | int node; |
2379 | ||
7d76f856 | 2380 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2381 | if (node == NUMA_NO_NODE) |
2382 | node = set->numa_node; | |
cc71a6f4 JA |
2383 | |
2384 | INIT_LIST_HEAD(&tags->page_list); | |
2385 | ||
320ae51f JA |
2386 | /* |
2387 | * rq_size is the size of the request plus driver payload, rounded | |
2388 | * to the cacheline size | |
2389 | */ | |
24d2f903 | 2390 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2391 | cache_line_size()); |
cc71a6f4 | 2392 | left = rq_size * depth; |
320ae51f | 2393 | |
cc71a6f4 | 2394 | for (i = 0; i < depth; ) { |
320ae51f JA |
2395 | int this_order = max_order; |
2396 | struct page *page; | |
2397 | int to_do; | |
2398 | void *p; | |
2399 | ||
b3a834b1 | 2400 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2401 | this_order--; |
2402 | ||
2403 | do { | |
59f082e4 | 2404 | page = alloc_pages_node(node, |
36e1f3d1 | 2405 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2406 | this_order); |
320ae51f JA |
2407 | if (page) |
2408 | break; | |
2409 | if (!this_order--) | |
2410 | break; | |
2411 | if (order_to_size(this_order) < rq_size) | |
2412 | break; | |
2413 | } while (1); | |
2414 | ||
2415 | if (!page) | |
24d2f903 | 2416 | goto fail; |
320ae51f JA |
2417 | |
2418 | page->private = this_order; | |
24d2f903 | 2419 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2420 | |
2421 | p = page_address(page); | |
f75782e4 CM |
2422 | /* |
2423 | * Allow kmemleak to scan these pages as they contain pointers | |
2424 | * to additional allocations like via ops->init_request(). | |
2425 | */ | |
36e1f3d1 | 2426 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2427 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2428 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2429 | left -= to_do * rq_size; |
2430 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2431 | struct request *rq = p; |
2432 | ||
2433 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2434 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2435 | tags->static_rqs[i] = NULL; | |
2436 | goto fail; | |
e9b267d9 CH |
2437 | } |
2438 | ||
320ae51f JA |
2439 | p += rq_size; |
2440 | i++; | |
2441 | } | |
2442 | } | |
cc71a6f4 | 2443 | return 0; |
320ae51f | 2444 | |
24d2f903 | 2445 | fail: |
cc71a6f4 JA |
2446 | blk_mq_free_rqs(set, tags, hctx_idx); |
2447 | return -ENOMEM; | |
320ae51f JA |
2448 | } |
2449 | ||
bf0beec0 ML |
2450 | struct rq_iter_data { |
2451 | struct blk_mq_hw_ctx *hctx; | |
2452 | bool has_rq; | |
2453 | }; | |
2454 | ||
2455 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2456 | { | |
2457 | struct rq_iter_data *iter_data = data; | |
2458 | ||
2459 | if (rq->mq_hctx != iter_data->hctx) | |
2460 | return true; | |
2461 | iter_data->has_rq = true; | |
2462 | return false; | |
2463 | } | |
2464 | ||
2465 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2466 | { | |
2467 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2468 | hctx->sched_tags : hctx->tags; | |
2469 | struct rq_iter_data data = { | |
2470 | .hctx = hctx, | |
2471 | }; | |
2472 | ||
2473 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2474 | return data.has_rq; | |
2475 | } | |
2476 | ||
2477 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2478 | struct blk_mq_hw_ctx *hctx) | |
2479 | { | |
2480 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2481 | return false; | |
2482 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2483 | return false; | |
2484 | return true; | |
2485 | } | |
2486 | ||
2487 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2488 | { | |
2489 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2490 | struct blk_mq_hw_ctx, cpuhp_online); | |
2491 | ||
2492 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2493 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2494 | return 0; | |
2495 | ||
2496 | /* | |
2497 | * Prevent new request from being allocated on the current hctx. | |
2498 | * | |
2499 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2500 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2501 | * seen once we return from the tag allocator. | |
2502 | */ | |
2503 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2504 | smp_mb__after_atomic(); | |
2505 | ||
2506 | /* | |
2507 | * Try to grab a reference to the queue and wait for any outstanding | |
2508 | * requests. If we could not grab a reference the queue has been | |
2509 | * frozen and there are no requests. | |
2510 | */ | |
2511 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2512 | while (blk_mq_hctx_has_requests(hctx)) | |
2513 | msleep(5); | |
2514 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2515 | } | |
2516 | ||
2517 | return 0; | |
2518 | } | |
2519 | ||
2520 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2521 | { | |
2522 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2523 | struct blk_mq_hw_ctx, cpuhp_online); | |
2524 | ||
2525 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2526 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2527 | return 0; | |
2528 | } | |
2529 | ||
e57690fe JA |
2530 | /* |
2531 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2532 | * software queue to the hw queue dispatch list, and ensure that it | |
2533 | * gets run. | |
2534 | */ | |
9467f859 | 2535 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2536 | { |
9467f859 | 2537 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2538 | struct blk_mq_ctx *ctx; |
2539 | LIST_HEAD(tmp); | |
c16d6b5a | 2540 | enum hctx_type type; |
484b4061 | 2541 | |
9467f859 | 2542 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2543 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2544 | return 0; | |
2545 | ||
e57690fe | 2546 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2547 | type = hctx->type; |
484b4061 JA |
2548 | |
2549 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2550 | if (!list_empty(&ctx->rq_lists[type])) { |
2551 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2552 | blk_mq_hctx_clear_pending(hctx, ctx); |
2553 | } | |
2554 | spin_unlock(&ctx->lock); | |
2555 | ||
2556 | if (list_empty(&tmp)) | |
9467f859 | 2557 | return 0; |
484b4061 | 2558 | |
e57690fe JA |
2559 | spin_lock(&hctx->lock); |
2560 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2561 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2562 | |
2563 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2564 | return 0; |
484b4061 JA |
2565 | } |
2566 | ||
9467f859 | 2567 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2568 | { |
bf0beec0 ML |
2569 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2570 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2571 | &hctx->cpuhp_online); | |
9467f859 TG |
2572 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2573 | &hctx->cpuhp_dead); | |
484b4061 JA |
2574 | } |
2575 | ||
c3b4afca | 2576 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2577 | static void blk_mq_exit_hctx(struct request_queue *q, |
2578 | struct blk_mq_tag_set *set, | |
