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> |
320ae51f JA |
29 | |
30 | #include <trace/events/block.h> | |
31 | ||
32 | #include <linux/blk-mq.h> | |
33 | #include "blk.h" | |
34 | #include "blk-mq.h" | |
9c1051aa | 35 | #include "blk-mq-debugfs.h" |
320ae51f | 36 | #include "blk-mq-tag.h" |
986d413b | 37 | #include "blk-pm.h" |
cf43e6be | 38 | #include "blk-stat.h" |
bd166ef1 | 39 | #include "blk-mq-sched.h" |
c1c80384 | 40 | #include "blk-rq-qos.h" |
320ae51f | 41 | |
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
720b8ccc SB |
45 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
46 | { | |
3d244306 | 47 | int ddir, sectors, bucket; |
720b8ccc | 48 | |
99c749a4 | 49 | ddir = rq_data_dir(rq); |
3d244306 | 50 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 51 | |
3d244306 | 52 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
53 | |
54 | if (bucket < 0) | |
55 | return -1; | |
56 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
57 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
58 | ||
59 | return bucket; | |
60 | } | |
61 | ||
320ae51f | 62 | /* |
85fae294 YY |
63 | * Check if any of the ctx, dispatch list or elevator |
64 | * have pending work in this hardware queue. | |
320ae51f | 65 | */ |
79f720a7 | 66 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 67 | { |
79f720a7 JA |
68 | return !list_empty_careful(&hctx->dispatch) || |
69 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 70 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
71 | } |
72 | ||
320ae51f JA |
73 | /* |
74 | * Mark this ctx as having pending work in this hardware queue | |
75 | */ | |
76 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
77 | struct blk_mq_ctx *ctx) | |
78 | { | |
f31967f0 JA |
79 | const int bit = ctx->index_hw[hctx->type]; |
80 | ||
81 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
82 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
83 | } |
84 | ||
85 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
86 | struct blk_mq_ctx *ctx) | |
87 | { | |
f31967f0 JA |
88 | const int bit = ctx->index_hw[hctx->type]; |
89 | ||
90 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
91 | } |
92 | ||
f299b7c7 JA |
93 | struct mq_inflight { |
94 | struct hd_struct *part; | |
95 | unsigned int *inflight; | |
96 | }; | |
97 | ||
7baa8572 | 98 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
99 | struct request *rq, void *priv, |
100 | bool reserved) | |
101 | { | |
102 | struct mq_inflight *mi = priv; | |
103 | ||
6131837b | 104 | /* |
e016b782 | 105 | * index[0] counts the specific partition that was asked for. |
6131837b OS |
106 | */ |
107 | if (rq->part == mi->part) | |
108 | mi->inflight[0]++; | |
7baa8572 JA |
109 | |
110 | return true; | |
f299b7c7 JA |
111 | } |
112 | ||
e016b782 | 113 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) |
f299b7c7 | 114 | { |
e016b782 | 115 | unsigned inflight[2]; |
f299b7c7 JA |
116 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; |
117 | ||
b8d62b3a | 118 | inflight[0] = inflight[1] = 0; |
f299b7c7 | 119 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 MP |
120 | |
121 | return inflight[0]; | |
f299b7c7 JA |
122 | } |
123 | ||
7baa8572 | 124 | static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx, |
bf0ddaba OS |
125 | struct request *rq, void *priv, |
126 | bool reserved) | |
127 | { | |
128 | struct mq_inflight *mi = priv; | |
129 | ||
130 | if (rq->part == mi->part) | |
131 | mi->inflight[rq_data_dir(rq)]++; | |
7baa8572 JA |
132 | |
133 | return true; | |
bf0ddaba OS |
134 | } |
135 | ||
136 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
137 | unsigned int inflight[2]) | |
138 | { | |
139 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
140 | ||
141 | inflight[0] = inflight[1] = 0; | |
142 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi); | |
143 | } | |
144 | ||
1671d522 | 145 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 146 | { |
7996a8b5 BL |
147 | mutex_lock(&q->mq_freeze_lock); |
148 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 149 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 150 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 151 | if (queue_is_mq(q)) |
055f6e18 | 152 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
153 | } else { |
154 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 155 | } |
f3af020b | 156 | } |
1671d522 | 157 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 158 | |
6bae363e | 159 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 160 | { |
3ef28e83 | 161 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 162 | } |
6bae363e | 163 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 164 | |
f91328c4 KB |
165 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
166 | unsigned long timeout) | |
167 | { | |
168 | return wait_event_timeout(q->mq_freeze_wq, | |
169 | percpu_ref_is_zero(&q->q_usage_counter), | |
170 | timeout); | |
171 | } | |
172 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 173 | |
f3af020b TH |
174 | /* |
175 | * Guarantee no request is in use, so we can change any data structure of | |
176 | * the queue afterward. | |
177 | */ | |
3ef28e83 | 178 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 179 | { |
3ef28e83 DW |
180 | /* |
181 | * In the !blk_mq case we are only calling this to kill the | |
182 | * q_usage_counter, otherwise this increases the freeze depth | |
183 | * and waits for it to return to zero. For this reason there is | |
184 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
185 | * exported to drivers as the only user for unfreeze is blk_mq. | |
186 | */ | |
1671d522 | 187 | blk_freeze_queue_start(q); |
f3af020b TH |
188 | blk_mq_freeze_queue_wait(q); |
189 | } | |
3ef28e83 DW |
190 | |
191 | void blk_mq_freeze_queue(struct request_queue *q) | |
192 | { | |
193 | /* | |
194 | * ...just an alias to keep freeze and unfreeze actions balanced | |
195 | * in the blk_mq_* namespace | |
196 | */ | |
197 | blk_freeze_queue(q); | |
198 | } | |
c761d96b | 199 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 200 | |
b4c6a028 | 201 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 202 | { |
7996a8b5 BL |
203 | mutex_lock(&q->mq_freeze_lock); |
204 | q->mq_freeze_depth--; | |
205 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
206 | if (!q->mq_freeze_depth) { | |
bdd63160 | 207 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 208 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 209 | } |
7996a8b5 | 210 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 211 | } |
b4c6a028 | 212 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 213 | |
852ec809 BVA |
214 | /* |
215 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
216 | * mpt3sas driver such that this function can be removed. | |
217 | */ | |
218 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
219 | { | |
8814ce8a | 220 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
221 | } |
222 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
223 | ||
6a83e74d | 224 | /** |
69e07c4a | 225 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
226 | * @q: request queue. |
227 | * | |
228 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
229 | * callback function is invoked. Once this function is returned, we make |
230 | * sure no dispatch can happen until the queue is unquiesced via | |
231 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
232 | */ |
233 | void blk_mq_quiesce_queue(struct request_queue *q) | |
234 | { | |
235 | struct blk_mq_hw_ctx *hctx; | |
236 | unsigned int i; | |
237 | bool rcu = false; | |
238 | ||
1d9e9bc6 | 239 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 240 | |
6a83e74d BVA |
241 | queue_for_each_hw_ctx(q, hctx, i) { |
242 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 243 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
244 | else |
245 | rcu = true; | |
246 | } | |
247 | if (rcu) | |
248 | synchronize_rcu(); | |
249 | } | |
250 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
251 | ||
e4e73913 ML |
252 | /* |
253 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
254 | * @q: request queue. | |
255 | * | |
256 | * This function recovers queue into the state before quiescing | |
257 | * which is done by blk_mq_quiesce_queue. | |
258 | */ | |
259 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
260 | { | |
8814ce8a | 261 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 262 | |
1d9e9bc6 ML |
263 | /* dispatch requests which are inserted during quiescing */ |
264 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
265 | } |
266 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
267 | ||
aed3ea94 JA |
268 | void blk_mq_wake_waiters(struct request_queue *q) |
269 | { | |
270 | struct blk_mq_hw_ctx *hctx; | |
271 | unsigned int i; | |
272 | ||
273 | queue_for_each_hw_ctx(q, hctx, i) | |
274 | if (blk_mq_hw_queue_mapped(hctx)) | |
275 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
276 | } | |
277 | ||
320ae51f JA |
278 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
279 | { | |
280 | return blk_mq_has_free_tags(hctx->tags); | |
281 | } | |
282 | EXPORT_SYMBOL(blk_mq_can_queue); | |
283 | ||
fe1f4526 | 284 | /* |
9a91b05b HT |
285 | * Only need start/end time stamping if we have iostat or |
286 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
287 | */ |
288 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
289 | { | |
9a91b05b | 290 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
291 | } |
292 | ||
e4cdf1a1 | 293 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
6f816b4b | 294 | unsigned int tag, unsigned int op, u64 alloc_time_ns) |
320ae51f | 295 | { |
e4cdf1a1 CH |
296 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
297 | struct request *rq = tags->static_rqs[tag]; | |
bf9ae8c5 | 298 | req_flags_t rq_flags = 0; |
c3a148d2 | 299 | |
e4cdf1a1 CH |
300 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
301 | rq->tag = -1; | |
302 | rq->internal_tag = tag; | |
303 | } else { | |
d263ed99 | 304 | if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) { |
bf9ae8c5 | 305 | rq_flags = RQF_MQ_INFLIGHT; |
e4cdf1a1 CH |
306 | atomic_inc(&data->hctx->nr_active); |
307 | } | |
308 | rq->tag = tag; | |
309 | rq->internal_tag = -1; | |
310 | data->hctx->tags->rqs[rq->tag] = rq; | |
311 | } | |
312 | ||
af76e555 | 313 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
314 | rq->q = data->q; |
315 | rq->mq_ctx = data->ctx; | |
ea4f995e | 316 | rq->mq_hctx = data->hctx; |
bf9ae8c5 | 317 | rq->rq_flags = rq_flags; |
ef295ecf | 318 | rq->cmd_flags = op; |
1b6d65a0 BVA |
319 | if (data->flags & BLK_MQ_REQ_PREEMPT) |
320 | rq->rq_flags |= RQF_PREEMPT; | |
e4cdf1a1 | 321 | if (blk_queue_io_stat(data->q)) |
e8064021 | 322 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 323 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
324 | INIT_HLIST_NODE(&rq->hash); |
325 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
326 | rq->rq_disk = NULL; |
327 | rq->part = NULL; | |
6f816b4b TH |
328 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
329 | rq->alloc_time_ns = alloc_time_ns; | |
330 | #endif | |
fe1f4526 JA |
331 | if (blk_mq_need_time_stamp(rq)) |
332 | rq->start_time_ns = ktime_get_ns(); | |
333 | else | |
334 | rq->start_time_ns = 0; | |
544ccc8d | 335 | rq->io_start_time_ns = 0; |
3d244306 | 336 | rq->stats_sectors = 0; |
af76e555 CH |
337 | rq->nr_phys_segments = 0; |
338 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
339 | rq->nr_integrity_segments = 0; | |
340 | #endif | |
af76e555 | 341 | /* tag was already set */ |
af76e555 | 342 | rq->extra_len = 0; |
079076b3 | 343 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 344 | |
f6be4fb4 JA |
345 | rq->timeout = 0; |
346 | ||
af76e555 CH |
347 | rq->end_io = NULL; |
348 | rq->end_io_data = NULL; | |
af76e555 | 349 | |
e4cdf1a1 | 350 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
12f5b931 | 351 | refcount_set(&rq->ref, 1); |
e4cdf1a1 | 352 | return rq; |
5dee8577 CH |
353 | } |
354 | ||
d2c0d383 | 355 | static struct request *blk_mq_get_request(struct request_queue *q, |
f9afca4d JA |
356 | struct bio *bio, |
357 | struct blk_mq_alloc_data *data) | |
d2c0d383 CH |
358 | { |
359 | struct elevator_queue *e = q->elevator; | |
360 | struct request *rq; | |
e4cdf1a1 | 361 | unsigned int tag; |
c05f4220 | 362 | bool clear_ctx_on_error = false; |
6f816b4b | 363 | u64 alloc_time_ns = 0; |
d2c0d383 CH |
364 | |
365 | blk_queue_enter_live(q); | |
6f816b4b TH |
366 | |
367 | /* alloc_time includes depth and tag waits */ | |
368 | if (blk_queue_rq_alloc_time(q)) | |
369 | alloc_time_ns = ktime_get_ns(); | |
370 | ||
d2c0d383 | 371 | data->q = q; |
21e768b4 BVA |
372 | if (likely(!data->ctx)) { |
373 | data->ctx = blk_mq_get_ctx(q); | |
c05f4220 | 374 | clear_ctx_on_error = true; |
21e768b4 | 375 | } |
d2c0d383 | 376 | if (likely(!data->hctx)) |
f9afca4d | 377 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, |
8ccdf4a3 | 378 | data->ctx); |
f9afca4d | 379 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 380 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
381 | |
382 | if (e) { | |
383 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
384 | ||
385 | /* | |
386 | * Flush requests are special and go directly to the | |
17a51199 JA |
387 | * dispatch list. Don't include reserved tags in the |
388 | * limiting, as it isn't useful. | |
d2c0d383 | 389 | */ |
f9afca4d JA |
390 | if (!op_is_flush(data->cmd_flags) && |
391 | e->type->ops.limit_depth && | |
17a51199 | 392 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 393 | e->type->ops.limit_depth(data->cmd_flags, data); |
d263ed99 JW |
394 | } else { |
395 | blk_mq_tag_busy(data->hctx); | |
d2c0d383 CH |
396 | } |
397 | ||
e4cdf1a1 CH |
398 | tag = blk_mq_get_tag(data); |
399 | if (tag == BLK_MQ_TAG_FAIL) { | |
c05f4220 | 400 | if (clear_ctx_on_error) |
1ad43c00 | 401 | data->ctx = NULL; |
037cebb8 CH |
402 | blk_queue_exit(q); |
403 | return NULL; | |
d2c0d383 CH |
404 | } |
405 | ||
6f816b4b | 406 | rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns); |
f9afca4d | 407 | if (!op_is_flush(data->cmd_flags)) { |
037cebb8 | 408 | rq->elv.icq = NULL; |
f9cd4bfe | 409 | if (e && e->type->ops.prepare_request) { |
e2b3fa5a DLM |
410 | if (e->type->icq_cache) |
411 | blk_mq_sched_assign_ioc(rq); | |
44e8c2bf | 412 | |
f9cd4bfe | 413 | e->type->ops.prepare_request(rq, bio); |
5bbf4e5a | 414 | rq->rq_flags |= RQF_ELVPRIV; |
44e8c2bf | 415 | } |
037cebb8 CH |
416 | } |
417 | data->hctx->queued++; | |
418 | return rq; | |
d2c0d383 CH |
419 | } |
420 | ||
cd6ce148 | 421 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 422 | blk_mq_req_flags_t flags) |
320ae51f | 423 | { |
f9afca4d | 424 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
bd166ef1 | 425 | struct request *rq; |
a492f075 | 426 | int ret; |
320ae51f | 427 | |
3a0a5299 | 428 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
429 | if (ret) |
430 | return ERR_PTR(ret); | |
320ae51f | 431 | |
f9afca4d | 432 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 433 | blk_queue_exit(q); |
841bac2c | 434 | |
bd166ef1 | 435 | if (!rq) |
a492f075 | 436 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
437 | |
438 | rq->__data_len = 0; | |
439 | rq->__sector = (sector_t) -1; | |
440 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
441 | return rq; |
442 | } | |
4bb659b1 | 443 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 444 | |
cd6ce148 | 445 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 446 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 447 | { |
f9afca4d | 448 | struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op }; |
1f5bd336 | 449 | struct request *rq; |
6d2809d5 | 450 | unsigned int cpu; |
1f5bd336 ML |
451 | int ret; |
452 | ||
453 | /* | |
454 | * If the tag allocator sleeps we could get an allocation for a | |
455 | * different hardware context. No need to complicate the low level | |
456 | * allocator for this for the rare use case of a command tied to | |
457 | * a specific queue. | |
458 | */ | |
459 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
460 | return ERR_PTR(-EINVAL); | |
461 | ||
462 | if (hctx_idx >= q->nr_hw_queues) | |
463 | return ERR_PTR(-EIO); | |
464 | ||
