]>
Commit | Line | Data |
---|---|---|
3dcf60bc | 1 | // SPDX-License-Identifier: GPL-2.0 |
75bb4625 JA |
2 | /* |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
320ae51f JA |
8 | #include <linux/kernel.h> |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
f75782e4 | 13 | #include <linux/kmemleak.h> |
320ae51f JA |
14 | #include <linux/mm.h> |
15 | #include <linux/init.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/smp.h> | |
19 | #include <linux/llist.h> | |
20 | #include <linux/list_sort.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/cache.h> | |
23 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 24 | #include <linux/sched/topology.h> |
174cd4b1 | 25 | #include <linux/sched/signal.h> |
320ae51f | 26 | #include <linux/delay.h> |
aedcd72f | 27 | #include <linux/crash_dump.h> |
88c7b2b7 | 28 | #include <linux/prefetch.h> |
a892c8d5 | 29 | #include <linux/blk-crypto.h> |
320ae51f JA |
30 | |
31 | #include <trace/events/block.h> | |
32 | ||
33 | #include <linux/blk-mq.h> | |
54d4e6ab | 34 | #include <linux/t10-pi.h> |
320ae51f JA |
35 | #include "blk.h" |
36 | #include "blk-mq.h" | |
9c1051aa | 37 | #include "blk-mq-debugfs.h" |
320ae51f | 38 | #include "blk-mq-tag.h" |
986d413b | 39 | #include "blk-pm.h" |
cf43e6be | 40 | #include "blk-stat.h" |
bd166ef1 | 41 | #include "blk-mq-sched.h" |
c1c80384 | 42 | #include "blk-rq-qos.h" |
320ae51f | 43 | |
f9ab4918 | 44 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 45 | |
34dbad5d OS |
46 | static void blk_mq_poll_stats_start(struct request_queue *q); |
47 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
48 | ||
720b8ccc SB |
49 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
50 | { | |
3d244306 | 51 | int ddir, sectors, bucket; |
720b8ccc | 52 | |
99c749a4 | 53 | ddir = rq_data_dir(rq); |
3d244306 | 54 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 55 | |
3d244306 | 56 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
57 | |
58 | if (bucket < 0) | |
59 | return -1; | |
60 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
61 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
62 | ||
63 | return bucket; | |
64 | } | |
65 | ||
320ae51f | 66 | /* |
85fae294 YY |
67 | * Check if any of the ctx, dispatch list or elevator |
68 | * have pending work in this hardware queue. | |
320ae51f | 69 | */ |
79f720a7 | 70 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 71 | { |
79f720a7 JA |
72 | return !list_empty_careful(&hctx->dispatch) || |
73 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 74 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
75 | } |
76 | ||
320ae51f JA |
77 | /* |
78 | * Mark this ctx as having pending work in this hardware queue | |
79 | */ | |
80 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
f31967f0 JA |
83 | const int bit = ctx->index_hw[hctx->type]; |
84 | ||
85 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
86 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
87 | } |
88 | ||
89 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
90 | struct blk_mq_ctx *ctx) | |
91 | { | |
f31967f0 JA |
92 | const int bit = ctx->index_hw[hctx->type]; |
93 | ||
94 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
95 | } |
96 | ||
f299b7c7 | 97 | struct mq_inflight { |
8446fe92 | 98 | struct block_device *part; |
a2e80f6f | 99 | unsigned int inflight[2]; |
f299b7c7 JA |
100 | }; |
101 | ||
7baa8572 | 102 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
103 | struct request *rq, void *priv, |
104 | bool reserved) | |
105 | { | |
106 | struct mq_inflight *mi = priv; | |
107 | ||
b0d97557 JX |
108 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
109 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 110 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
111 | |
112 | return true; | |
f299b7c7 JA |
113 | } |
114 | ||
8446fe92 CH |
115 | unsigned int blk_mq_in_flight(struct request_queue *q, |
116 | struct block_device *part) | |
f299b7c7 | 117 | { |
a2e80f6f | 118 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 119 | |
f299b7c7 | 120 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 121 | |
a2e80f6f | 122 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
123 | } |
124 | ||
8446fe92 CH |
125 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
126 | unsigned int inflight[2]) | |
bf0ddaba | 127 | { |
a2e80f6f | 128 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 129 | |
bb4e6b14 | 130 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
131 | inflight[0] = mi.inflight[0]; |
132 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
133 | } |
134 | ||
1671d522 | 135 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 136 | { |
7996a8b5 BL |
137 | mutex_lock(&q->mq_freeze_lock); |
138 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 139 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 140 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 141 | if (queue_is_mq(q)) |
055f6e18 | 142 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
143 | } else { |
144 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 145 | } |
f3af020b | 146 | } |
1671d522 | 147 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 148 | |
6bae363e | 149 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 150 | { |
3ef28e83 | 151 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 152 | } |
6bae363e | 153 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 154 | |
f91328c4 KB |
155 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
156 | unsigned long timeout) | |
157 | { | |
158 | return wait_event_timeout(q->mq_freeze_wq, | |
159 | percpu_ref_is_zero(&q->q_usage_counter), | |
160 | timeout); | |
161 | } | |
162 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 163 | |
f3af020b TH |
164 | /* |
165 | * Guarantee no request is in use, so we can change any data structure of | |
166 | * the queue afterward. | |
167 | */ | |
3ef28e83 | 168 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 169 | { |
3ef28e83 DW |
170 | /* |
171 | * In the !blk_mq case we are only calling this to kill the | |
172 | * q_usage_counter, otherwise this increases the freeze depth | |
173 | * and waits for it to return to zero. For this reason there is | |
174 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
175 | * exported to drivers as the only user for unfreeze is blk_mq. | |
176 | */ | |
1671d522 | 177 | blk_freeze_queue_start(q); |
f3af020b TH |
178 | blk_mq_freeze_queue_wait(q); |
179 | } | |
3ef28e83 DW |
180 | |
181 | void blk_mq_freeze_queue(struct request_queue *q) | |
182 | { | |
183 | /* | |
184 | * ...just an alias to keep freeze and unfreeze actions balanced | |
185 | * in the blk_mq_* namespace | |
186 | */ | |
187 | blk_freeze_queue(q); | |
188 | } | |
c761d96b | 189 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 190 | |
b4c6a028 | 191 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 192 | { |
7996a8b5 BL |
193 | mutex_lock(&q->mq_freeze_lock); |
194 | q->mq_freeze_depth--; | |
195 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
196 | if (!q->mq_freeze_depth) { | |
bdd63160 | 197 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 198 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 199 | } |
7996a8b5 | 200 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 201 | } |
b4c6a028 | 202 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 203 | |
852ec809 BVA |
204 | /* |
205 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
206 | * mpt3sas driver such that this function can be removed. | |
207 | */ | |
208 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
209 | { | |
8814ce8a | 210 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
211 | } |
212 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
213 | ||
6a83e74d | 214 | /** |
69e07c4a | 215 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
216 | * @q: request queue. |
217 | * | |
218 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
219 | * callback function is invoked. Once this function is returned, we make |
220 | * sure no dispatch can happen until the queue is unquiesced via | |
221 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
222 | */ |
223 | void blk_mq_quiesce_queue(struct request_queue *q) | |
224 | { | |
225 | struct blk_mq_hw_ctx *hctx; | |
226 | unsigned int i; | |
227 | bool rcu = false; | |
228 | ||
1d9e9bc6 | 229 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 230 | |
6a83e74d BVA |
231 | queue_for_each_hw_ctx(q, hctx, i) { |
232 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 233 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
234 | else |
235 | rcu = true; | |
236 | } | |
237 | if (rcu) | |
238 | synchronize_rcu(); | |
239 | } | |
240 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
241 | ||
e4e73913 ML |
242 | /* |
243 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
244 | * @q: request queue. | |
245 | * | |
246 | * This function recovers queue into the state before quiescing | |
247 | * which is done by blk_mq_quiesce_queue. | |
248 | */ | |
249 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
250 | { | |
8814ce8a | 251 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 252 | |
1d9e9bc6 ML |
253 | /* dispatch requests which are inserted during quiescing */ |
254 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
255 | } |
256 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
257 | ||
aed3ea94 JA |
258 | void blk_mq_wake_waiters(struct request_queue *q) |
259 | { | |
260 | struct blk_mq_hw_ctx *hctx; | |
261 | unsigned int i; | |
262 | ||
263 | queue_for_each_hw_ctx(q, hctx, i) | |
264 | if (blk_mq_hw_queue_mapped(hctx)) | |
265 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
266 | } | |
267 | ||
fe1f4526 | 268 | /* |
9a91b05b HT |
269 | * Only need start/end time stamping if we have iostat or |
270 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
271 | */ |
272 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
273 | { | |
9a91b05b | 274 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
275 | } |
276 | ||
e4cdf1a1 | 277 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 278 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 279 | { |
e4cdf1a1 CH |
280 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
281 | struct request *rq = tags->static_rqs[tag]; | |
c3a148d2 | 282 | |
42fdc5e4 | 283 | if (data->q->elevator) { |
76647368 | 284 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
285 | rq->internal_tag = tag; |
286 | } else { | |
e4cdf1a1 | 287 | rq->tag = tag; |
76647368 | 288 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
289 | } |
290 | ||
af76e555 | 291 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
292 | rq->q = data->q; |
293 | rq->mq_ctx = data->ctx; | |
ea4f995e | 294 | rq->mq_hctx = data->hctx; |
568f2700 | 295 | rq->rq_flags = 0; |
7ea4d8a4 | 296 | rq->cmd_flags = data->cmd_flags; |
0854bcdc BVA |
297 | if (data->flags & BLK_MQ_REQ_PM) |
298 | rq->rq_flags |= RQF_PM; | |
e4cdf1a1 | 299 | if (blk_queue_io_stat(data->q)) |
e8064021 | 300 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 301 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
302 | INIT_HLIST_NODE(&rq->hash); |
303 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
304 | rq->rq_disk = NULL; |
305 | rq->part = NULL; | |
6f816b4b TH |
306 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
307 | rq->alloc_time_ns = alloc_time_ns; | |
308 | #endif | |
fe1f4526 JA |
309 | if (blk_mq_need_time_stamp(rq)) |
310 | rq->start_time_ns = ktime_get_ns(); | |
311 | else | |
312 | rq->start_time_ns = 0; | |
544ccc8d | 313 | rq->io_start_time_ns = 0; |
3d244306 | 314 | rq->stats_sectors = 0; |
af76e555 CH |
315 | rq->nr_phys_segments = 0; |
316 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
317 | rq->nr_integrity_segments = 0; | |
318 | #endif | |
a892c8d5 | 319 | blk_crypto_rq_set_defaults(rq); |
af76e555 | 320 | /* tag was already set */ |
079076b3 | 321 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 322 | |
f6be4fb4 JA |
323 | rq->timeout = 0; |
324 | ||
af76e555 CH |
325 | rq->end_io = NULL; |
326 | rq->end_io_data = NULL; | |
af76e555 | 327 | |
7ea4d8a4 | 328 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; |
12f5b931 | 329 | refcount_set(&rq->ref, 1); |
7ea4d8a4 CH |
330 | |
331 | if (!op_is_flush(data->cmd_flags)) { | |
332 | struct elevator_queue *e = data->q->elevator; | |
333 | ||
334 | rq->elv.icq = NULL; | |
335 | if (e && e->type->ops.prepare_request) { | |
336 | if (e->type->icq_cache) | |
337 | blk_mq_sched_assign_ioc(rq); | |
338 | ||
339 | e->type->ops.prepare_request(rq); | |
340 | rq->rq_flags |= RQF_ELVPRIV; | |
341 | } | |
342 | } | |
343 | ||
344 | data->hctx->queued++; | |
e4cdf1a1 | 345 | return rq; |
5dee8577 CH |
346 | } |
347 | ||
e6e7abff | 348 | static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data) |
d2c0d383 | 349 | { |
e6e7abff | 350 | struct request_queue *q = data->q; |
d2c0d383 | 351 | struct elevator_queue *e = q->elevator; |
6f816b4b | 352 | u64 alloc_time_ns = 0; |
600c3b0c | 353 | unsigned int tag; |
d2c0d383 | 354 | |
6f816b4b TH |
355 | /* alloc_time includes depth and tag waits */ |
356 | if (blk_queue_rq_alloc_time(q)) | |
357 | alloc_time_ns = ktime_get_ns(); | |
358 | ||
f9afca4d | 359 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 360 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
361 | |
362 | if (e) { | |
d2c0d383 | 363 | /* |
8d663f34 | 364 | * Flush/passthrough 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 | 368 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 369 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 370 | e->type->ops.limit_depth && |
17a51199 | 371 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 372 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
373 | } |
374 | ||
bf0beec0 | 375 | retry: |
600c3b0c CH |
376 | data->ctx = blk_mq_get_ctx(q); |
377 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 378 | if (!e) |
600c3b0c CH |
379 | blk_mq_tag_busy(data->hctx); |
380 | ||
bf0beec0 ML |
381 | /* |
382 | * Waiting allocations only fail because of an inactive hctx. In that | |
383 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
384 | * should have migrated us to an online CPU by now. | |
385 | */ | |
e4cdf1a1 | 386 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
387 | if (tag == BLK_MQ_NO_TAG) { |
388 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
389 | return NULL; | |
390 | ||
391 | /* | |
392 | * Give up the CPU and sleep for a random short time to ensure | |
393 | * that thread using a realtime scheduling class are migrated | |
70f15a4f | 394 | * off the CPU, and thus off the hctx that is going away. |
bf0beec0 ML |
395 | */ |
396 | msleep(3); | |
397 | goto retry; | |
398 | } | |
7ea4d8a4 | 399 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
400 | } |
401 | ||
cd6ce148 | 402 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 403 | blk_mq_req_flags_t flags) |
320ae51f | 404 | { |
e6e7abff CH |
405 | struct blk_mq_alloc_data data = { |
406 | .q = q, | |
407 | .flags = flags, | |
408 | .cmd_flags = op, | |
409 | }; | |
bd166ef1 | 410 | struct request *rq; |
a492f075 | 411 | int ret; |
320ae51f | 412 | |
3a0a5299 | 413 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
414 | if (ret) |
415 | return ERR_PTR(ret); | |
320ae51f | 416 | |
e6e7abff | 417 | rq = __blk_mq_alloc_request(&data); |
bd166ef1 | 418 | if (!rq) |
a5ea5811 | 419 | goto out_queue_exit; |
0c4de0f3 CH |
420 | rq->__data_len = 0; |
421 | rq->__sector = (sector_t) -1; | |
422 | rq->bio = rq->biotail = NULL; | |
320ae51f | 423 | return rq; |
a5ea5811 CH |
424 | out_queue_exit: |
425 | blk_queue_exit(q); | |
426 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 427 | } |
4bb659b1 | 428 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 429 | |
cd6ce148 | 430 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 431 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 432 | { |
e6e7abff CH |
433 | struct blk_mq_alloc_data data = { |
434 | .q = q, | |
435 | .flags = flags, | |
436 | .cmd_flags = op, | |
437 | }; | |
600c3b0c | 438 | u64 alloc_time_ns = 0; |
6d2809d5 | 439 | unsigned int cpu; |
600c3b0c | 440 | unsigned int tag; |
1f5bd336 ML |
441 | int ret; |
442 | ||
600c3b0c CH |
443 | /* alloc_time includes depth and tag waits */ |
444 | if (blk_queue_rq_alloc_time(q)) | |
445 | alloc_time_ns = ktime_get_ns(); | |
446 | ||
1f5bd336 ML |
447 | /* |
448 | * If the tag allocator sleeps we could get an allocation for a | |
449 | * different hardware context. No need to complicate the low level | |
450 | * allocator for this for the rare use case of a command tied to | |
451 | * a specific queue. | |
452 | */ | |
600c3b0c | 453 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
454 | return ERR_PTR(-EINVAL); |
455 | ||
456 | if (hctx_idx >= q->nr_hw_queues) | |
457 | return ERR_PTR(-EIO); | |
458 | ||
3a0a5299 | 459 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
460 | if (ret) |
461 | return ERR_PTR(ret); | |
462 | ||
c8712c6a CH |
463 | /* |
464 | * Check if the hardware context is actually mapped to anything. | |
465 | * If not tell the caller that it should skip this queue. | |
466 | */ | |
a5ea5811 | 467 | ret = -EXDEV; |
e6e7abff CH |
468 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
469 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 470 | goto out_queue_exit; |
e6e7abff CH |
471 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
472 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 473 | |
42fdc5e4 | 474 | if (!q->elevator) |
600c3b0c CH |
475 | blk_mq_tag_busy(data.hctx); |
476 | ||
a5ea5811 | 477 | ret = -EWOULDBLOCK; |
600c3b0c CH |
478 | tag = blk_mq_get_tag(&data); |
479 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 480 | goto out_queue_exit; |
600c3b0c CH |
481 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