2579 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2580 | { | |
8ab0b7dc ML |
2581 | if (blk_mq_hw_queue_mapped(hctx)) |
2582 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2583 | |
f70ced09 | 2584 | if (set->ops->exit_request) |
d6296d39 | 2585 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2586 | |
08e98fc6 ML |
2587 | if (set->ops->exit_hctx) |
2588 | set->ops->exit_hctx(hctx, hctx_idx); | |
2589 | ||
9467f859 | 2590 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2591 | |
2592 | spin_lock(&q->unused_hctx_lock); | |
2593 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2594 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2595 | } |
2596 | ||
624dbe47 ML |
2597 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2598 | struct blk_mq_tag_set *set, int nr_queue) | |
2599 | { | |
2600 | struct blk_mq_hw_ctx *hctx; | |
2601 | unsigned int i; | |
2602 | ||
2603 | queue_for_each_hw_ctx(q, hctx, i) { | |
2604 | if (i == nr_queue) | |
2605 | break; | |
477e19de | 2606 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2607 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2608 | } |
624dbe47 ML |
2609 | } |
2610 | ||
7c6c5b7c ML |
2611 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2612 | { | |
2613 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2614 | ||
2615 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2616 | __alignof__(struct blk_mq_hw_ctx)) != | |
2617 | sizeof(struct blk_mq_hw_ctx)); | |
2618 | ||
2619 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2620 | hw_ctx_size += sizeof(struct srcu_struct); | |
2621 | ||
2622 | return hw_ctx_size; | |
2623 | } | |
2624 | ||
08e98fc6 ML |
2625 | static int blk_mq_init_hctx(struct request_queue *q, |
2626 | struct blk_mq_tag_set *set, | |
2627 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2628 | { |
7c6c5b7c ML |
2629 | hctx->queue_num = hctx_idx; |
2630 | ||
bf0beec0 ML |
2631 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2632 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2633 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2634 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2635 | ||
2636 | hctx->tags = set->tags[hctx_idx]; | |
2637 | ||
2638 | if (set->ops->init_hctx && | |
2639 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2640 | goto unregister_cpu_notifier; | |
08e98fc6 | 2641 | |
7c6c5b7c ML |
2642 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2643 | hctx->numa_node)) | |
2644 | goto exit_hctx; | |
2645 | return 0; | |
2646 | ||
2647 | exit_hctx: | |
2648 | if (set->ops->exit_hctx) | |
2649 | set->ops->exit_hctx(hctx, hctx_idx); | |
2650 | unregister_cpu_notifier: | |
2651 | blk_mq_remove_cpuhp(hctx); | |
2652 | return -1; | |
2653 | } | |
2654 | ||
2655 | static struct blk_mq_hw_ctx * | |
2656 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2657 | int node) | |
2658 | { | |
2659 | struct blk_mq_hw_ctx *hctx; | |
2660 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2661 | ||
2662 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2663 | if (!hctx) | |
2664 | goto fail_alloc_hctx; | |
2665 | ||
2666 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2667 | goto free_hctx; | |
2668 | ||
2669 | atomic_set(&hctx->nr_active, 0); | |
b445547e | 2670 | atomic_set(&hctx->elevator_queued, 0); |
08e98fc6 | 2671 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2672 | node = set->numa_node; |
2673 | hctx->numa_node = node; | |
08e98fc6 | 2674 | |
9f993737 | 2675 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2676 | spin_lock_init(&hctx->lock); |
2677 | INIT_LIST_HEAD(&hctx->dispatch); | |
2678 | hctx->queue = q; | |
51db1c37 | 2679 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2680 | |
2f8f1336 ML |
2681 | INIT_LIST_HEAD(&hctx->hctx_list); |
2682 | ||
320ae51f | 2683 | /* |
08e98fc6 ML |
2684 | * Allocate space for all possible cpus to avoid allocation at |
2685 | * runtime | |
320ae51f | 2686 | */ |
d904bfa7 | 2687 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2688 | gfp, node); |
08e98fc6 | 2689 | if (!hctx->ctxs) |
7c6c5b7c | 2690 | goto free_cpumask; |
320ae51f | 2691 | |
5b202853 | 2692 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2693 | gfp, node)) |
08e98fc6 | 2694 | goto free_ctxs; |
08e98fc6 | 2695 | hctx->nr_ctx = 0; |
320ae51f | 2696 | |
5815839b | 2697 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2698 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2699 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2700 | ||
754a1572 | 2701 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2702 | if (!hctx->fq) |
7c6c5b7c | 2703 | goto free_bitmap; |
320ae51f | 2704 | |
6a83e74d | 2705 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2706 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2707 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2708 | |
7c6c5b7c | 2709 | return hctx; |
320ae51f | 2710 | |
08e98fc6 | 2711 | free_bitmap: |
88459642 | 2712 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2713 | free_ctxs: |
2714 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2715 | free_cpumask: |
2716 | free_cpumask_var(hctx->cpumask); | |
2717 | free_hctx: | |
2718 | kfree(hctx); | |
2719 | fail_alloc_hctx: | |
2720 | return NULL; | |
08e98fc6 | 2721 | } |
320ae51f | 2722 | |
320ae51f JA |
2723 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2724 | unsigned int nr_hw_queues) | |
2725 | { | |
b3c661b1 JA |
2726 | struct blk_mq_tag_set *set = q->tag_set; |
2727 | unsigned int i, j; | |
320ae51f JA |
2728 | |
2729 | for_each_possible_cpu(i) { | |
2730 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2731 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2732 | int k; |
320ae51f | 2733 | |
320ae51f JA |
2734 | __ctx->cpu = i; |
2735 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2736 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2737 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2738 | ||
320ae51f JA |
2739 | __ctx->queue = q; |
2740 | ||
320ae51f JA |
2741 | /* |
2742 | * Set local node, IFF we have more than one hw queue. If | |
2743 | * not, we remain on the home node of the device | |
2744 | */ | |
b3c661b1 JA |
2745 | for (j = 0; j < set->nr_maps; j++) { |
2746 | hctx = blk_mq_map_queue_type(q, j, i); | |
2747 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2748 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2749 | } |
320ae51f JA |
2750 | } |
2751 | } | |
2752 | ||
03b63b02 WZ |
2753 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, |
2754 | int hctx_idx) | |
cc71a6f4 | 2755 | { |
1c0706a7 | 2756 | unsigned int flags = set->flags; |
cc71a6f4 JA |
2757 | int ret = 0; |
2758 | ||
2759 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1c0706a7 | 2760 | set->queue_depth, set->reserved_tags, flags); |
cc71a6f4 JA |
2761 | if (!set->tags[hctx_idx]) |
2762 | return false; | |
2763 | ||
2764 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2765 | set->queue_depth); | |
2766 | if (!ret) | |
2767 | return true; | |
