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 | ||
320ae51f | 68 | /* |
85fae294 YY |
69 | * Check if any of the ctx, dispatch list or elevator |
70 | * have pending work in this hardware queue. | |
320ae51f | 71 | */ |
79f720a7 | 72 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 73 | { |
79f720a7 JA |
74 | return !list_empty_careful(&hctx->dispatch) || |
75 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
bd166ef1 | 76 | blk_mq_sched_has_work(hctx); |
1429d7c9 JA |
77 | } |
78 | ||
320ae51f JA |
79 | /* |
80 | * Mark this ctx as having pending work in this hardware queue | |
81 | */ | |
82 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
83 | struct blk_mq_ctx *ctx) | |
84 | { | |
f31967f0 JA |
85 | const int bit = ctx->index_hw[hctx->type]; |
86 | ||
87 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
88 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
1429d7c9 JA |
89 | } |
90 | ||
91 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
92 | struct blk_mq_ctx *ctx) | |
93 | { | |
f31967f0 JA |
94 | const int bit = ctx->index_hw[hctx->type]; |
95 | ||
96 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
320ae51f JA |
97 | } |
98 | ||
f299b7c7 | 99 | struct mq_inflight { |
8446fe92 | 100 | struct block_device *part; |
a2e80f6f | 101 | unsigned int inflight[2]; |
f299b7c7 JA |
102 | }; |
103 | ||
7baa8572 | 104 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, |
f299b7c7 JA |
105 | struct request *rq, void *priv, |
106 | bool reserved) | |
107 | { | |
108 | struct mq_inflight *mi = priv; | |
109 | ||
b0d97557 JX |
110 | if ((!mi->part->bd_partno || rq->part == mi->part) && |
111 | blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) | |
bb4e6b14 | 112 | mi->inflight[rq_data_dir(rq)]++; |
7baa8572 JA |
113 | |
114 | return true; | |
f299b7c7 JA |
115 | } |
116 | ||
8446fe92 CH |
117 | unsigned int blk_mq_in_flight(struct request_queue *q, |
118 | struct block_device *part) | |
f299b7c7 | 119 | { |
a2e80f6f | 120 | struct mq_inflight mi = { .part = part }; |
f299b7c7 | 121 | |
f299b7c7 | 122 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
e016b782 | 123 | |
a2e80f6f | 124 | return mi.inflight[0] + mi.inflight[1]; |
bf0ddaba OS |
125 | } |
126 | ||
8446fe92 CH |
127 | void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, |
128 | unsigned int inflight[2]) | |
bf0ddaba | 129 | { |
a2e80f6f | 130 | struct mq_inflight mi = { .part = part }; |
bf0ddaba | 131 | |
bb4e6b14 | 132 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
a2e80f6f PB |
133 | inflight[0] = mi.inflight[0]; |
134 | inflight[1] = mi.inflight[1]; | |
bf0ddaba OS |
135 | } |
136 | ||
1671d522 | 137 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 138 | { |
7996a8b5 BL |
139 | mutex_lock(&q->mq_freeze_lock); |
140 | if (++q->mq_freeze_depth == 1) { | |
3ef28e83 | 141 | percpu_ref_kill(&q->q_usage_counter); |
7996a8b5 | 142 | mutex_unlock(&q->mq_freeze_lock); |
344e9ffc | 143 | if (queue_is_mq(q)) |
055f6e18 | 144 | blk_mq_run_hw_queues(q, false); |
7996a8b5 BL |
145 | } else { |
146 | mutex_unlock(&q->mq_freeze_lock); | |
cddd5d17 | 147 | } |
f3af020b | 148 | } |
1671d522 | 149 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 150 | |
6bae363e | 151 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 152 | { |
3ef28e83 | 153 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 154 | } |
6bae363e | 155 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 156 | |
f91328c4 KB |
157 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
158 | unsigned long timeout) | |
159 | { | |
160 | return wait_event_timeout(q->mq_freeze_wq, | |
161 | percpu_ref_is_zero(&q->q_usage_counter), | |
162 | timeout); | |
163 | } | |
164 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 165 | |
f3af020b TH |
166 | /* |
167 | * Guarantee no request is in use, so we can change any data structure of | |
168 | * the queue afterward. | |
169 | */ | |
3ef28e83 | 170 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 171 | { |
3ef28e83 DW |
172 | /* |
173 | * In the !blk_mq case we are only calling this to kill the | |
174 | * q_usage_counter, otherwise this increases the freeze depth | |
175 | * and waits for it to return to zero. For this reason there is | |
176 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
177 | * exported to drivers as the only user for unfreeze is blk_mq. | |
178 | */ | |
1671d522 | 179 | blk_freeze_queue_start(q); |
f3af020b TH |
180 | blk_mq_freeze_queue_wait(q); |
181 | } | |
3ef28e83 DW |
182 | |
183 | void blk_mq_freeze_queue(struct request_queue *q) | |
184 | { | |
185 | /* | |
186 | * ...just an alias to keep freeze and unfreeze actions balanced | |
187 | * in the blk_mq_* namespace | |
188 | */ | |
189 | blk_freeze_queue(q); | |
190 | } | |
c761d96b | 191 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 192 | |
aec89dc5 | 193 | void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) |
320ae51f | 194 | { |
7996a8b5 | 195 | mutex_lock(&q->mq_freeze_lock); |
aec89dc5 CH |
196 | if (force_atomic) |
197 | q->q_usage_counter.data->force_atomic = true; | |
7996a8b5 BL |
198 | q->mq_freeze_depth--; |
199 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
200 | if (!q->mq_freeze_depth) { | |
bdd63160 | 201 | percpu_ref_resurrect(&q->q_usage_counter); |
320ae51f | 202 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 203 | } |
7996a8b5 | 204 | mutex_unlock(&q->mq_freeze_lock); |
320ae51f | 205 | } |
aec89dc5 CH |
206 | |
207 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
208 | { | |
209 | __blk_mq_unfreeze_queue(q, false); | |
210 | } | |
b4c6a028 | 211 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 212 | |
852ec809 BVA |
213 | /* |
214 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
215 | * mpt3sas driver such that this function can be removed. | |
216 | */ | |
217 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
218 | { | |
8814ce8a | 219 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); |
852ec809 BVA |
220 | } |
221 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
222 | ||
6a83e74d | 223 | /** |
69e07c4a | 224 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
225 | * @q: request queue. |
226 | * | |
227 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
228 | * callback function is invoked. Once this function is returned, we make |
229 | * sure no dispatch can happen until the queue is unquiesced via | |
230 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
231 | */ |
232 | void blk_mq_quiesce_queue(struct request_queue *q) | |
233 | { | |
234 | struct blk_mq_hw_ctx *hctx; | |
235 | unsigned int i; | |
236 | bool rcu = false; | |
237 | ||
1d9e9bc6 | 238 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 239 | |
6a83e74d BVA |
240 | queue_for_each_hw_ctx(q, hctx, i) { |
241 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
05707b64 | 242 | synchronize_srcu(hctx->srcu); |
6a83e74d BVA |
243 | else |
244 | rcu = true; | |
245 | } | |
246 | if (rcu) | |
247 | synchronize_rcu(); | |
248 | } | |
249 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
250 | ||
e4e73913 ML |
251 | /* |
252 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
253 | * @q: request queue. | |
254 | * | |
255 | * This function recovers queue into the state before quiescing | |
256 | * which is done by blk_mq_quiesce_queue. | |
257 | */ | |
258 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
259 | { | |
8814ce8a | 260 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
f4560ffe | 261 | |
1d9e9bc6 ML |
262 | /* dispatch requests which are inserted during quiescing */ |
263 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
264 | } |
265 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
266 | ||
aed3ea94 JA |
267 | void blk_mq_wake_waiters(struct request_queue *q) |
268 | { | |
269 | struct blk_mq_hw_ctx *hctx; | |
270 | unsigned int i; | |
271 | ||
272 | queue_for_each_hw_ctx(q, hctx, i) | |
273 | if (blk_mq_hw_queue_mapped(hctx)) | |
274 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
275 | } | |
276 | ||
fe1f4526 | 277 | /* |
9a91b05b HT |
278 | * Only need start/end time stamping if we have iostat or |
279 | * blk stats enabled, or using an IO scheduler. | |
fe1f4526 JA |
280 | */ |
281 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
282 | { | |
9a91b05b | 283 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; |
fe1f4526 JA |
284 | } |
285 | ||
e4cdf1a1 | 286 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
7ea4d8a4 | 287 | unsigned int tag, u64 alloc_time_ns) |
320ae51f | 288 | { |
e4cdf1a1 CH |
289 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
290 | struct request *rq = tags->static_rqs[tag]; | |
c3a148d2 | 291 | |
42fdc5e4 | 292 | if (data->q->elevator) { |
76647368 | 293 | rq->tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
294 | rq->internal_tag = tag; |
295 | } else { | |
e4cdf1a1 | 296 | rq->tag = tag; |
76647368 | 297 | rq->internal_tag = BLK_MQ_NO_TAG; |
e4cdf1a1 CH |
298 | } |
299 | ||
af76e555 | 300 | /* csd/requeue_work/fifo_time is initialized before use */ |
e4cdf1a1 CH |
301 | rq->q = data->q; |
302 | rq->mq_ctx = data->ctx; | |
ea4f995e | 303 | rq->mq_hctx = data->hctx; |
568f2700 | 304 | rq->rq_flags = 0; |
7ea4d8a4 | 305 | rq->cmd_flags = data->cmd_flags; |
0854bcdc BVA |
306 | if (data->flags & BLK_MQ_REQ_PM) |
307 | rq->rq_flags |= RQF_PM; | |
e4cdf1a1 | 308 | if (blk_queue_io_stat(data->q)) |
e8064021 | 309 | rq->rq_flags |= RQF_IO_STAT; |
7c3fb70f | 310 | INIT_LIST_HEAD(&rq->queuelist); |
af76e555 CH |
311 | INIT_HLIST_NODE(&rq->hash); |
312 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
313 | rq->rq_disk = NULL; |
314 | rq->part = NULL; | |
6f816b4b TH |
315 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME |
316 | rq->alloc_time_ns = alloc_time_ns; | |
317 | #endif | |
fe1f4526 JA |
318 | if (blk_mq_need_time_stamp(rq)) |
319 | rq->start_time_ns = ktime_get_ns(); | |
320 | else | |
321 | rq->start_time_ns = 0; | |
544ccc8d | 322 | rq->io_start_time_ns = 0; |
3d244306 | 323 | rq->stats_sectors = 0; |
af76e555 CH |
324 | rq->nr_phys_segments = 0; |
325 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
326 | rq->nr_integrity_segments = 0; | |
327 | #endif | |
a892c8d5 | 328 | blk_crypto_rq_set_defaults(rq); |
af76e555 | 329 | /* tag was already set */ |
079076b3 | 330 | WRITE_ONCE(rq->deadline, 0); |
af76e555 | 331 | |
f6be4fb4 JA |
332 | rq->timeout = 0; |
333 | ||
af76e555 CH |
334 | rq->end_io = NULL; |
335 | rq->end_io_data = NULL; | |
af76e555 | 336 | |
7ea4d8a4 | 337 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; |
12f5b931 | 338 | refcount_set(&rq->ref, 1); |
7ea4d8a4 CH |
339 | |
340 | if (!op_is_flush(data->cmd_flags)) { | |
341 | struct elevator_queue *e = data->q->elevator; | |
342 | ||
343 | rq->elv.icq = NULL; | |
344 | if (e && e->type->ops.prepare_request) { | |
345 | if (e->type->icq_cache) | |
346 | blk_mq_sched_assign_ioc(rq); | |
347 | ||
348 | e->type->ops.prepare_request(rq); | |
349 | rq->rq_flags |= RQF_ELVPRIV; | |
350 | } | |
351 | } | |
352 | ||
353 | data->hctx->queued++; | |
e4cdf1a1 | 354 | return rq; |
5dee8577 CH |
355 | } |
356 | ||
b90cfaed | 357 | static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) |
d2c0d383 | 358 | { |
e6e7abff | 359 | struct request_queue *q = data->q; |
d2c0d383 | 360 | struct elevator_queue *e = q->elevator; |
6f816b4b | 361 | u64 alloc_time_ns = 0; |
47c122e3 | 362 | struct request *rq; |
600c3b0c | 363 | unsigned int tag; |
d2c0d383 | 364 | |
6f816b4b TH |
365 | /* alloc_time includes depth and tag waits */ |
366 | if (blk_queue_rq_alloc_time(q)) | |
367 | alloc_time_ns = ktime_get_ns(); | |
368 | ||
f9afca4d | 369 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 370 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
371 | |
372 | if (e) { | |
d2c0d383 | 373 | /* |
8d663f34 | 374 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
375 | * dispatch list. Don't include reserved tags in the |
376 | * limiting, as it isn't useful. | |
d2c0d383 | 377 | */ |
f9afca4d | 378 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 379 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 380 | e->type->ops.limit_depth && |
17a51199 | 381 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 382 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
383 | } |
384 | ||
bf0beec0 | 385 | retry: |
600c3b0c CH |
386 | data->ctx = blk_mq_get_ctx(q); |
387 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 388 | if (!e) |
600c3b0c CH |
389 | blk_mq_tag_busy(data->hctx); |
390 | ||
bf0beec0 ML |
391 | /* |
392 | * Waiting allocations only fail because of an inactive hctx. In that | |
393 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
394 | * should have migrated us to an online CPU by now. | |
395 | */ | |
47c122e3 JA |
396 | do { |
397 | tag = blk_mq_get_tag(data); | |
b90cfaed CH |
398 | if (tag == BLK_MQ_NO_TAG) { |
399 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
400 | break; | |
401 | /* | |
402 | * Give up the CPU and sleep for a random short time to | |
403 | * ensure that thread using a realtime scheduling class | |
404 | * are migrated off the CPU, and thus off the hctx that | |
405 | * is going away. | |
406 | */ | |
407 | msleep(3); | |
408 | goto retry; | |
47c122e3 | 409 | } |
bf0beec0 | 410 | |
b90cfaed CH |
411 | rq = blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
412 | if (!--data->nr_tags || e || | |
413 | (data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) | |
414 | return rq; | |
415 | ||
416 | /* link into the cached list */ | |
417 | rq->rq_next = *data->cached_rq; | |
418 | *data->cached_rq = rq; | |
419 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
47c122e3 JA |
420 | } while (1); |
421 | ||
b90cfaed CH |
422 | if (!data->cached_rq) |
423 | return NULL; | |
47c122e3 | 424 | |
b90cfaed CH |
425 | rq = *data->cached_rq; |
426 | *data->cached_rq = rq->rq_next; | |
427 | return rq; | |
d2c0d383 CH |
428 | } |
429 | ||
cd6ce148 | 430 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 431 | blk_mq_req_flags_t flags) |
320ae51f | 432 | { |
e6e7abff CH |
433 | struct blk_mq_alloc_data data = { |
434 | .q = q, | |
435 | .flags = flags, | |
436 | .cmd_flags = op, | |
47c122e3 | 437 | .nr_tags = 1, |
e6e7abff | 438 | }; |
bd166ef1 | 439 | struct request *rq; |
a492f075 | 440 | int ret; |
320ae51f | 441 | |
3a0a5299 | 442 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
443 | if (ret) |
444 | return ERR_PTR(ret); | |
320ae51f | 445 | |
b90cfaed | 446 | rq = __blk_mq_alloc_requests(&data); |
bd166ef1 | 447 | if (!rq) |
a5ea5811 | 448 | goto out_queue_exit; |
0c4de0f3 CH |
449 | rq->__data_len = 0; |
450 | rq->__sector = (sector_t) -1; | |
451 | rq->bio = rq->biotail = NULL; | |
320ae51f | 452 | return rq; |
a5ea5811 CH |
453 | out_queue_exit: |
454 | blk_queue_exit(q); | |
455 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 456 | } |
4bb659b1 | 457 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 458 | |
cd6ce148 | 459 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 460 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 461 | { |
e6e7abff CH |
462 | struct blk_mq_alloc_data data = { |
463 | .q = q, | |
464 | .flags = flags, | |
465 | .cmd_flags = op, | |
47c122e3 | 466 | .nr_tags = 1, |
e6e7abff | 467 | }; |
600c3b0c | 468 | u64 alloc_time_ns = 0; |
6d2809d5 | 469 | unsigned int cpu; |
600c3b0c | 470 | unsigned int tag; |
1f5bd336 ML |
471 | int ret; |
472 | ||
600c3b0c CH |
473 | /* alloc_time includes depth and tag waits */ |
474 | if (blk_queue_rq_alloc_time(q)) | |
475 | alloc_time_ns = ktime_get_ns(); | |
476 | ||
1f5bd336 ML |
477 | /* |
478 | * If the tag allocator sleeps we could get an allocation for a | |
479 | * different hardware context. No need to complicate the low level | |
480 | * allocator for this for the rare use case of a command tied to | |
481 | * a specific queue. | |
482 | */ | |
600c3b0c | 483 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
484 | return ERR_PTR(-EINVAL); |
485 | ||
486 | if (hctx_idx >= q->nr_hw_queues) | |
487 | return ERR_PTR(-EIO); | |
488 | ||
3a0a5299 | 489 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
490 | if (ret) |
491 | return ERR_PTR(ret); | |
492 | ||
c8712c6a CH |
493 | /* |
494 | * Check if the hardware context is actually mapped to anything. | |
495 | * If not tell the caller that it should skip this queue. | |
496 | */ | |
a5ea5811 | 497 | ret = -EXDEV; |
e6e7abff CH |
498 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
499 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 500 | goto out_queue_exit; |
e6e7abff CH |
501 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
502 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 503 | |
42fdc5e4 | 504 | if (!q->elevator) |
600c3b0c CH |
505 | blk_mq_tag_busy(data.hctx); |
506 | ||
a5ea5811 | 507 | ret = -EWOULDBLOCK; |
600c3b0c CH |
508 | tag = blk_mq_get_tag(&data); |
509 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 510 | goto out_queue_exit; |
600c3b0c CH |
511 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
512 | ||
a5ea5811 CH |
513 | out_queue_exit: |
514 | blk_queue_exit(q); | |
515 | return ERR_PTR(ret); | |
1f5bd336 ML |
516 | } |
517 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
518 | ||
12f5b931 KB |
519 | static void __blk_mq_free_request(struct request *rq) |
520 | { | |
521 | struct request_queue *q = rq->q; | |
522 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 523 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
524 | const int sched_tag = rq->internal_tag; |
525 | ||
a892c8d5 | 526 | blk_crypto_free_request(rq); |
986d413b | 527 | blk_pm_mark_last_busy(rq); |
ea4f995e | 528 | rq->mq_hctx = NULL; |
76647368 | 529 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 530 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 531 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 532 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
533 | blk_mq_sched_restart(hctx); |
534 | blk_queue_exit(q); | |
535 | } | |
536 | ||
6af54051 | 537 | void blk_mq_free_request(struct request *rq) |
320ae51f | 538 | { |
320ae51f | 539 | struct request_queue *q = rq->q; |
6af54051 CH |
540 | struct elevator_queue *e = q->elevator; |
541 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 542 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 543 | |
5bbf4e5a | 544 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
545 | if (e && e->type->ops.finish_request) |
546 | e->type->ops.finish_request(rq); | |
6af54051 CH |
547 | if (rq->elv.icq) { |
548 | put_io_context(rq->elv.icq->ioc); | |
549 | rq->elv.icq = NULL; | |
550 | } | |
551 | } | |
320ae51f | 552 | |
6af54051 | 553 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 554 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 555 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 556 | |
