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 | ||
e6e7abff | 357 | static struct request *__blk_mq_alloc_request(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; |
600c3b0c | 362 | unsigned int tag; |
d2c0d383 | 363 | |
6f816b4b TH |
364 | /* alloc_time includes depth and tag waits */ |
365 | if (blk_queue_rq_alloc_time(q)) | |
366 | alloc_time_ns = ktime_get_ns(); | |
367 | ||
f9afca4d | 368 | if (data->cmd_flags & REQ_NOWAIT) |
03a07c92 | 369 | data->flags |= BLK_MQ_REQ_NOWAIT; |
d2c0d383 CH |
370 | |
371 | if (e) { | |
d2c0d383 | 372 | /* |
8d663f34 | 373 | * Flush/passthrough requests are special and go directly to the |
17a51199 JA |
374 | * dispatch list. Don't include reserved tags in the |
375 | * limiting, as it isn't useful. | |
d2c0d383 | 376 | */ |
f9afca4d | 377 | if (!op_is_flush(data->cmd_flags) && |
8d663f34 | 378 | !blk_op_is_passthrough(data->cmd_flags) && |
f9afca4d | 379 | e->type->ops.limit_depth && |
17a51199 | 380 | !(data->flags & BLK_MQ_REQ_RESERVED)) |
f9afca4d | 381 | e->type->ops.limit_depth(data->cmd_flags, data); |
d2c0d383 CH |
382 | } |
383 | ||
bf0beec0 | 384 | retry: |
600c3b0c CH |
385 | data->ctx = blk_mq_get_ctx(q); |
386 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
42fdc5e4 | 387 | if (!e) |
600c3b0c CH |
388 | blk_mq_tag_busy(data->hctx); |
389 | ||
bf0beec0 ML |
390 | /* |
391 | * Waiting allocations only fail because of an inactive hctx. In that | |
392 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
393 | * should have migrated us to an online CPU by now. | |
394 | */ | |
e4cdf1a1 | 395 | tag = blk_mq_get_tag(data); |
bf0beec0 ML |
396 | if (tag == BLK_MQ_NO_TAG) { |
397 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
398 | return NULL; | |
399 | ||
400 | /* | |
401 | * Give up the CPU and sleep for a random short time to ensure | |
402 | * that thread using a realtime scheduling class are migrated | |
70f15a4f | 403 | * off the CPU, and thus off the hctx that is going away. |
bf0beec0 ML |
404 | */ |
405 | msleep(3); | |
406 | goto retry; | |
407 | } | |
7ea4d8a4 | 408 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); |
d2c0d383 CH |
409 | } |
410 | ||
cd6ce148 | 411 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
9a95e4ef | 412 | blk_mq_req_flags_t flags) |
320ae51f | 413 | { |
e6e7abff CH |
414 | struct blk_mq_alloc_data data = { |
415 | .q = q, | |
416 | .flags = flags, | |
417 | .cmd_flags = op, | |
418 | }; | |
bd166ef1 | 419 | struct request *rq; |
a492f075 | 420 | int ret; |
320ae51f | 421 | |
3a0a5299 | 422 | ret = blk_queue_enter(q, flags); |
a492f075 JL |
423 | if (ret) |
424 | return ERR_PTR(ret); | |
320ae51f | 425 | |
e6e7abff | 426 | rq = __blk_mq_alloc_request(&data); |
bd166ef1 | 427 | if (!rq) |
a5ea5811 | 428 | goto out_queue_exit; |
0c4de0f3 CH |
429 | rq->__data_len = 0; |
430 | rq->__sector = (sector_t) -1; | |
431 | rq->bio = rq->biotail = NULL; | |
320ae51f | 432 | return rq; |
a5ea5811 CH |
433 | out_queue_exit: |
434 | blk_queue_exit(q); | |
435 | return ERR_PTR(-EWOULDBLOCK); | |
320ae51f | 436 | } |
4bb659b1 | 437 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 438 | |
cd6ce148 | 439 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
9a95e4ef | 440 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) |
1f5bd336 | 441 | { |
e6e7abff CH |
442 | struct blk_mq_alloc_data data = { |
443 | .q = q, | |
444 | .flags = flags, | |
445 | .cmd_flags = op, | |
446 | }; | |
600c3b0c | 447 | u64 alloc_time_ns = 0; |
6d2809d5 | 448 | unsigned int cpu; |
600c3b0c | 449 | unsigned int tag; |
1f5bd336 ML |
450 | int ret; |
451 | ||
600c3b0c CH |
452 | /* alloc_time includes depth and tag waits */ |
453 | if (blk_queue_rq_alloc_time(q)) | |
454 | alloc_time_ns = ktime_get_ns(); | |
455 | ||
1f5bd336 ML |
456 | /* |
457 | * If the tag allocator sleeps we could get an allocation for a | |
458 | * different hardware context. No need to complicate the low level | |
459 | * allocator for this for the rare use case of a command tied to | |
460 | * a specific queue. | |
461 | */ | |
600c3b0c | 462 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) |
1f5bd336 ML |
463 | return ERR_PTR(-EINVAL); |
464 | ||
465 | if (hctx_idx >= q->nr_hw_queues) | |
466 | return ERR_PTR(-EIO); | |
467 | ||
3a0a5299 | 468 | ret = blk_queue_enter(q, flags); |
1f5bd336 ML |
469 | if (ret) |
470 | return ERR_PTR(ret); | |
471 | ||
c8712c6a CH |
472 | /* |
473 | * Check if the hardware context is actually mapped to anything. | |
474 | * If not tell the caller that it should skip this queue. | |
475 | */ | |
a5ea5811 | 476 | ret = -EXDEV; |
e6e7abff CH |
477 | data.hctx = q->queue_hw_ctx[hctx_idx]; |
478 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
a5ea5811 | 479 | goto out_queue_exit; |
e6e7abff CH |
480 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); |
481 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 482 | |
42fdc5e4 | 483 | if (!q->elevator) |
600c3b0c CH |
484 | blk_mq_tag_busy(data.hctx); |
485 | ||
a5ea5811 | 486 | ret = -EWOULDBLOCK; |
600c3b0c CH |
487 | tag = blk_mq_get_tag(&data); |
488 | if (tag == BLK_MQ_NO_TAG) | |
a5ea5811 | 489 | goto out_queue_exit; |
600c3b0c CH |
490 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); |
491 | ||
a5ea5811 CH |
492 | out_queue_exit: |
493 | blk_queue_exit(q); | |
494 | return ERR_PTR(ret); | |
1f5bd336 ML |
495 | } |
496 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
497 | ||
12f5b931 KB |
498 | static void __blk_mq_free_request(struct request *rq) |
499 | { | |
500 | struct request_queue *q = rq->q; | |
501 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 502 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
12f5b931 KB |
503 | const int sched_tag = rq->internal_tag; |
504 | ||
a892c8d5 | 505 | blk_crypto_free_request(rq); |
986d413b | 506 | blk_pm_mark_last_busy(rq); |
ea4f995e | 507 | rq->mq_hctx = NULL; |
76647368 | 508 | if (rq->tag != BLK_MQ_NO_TAG) |
cae740a0 | 509 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); |
76647368 | 510 | if (sched_tag != BLK_MQ_NO_TAG) |
cae740a0 | 511 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); |
12f5b931 KB |
512 | blk_mq_sched_restart(hctx); |
513 | blk_queue_exit(q); | |
514 | } | |
515 | ||
6af54051 | 516 | void blk_mq_free_request(struct request *rq) |
320ae51f | 517 | { |
320ae51f | 518 | struct request_queue *q = rq->q; |
6af54051 CH |
519 | struct elevator_queue *e = q->elevator; |
520 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
ea4f995e | 521 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
6af54051 | 522 | |
5bbf4e5a | 523 | if (rq->rq_flags & RQF_ELVPRIV) { |
f9cd4bfe JA |
524 | if (e && e->type->ops.finish_request) |
525 | e->type->ops.finish_request(rq); | |
6af54051 CH |
526 | if (rq->elv.icq) { |
527 | put_io_context(rq->elv.icq->ioc); | |
528 | rq->elv.icq = NULL; | |
529 | } | |
530 | } | |
320ae51f | 531 | |
6af54051 | 532 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 533 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
bccf5e26 | 534 | __blk_mq_dec_active_requests(hctx); |
87760e5e | 535 | |
7beb2f84 | 536 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
d152c682 | 537 | laptop_io_completion(q->disk->bdi); |
7beb2f84 | 538 | |
a7905043 | 539 | rq_qos_done(q, rq); |
0d2602ca | 540 | |
12f5b931 KB |
541 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
542 | if (refcount_dec_and_test(&rq->ref)) | |
543 | __blk_mq_free_request(rq); | |
320ae51f | 544 | } |
1a3b595a | 545 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 546 | |
2a842aca | 547 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 548 | { |
fe1f4526 JA |
549 | u64 now = 0; |
550 | ||
551 | if (blk_mq_need_time_stamp(rq)) | |
552 | now = ktime_get_ns(); | |
522a7775 | 553 | |
4bc6339a OS |
554 | if (rq->rq_flags & RQF_STATS) { |
555 | blk_mq_poll_stats_start(rq->q); | |
522a7775 | 556 | blk_stat_add(rq, now); |
4bc6339a OS |
557 | } |
558 | ||
87890092 | 559 | blk_mq_sched_completed_request(rq, now); |
ed88660a | 560 | |
522a7775 | 561 | blk_account_io_done(rq, now); |
0d11e6ac | 562 | |
91b63639 | 563 | if (rq->end_io) { |
a7905043 | 564 | rq_qos_done(rq->q, rq); |
320ae51f | 565 | rq->end_io(rq, error); |
91b63639 | 566 | } else { |
320ae51f | 567 | blk_mq_free_request(rq); |
91b63639 | 568 | } |
320ae51f | 569 | } |
c8a446ad | 570 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 571 | |
2a842aca | 572 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
573 | { |
574 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
575 | BUG(); | |
c8a446ad | 576 | __blk_mq_end_request(rq, error); |
63151a44 | 577 | } |
c8a446ad | 578 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 579 | |
f9ab4918 | 580 | static void blk_complete_reqs(struct llist_head *list) |
320ae51f | 581 | { |
f9ab4918 SAS |
582 | struct llist_node *entry = llist_reverse_order(llist_del_all(list)); |
583 | struct request *rq, *next; | |
c3077b5d | 584 | |
f9ab4918 | 585 | llist_for_each_entry_safe(rq, next, entry, ipi_list) |
c3077b5d | 586 | rq->q->mq_ops->complete(rq); |
320ae51f | 587 | } |
320ae51f | 588 | |
f9ab4918 | 589 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) |
320ae51f | 590 | { |
f9ab4918 | 591 | blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); |
115243f5 CH |
592 | } |
593 | ||
c3077b5d CH |
594 | static int blk_softirq_cpu_dead(unsigned int cpu) |
595 | { | |
f9ab4918 | 596 | blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); |
c3077b5d CH |
597 | return 0; |
598 | } | |
599 | ||
40d09b53 | 600 | static void __blk_mq_complete_request_remote(void *data) |
c3077b5d | 601 | { |
f9ab4918 | 602 | __raise_softirq_irqoff(BLOCK_SOFTIRQ); |
c3077b5d CH |
603 | } |
604 | ||
96339526 CH |
605 | static inline bool blk_mq_complete_need_ipi(struct request *rq) |
606 | { | |
607 | int cpu = raw_smp_processor_id(); | |
608 | ||
609 | if (!IS_ENABLED(CONFIG_SMP) || | |
610 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
611 | return false; | |
71425189 SAS |
612 | /* |
613 | * With force threaded interrupts enabled, raising softirq from an SMP | |
614 | * function call will always result in waking the ksoftirqd thread. | |
615 | * This is probably worse than completing the request on a different | |
616 | * cache domain. | |
617 | */ | |
91cc470e | 618 | if (force_irqthreads()) |
71425189 | 619 | return false; |
96339526 CH |
620 | |
621 | /* same CPU or cache domain? Complete locally */ | |
622 | if (cpu == rq->mq_ctx->cpu || | |
623 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
624 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
625 | return false; | |
626 | ||
627 | /* don't try to IPI to an offline CPU */ | |
628 | return cpu_online(rq->mq_ctx->cpu); | |
629 | } | |
630 | ||
f9ab4918 SAS |
631 | static void blk_mq_complete_send_ipi(struct request *rq) |
632 | { | |
633 | struct llist_head *list; | |
634 | unsigned int cpu; | |
635 | ||
636 | cpu = rq->mq_ctx->cpu; | |
637 | list = &per_cpu(blk_cpu_done, cpu); | |
638 | if (llist_add(&rq->ipi_list, list)) { | |
639 | INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq); | |
640 | smp_call_function_single_async(cpu, &rq->csd); | |
641 | } | |
642 | } | |
643 | ||
644 | static void blk_mq_raise_softirq(struct request *rq) | |
645 | { | |
646 | struct llist_head *list; | |
647 | ||
648 | preempt_disable(); | |
649 | list = this_cpu_ptr(&blk_cpu_done); | |
650 | if (llist_add(&rq->ipi_list, list)) | |
651 | raise_softirq(BLOCK_SOFTIRQ); | |
652 | preempt_enable(); | |
653 | } | |
654 | ||
40d09b53 | 655 | bool blk_mq_complete_request_remote(struct request *rq) |
320ae51f | 656 | { |
af78ff7c | 657 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); |
36e76539 | 658 | |
4ab32bf3 JA |
659 | /* |
660 | * For a polled request, always complete locallly, it's pointless | |
661 | * to redirect the completion. | |
662 | */ | |
40d09b53 CH |
663 | if (rq->cmd_flags & REQ_HIPRI) |
664 | return false; | |
38535201 | 665 | |
96339526 | 666 | if (blk_mq_complete_need_ipi(rq)) { |
f9ab4918 SAS |
667 | blk_mq_complete_send_ipi(rq); |
668 | return true; | |
3d6efbf6 | 669 | } |
40d09b53 | 670 | |
f9ab4918 SAS |
671 | if (rq->q->nr_hw_queues == 1) { |
672 | blk_mq_raise_softirq(rq); | |
673 | return true; | |
674 | } | |
675 | return false; | |
40d09b53 CH |
676 | } |
677 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
678 | ||
679 | /** | |
680 | * blk_mq_complete_request - end I/O on a request | |
681 | * @rq: the request being processed | |
682 | * | |
683 | * Description: | |
684 | * Complete a request by scheduling the ->complete_rq operation. | |
685 | **/ | |
686 | void blk_mq_complete_request(struct request *rq) | |
687 | { | |
688 | if (!blk_mq_complete_request_remote(rq)) | |
689 | rq->q->mq_ops->complete(rq); | |
320ae51f | 690 | } |
15f73f5b | 691 | EXPORT_SYMBOL(blk_mq_complete_request); |
30a91cb4 | 692 | |
04ced159 | 693 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) |
b7435db8 | 694 | __releases(hctx->srcu) |
04ced159 JA |
695 | { |
696 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
697 | rcu_read_unlock(); | |
698 | else | |
05707b64 | 699 | srcu_read_unlock(hctx->srcu, srcu_idx); |
04ced159 JA |
700 | } |
701 | ||
702 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
b7435db8 | 703 | __acquires(hctx->srcu) |
04ced159 | 704 | { |
08b5a6e2 JA |
705 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
706 | /* shut up gcc false positive */ | |
707 | *srcu_idx = 0; | |
04ced159 | 708 | rcu_read_lock(); |
08b5a6e2 | 709 | } else |
05707b64 | 710 | *srcu_idx = srcu_read_lock(hctx->srcu); |
04ced159 JA |
711 | } |
712 | ||
105663f7 AA |
713 | /** |
714 | * blk_mq_start_request - Start processing a request | |
715 | * @rq: Pointer to request to be started | |
716 | * | |
717 | * Function used by device drivers to notify the block layer that a request | |
718 | * is going to be processed now, so blk layer can do proper initializations | |
719 | * such as starting the timeout timer. | |
720 | */ | |
e2490073 | 721 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
722 | { |
723 | struct request_queue *q = rq->q; | |
724 | ||
a54895fa | 725 | trace_block_rq_issue(rq); |
320ae51f | 726 | |
cf43e6be | 727 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
544ccc8d | 728 | rq->io_start_time_ns = ktime_get_ns(); |
3d244306 | 729 | rq->stats_sectors = blk_rq_sectors(rq); |
cf43e6be | 730 | rq->rq_flags |= RQF_STATS; |
a7905043 | 731 | rq_qos_issue(q, rq); |
cf43e6be JA |
732 | } |
733 | ||
1d9bd516 | 734 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); |
538b7534 | 735 | |
1d9bd516 | 736 | blk_add_timer(rq); |
12f5b931 | 737 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); |
49f5baa5 | 738 | |
54d4e6ab MG |
739 | #ifdef CONFIG_BLK_DEV_INTEGRITY |
740 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
741 | q->integrity.profile->prepare_fn(rq); | |
742 | #endif | |
320ae51f | 743 | } |
e2490073 | 744 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 745 | |
ed0791b2 | 746 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
747 | { |
748 | struct request_queue *q = rq->q; | |
749 | ||
923218f6 ML |
750 | blk_mq_put_driver_tag(rq); |
751 | ||
a54895fa | 752 | trace_block_rq_requeue(rq); |
a7905043 | 753 | rq_qos_requeue(q, rq); |
49f5baa5 | 754 | |
12f5b931 KB |
755 | if (blk_mq_request_started(rq)) { |
756 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
da661267 | 757 | rq->rq_flags &= ~RQF_TIMED_OUT; |
e2490073 | 758 | } |
320ae51f JA |
759 | } |
760 | ||
2b053aca | 761 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 762 | { |
ed0791b2 | 763 | __blk_mq_requeue_request(rq); |
ed0791b2 | 764 | |
105976f5 ML |
765 | /* this request will be re-inserted to io scheduler queue */ |
766 | blk_mq_sched_requeue_request(rq); | |
767 | ||
7d692330 | 768 | BUG_ON(!list_empty(&rq->queuelist)); |
2b053aca | 769 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
770 | } |
771 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
772 | ||
6fca6a61 CH |
773 | static void blk_mq_requeue_work(struct work_struct *work) |
774 | { | |
775 | struct request_queue *q = | |
2849450a | 776 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
777 | LIST_HEAD(rq_list); |
778 | struct request *rq, *next; | |
6fca6a61 | 779 | |
18e9781d | 780 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 781 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 782 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
783 | |
784 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
aef1897c | 785 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) |
6fca6a61 CH |
786 | continue; |
787 | ||
e8064021 | 788 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 789 | list_del_init(&rq->queuelist); |
aef1897c JW |
790 | /* |
791 | * If RQF_DONTPREP, rq has contained some driver specific | |
792 | * data, so insert it to hctx dispatch list to avoid any | |
793 | * merge. | |
794 | */ | |
795 | if (rq->rq_flags & RQF_DONTPREP) | |
01e99aec | 796 | blk_mq_request_bypass_insert(rq, false, false); |
aef1897c JW |
797 | else |
798 | blk_mq_sched_insert_request(rq, true, false, false); | |
6fca6a61 CH |
799 | } |
800 | ||
801 | while (!list_empty(&rq_list)) { | |
802 | rq = list_entry(rq_list.next, struct request, queuelist); | |
803 | list_del_init(&rq->queuelist); | |
9e97d295 | 804 | blk_mq_sched_insert_request(rq, false, false, false); |
6fca6a61 CH |
805 | } |
806 | ||
52d7f1b5 | 807 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