482 | ||
a5ea5811 CH |
483 | out_queue_exit: |
484 | blk_queue_exit(q); | |
485 | return ERR_PTR(ret); | |
1f5bd336 ML |
486 | } |
487 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
488 | ||
12f5b931 KB |
489 | static void __blk_mq_free_request(struct request *rq) |
490 | { | |
491 | struct request_queue *q = rq->q; | |
492 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 493 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
494 | const int sched_tag = rq->internal_tag; |
495 | ||
a892c8d5 | 496 | blk_crypto_free_request(rq); |
986d413b | 497 | blk_pm_mark_last_busy(rq); |
ea4f995e | 498 | rq->mq_hctx = NULL; |
76647368 | 499 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 500 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 501 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 502 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
503 | blk_mq_sched_restart(hctx); |
504 | blk_queue_exit(q); | |
505 | } | |
506 | ||
6af54051 | 507 | void blk_mq_free_request(struct request *rq) |
320ae51f | 508 | { |
320ae51f | 509 | struct request_queue *q = rq->q; |
6af54051 CH |
510 | struct elevator_queue *e = q->elevator; |
511 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 512 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 513 | |
5bbf4e5a | 514 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
515 | if (e && e->type->ops.finish_request) |
516 | e->type->ops.finish_request(rq); | |
6af54051 CH |
517 | if (rq->elv.icq) { |
518 | put_io_context(rq->elv.icq->ioc); | |
519 | rq->elv.icq = NULL; | |
520 | } | |
521 | } | |
320ae51f | 522 | |
6af54051 | 523 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 524 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 525 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 526 | |
7beb2f84 | 527 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 528 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 529 | |
a7905043 | 530 | rq_qos_done(q, rq); |
0d2602ca | 531 | |
12f5b931 KB |
532 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
533 | if (refcount_dec_and_test(&rq->ref)) | |
534 | __blk_mq_free_request(rq); | |
320ae51f | 535 | } |
1a3b595a | 536 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 537 | |
2a842aca | 538 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 539 | { |
fe1f4526 JA |
540 | u64 now = 0; |
541 | ||
542 | if (blk_mq_need_time_stamp(rq)) | |
543 | now = ktime_get_ns(); | |
522a7775 | 544 | |
4bc6339a OS |
545 | if (rq->rq_flags & RQF_STATS) { |
546 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 547 | blk_stat_add(rq, now); |
4bc6339a OS |
548 | } |
549 | ||
87890092 | 550 | blk_mq_sched_completed_request(rq, now); |
ed88660a | 551 | |
522a7775 | 552 | blk_account_io_done(rq, now); |
0d11e6ac | 553 | |
91b63639 | 554 | if (rq->end_io) { |
a7905043 | 555 | rq_qos_done(rq->q, rq); |
320ae51f | 556 | rq->end_io(rq, error); |
91b63639 | 557 | } else { |
320ae51f | 558 | blk_mq_free_request(rq); |
91b63639 | 559 | } |
320ae51f | 560 | } |
c8a446ad | 561 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 562 | |
2a842aca | 563 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
564 | { |
565 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
566 | BUG(); | |
c8a446ad | 567 | __blk_mq_end_request(rq, error); |
63151a44 | 568 | } |
c8a446ad | 569 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 570 | |
f9ab4918 | 571 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 572 | { |
f9ab4918 SAS |
573 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
574 | struct request *rq, *next; | |
c3077b5d | 575 | |
f9ab4918 | 576 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 577 | rq->q->mq_ops->complete(rq); |
320ae51f | 578 | } |
320ae51f | 579 | |
f9ab4918 | 580 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 581 | { |
f9ab4918 | 582 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
583 | } |
584 | ||
c3077b5d CH |
585 | static int blk_softirq_cpu_dead(unsigned int cpu) |
586 | { | |
f9ab4918 | 587 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
588 | return 0; |
589 | } | |
590 | ||
40d09b53 | 591 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 592 | { |
f9ab4918 | 593 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
594 | } |
595 | ||
96339526 CH |
596 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
597 | { | |
598 | int cpu = raw_smp_processor_id(); | |
599 | ||
600 | if (!IS_ENABLED(CONFIG_SMP) || | |
601 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
602 | return false; | |
71425189 SAS |
603 | /* |
604 | * With force threaded interrupts enabled, raising softirq from an SMP | |
605 | * function call will always result in waking the ksoftirqd thread. | |
606 | * This is probably worse than completing the request on a different | |
607 | * cache domain. | |
608 | */ | |
91cc470e | 609 | if (force_irqthreads()) |
71425189 | 610 | return false; |
96339526 CH |
611 | |
612 | /* same CPU or cache domain? Complete locally */ | |
613 | if (cpu == rq->mq_ctx->cpu || | |
614 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
615 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
616 | return false; | |
617 | ||
618 | /* don't try to IPI to an offline CPU */ | |
619 | return cpu_online(rq->mq_ctx->cpu); | |
620 | } | |
621 | ||
f9ab4918 SAS |
622 | static void blk_mq_complete_send_ipi(struct request *rq) |
623 | { | |
624 | struct llist_head *list; | |
625 | unsigned int cpu; | |
626 | ||
627 | cpu = rq->mq_ctx->cpu; | |
628 | list = &per_cpu(blk_cpu_done, cpu); | |
629 | if (llist_add(&rq->ipi_list, list)) { | |
630 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
631 | smp_call_function_single_async(cpu, &rq->csd); | |
632 | } | |
633 | } | |
634 | ||
635 | static void blk_mq_raise_softirq(struct request *rq) | |
636 | { | |
637 | struct llist_head *list; | |
638 | ||
639 | preempt_disable(); | |
640 | list = this_cpu_ptr(&blk_cpu_done); | |
641 | if (llist_add(&rq->ipi_list, list)) | |
642 | raise_softirq(BLOCK_SOFTIRQ); | |
643 | preempt_enable(); | |
644 | } | |
645 | ||
40d09b53 | 646 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 647 | { |
af78ff7c | 648 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 649 | |
4ab32bf3 JA |
650 | /* |
651 | * For a polled request, always complete locallly, it's pointless | |
652 | * to redirect the completion. | |
653 | */ | |
40d09b53 CH |
654 | if (rq->cmd_flags & REQ_HIPRI) |
655 | return false; | |
38535201 | 656 | |
96339526 | 657 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
658 | blk_mq_complete_send_ipi(rq); |
659 | return true; | |
3d6efbf6 | 660 | } |
40d09b53 | 661 | |
f9ab4918 SAS |
662 | if (rq->q->nr_hw_queues == 1) { |
663 | blk_mq_raise_softirq(rq); | |
664 | return true; | |
665 | } | |
666 | return false; | |
40d09b53 CH |
667 | } |
668 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
669 | ||
670 | /** | |
671 | * blk_mq_complete_request - end I/O on a request | |
672 | * @rq: the request being processed | |
673 | * | |
674 | * Description: | |
675 | * Complete a request by scheduling the ->complete_rq operation. | |
676 | **/ | |
677 | void blk_mq_complete_request(struct request *rq) | |
678 | { | |
679 | if (!blk_mq_complete_request_remote(rq)) | |
680 | rq->q->mq_ops->complete(rq); | |
320ae51f | 681 | } |
15f73f5b | 682 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 683 | |
04ced159 | 684 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 685 | __releases(hctx->srcu) |
04ced159 JA |
686 | { |
687 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
688 | rcu_read_unlock(); | |
689 | else | |
05707b64 | 690 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
691 | } |
692 | ||
693 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 694 | __acquires(hctx->srcu) |
04ced159 | 695 | { |
08b5a6e2 JA |
696 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
697 | /* shut up gcc false positive */ | |
698 | *srcu_idx = 0; | |
04ced159 | 699 | rcu_read_lock(); |
08b5a6e2 | 700 | } else |
05707b64 | 701 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
702 | } |
703 | ||
105663f7 AA |
704 | /** |
705 | * blk_mq_start_request - Start processing a request | |
706 | * @rq: Pointer to request to be started | |
707 | * | |
708 | * Function used by device drivers to notify the block layer that a request | |
709 | * is going to be processed now, so blk layer can do proper initializations | |
710 | * such as starting the timeout timer. | |
711 | */ | |
e2490073 | 712 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
713 | { |
714 | struct request_queue *q = rq->q; | |
715 | ||
a54895fa | 716 | trace_block_rq_issue(rq); |
320ae51f | 717 | |
cf43e6be | 718 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 719 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 720 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 721 | rq->rq_flags |= RQF_STATS; |
a7905043 | 722 | rq_qos_issue(q, rq); |
cf43e6be JA |
723 | } |
724 | ||
1d9bd516 | 725 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 726 | |
1d9bd516 | 727 | blk_add_timer(rq); |
12f5b931 | 728 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 729 | |
54d4e6ab MG |
730 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
731 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
732 | q->integrity.profile->prepare_fn(rq); | |
733 | #endif | |
320ae51f | 734 | } |
e2490073 | 735 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 736 | |
ed0791b2 | 737 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
738 | { |
739 | struct request_queue *q = rq->q; | |
740 | ||
923218f6 ML |
741 | blk_mq_put_driver_tag(rq); |
742 | ||
a54895fa | 743 | trace_block_rq_requeue(rq); |
a7905043 | 744 | rq_qos_requeue(q, rq); |
49f5baa5 | 745 | |
12f5b931 KB |
746 | if (blk_mq_request_started(rq)) { |
747 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 748 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 749 | } |
320ae51f JA |
750 | } |
751 | ||
2b053aca | 752 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 753 | { |
ed0791b2 | 754 | __blk_mq_requeue_request(rq); |
ed0791b2 | 755 | |
105976f5 ML |
756 | /* this request will be re-inserted to io scheduler queue */ |
757 | blk_mq_sched_requeue_request(rq); | |
758 | ||
7d692330 | 759 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 760 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
761 | } |
762 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
763 | ||
6fca6a61 CH |
764 | static void blk_mq_requeue_work(struct work_struct *work) |
765 | { | |
766 | struct request_queue *q = | |
2849450a | 767 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
768 | LIST_HEAD(rq_list); |
769 | struct request *rq, *next; | |
6fca6a61 | 770 | |
18e9781d | 771 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 772 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 773 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
774 | |
775 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 776 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
777 | continue; |
778 | ||
e8064021 | 779 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 780 | list_del_init(&rq->queuelist); |
aef1897c JW |
781 | /* |
782 | * If RQF_DONTPREP, rq has contained some driver specific | |
783 | * data, so insert it to hctx dispatch list to avoid any | |
784 | * merge. | |
785 | */ | |
786 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 787 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
788 | else |
789 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
790 | } |
791 | ||
792 | while (!list_empty(&rq_list)) { | |
793 | rq = list_entry(rq_list.next, struct request, queuelist); | |
794 | list_del_init(&rq->queuelist); | |
9e97d295 | 795 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
796 | } |
797 | ||
52d7f1b5 | 798 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
799 | } |
800 | ||
2b053aca BVA |
801 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
802 | bool kick_requeue_list) | |
6fca6a61 CH |
803 | { |
804 | struct request_queue *q = rq->q; | |
805 | unsigned long flags; | |
806 | ||
807 | /* | |
808 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 809 | * request head insertion from the workqueue. |
6fca6a61 | 810 | */ |
e8064021 | 811 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
812 | |
813 | spin_lock_irqsave(&q->requeue_lock, flags); | |
814 | if (at_head) { | |
e8064021 | 815 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
816 | list_add(&rq->queuelist, &q->requeue_list); |
817 | } else { | |
818 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
819 | } | |
820 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
821 | |
822 | if (kick_requeue_list) | |
823 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 824 | } |
6fca6a61 CH |
825 | |
826 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
827 | { | |
ae943d20 | 828 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
829 | } |
830 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
831 | ||
2849450a MS |
832 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
833 | unsigned long msecs) | |
834 | { | |
d4acf365 BVA |
835 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
836 | msecs_to_jiffies(msecs)); | |
2849450a MS |
837 | } |
838 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
839 | ||
0e62f51f JA |
840 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
841 | { | |
88c7b2b7 JA |
842 | if (tag < tags->nr_tags) { |
843 | prefetch(tags->rqs[tag]); | |
4ee86bab | 844 | return tags->rqs[tag]; |
88c7b2b7 | 845 | } |
4ee86bab HR |
846 | |
847 | return NULL; | |
24d2f903 CH |
848 | } |
849 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
850 | ||
3c94d83c JA |
851 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
852 | void *priv, bool reserved) | |
ae879912 JA |
853 | { |
854 | /* | |
05a4fed6 | 855 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 856 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 857 | */ |
05a4fed6 | 858 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
859 | bool *busy = priv; |
860 | ||
861 | *busy = true; | |
862 | return false; | |
863 | } | |
864 | ||
865 | return true; | |
866 | } | |
867 | ||
3c94d83c | 868 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
869 | { |
870 | bool busy = false; | |
871 | ||
3c94d83c | 872 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
873 | return busy; |
874 | } | |
3c94d83c | 875 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 876 | |
358f70da | 877 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 878 | { |
da661267 | 879 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
880 | if (req->q->mq_ops->timeout) { |
881 | enum blk_eh_timer_return ret; | |
882 | ||
883 | ret = req->q->mq_ops->timeout(req, reserved); | |
884 | if (ret == BLK_EH_DONE) | |
885 | return; | |
886 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 887 | } |
d1210d5a CH |
888 | |
889 | blk_add_timer(req); | |
87ee7b11 | 890 | } |
5b3f25fc | 891 | |
12f5b931 | 892 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 893 | { |
12f5b931 | 894 | unsigned long deadline; |
87ee7b11 | 895 | |
12f5b931 KB |
896 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
897 | return false; | |
da661267 CH |
898 | if (rq->rq_flags & RQF_TIMED_OUT) |
899 | return false; | |
a7af0af3 | 900 | |
079076b3 | 901 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
902 | if (time_after_eq(jiffies, deadline)) |
903 | return true; | |
a7af0af3 | 904 | |
12f5b931 KB |
905 | if (*next == 0) |
906 | *next = deadline; | |
907 | else if (time_after(*next, deadline)) | |
908 | *next = deadline; | |
909 | return false; | |
87ee7b11 JA |
910 | } |
911 | ||
2e315dc0 ML |
912 | void blk_mq_put_rq_ref(struct request *rq) |
913 | { | |
a9ed27a7 | 914 | if (is_flush_rq(rq)) |
2e315dc0 ML |
915 | rq->end_io(rq, 0); |
916 | else if (refcount_dec_and_test(&rq->ref)) | |
917 | __blk_mq_free_request(rq); | |
918 | } | |
919 | ||
7baa8572 | 920 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
921 | struct request *rq, void *priv, bool reserved) |
922 | { | |
12f5b931 KB |
923 | unsigned long *next = priv; |
924 | ||
925 | /* | |
c797b40c ML |
926 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
927 | * be reallocated underneath the timeout handler's processing, then | |
928 | * the expire check is reliable. If the request is not expired, then | |
929 | * it was completed and reallocated as a new request after returning | |
930 | * from blk_mq_check_expired(). | |
1d9bd516 | 931 | */ |
12f5b931 | 932 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 933 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 934 | return true; |
1d9bd516 TH |
935 | } |
936 | ||
287922eb | 937 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 938 | { |
287922eb CH |
939 | struct request_queue *q = |
940 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 941 | unsigned long next = 0; |
1d9bd516 | 942 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 943 | int i; |
320ae51f | 944 | |
71f79fb3 GKB |
945 | /* A deadlock might occur if a request is stuck requiring a |
946 | * timeout at the same time a queue freeze is waiting | |
947 | * completion, since the timeout code would not be able to | |
948 | * acquire the queue reference here. | |
949 | * | |
950 | * That's why we don't use blk_queue_enter here; instead, we use | |
951 | * percpu_ref_tryget directly, because we need to be able to | |
952 | * obtain a reference even in the short window between the queue | |
953 | * starting to freeze, by dropping the first reference in | |
1671d522 | 954 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
955 | * consumed, marked by the instant q_usage_counter reaches |
956 | * zero. | |
957 | */ | |
958 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
959 | return; |
960 | ||
