]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
34 | #include "blk-mq-tag.h" | |
cf43e6be | 35 | #include "blk-stat.h" |
87760e5e | 36 | #include "blk-wbt.h" |
bd166ef1 | 37 | #include "blk-mq-sched.h" |
320ae51f JA |
38 | |
39 | static DEFINE_MUTEX(all_q_mutex); | |
40 | static LIST_HEAD(all_q_list); | |
41 | ||
320ae51f JA |
42 | /* |
43 | * Check if any of the ctx's have pending work in this hardware queue | |
44 | */ | |
50e1dab8 | 45 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 46 | { |
bd166ef1 JA |
47 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
48 | !list_empty_careful(&hctx->dispatch) || | |
49 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
50 | } |
51 | ||
320ae51f JA |
52 | /* |
53 | * Mark this ctx as having pending work in this hardware queue | |
54 | */ | |
55 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
56 | struct blk_mq_ctx *ctx) | |
57 | { | |
88459642 OS |
58 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
59 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
60 | } |
61 | ||
62 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
63 | struct blk_mq_ctx *ctx) | |
64 | { | |
88459642 | 65 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
66 | } |
67 | ||
b4c6a028 | 68 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 69 | { |
4ecd4fef | 70 | int freeze_depth; |
cddd5d17 | 71 | |
4ecd4fef CH |
72 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
73 | if (freeze_depth == 1) { | |
3ef28e83 | 74 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 75 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 76 | } |
f3af020b | 77 | } |
b4c6a028 | 78 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b | 79 | |
6bae363e | 80 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 81 | { |
3ef28e83 | 82 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 83 | } |
6bae363e | 84 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 85 | |
f91328c4 KB |
86 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
87 | unsigned long timeout) | |
88 | { | |
89 | return wait_event_timeout(q->mq_freeze_wq, | |
90 | percpu_ref_is_zero(&q->q_usage_counter), | |
91 | timeout); | |
92 | } | |
93 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 94 | |
f3af020b TH |
95 | /* |
96 | * Guarantee no request is in use, so we can change any data structure of | |
97 | * the queue afterward. | |
98 | */ | |
3ef28e83 | 99 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 100 | { |
3ef28e83 DW |
101 | /* |
102 | * In the !blk_mq case we are only calling this to kill the | |
103 | * q_usage_counter, otherwise this increases the freeze depth | |
104 | * and waits for it to return to zero. For this reason there is | |
105 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
106 | * exported to drivers as the only user for unfreeze is blk_mq. | |
107 | */ | |
f3af020b TH |
108 | blk_mq_freeze_queue_start(q); |
109 | blk_mq_freeze_queue_wait(q); | |
110 | } | |
3ef28e83 DW |
111 | |
112 | void blk_mq_freeze_queue(struct request_queue *q) | |
113 | { | |
114 | /* | |
115 | * ...just an alias to keep freeze and unfreeze actions balanced | |
116 | * in the blk_mq_* namespace | |
117 | */ | |
118 | blk_freeze_queue(q); | |
119 | } | |
c761d96b | 120 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 121 | |
b4c6a028 | 122 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 123 | { |
4ecd4fef | 124 | int freeze_depth; |
320ae51f | 125 | |
4ecd4fef CH |
126 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
127 | WARN_ON_ONCE(freeze_depth < 0); | |
128 | if (!freeze_depth) { | |
3ef28e83 | 129 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 130 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 131 | } |
320ae51f | 132 | } |
b4c6a028 | 133 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 134 | |
6a83e74d BVA |
135 | /** |
136 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
137 | * @q: request queue. | |
138 | * | |
139 | * Note: this function does not prevent that the struct request end_io() | |
140 | * callback function is invoked. Additionally, it is not prevented that | |
141 | * new queue_rq() calls occur unless the queue has been stopped first. | |
142 | */ | |
143 | void blk_mq_quiesce_queue(struct request_queue *q) | |
144 | { | |
145 | struct blk_mq_hw_ctx *hctx; | |
146 | unsigned int i; | |
147 | bool rcu = false; | |
148 | ||
149 | blk_mq_stop_hw_queues(q); | |
150 | ||
151 | queue_for_each_hw_ctx(q, hctx, i) { | |
152 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
153 | synchronize_srcu(&hctx->queue_rq_srcu); | |
154 | else | |
155 | rcu = true; | |
156 | } | |
157 | if (rcu) | |
158 | synchronize_rcu(); | |
159 | } | |
160 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
161 | ||
aed3ea94 JA |
162 | void blk_mq_wake_waiters(struct request_queue *q) |
163 | { | |
164 | struct blk_mq_hw_ctx *hctx; | |
165 | unsigned int i; | |
166 | ||
167 | queue_for_each_hw_ctx(q, hctx, i) | |
168 | if (blk_mq_hw_queue_mapped(hctx)) | |