808 | } |
809 | ||
2b053aca BVA |
810 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
811 | bool kick_requeue_list) | |
6fca6a61 CH |
812 | { |
813 | struct request_queue *q = rq->q; | |
814 | unsigned long flags; | |
815 | ||
816 | /* | |
817 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
ff821d27 | 818 | * request head insertion from the workqueue. |
6fca6a61 | 819 | */ |
e8064021 | 820 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
821 | |
822 | spin_lock_irqsave(&q->requeue_lock, flags); | |
823 | if (at_head) { | |
e8064021 | 824 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
825 | list_add(&rq->queuelist, &q->requeue_list); |
826 | } else { | |
827 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
828 | } | |
829 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
830 | |
831 | if (kick_requeue_list) | |
832 | blk_mq_kick_requeue_list(q); | |
6fca6a61 | 833 | } |
6fca6a61 CH |
834 | |
835 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
836 | { | |
ae943d20 | 837 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); |
6fca6a61 CH |
838 | } |
839 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
840 | ||
2849450a MS |
841 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
842 | unsigned long msecs) | |
843 | { | |
d4acf365 BVA |
844 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
845 | msecs_to_jiffies(msecs)); | |
2849450a MS |
846 | } |
847 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
848 | ||
0e62f51f JA |
849 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
850 | { | |
88c7b2b7 JA |
851 | if (tag < tags->nr_tags) { |
852 | prefetch(tags->rqs[tag]); | |
4ee86bab | 853 | return tags->rqs[tag]; |
88c7b2b7 | 854 | } |
4ee86bab HR |
855 | |
856 | return NULL; | |
24d2f903 CH |
857 | } |
858 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
859 | ||
3c94d83c JA |
860 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, |
861 | void *priv, bool reserved) | |
ae879912 JA |
862 | { |
863 | /* | |
05a4fed6 | 864 | * If we find a request that isn't idle and the queue matches, |
3c94d83c | 865 | * we know the queue is busy. Return false to stop the iteration. |
ae879912 | 866 | */ |
05a4fed6 | 867 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { |
ae879912 JA |
868 | bool *busy = priv; |
869 | ||
870 | *busy = true; | |
871 | return false; | |
872 | } | |
873 | ||
874 | return true; | |
875 | } | |
876 | ||
3c94d83c | 877 | bool blk_mq_queue_inflight(struct request_queue *q) |
ae879912 JA |
878 | { |
879 | bool busy = false; | |
880 | ||
3c94d83c | 881 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); |
ae879912 JA |
882 | return busy; |
883 | } | |
3c94d83c | 884 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); |
ae879912 | 885 | |
358f70da | 886 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 887 | { |
da661267 | 888 | req->rq_flags |= RQF_TIMED_OUT; |
d1210d5a CH |
889 | if (req->q->mq_ops->timeout) { |
890 | enum blk_eh_timer_return ret; | |
891 | ||
892 | ret = req->q->mq_ops->timeout(req, reserved); | |
893 | if (ret == BLK_EH_DONE) | |
894 | return; | |
895 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
46f92d42 | 896 | } |
d1210d5a CH |
897 | |
898 | blk_add_timer(req); | |
87ee7b11 | 899 | } |
5b3f25fc | 900 | |
12f5b931 | 901 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) |
81481eb4 | 902 | { |
12f5b931 | 903 | unsigned long deadline; |
87ee7b11 | 904 | |
12f5b931 KB |
905 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) |
906 | return false; | |
da661267 CH |
907 | if (rq->rq_flags & RQF_TIMED_OUT) |
908 | return false; | |
a7af0af3 | 909 | |
079076b3 | 910 | deadline = READ_ONCE(rq->deadline); |
12f5b931 KB |
911 | if (time_after_eq(jiffies, deadline)) |
912 | return true; | |
a7af0af3 | 913 | |
12f5b931 KB |
914 | if (*next == 0) |
915 | *next = deadline; | |
916 | else if (time_after(*next, deadline)) | |
917 | *next = deadline; | |
918 | return false; | |
87ee7b11 JA |
919 | } |
920 | ||
2e315dc0 ML |
921 | void blk_mq_put_rq_ref(struct request *rq) |
922 | { | |
a9ed27a7 | 923 | if (is_flush_rq(rq)) |
2e315dc0 ML |
924 | rq->end_io(rq, 0); |
925 | else if (refcount_dec_and_test(&rq->ref)) | |
926 | __blk_mq_free_request(rq); | |
927 | } | |
928 | ||
7baa8572 | 929 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
1d9bd516 TH |
930 | struct request *rq, void *priv, bool reserved) |
931 | { | |
12f5b931 KB |
932 | unsigned long *next = priv; |
933 | ||
934 | /* | |
c797b40c ML |
935 | * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot |
936 | * be reallocated underneath the timeout handler's processing, then | |
937 | * the expire check is reliable. If the request is not expired, then | |
938 | * it was completed and reallocated as a new request after returning | |
939 | * from blk_mq_check_expired(). | |
1d9bd516 | 940 | */ |
12f5b931 | 941 | if (blk_mq_req_expired(rq, next)) |
1d9bd516 | 942 | blk_mq_rq_timed_out(rq, reserved); |
7baa8572 | 943 | return true; |
1d9bd516 TH |
944 | } |
945 | ||
287922eb | 946 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 947 | { |
287922eb CH |
948 | struct request_queue *q = |
949 | container_of(work, struct request_queue, timeout_work); | |
12f5b931 | 950 | unsigned long next = 0; |
1d9bd516 | 951 | struct blk_mq_hw_ctx *hctx; |
81481eb4 | 952 | int i; |
320ae51f | 953 | |
71f79fb3 GKB |
954 | /* A deadlock might occur if a request is stuck requiring a |
955 | * timeout at the same time a queue freeze is waiting | |
956 | * completion, since the timeout code would not be able to | |
957 | * acquire the queue reference here. | |
958 | * | |
959 | * That's why we don't use blk_queue_enter here; instead, we use | |
960 | * percpu_ref_tryget directly, because we need to be able to | |
961 | * obtain a reference even in the short window between the queue | |
962 | * starting to freeze, by dropping the first reference in | |
1671d522 | 963 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
964 | * consumed, marked by the instant q_usage_counter reaches |
965 | * zero. | |
966 | */ | |
967 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
968 | return; |
969 | ||
12f5b931 | 970 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); |
320ae51f | 971 | |
12f5b931 KB |
972 | if (next != 0) { |
973 | mod_timer(&q->timeout, next); | |
0d2602ca | 974 | } else { |
fcd36c36 BVA |
975 | /* |
976 | * Request timeouts are handled as a forward rolling timer. If | |
977 | * we end up here it means that no requests are pending and | |
978 | * also that no request has been pending for a while. Mark | |
979 | * each hctx as idle. | |
980 | */ | |
f054b56c ML |
981 | queue_for_each_hw_ctx(q, hctx, i) { |
982 | /* the hctx may be unmapped, so check it here */ | |
983 | if (blk_mq_hw_queue_mapped(hctx)) | |
984 | blk_mq_tag_idle(hctx); | |
985 | } | |
0d2602ca | 986 | } |
287922eb | 987 | blk_queue_exit(q); |
320ae51f JA |
988 | } |
989 | ||
88459642 OS |
990 | struct flush_busy_ctx_data { |
991 | struct blk_mq_hw_ctx *hctx; | |
992 | struct list_head *list; | |
993 | }; | |
994 | ||
995 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
996 | { | |
997 | struct flush_busy_ctx_data *flush_data = data; | |
998 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
999 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1000 | enum hctx_type type = hctx->type; |
88459642 | 1001 | |
88459642 | 1002 | spin_lock(&ctx->lock); |
c16d6b5a | 1003 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); |
e9a99a63 | 1004 | sbitmap_clear_bit(sb, bitnr); |
88459642 OS |
1005 | spin_unlock(&ctx->lock); |
1006 | return true; | |
1007 | } | |
1008 | ||
1429d7c9 JA |
1009 | /* |
1010 | * Process software queues that have been marked busy, splicing them | |
1011 | * to the for-dispatch | |
1012 | */ | |
2c3ad667 | 1013 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 1014 | { |
88459642 OS |
1015 | struct flush_busy_ctx_data data = { |
1016 | .hctx = hctx, | |
1017 | .list = list, | |
1018 | }; | |
1429d7c9 | 1019 | |
88459642 | 1020 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 1021 | } |
2c3ad667 | 1022 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 1023 | |
b347689f ML |
1024 | struct dispatch_rq_data { |
1025 | struct blk_mq_hw_ctx *hctx; | |
1026 | struct request *rq; | |
1027 | }; | |
1028 | ||
1029 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1030 | void *data) | |
1031 | { | |
1032 | struct dispatch_rq_data *dispatch_data = data; | |
1033 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1034 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
c16d6b5a | 1035 | enum hctx_type type = hctx->type; |
b347689f ML |
1036 | |
1037 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
1038 | if (!list_empty(&ctx->rq_lists[type])) { |
1039 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
b347689f | 1040 | list_del_init(&dispatch_data->rq->queuelist); |
c16d6b5a | 1041 | if (list_empty(&ctx->rq_lists[type])) |
b347689f ML |
1042 | sbitmap_clear_bit(sb, bitnr); |
1043 | } | |
1044 | spin_unlock(&ctx->lock); | |
1045 | ||
1046 | return !dispatch_data->rq; | |
1047 | } | |
1048 | ||
1049 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1050 | struct blk_mq_ctx *start) | |
1051 | { | |
f31967f0 | 1052 | unsigned off = start ? start->index_hw[hctx->type] : 0; |
b347689f ML |
1053 | struct dispatch_rq_data data = { |
1054 | .hctx = hctx, | |
1055 | .rq = NULL, | |
1056 | }; | |
1057 | ||
1058 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1059 | dispatch_rq_from_ctx, &data); | |
1060 | ||
1061 | return data.rq; | |
1062 | } | |
1063 | ||
703fd1c0 JA |
1064 | static inline unsigned int queued_to_index(unsigned int queued) |
1065 | { | |
1066 | if (!queued) | |
1067 | return 0; | |
1429d7c9 | 1068 | |
703fd1c0 | 1069 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
1070 | } |
1071 | ||
570e9b73 ML |
1072 | static bool __blk_mq_get_driver_tag(struct request *rq) |
1073 | { | |
222a5ae0 | 1074 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; |
570e9b73 | 1075 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; |
570e9b73 ML |
1076 | int tag; |
1077 | ||
568f2700 ML |
1078 | blk_mq_tag_busy(rq->mq_hctx); |
1079 | ||
570e9b73 | 1080 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { |
222a5ae0 | 1081 | bt = rq->mq_hctx->tags->breserved_tags; |
570e9b73 | 1082 | tag_offset = 0; |
28500850 ML |
1083 | } else { |
1084 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1085 | return false; | |
570e9b73 ML |
1086 | } |
1087 | ||
570e9b73 ML |
1088 | tag = __sbitmap_queue_get(bt); |
1089 | if (tag == BLK_MQ_NO_TAG) | |
1090 | return false; | |
1091 | ||
1092 | rq->tag = tag + tag_offset; | |
570e9b73 ML |
1093 | return true; |
1094 | } | |
1095 | ||
61347154 | 1096 | bool blk_mq_get_driver_tag(struct request *rq) |
570e9b73 | 1097 | { |
568f2700 ML |
1098 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
1099 | ||
1100 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1101 | return false; | |
1102 | ||
51db1c37 | 1103 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && |
568f2700 ML |
1104 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { |
1105 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
bccf5e26 | 1106 | __blk_mq_inc_active_requests(hctx); |
568f2700 ML |
1107 | } |
1108 | hctx->tags->rqs[rq->tag] = rq; | |
1109 | return true; | |
570e9b73 ML |
1110 | } |
1111 | ||
eb619fdb JA |
1112 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, |
1113 | int flags, void *key) | |
da55f2cc OS |
1114 | { |
1115 | struct blk_mq_hw_ctx *hctx; | |
1116 | ||
1117 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1118 | ||
5815839b | 1119 | spin_lock(&hctx->dispatch_wait_lock); |
e8618575 JA |
1120 | if (!list_empty(&wait->entry)) { |
1121 | struct sbitmap_queue *sbq; | |
1122 | ||
1123 | list_del_init(&wait->entry); | |
222a5ae0 | 1124 | sbq = hctx->tags->bitmap_tags; |
e8618575 JA |
1125 | atomic_dec(&sbq->ws_active); |
1126 | } | |
5815839b ML |
1127 | spin_unlock(&hctx->dispatch_wait_lock); |
1128 | ||
da55f2cc OS |
1129 | blk_mq_run_hw_queue(hctx, true); |
1130 | return 1; | |
1131 | } | |
1132 | ||
f906a6a0 JA |
1133 | /* |
1134 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
ee3e4de5 BVA |
1135 | * the tag wakeups. For non-shared tags, we can simply mark us needing a |
1136 | * restart. For both cases, take care to check the condition again after | |
f906a6a0 JA |
1137 | * marking us as waiting. |
1138 | */ | |
2278d69f | 1139 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, |
f906a6a0 | 1140 | struct request *rq) |
da55f2cc | 1141 | { |
222a5ae0 | 1142 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; |
5815839b | 1143 | struct wait_queue_head *wq; |
f906a6a0 JA |
1144 | wait_queue_entry_t *wait; |
1145 | bool ret; | |
da55f2cc | 1146 | |
51db1c37 | 1147 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
684b7324 | 1148 | blk_mq_sched_mark_restart_hctx(hctx); |
f906a6a0 | 1149 | |
c27d53fb BVA |
1150 | /* |
1151 | * It's possible that a tag was freed in the window between the | |
1152 | * allocation failure and adding the hardware queue to the wait | |
1153 | * queue. | |
1154 | * | |
1155 | * Don't clear RESTART here, someone else could have set it. | |
1156 | * At most this will cost an extra queue run. | |
1157 | */ | |
8ab6bb9e | 1158 | return blk_mq_get_driver_tag(rq); |
eb619fdb | 1159 | } |
eb619fdb | 1160 | |
2278d69f | 1161 | wait = &hctx->dispatch_wait; |
c27d53fb BVA |
1162 | if (!list_empty_careful(&wait->entry)) |
1163 | return false; | |
1164 | ||
e8618575 | 1165 | wq = &bt_wait_ptr(sbq, hctx)->wait; |
5815839b ML |
1166 | |
1167 | spin_lock_irq(&wq->lock); | |
1168 | spin_lock(&hctx->dispatch_wait_lock); | |
c27d53fb | 1169 | if (!list_empty(&wait->entry)) { |
5815839b ML |
1170 | spin_unlock(&hctx->dispatch_wait_lock); |
1171 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1172 | return false; |
eb619fdb JA |
1173 | } |
1174 | ||
e8618575 | 1175 | atomic_inc(&sbq->ws_active); |
5815839b ML |
1176 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; |
1177 | __add_wait_queue(wq, wait); | |
c27d53fb | 1178 | |
da55f2cc | 1179 | /* |
eb619fdb JA |
1180 | * It's possible that a tag was freed in the window between the |
1181 | * allocation failure and adding the hardware queue to the wait | |
1182 | * queue. | |
da55f2cc | 1183 | */ |
8ab6bb9e | 1184 | ret = blk_mq_get_driver_tag(rq); |
c27d53fb | 1185 | if (!ret) { |
5815839b ML |
1186 | spin_unlock(&hctx->dispatch_wait_lock); |
1187 | spin_unlock_irq(&wq->lock); | |
c27d53fb | 1188 | return false; |
eb619fdb | 1189 | } |
c27d53fb BVA |
1190 | |
1191 | /* | |
1192 | * We got a tag, remove ourselves from the wait queue to ensure | |
1193 | * someone else gets the wakeup. | |
1194 | */ | |
c27d53fb | 1195 | list_del_init(&wait->entry); |
e8618575 | 1196 | atomic_dec(&sbq->ws_active); |
5815839b ML |
1197 | spin_unlock(&hctx->dispatch_wait_lock); |
1198 | spin_unlock_irq(&wq->lock); | |
c27d53fb BVA |
1199 | |
1200 | return true; | |
da55f2cc OS |
1201 | } |
1202 | ||
6e768717 ML |
1203 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 |
1204 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1205 | /* | |
1206 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1207 | * - EWMA is one simple way to compute running average value | |
1208 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1209 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1210 | * factor doesn't matter because EWMA decreases exponentially | |
1211 | */ | |
1212 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1213 | { | |
1214 | unsigned int ewma; | |
1215 | ||
6e768717 ML |
1216 | ewma = hctx->dispatch_busy; |
1217 | ||
1218 | if (!ewma && !busy) | |
1219 | return; | |
1220 | ||
1221 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1222 | if (busy) | |
1223 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1224 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1225 | ||
1226 | hctx->dispatch_busy = ewma; | |
1227 | } | |
1228 | ||
86ff7c2a ML |
1229 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ |
1230 | ||
c92a4103 JT |
1231 | static void blk_mq_handle_dev_resource(struct request *rq, |
1232 | struct list_head *list) | |
1233 | { | |
1234 | struct request *next = | |
1235 | list_first_entry_or_null(list, struct request, queuelist); | |
1236 | ||
1237 | /* | |
1238 | * If an I/O scheduler has been configured and we got a driver tag for | |
1239 | * the next request already, free it. | |
1240 | */ | |
1241 | if (next) | |
1242 | blk_mq_put_driver_tag(next); | |
1243 | ||
1244 | list_add(&rq->queuelist, list); | |
1245 | __blk_mq_requeue_request(rq); | |
1246 | } | |
1247 | ||
0512a75b KB |
1248 | static void blk_mq_handle_zone_resource(struct request *rq, |
1249 | struct list_head *zone_list) | |
1250 | { | |
1251 | /* | |
1252 | * If we end up here it is because we cannot dispatch a request to a | |
1253 | * specific zone due to LLD level zone-write locking or other zone | |
1254 | * related resource not being available. In this case, set the request | |
1255 | * aside in zone_list for retrying it later. | |
1256 | */ | |
1257 | list_add(&rq->queuelist, zone_list); | |
1258 | __blk_mq_requeue_request(rq); | |
1259 | } | |
1260 | ||
75383524 ML |
1261 | enum prep_dispatch { |
1262 | PREP_DISPATCH_OK, | |
1263 | PREP_DISPATCH_NO_TAG, | |
1264 | PREP_DISPATCH_NO_BUDGET, | |
1265 | }; | |
1266 | ||
1267 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1268 | bool need_budget) | |