2768 | ||
1c0706a7 | 2769 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
cc71a6f4 JA |
2770 | set->tags[hctx_idx] = NULL; |
2771 | return false; | |
2772 | } | |
2773 | ||
2774 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2775 | unsigned int hctx_idx) | |
2776 | { | |
1c0706a7 JG |
2777 | unsigned int flags = set->flags; |
2778 | ||
4e6db0f2 | 2779 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 | 2780 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
1c0706a7 | 2781 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
bd166ef1 JA |
2782 | set->tags[hctx_idx] = NULL; |
2783 | } | |
cc71a6f4 JA |
2784 | } |
2785 | ||
4b855ad3 | 2786 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2787 | { |
b3c661b1 | 2788 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2789 | struct blk_mq_hw_ctx *hctx; |
2790 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2791 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2792 | |
2793 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2794 | cpumask_clear(hctx->cpumask); |
320ae51f | 2795 | hctx->nr_ctx = 0; |
d416c92c | 2796 | hctx->dispatch_from = NULL; |
320ae51f JA |
2797 | } |
2798 | ||
2799 | /* | |
4b855ad3 | 2800 | * Map software to hardware queues. |
4412efec ML |
2801 | * |
2802 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2803 | */ |
20e4d813 | 2804 | for_each_possible_cpu(i) { |
4412efec | 2805 | |
897bb0c7 | 2806 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2807 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2808 | if (!set->map[j].nr_queues) { |
2809 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2810 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2811 | continue; |
bb94aea1 | 2812 | } |
fd689871 ML |
2813 | hctx_idx = set->map[j].mq_map[i]; |
2814 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2815 | if (!set->tags[hctx_idx] && | |
03b63b02 | 2816 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { |
fd689871 ML |
2817 | /* |
2818 | * If tags initialization fail for some hctx, | |
2819 | * that hctx won't be brought online. In this | |
2820 | * case, remap the current ctx to hctx[0] which | |
2821 | * is guaranteed to always have tags allocated | |
2822 | */ | |
2823 | set->map[j].mq_map[i] = 0; | |
2824 | } | |
e5edd5f2 | 2825 | |
b3c661b1 | 2826 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2827 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2828 | /* |
2829 | * If the CPU is already set in the mask, then we've | |
2830 | * mapped this one already. This can happen if | |
2831 | * devices share queues across queue maps. | |
2832 | */ | |
2833 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2834 | continue; | |
2835 | ||
2836 | cpumask_set_cpu(i, hctx->cpumask); | |
2837 | hctx->type = j; | |
2838 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2839 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2840 | ||
2841 | /* | |
2842 | * If the nr_ctx type overflows, we have exceeded the | |
2843 | * amount of sw queues we can support. | |
2844 | */ | |
2845 | BUG_ON(!hctx->nr_ctx); | |
2846 | } | |
bb94aea1 JW |
2847 | |
2848 | for (; j < HCTX_MAX_TYPES; j++) | |
2849 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2850 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2851 | } |
506e931f JA |
2852 | |
2853 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2854 | /* |
2855 | * If no software queues are mapped to this hardware queue, | |
2856 | * disable it and free the request entries. | |
2857 | */ | |
2858 | if (!hctx->nr_ctx) { | |
2859 | /* Never unmap queue 0. We need it as a | |
2860 | * fallback in case of a new remap fails | |
2861 | * allocation | |
2862 | */ | |
2863 | if (i && set->tags[i]) | |
2864 | blk_mq_free_map_and_requests(set, i); | |
2865 | ||
2866 | hctx->tags = NULL; | |
2867 | continue; | |
2868 | } | |
484b4061 | 2869 | |
2a34c087 ML |
2870 | hctx->tags = set->tags[i]; |
2871 | WARN_ON(!hctx->tags); | |
2872 | ||
889fa31f CY |
2873 | /* |
2874 | * Set the map size to the number of mapped software queues. | |
2875 | * This is more accurate and more efficient than looping | |
2876 | * over all possibly mapped software queues. | |
2877 | */ | |
88459642 | 2878 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2879 | |
484b4061 JA |
2880 | /* |
2881 | * Initialize batch roundrobin counts | |
2882 | */ | |
f82ddf19 | 2883 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2884 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2885 | } | |
320ae51f JA |
2886 | } |
2887 | ||
8e8320c9 JA |
2888 | /* |
2889 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2890 | * the queue isn't live yet. | |
2891 | */ | |
2404e607 | 2892 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2893 | { |
2894 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2895 | int i; |
2896 | ||
2404e607 | 2897 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2898 | if (shared) |
51db1c37 | 2899 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
97889f9a | 2900 | else |
51db1c37 | 2901 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 JM |
2902 | } |
2903 | } | |
2904 | ||
655ac300 HR |
2905 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
2906 | bool shared) | |
2404e607 JM |
2907 | { |
2908 | struct request_queue *q; | |
0d2602ca | 2909 | |
705cda97 BVA |
2910 | lockdep_assert_held(&set->tag_list_lock); |
2911 | ||
0d2602ca JA |
2912 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2913 | blk_mq_freeze_queue(q); | |
2404e607 | 2914 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2915 | blk_mq_unfreeze_queue(q); |
2916 | } | |
2917 | } | |
2918 | ||
2919 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2920 | { | |
2921 | struct blk_mq_tag_set *set = q->tag_set; | |
2922 | ||
0d2602ca | 2923 | mutex_lock(&set->tag_list_lock); |
08c875cb | 2924 | list_del(&q->tag_set_list); |
2404e607 JM |
2925 | if (list_is_singular(&set->tag_list)) { |
2926 | /* just transitioned to unshared */ | |
51db1c37 | 2927 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 2928 | /* update existing queue */ |
655ac300 | 2929 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 2930 | } |
0d2602ca | 2931 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2932 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2933 | } |
2934 | ||
2935 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2936 | struct request_queue *q) | |
2937 | { | |
0d2602ca | 2938 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2939 | |
ff821d27 JA |
2940 | /* |
2941 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2942 | */ | |
2943 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
2944 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
2945 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 2946 | /* update existing queue */ |
655ac300 | 2947 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 2948 | } |
51db1c37 | 2949 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 2950 | queue_set_hctx_shared(q, true); |