3a0a5299 | 465 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
466 | if (ret) |
467 | return ERR_PTR(ret); | |
468 | ||
c8712c6a CH |
469 | /* |
470 | * Check if the hardware context is actually mapped to anything. | |
471 | * If not tell the caller that it should skip this queue. | |
472 | */ | |
6d2809d5 OS |
473 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
474 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
475 | blk_queue_exit(q); | |
476 | return ERR_PTR(-EXDEV); | |
c8712c6a | 477 | } |
20e4d813 | 478 | cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask); |
6d2809d5 | 479 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); |
1f5bd336 | 480 | |
f9afca4d | 481 | rq = blk_mq_get_request(q, NULL, &alloc_data); |
3280d66a | 482 | blk_queue_exit(q); |
c8712c6a | 483 | |
6d2809d5 OS |
484 | if (!rq) |
485 | return ERR_PTR(-EWOULDBLOCK); | |
486 | ||
487 | return rq; | |
1f5bd336 ML |
488 | } |
489 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
490 | ||
12f5b931 KB |
491 | static void __blk_mq_free_request(struct request *rq) |
492 | { | |
493 | struct request_queue *q = rq->q; | |
494 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 495 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
496 | const int sched_tag = rq->internal_tag; |
497 | ||
986d413b | 498 | blk_pm_mark_last_busy(rq); |
ea4f995e | 499 | rq->mq_hctx = NULL; |
12f5b931 KB |
500 | if (rq->tag != -1) |
501 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
502 | if (sched_tag != -1) | |
503 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); | |
504 | blk_mq_sched_restart(hctx); | |
505 | blk_queue_exit(q); | |
506 | } | |
507 | ||
6af54051 | 508 | void blk_mq_free_request(struct request *rq) |
320ae51f | 509 | { |
320ae51f | 510 | struct request_queue *q = rq->q; |
6af54051 CH |
511 | struct elevator_queue *e = q->elevator; |
512 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 513 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 514 | |
5bbf4e5a | 515 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
516 | if (e && e->type->ops.finish_request) |
517 | e->type->ops.finish_request(rq); | |
6af54051 CH |
518 | if (rq->elv.icq) { |
519 | put_io_context(rq->elv.icq->ioc); | |
520 | rq->elv.icq = NULL; | |
521 | } | |
522 | } | |
320ae51f | 523 | |
6af54051 | 524 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 525 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 526 | atomic_dec(&hctx->nr_active); |
87760e5e | 527 | |
7beb2f84 JA |
528 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
529 | laptop_io_completion(q->backing_dev_info); | |
530 | ||
a7905043 | 531 | rq_qos_done(q, rq); |
0d2602ca | 532 | |
12f5b931 KB |
533 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
534 | if (refcount_dec_and_test(&rq->ref)) | |
535 | __blk_mq_free_request(rq); | |
320ae51f | 536 | } |
1a3b595a | 537 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 538 | |
2a842aca | 539 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 540 | { |
fe1f4526 JA |
541 | u64 now = 0; |
542 | ||
543 | if (blk_mq_need_time_stamp(rq)) | |
544 | now = ktime_get_ns(); | |
522a7775 | 545 | |
4bc6339a OS |
546 | if (rq->rq_flags & RQF_STATS) { |
547 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 548 | blk_stat_add(rq, now); |
4bc6339a OS |
549 | } |
550 | ||
ed88660a OS |
551 | if (rq->internal_tag != -1) |
552 | blk_mq_sched_completed_request(rq, now); | |
553 | ||
522a7775 | 554 | blk_account_io_done(rq, now); |
0d11e6ac | 555 | |
91b63639 | 556 | if (rq->end_io) { |
a7905043 | 557 | rq_qos_done(rq->q, rq); |
320ae51f | 558 | rq->end_io(rq, error); |
91b63639 | 559 | } else { |
320ae51f | 560 | blk_mq_free_request(rq); |
91b63639 | 561 | } |
320ae51f | 562 | } |
c8a446ad | 563 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 564 | |
2a842aca | 565 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
566 | { |
567 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
568 | BUG(); | |
c8a446ad | 569 | __blk_mq_end_request(rq, error); |
63151a44 | 570 | } |
c8a446ad | 571 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 572 | |
30a91cb4 | 573 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 574 | { |
3d6efbf6 | 575 | struct request *rq = data; |
c7bb9ad1 | 576 | struct request_queue *q = rq->q; |
320ae51f | 577 | |
c7bb9ad1 | 578 | q->mq_ops->complete(rq); |
320ae51f | 579 | } |
320ae51f | 580 | |
453f8341 | 581 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
582 | { |
583 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
c7bb9ad1 | 584 | struct request_queue *q = rq->q; |
38535201 | 585 | bool shared = false; |
320ae51f JA |
586 | int cpu; |
587 | ||
af78ff7c | 588 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 ML |
589 | /* |
590 | * Most of single queue controllers, there is only one irq vector | |
591 | * for handling IO completion, and the only irq's affinity is set | |
592 | * as all possible CPUs. On most of ARCHs, this affinity means the | |
593 | * irq is handled on one specific CPU. | |
594 | * | |
595 | * So complete IO reqeust in softirq context in case of single queue | |
596 | * for not degrading IO performance by irqsoff latency. | |
597 | */ | |
c7bb9ad1 | 598 | if (q->nr_hw_queues == 1) { |
36e76539 ML |
599 | __blk_complete_request(rq); |
600 | return; | |
601 | } | |
602 | ||
4ab32bf3 JA |
603 | /* |
604 | * For a polled request, always complete locallly, it's pointless | |
605 | * to redirect the completion. | |
606 | */ | |
607 | if ((rq->cmd_flags & REQ_HIPRI) || | |
608 | !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) { | |
c7bb9ad1 | 609 | q->mq_ops->complete(rq); |
30a91cb4 CH |
610 | return; |
611 | } | |
320ae51f JA |
612 | |
613 | cpu = get_cpu(); | |
c7bb9ad1 | 614 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags)) |
38535201 CH |
615 | shared = cpus_share_cache(cpu, ctx->cpu); |
616 | ||
617 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 618 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
619 | rq->csd.info = rq; |
620 | rq->csd.flags = 0; | |
c46fff2a | 621 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 622 | } else { |
c7bb9ad1 | 623 | q->mq_ops->complete(rq); |
3d6efbf6 | 624 | } |
320ae51f JA |
625 | put_cpu(); |
626 | } | |
30a91cb4 | 627 | |
04ced159 | 628 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 629 | __releases(hctx->srcu) |
04ced159 JA |
630 | { |
631 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
632 | rcu_read_unlock(); | |
633 | else | |
05707b64 | 634 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
635 | } |
636 | ||
637 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 638 | __acquires(hctx->srcu) |
04ced159 | 639 | { |
08b5a6e2 JA |
640 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
641 | /* shut up gcc false positive */ | |
642 | *srcu_idx = 0; | |
04ced159 | 643 | rcu_read_lock(); |
08b5a6e2 | 644 | } else |
05707b64 | 645 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
646 | } |
647 | ||
30a91cb4 CH |
648 | /** |
649 | * blk_mq_complete_request - end I/O on a request | |
650 | * @rq: the request being processed | |
651 | * | |
652 | * Description: | |
653 | * Ends all I/O on a request. It does not handle partial completions. | |
654 | * The actual completion happens out-of-order, through a IPI handler. | |
655 | **/ | |
16c15eb1 | 656 | bool blk_mq_complete_request(struct request *rq) |
30a91cb4 | 657 | { |
12f5b931 | 658 | if (unlikely(blk_should_fake_timeout(rq->q))) |
16c15eb1 | 659 | return false; |
12f5b931 | 660 | __blk_mq_complete_request(rq); |
16c15eb1 | 661 | return true; |
30a91cb4 CH |
662 | } |
663 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 664 | |
973c0191 KB |
665 | int blk_mq_request_started(struct request *rq) |
666 | { | |
5a61c363 | 667 | return blk_mq_rq_state(rq) != MQ_RQ_IDLE; |
973c0191 KB |
668 | } |
669 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
670 | ||
aa306ab7 ML |
671 | int blk_mq_request_completed(struct request *rq) |
672 | { | |
673 | return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE; | |
674 | } | |
675 | EXPORT_SYMBOL_GPL(blk_mq_request_completed); | |
676 | ||
e2490073 | 677 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
678 | { |
679 | struct request_queue *q = rq->q; | |
680 | ||
681 | trace_block_rq_issue(q, rq); | |
682 | ||
cf43e6be | 683 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 684 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 685 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 686 | rq->rq_flags |= RQF_STATS; |
a7905043 | 687 | rq_qos_issue(q, rq); |
cf43e6be JA |
688 | } |
689 | ||
1d9bd516 | 690 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 691 | |
1d9bd516 | 692 | blk_add_timer(rq); |
12f5b931 | 693 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 CH |
694 | |
695 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
696 | /* | |
697 | * Make sure space for the drain appears. We know we can do | |
698 | * this because max_hw_segments has been adjusted to be one | |
699 | * fewer than the device can handle. | |
700 | */ | |
701 | rq->nr_phys_segments++; | |
702 | } | |
320ae51f | 703 | } |
e2490073 | 704 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 705 | |
ed0791b2 | 706 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
707 | { |
708 | struct request_queue *q = rq->q; | |
709 | ||
923218f6 ML |
710 | blk_mq_put_driver_tag(rq); |
711 | ||
320ae51f | 712 | trace_block_rq_requeue(q, rq); |
a7905043 | 713 | rq_qos_requeue(q, rq); |
49f5baa5 | 714 | |
12f5b931 KB |
715 | if (blk_mq_request_started(rq)) { |
716 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 717 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 CH |
718 | if (q->dma_drain_size && blk_rq_bytes(rq)) |
719 | rq->nr_phys_segments--; | |
720 | } | |
320ae51f JA |
721 | } |
722 | ||
2b053aca | 723 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 724 | { |
ed0791b2 | 725 | __blk_mq_requeue_request(rq); |
ed0791b2 | 726 | |
105976f5 ML |
727 | /* this request will be re-inserted to io scheduler queue */ |
728 | blk_mq_sched_requeue_request(rq); | |
729 | ||
7d692330 | 730 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 731 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
732 | } |
733 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
734 | ||
6fca6a61 CH |
735 | static void blk_mq_requeue_work(struct work_struct *work) |
736 | { | |
737 | struct request_queue *q = | |
2849450a | 738 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
739 | LIST_HEAD(rq_list); |
740 | struct request *rq, *next; | |
6fca6a61 | 741 | |
18e9781d | 742 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 743 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 744 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
745 | |
746 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 747 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
748 | continue; |
749 | ||
e8064021 | 750 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 751 | list_del_init(&rq->queuelist); |
aef1897c JW |
752 | /* |
753 | * If RQF_DONTPREP, rq has contained some driver specific | |
754 | * data, so insert it to hctx dispatch list to avoid any | |
755 | * merge. | |
756 | */ | |
757 | if (rq->rq_flags & RQF_DONTPREP) | |
758 | blk_mq_request_bypass_insert(rq, false); | |
759 | else | |
760 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
761 | } |
762 | ||
763 | while (!list_empty(&rq_list)) { | |
764 | rq = list_entry(rq_list.next, struct request, queuelist); | |
765 | list_del_init(&rq->queuelist); | |
9e97d295 | 766 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
767 | } |
768 | ||
52d7f1b5 | 769 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
770 | } |
771 | ||
2b053aca BVA |
772 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
773 | bool kick_requeue_list) | |
6fca6a61 CH |
774 | { |
775 | struct request_queue *q = rq->q; | |
776 | unsigned long flags; | |
777 | ||
778 | /* | |
779 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 780 | * request head insertion from the workqueue. |
6fca6a61 | 781 | */ |
e8064021 | 782 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
783 | |
784 | spin_lock_irqsave(&q->requeue_lock, flags); | |
785 | if (at_head) { | |
e8064021 | 786 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
787 | list_add(&rq->queuelist, &q->requeue_list); |
788 | } else { | |
789 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
790 | } | |
791 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
792 | |
793 | if (kick_requeue_list) | |
794 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 795 | } |
6fca6a61 CH |
796 | |
797 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
798 | { | |
ae943d20 | 799 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
800 | } |
801 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
802 | ||
2849450a MS |
803 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
804 | unsigned long msecs) | |
805 | { | |
d4acf365 BVA |
806 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
807 | msecs_to_jiffies(msecs)); | |
2849450a MS |
808 | } |
809 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
810 | ||
0e62f51f JA |
811 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
812 | { | |
88c7b2b7 JA |
813 | if (tag < tags->nr_tags) { |
814 | prefetch(tags->rqs[tag]); | |
4ee86bab | 815 | return tags->rqs[tag]; |
88c7b2b7 | 816 | } |
4ee86bab HR |
817 | |
818 | return NULL; | |
24d2f903 CH |
819 | } |
820 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
821 | ||
3c94d83c JA |
822 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
823 | void *priv, bool reserved) | |
ae879912 JA |
824 | { |
825 | /* | |
3c94d83c JA |
826 | * If we find a request that is inflight and the queue matches, |
827 | * we know the queue is busy. Return false to stop the iteration. | |
ae879912 | 828 | */ |
3c94d83c | 829 | if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) { |
ae879912 JA |
830 | bool *busy = priv; |
831 | ||
832 | *busy = true; | |
833 | return false; | |
834 | } | |
835 | ||
836 | return true; | |
837 | } | |
838 | ||
3c94d83c | 839 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
840 | { |
841 | bool busy = false; | |
842 | ||
3c94d83c | 843 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
844 | return busy; |
845 | } | |
3c94d83c | 846 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 847 | |
358f70da | 848 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 849 | { |
da661267 | 850 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
851 | if (req->q->mq_ops->timeout) { |
852 | enum blk_eh_timer_return ret; | |
853 | ||
854 | ret = req->q->mq_ops->timeout(req, reserved); | |
855 | if (ret == BLK_EH_DONE) | |
856 | return; | |
857 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 858 | } |
d1210d5a CH |
859 | |
860 | blk_add_timer(req); | |
87ee7b11 | 861 | } |
5b3f25fc | 862 | |
12f5b931 | 863 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 864 | { |
12f5b931 | 865 | unsigned long deadline; |
87ee7b11 | 866 | |
12f5b931 KB |
867 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
868 | return false; | |
da661267 CH |
869 | if (rq->rq_flags & RQF_TIMED_OUT) |
870 | return false; | |
a7af0af3 | 871 | |
079076b3 | 872 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
873 | if (time_after_eq(jiffies, deadline)) |
874 | return true; | |
a7af0af3 | 875 | |
12f5b931 KB |
876 | if (*next == 0) |
877 | *next = deadline; | |
878 | else if (time_after(*next, deadline)) | |
879 | *next = deadline; | |
880 | return false; | |
87ee7b11 JA |
881 | } |
882 | ||
7baa8572 | 883 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
884 | struct request *rq, void *priv, bool reserved) |
885 | { | |
12f5b931 KB |
886 | unsigned long *next = priv; |
887 | ||
888 | /* | |
889 | * Just do a quick check if it is expired before locking the request in | |
890 | * so we're not unnecessarilly synchronizing across CPUs. | |
891 | */ | |
892 | if (!blk_mq_req_expired(rq, next)) | |
7baa8572 | 893 | return true; |
12f5b931 KB |
894 | |
895 | /* | |
896 | * We have reason to believe the request may be expired. Take a | |
897 | * reference on the request to lock this request lifetime into its | |
898 | * currently allocated context to prevent it from being reallocated in | |