1269 | { | |
1270 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2a5a24aa | 1271 | int budget_token = -1; |
75383524 | 1272 | |
2a5a24aa ML |
1273 | if (need_budget) { |
1274 | budget_token = blk_mq_get_dispatch_budget(rq->q); | |
1275 | if (budget_token < 0) { | |
1276 | blk_mq_put_driver_tag(rq); | |
1277 | return PREP_DISPATCH_NO_BUDGET; | |
1278 | } | |
1279 | blk_mq_set_rq_budget_token(rq, budget_token); | |
75383524 ML |
1280 | } |
1281 | ||
1282 | if (!blk_mq_get_driver_tag(rq)) { | |
1283 | /* | |
1284 | * The initial allocation attempt failed, so we need to | |
1285 | * rerun the hardware queue when a tag is freed. The | |
1286 | * waitqueue takes care of that. If the queue is run | |
1287 | * before we add this entry back on the dispatch list, | |
1288 | * we'll re-run it below. | |
1289 | */ | |
1290 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1fd40b5e ML |
1291 | /* |
1292 | * All budgets not got from this function will be put | |
1293 | * together during handling partial dispatch | |
1294 | */ | |
1295 | if (need_budget) | |
2a5a24aa | 1296 | blk_mq_put_dispatch_budget(rq->q, budget_token); |
75383524 ML |
1297 | return PREP_DISPATCH_NO_TAG; |
1298 | } | |
1299 | } | |
1300 | ||
1301 | return PREP_DISPATCH_OK; | |
1302 | } | |
1303 | ||
1fd40b5e ML |
1304 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ |
1305 | static void blk_mq_release_budgets(struct request_queue *q, | |
2a5a24aa | 1306 | struct list_head *list) |
1fd40b5e | 1307 | { |
2a5a24aa | 1308 | struct request *rq; |
1fd40b5e | 1309 | |
2a5a24aa ML |
1310 | list_for_each_entry(rq, list, queuelist) { |
1311 | int budget_token = blk_mq_get_rq_budget_token(rq); | |
1fd40b5e | 1312 | |
2a5a24aa ML |
1313 | if (budget_token >= 0) |
1314 | blk_mq_put_dispatch_budget(q, budget_token); | |
1315 | } | |
1fd40b5e ML |
1316 | } |
1317 | ||
1f57f8d4 JA |
1318 | /* |
1319 | * Returns true if we did some work AND can potentially do more. | |
1320 | */ | |
445874e8 | 1321 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, |
1fd40b5e | 1322 | unsigned int nr_budgets) |
320ae51f | 1323 | { |
75383524 | 1324 | enum prep_dispatch prep; |
445874e8 | 1325 | struct request_queue *q = hctx->queue; |
6d6f167c | 1326 | struct request *rq, *nxt; |
fc17b653 | 1327 | int errors, queued; |
86ff7c2a | 1328 | blk_status_t ret = BLK_STS_OK; |
0512a75b | 1329 | LIST_HEAD(zone_list); |
320ae51f | 1330 | |
81380ca1 OS |
1331 | if (list_empty(list)) |
1332 | return false; | |
1333 | ||
320ae51f JA |
1334 | /* |
1335 | * Now process all the entries, sending them to the driver. | |
1336 | */ | |
93efe981 | 1337 | errors = queued = 0; |
81380ca1 | 1338 | do { |
74c45052 | 1339 | struct blk_mq_queue_data bd; |
320ae51f | 1340 | |
f04c3df3 | 1341 | rq = list_first_entry(list, struct request, queuelist); |
0bca799b | 1342 | |
445874e8 | 1343 | WARN_ON_ONCE(hctx != rq->mq_hctx); |
1fd40b5e | 1344 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); |
75383524 | 1345 | if (prep != PREP_DISPATCH_OK) |
0bca799b | 1346 | break; |
de148297 | 1347 | |
320ae51f | 1348 | list_del_init(&rq->queuelist); |
320ae51f | 1349 | |
74c45052 | 1350 | bd.rq = rq; |
113285b4 JA |
1351 | |
1352 | /* | |
1353 | * Flag last if we have no more requests, or if we have more | |
1354 | * but can't assign a driver tag to it. | |
1355 | */ | |
1356 | if (list_empty(list)) | |
1357 | bd.last = true; | |
1358 | else { | |
113285b4 | 1359 | nxt = list_first_entry(list, struct request, queuelist); |
8ab6bb9e | 1360 | bd.last = !blk_mq_get_driver_tag(nxt); |
113285b4 | 1361 | } |
74c45052 | 1362 | |
1fd40b5e ML |
1363 | /* |
1364 | * once the request is queued to lld, no need to cover the | |
1365 | * budget any more | |
1366 | */ | |
1367 | if (nr_budgets) | |
1368 | nr_budgets--; | |
74c45052 | 1369 | ret = q->mq_ops->queue_rq(hctx, &bd); |
7bf13729 ML |
1370 | switch (ret) { |
1371 | case BLK_STS_OK: | |
1372 | queued++; | |
320ae51f | 1373 | break; |
7bf13729 ML |
1374 | case BLK_STS_RESOURCE: |
1375 | case BLK_STS_DEV_RESOURCE: | |
1376 | blk_mq_handle_dev_resource(rq, list); | |
1377 | goto out; | |
1378 | case BLK_STS_ZONE_RESOURCE: | |
0512a75b KB |
1379 | /* |
1380 | * Move the request to zone_list and keep going through | |
1381 | * the dispatch list to find more requests the drive can | |
1382 | * accept. | |
1383 | */ | |
1384 | blk_mq_handle_zone_resource(rq, &zone_list); | |
7bf13729 ML |
1385 | break; |
1386 | default: | |
93efe981 | 1387 | errors++; |
e21ee5a6 | 1388 | blk_mq_end_request(rq, ret); |
320ae51f | 1389 | } |
81380ca1 | 1390 | } while (!list_empty(list)); |
7bf13729 | 1391 | out: |
0512a75b KB |
1392 | if (!list_empty(&zone_list)) |
1393 | list_splice_tail_init(&zone_list, list); | |
1394 | ||
703fd1c0 | 1395 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f | 1396 | |
632bfb63 | 1397 | /* If we didn't flush the entire list, we could have told the driver |
1398 | * there was more coming, but that turned out to be a lie. | |
1399 | */ | |
1400 | if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued) | |
1401 | q->mq_ops->commit_rqs(hctx); | |
320ae51f JA |
1402 | /* |
1403 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1404 | * that is where we will continue on next queue run. | |
1405 | */ | |
f04c3df3 | 1406 | if (!list_empty(list)) { |
86ff7c2a | 1407 | bool needs_restart; |
75383524 ML |
1408 | /* For non-shared tags, the RESTART check will suffice */ |
1409 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
51db1c37 | 1410 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); |
75383524 | 1411 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; |
86ff7c2a | 1412 | |
2a5a24aa ML |
1413 | if (nr_budgets) |
1414 | blk_mq_release_budgets(q, list); | |
86ff7c2a | 1415 | |
320ae51f | 1416 | spin_lock(&hctx->lock); |
01e99aec | 1417 | list_splice_tail_init(list, &hctx->dispatch); |
320ae51f | 1418 | spin_unlock(&hctx->lock); |
f04c3df3 | 1419 | |
d7d8535f ML |
1420 | /* |
1421 | * Order adding requests to hctx->dispatch and checking | |
1422 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1423 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1424 | * miss the new added requests to hctx->dispatch, meantime | |
1425 | * SCHED_RESTART is observed here. | |
1426 | */ | |
1427 | smp_mb(); | |
1428 | ||
9ba52e58 | 1429 | /* |
710c785f BVA |
1430 | * If SCHED_RESTART was set by the caller of this function and |
1431 | * it is no longer set that means that it was cleared by another | |
1432 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1433 | * |
eb619fdb JA |
1434 | * If 'no_tag' is set, that means that we failed getting |
1435 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1436 | * waitqueue is no longer active, ensure that we run the queue | |
1437 | * AFTER adding our entries back to the list. | |
bd166ef1 | 1438 | * |
710c785f BVA |
1439 | * If no I/O scheduler has been configured it is possible that |
1440 | * the hardware queue got stopped and restarted before requests | |
1441 | * were pushed back onto the dispatch list. Rerun the queue to | |
1442 | * avoid starvation. Notes: | |
1443 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1444 | * been stopped before rerunning a queue. | |
1445 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1446 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1447 | * and dm-rq. |
86ff7c2a ML |
1448 | * |
1449 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1450 | * bit is set, run queue after a delay to avoid IO stalls | |
ab3cee37 DA |
1451 | * that could otherwise occur if the queue is idle. We'll do |
1452 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
bd166ef1 | 1453 | */ |
86ff7c2a ML |
1454 | needs_restart = blk_mq_sched_needs_restart(hctx); |
1455 | if (!needs_restart || | |
eb619fdb | 1456 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) |
bd166ef1 | 1457 | blk_mq_run_hw_queue(hctx, true); |
ab3cee37 DA |
1458 | else if (needs_restart && (ret == BLK_STS_RESOURCE || |
1459 | no_budget_avail)) | |
86ff7c2a | 1460 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); |
1f57f8d4 | 1461 | |
6e768717 | 1462 | blk_mq_update_dispatch_busy(hctx, true); |
1f57f8d4 | 1463 | return false; |
6e768717 ML |
1464 | } else |
1465 | blk_mq_update_dispatch_busy(hctx, false); | |
f04c3df3 | 1466 | |
93efe981 | 1467 | return (queued + errors) != 0; |
f04c3df3 JA |
1468 | } |
1469 | ||
105663f7 AA |
1470 | /** |
1471 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1472 | * @hctx: Pointer to the hardware queue to run. | |
1473 | * | |
1474 | * Send pending requests to the hardware. | |
1475 | */ | |
6a83e74d BVA |
1476 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1477 | { | |
1478 | int srcu_idx; | |
1479 | ||
b7a71e66 JA |
1480 | /* |
1481 | * We can't run the queue inline with ints disabled. Ensure that | |
1482 | * we catch bad users of this early. | |
1483 | */ | |
1484 | WARN_ON_ONCE(in_interrupt()); | |
1485 | ||
04ced159 | 1486 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); |
bf4907c0 | 1487 | |
04ced159 JA |
1488 | hctx_lock(hctx, &srcu_idx); |
1489 | blk_mq_sched_dispatch_requests(hctx); | |
1490 | hctx_unlock(hctx, srcu_idx); | |
6a83e74d BVA |
1491 | } |
1492 | ||
f82ddf19 ML |
1493 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) |
1494 | { | |
1495 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1496 | ||
1497 | if (cpu >= nr_cpu_ids) | |
1498 | cpu = cpumask_first(hctx->cpumask); | |
1499 | return cpu; | |
1500 | } | |
1501 | ||
506e931f JA |
1502 | /* |
1503 | * It'd be great if the workqueue API had a way to pass | |
1504 | * in a mask and had some smarts for more clever placement. | |
1505 | * For now we just round-robin here, switching for every | |
1506 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1507 | */ | |
1508 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1509 | { | |
7bed4595 | 1510 | bool tried = false; |
476f8c98 | 1511 | int next_cpu = hctx->next_cpu; |
7bed4595 | 1512 | |
b657d7e6 CH |
1513 | if (hctx->queue->nr_hw_queues == 1) |
1514 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1515 | |
1516 | if (--hctx->next_cpu_batch <= 0) { | |
7bed4595 | 1517 | select_cpu: |
476f8c98 | 1518 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, |
20e4d813 | 1519 | cpu_online_mask); |
506e931f | 1520 | if (next_cpu >= nr_cpu_ids) |
f82ddf19 | 1521 | next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
1522 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
1523 | } | |
1524 | ||
7bed4595 ML |
1525 | /* |
1526 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1527 | * and it should only happen in the path of handling CPU DEAD. | |
1528 | */ | |
476f8c98 | 1529 | if (!cpu_online(next_cpu)) { |
7bed4595 ML |
1530 | if (!tried) { |
1531 | tried = true; | |
1532 | goto select_cpu; | |
1533 | } | |
1534 | ||
1535 | /* | |
1536 | * Make sure to re-select CPU next time once after CPUs | |
1537 | * in hctx->cpumask become online again. | |
1538 | */ | |
476f8c98 | 1539 | hctx->next_cpu = next_cpu; |
7bed4595 ML |
1540 | hctx->next_cpu_batch = 1; |
1541 | return WORK_CPU_UNBOUND; | |
1542 | } | |
476f8c98 ML |
1543 | |
1544 | hctx->next_cpu = next_cpu; | |
1545 | return next_cpu; | |
506e931f JA |
1546 | } |
1547 | ||
105663f7 AA |
1548 | /** |
1549 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1550 | * @hctx: Pointer to the hardware queue to run. | |
1551 | * @async: If we want to run the queue asynchronously. | |
fa94ba8a | 1552 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1553 | * |
1554 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1555 | * with a delay of @msecs. | |
1556 | */ | |
7587a5ae BVA |
1557 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1558 | unsigned long msecs) | |
320ae51f | 1559 | { |
5435c023 | 1560 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
1561 | return; |
1562 | ||
1b792f2f | 1563 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1564 | int cpu = get_cpu(); |
1565 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1566 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1567 | put_cpu(); |
398205b8 PB |
1568 | return; |
1569 | } | |
e4043dcf | 1570 | |
2a90d4aa | 1571 | put_cpu(); |
e4043dcf | 1572 | } |
398205b8 | 1573 | |
ae943d20 BVA |
1574 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, |
1575 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1576 | } |
1577 | ||
105663f7 AA |
1578 | /** |
1579 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1580 | * @hctx: Pointer to the hardware queue to run. | |
fa94ba8a | 1581 | * @msecs: Milliseconds of delay to wait before running the queue. |
105663f7 AA |
1582 | * |
1583 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1584 | */ | |
7587a5ae BVA |
1585 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) |
1586 | { | |
1587 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1588 | } | |
1589 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1590 | ||
105663f7 AA |
1591 | /** |
1592 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1593 | * @hctx: Pointer to the hardware queue to run. | |
1594 | * @async: If we want to run the queue asynchronously. | |
1595 | * | |
1596 | * Check if the request queue is not in a quiesced state and if there are | |
1597 | * pending requests to be sent. If this is true, run the queue to send requests | |
1598 | * to hardware. | |
1599 | */ | |
626fb735 | 1600 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
7587a5ae | 1601 | { |
24f5a90f ML |
1602 | int srcu_idx; |
1603 | bool need_run; | |
1604 | ||
1605 | /* | |
1606 | * When queue is quiesced, we may be switching io scheduler, or | |
1607 | * updating nr_hw_queues, or other things, and we can't run queue | |
1608 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1609 | * | |
1610 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1611 | * quiesced. | |
1612 | */ | |
04ced159 JA |
1613 | hctx_lock(hctx, &srcu_idx); |
1614 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1615 | blk_mq_hctx_has_pending(hctx); | |
1616 | hctx_unlock(hctx, srcu_idx); | |
24f5a90f | 1617 | |
626fb735 | 1618 | if (need_run) |
79f720a7 | 1619 | __blk_mq_delay_run_hw_queue(hctx, async, 0); |
320ae51f | 1620 | } |
5b727272 | 1621 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1622 | |
b6e68ee8 JK |
1623 | /* |
1624 | * Is the request queue handled by an IO scheduler that does not respect | |
1625 | * hardware queues when dispatching? | |
1626 | */ | |
1627 | static bool blk_mq_has_sqsched(struct request_queue *q) | |
1628 | { | |
1629 | struct elevator_queue *e = q->elevator; | |
1630 | ||
1631 | if (e && e->type->ops.dispatch_request && | |
1632 | !(e->type->elevator_features & ELEVATOR_F_MQ_AWARE)) | |
1633 | return true; | |
1634 | return false; | |
1635 | } | |
1636 | ||
1637 | /* | |
1638 | * Return prefered queue to dispatch from (if any) for non-mq aware IO | |
1639 | * scheduler. | |
1640 | */ | |
1641 | static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) | |
1642 | { | |
1643 | struct blk_mq_hw_ctx *hctx; | |
1644 | ||
1645 | /* | |
1646 | * If the IO scheduler does not respect hardware queues when | |
1647 | * dispatching, we just don't bother with multiple HW queues and | |
1648 | * dispatch from hctx for the current CPU since running multiple queues | |
1649 | * just causes lock contention inside the scheduler and pointless cache | |
1650 | * bouncing. | |
1651 | */ | |
1652 | hctx = blk_mq_map_queue_type(q, HCTX_TYPE_DEFAULT, | |
1653 | raw_smp_processor_id()); | |
1654 | if (!blk_mq_hctx_stopped(hctx)) | |
1655 | return hctx; | |
1656 | return NULL; | |
1657 | } | |
1658 | ||
105663f7 | 1659 | /** |
24f7bb88 | 1660 | * blk_mq_run_hw_queues - Run all hardware queues in a request queue. |
105663f7 AA |
1661 | * @q: Pointer to the request queue to run. |
1662 | * @async: If we want to run the queue asynchronously. | |
1663 | */ | |
b94ec296 | 1664 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f | 1665 | { |
b6e68ee8 | 1666 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
320ae51f JA |
1667 | int i; |
1668 | ||
b6e68ee8 JK |
1669 | sq_hctx = NULL; |
1670 | if (blk_mq_has_sqsched(q)) | |
1671 | sq_hctx = blk_mq_get_sq_hctx(q); | |
320ae51f | 1672 | queue_for_each_hw_ctx(q, hctx, i) { |
79f720a7 | 1673 | if (blk_mq_hctx_stopped(hctx)) |
320ae51f | 1674 | continue; |
b6e68ee8 JK |
1675 | /* |
1676 | * Dispatch from this hctx either if there's no hctx preferred | |
1677 | * by IO scheduler or if it has requests that bypass the | |
1678 | * scheduler. | |
1679 | */ | |
1680 | if (!sq_hctx || sq_hctx == hctx || | |
1681 | !list_empty_careful(&hctx->dispatch)) | |
1682 | blk_mq_run_hw_queue(hctx, async); | |
320ae51f JA |
1683 | } |
1684 | } | |
b94ec296 | 1685 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1686 | |
b9151e7b DA |
1687 | /** |
1688 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1689 | * @q: Pointer to the request queue to run. | |
fa94ba8a | 1690 | * @msecs: Milliseconds of delay to wait before running the queues. |
b9151e7b DA |
1691 | */ |
1692 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1693 | { | |
b6e68ee8 | 1694 | struct blk_mq_hw_ctx *hctx, *sq_hctx; |
b9151e7b DA |
1695 | int i; |
1696 | ||
b6e68ee8 JK |
1697 | sq_hctx = NULL; |
1698 | if (blk_mq_has_sqsched(q)) | |
1699 | sq_hctx = blk_mq_get_sq_hctx(q); | |
b9151e7b DA |
1700 | queue_for_each_hw_ctx(q, hctx, i) { |
1701 | if (blk_mq_hctx_stopped(hctx)) | |
1702 | continue; | |
b6e68ee8 JK |
1703 | /* |
