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> | |
fe45e630 | 13 | #include <linux/blk-integrity.h> |
f75782e4 | 14 | #include <linux/kmemleak.h> |
320ae51f JA |
15 | #include <linux/mm.h> |
16 | #include <linux/init.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/workqueue.h> | |
19 | #include <linux/smp.h> | |
e41d12f5 | 20 | #include <linux/interrupt.h> |
320ae51f JA |
21 | #include <linux/llist.h> |
22 | #include <linux/list_sort.h> | |
23 | #include <linux/cpu.h> | |
24 | #include <linux/cache.h> | |
25 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 26 | #include <linux/sched/topology.h> |
174cd4b1 | 27 | #include <linux/sched/signal.h> |
320ae51f | 28 | #include <linux/delay.h> |
aedcd72f | 29 | #include <linux/crash_dump.h> |
88c7b2b7 | 30 | #include <linux/prefetch.h> |
a892c8d5 | 31 | #include <linux/blk-crypto.h> |
320ae51f JA |
32 | |
33 | #include <trace/events/block.h> | |
34 | ||
35 | #include <linux/blk-mq.h> | |
54d4e6ab | 36 | #include <linux/t10-pi.h> |
320ae51f JA |
37 | #include "blk.h" |
38 | #include "blk-mq.h" | |
9c1051aa | 39 | #include "blk-mq-debugfs.h" |
320ae51f | 40 | #include "blk-mq-tag.h" |
986d413b | 41 | #include "blk-pm.h" |
cf43e6be | 42 | #include "blk-stat.h" |
bd166ef1 | 43 | #include "blk-mq-sched.h" |
c1c80384 | 44 | #include "blk-rq-qos.h" |
320ae51f | 45 | |
f9ab4918 | 46 | static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); |
c3077b5d | 47 | |
34dbad5d OS |
48 | static void blk_mq_poll_stats_start(struct request_queue *q); |
49 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
50 | ||
720b8ccc SB |
51 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
52 | { | |
3d244306 | 53 | int ddir, sectors, bucket; |
720b8ccc | 54 | |
99c749a4 | 55 | ddir = rq_data_dir(rq); |
3d244306 | 56 | sectors = blk_rq_stats_sectors(rq); |
720b8ccc | 57 | |
3d244306 | 58 | bucket = ddir + 2 * ilog2(sectors); |
720b8ccc SB |
59 | |
60 | if (bucket < 0) | |
61 | return -1; | |
62 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
63 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
64 | ||
65 | return bucket; | |
66 | } | |
67 | ||
3e08773c CH |
68 | #define BLK_QC_T_SHIFT 16 |
69 | #define BLK_QC_T_INTERNAL (1U << 31) | |
70 | ||
f70299f0 CH |
71 | static inline struct blk_mq_hw_ctx *blk_qc_to_hctx(struct request_queue *q, |
72 | blk_qc_t qc) | |
73 | { | |
74 | return q->queue_hw_ctx[(qc & ~BLK_QC_T_INTERNAL) >> BLK_QC_T_SHIFT]; | |
75 | } | |
76 | ||
c6699d6f CH |
77 | static inline struct request *blk_qc_to_rq(struct blk_mq_hw_ctx *hctx, |
78 | blk_qc_t qc) | |
79 | { | |
efbabbe1 CH |
80 | unsigned int tag = qc & ((1U << BLK_QC_T_SHIFT) - 1); |
81 | ||
82 | if (qc & BLK_QC_T_INTERNAL) | |
83 | return blk_mq_tag_to_rq(hctx->sched_tags, tag); | |
84 | return blk_mq_tag_to_rq(hctx->tags, tag); | |
c6699d6f CH |
85 | } |
86 | ||
3e08773c CH |
87 | static inline blk_qc_t blk_rq_to_qc(struct request *rq) |
88 | { | |
89 | return (rq->mq_hctx->queue_num << BLK_QC_T_SHIFT) | | |
90 | (rq->tag != -1 ? | |
91 | rq->tag : (rq->internal_tag | BLK_QC_T_INTERNAL)); | |
92 | } | |
93 | ||
320ae51f | 94 | /* |
85fae294 YY |
95 | * Check if any of the ctx, dispatch list or elevator |
96 | * have pending work in this hardware queue. | |
320ae51f | 97 | */ |
79f720a7 | 98 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 99 | { |
79f720a7 JA |
100 | return !list_empty_careful(&hctx->dispatch) || |
101 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 102 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
103 | } |
104 | ||
320ae51f JA |
105 | /* |
106 | * Mark this ctx as having pending work in this hardware queue | |
107 | */ | |
108 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
109 | struct blk_mq_ctx *ctx) | |
110 | { | |
f31967f0 JA |
111 | const int bit = ctx->index_hw[hctx->type]; |
112 | ||
113 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
114 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
115 | } |
116 | ||
117 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
118 | struct blk_mq_ctx *ctx) | |
119 | { | |
f31967f0 JA |
120 | const int bit = ctx->index_hw[hctx->type]; |
121 | ||
122 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
123 | } |
124 | ||
f299b7c7 | 125 | struct mq_inflight { |
8446fe92 | 126 | struct block_device *part; |
a2e80f6f | 127 | unsigned int inflight[2]; |
f299b7c7 JA |
128 | }; |
129 | ||
7baa8572 | 130 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
131 | struct request *rq, void *priv, |
132 | bool reserved) | |
133 | { | |
134 | struct mq_inflight *mi = priv; | |
135 | ||
b0d97557 JX |
136 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
137 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 138 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
139 | |
140 | return true; | |
f299b7c7 JA |
141 | } |
142 | ||
8446fe92 CH |
143 | unsigned int blk_mq_in_flight(struct request_queue *q, |
144 | struct block_device *part) | |
f299b7c7 | 145 | { |
a2e80f6f | 146 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 147 | |
f299b7c7 | 148 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 149 | |
a2e80f6f | 150 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
151 | } |
152 | ||
8446fe92 CH |
153 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
154 | unsigned int inflight[2]) | |
bf0ddaba | 155 | { |
a2e80f6f | 156 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 157 | |
bb4e6b14 | 158 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
159 | inflight[0] = mi.inflight[0]; |
160 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
161 | } |
162 | ||
1671d522 | 163 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 164 | { |
7996a8b5 BL |
165 | mutex_lock(&q->mq_freeze_lock); |
166 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 167 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 168 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 169 | if (queue_is_mq(q)) |
055f6e18 | 170 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
171 | } else { |
172 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 173 | } |
f3af020b | 174 | } |
1671d522 | 175 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 176 | |
6bae363e | 177 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 178 | { |
3ef28e83 | 179 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 180 | } |
6bae363e | 181 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 182 | |
f91328c4 KB |
183 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
184 | unsigned long timeout) | |
185 | { | |
186 | return wait_event_timeout(q->mq_freeze_wq, | |
187 | percpu_ref_is_zero(&q->q_usage_counter), | |
188 | timeout); | |
189 | } | |
190 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 191 | |
f3af020b TH |
192 | /* |
193 | * Guarantee no request is in use, so we can change any data structure of | |
194 | * the queue afterward. | |
195 | */ | |
3ef28e83 | 196 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 197 | { |
3ef28e83 DW |
198 | /* |
199 | * In the !blk_mq case we are only calling this to kill the | |
200 | * q_usage_counter, otherwise this increases the freeze depth | |
201 | * and waits for it to return to zero. For this reason there is | |
202 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
203 | * exported to drivers as the only user for unfreeze is blk_mq. | |
204 | */ | |
1671d522 | 205 | blk_freeze_queue_start(q); |
f3af020b TH |
206 | blk_mq_freeze_queue_wait(q); |
207 | } | |
3ef28e83 DW |
208 | |
209 | void blk_mq_freeze_queue(struct request_queue *q) | |
210 | { | |
211 | /* | |
212 | * ...just an alias to keep freeze and unfreeze actions balanced | |
213 | * in the blk_mq_* namespace | |
214 | */ | |
215 | blk_freeze_queue(q); | |
216 | } | |
c761d96b | 217 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 218 | |
aec89dc5 | 219 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 220 | { |
7996a8b5 | 221 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
222 | if (force_atomic) |
223 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
224 | q->mq_freeze_depth--; |
225 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
226 | if (!q->mq_freeze_depth) { | |
bdd63160 | 227 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 228 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 229 | } |
7996a8b5 | 230 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 231 | } |
aec89dc5 CH |
232 | |
233 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
234 | { | |
235 | __blk_mq_unfreeze_queue(q, false); | |
236 | } | |
b4c6a028 | 237 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 238 | |
852ec809 BVA |
239 | /* |
240 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
241 | * mpt3sas driver such that this function can be removed. | |
242 | */ | |
243 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
244 | { | |
8814ce8a | 245 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
246 | } |
247 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
248 | ||
6a83e74d | 249 | /** |
69e07c4a | 250 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
251 | * @q: request queue. |
252 | * | |
253 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
254 | * callback function is invoked. Once this function is returned, we make |
255 | * sure no dispatch can happen until the queue is unquiesced via | |
256 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
257 | */ |
258 | void blk_mq_quiesce_queue(struct request_queue *q) | |
259 | { | |
260 | struct blk_mq_hw_ctx *hctx; | |
261 | unsigned int i; | |
262 | bool rcu = false; | |
263 | ||
1d9e9bc6 | 264 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 265 | |
6a83e74d BVA |
266 | queue_for_each_hw_ctx(q, hctx, i) { |
267 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 268 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
269 | else |
270 | rcu = true; | |
271 | } | |
272 | if (rcu) | |
273 | synchronize_rcu(); | |
274 | } | |
275 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
276 | ||
e4e73913 ML |
277 | /* |
278 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
279 | * @q: request queue. | |
280 | * | |
281 | * This function recovers queue into the state before quiescing | |
282 | * which is done by blk_mq_quiesce_queue. | |
283 | */ | |
284 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
285 | { | |
8814ce8a | 286 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 287 | |
1d9e9bc6 ML |
288 | /* dispatch requests which are inserted during quiescing */ |
289 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
290 | } |
291 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
292 | ||
aed3ea94 JA |
293 | void blk_mq_wake_waiters(struct request_queue *q) |
294 | { | |
295 | struct blk_mq_hw_ctx *hctx; | |
296 | unsigned int i; | |
297 | ||
298 | queue_for_each_hw_ctx(q, hctx, i) | |
299 | if (blk_mq_hw_queue_mapped(hctx)) | |
300 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
301 | } | |
302 | ||
fe1f4526 | 303 | /* |
9a91b05b HT |
304 | * Only need start/end time stamping if we have iostat or |
305 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
306 | */ |
307 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
308 | { | |
2ff0682d | 309 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS | RQF_ELV)); |
fe1f4526 JA |
310 | } |
311 | ||
e4cdf1a1 | 312 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 313 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 314 | { |
605f784e PB |
315 | struct blk_mq_ctx *ctx = data->ctx; |
316 | struct blk_mq_hw_ctx *hctx = data->hctx; | |
317 | struct request_queue *q = data->q; | |
318 | struct elevator_queue *e = q->elevator; | |
e4cdf1a1 CH |
319 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
320 | struct request *rq = tags->static_rqs[tag]; | |
12845906 | 321 | unsigned int rq_flags = 0; |
c3a148d2 | 322 | |
605f784e | 323 | if (e) { |
12845906 | 324 | rq_flags = RQF_ELV; |
76647368 | 325 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
326 | rq->internal_tag = tag; |
327 | } else { | |
e4cdf1a1 | 328 | rq->tag = tag; |
76647368 | 329 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
330 | } |
331 | ||
12845906 PB |
332 | if (data->flags & BLK_MQ_REQ_PM) |
333 | rq_flags |= RQF_PM; | |
334 | if (blk_queue_io_stat(q)) | |
335 | rq_flags |= RQF_IO_STAT; | |
336 | rq->rq_flags = rq_flags; | |
337 | ||
4f266f2b PB |
338 | if (blk_mq_need_time_stamp(rq)) |
339 | rq->start_time_ns = ktime_get_ns(); | |
340 | else | |
341 | rq->start_time_ns = 0; | |
af76e555 | 342 | /* csd/requeue_work/fifo_time is initialized before use */ |
605f784e PB |
343 | rq->q = q; |
344 | rq->mq_ctx = ctx; | |
345 | rq->mq_hctx = hctx; | |
7ea4d8a4 | 346 | rq->cmd_flags = data->cmd_flags; |
af76e555 CH |
347 | rq->rq_disk = NULL; |
348 | rq->part = NULL; | |
6f816b4b TH |
349 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
350 | rq->alloc_time_ns = alloc_time_ns; | |
351 | #endif | |
544ccc8d | 352 | rq->io_start_time_ns = 0; |
3d244306 | 353 | rq->stats_sectors = 0; |
af76e555 CH |
354 | rq->nr_phys_segments = 0; |
355 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
356 | rq->nr_integrity_segments = 0; | |
357 | #endif | |
f6be4fb4 | 358 | rq->timeout = 0; |
af76e555 CH |
359 | rq->end_io = NULL; |
360 | rq->end_io_data = NULL; | |
af76e555 | 361 | |
4f266f2b PB |
362 | blk_crypto_rq_set_defaults(rq); |
363 | INIT_LIST_HEAD(&rq->queuelist); | |
364 | /* tag was already set */ | |
365 | WRITE_ONCE(rq->deadline, 0); | |
12f5b931 | 366 | refcount_set(&rq->ref, 1); |
7ea4d8a4 | 367 | |
4f266f2b | 368 | if (rq->rq_flags & RQF_ELV) { |
7ea4d8a4 CH |
369 | struct elevator_queue *e = data->q->elevator; |
370 | ||
371 | rq->elv.icq = NULL; | |
4f266f2b PB |
372 | INIT_HLIST_NODE(&rq->hash); |
373 | RB_CLEAR_NODE(&rq->rb_node); | |
374 | ||
375 | if (!op_is_flush(data->cmd_flags) && | |
376 | e->type->ops.prepare_request) { | |
7ea4d8a4 CH |
377 | if (e->type->icq_cache) |
378 | blk_mq_sched_assign_ioc(rq); | |
379 | ||
380 | e->type->ops.prepare_request(rq); | |
381 | rq->rq_flags |= RQF_ELVPRIV; | |
382 | } | |
383 | } | |
384 | ||
e4cdf1a1 | 385 | return rq; |
5dee8577 CH |
386 | } |
387 | ||
349302da JA |
388 | static inline struct request * |
389 | __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data, | |
390 | u64 alloc_time_ns) | |
391 | { | |
392 | unsigned int tag, tag_offset; | |
393 | struct request *rq; | |
394 | unsigned long tags; | |
395 | int i, nr = 0; | |
396 | ||
397 | tags = blk_mq_get_tags(data, data->nr_tags, &tag_offset); | |
398 | if (unlikely(!tags)) | |
399 | return NULL; | |
400 | ||
401 | for (i = 0; tags; i++) { | |
402 | if (!(tags & (1UL << i))) | |
403 | continue; | |
404 | tag = tag_offset + i; | |
405 | tags &= ~(1UL << i); | |
406 | rq = blk_mq_rq_ctx_init(data, tag, alloc_time_ns); | |
013a7f95 | 407 | rq_list_add(data->cached_rq, rq); |
349302da JA |
408 | } |
409 | data->nr_tags -= nr; | |
410 | ||
013a7f95 | 411 | return rq_list_pop(data->cached_rq); |
349302da JA |
412 | } |
413 | ||
b90cfaed | 414 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 415 | { |
e6e7abff | 416 | struct request_queue *q = data->q; |
d2c0d383 | 417 | struct elevator_queue *e = q->elevator; |
6f816b4b | 418 | u64 alloc_time_ns = 0; |
47c122e3 | 419 | struct request *rq; |
600c3b0c | 420 | unsigned int tag; |
d2c0d383 | 421 | |
6f816b4b TH |
422 | /* alloc_time includes depth and tag waits */ |
423 | if (blk_queue_rq_alloc_time(q)) | |
424 | alloc_time_ns = ktime_get_ns(); | |
425 | ||
f9afca4d | 426 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 427 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
428 | |
429 | if (e) { | |
d2c0d383 | 430 | /* |
8d663f34 | 431 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
432 | * dispatch list. Don't include reserved tags in the |
433 | * limiting, as it isn't useful. | |
d2c0d383 | 434 | */ |
f9afca4d | 435 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 436 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 437 | e->type->ops.limit_depth && |
17a51199 | 438 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 439 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
440 | } |
441 | ||
bf0beec0 | 442 | retry: |
600c3b0c CH |
443 | data->ctx = blk_mq_get_ctx(q); |
444 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 445 | if (!e) |
600c3b0c CH |
446 | blk_mq_tag_busy(data->hctx); |
447 | ||
349302da JA |
448 | /* |
449 | * Try batched alloc if we want more than 1 tag. | |
450 | */ | |
451 | if (data->nr_tags > 1) { | |
452 | rq = __blk_mq_alloc_requests_batch(data, alloc_time_ns); | |
453 | if (rq) | |
454 | return rq; | |
455 | data->nr_tags = 1; | |
456 | } | |
457 | ||
bf0beec0 ML |
458 | /* |
459 | * Waiting allocations only fail because of an inactive hctx. In that | |
460 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
461 | * should have migrated us to an online CPU by now. | |
462 | */ | |
349302da JA |
463 | tag = blk_mq_get_tag(data); |
464 | if (tag == BLK_MQ_NO_TAG) { | |
465 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
466 | return NULL; | |
467 | /* | |
468 | * Give up the CPU and sleep for a random short time to | |
469 | * ensure that thread using a realtime scheduling class | |
470 | * are migrated off the CPU, and thus off the hctx that | |
471 | * is going away. | |
472 | */ | |
473 | msleep(3); | |
474 | goto retry; | |
475 | } | |
47c122e3 | 476 | |
349302da | 477 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
478 | } |
479 | ||
cd6ce148 | 480 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 481 | blk_mq_req_flags_t flags) |
320ae51f | 482 | { |
e6e7abff CH |
483 | struct blk_mq_alloc_data data = { |
484 | .q = q, | |
485 | .flags = flags, | |
486 | .cmd_flags = op, | |
47c122e3 | 487 | .nr_tags = 1, |
e6e7abff | 488 | }; |
bd166ef1 | 489 | struct request *rq; |
a492f075 | 490 | int ret; |
320ae51f | 491 | |
3a0a5299 | 492 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
493 | if (ret) |
494 | return ERR_PTR(ret); | |
320ae51f | 495 | |
b90cfaed | 496 | rq = __blk_mq_alloc_requests(&data); |
bd166ef1 | 497 | if (!rq) |
a5ea5811 | 498 | goto out_queue_exit; |
0c4de0f3 CH |
499 | rq->__data_len = 0; |
500 | rq->__sector = (sector_t) -1; | |
501 | rq->bio = rq->biotail = NULL; | |
320ae51f | 502 | return rq; |
a5ea5811 CH |
503 | out_queue_exit: |
504 | blk_queue_exit(q); | |
505 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 506 | } |
4bb659b1 | 507 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 508 | |
cd6ce148 | 509 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 510 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 511 | { |
e6e7abff CH |
512 | struct blk_mq_alloc_data data = { |
513 | .q = q, | |
514 | .flags = flags, | |
515 | .cmd_flags = op, | |
47c122e3 | 516 | .nr_tags = 1, |
e6e7abff | 517 | }; |
600c3b0c | 518 | u64 alloc_time_ns = 0; |
6d2809d5 | 519 | unsigned int cpu; |
600c3b0c | 520 | unsigned int tag; |
1f5bd336 ML |
521 | int ret; |
522 | ||
600c3b0c CH |
523 | /* alloc_time includes depth and tag waits */ |
524 | if (blk_queue_rq_alloc_time(q)) | |
525 | alloc_time_ns = ktime_get_ns(); | |
526 | ||
1f5bd336 ML |
527 | /* |
528 | * If the tag allocator sleeps we could get an allocation for a | |
529 | * different hardware context. No need to complicate the low level | |
530 | * allocator for this for the rare use case of a command tied to | |
531 | * a specific queue. | |
532 | */ | |
600c3b0c | 533 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
534 | return ERR_PTR(-EINVAL); |
535 | ||
536 | if (hctx_idx >= q->nr_hw_queues) | |
537 | return ERR_PTR(-EIO); | |
538 | ||
3a0a5299 | 539 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