899 | * the event the completion by-passes this timeout handler. | |
900 | * | |
901 | * If the reference was already released, then the driver beat the | |
902 | * timeout handler to posting a natural completion. | |
903 | */ | |
904 | if (!refcount_inc_not_zero(&rq->ref)) | |
7baa8572 | 905 | return true; |
12f5b931 | 906 | |
1d9bd516 | 907 | /* |
12f5b931 KB |
908 | * The request is now locked and cannot be reallocated underneath the |
909 | * timeout handler's processing. Re-verify this exact request is truly | |
910 | * expired; if it is not expired, then the request was completed and | |
911 | * reallocated as a new request. | |
1d9bd516 | 912 | */ |
12f5b931 | 913 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 914 | blk_mq_rq_timed_out(rq, reserved); |
12f5b931 KB |
915 | if (refcount_dec_and_test(&rq->ref)) |
916 | __blk_mq_free_request(rq); | |
7baa8572 JA |
917 | |
918 | return true; | |
1d9bd516 TH |
919 | } |
920 | ||
287922eb | 921 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 922 | { |
287922eb CH |
923 | struct request_queue *q = |
924 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 925 | unsigned long next = 0; |
1d9bd516 | 926 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 927 | int i; |
320ae51f | 928 | |
71f79fb3 GKB |
929 | /* A deadlock might occur if a request is stuck requiring a |
930 | * timeout at the same time a queue freeze is waiting | |
931 | * completion, since the timeout code would not be able to | |
932 | * acquire the queue reference here. | |
933 | * | |
934 | * That's why we don't use blk_queue_enter here; instead, we use | |
935 | * percpu_ref_tryget directly, because we need to be able to | |
936 | * obtain a reference even in the short window between the queue | |
937 | * starting to freeze, by dropping the first reference in | |
1671d522 | 938 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
939 | * consumed, marked by the instant q_usage_counter reaches |
940 | * zero. | |
941 | */ | |
942 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
943 | return; |
944 | ||
12f5b931 | 945 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 946 | |
12f5b931 KB |
947 | if (next != 0) { |
948 | mod_timer(&q->timeout, next); | |
0d2602ca | 949 | } else { |
fcd36c36 BVA |
950 | /* |
951 | * Request timeouts are handled as a forward rolling timer. If | |
952 | * we end up here it means that no requests are pending and | |
953 | * also that no request has been pending for a while. Mark | |
954 | * each hctx as idle. | |
955 | */ | |
f054b56c ML |
956 | queue_for_each_hw_ctx(q, hctx, i) { |
957 | /* the hctx may be unmapped, so check it here */ | |
958 | if (blk_mq_hw_queue_mapped(hctx)) | |
959 | blk_mq_tag_idle(hctx); | |
960 | } | |
0d2602ca | 961 | } |
287922eb | 962 | blk_queue_exit(q); |
320ae51f JA |
963 | } |
964 | ||
88459642 OS |
965 | struct flush_busy_ctx_data { |
966 | struct blk_mq_hw_ctx *hctx; | |
967 | struct list_head *list; | |
968 | }; | |
969 | ||
970 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
971 | { | |
972 | struct flush_busy_ctx_data *flush_data = data; | |
973 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
974 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 975 | enum hctx_type type = hctx->type; |
88459642 | 976 | |
88459642 | 977 | spin_lock(&ctx->lock); |
c16d6b5a | 978 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 979 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
980 | spin_unlock(&ctx->lock); |
981 | return true; | |
982 | } | |
983 | ||
1429d7c9 JA |
984 | /* |
985 | * Process software queues that have been marked busy, splicing them | |
986 | * to the for-dispatch | |
987 | */ | |
2c3ad667 | 988 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 989 | { |
88459642 OS |
990 | struct flush_busy_ctx_data data = { |
991 | .hctx = hctx, | |
992 | .list = list, | |
993 | }; | |
1429d7c9 | 994 | |
88459642 | 995 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 996 | } |
2c3ad667 | 997 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 998 | |
b347689f ML |
999 | struct dispatch_rq_data { |
1000 | struct blk_mq_hw_ctx *hctx; | |
1001 | struct request *rq; | |
1002 | }; | |
1003 | ||
1004 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1005 | void *data) | |
1006 | { | |
1007 | struct dispatch_rq_data *dispatch_data = data; | |
1008 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1009 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1010 | enum hctx_type type = hctx->type; |
b347689f ML |
1011 | |
1012 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1013 | if (!list_empty(&ctx->rq_lists[type])) { |
1014 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1015 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1016 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1017 | sbitmap_clear_bit(sb, bitnr); |
1018 | } | |
1019 | spin_unlock(&ctx->lock); | |
1020 | ||
1021 | return !dispatch_data->rq; | |
1022 | } | |
1023 | ||
1024 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1025 | struct blk_mq_ctx *start) | |
1026 | { | |
f31967f0 | 1027 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1028 | struct dispatch_rq_data data = { |
1029 | .hctx = hctx, | |
1030 | .rq = NULL, | |
1031 | }; | |
1032 | ||
1033 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1034 | dispatch_rq_from_ctx, &data); | |
1035 | ||
1036 | return data.rq; | |
1037 | } | |
1038 | ||
703fd1c0 JA |
1039 | static inline unsigned int queued_to_index(unsigned int queued) |
1040 | { | |
1041 | if (!queued) | |
1042 | return 0; | |
1429d7c9 | 1043 | |
703fd1c0 | 1044 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1045 | } |
1046 | ||
8ab6bb9e | 1047 | bool blk_mq_get_driver_tag(struct request *rq) |
bd166ef1 JA |
1048 | { |
1049 | struct blk_mq_alloc_data data = { | |
1050 | .q = rq->q, | |
ea4f995e | 1051 | .hctx = rq->mq_hctx, |
8ab6bb9e | 1052 | .flags = BLK_MQ_REQ_NOWAIT, |
f9afca4d | 1053 | .cmd_flags = rq->cmd_flags, |
bd166ef1 | 1054 | }; |
d263ed99 | 1055 | bool shared; |
5feeacdd | 1056 | |
81380ca1 OS |
1057 | if (rq->tag != -1) |
1058 | goto done; | |
bd166ef1 | 1059 | |
415b806d SG |
1060 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
1061 | data.flags |= BLK_MQ_REQ_RESERVED; | |
1062 | ||
d263ed99 | 1063 | shared = blk_mq_tag_busy(data.hctx); |
bd166ef1 JA |
1064 | rq->tag = blk_mq_get_tag(&data); |
1065 | if (rq->tag >= 0) { | |
d263ed99 | 1066 | if (shared) { |
200e86b3 JA |
1067 | rq->rq_flags |= RQF_MQ_INFLIGHT; |
1068 | atomic_inc(&data.hctx->nr_active); | |
1069 | } | |
bd166ef1 | 1070 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
1071 | } |
1072 | ||
81380ca1 | 1073 | done: |
81380ca1 | 1074 | return rq->tag != -1; |
bd166ef1 JA |
1075 | } |
1076 | ||
eb619fdb JA |
1077 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1078 | int flags, void *key) | |
da55f2cc OS |
1079 | { |
1080 | struct blk_mq_hw_ctx *hctx; | |
1081 | ||
1082 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1083 | ||
5815839b | 1084 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1085 | if (!list_empty(&wait->entry)) { |
1086 | struct sbitmap_queue *sbq; | |
1087 | ||
1088 | list_del_init(&wait->entry); | |
1089 | sbq = &hctx->tags->bitmap_tags; | |
1090 | atomic_dec(&sbq->ws_active); | |
1091 | } | |
5815839b ML |
1092 | spin_unlock(&hctx->dispatch_wait_lock); |
1093 | ||
da55f2cc OS |
1094 | blk_mq_run_hw_queue(hctx, true); |
1095 | return 1; | |
1096 | } | |
1097 | ||
f906a6a0 JA |
1098 | /* |
1099 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1100 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1101 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1102 | * marking us as waiting. |
1103 | */ | |
2278d69f | 1104 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1105 | struct request *rq) |
da55f2cc | 1106 | { |
e8618575 | 1107 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1108 | struct wait_queue_head *wq; |
f906a6a0 JA |
1109 | wait_queue_entry_t *wait; |
1110 | bool ret; | |
da55f2cc | 1111 | |
2278d69f | 1112 | if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) { |
684b7324 | 1113 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1114 | |
c27d53fb BVA |
1115 | /* |
1116 | * It's possible that a tag was freed in the window between the | |
1117 | * allocation failure and adding the hardware queue to the wait | |
1118 | * queue. | |
1119 | * | |
1120 | * Don't clear RESTART here, someone else could have set it. | |
1121 | * At most this will cost an extra queue run. | |
1122 | */ | |
8ab6bb9e | 1123 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1124 | } |
eb619fdb | 1125 | |
2278d69f | 1126 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1127 | if (!list_empty_careful(&wait->entry)) |
1128 | return false; | |
1129 | ||
e8618575 | 1130 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1131 | |
1132 | spin_lock_irq(&wq->lock); | |
1133 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1134 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1135 | spin_unlock(&hctx->dispatch_wait_lock); |
1136 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1137 | return false; |
eb619fdb JA |
1138 | } |
1139 | ||
e8618575 | 1140 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1141 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1142 | __add_wait_queue(wq, wait); | |
c27d53fb | 1143 | |
da55f2cc | 1144 | /* |
eb619fdb JA |
1145 | * It's possible that a tag was freed in the window between the |
1146 | * allocation failure and adding the hardware queue to the wait | |
1147 | * queue. | |
da55f2cc | 1148 | */ |
8ab6bb9e | 1149 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1150 | if (!ret) { |
5815839b ML |
1151 | spin_unlock(&hctx->dispatch_wait_lock); |
1152 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1153 | return false; |
eb619fdb | 1154 | } |
c27d53fb BVA |
1155 | |
1156 | /* | |
1157 | * We got a tag, remove ourselves from the wait queue to ensure | |
1158 | * someone else gets the wakeup. | |
1159 | */ | |
c27d53fb | 1160 | list_del_init(&wait->entry); |
e8618575 | 1161 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1162 | spin_unlock(&hctx->dispatch_wait_lock); |
1163 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1164 | |
1165 | return true; | |
da55f2cc OS |
1166 | } |
1167 | ||
6e768717 ML |
1168 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1169 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1170 | /* | |
1171 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1172 | * - EWMA is one simple way to compute running average value | |
1173 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1174 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1175 | * factor doesn't matter because EWMA decreases exponentially | |
1176 | */ | |
1177 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1178 | { | |
1179 | unsigned int ewma; | |
1180 | ||
1181 | if (hctx->queue->elevator) | |
1182 | return; | |
1183 | ||
1184 | ewma = hctx->dispatch_busy; | |
1185 | ||
1186 | if (!ewma && !busy) | |
1187 | return; | |
1188 | ||
1189 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1190 | if (busy) | |
1191 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1192 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1193 | ||
1194 | hctx->dispatch_busy = ewma; | |
1195 | } | |
1196 | ||
86ff7c2a ML |
1197 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1198 | ||
1f57f8d4 JA |
1199 | /* |
1200 | * Returns true if we did some work AND can potentially do more. | |
1201 | */ | |
de148297 | 1202 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
eb619fdb | 1203 | bool got_budget) |
320ae51f | 1204 | { |
81380ca1 | 1205 | struct blk_mq_hw_ctx *hctx; |
6d6f167c | 1206 | struct request *rq, *nxt; |
eb619fdb | 1207 | bool no_tag = false; |
fc17b653 | 1208 | int errors, queued; |
86ff7c2a | 1209 | blk_status_t ret = BLK_STS_OK; |
320ae51f | 1210 | |
81380ca1 OS |
1211 | if (list_empty(list)) |
1212 | return false; | |
1213 | ||
de148297 ML |
1214 | WARN_ON(!list_is_singular(list) && got_budget); |
1215 | ||
320ae51f JA |
1216 | /* |
1217 | * Now process all the entries, sending them to the driver. | |
1218 | */ | |
93efe981 | 1219 | errors = queued = 0; |
81380ca1 | 1220 | do { |
74c45052 | 1221 | struct blk_mq_queue_data bd; |
320ae51f | 1222 | |
f04c3df3 | 1223 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1224 | |
ea4f995e | 1225 | hctx = rq->mq_hctx; |
0bca799b ML |
1226 | if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) |
1227 | break; | |
1228 | ||
8ab6bb9e | 1229 | if (!blk_mq_get_driver_tag(rq)) { |
3c782d67 | 1230 | /* |
da55f2cc | 1231 | * The initial allocation attempt failed, so we need to |
eb619fdb JA |
1232 | * rerun the hardware queue when a tag is freed. The |
1233 | * waitqueue takes care of that. If the queue is run | |
1234 | * before we add this entry back on the dispatch list, | |
1235 | * we'll re-run it below. | |
3c782d67 | 1236 | */ |
2278d69f | 1237 | if (!blk_mq_mark_tag_wait(hctx, rq)) { |
0bca799b | 1238 | blk_mq_put_dispatch_budget(hctx); |
f906a6a0 JA |
1239 | /* |
1240 | * For non-shared tags, the RESTART check | |
1241 | * will suffice. | |
1242 | */ | |
1243 | if (hctx->flags & BLK_MQ_F_TAG_SHARED) | |
1244 | no_tag = true; | |
de148297 ML |
1245 | break; |
1246 | } | |
1247 | } | |
1248 | ||
320ae51f | 1249 | list_del_init(&rq->queuelist); |
320ae51f | 1250 | |
74c45052 | 1251 | bd.rq = rq; |
113285b4 JA |
1252 | |
1253 | /* | |
1254 | * Flag last if we have no more requests, or if we have more | |
1255 | * but can't assign a driver tag to it. | |
1256 | */ | |
1257 | if (list_empty(list)) | |
1258 | bd.last = true; | |
1259 | else { | |
113285b4 | 1260 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1261 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1262 | } |
74c45052 JA |
1263 | |
1264 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
86ff7c2a | 1265 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) { |
6d6f167c JW |
1266 | /* |
1267 | * If an I/O scheduler has been configured and we got a | |
ff821d27 JA |
1268 | * driver tag for the next request already, free it |
1269 | * again. | |
6d6f167c JW |
1270 | */ |
1271 | if (!list_empty(list)) { | |
1272 | nxt = list_first_entry(list, struct request, queuelist); | |
1273 | blk_mq_put_driver_tag(nxt); | |
1274 | } | |
f04c3df3 | 1275 | list_add(&rq->queuelist, list); |
ed0791b2 | 1276 | __blk_mq_requeue_request(rq); |
320ae51f | 1277 | break; |
fc17b653 CH |
1278 | } |
1279 | ||
1280 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1281 | errors++; |
2a842aca | 1282 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1283 | continue; |
320ae51f JA |
1284 | } |
1285 | ||
fc17b653 | 1286 | queued++; |
81380ca1 | 1287 | } while (!list_empty(list)); |
320ae51f | 1288 | |
703fd1c0 | 1289 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1290 | |
1291 | /* | |
1292 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1293 | * that is where we will continue on next queue run. | |
1294 | */ | |
f04c3df3 | 1295 | if (!list_empty(list)) { |
86ff7c2a ML |
1296 | bool needs_restart; |
1297 | ||
d666ba98 JA |
1298 | /* |
1299 | * If we didn't flush the entire list, we could have told | |
1300 | * the driver there was more coming, but that turned out to | |
1301 | * be a lie. | |
1302 | */ | |
1303 | if (q->mq_ops->commit_rqs) | |
1304 | q->mq_ops->commit_rqs(hctx); | |
1305 | ||
320ae51f | 1306 | spin_lock(&hctx->lock); |
c13660a0 | 1307 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1308 | spin_unlock(&hctx->lock); |
f04c3df3 | 1309 | |
9ba52e58 | 1310 | /* |
710c785f BVA |
1311 | * If SCHED_RESTART was set by the caller of this function and |
1312 | * it is no longer set that means that it was cleared by another | |
1313 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1314 | * |
eb619fdb JA |
1315 | * If 'no_tag' is set, that means that we failed getting |
1316 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1317 | * waitqueue is no longer active, ensure that we run the queue | |
1318 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1319 | * |
710c785f BVA |
1320 | * If no I/O scheduler has been configured it is possible that |
1321 | * the hardware queue got stopped and restarted before requests | |