7beb2f84 | 557 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 558 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 559 | |
a7905043 | 560 | rq_qos_done(q, rq); |
0d2602ca | 561 | |
12f5b931 KB |
562 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
563 | if (refcount_dec_and_test(&rq->ref)) | |
564 | __blk_mq_free_request(rq); | |
320ae51f | 565 | } |
1a3b595a | 566 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 567 | |
47c122e3 JA |
568 | void blk_mq_free_plug_rqs(struct blk_plug *plug) |
569 | { | |
570 | while (plug->cached_rq) { | |
571 | struct request *rq; | |
572 | ||
573 | rq = plug->cached_rq; | |
574 | plug->cached_rq = rq->rq_next; | |
575 | percpu_ref_get(&rq->q->q_usage_counter); | |
576 | blk_mq_free_request(rq); | |
577 | } | |
578 | } | |
579 | ||
2a842aca | 580 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 581 | { |
8971a3b7 PB |
582 | if (blk_mq_need_time_stamp(rq)) { |
583 | u64 now = ktime_get_ns(); | |
fe1f4526 | 584 | |
8971a3b7 PB |
585 | if (rq->rq_flags & RQF_STATS) { |
586 | blk_mq_poll_stats_start(rq->q); | |
587 | blk_stat_add(rq, now); | |
588 | } | |
522a7775 | 589 | |
8971a3b7 PB |
590 | blk_mq_sched_completed_request(rq, now); |
591 | blk_account_io_done(rq, now); | |
4bc6339a OS |
592 | } |
593 | ||
91b63639 | 594 | if (rq->end_io) { |
a7905043 | 595 | rq_qos_done(rq->q, rq); |
320ae51f | 596 | rq->end_io(rq, error); |
91b63639 | 597 | } else { |
320ae51f | 598 | blk_mq_free_request(rq); |
91b63639 | 599 | } |
320ae51f | 600 | } |
c8a446ad | 601 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 602 | |
2a842aca | 603 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
604 | { |
605 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
606 | BUG(); | |
c8a446ad | 607 | __blk_mq_end_request(rq, error); |
63151a44 | 608 | } |
c8a446ad | 609 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 610 | |
f9ab4918 | 611 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 612 | { |
f9ab4918 SAS |
613 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
614 | struct request *rq, *next; | |
c3077b5d | 615 | |
f9ab4918 | 616 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 617 | rq->q->mq_ops->complete(rq); |
320ae51f | 618 | } |
320ae51f | 619 | |
f9ab4918 | 620 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 621 | { |
f9ab4918 | 622 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
623 | } |
624 | ||
c3077b5d CH |
625 | static int blk_softirq_cpu_dead(unsigned int cpu) |
626 | { | |
f9ab4918 | 627 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
628 | return 0; |
629 | } | |
630 | ||
40d09b53 | 631 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 632 | { |
f9ab4918 | 633 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
634 | } |
635 | ||
96339526 CH |
636 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
637 | { | |
638 | int cpu = raw_smp_processor_id(); | |
639 | ||
640 | if (!IS_ENABLED(CONFIG_SMP) || | |
641 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
642 | return false; | |
71425189 SAS |
643 | /* |
644 | * With force threaded interrupts enabled, raising softirq from an SMP | |
645 | * function call will always result in waking the ksoftirqd thread. | |
646 | * This is probably worse than completing the request on a different | |
647 | * cache domain. | |
648 | */ | |
91cc470e | 649 | if (force_irqthreads()) |
71425189 | 650 | return false; |
96339526 CH |
651 | |
652 | /* same CPU or cache domain? Complete locally */ | |
653 | if (cpu == rq->mq_ctx->cpu || | |
654 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
655 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
656 | return false; | |
657 | ||
658 | /* don't try to IPI to an offline CPU */ | |
659 | return cpu_online(rq->mq_ctx->cpu); | |
660 | } | |
661 | ||
f9ab4918 SAS |
662 | static void blk_mq_complete_send_ipi(struct request *rq) |
663 | { | |
664 | struct llist_head *list; | |
665 | unsigned int cpu; | |
666 | ||
667 | cpu = rq->mq_ctx->cpu; | |
668 | list = &per_cpu(blk_cpu_done, cpu); | |
669 | if (llist_add(&rq->ipi_list, list)) { | |
670 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
671 | smp_call_function_single_async(cpu, &rq->csd); | |
672 | } | |
673 | } | |
674 | ||
675 | static void blk_mq_raise_softirq(struct request *rq) | |
676 | { | |
677 | struct llist_head *list; | |
678 | ||
679 | preempt_disable(); | |
680 | list = this_cpu_ptr(&blk_cpu_done); | |
681 | if (llist_add(&rq->ipi_list, list)) | |
682 | raise_softirq(BLOCK_SOFTIRQ); | |
683 | preempt_enable(); | |
684 | } | |
685 | ||
40d09b53 | 686 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 687 | { |
af78ff7c | 688 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 689 | |
4ab32bf3 JA |
690 | /* |
691 | * For a polled request, always complete locallly, it's pointless | |
692 | * to redirect the completion. | |
693 | */ | |
40d09b53 CH |
694 | if (rq->cmd_flags & REQ_HIPRI) |
695 | return false; | |
38535201 | 696 | |
96339526 | 697 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
698 | blk_mq_complete_send_ipi(rq); |
699 | return true; | |
3d6efbf6 | 700 | } |
40d09b53 | 701 | |
f9ab4918 SAS |
702 | if (rq->q->nr_hw_queues == 1) { |
703 | blk_mq_raise_softirq(rq); | |
704 | return true; | |
705 | } | |
706 | return false; | |
40d09b53 CH |
707 | } |
708 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
709 | ||
710 | /** | |
711 | * blk_mq_complete_request - end I/O on a request | |
712 | * @rq: the request being processed | |
713 | * | |
714 | * Description: | |
715 | * Complete a request by scheduling the ->complete_rq operation. | |
716 | **/ | |
717 | void blk_mq_complete_request(struct request *rq) | |
718 | { | |
719 | if (!blk_mq_complete_request_remote(rq)) | |
720 | rq->q->mq_ops->complete(rq); | |
320ae51f | 721 | } |
15f73f5b | 722 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 723 | |
04ced159 | 724 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 725 | __releases(hctx->srcu) |
04ced159 JA |
726 | { |
727 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
728 | rcu_read_unlock(); | |
729 | else | |
05707b64 | 730 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
731 | } |
732 | ||
733 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 734 | __acquires(hctx->srcu) |
04ced159 | 735 | { |
08b5a6e2 JA |
736 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
737 | /* shut up gcc false positive */ | |
738 | *srcu_idx = 0; | |
04ced159 | 739 | rcu_read_lock(); |
08b5a6e2 | 740 | } else |
05707b64 | 741 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
742 | } |
743 | ||
105663f7 AA |
744 | /** |
745 | * blk_mq_start_request - Start processing a request | |
746 | * @rq: Pointer to request to be started | |
747 | * | |
748 | * Function used by device drivers to notify the block layer that a request | |
749 | * is going to be processed now, so blk layer can do proper initializations | |
750 | * such as starting the timeout timer. | |
751 | */ | |
e2490073 | 752 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
753 | { |
754 | struct request_queue *q = rq->q; | |
755 | ||
a54895fa | 756 | trace_block_rq_issue(rq); |
320ae51f | 757 | |
cf43e6be | 758 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
00067077 JA |
759 | u64 start_time; |
760 | #ifdef CONFIG_BLK_CGROUP | |
761 | if (rq->bio) | |
762 | start_time = bio_issue_time(&rq->bio->bi_issue); | |
763 | else | |
764 | #endif | |
765 | start_time = ktime_get_ns(); | |
766 | rq->io_start_time_ns = start_time; | |
3d244306 | 767 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 768 | rq->rq_flags |= RQF_STATS; |
a7905043 | 769 | rq_qos_issue(q, rq); |
cf43e6be JA |
770 | } |
771 | ||
1d9bd516 | 772 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 773 | |
1d9bd516 | 774 | blk_add_timer(rq); |
12f5b931 | 775 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 776 | |
54d4e6ab MG |
777 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
778 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
779 | q->integrity.profile->prepare_fn(rq); | |
780 | #endif | |
320ae51f | 781 | } |
e2490073 | 782 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 783 | |
ed0791b2 | 784 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
785 | { |
786 | struct request_queue *q = rq->q; | |
787 | ||
923218f6 ML |
788 | blk_mq_put_driver_tag(rq); |
789 | ||
a54895fa | 790 | trace_block_rq_requeue(rq); |
a7905043 | 791 | rq_qos_requeue(q, rq); |
49f5baa5 | 792 | |
12f5b931 KB |
793 | if (blk_mq_request_started(rq)) { |
794 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 795 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 796 | } |
320ae51f JA |
797 | } |
798 | ||
2b053aca | 799 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 800 | { |
ed0791b2 | 801 | __blk_mq_requeue_request(rq); |
ed0791b2 | 802 | |
105976f5 ML |
803 | /* this request will be re-inserted to io scheduler queue */ |
804 | blk_mq_sched_requeue_request(rq); | |
805 | ||
7d692330 | 806 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 807 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
808 | } |
809 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
810 | ||
6fca6a61 CH |
811 | static void blk_mq_requeue_work(struct work_struct *work) |
812 | { | |
813 | struct request_queue *q = | |
2849450a | 814 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
815 | LIST_HEAD(rq_list); |
816 | struct request *rq, *next; | |
6fca6a61 | 817 | |
18e9781d | 818 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 819 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 820 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
821 | |
822 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 823 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
824 | continue; |
825 | ||
e8064021 | 826 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 827 | list_del_init(&rq->queuelist); |
aef1897c JW |
828 | /* |
829 | * If RQF_DONTPREP, rq has contained some driver specific | |
830 | * data, so insert it to hctx dispatch list to avoid any | |
831 | * merge. | |
832 | */ | |
833 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 834 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
835 | else |
836 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
837 | } |
838 | ||
839 | while (!list_empty(&rq_list)) { | |
840 | rq = list_entry(rq_list.next, struct request, queuelist); | |
841 | list_del_init(&rq->queuelist); | |
9e97d295 | 842 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
843 | } |
844 | ||
52d7f1b5 | 845 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
846 | } |
847 | ||
2b053aca BVA |
848 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
849 | bool kick_requeue_list) | |
6fca6a61 CH |
850 | { |
851 | struct request_queue *q = rq->q; | |
852 | unsigned long flags; | |
853 | ||
854 | /* | |
855 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 856 | * request head insertion from the workqueue. |
6fca6a61 | 857 | */ |
e8064021 | 858 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
859 | |
860 | spin_lock_irqsave(&q->requeue_lock, flags); | |
861 | if (at_head) { | |
e8064021 | 862 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
863 | list_add(&rq->queuelist, &q->requeue_list); |
864 | } else { | |
865 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
866 | } | |
867 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
868 | |
869 | if (kick_requeue_list) | |
870 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 871 | } |
6fca6a61 CH |
872 | |
873 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
874 | { | |
ae943d20 | 875 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
876 | } |
877 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
878 | ||
2849450a MS |
879 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
880 | unsigned long msecs) | |
881 | { | |
d4acf365 BVA |
882 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
883 | msecs_to_jiffies(msecs)); | |
2849450a MS |
884 | } |
885 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
886 | ||
0e62f51f JA |
887 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
888 | { | |
88c7b2b7 JA |
889 | if (tag < tags->nr_tags) { |
890 | prefetch(tags->rqs[tag]); | |
4ee86bab | 891 | return tags->rqs[tag]; |
88c7b2b7 | 892 | } |
4ee86bab HR |
893 | |
894 | return NULL; | |
24d2f903 CH |
895 | } |
896 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
897 | ||
3c94d83c JA |
898 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
899 | void *priv, bool reserved) | |
ae879912 JA |
900 | { |
901 | /* | |
05a4fed6 | 902 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 903 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 904 | */ |
05a4fed6 | 905 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
906 | bool *busy = priv; |
907 | ||
908 | *busy = true; | |
909 | return false; | |
910 | } | |
911 | ||
912 | return true; | |
913 | } | |
914 | ||
3c94d83c | 915 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
916 | { |
917 | bool busy = false; | |
918 | ||
3c94d83c | 919 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
920 | return busy; |
921 | } | |
3c94d83c | 922 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 923 | |
358f70da | 924 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 925 | { |
da661267 | 926 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
927 | if (req->q->mq_ops->timeout) { |
928 | enum blk_eh_timer_return ret; | |
929 | ||
930 | ret = req->q->mq_ops->timeout(req, reserved); | |
931 | if (ret == BLK_EH_DONE) | |
932 | return; | |
933 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 934 | } |
d1210d5a CH |
935 | |
936 | blk_add_timer(req); | |
87ee7b11 | 937 | } |
5b3f25fc | 938 | |
12f5b931 | 939 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 940 | { |
12f5b931 | 941 | unsigned long deadline; |
87ee7b11 | 942 | |
12f5b931 KB |
943 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
944 | return false; | |
da661267 CH |
945 | if (rq->rq_flags & RQF_TIMED_OUT) |
946 | return false; | |
a7af0af3 | 947 | |
079076b3 | 948 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
949 | if (time_after_eq(jiffies, deadline)) |
950 | return true; | |
a7af0af3 | 951 | |
12f5b931 KB |
952 | if (*next == 0) |
953 | *next = deadline; | |
954 | else if (time_after(*next, deadline)) | |
955 | *next = deadline; | |
956 | return false; | |
87ee7b11 JA |
957 | } |
958 | ||
2e315dc0 ML |
959 | void blk_mq_put_rq_ref(struct request *rq) |
960 | { | |
a9ed27a7 | 961 | if (is_flush_rq(rq)) |
2e315dc0 ML |
962 | rq->end_io(rq, 0); |
963 | else if (refcount_dec_and_test(&rq->ref)) | |
964 | __blk_mq_free_request(rq); | |
965 | } | |
966 | ||
7baa8572 | 967 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
968 | struct request *rq, void *priv, bool reserved) |
969 | { | |
12f5b931 KB |
970 | unsigned long *next = priv; |
971 | ||
972 | /* | |
c797b40c ML |
973 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
974 | * be reallocated underneath the timeout handler's processing, then | |
975 | * the expire check is reliable. If the request is not expired, then | |
976 | * it was completed and reallocated as a new request after returning | |
977 | * from blk_mq_check_expired(). | |
1d9bd516 | 978 | */ |
12f5b931 | 979 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 980 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 981 | return true; |
1d9bd516 TH |
982 | } |
983 | ||
287922eb | 984 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 985 | { |
287922eb CH |
986 | struct request_queue *q = |
987 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 988 | unsigned long next = 0; |
1d9bd516 | 989 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 990 | int i; |
320ae51f | 991 | |
71f79fb3 GKB |
992 | /* A deadlock might occur if a request is stuck requiring a |
993 | * timeout at the same time a queue freeze is waiting | |
994 | * completion, since the timeout code would not be able to | |
995 | * acquire the queue reference here. | |
996 | * | |
997 | * That's why we don't use blk_queue_enter here; instead, we use | |
998 | * percpu_ref_tryget directly, because we need to be able to | |
999 | * obtain a reference even in the short window between the queue | |
1000 | * starting to freeze, by dropping the first reference in | |
1671d522 | 1001 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
1002 | * consumed, marked by the instant q_usage_counter reaches |
1003 | * zero. | |
1004 | */ | |
1005 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
1006 | return; |
1007 | ||
12f5b931 | 1008 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 1009 | |
12f5b931 KB |
1010 | if (next != 0) { |
1011 | mod_timer(&q->timeout, next); | |
0d2602ca | 1012 | } else { |
fcd36c36 BVA |
1013 | /* |
1014 | * Request timeouts are handled as a forward rolling timer. If | |
1015 | * we end up here it means that no requests are pending and | |
1016 | * also that no request has been pending for a while. Mark | |
1017 | * each hctx as idle. | |
1018 | */ | |
f054b56c ML |
1019 | queue_for_each_hw_ctx(q, hctx, i) { |
1020 | /* the hctx may be unmapped, so check it here */ | |
1021 | if (blk_mq_hw_queue_mapped(hctx)) | |
1022 | blk_mq_tag_idle(hctx); | |
1023 | } | |
0d2602ca | 1024 | } |
287922eb | 1025 | blk_queue_exit(q); |
320ae51f JA |
1026 | } |
1027 | ||
88459642 OS |
1028 | struct flush_busy_ctx_data { |
1029 | struct blk_mq_hw_ctx *hctx; | |
1030 | struct list_head *list; | |
1031 | }; | |
1032 | ||
1033 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1034 | { | |
1035 | struct flush_busy_ctx_data *flush_data = data; | |
1036 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1037 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1038 | enum hctx_type type = hctx->type; |
88459642 | 1039 | |
88459642 | 1040 | spin_lock(&ctx->lock); |
c16d6b5a | 1041 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1042 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1043 | spin_unlock(&ctx->lock); |
1044 | return true; | |
1045 | } | |
1046 | ||
1429d7c9 JA |
1047 | /* |
1048 | * Process software queues that have been marked busy, splicing them | |
1049 | * to the for-dispatch | |
1050 | */ | |
2c3ad667 | 1051 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1052 | { |
88459642 OS |
1053 | struct flush_busy_ctx_data data = { |
1054 | .hctx = hctx, | |
1055 | .list = list, | |
1056 | }; | |
1429d7c9 | 1057 | |
88459642 | 1058 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1059 | } |
2c3ad667 | 1060 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1061 | |
b347689f ML |
1062 | struct dispatch_rq_data { |
1063 | struct blk_mq_hw_ctx *hctx; | |
1064 | struct request *rq; | |
1065 | }; | |
1066 | ||
1067 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1068 | void *data) | |
1069 | { | |
1070 | struct dispatch_rq_data *dispatch_data = data; | |