540 | if (ret) |
541 | return ERR_PTR(ret); | |
542 | ||
c8712c6a CH |
543 | /* |
544 | * Check if the hardware context is actually mapped to anything. | |
545 | * If not tell the caller that it should skip this queue. | |
546 | */ | |
a5ea5811 | 547 | ret = -EXDEV; |
e6e7abff CH |
548 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
549 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 550 | goto out_queue_exit; |
e6e7abff CH |
551 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
552 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 553 | |
42fdc5e4 | 554 | if (!q->elevator) |
600c3b0c CH |
555 | blk_mq_tag_busy(data.hctx); |
556 | ||
a5ea5811 | 557 | ret = -EWOULDBLOCK; |
600c3b0c CH |
558 | tag = blk_mq_get_tag(&data); |
559 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 560 | goto out_queue_exit; |
600c3b0c CH |
561 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
562 | ||
a5ea5811 CH |
563 | out_queue_exit: |
564 | blk_queue_exit(q); | |
565 | return ERR_PTR(ret); | |
1f5bd336 ML |
566 | } |
567 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
568 | ||
12f5b931 KB |
569 | static void __blk_mq_free_request(struct request *rq) |
570 | { | |
571 | struct request_queue *q = rq->q; | |
572 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 573 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
574 | const int sched_tag = rq->internal_tag; |
575 | ||
a892c8d5 | 576 | blk_crypto_free_request(rq); |
986d413b | 577 | blk_pm_mark_last_busy(rq); |
ea4f995e | 578 | rq->mq_hctx = NULL; |
76647368 | 579 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 580 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 581 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 582 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
583 | blk_mq_sched_restart(hctx); |
584 | blk_queue_exit(q); | |
585 | } | |
586 | ||
6af54051 | 587 | void blk_mq_free_request(struct request *rq) |
320ae51f | 588 | { |
320ae51f | 589 | struct request_queue *q = rq->q; |
ea4f995e | 590 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 591 | |
2ff0682d JA |
592 | if (rq->rq_flags & (RQF_ELVPRIV | RQF_ELV)) { |
593 | struct elevator_queue *e = q->elevator; | |
594 | ||
595 | if (e->type->ops.finish_request) | |
f9cd4bfe | 596 | e->type->ops.finish_request(rq); |
6af54051 CH |
597 | if (rq->elv.icq) { |
598 | put_io_context(rq->elv.icq->ioc); | |
599 | rq->elv.icq = NULL; | |
600 | } | |
601 | } | |
320ae51f | 602 | |
e8064021 | 603 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 604 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 605 | |
7beb2f84 | 606 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 607 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 608 | |
a7905043 | 609 | rq_qos_done(q, rq); |
0d2602ca | 610 | |
12f5b931 KB |
611 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
612 | if (refcount_dec_and_test(&rq->ref)) | |
613 | __blk_mq_free_request(rq); | |
320ae51f | 614 | } |
1a3b595a | 615 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 616 | |
47c122e3 JA |
617 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
618 | { | |
013a7f95 | 619 | struct request *rq; |
47c122e3 | 620 | |
013a7f95 | 621 | while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) { |
47c122e3 JA |
622 | percpu_ref_get(&rq->q->q_usage_counter); |
623 | blk_mq_free_request(rq); | |
624 | } | |
625 | } | |
626 | ||
9be3e06f JA |
627 | static void req_bio_endio(struct request *rq, struct bio *bio, |
628 | unsigned int nbytes, blk_status_t error) | |
629 | { | |
630 | if (error) | |
631 | bio->bi_status = error; | |
632 | ||
633 | if (unlikely(rq->rq_flags & RQF_QUIET)) | |
634 | bio_set_flag(bio, BIO_QUIET); | |
635 | ||
636 | bio_advance(bio, nbytes); | |
637 | ||
638 | if (req_op(rq) == REQ_OP_ZONE_APPEND && error == BLK_STS_OK) { | |
639 | /* | |
640 | * Partial zone append completions cannot be supported as the | |
641 | * BIO fragments may end up not being written sequentially. | |
642 | */ | |
643 | if (bio->bi_iter.bi_size) | |
644 | bio->bi_status = BLK_STS_IOERR; | |
645 | else | |
646 | bio->bi_iter.bi_sector = rq->__sector; | |
647 | } | |
648 | ||
649 | /* don't actually finish bio if it's part of flush sequence */ | |
650 | if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) | |
651 | bio_endio(bio); | |
652 | } | |
653 | ||
654 | static void blk_account_io_completion(struct request *req, unsigned int bytes) | |
655 | { | |
656 | if (req->part && blk_do_io_stat(req)) { | |
657 | const int sgrp = op_stat_group(req_op(req)); | |
658 | ||
659 | part_stat_lock(); | |
660 | part_stat_add(req->part, sectors[sgrp], bytes >> 9); | |
661 | part_stat_unlock(); | |
662 | } | |
663 | } | |
664 | ||
665 | /** | |
666 | * blk_update_request - Complete multiple bytes without completing the request | |
667 | * @req: the request being processed | |
668 | * @error: block status code | |
669 | * @nr_bytes: number of bytes to complete for @req | |
670 | * | |
671 | * Description: | |
672 | * Ends I/O on a number of bytes attached to @req, but doesn't complete | |
673 | * the request structure even if @req doesn't have leftover. | |
674 | * If @req has leftover, sets it up for the next range of segments. | |
675 | * | |
676 | * Passing the result of blk_rq_bytes() as @nr_bytes guarantees | |
677 | * %false return from this function. | |
678 | * | |
679 | * Note: | |
680 | * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function | |
681 | * except in the consistency check at the end of this function. | |
682 | * | |
683 | * Return: | |
684 | * %false - this request doesn't have any more data | |
685 | * %true - this request has more data | |
686 | **/ | |
687 | bool blk_update_request(struct request *req, blk_status_t error, | |
688 | unsigned int nr_bytes) | |
689 | { | |
690 | int total_bytes; | |
691 | ||
692 | trace_block_rq_complete(req, blk_status_to_errno(error), nr_bytes); | |
693 | ||
694 | if (!req->bio) | |
695 | return false; | |
696 | ||
697 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
698 | if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && | |
699 | error == BLK_STS_OK) | |
700 | req->q->integrity.profile->complete_fn(req, nr_bytes); | |
701 | #endif | |
702 | ||
703 | if (unlikely(error && !blk_rq_is_passthrough(req) && | |
704 | !(req->rq_flags & RQF_QUIET))) | |
705 | blk_print_req_error(req, error); | |
706 | ||
707 | blk_account_io_completion(req, nr_bytes); | |
708 | ||
709 | total_bytes = 0; | |
710 | while (req->bio) { | |
711 | struct bio *bio = req->bio; | |
712 | unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); | |
713 | ||
714 | if (bio_bytes == bio->bi_iter.bi_size) | |
715 | req->bio = bio->bi_next; | |
716 | ||
717 | /* Completion has already been traced */ | |
718 | bio_clear_flag(bio, BIO_TRACE_COMPLETION); | |
719 | req_bio_endio(req, bio, bio_bytes, error); | |
720 | ||
721 | total_bytes += bio_bytes; | |
722 | nr_bytes -= bio_bytes; | |
723 | ||
724 | if (!nr_bytes) | |
725 | break; | |
726 | } | |
727 | ||
728 | /* | |
729 | * completely done | |
730 | */ | |
731 | if (!req->bio) { | |
732 | /* | |
733 | * Reset counters so that the request stacking driver | |
734 | * can find how many bytes remain in the request | |
735 | * later. | |
736 | */ | |
737 | req->__data_len = 0; | |
738 | return false; | |
739 | } | |
740 | ||
741 | req->__data_len -= total_bytes; | |
742 | ||
743 | /* update sector only for requests with clear definition of sector */ | |
744 | if (!blk_rq_is_passthrough(req)) | |
745 | req->__sector += total_bytes >> 9; | |
746 | ||
747 | /* mixed attributes always follow the first bio */ | |
748 | if (req->rq_flags & RQF_MIXED_MERGE) { | |
749 | req->cmd_flags &= ~REQ_FAILFAST_MASK; | |
750 | req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; | |
751 | } | |
752 | ||
753 | if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { | |
754 | /* | |
755 | * If total number of sectors is less than the first segment | |
756 | * size, something has gone terribly wrong. | |
757 | */ | |
758 | if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { | |
759 | blk_dump_rq_flags(req, "request botched"); | |
760 | req->__data_len = blk_rq_cur_bytes(req); | |
761 | } | |
762 | ||
763 | /* recalculate the number of segments */ | |
764 | req->nr_phys_segments = blk_recalc_rq_segments(req); | |
765 | } | |
766 | ||
767 | return true; | |
768 | } | |
769 | EXPORT_SYMBOL_GPL(blk_update_request); | |
770 | ||
2a842aca | 771 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 772 | { |
8971a3b7 PB |
773 | if (blk_mq_need_time_stamp(rq)) { |
774 | u64 now = ktime_get_ns(); | |
fe1f4526 | 775 | |
8971a3b7 PB |
776 | if (rq->rq_flags & RQF_STATS) { |
777 | blk_mq_poll_stats_start(rq->q); | |
778 | blk_stat_add(rq, now); | |
779 | } | |
522a7775 | 780 | |
8971a3b7 PB |
781 | blk_mq_sched_completed_request(rq, now); |
782 | blk_account_io_done(rq, now); | |
4bc6339a OS |
783 | } |
784 | ||
91b63639 | 785 | if (rq->end_io) { |
a7905043 | 786 | rq_qos_done(rq->q, rq); |
320ae51f | 787 | rq->end_io(rq, error); |
91b63639 | 788 | } else { |
320ae51f | 789 | blk_mq_free_request(rq); |
91b63639 | 790 | } |
320ae51f | 791 | } |
c8a446ad | 792 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 793 | |
2a842aca | 794 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
795 | { |
796 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
797 | BUG(); | |
c8a446ad | 798 | __blk_mq_end_request(rq, error); |
63151a44 | 799 | } |
c8a446ad | 800 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 801 | |
f9ab4918 | 802 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 803 | { |
f9ab4918 SAS |
804 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
805 | struct request *rq, *next; | |
c3077b5d | 806 | |
f9ab4918 | 807 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 808 | rq->q->mq_ops->complete(rq); |
320ae51f | 809 | } |
320ae51f | 810 | |
f9ab4918 | 811 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 812 | { |
f9ab4918 | 813 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
814 | } |
815 | ||
c3077b5d CH |
816 | static int blk_softirq_cpu_dead(unsigned int cpu) |
817 | { | |
f9ab4918 | 818 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
819 | return 0; |
820 | } | |
821 | ||
40d09b53 | 822 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 823 | { |
f9ab4918 | 824 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
825 | } |
826 | ||
96339526 CH |
827 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
828 | { | |
829 | int cpu = raw_smp_processor_id(); | |
830 | ||
831 | if (!IS_ENABLED(CONFIG_SMP) || | |
832 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
833 | return false; | |
71425189 SAS |
834 | /* |
835 | * With force threaded interrupts enabled, raising softirq from an SMP | |
836 | * function call will always result in waking the ksoftirqd thread. | |
837 | * This is probably worse than completing the request on a different | |
838 | * cache domain. | |
839 | */ | |
91cc470e | 840 | if (force_irqthreads()) |
71425189 | 841 | return false; |
96339526 CH |
842 | |
843 | /* same CPU or cache domain? Complete locally */ | |
844 | if (cpu == rq->mq_ctx->cpu || | |
845 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
846 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
847 | return false; | |
848 | ||
849 | /* don't try to IPI to an offline CPU */ | |
850 | return cpu_online(rq->mq_ctx->cpu); | |
851 | } | |
852 | ||
f9ab4918 SAS |
853 | static void blk_mq_complete_send_ipi(struct request *rq) |
854 | { | |
855 | struct llist_head *list; | |
856 | unsigned int cpu; | |
857 | ||
858 | cpu = rq->mq_ctx->cpu; | |
859 | list = &per_cpu(blk_cpu_done, cpu); | |
860 | if (llist_add(&rq->ipi_list, list)) { | |
861 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
862 | smp_call_function_single_async(cpu, &rq->csd); | |
863 | } | |
864 | } | |
865 | ||
866 | static void blk_mq_raise_softirq(struct request *rq) | |
867 | { | |
868 | struct llist_head *list; | |
869 | ||
870 | preempt_disable(); | |
871 | list = this_cpu_ptr(&blk_cpu_done); | |
872 | if (llist_add(&rq->ipi_list, list)) | |
873 | raise_softirq(BLOCK_SOFTIRQ); | |
874 | preempt_enable(); | |
875 | } | |
876 | ||
40d09b53 | 877 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 878 | { |
af78ff7c | 879 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 880 | |
4ab32bf3 JA |
881 | /* |
882 | * For a polled request, always complete locallly, it's pointless | |
883 | * to redirect the completion. | |
884 | */ | |
6ce913fe | 885 | if (rq->cmd_flags & REQ_POLLED) |
40d09b53 | 886 | return false; |
38535201 | 887 | |
96339526 | 888 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
889 | blk_mq_complete_send_ipi(rq); |
890 | return true; | |
3d6efbf6 | 891 | } |
40d09b53 | 892 | |
f9ab4918 SAS |
893 | if (rq->q->nr_hw_queues == 1) { |
894 | blk_mq_raise_softirq(rq); | |
895 | return true; | |
896 | } | |
897 | return false; | |
40d09b53 CH |
898 | } |
899 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
900 | ||
901 | /** | |
902 | * blk_mq_complete_request - end I/O on a request | |
903 | * @rq: the request being processed | |
904 | * | |
905 | * Description: | |
906 | * Complete a request by scheduling the ->complete_rq operation. | |
907 | **/ | |
908 | void blk_mq_complete_request(struct request *rq) | |
909 | { | |
910 | if (!blk_mq_complete_request_remote(rq)) | |
911 | rq->q->mq_ops->complete(rq); | |
320ae51f | 912 | } |
15f73f5b | 913 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 914 | |
04ced159 | 915 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 916 | __releases(hctx->srcu) |
04ced159 JA |
917 | { |
918 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
919 | rcu_read_unlock(); | |
920 | else | |
05707b64 | 921 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
922 | } |
923 | ||
924 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 925 | __acquires(hctx->srcu) |
04ced159 | 926 | { |
08b5a6e2 JA |
927 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
928 | /* shut up gcc false positive */ | |
929 | *srcu_idx = 0; | |
04ced159 | 930 | rcu_read_lock(); |
08b5a6e2 | 931 | } else |
05707b64 | 932 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
933 | } |
934 | ||
105663f7 AA |
935 | /** |
936 | * blk_mq_start_request - Start processing a request | |
937 | * @rq: Pointer to request to be started | |
938 | * | |
939 | * Function used by device drivers to notify the block layer that a request | |
940 | * is going to be processed now, so blk layer can do proper initializations | |
941 | * such as starting the timeout timer. | |
942 | */ | |
e2490073 | 943 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
944 | { |
945 | struct request_queue *q = rq->q; | |
946 | ||
a54895fa | 947 | trace_block_rq_issue(rq); |
320ae51f | 948 | |
cf43e6be | 949 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
00067077 JA |
950 | u64 start_time; |
951 | #ifdef CONFIG_BLK_CGROUP | |
952 | if (rq->bio) | |
953 | start_time = bio_issue_time(&rq->bio->bi_issue); | |
954 | else | |
955 | #endif | |
956 | start_time = ktime_get_ns(); | |
957 | rq->io_start_time_ns = start_time; | |
3d244306 | 958 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 959 | rq->rq_flags |= RQF_STATS; |
a7905043 | 960 | rq_qos_issue(q, rq); |
cf43e6be JA |
961 | } |
962 | ||
1d9bd516 | 963 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 964 | |
1d9bd516 | 965 | blk_add_timer(rq); |
12f5b931 | 966 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 967 | |
54d4e6ab MG |
968 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
969 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
970 | q->integrity.profile->prepare_fn(rq); | |
971 | #endif | |
3e08773c CH |
972 | if (rq->bio && rq->bio->bi_opf & REQ_POLLED) |
973 | WRITE_ONCE(rq->bio->bi_cookie, blk_rq_to_qc(rq)); | |
320ae51f | 974 | } |
e2490073 | 975 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 976 | |
ed0791b2 | 977 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
978 | { |
979 | struct request_queue *q = rq->q; | |
980 | ||
923218f6 ML |
981 | blk_mq_put_driver_tag(rq); |
982 | ||
a54895fa | 983 | trace_block_rq_requeue(rq); |
a7905043 | 984 | rq_qos_requeue(q, rq); |
49f5baa5 | 985 | |
12f5b931 KB |
986 | if (blk_mq_request_started(rq)) { |
987 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 988 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 989 | } |
320ae51f JA |
990 | } |
991 | ||
2b053aca | 992 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 993 | { |
ed0791b2 | 994 | __blk_mq_requeue_request(rq); |
ed0791b2 | 995 | |
105976f5 ML |
996 | /* this request will be re-inserted to io scheduler queue */ |
997 | blk_mq_sched_requeue_request(rq); | |
998 | ||
7d692330 | 999 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 1000 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
1001 | } |
1002 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
1003 | ||
6fca6a61 CH |
1004 | static void blk_mq_requeue_work(struct work_struct *work) |
1005 | { | |
1006 | struct request_queue *q = | |
2849450a | 1007 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
1008 | LIST_HEAD(rq_list); |
1009 | struct request *rq, *next; | |
6fca6a61 | 1010 | |
18e9781d | 1011 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 1012 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 1013 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
1014 | |
1015 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 1016 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
1017 | continue; |
1018 | ||
e8064021 | 1019 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 1020 | list_del_init(&rq->queuelist); |
aef1897c JW |
1021 | /* |
1022 | * If RQF_DONTPREP, rq has contained some driver specific | |
1023 | * data, so insert it to hctx dispatch list to avoid any | |
1024 | * merge. | |
1025 | */ | |
1026 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 1027 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
1028 | else |
1029 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
1030 | } |
1031 | ||
1032 | while (!list_empty(&rq_list)) { | |
1033 | rq = list_entry(rq_list.next, struct request, queuelist); | |
1034 | list_del_init(&rq->queuelist); | |
9e97d295 | 1035 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
1036 | } |
1037 | ||
52d7f1b5 | 1038 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
1039 | } |
1040 | ||
2b053aca BVA |
1041 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
1042 | bool kick_requeue_list) | |
6fca6a61 CH |
1043 | { |
1044 | struct request_queue *q = rq->q; | |
1045 | unsigned long flags; | |
1046 | ||
1047 | /* | |
1048 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 1049 | * request head insertion from the workqueue. |
6fca6a61 | 1050 | */ |
e8064021 | 1051 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
1052 | |
1053 | spin_lock_irqsave(&q->requeue_lock, flags); | |
1054 | if (at_head) { | |
e8064021 | 1055 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
1056 | list_add(&rq->queuelist, &q->requeue_list); |
1057 | } else { | |
1058 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
1059 | } | |
1060 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
1061 | |
1062 | if (kick_requeue_list) | |
1063 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 1064 | } |
6fca6a61 CH |
1065 | |
1066 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
1067 | { | |
ae943d20 | 1068 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
1069 | } |
1070 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
1071 | ||
2849450a MS |