12f5b931 | 961 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 962 | |
12f5b931 KB |
963 | if (next != 0) { |
964 | mod_timer(&q->timeout, next); | |
0d2602ca | 965 | } else { |
fcd36c36 BVA |
966 | /* |
967 | * Request timeouts are handled as a forward rolling timer. If | |
968 | * we end up here it means that no requests are pending and | |
969 | * also that no request has been pending for a while. Mark | |
970 | * each hctx as idle. | |
971 | */ | |
f054b56c ML |
972 | queue_for_each_hw_ctx(q, hctx, i) { |
973 | /* the hctx may be unmapped, so check it here */ | |
974 | if (blk_mq_hw_queue_mapped(hctx)) | |
975 | blk_mq_tag_idle(hctx); | |
976 | } | |
0d2602ca | 977 | } |
287922eb | 978 | blk_queue_exit(q); |
320ae51f JA |
979 | } |
980 | ||
88459642 OS |
981 | struct flush_busy_ctx_data { |
982 | struct blk_mq_hw_ctx *hctx; | |
983 | struct list_head *list; | |
984 | }; | |
985 | ||
986 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
987 | { | |
988 | struct flush_busy_ctx_data *flush_data = data; | |
989 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
990 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 991 | enum hctx_type type = hctx->type; |
88459642 | 992 | |
88459642 | 993 | spin_lock(&ctx->lock); |
c16d6b5a | 994 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 995 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
996 | spin_unlock(&ctx->lock); |
997 | return true; | |
998 | } | |
999 | ||
1429d7c9 JA |
1000 | /* |
1001 | * Process software queues that have been marked busy, splicing them | |
1002 | * to the for-dispatch | |
1003 | */ | |
2c3ad667 | 1004 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1005 | { |
88459642 OS |
1006 | struct flush_busy_ctx_data data = { |
1007 | .hctx = hctx, | |
1008 | .list = list, | |
1009 | }; | |
1429d7c9 | 1010 | |
88459642 | 1011 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1012 | } |
2c3ad667 | 1013 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1014 | |
b347689f ML |
1015 | struct dispatch_rq_data { |
1016 | struct blk_mq_hw_ctx *hctx; | |
1017 | struct request *rq; | |
1018 | }; | |
1019 | ||
1020 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1021 | void *data) | |
1022 | { | |
1023 | struct dispatch_rq_data *dispatch_data = data; | |
1024 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1025 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1026 | enum hctx_type type = hctx->type; |
b347689f ML |
1027 | |
1028 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1029 | if (!list_empty(&ctx->rq_lists[type])) { |
1030 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1031 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1032 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1033 | sbitmap_clear_bit(sb, bitnr); |
1034 | } | |
1035 | spin_unlock(&ctx->lock); | |
1036 | ||
1037 | return !dispatch_data->rq; | |
1038 | } | |
1039 | ||
1040 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1041 | struct blk_mq_ctx *start) | |
1042 | { | |
f31967f0 | 1043 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1044 | struct dispatch_rq_data data = { |
1045 | .hctx = hctx, | |
1046 | .rq = NULL, | |
1047 | }; | |
1048 | ||
1049 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1050 | dispatch_rq_from_ctx, &data); | |
1051 | ||
1052 | return data.rq; | |
1053 | } | |
1054 | ||
703fd1c0 JA |
1055 | static inline unsigned int queued_to_index(unsigned int queued) |
1056 | { | |
1057 | if (!queued) | |
1058 | return 0; | |
1429d7c9 | 1059 | |
703fd1c0 | 1060 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1061 | } |
1062 | ||
570e9b73 ML |
1063 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1064 | { | |
222a5ae0 | 1065 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1066 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1067 | int tag; |
1068 | ||
568f2700 ML |
1069 | blk_mq_tag_busy(rq->mq_hctx); |
1070 | ||
570e9b73 | 1071 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
222a5ae0 | 1072 | bt = rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1073 | tag_offset = 0; |
28500850 ML |
1074 | } else { |
1075 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1076 | return false; | |
570e9b73 ML |
1077 | } |
1078 | ||
570e9b73 ML |
1079 | tag = __sbitmap_queue_get(bt); |
1080 | if (tag == BLK_MQ_NO_TAG) | |
1081 | return false; | |
1082 | ||
1083 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1084 | return true; |
1085 | } | |
1086 | ||
61347154 | 1087 | bool blk_mq_get_driver_tag(struct request *rq) |
570e9b73 | 1088 | { |
568f2700 ML |
1089 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1090 | ||
1091 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1092 | return false; | |
1093 | ||
51db1c37 | 1094 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1095 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1096 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1097 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1098 | } |
1099 | hctx->tags->rqs[rq->tag] = rq; | |
1100 | return true; | |
570e9b73 ML |
1101 | } |
1102 | ||
eb619fdb JA |
1103 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1104 | int flags, void *key) | |
da55f2cc OS |
1105 | { |
1106 | struct blk_mq_hw_ctx *hctx; | |
1107 | ||
1108 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1109 | ||
5815839b | 1110 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1111 | if (!list_empty(&wait->entry)) { |
1112 | struct sbitmap_queue *sbq; | |
1113 | ||
1114 | list_del_init(&wait->entry); | |
222a5ae0 | 1115 | sbq = hctx->tags->bitmap_tags; |
e8618575 JA |
1116 | atomic_dec(&sbq->ws_active); |
1117 | } | |
5815839b ML |
1118 | spin_unlock(&hctx->dispatch_wait_lock); |
1119 | ||
da55f2cc OS |
1120 | blk_mq_run_hw_queue(hctx, true); |
1121 | return 1; | |
1122 | } | |
1123 | ||
f906a6a0 JA |
1124 | /* |
1125 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1126 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1127 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1128 | * marking us as waiting. |
1129 | */ | |
2278d69f | 1130 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1131 | struct request *rq) |
da55f2cc | 1132 | { |
222a5ae0 | 1133 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; |
5815839b | 1134 | struct wait_queue_head *wq; |
f906a6a0 JA |
1135 | wait_queue_entry_t *wait; |
1136 | bool ret; | |
da55f2cc | 1137 | |
51db1c37 | 1138 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1139 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1140 | |
c27d53fb BVA |
1141 | /* |
1142 | * It's possible that a tag was freed in the window between the | |
1143 | * allocation failure and adding the hardware queue to the wait | |
1144 | * queue. | |
1145 | * | |
1146 | * Don't clear RESTART here, someone else could have set it. | |
1147 | * At most this will cost an extra queue run. | |
1148 | */ | |
8ab6bb9e | 1149 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1150 | } |
eb619fdb | 1151 | |
2278d69f | 1152 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1153 | if (!list_empty_careful(&wait->entry)) |
1154 | return false; | |
1155 | ||
e8618575 | 1156 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1157 | |
1158 | spin_lock_irq(&wq->lock); | |
1159 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1160 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1161 | spin_unlock(&hctx->dispatch_wait_lock); |
1162 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1163 | return false; |
eb619fdb JA |
1164 | } |
1165 | ||
e8618575 | 1166 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1167 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1168 | __add_wait_queue(wq, wait); | |
c27d53fb | 1169 | |
da55f2cc | 1170 | /* |
eb619fdb JA |
1171 | * It's possible that a tag was freed in the window between the |
1172 | * allocation failure and adding the hardware queue to the wait | |
1173 | * queue. | |
da55f2cc | 1174 | */ |
8ab6bb9e | 1175 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1176 | if (!ret) { |
5815839b ML |
1177 | spin_unlock(&hctx->dispatch_wait_lock); |
1178 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1179 | return false; |
eb619fdb | 1180 | } |
c27d53fb BVA |
1181 | |
1182 | /* | |
1183 | * We got a tag, remove ourselves from the wait queue to ensure | |
1184 | * someone else gets the wakeup. | |
1185 | */ | |
c27d53fb | 1186 | list_del_init(&wait->entry); |
e8618575 | 1187 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1188 | spin_unlock(&hctx->dispatch_wait_lock); |
1189 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1190 | |
1191 | return true; | |
da55f2cc OS |
1192 | } |
1193 | ||
6e768717 ML |
1194 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1195 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1196 | /* | |
1197 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1198 | * - EWMA is one simple way to compute running average value | |
1199 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1200 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1201 | * factor doesn't matter because EWMA decreases exponentially | |
1202 | */ | |
1203 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1204 | { | |
1205 | unsigned int ewma; | |
1206 | ||
6e768717 ML |
1207 | ewma = hctx->dispatch_busy; |
1208 | ||
1209 | if (!ewma && !busy) | |
1210 | return; | |
1211 | ||
1212 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1213 | if (busy) | |
1214 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1215 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1216 | ||
1217 | hctx->dispatch_busy = ewma; | |
1218 | } | |
1219 | ||
86ff7c2a ML |
1220 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1221 | ||
c92a4103 JT |
1222 | static void blk_mq_handle_dev_resource(struct request *rq, |
1223 | struct list_head *list) | |
1224 | { | |
1225 | struct request *next = | |
1226 | list_first_entry_or_null(list, struct request, queuelist); | |
1227 | ||
1228 | /* | |
1229 | * If an I/O scheduler has been configured and we got a driver tag for | |
1230 | * the next request already, free it. | |
1231 | */ | |
1232 | if (next) | |
1233 | blk_mq_put_driver_tag(next); | |
1234 | ||
1235 | list_add(&rq->queuelist, list); | |
1236 | __blk_mq_requeue_request(rq); | |
1237 | } | |
1238 | ||
0512a75b KB |
1239 | static void blk_mq_handle_zone_resource(struct request *rq, |
1240 | struct list_head *zone_list) | |
1241 | { | |
1242 | /* | |
1243 | * If we end up here it is because we cannot dispatch a request to a | |
1244 | * specific zone due to LLD level zone-write locking or other zone | |
1245 | * related resource not being available. In this case, set the request | |
1246 | * aside in zone_list for retrying it later. | |
1247 | */ | |
1248 | list_add(&rq->queuelist, zone_list); | |
1249 | __blk_mq_requeue_request(rq); | |
1250 | } | |
1251 | ||
75383524 ML |
1252 | enum prep_dispatch { |
1253 | PREP_DISPATCH_OK, | |
1254 | PREP_DISPATCH_NO_TAG, | |
1255 | PREP_DISPATCH_NO_BUDGET, | |
1256 | }; | |
1257 | ||
1258 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1259 | bool need_budget) | |
1260 | { | |
1261 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1262 | int budget_token = -1; |
75383524 | 1263 | |
2a5a24aa ML |
1264 | if (need_budget) { |
1265 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1266 | if (budget_token < 0) { | |
1267 | blk_mq_put_driver_tag(rq); | |
1268 | return PREP_DISPATCH_NO_BUDGET; | |
1269 | } | |
1270 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1271 | } |
1272 | ||
1273 | if (!blk_mq_get_driver_tag(rq)) { | |
1274 | /* | |
1275 | * The initial allocation attempt failed, so we need to | |
1276 | * rerun the hardware queue when a tag is freed. The | |
1277 | * waitqueue takes care of that. If the queue is run | |
1278 | * before we add this entry back on the dispatch list, | |
1279 | * we'll re-run it below. | |
1280 | */ | |
1281 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1282 | /* |
1283 | * All budgets not got from this function will be put | |
1284 | * together during handling partial dispatch | |
1285 | */ | |
1286 | if (need_budget) | |
2a5a24aa | 1287 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1288 | return PREP_DISPATCH_NO_TAG; |
1289 | } | |
1290 | } | |
1291 | ||
1292 | return PREP_DISPATCH_OK; | |
1293 | } | |
1294 | ||
1fd40b5e ML |
1295 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1296 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1297 | struct list_head *list) |
1fd40b5e | 1298 | { |
2a5a24aa | 1299 | struct request *rq; |
1fd40b5e | 1300 | |
2a5a24aa ML |
1301 | list_for_each_entry(rq, list, queuelist) { |
1302 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1303 | |
2a5a24aa ML |
1304 | if (budget_token >= 0) |
1305 | blk_mq_put_dispatch_budget(q, budget_token); | |
1306 | } | |
1fd40b5e ML |
1307 | } |
1308 | ||
1f57f8d4 JA |
1309 | /* |
1310 | * Returns true if we did some work AND can potentially do more. | |
1311 | */ | |
445874e8 | 1312 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1313 | unsigned int nr_budgets) |
320ae51f | 1314 | { |
75383524 | 1315 | enum prep_dispatch prep; |
445874e8 | 1316 | struct request_queue *q = hctx->queue; |
6d6f167c | 1317 | struct request *rq, *nxt; |
fc17b653 | 1318 | int errors, queued; |
86ff7c2a | 1319 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1320 | LIST_HEAD(zone_list); |
320ae51f | 1321 | |
81380ca1 OS |
1322 | if (list_empty(list)) |
1323 | return false; | |
1324 | ||
320ae51f JA |
1325 | /* |
1326 | * Now process all the entries, sending them to the driver. | |
1327 | */ | |
93efe981 | 1328 | errors = queued = 0; |
81380ca1 | 1329 | do { |
74c45052 | 1330 | struct blk_mq_queue_data bd; |
320ae51f | 1331 | |
f04c3df3 | 1332 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1333 | |
445874e8 | 1334 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1335 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1336 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1337 | break; |
de148297 | 1338 | |
320ae51f | 1339 | list_del_init(&rq->queuelist); |
320ae51f | 1340 | |
74c45052 | 1341 | bd.rq = rq; |
113285b4 JA |
1342 | |
1343 | /* | |
1344 | * Flag last if we have no more requests, or if we have more | |
1345 | * but can't assign a driver tag to it. | |
1346 | */ | |
1347 | if (list_empty(list)) | |
1348 | bd.last = true; | |
1349 | else { | |
113285b4 | 1350 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1351 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1352 | } |
74c45052 | 1353 | |
1fd40b5e ML |
1354 | /* |
1355 | * once the request is queued to lld, no need to cover the | |
1356 | * budget any more | |
1357 | */ | |
1358 | if (nr_budgets) | |
1359 | nr_budgets--; | |
74c45052 | 1360 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1361 | switch (ret) { |
1362 | case BLK_STS_OK: | |
1363 | queued++; | |
320ae51f | 1364 | break; |
7bf13729 ML |
1365 | case BLK_STS_RESOURCE: |
1366 | case BLK_STS_DEV_RESOURCE: | |
1367 | blk_mq_handle_dev_resource(rq, list); | |
1368 | goto out; | |
1369 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1370 | /* |
1371 | * Move the request to zone_list and keep going through | |
1372 | * the dispatch list to find more requests the drive can | |
1373 | * accept. | |
1374 | */ | |
1375 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1376 | break; |
1377 | default: | |
93efe981 | 1378 | errors++; |
e21ee5a6 | 1379 | blk_mq_end_request(rq, ret); |
320ae51f | 1380 | } |
81380ca1 | 1381 | } while (!list_empty(list)); |
7bf13729 | 1382 | out: |
0512a75b KB |
1383 | if (!list_empty(&zone_list)) |
1384 | list_splice_tail_init(&zone_list, list); | |
1385 | ||
703fd1c0 | 1386 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1387 | |
632bfb63 | 1388 | /* If we didn't flush the entire list, we could have told the driver |
1389 | * there was more coming, but that turned out to be a lie. | |
1390 | */ | |
1391 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1392 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1393 | /* |
1394 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1395 | * that is where we will continue on next queue run. | |
1396 | */ | |
f04c3df3 | 1397 | if (!list_empty(list)) { |
86ff7c2a | 1398 | bool needs_restart; |
75383524 ML |
1399 | /* For non-shared tags, the RESTART check will suffice */ |
1400 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1401 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1402 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1403 | |
2a5a24aa ML |
1404 | if (nr_budgets) |
1405 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1406 | |
320ae51f | 1407 | spin_lock(&hctx->lock); |
01e99aec | 1408 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1409 | spin_unlock(&hctx->lock); |
f04c3df3 | 1410 | |
d7d8535f ML |
1411 | /* |
1412 | * Order adding requests to hctx->dispatch and checking | |
1413 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1414 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1415 | * miss the new added requests to hctx->dispatch, meantime | |
1416 | * SCHED_RESTART is observed here. | |
1417 | */ | |
1418 | smp_mb(); | |
1419 | ||
9ba52e58 | 1420 | /* |
710c785f BVA |
1421 | * If SCHED_RESTART was set by the caller of this function and |
1422 | * it is no longer set that means that it was cleared by another | |
1423 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1424 | * |
eb619fdb JA |
1425 | * If 'no_tag' is set, that means that we failed getting |
1426 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1427 | * waitqueue is no longer active, ensure that we run the queue | |
1428 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1429 | * |
710c785f BVA |
1430 | * If no I/O scheduler has been configured it is possible that |
1431 | * the hardware queue got stopped and restarted before requests | |