1071 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1072 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1073 | enum hctx_type type = hctx->type; |
b347689f ML |
1074 | |
1075 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1076 | if (!list_empty(&ctx->rq_lists[type])) { |
1077 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1078 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1079 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1080 | sbitmap_clear_bit(sb, bitnr); |
1081 | } | |
1082 | spin_unlock(&ctx->lock); | |
1083 | ||
1084 | return !dispatch_data->rq; | |
1085 | } | |
1086 | ||
1087 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1088 | struct blk_mq_ctx *start) | |
1089 | { | |
f31967f0 | 1090 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1091 | struct dispatch_rq_data data = { |
1092 | .hctx = hctx, | |
1093 | .rq = NULL, | |
1094 | }; | |
1095 | ||
1096 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1097 | dispatch_rq_from_ctx, &data); | |
1098 | ||
1099 | return data.rq; | |
1100 | } | |
1101 | ||
703fd1c0 JA |
1102 | static inline unsigned int queued_to_index(unsigned int queued) |
1103 | { | |
1104 | if (!queued) | |
1105 | return 0; | |
1429d7c9 | 1106 | |
703fd1c0 | 1107 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1108 | } |
1109 | ||
570e9b73 ML |
1110 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1111 | { | |
ae0f1a73 | 1112 | struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1113 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1114 | int tag; |
1115 | ||
568f2700 ML |
1116 | blk_mq_tag_busy(rq->mq_hctx); |
1117 | ||
570e9b73 | 1118 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
ae0f1a73 | 1119 | bt = &rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1120 | tag_offset = 0; |
28500850 ML |
1121 | } else { |
1122 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1123 | return false; | |
570e9b73 ML |
1124 | } |
1125 | ||
570e9b73 ML |
1126 | tag = __sbitmap_queue_get(bt); |
1127 | if (tag == BLK_MQ_NO_TAG) | |
1128 | return false; | |
1129 | ||
1130 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1131 | return true; |
1132 | } | |
1133 | ||
61347154 | 1134 | bool blk_mq_get_driver_tag(struct request *rq) |
570e9b73 | 1135 | { |
568f2700 ML |
1136 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1137 | ||
1138 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1139 | return false; | |
1140 | ||
51db1c37 | 1141 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1142 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1143 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1144 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1145 | } |
1146 | hctx->tags->rqs[rq->tag] = rq; | |
1147 | return true; | |
570e9b73 ML |
1148 | } |
1149 | ||
eb619fdb JA |
1150 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1151 | int flags, void *key) | |
da55f2cc OS |
1152 | { |
1153 | struct blk_mq_hw_ctx *hctx; | |
1154 | ||
1155 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1156 | ||
5815839b | 1157 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1158 | if (!list_empty(&wait->entry)) { |
1159 | struct sbitmap_queue *sbq; | |
1160 | ||
1161 | list_del_init(&wait->entry); | |
ae0f1a73 | 1162 | sbq = &hctx->tags->bitmap_tags; |
e8618575 JA |
1163 | atomic_dec(&sbq->ws_active); |
1164 | } | |
5815839b ML |
1165 | spin_unlock(&hctx->dispatch_wait_lock); |
1166 | ||
da55f2cc OS |
1167 | blk_mq_run_hw_queue(hctx, true); |
1168 | return 1; | |
1169 | } | |
1170 | ||
f906a6a0 JA |
1171 | /* |
1172 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1173 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1174 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1175 | * marking us as waiting. |
1176 | */ | |
2278d69f | 1177 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1178 | struct request *rq) |
da55f2cc | 1179 | { |
ae0f1a73 | 1180 | struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags; |
5815839b | 1181 | struct wait_queue_head *wq; |
f906a6a0 JA |
1182 | wait_queue_entry_t *wait; |
1183 | bool ret; | |
da55f2cc | 1184 | |
51db1c37 | 1185 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1186 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1187 | |
c27d53fb BVA |
1188 | /* |
1189 | * It's possible that a tag was freed in the window between the | |
1190 | * allocation failure and adding the hardware queue to the wait | |
1191 | * queue. | |
1192 | * | |
1193 | * Don't clear RESTART here, someone else could have set it. | |
1194 | * At most this will cost an extra queue run. | |
1195 | */ | |
8ab6bb9e | 1196 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1197 | } |
eb619fdb | 1198 | |
2278d69f | 1199 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1200 | if (!list_empty_careful(&wait->entry)) |
1201 | return false; | |
1202 | ||
e8618575 | 1203 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1204 | |
1205 | spin_lock_irq(&wq->lock); | |
1206 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1207 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1208 | spin_unlock(&hctx->dispatch_wait_lock); |
1209 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1210 | return false; |
eb619fdb JA |
1211 | } |
1212 | ||
e8618575 | 1213 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1214 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1215 | __add_wait_queue(wq, wait); | |
c27d53fb | 1216 | |
da55f2cc | 1217 | /* |
eb619fdb JA |
1218 | * It's possible that a tag was freed in the window between the |
1219 | * allocation failure and adding the hardware queue to the wait | |
1220 | * queue. | |
da55f2cc | 1221 | */ |
8ab6bb9e | 1222 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1223 | if (!ret) { |
5815839b ML |
1224 | spin_unlock(&hctx->dispatch_wait_lock); |
1225 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1226 | return false; |
eb619fdb | 1227 | } |
c27d53fb BVA |
1228 | |
1229 | /* | |
1230 | * We got a tag, remove ourselves from the wait queue to ensure | |
1231 | * someone else gets the wakeup. | |
1232 | */ | |
c27d53fb | 1233 | list_del_init(&wait->entry); |
e8618575 | 1234 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1235 | spin_unlock(&hctx->dispatch_wait_lock); |
1236 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1237 | |
1238 | return true; | |
da55f2cc OS |
1239 | } |
1240 | ||
6e768717 ML |
1241 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1242 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1243 | /* | |
1244 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1245 | * - EWMA is one simple way to compute running average value | |
1246 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1247 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1248 | * factor doesn't matter because EWMA decreases exponentially | |
1249 | */ | |
1250 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1251 | { | |
1252 | unsigned int ewma; | |
1253 | ||
6e768717 ML |
1254 | ewma = hctx->dispatch_busy; |
1255 | ||
1256 | if (!ewma && !busy) | |
1257 | return; | |
1258 | ||
1259 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1260 | if (busy) | |
1261 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1262 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1263 | ||
1264 | hctx->dispatch_busy = ewma; | |
1265 | } | |
1266 | ||
86ff7c2a ML |
1267 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1268 | ||
c92a4103 JT |
1269 | static void blk_mq_handle_dev_resource(struct request *rq, |
1270 | struct list_head *list) | |
1271 | { | |
1272 | struct request *next = | |
1273 | list_first_entry_or_null(list, struct request, queuelist); | |
1274 | ||
1275 | /* | |
1276 | * If an I/O scheduler has been configured and we got a driver tag for | |
1277 | * the next request already, free it. | |
1278 | */ | |
1279 | if (next) | |
1280 | blk_mq_put_driver_tag(next); | |
1281 | ||
1282 | list_add(&rq->queuelist, list); | |
1283 | __blk_mq_requeue_request(rq); | |
1284 | } | |
1285 | ||
0512a75b KB |
1286 | static void blk_mq_handle_zone_resource(struct request *rq, |
1287 | struct list_head *zone_list) | |
1288 | { | |
1289 | /* | |
1290 | * If we end up here it is because we cannot dispatch a request to a | |
1291 | * specific zone due to LLD level zone-write locking or other zone | |
1292 | * related resource not being available. In this case, set the request | |
1293 | * aside in zone_list for retrying it later. | |
1294 | */ | |
1295 | list_add(&rq->queuelist, zone_list); | |
1296 | __blk_mq_requeue_request(rq); | |
1297 | } | |
1298 | ||
75383524 ML |
1299 | enum prep_dispatch { |
1300 | PREP_DISPATCH_OK, | |
1301 | PREP_DISPATCH_NO_TAG, | |
1302 | PREP_DISPATCH_NO_BUDGET, | |
1303 | }; | |
1304 | ||
1305 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1306 | bool need_budget) | |
1307 | { | |
1308 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1309 | int budget_token = -1; |
75383524 | 1310 | |
2a5a24aa ML |
1311 | if (need_budget) { |
1312 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1313 | if (budget_token < 0) { | |
1314 | blk_mq_put_driver_tag(rq); | |
1315 | return PREP_DISPATCH_NO_BUDGET; | |
1316 | } | |
1317 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1318 | } |
1319 | ||
1320 | if (!blk_mq_get_driver_tag(rq)) { | |
1321 | /* | |
1322 | * The initial allocation attempt failed, so we need to | |
1323 | * rerun the hardware queue when a tag is freed. The | |
1324 | * waitqueue takes care of that. If the queue is run | |
1325 | * before we add this entry back on the dispatch list, | |
1326 | * we'll re-run it below. | |
1327 | */ | |
1328 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1329 | /* |
1330 | * All budgets not got from this function will be put | |
1331 | * together during handling partial dispatch | |
1332 | */ | |
1333 | if (need_budget) | |
2a5a24aa | 1334 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1335 | return PREP_DISPATCH_NO_TAG; |
1336 | } | |
1337 | } | |
1338 | ||
1339 | return PREP_DISPATCH_OK; | |
1340 | } | |
1341 | ||
1fd40b5e ML |
1342 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1343 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1344 | struct list_head *list) |
1fd40b5e | 1345 | { |
2a5a24aa | 1346 | struct request *rq; |
1fd40b5e | 1347 | |
2a5a24aa ML |
1348 | list_for_each_entry(rq, list, queuelist) { |
1349 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1350 | |
2a5a24aa ML |
1351 | if (budget_token >= 0) |
1352 | blk_mq_put_dispatch_budget(q, budget_token); | |
1353 | } | |
1fd40b5e ML |
1354 | } |
1355 | ||
1f57f8d4 JA |
1356 | /* |
1357 | * Returns true if we did some work AND can potentially do more. | |
1358 | */ | |
445874e8 | 1359 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1360 | unsigned int nr_budgets) |
320ae51f | 1361 | { |
75383524 | 1362 | enum prep_dispatch prep; |
445874e8 | 1363 | struct request_queue *q = hctx->queue; |
6d6f167c | 1364 | struct request *rq, *nxt; |
fc17b653 | 1365 | int errors, queued; |
86ff7c2a | 1366 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1367 | LIST_HEAD(zone_list); |
320ae51f | 1368 | |
81380ca1 OS |
1369 | if (list_empty(list)) |
1370 | return false; | |
1371 | ||
320ae51f JA |
1372 | /* |
1373 | * Now process all the entries, sending them to the driver. | |
1374 | */ | |
93efe981 | 1375 | errors = queued = 0; |
81380ca1 | 1376 | do { |
74c45052 | 1377 | struct blk_mq_queue_data bd; |
320ae51f | 1378 | |
f04c3df3 | 1379 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1380 | |
445874e8 | 1381 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1382 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1383 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1384 | break; |
de148297 | 1385 | |
320ae51f | 1386 | list_del_init(&rq->queuelist); |
320ae51f | 1387 | |
74c45052 | 1388 | bd.rq = rq; |
113285b4 JA |
1389 | |
1390 | /* | |
1391 | * Flag last if we have no more requests, or if we have more | |
1392 | * but can't assign a driver tag to it. | |
1393 | */ | |
1394 | if (list_empty(list)) | |
1395 | bd.last = true; | |
1396 | else { | |
113285b4 | 1397 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1398 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1399 | } |
74c45052 | 1400 | |
1fd40b5e ML |
1401 | /* |
1402 | * once the request is queued to lld, no need to cover the | |
1403 | * budget any more | |
1404 | */ | |
1405 | if (nr_budgets) | |
1406 | nr_budgets--; | |
74c45052 | 1407 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1408 | switch (ret) { |
1409 | case BLK_STS_OK: | |
1410 | queued++; | |
320ae51f | 1411 | break; |
7bf13729 ML |
1412 | case BLK_STS_RESOURCE: |
1413 | case BLK_STS_DEV_RESOURCE: | |
1414 | blk_mq_handle_dev_resource(rq, list); | |
1415 | goto out; | |
1416 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1417 | /* |
1418 | * Move the request to zone_list and keep going through | |
1419 | * the dispatch list to find more requests the drive can | |
1420 | * accept. | |
1421 | */ | |
1422 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1423 | break; |
1424 | default: | |
93efe981 | 1425 | errors++; |
e21ee5a6 | 1426 | blk_mq_end_request(rq, ret); |
320ae51f | 1427 | } |
81380ca1 | 1428 | } while (!list_empty(list)); |
7bf13729 | 1429 | out: |
0512a75b KB |
1430 | if (!list_empty(&zone_list)) |
1431 | list_splice_tail_init(&zone_list, list); | |
1432 | ||
703fd1c0 | 1433 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1434 | |
632bfb63 | 1435 | /* If we didn't flush the entire list, we could have told the driver |
1436 | * there was more coming, but that turned out to be a lie. | |
1437 | */ | |
1438 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1439 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1440 | /* |
1441 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1442 | * that is where we will continue on next queue run. | |
1443 | */ | |
f04c3df3 | 1444 | if (!list_empty(list)) { |
86ff7c2a | 1445 | bool needs_restart; |
75383524 ML |
1446 | /* For non-shared tags, the RESTART check will suffice */ |
1447 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1448 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1449 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1450 | |
2a5a24aa ML |
1451 | if (nr_budgets) |
1452 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1453 | |
320ae51f | 1454 | spin_lock(&hctx->lock); |
01e99aec | 1455 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1456 | spin_unlock(&hctx->lock); |
f04c3df3 | 1457 | |
d7d8535f ML |
1458 | /* |
1459 | * Order adding requests to hctx->dispatch and checking | |
1460 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1461 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1462 | * miss the new added requests to hctx->dispatch, meantime | |
1463 | * SCHED_RESTART is observed here. | |
1464 | */ | |
1465 | smp_mb(); | |
1466 | ||
9ba52e58 | 1467 | /* |
710c785f BVA |
1468 | * If SCHED_RESTART was set by the caller of this function and |
1469 | * it is no longer set that means that it was cleared by another | |
1470 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1471 | * |
eb619fdb JA |
1472 | * If 'no_tag' is set, that means that we failed getting |
1473 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1474 | * waitqueue is no longer active, ensure that we run the queue | |
1475 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1476 | * |
710c785f BVA |
1477 | * If no I/O scheduler has been configured it is possible that |
1478 | * the hardware queue got stopped and restarted before requests | |
1479 | * were pushed back onto the dispatch list. Rerun the queue to | |
1480 | * avoid starvation. Notes: | |
1481 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1482 | * been stopped before rerunning a queue. | |
1483 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1484 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1485 | * and dm-rq. |
86ff7c2a ML |
1486 | * |
1487 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1488 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1489 | * that could otherwise occur if the queue is idle. We'll do |
1490 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1491 | */ |
86ff7c2a ML |
1492 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1493 | if (!needs_restart || | |
eb619fdb | 1494 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1495 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1496 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1497 | no_budget_avail)) | |
86ff7c2a | 1498 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1499 | |
6e768717 | 1500 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1501 | return false; |
6e768717 ML |
1502 | } else |
1503 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1504 | |
93efe981 | 1505 | return (queued + errors) != 0; |
f04c3df3 JA |
1506 | } |
1507 | ||
105663f7 AA |
1508 | /** |
1509 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1510 | * @hctx: Pointer to the hardware queue to run. | |
1511 | * | |
1512 | * Send pending requests to the hardware. | |
1513 | */ | |
6a83e74d BVA |
1514 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1515 | { | |
1516 | int srcu_idx; | |
1517 | ||
b7a71e66 JA |
1518 | /* |
1519 | * We can't run the queue inline with ints disabled. Ensure that | |
1520 | * we catch bad users of this early. | |
1521 | */ | |
1522 | WARN_ON_ONCE(in_interrupt()); | |
1523 | ||
04ced159 | 1524 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1525 | |
04ced159 JA |
1526 | hctx_lock(hctx, &srcu_idx); |
1527 | blk_mq_sched_dispatch_requests(hctx); | |
1528 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1529 | } |
1530 | ||
f82ddf19 ML |
1531 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1532 | { | |
1533 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1534 | ||
1535 | if (cpu >= nr_cpu_ids) | |
1536 | cpu = cpumask_first(hctx->cpumask); | |
1537 | return cpu; | |
1538 | } | |
1539 | ||
506e931f JA |
1540 | /* |
1541 | * It'd be great if the workqueue API had a way to pass | |
1542 | * in a mask and had some smarts for more clever placement. | |
1543 | * For now we just round-robin here, switching for every | |
1544 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1545 | */ | |
1546 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1547 | { | |
7bed4595 | 1548 | bool tried = false; |
476f8c98 | 1549 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1550 | |
b657d7e6 CH |
1551 | if (hctx->queue->nr_hw_queues == 1) |
1552 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1553 | |
1554 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1555 | select_cpu: |
476f8c98 | 1556 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1557 | cpu_online_mask); |
506e931f | 1558 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1559 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1560 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1561 | } | |