1322 | * were pushed back onto the dispatch list. Rerun the queue to | |
1323 | * avoid starvation. Notes: | |
1324 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1325 | * been stopped before rerunning a queue. | |
1326 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1327 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1328 | * and dm-rq. |
86ff7c2a ML |
1329 | * |
1330 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1331 | * bit is set, run queue after a delay to avoid IO stalls | |
1332 | * that could otherwise occur if the queue is idle. | |
bd166ef1 | 1333 | */ |
86ff7c2a ML |
1334 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1335 | if (!needs_restart || | |
eb619fdb | 1336 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1337 | blk_mq_run_hw_queue(hctx, true); |
86ff7c2a ML |
1338 | else if (needs_restart && (ret == BLK_STS_RESOURCE)) |
1339 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1f57f8d4 | 1340 | |
6e768717 | 1341 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1342 | return false; |
6e768717 ML |
1343 | } else |
1344 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1345 | |
1f57f8d4 JA |
1346 | /* |
1347 | * If the host/device is unable to accept more work, inform the | |
1348 | * caller of that. | |
1349 | */ | |
1350 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1351 | return false; | |
1352 | ||
93efe981 | 1353 | return (queued + errors) != 0; |
f04c3df3 JA |
1354 | } |
1355 | ||
6a83e74d BVA |
1356 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1357 | { | |
1358 | int srcu_idx; | |
1359 | ||
b7a71e66 JA |
1360 | /* |
1361 | * We should be running this queue from one of the CPUs that | |
1362 | * are mapped to it. | |
7df938fb ML |
1363 | * |
1364 | * There are at least two related races now between setting | |
1365 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1366 | * __blk_mq_run_hw_queue(): | |
1367 | * | |
1368 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1369 | * but later it becomes online, then this warning is harmless | |
1370 | * at all | |
1371 | * | |
1372 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1373 | * but later it becomes offline, then the warning can't be | |
1374 | * triggered, and we depend on blk-mq timeout handler to | |
1375 | * handle dispatched requests to this hctx | |
b7a71e66 | 1376 | */ |
7df938fb ML |
1377 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1378 | cpu_online(hctx->next_cpu)) { | |
1379 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1380 | raw_smp_processor_id(), | |
1381 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1382 | dump_stack(); | |
1383 | } | |
6a83e74d | 1384 | |
b7a71e66 JA |
1385 | /* |
1386 | * We can't run the queue inline with ints disabled. Ensure that | |
1387 | * we catch bad users of this early. | |
1388 | */ | |
1389 | WARN_ON_ONCE(in_interrupt()); | |
1390 | ||
04ced159 | 1391 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1392 | |
04ced159 JA |
1393 | hctx_lock(hctx, &srcu_idx); |
1394 | blk_mq_sched_dispatch_requests(hctx); | |
1395 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1396 | } |
1397 | ||
f82ddf19 ML |
1398 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1399 | { | |
1400 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1401 | ||
1402 | if (cpu >= nr_cpu_ids) | |
1403 | cpu = cpumask_first(hctx->cpumask); | |
1404 | return cpu; | |
1405 | } | |
1406 | ||
506e931f JA |
1407 | /* |
1408 | * It'd be great if the workqueue API had a way to pass | |
1409 | * in a mask and had some smarts for more clever placement. | |
1410 | * For now we just round-robin here, switching for every | |
1411 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1412 | */ | |
1413 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1414 | { | |
7bed4595 | 1415 | bool tried = false; |
476f8c98 | 1416 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1417 | |
b657d7e6 CH |
1418 | if (hctx->queue->nr_hw_queues == 1) |
1419 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1420 | |
1421 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1422 | select_cpu: |
476f8c98 | 1423 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1424 | cpu_online_mask); |
506e931f | 1425 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1426 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1427 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1428 | } | |
1429 | ||
7bed4595 ML |
1430 | /* |
1431 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1432 | * and it should only happen in the path of handling CPU DEAD. | |
1433 | */ | |
476f8c98 | 1434 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1435 | if (!tried) { |
1436 | tried = true; | |
1437 | goto select_cpu; | |
1438 | } | |
1439 | ||
1440 | /* | |
1441 | * Make sure to re-select CPU next time once after CPUs | |
1442 | * in hctx->cpumask become online again. | |
1443 | */ | |
476f8c98 | 1444 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1445 | hctx->next_cpu_batch = 1; |
1446 | return WORK_CPU_UNBOUND; | |
1447 | } | |
476f8c98 ML |
1448 | |
1449 | hctx->next_cpu = next_cpu; | |
1450 | return next_cpu; | |
506e931f JA |
1451 | } |
1452 | ||
7587a5ae BVA |
1453 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1454 | unsigned long msecs) | |
320ae51f | 1455 | { |
5435c023 | 1456 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1457 | return; |
1458 | ||
1b792f2f | 1459 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1460 | int cpu = get_cpu(); |
1461 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1462 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1463 | put_cpu(); |
398205b8 PB |
1464 | return; |
1465 | } | |
e4043dcf | 1466 | |
2a90d4aa | 1467 | put_cpu(); |
e4043dcf | 1468 | } |
398205b8 | 1469 | |
ae943d20 BVA |
1470 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1471 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1472 | } |
1473 | ||
1474 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1475 | { | |
1476 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1477 | } | |
1478 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1479 | ||
79f720a7 | 1480 | bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1481 | { |
24f5a90f ML |
1482 | int srcu_idx; |
1483 | bool need_run; | |
1484 | ||
1485 | /* | |
1486 | * When queue is quiesced, we may be switching io scheduler, or | |
1487 | * updating nr_hw_queues, or other things, and we can't run queue | |
1488 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1489 | * | |
1490 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1491 | * quiesced. | |
1492 | */ | |
04ced159 JA |
1493 | hctx_lock(hctx, &srcu_idx); |
1494 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1495 | blk_mq_hctx_has_pending(hctx); | |
1496 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f ML |
1497 | |
1498 | if (need_run) { | |
79f720a7 JA |
1499 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
1500 | return true; | |
1501 | } | |
1502 | ||
1503 | return false; | |
320ae51f | 1504 | } |
5b727272 | 1505 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1506 | |
b94ec296 | 1507 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1508 | { |
1509 | struct blk_mq_hw_ctx *hctx; | |
1510 | int i; | |
1511 | ||
1512 | queue_for_each_hw_ctx(q, hctx, i) { | |
79f720a7 | 1513 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1514 | continue; |
1515 | ||
b94ec296 | 1516 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1517 | } |
1518 | } | |
b94ec296 | 1519 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1520 | |
fd001443 BVA |
1521 | /** |
1522 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1523 | * @q: request queue. | |
1524 | * | |
1525 | * The caller is responsible for serializing this function against | |
1526 | * blk_mq_{start,stop}_hw_queue(). | |
1527 | */ | |
1528 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1529 | { | |
1530 | struct blk_mq_hw_ctx *hctx; | |
1531 | int i; | |
1532 | ||
1533 | queue_for_each_hw_ctx(q, hctx, i) | |
1534 | if (blk_mq_hctx_stopped(hctx)) | |
1535 | return true; | |
1536 | ||
1537 | return false; | |
1538 | } | |
1539 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1540 | ||
39a70c76 ML |
1541 | /* |
1542 | * This function is often used for pausing .queue_rq() by driver when | |
1543 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1544 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1545 | * |
1546 | * We do not guarantee that dispatch can be drained or blocked | |
1547 | * after blk_mq_stop_hw_queue() returns. Please use | |
1548 | * blk_mq_quiesce_queue() for that requirement. | |
1549 | */ | |
2719aa21 JA |
1550 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1551 | { | |
641a9ed6 | 1552 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1553 | |
641a9ed6 | 1554 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1555 | } |
641a9ed6 | 1556 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1557 | |
39a70c76 ML |
1558 | /* |
1559 | * This function is often used for pausing .queue_rq() by driver when | |
1560 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1561 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1562 | * |
1563 | * We do not guarantee that dispatch can be drained or blocked | |
1564 | * after blk_mq_stop_hw_queues() returns. Please use | |
1565 | * blk_mq_quiesce_queue() for that requirement. | |
1566 | */ | |
2719aa21 JA |
1567 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1568 | { | |
641a9ed6 ML |
1569 | struct blk_mq_hw_ctx *hctx; |
1570 | int i; | |
1571 | ||
1572 | queue_for_each_hw_ctx(q, hctx, i) | |
1573 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1574 | } |
1575 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1576 | ||
320ae51f JA |
1577 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1578 | { | |
1579 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1580 | |
0ffbce80 | 1581 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1582 | } |
1583 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1584 | ||
2f268556 CH |
1585 | void blk_mq_start_hw_queues(struct request_queue *q) |
1586 | { | |
1587 | struct blk_mq_hw_ctx *hctx; | |
1588 | int i; | |
1589 | ||
1590 | queue_for_each_hw_ctx(q, hctx, i) | |
1591 | blk_mq_start_hw_queue(hctx); | |
1592 | } | |
1593 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1594 | ||
ae911c5e JA |
1595 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1596 | { | |
1597 | if (!blk_mq_hctx_stopped(hctx)) | |
1598 | return; | |
1599 | ||
1600 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1601 | blk_mq_run_hw_queue(hctx, async); | |
1602 | } | |
1603 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1604 | ||
1b4a3258 | 1605 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1606 | { |
1607 | struct blk_mq_hw_ctx *hctx; | |
1608 | int i; | |
1609 | ||
ae911c5e JA |
1610 | queue_for_each_hw_ctx(q, hctx, i) |
1611 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1612 | } |
1613 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1614 | ||
70f4db63 | 1615 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1616 | { |
1617 | struct blk_mq_hw_ctx *hctx; | |
1618 | ||
9f993737 | 1619 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1620 | |
21c6e939 | 1621 | /* |
15fe8a90 | 1622 | * If we are stopped, don't run the queue. |
21c6e939 | 1623 | */ |
15fe8a90 | 1624 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
0196d6b4 | 1625 | return; |
7587a5ae BVA |
1626 | |
1627 | __blk_mq_run_hw_queue(hctx); | |
1628 | } | |
1629 | ||
cfd0c552 | 1630 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1631 | struct request *rq, |
1632 | bool at_head) | |
320ae51f | 1633 | { |
e57690fe | 1634 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1635 | enum hctx_type type = hctx->type; |
e57690fe | 1636 | |
7b607814 BVA |
1637 | lockdep_assert_held(&ctx->lock); |
1638 | ||
01b983c9 JA |
1639 | trace_block_rq_insert(hctx->queue, rq); |
1640 | ||
72a0a36e | 1641 | if (at_head) |
c16d6b5a | 1642 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1643 | else |
c16d6b5a | 1644 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1645 | } |
4bb659b1 | 1646 | |
2c3ad667 JA |
1647 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1648 | bool at_head) | |
cfd0c552 ML |
1649 | { |
1650 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1651 | ||
7b607814 BVA |
1652 | lockdep_assert_held(&ctx->lock); |
1653 | ||
e57690fe | 1654 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1655 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1656 | } |
1657 | ||
157f377b JA |
1658 | /* |
1659 | * Should only be used carefully, when the caller knows we want to | |
1660 | * bypass a potential IO scheduler on the target device. | |
1661 | */ | |
b0850297 | 1662 | void blk_mq_request_bypass_insert(struct request *rq, bool run_queue) |
157f377b | 1663 | { |
ea4f995e | 1664 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1665 | |
1666 | spin_lock(&hctx->lock); | |
1667 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1668 | spin_unlock(&hctx->lock); | |
1669 | ||
b0850297 ML |
1670 | if (run_queue) |
1671 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1672 | } |
1673 | ||
bd166ef1 JA |
1674 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1675 | struct list_head *list) | |
320ae51f JA |
1676 | |
1677 | { | |
3f0cedc7 | 1678 | struct request *rq; |
c16d6b5a | 1679 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1680 | |
320ae51f JA |
1681 | /* |
1682 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1683 | * offline now | |
1684 | */ | |
3f0cedc7 | 1685 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1686 | BUG_ON(rq->mq_ctx != ctx); |
3f0cedc7 | 1687 | trace_block_rq_insert(hctx->queue, rq); |
320ae51f | 1688 | } |
3f0cedc7 ML |
1689 | |
1690 | spin_lock(&ctx->lock); | |
c16d6b5a | 1691 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1692 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1693 | spin_unlock(&ctx->lock); |
320ae51f JA |
1694 | } |
1695 | ||
3110fc79 | 1696 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) |
320ae51f JA |
1697 | { |
1698 | struct request *rqa = container_of(a, struct request, queuelist); | |
1699 | struct request *rqb = container_of(b, struct request, queuelist); | |
1700 | ||
3110fc79 JA |
1701 | if (rqa->mq_ctx < rqb->mq_ctx) |
1702 | return -1; | |
1703 | else if (rqa->mq_ctx > rqb->mq_ctx) | |
1704 | return 1; | |
1705 | else if (rqa->mq_hctx < rqb->mq_hctx) | |
1706 | return -1; | |
1707 | else if (rqa->mq_hctx > rqb->mq_hctx) | |
1708 | return 1; | |
1709 | ||
1710 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1711 | } |
1712 | ||
1713 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1714 | { | |
67cae4c9 | 1715 | struct blk_mq_hw_ctx *this_hctx; |
320ae51f JA |
1716 | struct blk_mq_ctx *this_ctx; |
1717 | struct request_queue *this_q; | |
1718 | struct request *rq; | |
1719 | LIST_HEAD(list); | |
67cae4c9 | 1720 | LIST_HEAD(rq_list); |
320ae51f JA |
1721 | unsigned int depth; |
1722 | ||
1723 | list_splice_init(&plug->mq_list, &list); | |
1724 | ||
ce5b009c JA |
1725 | if (plug->rq_count > 2 && plug->multiple_queues) |
1726 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1727 | |
bcc816df DZ |
1728 | plug->rq_count = 0; |
1729 | ||
320ae51f | 1730 | this_q = NULL; |
67cae4c9 | 1731 | this_hctx = NULL; |
320ae51f JA |
1732 | this_ctx = NULL; |
1733 | depth = 0; | |
1734 | ||
1735 | while (!list_empty(&list)) { | |
1736 | rq = list_entry_rq(list.next); | |
1737 | list_del_init(&rq->queuelist); | |
1738 | BUG_ON(!rq->q); | |
67cae4c9 JA |
1739 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) { |
1740 | if (this_hctx) { | |
587562d0 | 1741 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 JA |
1742 | blk_mq_sched_insert_requests(this_hctx, this_ctx, |
1743 | &rq_list, | |
bd166ef1 | 1744 | from_schedule); |
320ae51f JA |
1745 | } |
1746 | ||
320ae51f | 1747 | this_q = rq->q; |
67cae4c9 JA |
1748 | this_ctx = rq->mq_ctx; |
1749 | this_hctx = rq->mq_hctx; | |
320ae51f JA |
1750 | depth = 0; |
1751 | } | |
1752 | ||
1753 | depth++; | |
67cae4c9 | 1754 | list_add_tail(&rq->queuelist, &rq_list); |
320ae51f JA |
1755 | } |
1756 | ||