1704 | * Dispatch from this hctx either if there's no hctx preferred | |
1705 | * by IO scheduler or if it has requests that bypass the | |
1706 | * scheduler. | |
1707 | */ | |
1708 | if (!sq_hctx || sq_hctx == hctx || | |
1709 | !list_empty_careful(&hctx->dispatch)) | |
1710 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
b9151e7b DA |
1711 | } |
1712 | } | |
1713 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1714 | ||
fd001443 BVA |
1715 | /** |
1716 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1717 | * @q: request queue. | |
1718 | * | |
1719 | * The caller is responsible for serializing this function against | |
1720 | * blk_mq_{start,stop}_hw_queue(). | |
1721 | */ | |
1722 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1723 | { | |
1724 | struct blk_mq_hw_ctx *hctx; | |
1725 | int i; | |
1726 | ||
1727 | queue_for_each_hw_ctx(q, hctx, i) | |
1728 | if (blk_mq_hctx_stopped(hctx)) | |
1729 | return true; | |
1730 | ||
1731 | return false; | |
1732 | } | |
1733 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1734 | ||
39a70c76 ML |
1735 | /* |
1736 | * This function is often used for pausing .queue_rq() by driver when | |
1737 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1738 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1739 | * |
1740 | * We do not guarantee that dispatch can be drained or blocked | |
1741 | * after blk_mq_stop_hw_queue() returns. Please use | |
1742 | * blk_mq_quiesce_queue() for that requirement. | |
1743 | */ | |
2719aa21 JA |
1744 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1745 | { | |
641a9ed6 | 1746 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1747 | |
641a9ed6 | 1748 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1749 | } |
641a9ed6 | 1750 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1751 | |
39a70c76 ML |
1752 | /* |
1753 | * This function is often used for pausing .queue_rq() by driver when | |
1754 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1755 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1756 | * |
1757 | * We do not guarantee that dispatch can be drained or blocked | |
1758 | * after blk_mq_stop_hw_queues() returns. Please use | |
1759 | * blk_mq_quiesce_queue() for that requirement. | |
1760 | */ | |
2719aa21 JA |
1761 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1762 | { | |
641a9ed6 ML |
1763 | struct blk_mq_hw_ctx *hctx; |
1764 | int i; | |
1765 | ||
1766 | queue_for_each_hw_ctx(q, hctx, i) | |
1767 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1768 | } |
1769 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1770 | ||
320ae51f JA |
1771 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1772 | { | |
1773 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1774 | |
0ffbce80 | 1775 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1776 | } |
1777 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1778 | ||
2f268556 CH |
1779 | void blk_mq_start_hw_queues(struct request_queue *q) |
1780 | { | |
1781 | struct blk_mq_hw_ctx *hctx; | |
1782 | int i; | |
1783 | ||
1784 | queue_for_each_hw_ctx(q, hctx, i) | |
1785 | blk_mq_start_hw_queue(hctx); | |
1786 | } | |
1787 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1788 | ||
ae911c5e JA |
1789 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1790 | { | |
1791 | if (!blk_mq_hctx_stopped(hctx)) | |
1792 | return; | |
1793 | ||
1794 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1795 | blk_mq_run_hw_queue(hctx, async); | |
1796 | } | |
1797 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1798 | ||
1b4a3258 | 1799 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1800 | { |
1801 | struct blk_mq_hw_ctx *hctx; | |
1802 | int i; | |
1803 | ||
ae911c5e JA |
1804 | queue_for_each_hw_ctx(q, hctx, i) |
1805 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1806 | } |
1807 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1808 | ||
70f4db63 | 1809 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1810 | { |
1811 | struct blk_mq_hw_ctx *hctx; | |
1812 | ||
9f993737 | 1813 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1814 | |
21c6e939 | 1815 | /* |
15fe8a90 | 1816 | * If we are stopped, don't run the queue. |
21c6e939 | 1817 | */ |
0841031a | 1818 | if (blk_mq_hctx_stopped(hctx)) |
0196d6b4 | 1819 | return; |
7587a5ae BVA |
1820 | |
1821 | __blk_mq_run_hw_queue(hctx); | |
1822 | } | |
1823 | ||
cfd0c552 | 1824 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1825 | struct request *rq, |
1826 | bool at_head) | |
320ae51f | 1827 | { |
e57690fe | 1828 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
c16d6b5a | 1829 | enum hctx_type type = hctx->type; |
e57690fe | 1830 | |
7b607814 BVA |
1831 | lockdep_assert_held(&ctx->lock); |
1832 | ||
a54895fa | 1833 | trace_block_rq_insert(rq); |
01b983c9 | 1834 | |
72a0a36e | 1835 | if (at_head) |
c16d6b5a | 1836 | list_add(&rq->queuelist, &ctx->rq_lists[type]); |
72a0a36e | 1837 | else |
c16d6b5a | 1838 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); |
cfd0c552 | 1839 | } |
4bb659b1 | 1840 | |
2c3ad667 JA |
1841 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1842 | bool at_head) | |
cfd0c552 ML |
1843 | { |
1844 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1845 | ||
7b607814 BVA |
1846 | lockdep_assert_held(&ctx->lock); |
1847 | ||
e57690fe | 1848 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1849 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1850 | } |
1851 | ||
105663f7 AA |
1852 | /** |
1853 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1854 | * @rq: Pointer to request to be inserted. | |
26bfeb26 | 1855 | * @at_head: true if the request should be inserted at the head of the list. |
105663f7 AA |
1856 | * @run_queue: If we should run the hardware queue after inserting the request. |
1857 | * | |
157f377b JA |
1858 | * Should only be used carefully, when the caller knows we want to |
1859 | * bypass a potential IO scheduler on the target device. | |
1860 | */ | |
01e99aec ML |
1861 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, |
1862 | bool run_queue) | |
157f377b | 1863 | { |
ea4f995e | 1864 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; |
157f377b JA |
1865 | |
1866 | spin_lock(&hctx->lock); | |
01e99aec ML |
1867 | if (at_head) |
1868 | list_add(&rq->queuelist, &hctx->dispatch); | |
1869 | else | |
1870 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
157f377b JA |
1871 | spin_unlock(&hctx->lock); |
1872 | ||
b0850297 ML |
1873 | if (run_queue) |
1874 | blk_mq_run_hw_queue(hctx, false); | |
157f377b JA |
1875 | } |
1876 | ||
bd166ef1 JA |
1877 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1878 | struct list_head *list) | |
320ae51f JA |
1879 | |
1880 | { | |
3f0cedc7 | 1881 | struct request *rq; |
c16d6b5a | 1882 | enum hctx_type type = hctx->type; |
3f0cedc7 | 1883 | |
320ae51f JA |
1884 | /* |
1885 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1886 | * offline now | |
1887 | */ | |
3f0cedc7 | 1888 | list_for_each_entry(rq, list, queuelist) { |
e57690fe | 1889 | BUG_ON(rq->mq_ctx != ctx); |
a54895fa | 1890 | trace_block_rq_insert(rq); |
320ae51f | 1891 | } |
3f0cedc7 ML |
1892 | |
1893 | spin_lock(&ctx->lock); | |
c16d6b5a | 1894 | list_splice_tail_init(list, &ctx->rq_lists[type]); |
cfd0c552 | 1895 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1896 | spin_unlock(&ctx->lock); |
320ae51f JA |
1897 | } |
1898 | ||
4f0f586b ST |
1899 | static int plug_rq_cmp(void *priv, const struct list_head *a, |
1900 | const struct list_head *b) | |
320ae51f JA |
1901 | { |
1902 | struct request *rqa = container_of(a, struct request, queuelist); | |
1903 | struct request *rqb = container_of(b, struct request, queuelist); | |
1904 | ||
7d30a621 PB |
1905 | if (rqa->mq_ctx != rqb->mq_ctx) |
1906 | return rqa->mq_ctx > rqb->mq_ctx; | |
1907 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1908 | return rqa->mq_hctx > rqb->mq_hctx; | |
3110fc79 JA |
1909 | |
1910 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
320ae51f JA |
1911 | } |
1912 | ||
1913 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1914 | { | |
320ae51f | 1915 | LIST_HEAD(list); |
320ae51f | 1916 | |
95ed0c5b PB |
1917 | if (list_empty(&plug->mq_list)) |
1918 | return; | |
320ae51f JA |
1919 | list_splice_init(&plug->mq_list, &list); |
1920 | ||
ce5b009c JA |
1921 | if (plug->rq_count > 2 && plug->multiple_queues) |
1922 | list_sort(NULL, &list, plug_rq_cmp); | |
320ae51f | 1923 | |
bcc816df DZ |
1924 | plug->rq_count = 0; |
1925 | ||
95ed0c5b PB |
1926 | do { |
1927 | struct list_head rq_list; | |
1928 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1929 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1930 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1931 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1932 | unsigned int depth = 1; | |
1933 | ||
1934 | list_for_each_continue(pos, &list) { | |
1935 | rq = list_entry_rq(pos); | |
1936 | BUG_ON(!rq->q); | |
1937 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1938 | break; | |
1939 | depth++; | |
320ae51f JA |
1940 | } |
1941 | ||
95ed0c5b PB |
1942 | list_cut_before(&rq_list, &list, pos); |
1943 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
67cae4c9 | 1944 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, |
bd166ef1 | 1945 | from_schedule); |
95ed0c5b | 1946 | } while(!list_empty(&list)); |
320ae51f JA |
1947 | } |
1948 | ||
14ccb66b CH |
1949 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, |
1950 | unsigned int nr_segs) | |
320ae51f | 1951 | { |
93f221ae EB |
1952 | int err; |
1953 | ||
f924cdde CH |
1954 | if (bio->bi_opf & REQ_RAHEAD) |
1955 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1956 | ||
1957 | rq->__sector = bio->bi_iter.bi_sector; | |
1958 | rq->write_hint = bio->bi_write_hint; | |
14ccb66b | 1959 | blk_rq_bio_prep(rq, bio, nr_segs); |
93f221ae EB |
1960 | |
1961 | /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ | |
1962 | err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1963 | WARN_ON_ONCE(err); | |
4b570521 | 1964 | |
b5af37ab | 1965 | blk_account_io_start(rq); |
320ae51f JA |
1966 | } |
1967 | ||
0f95549c MS |
1968 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, |
1969 | struct request *rq, | |
be94f058 | 1970 | blk_qc_t *cookie, bool last) |
f984df1f | 1971 | { |
f984df1f | 1972 | struct request_queue *q = rq->q; |
f984df1f SL |
1973 | struct blk_mq_queue_data bd = { |
1974 | .rq = rq, | |
be94f058 | 1975 | .last = last, |
f984df1f | 1976 | }; |
bd166ef1 | 1977 | blk_qc_t new_cookie; |
f06345ad | 1978 | blk_status_t ret; |
0f95549c MS |
1979 | |
1980 | new_cookie = request_to_qc_t(hctx, rq); | |
1981 | ||
1982 | /* | |
1983 | * For OK queue, we are done. For error, caller may kill it. | |
1984 | * Any other error (busy), just add it to our list as we | |
1985 | * previously would have done. | |
1986 | */ | |
1987 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1988 | switch (ret) { | |
1989 | case BLK_STS_OK: | |
6ce3dd6e | 1990 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
1991 | *cookie = new_cookie; |
1992 | break; | |
1993 | case BLK_STS_RESOURCE: | |
86ff7c2a | 1994 | case BLK_STS_DEV_RESOURCE: |
6ce3dd6e | 1995 | blk_mq_update_dispatch_busy(hctx, true); |
0f95549c MS |
1996 | __blk_mq_requeue_request(rq); |
1997 | break; | |
1998 | default: | |
6ce3dd6e | 1999 | blk_mq_update_dispatch_busy(hctx, false); |
0f95549c MS |
2000 | *cookie = BLK_QC_T_NONE; |
2001 | break; | |
2002 | } | |
2003 | ||
2004 | return ret; | |
2005 | } | |
2006 | ||
fd9c40f6 | 2007 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
0f95549c | 2008 | struct request *rq, |
396eaf21 | 2009 | blk_qc_t *cookie, |
fd9c40f6 | 2010 | bool bypass_insert, bool last) |
0f95549c MS |
2011 | { |
2012 | struct request_queue *q = rq->q; | |
d964f04a | 2013 | bool run_queue = true; |
2a5a24aa | 2014 | int budget_token; |
d964f04a | 2015 | |
23d4ee19 | 2016 | /* |
fd9c40f6 | 2017 | * RCU or SRCU read lock is needed before checking quiesced flag. |
23d4ee19 | 2018 | * |
fd9c40f6 BVA |
2019 | * When queue is stopped or quiesced, ignore 'bypass_insert' from |
2020 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2021 | * and avoid driver to try to dispatch again. | |
23d4ee19 | 2022 | */ |
fd9c40f6 | 2023 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { |
d964f04a | 2024 | run_queue = false; |
fd9c40f6 BVA |
2025 | bypass_insert = false; |
2026 | goto insert; | |
d964f04a | 2027 | } |
f984df1f | 2028 | |
fd9c40f6 BVA |
2029 | if (q->elevator && !bypass_insert) |
2030 | goto insert; | |
2253efc8 | 2031 | |
2a5a24aa ML |
2032 | budget_token = blk_mq_get_dispatch_budget(q); |
2033 | if (budget_token < 0) | |
fd9c40f6 | 2034 | goto insert; |
bd166ef1 | 2035 | |
2a5a24aa ML |
2036 | blk_mq_set_rq_budget_token(rq, budget_token); |
2037 | ||
8ab6bb9e | 2038 | if (!blk_mq_get_driver_tag(rq)) { |
2a5a24aa | 2039 | blk_mq_put_dispatch_budget(q, budget_token); |
fd9c40f6 | 2040 | goto insert; |
88022d72 | 2041 | } |
de148297 | 2042 | |
fd9c40f6 BVA |
2043 | return __blk_mq_issue_directly(hctx, rq, cookie, last); |
2044 | insert: | |
2045 | if (bypass_insert) | |
2046 | return BLK_STS_RESOURCE; | |
2047 | ||
db03f88f ML |
2048 | blk_mq_sched_insert_request(rq, false, run_queue, false); |
2049 | ||
fd9c40f6 BVA |
2050 | return BLK_STS_OK; |
2051 | } | |
2052 | ||
105663f7 AA |
2053 | /** |
2054 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2055 | * @hctx: Pointer of the associated hardware queue. | |
2056 | * @rq: Pointer to request to be sent. | |
2057 | * @cookie: Request queue cookie. | |
2058 | * | |
2059 | * If the device has enough resources to accept a new request now, send the | |
2060 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2061 | * we can try send it another time in the future. Requests inserted at this | |
2062 | * queue have higher priority. | |
2063 | */ | |
fd9c40f6 BVA |
2064 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
2065 | struct request *rq, blk_qc_t *cookie) | |
2066 | { | |
2067 | blk_status_t ret; | |
2068 | int srcu_idx; | |
2069 | ||
2070 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2071 | ||
2072 | hctx_lock(hctx, &srcu_idx); | |
2073 | ||
2074 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2075 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
01e99aec | 2076 | blk_mq_request_bypass_insert(rq, false, true); |
fd9c40f6 BVA |
2077 | else if (ret != BLK_STS_OK) |
2078 | blk_mq_end_request(rq, ret); | |
2079 | ||
2080 | hctx_unlock(hctx, srcu_idx); | |
2081 | } | |
2082 | ||
2083 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2084 | { | |
2085 | blk_status_t ret; | |
2086 | int srcu_idx; | |
2087 | blk_qc_t unused_cookie; | |
2088 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2089 | ||
2090 | hctx_lock(hctx, &srcu_idx); | |
2091 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
04ced159 | 2092 | hctx_unlock(hctx, srcu_idx); |
7f556a44 JW |
2093 | |
2094 | return ret; | |
5eb6126e CH |
2095 | } |
2096 | ||
6ce3dd6e ML |
2097 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, |
2098 | struct list_head *list) | |
2099 | { | |
536167d4 | 2100 | int queued = 0; |
632bfb63 | 2101 | int errors = 0; |
536167d4 | 2102 | |
6ce3dd6e | 2103 | while (!list_empty(list)) { |
fd9c40f6 | 2104 | blk_status_t ret; |
6ce3dd6e ML |
2105 | struct request *rq = list_first_entry(list, struct request, |
2106 | queuelist); | |
2107 | ||
2108 | list_del_init(&rq->queuelist); | |
fd9c40f6 BVA |
2109 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); |
2110 | if (ret != BLK_STS_OK) { | |
2111 | if (ret == BLK_STS_RESOURCE || | |
2112 | ret == BLK_STS_DEV_RESOURCE) { | |
01e99aec | 2113 | blk_mq_request_bypass_insert(rq, false, |
c616cbee | 2114 | list_empty(list)); |
fd9c40f6 BVA |
2115 | break; |
2116 | } | |
2117 | blk_mq_end_request(rq, ret); | |
632bfb63 | 2118 | errors++; |
536167d4 KB |
2119 | } else |
2120 | queued++; | |
6ce3dd6e | 2121 | } |
d666ba98 JA |
2122 | |
2123 | /* | |
2124 | * If we didn't flush the entire list, we could have told | |
2125 | * the driver there was more coming, but that turned out to | |
2126 | * be a lie. | |
2127 | */ | |
632bfb63 | 2128 | if ((!list_empty(list) || errors) && |
2129 | hctx->queue->mq_ops->commit_rqs && queued) | |
d666ba98 | 2130 | hctx->queue->mq_ops->commit_rqs(hctx); |
6ce3dd6e ML |
2131 | } |
2132 | ||
ce5b009c JA |
2133 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) |
2134 | { | |
2135 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2136 | plug->rq_count++; | |
2137 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2138 | struct request *tmp; | |
2139 | ||
2140 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2141 | queuelist); | |
2142 | if (tmp->q != rq->q) | |
2143 | plug->multiple_queues = true; | |
2144 | } | |
2145 | } | |
2146 | ||
7f2a6a69 SL |
2147 | /* |
2148 | * Allow 4x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple | |
2149 | * queues. This is important for md arrays to benefit from merging | |
2150 | * requests. | |
2151 | */ | |
2152 | static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) | |
2153 | { | |
2154 | if (plug->multiple_queues) | |
2155 | return BLK_MAX_REQUEST_COUNT * 4; | |
2156 | return BLK_MAX_REQUEST_COUNT; | |
2157 | } | |
2158 | ||
105663f7 | 2159 | /** |
c62b37d9 | 2160 | * blk_mq_submit_bio - Create and send a request to block device. |
105663f7 AA |
2161 | * @bio: Bio pointer. |
2162 | * | |
2163 | * Builds up a request structure from @q and @bio and send to the device. The | |