08c875cb | 2951 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2952 | |
0d2602ca JA |
2953 | mutex_unlock(&set->tag_list_lock); |
2954 | } | |
2955 | ||
1db4909e ML |
2956 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2957 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2958 | { | |
2959 | struct blk_mq_ctxs *ctxs; | |
2960 | int cpu; | |
2961 | ||
2962 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2963 | if (!ctxs) | |
2964 | return -ENOMEM; | |
2965 | ||
2966 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2967 | if (!ctxs->queue_ctx) | |
2968 | goto fail; | |
2969 | ||
2970 | for_each_possible_cpu(cpu) { | |
2971 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2972 | ctx->ctxs = ctxs; | |
2973 | } | |
2974 | ||
2975 | q->mq_kobj = &ctxs->kobj; | |
2976 | q->queue_ctx = ctxs->queue_ctx; | |
2977 | ||
2978 | return 0; | |
2979 | fail: | |
2980 | kfree(ctxs); | |
2981 | return -ENOMEM; | |
2982 | } | |
2983 | ||
e09aae7e ML |
2984 | /* |
2985 | * It is the actual release handler for mq, but we do it from | |
2986 | * request queue's release handler for avoiding use-after-free | |
2987 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2988 | * but we can't group ctx/kctx kobj without it. | |
2989 | */ | |
2990 | void blk_mq_release(struct request_queue *q) | |
2991 | { | |
2f8f1336 ML |
2992 | struct blk_mq_hw_ctx *hctx, *next; |
2993 | int i; | |
e09aae7e | 2994 | |
2f8f1336 ML |
2995 | queue_for_each_hw_ctx(q, hctx, i) |
2996 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2997 | ||
2998 | /* all hctx are in .unused_hctx_list now */ | |
2999 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3000 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3001 | kobject_put(&hctx->kobj); |
c3b4afca | 3002 | } |
e09aae7e ML |
3003 | |
3004 | kfree(q->queue_hw_ctx); | |
3005 | ||
7ea5fe31 ML |
3006 | /* |
3007 | * release .mq_kobj and sw queue's kobject now because | |
3008 | * both share lifetime with request queue. | |
3009 | */ | |
3010 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3011 | } |
3012 | ||
2f227bb9 CH |
3013 | struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
3014 | void *queuedata) | |
b62c21b7 MS |
3015 | { |
3016 | struct request_queue *uninit_q, *q; | |
3017 | ||
c62b37d9 | 3018 | uninit_q = blk_alloc_queue(set->numa_node); |
b62c21b7 MS |
3019 | if (!uninit_q) |
3020 | return ERR_PTR(-ENOMEM); | |
2f227bb9 | 3021 | uninit_q->queuedata = queuedata; |
b62c21b7 | 3022 | |
737eb78e DLM |
3023 | /* |
3024 | * Initialize the queue without an elevator. device_add_disk() will do | |
3025 | * the initialization. | |
3026 | */ | |
3027 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
b62c21b7 MS |
3028 | if (IS_ERR(q)) |
3029 | blk_cleanup_queue(uninit_q); | |
3030 | ||
3031 | return q; | |
3032 | } | |
2f227bb9 CH |
3033 | EXPORT_SYMBOL_GPL(blk_mq_init_queue_data); |
3034 | ||
3035 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3036 | { | |
3037 | return blk_mq_init_queue_data(set, NULL); | |
3038 | } | |
b62c21b7 MS |
3039 | EXPORT_SYMBOL(blk_mq_init_queue); |
3040 | ||
9316a9ed JA |
3041 | /* |
3042 | * Helper for setting up a queue with mq ops, given queue depth, and | |
3043 | * the passed in mq ops flags. | |
3044 | */ | |
3045 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
3046 | const struct blk_mq_ops *ops, | |
3047 | unsigned int queue_depth, | |
3048 | unsigned int set_flags) | |
3049 | { | |
3050 | struct request_queue *q; | |
3051 | int ret; | |
3052 | ||
3053 | memset(set, 0, sizeof(*set)); | |
3054 | set->ops = ops; | |
3055 | set->nr_hw_queues = 1; | |
b3c661b1 | 3056 | set->nr_maps = 1; |
9316a9ed JA |
3057 | set->queue_depth = queue_depth; |
3058 | set->numa_node = NUMA_NO_NODE; | |
3059 | set->flags = set_flags; | |
3060 | ||
3061 | ret = blk_mq_alloc_tag_set(set); | |
3062 | if (ret) | |
3063 | return ERR_PTR(ret); | |
3064 | ||
3065 | q = blk_mq_init_queue(set); | |
3066 | if (IS_ERR(q)) { | |
3067 | blk_mq_free_tag_set(set); | |
3068 | return q; | |
3069 | } | |
3070 | ||
3071 | return q; | |
3072 | } | |
3073 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
3074 | ||
34d11ffa JW |
3075 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3076 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3077 | int hctx_idx, int node) | |
3078 | { | |
2f8f1336 | 3079 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3080 | |
2f8f1336 ML |
3081 | /* reuse dead hctx first */ |
3082 | spin_lock(&q->unused_hctx_lock); | |
3083 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3084 | if (tmp->numa_node == node) { | |
3085 | hctx = tmp; | |
3086 | break; | |
3087 | } | |
3088 | } | |
3089 | if (hctx) | |
3090 | list_del_init(&hctx->hctx_list); | |
3091 | spin_unlock(&q->unused_hctx_lock); | |
3092 | ||
3093 | if (!hctx) | |
3094 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3095 | if (!hctx) |
7c6c5b7c | 3096 | goto fail; |
34d11ffa | 3097 | |
7c6c5b7c ML |
3098 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3099 | goto free_hctx; | |
34d11ffa JW |
3100 | |
3101 | return hctx; | |
7c6c5b7c ML |
3102 | |
3103 | free_hctx: | |
3104 | kobject_put(&hctx->kobj); | |
3105 | fail: | |
3106 | return NULL; | |
34d11ffa JW |
3107 | } |
3108 | ||
868f2f0b KB |
3109 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3110 | struct request_queue *q) | |
320ae51f | 3111 | { |
e01ad46d | 3112 | int i, j, end; |
868f2f0b | 3113 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3114 | |
ac0d6b92 BVA |
3115 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3116 | struct blk_mq_hw_ctx **new_hctxs; | |
3117 | ||
3118 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3119 | sizeof(*new_hctxs), GFP_KERNEL, | |
3120 | set->numa_node); | |
3121 | if (!new_hctxs) | |
3122 | return; | |
3123 | if (hctxs) | |
3124 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3125 | sizeof(*hctxs)); | |
3126 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3127 | kfree(hctxs); |
3128 | hctxs = new_hctxs; | |
3129 | } | |
3130 | ||
fb350e0a ML |
3131 | /* protect against switching io scheduler */ |
3132 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3133 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3134 | int node; |
34d11ffa | 3135 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3136 | |
7d76f856 | 3137 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3138 | /* |
3139 | * If the hw queue has been mapped to another numa node, | |
3140 | * we need to realloc the hctx. If allocation fails, fallback | |
3141 | * to use the previous one. | |
3142 | */ | |
3143 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3144 | continue; | |
868f2f0b | 3145 | |
34d11ffa JW |
3146 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3147 | if (hctx) { | |
2f8f1336 | 3148 | if (hctxs[i]) |
34d11ffa | 3149 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3150 | hctxs[i] = hctx; |
3151 | } else { | |
3152 | if (hctxs[i]) | |
3153 | pr_warn("Allocate new hctx on node %d fails,\ | |