1562 | ||
7bed4595 ML |
1563 | /* |
1564 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1565 | * and it should only happen in the path of handling CPU DEAD. | |
1566 | */ | |
476f8c98 | 1567 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1568 | if (!tried) { |
1569 | tried = true; | |
1570 | goto select_cpu; | |
1571 | } | |
1572 | ||
1573 | /* | |
1574 | * Make sure to re-select CPU next time once after CPUs | |
1575 | * in hctx->cpumask become online again. | |
1576 | */ | |
476f8c98 | 1577 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1578 | hctx->next_cpu_batch = 1; |
1579 | return WORK_CPU_UNBOUND; | |
1580 | } | |
476f8c98 ML |
1581 | |
1582 | hctx->next_cpu = next_cpu; | |
1583 | return next_cpu; | |
506e931f JA |
1584 | } |
1585 | ||
105663f7 AA |
1586 | /** |
1587 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1588 | * @hctx: Pointer to the hardware queue to run. | |
1589 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1590 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1591 | * |
1592 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1593 | * with a delay of @msecs. | |
1594 | */ | |
7587a5ae BVA |
1595 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1596 | unsigned long msecs) | |
320ae51f | 1597 | { |
5435c023 | 1598 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1599 | return; |
1600 | ||
1b792f2f | 1601 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1602 | int cpu = get_cpu(); |
1603 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1604 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1605 | put_cpu(); |
398205b8 PB |
1606 | return; |
1607 | } | |
e4043dcf | 1608 | |
2a90d4aa | 1609 | put_cpu(); |
e4043dcf | 1610 | } |
398205b8 | 1611 | |
ae943d20 BVA |
1612 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1613 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1614 | } |
1615 | ||
105663f7 AA |
1616 | /** |
1617 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1618 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1619 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1620 | * |
1621 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1622 | */ | |
7587a5ae BVA |
1623 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1624 | { | |
1625 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1626 | } | |
1627 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1628 | ||
105663f7 AA |
1629 | /** |
1630 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1631 | * @hctx: Pointer to the hardware queue to run. | |
1632 | * @async: If we want to run the queue asynchronously. | |
1633 | * | |
1634 | * Check if the request queue is not in a quiesced state and if there are | |
1635 | * pending requests to be sent. If this is true, run the queue to send requests | |
1636 | * to hardware. | |
1637 | */ | |
626fb735 | 1638 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1639 | { |
24f5a90f ML |
1640 | int srcu_idx; |
1641 | bool need_run; | |
1642 | ||
1643 | /* | |
1644 | * When queue is quiesced, we may be switching io scheduler, or | |
1645 | * updating nr_hw_queues, or other things, and we can't run queue | |
1646 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1647 | * | |
1648 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1649 | * quiesced. | |
1650 | */ | |
04ced159 JA |
1651 | hctx_lock(hctx, &srcu_idx); |
1652 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1653 | blk_mq_hctx_has_pending(hctx); | |
1654 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1655 | |
626fb735 | 1656 | if (need_run) |
79f720a7 | 1657 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1658 | } |
5b727272 | 1659 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1660 | |
b6e68ee8 JK |
1661 | /* |
1662 | * Is the request queue handled by an IO scheduler that does not respect | |
1663 | * hardware queues when dispatching? | |
1664 | */ | |
1665 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
1666 | { | |
1667 | struct elevator_queue *e = q->elevator; | |
1668 | ||
1669 | if (e && e->type->ops.dispatch_request && | |
1670 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
1671 | return true; | |
1672 | return false; | |
1673 | } | |
1674 | ||
1675 | /* | |
1676 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
1677 | * scheduler. | |
1678 | */ | |
1679 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
1680 | { | |
1681 | struct blk_mq_hw_ctx *hctx; | |
1682 | ||
1683 | /* | |
1684 | * If the IO scheduler does not respect hardware queues when | |
1685 | * dispatching, we just don't bother with multiple HW queues and | |
1686 | * dispatch from hctx for the current CPU since running multiple queues | |
1687 | * just causes lock contention inside the scheduler and pointless cache | |
1688 | * bouncing. | |
1689 | */ | |
1690 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
1691 | raw_smp_processor_id()); | |
1692 | if (!blk_mq_hctx_stopped(hctx)) | |
1693 | return hctx; | |
1694 | return NULL; | |
1695 | } | |
1696 | ||
105663f7 | 1697 | /** |
24f7bb88 | 1698 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1699 | * @q: Pointer to the request queue to run. |
1700 | * @async: If we want to run the queue asynchronously. | |
1701 | */ | |
b94ec296 | 1702 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 1703 | { |
b6e68ee8 | 1704 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
1705 | int i; |
1706 | ||
b6e68ee8 JK |
1707 | sq_hctx = NULL; |
1708 | if (blk_mq_has_sqsched(q)) | |
1709 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 1710 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 1711 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 1712 | continue; |
b6e68ee8 JK |
1713 | /* |
1714 | * Dispatch from this hctx either if there's no hctx preferred | |
1715 | * by IO scheduler or if it has requests that bypass the | |
1716 | * scheduler. | |
1717 | */ | |
1718 | if (!sq_hctx || sq_hctx == hctx || | |
1719 | !list_empty_careful(&hctx->dispatch)) | |
1720 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
1721 | } |
1722 | } | |
b94ec296 | 1723 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1724 | |
b9151e7b DA |
1725 | /** |
1726 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1727 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1728 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1729 | */ |
1730 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1731 | { | |
b6e68ee8 | 1732 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
1733 | int i; |
1734 | ||
b6e68ee8 JK |
1735 | sq_hctx = NULL; |
1736 | if (blk_mq_has_sqsched(q)) | |
1737 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
1738 | queue_for_each_hw_ctx(q, hctx, i) { |
1739 | if (blk_mq_hctx_stopped(hctx)) | |
1740 | continue; | |
b6e68ee8 JK |
1741 | /* |
1742 | * Dispatch from this hctx either if there's no hctx preferred | |
1743 | * by IO scheduler or if it has requests that bypass the | |
1744 | * scheduler. | |
1745 | */ | |
1746 | if (!sq_hctx || sq_hctx == hctx || | |
1747 | !list_empty_careful(&hctx->dispatch)) | |
1748 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
1749 | } |
1750 | } | |
1751 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1752 | ||
fd001443 BVA |
1753 | /** |
1754 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1755 | * @q: request queue. | |
1756 | * | |
1757 | * The caller is responsible for serializing this function against | |
1758 | * blk_mq_{start,stop}_hw_queue(). | |
1759 | */ | |
1760 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1761 | { | |
1762 | struct blk_mq_hw_ctx *hctx; | |
1763 | int i; | |
1764 | ||
1765 | queue_for_each_hw_ctx(q, hctx, i) | |
1766 | if (blk_mq_hctx_stopped(hctx)) | |
1767 | return true; | |
1768 | ||
1769 | return false; | |
1770 | } | |
1771 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1772 | ||
39a70c76 ML |
1773 | /* |
1774 | * This function is often used for pausing .queue_rq() by driver when | |
1775 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1776 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1777 | * |
1778 | * We do not guarantee that dispatch can be drained or blocked | |
1779 | * after blk_mq_stop_hw_queue() returns. Please use | |
1780 | * blk_mq_quiesce_queue() for that requirement. | |
1781 | */ | |
2719aa21 JA |
1782 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1783 | { | |
641a9ed6 | 1784 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1785 | |
641a9ed6 | 1786 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1787 | } |
641a9ed6 | 1788 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1789 | |
39a70c76 ML |
1790 | /* |
1791 | * This function is often used for pausing .queue_rq() by driver when | |
1792 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1793 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1794 | * |
1795 | * We do not guarantee that dispatch can be drained or blocked | |
1796 | * after blk_mq_stop_hw_queues() returns. Please use | |
1797 | * blk_mq_quiesce_queue() for that requirement. | |
1798 | */ | |
2719aa21 JA |
1799 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1800 | { | |
641a9ed6 ML |
1801 | struct blk_mq_hw_ctx *hctx; |
1802 | int i; | |
1803 | ||
1804 | queue_for_each_hw_ctx(q, hctx, i) | |
1805 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1806 | } |
1807 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1808 | ||
320ae51f JA |
1809 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1810 | { | |
1811 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1812 | |
0ffbce80 | 1813 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1814 | } |
1815 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1816 | ||
2f268556 CH |
1817 | void blk_mq_start_hw_queues(struct request_queue *q) |
1818 | { | |
1819 | struct blk_mq_hw_ctx *hctx; | |
1820 | int i; | |
1821 | ||
1822 | queue_for_each_hw_ctx(q, hctx, i) | |
1823 | blk_mq_start_hw_queue(hctx); | |
1824 | } | |
1825 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1826 | ||
ae911c5e JA |
1827 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1828 | { | |
1829 | if (!blk_mq_hctx_stopped(hctx)) | |
1830 | return; | |
1831 | ||
1832 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1833 | blk_mq_run_hw_queue(hctx, async); | |
1834 | } | |
1835 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1836 | ||
1b4a3258 | 1837 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1838 | { |
1839 | struct blk_mq_hw_ctx *hctx; | |
1840 | int i; | |
1841 | ||
ae911c5e JA |
1842 | queue_for_each_hw_ctx(q, hctx, i) |
1843 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1844 | } |
1845 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1846 | ||
70f4db63 | 1847 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1848 | { |
1849 | struct blk_mq_hw_ctx *hctx; | |
1850 | ||
9f993737 | 1851 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1852 | |
21c6e939 | 1853 | /* |
15fe8a90 | 1854 | * If we are stopped, don't run the queue. |
21c6e939 | 1855 | */ |
0841031a | 1856 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 1857 | return; |
7587a5ae BVA |
1858 | |
1859 | __blk_mq_run_hw_queue(hctx); | |
1860 | } | |
1861 | ||
cfd0c552 | 1862 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1863 | struct request *rq, |
1864 | bool at_head) | |
320ae51f | 1865 | { |
e57690fe | 1866 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1867 | enum hctx_type type = hctx->type; |
e57690fe | 1868 | |
7b607814 BVA |
1869 | lockdep_assert_held(&ctx->lock); |
1870 | ||
a54895fa | 1871 | trace_block_rq_insert(rq); |
01b983c9 | 1872 | |
72a0a36e | 1873 | if (at_head) |
c16d6b5a | 1874 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1875 | else |
c16d6b5a | 1876 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1877 | } |
4bb659b1 | 1878 | |
2c3ad667 JA |
1879 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1880 | bool at_head) | |
cfd0c552 ML |
1881 | { |
1882 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1883 | ||
7b607814 BVA |
1884 | lockdep_assert_held(&ctx->lock); |
1885 | ||
e57690fe | 1886 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1887 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1888 | } |
1889 | ||
105663f7 AA |
1890 | /** |
1891 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1892 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 1893 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
1894 | * @run_queue: If we should run the hardware queue after inserting the request. |
1895 | * | |
157f377b JA |
1896 | * Should only be used carefully, when the caller knows we want to |
1897 | * bypass a potential IO scheduler on the target device. | |
1898 | */ | |
01e99aec ML |
1899 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1900 | bool run_queue) | |
157f377b | 1901 | { |
ea4f995e | 1902 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1903 | |
1904 | spin_lock(&hctx->lock); | |
01e99aec ML |
1905 | if (at_head) |
1906 | list_add(&rq->queuelist, &hctx->dispatch); | |
1907 | else | |
1908 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1909 | spin_unlock(&hctx->lock); |
1910 | ||
b0850297 ML |
1911 | if (run_queue) |
1912 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1913 | } |
1914 | ||
bd166ef1 JA |
1915 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1916 | struct list_head *list) | |
320ae51f JA |
1917 | |
1918 | { | |
3f0cedc7 | 1919 | struct request *rq; |
c16d6b5a | 1920 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1921 | |
320ae51f JA |
1922 | /* |
1923 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1924 | * offline now | |
1925 | */ | |
3f0cedc7 | 1926 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1927 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 1928 | trace_block_rq_insert(rq); |
320ae51f | 1929 | } |
3f0cedc7 ML |
1930 | |
1931 | spin_lock(&ctx->lock); | |
c16d6b5a | 1932 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1933 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1934 | spin_unlock(&ctx->lock); |
320ae51f JA |
1935 | } |
1936 | ||
4f0f586b ST |
1937 | static int plug_rq_cmp(void *priv, const struct list_head *a, |
1938 | const struct list_head *b) | |
320ae51f JA |
1939 | { |
1940 | struct request *rqa = container_of(a, struct request, queuelist); | |
1941 | struct request *rqb = container_of(b, struct request, queuelist); | |
1942 | ||
7d30a621 PB |
1943 | if (rqa->mq_ctx != rqb->mq_ctx) |
1944 | return rqa->mq_ctx > rqb->mq_ctx; | |
1945 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1946 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1947 | |
1948 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1949 | } |
1950 | ||
1951 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1952 | { | |
320ae51f | 1953 | LIST_HEAD(list); |
320ae51f | 1954 | |
95ed0c5b PB |
1955 | if (list_empty(&plug->mq_list)) |
1956 | return; | |
320ae51f JA |
1957 | list_splice_init(&plug->mq_list, &list); |
1958 | ||
ce5b009c JA |
1959 | if (plug->rq_count > 2 && plug->multiple_queues) |
1960 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1961 | |
bcc816df DZ |
1962 | plug->rq_count = 0; |
1963 | ||
95ed0c5b PB |
1964 | do { |
1965 | struct list_head rq_list; | |
1966 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1967 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1968 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1969 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1970 | unsigned int depth = 1; | |
1971 | ||
1972 | list_for_each_continue(pos, &list) { | |
1973 | rq = list_entry_rq(pos); | |
1974 | BUG_ON(!rq->q); | |
1975 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1976 | break; | |
1977 | depth++; | |
320ae51f JA |
1978 | } |
1979 | ||
95ed0c5b PB |
1980 | list_cut_before(&rq_list, &list, pos); |
1981 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1982 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1983 | from_schedule); |
95ed0c5b | 1984 | } while(!list_empty(&list)); |
320ae51f JA |
1985 | } |
1986 | ||
14ccb66b CH |
1987 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1988 | unsigned int nr_segs) | |
320ae51f | 1989 | { |
93f221ae EB |
1990 | int err; |
1991 | ||
f924cdde CH |
1992 | if (bio->bi_opf & REQ_RAHEAD) |
1993 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1994 | ||
1995 | rq->__sector = bio->bi_iter.bi_sector; | |
1996 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1997 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1998 | |
1999 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
2000 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
2001 | WARN_ON_ONCE(err); | |
4b570521 | 2002 | |
b5af37ab | 2003 | blk_account_io_start(rq); |
320ae51f JA |
2004 | } |
2005 | ||
0f95549c MS |
2006 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
2007 | struct request *rq, | |
be94f058 | 2008 | blk_qc_t *cookie, bool last) |
f984df1f | 2009 | { |
f984df1f | 2010 | struct request_queue *q = rq->q; |
f984df1f SL |
2011 | struct blk_mq_queue_data bd = { |
2012 | .rq = rq, | |
be94f058 | 2013 | .last = last, |
f984df1f | 2014 | }; |
bd166ef1 | 2015 | blk_qc_t new_cookie; |
f06345ad | 2016 | blk_status_t ret; |
0f95549c MS |
2017 | |
2018 | new_cookie = request_to_qc_t(hctx, rq); | |
2019 | ||
2020 | /* | |
2021 | * For OK queue, we are done. For error, caller may kill it. | |
2022 | * Any other error (busy), just add it to our list as we | |
2023 | * previously would have done. | |
2024 | */ | |
2025 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
2026 | switch (ret) { | |
2027 | case BLK_STS_OK: | |
6ce3dd6e | 2028 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2029 | *cookie = new_cookie; |
2030 | break; | |
2031 | case BLK_STS_RESOURCE: | |
86ff7c2a | 2032 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 2033 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
2034 | __blk_mq_requeue_request(rq); |
2035 | break; | |
2036 | default: | |
6ce3dd6e | 2037 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2038 | *cookie = BLK_QC_T_NONE; |
2039 | break; | |
2040 | } | |
2041 | ||
2042 | return ret; | |
2043 | } | |
2044 | ||
fd9c40f6 | 2045 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2046 | struct request *rq, |
396eaf21 | 2047 | blk_qc_t *cookie, |
fd9c40f6 | 2048 | bool bypass_insert, bool last) |
0f95549c MS |
2049 | { |
2050 | struct request_queue *q = rq->q; | |
d964f04a | 2051 | bool run_queue = true; |
2a5a24aa | 2052 | int budget_token; |
d964f04a | 2053 | |
23d4ee19 | 2054 | /* |
fd9c40f6 | 2055 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2056 | * |
fd9c40f6 BVA |
2057 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2058 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2059 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2060 | */ |