169 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
170 | |
171 | /* | |
172 | * If we are called because the queue has now been marked as | |
173 | * dying, we need to ensure that processes currently waiting on | |
174 | * the queue are notified as well. | |
175 | */ | |
176 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
177 | } |
178 | ||
320ae51f JA |
179 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
180 | { | |
181 | return blk_mq_has_free_tags(hctx->tags); | |
182 | } | |
183 | EXPORT_SYMBOL(blk_mq_can_queue); | |
184 | ||
2c3ad667 JA |
185 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
186 | struct request *rq, unsigned int op) | |
320ae51f | 187 | { |
af76e555 CH |
188 | INIT_LIST_HEAD(&rq->queuelist); |
189 | /* csd/requeue_work/fifo_time is initialized before use */ | |
190 | rq->q = q; | |
320ae51f | 191 | rq->mq_ctx = ctx; |
ef295ecf | 192 | rq->cmd_flags = op; |
e8064021 CH |
193 | if (blk_queue_io_stat(q)) |
194 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
195 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
196 | rq->cpu = -1; | |
af76e555 CH |
197 | INIT_HLIST_NODE(&rq->hash); |
198 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
199 | rq->rq_disk = NULL; |
200 | rq->part = NULL; | |
3ee32372 | 201 | rq->start_time = jiffies; |
af76e555 CH |
202 | #ifdef CONFIG_BLK_CGROUP |
203 | rq->rl = NULL; | |
0fec08b4 | 204 | set_start_time_ns(rq); |
af76e555 CH |
205 | rq->io_start_time_ns = 0; |
206 | #endif | |
207 | rq->nr_phys_segments = 0; | |
208 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
209 | rq->nr_integrity_segments = 0; | |
210 | #endif | |
af76e555 CH |
211 | rq->special = NULL; |
212 | /* tag was already set */ | |
213 | rq->errors = 0; | |
af76e555 | 214 | rq->extra_len = 0; |
af76e555 | 215 | |
af76e555 | 216 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
217 | rq->timeout = 0; |
218 | ||
af76e555 CH |
219 | rq->end_io = NULL; |
220 | rq->end_io_data = NULL; | |
221 | rq->next_rq = NULL; | |
222 | ||
ef295ecf | 223 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 224 | } |
2c3ad667 | 225 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 226 | |
2c3ad667 JA |
227 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
228 | unsigned int op) | |
5dee8577 CH |
229 | { |
230 | struct request *rq; | |
231 | unsigned int tag; | |
232 | ||
cb96a42c | 233 | tag = blk_mq_get_tag(data); |
5dee8577 | 234 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
235 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
236 | ||
237 | rq = tags->static_rqs[tag]; | |
5dee8577 | 238 | |
bd166ef1 JA |
239 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
240 | rq->tag = -1; | |
241 | rq->internal_tag = tag; | |
242 | } else { | |
200e86b3 JA |
243 | if (blk_mq_tag_busy(data->hctx)) { |
244 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
245 | atomic_inc(&data->hctx->nr_active); | |
246 | } | |
bd166ef1 JA |
247 | rq->tag = tag; |
248 | rq->internal_tag = -1; | |
562bef42 | 249 | data->hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
250 | } |
251 | ||
ef295ecf | 252 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
253 | return rq; |
254 | } | |
255 | ||
256 | return NULL; | |
257 | } | |
2c3ad667 | 258 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 259 | |
6f3b0e8b CH |
260 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
261 | unsigned int flags) | |
320ae51f | 262 | { |
5a797e00 | 263 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 264 | struct request *rq; |
a492f075 | 265 | int ret; |
320ae51f | 266 | |
6f3b0e8b | 267 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
268 | if (ret) |
269 | return ERR_PTR(ret); | |
320ae51f | 270 | |
bd166ef1 | 271 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 272 | |
bd166ef1 JA |
273 | blk_mq_put_ctx(alloc_data.ctx); |
274 | blk_queue_exit(q); | |
275 | ||
276 | if (!rq) | |
a492f075 | 277 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
278 | |
279 | rq->__data_len = 0; | |
280 | rq->__sector = (sector_t) -1; | |
281 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
282 | return rq; |
283 | } | |
4bb659b1 | 284 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 285 | |
1f5bd336 ML |
286 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
287 | unsigned int flags, unsigned int hctx_idx) | |
288 | { | |
6d2809d5 | 289 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 290 | struct request *rq; |
6d2809d5 | 291 | unsigned int cpu; |
1f5bd336 ML |
292 | int ret; |
293 | ||
294 | /* | |
295 | * If the tag allocator sleeps we could get an allocation for a | |
296 | * different hardware context. No need to complicate the low level | |
297 | * allocator for this for the rare use case of a command tied to | |
298 | * a specific queue. | |
299 | */ | |
300 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
301 | return ERR_PTR(-EINVAL); | |
302 | ||
303 | if (hctx_idx >= q->nr_hw_queues) | |
304 | return ERR_PTR(-EIO); | |