1072 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
1073 | unsigned long msecs) | |
1074 | { | |
d4acf365 BVA |
1075 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
1076 | msecs_to_jiffies(msecs)); | |
2849450a MS |
1077 | } |
1078 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
1079 | ||
0e62f51f JA |
1080 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
1081 | { | |
88c7b2b7 JA |
1082 | if (tag < tags->nr_tags) { |
1083 | prefetch(tags->rqs[tag]); | |
4ee86bab | 1084 | return tags->rqs[tag]; |
88c7b2b7 | 1085 | } |
4ee86bab HR |
1086 | |
1087 | return NULL; | |
24d2f903 CH |
1088 | } |
1089 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
1090 | ||
3c94d83c JA |
1091 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1092 | void *priv, bool reserved) | |
ae879912 JA |
1093 | { |
1094 | /* | |
05a4fed6 | 1095 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 1096 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 1097 | */ |
05a4fed6 | 1098 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
1099 | bool *busy = priv; |
1100 | ||
1101 | *busy = true; | |
1102 | return false; | |
1103 | } | |
1104 | ||
1105 | return true; | |
1106 | } | |
1107 | ||
3c94d83c | 1108 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
1109 | { |
1110 | bool busy = false; | |
1111 | ||
3c94d83c | 1112 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
1113 | return busy; |
1114 | } | |
3c94d83c | 1115 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 1116 | |
358f70da | 1117 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 1118 | { |
da661267 | 1119 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
1120 | if (req->q->mq_ops->timeout) { |
1121 | enum blk_eh_timer_return ret; | |
1122 | ||
1123 | ret = req->q->mq_ops->timeout(req, reserved); | |
1124 | if (ret == BLK_EH_DONE) | |
1125 | return; | |
1126 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 1127 | } |
d1210d5a CH |
1128 | |
1129 | blk_add_timer(req); | |
87ee7b11 | 1130 | } |
5b3f25fc | 1131 | |
12f5b931 | 1132 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 1133 | { |
12f5b931 | 1134 | unsigned long deadline; |
87ee7b11 | 1135 | |
12f5b931 KB |
1136 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
1137 | return false; | |
da661267 CH |
1138 | if (rq->rq_flags & RQF_TIMED_OUT) |
1139 | return false; | |
a7af0af3 | 1140 | |
079076b3 | 1141 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
1142 | if (time_after_eq(jiffies, deadline)) |
1143 | return true; | |
a7af0af3 | 1144 | |
12f5b931 KB |
1145 | if (*next == 0) |
1146 | *next = deadline; | |
1147 | else if (time_after(*next, deadline)) | |
1148 | *next = deadline; | |
1149 | return false; | |
87ee7b11 JA |
1150 | } |
1151 | ||
2e315dc0 ML |
1152 | void blk_mq_put_rq_ref(struct request *rq) |
1153 | { | |
a9ed27a7 | 1154 | if (is_flush_rq(rq)) |
2e315dc0 ML |
1155 | rq->end_io(rq, 0); |
1156 | else if (refcount_dec_and_test(&rq->ref)) | |
1157 | __blk_mq_free_request(rq); | |
1158 | } | |
1159 | ||
7baa8572 | 1160 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
1161 | struct request *rq, void *priv, bool reserved) |
1162 | { | |
12f5b931 KB |
1163 | unsigned long *next = priv; |
1164 | ||
1165 | /* | |
c797b40c ML |
1166 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
1167 | * be reallocated underneath the timeout handler's processing, then | |
1168 | * the expire check is reliable. If the request is not expired, then | |
1169 | * it was completed and reallocated as a new request after returning | |
1170 | * from blk_mq_check_expired(). | |
1d9bd516 | 1171 | */ |
12f5b931 | 1172 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 1173 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 1174 | return true; |
1d9bd516 TH |
1175 | } |
1176 | ||
287922eb | 1177 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 1178 | { |
287922eb CH |
1179 | struct request_queue *q = |
1180 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 1181 | unsigned long next = 0; |
1d9bd516 | 1182 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 1183 | int i; |
320ae51f | 1184 | |
71f79fb3 GKB |
1185 | /* A deadlock might occur if a request is stuck requiring a |
1186 | * timeout at the same time a queue freeze is waiting | |
1187 | * completion, since the timeout code would not be able to | |
1188 | * acquire the queue reference here. | |
1189 | * | |
1190 | * That's why we don't use blk_queue_enter here; instead, we use | |
1191 | * percpu_ref_tryget directly, because we need to be able to | |
1192 | * obtain a reference even in the short window between the queue | |
1193 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1194 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1195 | * consumed, marked by the instant q_usage_counter reaches |
1196 | * zero. | |
1197 | */ | |
1198 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1199 | return; |
1200 | ||
12f5b931 | 1201 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1202 | |
12f5b931 KB |
1203 | if (next != 0) { |
1204 | mod_timer(&q->timeout, next); | |
0d2602ca | 1205 | } else { |
fcd36c36 BVA |
1206 | /* |
1207 | * Request timeouts are handled as a forward rolling timer. If | |
1208 | * we end up here it means that no requests are pending and | |
1209 | * also that no request has been pending for a while. Mark | |
1210 | * each hctx as idle. | |
1211 | */ | |
f054b56c ML |
1212 | queue_for_each_hw_ctx(q, hctx, i) { |
1213 | /* the hctx may be unmapped, so check it here */ | |
1214 | if (blk_mq_hw_queue_mapped(hctx)) | |
1215 | blk_mq_tag_idle(hctx); | |
1216 | } | |
0d2602ca | 1217 | } |
287922eb | 1218 | blk_queue_exit(q); |
320ae51f JA |
1219 | } |
1220 | ||
88459642 OS |
1221 | struct flush_busy_ctx_data { |
1222 | struct blk_mq_hw_ctx *hctx; | |
1223 | struct list_head *list; | |
1224 | }; | |
1225 | ||
1226 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1227 | { | |
1228 | struct flush_busy_ctx_data *flush_data = data; | |
1229 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1230 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1231 | enum hctx_type type = hctx->type; |
88459642 | 1232 | |
88459642 | 1233 | spin_lock(&ctx->lock); |
c16d6b5a | 1234 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1235 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1236 | spin_unlock(&ctx->lock); |
1237 | return true; | |
1238 | } | |
1239 | ||
1429d7c9 JA |
1240 | /* |
1241 | * Process software queues that have been marked busy, splicing them | |
1242 | * to the for-dispatch | |
1243 | */ | |
2c3ad667 | 1244 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1245 | { |
88459642 OS |
1246 | struct flush_busy_ctx_data data = { |
1247 | .hctx = hctx, | |
1248 | .list = list, | |
1249 | }; | |
1429d7c9 | 1250 | |
88459642 | 1251 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1252 | } |
2c3ad667 | 1253 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1254 | |
b347689f ML |
1255 | struct dispatch_rq_data { |
1256 | struct blk_mq_hw_ctx *hctx; | |
1257 | struct request *rq; | |
1258 | }; | |
1259 | ||
1260 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1261 | void *data) | |
1262 | { | |
1263 | struct dispatch_rq_data *dispatch_data = data; | |
1264 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1265 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1266 | enum hctx_type type = hctx->type; |
b347689f ML |
1267 | |
1268 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1269 | if (!list_empty(&ctx->rq_lists[type])) { |
1270 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1271 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1272 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1273 | sbitmap_clear_bit(sb, bitnr); |
1274 | } | |
1275 | spin_unlock(&ctx->lock); | |
1276 | ||
1277 | return !dispatch_data->rq; | |
1278 | } | |
1279 | ||
1280 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1281 | struct blk_mq_ctx *start) | |
1282 | { | |
f31967f0 | 1283 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1284 | struct dispatch_rq_data data = { |
1285 | .hctx = hctx, | |
1286 | .rq = NULL, | |
1287 | }; | |
1288 | ||
1289 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1290 | dispatch_rq_from_ctx, &data); | |
1291 | ||
1292 | return data.rq; | |
1293 | } | |
1294 | ||
570e9b73 ML |
1295 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1296 | { | |
ae0f1a73 | 1297 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1298 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1299 | int tag; |
1300 | ||
568f2700 ML |
1301 | blk_mq_tag_busy(rq->mq_hctx); |
1302 | ||
570e9b73 | 1303 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1304 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1305 | tag_offset = 0; |
28500850 ML |
1306 | } else { |
1307 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1308 | return false; | |
570e9b73 ML |
1309 | } |
1310 | ||
570e9b73 ML |
1311 | tag = __sbitmap_queue_get(bt); |
1312 | if (tag == BLK_MQ_NO_TAG) | |
1313 | return false; | |
1314 | ||
1315 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1316 | return true; |
1317 | } | |
1318 | ||
61347154 | 1319 | bool blk_mq_get_driver_tag(struct request *rq) |
570e9b73 | 1320 | { |
568f2700 ML |
1321 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1322 | ||
1323 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1324 | return false; | |
1325 | ||
51db1c37 | 1326 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1327 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1328 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1329 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1330 | } |
1331 | hctx->tags->rqs[rq->tag] = rq; | |
1332 | return true; | |
570e9b73 ML |
1333 | } |
1334 | ||
eb619fdb JA |
1335 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1336 | int flags, void *key) | |
da55f2cc OS |
1337 | { |
1338 | struct blk_mq_hw_ctx *hctx; | |
1339 | ||
1340 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1341 | ||
5815839b | 1342 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1343 | if (!list_empty(&wait->entry)) { |
1344 | struct sbitmap_queue *sbq; | |
1345 | ||
1346 | list_del_init(&wait->entry); | |
ae0f1a73 | 1347 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1348 | atomic_dec(&sbq->ws_active); |
1349 | } | |
5815839b ML |
1350 | spin_unlock(&hctx->dispatch_wait_lock); |
1351 | ||
da55f2cc OS |
1352 | blk_mq_run_hw_queue(hctx, true); |
1353 | return 1; | |
1354 | } | |
1355 | ||
f906a6a0 JA |
1356 | /* |
1357 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1358 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1359 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1360 | * marking us as waiting. |
1361 | */ | |
2278d69f | 1362 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1363 | struct request *rq) |
da55f2cc | 1364 | { |
ae0f1a73 | 1365 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1366 | struct wait_queue_head *wq; |
f906a6a0 JA |
1367 | wait_queue_entry_t *wait; |
1368 | bool ret; | |
da55f2cc | 1369 | |
51db1c37 | 1370 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1371 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1372 | |
c27d53fb BVA |
1373 | /* |
1374 | * It's possible that a tag was freed in the window between the | |
1375 | * allocation failure and adding the hardware queue to the wait | |
1376 | * queue. | |
1377 | * | |
1378 | * Don't clear RESTART here, someone else could have set it. | |
1379 | * At most this will cost an extra queue run. | |
1380 | */ | |
8ab6bb9e | 1381 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1382 | } |
eb619fdb | 1383 | |
2278d69f | 1384 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1385 | if (!list_empty_careful(&wait->entry)) |
1386 | return false; | |
1387 | ||
e8618575 | 1388 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1389 | |
1390 | spin_lock_irq(&wq->lock); | |
1391 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1392 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1393 | spin_unlock(&hctx->dispatch_wait_lock); |
1394 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1395 | return false; |
eb619fdb JA |
1396 | } |
1397 | ||
e8618575 | 1398 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1399 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1400 | __add_wait_queue(wq, wait); | |
c27d53fb | 1401 | |
da55f2cc | 1402 | /* |
eb619fdb JA |
1403 | * It's possible that a tag was freed in the window between the |
1404 | * allocation failure and adding the hardware queue to the wait | |
1405 | * queue. | |
da55f2cc | 1406 | */ |
8ab6bb9e | 1407 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1408 | if (!ret) { |
5815839b ML |
1409 | spin_unlock(&hctx->dispatch_wait_lock); |
1410 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1411 | return false; |
eb619fdb | 1412 | } |
c27d53fb BVA |
1413 | |
1414 | /* | |
1415 | * We got a tag, remove ourselves from the wait queue to ensure | |
1416 | * someone else gets the wakeup. | |
1417 | */ | |
c27d53fb | 1418 | list_del_init(&wait->entry); |
e8618575 | 1419 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1420 | spin_unlock(&hctx->dispatch_wait_lock); |
1421 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1422 | |
1423 | return true; | |
da55f2cc OS |
1424 | } |
1425 | ||
6e768717 ML |
1426 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1427 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1428 | /* | |
1429 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1430 | * - EWMA is one simple way to compute running average value | |
1431 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1432 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1433 | * factor doesn't matter because EWMA decreases exponentially | |
1434 | */ | |
1435 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1436 | { | |
1437 | unsigned int ewma; | |
1438 | ||
6e768717 ML |
1439 | ewma = hctx->dispatch_busy; |
1440 | ||
1441 | if (!ewma && !busy) | |
1442 | return; | |
1443 | ||
1444 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1445 | if (busy) | |
1446 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1447 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1448 | ||
1449 | hctx->dispatch_busy = ewma; | |
1450 | } | |
1451 | ||
86ff7c2a ML |
1452 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1453 | ||
c92a4103 JT |
1454 | static void blk_mq_handle_dev_resource(struct request *rq, |
1455 | struct list_head *list) | |
1456 | { | |
1457 | struct request *next = | |
1458 | list_first_entry_or_null(list, struct request, queuelist); | |
1459 | ||
1460 | /* | |
1461 | * If an I/O scheduler has been configured and we got a driver tag for | |
1462 | * the next request already, free it. | |
1463 | */ | |
1464 | if (next) | |
1465 | blk_mq_put_driver_tag(next); | |
1466 | ||
1467 | list_add(&rq->queuelist, list); | |
1468 | __blk_mq_requeue_request(rq); | |
1469 | } | |
1470 | ||
0512a75b KB |
1471 | static void blk_mq_handle_zone_resource(struct request *rq, |
1472 | struct list_head *zone_list) | |
1473 | { | |
1474 | /* | |
1475 | * If we end up here it is because we cannot dispatch a request to a | |
1476 | * specific zone due to LLD level zone-write locking or other zone | |
1477 | * related resource not being available. In this case, set the request | |
1478 | * aside in zone_list for retrying it later. | |
1479 | */ | |
1480 | list_add(&rq->queuelist, zone_list); | |
1481 | __blk_mq_requeue_request(rq); | |
1482 | } | |
1483 | ||
75383524 ML |
1484 | enum prep_dispatch { |
1485 | PREP_DISPATCH_OK, | |
1486 | PREP_DISPATCH_NO_TAG, | |
1487 | PREP_DISPATCH_NO_BUDGET, | |
1488 | }; | |
1489 | ||
1490 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1491 | bool need_budget) | |
1492 | { | |
1493 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1494 | int budget_token = -1; |
75383524 | 1495 | |
2a5a24aa ML |
1496 | if (need_budget) { |
1497 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1498 | if (budget_token < 0) { | |
1499 | blk_mq_put_driver_tag(rq); | |
1500 | return PREP_DISPATCH_NO_BUDGET; | |
1501 | } | |
1502 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1503 | } |
1504 | ||
1505 | if (!blk_mq_get_driver_tag(rq)) { | |
1506 | /* | |
1507 | * The initial allocation attempt failed, so we need to | |
1508 | * rerun the hardware queue when a tag is freed. The | |
1509 | * waitqueue takes care of that. If the queue is run | |
1510 | * before we add this entry back on the dispatch list, | |
1511 | * we'll re-run it below. | |
1512 | */ | |
1513 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1514 | /* |
1515 | * All budgets not got from this function will be put | |
1516 | * together during handling partial dispatch | |
1517 | */ | |
1518 | if (need_budget) | |
2a5a24aa | 1519 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1520 | return PREP_DISPATCH_NO_TAG; |
1521 | } | |
1522 | } | |
1523 | ||
1524 | return PREP_DISPATCH_OK; | |
1525 | } | |
1526 | ||
1fd40b5e ML |
1527 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1528 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1529 | struct list_head *list) |
1fd40b5e | 1530 | { |
2a5a24aa | 1531 | struct request *rq; |
1fd40b5e | 1532 | |
2a5a24aa ML |
1533 | list_for_each_entry(rq, list, queuelist) { |
1534 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1535 | |
2a5a24aa ML |
1536 | if (budget_token >= 0) |
1537 | blk_mq_put_dispatch_budget(q, budget_token); | |
1538 | } | |
1fd40b5e ML |
1539 | } |
1540 | ||
1f57f8d4 JA |
1541 | /* |
1542 | * Returns true if we did some work AND can potentially do more. | |
1543 | */ | |
445874e8 | 1544 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1545 | unsigned int nr_budgets) |
320ae51f | 1546 | { |
75383524 | 1547 | enum prep_dispatch prep; |
445874e8 | 1548 | struct request_queue *q = hctx->queue; |
6d6f167c | 1549 | struct request *rq, *nxt; |
fc17b653 | 1550 | int errors, queued; |
86ff7c2a | 1551 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1552 | LIST_HEAD(zone_list); |
320ae51f | 1553 | |
81380ca1 OS |
1554 | if (list_empty(list)) |
1555 | return false; | |
1556 | ||
320ae51f JA |
1557 | /* |
1558 | * Now process all the entries, sending them to the driver. | |
1559 | */ | |
93efe981 | 1560 | errors = queued = 0; |
81380ca1 | 1561 | do { |
74c45052 | 1562 | struct blk_mq_queue_data bd; |
320ae51f | 1563 | |
f04c3df3 | 1564 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1565 | |
445874e8 | 1566 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1567 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1568 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1569 | break; |
de148297 | 1570 | |
320ae51f | 1571 | list_del_init(&rq->queuelist); |
320ae51f | 1572 | |
74c45052 | 1573 | bd.rq = rq; |
113285b4 JA |
1574 | |
1575 | /* | |
1576 | * Flag last if we have no more requests, or if we have more | |
1577 | * but can't assign a driver tag to it. | |
1578 | */ | |
1579 | if (list_empty(list)) | |
1580 | bd.last = true; | |
1581 | else { | |
113285b4 | 1582 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1583 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1584 | } |
74c45052 | 1585 | |
1fd40b5e ML |
1586 | /* |
1587 | * once the request is queued to lld, no need to cover the | |
1588 | * budget any more | |
1589 | */ | |
1590 | if (nr_budgets) | |
1591 | nr_budgets--; | |
74c45052 | 1592 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1593 | switch (ret) { |
1594 | case BLK_STS_OK: | |
1595 | queued++; | |
320ae51f | 1596 | break; |
7bf13729 ML |
1597 | case BLK_STS_RESOURCE: |
1598 | case BLK_STS_DEV_RESOURCE: | |
1599 | blk_mq_handle_dev_resource(rq, list); | |