fd9c40f6 | 2061 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2062 | run_queue = false; |
fd9c40f6 BVA |
2063 | bypass_insert = false; |
2064 | goto insert; | |
d964f04a | 2065 | } |
f984df1f | 2066 | |
fd9c40f6 BVA |
2067 | if (q->elevator && !bypass_insert) |
2068 | goto insert; | |
2253efc8 | 2069 | |
2a5a24aa ML |
2070 | budget_token = blk_mq_get_dispatch_budget(q); |
2071 | if (budget_token < 0) | |
fd9c40f6 | 2072 | goto insert; |
bd166ef1 | 2073 | |
2a5a24aa ML |
2074 | blk_mq_set_rq_budget_token(rq, budget_token); |
2075 | ||
8ab6bb9e | 2076 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2077 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2078 | goto insert; |
88022d72 | 2079 | } |
de148297 | 2080 | |
fd9c40f6 BVA |
2081 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2082 | insert: | |
2083 | if (bypass_insert) | |
2084 | return BLK_STS_RESOURCE; | |
2085 | ||
db03f88f ML |
2086 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2087 | ||
fd9c40f6 BVA |
2088 | return BLK_STS_OK; |
2089 | } | |
2090 | ||
105663f7 AA |
2091 | /** |
2092 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2093 | * @hctx: Pointer of the associated hardware queue. | |
2094 | * @rq: Pointer to request to be sent. | |
2095 | * @cookie: Request queue cookie. | |
2096 | * | |
2097 | * If the device has enough resources to accept a new request now, send the | |
2098 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2099 | * we can try send it another time in the future. Requests inserted at this | |
2100 | * queue have higher priority. | |
2101 | */ | |
fd9c40f6 BVA |
2102 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2103 | struct request *rq, blk_qc_t *cookie) | |
2104 | { | |
2105 | blk_status_t ret; | |
2106 | int srcu_idx; | |
2107 | ||
2108 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2109 | ||
2110 | hctx_lock(hctx, &srcu_idx); | |
2111 | ||
2112 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2113 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2114 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2115 | else if (ret != BLK_STS_OK) |
2116 | blk_mq_end_request(rq, ret); | |
2117 | ||
2118 | hctx_unlock(hctx, srcu_idx); | |
2119 | } | |
2120 | ||
2121 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2122 | { | |
2123 | blk_status_t ret; | |
2124 | int srcu_idx; | |
2125 | blk_qc_t unused_cookie; | |
2126 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2127 | ||
2128 | hctx_lock(hctx, &srcu_idx); | |
2129 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2130 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2131 | |
2132 | return ret; | |
5eb6126e CH |
2133 | } |
2134 | ||
6ce3dd6e ML |
2135 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2136 | struct list_head *list) | |
2137 | { | |
536167d4 | 2138 | int queued = 0; |
632bfb63 | 2139 | int errors = 0; |
536167d4 | 2140 | |
6ce3dd6e | 2141 | while (!list_empty(list)) { |
fd9c40f6 | 2142 | blk_status_t ret; |
6ce3dd6e ML |
2143 | struct request *rq = list_first_entry(list, struct request, |
2144 | queuelist); | |
2145 | ||
2146 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2147 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2148 | if (ret != BLK_STS_OK) { | |
2149 | if (ret == BLK_STS_RESOURCE || | |
2150 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2151 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2152 | list_empty(list)); |
fd9c40f6 BVA |
2153 | break; |
2154 | } | |
2155 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2156 | errors++; |
536167d4 KB |
2157 | } else |
2158 | queued++; | |
6ce3dd6e | 2159 | } |
d666ba98 JA |
2160 | |
2161 | /* | |
2162 | * If we didn't flush the entire list, we could have told | |
2163 | * the driver there was more coming, but that turned out to | |
2164 | * be a lie. | |
2165 | */ | |
632bfb63 | 2166 | if ((!list_empty(list) || errors) && |
2167 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2168 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2169 | } |
2170 | ||
ce5b009c JA |
2171 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2172 | { | |
2173 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2174 | plug->rq_count++; | |
2175 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2176 | struct request *tmp; | |
2177 | ||
2178 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2179 | queuelist); | |
2180 | if (tmp->q != rq->q) | |
2181 | plug->multiple_queues = true; | |
2182 | } | |
2183 | } | |
2184 | ||
7f2a6a69 | 2185 | /* |
ba0ffdd8 | 2186 | * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple |
7f2a6a69 SL |
2187 | * queues. This is important for md arrays to benefit from merging |
2188 | * requests. | |
2189 | */ | |
2190 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2191 | { | |
2192 | if (plug->multiple_queues) | |
ba0ffdd8 | 2193 | return BLK_MAX_REQUEST_COUNT * 2; |
7f2a6a69 SL |
2194 | return BLK_MAX_REQUEST_COUNT; |
2195 | } | |
2196 | ||
105663f7 | 2197 | /** |
c62b37d9 | 2198 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2199 | * @bio: Bio pointer. |
2200 | * | |
2201 | * Builds up a request structure from @q and @bio and send to the device. The | |
2202 | * request may not be queued directly to hardware if: | |
2203 | * * This request can be merged with another one | |
2204 | * * We want to place request at plug queue for possible future merging | |
2205 | * * There is an IO scheduler active at this queue | |
2206 | * | |
2207 | * It will not queue the request if there is an error with the bio, or at the | |
2208 | * request creation. | |
2209 | * | |
2210 | * Returns: Request queue cookie. | |
2211 | */ | |
c62b37d9 | 2212 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2213 | { |
309dca30 | 2214 | struct request_queue *q = bio->bi_bdev->bd_disk->queue; |
ef295ecf | 2215 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2216 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
07068d5b | 2217 | struct request *rq; |
f984df1f | 2218 | struct blk_plug *plug; |
5b3f341f | 2219 | struct request *same_queue_rq = NULL; |
14ccb66b | 2220 | unsigned int nr_segs; |
7b371636 | 2221 | blk_qc_t cookie; |
a892c8d5 | 2222 | blk_status_t ret; |
cc29e1bf | 2223 | bool hipri; |
07068d5b JA |
2224 | |
2225 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2226 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2227 | |
e23947bd | 2228 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2229 | goto queue_exit; |
07068d5b | 2230 | |
87c279e6 | 2231 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2232 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2233 | goto queue_exit; |
f984df1f | 2234 | |
14ccb66b | 2235 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2236 | goto queue_exit; |
bd166ef1 | 2237 | |
d5337560 | 2238 | rq_qos_throttle(q, bio); |
87760e5e | 2239 | |
cc29e1bf JX |
2240 | hipri = bio->bi_opf & REQ_HIPRI; |
2241 | ||
47c122e3 JA |
2242 | plug = blk_mq_plug(q, bio); |
2243 | if (plug && plug->cached_rq) { | |
2244 | rq = plug->cached_rq; | |
2245 | plug->cached_rq = rq->rq_next; | |
2246 | INIT_LIST_HEAD(&rq->queuelist); | |
47c122e3 | 2247 | } else { |
0f38d766 CH |
2248 | struct blk_mq_alloc_data data = { |
2249 | .q = q, | |
2250 | .nr_tags = 1, | |
2251 | .cmd_flags = bio->bi_opf, | |
2252 | }; | |
2253 | ||
47c122e3 JA |
2254 | if (plug) { |
2255 | data.nr_tags = plug->nr_ios; | |
2256 | plug->nr_ios = 1; | |
2257 | data.cached_rq = &plug->cached_rq; | |
2258 | } | |
b90cfaed | 2259 | rq = __blk_mq_alloc_requests(&data); |
47c122e3 JA |
2260 | if (unlikely(!rq)) { |
2261 | rq_qos_cleanup(q, bio); | |
2262 | if (bio->bi_opf & REQ_NOWAIT) | |
2263 | bio_wouldblock_error(bio); | |
2264 | goto queue_exit; | |
2265 | } | |
87760e5e JA |
2266 | } |
2267 | ||
e8a676d6 | 2268 | trace_block_getrq(bio); |
d6f1dda2 | 2269 | |
c1c80384 | 2270 | rq_qos_track(q, rq, bio); |
07068d5b | 2271 | |
0f38d766 | 2272 | cookie = request_to_qc_t(rq->mq_hctx, rq); |
07068d5b | 2273 | |
970d168d BVA |
2274 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2275 | ||
a892c8d5 ST |
2276 | ret = blk_crypto_init_request(rq); |
2277 | if (ret != BLK_STS_OK) { | |
2278 | bio->bi_status = ret; | |
2279 | bio_endio(bio); | |
2280 | blk_mq_free_request(rq); | |
2281 | return BLK_QC_T_NONE; | |
2282 | } | |
2283 | ||
07068d5b | 2284 | if (unlikely(is_flush_fua)) { |
4a60f360 | 2285 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
105663f7 | 2286 | /* Bypass scheduler for flush requests */ |
923218f6 | 2287 | blk_insert_flush(rq); |
4a60f360 | 2288 | blk_mq_run_hw_queue(hctx, true); |
03f26d8f | 2289 | } else if (plug && (q->nr_hw_queues == 1 || |
079a2e3e | 2290 | blk_mq_is_shared_tags(rq->mq_hctx->flags) || |
03f26d8f | 2291 | q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) { |
b2c5d16b JA |
2292 | /* |
2293 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2294 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2295 | * |
2296 | * Use normal plugging if this disk is slow HDD, as sequential | |
2297 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2298 | */ |
5f0ed774 | 2299 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2300 | struct request *last = NULL; |
2301 | ||
676d0607 | 2302 | if (!request_count) |
e6c4438b | 2303 | trace_block_plug(q); |
600271d9 SL |
2304 | else |
2305 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2306 | |
7f2a6a69 | 2307 | if (request_count >= blk_plug_max_rq_count(plug) || (last && |
600271d9 | 2308 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
e6c4438b JM |
2309 | blk_flush_plug_list(plug, false); |
2310 | trace_block_plug(q); | |
320ae51f | 2311 | } |
b094f89c | 2312 | |
ce5b009c | 2313 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2314 | } else if (q->elevator) { |
105663f7 | 2315 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2316 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2317 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2318 | /* |
6a83e74d | 2319 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2320 | * Otherwise the existing request in the plug list will be |
2321 | * issued. So the plug list will have one request at most | |
2299722c CH |
2322 | * The plug list might get flushed before this. If that happens, |
2323 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2324 | */ |
2299722c CH |
2325 | if (list_empty(&plug->mq_list)) |
2326 | same_queue_rq = NULL; | |
4711b573 | 2327 | if (same_queue_rq) { |
2299722c | 2328 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2329 | plug->rq_count--; |
2330 | } | |
ce5b009c | 2331 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2332 | trace_block_plug(q); |
2299722c | 2333 | |
dad7a3be | 2334 | if (same_queue_rq) { |
ff3b74b8 | 2335 | trace_block_unplug(q, 1, true); |
0f38d766 CH |
2336 | blk_mq_try_issue_directly(same_queue_rq->mq_hctx, |
2337 | same_queue_rq, &cookie); | |
dad7a3be | 2338 | } |
a12de1d4 | 2339 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
0f38d766 | 2340 | !rq->mq_hctx->dispatch_busy) { |
105663f7 AA |
2341 | /* |
2342 | * There is no scheduler and we can try to send directly | |
2343 | * to the hardware. | |
2344 | */ | |
0f38d766 | 2345 | blk_mq_try_issue_directly(rq->mq_hctx, rq, &cookie); |
ab42f35d | 2346 | } else { |
105663f7 | 2347 | /* Default case. */ |
8fa9f556 | 2348 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2349 | } |
320ae51f | 2350 | |
cc29e1bf JX |
2351 | if (!hipri) |
2352 | return BLK_QC_T_NONE; | |
7b371636 | 2353 | return cookie; |
ac7c5675 CH |
2354 | queue_exit: |
2355 | blk_queue_exit(q); | |
2356 | return BLK_QC_T_NONE; | |
320ae51f JA |
2357 | } |
2358 | ||
bd63141d ML |
2359 | static size_t order_to_size(unsigned int order) |
2360 | { | |
2361 | return (size_t)PAGE_SIZE << order; | |
2362 | } | |
2363 | ||
2364 | /* called before freeing request pool in @tags */ | |
f32e4eaf JG |
2365 | static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, |
2366 | struct blk_mq_tags *tags) | |
bd63141d | 2367 | { |
bd63141d ML |
2368 | struct page *page; |
2369 | unsigned long flags; | |
2370 | ||
4f245d5b JG |
2371 | /* There is no need to clear a driver tags own mapping */ |
2372 | if (drv_tags == tags) | |
2373 | return; | |
2374 | ||
bd63141d ML |
2375 | list_for_each_entry(page, &tags->page_list, lru) { |
2376 | unsigned long start = (unsigned long)page_address(page); | |
2377 | unsigned long end = start + order_to_size(page->private); | |
2378 | int i; | |
2379 | ||
f32e4eaf | 2380 | for (i = 0; i < drv_tags->nr_tags; i++) { |
bd63141d ML |
2381 | struct request *rq = drv_tags->rqs[i]; |
2382 | unsigned long rq_addr = (unsigned long)rq; | |
2383 | ||
2384 | if (rq_addr >= start && rq_addr < end) { | |
2385 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
2386 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
2387 | } | |
2388 | } | |
2389 | } | |
2390 | ||
2391 | /* | |
2392 | * Wait until all pending iteration is done. | |
2393 | * | |
2394 | * Request reference is cleared and it is guaranteed to be observed | |
2395 | * after the ->lock is released. | |
2396 | */ | |
2397 | spin_lock_irqsave(&drv_tags->lock, flags); | |
2398 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
2399 | } | |
2400 | ||
cc71a6f4 JA |
2401 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2402 | unsigned int hctx_idx) | |
95363efd | 2403 | { |
f32e4eaf | 2404 | struct blk_mq_tags *drv_tags; |
e9b267d9 | 2405 | struct page *page; |
320ae51f | 2406 | |
079a2e3e JG |
2407 | if (blk_mq_is_shared_tags(set->flags)) |
2408 | drv_tags = set->shared_tags; | |
e155b0c2 JG |
2409 | else |
2410 | drv_tags = set->tags[hctx_idx]; | |
f32e4eaf | 2411 | |
65de57bb | 2412 | if (tags->static_rqs && set->ops->exit_request) { |
e9b267d9 | 2413 | int i; |
320ae51f | 2414 | |
24d2f903 | 2415 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2416 | struct request *rq = tags->static_rqs[i]; |
2417 | ||
2418 | if (!rq) | |
e9b267d9 | 2419 | continue; |
d6296d39 | 2420 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2421 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2422 | } |
320ae51f | 2423 | } |
320ae51f | 2424 | |
f32e4eaf | 2425 | blk_mq_clear_rq_mapping(drv_tags, tags); |
bd63141d | 2426 | |
24d2f903 CH |
2427 | while (!list_empty(&tags->page_list)) { |
2428 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2429 | list_del_init(&page->lru); |
f75782e4 CM |
2430 | /* |
2431 | * Remove kmemleak object previously allocated in | |
273938bf | 2432 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2433 | */ |
2434 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2435 | __free_pages(page, page->private); |
2436 | } | |
cc71a6f4 | 2437 | } |
320ae51f | 2438 | |
e155b0c2 | 2439 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
cc71a6f4 | 2440 | { |
24d2f903 | 2441 | kfree(tags->rqs); |
cc71a6f4 | 2442 | tags->rqs = NULL; |
2af8cbe3 JA |
2443 | kfree(tags->static_rqs); |
2444 | tags->static_rqs = NULL; | |
320ae51f | 2445 | |
e155b0c2 | 2446 | blk_mq_free_tags(tags); |
320ae51f JA |
2447 | } |
2448 | ||
63064be1 JG |
2449 | static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2450 | unsigned int hctx_idx, | |
2451 | unsigned int nr_tags, | |
e155b0c2 | 2452 | unsigned int reserved_tags) |
320ae51f | 2453 | { |
24d2f903 | 2454 | struct blk_mq_tags *tags; |
59f082e4 | 2455 | int node; |
320ae51f | 2456 | |
7d76f856 | 2457 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2458 | if (node == NUMA_NO_NODE) |
2459 | node = set->numa_node; | |
2460 | ||
e155b0c2 JG |
2461 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, |
2462 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
2463 | if (!tags) |
2464 | return NULL; | |
320ae51f | 2465 | |
590b5b7d | 2466 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2467 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2468 | node); |
24d2f903 | 2469 | if (!tags->rqs) { |
e155b0c2 | 2470 | blk_mq_free_tags(tags); |
24d2f903 CH |
2471 | return NULL; |
2472 | } | |
320ae51f | 2473 | |
590b5b7d KC |
2474 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2475 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2476 | node); | |
2af8cbe3 JA |
2477 | if (!tags->static_rqs) { |
2478 | kfree(tags->rqs); | |
e155b0c2 | 2479 | blk_mq_free_tags(tags); |
2af8cbe3 JA |
2480 | return NULL; |
2481 | } | |
2482 | ||
cc71a6f4 JA |
2483 | return tags; |
2484 | } | |
2485 | ||
1d9bd516 TH |
2486 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2487 | unsigned int hctx_idx, int node) | |
2488 | { | |
2489 | int ret; | |
2490 | ||
2491 | if (set->ops->init_request) { | |
2492 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2493 | if (ret) | |
2494 | return ret; | |
2495 | } | |
2496 | ||
12f5b931 | 2497 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2498 | return 0; |
2499 | } | |
2500 | ||
63064be1 JG |
2501 | static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, |
2502 | struct blk_mq_tags *tags, | |
2503 | unsigned int hctx_idx, unsigned int depth) | |
cc71a6f4 JA |
2504 | { |
2505 | unsigned int i, j, entries_per_page, max_order = 4; | |
2506 | size_t rq_size, left; | |
59f082e4 SL |
2507 | int node; |
2508 | ||
7d76f856 | 2509 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2510 | if (node == NUMA_NO_NODE) |
2511 | node = set->numa_node; | |
cc71a6f4 JA |
2512 | |
2513 | INIT_LIST_HEAD(&tags->page_list); | |
2514 | ||
320ae51f JA |
2515 | /* |
2516 | * rq_size is the size of the request plus driver payload, rounded | |
2517 | * to the cacheline size | |
2518 | */ | |
24d2f903 | 2519 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2520 | cache_line_size()); |
cc71a6f4 | 2521 | left = rq_size * depth; |
320ae51f | 2522 | |
cc71a6f4 | 2523 | for (i = 0; i < depth; ) { |
320ae51f JA |
2524 | int this_order = max_order; |
2525 | struct page *page; | |
2526 | int to_do; | |
2527 | void *p; | |
2528 | ||
b3a834b1 | 2529 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2530 | this_order--; |
2531 | ||
2532 | do { | |
59f082e4 | 2533 | page = alloc_pages_node(node, |
36e1f3d1 | 2534 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2535 | this_order); |
320ae51f JA |
2536 | if (page) |
2537 | break; | |
2538 | if (!this_order--) | |
2539 | break; | |
2540 | if (order_to_size(this_order) < rq_size) | |
2541 | break; | |
2542 | } while (1); | |
2543 | ||
2544 | if (!page) | |
24d2f903 | 2545 | goto fail; |
320ae51f JA |
2546 | |
2547 | page->private = this_order; | |
24d2f903 | 2548 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2549 | |
2550 | p = page_address(page); | |
f75782e4 CM |
2551 | /* |
2552 | * Allow kmemleak to scan these pages as they contain pointers | |
2553 | * to additional allocations like via ops->init_request(). | |
2554 | */ | |
36e1f3d1 | 2555 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2556 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2557 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2558 | left -= to_do * rq_size; |