3154 | fallback to previous one on node %d\n", | |
3155 | node, hctxs[i]->numa_node); | |
3156 | else | |
3157 | break; | |
868f2f0b | 3158 | } |
320ae51f | 3159 | } |
e01ad46d JW |
3160 | /* |
3161 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3162 | * hctxs and keep the previous q->nr_hw_queues. | |
3163 | */ | |
3164 | if (i != set->nr_hw_queues) { | |
3165 | j = q->nr_hw_queues; | |
3166 | end = i; | |
3167 | } else { | |
3168 | j = i; | |
3169 | end = q->nr_hw_queues; | |
3170 | q->nr_hw_queues = set->nr_hw_queues; | |
3171 | } | |
34d11ffa | 3172 | |
e01ad46d | 3173 | for (; j < end; j++) { |
868f2f0b KB |
3174 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3175 | ||
3176 | if (hctx) { | |
cc71a6f4 JA |
3177 | if (hctx->tags) |
3178 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 3179 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3180 | hctxs[j] = NULL; |
868f2f0b KB |
3181 | } |
3182 | } | |
fb350e0a | 3183 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3184 | } |
3185 | ||
3186 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
737eb78e DLM |
3187 | struct request_queue *q, |
3188 | bool elevator_init) | |
868f2f0b | 3189 | { |
66841672 ML |
3190 | /* mark the queue as mq asap */ |
3191 | q->mq_ops = set->ops; | |
3192 | ||
34dbad5d | 3193 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3194 | blk_mq_poll_stats_bkt, |
3195 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3196 | if (!q->poll_cb) |
3197 | goto err_exit; | |
3198 | ||
1db4909e | 3199 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3200 | goto err_poll; |
868f2f0b | 3201 | |
737f98cf ML |
3202 | /* init q->mq_kobj and sw queues' kobjects */ |
3203 | blk_mq_sysfs_init(q); | |
3204 | ||
2f8f1336 ML |
3205 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3206 | spin_lock_init(&q->unused_hctx_lock); | |
3207 | ||
868f2f0b KB |
3208 | blk_mq_realloc_hw_ctxs(set, q); |
3209 | if (!q->nr_hw_queues) | |
3210 | goto err_hctxs; | |
320ae51f | 3211 | |
287922eb | 3212 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3213 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3214 | |
a8908939 | 3215 | q->tag_set = set; |
320ae51f | 3216 | |
94eddfbe | 3217 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3218 | if (set->nr_maps > HCTX_TYPE_POLL && |
3219 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3220 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3221 | |
1be036e9 CH |
3222 | q->sg_reserved_size = INT_MAX; |
3223 | ||
2849450a | 3224 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3225 | INIT_LIST_HEAD(&q->requeue_list); |
3226 | spin_lock_init(&q->requeue_lock); | |
3227 | ||
eba71768 JA |
3228 | q->nr_requests = set->queue_depth; |
3229 | ||
64f1c21e JA |
3230 | /* |
3231 | * Default to classic polling | |
3232 | */ | |
29ece8b4 | 3233 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3234 | |
24d2f903 | 3235 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3236 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3237 | blk_mq_map_swqueue(q); |
4593fdbe | 3238 | |
737eb78e DLM |
3239 | if (elevator_init) |
3240 | elevator_init_mq(q); | |
d3484991 | 3241 | |
320ae51f | 3242 | return q; |
18741986 | 3243 | |
320ae51f | 3244 | err_hctxs: |
868f2f0b | 3245 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3246 | q->nr_hw_queues = 0; |
1db4909e | 3247 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3248 | err_poll: |
3249 | blk_stat_free_callback(q->poll_cb); | |
3250 | q->poll_cb = NULL; | |
c7de5726 ML |
3251 | err_exit: |
3252 | q->mq_ops = NULL; | |
320ae51f JA |
3253 | return ERR_PTR(-ENOMEM); |
3254 | } | |
b62c21b7 | 3255 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3256 | |
c7e2d94b ML |
3257 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3258 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3259 | { |
624dbe47 | 3260 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3261 | |
0d2602ca | 3262 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 3263 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 3264 | } |
320ae51f | 3265 | |
a5164405 JA |
3266 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3267 | { | |
3268 | int i; | |
3269 | ||
8229cca8 | 3270 | for (i = 0; i < set->nr_hw_queues; i++) { |
03b63b02 | 3271 | if (!__blk_mq_alloc_map_and_request(set, i)) |
a5164405 | 3272 | goto out_unwind; |
8229cca8 XT |
3273 | cond_resched(); |
3274 | } | |
a5164405 JA |
3275 | |
3276 | return 0; | |
3277 | ||
3278 | out_unwind: | |
3279 | while (--i >= 0) | |
2e194422 | 3280 | blk_mq_free_map_and_requests(set, i); |
a5164405 | 3281 | |
a5164405 JA |
3282 | return -ENOMEM; |
3283 | } | |
3284 | ||
3285 | /* | |
3286 | * Allocate the request maps associated with this tag_set. Note that this | |
3287 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3288 | * will be updated to reflect the allocated depth. | |
3289 | */ | |
79fab528 | 3290 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) |
a5164405 JA |
3291 | { |
3292 | unsigned int depth; | |
3293 | int err; | |
3294 | ||
3295 | depth = set->queue_depth; | |
3296 | do { | |
3297 | err = __blk_mq_alloc_rq_maps(set); | |
3298 | if (!err) | |
3299 | break; | |
3300 | ||
3301 | set->queue_depth >>= 1; | |
3302 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3303 | err = -ENOMEM; | |
3304 | break; | |
3305 | } | |
3306 | } while (set->queue_depth); | |
3307 | ||
3308 | if (!set->queue_depth || err) { | |
3309 | pr_err("blk-mq: failed to allocate request map\n"); | |
3310 | return -ENOMEM; | |
3311 | } | |
3312 | ||
3313 | if (depth != set->queue_depth) | |
3314 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3315 | depth, set->queue_depth); | |
3316 | ||
3317 | return 0; | |
3318 | } | |
3319 | ||
ebe8bddb OS |
3320 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3321 | { | |
6e66b493 BVA |
3322 | /* |
3323 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3324 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3325 | * number of hardware queues. | |
3326 | */ | |
3327 | if (set->nr_maps == 1) | |
3328 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3329 | ||
59388702 | 3330 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3331 | int i; |
3332 | ||
7d4901a9 ML |
3333 | /* |
3334 | * transport .map_queues is usually done in the following | |
3335 | * way: | |
3336 | * | |
3337 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3338 | * mask = get_cpu_mask(queue) | |
3339 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3340 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3341 | * } |
3342 | * | |
3343 | * When we need to remap, the table has to be cleared for | |
3344 | * killing stale mapping since one CPU may not be mapped | |
3345 | * to any hw queue. | |
3346 | */ | |
b3c661b1 JA |
3347 | for (i = 0; i < set->nr_maps; i++) |
3348 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3349 | |
ebe8bddb | 3350 | return set->ops->map_queues(set); |
b3c661b1 JA |
3351 | } else { |