2164 | * request may not be queued directly to hardware if: | |
2165 | * * This request can be merged with another one | |
2166 | * * We want to place request at plug queue for possible future merging | |
2167 | * * There is an IO scheduler active at this queue | |
2168 | * | |
2169 | * It will not queue the request if there is an error with the bio, or at the | |
2170 | * request creation. | |
2171 | * | |
2172 | * Returns: Request queue cookie. | |
2173 | */ | |
c62b37d9 | 2174 | blk_qc_t blk_mq_submit_bio(struct bio *bio) |
07068d5b | 2175 | { |
309dca30 | 2176 | struct request_queue *q = bio->bi_bdev->bd_disk->queue; |
ef295ecf | 2177 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 2178 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6e7abff CH |
2179 | struct blk_mq_alloc_data data = { |
2180 | .q = q, | |
2181 | }; | |
07068d5b | 2182 | struct request *rq; |
f984df1f | 2183 | struct blk_plug *plug; |
5b3f341f | 2184 | struct request *same_queue_rq = NULL; |
14ccb66b | 2185 | unsigned int nr_segs; |
7b371636 | 2186 | blk_qc_t cookie; |
a892c8d5 | 2187 | blk_status_t ret; |
cc29e1bf | 2188 | bool hipri; |
07068d5b JA |
2189 | |
2190 | blk_queue_bounce(q, &bio); | |
f695ca38 | 2191 | __blk_queue_split(&bio, &nr_segs); |
f36ea50c | 2192 | |
e23947bd | 2193 | if (!bio_integrity_prep(bio)) |
ac7c5675 | 2194 | goto queue_exit; |
07068d5b | 2195 | |
87c279e6 | 2196 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
14ccb66b | 2197 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) |
ac7c5675 | 2198 | goto queue_exit; |
f984df1f | 2199 | |
14ccb66b | 2200 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) |
ac7c5675 | 2201 | goto queue_exit; |
bd166ef1 | 2202 | |
d5337560 | 2203 | rq_qos_throttle(q, bio); |
87760e5e | 2204 | |
cc29e1bf JX |
2205 | hipri = bio->bi_opf & REQ_HIPRI; |
2206 | ||
7809167d | 2207 | data.cmd_flags = bio->bi_opf; |
e6e7abff | 2208 | rq = __blk_mq_alloc_request(&data); |
87760e5e | 2209 | if (unlikely(!rq)) { |
c1c80384 | 2210 | rq_qos_cleanup(q, bio); |
7b6620d7 | 2211 | if (bio->bi_opf & REQ_NOWAIT) |
03a07c92 | 2212 | bio_wouldblock_error(bio); |
ac7c5675 | 2213 | goto queue_exit; |
87760e5e JA |
2214 | } |
2215 | ||
e8a676d6 | 2216 | trace_block_getrq(bio); |
d6f1dda2 | 2217 | |
c1c80384 | 2218 | rq_qos_track(q, rq, bio); |
07068d5b | 2219 | |
fd2d3326 | 2220 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 2221 | |
970d168d BVA |
2222 | blk_mq_bio_to_request(rq, bio, nr_segs); |
2223 | ||
a892c8d5 ST |
2224 | ret = blk_crypto_init_request(rq); |
2225 | if (ret != BLK_STS_OK) { | |
2226 | bio->bi_status = ret; | |
2227 | bio_endio(bio); | |
2228 | blk_mq_free_request(rq); | |
2229 | return BLK_QC_T_NONE; | |
2230 | } | |
2231 | ||
b49773e7 | 2232 | plug = blk_mq_plug(q, bio); |
07068d5b | 2233 | if (unlikely(is_flush_fua)) { |
105663f7 | 2234 | /* Bypass scheduler for flush requests */ |
923218f6 ML |
2235 | blk_insert_flush(rq); |
2236 | blk_mq_run_hw_queue(data.hctx, true); | |
03f26d8f ML |
2237 | } else if (plug && (q->nr_hw_queues == 1 || |
2238 | blk_mq_is_sbitmap_shared(rq->mq_hctx->flags) || | |
2239 | q->mq_ops->commit_rqs || !blk_queue_nonrot(q))) { | |
b2c5d16b JA |
2240 | /* |
2241 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2242 | * we know the driver uses bd->last in a smart fashion. | |
3154df26 ML |
2243 | * |
2244 | * Use normal plugging if this disk is slow HDD, as sequential | |
2245 | * IO may benefit a lot from plug merging. | |
b2c5d16b | 2246 | */ |
5f0ed774 | 2247 | unsigned int request_count = plug->rq_count; |
600271d9 SL |
2248 | struct request *last = NULL; |
2249 | ||
676d0607 | 2250 | if (!request_count) |
e6c4438b | 2251 | trace_block_plug(q); |
600271d9 SL |
2252 | else |
2253 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 2254 | |
7f2a6a69 | 2255 | if (request_count >= blk_plug_max_rq_count(plug) || (last && |
600271d9 | 2256 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { |
e6c4438b JM |
2257 | blk_flush_plug_list(plug, false); |
2258 | trace_block_plug(q); | |
320ae51f | 2259 | } |
b094f89c | 2260 | |
ce5b009c | 2261 | blk_add_rq_to_plug(plug, rq); |
a12de1d4 | 2262 | } else if (q->elevator) { |
105663f7 | 2263 | /* Insert the request at the IO scheduler queue */ |
a12de1d4 | 2264 | blk_mq_sched_insert_request(rq, false, true, true); |
2299722c | 2265 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 2266 | /* |
6a83e74d | 2267 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
2268 | * Otherwise the existing request in the plug list will be |
2269 | * issued. So the plug list will have one request at most | |
2299722c CH |
2270 | * The plug list might get flushed before this. If that happens, |
2271 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 2272 | */ |
2299722c CH |
2273 | if (list_empty(&plug->mq_list)) |
2274 | same_queue_rq = NULL; | |
4711b573 | 2275 | if (same_queue_rq) { |
2299722c | 2276 | list_del_init(&same_queue_rq->queuelist); |
4711b573 JA |
2277 | plug->rq_count--; |
2278 | } | |
ce5b009c | 2279 | blk_add_rq_to_plug(plug, rq); |
ff3b74b8 | 2280 | trace_block_plug(q); |
2299722c | 2281 | |
dad7a3be | 2282 | if (same_queue_rq) { |
ea4f995e | 2283 | data.hctx = same_queue_rq->mq_hctx; |
ff3b74b8 | 2284 | trace_block_unplug(q, 1, true); |
2299722c | 2285 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
fd9c40f6 | 2286 | &cookie); |
dad7a3be | 2287 | } |
a12de1d4 ML |
2288 | } else if ((q->nr_hw_queues > 1 && is_sync) || |
2289 | !data.hctx->dispatch_busy) { | |
105663f7 AA |
2290 | /* |
2291 | * There is no scheduler and we can try to send directly | |
2292 | * to the hardware. | |
2293 | */ | |
fd9c40f6 | 2294 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
ab42f35d | 2295 | } else { |
105663f7 | 2296 | /* Default case. */ |
8fa9f556 | 2297 | blk_mq_sched_insert_request(rq, false, true, true); |
ab42f35d | 2298 | } |
320ae51f | 2299 | |
cc29e1bf JX |
2300 | if (!hipri) |
2301 | return BLK_QC_T_NONE; | |
7b371636 | 2302 | return cookie; |
ac7c5675 CH |
2303 | queue_exit: |
2304 | blk_queue_exit(q); | |
2305 | return BLK_QC_T_NONE; | |
320ae51f JA |
2306 | } |
2307 | ||
bd63141d ML |
2308 | static size_t order_to_size(unsigned int order) |
2309 | { | |
2310 | return (size_t)PAGE_SIZE << order; | |
2311 | } | |
2312 | ||
2313 | /* called before freeing request pool in @tags */ | |
2314 | static void blk_mq_clear_rq_mapping(struct blk_mq_tag_set *set, | |
2315 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
2316 | { | |
2317 | struct blk_mq_tags *drv_tags = set->tags[hctx_idx]; | |
2318 | struct page *page; | |
2319 | unsigned long flags; | |
2320 | ||
2321 | list_for_each_entry(page, &tags->page_list, lru) { | |
2322 | unsigned long start = (unsigned long)page_address(page); | |
2323 | unsigned long end = start + order_to_size(page->private); | |
2324 | int i; | |
2325 | ||
2326 | for (i = 0; i < set->queue_depth; i++) { | |
2327 | struct request *rq = drv_tags->rqs[i]; | |
2328 | unsigned long rq_addr = (unsigned long)rq; | |
2329 | ||
2330 | if (rq_addr >= start && rq_addr < end) { | |
2331 | WARN_ON_ONCE(refcount_read(&rq->ref) != 0); | |
2332 | cmpxchg(&drv_tags->rqs[i], rq, NULL); | |
2333 | } | |
2334 | } | |
2335 | } | |
2336 | ||
2337 | /* | |
2338 | * Wait until all pending iteration is done. | |
2339 | * | |
2340 | * Request reference is cleared and it is guaranteed to be observed | |
2341 | * after the ->lock is released. | |
2342 | */ | |
2343 | spin_lock_irqsave(&drv_tags->lock, flags); | |
2344 | spin_unlock_irqrestore(&drv_tags->lock, flags); | |
2345 | } | |
2346 | ||
cc71a6f4 JA |
2347 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2348 | unsigned int hctx_idx) | |
95363efd | 2349 | { |
e9b267d9 | 2350 | struct page *page; |
320ae51f | 2351 | |
24d2f903 | 2352 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 2353 | int i; |
320ae51f | 2354 | |
24d2f903 | 2355 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
2356 | struct request *rq = tags->static_rqs[i]; |
2357 | ||
2358 | if (!rq) | |
e9b267d9 | 2359 | continue; |
d6296d39 | 2360 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 2361 | tags->static_rqs[i] = NULL; |
e9b267d9 | 2362 | } |
320ae51f | 2363 | } |
320ae51f | 2364 | |
bd63141d ML |
2365 | blk_mq_clear_rq_mapping(set, tags, hctx_idx); |
2366 | ||
24d2f903 CH |
2367 | while (!list_empty(&tags->page_list)) { |
2368 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 2369 | list_del_init(&page->lru); |
f75782e4 CM |
2370 | /* |
2371 | * Remove kmemleak object previously allocated in | |
273938bf | 2372 | * blk_mq_alloc_rqs(). |
f75782e4 CM |
2373 | */ |
2374 | kmemleak_free(page_address(page)); | |
320ae51f JA |
2375 | __free_pages(page, page->private); |
2376 | } | |
cc71a6f4 | 2377 | } |
320ae51f | 2378 | |
1c0706a7 | 2379 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) |
cc71a6f4 | 2380 | { |
24d2f903 | 2381 | kfree(tags->rqs); |
cc71a6f4 | 2382 | tags->rqs = NULL; |
2af8cbe3 JA |
2383 | kfree(tags->static_rqs); |
2384 | tags->static_rqs = NULL; | |
320ae51f | 2385 | |
1c0706a7 | 2386 | blk_mq_free_tags(tags, flags); |
320ae51f JA |
2387 | } |
2388 | ||
cc71a6f4 JA |
2389 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
2390 | unsigned int hctx_idx, | |
2391 | unsigned int nr_tags, | |
1c0706a7 JG |
2392 | unsigned int reserved_tags, |
2393 | unsigned int flags) | |
320ae51f | 2394 | { |
24d2f903 | 2395 | struct blk_mq_tags *tags; |
59f082e4 | 2396 | int node; |
320ae51f | 2397 | |
7d76f856 | 2398 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2399 | if (node == NUMA_NO_NODE) |
2400 | node = set->numa_node; | |
2401 | ||
1c0706a7 | 2402 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); |
24d2f903 CH |
2403 | if (!tags) |
2404 | return NULL; | |
320ae51f | 2405 | |
590b5b7d | 2406 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
36e1f3d1 | 2407 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 2408 | node); |
24d2f903 | 2409 | if (!tags->rqs) { |
1c0706a7 | 2410 | blk_mq_free_tags(tags, flags); |
24d2f903 CH |
2411 | return NULL; |
2412 | } | |
320ae51f | 2413 | |
590b5b7d KC |
2414 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), |
2415 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2416 | node); | |
2af8cbe3 JA |
2417 | if (!tags->static_rqs) { |
2418 | kfree(tags->rqs); | |
1c0706a7 | 2419 | blk_mq_free_tags(tags, flags); |
2af8cbe3 JA |
2420 | return NULL; |
2421 | } | |
2422 | ||
cc71a6f4 JA |
2423 | return tags; |
2424 | } | |
2425 | ||
1d9bd516 TH |
2426 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, |
2427 | unsigned int hctx_idx, int node) | |
2428 | { | |
2429 | int ret; | |
2430 | ||
2431 | if (set->ops->init_request) { | |
2432 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2433 | if (ret) | |
2434 | return ret; | |
2435 | } | |
2436 | ||
12f5b931 | 2437 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); |
1d9bd516 TH |
2438 | return 0; |
2439 | } | |
2440 | ||
cc71a6f4 JA |
2441 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
2442 | unsigned int hctx_idx, unsigned int depth) | |
2443 | { | |
2444 | unsigned int i, j, entries_per_page, max_order = 4; | |
2445 | size_t rq_size, left; | |
59f082e4 SL |
2446 | int node; |
2447 | ||
7d76f856 | 2448 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); |
59f082e4 SL |
2449 | if (node == NUMA_NO_NODE) |
2450 | node = set->numa_node; | |
cc71a6f4 JA |
2451 | |
2452 | INIT_LIST_HEAD(&tags->page_list); | |
2453 | ||
320ae51f JA |
2454 | /* |
2455 | * rq_size is the size of the request plus driver payload, rounded | |
2456 | * to the cacheline size | |
2457 | */ | |
24d2f903 | 2458 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 2459 | cache_line_size()); |
cc71a6f4 | 2460 | left = rq_size * depth; |
320ae51f | 2461 | |
cc71a6f4 | 2462 | for (i = 0; i < depth; ) { |
320ae51f JA |
2463 | int this_order = max_order; |
2464 | struct page *page; | |
2465 | int to_do; | |
2466 | void *p; | |
2467 | ||
b3a834b1 | 2468 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
2469 | this_order--; |
2470 | ||
2471 | do { | |
59f082e4 | 2472 | page = alloc_pages_node(node, |
36e1f3d1 | 2473 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 2474 | this_order); |
320ae51f JA |
2475 | if (page) |
2476 | break; | |
2477 | if (!this_order--) | |
2478 | break; | |
2479 | if (order_to_size(this_order) < rq_size) | |
2480 | break; | |
2481 | } while (1); | |
2482 | ||
2483 | if (!page) | |
24d2f903 | 2484 | goto fail; |
320ae51f JA |
2485 | |
2486 | page->private = this_order; | |
24d2f903 | 2487 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
2488 | |
2489 | p = page_address(page); | |
f75782e4 CM |
2490 | /* |
2491 | * Allow kmemleak to scan these pages as they contain pointers | |
2492 | * to additional allocations like via ops->init_request(). | |
2493 | */ | |
36e1f3d1 | 2494 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 2495 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 2496 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
2497 | left -= to_do * rq_size; |
2498 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
2499 | struct request *rq = p; |
2500 | ||
2501 | tags->static_rqs[i] = rq; | |
1d9bd516 TH |
2502 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { |
2503 | tags->static_rqs[i] = NULL; | |
2504 | goto fail; | |
e9b267d9 CH |
2505 | } |
2506 | ||
320ae51f JA |
2507 | p += rq_size; |
2508 | i++; | |
2509 | } | |
2510 | } | |
cc71a6f4 | 2511 | return 0; |
320ae51f | 2512 | |
24d2f903 | 2513 | fail: |
cc71a6f4 JA |
2514 | blk_mq_free_rqs(set, tags, hctx_idx); |
2515 | return -ENOMEM; | |
320ae51f JA |
2516 | } |
2517 | ||
bf0beec0 ML |
2518 | struct rq_iter_data { |
2519 | struct blk_mq_hw_ctx *hctx; | |
2520 | bool has_rq; | |
2521 | }; | |
2522 | ||
2523 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2524 | { | |
2525 | struct rq_iter_data *iter_data = data; | |
2526 | ||
2527 | if (rq->mq_hctx != iter_data->hctx) | |
2528 | return true; | |
2529 | iter_data->has_rq = true; | |
2530 | return false; | |
2531 | } | |
2532 | ||
2533 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2534 | { | |
2535 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2536 | hctx->sched_tags : hctx->tags; | |
2537 | struct rq_iter_data data = { | |
2538 | .hctx = hctx, | |
2539 | }; | |
2540 | ||
2541 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2542 | return data.has_rq; | |
2543 | } | |
2544 | ||
2545 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2546 | struct blk_mq_hw_ctx *hctx) | |
2547 | { | |
2548 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2549 | return false; | |
2550 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2551 | return false; | |
2552 | return true; | |
2553 | } | |
2554 | ||
2555 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2556 | { | |
2557 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2558 | struct blk_mq_hw_ctx, cpuhp_online); | |
2559 | ||
2560 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2561 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2562 | return 0; | |
2563 | ||
2564 | /* | |
2565 | * Prevent new request from being allocated on the current hctx. | |
2566 | * | |
2567 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2568 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2569 | * seen once we return from the tag allocator. | |
2570 | */ | |
2571 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2572 | smp_mb__after_atomic(); | |
2573 | ||
2574 | /* | |
2575 | * Try to grab a reference to the queue and wait for any outstanding | |
2576 | * requests. If we could not grab a reference the queue has been | |
2577 | * frozen and there are no requests. | |
2578 | */ | |
2579 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2580 | while (blk_mq_hctx_has_requests(hctx)) | |
2581 | msleep(5); | |
2582 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2583 | } | |
2584 | ||
2585 | return 0; | |
2586 | } | |
2587 | ||
2588 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2589 | { | |
2590 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2591 | struct blk_mq_hw_ctx, cpuhp_online); | |
2592 | ||
2593 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2594 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2595 | return 0; | |
2596 | } | |
2597 | ||
e57690fe JA |
2598 | /* |
2599 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2600 | * software queue to the hw queue dispatch list, and ensure that it | |
2601 | * gets run. | |
2602 | */ | |
9467f859 | 2603 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 2604 | { |
9467f859 | 2605 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
2606 | struct blk_mq_ctx *ctx; |
2607 | LIST_HEAD(tmp); | |
c16d6b5a | 2608 | enum hctx_type type; |
484b4061 | 2609 | |
9467f859 | 2610 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
bf0beec0 ML |
2611 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) |
2612 | return 0; | |
2613 | ||
e57690fe | 2614 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
c16d6b5a | 2615 | type = hctx->type; |
484b4061 JA |
2616 | |
2617 | spin_lock(&ctx->lock); | |
c16d6b5a ML |
2618 | if (!list_empty(&ctx->rq_lists[type])) { |
2619 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
484b4061 JA |
2620 | blk_mq_hctx_clear_pending(hctx, ctx); |
2621 | } | |
2622 | spin_unlock(&ctx->lock); | |
2623 | ||
2624 | if (list_empty(&tmp)) | |
9467f859 | 2625 | return 0; |
484b4061 | 2626 | |
e57690fe JA |
2627 | spin_lock(&hctx->lock); |