1432 | * were pushed back onto the dispatch list. Rerun the queue to | |
1433 | * avoid starvation. Notes: | |
1434 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1435 | * been stopped before rerunning a queue. | |
1436 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1437 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1438 | * and dm-rq. |
86ff7c2a ML |
1439 | * |
1440 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1441 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1442 | * that could otherwise occur if the queue is idle. We'll do |
1443 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1444 | */ |
86ff7c2a ML |
1445 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1446 | if (!needs_restart || | |
eb619fdb | 1447 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1448 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1449 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1450 | no_budget_avail)) | |
86ff7c2a | 1451 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1452 | |
6e768717 | 1453 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1454 | return false; |
6e768717 ML |
1455 | } else |
1456 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1457 | |
93efe981 | 1458 | return (queued + errors) != 0; |
f04c3df3 JA |
1459 | } |
1460 | ||
105663f7 AA |
1461 | /** |
1462 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1463 | * @hctx: Pointer to the hardware queue to run. | |
1464 | * | |
1465 | * Send pending requests to the hardware. | |
1466 | */ | |
6a83e74d BVA |
1467 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1468 | { | |
1469 | int srcu_idx; | |
1470 | ||
b7a71e66 JA |
1471 | /* |
1472 | * We can't run the queue inline with ints disabled. Ensure that | |
1473 | * we catch bad users of this early. | |
1474 | */ | |
1475 | WARN_ON_ONCE(in_interrupt()); | |
1476 | ||
04ced159 | 1477 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1478 | |
04ced159 JA |
1479 | hctx_lock(hctx, &srcu_idx); |
1480 | blk_mq_sched_dispatch_requests(hctx); | |
1481 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1482 | } |
1483 | ||
f82ddf19 ML |
1484 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1485 | { | |
1486 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1487 | ||
1488 | if (cpu >= nr_cpu_ids) | |
1489 | cpu = cpumask_first(hctx->cpumask); | |
1490 | return cpu; | |
1491 | } | |
1492 | ||
506e931f JA |
1493 | /* |
1494 | * It'd be great if the workqueue API had a way to pass | |
1495 | * in a mask and had some smarts for more clever placement. | |
1496 | * For now we just round-robin here, switching for every | |
1497 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1498 | */ | |
1499 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1500 | { | |
7bed4595 | 1501 | bool tried = false; |
476f8c98 | 1502 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1503 | |
b657d7e6 CH |
1504 | if (hctx->queue->nr_hw_queues == 1) |
1505 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1506 | |
1507 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1508 | select_cpu: |
476f8c98 | 1509 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1510 | cpu_online_mask); |
506e931f | 1511 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1512 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1513 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1514 | } | |
1515 | ||
7bed4595 ML |
1516 | /* |
1517 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1518 | * and it should only happen in the path of handling CPU DEAD. | |
1519 | */ | |
476f8c98 | 1520 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1521 | if (!tried) { |
1522 | tried = true; | |
1523 | goto select_cpu; | |
1524 | } | |
1525 | ||
1526 | /* | |
1527 | * Make sure to re-select CPU next time once after CPUs | |
1528 | * in hctx->cpumask become online again. | |
1529 | */ | |
476f8c98 | 1530 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1531 | hctx->next_cpu_batch = 1; |
1532 | return WORK_CPU_UNBOUND; | |
1533 | } | |
476f8c98 ML |
1534 | |
1535 | hctx->next_cpu = next_cpu; | |
1536 | return next_cpu; | |
506e931f JA |
1537 | } |
1538 | ||
105663f7 AA |
1539 | /** |
1540 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1541 | * @hctx: Pointer to the hardware queue to run. | |
1542 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1543 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1544 | * |
1545 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1546 | * with a delay of @msecs. | |
1547 | */ | |
7587a5ae BVA |
1548 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1549 | unsigned long msecs) | |
320ae51f | 1550 | { |
5435c023 | 1551 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1552 | return; |
1553 | ||
1b792f2f | 1554 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1555 | int cpu = get_cpu(); |
1556 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1557 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1558 | put_cpu(); |
398205b8 PB |
1559 | return; |
1560 | } | |
e4043dcf | 1561 | |
2a90d4aa | 1562 | put_cpu(); |
e4043dcf | 1563 | } |
398205b8 | 1564 | |
ae943d20 BVA |
1565 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1566 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1567 | } |
1568 | ||
105663f7 AA |
1569 | /** |
1570 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1571 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1572 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1573 | * |
1574 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1575 | */ | |
7587a5ae BVA |
1576 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1577 | { | |
1578 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1579 | } | |
1580 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1581 | ||
105663f7 AA |
1582 | /** |
1583 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1584 | * @hctx: Pointer to the hardware queue to run. | |
1585 | * @async: If we want to run the queue asynchronously. | |
1586 | * | |
1587 | * Check if the request queue is not in a quiesced state and if there are | |
1588 | * pending requests to be sent. If this is true, run the queue to send requests | |
1589 | * to hardware. | |
1590 | */ | |
626fb735 | 1591 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1592 | { |
24f5a90f ML |
1593 | int srcu_idx; |
1594 | bool need_run; | |
1595 | ||
1596 | /* | |
1597 | * When queue is quiesced, we may be switching io scheduler, or | |
1598 | * updating nr_hw_queues, or other things, and we can't run queue | |
1599 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1600 | * | |
1601 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1602 | * quiesced. | |
1603 | */ | |
04ced159 JA |
1604 | hctx_lock(hctx, &srcu_idx); |
1605 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1606 | blk_mq_hctx_has_pending(hctx); | |
1607 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1608 | |
626fb735 | 1609 | if (need_run) |
79f720a7 | 1610 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1611 | } |
5b727272 | 1612 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1613 | |
b6e68ee8 JK |
1614 | /* |
1615 | * Is the request queue handled by an IO scheduler that does not respect | |
1616 | * hardware queues when dispatching? | |
1617 | */ | |
1618 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
1619 | { | |
1620 | struct elevator_queue *e = q->elevator; | |
1621 | ||
1622 | if (e && e->type->ops.dispatch_request && | |
1623 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
1624 | return true; | |
1625 | return false; | |
1626 | } | |
1627 | ||
1628 | /* | |
1629 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
1630 | * scheduler. | |
1631 | */ | |
1632 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
1633 | { | |
1634 | struct blk_mq_hw_ctx *hctx; | |
1635 | ||
1636 | /* | |
1637 | * If the IO scheduler does not respect hardware queues when | |
1638 | * dispatching, we just don't bother with multiple HW queues and | |
1639 | * dispatch from hctx for the current CPU since running multiple queues | |
1640 | * just causes lock contention inside the scheduler and pointless cache | |
1641 | * bouncing. | |
1642 | */ | |
1643 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
1644 | raw_smp_processor_id()); | |
1645 | if (!blk_mq_hctx_stopped(hctx)) | |
1646 | return hctx; | |
1647 | return NULL; | |
1648 | } | |
1649 | ||
105663f7 | 1650 | /** |
24f7bb88 | 1651 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1652 | * @q: Pointer to the request queue to run. |
1653 | * @async: If we want to run the queue asynchronously. | |
1654 | */ | |
b94ec296 | 1655 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 1656 | { |
b6e68ee8 | 1657 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
1658 | int i; |
1659 | ||
b6e68ee8 JK |
1660 | sq_hctx = NULL; |
1661 | if (blk_mq_has_sqsched(q)) | |
1662 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 1663 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 1664 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 1665 | continue; |
b6e68ee8 JK |
1666 | /* |
1667 | * Dispatch from this hctx either if there's no hctx preferred | |
1668 | * by IO scheduler or if it has requests that bypass the | |
1669 | * scheduler. | |
1670 | */ | |
1671 | if (!sq_hctx || sq_hctx == hctx || | |
1672 | !list_empty_careful(&hctx->dispatch)) | |
1673 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
1674 | } |
1675 | } | |
b94ec296 | 1676 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1677 | |
b9151e7b DA |
1678 | /** |
1679 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1680 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1681 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1682 | */ |
1683 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1684 | { | |
b6e68ee8 | 1685 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
1686 | int i; |
1687 | ||
b6e68ee8 JK |
1688 | sq_hctx = NULL; |
1689 | if (blk_mq_has_sqsched(q)) | |
1690 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
1691 | queue_for_each_hw_ctx(q, hctx, i) { |
1692 | if (blk_mq_hctx_stopped(hctx)) | |
1693 | continue; | |
b6e68ee8 JK |
1694 | /* |
1695 | * Dispatch from this hctx either if there's no hctx preferred | |
1696 | * by IO scheduler or if it has requests that bypass the | |
1697 | * scheduler. | |
1698 | */ | |
1699 | if (!sq_hctx || sq_hctx == hctx || | |
1700 | !list_empty_careful(&hctx->dispatch)) | |
1701 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
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 | ||
a54895fa | 1824 | trace_block_rq_insert(rq); |
01b983c9 | 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); |
a54895fa | 1881 | trace_block_rq_insert(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 | ||
4f0f586b ST |
1890 | static int plug_rq_cmp(void *priv, const struct list_head *a, |
1891 | const struct list_head *b) | |
320ae51f JA |
1892 | { |
1893 | struct request *rqa = container_of(a, struct request, queuelist); | |
1894 | struct request *rqb = container_of(b, struct request, queuelist); | |
1895 | ||
7d30a621 PB |
1896 | if (rqa->mq_ctx != rqb->mq_ctx) |
1897 | return rqa->mq_ctx > rqb->mq_ctx; | |
1898 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1899 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1900 | |
1901 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1902 | } |
1903 | ||
1904 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1905 | { | |
320ae51f | 1906 | LIST_HEAD(list); |
320ae51f | 1907 | |
95ed0c5b PB |
1908 | if (list_empty(&plug->mq_list)) |
1909 | return; | |
320ae51f JA |
1910 | list_splice_init(&plug->mq_list, &list); |
1911 | ||
ce5b009c JA |
1912 | if (plug->rq_count > 2 && plug->multiple_queues) |
1913 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1914 | |
bcc816df DZ |
1915 | plug->rq_count = 0; |
1916 | ||
95ed0c5b PB |
1917 | do { |
1918 | struct list_head rq_list; | |
1919 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1920 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1921 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1922 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1923 | unsigned int depth = 1; | |
1924 | ||
1925 | list_for_each_continue(pos, &list) { | |
1926 | rq = list_entry_rq(pos); | |
1927 | BUG_ON(!rq->q); | |
1928 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1929 | break; | |
1930 | depth++; | |
320ae51f JA |
1931 | } |
1932 | ||
95ed0c5b PB |
1933 | list_cut_before(&rq_list, &list, pos); |
1934 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1935 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1936 | from_schedule); |
95ed0c5b | 1937 | } while(!list_empty(&list)); |
320ae51f JA |
1938 | } |
1939 | ||
14ccb66b CH |
1940 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1941 | unsigned int nr_segs) | |
320ae51f | 1942 | { |
93f221ae EB |
1943 | int err; |
1944 | ||
f924cdde CH |
1945 | if (bio->bi_opf & REQ_RAHEAD) |
1946 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1947 | ||
1948 | rq->__sector = bio->bi_iter.bi_sector; | |
1949 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1950 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1951 | |
1952 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
1953 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1954 | WARN_ON_ONCE(err); | |
4b570521 | 1955 | |
b5af37ab | 1956 | blk_account_io_start(rq); |
320ae51f JA |
1957 | } |
1958 | ||
0f95549c MS |
1959 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1960 | struct request *rq, | |
be94f058 | 1961 | blk_qc_t *cookie, bool last) |
f984df1f | 1962 | { |
f984df1f | 1963 | struct request_queue *q = rq->q; |
f984df1f SL |
1964 | struct blk_mq_queue_data bd = { |
1965 | .rq = rq, | |
be94f058 | 1966 | .last = last, |
f984df1f | 1967 | }; |
bd166ef1 | 1968 | blk_qc_t new_cookie; |
f06345ad | 1969 | blk_status_t ret; |
0f95549c MS |
1970 | |
1971 | new_cookie = request_to_qc_t(hctx, rq); | |
1972 | ||
1973 | /* | |
1974 | * For OK queue, we are done. For error, caller may kill it. | |
1975 | * Any other error (busy), just add it to our list as we | |
1976 | * previously would have done. | |
1977 | */ | |
1978 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1979 | switch (ret) { | |
1980 | case BLK_STS_OK: | |
6ce3dd6e | 1981 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1982 | *cookie = new_cookie; |
1983 | break; | |
1984 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1985 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1986 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1987 | __blk_mq_requeue_request(rq); |
1988 | break; | |
1989 | default: | |
6ce3dd6e | 1990 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1991 | *cookie = BLK_QC_T_NONE; |
1992 | break; | |
1993 | } | |
1994 | ||
1995 | return ret; | |
1996 | } | |
1997 | ||
fd9c40f6 | 1998 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1999 | struct request *rq, |
396eaf21 | 2000 | blk_qc_t *cookie, |
fd9c40f6 | 2001 | bool bypass_insert, bool last) |
0f95549c MS |
2002 | { |
2003 | struct request_queue *q = rq->q; | |
d964f04a | 2004 | bool run_queue = true; |
2a5a24aa | 2005 | int budget_token; |
d964f04a | 2006 | |
23d4ee19 | 2007 | /* |
fd9c40f6 | 2008 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2009 | * |
fd9c40f6 BVA |
2010 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2011 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2012 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2013 | */ |
fd9c40f6 | 2014 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2015 | run_queue = false; |
fd9c40f6 BVA |
2016 | bypass_insert = false; |
2017 | goto insert; | |
d964f04a | 2018 | } |
f984df1f | 2019 | |
fd9c40f6 BVA |
2020 | if (q->elevator && !bypass_insert) |
2021 | goto insert; | |
2253efc8 | 2022 | |
2a5a24aa ML |
2023 | budget_token = blk_mq_get_dispatch_budget(q); |
2024 | if (budget_token < 0) | |
fd9c40f6 | 2025 | goto insert; |
bd166ef1 | 2026 | |
2a5a24aa ML |
2027 | blk_mq_set_rq_budget_token(rq, budget_token); |
2028 | ||
8ab6bb9e | 2029 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2030 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2031 | goto insert; |
88022d72 | 2032 | } |
de148297 | 2033 | |
fd9c40f6 BVA |
2034 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2035 | insert: | |
2036 | if (bypass_insert) | |
2037 | return BLK_STS_RESOURCE; | |
2038 | ||
db03f88f ML |
2039 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2040 | ||
fd9c40f6 BVA |
2041 | return BLK_STS_OK; |
2042 | } | |
2043 | ||
105663f7 AA |
2044 | /** |
2045 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2046 | * @hctx: Pointer of the associated hardware queue. | |
2047 | * @rq: Pointer to request to be sent. | |
2048 | * @cookie: Request queue cookie. | |
2049 | * | |
2050 | * If the device has enough resources to accept a new request now, send the | |
2051 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2052 | * we can try send it another time in the future. Requests inserted at this | |
2053 | * queue have higher priority. | |
2054 | */ | |
fd9c40f6 BVA |
2055 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2056 | struct request *rq, blk_qc_t *cookie) | |
2057 | { | |
2058 | blk_status_t ret; | |
2059 | int srcu_idx; | |
2060 | ||
2061 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2062 | ||
2063 | hctx_lock(hctx, &srcu_idx); | |
2064 | ||
2065 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2066 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2067 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2068 | else if (ret != BLK_STS_OK) |
2069 | blk_mq_end_request(rq, ret); | |
2070 | ||
2071 | hctx_unlock(hctx, srcu_idx); | |
2072 | } | |
2073 | ||
2074 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2075 | { | |
2076 | blk_status_t ret; | |
2077 | int srcu_idx; | |