1600 | goto out; | |
1601 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1602 | /* |
1603 | * Move the request to zone_list and keep going through | |
1604 | * the dispatch list to find more requests the drive can | |
1605 | * accept. | |
1606 | */ | |
1607 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1608 | break; |
1609 | default: | |
93efe981 | 1610 | errors++; |
e21ee5a6 | 1611 | blk_mq_end_request(rq, ret); |
320ae51f | 1612 | } |
81380ca1 | 1613 | } while (!list_empty(list)); |
7bf13729 | 1614 | out: |
0512a75b KB |
1615 | if (!list_empty(&zone_list)) |
1616 | list_splice_tail_init(&zone_list, list); | |
1617 | ||
632bfb63 | 1618 | /* If we didn't flush the entire list, we could have told the driver |
1619 | * there was more coming, but that turned out to be a lie. | |
1620 | */ | |
1621 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1622 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1623 | /* |
1624 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1625 | * that is where we will continue on next queue run. | |
1626 | */ | |
f04c3df3 | 1627 | if (!list_empty(list)) { |
86ff7c2a | 1628 | bool needs_restart; |
75383524 ML |
1629 | /* For non-shared tags, the RESTART check will suffice */ |
1630 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1631 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1632 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1633 | |
2a5a24aa ML |
1634 | if (nr_budgets) |
1635 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1636 | |
320ae51f | 1637 | spin_lock(&hctx->lock); |
01e99aec | 1638 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1639 | spin_unlock(&hctx->lock); |
f04c3df3 | 1640 | |
d7d8535f ML |
1641 | /* |
1642 | * Order adding requests to hctx->dispatch and checking | |
1643 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1644 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1645 | * miss the new added requests to hctx->dispatch, meantime | |
1646 | * SCHED_RESTART is observed here. | |
1647 | */ | |
1648 | smp_mb(); | |
1649 | ||
9ba52e58 | 1650 | /* |
710c785f BVA |
1651 | * If SCHED_RESTART was set by the caller of this function and |
1652 | * it is no longer set that means that it was cleared by another | |
1653 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1654 | * |
eb619fdb JA |
1655 | * If 'no_tag' is set, that means that we failed getting |
1656 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1657 | * waitqueue is no longer active, ensure that we run the queue | |
1658 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1659 | * |
710c785f BVA |
1660 | * If no I/O scheduler has been configured it is possible that |
1661 | * the hardware queue got stopped and restarted before requests | |
1662 | * were pushed back onto the dispatch list. Rerun the queue to | |
1663 | * avoid starvation. Notes: | |
1664 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1665 | * been stopped before rerunning a queue. | |
1666 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1667 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1668 | * and dm-rq. |
86ff7c2a ML |
1669 | * |
1670 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1671 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1672 | * that could otherwise occur if the queue is idle. We'll do |
1673 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1674 | */ |
86ff7c2a ML |
1675 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1676 | if (!needs_restart || | |
eb619fdb | 1677 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1678 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1679 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1680 | no_budget_avail)) | |
86ff7c2a | 1681 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1682 | |
6e768717 | 1683 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1684 | return false; |
6e768717 ML |
1685 | } else |
1686 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1687 | |
93efe981 | 1688 | return (queued + errors) != 0; |
f04c3df3 JA |
1689 | } |
1690 | ||
105663f7 AA |
1691 | /** |
1692 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1693 | * @hctx: Pointer to the hardware queue to run. | |
1694 | * | |
1695 | * Send pending requests to the hardware. | |
1696 | */ | |
6a83e74d BVA |
1697 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1698 | { | |
1699 | int srcu_idx; | |
1700 | ||
b7a71e66 JA |
1701 | /* |
1702 | * We can't run the queue inline with ints disabled. Ensure that | |
1703 | * we catch bad users of this early. | |
1704 | */ | |
1705 | WARN_ON_ONCE(in_interrupt()); | |
1706 | ||
04ced159 | 1707 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1708 | |
04ced159 JA |
1709 | hctx_lock(hctx, &srcu_idx); |
1710 | blk_mq_sched_dispatch_requests(hctx); | |
1711 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1712 | } |
1713 | ||
f82ddf19 ML |
1714 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1715 | { | |
1716 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1717 | ||
1718 | if (cpu >= nr_cpu_ids) | |
1719 | cpu = cpumask_first(hctx->cpumask); | |
1720 | return cpu; | |
1721 | } | |
1722 | ||
506e931f JA |
1723 | /* |
1724 | * It'd be great if the workqueue API had a way to pass | |
1725 | * in a mask and had some smarts for more clever placement. | |
1726 | * For now we just round-robin here, switching for every | |
1727 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1728 | */ | |
1729 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1730 | { | |
7bed4595 | 1731 | bool tried = false; |
476f8c98 | 1732 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1733 | |
b657d7e6 CH |
1734 | if (hctx->queue->nr_hw_queues == 1) |
1735 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1736 | |
1737 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1738 | select_cpu: |
476f8c98 | 1739 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1740 | cpu_online_mask); |
506e931f | 1741 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1742 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1743 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1744 | } | |
1745 | ||
7bed4595 ML |
1746 | /* |
1747 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1748 | * and it should only happen in the path of handling CPU DEAD. | |
1749 | */ | |
476f8c98 | 1750 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1751 | if (!tried) { |
1752 | tried = true; | |
1753 | goto select_cpu; | |
1754 | } | |
1755 | ||
1756 | /* | |
1757 | * Make sure to re-select CPU next time once after CPUs | |
1758 | * in hctx->cpumask become online again. | |
1759 | */ | |
476f8c98 | 1760 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1761 | hctx->next_cpu_batch = 1; |
1762 | return WORK_CPU_UNBOUND; | |
1763 | } | |
476f8c98 ML |
1764 | |
1765 | hctx->next_cpu = next_cpu; | |
1766 | return next_cpu; | |
506e931f JA |
1767 | } |
1768 | ||
105663f7 AA |
1769 | /** |
1770 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1771 | * @hctx: Pointer to the hardware queue to run. | |
1772 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1773 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1774 | * |
1775 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1776 | * with a delay of @msecs. | |
1777 | */ | |
7587a5ae BVA |
1778 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1779 | unsigned long msecs) | |
320ae51f | 1780 | { |
5435c023 | 1781 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1782 | return; |
1783 | ||
1b792f2f | 1784 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1785 | int cpu = get_cpu(); |
1786 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1787 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1788 | put_cpu(); |
398205b8 PB |
1789 | return; |
1790 | } | |
e4043dcf | 1791 | |
2a90d4aa | 1792 | put_cpu(); |
e4043dcf | 1793 | } |
398205b8 | 1794 | |
ae943d20 BVA |
1795 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1796 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1797 | } |
1798 | ||
105663f7 AA |
1799 | /** |
1800 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1801 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1802 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1803 | * |
1804 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1805 | */ | |
7587a5ae BVA |
1806 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1807 | { | |
1808 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1809 | } | |
1810 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1811 | ||
105663f7 AA |
1812 | /** |
1813 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1814 | * @hctx: Pointer to the hardware queue to run. | |
1815 | * @async: If we want to run the queue asynchronously. | |
1816 | * | |
1817 | * Check if the request queue is not in a quiesced state and if there are | |
1818 | * pending requests to be sent. If this is true, run the queue to send requests | |
1819 | * to hardware. | |
1820 | */ | |
626fb735 | 1821 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1822 | { |
24f5a90f ML |
1823 | int srcu_idx; |
1824 | bool need_run; | |
1825 | ||
1826 | /* | |
1827 | * When queue is quiesced, we may be switching io scheduler, or | |
1828 | * updating nr_hw_queues, or other things, and we can't run queue | |
1829 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1830 | * | |
1831 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1832 | * quiesced. | |
1833 | */ | |
04ced159 JA |
1834 | hctx_lock(hctx, &srcu_idx); |
1835 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1836 | blk_mq_hctx_has_pending(hctx); | |
1837 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1838 | |
626fb735 | 1839 | if (need_run) |
79f720a7 | 1840 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1841 | } |
5b727272 | 1842 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1843 | |
b6e68ee8 JK |
1844 | /* |
1845 | * Is the request queue handled by an IO scheduler that does not respect | |
1846 | * hardware queues when dispatching? | |
1847 | */ | |
1848 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
1849 | { | |
1850 | struct elevator_queue *e = q->elevator; | |
1851 | ||
1852 | if (e && e->type->ops.dispatch_request && | |
1853 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
1854 | return true; | |
1855 | return false; | |
1856 | } | |
1857 | ||
1858 | /* | |
1859 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
1860 | * scheduler. | |
1861 | */ | |
1862 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
1863 | { | |
1864 | struct blk_mq_hw_ctx *hctx; | |
1865 | ||
1866 | /* | |
1867 | * If the IO scheduler does not respect hardware queues when | |
1868 | * dispatching, we just don't bother with multiple HW queues and | |
1869 | * dispatch from hctx for the current CPU since running multiple queues | |
1870 | * just causes lock contention inside the scheduler and pointless cache | |
1871 | * bouncing. | |
1872 | */ | |
1873 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
1874 | raw_smp_processor_id()); | |
1875 | if (!blk_mq_hctx_stopped(hctx)) | |
1876 | return hctx; | |
1877 | return NULL; | |
1878 | } | |
1879 | ||
105663f7 | 1880 | /** |
24f7bb88 | 1881 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1882 | * @q: Pointer to the request queue to run. |
1883 | * @async: If we want to run the queue asynchronously. | |
1884 | */ | |
b94ec296 | 1885 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 1886 | { |
b6e68ee8 | 1887 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
1888 | int i; |
1889 | ||
b6e68ee8 JK |
1890 | sq_hctx = NULL; |
1891 | if (blk_mq_has_sqsched(q)) | |
1892 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 1893 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 1894 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 1895 | continue; |
b6e68ee8 JK |
1896 | /* |
1897 | * Dispatch from this hctx either if there's no hctx preferred | |
1898 | * by IO scheduler or if it has requests that bypass the | |
1899 | * scheduler. | |
1900 | */ | |
1901 | if (!sq_hctx || sq_hctx == hctx || | |
1902 | !list_empty_careful(&hctx->dispatch)) | |
1903 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
1904 | } |
1905 | } | |
b94ec296 | 1906 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1907 | |
b9151e7b DA |
1908 | /** |
1909 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1910 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1911 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1912 | */ |
1913 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1914 | { | |
b6e68ee8 | 1915 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
1916 | int i; |
1917 | ||
b6e68ee8 JK |
1918 | sq_hctx = NULL; |
1919 | if (blk_mq_has_sqsched(q)) | |
1920 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
1921 | queue_for_each_hw_ctx(q, hctx, i) { |
1922 | if (blk_mq_hctx_stopped(hctx)) | |
1923 | continue; | |
b6e68ee8 JK |
1924 | /* |
1925 | * Dispatch from this hctx either if there's no hctx preferred | |
1926 | * by IO scheduler or if it has requests that bypass the | |
1927 | * scheduler. | |
1928 | */ | |
1929 | if (!sq_hctx || sq_hctx == hctx || | |
1930 | !list_empty_careful(&hctx->dispatch)) | |
1931 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
1932 | } |
1933 | } | |
1934 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1935 | ||
fd001443 BVA |
1936 | /** |
1937 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1938 | * @q: request queue. | |
1939 | * | |
1940 | * The caller is responsible for serializing this function against | |
1941 | * blk_mq_{start,stop}_hw_queue(). | |
1942 | */ | |
1943 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1944 | { | |
1945 | struct blk_mq_hw_ctx *hctx; | |
1946 | int i; | |
1947 | ||
1948 | queue_for_each_hw_ctx(q, hctx, i) | |
1949 | if (blk_mq_hctx_stopped(hctx)) | |
1950 | return true; | |
1951 | ||
1952 | return false; | |
1953 | } | |
1954 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1955 | ||
39a70c76 ML |
1956 | /* |
1957 | * This function is often used for pausing .queue_rq() by driver when | |
1958 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1959 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1960 | * |
1961 | * We do not guarantee that dispatch can be drained or blocked | |
1962 | * after blk_mq_stop_hw_queue() returns. Please use | |
1963 | * blk_mq_quiesce_queue() for that requirement. | |
1964 | */ | |
2719aa21 JA |
1965 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1966 | { | |
641a9ed6 | 1967 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1968 | |
641a9ed6 | 1969 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1970 | } |
641a9ed6 | 1971 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1972 | |
39a70c76 ML |
1973 | /* |
1974 | * This function is often used for pausing .queue_rq() by driver when | |
1975 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1976 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1977 | * |
1978 | * We do not guarantee that dispatch can be drained or blocked | |
1979 | * after blk_mq_stop_hw_queues() returns. Please use | |
1980 | * blk_mq_quiesce_queue() for that requirement. | |
1981 | */ | |
2719aa21 JA |
1982 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1983 | { | |
641a9ed6 ML |
1984 | struct blk_mq_hw_ctx *hctx; |
1985 | int i; | |
1986 | ||
1987 | queue_for_each_hw_ctx(q, hctx, i) | |
1988 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1989 | } |
1990 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1991 | ||
320ae51f JA |
1992 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1993 | { | |
1994 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1995 | |
0ffbce80 | 1996 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1997 | } |
1998 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1999 | ||
2f268556 CH |
2000 | void blk_mq_start_hw_queues(struct request_queue *q) |
2001 | { | |
2002 | struct blk_mq_hw_ctx *hctx; | |
2003 | int i; | |
2004 | ||
2005 | queue_for_each_hw_ctx(q, hctx, i) | |
2006 | blk_mq_start_hw_queue(hctx); | |
2007 | } | |
2008 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
2009 | ||
ae911c5e JA |
2010 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
2011 | { | |
2012 | if (!blk_mq_hctx_stopped(hctx)) | |
2013 | return; | |
2014 | ||
2015 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
2016 | blk_mq_run_hw_queue(hctx, async); | |
2017 | } | |
2018 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
2019 | ||
1b4a3258 | 2020 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
2021 | { |
2022 | struct blk_mq_hw_ctx *hctx; | |
2023 | int i; | |
2024 | ||
ae911c5e JA |
2025 | queue_for_each_hw_ctx(q, hctx, i) |
2026 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
2027 | } |
2028 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
2029 | ||
70f4db63 | 2030 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
2031 | { |
2032 | struct blk_mq_hw_ctx *hctx; | |
2033 | ||
9f993737 | 2034 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 2035 | |
21c6e939 | 2036 | /* |
15fe8a90 | 2037 | * If we are stopped, don't run the queue. |
21c6e939 | 2038 | */ |
0841031a | 2039 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 2040 | return; |
7587a5ae BVA |
2041 | |
2042 | __blk_mq_run_hw_queue(hctx); | |
2043 | } | |
2044 | ||
cfd0c552 | 2045 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
2046 | struct request *rq, |
2047 | bool at_head) | |
320ae51f | 2048 | { |
e57690fe | 2049 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 2050 | enum hctx_type type = hctx->type; |
e57690fe | 2051 | |
7b607814 BVA |
2052 | lockdep_assert_held(&ctx->lock); |
2053 | ||
a54895fa | 2054 | trace_block_rq_insert(rq); |
01b983c9 | 2055 | |
72a0a36e | 2056 | if (at_head) |
c16d6b5a | 2057 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 2058 | else |
c16d6b5a | 2059 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 2060 | } |
4bb659b1 | 2061 | |
2c3ad667 JA |
2062 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
2063 | bool at_head) | |
cfd0c552 ML |
2064 | { |
2065 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
2066 | ||
7b607814 BVA |
2067 | lockdep_assert_held(&ctx->lock); |
2068 | ||
e57690fe | 2069 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 2070 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
2071 | } |
2072 | ||
105663f7 AA |
2073 | /** |
2074 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
2075 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 2076 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
2077 | * @run_queue: If we should run the hardware queue after inserting the request. |
2078 | * | |
157f377b JA |
2079 | * Should only be used carefully, when the caller knows we want to |
2080 | * bypass a potential IO scheduler on the target device. | |
2081 | */ | |
01e99aec ML |
2082 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
2083 | bool run_queue) | |
157f377b | 2084 | { |
ea4f995e | 2085 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
2086 | |
2087 | spin_lock(&hctx->lock); | |
01e99aec ML |
2088 | if (at_head) |
2089 | list_add(&rq->queuelist, &hctx->dispatch); | |
2090 | else | |
2091 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
2092 | spin_unlock(&hctx->lock); |
2093 | ||
b0850297 ML |
2094 | if (run_queue) |
2095 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
2096 | } |
2097 | ||
bd166ef1 JA |
2098 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
2099 | struct list_head *list) | |
320ae51f JA |
2100 | |
2101 | { | |
3f0cedc7 | 2102 | struct request *rq; |
c16d6b5a | 2103 | enum hctx_type type = hctx->type; |
3f0cedc7 | 2104 | |