2628 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2629 | spin_unlock(&hctx->lock); | |
484b4061 JA |
2630 | |
2631 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 2632 | return 0; |
484b4061 JA |
2633 | } |
2634 | ||
9467f859 | 2635 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 2636 | { |
bf0beec0 ML |
2637 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2638 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2639 | &hctx->cpuhp_online); | |
9467f859 TG |
2640 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
2641 | &hctx->cpuhp_dead); | |
484b4061 JA |
2642 | } |
2643 | ||
364b6181 ML |
2644 | /* |
2645 | * Before freeing hw queue, clearing the flush request reference in | |
2646 | * tags->rqs[] for avoiding potential UAF. | |
2647 | */ | |
2648 | static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, | |
2649 | unsigned int queue_depth, struct request *flush_rq) | |
2650 | { | |
2651 | int i; | |
2652 | unsigned long flags; | |
2653 | ||
2654 | /* The hw queue may not be mapped yet */ | |
2655 | if (!tags) | |
2656 | return; | |
2657 | ||
2658 | WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0); | |
2659 | ||
2660 | for (i = 0; i < queue_depth; i++) | |
2661 | cmpxchg(&tags->rqs[i], flush_rq, NULL); | |
2662 | ||
2663 | /* | |
2664 | * Wait until all pending iteration is done. | |
2665 | * | |
2666 | * Request reference is cleared and it is guaranteed to be observed | |
2667 | * after the ->lock is released. | |
2668 | */ | |
2669 | spin_lock_irqsave(&tags->lock, flags); | |
2670 | spin_unlock_irqrestore(&tags->lock, flags); | |
2671 | } | |
2672 | ||
c3b4afca | 2673 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
2674 | static void blk_mq_exit_hctx(struct request_queue *q, |
2675 | struct blk_mq_tag_set *set, | |
2676 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2677 | { | |
364b6181 ML |
2678 | struct request *flush_rq = hctx->fq->flush_rq; |
2679 | ||
8ab0b7dc ML |
2680 | if (blk_mq_hw_queue_mapped(hctx)) |
2681 | blk_mq_tag_idle(hctx); | |
08e98fc6 | 2682 | |
364b6181 ML |
2683 | blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], |
2684 | set->queue_depth, flush_rq); | |
f70ced09 | 2685 | if (set->ops->exit_request) |
364b6181 | 2686 | set->ops->exit_request(set, flush_rq, hctx_idx); |
f70ced09 | 2687 | |
08e98fc6 ML |
2688 | if (set->ops->exit_hctx) |
2689 | set->ops->exit_hctx(hctx, hctx_idx); | |
2690 | ||
9467f859 | 2691 | blk_mq_remove_cpuhp(hctx); |
2f8f1336 ML |
2692 | |
2693 | spin_lock(&q->unused_hctx_lock); | |
2694 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2695 | spin_unlock(&q->unused_hctx_lock); | |
08e98fc6 ML |
2696 | } |
2697 | ||
624dbe47 ML |
2698 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2699 | struct blk_mq_tag_set *set, int nr_queue) | |
2700 | { | |
2701 | struct blk_mq_hw_ctx *hctx; | |
2702 | unsigned int i; | |
2703 | ||
2704 | queue_for_each_hw_ctx(q, hctx, i) { | |
2705 | if (i == nr_queue) | |
2706 | break; | |
477e19de | 2707 | blk_mq_debugfs_unregister_hctx(hctx); |
08e98fc6 | 2708 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2709 | } |
624dbe47 ML |
2710 | } |
2711 | ||
7c6c5b7c ML |
2712 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2713 | { | |
2714 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2715 | ||
2716 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2717 | __alignof__(struct blk_mq_hw_ctx)) != | |
2718 | sizeof(struct blk_mq_hw_ctx)); | |
2719 | ||
2720 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2721 | hw_ctx_size += sizeof(struct srcu_struct); | |
2722 | ||
2723 | return hw_ctx_size; | |
2724 | } | |
2725 | ||
08e98fc6 ML |
2726 | static int blk_mq_init_hctx(struct request_queue *q, |
2727 | struct blk_mq_tag_set *set, | |
2728 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2729 | { |
7c6c5b7c ML |
2730 | hctx->queue_num = hctx_idx; |
2731 | ||
bf0beec0 ML |
2732 | if (!(hctx->flags & BLK_MQ_F_STACKING)) |
2733 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2734 | &hctx->cpuhp_online); | |
7c6c5b7c ML |
2735 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
2736 | ||
2737 | hctx->tags = set->tags[hctx_idx]; | |
2738 | ||
2739 | if (set->ops->init_hctx && | |
2740 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2741 | goto unregister_cpu_notifier; | |
08e98fc6 | 2742 | |
7c6c5b7c ML |
2743 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, |
2744 | hctx->numa_node)) | |
2745 | goto exit_hctx; | |
2746 | return 0; | |
2747 | ||
2748 | exit_hctx: | |
2749 | if (set->ops->exit_hctx) | |
2750 | set->ops->exit_hctx(hctx, hctx_idx); | |
2751 | unregister_cpu_notifier: | |
2752 | blk_mq_remove_cpuhp(hctx); | |
2753 | return -1; | |
2754 | } | |
2755 | ||
2756 | static struct blk_mq_hw_ctx * | |
2757 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2758 | int node) | |
2759 | { | |
2760 | struct blk_mq_hw_ctx *hctx; | |
2761 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2762 | ||
2763 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2764 | if (!hctx) | |
2765 | goto fail_alloc_hctx; | |
2766 | ||
2767 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2768 | goto free_hctx; | |
2769 | ||
2770 | atomic_set(&hctx->nr_active, 0); | |
08e98fc6 | 2771 | if (node == NUMA_NO_NODE) |
7c6c5b7c ML |
2772 | node = set->numa_node; |
2773 | hctx->numa_node = node; | |
08e98fc6 | 2774 | |
9f993737 | 2775 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2776 | spin_lock_init(&hctx->lock); |
2777 | INIT_LIST_HEAD(&hctx->dispatch); | |
2778 | hctx->queue = q; | |
51db1c37 | 2779 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; |
08e98fc6 | 2780 | |
2f8f1336 ML |
2781 | INIT_LIST_HEAD(&hctx->hctx_list); |
2782 | ||
320ae51f | 2783 | /* |
08e98fc6 ML |
2784 | * Allocate space for all possible cpus to avoid allocation at |
2785 | * runtime | |
320ae51f | 2786 | */ |
d904bfa7 | 2787 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), |
7c6c5b7c | 2788 | gfp, node); |
08e98fc6 | 2789 | if (!hctx->ctxs) |
7c6c5b7c | 2790 | goto free_cpumask; |
320ae51f | 2791 | |
5b202853 | 2792 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), |
c548e62b | 2793 | gfp, node, false, false)) |
08e98fc6 | 2794 | goto free_ctxs; |
08e98fc6 | 2795 | hctx->nr_ctx = 0; |
320ae51f | 2796 | |
5815839b | 2797 | spin_lock_init(&hctx->dispatch_wait_lock); |
eb619fdb JA |
2798 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); |
2799 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2800 | ||
754a1572 | 2801 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); |
f70ced09 | 2802 | if (!hctx->fq) |
7c6c5b7c | 2803 | goto free_bitmap; |
320ae51f | 2804 | |
6a83e74d | 2805 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
05707b64 | 2806 | init_srcu_struct(hctx->srcu); |
7c6c5b7c | 2807 | blk_mq_hctx_kobj_init(hctx); |
6a83e74d | 2808 | |
7c6c5b7c | 2809 | return hctx; |
320ae51f | 2810 | |
08e98fc6 | 2811 | free_bitmap: |
88459642 | 2812 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2813 | free_ctxs: |
2814 | kfree(hctx->ctxs); | |
7c6c5b7c ML |
2815 | free_cpumask: |
2816 | free_cpumask_var(hctx->cpumask); | |
2817 | free_hctx: | |
2818 | kfree(hctx); | |
2819 | fail_alloc_hctx: | |
2820 | return NULL; | |
08e98fc6 | 2821 | } |
320ae51f | 2822 | |
320ae51f JA |
2823 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2824 | unsigned int nr_hw_queues) | |
2825 | { | |
b3c661b1 JA |
2826 | struct blk_mq_tag_set *set = q->tag_set; |
2827 | unsigned int i, j; | |
320ae51f JA |
2828 | |
2829 | for_each_possible_cpu(i) { | |
2830 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2831 | struct blk_mq_hw_ctx *hctx; | |
c16d6b5a | 2832 | int k; |
320ae51f | 2833 | |
320ae51f JA |
2834 | __ctx->cpu = i; |
2835 | spin_lock_init(&__ctx->lock); | |
c16d6b5a ML |
2836 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) |
2837 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2838 | ||
320ae51f JA |
2839 | __ctx->queue = q; |
2840 | ||
320ae51f JA |
2841 | /* |
2842 | * Set local node, IFF we have more than one hw queue. If | |
2843 | * not, we remain on the home node of the device | |
2844 | */ | |
b3c661b1 JA |
2845 | for (j = 0; j < set->nr_maps; j++) { |
2846 | hctx = blk_mq_map_queue_type(q, j, i); | |
2847 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
576e85c5 | 2848 | hctx->numa_node = cpu_to_node(i); |
b3c661b1 | 2849 | } |
320ae51f JA |
2850 | } |
2851 | } | |
2852 | ||
03b63b02 WZ |
2853 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, |
2854 | int hctx_idx) | |
cc71a6f4 | 2855 | { |
1c0706a7 | 2856 | unsigned int flags = set->flags; |
cc71a6f4 JA |
2857 | int ret = 0; |
2858 | ||
2859 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1c0706a7 | 2860 | set->queue_depth, set->reserved_tags, flags); |
cc71a6f4 JA |
2861 | if (!set->tags[hctx_idx]) |
2862 | return false; | |
2863 | ||
2864 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2865 | set->queue_depth); | |
2866 | if (!ret) | |
2867 | return true; | |
2868 | ||
1c0706a7 | 2869 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
cc71a6f4 JA |
2870 | set->tags[hctx_idx] = NULL; |
2871 | return false; | |
2872 | } | |
2873 | ||
2874 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2875 | unsigned int hctx_idx) | |
2876 | { | |
1c0706a7 JG |
2877 | unsigned int flags = set->flags; |
2878 | ||
4e6db0f2 | 2879 | if (set->tags && set->tags[hctx_idx]) { |
bd166ef1 | 2880 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); |
1c0706a7 | 2881 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); |
bd166ef1 JA |
2882 | set->tags[hctx_idx] = NULL; |
2883 | } | |
cc71a6f4 JA |
2884 | } |
2885 | ||
4b855ad3 | 2886 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2887 | { |
b3c661b1 | 2888 | unsigned int i, j, hctx_idx; |
320ae51f JA |
2889 | struct blk_mq_hw_ctx *hctx; |
2890 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2891 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f JA |
2892 | |
2893 | queue_for_each_hw_ctx(q, hctx, i) { | |
e4043dcf | 2894 | cpumask_clear(hctx->cpumask); |
320ae51f | 2895 | hctx->nr_ctx = 0; |
d416c92c | 2896 | hctx->dispatch_from = NULL; |
320ae51f JA |
2897 | } |
2898 | ||
2899 | /* | |
4b855ad3 | 2900 | * Map software to hardware queues. |
4412efec ML |
2901 | * |
2902 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2903 | */ |
20e4d813 | 2904 | for_each_possible_cpu(i) { |
4412efec | 2905 | |
897bb0c7 | 2906 | ctx = per_cpu_ptr(q->queue_ctx, i); |
b3c661b1 | 2907 | for (j = 0; j < set->nr_maps; j++) { |
bb94aea1 JW |
2908 | if (!set->map[j].nr_queues) { |
2909 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2910 | HCTX_TYPE_DEFAULT, i); | |
e5edd5f2 | 2911 | continue; |
bb94aea1 | 2912 | } |
fd689871 ML |
2913 | hctx_idx = set->map[j].mq_map[i]; |
2914 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2915 | if (!set->tags[hctx_idx] && | |
03b63b02 | 2916 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { |
fd689871 ML |
2917 | /* |
2918 | * If tags initialization fail for some hctx, | |
2919 | * that hctx won't be brought online. In this | |
2920 | * case, remap the current ctx to hctx[0] which | |
2921 | * is guaranteed to always have tags allocated | |
2922 | */ | |
2923 | set->map[j].mq_map[i] = 0; | |
2924 | } | |
e5edd5f2 | 2925 | |
b3c661b1 | 2926 | hctx = blk_mq_map_queue_type(q, j, i); |
8ccdf4a3 | 2927 | ctx->hctxs[j] = hctx; |
b3c661b1 JA |
2928 | /* |
2929 | * If the CPU is already set in the mask, then we've | |
2930 | * mapped this one already. This can happen if | |
2931 | * devices share queues across queue maps. | |
2932 | */ | |
2933 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2934 | continue; | |
2935 | ||
2936 | cpumask_set_cpu(i, hctx->cpumask); | |
2937 | hctx->type = j; | |
2938 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2939 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2940 | ||
2941 | /* | |
2942 | * If the nr_ctx type overflows, we have exceeded the | |
2943 | * amount of sw queues we can support. | |
2944 | */ | |
2945 | BUG_ON(!hctx->nr_ctx); | |
2946 | } | |
bb94aea1 JW |
2947 | |
2948 | for (; j < HCTX_MAX_TYPES; j++) | |
2949 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2950 | HCTX_TYPE_DEFAULT, i); | |
320ae51f | 2951 | } |
506e931f JA |
2952 | |
2953 | queue_for_each_hw_ctx(q, hctx, i) { | |
4412efec ML |
2954 | /* |
2955 | * If no software queues are mapped to this hardware queue, | |
2956 | * disable it and free the request entries. | |
2957 | */ | |
2958 | if (!hctx->nr_ctx) { | |
2959 | /* Never unmap queue 0. We need it as a | |
2960 | * fallback in case of a new remap fails | |
2961 | * allocation | |
2962 | */ | |
2963 | if (i && set->tags[i]) | |
2964 | blk_mq_free_map_and_requests(set, i); | |
2965 | ||
2966 | hctx->tags = NULL; | |
2967 | continue; | |
2968 | } | |
484b4061 | 2969 | |
2a34c087 ML |
2970 | hctx->tags = set->tags[i]; |
2971 | WARN_ON(!hctx->tags); | |
2972 | ||
889fa31f CY |
2973 | /* |
2974 | * Set the map size to the number of mapped software queues. | |
2975 | * This is more accurate and more efficient than looping | |
2976 | * over all possibly mapped software queues. | |
2977 | */ | |
88459642 | 2978 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2979 | |
484b4061 JA |
2980 | /* |
2981 | * Initialize batch roundrobin counts | |
2982 | */ | |
f82ddf19 | 2983 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); |
506e931f JA |
2984 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; |
2985 | } | |
320ae51f JA |
2986 | } |
2987 | ||
8e8320c9 JA |
2988 | /* |
2989 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2990 | * the queue isn't live yet. | |
2991 | */ | |
2404e607 | 2992 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2993 | { |
2994 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2995 | int i; |
2996 | ||
2404e607 | 2997 | queue_for_each_hw_ctx(q, hctx, i) { |
454bb677 | 2998 | if (shared) { |
51db1c37 | 2999 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 YK |
3000 | } else { |
3001 | blk_mq_tag_idle(hctx); | |
51db1c37 | 3002 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
454bb677 | 3003 | } |
2404e607 JM |
3004 | } |
3005 | } | |
3006 | ||
655ac300 HR |
3007 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, |
3008 | bool shared) | |
2404e607 JM |
3009 | { |
3010 | struct request_queue *q; | |
0d2602ca | 3011 | |
705cda97 BVA |
3012 | lockdep_assert_held(&set->tag_list_lock); |
3013 | ||
0d2602ca JA |
3014 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3015 | blk_mq_freeze_queue(q); | |
2404e607 | 3016 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
3017 | blk_mq_unfreeze_queue(q); |
3018 | } | |
3019 | } | |
3020 | ||
3021 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
3022 | { | |
3023 | struct blk_mq_tag_set *set = q->tag_set; | |
3024 | ||
0d2602ca | 3025 | mutex_lock(&set->tag_list_lock); |
08c875cb | 3026 | list_del(&q->tag_set_list); |
2404e607 JM |
3027 | if (list_is_singular(&set->tag_list)) { |
3028 | /* just transitioned to unshared */ | |
51db1c37 | 3029 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; |
2404e607 | 3030 | /* update existing queue */ |
655ac300 | 3031 | blk_mq_update_tag_set_shared(set, false); |
2404e607 | 3032 | } |
0d2602ca | 3033 | mutex_unlock(&set->tag_list_lock); |
a347c7ad | 3034 | INIT_LIST_HEAD(&q->tag_set_list); |
0d2602ca JA |
3035 | } |
3036 | ||
3037 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
3038 | struct request_queue *q) | |
3039 | { | |
0d2602ca | 3040 | mutex_lock(&set->tag_list_lock); |
2404e607 | 3041 | |
ff821d27 JA |
3042 | /* |
3043 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
3044 | */ | |
3045 | if (!list_empty(&set->tag_list) && | |
51db1c37 ML |
3046 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { |
3047 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2404e607 | 3048 | /* update existing queue */ |
655ac300 | 3049 | blk_mq_update_tag_set_shared(set, true); |
2404e607 | 3050 | } |
51db1c37 | 3051 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) |
2404e607 | 3052 | queue_set_hctx_shared(q, true); |
08c875cb | 3053 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 3054 | |
0d2602ca JA |
3055 | mutex_unlock(&set->tag_list_lock); |
3056 | } | |
3057 | ||
1db4909e ML |
3058 | /* All allocations will be freed in release handler of q->mq_kobj */ |
3059 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
3060 | { | |
3061 | struct blk_mq_ctxs *ctxs; | |
3062 | int cpu; | |
3063 | ||
3064 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
3065 | if (!ctxs) | |
3066 | return -ENOMEM; | |
3067 | ||
3068 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
3069 | if (!ctxs->queue_ctx) | |
3070 | goto fail; | |
3071 | ||
3072 | for_each_possible_cpu(cpu) { | |
3073 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
3074 | ctx->ctxs = ctxs; | |
3075 | } | |
3076 | ||
3077 | q->mq_kobj = &ctxs->kobj; | |
3078 | q->queue_ctx = ctxs->queue_ctx; | |
3079 | ||
3080 | return 0; | |
3081 | fail: | |
3082 | kfree(ctxs); | |
3083 | return -ENOMEM; | |
3084 | } | |
3085 | ||
e09aae7e ML |
3086 | /* |
3087 | * It is the actual release handler for mq, but we do it from | |
3088 | * request queue's release handler for avoiding use-after-free | |
3089 | * and headache because q->mq_kobj shouldn't have been introduced, | |
3090 | * but we can't group ctx/kctx kobj without it. | |
3091 | */ | |
3092 | void blk_mq_release(struct request_queue *q) | |
3093 | { | |
2f8f1336 ML |