3352 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3353 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3354 | } |
ebe8bddb OS |
3355 | } |
3356 | ||
f7e76dbc BVA |
3357 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3358 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3359 | { | |
3360 | struct blk_mq_tags **new_tags; | |
3361 | ||
3362 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3363 | return 0; | |
3364 | ||
3365 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3366 | GFP_KERNEL, set->numa_node); | |
3367 | if (!new_tags) | |
3368 | return -ENOMEM; | |
3369 | ||
3370 | if (set->tags) | |
3371 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3372 | sizeof(*set->tags)); | |
3373 | kfree(set->tags); | |
3374 | set->tags = new_tags; | |
3375 | set->nr_hw_queues = new_nr_hw_queues; | |
3376 | ||
3377 | return 0; | |
3378 | } | |
3379 | ||
a4391c64 JA |
3380 | /* |
3381 | * Alloc a tag set to be associated with one or more request queues. | |
3382 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3383 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3384 | * value will be stored in set->queue_depth. |
3385 | */ | |
24d2f903 CH |
3386 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3387 | { | |
b3c661b1 | 3388 | int i, ret; |
da695ba2 | 3389 | |
205fb5f5 BVA |
3390 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3391 | ||
24d2f903 CH |
3392 | if (!set->nr_hw_queues) |
3393 | return -EINVAL; | |
a4391c64 | 3394 | if (!set->queue_depth) |
24d2f903 CH |
3395 | return -EINVAL; |
3396 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3397 | return -EINVAL; | |
3398 | ||
7d7e0f90 | 3399 | if (!set->ops->queue_rq) |
24d2f903 CH |
3400 | return -EINVAL; |
3401 | ||
de148297 ML |
3402 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3403 | return -EINVAL; | |
3404 | ||
a4391c64 JA |
3405 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3406 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3407 | BLK_MQ_MAX_DEPTH); | |
3408 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3409 | } | |
24d2f903 | 3410 | |
b3c661b1 JA |
3411 | if (!set->nr_maps) |
3412 | set->nr_maps = 1; | |
3413 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3414 | return -EINVAL; | |
3415 | ||
6637fadf SL |
3416 | /* |
3417 | * If a crashdump is active, then we are potentially in a very | |
3418 | * memory constrained environment. Limit us to 1 queue and | |
3419 | * 64 tags to prevent using too much memory. | |
3420 | */ | |
3421 | if (is_kdump_kernel()) { | |
3422 | set->nr_hw_queues = 1; | |
59388702 | 3423 | set->nr_maps = 1; |
6637fadf SL |
3424 | set->queue_depth = min(64U, set->queue_depth); |
3425 | } | |
868f2f0b | 3426 | /* |
392546ae JA |
3427 | * There is no use for more h/w queues than cpus if we just have |
3428 | * a single map | |
868f2f0b | 3429 | */ |
392546ae | 3430 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3431 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3432 | |
f7e76dbc | 3433 | if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0) |
a5164405 | 3434 | return -ENOMEM; |
24d2f903 | 3435 | |
da695ba2 | 3436 | ret = -ENOMEM; |
b3c661b1 JA |
3437 | for (i = 0; i < set->nr_maps; i++) { |
3438 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3439 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3440 | GFP_KERNEL, set->numa_node); |
3441 | if (!set->map[i].mq_map) | |
3442 | goto out_free_mq_map; | |
59388702 | 3443 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3444 | } |
bdd17e75 | 3445 | |
ebe8bddb | 3446 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3447 | if (ret) |
3448 | goto out_free_mq_map; | |
3449 | ||
79fab528 | 3450 | ret = blk_mq_alloc_map_and_requests(set); |
da695ba2 | 3451 | if (ret) |
bdd17e75 | 3452 | goto out_free_mq_map; |
24d2f903 | 3453 | |
32bc15af | 3454 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
f1b49fdc JG |
3455 | atomic_set(&set->active_queues_shared_sbitmap, 0); |
3456 | ||
32bc15af JG |
3457 | if (blk_mq_init_shared_sbitmap(set, set->flags)) { |
3458 | ret = -ENOMEM; | |
3459 | goto out_free_mq_rq_maps; | |
3460 | } | |
3461 | } | |
3462 | ||
0d2602ca JA |
3463 | mutex_init(&set->tag_list_lock); |
3464 | INIT_LIST_HEAD(&set->tag_list); | |
3465 | ||
24d2f903 | 3466 | return 0; |
bdd17e75 | 3467 | |
32bc15af JG |
3468 | out_free_mq_rq_maps: |
3469 | for (i = 0; i < set->nr_hw_queues; i++) | |
3470 | blk_mq_free_map_and_requests(set, i); | |
bdd17e75 | 3471 | out_free_mq_map: |
b3c661b1 JA |
3472 | for (i = 0; i < set->nr_maps; i++) { |
3473 | kfree(set->map[i].mq_map); | |
3474 | set->map[i].mq_map = NULL; | |
3475 | } | |
5676e7b6 RE |
3476 | kfree(set->tags); |
3477 | set->tags = NULL; | |
da695ba2 | 3478 | return ret; |
24d2f903 CH |
3479 | } |
3480 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3481 | ||
3482 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3483 | { | |
b3c661b1 | 3484 | int i, j; |
24d2f903 | 3485 | |
f7e76dbc | 3486 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3487 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3488 | |
32bc15af JG |
3489 | if (blk_mq_is_sbitmap_shared(set->flags)) |
3490 | blk_mq_exit_shared_sbitmap(set); | |
3491 | ||
b3c661b1 JA |
3492 | for (j = 0; j < set->nr_maps; j++) { |
3493 | kfree(set->map[j].mq_map); | |
3494 | set->map[j].mq_map = NULL; | |
3495 | } | |
bdd17e75 | 3496 | |
981bd189 | 3497 | kfree(set->tags); |
5676e7b6 | 3498 | set->tags = NULL; |
24d2f903 CH |
3499 | } |
3500 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3501 | ||
e3a2b3f9 JA |
3502 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3503 | { | |
3504 | struct blk_mq_tag_set *set = q->tag_set; | |
3505 | struct blk_mq_hw_ctx *hctx; | |
3506 | int i, ret; | |
3507 | ||
bd166ef1 | 3508 | if (!set) |
e3a2b3f9 JA |
3509 | return -EINVAL; |
3510 | ||
e5fa8140 AZ |
3511 | if (q->nr_requests == nr) |
3512 | return 0; | |
3513 | ||
70f36b60 | 3514 | blk_mq_freeze_queue(q); |
24f5a90f | 3515 | blk_mq_quiesce_queue(q); |
70f36b60 | 3516 | |
e3a2b3f9 JA |
3517 | ret = 0; |
3518 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3519 | if (!hctx->tags) |
3520 | continue; | |
bd166ef1 JA |
3521 | /* |
3522 | * If we're using an MQ scheduler, just update the scheduler | |
3523 | * queue depth. This is similar to what the old code would do. | |
3524 | */ | |
70f36b60 | 3525 | if (!hctx->sched_tags) { |
c2e82a23 | 3526 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 | 3527 | false); |
32bc15af JG |
3528 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) |
3529 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
70f36b60 JA |
3530 | } else { |
3531 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3532 | nr, true); | |
3533 | } | |
e3a2b3f9 JA |
3534 | if (ret) |
3535 | break; | |
77f1e0a5 JA |
3536 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3537 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3538 | } |
3539 | ||
3540 | if (!ret) | |
3541 | q->nr_requests = nr; | |
3542 | ||