305 | ||
306 | ret = blk_queue_enter(q, true); | |
307 | if (ret) | |
308 | return ERR_PTR(ret); | |
309 | ||
c8712c6a CH |
310 | /* |
311 | * Check if the hardware context is actually mapped to anything. | |
312 | * If not tell the caller that it should skip this queue. | |
313 | */ | |
6d2809d5 OS |
314 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
315 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
316 | blk_queue_exit(q); | |
317 | return ERR_PTR(-EXDEV); | |
c8712c6a | 318 | } |
6d2809d5 OS |
319 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
320 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 321 | |
6d2809d5 | 322 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
c8712c6a | 323 | |
6d2809d5 | 324 | blk_mq_put_ctx(alloc_data.ctx); |
c8712c6a | 325 | blk_queue_exit(q); |
6d2809d5 OS |
326 | |
327 | if (!rq) | |
328 | return ERR_PTR(-EWOULDBLOCK); | |
329 | ||
330 | return rq; | |
1f5bd336 ML |
331 | } |
332 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
333 | ||
bd166ef1 JA |
334 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
335 | struct request *rq) | |
320ae51f | 336 | { |
bd166ef1 | 337 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
338 | struct request_queue *q = rq->q; |
339 | ||
e8064021 | 340 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 341 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
342 | |
343 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 344 | rq->rq_flags = 0; |
0d2602ca | 345 | |
af76e555 | 346 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 347 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
348 | if (rq->tag != -1) |
349 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
350 | if (sched_tag != -1) | |
351 | blk_mq_sched_completed_request(hctx, rq); | |
50e1dab8 | 352 | blk_mq_sched_restart_queues(hctx); |
3ef28e83 | 353 | blk_queue_exit(q); |
320ae51f JA |
354 | } |
355 | ||
bd166ef1 | 356 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 357 | struct request *rq) |
320ae51f JA |
358 | { |
359 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
360 | |
361 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
362 | __blk_mq_finish_request(hctx, ctx, rq); |
363 | } | |
364 | ||
365 | void blk_mq_finish_request(struct request *rq) | |
366 | { | |
367 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 368 | } |
7c7f2f2b JA |
369 | |
370 | void blk_mq_free_request(struct request *rq) | |
371 | { | |
bd166ef1 | 372 | blk_mq_sched_put_request(rq); |
320ae51f | 373 | } |
1a3b595a | 374 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 375 | |
c8a446ad | 376 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 377 | { |
0d11e6ac ML |
378 | blk_account_io_done(rq); |
379 | ||
91b63639 | 380 | if (rq->end_io) { |
87760e5e | 381 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 382 | rq->end_io(rq, error); |
91b63639 CH |
383 | } else { |
384 | if (unlikely(blk_bidi_rq(rq))) | |
385 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 386 | blk_mq_free_request(rq); |
91b63639 | 387 | } |
320ae51f | 388 | } |
c8a446ad | 389 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 390 | |
c8a446ad | 391 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
392 | { |
393 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
394 | BUG(); | |
c8a446ad | 395 | __blk_mq_end_request(rq, error); |
63151a44 | 396 | } |
c8a446ad | 397 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 398 | |
30a91cb4 | 399 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 400 | { |
3d6efbf6 | 401 | struct request *rq = data; |
320ae51f | 402 | |
30a91cb4 | 403 | rq->q->softirq_done_fn(rq); |
320ae51f | 404 | } |
320ae51f | 405 | |
ed851860 | 406 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
407 | { |
408 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 409 | bool shared = false; |
320ae51f JA |
410 | int cpu; |
411 | ||
38535201 | 412 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
413 | rq->q->softirq_done_fn(rq); |
414 | return; | |
415 | } | |
320ae51f JA |
416 | |
417 | cpu = get_cpu(); | |
38535201 CH |
418 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
419 | shared = cpus_share_cache(cpu, ctx->cpu); | |
420 | ||
421 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 422 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
423 | rq->csd.info = rq; |
424 | rq->csd.flags = 0; | |
c46fff2a | 425 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 426 | } else { |
30a91cb4 | 427 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 428 | } |
320ae51f JA |
429 | put_cpu(); |
430 | } | |
30a91cb4 | 431 | |
cf43e6be JA |
432 | static void blk_mq_stat_add(struct request *rq) |
433 | { | |
434 | if (rq->rq_flags & RQF_STATS) { | |
435 | /* | |
436 | * We could rq->mq_ctx here, but there's less of a risk | |
437 | * of races if we have the completion event add the stats | |
438 | * to the local software queue. | |
439 | */ | |
440 | struct blk_mq_ctx *ctx; | |
441 | ||