320ae51f JA |
2105 | /* |
2106 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
2107 | * offline now | |
2108 | */ | |
3f0cedc7 | 2109 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 2110 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 2111 | trace_block_rq_insert(rq); |
320ae51f | 2112 | } |
3f0cedc7 ML |
2113 | |
2114 | spin_lock(&ctx->lock); | |
c16d6b5a | 2115 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 2116 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 2117 | spin_unlock(&ctx->lock); |
320ae51f JA |
2118 | } |
2119 | ||
4f0f586b ST |
2120 | static int plug_rq_cmp(void *priv, const struct list_head *a, |
2121 | const struct list_head *b) | |
320ae51f JA |
2122 | { |
2123 | struct request *rqa = container_of(a, struct request, queuelist); | |
2124 | struct request *rqb = container_of(b, struct request, queuelist); | |
2125 | ||
7d30a621 PB |
2126 | if (rqa->mq_ctx != rqb->mq_ctx) |
2127 | return rqa->mq_ctx > rqb->mq_ctx; | |
2128 | if (rqa->mq_hctx != rqb->mq_hctx) | |
2129 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
2130 | |
2131 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
2132 | } |
2133 | ||
2134 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
2135 | { | |
320ae51f | 2136 | LIST_HEAD(list); |
320ae51f | 2137 | |
95ed0c5b PB |
2138 | if (list_empty(&plug->mq_list)) |
2139 | return; | |
320ae51f JA |
2140 | list_splice_init(&plug->mq_list, &list); |
2141 | ||
ce5b009c JA |
2142 | if (plug->rq_count > 2 && plug->multiple_queues) |
2143 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 2144 | |
bcc816df DZ |
2145 | plug->rq_count = 0; |
2146 | ||
95ed0c5b PB |
2147 | do { |
2148 | struct list_head rq_list; | |
2149 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
2150 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
2151 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
2152 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
2153 | unsigned int depth = 1; | |
2154 | ||
2155 | list_for_each_continue(pos, &list) { | |
2156 | rq = list_entry_rq(pos); | |
2157 | BUG_ON(!rq->q); | |
2158 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
2159 | break; | |
2160 | depth++; | |
320ae51f JA |
2161 | } |
2162 | ||
95ed0c5b PB |
2163 | list_cut_before(&rq_list, &list, pos); |
2164 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 2165 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 2166 | from_schedule); |
95ed0c5b | 2167 | } while(!list_empty(&list)); |
320ae51f JA |
2168 | } |
2169 | ||
14ccb66b CH |
2170 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
2171 | unsigned int nr_segs) | |
320ae51f | 2172 | { |
93f221ae EB |
2173 | int err; |
2174 | ||
f924cdde CH |
2175 | if (bio->bi_opf & REQ_RAHEAD) |
2176 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
2177 | ||
2178 | rq->__sector = bio->bi_iter.bi_sector; | |
2179 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 2180 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
2181 | |
2182 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2183 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2184 | WARN_ON_ONCE(err); | |
4b570521 | 2185 | |
b5af37ab | 2186 | blk_account_io_start(rq); |
320ae51f JA |
2187 | } |
2188 | ||
0f95549c | 2189 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2190 | struct request *rq, bool last) |
f984df1f | 2191 | { |
f984df1f | 2192 | struct request_queue *q = rq->q; |
f984df1f SL |
2193 | struct blk_mq_queue_data bd = { |
2194 | .rq = rq, | |
be94f058 | 2195 | .last = last, |
f984df1f | 2196 | }; |
f06345ad | 2197 | blk_status_t ret; |
0f95549c | 2198 | |
0f95549c MS |
2199 | /* |
2200 | * For OK queue, we are done. For error, caller may kill it. | |
2201 | * Any other error (busy), just add it to our list as we | |
2202 | * previously would have done. | |
2203 | */ | |
2204 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2205 | switch (ret) { | |
2206 | case BLK_STS_OK: | |
6ce3dd6e | 2207 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2208 | break; |
2209 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2210 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2211 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2212 | __blk_mq_requeue_request(rq); |
2213 | break; | |
2214 | default: | |
6ce3dd6e | 2215 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2216 | break; |
2217 | } | |
2218 | ||
2219 | return ret; | |
2220 | } | |
2221 | ||
fd9c40f6 | 2222 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2223 | struct request *rq, |
fd9c40f6 | 2224 | bool bypass_insert, bool last) |
0f95549c MS |
2225 | { |
2226 | struct request_queue *q = rq->q; | |
d964f04a | 2227 | bool run_queue = true; |
2a5a24aa | 2228 | int budget_token; |
d964f04a | 2229 | |
23d4ee19 | 2230 | /* |
fd9c40f6 | 2231 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2232 | * |
fd9c40f6 BVA |
2233 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2234 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2235 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2236 | */ |
fd9c40f6 | 2237 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2238 | run_queue = false; |
fd9c40f6 BVA |
2239 | bypass_insert = false; |
2240 | goto insert; | |
d964f04a | 2241 | } |
f984df1f | 2242 | |
2ff0682d | 2243 | if ((rq->rq_flags & RQF_ELV) && !bypass_insert) |
fd9c40f6 | 2244 | goto insert; |
2253efc8 | 2245 | |
2a5a24aa ML |
2246 | budget_token = blk_mq_get_dispatch_budget(q); |
2247 | if (budget_token < 0) | |
fd9c40f6 | 2248 | goto insert; |
bd166ef1 | 2249 | |
2a5a24aa ML |
2250 | blk_mq_set_rq_budget_token(rq, budget_token); |
2251 | ||
8ab6bb9e | 2252 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2253 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2254 | goto insert; |
88022d72 | 2255 | } |
de148297 | 2256 | |
3e08773c | 2257 | return __blk_mq_issue_directly(hctx, rq, last); |
fd9c40f6 BVA |
2258 | insert: |
2259 | if (bypass_insert) | |
2260 | return BLK_STS_RESOURCE; | |
2261 | ||
db03f88f ML |
2262 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2263 | ||
fd9c40f6 BVA |
2264 | return BLK_STS_OK; |
2265 | } | |
2266 | ||
105663f7 AA |
2267 | /** |
2268 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2269 | * @hctx: Pointer of the associated hardware queue. | |
2270 | * @rq: Pointer to request to be sent. | |
105663f7 AA |
2271 | * |
2272 | * If the device has enough resources to accept a new request now, send the | |
2273 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2274 | * we can try send it another time in the future. Requests inserted at this | |
2275 | * queue have higher priority. | |
2276 | */ | |
fd9c40f6 | 2277 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
3e08773c | 2278 | struct request *rq) |
fd9c40f6 BVA |
2279 | { |
2280 | blk_status_t ret; | |
2281 | int srcu_idx; | |
2282 | ||
2283 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2284 | ||
2285 | hctx_lock(hctx, &srcu_idx); | |
2286 | ||
3e08773c | 2287 | ret = __blk_mq_try_issue_directly(hctx, rq, false, true); |
fd9c40f6 | 2288 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) |
01e99aec | 2289 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2290 | else if (ret != BLK_STS_OK) |
2291 | blk_mq_end_request(rq, ret); | |
2292 | ||
2293 | hctx_unlock(hctx, srcu_idx); | |
2294 | } | |
2295 | ||
2296 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2297 | { | |
2298 | blk_status_t ret; | |
2299 | int srcu_idx; | |
fd9c40f6 BVA |
2300 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
2301 | ||
2302 | hctx_lock(hctx, &srcu_idx); | |
3e08773c | 2303 | ret = __blk_mq_try_issue_directly(hctx, rq, true, last); |
04ced159 | 2304 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2305 | |
2306 | return ret; | |
5eb6126e CH |
2307 | } |
2308 | ||
6ce3dd6e ML |
2309 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2310 | struct list_head *list) | |
2311 | { | |
536167d4 | 2312 | int queued = 0; |
632bfb63 | 2313 | int errors = 0; |
536167d4 | 2314 | |
6ce3dd6e | 2315 | while (!list_empty(list)) { |
fd9c40f6 | 2316 | blk_status_t ret; |
6ce3dd6e ML |
2317 | struct request *rq = list_first_entry(list, struct request, |
2318 | queuelist); | |
2319 | ||
2320 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2321 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2322 | if (ret != BLK_STS_OK) { | |
2323 | if (ret == BLK_STS_RESOURCE || | |
2324 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2325 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2326 | list_empty(list)); |
fd9c40f6 BVA |
2327 | break; |
2328 | } | |
2329 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2330 | errors++; |
536167d4 KB |
2331 | } else |
2332 | queued++; | |
6ce3dd6e | 2333 | } |
d666ba98 JA |
2334 | |
2335 | /* | |
2336 | * If we didn't flush the entire list, we could have told | |
2337 | * the driver there was more coming, but that turned out to | |
2338 | * be a lie. | |
2339 | */ | |
632bfb63 | 2340 | if ((!list_empty(list) || errors) && |
2341 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2342 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2343 | } |
2344 | ||
ce5b009c JA |
2345 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2346 | { | |
2347 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2348 | plug->rq_count++; | |
2349 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2350 | struct request *tmp; | |
2351 | ||
2352 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2353 | queuelist); | |
2354 | if (tmp->q != rq->q) | |
2355 | plug->multiple_queues = true; | |
2356 | } | |
2357 | } | |
2358 | ||
7f2a6a69 | 2359 | /* |
ba0ffdd8 | 2360 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
7f2a6a69 SL |
2361 | * queues. This is important for md arrays to benefit from merging |
2362 | * requests. | |
2363 | */ | |
2364 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2365 | { | |
2366 | if (plug->multiple_queues) | |
ba0ffdd8 | 2367 | return BLK_MAX_REQUEST_COUNT * 2; |
7f2a6a69 SL |
2368 | return BLK_MAX_REQUEST_COUNT; |
2369 | } | |
2370 | ||
105663f7 | 2371 | /** |
c62b37d9 | 2372 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2373 | * @bio: Bio pointer. |
2374 | * | |
2375 | * Builds up a request structure from @q and @bio and send to the device. The | |
2376 | * request may not be queued directly to hardware if: | |
2377 | * * This request can be merged with another one | |
2378 | * * We want to place request at plug queue for possible future merging | |
2379 | * * There is an IO scheduler active at this queue | |
2380 | * | |
2381 | * It will not queue the request if there is an error with the bio, or at the | |
2382 | * request creation. | |
105663f7 | 2383 | */ |
3e08773c | 2384 | void blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2385 | { |
ed6cddef | 2386 | struct request_queue *q = bdev_get_queue(bio->bi_bdev); |
ef295ecf | 2387 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2388 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
07068d5b | 2389 | struct request *rq; |
f984df1f | 2390 | struct blk_plug *plug; |
5b3f341f | 2391 | struct request *same_queue_rq = NULL; |
abd45c15 | 2392 | unsigned int nr_segs = 1; |
a892c8d5 | 2393 | blk_status_t ret; |
07068d5b JA |
2394 | |
2395 | blk_queue_bounce(q, &bio); | |
abd45c15 JA |
2396 | if (blk_may_split(q, bio)) |
2397 | __blk_queue_split(q, &bio, &nr_segs); | |
f36ea50c | 2398 | |
e23947bd | 2399 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2400 | goto queue_exit; |
07068d5b | 2401 | |
87c279e6 | 2402 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2403 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2404 | goto queue_exit; |
f984df1f | 2405 | |
14ccb66b | 2406 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2407 | goto queue_exit; |
bd166ef1 | 2408 | |
d5337560 | 2409 | rq_qos_throttle(q, bio); |
87760e5e | 2410 | |
47c122e3 JA |
2411 | plug = blk_mq_plug(q, bio); |
2412 | if (plug && plug->cached_rq) { | |
013a7f95 | 2413 | rq = rq_list_pop(&plug->cached_rq); |
47c122e3 | 2414 | INIT_LIST_HEAD(&rq->queuelist); |
47c122e3 | 2415 | } else { |
0f38d766 CH |
2416 | struct blk_mq_alloc_data data = { |
2417 | .q = q, | |
2418 | .nr_tags = 1, | |
2419 | .cmd_flags = bio->bi_opf, | |
2420 | }; | |
2421 | ||
47c122e3 JA |
2422 | if (plug) { |
2423 | data.nr_tags = plug->nr_ios; | |
2424 | plug->nr_ios = 1; | |
2425 | data.cached_rq = &plug->cached_rq; | |
2426 | } | |
b90cfaed | 2427 | rq = __blk_mq_alloc_requests(&data); |
47c122e3 JA |
2428 | if (unlikely(!rq)) { |
2429 | rq_qos_cleanup(q, bio); | |
2430 | if (bio->bi_opf & REQ_NOWAIT) | |
2431 | bio_wouldblock_error(bio); | |
2432 | goto queue_exit; | |
2433 | } | |
87760e5e JA |
2434 | } |
2435 | ||
e8a676d6 | 2436 | trace_block_getrq(bio); |
d6f1dda2 | 2437 | |
c1c80384 | 2438 | rq_qos_track(q, rq, bio); |
07068d5b | 2439 | |
970d168d BVA |
2440 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2441 | ||
a892c8d5 ST |
2442 | ret = blk_crypto_init_request(rq); |
2443 | if (ret != BLK_STS_OK) { | |
2444 | bio->bi_status = ret; | |
2445 | bio_endio(bio); | |
2446 | blk_mq_free_request(rq); | |
3e08773c | 2447 | return; |
a892c8d5 ST |
2448 | } |
2449 | ||
07068d5b | 2450 | if (unlikely(is_flush_fua)) { |
4a60f360 | 2451 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
105663f7 | 2452 | /* Bypass scheduler for flush requests */ |
923218f6 | 2453 | blk_insert_flush(rq); |
4a60f360 | 2454 | blk_mq_run_hw_queue(hctx, true); |
03f26d8f | 2455 | } else if (plug && (q->nr_hw_queues == 1 || |
079a2e3e | 2456 | blk_mq_is_shared_tags(rq->mq_hctx->flags) || |
03f26d8f | 2457 | q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) { |
b2c5d16b JA |
2458 | /* |
2459 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2460 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2461 | * |
2462 | * Use normal plugging if this disk is slow HDD, as sequential | |
2463 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2464 | */ |
5f0ed774 | 2465 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2466 | struct request *last = NULL; |
2467 | ||
676d0607 | 2468 | if (!request_count) |
e6c4438b | 2469 | trace_block_plug(q); |
600271d9 SL |
2470 | else |
2471 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2472 | |
7f2a6a69 | 2473 | if (request_count >= blk_plug_max_rq_count(plug) || (last && |
600271d9 | 2474 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
e6c4438b JM |
2475 | blk_flush_plug_list(plug, false); |
2476 | trace_block_plug(q); | |
320ae51f | 2477 | } |
b094f89c | 2478 | |
ce5b009c | 2479 | blk_add_rq_to_plug(plug, rq); |
2ff0682d | 2480 | } else if (rq->rq_flags & RQF_ELV) { |
105663f7 | 2481 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2482 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2483 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2484 | /* |
6a83e74d | 2485 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2486 | * Otherwise the existing request in the plug list will be |
2487 | * issued. So the plug list will have one request at most | |
2299722c CH |
2488 | * The plug list might get flushed before this. If that happens, |
2489 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2490 | */ |
2299722c CH |
2491 | if (list_empty(&plug->mq_list)) |
2492 | same_queue_rq = NULL; | |
4711b573 | 2493 | if (same_queue_rq) { |
2299722c | 2494 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2495 | plug->rq_count--; |
2496 | } | |
ce5b009c | 2497 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2498 | trace_block_plug(q); |
2299722c | 2499 | |
dad7a3be | 2500 | if (same_queue_rq) { |
ff3b74b8 | 2501 | trace_block_unplug(q, 1, true); |
0f38d766 | 2502 | blk_mq_try_issue_directly(same_queue_rq->mq_hctx, |
3e08773c | 2503 | same_queue_rq); |
dad7a3be | 2504 | } |
a12de1d4 | 2505 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
0f38d766 | 2506 | !rq->mq_hctx->dispatch_busy) { |
105663f7 AA |
2507 | /* |
2508 | * There is no scheduler and we can try to send directly | |
2509 | * to the hardware. | |
2510 | */ | |
3e08773c | 2511 | blk_mq_try_issue_directly(rq->mq_hctx, rq); |
ab42f35d | 2512 | } else { |
105663f7 | 2513 | /* Default case. */ |
8fa9f556 | 2514 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2515 | } |
320ae51f | 2516 | |
3e08773c | 2517 | return; |
ac7c5675 CH |
2518 | queue_exit: |
2519 | blk_queue_exit(q); | |
320ae51f JA |
2520 | } |
2521 | ||
bd63141d ML |
2522 | static size_t order_to_size(unsigned int order) |
2523 | { | |
2524 | return (size_t)PAGE_SIZE << order; | |
2525 | } | |
2526 | ||
2527 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
2528 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
2529 | struct blk_mq_tags *tags) | |
bd63141d | 2530 | { |
bd63141d ML |
2531 | struct page *page; |
2532 | unsigned long flags; | |
2533 | ||
4f245d5b JG |
2534 | /* There is no need to clear a driver tags own mapping */ |
2535 | if (drv_tags == tags) | |
2536 | return; | |
2537 | ||
bd63141d ML |
2538 | list_for_each_entry(page, &tags->page_list, lru) { |
2539 | unsigned long start = (unsigned long)page_address(page); | |
2540 | unsigned long end = start + order_to_size(page->private); | |
2541 | int i; | |
2542 | ||
f32e4eaf | 2543 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
2544 | struct request *rq = drv_tags->rqs[i]; |
2545 | unsigned long rq_addr = (unsigned long)rq; | |
2546 | ||
2547 | if (rq_addr >= start && rq_addr < end) { | |
2548 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
2549 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
2550 | } | |
2551 | } | |
2552 | } | |
2553 | ||
2554 | /* | |
2555 | * Wait until all pending iteration is done. | |
2556 | * | |
2557 | * Request reference is cleared and it is guaranteed to be observed | |
2558 | * after the ->lock is released. | |
2559 | */ | |
2560 | spin_lock_irqsave(&drv_tags->lock, flags); | |
2561 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
2562 | } | |
2563 | ||
cc71a6f4 JA |
2564 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2565 | unsigned int hctx_idx) | |
95363efd | 2566 | { |
f32e4eaf | 2567 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 2568 | struct page *page; |
320ae51f | 2569 | |
079a2e3e JG |
2570 | if (blk_mq_is_shared_tags(set->flags)) |
2571 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
2572 | else |
2573 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 2574 | |
65de57bb | 2575 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 2576 | int i; |
320ae51f | 2577 | |
24d2f903 | 2578 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2579 | struct request *rq = tags->static_rqs[i]; |
2580 | ||
2581 | if (!rq) | |
e9b267d9 | 2582 | continue; |
d6296d39 | 2583 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2584 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2585 | } |
320ae51f | 2586 | } |
320ae51f | 2587 | |
f32e4eaf | 2588 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 2589 | |
24d2f903 CH |
2590 | while (!list_empty(&tags->page_list)) { |
2591 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2592 | list_del_init(&page->lru); |
f75782e4 CM |
2593 | /* |
2594 | * Remove kmemleak object previously allocated in | |
273938bf | 2595 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2596 | */ |
2597 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2598 | __free_pages(page, page->private); |
2599 | } | |
cc71a6f4 | 2600 | } |
320ae51f | 2601 | |
e155b0c2 | 2602 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 2603 | { |
24d2f903 | 2604 | kfree(tags->rqs); |
cc71a6f4 | 2605 | tags->rqs = NULL; |
2af8cbe3 JA |
2606 | kfree(tags->static_rqs); |
2607 | tags->static_rqs = NULL; | |
320ae51f | 2608 | |