3094 | struct blk_mq_hw_ctx *hctx, *next; |
3095 | int i; | |
e09aae7e | 3096 | |
2f8f1336 ML |
3097 | queue_for_each_hw_ctx(q, hctx, i) |
3098 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
3099 | ||
3100 | /* all hctx are in .unused_hctx_list now */ | |
3101 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
3102 | list_del_init(&hctx->hctx_list); | |
6c8b232e | 3103 | kobject_put(&hctx->kobj); |
c3b4afca | 3104 | } |
e09aae7e ML |
3105 | |
3106 | kfree(q->queue_hw_ctx); | |
3107 | ||
7ea5fe31 ML |
3108 | /* |
3109 | * release .mq_kobj and sw queue's kobject now because | |
3110 | * both share lifetime with request queue. | |
3111 | */ | |
3112 | blk_mq_sysfs_deinit(q); | |
e09aae7e ML |
3113 | } |
3114 | ||
5ec780a6 | 3115 | static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, |
2f227bb9 | 3116 | void *queuedata) |
b62c21b7 | 3117 | { |
26a9750a CH |
3118 | struct request_queue *q; |
3119 | int ret; | |
b62c21b7 | 3120 | |
26a9750a CH |
3121 | q = blk_alloc_queue(set->numa_node); |
3122 | if (!q) | |
b62c21b7 | 3123 | return ERR_PTR(-ENOMEM); |
26a9750a CH |
3124 | q->queuedata = queuedata; |
3125 | ret = blk_mq_init_allocated_queue(set, q); | |
3126 | if (ret) { | |
3127 | blk_cleanup_queue(q); | |
3128 | return ERR_PTR(ret); | |
3129 | } | |
b62c21b7 MS |
3130 | return q; |
3131 | } | |
2f227bb9 CH |
3132 | |
3133 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3134 | { | |
3135 | return blk_mq_init_queue_data(set, NULL); | |
3136 | } | |
b62c21b7 MS |
3137 | EXPORT_SYMBOL(blk_mq_init_queue); |
3138 | ||
4dcc4874 CH |
3139 | struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, |
3140 | struct lock_class_key *lkclass) | |
9316a9ed JA |
3141 | { |
3142 | struct request_queue *q; | |
b461dfc4 | 3143 | struct gendisk *disk; |
9316a9ed | 3144 | |
b461dfc4 CH |
3145 | q = blk_mq_init_queue_data(set, queuedata); |
3146 | if (IS_ERR(q)) | |
3147 | return ERR_CAST(q); | |
9316a9ed | 3148 | |
4a1fa41d | 3149 | disk = __alloc_disk_node(q, set->numa_node, lkclass); |
b461dfc4 CH |
3150 | if (!disk) { |
3151 | blk_cleanup_queue(q); | |
3152 | return ERR_PTR(-ENOMEM); | |
9316a9ed | 3153 | } |
b461dfc4 | 3154 | return disk; |
9316a9ed | 3155 | } |
b461dfc4 | 3156 | EXPORT_SYMBOL(__blk_mq_alloc_disk); |
9316a9ed | 3157 | |
34d11ffa JW |
3158 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( |
3159 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3160 | int hctx_idx, int node) | |
3161 | { | |
2f8f1336 | 3162 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; |
34d11ffa | 3163 | |
2f8f1336 ML |
3164 | /* reuse dead hctx first */ |
3165 | spin_lock(&q->unused_hctx_lock); | |
3166 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3167 | if (tmp->numa_node == node) { | |
3168 | hctx = tmp; | |
3169 | break; | |
3170 | } | |
3171 | } | |
3172 | if (hctx) | |
3173 | list_del_init(&hctx->hctx_list); | |
3174 | spin_unlock(&q->unused_hctx_lock); | |
3175 | ||
3176 | if (!hctx) | |
3177 | hctx = blk_mq_alloc_hctx(q, set, node); | |
34d11ffa | 3178 | if (!hctx) |
7c6c5b7c | 3179 | goto fail; |
34d11ffa | 3180 | |
7c6c5b7c ML |
3181 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) |
3182 | goto free_hctx; | |
34d11ffa JW |
3183 | |
3184 | return hctx; | |
7c6c5b7c ML |
3185 | |
3186 | free_hctx: | |
3187 | kobject_put(&hctx->kobj); | |
3188 | fail: | |
3189 | return NULL; | |
34d11ffa JW |
3190 | } |
3191 | ||
868f2f0b KB |
3192 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
3193 | struct request_queue *q) | |
320ae51f | 3194 | { |
e01ad46d | 3195 | int i, j, end; |
868f2f0b | 3196 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; |
f14bbe77 | 3197 | |
ac0d6b92 BVA |
3198 | if (q->nr_hw_queues < set->nr_hw_queues) { |
3199 | struct blk_mq_hw_ctx **new_hctxs; | |
3200 | ||
3201 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3202 | sizeof(*new_hctxs), GFP_KERNEL, | |
3203 | set->numa_node); | |
3204 | if (!new_hctxs) | |
3205 | return; | |
3206 | if (hctxs) | |
3207 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3208 | sizeof(*hctxs)); | |
3209 | q->queue_hw_ctx = new_hctxs; | |
ac0d6b92 BVA |
3210 | kfree(hctxs); |
3211 | hctxs = new_hctxs; | |
3212 | } | |
3213 | ||
fb350e0a ML |
3214 | /* protect against switching io scheduler */ |
3215 | mutex_lock(&q->sysfs_lock); | |
24d2f903 | 3216 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 3217 | int node; |
34d11ffa | 3218 | struct blk_mq_hw_ctx *hctx; |
868f2f0b | 3219 | |
7d76f856 | 3220 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); |
34d11ffa JW |
3221 | /* |
3222 | * If the hw queue has been mapped to another numa node, | |
3223 | * we need to realloc the hctx. If allocation fails, fallback | |
3224 | * to use the previous one. | |
3225 | */ | |
3226 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3227 | continue; | |
868f2f0b | 3228 | |
34d11ffa JW |
3229 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); |
3230 | if (hctx) { | |
2f8f1336 | 3231 | if (hctxs[i]) |
34d11ffa | 3232 | blk_mq_exit_hctx(q, set, hctxs[i], i); |
34d11ffa JW |
3233 | hctxs[i] = hctx; |
3234 | } else { | |
3235 | if (hctxs[i]) | |
3236 | pr_warn("Allocate new hctx on node %d fails,\ | |
3237 | fallback to previous one on node %d\n", | |
3238 | node, hctxs[i]->numa_node); | |
3239 | else | |
3240 | break; | |
868f2f0b | 3241 | } |
320ae51f | 3242 | } |
e01ad46d JW |
3243 | /* |
3244 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3245 | * hctxs and keep the previous q->nr_hw_queues. | |
3246 | */ | |
3247 | if (i != set->nr_hw_queues) { | |
3248 | j = q->nr_hw_queues; | |
3249 | end = i; | |
3250 | } else { | |
3251 | j = i; | |
3252 | end = q->nr_hw_queues; | |
3253 | q->nr_hw_queues = set->nr_hw_queues; | |
3254 | } | |
34d11ffa | 3255 | |
e01ad46d | 3256 | for (; j < end; j++) { |
868f2f0b KB |
3257 | struct blk_mq_hw_ctx *hctx = hctxs[j]; |
3258 | ||
3259 | if (hctx) { | |
cc71a6f4 JA |
3260 | if (hctx->tags) |
3261 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 3262 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 3263 | hctxs[j] = NULL; |
868f2f0b KB |
3264 | } |
3265 | } | |
fb350e0a | 3266 | mutex_unlock(&q->sysfs_lock); |
868f2f0b KB |
3267 | } |
3268 | ||
26a9750a CH |
3269 | int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, |
3270 | struct request_queue *q) | |
868f2f0b | 3271 | { |
66841672 ML |
3272 | /* mark the queue as mq asap */ |
3273 | q->mq_ops = set->ops; | |
3274 | ||
34dbad5d | 3275 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
3276 | blk_mq_poll_stats_bkt, |
3277 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
3278 | if (!q->poll_cb) |
3279 | goto err_exit; | |
3280 | ||
1db4909e | 3281 | if (blk_mq_alloc_ctxs(q)) |
41de54c6 | 3282 | goto err_poll; |
868f2f0b | 3283 | |
737f98cf ML |
3284 | /* init q->mq_kobj and sw queues' kobjects */ |
3285 | blk_mq_sysfs_init(q); | |
3286 | ||
2f8f1336 ML |
3287 | INIT_LIST_HEAD(&q->unused_hctx_list); |
3288 | spin_lock_init(&q->unused_hctx_lock); | |
3289 | ||
868f2f0b KB |
3290 | blk_mq_realloc_hw_ctxs(set, q); |
3291 | if (!q->nr_hw_queues) | |
3292 | goto err_hctxs; | |
320ae51f | 3293 | |
287922eb | 3294 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 3295 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f | 3296 | |
a8908939 | 3297 | q->tag_set = set; |
320ae51f | 3298 | |
94eddfbe | 3299 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
cd19181b ML |
3300 | if (set->nr_maps > HCTX_TYPE_POLL && |
3301 | set->map[HCTX_TYPE_POLL].nr_queues) | |
6544d229 | 3302 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); |
320ae51f | 3303 | |
2849450a | 3304 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
3305 | INIT_LIST_HEAD(&q->requeue_list); |
3306 | spin_lock_init(&q->requeue_lock); | |
3307 | ||
eba71768 JA |
3308 | q->nr_requests = set->queue_depth; |
3309 | ||
64f1c21e JA |
3310 | /* |
3311 | * Default to classic polling | |
3312 | */ | |
29ece8b4 | 3313 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; |
64f1c21e | 3314 | |
24d2f903 | 3315 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 3316 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 3317 | blk_mq_map_swqueue(q); |
26a9750a | 3318 | return 0; |
18741986 | 3319 | |
320ae51f | 3320 | err_hctxs: |
868f2f0b | 3321 | kfree(q->queue_hw_ctx); |
73d9c8d4 | 3322 | q->nr_hw_queues = 0; |
1db4909e | 3323 | blk_mq_sysfs_deinit(q); |
41de54c6 JS |
3324 | err_poll: |
3325 | blk_stat_free_callback(q->poll_cb); | |
3326 | q->poll_cb = NULL; | |
c7de5726 ML |
3327 | err_exit: |
3328 | q->mq_ops = NULL; | |
26a9750a | 3329 | return -ENOMEM; |
320ae51f | 3330 | } |
b62c21b7 | 3331 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f | 3332 | |
c7e2d94b ML |
3333 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ |
3334 | void blk_mq_exit_queue(struct request_queue *q) | |
320ae51f | 3335 | { |
630ef623 | 3336 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 3337 | |
630ef623 | 3338 | /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ |
624dbe47 | 3339 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
630ef623 BVA |
3340 | /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ |
3341 | blk_mq_del_queue_tag_set(q); | |
320ae51f | 3342 | } |
320ae51f | 3343 | |
a5164405 JA |
3344 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
3345 | { | |
3346 | int i; | |
3347 | ||
8229cca8 | 3348 | for (i = 0; i < set->nr_hw_queues; i++) { |
03b63b02 | 3349 | if (!__blk_mq_alloc_map_and_request(set, i)) |
a5164405 | 3350 | goto out_unwind; |
8229cca8 XT |
3351 | cond_resched(); |
3352 | } | |
a5164405 JA |
3353 | |
3354 | return 0; | |
3355 | ||
3356 | out_unwind: | |
3357 | while (--i >= 0) | |
2e194422 | 3358 | blk_mq_free_map_and_requests(set, i); |
a5164405 | 3359 | |
a5164405 JA |
3360 | return -ENOMEM; |
3361 | } | |
3362 | ||
3363 | /* | |
3364 | * Allocate the request maps associated with this tag_set. Note that this | |
3365 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3366 | * will be updated to reflect the allocated depth. | |
3367 | */ | |
79fab528 | 3368 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) |
a5164405 JA |
3369 | { |
3370 | unsigned int depth; | |
3371 | int err; | |
3372 | ||
3373 | depth = set->queue_depth; | |
3374 | do { | |
3375 | err = __blk_mq_alloc_rq_maps(set); | |
3376 | if (!err) | |
3377 | break; | |
3378 | ||
3379 | set->queue_depth >>= 1; | |
3380 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3381 | err = -ENOMEM; | |
3382 | break; | |
3383 | } | |
3384 | } while (set->queue_depth); | |
3385 | ||
3386 | if (!set->queue_depth || err) { | |
3387 | pr_err("blk-mq: failed to allocate request map\n"); | |
3388 | return -ENOMEM; | |
3389 | } | |
3390 | ||
3391 | if (depth != set->queue_depth) | |
3392 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3393 | depth, set->queue_depth); | |
3394 | ||
3395 | return 0; | |
3396 | } | |
3397 | ||
ebe8bddb OS |
3398 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
3399 | { | |
6e66b493 BVA |
3400 | /* |
3401 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3402 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3403 | * number of hardware queues. | |
3404 | */ | |
3405 | if (set->nr_maps == 1) | |
3406 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3407 | ||
59388702 | 3408 | if (set->ops->map_queues && !is_kdump_kernel()) { |
b3c661b1 JA |
3409 | int i; |
3410 | ||
7d4901a9 ML |
3411 | /* |
3412 | * transport .map_queues is usually done in the following | |
3413 | * way: | |
3414 | * | |
3415 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3416 | * mask = get_cpu_mask(queue) | |
3417 | * for_each_cpu(cpu, mask) | |
b3c661b1 | 3418 | * set->map[x].mq_map[cpu] = queue; |
7d4901a9 ML |
3419 | * } |
3420 | * | |
3421 | * When we need to remap, the table has to be cleared for | |
3422 | * killing stale mapping since one CPU may not be mapped | |
3423 | * to any hw queue. | |
3424 | */ | |
b3c661b1 JA |
3425 | for (i = 0; i < set->nr_maps; i++) |
3426 | blk_mq_clear_mq_map(&set->map[i]); | |
7d4901a9 | 3427 | |
ebe8bddb | 3428 | return set->ops->map_queues(set); |
b3c661b1 JA |
3429 | } else { |
3430 | BUG_ON(set->nr_maps > 1); | |
7d76f856 | 3431 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
b3c661b1 | 3432 | } |
ebe8bddb OS |
3433 | } |
3434 | ||
f7e76dbc BVA |
3435 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, |
3436 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3437 | { | |
3438 | struct blk_mq_tags **new_tags; | |
3439 | ||
3440 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3441 | return 0; | |
3442 | ||
3443 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3444 | GFP_KERNEL, set->numa_node); | |
3445 | if (!new_tags) | |
3446 | return -ENOMEM; | |
3447 | ||
3448 | if (set->tags) | |
3449 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3450 | sizeof(*set->tags)); | |
3451 | kfree(set->tags); | |
3452 | set->tags = new_tags; | |
3453 | set->nr_hw_queues = new_nr_hw_queues; | |
3454 | ||
3455 | return 0; | |
3456 | } | |
3457 | ||
91cdf265 MI |
3458 | static int blk_mq_alloc_tag_set_tags(struct blk_mq_tag_set *set, |
3459 | int new_nr_hw_queues) | |
3460 | { | |
3461 | return blk_mq_realloc_tag_set_tags(set, 0, new_nr_hw_queues); | |
3462 | } | |
3463 | ||
a4391c64 JA |
3464 | /* |
3465 | * Alloc a tag set to be associated with one or more request queues. | |
3466 | * May fail with EINVAL for various error conditions. May adjust the | |
c018c84f | 3467 | * requested depth down, if it's too large. In that case, the set |
a4391c64 JA |
3468 | * value will be stored in set->queue_depth. |
3469 | */ | |
24d2f903 CH |
3470 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
3471 | { | |
b3c661b1 | 3472 | int i, ret; |
da695ba2 | 3473 | |
205fb5f5 BVA |
3474 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
3475 | ||
24d2f903 CH |
3476 | if (!set->nr_hw_queues) |
3477 | return -EINVAL; | |
a4391c64 | 3478 | if (!set->queue_depth) |
24d2f903 CH |
3479 | return -EINVAL; |
3480 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3481 | return -EINVAL; | |
3482 | ||
7d7e0f90 | 3483 | if (!set->ops->queue_rq) |
24d2f903 CH |
3484 | return -EINVAL; |
3485 | ||
de148297 ML |
3486 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
3487 | return -EINVAL; | |
3488 | ||
a4391c64 JA |
3489 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
3490 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3491 | BLK_MQ_MAX_DEPTH); | |
3492 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3493 | } | |
24d2f903 | 3494 | |
b3c661b1 JA |
3495 | if (!set->nr_maps) |
3496 | set->nr_maps = 1; | |
3497 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3498 | return -EINVAL; | |
3499 | ||
6637fadf SL |
3500 | /* |
3501 | * If a crashdump is active, then we are potentially in a very | |
3502 | * memory constrained environment. Limit us to 1 queue and | |
3503 | * 64 tags to prevent using too much memory. | |
3504 | */ | |
3505 | if (is_kdump_kernel()) { | |
3506 | set->nr_hw_queues = 1; | |
59388702 | 3507 | set->nr_maps = 1; |
6637fadf SL |
3508 | set->queue_depth = min(64U, set->queue_depth); |
3509 | } | |
868f2f0b | 3510 | /* |
392546ae JA |
3511 | * There is no use for more h/w queues than cpus if we just have |
3512 | * a single map | |
868f2f0b | 3513 | */ |
392546ae | 3514 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3515 | set->nr_hw_queues = nr_cpu_ids; |
6637fadf | 3516 | |
91cdf265 | 3517 | if (blk_mq_alloc_tag_set_tags(set, set->nr_hw_queues) < 0) |
a5164405 | 3518 | return -ENOMEM; |
24d2f903 | 3519 | |
da695ba2 | 3520 | ret = -ENOMEM; |
b3c661b1 JA |
3521 | for (i = 0; i < set->nr_maps; i++) { |
3522 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
07b35eb5 | 3523 | sizeof(set->map[i].mq_map[0]), |
b3c661b1 JA |
3524 | GFP_KERNEL, set->numa_node); |
3525 | if (!set->map[i].mq_map) | |
3526 | goto out_free_mq_map; | |
59388702 | 3527 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; |
b3c661b1 | 3528 | } |
bdd17e75 | 3529 | |
ebe8bddb | 3530 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
3531 | if (ret) |
3532 | goto out_free_mq_map; | |
3533 | ||
79fab528 | 3534 | ret = blk_mq_alloc_map_and_requests(set); |
da695ba2 | 3535 | if (ret) |
bdd17e75 | 3536 | goto out_free_mq_map; |
24d2f903 | 3537 | |
32bc15af | 3538 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
f1b49fdc JG |
3539 | atomic_set(&set->active_queues_shared_sbitmap, 0); |
3540 | ||
56b68085 | 3541 | if (blk_mq_init_shared_sbitmap(set)) { |
32bc15af JG |
3542 | ret = -ENOMEM; |
3543 | goto out_free_mq_rq_maps; | |
3544 | } | |
3545 | } | |
3546 | ||
0d2602ca JA |
3547 | mutex_init(&set->tag_list_lock); |
3548 | INIT_LIST_HEAD(&set->tag_list); | |
3549 | ||
24d2f903 | 3550 | return 0; |
bdd17e75 | 3551 | |
32bc15af JG |
3552 | out_free_mq_rq_maps: |
3553 | for (i = 0; i < set->nr_hw_queues; i++) | |
3554 | blk_mq_free_map_and_requests(set, i); | |
bdd17e75 | 3555 | out_free_mq_map: |
b3c661b1 JA |
3556 | for (i = 0; i < set->nr_maps; i++) { |
3557 | kfree(set->map[i].mq_map); | |
3558 | set->map[i].mq_map = NULL; | |
3559 | } | |
5676e7b6 RE |
3560 | kfree(set->tags); |
3561 | set->tags = NULL; | |
da695ba2 | 3562 | return ret; |
24d2f903 CH |
3563 | } |
3564 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3565 | ||
cdb14e0f CH |
3566 | /* allocate and initialize a tagset for a simple single-queue device */ |
3567 | int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, | |
3568 | const struct blk_mq_ops *ops, unsigned int queue_depth, | |