442 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
443 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
444 | } | |
445 | } | |
446 | ||
1fa8cc52 | 447 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
448 | { |
449 | struct request_queue *q = rq->q; | |
450 | ||
cf43e6be JA |
451 | blk_mq_stat_add(rq); |
452 | ||
ed851860 | 453 | if (!q->softirq_done_fn) |
c8a446ad | 454 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
455 | else |
456 | blk_mq_ipi_complete_request(rq); | |
457 | } | |
458 | ||
30a91cb4 CH |
459 | /** |
460 | * blk_mq_complete_request - end I/O on a request | |
461 | * @rq: the request being processed | |
462 | * | |
463 | * Description: | |
464 | * Ends all I/O on a request. It does not handle partial completions. | |
465 | * The actual completion happens out-of-order, through a IPI handler. | |
466 | **/ | |
f4829a9b | 467 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 468 | { |
95f09684 JA |
469 | struct request_queue *q = rq->q; |
470 | ||
471 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 472 | return; |
f4829a9b CH |
473 | if (!blk_mark_rq_complete(rq)) { |
474 | rq->errors = error; | |
ed851860 | 475 | __blk_mq_complete_request(rq); |
f4829a9b | 476 | } |
30a91cb4 CH |
477 | } |
478 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 479 | |
973c0191 KB |
480 | int blk_mq_request_started(struct request *rq) |
481 | { | |
482 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
483 | } | |
484 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
485 | ||
e2490073 | 486 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
487 | { |
488 | struct request_queue *q = rq->q; | |
489 | ||
bd166ef1 JA |
490 | blk_mq_sched_started_request(rq); |
491 | ||
320ae51f JA |
492 | trace_block_rq_issue(q, rq); |
493 | ||
cf43e6be JA |
494 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
495 | blk_stat_set_issue_time(&rq->issue_stat); | |
496 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 497 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
498 | } |
499 | ||
2b8393b4 | 500 | blk_add_timer(rq); |
87ee7b11 | 501 | |
538b7534 JA |
502 | /* |
503 | * Ensure that ->deadline is visible before set the started | |
504 | * flag and clear the completed flag. | |
505 | */ | |
506 | smp_mb__before_atomic(); | |
507 | ||
87ee7b11 JA |
508 | /* |
509 | * Mark us as started and clear complete. Complete might have been | |
510 | * set if requeue raced with timeout, which then marked it as | |
511 | * complete. So be sure to clear complete again when we start | |
512 | * the request, otherwise we'll ignore the completion event. | |
513 | */ | |
4b570521 JA |
514 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
515 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
516 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
517 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
518 | |
519 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
520 | /* | |
521 | * Make sure space for the drain appears. We know we can do | |
522 | * this because max_hw_segments has been adjusted to be one | |
523 | * fewer than the device can handle. | |
524 | */ | |
525 | rq->nr_phys_segments++; | |
526 | } | |
320ae51f | 527 | } |
e2490073 | 528 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 529 | |
ed0791b2 | 530 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
531 | { |
532 | struct request_queue *q = rq->q; | |
533 | ||
534 | trace_block_rq_requeue(q, rq); | |
87760e5e | 535 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 536 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 537 | |
e2490073 CH |
538 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
539 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
540 | rq->nr_phys_segments--; | |
541 | } | |
320ae51f JA |
542 | } |
543 | ||
2b053aca | 544 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 545 | { |
ed0791b2 | 546 | __blk_mq_requeue_request(rq); |
ed0791b2 | 547 | |
ed0791b2 | 548 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 549 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
550 | } |
551 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
552 | ||
6fca6a61 CH |
553 | static void blk_mq_requeue_work(struct work_struct *work) |
554 | { | |
555 | struct request_queue *q = | |
2849450a | 556 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
557 | LIST_HEAD(rq_list); |
558 | struct request *rq, *next; | |
559 | unsigned long flags; | |
560 | ||
561 | spin_lock_irqsave(&q->requeue_lock, flags); | |
562 | list_splice_init(&q->requeue_list, &rq_list); | |
563 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
564 | ||
565 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 566 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
567 | continue; |
568 | ||
e8064021 | 569 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 570 | list_del_init(&rq->queuelist); |
bd6737f1 | 571 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
572 | } |
573 | ||