2078 | blk_qc_t unused_cookie; | |
2079 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2080 | ||
2081 | hctx_lock(hctx, &srcu_idx); | |
2082 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2083 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2084 | |
2085 | return ret; | |
5eb6126e CH |
2086 | } |
2087 | ||
6ce3dd6e ML |
2088 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2089 | struct list_head *list) | |
2090 | { | |
536167d4 | 2091 | int queued = 0; |
632bfb63 | 2092 | int errors = 0; |
536167d4 | 2093 | |
6ce3dd6e | 2094 | while (!list_empty(list)) { |
fd9c40f6 | 2095 | blk_status_t ret; |
6ce3dd6e ML |
2096 | struct request *rq = list_first_entry(list, struct request, |
2097 | queuelist); | |
2098 | ||
2099 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2100 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2101 | if (ret != BLK_STS_OK) { | |
2102 | if (ret == BLK_STS_RESOURCE || | |
2103 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2104 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2105 | list_empty(list)); |
fd9c40f6 BVA |
2106 | break; |
2107 | } | |
2108 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2109 | errors++; |
536167d4 KB |
2110 | } else |
2111 | queued++; | |
6ce3dd6e | 2112 | } |
d666ba98 JA |
2113 | |
2114 | /* | |
2115 | * If we didn't flush the entire list, we could have told | |
2116 | * the driver there was more coming, but that turned out to | |
2117 | * be a lie. | |
2118 | */ | |
632bfb63 | 2119 | if ((!list_empty(list) || errors) && |
2120 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2121 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2122 | } |
2123 | ||
ce5b009c JA |
2124 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2125 | { | |
2126 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2127 | plug->rq_count++; | |
2128 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2129 | struct request *tmp; | |
2130 | ||
2131 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2132 | queuelist); | |
2133 | if (tmp->q != rq->q) | |
2134 | plug->multiple_queues = true; | |
2135 | } | |
2136 | } | |
2137 | ||
105663f7 | 2138 | /** |
c62b37d9 | 2139 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2140 | * @bio: Bio pointer. |
2141 | * | |
2142 | * Builds up a request structure from @q and @bio and send to the device. The | |
2143 | * request may not be queued directly to hardware if: | |
2144 | * * This request can be merged with another one | |
2145 | * * We want to place request at plug queue for possible future merging | |
2146 | * * There is an IO scheduler active at this queue | |
2147 | * | |
2148 | * It will not queue the request if there is an error with the bio, or at the | |
2149 | * request creation. | |
2150 | * | |
2151 | * Returns: Request queue cookie. | |
2152 | */ | |
c62b37d9 | 2153 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2154 | { |
309dca30 | 2155 | struct request_queue *q = bio->bi_bdev->bd_disk->queue; |
ef295ecf | 2156 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2157 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6e7abff CH |
2158 | struct blk_mq_alloc_data data = { |
2159 | .q = q, | |
2160 | }; | |
07068d5b | 2161 | struct request *rq; |
f984df1f | 2162 | struct blk_plug *plug; |
5b3f341f | 2163 | struct request *same_queue_rq = NULL; |
14ccb66b | 2164 | unsigned int nr_segs; |
7b371636 | 2165 | blk_qc_t cookie; |
a892c8d5 | 2166 | blk_status_t ret; |
cc29e1bf | 2167 | bool hipri; |
07068d5b JA |
2168 | |
2169 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2170 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2171 | |
e23947bd | 2172 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2173 | goto queue_exit; |
07068d5b | 2174 | |
87c279e6 | 2175 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2176 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2177 | goto queue_exit; |
f984df1f | 2178 | |
14ccb66b | 2179 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2180 | goto queue_exit; |
bd166ef1 | 2181 | |
d5337560 | 2182 | rq_qos_throttle(q, bio); |
87760e5e | 2183 | |
cc29e1bf JX |
2184 | hipri = bio->bi_opf & REQ_HIPRI; |
2185 | ||
7809167d | 2186 | data.cmd_flags = bio->bi_opf; |
e6e7abff | 2187 | rq = __blk_mq_alloc_request(&data); |
87760e5e | 2188 | if (unlikely(!rq)) { |
c1c80384 | 2189 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2190 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2191 | bio_wouldblock_error(bio); |
ac7c5675 | 2192 | goto queue_exit; |
87760e5e JA |
2193 | } |
2194 | ||
e8a676d6 | 2195 | trace_block_getrq(bio); |
d6f1dda2 | 2196 | |
c1c80384 | 2197 | rq_qos_track(q, rq, bio); |
07068d5b | 2198 | |
fd2d3326 | 2199 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2200 | |
970d168d BVA |
2201 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2202 | ||
a892c8d5 ST |
2203 | ret = blk_crypto_init_request(rq); |
2204 | if (ret != BLK_STS_OK) { | |
2205 | bio->bi_status = ret; | |
2206 | bio_endio(bio); | |
2207 | blk_mq_free_request(rq); | |
2208 | return BLK_QC_T_NONE; | |
2209 | } | |
2210 | ||
b49773e7 | 2211 | plug = blk_mq_plug(q, bio); |
07068d5b | 2212 | if (unlikely(is_flush_fua)) { |
105663f7 | 2213 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2214 | blk_insert_flush(rq); |
2215 | blk_mq_run_hw_queue(data.hctx, true); | |
03f26d8f ML |
2216 | } else if (plug && (q->nr_hw_queues == 1 || |
2217 | blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) || | |
2218 | q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2219 | /* |
2220 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2221 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2222 | * |
2223 | * Use normal plugging if this disk is slow HDD, as sequential | |
2224 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2225 | */ |
5f0ed774 | 2226 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2227 | struct request *last = NULL; |
2228 | ||
676d0607 | 2229 | if (!request_count) |
e6c4438b | 2230 | trace_block_plug(q); |
600271d9 SL |
2231 | else |
2232 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2233 | |
600271d9 SL |
2234 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
2235 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
2236 | blk_flush_plug_list(plug, false); |
2237 | trace_block_plug(q); | |
320ae51f | 2238 | } |
b094f89c | 2239 | |
ce5b009c | 2240 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2241 | } else if (q->elevator) { |
105663f7 | 2242 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2243 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2244 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2245 | /* |
6a83e74d | 2246 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2247 | * Otherwise the existing request in the plug list will be |
2248 | * issued. So the plug list will have one request at most | |
2299722c CH |
2249 | * The plug list might get flushed before this. If that happens, |
2250 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2251 | */ |
2299722c CH |
2252 | if (list_empty(&plug->mq_list)) |
2253 | same_queue_rq = NULL; | |
4711b573 | 2254 | if (same_queue_rq) { |
2299722c | 2255 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2256 | plug->rq_count--; |
2257 | } | |
ce5b009c | 2258 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2259 | trace_block_plug(q); |
2299722c | 2260 | |
dad7a3be | 2261 | if (same_queue_rq) { |
ea4f995e | 2262 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2263 | trace_block_unplug(q, 1, true); |
2299722c | 2264 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2265 | &cookie); |
dad7a3be | 2266 | } |
a12de1d4 ML |
2267 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2268 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2269 | /* |
2270 | * There is no scheduler and we can try to send directly | |
2271 | * to the hardware. | |
2272 | */ | |
fd9c40f6 | 2273 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2274 | } else { |
105663f7 | 2275 | /* Default case. */ |
8fa9f556 | 2276 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2277 | } |
320ae51f | 2278 | |
cc29e1bf JX |
2279 | if (!hipri) |
2280 | return BLK_QC_T_NONE; | |
7b371636 | 2281 | return cookie; |
ac7c5675 CH |
2282 | queue_exit: |
2283 | blk_queue_exit(q); | |
2284 | return BLK_QC_T_NONE; | |
320ae51f JA |
2285 | } |
2286 | ||
bd63141d ML |
2287 | static size_t order_to_size(unsigned int order) |
2288 | { | |
2289 | return (size_t)PAGE_SIZE << order; | |
2290 | } | |
2291 | ||
2292 | /* called before freeing request pool in @tags */ | |
2293 | static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set, | |
2294 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
2295 | { | |
2296 | struct blk_mq_tags *drv_tags = set->tags[hctx_idx]; | |
2297 | struct page *page; | |
2298 | unsigned long flags; | |
2299 | ||
2300 | list_for_each_entry(page, &tags->page_list, lru) { | |
2301 | unsigned long start = (unsigned long)page_address(page); | |
2302 | unsigned long end = start + order_to_size(page->private); | |
2303 | int i; | |
2304 | ||
2305 | for (i = 0; i < set->queue_depth; i++) { | |
2306 | struct request *rq = drv_tags->rqs[i]; | |
2307 | unsigned long rq_addr = (unsigned long)rq; | |
2308 | ||
2309 | if (rq_addr >= start && rq_addr < end) { | |
2310 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
2311 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
2312 | } | |
2313 | } | |
2314 | } | |
2315 | ||
2316 | /* | |
2317 | * Wait until all pending iteration is done. | |
2318 | * | |
2319 | * Request reference is cleared and it is guaranteed to be observed | |
2320 | * after the ->lock is released. | |
2321 | */ | |
2322 | spin_lock_irqsave(&drv_tags->lock, flags); | |
2323 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
2324 | } | |
2325 | ||
cc71a6f4 JA |
2326 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2327 | unsigned int hctx_idx) | |
95363efd | 2328 | { |
e9b267d9 | 2329 | struct page *page; |
320ae51f | 2330 | |
24d2f903 | 2331 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2332 | int i; |
320ae51f | 2333 | |
24d2f903 | 2334 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2335 | struct request *rq = tags->static_rqs[i]; |
2336 | ||
2337 | if (!rq) | |
e9b267d9 | 2338 | continue; |
d6296d39 | 2339 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2340 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2341 | } |
320ae51f | 2342 | } |
320ae51f | 2343 | |
bd63141d ML |
2344 | blk_mq_clear_rq_mapping(set, tags, hctx_idx); |
2345 | ||
24d2f903 CH |
2346 | while (!list_empty(&tags->page_list)) { |
2347 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2348 | list_del_init(&page->lru); |
f75782e4 CM |
2349 | /* |
2350 | * Remove kmemleak object previously allocated in | |
273938bf | 2351 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2352 | */ |
2353 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2354 | __free_pages(page, page->private); |
2355 | } | |
cc71a6f4 | 2356 | } |
320ae51f | 2357 | |
1c0706a7 | 2358 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) |
cc71a6f4 | 2359 | { |
24d2f903 | 2360 | kfree(tags->rqs); |
cc71a6f4 | 2361 | tags->rqs = NULL; |
2af8cbe3 JA |
2362 | kfree(tags->static_rqs); |
2363 | tags->static_rqs = NULL; | |
320ae51f | 2364 | |
1c0706a7 | 2365 | blk_mq_free_tags(tags, flags); |
320ae51f JA |
2366 | } |
2367 | ||
cc71a6f4 JA |
2368 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2369 | unsigned int hctx_idx, | |
2370 | unsigned int nr_tags, | |
1c0706a7 JG |
2371 | unsigned int reserved_tags, |
2372 | unsigned int flags) | |
320ae51f | 2373 | { |
24d2f903 | 2374 | struct blk_mq_tags *tags; |
59f082e4 | 2375 | int node; |
320ae51f | 2376 | |
7d76f856 | 2377 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2378 | if (node == NUMA_NO_NODE) |
2379 | node = set->numa_node; | |
2380 | ||
1c0706a7 | 2381 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); |
24d2f903 CH |
2382 | if (!tags) |
2383 | return NULL; | |
320ae51f | 2384 | |
590b5b7d | 2385 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2386 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2387 | node); |
24d2f903 | 2388 | if (!tags->rqs) { |
1c0706a7 | 2389 | blk_mq_free_tags(tags, flags); |
24d2f903 CH |
2390 | return NULL; |
2391 | } | |
320ae51f | 2392 | |
590b5b7d KC |
2393 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2394 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2395 | node); | |
2af8cbe3 JA |
2396 | if (!tags->static_rqs) { |
2397 | kfree(tags->rqs); | |
1c0706a7 | 2398 | blk_mq_free_tags(tags, flags); |
2af8cbe3 JA |
2399 | return NULL; |
2400 | } | |
2401 | ||
cc71a6f4 JA |
2402 | return tags; |
2403 | } | |
2404 | ||
1d9bd516 TH |
2405 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2406 | unsigned int hctx_idx, int node) | |
2407 | { | |
2408 | int ret; | |
2409 | ||
2410 | if (set->ops->init_request) { | |
2411 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2412 | if (ret) | |
2413 | return ret; | |
2414 | } | |
2415 | ||
12f5b931 | 2416 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2417 | return 0; |
2418 | } | |
2419 | ||
cc71a6f4 JA |
2420 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2421 | unsigned int hctx_idx, unsigned int depth) | |
2422 | { | |
2423 | unsigned int i, j, entries_per_page, max_order = 4; | |
2424 | size_t rq_size, left; | |
59f082e4 SL |
2425 | int node; |
2426 | ||
7d76f856 | 2427 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2428 | if (node == NUMA_NO_NODE) |
2429 | node = set->numa_node; | |
cc71a6f4 JA |
2430 | |
2431 | INIT_LIST_HEAD(&tags->page_list); | |
2432 | ||
320ae51f JA |
2433 | /* |
2434 | * rq_size is the size of the request plus driver payload, rounded | |
2435 | * to the cacheline size | |
2436 | */ | |
24d2f903 | 2437 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2438 | cache_line_size()); |
cc71a6f4 | 2439 | left = rq_size * depth; |
320ae51f | 2440 | |
cc71a6f4 | 2441 | for (i = 0; i < depth; ) { |
320ae51f JA |
2442 | int this_order = max_order; |
2443 | struct page *page; | |
2444 | int to_do; | |
2445 | void *p; | |
2446 | ||
b3a834b1 | 2447 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2448 | this_order--; |
2449 | ||
2450 | do { | |
59f082e4 | 2451 | page = alloc_pages_node(node, |
36e1f3d1 | 2452 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2453 | this_order); |
320ae51f JA |
2454 | if (page) |
2455 | break; | |
2456 | if (!this_order--) | |
2457 | break; | |
2458 | if (order_to_size(this_order) < rq_size) | |
2459 | break; | |
2460 | } while (1); | |
2461 | ||
2462 | if (!page) | |
24d2f903 | 2463 | goto fail; |
320ae51f JA |
2464 | |
2465 | page->private = this_order; | |
24d2f903 | 2466 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2467 | |
2468 | p = page_address(page); | |
f75782e4 CM |
2469 | /* |
2470 | * Allow kmemleak to scan these pages as they contain pointers | |
2471 | * to additional allocations like via ops->init_request(). | |
2472 | */ | |
36e1f3d1 | 2473 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2474 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2475 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2476 | left -= to_do * rq_size; |
2477 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2478 | struct request *rq = p; |
2479 | ||
2480 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2481 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2482 | tags->static_rqs[i] = NULL; | |
2483 | goto fail; | |
e9b267d9 CH |
2484 | } |
2485 | ||
320ae51f JA |
2486 | p += rq_size; |
2487 | i++; | |
2488 | } | |
2489 | } | |
cc71a6f4 | 2490 | return 0; |
320ae51f | 2491 | |
24d2f903 | 2492 | fail: |
cc71a6f4 JA |
2493 | blk_mq_free_rqs(set, tags, hctx_idx); |
2494 | return -ENOMEM; | |
320ae51f JA |
2495 | } |
2496 | ||
bf0beec0 ML |
2497 | struct rq_iter_data { |
2498 | struct blk_mq_hw_ctx *hctx; | |
2499 | bool has_rq; | |
2500 | }; | |
2501 | ||
2502 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2503 | { | |
2504 | struct rq_iter_data *iter_data = data; | |
2505 | ||
2506 | if (rq->mq_hctx != iter_data->hctx) | |
2507 | return true; | |
2508 | iter_data->has_rq = true; | |
2509 | return false; | |
2510 | } | |
2511 | ||
2512 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2513 | { | |
2514 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2515 | hctx->sched_tags : hctx->tags; | |
2516 | struct rq_iter_data data = { | |
2517 | .hctx = hctx, | |
2518 | }; | |
2519 | ||
2520 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2521 | return data.has_rq; | |
2522 | } | |
2523 | ||
2524 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2525 | struct blk_mq_hw_ctx *hctx) | |
2526 | { | |
2527 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2528 | return false; | |
2529 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2530 | return false; | |
2531 | return true; | |
2532 | } | |
2533 | ||
2534 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2535 | { | |
2536 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2537 | struct blk_mq_hw_ctx, cpuhp_online); | |
2538 | ||
2539 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2540 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2541 | return 0; | |
2542 | ||
2543 | /* | |
2544 | * Prevent new request from being allocated on the current hctx. | |
2545 | * | |
2546 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2547 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2548 | * seen once we return from the tag allocator. | |
2549 | */ | |
2550 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2551 | smp_mb__after_atomic(); | |
2552 | ||
2553 | /* | |
2554 | * Try to grab a reference to the queue and wait for any outstanding | |
2555 | * requests. If we could not grab a reference the queue has been | |
2556 | * frozen and there are no requests. | |