e155b0c2 | 2609 | blk_mq_free_tags(tags); |
320ae51f JA |
2610 | } |
2611 | ||
63064be1 JG |
2612 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2613 | unsigned int hctx_idx, | |
2614 | unsigned int nr_tags, | |
e155b0c2 | 2615 | unsigned int reserved_tags) |
320ae51f | 2616 | { |
24d2f903 | 2617 | struct blk_mq_tags *tags; |
59f082e4 | 2618 | int node; |
320ae51f | 2619 | |
7d76f856 | 2620 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2621 | if (node == NUMA_NO_NODE) |
2622 | node = set->numa_node; | |
2623 | ||
e155b0c2 JG |
2624 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
2625 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
2626 | if (!tags) |
2627 | return NULL; | |
320ae51f | 2628 | |
590b5b7d | 2629 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2630 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2631 | node); |
24d2f903 | 2632 | if (!tags->rqs) { |
e155b0c2 | 2633 | blk_mq_free_tags(tags); |
24d2f903 CH |
2634 | return NULL; |
2635 | } | |
320ae51f | 2636 | |
590b5b7d KC |
2637 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2638 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2639 | node); | |
2af8cbe3 JA |
2640 | if (!tags->static_rqs) { |
2641 | kfree(tags->rqs); | |
e155b0c2 | 2642 | blk_mq_free_tags(tags); |
2af8cbe3 JA |
2643 | return NULL; |
2644 | } | |
2645 | ||
cc71a6f4 JA |
2646 | return tags; |
2647 | } | |
2648 | ||
1d9bd516 TH |
2649 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2650 | unsigned int hctx_idx, int node) | |
2651 | { | |
2652 | int ret; | |
2653 | ||
2654 | if (set->ops->init_request) { | |
2655 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2656 | if (ret) | |
2657 | return ret; | |
2658 | } | |
2659 | ||
12f5b931 | 2660 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2661 | return 0; |
2662 | } | |
2663 | ||
63064be1 JG |
2664 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
2665 | struct blk_mq_tags *tags, | |
2666 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
2667 | { |
2668 | unsigned int i, j, entries_per_page, max_order = 4; | |
2669 | size_t rq_size, left; | |
59f082e4 SL |
2670 | int node; |
2671 | ||
7d76f856 | 2672 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2673 | if (node == NUMA_NO_NODE) |
2674 | node = set->numa_node; | |
cc71a6f4 JA |
2675 | |
2676 | INIT_LIST_HEAD(&tags->page_list); | |
2677 | ||
320ae51f JA |
2678 | /* |
2679 | * rq_size is the size of the request plus driver payload, rounded | |
2680 | * to the cacheline size | |
2681 | */ | |
24d2f903 | 2682 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2683 | cache_line_size()); |
cc71a6f4 | 2684 | left = rq_size * depth; |
320ae51f | 2685 | |
cc71a6f4 | 2686 | for (i = 0; i < depth; ) { |
320ae51f JA |
2687 | int this_order = max_order; |
2688 | struct page *page; | |
2689 | int to_do; | |
2690 | void *p; | |
2691 | ||
b3a834b1 | 2692 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2693 | this_order--; |
2694 | ||
2695 | do { | |
59f082e4 | 2696 | page = alloc_pages_node(node, |
36e1f3d1 | 2697 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2698 | this_order); |
320ae51f JA |
2699 | if (page) |
2700 | break; | |
2701 | if (!this_order--) | |
2702 | break; | |
2703 | if (order_to_size(this_order) < rq_size) | |
2704 | break; | |
2705 | } while (1); | |
2706 | ||
2707 | if (!page) | |
24d2f903 | 2708 | goto fail; |
320ae51f JA |
2709 | |
2710 | page->private = this_order; | |
24d2f903 | 2711 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2712 | |
2713 | p = page_address(page); | |
f75782e4 CM |
2714 | /* |
2715 | * Allow kmemleak to scan these pages as they contain pointers | |
2716 | * to additional allocations like via ops->init_request(). | |
2717 | */ | |
36e1f3d1 | 2718 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2719 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2720 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2721 | left -= to_do * rq_size; |
2722 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2723 | struct request *rq = p; |
2724 | ||
2725 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2726 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2727 | tags->static_rqs[i] = NULL; | |
2728 | goto fail; | |
e9b267d9 CH |
2729 | } |
2730 | ||
320ae51f JA |
2731 | p += rq_size; |
2732 | i++; | |
2733 | } | |
2734 | } | |
cc71a6f4 | 2735 | return 0; |
320ae51f | 2736 | |
24d2f903 | 2737 | fail: |
cc71a6f4 JA |
2738 | blk_mq_free_rqs(set, tags, hctx_idx); |
2739 | return -ENOMEM; | |
320ae51f JA |
2740 | } |
2741 | ||
bf0beec0 ML |
2742 | struct rq_iter_data { |
2743 | struct blk_mq_hw_ctx *hctx; | |
2744 | bool has_rq; | |
2745 | }; | |
2746 | ||
2747 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2748 | { | |
2749 | struct rq_iter_data *iter_data = data; | |
2750 | ||
2751 | if (rq->mq_hctx != iter_data->hctx) | |
2752 | return true; | |
2753 | iter_data->has_rq = true; | |
2754 | return false; | |
2755 | } | |
2756 | ||
2757 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2758 | { | |
2759 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2760 | hctx->sched_tags : hctx->tags; | |
2761 | struct rq_iter_data data = { | |
2762 | .hctx = hctx, | |
2763 | }; | |
2764 | ||
2765 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2766 | return data.has_rq; | |
2767 | } | |
2768 | ||
2769 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2770 | struct blk_mq_hw_ctx *hctx) | |
2771 | { | |
2772 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2773 | return false; | |
2774 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2775 | return false; | |
2776 | return true; | |
2777 | } | |
2778 | ||
2779 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2780 | { | |
2781 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2782 | struct blk_mq_hw_ctx, cpuhp_online); | |
2783 | ||
2784 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2785 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2786 | return 0; | |
2787 | ||
2788 | /* | |
2789 | * Prevent new request from being allocated on the current hctx. | |
2790 | * | |
2791 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2792 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2793 | * seen once we return from the tag allocator. | |
2794 | */ | |
2795 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2796 | smp_mb__after_atomic(); | |
2797 | ||
2798 | /* | |
2799 | * Try to grab a reference to the queue and wait for any outstanding | |
2800 | * requests. If we could not grab a reference the queue has been | |
2801 | * frozen and there are no requests. | |
2802 | */ | |
2803 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2804 | while (blk_mq_hctx_has_requests(hctx)) | |
2805 | msleep(5); | |
2806 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2807 | } | |
2808 | ||
2809 | return 0; | |
2810 | } | |
2811 | ||
2812 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2813 | { | |
2814 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2815 | struct blk_mq_hw_ctx, cpuhp_online); | |
2816 | ||
2817 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2818 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2819 | return 0; | |
2820 | } | |
2821 | ||
e57690fe JA |
2822 | /* |
2823 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2824 | * software queue to the hw queue dispatch list, and ensure that it | |
2825 | * gets run. | |
2826 | */ | |
9467f859 | 2827 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2828 | { |
9467f859 | 2829 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2830 | struct blk_mq_ctx *ctx; |
2831 | LIST_HEAD(tmp); | |
c16d6b5a | 2832 | enum hctx_type type; |
484b4061 | 2833 | |
9467f859 | 2834 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2835 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2836 | return 0; | |
2837 | ||
e57690fe | 2838 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2839 | type = hctx->type; |
484b4061 JA |
2840 | |
2841 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2842 | if (!list_empty(&ctx->rq_lists[type])) { |
2843 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2844 | blk_mq_hctx_clear_pending(hctx, ctx); |
2845 | } | |
2846 | spin_unlock(&ctx->lock); | |
2847 | ||
2848 | if (list_empty(&tmp)) | |
9467f859 | 2849 | return 0; |
484b4061 | 2850 | |
e57690fe JA |
2851 | spin_lock(&hctx->lock); |
2852 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2853 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2854 | |
2855 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2856 | return 0; |
484b4061 JA |
2857 | } |
2858 | ||
9467f859 | 2859 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2860 | { |
bf0beec0 ML |
2861 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2862 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2863 | &hctx->cpuhp_online); | |
9467f859 TG |
2864 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2865 | &hctx->cpuhp_dead); | |
484b4061 JA |
2866 | } |
2867 | ||
364b6181 ML |
2868 | /* |
2869 | * Before freeing hw queue, clearing the flush request reference in | |
2870 | * tags->rqs[] for avoiding potential UAF. | |
2871 | */ | |
2872 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
2873 | unsigned int queue_depth, struct request *flush_rq) | |
2874 | { | |
2875 | int i; | |
2876 | unsigned long flags; | |
2877 | ||
2878 | /* The hw queue may not be mapped yet */ | |
2879 | if (!tags) | |
2880 | return; | |
2881 | ||
2882 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
2883 | ||
2884 | for (i = 0; i < queue_depth; i++) | |
2885 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
2886 | ||
2887 | /* | |
2888 | * Wait until all pending iteration is done. | |
2889 | * | |
2890 | * Request reference is cleared and it is guaranteed to be observed | |
2891 | * after the ->lock is released. | |
2892 | */ | |
2893 | spin_lock_irqsave(&tags->lock, flags); | |
2894 | spin_unlock_irqrestore(&tags->lock, flags); | |
2895 | } | |
2896 | ||
c3b4afca | 2897 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2898 | static void blk_mq_exit_hctx(struct request_queue *q, |
2899 | struct blk_mq_tag_set *set, | |
2900 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2901 | { | |
364b6181 ML |
2902 | struct request *flush_rq = hctx->fq->flush_rq; |
2903 | ||
8ab0b7dc ML |
2904 | if (blk_mq_hw_queue_mapped(hctx)) |
2905 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2906 | |
364b6181 ML |
2907 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
2908 | set->queue_depth, flush_rq); | |
f70ced09 | 2909 | if (set->ops->exit_request) |
364b6181 | 2910 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 2911 | |
08e98fc6 ML |
2912 | if (set->ops->exit_hctx) |
2913 | set->ops->exit_hctx(hctx, hctx_idx); | |
2914 | ||
9467f859 | 2915 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2916 | |
2917 | spin_lock(&q->unused_hctx_lock); | |
2918 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2919 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2920 | } |
2921 | ||
624dbe47 ML |
2922 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2923 | struct blk_mq_tag_set *set, int nr_queue) | |
2924 | { | |
2925 | struct blk_mq_hw_ctx *hctx; | |
2926 | unsigned int i; | |
2927 | ||
2928 | queue_for_each_hw_ctx(q, hctx, i) { | |
2929 | if (i == nr_queue) | |
2930 | break; | |
477e19de | 2931 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2932 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2933 | } |
624dbe47 ML |
2934 | } |
2935 | ||
7c6c5b7c ML |
2936 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2937 | { | |
2938 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2939 | ||
2940 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2941 | __alignof__(struct blk_mq_hw_ctx)) != | |
2942 | sizeof(struct blk_mq_hw_ctx)); | |
2943 | ||
2944 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2945 | hw_ctx_size += sizeof(struct srcu_struct); | |
2946 | ||
2947 | return hw_ctx_size; | |
2948 | } | |
2949 | ||
08e98fc6 ML |
2950 | static int blk_mq_init_hctx(struct request_queue *q, |
2951 | struct blk_mq_tag_set *set, | |
2952 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2953 | { |
7c6c5b7c ML |
2954 | hctx->queue_num = hctx_idx; |
2955 | ||
bf0beec0 ML |
2956 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2957 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2958 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2959 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2960 | ||
2961 | hctx->tags = set->tags[hctx_idx]; | |
2962 | ||
2963 | if (set->ops->init_hctx && | |
2964 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2965 | goto unregister_cpu_notifier; | |
08e98fc6 | 2966 | |
7c6c5b7c ML |
2967 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2968 | hctx->numa_node)) | |
2969 | goto exit_hctx; | |
2970 | return 0; | |
2971 | ||
2972 | exit_hctx: | |
2973 | if (set->ops->exit_hctx) | |
2974 | set->ops->exit_hctx(hctx, hctx_idx); | |
2975 | unregister_cpu_notifier: | |
2976 | blk_mq_remove_cpuhp(hctx); | |
2977 | return -1; | |
2978 | } | |
2979 | ||
2980 | static struct blk_mq_hw_ctx * | |
2981 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2982 | int node) | |
2983 | { | |
2984 | struct blk_mq_hw_ctx *hctx; | |
2985 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2986 | ||
2987 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2988 | if (!hctx) | |
2989 | goto fail_alloc_hctx; | |
2990 | ||
2991 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2992 | goto free_hctx; | |
2993 | ||
2994 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2995 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2996 | node = set->numa_node; |
2997 | hctx->numa_node = node; | |
08e98fc6 | 2998 | |
9f993737 | 2999 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
3000 | spin_lock_init(&hctx->lock); |
3001 | INIT_LIST_HEAD(&hctx->dispatch); | |
3002 | hctx->queue = q; | |
51db1c37 | 3003 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 3004 | |
2f8f1336 ML |
3005 | INIT_LIST_HEAD(&hctx->hctx_list); |
3006 | ||
320ae51f | 3007 | /* |
08e98fc6 ML |
3008 | * Allocate space for all possible cpus to avoid allocation at |
3009 | * runtime | |
320ae51f | 3010 | */ |
d904bfa7 | 3011 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 3012 | gfp, node); |
08e98fc6 | 3013 | if (!hctx->ctxs) |
7c6c5b7c | 3014 | goto free_cpumask; |
320ae51f | 3015 | |
5b202853 | 3016 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 3017 | gfp, node, false, false)) |
08e98fc6 | 3018 | goto free_ctxs; |
08e98fc6 | 3019 | hctx->nr_ctx = 0; |
320ae51f | 3020 | |
5815839b | 3021 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
3022 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
3023 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
3024 | ||
754a1572 | 3025 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 3026 | if (!hctx->fq) |
7c6c5b7c | 3027 | goto free_bitmap; |
320ae51f | 3028 | |
6a83e74d | 3029 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 3030 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 3031 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 3032 | |
7c6c5b7c | 3033 | return hctx; |
320ae51f | 3034 | |
08e98fc6 | 3035 | free_bitmap: |
88459642 | 3036 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
3037 | free_ctxs: |
3038 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
3039 | free_cpumask: |
3040 | free_cpumask_var(hctx->cpumask); | |
3041 | free_hctx: | |
3042 | kfree(hctx); | |
3043 | fail_alloc_hctx: | |
3044 | return NULL; | |
08e98fc6 | 3045 | } |
320ae51f | 3046 | |
320ae51f JA |
3047 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
3048 | unsigned int nr_hw_queues) | |
3049 | { | |
b3c661b1 JA |
3050 | struct blk_mq_tag_set *set = q->tag_set; |
3051 | unsigned int i, j; | |
320ae51f JA |
3052 | |
3053 | for_each_possible_cpu(i) { | |
3054 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
3055 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 3056 | int k; |
320ae51f | 3057 | |
320ae51f JA |
3058 | __ctx->cpu = i; |
3059 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
3060 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
3061 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
3062 | ||
320ae51f JA |
3063 | __ctx->queue = q; |
3064 | ||
320ae51f JA |
3065 | /* |
3066 | * Set local node, IFF we have more than one hw queue. If | |
3067 | * not, we remain on the home node of the device | |
3068 | */ | |
b3c661b1 JA |
3069 | for (j = 0; j < set->nr_maps; j++) { |
3070 | hctx = blk_mq_map_queue_type(q, j, i); | |
3071 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 3072 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 3073 | } |
320ae51f JA |
3074 | } |
3075 | } | |
3076 | ||
63064be1 JG |
3077 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
3078 | unsigned int hctx_idx, | |
3079 | unsigned int depth) | |
cc71a6f4 | 3080 | { |
63064be1 JG |
3081 | struct blk_mq_tags *tags; |
3082 | int ret; | |
cc71a6f4 | 3083 | |
e155b0c2 | 3084 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
3085 | if (!tags) |
3086 | return NULL; | |
cc71a6f4 | 3087 | |
63064be1 JG |
3088 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
3089 | if (ret) { | |
e155b0c2 | 3090 | blk_mq_free_rq_map(tags); |
63064be1 JG |
3091 | return NULL; |
3092 | } | |
cc71a6f4 | 3093 | |
63064be1 JG |
3094 | return tags; |
3095 | } | |
3096 | ||
3097 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, | |
3098 | int hctx_idx) | |
3099 | { | |
079a2e3e JG |
3100 | if (blk_mq_is_shared_tags(set->flags)) { |
3101 | set->tags[hctx_idx] = set->shared_tags; | |
e155b0c2 JG |
3102 | |
3103 | return true; | |
3104 | } | |
3105 | ||
63064be1 JG |
3106 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
3107 | set->queue_depth); | |
3108 | ||
3109 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
3110 | } |
3111 | ||
645db34e JG |
3112 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3113 | struct blk_mq_tags *tags, | |
3114 | unsigned int hctx_idx) | |
cc71a6f4 | 3115 | { |
645db34e JG |
3116 | if (tags) { |
3117 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 3118 | blk_mq_free_rq_map(tags); |
bd166ef1 | 3119 | } |
cc71a6f4 JA |
3120 | } |
3121 | ||
e155b0c2 JG |
3122 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
3123 | unsigned int hctx_idx) | |
3124 | { | |
079a2e3e | 3125 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
3126 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
3127 | ||
3128 | set->tags[hctx_idx] = NULL; | |
3129 | } | |
3130 | ||
4b855ad3 | 3131 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 3132 | { |
b3c661b1 | 3133 | unsigned int i, j, hctx_idx; |
320ae51f JA |
3134 | struct blk_mq_hw_ctx *hctx; |
3135 | struct blk_mq_ctx *ctx; | |
2a34c087 | 3136 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
3137 | |
3138 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 3139 | cpumask_clear(hctx->cpumask); |
320ae51f | 3140 | hctx->nr_ctx = 0; |
d416c92c | 3141 | hctx->dispatch_from = NULL; |
320ae51f JA |
3142 | } |
3143 | ||
3144 | /* | |
4b855ad3 | 3145 | * Map software to hardware queues. |
4412efec ML |
3146 | * |
3147 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 3148 | */ |
20e4d813 | 3149 | for_each_possible_cpu(i) { |
4412efec | 3150 | |
897bb0c7 | 3151 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 3152 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
3153 | if (!set->map[j].nr_queues) { |
3154 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3155 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 3156 | continue; |
bb94aea1 | 3157 | } |
fd689871 ML |