3569 | unsigned int set_flags) | |
3570 | { | |
3571 | memset(set, 0, sizeof(*set)); | |
3572 | set->ops = ops; | |
3573 | set->nr_hw_queues = 1; | |
3574 | set->nr_maps = 1; | |
3575 | set->queue_depth = queue_depth; | |
3576 | set->numa_node = NUMA_NO_NODE; | |
3577 | set->flags = set_flags; | |
3578 | return blk_mq_alloc_tag_set(set); | |
3579 | } | |
3580 | EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); | |
3581 | ||
24d2f903 CH |
3582 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) |
3583 | { | |
b3c661b1 | 3584 | int i, j; |
24d2f903 | 3585 | |
f7e76dbc | 3586 | for (i = 0; i < set->nr_hw_queues; i++) |
cc71a6f4 | 3587 | blk_mq_free_map_and_requests(set, i); |
484b4061 | 3588 | |
32bc15af JG |
3589 | if (blk_mq_is_sbitmap_shared(set->flags)) |
3590 | blk_mq_exit_shared_sbitmap(set); | |
3591 | ||
b3c661b1 JA |
3592 | for (j = 0; j < set->nr_maps; j++) { |
3593 | kfree(set->map[j].mq_map); | |
3594 | set->map[j].mq_map = NULL; | |
3595 | } | |
bdd17e75 | 3596 | |
981bd189 | 3597 | kfree(set->tags); |
5676e7b6 | 3598 | set->tags = NULL; |
24d2f903 CH |
3599 | } |
3600 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3601 | ||
e3a2b3f9 JA |
3602 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
3603 | { | |
3604 | struct blk_mq_tag_set *set = q->tag_set; | |
3605 | struct blk_mq_hw_ctx *hctx; | |
3606 | int i, ret; | |
3607 | ||
bd166ef1 | 3608 | if (!set) |
e3a2b3f9 JA |
3609 | return -EINVAL; |
3610 | ||
e5fa8140 AZ |
3611 | if (q->nr_requests == nr) |
3612 | return 0; | |
3613 | ||
70f36b60 | 3614 | blk_mq_freeze_queue(q); |
24f5a90f | 3615 | blk_mq_quiesce_queue(q); |
70f36b60 | 3616 | |
e3a2b3f9 JA |
3617 | ret = 0; |
3618 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
3619 | if (!hctx->tags) |
3620 | continue; | |
bd166ef1 JA |
3621 | /* |
3622 | * If we're using an MQ scheduler, just update the scheduler | |
3623 | * queue depth. This is similar to what the old code would do. | |
3624 | */ | |
70f36b60 | 3625 | if (!hctx->sched_tags) { |
c2e82a23 | 3626 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, |
70f36b60 | 3627 | false); |
32bc15af JG |
3628 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) |
3629 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
70f36b60 JA |
3630 | } else { |
3631 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3632 | nr, true); | |
d97e594c JG |
3633 | if (blk_mq_is_sbitmap_shared(set->flags)) { |
3634 | hctx->sched_tags->bitmap_tags = | |
3635 | &q->sched_bitmap_tags; | |
3636 | hctx->sched_tags->breserved_tags = | |
3637 | &q->sched_breserved_tags; | |
3638 | } | |
70f36b60 | 3639 | } |
e3a2b3f9 JA |
3640 | if (ret) |
3641 | break; | |
77f1e0a5 JA |
3642 | if (q->elevator && q->elevator->type->ops.depth_updated) |
3643 | q->elevator->type->ops.depth_updated(hctx); | |
e3a2b3f9 | 3644 | } |
d97e594c | 3645 | if (!ret) { |
e3a2b3f9 | 3646 | q->nr_requests = nr; |
d97e594c JG |
3647 | if (q->elevator && blk_mq_is_sbitmap_shared(set->flags)) |
3648 | sbitmap_queue_resize(&q->sched_bitmap_tags, | |
3649 | nr - set->reserved_tags); | |
3650 | } | |
e3a2b3f9 | 3651 | |
24f5a90f | 3652 | blk_mq_unquiesce_queue(q); |
70f36b60 | 3653 | blk_mq_unfreeze_queue(q); |
70f36b60 | 3654 | |
e3a2b3f9 JA |
3655 | return ret; |
3656 | } | |
3657 | ||
d48ece20 JW |
3658 | /* |
3659 | * request_queue and elevator_type pair. | |
3660 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3661 | * the elevator_type associated with a request_queue. | |
3662 | */ | |
3663 | struct blk_mq_qe_pair { | |
3664 | struct list_head node; | |
3665 | struct request_queue *q; | |
3666 | struct elevator_type *type; | |
3667 | }; | |
3668 | ||
3669 | /* | |
3670 | * Cache the elevator_type in qe pair list and switch the | |
3671 | * io scheduler to 'none' | |
3672 | */ | |
3673 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3674 | struct request_queue *q) | |
3675 | { | |
3676 | struct blk_mq_qe_pair *qe; | |
3677 | ||
3678 | if (!q->elevator) | |
3679 | return true; | |
3680 | ||
3681 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3682 | if (!qe) | |
3683 | return false; | |
3684 | ||
3685 | INIT_LIST_HEAD(&qe->node); | |
3686 | qe->q = q; | |
3687 | qe->type = q->elevator->type; | |
3688 | list_add(&qe->node, head); | |
3689 | ||
3690 | mutex_lock(&q->sysfs_lock); | |
3691 | /* | |
3692 | * After elevator_switch_mq, the previous elevator_queue will be | |
3693 | * released by elevator_release. The reference of the io scheduler | |
3694 | * module get by elevator_get will also be put. So we need to get | |
3695 | * a reference of the io scheduler module here to prevent it to be | |
3696 | * removed. | |
3697 | */ | |
3698 | __module_get(qe->type->elevator_owner); | |
3699 | elevator_switch_mq(q, NULL); | |
3700 | mutex_unlock(&q->sysfs_lock); | |
3701 | ||
3702 | return true; | |
3703 | } | |
3704 | ||
3705 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3706 | struct request_queue *q) | |
3707 | { | |
3708 | struct blk_mq_qe_pair *qe; | |
3709 | struct elevator_type *t = NULL; | |
3710 | ||
3711 | list_for_each_entry(qe, head, node) | |
3712 | if (qe->q == q) { | |
3713 | t = qe->type; | |
3714 | break; | |
3715 | } | |
3716 | ||
3717 | if (!t) | |
3718 | return; | |
3719 | ||
3720 | list_del(&qe->node); | |
3721 | kfree(qe); | |
3722 | ||
3723 | mutex_lock(&q->sysfs_lock); | |
3724 | elevator_switch_mq(q, t); | |
3725 | mutex_unlock(&q->sysfs_lock); | |
3726 | } | |
3727 | ||
e4dc2b32 KB |
3728 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
3729 | int nr_hw_queues) | |
868f2f0b KB |
3730 | { |
3731 | struct request_queue *q; | |
d48ece20 | 3732 | LIST_HEAD(head); |
e01ad46d | 3733 | int prev_nr_hw_queues; |
868f2f0b | 3734 | |
705cda97 BVA |
3735 | lockdep_assert_held(&set->tag_list_lock); |
3736 | ||
392546ae | 3737 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) |
868f2f0b | 3738 | nr_hw_queues = nr_cpu_ids; |
fe35ec58 WZ |
3739 | if (nr_hw_queues < 1) |
3740 | return; | |
3741 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
868f2f0b KB |
3742 | return; |
3743 | ||
3744 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3745 | blk_mq_freeze_queue(q); | |
d48ece20 JW |
3746 | /* |
3747 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3748 | * with the previous scheduler. We will switch back once we are done | |
3749 | * updating the new sw to hw queue mappings. | |
3750 | */ | |
3751 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3752 | if (!blk_mq_elv_switch_none(&head, q)) | |
3753 | goto switch_back; | |
868f2f0b | 3754 | |
477e19de JW |
3755 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3756 | blk_mq_debugfs_unregister_hctxs(q); | |
3757 | blk_mq_sysfs_unregister(q); | |
3758 | } | |
3759 | ||
a2584e43 | 3760 | prev_nr_hw_queues = set->nr_hw_queues; |
f7e76dbc BVA |
3761 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < |
3762 | 0) | |
3763 | goto reregister; | |
3764 | ||
868f2f0b | 3765 | set->nr_hw_queues = nr_hw_queues; |
e01ad46d | 3766 | fallback: |
aa880ad6 | 3767 | blk_mq_update_queue_map(set); |
868f2f0b KB |
3768 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3769 | blk_mq_realloc_hw_ctxs(set, q); | |
e01ad46d JW |
3770 | if (q->nr_hw_queues != set->nr_hw_queues) { |
3771 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3772 | nr_hw_queues, prev_nr_hw_queues); | |
3773 | set->nr_hw_queues = prev_nr_hw_queues; | |
7d76f856 | 3774 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); |
e01ad46d JW |
3775 | goto fallback; |
3776 | } | |
477e19de JW |
3777 | blk_mq_map_swqueue(q); |
3778 | } | |
3779 | ||
f7e76dbc | 3780 | reregister: |
477e19de JW |
3781 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
3782 | blk_mq_sysfs_register(q); | |
3783 | blk_mq_debugfs_register_hctxs(q); | |
868f2f0b KB |
3784 | } |
3785 | ||
d48ece20 JW |
3786 | switch_back: |
3787 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3788 | blk_mq_elv_switch_back(&head, q); | |
3789 | ||
868f2f0b KB |
3790 | list_for_each_entry(q, &set->tag_list, tag_set_list) |
3791 | blk_mq_unfreeze_queue(q); | |
3792 | } | |
e4dc2b32 KB |
3793 | |
3794 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3795 | { | |
3796 | mutex_lock(&set->tag_list_lock); | |
3797 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3798 | mutex_unlock(&set->tag_list_lock); | |
3799 | } | |
868f2f0b KB |
3800 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
3801 | ||
34dbad5d OS |
3802 | /* Enable polling stats and return whether they were already enabled. */ |
3803 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3804 | { | |
3805 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
7dfdbc73 | 3806 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) |
34dbad5d OS |
3807 | return true; |
3808 | blk_stat_add_callback(q, q->poll_cb); | |
3809 | return false; | |
3810 | } | |
3811 | ||
3812 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3813 | { | |
3814 | /* | |
3815 | * We don't arm the callback if polling stats are not enabled or the | |
3816 | * callback is already active. | |
3817 | */ | |
3818 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3819 | blk_stat_is_active(q->poll_cb)) | |
3820 | return; | |
3821 | ||
3822 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3823 | } | |
3824 | ||
3825 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3826 | { | |
3827 | struct request_queue *q = cb->data; | |
720b8ccc | 3828 | int bucket; |
34dbad5d | 3829 | |
720b8ccc SB |
3830 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
3831 | if (cb->stat[bucket].nr_samples) | |
3832 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3833 | } | |
34dbad5d OS |
3834 | } |
3835 | ||
64f1c21e | 3836 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
64f1c21e JA |
3837 | struct request *rq) |
3838 | { | |
64f1c21e | 3839 | unsigned long ret = 0; |
720b8ccc | 3840 | int bucket; |
64f1c21e JA |
3841 | |
3842 | /* | |
3843 | * If stats collection isn't on, don't sleep but turn it on for | |
3844 | * future users | |
3845 | */ | |
34dbad5d | 3846 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
3847 | return 0; |
3848 | ||
64f1c21e JA |
3849 | /* |
3850 | * As an optimistic guess, use half of the mean service time | |
3851 | * for this type of request. We can (and should) make this smarter. | |
3852 | * For instance, if the completion latencies are tight, we can | |
3853 | * get closer than just half the mean. This is especially | |
3854 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
3855 | * than ~10 usec. We do use the stats for the relevant IO size |
3856 | * if available which does lead to better estimates. | |
64f1c21e | 3857 | */ |
720b8ccc SB |
3858 | bucket = blk_mq_poll_stats_bkt(rq); |
3859 | if (bucket < 0) | |
3860 | return ret; | |
3861 | ||
3862 | if (q->poll_stat[bucket].nr_samples) | |
3863 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
3864 | |
3865 | return ret; | |
3866 | } | |
3867 | ||
06426adf JA |
3868 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
3869 | struct request *rq) | |
3870 | { | |
3871 | struct hrtimer_sleeper hs; | |
3872 | enum hrtimer_mode mode; | |
64f1c21e | 3873 | unsigned int nsecs; |
06426adf JA |
3874 | ktime_t kt; |
3875 | ||
76a86f9d | 3876 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) |
64f1c21e JA |
3877 | return false; |
3878 | ||
3879 | /* | |
1052b8ac | 3880 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: |
64f1c21e | 3881 | * |
64f1c21e JA |
3882 | * 0: use half of prev avg |
3883 | * >0: use this specific value | |
3884 | */ | |
1052b8ac | 3885 | if (q->poll_nsec > 0) |
64f1c21e JA |
3886 | nsecs = q->poll_nsec; |
3887 | else | |
cae740a0 | 3888 | nsecs = blk_mq_poll_nsecs(q, rq); |
64f1c21e JA |
3889 | |
3890 | if (!nsecs) | |
06426adf JA |
3891 | return false; |
3892 | ||
76a86f9d | 3893 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; |
06426adf JA |
3894 | |
3895 | /* | |
3896 | * This will be replaced with the stats tracking code, using | |
3897 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3898 | */ | |
8b0e1953 | 3899 | kt = nsecs; |
06426adf JA |
3900 | |
3901 | mode = HRTIMER_MODE_REL; | |
dbc1625f | 3902 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); |
06426adf JA |
3903 | hrtimer_set_expires(&hs.timer, kt); |
3904 | ||
06426adf | 3905 | do { |
5a61c363 | 3906 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) |
06426adf JA |
3907 | break; |
3908 | set_current_state(TASK_UNINTERRUPTIBLE); | |
9dd8813e | 3909 | hrtimer_sleeper_start_expires(&hs, mode); |
06426adf JA |
3910 | if (hs.task) |
3911 | io_schedule(); | |
3912 | hrtimer_cancel(&hs.timer); | |
3913 | mode = HRTIMER_MODE_ABS; | |
3914 | } while (hs.task && !signal_pending(current)); | |
3915 | ||
3916 | __set_current_state(TASK_RUNNING); | |
3917 | destroy_hrtimer_on_stack(&hs.timer); | |
3918 | return true; | |
3919 | } | |
3920 | ||
1052b8ac JA |
3921 | static bool blk_mq_poll_hybrid(struct request_queue *q, |
3922 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
bbd7bb70 | 3923 | { |
1052b8ac JA |
3924 | struct request *rq; |
3925 | ||
29ece8b4 | 3926 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) |
1052b8ac JA |
3927 | return false; |
3928 | ||
3929 | if (!blk_qc_t_is_internal(cookie)) | |
3930 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3931 | else { | |
3932 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3933 | /* | |
3934 | * With scheduling, if the request has completed, we'll | |
3935 | * get a NULL return here, as we clear the sched tag when | |
3936 | * that happens. The request still remains valid, like always, | |
3937 | * so we should be safe with just the NULL check. | |
3938 | */ | |
3939 | if (!rq) | |
3940 | return false; | |
3941 | } | |
3942 | ||
cae740a0 | 3943 | return blk_mq_poll_hybrid_sleep(q, rq); |
1052b8ac JA |
3944 | } |
3945 | ||
529262d5 CH |
3946 | /** |
3947 | * blk_poll - poll for IO completions | |
3948 | * @q: the queue | |
3949 | * @cookie: cookie passed back at IO submission time | |
3950 | * @spin: whether to spin for completions | |
3951 | * | |
3952 | * Description: | |
3953 | * Poll for completions on the passed in queue. Returns number of | |
3954 | * completed entries found. If @spin is true, then blk_poll will continue | |
3955 | * looping until at least one completion is found, unless the task is | |
3956 | * otherwise marked running (or we need to reschedule). | |
3957 | */ | |
3958 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
1052b8ac JA |
3959 | { |
3960 | struct blk_mq_hw_ctx *hctx; | |
2f064a59 | 3961 | unsigned int state; |
bbd7bb70 | 3962 | |
529262d5 CH |
3963 | if (!blk_qc_t_valid(cookie) || |
3964 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
1052b8ac JA |
3965 | return 0; |
3966 | ||
529262d5 CH |
3967 | if (current->plug) |
3968 | blk_flush_plug_list(current->plug, false); | |
3969 | ||
1052b8ac JA |
3970 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; |
3971 | ||
06426adf JA |
3972 | /* |
3973 | * If we sleep, have the caller restart the poll loop to reset | |
3974 | * the state. Like for the other success return cases, the | |
3975 | * caller is responsible for checking if the IO completed. If | |
3976 | * the IO isn't complete, we'll get called again and will go | |
f6f371f7 PB |
3977 | * straight to the busy poll loop. If specified not to spin, |
3978 | * we also should not sleep. | |
06426adf | 3979 | */ |
f6f371f7 | 3980 | if (spin && blk_mq_poll_hybrid(q, hctx, cookie)) |
85f4d4b6 | 3981 | return 1; |
06426adf | 3982 | |
bbd7bb70 JA |
3983 | hctx->poll_considered++; |
3984 | ||
d6c23bb3 | 3985 | state = get_current_state(); |
aa61bec3 | 3986 | do { |
bbd7bb70 JA |
3987 | int ret; |
3988 | ||
3989 | hctx->poll_invoked++; | |
3990 | ||
9743139c | 3991 | ret = q->mq_ops->poll(hctx); |
bbd7bb70 JA |
3992 | if (ret > 0) { |
3993 | hctx->poll_success++; | |
849a3700 | 3994 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 3995 | return ret; |
bbd7bb70 JA |
3996 | } |
3997 | ||
3998 | if (signal_pending_state(state, current)) | |
849a3700 | 3999 | __set_current_state(TASK_RUNNING); |
bbd7bb70 | 4000 | |
b03fbd4f | 4001 | if (task_is_running(current)) |
85f4d4b6 | 4002 | return 1; |
0a1b8b87 | 4003 | if (ret < 0 || !spin) |
bbd7bb70 JA |
4004 | break; |
4005 | cpu_relax(); | |
aa61bec3 | 4006 | } while (!need_resched()); |
bbd7bb70 | 4007 | |
67b4110f | 4008 | __set_current_state(TASK_RUNNING); |
85f4d4b6 | 4009 | return 0; |
bbd7bb70 | 4010 | } |
529262d5 | 4011 | EXPORT_SYMBOL_GPL(blk_poll); |
bbd7bb70 | 4012 | |
9cf2bab6 JA |
4013 | unsigned int blk_mq_rq_cpu(struct request *rq) |
4014 | { | |
4015 | return rq->mq_ctx->cpu; | |
4016 | } | |
4017 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
4018 | ||
320ae51f JA |
4019 | static int __init blk_mq_init(void) |
4020 | { | |
c3077b5d CH |
4021 | int i; |
4022 | ||
4023 | for_each_possible_cpu(i) | |
f9ab4918 | 4024 | init_llist_head(&per_cpu(blk_cpu_done, i)); |
c3077b5d CH |
4025 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); |
4026 | ||
4027 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
4028 | "block/softirq:dead", NULL, | |
4029 | blk_softirq_cpu_dead); | |
9467f859 TG |
4030 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
4031 | blk_mq_hctx_notify_dead); | |
bf0beec0 ML |
4032 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", |
4033 | blk_mq_hctx_notify_online, | |
4034 | blk_mq_hctx_notify_offline); | |
320ae51f JA |
4035 | return 0; |
4036 | } | |
4037 | subsys_initcall(blk_mq_init); |