2557 | */ | |
2558 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2559 | while (blk_mq_hctx_has_requests(hctx)) | |
2560 | msleep(5); | |
2561 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2562 | } | |
2563 | ||
2564 | return 0; | |
2565 | } | |
2566 | ||
2567 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2568 | { | |
2569 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2570 | struct blk_mq_hw_ctx, cpuhp_online); | |
2571 | ||
2572 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2573 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2574 | return 0; | |
2575 | } | |
2576 | ||
e57690fe JA |
2577 | /* |
2578 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2579 | * software queue to the hw queue dispatch list, and ensure that it | |
2580 | * gets run. | |
2581 | */ | |
9467f859 | 2582 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2583 | { |
9467f859 | 2584 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2585 | struct blk_mq_ctx *ctx; |
2586 | LIST_HEAD(tmp); | |
c16d6b5a | 2587 | enum hctx_type type; |
484b4061 | 2588 | |
9467f859 | 2589 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2590 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2591 | return 0; | |
2592 | ||
e57690fe | 2593 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2594 | type = hctx->type; |
484b4061 JA |
2595 | |
2596 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2597 | if (!list_empty(&ctx->rq_lists[type])) { |
2598 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2599 | blk_mq_hctx_clear_pending(hctx, ctx); |
2600 | } | |
2601 | spin_unlock(&ctx->lock); | |
2602 | ||
2603 | if (list_empty(&tmp)) | |
9467f859 | 2604 | return 0; |
484b4061 | 2605 | |
e57690fe JA |
2606 | spin_lock(&hctx->lock); |
2607 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2608 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2609 | |
2610 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2611 | return 0; |
484b4061 JA |
2612 | } |
2613 | ||
9467f859 | 2614 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2615 | { |
bf0beec0 ML |
2616 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2617 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2618 | &hctx->cpuhp_online); | |
9467f859 TG |
2619 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2620 | &hctx->cpuhp_dead); | |
484b4061 JA |
2621 | } |
2622 | ||
364b6181 ML |
2623 | /* |
2624 | * Before freeing hw queue, clearing the flush request reference in | |
2625 | * tags->rqs[] for avoiding potential UAF. | |
2626 | */ | |
2627 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
2628 | unsigned int queue_depth, struct request *flush_rq) | |
2629 | { | |
2630 | int i; | |
2631 | unsigned long flags; | |
2632 | ||
2633 | /* The hw queue may not be mapped yet */ | |
2634 | if (!tags) | |
2635 | return; | |
2636 | ||
2637 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
2638 | ||
2639 | for (i = 0; i < queue_depth; i++) | |
2640 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
2641 | ||
2642 | /* | |
2643 | * Wait until all pending iteration is done. | |
2644 | * | |
2645 | * Request reference is cleared and it is guaranteed to be observed | |
2646 | * after the ->lock is released. | |
2647 | */ | |
2648 | spin_lock_irqsave(&tags->lock, flags); | |
2649 | spin_unlock_irqrestore(&tags->lock, flags); | |
2650 | } | |
2651 | ||
c3b4afca | 2652 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2653 | static void blk_mq_exit_hctx(struct request_queue *q, |
2654 | struct blk_mq_tag_set *set, | |
2655 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2656 | { | |
364b6181 ML |
2657 | struct request *flush_rq = hctx->fq->flush_rq; |
2658 | ||
8ab0b7dc ML |
2659 | if (blk_mq_hw_queue_mapped(hctx)) |
2660 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2661 | |
364b6181 ML |
2662 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
2663 | set->queue_depth, flush_rq); | |
f70ced09 | 2664 | if (set->ops->exit_request) |
364b6181 | 2665 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 2666 | |
08e98fc6 ML |
2667 | if (set->ops->exit_hctx) |
2668 | set->ops->exit_hctx(hctx, hctx_idx); | |
2669 | ||
9467f859 | 2670 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2671 | |
2672 | spin_lock(&q->unused_hctx_lock); | |
2673 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2674 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2675 | } |
2676 | ||
624dbe47 ML |
2677 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2678 | struct blk_mq_tag_set *set, int nr_queue) | |
2679 | { | |
2680 | struct blk_mq_hw_ctx *hctx; | |
2681 | unsigned int i; | |
2682 | ||
2683 | queue_for_each_hw_ctx(q, hctx, i) { | |
2684 | if (i == nr_queue) | |
2685 | break; | |
477e19de | 2686 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2687 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2688 | } |
624dbe47 ML |
2689 | } |
2690 | ||
7c6c5b7c ML |
2691 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2692 | { | |
2693 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2694 | ||
2695 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2696 | __alignof__(struct blk_mq_hw_ctx)) != | |
2697 | sizeof(struct blk_mq_hw_ctx)); | |
2698 | ||
2699 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2700 | hw_ctx_size += sizeof(struct srcu_struct); | |
2701 | ||
2702 | return hw_ctx_size; | |
2703 | } | |
2704 | ||
08e98fc6 ML |
2705 | static int blk_mq_init_hctx(struct request_queue *q, |
2706 | struct blk_mq_tag_set *set, | |
2707 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2708 | { |
7c6c5b7c ML |
2709 | hctx->queue_num = hctx_idx; |
2710 | ||
bf0beec0 ML |
2711 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2712 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2713 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2714 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2715 | ||
2716 | hctx->tags = set->tags[hctx_idx]; | |
2717 | ||
2718 | if (set->ops->init_hctx && | |
2719 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2720 | goto unregister_cpu_notifier; | |
08e98fc6 | 2721 | |
7c6c5b7c ML |
2722 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2723 | hctx->numa_node)) | |
2724 | goto exit_hctx; | |
2725 | return 0; | |
2726 | ||
2727 | exit_hctx: | |
2728 | if (set->ops->exit_hctx) | |
2729 | set->ops->exit_hctx(hctx, hctx_idx); | |
2730 | unregister_cpu_notifier: | |
2731 | blk_mq_remove_cpuhp(hctx); | |
2732 | return -1; | |
2733 | } | |
2734 | ||
2735 | static struct blk_mq_hw_ctx * | |
2736 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2737 | int node) | |
2738 | { | |
2739 | struct blk_mq_hw_ctx *hctx; | |
2740 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2741 | ||
2742 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2743 | if (!hctx) | |
2744 | goto fail_alloc_hctx; | |
2745 | ||
2746 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2747 | goto free_hctx; | |
2748 | ||
2749 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2750 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2751 | node = set->numa_node; |
2752 | hctx->numa_node = node; | |
08e98fc6 | 2753 | |
9f993737 | 2754 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2755 | spin_lock_init(&hctx->lock); |
2756 | INIT_LIST_HEAD(&hctx->dispatch); | |
2757 | hctx->queue = q; | |
51db1c37 | 2758 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2759 | |
2f8f1336 ML |
2760 | INIT_LIST_HEAD(&hctx->hctx_list); |
2761 | ||
320ae51f | 2762 | /* |
08e98fc6 ML |
2763 | * Allocate space for all possible cpus to avoid allocation at |
2764 | * runtime | |
320ae51f | 2765 | */ |
d904bfa7 | 2766 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2767 | gfp, node); |
08e98fc6 | 2768 | if (!hctx->ctxs) |
7c6c5b7c | 2769 | goto free_cpumask; |
320ae51f | 2770 | |
5b202853 | 2771 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 2772 | gfp, node, false, false)) |
08e98fc6 | 2773 | goto free_ctxs; |
08e98fc6 | 2774 | hctx->nr_ctx = 0; |
320ae51f | 2775 | |
5815839b | 2776 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2777 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2778 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2779 | ||
754a1572 | 2780 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2781 | if (!hctx->fq) |
7c6c5b7c | 2782 | goto free_bitmap; |
320ae51f | 2783 | |
6a83e74d | 2784 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2785 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2786 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2787 | |
7c6c5b7c | 2788 | return hctx; |
320ae51f | 2789 | |
08e98fc6 | 2790 | free_bitmap: |
88459642 | 2791 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2792 | free_ctxs: |
2793 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2794 | free_cpumask: |
2795 | free_cpumask_var(hctx->cpumask); | |
2796 | free_hctx: | |
2797 | kfree(hctx); | |
2798 | fail_alloc_hctx: | |
2799 | return NULL; | |
08e98fc6 | 2800 | } |
320ae51f | 2801 | |
320ae51f JA |
2802 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2803 | unsigned int nr_hw_queues) | |
2804 | { | |
b3c661b1 JA |
2805 | struct blk_mq_tag_set *set = q->tag_set; |
2806 | unsigned int i, j; | |
320ae51f JA |
2807 | |
2808 | for_each_possible_cpu(i) { | |
2809 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2810 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2811 | int k; |
320ae51f | 2812 | |
320ae51f JA |
2813 | __ctx->cpu = i; |
2814 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2815 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2816 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2817 | ||
320ae51f JA |
2818 | __ctx->queue = q; |
2819 | ||
320ae51f JA |
2820 | /* |
2821 | * Set local node, IFF we have more than one hw queue. If | |
2822 | * not, we remain on the home node of the device | |
2823 | */ | |
b3c661b1 JA |
2824 | for (j = 0; j < set->nr_maps; j++) { |
2825 | hctx = blk_mq_map_queue_type(q, j, i); | |
2826 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2827 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2828 | } |
320ae51f JA |
2829 | } |
2830 | } | |
2831 | ||
03b63b02 WZ |
2832 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, |
2833 | int hctx_idx) | |
cc71a6f4 | 2834 | { |
1c0706a7 | 2835 | unsigned int flags = set->flags; |
cc71a6f4 JA |
2836 | int ret = 0; |
2837 | ||
2838 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1c0706a7 | 2839 | set->queue_depth, set->reserved_tags, flags); |
cc71a6f4 JA |
2840 | if (!set->tags[hctx_idx]) |
2841 | return false; | |
2842 | ||
2843 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2844 | set->queue_depth); | |
2845 | if (!ret) | |
2846 | return true; | |
2847 | ||
1c0706a7 | 2848 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
cc71a6f4 JA |
2849 | set->tags[hctx_idx] = NULL; |
2850 | return false; | |
2851 | } | |
2852 | ||
2853 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2854 | unsigned int hctx_idx) | |
2855 | { | |
1c0706a7 JG |
2856 | unsigned int flags = set->flags; |
2857 | ||
4e6db0f2 | 2858 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 | 2859 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
1c0706a7 | 2860 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
bd166ef1 JA |
2861 | set->tags[hctx_idx] = NULL; |
2862 | } | |
cc71a6f4 JA |
2863 | } |
2864 | ||
4b855ad3 | 2865 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2866 | { |
b3c661b1 | 2867 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2868 | struct blk_mq_hw_ctx *hctx; |
2869 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2870 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2871 | |
2872 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2873 | cpumask_clear(hctx->cpumask); |
320ae51f | 2874 | hctx->nr_ctx = 0; |
d416c92c | 2875 | hctx->dispatch_from = NULL; |
320ae51f JA |
2876 | } |
2877 | ||
2878 | /* | |
4b855ad3 | 2879 | * Map software to hardware queues. |
4412efec ML |
2880 | * |
2881 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2882 | */ |
20e4d813 | 2883 | for_each_possible_cpu(i) { |
4412efec | 2884 | |
897bb0c7 | 2885 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2886 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2887 | if (!set->map[j].nr_queues) { |
2888 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2889 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2890 | continue; |
bb94aea1 | 2891 | } |
fd689871 ML |
2892 | hctx_idx = set->map[j].mq_map[i]; |
2893 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2894 | if (!set->tags[hctx_idx] && | |
03b63b02 | 2895 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { |
fd689871 ML |
2896 | /* |
2897 | * If tags initialization fail for some hctx, | |
2898 | * that hctx won't be brought online. In this | |
2899 | * case, remap the current ctx to hctx[0] which | |
2900 | * is guaranteed to always have tags allocated | |
2901 | */ | |
2902 | set->map[j].mq_map[i] = 0; | |
2903 | } | |
e5edd5f2 | 2904 | |
b3c661b1 | 2905 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2906 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2907 | /* |
2908 | * If the CPU is already set in the mask, then we've | |
2909 | * mapped this one already. This can happen if | |
2910 | * devices share queues across queue maps. | |
2911 | */ | |
2912 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2913 | continue; | |
2914 | ||
2915 | cpumask_set_cpu(i, hctx->cpumask); | |
2916 | hctx->type = j; | |
2917 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2918 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2919 | ||
2920 | /* | |
2921 | * If the nr_ctx type overflows, we have exceeded the | |
2922 | * amount of sw queues we can support. | |
2923 | */ | |
2924 | BUG_ON(!hctx->nr_ctx); | |
2925 | } | |
bb94aea1 JW |
2926 | |
2927 | for (; j < HCTX_MAX_TYPES; j++) | |
2928 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2929 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2930 | } |
506e931f JA |
2931 | |
2932 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2933 | /* |
2934 | * If no software queues are mapped to this hardware queue, | |
2935 | * disable it and free the request entries. | |
2936 | */ | |
2937 | if (!hctx->nr_ctx) { | |
2938 | /* Never unmap queue 0. We need it as a | |
2939 | * fallback in case of a new remap fails | |
2940 | * allocation | |
2941 | */ | |
2942 | if (i && set->tags[i]) | |
2943 | blk_mq_free_map_and_requests(set, i); | |
2944 | ||
2945 | hctx->tags = NULL; | |
2946 | continue; | |
2947 | } | |
484b4061 | 2948 | |
2a34c087 ML |
2949 | hctx->tags = set->tags[i]; |
2950 | WARN_ON(!hctx->tags); | |
2951 | ||
889fa31f CY |
2952 | /* |
2953 | * Set the map size to the number of mapped software queues. | |
2954 | * This is more accurate and more efficient than looping | |
2955 | * over all possibly mapped software queues. | |
2956 | */ | |
88459642 | 2957 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2958 | |
484b4061 JA |
2959 | /* |
2960 | * Initialize batch roundrobin counts | |
2961 | */ | |
f82ddf19 | 2962 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2963 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2964 | } | |
320ae51f JA |
2965 | } |
2966 | ||
8e8320c9 JA |
2967 | /* |
2968 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2969 | * the queue isn't live yet. | |
2970 | */ | |
2404e607 | 2971 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2972 | { |
2973 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2974 | int i; |
2975 | ||
2404e607 | 2976 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 2977 | if (shared) { |
51db1c37 | 2978 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
2979 | } else { |
2980 | blk_mq_tag_idle(hctx); | |
51db1c37 | 2981 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 2982 | } |
2404e607 JM |
2983 | } |
2984 | } | |
2985 | ||
655ac300 HR |
2986 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
2987 | bool shared) | |
2404e607 JM |
2988 | { |
2989 | struct request_queue *q; | |
0d2602ca | 2990 | |
705cda97 BVA |
2991 | lockdep_assert_held(&set->tag_list_lock); |
2992 | ||
0d2602ca JA |
2993 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2994 | blk_mq_freeze_queue(q); | |
2404e607 | 2995 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2996 | blk_mq_unfreeze_queue(q); |
2997 | } | |
2998 | } | |
2999 | ||
3000 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3001 | { | |
3002 | struct blk_mq_tag_set *set = q->tag_set; | |
3003 | ||
0d2602ca | 3004 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3005 | list_del(&q->tag_set_list); |
2404e607 JM |
3006 | if (list_is_singular(&set->tag_list)) { |
3007 | /* just transitioned to unshared */ | |
51db1c37 | 3008 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3009 | /* update existing queue */ |
655ac300 | 3010 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3011 | } |
0d2602ca | 3012 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3013 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3014 | } |
3015 | ||
3016 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3017 | struct request_queue *q) | |
3018 | { | |
0d2602ca | 3019 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3020 | |
ff821d27 JA |
3021 | /* |
3022 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3023 | */ | |
3024 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3025 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3026 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3027 | /* update existing queue */ |
655ac300 | 3028 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3029 | } |
51db1c37 | 3030 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3031 | queue_set_hctx_shared(q, true); |
08c875cb | 3032 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3033 | |
0d2602ca JA |