1757 | /* | |
67cae4c9 JA |
1758 | * If 'this_hctx' is set, we know we have entries to complete |
1759 | * on 'rq_list'. Do those. | |
320ae51f | 1760 | */ |
67cae4c9 | 1761 | if (this_hctx) { |
587562d0 | 1762 | trace_block_unplug(this_q, depth, !from_schedule); |
67cae4c9 | 1763 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1764 | from_schedule); |
320ae51f JA |
1765 | } |
1766 | } | |
1767 | ||
14ccb66b CH |
1768 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1769 | unsigned int nr_segs) | |
320ae51f | 1770 | { |
f924cdde CH |
1771 | if (bio->bi_opf & REQ_RAHEAD) |
1772 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1773 | ||
1774 | rq->__sector = bio->bi_iter.bi_sector; | |
1775 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1776 | blk_rq_bio_prep(rq, bio, nr_segs); |
4b570521 | 1777 | |
6e85eaf3 | 1778 | blk_account_io_start(rq, true); |
320ae51f JA |
1779 | } |
1780 | ||
0f95549c MS |
1781 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1782 | struct request *rq, | |
be94f058 | 1783 | blk_qc_t *cookie, bool last) |
f984df1f | 1784 | { |
f984df1f | 1785 | struct request_queue *q = rq->q; |
f984df1f SL |
1786 | struct blk_mq_queue_data bd = { |
1787 | .rq = rq, | |
be94f058 | 1788 | .last = last, |
f984df1f | 1789 | }; |
bd166ef1 | 1790 | blk_qc_t new_cookie; |
f06345ad | 1791 | blk_status_t ret; |
0f95549c MS |
1792 | |
1793 | new_cookie = request_to_qc_t(hctx, rq); | |
1794 | ||
1795 | /* | |
1796 | * For OK queue, we are done. For error, caller may kill it. | |
1797 | * Any other error (busy), just add it to our list as we | |
1798 | * previously would have done. | |
1799 | */ | |
1800 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1801 | switch (ret) { | |
1802 | case BLK_STS_OK: | |
6ce3dd6e | 1803 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1804 | *cookie = new_cookie; |
1805 | break; | |
1806 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1807 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1808 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1809 | __blk_mq_requeue_request(rq); |
1810 | break; | |
1811 | default: | |
6ce3dd6e | 1812 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1813 | *cookie = BLK_QC_T_NONE; |
1814 | break; | |
1815 | } | |
1816 | ||
1817 | return ret; | |
1818 | } | |
1819 | ||
fd9c40f6 | 1820 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 1821 | struct request *rq, |
396eaf21 | 1822 | blk_qc_t *cookie, |
fd9c40f6 | 1823 | bool bypass_insert, bool last) |
0f95549c MS |
1824 | { |
1825 | struct request_queue *q = rq->q; | |
d964f04a ML |
1826 | bool run_queue = true; |
1827 | ||
23d4ee19 | 1828 | /* |
fd9c40f6 | 1829 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 1830 | * |
fd9c40f6 BVA |
1831 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
1832 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
1833 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 1834 | */ |
fd9c40f6 | 1835 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 1836 | run_queue = false; |
fd9c40f6 BVA |
1837 | bypass_insert = false; |
1838 | goto insert; | |
d964f04a | 1839 | } |
f984df1f | 1840 | |
fd9c40f6 BVA |
1841 | if (q->elevator && !bypass_insert) |
1842 | goto insert; | |
2253efc8 | 1843 | |
0bca799b | 1844 | if (!blk_mq_get_dispatch_budget(hctx)) |
fd9c40f6 | 1845 | goto insert; |
bd166ef1 | 1846 | |
8ab6bb9e | 1847 | if (!blk_mq_get_driver_tag(rq)) { |
0bca799b | 1848 | blk_mq_put_dispatch_budget(hctx); |
fd9c40f6 | 1849 | goto insert; |
88022d72 | 1850 | } |
de148297 | 1851 | |
fd9c40f6 BVA |
1852 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
1853 | insert: | |
1854 | if (bypass_insert) | |
1855 | return BLK_STS_RESOURCE; | |
1856 | ||
1857 | blk_mq_request_bypass_insert(rq, run_queue); | |
1858 | return BLK_STS_OK; | |
1859 | } | |
1860 | ||
1861 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1862 | struct request *rq, blk_qc_t *cookie) | |
1863 | { | |
1864 | blk_status_t ret; | |
1865 | int srcu_idx; | |
1866 | ||
1867 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1868 | ||
1869 | hctx_lock(hctx, &srcu_idx); | |
1870 | ||
1871 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
1872 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
1873 | blk_mq_request_bypass_insert(rq, true); | |
1874 | else if (ret != BLK_STS_OK) | |
1875 | blk_mq_end_request(rq, ret); | |
1876 | ||
1877 | hctx_unlock(hctx, srcu_idx); | |
1878 | } | |
1879 | ||
1880 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
1881 | { | |
1882 | blk_status_t ret; | |
1883 | int srcu_idx; | |
1884 | blk_qc_t unused_cookie; | |
1885 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1886 | ||
1887 | hctx_lock(hctx, &srcu_idx); | |
1888 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 1889 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
1890 | |
1891 | return ret; | |
5eb6126e CH |
1892 | } |
1893 | ||
6ce3dd6e ML |
1894 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
1895 | struct list_head *list) | |
1896 | { | |
1897 | while (!list_empty(list)) { | |
fd9c40f6 | 1898 | blk_status_t ret; |
6ce3dd6e ML |
1899 | struct request *rq = list_first_entry(list, struct request, |
1900 | queuelist); | |
1901 | ||
1902 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
1903 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
1904 | if (ret != BLK_STS_OK) { | |
1905 | if (ret == BLK_STS_RESOURCE || | |
1906 | ret == BLK_STS_DEV_RESOURCE) { | |
1907 | blk_mq_request_bypass_insert(rq, | |
c616cbee | 1908 | list_empty(list)); |
fd9c40f6 BVA |
1909 | break; |
1910 | } | |
1911 | blk_mq_end_request(rq, ret); | |
1912 | } | |
6ce3dd6e | 1913 | } |
d666ba98 JA |
1914 | |
1915 | /* | |
1916 | * If we didn't flush the entire list, we could have told | |
1917 | * the driver there was more coming, but that turned out to | |
1918 | * be a lie. | |
1919 | */ | |
fd9c40f6 | 1920 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs) |
d666ba98 | 1921 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
1922 | } |
1923 | ||
ce5b009c JA |
1924 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
1925 | { | |
1926 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1927 | plug->rq_count++; | |
1928 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
1929 | struct request *tmp; | |
1930 | ||
1931 | tmp = list_first_entry(&plug->mq_list, struct request, | |
1932 | queuelist); | |
1933 | if (tmp->q != rq->q) | |
1934 | plug->multiple_queues = true; | |
1935 | } | |
1936 | } | |
1937 | ||
dece1635 | 1938 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1939 | { |
ef295ecf | 1940 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1941 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
7809167d | 1942 | struct blk_mq_alloc_data data = { .flags = 0}; |
07068d5b | 1943 | struct request *rq; |
f984df1f | 1944 | struct blk_plug *plug; |
5b3f341f | 1945 | struct request *same_queue_rq = NULL; |
14ccb66b | 1946 | unsigned int nr_segs; |
7b371636 | 1947 | blk_qc_t cookie; |
07068d5b JA |
1948 | |
1949 | blk_queue_bounce(q, &bio); | |
14ccb66b | 1950 | __blk_queue_split(q, &bio, &nr_segs); |
f36ea50c | 1951 | |
e23947bd | 1952 | if (!bio_integrity_prep(bio)) |
dece1635 | 1953 | return BLK_QC_T_NONE; |
07068d5b | 1954 | |
87c279e6 | 1955 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 1956 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
87c279e6 | 1957 | return BLK_QC_T_NONE; |
f984df1f | 1958 | |
14ccb66b | 1959 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
bd166ef1 JA |
1960 | return BLK_QC_T_NONE; |
1961 | ||
d5337560 | 1962 | rq_qos_throttle(q, bio); |
87760e5e | 1963 | |
7809167d | 1964 | data.cmd_flags = bio->bi_opf; |
f9afca4d | 1965 | rq = blk_mq_get_request(q, bio, &data); |
87760e5e | 1966 | if (unlikely(!rq)) { |
c1c80384 | 1967 | rq_qos_cleanup(q, bio); |
893a1c97 JA |
1968 | |
1969 | cookie = BLK_QC_T_NONE; | |
1970 | if (bio->bi_opf & REQ_NOWAIT_INLINE) | |
1971 | cookie = BLK_QC_T_EAGAIN; | |
1972 | else if (bio->bi_opf & REQ_NOWAIT) | |
03a07c92 | 1973 | bio_wouldblock_error(bio); |
893a1c97 | 1974 | return cookie; |
87760e5e JA |
1975 | } |
1976 | ||
d6f1dda2 XW |
1977 | trace_block_getrq(q, bio, bio->bi_opf); |
1978 | ||
c1c80384 | 1979 | rq_qos_track(q, rq, bio); |
07068d5b | 1980 | |
fd2d3326 | 1981 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1982 | |
970d168d BVA |
1983 | blk_mq_bio_to_request(rq, bio, nr_segs); |
1984 | ||
b49773e7 | 1985 | plug = blk_mq_plug(q, bio); |
07068d5b | 1986 | if (unlikely(is_flush_fua)) { |
923218f6 ML |
1987 | /* bypass scheduler for flush rq */ |
1988 | blk_insert_flush(rq); | |
1989 | blk_mq_run_hw_queue(data.hctx, true); | |
b2c5d16b JA |
1990 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) { |
1991 | /* | |
1992 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
1993 | * we know the driver uses bd->last in a smart fashion. | |
1994 | */ | |
5f0ed774 | 1995 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
1996 | struct request *last = NULL; |
1997 | ||
676d0607 | 1998 | if (!request_count) |
e6c4438b | 1999 | trace_block_plug(q); |
600271d9 SL |
2000 | else |
2001 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2002 | |
600271d9 SL |
2003 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
2004 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
2005 | blk_flush_plug_list(plug, false); |
2006 | trace_block_plug(q); | |
320ae51f | 2007 | } |
b094f89c | 2008 | |
ce5b009c | 2009 | blk_add_rq_to_plug(plug, rq); |
2299722c | 2010 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2011 | /* |
6a83e74d | 2012 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2013 | * Otherwise the existing request in the plug list will be |
2014 | * issued. So the plug list will have one request at most | |
2299722c CH |
2015 | * The plug list might get flushed before this. If that happens, |
2016 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2017 | */ |
2299722c CH |
2018 | if (list_empty(&plug->mq_list)) |
2019 | same_queue_rq = NULL; | |
4711b573 | 2020 | if (same_queue_rq) { |
2299722c | 2021 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2022 | plug->rq_count--; |
2023 | } | |
ce5b009c | 2024 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2025 | trace_block_plug(q); |
2299722c | 2026 | |
dad7a3be | 2027 | if (same_queue_rq) { |
ea4f995e | 2028 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2029 | trace_block_unplug(q, 1, true); |
2299722c | 2030 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2031 | &cookie); |
dad7a3be | 2032 | } |
6ce3dd6e ML |
2033 | } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator && |
2034 | !data.hctx->dispatch_busy)) { | |
fd9c40f6 | 2035 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2036 | } else { |
8fa9f556 | 2037 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2038 | } |
320ae51f | 2039 | |
7b371636 | 2040 | return cookie; |
320ae51f JA |
2041 | } |
2042 | ||
cc71a6f4 JA |
2043 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2044 | unsigned int hctx_idx) | |
95363efd | 2045 | { |
e9b267d9 | 2046 | struct page *page; |
320ae51f | 2047 | |
24d2f903 | 2048 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2049 | int i; |
320ae51f | 2050 | |
24d2f903 | 2051 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2052 | struct request *rq = tags->static_rqs[i]; |
2053 | ||
2054 | if (!rq) | |
e9b267d9 | 2055 | continue; |
d6296d39 | 2056 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2057 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2058 | } |
320ae51f | 2059 | } |
320ae51f | 2060 | |
24d2f903 CH |
2061 | while (!list_empty(&tags->page_list)) { |
2062 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2063 | list_del_init(&page->lru); |
f75782e4 CM |
2064 | /* |
2065 | * Remove kmemleak object previously allocated in | |
273938bf | 2066 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2067 | */ |
2068 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2069 | __free_pages(page, page->private); |
2070 | } | |
cc71a6f4 | 2071 | } |
320ae51f | 2072 | |
cc71a6f4 JA |
2073 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
2074 | { | |
24d2f903 | 2075 | kfree(tags->rqs); |
cc71a6f4 | 2076 | tags->rqs = NULL; |
2af8cbe3 JA |
2077 | kfree(tags->static_rqs); |
2078 | tags->static_rqs = NULL; | |
320ae51f | 2079 | |
24d2f903 | 2080 | blk_mq_free_tags(tags); |
320ae51f JA |
2081 | } |
2082 | ||
cc71a6f4 JA |
2083 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2084 | unsigned int hctx_idx, | |
2085 | unsigned int nr_tags, | |
2086 | unsigned int reserved_tags) | |
320ae51f | 2087 | { |
24d2f903 | 2088 | struct blk_mq_tags *tags; |
59f082e4 | 2089 | int node; |
320ae51f | 2090 | |
7d76f856 | 2091 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2092 | if (node == NUMA_NO_NODE) |
2093 | node = set->numa_node; | |
2094 | ||
2095 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 2096 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
2097 | if (!tags) |
2098 | return NULL; | |
320ae51f | 2099 | |
590b5b7d | 2100 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2101 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2102 | node); |
24d2f903 CH |
2103 | if (!tags->rqs) { |
2104 | blk_mq_free_tags(tags); | |
2105 | return NULL; | |
2106 | } | |
320ae51f | 2107 | |
590b5b7d KC |
2108 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2109 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2110 | node); | |
2af8cbe3 JA |
2111 | if (!tags->static_rqs) { |
2112 | kfree(tags->rqs); | |
2113 | blk_mq_free_tags(tags); | |
2114 | return NULL; | |
2115 | } | |
2116 | ||
cc71a6f4 JA |
2117 | return tags; |
2118 | } | |
2119 | ||
2120 | static size_t order_to_size(unsigned int order) | |
2121 | { | |
2122 | return (size_t)PAGE_SIZE << order; | |
2123 | } | |
2124 | ||
1d9bd516 TH |
2125 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2126 | unsigned int hctx_idx, int node) | |
2127 | { | |
2128 | int ret; | |
2129 | ||
2130 | if (set->ops->init_request) { | |
2131 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2132 | if (ret) | |
2133 | return ret; | |
2134 | } | |
2135 | ||
12f5b931 | 2136 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2137 | return 0; |
2138 | } | |
2139 | ||
cc71a6f4 JA |
2140 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2141 | unsigned int hctx_idx, unsigned int depth) | |
2142 | { | |
2143 | unsigned int i, j, entries_per_page, max_order = 4; | |
2144 | size_t rq_size, left; | |
59f082e4 SL |
2145 | int node; |
2146 | ||
7d76f856 | 2147 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2148 | if (node == NUMA_NO_NODE) |
2149 | node = set->numa_node; | |
cc71a6f4 JA |
2150 | |
2151 | INIT_LIST_HEAD(&tags->page_list); | |
2152 | ||
320ae51f JA |
2153 | /* |
2154 | * rq_size is the size of the request plus driver payload, rounded | |
2155 | * to the cacheline size | |
2156 | */ | |
24d2f903 | 2157 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2158 | cache_line_size()); |
cc71a6f4 | 2159 | left = rq_size * depth; |
320ae51f | 2160 | |
cc71a6f4 | 2161 | for (i = 0; i < depth; ) { |
320ae51f JA |
2162 | int this_order = max_order; |
2163 | struct page *page; | |
2164 | int to_do; | |
2165 | void *p; | |
2166 | ||
b3a834b1 | 2167 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2168 | this_order--; |
2169 | ||
2170 | do { | |
59f082e4 | 2171 | page = alloc_pages_node(node, |
36e1f3d1 | 2172 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2173 | this_order); |
320ae51f JA |
2174 | if (page) |
2175 | break; | |
2176 | if (!this_order--) | |
2177 | break; | |
2178 | if (order_to_size(this_order) < rq_size) | |
2179 | break; | |
2180 | } while (1); | |
2181 | ||
2182 | if (!page) | |
24d2f903 | 2183 | goto fail; |
320ae51f JA |
2184 | |
2185 | page->private = this_order; | |
24d2f903 | 2186 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2187 | |
2188 | p = page_address(page); | |
f75782e4 CM |
2189 | /* |
2190 | * Allow kmemleak to scan these pages as they contain pointers | |
2191 | * to additional allocations like via ops->init_request(). | |
2192 | */ | |
36e1f3d1 | 2193 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2194 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2195 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2196 | left -= to_do * rq_size; |