3158 | hctx_idx = set->map[j].mq_map[i]; |
3159 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
3160 | if (!set->tags[hctx_idx] && | |
63064be1 | 3161 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
3162 | /* |
3163 | * If tags initialization fail for some hctx, | |
3164 | * that hctx won't be brought online. In this | |
3165 | * case, remap the current ctx to hctx[0] which | |
3166 | * is guaranteed to always have tags allocated | |
3167 | */ | |
3168 | set->map[j].mq_map[i] = 0; | |
3169 | } | |
e5edd5f2 | 3170 | |
b3c661b1 | 3171 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3172 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3173 | /* |
3174 | * If the CPU is already set in the mask, then we've | |
3175 | * mapped this one already. This can happen if | |
3176 | * devices share queues across queue maps. | |
3177 | */ | |
3178 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3179 | continue; | |
3180 | ||
3181 | cpumask_set_cpu(i, hctx->cpumask); | |
3182 | hctx->type = j; | |
3183 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3184 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3185 | ||
3186 | /* | |
3187 | * If the nr_ctx type overflows, we have exceeded the | |
3188 | * amount of sw queues we can support. | |
3189 | */ | |
3190 | BUG_ON(!hctx->nr_ctx); | |
3191 | } | |
bb94aea1 JW |
3192 | |
3193 | for (; j < HCTX_MAX_TYPES; j++) | |
3194 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3195 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3196 | } |
506e931f JA |
3197 | |
3198 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3199 | /* |
3200 | * If no software queues are mapped to this hardware queue, | |
3201 | * disable it and free the request entries. | |
3202 | */ | |
3203 | if (!hctx->nr_ctx) { | |
3204 | /* Never unmap queue 0. We need it as a | |
3205 | * fallback in case of a new remap fails | |
3206 | * allocation | |
3207 | */ | |
e155b0c2 JG |
3208 | if (i) |
3209 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3210 | |
3211 | hctx->tags = NULL; | |
3212 | continue; | |
3213 | } | |
484b4061 | 3214 | |
2a34c087 ML |
3215 | hctx->tags = set->tags[i]; |
3216 | WARN_ON(!hctx->tags); | |
3217 | ||
889fa31f CY |
3218 | /* |
3219 | * Set the map size to the number of mapped software queues. | |
3220 | * This is more accurate and more efficient than looping | |
3221 | * over all possibly mapped software queues. | |
3222 | */ | |
88459642 | 3223 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3224 | |
484b4061 JA |
3225 | /* |
3226 | * Initialize batch roundrobin counts | |
3227 | */ | |
f82ddf19 | 3228 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3229 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3230 | } | |
320ae51f JA |
3231 | } |
3232 | ||
8e8320c9 JA |
3233 | /* |
3234 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3235 | * the queue isn't live yet. | |
3236 | */ | |
2404e607 | 3237 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3238 | { |
3239 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
3240 | int i; |
3241 | ||
2404e607 | 3242 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3243 | if (shared) { |
51db1c37 | 3244 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3245 | } else { |
3246 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3247 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3248 | } |
2404e607 JM |
3249 | } |
3250 | } | |
3251 | ||
655ac300 HR |
3252 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3253 | bool shared) | |
2404e607 JM |
3254 | { |
3255 | struct request_queue *q; | |
0d2602ca | 3256 | |
705cda97 BVA |
3257 | lockdep_assert_held(&set->tag_list_lock); |
3258 | ||
0d2602ca JA |
3259 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3260 | blk_mq_freeze_queue(q); | |
2404e607 | 3261 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3262 | blk_mq_unfreeze_queue(q); |
3263 | } | |
3264 | } | |
3265 | ||
3266 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3267 | { | |
3268 | struct blk_mq_tag_set *set = q->tag_set; | |
3269 | ||
0d2602ca | 3270 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3271 | list_del(&q->tag_set_list); |
2404e607 JM |
3272 | if (list_is_singular(&set->tag_list)) { |
3273 | /* just transitioned to unshared */ | |
51db1c37 | 3274 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3275 | /* update existing queue */ |
655ac300 | 3276 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3277 | } |
0d2602ca | 3278 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3279 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3280 | } |
3281 | ||
3282 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3283 | struct request_queue *q) | |
3284 | { | |
0d2602ca | 3285 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3286 | |
ff821d27 JA |
3287 | /* |
3288 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3289 | */ | |
3290 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3291 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3292 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3293 | /* update existing queue */ |
655ac300 | 3294 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3295 | } |
51db1c37 | 3296 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3297 | queue_set_hctx_shared(q, true); |
08c875cb | 3298 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3299 | |
0d2602ca JA |
3300 | mutex_unlock(&set->tag_list_lock); |
3301 | } | |
3302 | ||
1db4909e ML |
3303 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3304 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3305 | { | |
3306 | struct blk_mq_ctxs *ctxs; | |
3307 | int cpu; | |
3308 | ||
3309 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3310 | if (!ctxs) | |
3311 | return -ENOMEM; | |
3312 | ||
3313 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3314 | if (!ctxs->queue_ctx) | |
3315 | goto fail; | |
3316 | ||
3317 | for_each_possible_cpu(cpu) { | |
3318 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3319 | ctx->ctxs = ctxs; | |
3320 | } | |
3321 | ||
3322 | q->mq_kobj = &ctxs->kobj; | |
3323 | q->queue_ctx = ctxs->queue_ctx; | |
3324 | ||
3325 | return 0; | |
3326 | fail: | |
3327 | kfree(ctxs); | |
3328 | return -ENOMEM; | |
3329 | } | |
3330 | ||
e09aae7e ML |
3331 | /* |
3332 | * It is the actual release handler for mq, but we do it from | |
3333 | * request queue's release handler for avoiding use-after-free | |
3334 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3335 | * but we can't group ctx/kctx kobj without it. | |
3336 | */ | |
3337 | void blk_mq_release(struct request_queue *q) | |
3338 | { | |
2f8f1336 ML |
3339 | struct blk_mq_hw_ctx *hctx, *next; |
3340 | int i; | |
e09aae7e | 3341 | |
2f8f1336 ML |
3342 | queue_for_each_hw_ctx(q, hctx, i) |
3343 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3344 | ||
3345 | /* all hctx are in .unused_hctx_list now */ | |
3346 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3347 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3348 | kobject_put(&hctx->kobj); |
c3b4afca | 3349 | } |
e09aae7e ML |
3350 | |
3351 | kfree(q->queue_hw_ctx); | |
3352 | ||
7ea5fe31 ML |
3353 | /* |
3354 | * release .mq_kobj and sw queue's kobject now because | |
3355 | * both share lifetime with request queue. | |
3356 | */ | |
3357 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3358 | } |
3359 | ||
5ec780a6 | 3360 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3361 | void *queuedata) |
b62c21b7 | 3362 | { |
26a9750a CH |
3363 | struct request_queue *q; |
3364 | int ret; | |
b62c21b7 | 3365 | |
26a9750a CH |
3366 | q = blk_alloc_queue(set->numa_node); |
3367 | if (!q) | |
b62c21b7 | 3368 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3369 | q->queuedata = queuedata; |
3370 | ret = blk_mq_init_allocated_queue(set, q); | |
3371 | if (ret) { | |
3372 | blk_cleanup_queue(q); | |
3373 | return ERR_PTR(ret); | |
3374 | } | |
b62c21b7 MS |
3375 | return q; |
3376 | } | |
2f227bb9 CH |
3377 | |
3378 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3379 | { | |
3380 | return blk_mq_init_queue_data(set, NULL); | |
3381 | } | |
b62c21b7 MS |
3382 | EXPORT_SYMBOL(blk_mq_init_queue); |
3383 | ||
4dcc4874 CH |
3384 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3385 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3386 | { |
3387 | struct request_queue *q; | |
b461dfc4 | 3388 | struct gendisk *disk; |
9316a9ed | 3389 | |
b461dfc4 CH |
3390 | q = blk_mq_init_queue_data(set, queuedata); |
3391 | if (IS_ERR(q)) | |
3392 | return ERR_CAST(q); | |
9316a9ed | 3393 | |
4a1fa41d | 3394 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3395 | if (!disk) { |
3396 | blk_cleanup_queue(q); | |
3397 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3398 | } |
b461dfc4 | 3399 | return disk; |
9316a9ed | 3400 | } |
b461dfc4 | 3401 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3402 | |
34d11ffa JW |
3403 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3404 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3405 | int hctx_idx, int node) | |
3406 | { | |
2f8f1336 | 3407 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3408 | |
2f8f1336 ML |
3409 | /* reuse dead hctx first */ |
3410 | spin_lock(&q->unused_hctx_lock); | |
3411 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3412 | if (tmp->numa_node == node) { | |
3413 | hctx = tmp; | |
3414 | break; | |
3415 | } | |
3416 | } | |
3417 | if (hctx) | |
3418 | list_del_init(&hctx->hctx_list); | |
3419 | spin_unlock(&q->unused_hctx_lock); | |
3420 | ||
3421 | if (!hctx) | |
3422 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3423 | if (!hctx) |
7c6c5b7c | 3424 | goto fail; |
34d11ffa | 3425 | |
7c6c5b7c ML |
3426 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3427 | goto free_hctx; | |
34d11ffa JW |
3428 | |
3429 | return hctx; | |
7c6c5b7c ML |
3430 | |
3431 | free_hctx: | |
3432 | kobject_put(&hctx->kobj); | |
3433 | fail: | |
3434 | return NULL; | |
34d11ffa JW |
3435 | } |
3436 | ||
868f2f0b KB |
3437 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3438 | struct request_queue *q) | |
320ae51f | 3439 | { |
e01ad46d | 3440 | int i, j, end; |
868f2f0b | 3441 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3442 | |
ac0d6b92 BVA |
3443 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3444 | struct blk_mq_hw_ctx **new_hctxs; | |
3445 | ||
3446 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3447 | sizeof(*new_hctxs), GFP_KERNEL, | |
3448 | set->numa_node); | |
3449 | if (!new_hctxs) | |
3450 | return; | |
3451 | if (hctxs) | |
3452 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3453 | sizeof(*hctxs)); | |
3454 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3455 | kfree(hctxs); |
3456 | hctxs = new_hctxs; | |
3457 | } | |
3458 | ||
fb350e0a ML |
3459 | /* protect against switching io scheduler */ |
3460 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3461 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3462 | int node; |
34d11ffa | 3463 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3464 | |
7d76f856 | 3465 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3466 | /* |
3467 | * If the hw queue has been mapped to another numa node, | |
3468 | * we need to realloc the hctx. If allocation fails, fallback | |
3469 | * to use the previous one. | |
3470 | */ | |
3471 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3472 | continue; | |
868f2f0b | 3473 | |
34d11ffa JW |
3474 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3475 | if (hctx) { | |
2f8f1336 | 3476 | if (hctxs[i]) |
34d11ffa | 3477 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3478 | hctxs[i] = hctx; |
3479 | } else { | |
3480 | if (hctxs[i]) | |
3481 | pr_warn("Allocate new hctx on node %d fails,\ | |
3482 | fallback to previous one on node %d\n", | |
3483 | node, hctxs[i]->numa_node); | |
3484 | else | |
3485 | break; | |
868f2f0b | 3486 | } |
320ae51f | 3487 | } |
e01ad46d JW |
3488 | /* |
3489 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3490 | * hctxs and keep the previous q->nr_hw_queues. | |
3491 | */ | |
3492 | if (i != set->nr_hw_queues) { | |
3493 | j = q->nr_hw_queues; | |
3494 | end = i; | |
3495 | } else { | |
3496 | j = i; | |
3497 | end = q->nr_hw_queues; | |
3498 | q->nr_hw_queues = set->nr_hw_queues; | |
3499 | } | |
34d11ffa | 3500 | |
e01ad46d | 3501 | for (; j < end; j++) { |
868f2f0b KB |
3502 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3503 | ||
3504 | if (hctx) { | |
e155b0c2 | 3505 | __blk_mq_free_map_and_rqs(set, j); |
868f2f0b | 3506 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3507 | hctxs[j] = NULL; |
868f2f0b KB |
3508 | } |
3509 | } | |
fb350e0a | 3510 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3511 | } |
3512 | ||
26a9750a CH |
3513 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
3514 | struct request_queue *q) | |
868f2f0b | 3515 | { |
66841672 ML |
3516 | /* mark the queue as mq asap */ |
3517 | q->mq_ops = set->ops; | |
3518 | ||
34dbad5d | 3519 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3520 | blk_mq_poll_stats_bkt, |
3521 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3522 | if (!q->poll_cb) |
3523 | goto err_exit; | |
3524 | ||
1db4909e | 3525 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3526 | goto err_poll; |
868f2f0b | 3527 | |
737f98cf ML |
3528 | /* init q->mq_kobj and sw queues' kobjects */ |
3529 | blk_mq_sysfs_init(q); | |
3530 | ||
2f8f1336 ML |
3531 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3532 | spin_lock_init(&q->unused_hctx_lock); | |
3533 | ||
868f2f0b KB |
3534 | blk_mq_realloc_hw_ctxs(set, q); |
3535 | if (!q->nr_hw_queues) | |
3536 | goto err_hctxs; | |
320ae51f | 3537 | |
287922eb | 3538 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3539 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3540 | |
a8908939 | 3541 | q->tag_set = set; |
320ae51f | 3542 | |
94eddfbe | 3543 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3544 | if (set->nr_maps > HCTX_TYPE_POLL && |
3545 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3546 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3547 | |
2849450a | 3548 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3549 | INIT_LIST_HEAD(&q->requeue_list); |
3550 | spin_lock_init(&q->requeue_lock); | |
3551 | ||
eba71768 JA |
3552 | q->nr_requests = set->queue_depth; |
3553 | ||
64f1c21e JA |
3554 | /* |
3555 | * Default to classic polling | |
3556 | */ | |
29ece8b4 | 3557 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3558 | |
24d2f903 | 3559 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3560 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3561 | blk_mq_map_swqueue(q); |
26a9750a | 3562 | return 0; |
18741986 | 3563 | |
320ae51f | 3564 | err_hctxs: |
868f2f0b | 3565 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3566 | q->nr_hw_queues = 0; |
1db4909e | 3567 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3568 | err_poll: |
3569 | blk_stat_free_callback(q->poll_cb); | |
3570 | q->poll_cb = NULL; | |
c7de5726 ML |
3571 | err_exit: |
3572 | q->mq_ops = NULL; | |
26a9750a | 3573 | return -ENOMEM; |
320ae51f | 3574 | } |
b62c21b7 | 3575 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3576 | |
c7e2d94b ML |
3577 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3578 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3579 | { |
630ef623 | 3580 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3581 | |
630ef623 | 3582 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 3583 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
3584 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
3585 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 3586 | } |
320ae51f | 3587 | |
a5164405 JA |
3588 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3589 | { | |
3590 | int i; | |
3591 | ||
079a2e3e JG |
3592 | if (blk_mq_is_shared_tags(set->flags)) { |
3593 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
3594 | BLK_MQ_NO_HCTX_IDX, |
3595 | set->queue_depth); | |
079a2e3e | 3596 | if (!set->shared_tags) |
e155b0c2 JG |
3597 | return -ENOMEM; |
3598 | } | |
3599 | ||
8229cca8 | 3600 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 3601 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 3602 | goto out_unwind; |
8229cca8 XT |
3603 | cond_resched(); |
3604 | } | |
a5164405 JA |
3605 | |
3606 | return 0; | |
3607 | ||
3608 | out_unwind: | |
e155b0c2 JG |
3609 | while (--i >= 0) |
3610 | __blk_mq_free_map_and_rqs(set, i); | |
3611 | ||
079a2e3e JG |
3612 | if (blk_mq_is_shared_tags(set->flags)) { |
3613 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 3614 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 3615 | } |
a5164405 | 3616 | |
a5164405 JA |
3617 | return -ENOMEM; |
3618 | } | |
3619 | ||
3620 | /* | |
3621 | * Allocate the request maps associated with this tag_set. Note that this | |
3622 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3623 | * will be updated to reflect the allocated depth. | |
3624 | */ | |
63064be1 | 3625 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
3626 | { |
3627 | unsigned int depth; | |
3628 | int err; | |
3629 | ||
3630 | depth = set->queue_depth; | |
3631 | do { | |
3632 | err = __blk_mq_alloc_rq_maps(set); | |
3633 | if (!err) | |
3634 | break; | |
3635 | ||
3636 | set->queue_depth >>= 1; | |
3637 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3638 | err = -ENOMEM; | |
3639 | break; | |
3640 | } | |
3641 | } while (set->queue_depth); | |
3642 | ||
3643 | if (!set->queue_depth || err) { | |
3644 | pr_err("blk-mq: failed to allocate request map\n"); | |
3645 | return -ENOMEM; | |
3646 | } | |
3647 | ||
3648 | if (depth != set->queue_depth) | |
3649 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3650 | depth, set->queue_depth); | |
3651 | ||
3652 | return 0; | |
3653 | } | |
3654 | ||
ebe8bddb OS |
3655 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3656 | { | |
6e66b493 BVA |
3657 | /* |
3658 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3659 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3660 | * number of hardware queues. | |
3661 | */ | |
3662 | if (set->nr_maps == 1) | |
3663 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3664 | ||
59388702 | 3665 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3666 | int i; |
3667 | ||
7d4901a9 ML |
3668 | /* |
3669 | * transport .map_queues is usually done in the following | |
3670 | * way: | |
3671 | * | |
3672 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3673 | * mask = get_cpu_mask(queue) | |
3674 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3675 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3676 | * } |
3677 | * | |
3678 | * When we need to remap, the table has to be cleared for | |
3679 | * killing stale mapping since one CPU may not be mapped | |
3680 | * to any hw queue. | |
3681 | */ | |
b3c661b1 JA |
3682 | for (i = 0; i < set->nr_maps; i++) |
3683 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3684 | |
ebe8bddb | 3685 | return set->ops->map_queues(set); |
b3c661b1 JA |
3686 | } else { |
3687 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3688 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3689 | } |
ebe8bddb OS |
3690 | } |
3691 | ||
f7e76dbc BVA |
3692 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3693 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3694 | { | |
3695 | struct blk_mq_tags **new_tags; | |
3696 | ||
3697 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3698 | return 0; | |
3699 | ||
3700 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3701 | GFP_KERNEL, set->numa_node); | |
3702 | if (!new_tags) | |
3703 | return -ENOMEM; | |
3704 | ||