3034 | mutex_unlock(&set->tag_list_lock); |
3035 | } | |
3036 | ||
1db4909e ML |
3037 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3038 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3039 | { | |
3040 | struct blk_mq_ctxs *ctxs; | |
3041 | int cpu; | |
3042 | ||
3043 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3044 | if (!ctxs) | |
3045 | return -ENOMEM; | |
3046 | ||
3047 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3048 | if (!ctxs->queue_ctx) | |
3049 | goto fail; | |
3050 | ||
3051 | for_each_possible_cpu(cpu) { | |
3052 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3053 | ctx->ctxs = ctxs; | |
3054 | } | |
3055 | ||
3056 | q->mq_kobj = &ctxs->kobj; | |
3057 | q->queue_ctx = ctxs->queue_ctx; | |
3058 | ||
3059 | return 0; | |
3060 | fail: | |
3061 | kfree(ctxs); | |
3062 | return -ENOMEM; | |
3063 | } | |
3064 | ||
e09aae7e ML |
3065 | /* |
3066 | * It is the actual release handler for mq, but we do it from | |
3067 | * request queue's release handler for avoiding use-after-free | |
3068 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3069 | * but we can't group ctx/kctx kobj without it. | |
3070 | */ | |
3071 | void blk_mq_release(struct request_queue *q) | |
3072 | { | |
2f8f1336 ML |
3073 | struct blk_mq_hw_ctx *hctx, *next; |
3074 | int i; | |
e09aae7e | 3075 | |
2f8f1336 ML |
3076 | queue_for_each_hw_ctx(q, hctx, i) |
3077 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3078 | ||
3079 | /* all hctx are in .unused_hctx_list now */ | |
3080 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3081 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3082 | kobject_put(&hctx->kobj); |
c3b4afca | 3083 | } |
e09aae7e ML |
3084 | |
3085 | kfree(q->queue_hw_ctx); | |
3086 | ||
7ea5fe31 ML |
3087 | /* |
3088 | * release .mq_kobj and sw queue's kobject now because | |
3089 | * both share lifetime with request queue. | |
3090 | */ | |
3091 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3092 | } |
3093 | ||
5ec780a6 | 3094 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3095 | void *queuedata) |
b62c21b7 | 3096 | { |
26a9750a CH |
3097 | struct request_queue *q; |
3098 | int ret; | |
b62c21b7 | 3099 | |
26a9750a CH |
3100 | q = blk_alloc_queue(set->numa_node); |
3101 | if (!q) | |
b62c21b7 | 3102 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3103 | q->queuedata = queuedata; |
3104 | ret = blk_mq_init_allocated_queue(set, q); | |
3105 | if (ret) { | |
3106 | blk_cleanup_queue(q); | |
3107 | return ERR_PTR(ret); | |
3108 | } | |
b62c21b7 MS |
3109 | return q; |
3110 | } | |
2f227bb9 CH |
3111 | |
3112 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3113 | { | |
3114 | return blk_mq_init_queue_data(set, NULL); | |
3115 | } | |
b62c21b7 MS |
3116 | EXPORT_SYMBOL(blk_mq_init_queue); |
3117 | ||
4dcc4874 CH |
3118 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3119 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3120 | { |
3121 | struct request_queue *q; | |
b461dfc4 | 3122 | struct gendisk *disk; |
9316a9ed | 3123 | |
b461dfc4 CH |
3124 | q = blk_mq_init_queue_data(set, queuedata); |
3125 | if (IS_ERR(q)) | |
3126 | return ERR_CAST(q); | |
9316a9ed | 3127 | |
4a1fa41d | 3128 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3129 | if (!disk) { |
3130 | blk_cleanup_queue(q); | |
3131 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3132 | } |
b461dfc4 | 3133 | return disk; |
9316a9ed | 3134 | } |
b461dfc4 | 3135 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3136 | |
34d11ffa JW |
3137 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3138 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3139 | int hctx_idx, int node) | |
3140 | { | |
2f8f1336 | 3141 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3142 | |
2f8f1336 ML |
3143 | /* reuse dead hctx first */ |
3144 | spin_lock(&q->unused_hctx_lock); | |
3145 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3146 | if (tmp->numa_node == node) { | |
3147 | hctx = tmp; | |
3148 | break; | |
3149 | } | |
3150 | } | |
3151 | if (hctx) | |
3152 | list_del_init(&hctx->hctx_list); | |
3153 | spin_unlock(&q->unused_hctx_lock); | |
3154 | ||
3155 | if (!hctx) | |
3156 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3157 | if (!hctx) |
7c6c5b7c | 3158 | goto fail; |
34d11ffa | 3159 | |
7c6c5b7c ML |
3160 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3161 | goto free_hctx; | |
34d11ffa JW |
3162 | |
3163 | return hctx; | |
7c6c5b7c ML |
3164 | |
3165 | free_hctx: | |
3166 | kobject_put(&hctx->kobj); | |
3167 | fail: | |
3168 | return NULL; | |
34d11ffa JW |
3169 | } |
3170 | ||
868f2f0b KB |
3171 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3172 | struct request_queue *q) | |
320ae51f | 3173 | { |
e01ad46d | 3174 | int i, j, end; |
868f2f0b | 3175 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3176 | |
ac0d6b92 BVA |
3177 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3178 | struct blk_mq_hw_ctx **new_hctxs; | |
3179 | ||
3180 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3181 | sizeof(*new_hctxs), GFP_KERNEL, | |
3182 | set->numa_node); | |
3183 | if (!new_hctxs) | |
3184 | return; | |
3185 | if (hctxs) | |
3186 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3187 | sizeof(*hctxs)); | |
3188 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3189 | kfree(hctxs); |
3190 | hctxs = new_hctxs; | |
3191 | } | |
3192 | ||
fb350e0a ML |
3193 | /* protect against switching io scheduler */ |
3194 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3195 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3196 | int node; |
34d11ffa | 3197 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3198 | |
7d76f856 | 3199 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3200 | /* |
3201 | * If the hw queue has been mapped to another numa node, | |
3202 | * we need to realloc the hctx. If allocation fails, fallback | |
3203 | * to use the previous one. | |
3204 | */ | |
3205 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3206 | continue; | |
868f2f0b | 3207 | |
34d11ffa JW |
3208 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3209 | if (hctx) { | |
2f8f1336 | 3210 | if (hctxs[i]) |
34d11ffa | 3211 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3212 | hctxs[i] = hctx; |
3213 | } else { | |
3214 | if (hctxs[i]) | |
3215 | pr_warn("Allocate new hctx on node %d fails,\ | |
3216 | fallback to previous one on node %d\n", | |
3217 | node, hctxs[i]->numa_node); | |
3218 | else | |
3219 | break; | |
868f2f0b | 3220 | } |
320ae51f | 3221 | } |
e01ad46d JW |
3222 | /* |
3223 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3224 | * hctxs and keep the previous q->nr_hw_queues. | |
3225 | */ | |
3226 | if (i != set->nr_hw_queues) { | |
3227 | j = q->nr_hw_queues; | |
3228 | end = i; | |
3229 | } else { | |
3230 | j = i; | |
3231 | end = q->nr_hw_queues; | |
3232 | q->nr_hw_queues = set->nr_hw_queues; | |
3233 | } | |
34d11ffa | 3234 | |
e01ad46d | 3235 | for (; j < end; j++) { |
868f2f0b KB |
3236 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3237 | ||
3238 | if (hctx) { | |
cc71a6f4 JA |
3239 | if (hctx->tags) |
3240 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 3241 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3242 | hctxs[j] = NULL; |
868f2f0b KB |
3243 | } |
3244 | } | |
fb350e0a | 3245 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3246 | } |
3247 | ||
26a9750a CH |
3248 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
3249 | struct request_queue *q) | |
868f2f0b | 3250 | { |
66841672 ML |
3251 | /* mark the queue as mq asap */ |
3252 | q->mq_ops = set->ops; | |
3253 | ||
34dbad5d | 3254 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3255 | blk_mq_poll_stats_bkt, |
3256 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3257 | if (!q->poll_cb) |
3258 | goto err_exit; | |
3259 | ||
1db4909e | 3260 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3261 | goto err_poll; |
868f2f0b | 3262 | |
737f98cf ML |
3263 | /* init q->mq_kobj and sw queues' kobjects */ |
3264 | blk_mq_sysfs_init(q); | |
3265 | ||
2f8f1336 ML |
3266 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3267 | spin_lock_init(&q->unused_hctx_lock); | |
3268 | ||
868f2f0b KB |
3269 | blk_mq_realloc_hw_ctxs(set, q); |
3270 | if (!q->nr_hw_queues) | |
3271 | goto err_hctxs; | |
320ae51f | 3272 | |
287922eb | 3273 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3274 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3275 | |
a8908939 | 3276 | q->tag_set = set; |
320ae51f | 3277 | |
94eddfbe | 3278 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3279 | if (set->nr_maps > HCTX_TYPE_POLL && |
3280 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3281 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3282 | |
2849450a | 3283 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3284 | INIT_LIST_HEAD(&q->requeue_list); |
3285 | spin_lock_init(&q->requeue_lock); | |
3286 | ||
eba71768 JA |
3287 | q->nr_requests = set->queue_depth; |
3288 | ||
64f1c21e JA |
3289 | /* |
3290 | * Default to classic polling | |
3291 | */ | |
29ece8b4 | 3292 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3293 | |
24d2f903 | 3294 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3295 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3296 | blk_mq_map_swqueue(q); |
26a9750a | 3297 | return 0; |
18741986 | 3298 | |
320ae51f | 3299 | err_hctxs: |
868f2f0b | 3300 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3301 | q->nr_hw_queues = 0; |
1db4909e | 3302 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3303 | err_poll: |
3304 | blk_stat_free_callback(q->poll_cb); | |
3305 | q->poll_cb = NULL; | |
c7de5726 ML |
3306 | err_exit: |
3307 | q->mq_ops = NULL; | |
26a9750a | 3308 | return -ENOMEM; |
320ae51f | 3309 | } |
b62c21b7 | 3310 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3311 | |
c7e2d94b ML |
3312 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3313 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3314 | { |
630ef623 | 3315 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3316 | |
630ef623 | 3317 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 3318 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
3319 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
3320 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 3321 | } |
320ae51f | 3322 | |
a5164405 JA |
3323 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3324 | { | |
3325 | int i; | |
3326 | ||
8229cca8 | 3327 | for (i = 0; i < set->nr_hw_queues; i++) { |
03b63b02 | 3328 | if (!__blk_mq_alloc_map_and_request(set, i)) |
a5164405 | 3329 | goto out_unwind; |
8229cca8 XT |
3330 | cond_resched(); |
3331 | } | |
a5164405 JA |
3332 | |
3333 | return 0; | |
3334 | ||
3335 | out_unwind: | |
3336 | while (--i >= 0) | |
2e194422 | 3337 | blk_mq_free_map_and_requests(set, i); |
a5164405 | 3338 | |
a5164405 JA |
3339 | return -ENOMEM; |
3340 | } | |
3341 | ||
3342 | /* | |
3343 | * Allocate the request maps associated with this tag_set. Note that this | |
3344 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3345 | * will be updated to reflect the allocated depth. | |
3346 | */ | |
79fab528 | 3347 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) |
a5164405 JA |
3348 | { |
3349 | unsigned int depth; | |
3350 | int err; | |
3351 | ||
3352 | depth = set->queue_depth; | |
3353 | do { | |
3354 | err = __blk_mq_alloc_rq_maps(set); | |
3355 | if (!err) | |
3356 | break; | |
3357 | ||
3358 | set->queue_depth >>= 1; | |
3359 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3360 | err = -ENOMEM; | |
3361 | break; | |
3362 | } | |
3363 | } while (set->queue_depth); | |
3364 | ||
3365 | if (!set->queue_depth || err) { | |
3366 | pr_err("blk-mq: failed to allocate request map\n"); | |
3367 | return -ENOMEM; | |
3368 | } | |
3369 | ||
3370 | if (depth != set->queue_depth) | |
3371 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3372 | depth, set->queue_depth); | |
3373 | ||
3374 | return 0; | |
3375 | } | |
3376 | ||
ebe8bddb OS |
3377 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3378 | { | |
6e66b493 BVA |
3379 | /* |
3380 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3381 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3382 | * number of hardware queues. | |
3383 | */ | |
3384 | if (set->nr_maps == 1) | |
3385 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3386 | ||
59388702 | 3387 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3388 | int i; |
3389 | ||
7d4901a9 ML |
3390 | /* |
3391 | * transport .map_queues is usually done in the following | |
3392 | * way: | |
3393 | * | |
3394 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3395 | * mask = get_cpu_mask(queue) | |
3396 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3397 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3398 | * } |
3399 | * | |
3400 | * When we need to remap, the table has to be cleared for | |
3401 | * killing stale mapping since one CPU may not be mapped | |
3402 | * to any hw queue. | |
3403 | */ | |
b3c661b1 JA |
3404 | for (i = 0; i < set->nr_maps; i++) |
3405 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3406 | |
ebe8bddb | 3407 | return set->ops->map_queues(set); |
b3c661b1 JA |
3408 | } else { |
3409 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3410 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3411 | } |
ebe8bddb OS |
3412 | } |
3413 | ||
f7e76dbc BVA |
3414 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3415 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3416 | { | |
3417 | struct blk_mq_tags **new_tags; | |
3418 | ||
3419 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3420 | return 0; | |
3421 | ||
3422 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3423 | GFP_KERNEL, set->numa_node); | |
3424 | if (!new_tags) | |
3425 | return -ENOMEM; | |
3426 | ||
3427 | if (set->tags) | |
3428 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3429 | sizeof(*set->tags)); | |
3430 | kfree(set->tags); | |
3431 | set->tags = new_tags; | |
3432 | set->nr_hw_queues = new_nr_hw_queues; | |
3433 | ||
3434 | return 0; | |
3435 | } | |
3436 | ||
91cdf265 MI |
3437 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3438 | int new_nr_hw_queues) | |
3439 | { | |
3440 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3441 | } | |
3442 | ||
a4391c64 JA |
3443 | /* |
3444 | * Alloc a tag set to be associated with one or more request queues. | |
3445 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3446 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3447 | * value will be stored in set->queue_depth. |
3448 | */ | |
24d2f903 CH |
3449 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3450 | { | |
b3c661b1 | 3451 | int i, ret; |
da695ba2 | 3452 | |
205fb5f5 BVA |
3453 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3454 | ||
24d2f903 CH |
3455 | if (!set->nr_hw_queues) |
3456 | return -EINVAL; | |
a4391c64 | 3457 | if (!set->queue_depth) |
24d2f903 CH |
3458 | return -EINVAL; |
3459 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3460 | return -EINVAL; | |
3461 | ||
7d7e0f90 | 3462 | if (!set->ops->queue_rq) |
24d2f903 CH |
3463 | return -EINVAL; |
3464 | ||
de148297 ML |
3465 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3466 | return -EINVAL; | |
3467 | ||
a4391c64 JA |
3468 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3469 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3470 | BLK_MQ_MAX_DEPTH); | |
3471 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3472 | } | |
24d2f903 | 3473 | |
b3c661b1 JA |
3474 | if (!set->nr_maps) |
3475 | set->nr_maps = 1; | |
3476 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3477 | return -EINVAL; | |
3478 | ||
6637fadf SL |
3479 | /* |
3480 | * If a crashdump is active, then we are potentially in a very | |
3481 | * memory constrained environment. Limit us to 1 queue and | |
3482 | * 64 tags to prevent using too much memory. | |
3483 | */ | |
3484 | if (is_kdump_kernel()) { | |
3485 | set->nr_hw_queues = 1; | |
59388702 | 3486 | set->nr_maps = 1; |
6637fadf SL |
3487 | set->queue_depth = min(64U, set->queue_depth); |
3488 | } | |
868f2f0b | 3489 | /* |
392546ae JA |
3490 | * There is no use for more h/w queues than cpus if we just have |
3491 | * a single map | |
868f2f0b | 3492 | */ |
392546ae | 3493 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3494 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3495 | |
91cdf265 | 3496 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3497 | return -ENOMEM; |
24d2f903 | 3498 | |
da695ba2 | 3499 | ret = -ENOMEM; |
b3c661b1 JA |
3500 | for (i = 0; i < set->nr_maps; i++) { |
3501 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3502 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3503 | GFP_KERNEL, set->numa_node); |
3504 | if (!set->map[i].mq_map) | |
3505 | goto out_free_mq_map; | |
59388702 | 3506 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3507 | } |
bdd17e75 | 3508 | |
ebe8bddb | 3509 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3510 | if (ret) |
3511 | goto out_free_mq_map; | |
3512 | ||
79fab528 | 3513 | ret = blk_mq_alloc_map_and_requests(set); |
da695ba2 | 3514 | if (ret) |
bdd17e75 | 3515 | goto out_free_mq_map; |
24d2f903 | 3516 | |
32bc15af | 3517 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
f1b49fdc JG |
3518 | atomic_set(&set->active_queues_shared_sbitmap, 0); |
3519 | ||
56b68085 | 3520 | if (blk_mq_init_shared_sbitmap(set)) { |
32bc15af JG |
3521 | ret = -ENOMEM; |
3522 | goto out_free_mq_rq_maps; | |
3523 | } | |
3524 | } | |
3525 | ||
0d2602ca JA |
3526 | mutex_init(&set->tag_list_lock); |
3527 | INIT_LIST_HEAD(&set->tag_list); | |
3528 | ||
24d2f903 | 3529 | return 0; |
bdd17e75 | 3530 | |
32bc15af JG |
3531 | out_free_mq_rq_maps: |
3532 | for (i = 0; i < set->nr_hw_queues; i++) | |
3533 | blk_mq_free_map_and_requests(set, i); | |