2197 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2198 | struct request *rq = p; |
2199 | ||
2200 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2201 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2202 | tags->static_rqs[i] = NULL; | |
2203 | goto fail; | |
e9b267d9 CH |
2204 | } |
2205 | ||
320ae51f JA |
2206 | p += rq_size; |
2207 | i++; | |
2208 | } | |
2209 | } | |
cc71a6f4 | 2210 | return 0; |
320ae51f | 2211 | |
24d2f903 | 2212 | fail: |
cc71a6f4 JA |
2213 | blk_mq_free_rqs(set, tags, hctx_idx); |
2214 | return -ENOMEM; | |
320ae51f JA |
2215 | } |
2216 | ||
e57690fe JA |
2217 | /* |
2218 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2219 | * software queue to the hw queue dispatch list, and ensure that it | |
2220 | * gets run. | |
2221 | */ | |
9467f859 | 2222 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2223 | { |
9467f859 | 2224 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2225 | struct blk_mq_ctx *ctx; |
2226 | LIST_HEAD(tmp); | |
c16d6b5a | 2227 | enum hctx_type type; |
484b4061 | 2228 | |
9467f859 | 2229 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 2230 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2231 | type = hctx->type; |
484b4061 JA |
2232 | |
2233 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2234 | if (!list_empty(&ctx->rq_lists[type])) { |
2235 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2236 | blk_mq_hctx_clear_pending(hctx, ctx); |
2237 | } | |
2238 | spin_unlock(&ctx->lock); | |
2239 | ||
2240 | if (list_empty(&tmp)) | |
9467f859 | 2241 | return 0; |
484b4061 | 2242 | |
e57690fe JA |
2243 | spin_lock(&hctx->lock); |
2244 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2245 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2246 | |
2247 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2248 | return 0; |
484b4061 JA |
2249 | } |
2250 | ||
9467f859 | 2251 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2252 | { |
9467f859 TG |
2253 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2254 | &hctx->cpuhp_dead); | |
484b4061 JA |
2255 | } |
2256 | ||
c3b4afca | 2257 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2258 | static void blk_mq_exit_hctx(struct request_queue *q, |
2259 | struct blk_mq_tag_set *set, | |
2260 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2261 | { | |
8ab0b7dc ML |
2262 | if (blk_mq_hw_queue_mapped(hctx)) |
2263 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2264 | |
f70ced09 | 2265 | if (set->ops->exit_request) |
d6296d39 | 2266 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 2267 | |
08e98fc6 ML |
2268 | if (set->ops->exit_hctx) |
2269 | set->ops->exit_hctx(hctx, hctx_idx); | |
2270 | ||
9467f859 | 2271 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2272 | |
2273 | spin_lock(&q->unused_hctx_lock); | |
2274 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2275 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2276 | } |
2277 | ||
624dbe47 ML |
2278 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2279 | struct blk_mq_tag_set *set, int nr_queue) | |
2280 | { | |
2281 | struct blk_mq_hw_ctx *hctx; | |
2282 | unsigned int i; | |
2283 | ||
2284 | queue_for_each_hw_ctx(q, hctx, i) { | |
2285 | if (i == nr_queue) | |
2286 | break; | |
477e19de | 2287 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2288 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2289 | } |
624dbe47 ML |
2290 | } |
2291 | ||
7c6c5b7c ML |
2292 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2293 | { | |
2294 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2295 | ||
2296 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2297 | __alignof__(struct blk_mq_hw_ctx)) != | |
2298 | sizeof(struct blk_mq_hw_ctx)); | |
2299 | ||
2300 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2301 | hw_ctx_size += sizeof(struct srcu_struct); | |
2302 | ||
2303 | return hw_ctx_size; | |
2304 | } | |
2305 | ||
08e98fc6 ML |
2306 | static int blk_mq_init_hctx(struct request_queue *q, |
2307 | struct blk_mq_tag_set *set, | |
2308 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2309 | { |
7c6c5b7c ML |
2310 | hctx->queue_num = hctx_idx; |
2311 | ||
2312 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2313 | ||
2314 | hctx->tags = set->tags[hctx_idx]; | |
2315 | ||
2316 | if (set->ops->init_hctx && | |
2317 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2318 | goto unregister_cpu_notifier; | |
08e98fc6 | 2319 | |
7c6c5b7c ML |
2320 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2321 | hctx->numa_node)) | |
2322 | goto exit_hctx; | |
2323 | return 0; | |
2324 | ||
2325 | exit_hctx: | |
2326 | if (set->ops->exit_hctx) | |
2327 | set->ops->exit_hctx(hctx, hctx_idx); | |
2328 | unregister_cpu_notifier: | |
2329 | blk_mq_remove_cpuhp(hctx); | |
2330 | return -1; | |
2331 | } | |
2332 | ||
2333 | static struct blk_mq_hw_ctx * | |
2334 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2335 | int node) | |
2336 | { | |
2337 | struct blk_mq_hw_ctx *hctx; | |
2338 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2339 | ||
2340 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2341 | if (!hctx) | |
2342 | goto fail_alloc_hctx; | |
2343 | ||
2344 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2345 | goto free_hctx; | |
2346 | ||
2347 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2348 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2349 | node = set->numa_node; |
2350 | hctx->numa_node = node; | |
08e98fc6 | 2351 | |
9f993737 | 2352 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2353 | spin_lock_init(&hctx->lock); |
2354 | INIT_LIST_HEAD(&hctx->dispatch); | |
2355 | hctx->queue = q; | |
2404e607 | 2356 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2357 | |
2f8f1336 ML |
2358 | INIT_LIST_HEAD(&hctx->hctx_list); |
2359 | ||
320ae51f | 2360 | /* |
08e98fc6 ML |
2361 | * Allocate space for all possible cpus to avoid allocation at |
2362 | * runtime | |
320ae51f | 2363 | */ |
d904bfa7 | 2364 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2365 | gfp, node); |
08e98fc6 | 2366 | if (!hctx->ctxs) |
7c6c5b7c | 2367 | goto free_cpumask; |
320ae51f | 2368 | |
5b202853 | 2369 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
7c6c5b7c | 2370 | gfp, node)) |
08e98fc6 | 2371 | goto free_ctxs; |
08e98fc6 | 2372 | hctx->nr_ctx = 0; |
320ae51f | 2373 | |
5815839b | 2374 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2375 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2376 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2377 | ||
5b202853 | 2378 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size, |
7c6c5b7c | 2379 | gfp); |
f70ced09 | 2380 | if (!hctx->fq) |
7c6c5b7c | 2381 | goto free_bitmap; |
320ae51f | 2382 | |
6a83e74d | 2383 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2384 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2385 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2386 | |
7c6c5b7c | 2387 | return hctx; |
320ae51f | 2388 | |
08e98fc6 | 2389 | free_bitmap: |
88459642 | 2390 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2391 | free_ctxs: |
2392 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2393 | free_cpumask: |
2394 | free_cpumask_var(hctx->cpumask); | |
2395 | free_hctx: | |
2396 | kfree(hctx); | |
2397 | fail_alloc_hctx: | |
2398 | return NULL; | |
08e98fc6 | 2399 | } |
320ae51f | 2400 | |
320ae51f JA |
2401 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2402 | unsigned int nr_hw_queues) | |
2403 | { | |
b3c661b1 JA |
2404 | struct blk_mq_tag_set *set = q->tag_set; |
2405 | unsigned int i, j; | |
320ae51f JA |
2406 | |
2407 | for_each_possible_cpu(i) { | |
2408 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2409 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2410 | int k; |
320ae51f | 2411 | |
320ae51f JA |
2412 | __ctx->cpu = i; |
2413 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2414 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2415 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2416 | ||
320ae51f JA |
2417 | __ctx->queue = q; |
2418 | ||
320ae51f JA |
2419 | /* |
2420 | * Set local node, IFF we have more than one hw queue. If | |
2421 | * not, we remain on the home node of the device | |
2422 | */ | |
b3c661b1 JA |
2423 | for (j = 0; j < set->nr_maps; j++) { |
2424 | hctx = blk_mq_map_queue_type(q, j, i); | |
2425 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2426 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2427 | } | |
320ae51f JA |
2428 | } |
2429 | } | |
2430 | ||
cc71a6f4 JA |
2431 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2432 | { | |
2433 | int ret = 0; | |
2434 | ||
2435 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2436 | set->queue_depth, set->reserved_tags); | |
2437 | if (!set->tags[hctx_idx]) | |
2438 | return false; | |
2439 | ||
2440 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2441 | set->queue_depth); | |
2442 | if (!ret) | |
2443 | return true; | |
2444 | ||
2445 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2446 | set->tags[hctx_idx] = NULL; | |
2447 | return false; | |
2448 | } | |
2449 | ||
2450 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2451 | unsigned int hctx_idx) | |
2452 | { | |
4e6db0f2 | 2453 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 JA |
2454 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
2455 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2456 | set->tags[hctx_idx] = NULL; | |
2457 | } | |
cc71a6f4 JA |
2458 | } |
2459 | ||
4b855ad3 | 2460 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2461 | { |
b3c661b1 | 2462 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2463 | struct blk_mq_hw_ctx *hctx; |
2464 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2465 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2466 | |
2467 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2468 | cpumask_clear(hctx->cpumask); |
320ae51f | 2469 | hctx->nr_ctx = 0; |
d416c92c | 2470 | hctx->dispatch_from = NULL; |
320ae51f JA |
2471 | } |
2472 | ||
2473 | /* | |
4b855ad3 | 2474 | * Map software to hardware queues. |
4412efec ML |
2475 | * |
2476 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2477 | */ |
20e4d813 | 2478 | for_each_possible_cpu(i) { |
7d76f856 | 2479 | hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i]; |
4412efec ML |
2480 | /* unmapped hw queue can be remapped after CPU topo changed */ |
2481 | if (!set->tags[hctx_idx] && | |
2482 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
2483 | /* | |
2484 | * If tags initialization fail for some hctx, | |
2485 | * that hctx won't be brought online. In this | |
2486 | * case, remap the current ctx to hctx[0] which | |
2487 | * is guaranteed to always have tags allocated | |
2488 | */ | |
7d76f856 | 2489 | set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0; |
4412efec ML |
2490 | } |
2491 | ||
897bb0c7 | 2492 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2493 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2494 | if (!set->map[j].nr_queues) { |
2495 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2496 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2497 | continue; |
bb94aea1 | 2498 | } |
e5edd5f2 | 2499 | |
b3c661b1 | 2500 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2501 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2502 | /* |
2503 | * If the CPU is already set in the mask, then we've | |
2504 | * mapped this one already. This can happen if | |
2505 | * devices share queues across queue maps. | |
2506 | */ | |
2507 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2508 | continue; | |
2509 | ||
2510 | cpumask_set_cpu(i, hctx->cpumask); | |
2511 | hctx->type = j; | |
2512 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2513 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2514 | ||
2515 | /* | |
2516 | * If the nr_ctx type overflows, we have exceeded the | |
2517 | * amount of sw queues we can support. | |
2518 | */ | |
2519 | BUG_ON(!hctx->nr_ctx); | |
2520 | } | |
bb94aea1 JW |
2521 | |
2522 | for (; j < HCTX_MAX_TYPES; j++) | |
2523 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2524 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2525 | } |
506e931f JA |
2526 | |
2527 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2528 | /* |
2529 | * If no software queues are mapped to this hardware queue, | |
2530 | * disable it and free the request entries. | |
2531 | */ | |
2532 | if (!hctx->nr_ctx) { | |
2533 | /* Never unmap queue 0. We need it as a | |
2534 | * fallback in case of a new remap fails | |
2535 | * allocation | |
2536 | */ | |
2537 | if (i && set->tags[i]) | |
2538 | blk_mq_free_map_and_requests(set, i); | |
2539 | ||
2540 | hctx->tags = NULL; | |
2541 | continue; | |
2542 | } | |
484b4061 | 2543 | |
2a34c087 ML |
2544 | hctx->tags = set->tags[i]; |
2545 | WARN_ON(!hctx->tags); | |
2546 | ||
889fa31f CY |
2547 | /* |
2548 | * Set the map size to the number of mapped software queues. | |
2549 | * This is more accurate and more efficient than looping | |
2550 | * over all possibly mapped software queues. | |
2551 | */ | |
88459642 | 2552 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2553 | |
484b4061 JA |
2554 | /* |
2555 | * Initialize batch roundrobin counts | |
2556 | */ | |
f82ddf19 | 2557 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2558 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2559 | } | |
320ae51f JA |
2560 | } |
2561 | ||
8e8320c9 JA |
2562 | /* |
2563 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2564 | * the queue isn't live yet. | |
2565 | */ | |
2404e607 | 2566 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2567 | { |
2568 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2569 | int i; |
2570 | ||
2404e607 | 2571 | queue_for_each_hw_ctx(q, hctx, i) { |
97889f9a | 2572 | if (shared) |
2404e607 | 2573 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
97889f9a | 2574 | else |
2404e607 JM |
2575 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
2576 | } | |
2577 | } | |
2578 | ||
8e8320c9 JA |
2579 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2580 | bool shared) | |
2404e607 JM |
2581 | { |
2582 | struct request_queue *q; | |
0d2602ca | 2583 | |
705cda97 BVA |
2584 | lockdep_assert_held(&set->tag_list_lock); |
2585 | ||
0d2602ca JA |
2586 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2587 | blk_mq_freeze_queue(q); | |
2404e607 | 2588 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2589 | blk_mq_unfreeze_queue(q); |
2590 | } | |
2591 | } | |
2592 | ||
2593 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2594 | { | |
2595 | struct blk_mq_tag_set *set = q->tag_set; | |
2596 | ||
0d2602ca | 2597 | mutex_lock(&set->tag_list_lock); |
705cda97 | 2598 | list_del_rcu(&q->tag_set_list); |
2404e607 JM |
2599 | if (list_is_singular(&set->tag_list)) { |
2600 | /* just transitioned to unshared */ | |
2601 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2602 | /* update existing queue */ | |
2603 | blk_mq_update_tag_set_depth(set, false); | |
2604 | } | |
0d2602ca | 2605 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 2606 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
2607 | } |
2608 | ||
2609 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2610 | struct request_queue *q) | |
2611 | { | |
0d2602ca | 2612 | mutex_lock(&set->tag_list_lock); |
2404e607 | 2613 | |
ff821d27 JA |
2614 | /* |
2615 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2616 | */ | |
2617 | if (!list_empty(&set->tag_list) && | |
2618 | !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2404e607 JM |
2619 | set->flags |= BLK_MQ_F_TAG_SHARED; |
2620 | /* update existing queue */ | |
2621 | blk_mq_update_tag_set_depth(set, true); | |
2622 | } | |
2623 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2624 | queue_set_hctx_shared(q, true); | |
705cda97 | 2625 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2626 | |
0d2602ca JA |
2627 | mutex_unlock(&set->tag_list_lock); |
2628 | } | |
2629 | ||
1db4909e ML |
2630 | /* All allocations will be freed in release handler of q->mq_kobj */ |
2631 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2632 | { | |
2633 | struct blk_mq_ctxs *ctxs; | |
2634 | int cpu; | |
2635 | ||