24f5a90f | 3543 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3544 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3545 | |
e3a2b3f9 JA |
3546 | return ret; |
3547 | } | |
3548 | ||
d48ece20 JW |
3549 | /* |
3550 | * request_queue and elevator_type pair. | |
3551 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3552 | * the elevator_type associated with a request_queue. | |
3553 | */ | |
3554 | struct blk_mq_qe_pair { | |
3555 | struct list_head node; | |
3556 | struct request_queue *q; | |
3557 | struct elevator_type *type; | |
3558 | }; | |
3559 | ||
3560 | /* | |
3561 | * Cache the elevator_type in qe pair list and switch the | |
3562 | * io scheduler to 'none' | |
3563 | */ | |
3564 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3565 | struct request_queue *q) | |
3566 | { | |
3567 | struct blk_mq_qe_pair *qe; | |
3568 | ||
3569 | if (!q->elevator) | |
3570 | return true; | |
3571 | ||
3572 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3573 | if (!qe) | |
3574 | return false; | |
3575 | ||
3576 | INIT_LIST_HEAD(&qe->node); | |
3577 | qe->q = q; | |
3578 | qe->type = q->elevator->type; | |
3579 | list_add(&qe->node, head); | |
3580 | ||
3581 | mutex_lock(&q->sysfs_lock); | |
3582 | /* | |
3583 | * After elevator_switch_mq, the previous elevator_queue will be | |
3584 | * released by elevator_release. The reference of the io scheduler | |
3585 | * module get by elevator_get will also be put. So we need to get | |
3586 | * a reference of the io scheduler module here to prevent it to be | |
3587 | * removed. | |
3588 | */ | |
3589 | __module_get(qe->type->elevator_owner); | |
3590 | elevator_switch_mq(q, NULL); | |
3591 | mutex_unlock(&q->sysfs_lock); | |
3592 | ||
3593 | return true; | |
3594 | } | |
3595 | ||
3596 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3597 | struct request_queue *q) | |
3598 | { | |
3599 | struct blk_mq_qe_pair *qe; | |
3600 | struct elevator_type *t = NULL; | |
3601 | ||
3602 | list_for_each_entry(qe, head, node) | |
3603 | if (qe->q == q) { | |
3604 | t = qe->type; | |
3605 | break; | |
3606 | } | |
3607 | ||
3608 | if (!t) | |
3609 | return; | |
3610 | ||
3611 | list_del(&qe->node); | |
3612 | kfree(qe); | |
3613 | ||
3614 | mutex_lock(&q->sysfs_lock); | |
3615 | elevator_switch_mq(q, t); | |
3616 | mutex_unlock(&q->sysfs_lock); | |
3617 | } | |
3618 | ||
e4dc2b32 KB |
3619 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3620 | int nr_hw_queues) | |
868f2f0b KB |
3621 | { |
3622 | struct request_queue *q; | |
d48ece20 | 3623 | LIST_HEAD(head); |
e01ad46d | 3624 | int prev_nr_hw_queues; |
868f2f0b | 3625 | |
705cda97 BVA |
3626 | lockdep_assert_held(&set->tag_list_lock); |
3627 | ||
392546ae | 3628 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3629 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3630 | if (nr_hw_queues < 1) |
3631 | return; | |
3632 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3633 | return; |
3634 | ||
3635 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3636 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3637 | /* |
3638 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3639 | * with the previous scheduler. We will switch back once we are done | |
3640 | * updating the new sw to hw queue mappings. | |
3641 | */ | |
3642 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3643 | if (!blk_mq_elv_switch_none(&head, q)) | |
3644 | goto switch_back; | |
868f2f0b | 3645 | |
477e19de JW |
3646 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3647 | blk_mq_debugfs_unregister_hctxs(q); | |
3648 | blk_mq_sysfs_unregister(q); | |
3649 | } | |
3650 | ||
a2584e43 | 3651 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3652 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3653 | 0) | |
3654 | goto reregister; | |
3655 | ||
868f2f0b | 3656 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3657 | fallback: |
aa880ad6 | 3658 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3659 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3660 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3661 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3662 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3663 | nr_hw_queues, prev_nr_hw_queues); | |
3664 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3665 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3666 | goto fallback; |
3667 | } | |
477e19de JW |
3668 | blk_mq_map_swqueue(q); |
3669 | } | |
3670 | ||
f7e76dbc | 3671 | reregister: |
477e19de JW |
3672 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3673 | blk_mq_sysfs_register(q); | |
3674 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3675 | } |
3676 | ||
d48ece20 JW |
3677 | switch_back: |
3678 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3679 | blk_mq_elv_switch_back(&head, q); | |
3680 | ||
868f2f0b KB |
3681 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3682 | blk_mq_unfreeze_queue(q); | |
3683 | } | |
e4dc2b32 KB |
3684 | |
3685 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3686 | { | |
3687 | mutex_lock(&set->tag_list_lock); | |
3688 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3689 | mutex_unlock(&set->tag_list_lock); | |
3690 | } | |
868f2f0b KB |
3691 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3692 | ||
34dbad5d OS |
3693 | /* Enable polling stats and return whether they were already enabled. */ |
3694 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3695 | { | |
3696 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3697 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3698 | return true; |
3699 | blk_stat_add_callback(q, q->poll_cb); | |
3700 | return false; | |
3701 | } | |
3702 | ||
3703 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3704 | { | |
3705 | /* | |
3706 | * We don't arm the callback if polling stats are not enabled or the | |
3707 | * callback is already active. | |
3708 | */ | |
3709 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3710 | blk_stat_is_active(q->poll_cb)) | |
3711 | return; | |
3712 | ||
3713 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3714 | } | |
3715 | ||
3716 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3717 | { | |
3718 | struct request_queue *q = cb->data; | |
720b8ccc | 3719 | int bucket; |
34dbad5d | 3720 | |
720b8ccc SB |
3721 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3722 | if (cb->stat[bucket].nr_samples) | |
3723 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3724 | } | |
34dbad5d OS |
3725 | } |
3726 | ||
64f1c21e | 3727 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3728 | struct request *rq) |
3729 | { | |
64f1c21e | 3730 | unsigned long ret = 0; |
720b8ccc | 3731 | int bucket; |
64f1c21e JA |
3732 | |
3733 | /* | |
3734 | * If stats collection isn't on, don't sleep but turn it on for | |
3735 | * future users | |
3736 | */ | |
34dbad5d | 3737 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3738 | return 0; |
3739 | ||
64f1c21e JA |
3740 | /* |
3741 | * As an optimistic guess, use half of the mean service time | |