2559 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2560 | struct request *rq = p; |
2561 | ||
2562 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2563 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2564 | tags->static_rqs[i] = NULL; | |
2565 | goto fail; | |
e9b267d9 CH |
2566 | } |
2567 | ||
320ae51f JA |
2568 | p += rq_size; |
2569 | i++; | |
2570 | } | |
2571 | } | |
cc71a6f4 | 2572 | return 0; |
320ae51f | 2573 | |
24d2f903 | 2574 | fail: |
cc71a6f4 JA |
2575 | blk_mq_free_rqs(set, tags, hctx_idx); |
2576 | return -ENOMEM; | |
320ae51f JA |
2577 | } |
2578 | ||
bf0beec0 ML |
2579 | struct rq_iter_data { |
2580 | struct blk_mq_hw_ctx *hctx; | |
2581 | bool has_rq; | |
2582 | }; | |
2583 | ||
2584 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2585 | { | |
2586 | struct rq_iter_data *iter_data = data; | |
2587 | ||
2588 | if (rq->mq_hctx != iter_data->hctx) | |
2589 | return true; | |
2590 | iter_data->has_rq = true; | |
2591 | return false; | |
2592 | } | |
2593 | ||
2594 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2595 | { | |
2596 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2597 | hctx->sched_tags : hctx->tags; | |
2598 | struct rq_iter_data data = { | |
2599 | .hctx = hctx, | |
2600 | }; | |
2601 | ||
2602 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2603 | return data.has_rq; | |
2604 | } | |
2605 | ||
2606 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2607 | struct blk_mq_hw_ctx *hctx) | |
2608 | { | |
2609 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2610 | return false; | |
2611 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2612 | return false; | |
2613 | return true; | |
2614 | } | |
2615 | ||
2616 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2617 | { | |
2618 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2619 | struct blk_mq_hw_ctx, cpuhp_online); | |
2620 | ||
2621 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2622 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2623 | return 0; | |
2624 | ||
2625 | /* | |
2626 | * Prevent new request from being allocated on the current hctx. | |
2627 | * | |
2628 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2629 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2630 | * seen once we return from the tag allocator. | |
2631 | */ | |
2632 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2633 | smp_mb__after_atomic(); | |
2634 | ||
2635 | /* | |
2636 | * Try to grab a reference to the queue and wait for any outstanding | |
2637 | * requests. If we could not grab a reference the queue has been | |
2638 | * frozen and there are no requests. | |
2639 | */ | |
2640 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2641 | while (blk_mq_hctx_has_requests(hctx)) | |
2642 | msleep(5); | |
2643 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2644 | } | |
2645 | ||
2646 | return 0; | |
2647 | } | |
2648 | ||
2649 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2650 | { | |
2651 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2652 | struct blk_mq_hw_ctx, cpuhp_online); | |
2653 | ||
2654 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2655 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2656 | return 0; | |
2657 | } | |
2658 | ||
e57690fe JA |
2659 | /* |
2660 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2661 | * software queue to the hw queue dispatch list, and ensure that it | |
2662 | * gets run. | |
2663 | */ | |
9467f859 | 2664 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2665 | { |
9467f859 | 2666 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2667 | struct blk_mq_ctx *ctx; |
2668 | LIST_HEAD(tmp); | |
c16d6b5a | 2669 | enum hctx_type type; |
484b4061 | 2670 | |
9467f859 | 2671 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2672 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2673 | return 0; | |
2674 | ||
e57690fe | 2675 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2676 | type = hctx->type; |
484b4061 JA |
2677 | |
2678 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2679 | if (!list_empty(&ctx->rq_lists[type])) { |
2680 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2681 | blk_mq_hctx_clear_pending(hctx, ctx); |
2682 | } | |
2683 | spin_unlock(&ctx->lock); | |
2684 | ||
2685 | if (list_empty(&tmp)) | |
9467f859 | 2686 | return 0; |
484b4061 | 2687 | |
e57690fe JA |
2688 | spin_lock(&hctx->lock); |
2689 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2690 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2691 | |
2692 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2693 | return 0; |
484b4061 JA |
2694 | } |
2695 | ||
9467f859 | 2696 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2697 | { |
bf0beec0 ML |
2698 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2699 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2700 | &hctx->cpuhp_online); | |
9467f859 TG |
2701 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2702 | &hctx->cpuhp_dead); | |
484b4061 JA |
2703 | } |
2704 | ||
364b6181 ML |
2705 | /* |
2706 | * Before freeing hw queue, clearing the flush request reference in | |
2707 | * tags->rqs[] for avoiding potential UAF. | |
2708 | */ | |
2709 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
2710 | unsigned int queue_depth, struct request *flush_rq) | |
2711 | { | |
2712 | int i; | |
2713 | unsigned long flags; | |
2714 | ||
2715 | /* The hw queue may not be mapped yet */ | |
2716 | if (!tags) | |
2717 | return; | |
2718 | ||
2719 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
2720 | ||
2721 | for (i = 0; i < queue_depth; i++) | |
2722 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
2723 | ||
2724 | /* | |
2725 | * Wait until all pending iteration is done. | |
2726 | * | |
2727 | * Request reference is cleared and it is guaranteed to be observed | |
2728 | * after the ->lock is released. | |
2729 | */ | |
2730 | spin_lock_irqsave(&tags->lock, flags); | |
2731 | spin_unlock_irqrestore(&tags->lock, flags); | |
2732 | } | |
2733 | ||
c3b4afca | 2734 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2735 | static void blk_mq_exit_hctx(struct request_queue *q, |
2736 | struct blk_mq_tag_set *set, | |
2737 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2738 | { | |
364b6181 ML |
2739 | struct request *flush_rq = hctx->fq->flush_rq; |
2740 | ||
8ab0b7dc ML |
2741 | if (blk_mq_hw_queue_mapped(hctx)) |
2742 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2743 | |
364b6181 ML |
2744 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
2745 | set->queue_depth, flush_rq); | |
f70ced09 | 2746 | if (set->ops->exit_request) |
364b6181 | 2747 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 2748 | |
08e98fc6 ML |
2749 | if (set->ops->exit_hctx) |
2750 | set->ops->exit_hctx(hctx, hctx_idx); | |
2751 | ||
9467f859 | 2752 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2753 | |
2754 | spin_lock(&q->unused_hctx_lock); | |
2755 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2756 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2757 | } |
2758 | ||
624dbe47 ML |
2759 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2760 | struct blk_mq_tag_set *set, int nr_queue) | |
2761 | { | |
2762 | struct blk_mq_hw_ctx *hctx; | |
2763 | unsigned int i; | |
2764 | ||
2765 | queue_for_each_hw_ctx(q, hctx, i) { | |
2766 | if (i == nr_queue) | |
2767 | break; | |
477e19de | 2768 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2769 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2770 | } |
624dbe47 ML |
2771 | } |
2772 | ||
7c6c5b7c ML |
2773 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2774 | { | |
2775 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2776 | ||
2777 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2778 | __alignof__(struct blk_mq_hw_ctx)) != | |
2779 | sizeof(struct blk_mq_hw_ctx)); | |
2780 | ||
2781 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2782 | hw_ctx_size += sizeof(struct srcu_struct); | |
2783 | ||
2784 | return hw_ctx_size; | |
2785 | } | |
2786 | ||
08e98fc6 ML |
2787 | static int blk_mq_init_hctx(struct request_queue *q, |
2788 | struct blk_mq_tag_set *set, | |
2789 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2790 | { |
7c6c5b7c ML |
2791 | hctx->queue_num = hctx_idx; |
2792 | ||
bf0beec0 ML |
2793 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2794 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2795 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2796 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2797 | ||
2798 | hctx->tags = set->tags[hctx_idx]; | |
2799 | ||
2800 | if (set->ops->init_hctx && | |
2801 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2802 | goto unregister_cpu_notifier; | |
08e98fc6 | 2803 | |
7c6c5b7c ML |
2804 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2805 | hctx->numa_node)) | |
2806 | goto exit_hctx; | |
2807 | return 0; | |
2808 | ||
2809 | exit_hctx: | |
2810 | if (set->ops->exit_hctx) | |
2811 | set->ops->exit_hctx(hctx, hctx_idx); | |
2812 | unregister_cpu_notifier: | |
2813 | blk_mq_remove_cpuhp(hctx); | |
2814 | return -1; | |
2815 | } | |
2816 | ||
2817 | static struct blk_mq_hw_ctx * | |
2818 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2819 | int node) | |
2820 | { | |
2821 | struct blk_mq_hw_ctx *hctx; | |
2822 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2823 | ||
2824 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2825 | if (!hctx) | |
2826 | goto fail_alloc_hctx; | |
2827 | ||
2828 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2829 | goto free_hctx; | |
2830 | ||
2831 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2832 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2833 | node = set->numa_node; |
2834 | hctx->numa_node = node; | |
08e98fc6 | 2835 | |
9f993737 | 2836 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2837 | spin_lock_init(&hctx->lock); |
2838 | INIT_LIST_HEAD(&hctx->dispatch); | |
2839 | hctx->queue = q; | |
51db1c37 | 2840 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2841 | |
2f8f1336 ML |
2842 | INIT_LIST_HEAD(&hctx->hctx_list); |
2843 | ||
320ae51f | 2844 | /* |
08e98fc6 ML |
2845 | * Allocate space for all possible cpus to avoid allocation at |
2846 | * runtime | |
320ae51f | 2847 | */ |
d904bfa7 | 2848 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2849 | gfp, node); |
08e98fc6 | 2850 | if (!hctx->ctxs) |
7c6c5b7c | 2851 | goto free_cpumask; |
320ae51f | 2852 | |
5b202853 | 2853 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 2854 | gfp, node, false, false)) |
08e98fc6 | 2855 | goto free_ctxs; |
08e98fc6 | 2856 | hctx->nr_ctx = 0; |
320ae51f | 2857 | |
5815839b | 2858 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2859 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2860 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2861 | ||
754a1572 | 2862 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2863 | if (!hctx->fq) |
7c6c5b7c | 2864 | goto free_bitmap; |
320ae51f | 2865 | |
6a83e74d | 2866 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2867 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2868 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2869 | |
7c6c5b7c | 2870 | return hctx; |
320ae51f | 2871 | |
08e98fc6 | 2872 | free_bitmap: |
88459642 | 2873 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2874 | free_ctxs: |
2875 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2876 | free_cpumask: |
2877 | free_cpumask_var(hctx->cpumask); | |
2878 | free_hctx: | |
2879 | kfree(hctx); | |
2880 | fail_alloc_hctx: | |
2881 | return NULL; | |
08e98fc6 | 2882 | } |
320ae51f | 2883 | |
320ae51f JA |
2884 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2885 | unsigned int nr_hw_queues) | |
2886 | { | |
b3c661b1 JA |
2887 | struct blk_mq_tag_set *set = q->tag_set; |
2888 | unsigned int i, j; | |
320ae51f JA |
2889 | |
2890 | for_each_possible_cpu(i) { | |
2891 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2892 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2893 | int k; |
320ae51f | 2894 | |
320ae51f JA |
2895 | __ctx->cpu = i; |
2896 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2897 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2898 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2899 | ||
320ae51f JA |
2900 | __ctx->queue = q; |
2901 | ||
320ae51f JA |
2902 | /* |
2903 | * Set local node, IFF we have more than one hw queue. If | |
2904 | * not, we remain on the home node of the device | |
2905 | */ | |
b3c661b1 JA |
2906 | for (j = 0; j < set->nr_maps; j++) { |
2907 | hctx = blk_mq_map_queue_type(q, j, i); | |
2908 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2909 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2910 | } |
320ae51f JA |
2911 | } |
2912 | } | |
2913 | ||
63064be1 JG |
2914 | struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, |
2915 | unsigned int hctx_idx, | |
2916 | unsigned int depth) | |
cc71a6f4 | 2917 | { |
63064be1 JG |
2918 | struct blk_mq_tags *tags; |
2919 | int ret; | |
cc71a6f4 | 2920 | |
e155b0c2 | 2921 | tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); |
63064be1 JG |
2922 | if (!tags) |
2923 | return NULL; | |
cc71a6f4 | 2924 | |
63064be1 JG |
2925 | ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); |
2926 | if (ret) { | |
e155b0c2 | 2927 | blk_mq_free_rq_map(tags); |
63064be1 JG |
2928 | return NULL; |
2929 | } | |
cc71a6f4 | 2930 | |
63064be1 JG |
2931 | return tags; |
2932 | } | |
2933 | ||
2934 | static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, | |
2935 | int hctx_idx) | |
2936 | { | |
079a2e3e JG |
2937 | if (blk_mq_is_shared_tags(set->flags)) { |
2938 | set->tags[hctx_idx] = set->shared_tags; | |
e155b0c2 JG |
2939 | |
2940 | return true; | |
2941 | } | |
2942 | ||
63064be1 JG |
2943 | set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, |
2944 | set->queue_depth); | |
2945 | ||
2946 | return set->tags[hctx_idx]; | |
cc71a6f4 JA |
2947 | } |
2948 | ||
645db34e JG |
2949 | void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
2950 | struct blk_mq_tags *tags, | |
2951 | unsigned int hctx_idx) | |
cc71a6f4 | 2952 | { |
645db34e JG |
2953 | if (tags) { |
2954 | blk_mq_free_rqs(set, tags, hctx_idx); | |
e155b0c2 | 2955 | blk_mq_free_rq_map(tags); |
bd166ef1 | 2956 | } |
cc71a6f4 JA |
2957 | } |
2958 | ||
e155b0c2 JG |
2959 | static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, |
2960 | unsigned int hctx_idx) | |
2961 | { | |
079a2e3e | 2962 | if (!blk_mq_is_shared_tags(set->flags)) |
e155b0c2 JG |
2963 | blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); |
2964 | ||
2965 | set->tags[hctx_idx] = NULL; | |
2966 | } | |
2967 | ||
4b855ad3 | 2968 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2969 | { |
b3c661b1 | 2970 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2971 | struct blk_mq_hw_ctx *hctx; |
2972 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2973 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2974 | |
2975 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2976 | cpumask_clear(hctx->cpumask); |
320ae51f | 2977 | hctx->nr_ctx = 0; |
d416c92c | 2978 | hctx->dispatch_from = NULL; |
320ae51f JA |
2979 | } |
2980 | ||
2981 | /* | |
4b855ad3 | 2982 | * Map software to hardware queues. |
4412efec ML |
2983 | * |
2984 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2985 | */ |
20e4d813 | 2986 | for_each_possible_cpu(i) { |
4412efec | 2987 | |
897bb0c7 | 2988 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2989 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2990 | if (!set->map[j].nr_queues) { |
2991 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2992 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2993 | continue; |
bb94aea1 | 2994 | } |
fd689871 ML |
2995 | hctx_idx = set->map[j].mq_map[i]; |
2996 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2997 | if (!set->tags[hctx_idx] && | |
63064be1 | 2998 | !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { |
fd689871 ML |
2999 | /* |
3000 | * If tags initialization fail for some hctx, | |
3001 | * that hctx won't be brought online. In this | |
3002 | * case, remap the current ctx to hctx[0] which | |
3003 | * is guaranteed to always have tags allocated | |
3004 | */ | |
3005 | set->map[j].mq_map[i] = 0; | |
3006 | } | |
e5edd5f2 | 3007 | |
b3c661b1 | 3008 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 3009 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
3010 | /* |
3011 | * If the CPU is already set in the mask, then we've | |
3012 | * mapped this one already. This can happen if | |
3013 | * devices share queues across queue maps. | |
3014 | */ | |
3015 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
3016 | continue; | |
3017 | ||
3018 | cpumask_set_cpu(i, hctx->cpumask); | |
3019 | hctx->type = j; | |
3020 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
3021 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
3022 | ||
3023 | /* | |
3024 | * If the nr_ctx type overflows, we have exceeded the | |
3025 | * amount of sw queues we can support. | |
3026 | */ | |
3027 | BUG_ON(!hctx->nr_ctx); | |
3028 | } | |
bb94aea1 JW |
3029 | |
3030 | for (; j < HCTX_MAX_TYPES; j++) | |
3031 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
3032 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 3033 | } |
506e931f JA |
3034 | |
3035 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
3036 | /* |
3037 | * If no software queues are mapped to this hardware queue, | |
3038 | * disable it and free the request entries. | |
3039 | */ | |
3040 | if (!hctx->nr_ctx) { | |
3041 | /* Never unmap queue 0. We need it as a | |
3042 | * fallback in case of a new remap fails | |
3043 | * allocation | |
3044 | */ | |
e155b0c2 JG |
3045 | if (i) |
3046 | __blk_mq_free_map_and_rqs(set, i); | |
4412efec ML |
3047 | |
3048 | hctx->tags = NULL; | |
3049 | continue; | |
3050 | } | |
484b4061 | 3051 | |
2a34c087 ML |
3052 | hctx->tags = set->tags[i]; |
3053 | WARN_ON(!hctx->tags); | |
3054 | ||
889fa31f CY |
3055 | /* |
3056 | * Set the map size to the number of mapped software queues. | |
3057 | * This is more accurate and more efficient than looping | |
3058 | * over all possibly mapped software queues. | |
3059 | */ | |
88459642 | 3060 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 3061 | |
484b4061 JA |
3062 | /* |
3063 | * Initialize batch roundrobin counts | |
3064 | */ | |
f82ddf19 | 3065 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
3066 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
3067 | } | |
320ae51f JA |
3068 | } |
3069 | ||
8e8320c9 JA |
3070 | /* |
3071 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
3072 | * the queue isn't live yet. | |
3073 | */ | |
2404e607 | 3074 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
3075 | { |
3076 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
3077 | int i; |
3078 | ||