2636 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2637 | if (!ctxs) | |
2638 | return -ENOMEM; | |
2639 | ||
2640 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2641 | if (!ctxs->queue_ctx) | |
2642 | goto fail; | |
2643 | ||
2644 | for_each_possible_cpu(cpu) { | |
2645 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2646 | ctx->ctxs = ctxs; | |
2647 | } | |
2648 | ||
2649 | q->mq_kobj = &ctxs->kobj; | |
2650 | q->queue_ctx = ctxs->queue_ctx; | |
2651 | ||
2652 | return 0; | |
2653 | fail: | |
2654 | kfree(ctxs); | |
2655 | return -ENOMEM; | |
2656 | } | |
2657 | ||
e09aae7e ML |
2658 | /* |
2659 | * It is the actual release handler for mq, but we do it from | |
2660 | * request queue's release handler for avoiding use-after-free | |
2661 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2662 | * but we can't group ctx/kctx kobj without it. | |
2663 | */ | |
2664 | void blk_mq_release(struct request_queue *q) | |
2665 | { | |
2f8f1336 ML |
2666 | struct blk_mq_hw_ctx *hctx, *next; |
2667 | int i; | |
e09aae7e | 2668 | |
fbc2a15e ML |
2669 | cancel_delayed_work_sync(&q->requeue_work); |
2670 | ||
2f8f1336 ML |
2671 | queue_for_each_hw_ctx(q, hctx, i) |
2672 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2673 | ||
2674 | /* all hctx are in .unused_hctx_list now */ | |
2675 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2676 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 2677 | kobject_put(&hctx->kobj); |
c3b4afca | 2678 | } |
e09aae7e ML |
2679 | |
2680 | kfree(q->queue_hw_ctx); | |
2681 | ||
7ea5fe31 ML |
2682 | /* |
2683 | * release .mq_kobj and sw queue's kobject now because | |
2684 | * both share lifetime with request queue. | |
2685 | */ | |
2686 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
2687 | } |
2688 | ||
24d2f903 | 2689 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2690 | { |
2691 | struct request_queue *uninit_q, *q; | |
2692 | ||
6d469642 | 2693 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); |
b62c21b7 MS |
2694 | if (!uninit_q) |
2695 | return ERR_PTR(-ENOMEM); | |
2696 | ||
737eb78e DLM |
2697 | /* |
2698 | * Initialize the queue without an elevator. device_add_disk() will do | |
2699 | * the initialization. | |
2700 | */ | |
2701 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
b62c21b7 MS |
2702 | if (IS_ERR(q)) |
2703 | blk_cleanup_queue(uninit_q); | |
2704 | ||
2705 | return q; | |
2706 | } | |
2707 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2708 | ||
9316a9ed JA |
2709 | /* |
2710 | * Helper for setting up a queue with mq ops, given queue depth, and | |
2711 | * the passed in mq ops flags. | |
2712 | */ | |
2713 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
2714 | const struct blk_mq_ops *ops, | |
2715 | unsigned int queue_depth, | |
2716 | unsigned int set_flags) | |
2717 | { | |
2718 | struct request_queue *q; | |
2719 | int ret; | |
2720 | ||
2721 | memset(set, 0, sizeof(*set)); | |
2722 | set->ops = ops; | |
2723 | set->nr_hw_queues = 1; | |
b3c661b1 | 2724 | set->nr_maps = 1; |
9316a9ed JA |
2725 | set->queue_depth = queue_depth; |
2726 | set->numa_node = NUMA_NO_NODE; | |
2727 | set->flags = set_flags; | |
2728 | ||
2729 | ret = blk_mq_alloc_tag_set(set); | |
2730 | if (ret) | |
2731 | return ERR_PTR(ret); | |
2732 | ||
2733 | q = blk_mq_init_queue(set); | |
2734 | if (IS_ERR(q)) { | |
2735 | blk_mq_free_tag_set(set); | |
2736 | return q; | |
2737 | } | |
2738 | ||
2739 | return q; | |
2740 | } | |
2741 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
2742 | ||
34d11ffa JW |
2743 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
2744 | struct blk_mq_tag_set *set, struct request_queue *q, | |
2745 | int hctx_idx, int node) | |
2746 | { | |
2f8f1336 | 2747 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 2748 | |
2f8f1336 ML |
2749 | /* reuse dead hctx first */ |
2750 | spin_lock(&q->unused_hctx_lock); | |
2751 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
2752 | if (tmp->numa_node == node) { | |
2753 | hctx = tmp; | |
2754 | break; | |
2755 | } | |
2756 | } | |
2757 | if (hctx) | |
2758 | list_del_init(&hctx->hctx_list); | |
2759 | spin_unlock(&q->unused_hctx_lock); | |
2760 | ||
2761 | if (!hctx) | |
2762 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 2763 | if (!hctx) |
7c6c5b7c | 2764 | goto fail; |
34d11ffa | 2765 | |
7c6c5b7c ML |
2766 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
2767 | goto free_hctx; | |
34d11ffa JW |
2768 | |
2769 | return hctx; | |
7c6c5b7c ML |
2770 | |
2771 | free_hctx: | |
2772 | kobject_put(&hctx->kobj); | |
2773 | fail: | |
2774 | return NULL; | |
34d11ffa JW |
2775 | } |
2776 | ||
868f2f0b KB |
2777 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2778 | struct request_queue *q) | |
320ae51f | 2779 | { |
e01ad46d | 2780 | int i, j, end; |
868f2f0b | 2781 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 2782 | |
fb350e0a ML |
2783 | /* protect against switching io scheduler */ |
2784 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 2785 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2786 | int node; |
34d11ffa | 2787 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 2788 | |
7d76f856 | 2789 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
2790 | /* |
2791 | * If the hw queue has been mapped to another numa node, | |
2792 | * we need to realloc the hctx. If allocation fails, fallback | |
2793 | * to use the previous one. | |
2794 | */ | |
2795 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
2796 | continue; | |
868f2f0b | 2797 | |
34d11ffa JW |
2798 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
2799 | if (hctx) { | |
2f8f1336 | 2800 | if (hctxs[i]) |
34d11ffa | 2801 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
2802 | hctxs[i] = hctx; |
2803 | } else { | |
2804 | if (hctxs[i]) | |
2805 | pr_warn("Allocate new hctx on node %d fails,\ | |
2806 | fallback to previous one on node %d\n", | |
2807 | node, hctxs[i]->numa_node); | |
2808 | else | |
2809 | break; | |
868f2f0b | 2810 | } |
320ae51f | 2811 | } |
e01ad46d JW |
2812 | /* |
2813 | * Increasing nr_hw_queues fails. Free the newly allocated | |
2814 | * hctxs and keep the previous q->nr_hw_queues. | |
2815 | */ | |
2816 | if (i != set->nr_hw_queues) { | |
2817 | j = q->nr_hw_queues; | |
2818 | end = i; | |
2819 | } else { | |
2820 | j = i; | |
2821 | end = q->nr_hw_queues; | |
2822 | q->nr_hw_queues = set->nr_hw_queues; | |
2823 | } | |
34d11ffa | 2824 | |
e01ad46d | 2825 | for (; j < end; j++) { |
868f2f0b KB |
2826 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
2827 | ||
2828 | if (hctx) { | |
cc71a6f4 JA |
2829 | if (hctx->tags) |
2830 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2831 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2832 | hctxs[j] = NULL; |
868f2f0b KB |
2833 | } |
2834 | } | |
fb350e0a | 2835 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
2836 | } |
2837 | ||
392546ae JA |
2838 | /* |
2839 | * Maximum number of hardware queues we support. For single sets, we'll never | |
2840 | * have more than the CPUs (software queues). For multiple sets, the tag_set | |
2841 | * user may have set ->nr_hw_queues larger. | |
2842 | */ | |
2843 | static unsigned int nr_hw_queues(struct blk_mq_tag_set *set) | |
2844 | { | |
2845 | if (set->nr_maps == 1) | |
2846 | return nr_cpu_ids; | |
2847 | ||
2848 | return max(set->nr_hw_queues, nr_cpu_ids); | |
2849 | } | |
2850 | ||
868f2f0b | 2851 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
737eb78e DLM |
2852 | struct request_queue *q, |
2853 | bool elevator_init) | |
868f2f0b | 2854 | { |
66841672 ML |
2855 | /* mark the queue as mq asap */ |
2856 | q->mq_ops = set->ops; | |
2857 | ||
34dbad5d | 2858 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2859 | blk_mq_poll_stats_bkt, |
2860 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2861 | if (!q->poll_cb) |
2862 | goto err_exit; | |
2863 | ||
1db4909e | 2864 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 2865 | goto err_poll; |
868f2f0b | 2866 | |
737f98cf ML |
2867 | /* init q->mq_kobj and sw queues' kobjects */ |
2868 | blk_mq_sysfs_init(q); | |
2869 | ||
392546ae JA |
2870 | q->nr_queues = nr_hw_queues(set); |
2871 | q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)), | |
868f2f0b KB |
2872 | GFP_KERNEL, set->numa_node); |
2873 | if (!q->queue_hw_ctx) | |
1db4909e | 2874 | goto err_sys_init; |
868f2f0b | 2875 | |
2f8f1336 ML |
2876 | INIT_LIST_HEAD(&q->unused_hctx_list); |
2877 | spin_lock_init(&q->unused_hctx_lock); | |
2878 | ||
868f2f0b KB |
2879 | blk_mq_realloc_hw_ctxs(set, q); |
2880 | if (!q->nr_hw_queues) | |
2881 | goto err_hctxs; | |
320ae51f | 2882 | |
287922eb | 2883 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2884 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 2885 | |
a8908939 | 2886 | q->tag_set = set; |
320ae51f | 2887 | |
94eddfbe | 2888 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
2889 | if (set->nr_maps > HCTX_TYPE_POLL && |
2890 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 2891 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 2892 | |
1be036e9 CH |
2893 | q->sg_reserved_size = INT_MAX; |
2894 | ||
2849450a | 2895 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2896 | INIT_LIST_HEAD(&q->requeue_list); |
2897 | spin_lock_init(&q->requeue_lock); | |
2898 | ||
254d259d | 2899 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2900 | |
eba71768 JA |
2901 | /* |
2902 | * Do this after blk_queue_make_request() overrides it... | |
2903 | */ | |
2904 | q->nr_requests = set->queue_depth; | |
2905 | ||
64f1c21e JA |
2906 | /* |
2907 | * Default to classic polling | |
2908 | */ | |
29ece8b4 | 2909 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 2910 | |
24d2f903 | 2911 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2912 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2913 | blk_mq_map_swqueue(q); |
4593fdbe | 2914 | |
737eb78e DLM |
2915 | if (elevator_init) |
2916 | elevator_init_mq(q); | |
d3484991 | 2917 | |
320ae51f | 2918 | return q; |
18741986 | 2919 | |
320ae51f | 2920 | err_hctxs: |
868f2f0b | 2921 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 2922 | q->nr_hw_queues = 0; |
1db4909e ML |
2923 | err_sys_init: |
2924 | blk_mq_sysfs_deinit(q); | |
41de54c6 JS |
2925 | err_poll: |
2926 | blk_stat_free_callback(q->poll_cb); | |
2927 | q->poll_cb = NULL; | |
c7de5726 ML |
2928 | err_exit: |
2929 | q->mq_ops = NULL; | |
320ae51f JA |
2930 | return ERR_PTR(-ENOMEM); |
2931 | } | |
b62c21b7 | 2932 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 2933 | |
c7e2d94b ML |
2934 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
2935 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 2936 | { |
624dbe47 | 2937 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2938 | |
0d2602ca | 2939 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2940 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2941 | } |
320ae51f | 2942 | |
a5164405 JA |
2943 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2944 | { | |
2945 | int i; | |
2946 | ||
cc71a6f4 JA |
2947 | for (i = 0; i < set->nr_hw_queues; i++) |
2948 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2949 | goto out_unwind; |
a5164405 JA |
2950 | |
2951 | return 0; | |
2952 | ||
2953 | out_unwind: | |
2954 | while (--i >= 0) | |
cc71a6f4 | 2955 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2956 | |
a5164405 JA |
2957 | return -ENOMEM; |
2958 | } | |
2959 | ||
2960 | /* | |
2961 | * Allocate the request maps associated with this tag_set. Note that this | |
2962 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2963 | * will be updated to reflect the allocated depth. | |
2964 | */ | |
2965 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2966 | { | |
2967 | unsigned int depth; | |
2968 | int err; | |
2969 | ||
2970 | depth = set->queue_depth; | |
2971 | do { | |
2972 | err = __blk_mq_alloc_rq_maps(set); | |
2973 | if (!err) | |
2974 | break; | |
2975 | ||
2976 | set->queue_depth >>= 1; | |
2977 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2978 | err = -ENOMEM; | |
2979 | break; | |
2980 | } | |
2981 | } while (set->queue_depth); | |
2982 | ||
2983 | if (!set->queue_depth || err) { | |
2984 | pr_err("blk-mq: failed to allocate request map\n"); | |
2985 | return -ENOMEM; | |
2986 | } | |
2987 | ||
2988 | if (depth != set->queue_depth) | |
2989 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2990 | depth, set->queue_depth); | |
2991 | ||
2992 | return 0; | |
2993 | } | |
2994 | ||
ebe8bddb OS |
2995 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2996 | { | |
59388702 | 2997 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
2998 | int i; |
2999 | ||
7d4901a9 ML |
3000 | /* |
3001 | * transport .map_queues is usually done in the following | |
3002 | * way: | |
3003 | * | |
3004 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3005 | * mask = get_cpu_mask(queue) | |
3006 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3007 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3008 | * } |
3009 | * | |
3010 | * When we need to remap, the table has to be cleared for | |
3011 | * killing stale mapping since one CPU may not be mapped | |
3012 | * to any hw queue. | |
3013 | */ | |
b3c661b1 JA |
3014 | for (i = 0; i < set->nr_maps; i++) |
3015 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3016 | |
ebe8bddb | 3017 | return set->ops->map_queues(set); |
b3c661b1 JA |
3018 | } else { |
3019 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3020 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3021 | } |
ebe8bddb OS |
3022 | } |
3023 | ||
a4391c64 JA |
3024 | /* |
3025 | * Alloc a tag set to be associated with one or more request queues. | |
3026 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3027 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3028 | * value will be stored in set->queue_depth. |
3029 | */ | |
24d2f903 CH |
3030 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3031 | { | |
b3c661b1 | 3032 | int i, ret; |
da695ba2 | 3033 | |
205fb5f5 BVA |
3034 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3035 | ||
24d2f903 CH |
3036 | if (!set->nr_hw_queues) |
3037 | return -EINVAL; | |
a4391c64 | 3038 | if (!set->queue_depth) |
24d2f903 CH |
3039 | return -EINVAL; |
3040 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3041 | return -EINVAL; | |
3042 | ||
7d7e0f90 | 3043 | if (!set->ops->queue_rq) |
24d2f903 CH |
3044 | return -EINVAL; |
3045 | ||
de148297 ML |
3046 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3047 | return -EINVAL; | |
3048 | ||
a4391c64 JA |
3049 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3050 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3051 | BLK_MQ_MAX_DEPTH); | |
3052 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3053 | } | |
24d2f903 | 3054 | |
b3c661b1 JA |
3055 | if (!set->nr_maps) |
3056 | set->nr_maps = 1; | |
3057 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3058 | return -EINVAL; | |
3059 | ||
6637fadf SL |
3060 | /* |
3061 | * If a crashdump is active, then we are potentially in a very | |
3062 | * memory constrained environment. Limit us to 1 queue and | |
3063 | * 64 tags to prevent using too much memory. | |
3064 | */ | |
3065 | if (is_kdump_kernel()) { | |
3066 | set->nr_hw_queues = 1; | |
59388702 | 3067 | set->nr_maps = 1; |
6637fadf SL |
3068 | set->queue_depth = min(64U, set->queue_depth); |
3069 | } | |
868f2f0b | 3070 | /* |
392546ae JA |
3071 | * There is no use for more h/w queues than cpus if we just have |
3072 | * a single map | |
868f2f0b | 3073 | */ |
392546ae | 3074 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3075 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3076 | |
392546ae | 3077 | set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *), |
24d2f903 CH |
3078 | GFP_KERNEL, set->numa_node); |
3079 | if (!set->tags) | |
a5164405 | 3080 | return -ENOMEM; |
24d2f903 | 3081 | |
da695ba2 | 3082 | ret = -ENOMEM; |
b3c661b1 JA |
3083 | for (i = 0; i < set->nr_maps; i++) { |
3084 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3085 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3086 | GFP_KERNEL, set->numa_node); |
3087 | if (!set->map[i].mq_map) | |
3088 | goto out_free_mq_map; | |
59388702 | 3089 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3090 | } |