3742 | * for this type of request. We can (and should) make this smarter. | |
3743 | * For instance, if the completion latencies are tight, we can | |
3744 | * get closer than just half the mean. This is especially | |
3745 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3746 | * than ~10 usec. We do use the stats for the relevant IO size |
3747 | * if available which does lead to better estimates. | |
64f1c21e | 3748 | */ |
720b8ccc SB |
3749 | bucket = blk_mq_poll_stats_bkt(rq); |
3750 | if (bucket < 0) | |
3751 | return ret; | |
3752 | ||
3753 | if (q->poll_stat[bucket].nr_samples) | |
3754 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3755 | |
3756 | return ret; | |
3757 | } | |
3758 | ||
06426adf JA |
3759 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3760 | struct request *rq) | |
3761 | { | |
3762 | struct hrtimer_sleeper hs; | |
3763 | enum hrtimer_mode mode; | |
64f1c21e | 3764 | unsigned int nsecs; |
06426adf JA |
3765 | ktime_t kt; |
3766 | ||
76a86f9d | 3767 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3768 | return false; |
3769 | ||
3770 | /* | |
1052b8ac | 3771 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3772 | * |
64f1c21e JA |
3773 | * 0: use half of prev avg |
3774 | * >0: use this specific value | |
3775 | */ | |
1052b8ac | 3776 | if (q->poll_nsec > 0) |
64f1c21e JA |
3777 | nsecs = q->poll_nsec; |
3778 | else | |
cae740a0 | 3779 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3780 | |
3781 | if (!nsecs) | |
06426adf JA |
3782 | return false; |
3783 | ||
76a86f9d | 3784 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3785 | |
3786 | /* | |
3787 | * This will be replaced with the stats tracking code, using | |
3788 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3789 | */ | |
8b0e1953 | 3790 | kt = nsecs; |
06426adf JA |
3791 | |
3792 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3793 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3794 | hrtimer_set_expires(&hs.timer, kt); |
3795 | ||
06426adf | 3796 | do { |
5a61c363 | 3797 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3798 | break; |
3799 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3800 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3801 | if (hs.task) |
3802 | io_schedule(); | |
3803 | hrtimer_cancel(&hs.timer); | |
3804 | mode = HRTIMER_MODE_ABS; | |
3805 | } while (hs.task && !signal_pending(current)); | |
3806 | ||
3807 | __set_current_state(TASK_RUNNING); | |
3808 | destroy_hrtimer_on_stack(&hs.timer); | |
3809 | return true; | |
3810 | } | |
3811 | ||
1052b8ac JA |
3812 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3813 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3814 | { |
1052b8ac JA |
3815 | struct request *rq; |
3816 | ||
29ece8b4 | 3817 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3818 | return false; |
3819 | ||
3820 | if (!blk_qc_t_is_internal(cookie)) | |
3821 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3822 | else { | |
3823 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3824 | /* | |
3825 | * With scheduling, if the request has completed, we'll | |
3826 | * get a NULL return here, as we clear the sched tag when | |
3827 | * that happens. The request still remains valid, like always, | |
3828 | * so we should be safe with just the NULL check. | |
3829 | */ | |
3830 | if (!rq) | |
3831 | return false; | |
3832 | } | |
3833 | ||
cae740a0 | 3834 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3835 | } |
3836 | ||
529262d5 CH |
3837 | /** |
3838 | * blk_poll - poll for IO completions | |
3839 | * @q: the queue | |
3840 | * @cookie: cookie passed back at IO submission time | |
3841 | * @spin: whether to spin for completions | |
3842 | * | |
3843 | * Description: | |
3844 | * Poll for completions on the passed in queue. Returns number of | |
3845 | * completed entries found. If @spin is true, then blk_poll will continue | |
3846 | * looping until at least one completion is found, unless the task is | |
3847 | * otherwise marked running (or we need to reschedule). | |
3848 | */ | |
3849 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3850 | { |
3851 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3852 | long state; |
3853 | ||
529262d5 CH |
3854 | if (!blk_qc_t_valid(cookie) || |
3855 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3856 | return 0; |
3857 | ||
529262d5 CH |
3858 | if (current->plug) |
3859 | blk_flush_plug_list(current->plug, false); | |
3860 | ||
1052b8ac JA |
3861 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3862 | ||
06426adf JA |
3863 | /* |
3864 | * If we sleep, have the caller restart the poll loop to reset | |
3865 | * the state. Like for the other success return cases, the | |
3866 | * caller is responsible for checking if the IO completed. If | |
3867 | * the IO isn't complete, we'll get called again and will go | |
3868 | * straight to the busy poll loop. | |
3869 | */ | |
1052b8ac | 3870 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3871 | return 1; |
06426adf | 3872 | |
bbd7bb70 JA |
3873 | hctx->poll_considered++; |
3874 | ||
3875 | state = current->state; | |
aa61bec3 | 3876 | do { |
bbd7bb70 JA |
3877 | int ret; |
3878 | ||
3879 | hctx->poll_invoked++; | |
3880 | ||
9743139c | 3881 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3882 | if (ret > 0) { |
3883 | hctx->poll_success++; | |
849a3700 | 3884 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3885 | return ret; |
bbd7bb70 JA |
3886 | } |
3887 | ||
3888 | if (signal_pending_state(state, current)) | |
849a3700 | 3889 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3890 | |
3891 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3892 | return 1; |
0a1b8b87 | 3893 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3894 | break; |
3895 | cpu_relax(); | |
aa61bec3 | 3896 | } while (!need_resched()); |
bbd7bb70 | 3897 | |
67b4110f | 3898 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3899 | return 0; |
bbd7bb70 | 3900 | } |
529262d5 | 3901 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3902 | |
9cf2bab6 JA |
3903 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3904 | { | |
3905 | return rq->mq_ctx->cpu; | |
3906 | } | |
3907 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3908 | ||
320ae51f JA |
3909 | static int __init blk_mq_init(void) |
3910 | { | |
c3077b5d CH |
3911 | int i; |
3912 | ||
3913 | for_each_possible_cpu(i) | |
3914 | INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i)); | |
3915 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); | |
3916 | ||
3917 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
3918 | "block/softirq:dead", NULL, | |
3919 | blk_softirq_cpu_dead); | |
9467f859 TG |
3920 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3921 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
3922 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
3923 | blk_mq_hctx_notify_online, | |
3924 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
3925 | return 0; |
3926 | } | |
3927 | subsys_initcall(blk_mq_init); |