574 | while (!list_empty(&rq_list)) { | |
575 | rq = list_entry(rq_list.next, struct request, queuelist); | |
576 | list_del_init(&rq->queuelist); | |
bd6737f1 | 577 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
578 | } |
579 | ||
52d7f1b5 | 580 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
581 | } |
582 | ||
2b053aca BVA |
583 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
584 | bool kick_requeue_list) | |
6fca6a61 CH |
585 | { |
586 | struct request_queue *q = rq->q; | |
587 | unsigned long flags; | |
588 | ||
589 | /* | |
590 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
591 | * request head insertation from the workqueue. | |
592 | */ | |
e8064021 | 593 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
594 | |
595 | spin_lock_irqsave(&q->requeue_lock, flags); | |
596 | if (at_head) { | |
e8064021 | 597 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
598 | list_add(&rq->queuelist, &q->requeue_list); |
599 | } else { | |
600 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
601 | } | |
602 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
603 | |
604 | if (kick_requeue_list) | |
605 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
606 | } |
607 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
608 | ||
609 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
610 | { | |
2849450a | 611 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
612 | } |
613 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
614 | ||
2849450a MS |
615 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
616 | unsigned long msecs) | |
617 | { | |
618 | kblockd_schedule_delayed_work(&q->requeue_work, | |
619 | msecs_to_jiffies(msecs)); | |
620 | } | |
621 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
622 | ||
1885b24d JA |
623 | void blk_mq_abort_requeue_list(struct request_queue *q) |
624 | { | |
625 | unsigned long flags; | |
626 | LIST_HEAD(rq_list); | |
627 | ||
628 | spin_lock_irqsave(&q->requeue_lock, flags); | |
629 | list_splice_init(&q->requeue_list, &rq_list); | |
630 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
631 | ||
632 | while (!list_empty(&rq_list)) { | |
633 | struct request *rq; | |
634 | ||
635 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
636 | list_del_init(&rq->queuelist); | |
637 | rq->errors = -EIO; | |
638 | blk_mq_end_request(rq, rq->errors); | |
639 | } | |
640 | } | |
641 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
642 | ||
0e62f51f JA |
643 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
644 | { | |
88c7b2b7 JA |
645 | if (tag < tags->nr_tags) { |
646 | prefetch(tags->rqs[tag]); | |
4ee86bab | 647 | return tags->rqs[tag]; |
88c7b2b7 | 648 | } |
4ee86bab HR |
649 | |
650 | return NULL; | |
24d2f903 CH |
651 | } |
652 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
653 | ||
320ae51f | 654 | struct blk_mq_timeout_data { |
46f92d42 CH |
655 | unsigned long next; |
656 | unsigned int next_set; | |
320ae51f JA |
657 | }; |
658 | ||
90415837 | 659 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 660 | { |
f8a5b122 | 661 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 662 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
663 | |
664 | /* | |
665 | * We know that complete is set at this point. If STARTED isn't set | |
666 | * anymore, then the request isn't active and the "timeout" should | |
667 | * just be ignored. This can happen due to the bitflag ordering. | |
668 | * Timeout first checks if STARTED is set, and if it is, assumes | |
669 | * the request is active. But if we race with completion, then | |
670 | * we both flags will get cleared. So check here again, and ignore | |
671 | * a timeout event with a request that isn't active. | |
672 | */ | |
46f92d42 CH |
673 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
674 | return; | |
87ee7b11 | 675 | |
46f92d42 | 676 | if (ops->timeout) |
0152fb6b | 677 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
678 | |
679 | switch (ret) { | |
680 | case BLK_EH_HANDLED: | |
681 | __blk_mq_complete_request(req); | |
682 | break; | |
683 | case BLK_EH_RESET_TIMER: | |
684 | blk_add_timer(req); | |
685 | blk_clear_rq_complete(req); | |
686 | break; | |
687 | case BLK_EH_NOT_HANDLED: | |
688 | break; | |
689 | default: | |
690 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
691 | break; | |
692 | } | |
87ee7b11 | 693 | } |
5b3f25fc | 694 | |
81481eb4 CH |
695 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
696 | struct request *rq, void *priv, bool reserved) | |
697 | { | |
698 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 699 | |
eb130dbf KB |
700 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
701 | /* | |
702 | * If a request wasn't started before the queue was | |
703 | * marked dying, kill it here or it'll go unnoticed. | |
704 | */ | |
a59e0f57 KB |
705 | if (unlikely(blk_queue_dying(rq->q))) { |
706 | rq->errors = -EIO; | |
707 | blk_mq_end_request(rq, rq->errors); | |
708 | } | |
46f92d42 | 709 | return; |