bdd17e75 | 3091 | |
ebe8bddb | 3092 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3093 | if (ret) |
3094 | goto out_free_mq_map; | |
3095 | ||
3096 | ret = blk_mq_alloc_rq_maps(set); | |
3097 | if (ret) | |
bdd17e75 | 3098 | goto out_free_mq_map; |
24d2f903 | 3099 | |
0d2602ca JA |
3100 | mutex_init(&set->tag_list_lock); |
3101 | INIT_LIST_HEAD(&set->tag_list); | |
3102 | ||
24d2f903 | 3103 | return 0; |
bdd17e75 CH |
3104 | |
3105 | out_free_mq_map: | |
b3c661b1 JA |
3106 | for (i = 0; i < set->nr_maps; i++) { |
3107 | kfree(set->map[i].mq_map); | |
3108 | set->map[i].mq_map = NULL; | |
3109 | } | |
5676e7b6 RE |
3110 | kfree(set->tags); |
3111 | set->tags = NULL; | |
da695ba2 | 3112 | return ret; |
24d2f903 CH |
3113 | } |
3114 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3115 | ||
3116 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3117 | { | |
b3c661b1 | 3118 | int i, j; |
24d2f903 | 3119 | |
392546ae | 3120 | for (i = 0; i < nr_hw_queues(set); i++) |
cc71a6f4 | 3121 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3122 | |
b3c661b1 JA |
3123 | for (j = 0; j < set->nr_maps; j++) { |
3124 | kfree(set->map[j].mq_map); | |
3125 | set->map[j].mq_map = NULL; | |
3126 | } | |
bdd17e75 | 3127 | |
981bd189 | 3128 | kfree(set->tags); |
5676e7b6 | 3129 | set->tags = NULL; |
24d2f903 CH |
3130 | } |
3131 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3132 | ||
e3a2b3f9 JA |
3133 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3134 | { | |
3135 | struct blk_mq_tag_set *set = q->tag_set; | |
3136 | struct blk_mq_hw_ctx *hctx; | |
3137 | int i, ret; | |
3138 | ||
bd166ef1 | 3139 | if (!set) |
e3a2b3f9 JA |
3140 | return -EINVAL; |
3141 | ||
e5fa8140 AZ |
3142 | if (q->nr_requests == nr) |
3143 | return 0; | |
3144 | ||
70f36b60 | 3145 | blk_mq_freeze_queue(q); |
24f5a90f | 3146 | blk_mq_quiesce_queue(q); |
70f36b60 | 3147 | |
e3a2b3f9 JA |
3148 | ret = 0; |
3149 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3150 | if (!hctx->tags) |
3151 | continue; | |
bd166ef1 JA |
3152 | /* |
3153 | * If we're using an MQ scheduler, just update the scheduler | |
3154 | * queue depth. This is similar to what the old code would do. | |
3155 | */ | |
70f36b60 | 3156 | if (!hctx->sched_tags) { |
c2e82a23 | 3157 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 JA |
3158 | false); |
3159 | } else { | |
3160 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3161 | nr, true); | |
3162 | } | |
e3a2b3f9 JA |
3163 | if (ret) |
3164 | break; | |
77f1e0a5 JA |
3165 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3166 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 JA |
3167 | } |
3168 | ||
3169 | if (!ret) | |
3170 | q->nr_requests = nr; | |
3171 | ||
24f5a90f | 3172 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3173 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3174 | |
e3a2b3f9 JA |
3175 | return ret; |
3176 | } | |
3177 | ||
d48ece20 JW |
3178 | /* |
3179 | * request_queue and elevator_type pair. | |
3180 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3181 | * the elevator_type associated with a request_queue. | |
3182 | */ | |
3183 | struct blk_mq_qe_pair { | |
3184 | struct list_head node; | |
3185 | struct request_queue *q; | |
3186 | struct elevator_type *type; | |
3187 | }; | |
3188 | ||
3189 | /* | |
3190 | * Cache the elevator_type in qe pair list and switch the | |
3191 | * io scheduler to 'none' | |
3192 | */ | |
3193 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3194 | struct request_queue *q) | |
3195 | { | |
3196 | struct blk_mq_qe_pair *qe; | |
3197 | ||
3198 | if (!q->elevator) | |
3199 | return true; | |
3200 | ||
3201 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3202 | if (!qe) | |
3203 | return false; | |
3204 | ||
3205 | INIT_LIST_HEAD(&qe->node); | |
3206 | qe->q = q; | |
3207 | qe->type = q->elevator->type; | |
3208 | list_add(&qe->node, head); | |
3209 | ||
3210 | mutex_lock(&q->sysfs_lock); | |
3211 | /* | |
3212 | * After elevator_switch_mq, the previous elevator_queue will be | |
3213 | * released by elevator_release. The reference of the io scheduler | |
3214 | * module get by elevator_get will also be put. So we need to get | |
3215 | * a reference of the io scheduler module here to prevent it to be | |
3216 | * removed. | |
3217 | */ | |
3218 | __module_get(qe->type->elevator_owner); | |
3219 | elevator_switch_mq(q, NULL); | |
3220 | mutex_unlock(&q->sysfs_lock); | |
3221 | ||
3222 | return true; | |
3223 | } | |
3224 | ||
3225 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3226 | struct request_queue *q) | |
3227 | { | |
3228 | struct blk_mq_qe_pair *qe; | |
3229 | struct elevator_type *t = NULL; | |
3230 | ||
3231 | list_for_each_entry(qe, head, node) | |
3232 | if (qe->q == q) { | |
3233 | t = qe->type; | |
3234 | break; | |
3235 | } | |
3236 | ||
3237 | if (!t) | |
3238 | return; | |
3239 | ||
3240 | list_del(&qe->node); | |
3241 | kfree(qe); | |
3242 | ||
3243 | mutex_lock(&q->sysfs_lock); | |
3244 | elevator_switch_mq(q, t); | |
3245 | mutex_unlock(&q->sysfs_lock); | |
3246 | } | |
3247 | ||
e4dc2b32 KB |
3248 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3249 | int nr_hw_queues) | |
868f2f0b KB |
3250 | { |
3251 | struct request_queue *q; | |
d48ece20 | 3252 | LIST_HEAD(head); |
e01ad46d | 3253 | int prev_nr_hw_queues; |
868f2f0b | 3254 | |
705cda97 BVA |
3255 | lockdep_assert_held(&set->tag_list_lock); |
3256 | ||
392546ae | 3257 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b KB |
3258 | nr_hw_queues = nr_cpu_ids; |
3259 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
3260 | return; | |
3261 | ||
3262 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3263 | blk_mq_freeze_queue(q); | |
f5bbbbe4 JW |
3264 | /* |
3265 | * Sync with blk_mq_queue_tag_busy_iter. | |
3266 | */ | |
3267 | synchronize_rcu(); | |
d48ece20 JW |
3268 | /* |
3269 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3270 | * with the previous scheduler. We will switch back once we are done | |
3271 | * updating the new sw to hw queue mappings. | |
3272 | */ | |
3273 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3274 | if (!blk_mq_elv_switch_none(&head, q)) | |
3275 | goto switch_back; | |
868f2f0b | 3276 | |
477e19de JW |
3277 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3278 | blk_mq_debugfs_unregister_hctxs(q); | |
3279 | blk_mq_sysfs_unregister(q); | |
3280 | } | |
3281 | ||
e01ad46d | 3282 | prev_nr_hw_queues = set->nr_hw_queues; |
868f2f0b | 3283 | set->nr_hw_queues = nr_hw_queues; |
ebe8bddb | 3284 | blk_mq_update_queue_map(set); |
e01ad46d | 3285 | fallback: |
868f2f0b KB |
3286 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3287 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3288 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3289 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3290 | nr_hw_queues, prev_nr_hw_queues); | |
3291 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3292 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3293 | goto fallback; |
3294 | } | |
477e19de JW |
3295 | blk_mq_map_swqueue(q); |
3296 | } | |
3297 | ||
3298 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3299 | blk_mq_sysfs_register(q); | |
3300 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3301 | } |
3302 | ||
d48ece20 JW |
3303 | switch_back: |
3304 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3305 | blk_mq_elv_switch_back(&head, q); | |
3306 | ||
868f2f0b KB |
3307 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3308 | blk_mq_unfreeze_queue(q); | |
3309 | } | |
e4dc2b32 KB |
3310 | |
3311 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3312 | { | |
3313 | mutex_lock(&set->tag_list_lock); | |
3314 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3315 | mutex_unlock(&set->tag_list_lock); | |
3316 | } | |
868f2f0b KB |
3317 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3318 | ||
34dbad5d OS |
3319 | /* Enable polling stats and return whether they were already enabled. */ |
3320 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3321 | { | |
3322 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3323 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3324 | return true; |
3325 | blk_stat_add_callback(q, q->poll_cb); | |
3326 | return false; | |
3327 | } | |
3328 | ||
3329 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3330 | { | |
3331 | /* | |
3332 | * We don't arm the callback if polling stats are not enabled or the | |
3333 | * callback is already active. | |
3334 | */ | |
3335 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3336 | blk_stat_is_active(q->poll_cb)) | |
3337 | return; | |
3338 | ||
3339 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3340 | } | |
3341 | ||
3342 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3343 | { | |
3344 | struct request_queue *q = cb->data; | |
720b8ccc | 3345 | int bucket; |
34dbad5d | 3346 | |
720b8ccc SB |
3347 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3348 | if (cb->stat[bucket].nr_samples) | |
3349 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3350 | } | |
34dbad5d OS |
3351 | } |
3352 | ||
64f1c21e JA |
3353 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
3354 | struct blk_mq_hw_ctx *hctx, | |
3355 | struct request *rq) | |
3356 | { | |
64f1c21e | 3357 | unsigned long ret = 0; |
720b8ccc | 3358 | int bucket; |
64f1c21e JA |
3359 | |
3360 | /* | |
3361 | * If stats collection isn't on, don't sleep but turn it on for | |
3362 | * future users | |
3363 | */ | |
34dbad5d | 3364 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3365 | return 0; |
3366 | ||
64f1c21e JA |
3367 | /* |
3368 | * As an optimistic guess, use half of the mean service time | |
3369 | * for this type of request. We can (and should) make this smarter. | |
3370 | * For instance, if the completion latencies are tight, we can | |
3371 | * get closer than just half the mean. This is especially | |
3372 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3373 | * than ~10 usec. We do use the stats for the relevant IO size |
3374 | * if available which does lead to better estimates. | |
64f1c21e | 3375 | */ |
720b8ccc SB |
3376 | bucket = blk_mq_poll_stats_bkt(rq); |
3377 | if (bucket < 0) | |
3378 | return ret; | |
3379 | ||
3380 | if (q->poll_stat[bucket].nr_samples) | |
3381 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3382 | |
3383 | return ret; | |
3384 | } | |
3385 | ||
06426adf | 3386 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 3387 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
3388 | struct request *rq) |
3389 | { | |
3390 | struct hrtimer_sleeper hs; | |
3391 | enum hrtimer_mode mode; | |
64f1c21e | 3392 | unsigned int nsecs; |
06426adf JA |
3393 | ktime_t kt; |
3394 | ||
76a86f9d | 3395 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3396 | return false; |
3397 | ||
3398 | /* | |
1052b8ac | 3399 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3400 | * |
64f1c21e JA |
3401 | * 0: use half of prev avg |
3402 | * >0: use this specific value | |
3403 | */ | |
1052b8ac | 3404 | if (q->poll_nsec > 0) |
64f1c21e JA |
3405 | nsecs = q->poll_nsec; |
3406 | else | |
3407 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
3408 | ||
3409 | if (!nsecs) | |
06426adf JA |
3410 | return false; |
3411 | ||
76a86f9d | 3412 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3413 | |
3414 | /* | |
3415 | * This will be replaced with the stats tracking code, using | |
3416 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3417 | */ | |
8b0e1953 | 3418 | kt = nsecs; |
06426adf JA |
3419 | |
3420 | mode = HRTIMER_MODE_REL; | |
3421 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
3422 | hrtimer_set_expires(&hs.timer, kt); | |
3423 | ||
3424 | hrtimer_init_sleeper(&hs, current); | |
3425 | do { | |
5a61c363 | 3426 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3427 | break; |
3428 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3429 | hrtimer_start_expires(&hs.timer, mode); | |
3430 | if (hs.task) | |
3431 | io_schedule(); | |
3432 | hrtimer_cancel(&hs.timer); | |
3433 | mode = HRTIMER_MODE_ABS; | |
3434 | } while (hs.task && !signal_pending(current)); | |
3435 | ||
3436 | __set_current_state(TASK_RUNNING); | |
3437 | destroy_hrtimer_on_stack(&hs.timer); | |
3438 | return true; | |
3439 | } | |
3440 | ||
1052b8ac JA |
3441 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3442 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3443 | { |
1052b8ac JA |
3444 | struct request *rq; |
3445 | ||
29ece8b4 | 3446 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3447 | return false; |
3448 | ||
3449 | if (!blk_qc_t_is_internal(cookie)) | |
3450 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3451 | else { | |
3452 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3453 | /* | |
3454 | * With scheduling, if the request has completed, we'll | |
3455 | * get a NULL return here, as we clear the sched tag when | |
3456 | * that happens. The request still remains valid, like always, | |
3457 | * so we should be safe with just the NULL check. | |
3458 | */ | |
3459 | if (!rq) | |
3460 | return false; | |
3461 | } | |
3462 | ||
3463 | return blk_mq_poll_hybrid_sleep(q, hctx, rq); | |
3464 | } | |
3465 | ||
529262d5 CH |
3466 | /** |
3467 | * blk_poll - poll for IO completions | |
3468 | * @q: the queue | |
3469 | * @cookie: cookie passed back at IO submission time | |
3470 | * @spin: whether to spin for completions | |
3471 | * | |
3472 | * Description: | |
3473 | * Poll for completions on the passed in queue. Returns number of | |
3474 | * completed entries found. If @spin is true, then blk_poll will continue | |
3475 | * looping until at least one completion is found, unless the task is | |
3476 | * otherwise marked running (or we need to reschedule). | |
3477 | */ | |
3478 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3479 | { |
3480 | struct blk_mq_hw_ctx *hctx; | |
bbd7bb70 JA |
3481 | long state; |
3482 | ||
529262d5 CH |
3483 | if (!blk_qc_t_valid(cookie) || |
3484 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3485 | return 0; |
3486 | ||
529262d5 CH |
3487 | if (current->plug) |
3488 | blk_flush_plug_list(current->plug, false); | |
3489 | ||
1052b8ac JA |
3490 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3491 | ||
06426adf JA |
3492 | /* |
3493 | * If we sleep, have the caller restart the poll loop to reset | |
3494 | * the state. Like for the other success return cases, the | |
3495 | * caller is responsible for checking if the IO completed. If | |
3496 | * the IO isn't complete, we'll get called again and will go | |
3497 | * straight to the busy poll loop. | |
3498 | */ | |
1052b8ac | 3499 | if (blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3500 | return 1; |
06426adf | 3501 | |
bbd7bb70 JA |
3502 | hctx->poll_considered++; |
3503 | ||
3504 | state = current->state; | |
aa61bec3 | 3505 | do { |
bbd7bb70 JA |
3506 | int ret; |
3507 | ||
3508 | hctx->poll_invoked++; | |
3509 | ||
9743139c | 3510 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3511 | if (ret > 0) { |
3512 | hctx->poll_success++; | |
849a3700 | 3513 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3514 | return ret; |
bbd7bb70 JA |
3515 | } |
3516 | ||
3517 | if (signal_pending_state(state, current)) | |
849a3700 | 3518 | __set_current_state(TASK_RUNNING); |
bbd7bb70 JA |
3519 | |
3520 | if (current->state == TASK_RUNNING) | |
85f4d4b6 | 3521 | return 1; |
0a1b8b87 | 3522 | if (ret < 0 || !spin) |
bbd7bb70 JA |
3523 | break; |
3524 | cpu_relax(); | |
aa61bec3 | 3525 | } while (!need_resched()); |
bbd7bb70 | 3526 | |
67b4110f | 3527 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3528 | return 0; |
bbd7bb70 | 3529 | } |
529262d5 | 3530 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 3531 | |
9cf2bab6 JA |
3532 | unsigned int blk_mq_rq_cpu(struct request *rq) |
3533 | { | |
3534 | return rq->mq_ctx->cpu; | |
3535 | } | |
3536 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3537 | ||
320ae51f JA |
3538 | static int __init blk_mq_init(void) |
3539 | { | |
9467f859 TG |
3540 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
3541 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
3542 | return 0; |
3543 | } | |
3544 | subsys_initcall(blk_mq_init); |