bdd17e75 | 3534 | out_free_mq_map: |
b3c661b1 JA |
3535 | for (i = 0; i < set->nr_maps; i++) { |
3536 | kfree(set->map[i].mq_map); | |
3537 | set->map[i].mq_map = NULL; | |
3538 | } | |
5676e7b6 RE |
3539 | kfree(set->tags); |
3540 | set->tags = NULL; | |
da695ba2 | 3541 | return ret; |
24d2f903 CH |
3542 | } |
3543 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3544 | ||
cdb14e0f CH |
3545 | /* allocate and initialize a tagset for a simple single-queue device */ |
3546 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
3547 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
3548 | unsigned int set_flags) | |
3549 | { | |
3550 | memset(set, 0, sizeof(*set)); | |
3551 | set->ops = ops; | |
3552 | set->nr_hw_queues = 1; | |
3553 | set->nr_maps = 1; | |
3554 | set->queue_depth = queue_depth; | |
3555 | set->numa_node = NUMA_NO_NODE; | |
3556 | set->flags = set_flags; | |
3557 | return blk_mq_alloc_tag_set(set); | |
3558 | } | |
3559 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
3560 | ||
24d2f903 CH |
3561 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
3562 | { | |
b3c661b1 | 3563 | int i, j; |
24d2f903 | 3564 | |
f7e76dbc | 3565 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3566 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3567 | |
32bc15af JG |
3568 | if (blk_mq_is_sbitmap_shared(set->flags)) |
3569 | blk_mq_exit_shared_sbitmap(set); | |
3570 | ||
b3c661b1 JA |
3571 | for (j = 0; j < set->nr_maps; j++) { |
3572 | kfree(set->map[j].mq_map); | |
3573 | set->map[j].mq_map = NULL; | |
3574 | } | |
bdd17e75 | 3575 | |
981bd189 | 3576 | kfree(set->tags); |
5676e7b6 | 3577 | set->tags = NULL; |
24d2f903 CH |
3578 | } |
3579 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3580 | ||
e3a2b3f9 JA |
3581 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3582 | { | |
3583 | struct blk_mq_tag_set *set = q->tag_set; | |
3584 | struct blk_mq_hw_ctx *hctx; | |
3585 | int i, ret; | |
3586 | ||
bd166ef1 | 3587 | if (!set) |
e3a2b3f9 JA |
3588 | return -EINVAL; |
3589 | ||
e5fa8140 AZ |
3590 | if (q->nr_requests == nr) |
3591 | return 0; | |
3592 | ||
70f36b60 | 3593 | blk_mq_freeze_queue(q); |
24f5a90f | 3594 | blk_mq_quiesce_queue(q); |
70f36b60 | 3595 | |
e3a2b3f9 JA |
3596 | ret = 0; |
3597 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3598 | if (!hctx->tags) |
3599 | continue; | |
bd166ef1 JA |
3600 | /* |
3601 | * If we're using an MQ scheduler, just update the scheduler | |
3602 | * queue depth. This is similar to what the old code would do. | |
3603 | */ | |
70f36b60 | 3604 | if (!hctx->sched_tags) { |
c2e82a23 | 3605 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 | 3606 | false); |
32bc15af JG |
3607 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) |
3608 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
70f36b60 JA |
3609 | } else { |
3610 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3611 | nr, true); | |
d97e594c JG |
3612 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
3613 | hctx->sched_tags->bitmap_tags = | |
3614 | &q->sched_bitmap_tags; | |
3615 | hctx->sched_tags->breserved_tags = | |
3616 | &q->sched_breserved_tags; | |
3617 | } | |
70f36b60 | 3618 | } |
e3a2b3f9 JA |
3619 | if (ret) |
3620 | break; | |
77f1e0a5 JA |
3621 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3622 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 3623 | } |
d97e594c | 3624 | if (!ret) { |
e3a2b3f9 | 3625 | q->nr_requests = nr; |
d97e594c JG |
3626 | if (q->elevator && blk_mq_is_sbitmap_shared(set->flags)) |
3627 | sbitmap_queue_resize(&q->sched_bitmap_tags, | |
3628 | nr - set->reserved_tags); | |
3629 | } | |
e3a2b3f9 | 3630 | |
24f5a90f | 3631 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3632 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3633 | |
e3a2b3f9 JA |
3634 | return ret; |
3635 | } | |
3636 | ||
d48ece20 JW |
3637 | /* |
3638 | * request_queue and elevator_type pair. | |
3639 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3640 | * the elevator_type associated with a request_queue. | |
3641 | */ | |
3642 | struct blk_mq_qe_pair { | |
3643 | struct list_head node; | |
3644 | struct request_queue *q; | |
3645 | struct elevator_type *type; | |
3646 | }; | |
3647 | ||
3648 | /* | |
3649 | * Cache the elevator_type in qe pair list and switch the | |
3650 | * io scheduler to 'none' | |
3651 | */ | |
3652 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3653 | struct request_queue *q) | |
3654 | { | |
3655 | struct blk_mq_qe_pair *qe; | |
3656 | ||
3657 | if (!q->elevator) | |
3658 | return true; | |
3659 | ||
3660 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3661 | if (!qe) | |
3662 | return false; | |
3663 | ||
3664 | INIT_LIST_HEAD(&qe->node); | |
3665 | qe->q = q; | |
3666 | qe->type = q->elevator->type; | |
3667 | list_add(&qe->node, head); | |
3668 | ||
3669 | mutex_lock(&q->sysfs_lock); | |
3670 | /* | |
3671 | * After elevator_switch_mq, the previous elevator_queue will be | |
3672 | * released by elevator_release. The reference of the io scheduler | |
3673 | * module get by elevator_get will also be put. So we need to get | |
3674 | * a reference of the io scheduler module here to prevent it to be | |
3675 | * removed. | |
3676 | */ | |
3677 | __module_get(qe->type->elevator_owner); | |
3678 | elevator_switch_mq(q, NULL); | |
3679 | mutex_unlock(&q->sysfs_lock); | |
3680 | ||
3681 | return true; | |
3682 | } | |
3683 | ||
3684 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3685 | struct request_queue *q) | |
3686 | { | |
3687 | struct blk_mq_qe_pair *qe; | |
3688 | struct elevator_type *t = NULL; | |
3689 | ||
3690 | list_for_each_entry(qe, head, node) | |
3691 | if (qe->q == q) { | |
3692 | t = qe->type; | |
3693 | break; | |
3694 | } | |
3695 | ||
3696 | if (!t) | |
3697 | return; | |
3698 | ||
3699 | list_del(&qe->node); | |
3700 | kfree(qe); | |
3701 | ||
3702 | mutex_lock(&q->sysfs_lock); | |
3703 | elevator_switch_mq(q, t); | |
3704 | mutex_unlock(&q->sysfs_lock); | |
3705 | } | |
3706 | ||
e4dc2b32 KB |
3707 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3708 | int nr_hw_queues) | |
868f2f0b KB |
3709 | { |
3710 | struct request_queue *q; | |
d48ece20 | 3711 | LIST_HEAD(head); |
e01ad46d | 3712 | int prev_nr_hw_queues; |
868f2f0b | 3713 | |
705cda97 BVA |
3714 | lockdep_assert_held(&set->tag_list_lock); |
3715 | ||
392546ae | 3716 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3717 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3718 | if (nr_hw_queues < 1) |
3719 | return; | |
3720 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3721 | return; |
3722 | ||
3723 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3724 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3725 | /* |
3726 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3727 | * with the previous scheduler. We will switch back once we are done | |
3728 | * updating the new sw to hw queue mappings. | |
3729 | */ | |
3730 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3731 | if (!blk_mq_elv_switch_none(&head, q)) | |
3732 | goto switch_back; | |
868f2f0b | 3733 | |
477e19de JW |
3734 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3735 | blk_mq_debugfs_unregister_hctxs(q); | |
3736 | blk_mq_sysfs_unregister(q); | |
3737 | } | |
3738 | ||
a2584e43 | 3739 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3740 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3741 | 0) | |
3742 | goto reregister; | |
3743 | ||
868f2f0b | 3744 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3745 | fallback: |
aa880ad6 | 3746 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3747 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3748 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3749 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3750 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3751 | nr_hw_queues, prev_nr_hw_queues); | |
3752 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3753 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3754 | goto fallback; |
3755 | } | |
477e19de JW |
3756 | blk_mq_map_swqueue(q); |
3757 | } | |
3758 | ||
f7e76dbc | 3759 | reregister: |
477e19de JW |
3760 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3761 | blk_mq_sysfs_register(q); | |
3762 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3763 | } |
3764 | ||
d48ece20 JW |
3765 | switch_back: |
3766 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3767 | blk_mq_elv_switch_back(&head, q); | |
3768 | ||
868f2f0b KB |
3769 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3770 | blk_mq_unfreeze_queue(q); | |
3771 | } | |
e4dc2b32 KB |
3772 | |
3773 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3774 | { | |
3775 | mutex_lock(&set->tag_list_lock); | |
3776 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3777 | mutex_unlock(&set->tag_list_lock); | |
3778 | } | |
868f2f0b KB |
3779 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3780 | ||
34dbad5d OS |
3781 | /* Enable polling stats and return whether they were already enabled. */ |
3782 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3783 | { | |
3784 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3785 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3786 | return true; |
3787 | blk_stat_add_callback(q, q->poll_cb); | |
3788 | return false; | |
3789 | } | |
3790 | ||
3791 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3792 | { | |
3793 | /* | |
3794 | * We don't arm the callback if polling stats are not enabled or the | |
3795 | * callback is already active. | |
3796 | */ | |
3797 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3798 | blk_stat_is_active(q->poll_cb)) | |
3799 | return; | |
3800 | ||
3801 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3802 | } | |
3803 | ||
3804 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3805 | { | |
3806 | struct request_queue *q = cb->data; | |
720b8ccc | 3807 | int bucket; |
34dbad5d | 3808 | |
720b8ccc SB |
3809 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3810 | if (cb->stat[bucket].nr_samples) | |
3811 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3812 | } | |
34dbad5d OS |
3813 | } |
3814 | ||
64f1c21e | 3815 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3816 | struct request *rq) |
3817 | { | |
64f1c21e | 3818 | unsigned long ret = 0; |
720b8ccc | 3819 | int bucket; |
64f1c21e JA |
3820 | |
3821 | /* | |
3822 | * If stats collection isn't on, don't sleep but turn it on for | |
3823 | * future users | |
3824 | */ | |
34dbad5d | 3825 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3826 | return 0; |
3827 | ||
64f1c21e JA |
3828 | /* |
3829 | * As an optimistic guess, use half of the mean service time | |
3830 | * for this type of request. We can (and should) make this smarter. | |
3831 | * For instance, if the completion latencies are tight, we can | |
3832 | * get closer than just half the mean. This is especially | |
3833 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3834 | * than ~10 usec. We do use the stats for the relevant IO size |
3835 | * if available which does lead to better estimates. | |
64f1c21e | 3836 | */ |
720b8ccc SB |
3837 | bucket = blk_mq_poll_stats_bkt(rq); |
3838 | if (bucket < 0) | |
3839 | return ret; | |
3840 | ||
3841 | if (q->poll_stat[bucket].nr_samples) | |
3842 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3843 | |
3844 | return ret; | |
3845 | } | |
3846 | ||
06426adf JA |
3847 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3848 | struct request *rq) | |
3849 | { | |
3850 | struct hrtimer_sleeper hs; | |
3851 | enum hrtimer_mode mode; | |
64f1c21e | 3852 | unsigned int nsecs; |
06426adf JA |
3853 | ktime_t kt; |
3854 | ||
76a86f9d | 3855 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3856 | return false; |
3857 | ||
3858 | /* | |
1052b8ac | 3859 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3860 | * |
64f1c21e JA |
3861 | * 0: use half of prev avg |
3862 | * >0: use this specific value | |
3863 | */ | |
1052b8ac | 3864 | if (q->poll_nsec > 0) |
64f1c21e JA |
3865 | nsecs = q->poll_nsec; |
3866 | else | |
cae740a0 | 3867 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3868 | |
3869 | if (!nsecs) | |
06426adf JA |
3870 | return false; |
3871 | ||
76a86f9d | 3872 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3873 | |
3874 | /* | |
3875 | * This will be replaced with the stats tracking code, using | |
3876 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3877 | */ | |
8b0e1953 | 3878 | kt = nsecs; |
06426adf JA |
3879 | |
3880 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3881 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3882 | hrtimer_set_expires(&hs.timer, kt); |
3883 | ||
06426adf | 3884 | do { |
5a61c363 | 3885 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3886 | break; |
3887 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3888 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3889 | if (hs.task) |
3890 | io_schedule(); | |
3891 | hrtimer_cancel(&hs.timer); | |
3892 | mode = HRTIMER_MODE_ABS; | |
3893 | } while (hs.task && !signal_pending(current)); | |
3894 | ||
3895 | __set_current_state(TASK_RUNNING); | |
3896 | destroy_hrtimer_on_stack(&hs.timer); | |
3897 | return true; | |
3898 | } | |
3899 | ||
1052b8ac JA |
3900 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3901 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3902 | { |
1052b8ac JA |
3903 | struct request *rq; |
3904 | ||
29ece8b4 | 3905 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3906 | return false; |
3907 | ||
3908 | if (!blk_qc_t_is_internal(cookie)) | |
3909 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3910 | else { | |
3911 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3912 | /* | |
3913 | * With scheduling, if the request has completed, we'll | |
3914 | * get a NULL return here, as we clear the sched tag when | |
3915 | * that happens. The request still remains valid, like always, | |
3916 | * so we should be safe with just the NULL check. | |
3917 | */ | |
3918 | if (!rq) | |
3919 | return false; | |
3920 | } | |
3921 | ||
cae740a0 | 3922 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3923 | } |
3924 | ||
529262d5 CH |
3925 | /** |
3926 | * blk_poll - poll for IO completions | |
3927 | * @q: the queue | |
3928 | * @cookie: cookie passed back at IO submission time | |
3929 | * @spin: whether to spin for completions | |
3930 | * | |
3931 | * Description: | |
3932 | * Poll for completions on the passed in queue. Returns number of | |
3933 | * completed entries found. If @spin is true, then blk_poll will continue | |
3934 | * looping until at least one completion is found, unless the task is | |
3935 | * otherwise marked running (or we need to reschedule). | |
3936 | */ | |
3937 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3938 | { |
3939 | struct blk_mq_hw_ctx *hctx; | |
2f064a59 | 3940 | unsigned int state; |
bbd7bb70 | 3941 | |
529262d5 CH |
3942 | if (!blk_qc_t_valid(cookie) || |
3943 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3944 | return 0; |
3945 | ||
529262d5 CH |
3946 | if (current->plug) |
3947 | blk_flush_plug_list(current->plug, false); | |
3948 | ||
1052b8ac JA |
3949 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3950 | ||
06426adf JA |
3951 | /* |
3952 | * If we sleep, have the caller restart the poll loop to reset | |
3953 | * the state. Like for the other success return cases, the | |
3954 | * caller is responsible for checking if the IO completed. If | |
3955 | * the IO isn't complete, we'll get called again and will go | |
f6f371f7 PB |
3956 | * straight to the busy poll loop. If specified not to spin, |
3957 | * we also should not sleep. | |
06426adf | 3958 | */ |
f6f371f7 | 3959 | if (spin && blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3960 | return 1; |
06426adf | 3961 | |
bbd7bb70 JA |
3962 | hctx->poll_considered++; |
3963 | ||
d6c23bb3 | 3964 | state = get_current_state(); |
aa61bec3 | 3965 | do { |
bbd7bb70 JA |
3966 | int ret; |
3967 | ||
3968 | hctx->poll_invoked++; | |
3969 | ||
9743139c | 3970 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3971 | if (ret > 0) { |
3972 | hctx->poll_success++; | |
849a3700 | 3973 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3974 | return ret; |
bbd7bb70 JA |
3975 | } |
3976 | ||
3977 | if (signal_pending_state(state, current)) | |
849a3700 | 3978 | __set_current_state(TASK_RUNNING); |
bbd7bb70 | 3979 | |
b03fbd4f | 3980 | if (task_is_running(current)) |
85f4d4b6 | 3981 | return 1; |
0a1b8b87 | 3982 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3983 | break; |
3984 | cpu_relax(); | |
aa61bec3 | 3985 | } while (!need_resched()); |
bbd7bb70 | 3986 | |
67b4110f | 3987 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3988 | return 0; |
bbd7bb70 | 3989 | } |
529262d5 | 3990 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3991 | |
9cf2bab6 JA |
3992 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3993 | { | |
3994 | return rq->mq_ctx->cpu; | |
3995 | } | |
3996 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3997 | ||
320ae51f JA |
3998 | static int __init blk_mq_init(void) |
3999 | { | |
c3077b5d CH |
4000 | int i; |
4001 | ||
4002 | for_each_possible_cpu(i) | |
f9ab4918 | 4003 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4004 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4005 | ||
4006 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4007 | "block/softirq:dead", NULL, | |
4008 | blk_softirq_cpu_dead); | |
9467f859 TG |
4009 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4010 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4011 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4012 | blk_mq_hctx_notify_online, | |
4013 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4014 | return 0; |
4015 | } | |
4016 | subsys_initcall(blk_mq_init); |