eb130dbf | 710 | } |
87ee7b11 | 711 | |
46f92d42 CH |
712 | if (time_after_eq(jiffies, rq->deadline)) { |
713 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 714 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
715 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
716 | data->next = rq->deadline; | |
717 | data->next_set = 1; | |
718 | } | |
87ee7b11 JA |
719 | } |
720 | ||
287922eb | 721 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 722 | { |
287922eb CH |
723 | struct request_queue *q = |
724 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
725 | struct blk_mq_timeout_data data = { |
726 | .next = 0, | |
727 | .next_set = 0, | |
728 | }; | |
81481eb4 | 729 | int i; |
320ae51f | 730 | |
71f79fb3 GKB |
731 | /* A deadlock might occur if a request is stuck requiring a |
732 | * timeout at the same time a queue freeze is waiting | |
733 | * completion, since the timeout code would not be able to | |
734 | * acquire the queue reference here. | |
735 | * | |
736 | * That's why we don't use blk_queue_enter here; instead, we use | |
737 | * percpu_ref_tryget directly, because we need to be able to | |
738 | * obtain a reference even in the short window between the queue | |
739 | * starting to freeze, by dropping the first reference in | |
740 | * blk_mq_freeze_queue_start, and the moment the last request is | |
741 | * consumed, marked by the instant q_usage_counter reaches | |
742 | * zero. | |
743 | */ | |
744 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
745 | return; |
746 | ||
0bf6cd5b | 747 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 748 | |
81481eb4 CH |
749 | if (data.next_set) { |
750 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
751 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 752 | } else { |
0bf6cd5b CH |
753 | struct blk_mq_hw_ctx *hctx; |
754 | ||
f054b56c ML |
755 | queue_for_each_hw_ctx(q, hctx, i) { |
756 | /* the hctx may be unmapped, so check it here */ | |
757 | if (blk_mq_hw_queue_mapped(hctx)) | |
758 | blk_mq_tag_idle(hctx); | |
759 | } | |
0d2602ca | 760 | } |
287922eb | 761 | blk_queue_exit(q); |
320ae51f JA |
762 | } |
763 | ||
764 | /* | |
765 | * Reverse check our software queue for entries that we could potentially | |
766 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
767 | * too much time checking for merges. | |
768 | */ | |
769 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
770 | struct blk_mq_ctx *ctx, struct bio *bio) | |
771 | { | |
772 | struct request *rq; | |
773 | int checked = 8; | |
774 | ||
775 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 776 | bool merged = false; |
320ae51f JA |
777 | |
778 | if (!checked--) | |
779 | break; | |
780 | ||
781 | if (!blk_rq_merge_ok(rq, bio)) | |
782 | continue; | |
783 | ||
34fe7c05 CH |
784 | switch (blk_try_merge(rq, bio)) { |
785 | case ELEVATOR_BACK_MERGE: | |
786 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
787 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 788 | break; |
34fe7c05 CH |
789 | case ELEVATOR_FRONT_MERGE: |
790 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
791 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 792 | break; |
1e739730 CH |
793 | case ELEVATOR_DISCARD_MERGE: |
794 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 795 | break; |
34fe7c05 CH |
796 | default: |
797 | continue; | |
320ae51f | 798 | } |
34fe7c05 CH |
799 | |
800 | if (merged) | |
801 | ctx->rq_merged++; | |
802 | return merged; | |
320ae51f JA |
803 | } |
804 | ||
805 | return false; | |
806 | } | |
807 | ||
88459642 OS |
808 | struct flush_busy_ctx_data { |
809 | struct blk_mq_hw_ctx *hctx; | |
810 | struct list_head *list; | |
811 | }; | |
812 | ||
813 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
814 | { | |
815 | struct flush_busy_ctx_data *flush_data = data; | |
816 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
817 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
818 | ||
819 | sbitmap_clear_bit(sb, bitnr); | |
820 | spin_lock(&ctx->lock); | |
821 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
822 | spin_unlock(&ctx->lock); | |
823 | return true; | |
824 | } | |
825 | ||
1429d7c9 JA |
826 | /* |
827 | * Process software queues that have been marked busy, splicing them | |
828 | * to the for-dispatch | |
829 | */ | |
2c3ad667 | 830 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 831 | { |
88459642 OS |
832 | struct flush_busy_ctx_data data = { |
833 | .hctx = hctx, | |
834 | .list = list, | |
835 | }; | |
1429d7c9 | 836 | |
88459642 | 837 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 838 | } |
2c3ad667 | 839 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 840 | |
703fd1c0 JA |
841 | static inline unsigned int queued_to_index(unsigned int queued) |
842 | { | |
843 | if (!queued) | |
844 | return 0; | |
1429d7c9 | 845 | |
703fd1c0 | 846 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
847 | } |
848 | ||
bd6737f1 JA |
849 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