2404e607 | 3079 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 3080 | if (shared) { |
51db1c37 | 3081 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3082 | } else { |
3083 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3084 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3085 | } |
2404e607 JM |
3086 | } |
3087 | } | |
3088 | ||
655ac300 HR |
3089 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3090 | bool shared) | |
2404e607 JM |
3091 | { |
3092 | struct request_queue *q; | |
0d2602ca | 3093 | |
705cda97 BVA |
3094 | lockdep_assert_held(&set->tag_list_lock); |
3095 | ||
0d2602ca JA |
3096 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3097 | blk_mq_freeze_queue(q); | |
2404e607 | 3098 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3099 | blk_mq_unfreeze_queue(q); |
3100 | } | |
3101 | } | |
3102 | ||
3103 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3104 | { | |
3105 | struct blk_mq_tag_set *set = q->tag_set; | |
3106 | ||
0d2602ca | 3107 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3108 | list_del(&q->tag_set_list); |
2404e607 JM |
3109 | if (list_is_singular(&set->tag_list)) { |
3110 | /* just transitioned to unshared */ | |
51db1c37 | 3111 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3112 | /* update existing queue */ |
655ac300 | 3113 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3114 | } |
0d2602ca | 3115 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3116 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3117 | } |
3118 | ||
3119 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3120 | struct request_queue *q) | |
3121 | { | |
0d2602ca | 3122 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3123 | |
ff821d27 JA |
3124 | /* |
3125 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3126 | */ | |
3127 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3128 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3129 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3130 | /* update existing queue */ |
655ac300 | 3131 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3132 | } |
51db1c37 | 3133 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3134 | queue_set_hctx_shared(q, true); |
08c875cb | 3135 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3136 | |
0d2602ca JA |
3137 | mutex_unlock(&set->tag_list_lock); |
3138 | } | |
3139 | ||
1db4909e ML |
3140 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3141 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3142 | { | |
3143 | struct blk_mq_ctxs *ctxs; | |
3144 | int cpu; | |
3145 | ||
3146 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3147 | if (!ctxs) | |
3148 | return -ENOMEM; | |
3149 | ||
3150 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3151 | if (!ctxs->queue_ctx) | |
3152 | goto fail; | |
3153 | ||
3154 | for_each_possible_cpu(cpu) { | |
3155 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3156 | ctx->ctxs = ctxs; | |
3157 | } | |
3158 | ||
3159 | q->mq_kobj = &ctxs->kobj; | |
3160 | q->queue_ctx = ctxs->queue_ctx; | |
3161 | ||
3162 | return 0; | |
3163 | fail: | |
3164 | kfree(ctxs); | |
3165 | return -ENOMEM; | |
3166 | } | |
3167 | ||
e09aae7e ML |
3168 | /* |
3169 | * It is the actual release handler for mq, but we do it from | |
3170 | * request queue's release handler for avoiding use-after-free | |
3171 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3172 | * but we can't group ctx/kctx kobj without it. | |
3173 | */ | |
3174 | void blk_mq_release(struct request_queue *q) | |
3175 | { | |
2f8f1336 ML |
3176 | struct blk_mq_hw_ctx *hctx, *next; |
3177 | int i; | |
e09aae7e | 3178 | |
2f8f1336 ML |
3179 | queue_for_each_hw_ctx(q, hctx, i) |
3180 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3181 | ||
3182 | /* all hctx are in .unused_hctx_list now */ | |
3183 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3184 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3185 | kobject_put(&hctx->kobj); |
c3b4afca | 3186 | } |
e09aae7e ML |
3187 | |
3188 | kfree(q->queue_hw_ctx); | |
3189 | ||
7ea5fe31 ML |
3190 | /* |
3191 | * release .mq_kobj and sw queue's kobject now because | |
3192 | * both share lifetime with request queue. | |
3193 | */ | |
3194 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3195 | } |
3196 | ||
5ec780a6 | 3197 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3198 | void *queuedata) |
b62c21b7 | 3199 | { |
26a9750a CH |
3200 | struct request_queue *q; |
3201 | int ret; | |
b62c21b7 | 3202 | |
26a9750a CH |
3203 | q = blk_alloc_queue(set->numa_node); |
3204 | if (!q) | |
b62c21b7 | 3205 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3206 | q->queuedata = queuedata; |
3207 | ret = blk_mq_init_allocated_queue(set, q); | |
3208 | if (ret) { | |
3209 | blk_cleanup_queue(q); | |
3210 | return ERR_PTR(ret); | |
3211 | } | |
b62c21b7 MS |
3212 | return q; |
3213 | } | |
2f227bb9 CH |
3214 | |
3215 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3216 | { | |
3217 | return blk_mq_init_queue_data(set, NULL); | |
3218 | } | |
b62c21b7 MS |
3219 | EXPORT_SYMBOL(blk_mq_init_queue); |
3220 | ||
4dcc4874 CH |
3221 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3222 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3223 | { |
3224 | struct request_queue *q; | |
b461dfc4 | 3225 | struct gendisk *disk; |
9316a9ed | 3226 | |
b461dfc4 CH |
3227 | q = blk_mq_init_queue_data(set, queuedata); |
3228 | if (IS_ERR(q)) | |
3229 | return ERR_CAST(q); | |
9316a9ed | 3230 | |
4a1fa41d | 3231 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3232 | if (!disk) { |
3233 | blk_cleanup_queue(q); | |
3234 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3235 | } |
b461dfc4 | 3236 | return disk; |
9316a9ed | 3237 | } |
b461dfc4 | 3238 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3239 | |
34d11ffa JW |
3240 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3241 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3242 | int hctx_idx, int node) | |
3243 | { | |
2f8f1336 | 3244 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3245 | |
2f8f1336 ML |
3246 | /* reuse dead hctx first */ |
3247 | spin_lock(&q->unused_hctx_lock); | |
3248 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3249 | if (tmp->numa_node == node) { | |
3250 | hctx = tmp; | |
3251 | break; | |
3252 | } | |
3253 | } | |
3254 | if (hctx) | |
3255 | list_del_init(&hctx->hctx_list); | |
3256 | spin_unlock(&q->unused_hctx_lock); | |
3257 | ||
3258 | if (!hctx) | |
3259 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3260 | if (!hctx) |
7c6c5b7c | 3261 | goto fail; |
34d11ffa | 3262 | |
7c6c5b7c ML |
3263 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3264 | goto free_hctx; | |
34d11ffa JW |
3265 | |
3266 | return hctx; | |
7c6c5b7c ML |
3267 | |
3268 | free_hctx: | |
3269 | kobject_put(&hctx->kobj); | |
3270 | fail: | |
3271 | return NULL; | |
34d11ffa JW |
3272 | } |
3273 | ||
868f2f0b KB |
3274 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3275 | struct request_queue *q) | |
320ae51f | 3276 | { |
e01ad46d | 3277 | int i, j, end; |
868f2f0b | 3278 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3279 | |
ac0d6b92 BVA |
3280 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3281 | struct blk_mq_hw_ctx **new_hctxs; | |
3282 | ||
3283 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3284 | sizeof(*new_hctxs), GFP_KERNEL, | |
3285 | set->numa_node); | |
3286 | if (!new_hctxs) | |
3287 | return; | |
3288 | if (hctxs) | |
3289 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3290 | sizeof(*hctxs)); | |
3291 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3292 | kfree(hctxs); |
3293 | hctxs = new_hctxs; | |
3294 | } | |
3295 | ||
fb350e0a ML |
3296 | /* protect against switching io scheduler */ |
3297 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3298 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3299 | int node; |
34d11ffa | 3300 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3301 | |
7d76f856 | 3302 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3303 | /* |
3304 | * If the hw queue has been mapped to another numa node, | |
3305 | * we need to realloc the hctx. If allocation fails, fallback | |
3306 | * to use the previous one. | |
3307 | */ | |
3308 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3309 | continue; | |
868f2f0b | 3310 | |
34d11ffa JW |
3311 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3312 | if (hctx) { | |
2f8f1336 | 3313 | if (hctxs[i]) |
34d11ffa | 3314 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3315 | hctxs[i] = hctx; |
3316 | } else { | |
3317 | if (hctxs[i]) | |
3318 | pr_warn("Allocate new hctx on node %d fails,\ | |
3319 | fallback to previous one on node %d\n", | |
3320 | node, hctxs[i]->numa_node); | |
3321 | else | |
3322 | break; | |
868f2f0b | 3323 | } |
320ae51f | 3324 | } |
e01ad46d JW |
3325 | /* |
3326 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3327 | * hctxs and keep the previous q->nr_hw_queues. | |
3328 | */ | |
3329 | if (i != set->nr_hw_queues) { | |
3330 | j = q->nr_hw_queues; | |
3331 | end = i; | |
3332 | } else { | |
3333 | j = i; | |
3334 | end = q->nr_hw_queues; | |
3335 | q->nr_hw_queues = set->nr_hw_queues; | |
3336 | } | |
34d11ffa | 3337 | |
e01ad46d | 3338 | for (; j < end; j++) { |
868f2f0b KB |
3339 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3340 | ||
3341 | if (hctx) { | |
e155b0c2 | 3342 | __blk_mq_free_map_and_rqs(set, j); |
868f2f0b | 3343 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3344 | hctxs[j] = NULL; |
868f2f0b KB |
3345 | } |
3346 | } | |
fb350e0a | 3347 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3348 | } |
3349 | ||
26a9750a CH |
3350 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
3351 | struct request_queue *q) | |
868f2f0b | 3352 | { |
66841672 ML |
3353 | /* mark the queue as mq asap */ |
3354 | q->mq_ops = set->ops; | |
3355 | ||
34dbad5d | 3356 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3357 | blk_mq_poll_stats_bkt, |
3358 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3359 | if (!q->poll_cb) |
3360 | goto err_exit; | |
3361 | ||
1db4909e | 3362 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3363 | goto err_poll; |
868f2f0b | 3364 | |
737f98cf ML |
3365 | /* init q->mq_kobj and sw queues' kobjects */ |
3366 | blk_mq_sysfs_init(q); | |
3367 | ||
2f8f1336 ML |
3368 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3369 | spin_lock_init(&q->unused_hctx_lock); | |
3370 | ||
868f2f0b KB |
3371 | blk_mq_realloc_hw_ctxs(set, q); |
3372 | if (!q->nr_hw_queues) | |
3373 | goto err_hctxs; | |
320ae51f | 3374 | |
287922eb | 3375 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3376 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3377 | |
a8908939 | 3378 | q->tag_set = set; |
320ae51f | 3379 | |
94eddfbe | 3380 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3381 | if (set->nr_maps > HCTX_TYPE_POLL && |
3382 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3383 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3384 | |
2849450a | 3385 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3386 | INIT_LIST_HEAD(&q->requeue_list); |
3387 | spin_lock_init(&q->requeue_lock); | |
3388 | ||
eba71768 JA |
3389 | q->nr_requests = set->queue_depth; |
3390 | ||
64f1c21e JA |
3391 | /* |
3392 | * Default to classic polling | |
3393 | */ | |
29ece8b4 | 3394 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3395 | |
24d2f903 | 3396 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3397 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3398 | blk_mq_map_swqueue(q); |
26a9750a | 3399 | return 0; |
18741986 | 3400 | |
320ae51f | 3401 | err_hctxs: |
868f2f0b | 3402 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3403 | q->nr_hw_queues = 0; |
1db4909e | 3404 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3405 | err_poll: |
3406 | blk_stat_free_callback(q->poll_cb); | |
3407 | q->poll_cb = NULL; | |
c7de5726 ML |
3408 | err_exit: |
3409 | q->mq_ops = NULL; | |
26a9750a | 3410 | return -ENOMEM; |
320ae51f | 3411 | } |
b62c21b7 | 3412 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3413 | |
c7e2d94b ML |
3414 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3415 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3416 | { |
630ef623 | 3417 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3418 | |
630ef623 | 3419 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 3420 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
3421 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
3422 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 3423 | } |
320ae51f | 3424 | |
a5164405 JA |
3425 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3426 | { | |
3427 | int i; | |
3428 | ||
079a2e3e JG |
3429 | if (blk_mq_is_shared_tags(set->flags)) { |
3430 | set->shared_tags = blk_mq_alloc_map_and_rqs(set, | |
e155b0c2 JG |
3431 | BLK_MQ_NO_HCTX_IDX, |
3432 | set->queue_depth); | |
079a2e3e | 3433 | if (!set->shared_tags) |
e155b0c2 JG |
3434 | return -ENOMEM; |
3435 | } | |
3436 | ||
8229cca8 | 3437 | for (i = 0; i < set->nr_hw_queues; i++) { |
63064be1 | 3438 | if (!__blk_mq_alloc_map_and_rqs(set, i)) |
a5164405 | 3439 | goto out_unwind; |
8229cca8 XT |
3440 | cond_resched(); |
3441 | } | |
a5164405 JA |
3442 | |
3443 | return 0; | |
3444 | ||
3445 | out_unwind: | |
e155b0c2 JG |
3446 | while (--i >= 0) |
3447 | __blk_mq_free_map_and_rqs(set, i); | |
3448 | ||
079a2e3e JG |
3449 | if (blk_mq_is_shared_tags(set->flags)) { |
3450 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 | 3451 | BLK_MQ_NO_HCTX_IDX); |
645db34e | 3452 | } |
a5164405 | 3453 | |
a5164405 JA |
3454 | return -ENOMEM; |
3455 | } | |
3456 | ||
3457 | /* | |
3458 | * Allocate the request maps associated with this tag_set. Note that this | |
3459 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3460 | * will be updated to reflect the allocated depth. | |
3461 | */ | |
63064be1 | 3462 | static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) |
a5164405 JA |
3463 | { |
3464 | unsigned int depth; | |
3465 | int err; | |
3466 | ||
3467 | depth = set->queue_depth; | |
3468 | do { | |
3469 | err = __blk_mq_alloc_rq_maps(set); | |
3470 | if (!err) | |
3471 | break; | |
3472 | ||
3473 | set->queue_depth >>= 1; | |
3474 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3475 | err = -ENOMEM; | |
3476 | break; | |
3477 | } | |
3478 | } while (set->queue_depth); | |
3479 | ||
3480 | if (!set->queue_depth || err) { | |
3481 | pr_err("blk-mq: failed to allocate request map\n"); | |
3482 | return -ENOMEM; | |
3483 | } | |
3484 | ||
3485 | if (depth != set->queue_depth) | |
3486 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3487 | depth, set->queue_depth); | |
3488 | ||
3489 | return 0; | |
3490 | } | |
3491 | ||
ebe8bddb OS |
3492 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3493 | { | |
6e66b493 BVA |
3494 | /* |
3495 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3496 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3497 | * number of hardware queues. | |
3498 | */ | |
3499 | if (set->nr_maps == 1) | |
3500 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3501 | ||
59388702 | 3502 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3503 | int i; |
3504 | ||
7d4901a9 ML |
3505 | /* |
3506 | * transport .map_queues is usually done in the following | |
3507 | * way: | |
3508 | * | |
3509 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3510 | * mask = get_cpu_mask(queue) | |
3511 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3512 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3513 | * } |
3514 | * | |
3515 | * When we need to remap, the table has to be cleared for | |
3516 | * killing stale mapping since one CPU may not be mapped | |
3517 | * to any hw queue. | |
3518 | */ | |
b3c661b1 JA |
3519 | for (i = 0; i < set->nr_maps; i++) |
3520 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3521 | |
ebe8bddb | 3522 | return set->ops->map_queues(set); |
b3c661b1 JA |
3523 | } else { |
3524 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3525 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3526 | } |
ebe8bddb OS |
3527 | } |
3528 | ||
f7e76dbc BVA |
3529 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3530 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3531 | { | |
3532 | struct blk_mq_tags **new_tags; | |
3533 | ||
3534 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3535 | return 0; | |
3536 | ||
3537 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3538 | GFP_KERNEL, set->numa_node); | |
3539 | if (!new_tags) | |
3540 | return -ENOMEM; | |
3541 | ||
3542 | if (set->tags) | |
3543 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3544 | sizeof(*set->tags)); | |
3545 | kfree(set->tags); | |
3546 | set->tags = new_tags; | |
3547 | set->nr_hw_queues = new_nr_hw_queues; | |
3548 | ||
3549 | return 0; | |
3550 | } | |
3551 | ||
91cdf265 MI |
3552 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3553 | int new_nr_hw_queues) | |
3554 | { | |
3555 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3556 | } | |
3557 | ||
a4391c64 JA |
3558 | /* |
3559 | * Alloc a tag set to be associated with one or more request queues. | |
3560 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3561 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3562 | * value will be stored in set->queue_depth. |
3563 | */ | |
24d2f903 CH |
3564 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3565 | { | |
b3c661b1 | 3566 | int i, ret; |
da695ba2 | 3567 | |
205fb5f5 BVA |
3568 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3569 | ||
24d2f903 CH |
3570 | if (!set->nr_hw_queues) |
3571 | return -EINVAL; | |
a4391c64 | 3572 | if (!set->queue_depth) |
24d2f903 CH |
3573 | return -EINVAL; |
3574 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3575 | return -EINVAL; | |
3576 | ||
7d7e0f90 | 3577 | if (!set->ops->queue_rq) |
24d2f903 CH |
3578 | return -EINVAL; |
3579 | ||
de148297 ML |
3580 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3581 | return -EINVAL; | |
3582 | ||
a4391c64 JA |
3583 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3584 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3585 | BLK_MQ_MAX_DEPTH); | |
3586 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3587 | } | |
24d2f903 | 3588 | |
b3c661b1 JA |
3589 | if (!set->nr_maps) |
3590 | set->nr_maps = 1; | |
3591 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3592 | return -EINVAL; | |
3593 | ||
6637fadf SL |
3594 | /* |
3595 | * If a crashdump is active, then we are potentially in a very | |
3596 | * memory constrained environment. Limit us to 1 queue and | |
3597 | * 64 tags to prevent using too much memory. | |
3598 | */ | |
3599 | if (is_kdump_kernel()) { | |
3600 | set->nr_hw_queues = 1; | |
59388702 | 3601 | set->nr_maps = 1; |
6637fadf SL |
3602 | set->queue_depth = min(64U, set->queue_depth); |
3603 | } | |
868f2f0b | 3604 | /* |