3705 | if (set->tags) | |
3706 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3707 | sizeof(*set->tags)); | |
3708 | kfree(set->tags); | |
3709 | set->tags = new_tags; | |
3710 | set->nr_hw_queues = new_nr_hw_queues; | |
3711 | ||
3712 | return 0; | |
3713 | } | |
3714 | ||
91cdf265 MI |
3715 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3716 | int new_nr_hw_queues) | |
3717 | { | |
3718 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3719 | } | |
3720 | ||
a4391c64 JA |
3721 | /* |
3722 | * Alloc a tag set to be associated with one or more request queues. | |
3723 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3724 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3725 | * value will be stored in set->queue_depth. |
3726 | */ | |
24d2f903 CH |
3727 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3728 | { | |
b3c661b1 | 3729 | int i, ret; |
da695ba2 | 3730 | |
205fb5f5 BVA |
3731 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3732 | ||
24d2f903 CH |
3733 | if (!set->nr_hw_queues) |
3734 | return -EINVAL; | |
a4391c64 | 3735 | if (!set->queue_depth) |
24d2f903 CH |
3736 | return -EINVAL; |
3737 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3738 | return -EINVAL; | |
3739 | ||
7d7e0f90 | 3740 | if (!set->ops->queue_rq) |
24d2f903 CH |
3741 | return -EINVAL; |
3742 | ||
de148297 ML |
3743 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3744 | return -EINVAL; | |
3745 | ||
a4391c64 JA |
3746 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3747 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3748 | BLK_MQ_MAX_DEPTH); | |
3749 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3750 | } | |
24d2f903 | 3751 | |
b3c661b1 JA |
3752 | if (!set->nr_maps) |
3753 | set->nr_maps = 1; | |
3754 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3755 | return -EINVAL; | |
3756 | ||
6637fadf SL |
3757 | /* |
3758 | * If a crashdump is active, then we are potentially in a very | |
3759 | * memory constrained environment. Limit us to 1 queue and | |
3760 | * 64 tags to prevent using too much memory. | |
3761 | */ | |
3762 | if (is_kdump_kernel()) { | |
3763 | set->nr_hw_queues = 1; | |
59388702 | 3764 | set->nr_maps = 1; |
6637fadf SL |
3765 | set->queue_depth = min(64U, set->queue_depth); |
3766 | } | |
868f2f0b | 3767 | /* |
392546ae JA |
3768 | * There is no use for more h/w queues than cpus if we just have |
3769 | * a single map | |
868f2f0b | 3770 | */ |
392546ae | 3771 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3772 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3773 | |
91cdf265 | 3774 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3775 | return -ENOMEM; |
24d2f903 | 3776 | |
da695ba2 | 3777 | ret = -ENOMEM; |
b3c661b1 JA |
3778 | for (i = 0; i < set->nr_maps; i++) { |
3779 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3780 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3781 | GFP_KERNEL, set->numa_node); |
3782 | if (!set->map[i].mq_map) | |
3783 | goto out_free_mq_map; | |
59388702 | 3784 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3785 | } |
bdd17e75 | 3786 | |
ebe8bddb | 3787 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3788 | if (ret) |
3789 | goto out_free_mq_map; | |
3790 | ||
63064be1 | 3791 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 3792 | if (ret) |
bdd17e75 | 3793 | goto out_free_mq_map; |
24d2f903 | 3794 | |
0d2602ca JA |
3795 | mutex_init(&set->tag_list_lock); |
3796 | INIT_LIST_HEAD(&set->tag_list); | |
3797 | ||
24d2f903 | 3798 | return 0; |
bdd17e75 CH |
3799 | |
3800 | out_free_mq_map: | |
b3c661b1 JA |
3801 | for (i = 0; i < set->nr_maps; i++) { |
3802 | kfree(set->map[i].mq_map); | |
3803 | set->map[i].mq_map = NULL; | |
3804 | } | |
5676e7b6 RE |
3805 | kfree(set->tags); |
3806 | set->tags = NULL; | |
da695ba2 | 3807 | return ret; |
24d2f903 CH |
3808 | } |
3809 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3810 | ||
cdb14e0f CH |
3811 | /* allocate and initialize a tagset for a simple single-queue device */ |
3812 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
3813 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
3814 | unsigned int set_flags) | |
3815 | { | |
3816 | memset(set, 0, sizeof(*set)); | |
3817 | set->ops = ops; | |
3818 | set->nr_hw_queues = 1; | |
3819 | set->nr_maps = 1; | |
3820 | set->queue_depth = queue_depth; | |
3821 | set->numa_node = NUMA_NO_NODE; | |
3822 | set->flags = set_flags; | |
3823 | return blk_mq_alloc_tag_set(set); | |
3824 | } | |
3825 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
3826 | ||
24d2f903 CH |
3827 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
3828 | { | |
b3c661b1 | 3829 | int i, j; |
24d2f903 | 3830 | |
e155b0c2 JG |
3831 | for (i = 0; i < set->nr_hw_queues; i++) |
3832 | __blk_mq_free_map_and_rqs(set, i); | |
484b4061 | 3833 | |
079a2e3e JG |
3834 | if (blk_mq_is_shared_tags(set->flags)) { |
3835 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
3836 | BLK_MQ_NO_HCTX_IDX); |
3837 | } | |
32bc15af | 3838 | |
b3c661b1 JA |
3839 | for (j = 0; j < set->nr_maps; j++) { |
3840 | kfree(set->map[j].mq_map); | |
3841 | set->map[j].mq_map = NULL; | |
3842 | } | |
bdd17e75 | 3843 | |
981bd189 | 3844 | kfree(set->tags); |
5676e7b6 | 3845 | set->tags = NULL; |
24d2f903 CH |
3846 | } |
3847 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3848 | ||
e3a2b3f9 JA |
3849 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3850 | { | |
3851 | struct blk_mq_tag_set *set = q->tag_set; | |
3852 | struct blk_mq_hw_ctx *hctx; | |
3853 | int i, ret; | |
3854 | ||
bd166ef1 | 3855 | if (!set) |
e3a2b3f9 JA |
3856 | return -EINVAL; |
3857 | ||
e5fa8140 AZ |
3858 | if (q->nr_requests == nr) |
3859 | return 0; | |
3860 | ||
70f36b60 | 3861 | blk_mq_freeze_queue(q); |
24f5a90f | 3862 | blk_mq_quiesce_queue(q); |
70f36b60 | 3863 | |
e3a2b3f9 JA |
3864 | ret = 0; |
3865 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3866 | if (!hctx->tags) |
3867 | continue; | |
bd166ef1 JA |
3868 | /* |
3869 | * If we're using an MQ scheduler, just update the scheduler | |
3870 | * queue depth. This is similar to what the old code would do. | |
3871 | */ | |
f6adcef5 | 3872 | if (hctx->sched_tags) { |
70f36b60 | 3873 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 3874 | nr, true); |
f6adcef5 JG |
3875 | } else { |
3876 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
3877 | false); | |
70f36b60 | 3878 | } |
e3a2b3f9 JA |
3879 | if (ret) |
3880 | break; | |
77f1e0a5 JA |
3881 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3882 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 3883 | } |
d97e594c | 3884 | if (!ret) { |
e3a2b3f9 | 3885 | q->nr_requests = nr; |
079a2e3e | 3886 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 3887 | if (q->elevator) |
079a2e3e | 3888 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 3889 | else |
079a2e3e | 3890 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 3891 | } |
d97e594c | 3892 | } |
e3a2b3f9 | 3893 | |
24f5a90f | 3894 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3895 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3896 | |
e3a2b3f9 JA |
3897 | return ret; |
3898 | } | |
3899 | ||
d48ece20 JW |
3900 | /* |
3901 | * request_queue and elevator_type pair. | |
3902 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3903 | * the elevator_type associated with a request_queue. | |
3904 | */ | |
3905 | struct blk_mq_qe_pair { | |
3906 | struct list_head node; | |
3907 | struct request_queue *q; | |
3908 | struct elevator_type *type; | |
3909 | }; | |
3910 | ||
3911 | /* | |
3912 | * Cache the elevator_type in qe pair list and switch the | |
3913 | * io scheduler to 'none' | |
3914 | */ | |
3915 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3916 | struct request_queue *q) | |
3917 | { | |
3918 | struct blk_mq_qe_pair *qe; | |
3919 | ||
3920 | if (!q->elevator) | |
3921 | return true; | |
3922 | ||
3923 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3924 | if (!qe) | |
3925 | return false; | |
3926 | ||
3927 | INIT_LIST_HEAD(&qe->node); | |
3928 | qe->q = q; | |
3929 | qe->type = q->elevator->type; | |
3930 | list_add(&qe->node, head); | |
3931 | ||
3932 | mutex_lock(&q->sysfs_lock); | |
3933 | /* | |
3934 | * After elevator_switch_mq, the previous elevator_queue will be | |
3935 | * released by elevator_release. The reference of the io scheduler | |
3936 | * module get by elevator_get will also be put. So we need to get | |
3937 | * a reference of the io scheduler module here to prevent it to be | |
3938 | * removed. | |
3939 | */ | |
3940 | __module_get(qe->type->elevator_owner); | |
3941 | elevator_switch_mq(q, NULL); | |
3942 | mutex_unlock(&q->sysfs_lock); | |
3943 | ||
3944 | return true; | |
3945 | } | |
3946 | ||
3947 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3948 | struct request_queue *q) | |
3949 | { | |
3950 | struct blk_mq_qe_pair *qe; | |
3951 | struct elevator_type *t = NULL; | |
3952 | ||
3953 | list_for_each_entry(qe, head, node) | |
3954 | if (qe->q == q) { | |
3955 | t = qe->type; | |
3956 | break; | |
3957 | } | |
3958 | ||
3959 | if (!t) | |
3960 | return; | |
3961 | ||
3962 | list_del(&qe->node); | |
3963 | kfree(qe); | |
3964 | ||
3965 | mutex_lock(&q->sysfs_lock); | |
3966 | elevator_switch_mq(q, t); | |
3967 | mutex_unlock(&q->sysfs_lock); | |
3968 | } | |
3969 | ||
e4dc2b32 KB |
3970 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3971 | int nr_hw_queues) | |
868f2f0b KB |
3972 | { |
3973 | struct request_queue *q; | |
d48ece20 | 3974 | LIST_HEAD(head); |
e01ad46d | 3975 | int prev_nr_hw_queues; |
868f2f0b | 3976 | |
705cda97 BVA |
3977 | lockdep_assert_held(&set->tag_list_lock); |
3978 | ||
392546ae | 3979 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3980 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3981 | if (nr_hw_queues < 1) |
3982 | return; | |
3983 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3984 | return; |
3985 | ||
3986 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3987 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3988 | /* |
3989 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3990 | * with the previous scheduler. We will switch back once we are done | |
3991 | * updating the new sw to hw queue mappings. | |
3992 | */ | |
3993 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3994 | if (!blk_mq_elv_switch_none(&head, q)) | |
3995 | goto switch_back; | |
868f2f0b | 3996 | |
477e19de JW |
3997 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3998 | blk_mq_debugfs_unregister_hctxs(q); | |
3999 | blk_mq_sysfs_unregister(q); | |
4000 | } | |
4001 | ||
a2584e43 | 4002 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
4003 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
4004 | 0) | |
4005 | goto reregister; | |
4006 | ||
868f2f0b | 4007 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 4008 | fallback: |
aa880ad6 | 4009 | blk_mq_update_queue_map(set); |
868f2f0b KB |
4010 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4011 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
4012 | if (q->nr_hw_queues != set->nr_hw_queues) { |
4013 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
4014 | nr_hw_queues, prev_nr_hw_queues); | |
4015 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 4016 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
4017 | goto fallback; |
4018 | } | |
477e19de JW |
4019 | blk_mq_map_swqueue(q); |
4020 | } | |
4021 | ||
f7e76dbc | 4022 | reregister: |
477e19de JW |
4023 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
4024 | blk_mq_sysfs_register(q); | |
4025 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
4026 | } |
4027 | ||
d48ece20 JW |
4028 | switch_back: |
4029 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
4030 | blk_mq_elv_switch_back(&head, q); | |
4031 | ||
868f2f0b KB |
4032 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
4033 | blk_mq_unfreeze_queue(q); | |
4034 | } | |
e4dc2b32 KB |
4035 | |
4036 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
4037 | { | |
4038 | mutex_lock(&set->tag_list_lock); | |
4039 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
4040 | mutex_unlock(&set->tag_list_lock); | |
4041 | } | |
868f2f0b KB |
4042 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
4043 | ||
34dbad5d OS |
4044 | /* Enable polling stats and return whether they were already enabled. */ |
4045 | static bool blk_poll_stats_enable(struct request_queue *q) | |
4046 | { | |
4047 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 4048 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
4049 | return true; |
4050 | blk_stat_add_callback(q, q->poll_cb); | |
4051 | return false; | |
4052 | } | |
4053 | ||
4054 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
4055 | { | |
4056 | /* | |
4057 | * We don't arm the callback if polling stats are not enabled or the | |
4058 | * callback is already active. | |
4059 | */ | |
4060 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
4061 | blk_stat_is_active(q->poll_cb)) | |
4062 | return; | |
4063 | ||
4064 | blk_stat_activate_msecs(q->poll_cb, 100); | |
4065 | } | |
4066 | ||
4067 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
4068 | { | |
4069 | struct request_queue *q = cb->data; | |
720b8ccc | 4070 | int bucket; |
34dbad5d | 4071 | |
720b8ccc SB |
4072 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
4073 | if (cb->stat[bucket].nr_samples) | |
4074 | q->poll_stat[bucket] = cb->stat[bucket]; | |
4075 | } | |
34dbad5d OS |
4076 | } |
4077 | ||
64f1c21e | 4078 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
4079 | struct request *rq) |
4080 | { | |
64f1c21e | 4081 | unsigned long ret = 0; |
720b8ccc | 4082 | int bucket; |
64f1c21e JA |
4083 | |
4084 | /* | |
4085 | * If stats collection isn't on, don't sleep but turn it on for | |
4086 | * future users | |
4087 | */ | |
34dbad5d | 4088 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
4089 | return 0; |
4090 | ||
64f1c21e JA |
4091 | /* |
4092 | * As an optimistic guess, use half of the mean service time | |
4093 | * for this type of request. We can (and should) make this smarter. | |
4094 | * For instance, if the completion latencies are tight, we can | |
4095 | * get closer than just half the mean. This is especially | |
4096 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
4097 | * than ~10 usec. We do use the stats for the relevant IO size |
4098 | * if available which does lead to better estimates. | |
64f1c21e | 4099 | */ |
720b8ccc SB |
4100 | bucket = blk_mq_poll_stats_bkt(rq); |
4101 | if (bucket < 0) | |
4102 | return ret; | |
4103 | ||
4104 | if (q->poll_stat[bucket].nr_samples) | |
4105 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
4106 | |
4107 | return ret; | |
4108 | } | |
4109 | ||
c6699d6f | 4110 | static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc) |
06426adf | 4111 | { |
c6699d6f CH |
4112 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, qc); |
4113 | struct request *rq = blk_qc_to_rq(hctx, qc); | |
06426adf JA |
4114 | struct hrtimer_sleeper hs; |
4115 | enum hrtimer_mode mode; | |
64f1c21e | 4116 | unsigned int nsecs; |
06426adf JA |
4117 | ktime_t kt; |
4118 | ||
c6699d6f CH |
4119 | /* |
4120 | * If a request has completed on queue that uses an I/O scheduler, we | |
4121 | * won't get back a request from blk_qc_to_rq. | |
4122 | */ | |
4123 | if (!rq || (rq->rq_flags & RQF_MQ_POLL_SLEPT)) | |
64f1c21e JA |
4124 | return false; |
4125 | ||
4126 | /* | |
1052b8ac | 4127 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 4128 | * |
64f1c21e JA |
4129 | * 0: use half of prev avg |
4130 | * >0: use this specific value | |
4131 | */ | |
1052b8ac | 4132 | if (q->poll_nsec > 0) |
64f1c21e JA |
4133 | nsecs = q->poll_nsec; |
4134 | else | |
cae740a0 | 4135 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
4136 | |
4137 | if (!nsecs) | |
06426adf JA |
4138 | return false; |
4139 | ||
76a86f9d | 4140 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
4141 | |
4142 | /* | |
4143 | * This will be replaced with the stats tracking code, using | |
4144 | * 'avg_completion_time / 2' as the pre-sleep target. | |
4145 | */ | |
8b0e1953 | 4146 | kt = nsecs; |
06426adf JA |
4147 | |
4148 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 4149 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
4150 | hrtimer_set_expires(&hs.timer, kt); |
4151 | ||
06426adf | 4152 | do { |
5a61c363 | 4153 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
4154 | break; |
4155 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 4156 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
4157 | if (hs.task) |
4158 | io_schedule(); | |
4159 | hrtimer_cancel(&hs.timer); | |
4160 | mode = HRTIMER_MODE_ABS; | |
4161 | } while (hs.task && !signal_pending(current)); | |
4162 | ||
4163 | __set_current_state(TASK_RUNNING); | |
4164 | destroy_hrtimer_on_stack(&hs.timer); | |
c6699d6f CH |
4165 | |
4166 | /* | |
4167 | * If we sleep, have the caller restart the poll loop to reset the | |
4168 | * state. Like for the other success return cases, the caller is | |
4169 | * responsible for checking if the IO completed. If the IO isn't | |
4170 | * complete, we'll get called again and will go straight to the busy | |
4171 | * poll loop. | |
4172 | */ | |
06426adf JA |
4173 | return true; |
4174 | } | |
4175 | ||
c6699d6f | 4176 | static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie, |
ef99b2d3 | 4177 | unsigned int flags) |
bbd7bb70 | 4178 | { |
c6699d6f CH |
4179 | struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie); |
4180 | long state = get_current_state(); | |
4181 | int ret; | |
1052b8ac | 4182 | |
c6699d6f | 4183 | do { |
c6699d6f CH |
4184 | ret = q->mq_ops->poll(hctx); |
4185 | if (ret > 0) { | |
c6699d6f CH |
4186 | __set_current_state(TASK_RUNNING); |
4187 | return ret; | |
4188 | } | |
4189 | ||
4190 | if (signal_pending_state(state, current)) | |
4191 | __set_current_state(TASK_RUNNING); | |
4192 | if (task_is_running(current)) | |
4193 | return 1; | |
4194 | ||
ef99b2d3 | 4195 | if (ret < 0 || (flags & BLK_POLL_ONESHOT)) |
c6699d6f CH |
4196 | break; |
4197 | cpu_relax(); | |
4198 | } while (!need_resched()); | |
4199 | ||
4200 | __set_current_state(TASK_RUNNING); | |
4201 | return 0; | |
1052b8ac JA |
4202 | } |
4203 | ||
3e08773c | 4204 | int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, unsigned int flags) |
1052b8ac | 4205 | { |
d729cf9a CH |
4206 | if (!(flags & BLK_POLL_NOSLEEP) && |
4207 | q->poll_nsec != BLK_MQ_POLL_CLASSIC) { | |
c6699d6f | 4208 | if (blk_mq_poll_hybrid(q, cookie)) |
85f4d4b6 | 4209 | return 1; |
c6699d6f | 4210 | } |
ef99b2d3 | 4211 | return blk_mq_poll_classic(q, cookie, flags); |
bbd7bb70 JA |
4212 | } |
4213 | ||
9cf2bab6 JA |
4214 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4215 | { | |
4216 | return rq->mq_ctx->cpu; | |
4217 | } | |
4218 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4219 | ||
320ae51f JA |
4220 | static int __init blk_mq_init(void) |
4221 | { | |
c3077b5d CH |
4222 | int i; |
4223 | ||
4224 | for_each_possible_cpu(i) | |
f9ab4918 | 4225 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4226 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4227 | ||
4228 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4229 | "block/softirq:dead", NULL, | |
4230 | blk_softirq_cpu_dead); | |
9467f859 TG |
4231 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4232 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4233 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4234 | blk_mq_hctx_notify_online, | |
4235 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4236 | return 0; |
4237 | } | |
4238 | subsys_initcall(blk_mq_init); |