850 | bool wait) | |
bd166ef1 JA |
851 | { |
852 | struct blk_mq_alloc_data data = { | |
853 | .q = rq->q, | |
bd166ef1 JA |
854 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
855 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
856 | }; | |
857 | ||
bd166ef1 JA |
858 | if (rq->tag != -1) { |
859 | done: | |
860 | if (hctx) | |
861 | *hctx = data.hctx; | |
862 | return true; | |
863 | } | |
864 | ||
415b806d SG |
865 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
866 | data.flags |= BLK_MQ_REQ_RESERVED; | |
867 | ||
bd166ef1 JA |
868 | rq->tag = blk_mq_get_tag(&data); |
869 | if (rq->tag >= 0) { | |
200e86b3 JA |
870 | if (blk_mq_tag_busy(data.hctx)) { |
871 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
872 | atomic_inc(&data.hctx->nr_active); | |
873 | } | |
bd166ef1 JA |
874 | data.hctx->tags->rqs[rq->tag] = rq; |
875 | goto done; | |
876 | } | |
877 | ||
878 | return false; | |
879 | } | |
880 | ||
113285b4 JA |
881 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
882 | struct request *rq) | |
99cf1dc5 | 883 | { |
99cf1dc5 JA |
884 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
885 | rq->tag = -1; | |
886 | ||
887 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
888 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
889 | atomic_dec(&hctx->nr_active); | |
890 | } | |
891 | } | |
892 | ||
113285b4 JA |
893 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
894 | struct request *rq) | |
895 | { | |
896 | if (rq->tag == -1 || rq->internal_tag == -1) | |
897 | return; | |
898 | ||
899 | __blk_mq_put_driver_tag(hctx, rq); | |
900 | } | |
901 | ||
902 | static void blk_mq_put_driver_tag(struct request *rq) | |
903 | { | |
904 | struct blk_mq_hw_ctx *hctx; | |
905 | ||
906 | if (rq->tag == -1 || rq->internal_tag == -1) | |
907 | return; | |
908 | ||
909 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
910 | __blk_mq_put_driver_tag(hctx, rq); | |
911 | } | |
912 | ||
bd166ef1 JA |
913 | /* |
914 | * If we fail getting a driver tag because all the driver tags are already | |
915 | * assigned and on the dispatch list, BUT the first entry does not have a | |
916 | * tag, then we could deadlock. For that case, move entries with assigned | |
917 | * driver tags to the front, leaving the set of tagged requests in the | |
918 | * same order, and the untagged set in the same order. | |
919 | */ | |
920 | static bool reorder_tags_to_front(struct list_head *list) | |
921 | { | |
922 | struct request *rq, *tmp, *first = NULL; | |
923 | ||
924 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
925 | if (rq == first) | |
926 | break; | |
927 | if (rq->tag != -1) { | |
928 | list_move(&rq->queuelist, list); | |
929 | if (!first) | |
930 | first = rq; | |
931 | } | |
932 | } | |
933 | ||
934 | return first != NULL; | |
935 | } | |
936 | ||
da55f2cc OS |
937 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
938 | void *key) | |
939 | { | |
940 | struct blk_mq_hw_ctx *hctx; | |
941 | ||
942 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
943 | ||
944 | list_del(&wait->task_list); | |
945 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
946 | blk_mq_run_hw_queue(hctx, true); | |
947 | return 1; | |
948 | } | |
949 | ||
950 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
951 | { | |
952 | struct sbq_wait_state *ws; | |
953 | ||
954 | /* | |
955 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
956 | * The thread which wins the race to grab this bit adds the hardware | |
957 | * queue to the wait queue. | |
958 | */ | |
959 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
960 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
961 | return false; | |
962 | ||
963 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
964 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
965 | ||
966 | /* | |
967 | * As soon as this returns, it's no longer safe to fiddle with | |
968 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
969 | * and unlock the bit. | |
970 | */ | |
971 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
972 | return true; | |
973 | } | |
974 | ||
f04c3df3 | 975 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
976 | { |
977 | struct request_queue *q = hctx->queue; | |
320ae51f | 978 | struct request *rq; |
74c45052 JA |
979 | LIST_HEAD(driver_list); |
980 | struct list_head *dptr; | |
f04c3df3 | 981 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 982 | |
74c45052 JA |
983 | /* |
984 | * Start off with dptr being NULL, so we start the first request | |
985 | * immediately, even if we have more pending. | |
986 | */ | |
987 | dptr = NULL; | |
988 | ||
320ae51f JA |
989 | /* |
990 | * Now process all the entries, sending them to the driver. | |
991 | */ | |
1429d7c9 | 992 | queued = 0; |
f04c3df3 | 993 | while (!list_empty(list)) { |
74c45052 | 994 | struct blk_mq_queue_data bd; |
320ae51f | 995 | |
f04c3df3 | 996 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