392546ae JA |
3605 | * There is no use for more h/w queues than cpus if we just have |
3606 | * a single map | |
868f2f0b | 3607 | */ |
392546ae | 3608 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3609 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3610 | |
91cdf265 | 3611 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3612 | return -ENOMEM; |
24d2f903 | 3613 | |
da695ba2 | 3614 | ret = -ENOMEM; |
b3c661b1 JA |
3615 | for (i = 0; i < set->nr_maps; i++) { |
3616 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3617 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3618 | GFP_KERNEL, set->numa_node); |
3619 | if (!set->map[i].mq_map) | |
3620 | goto out_free_mq_map; | |
59388702 | 3621 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3622 | } |
bdd17e75 | 3623 | |
ebe8bddb | 3624 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3625 | if (ret) |
3626 | goto out_free_mq_map; | |
3627 | ||
63064be1 | 3628 | ret = blk_mq_alloc_set_map_and_rqs(set); |
da695ba2 | 3629 | if (ret) |
bdd17e75 | 3630 | goto out_free_mq_map; |
24d2f903 | 3631 | |
0d2602ca JA |
3632 | mutex_init(&set->tag_list_lock); |
3633 | INIT_LIST_HEAD(&set->tag_list); | |
3634 | ||
24d2f903 | 3635 | return 0; |
bdd17e75 CH |
3636 | |
3637 | out_free_mq_map: | |
b3c661b1 JA |
3638 | for (i = 0; i < set->nr_maps; i++) { |
3639 | kfree(set->map[i].mq_map); | |
3640 | set->map[i].mq_map = NULL; | |
3641 | } | |
5676e7b6 RE |
3642 | kfree(set->tags); |
3643 | set->tags = NULL; | |
da695ba2 | 3644 | return ret; |
24d2f903 CH |
3645 | } |
3646 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3647 | ||
cdb14e0f CH |
3648 | /* allocate and initialize a tagset for a simple single-queue device */ |
3649 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
3650 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
3651 | unsigned int set_flags) | |
3652 | { | |
3653 | memset(set, 0, sizeof(*set)); | |
3654 | set->ops = ops; | |
3655 | set->nr_hw_queues = 1; | |
3656 | set->nr_maps = 1; | |
3657 | set->queue_depth = queue_depth; | |
3658 | set->numa_node = NUMA_NO_NODE; | |
3659 | set->flags = set_flags; | |
3660 | return blk_mq_alloc_tag_set(set); | |
3661 | } | |
3662 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
3663 | ||
24d2f903 CH |
3664 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
3665 | { | |
b3c661b1 | 3666 | int i, j; |
24d2f903 | 3667 | |
e155b0c2 JG |
3668 | for (i = 0; i < set->nr_hw_queues; i++) |
3669 | __blk_mq_free_map_and_rqs(set, i); | |
484b4061 | 3670 | |
079a2e3e JG |
3671 | if (blk_mq_is_shared_tags(set->flags)) { |
3672 | blk_mq_free_map_and_rqs(set, set->shared_tags, | |
e155b0c2 JG |
3673 | BLK_MQ_NO_HCTX_IDX); |
3674 | } | |
32bc15af | 3675 | |
b3c661b1 JA |
3676 | for (j = 0; j < set->nr_maps; j++) { |
3677 | kfree(set->map[j].mq_map); | |
3678 | set->map[j].mq_map = NULL; | |
3679 | } | |
bdd17e75 | 3680 | |
981bd189 | 3681 | kfree(set->tags); |
5676e7b6 | 3682 | set->tags = NULL; |
24d2f903 CH |
3683 | } |
3684 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3685 | ||
e3a2b3f9 JA |
3686 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3687 | { | |
3688 | struct blk_mq_tag_set *set = q->tag_set; | |
3689 | struct blk_mq_hw_ctx *hctx; | |
3690 | int i, ret; | |
3691 | ||
bd166ef1 | 3692 | if (!set) |
e3a2b3f9 JA |
3693 | return -EINVAL; |
3694 | ||
e5fa8140 AZ |
3695 | if (q->nr_requests == nr) |
3696 | return 0; | |
3697 | ||
70f36b60 | 3698 | blk_mq_freeze_queue(q); |
24f5a90f | 3699 | blk_mq_quiesce_queue(q); |
70f36b60 | 3700 | |
e3a2b3f9 JA |
3701 | ret = 0; |
3702 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3703 | if (!hctx->tags) |
3704 | continue; | |
bd166ef1 JA |
3705 | /* |
3706 | * If we're using an MQ scheduler, just update the scheduler | |
3707 | * queue depth. This is similar to what the old code would do. | |
3708 | */ | |
f6adcef5 | 3709 | if (hctx->sched_tags) { |
70f36b60 | 3710 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, |
f6adcef5 | 3711 | nr, true); |
f6adcef5 JG |
3712 | } else { |
3713 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
3714 | false); | |
70f36b60 | 3715 | } |
e3a2b3f9 JA |
3716 | if (ret) |
3717 | break; | |
77f1e0a5 JA |
3718 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3719 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 3720 | } |
d97e594c | 3721 | if (!ret) { |
e3a2b3f9 | 3722 | q->nr_requests = nr; |
079a2e3e | 3723 | if (blk_mq_is_shared_tags(set->flags)) { |
8fa04464 | 3724 | if (q->elevator) |
079a2e3e | 3725 | blk_mq_tag_update_sched_shared_tags(q); |
8fa04464 | 3726 | else |
079a2e3e | 3727 | blk_mq_tag_resize_shared_tags(set, nr); |
8fa04464 | 3728 | } |
d97e594c | 3729 | } |
e3a2b3f9 | 3730 | |
24f5a90f | 3731 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3732 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3733 | |
e3a2b3f9 JA |
3734 | return ret; |
3735 | } | |
3736 | ||
d48ece20 JW |
3737 | /* |
3738 | * request_queue and elevator_type pair. | |
3739 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3740 | * the elevator_type associated with a request_queue. | |
3741 | */ | |
3742 | struct blk_mq_qe_pair { | |
3743 | struct list_head node; | |
3744 | struct request_queue *q; | |
3745 | struct elevator_type *type; | |
3746 | }; | |
3747 | ||
3748 | /* | |
3749 | * Cache the elevator_type in qe pair list and switch the | |
3750 | * io scheduler to 'none' | |
3751 | */ | |
3752 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3753 | struct request_queue *q) | |
3754 | { | |
3755 | struct blk_mq_qe_pair *qe; | |
3756 | ||
3757 | if (!q->elevator) | |
3758 | return true; | |
3759 | ||
3760 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3761 | if (!qe) | |
3762 | return false; | |
3763 | ||
3764 | INIT_LIST_HEAD(&qe->node); | |
3765 | qe->q = q; | |
3766 | qe->type = q->elevator->type; | |
3767 | list_add(&qe->node, head); | |
3768 | ||
3769 | mutex_lock(&q->sysfs_lock); | |
3770 | /* | |
3771 | * After elevator_switch_mq, the previous elevator_queue will be | |
3772 | * released by elevator_release. The reference of the io scheduler | |
3773 | * module get by elevator_get will also be put. So we need to get | |
3774 | * a reference of the io scheduler module here to prevent it to be | |
3775 | * removed. | |
3776 | */ | |
3777 | __module_get(qe->type->elevator_owner); | |
3778 | elevator_switch_mq(q, NULL); | |
3779 | mutex_unlock(&q->sysfs_lock); | |
3780 | ||
3781 | return true; | |
3782 | } | |
3783 | ||
3784 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3785 | struct request_queue *q) | |
3786 | { | |
3787 | struct blk_mq_qe_pair *qe; | |
3788 | struct elevator_type *t = NULL; | |
3789 | ||
3790 | list_for_each_entry(qe, head, node) | |
3791 | if (qe->q == q) { | |
3792 | t = qe->type; | |
3793 | break; | |
3794 | } | |
3795 | ||
3796 | if (!t) | |
3797 | return; | |
3798 | ||
3799 | list_del(&qe->node); | |
3800 | kfree(qe); | |
3801 | ||
3802 | mutex_lock(&q->sysfs_lock); | |
3803 | elevator_switch_mq(q, t); | |
3804 | mutex_unlock(&q->sysfs_lock); | |
3805 | } | |
3806 | ||
e4dc2b32 KB |
3807 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3808 | int nr_hw_queues) | |
868f2f0b KB |
3809 | { |
3810 | struct request_queue *q; | |
d48ece20 | 3811 | LIST_HEAD(head); |
e01ad46d | 3812 | int prev_nr_hw_queues; |
868f2f0b | 3813 | |
705cda97 BVA |
3814 | lockdep_assert_held(&set->tag_list_lock); |
3815 | ||
392546ae | 3816 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3817 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3818 | if (nr_hw_queues < 1) |
3819 | return; | |
3820 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3821 | return; |
3822 | ||
3823 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3824 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3825 | /* |
3826 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3827 | * with the previous scheduler. We will switch back once we are done | |
3828 | * updating the new sw to hw queue mappings. | |
3829 | */ | |
3830 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3831 | if (!blk_mq_elv_switch_none(&head, q)) | |
3832 | goto switch_back; | |
868f2f0b | 3833 | |
477e19de JW |
3834 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3835 | blk_mq_debugfs_unregister_hctxs(q); | |
3836 | blk_mq_sysfs_unregister(q); | |
3837 | } | |
3838 | ||
a2584e43 | 3839 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3840 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3841 | 0) | |
3842 | goto reregister; | |
3843 | ||
868f2f0b | 3844 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3845 | fallback: |
aa880ad6 | 3846 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3847 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3848 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3849 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3850 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3851 | nr_hw_queues, prev_nr_hw_queues); | |
3852 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3853 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3854 | goto fallback; |
3855 | } | |
477e19de JW |
3856 | blk_mq_map_swqueue(q); |
3857 | } | |
3858 | ||
f7e76dbc | 3859 | reregister: |
477e19de JW |
3860 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3861 | blk_mq_sysfs_register(q); | |
3862 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3863 | } |
3864 | ||
d48ece20 JW |
3865 | switch_back: |
3866 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3867 | blk_mq_elv_switch_back(&head, q); | |
3868 | ||
868f2f0b KB |
3869 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3870 | blk_mq_unfreeze_queue(q); | |
3871 | } | |
e4dc2b32 KB |
3872 | |
3873 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3874 | { | |
3875 | mutex_lock(&set->tag_list_lock); | |
3876 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3877 | mutex_unlock(&set->tag_list_lock); | |
3878 | } | |
868f2f0b KB |
3879 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3880 | ||
34dbad5d OS |
3881 | /* Enable polling stats and return whether they were already enabled. */ |
3882 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3883 | { | |
3884 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3885 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3886 | return true; |
3887 | blk_stat_add_callback(q, q->poll_cb); | |
3888 | return false; | |
3889 | } | |
3890 | ||
3891 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3892 | { | |
3893 | /* | |
3894 | * We don't arm the callback if polling stats are not enabled or the | |
3895 | * callback is already active. | |
3896 | */ | |
3897 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3898 | blk_stat_is_active(q->poll_cb)) | |
3899 | return; | |
3900 | ||
3901 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3902 | } | |
3903 | ||
3904 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3905 | { | |
3906 | struct request_queue *q = cb->data; | |
720b8ccc | 3907 | int bucket; |
34dbad5d | 3908 | |
720b8ccc SB |
3909 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3910 | if (cb->stat[bucket].nr_samples) | |
3911 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3912 | } | |
34dbad5d OS |
3913 | } |
3914 | ||
64f1c21e | 3915 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3916 | struct request *rq) |
3917 | { | |
64f1c21e | 3918 | unsigned long ret = 0; |
720b8ccc | 3919 | int bucket; |
64f1c21e JA |
3920 | |
3921 | /* | |
3922 | * If stats collection isn't on, don't sleep but turn it on for | |
3923 | * future users | |
3924 | */ | |
34dbad5d | 3925 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3926 | return 0; |
3927 | ||
64f1c21e JA |
3928 | /* |
3929 | * As an optimistic guess, use half of the mean service time | |
3930 | * for this type of request. We can (and should) make this smarter. | |
3931 | * For instance, if the completion latencies are tight, we can | |
3932 | * get closer than just half the mean. This is especially | |
3933 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3934 | * than ~10 usec. We do use the stats for the relevant IO size |
3935 | * if available which does lead to better estimates. | |
64f1c21e | 3936 | */ |
720b8ccc SB |
3937 | bucket = blk_mq_poll_stats_bkt(rq); |
3938 | if (bucket < 0) | |
3939 | return ret; | |
3940 | ||
3941 | if (q->poll_stat[bucket].nr_samples) | |
3942 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3943 | |
3944 | return ret; | |
3945 | } | |
3946 | ||
06426adf JA |
3947 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3948 | struct request *rq) | |
3949 | { | |
3950 | struct hrtimer_sleeper hs; | |
3951 | enum hrtimer_mode mode; | |
64f1c21e | 3952 | unsigned int nsecs; |
06426adf JA |
3953 | ktime_t kt; |
3954 | ||
76a86f9d | 3955 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3956 | return false; |
3957 | ||
3958 | /* | |
1052b8ac | 3959 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3960 | * |
64f1c21e JA |
3961 | * 0: use half of prev avg |
3962 | * >0: use this specific value | |
3963 | */ | |
1052b8ac | 3964 | if (q->poll_nsec > 0) |
64f1c21e JA |
3965 | nsecs = q->poll_nsec; |
3966 | else | |
cae740a0 | 3967 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3968 | |
3969 | if (!nsecs) | |
06426adf JA |
3970 | return false; |
3971 | ||
76a86f9d | 3972 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3973 | |
3974 | /* | |
3975 | * This will be replaced with the stats tracking code, using | |
3976 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3977 | */ | |
8b0e1953 | 3978 | kt = nsecs; |
06426adf JA |
3979 | |
3980 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3981 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3982 | hrtimer_set_expires(&hs.timer, kt); |
3983 | ||
06426adf | 3984 | do { |
5a61c363 | 3985 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3986 | break; |
3987 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3988 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3989 | if (hs.task) |
3990 | io_schedule(); | |
3991 | hrtimer_cancel(&hs.timer); | |
3992 | mode = HRTIMER_MODE_ABS; | |
3993 | } while (hs.task && !signal_pending(current)); | |
3994 | ||
3995 | __set_current_state(TASK_RUNNING); | |
3996 | destroy_hrtimer_on_stack(&hs.timer); | |
3997 | return true; | |
3998 | } | |
3999 | ||
1052b8ac JA |
4000 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
4001 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 4002 | { |
1052b8ac JA |
4003 | struct request *rq; |
4004 | ||
29ece8b4 | 4005 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
4006 | return false; |
4007 | ||
4008 | if (!blk_qc_t_is_internal(cookie)) | |
4009 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
4010 | else { | |
4011 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
4012 | /* | |
4013 | * With scheduling, if the request has completed, we'll | |
4014 | * get a NULL return here, as we clear the sched tag when | |
4015 | * that happens. The request still remains valid, like always, | |
4016 | * so we should be safe with just the NULL check. | |
4017 | */ | |
4018 | if (!rq) | |
4019 | return false; | |
4020 | } | |
4021 | ||
cae740a0 | 4022 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
4023 | } |
4024 | ||
529262d5 CH |
4025 | /** |
4026 | * blk_poll - poll for IO completions | |
4027 | * @q: the queue | |
4028 | * @cookie: cookie passed back at IO submission time | |
4029 | * @spin: whether to spin for completions | |
4030 | * | |
4031 | * Description: | |
4032 | * Poll for completions on the passed in queue. Returns number of | |
4033 | * completed entries found. If @spin is true, then blk_poll will continue | |
4034 | * looping until at least one completion is found, unless the task is | |
4035 | * otherwise marked running (or we need to reschedule). | |
4036 | */ | |
4037 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
4038 | { |
4039 | struct blk_mq_hw_ctx *hctx; | |
2f064a59 | 4040 | unsigned int state; |
bbd7bb70 | 4041 | |
529262d5 CH |
4042 | if (!blk_qc_t_valid(cookie) || |
4043 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
4044 | return 0; |
4045 | ||
529262d5 CH |
4046 | if (current->plug) |
4047 | blk_flush_plug_list(current->plug, false); | |
4048 | ||
1052b8ac JA |
4049 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
4050 | ||
06426adf JA |
4051 | /* |
4052 | * If we sleep, have the caller restart the poll loop to reset | |
4053 | * the state. Like for the other success return cases, the | |
4054 | * caller is responsible for checking if the IO completed. If | |
4055 | * the IO isn't complete, we'll get called again and will go | |
f6f371f7 PB |
4056 | * straight to the busy poll loop. If specified not to spin, |
4057 | * we also should not sleep. | |
06426adf | 4058 | */ |
f6f371f7 | 4059 | if (spin && blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 4060 | return 1; |
06426adf | 4061 | |
bbd7bb70 JA |
4062 | hctx->poll_considered++; |
4063 | ||
d6c23bb3 | 4064 | state = get_current_state(); |
aa61bec3 | 4065 | do { |
bbd7bb70 JA |
4066 | int ret; |
4067 | ||
4068 | hctx->poll_invoked++; | |
4069 | ||
9743139c | 4070 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
4071 | if (ret > 0) { |
4072 | hctx->poll_success++; | |
849a3700 | 4073 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4074 | return ret; |
bbd7bb70 JA |
4075 | } |
4076 | ||
4077 | if (signal_pending_state(state, current)) | |
849a3700 | 4078 | __set_current_state(TASK_RUNNING); |
bbd7bb70 | 4079 | |
b03fbd4f | 4080 | if (task_is_running(current)) |
85f4d4b6 | 4081 | return 1; |
0a1b8b87 | 4082 | if (ret < 0 || !spin) |
bbd7bb70 JA |
4083 | break; |
4084 | cpu_relax(); | |
aa61bec3 | 4085 | } while (!need_resched()); |
bbd7bb70 | 4086 | |
67b4110f | 4087 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4088 | return 0; |
bbd7bb70 | 4089 | } |
529262d5 | 4090 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 4091 | |
9cf2bab6 JA |
4092 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4093 | { | |
4094 | return rq->mq_ctx->cpu; | |
4095 | } | |
4096 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4097 | ||
320ae51f JA |
4098 | static int __init blk_mq_init(void) |
4099 | { | |
c3077b5d CH |
4100 | int i; |
4101 | ||
4102 | for_each_possible_cpu(i) | |
f9ab4918 | 4103 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4104 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4105 | ||
4106 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4107 | "block/softirq:dead", NULL, | |
4108 | blk_softirq_cpu_dead); | |
9467f859 TG |
4109 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4110 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4111 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4112 | blk_mq_hctx_notify_online, | |
4113 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4114 | return 0; |
4115 | } | |
4116 | subsys_initcall(blk_mq_init); |