997 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
998 | if (!queued && reorder_tags_to_front(list)) | |
999 | continue; | |
3c782d67 JA |
1000 | |
1001 | /* | |
da55f2cc OS |
1002 | * The initial allocation attempt failed, so we need to |
1003 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1004 | */ |
da55f2cc OS |
1005 | if (blk_mq_dispatch_wait_add(hctx)) { |
1006 | /* | |
1007 | * It's possible that a tag was freed in the | |
1008 | * window between the allocation failure and | |
1009 | * adding the hardware queue to the wait queue. | |
1010 | */ | |
1011 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
1012 | break; | |
1013 | } else { | |
3c782d67 | 1014 | break; |
da55f2cc | 1015 | } |
bd166ef1 | 1016 | } |
da55f2cc | 1017 | |
320ae51f | 1018 | list_del_init(&rq->queuelist); |
320ae51f | 1019 | |
74c45052 JA |
1020 | bd.rq = rq; |
1021 | bd.list = dptr; | |
113285b4 JA |
1022 | |
1023 | /* | |
1024 | * Flag last if we have no more requests, or if we have more | |
1025 | * but can't assign a driver tag to it. | |
1026 | */ | |
1027 | if (list_empty(list)) | |
1028 | bd.last = true; | |
1029 | else { | |
1030 | struct request *nxt; | |
1031 | ||
1032 | nxt = list_first_entry(list, struct request, queuelist); | |
1033 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1034 | } | |
74c45052 JA |
1035 | |
1036 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
1037 | switch (ret) { |
1038 | case BLK_MQ_RQ_QUEUE_OK: | |
1039 | queued++; | |
52b9c330 | 1040 | break; |
320ae51f | 1041 | case BLK_MQ_RQ_QUEUE_BUSY: |
113285b4 | 1042 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1043 | list_add(&rq->queuelist, list); |
ed0791b2 | 1044 | __blk_mq_requeue_request(rq); |
320ae51f JA |
1045 | break; |
1046 | default: | |
1047 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 1048 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 1049 | rq->errors = -EIO; |
c8a446ad | 1050 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
1051 | break; |
1052 | } | |
1053 | ||
1054 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1055 | break; | |
74c45052 JA |
1056 | |
1057 | /* | |
1058 | * We've done the first request. If we have more than 1 | |
1059 | * left in the list, set dptr to defer issue. | |
1060 | */ | |
f04c3df3 | 1061 | if (!dptr && list->next != list->prev) |
74c45052 | 1062 | dptr = &driver_list; |
320ae51f JA |
1063 | } |
1064 | ||
703fd1c0 | 1065 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1066 | |
1067 | /* | |
1068 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1069 | * that is where we will continue on next queue run. | |
1070 | */ | |
f04c3df3 | 1071 | if (!list_empty(list)) { |
113285b4 JA |
1072 | /* |
1073 | * If we got a driver tag for the next request already, | |
1074 | * free it again. | |
1075 | */ | |
1076 | rq = list_first_entry(list, struct request, queuelist); | |
1077 | blk_mq_put_driver_tag(rq); | |
1078 | ||
320ae51f | 1079 | spin_lock(&hctx->lock); |
c13660a0 | 1080 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1081 | spin_unlock(&hctx->lock); |
f04c3df3 | 1082 | |
9ba52e58 SL |
1083 | /* |
1084 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
1085 | * it's possible the queue is stopped and restarted again | |
1086 | * before this. Queue restart will dispatch requests. And since | |
1087 | * requests in rq_list aren't added into hctx->dispatch yet, | |
1088 | * the requests in rq_list might get lost. | |
1089 | * | |
1090 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 | 1091 | * |
da55f2cc OS |
1092 | * If RESTART or TAG_WAITING is set, then let completion restart |
1093 | * the queue instead of potentially looping here. | |
bd166ef1 | 1094 | */ |
da55f2cc OS |
1095 | if (!blk_mq_sched_needs_restart(hctx) && |
1096 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1097 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1098 | } |
f04c3df3 | 1099 | |
2aa0f21d | 1100 | return queued != 0; |
f04c3df3 JA |
1101 | } |
1102 | ||
6a83e74d BVA |
1103 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1104 | { | |
1105 | int srcu_idx; | |
1106 | ||
1107 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1108 | cpu_online(hctx->next_cpu)); | |
1109 | ||
1110 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1111 | rcu_read_lock(); | |
bd166ef1 | 1112 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1113 | rcu_read_unlock(); |
1114 | } else { | |
1115 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1116 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1117 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1118 | } | |
1119 | } | |
1120 | ||
506e931f JA |
1121 | /* |
1122 | * It'd be great if the workqueue API had a way to pass | |
1123 | * in a mask and had some smarts for more clever placement. | |
1124 | * For now we just round-robin here, switching for every | |
1125 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1126 | */ | |
1127 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1128 | { | |
b657d7e6 CH |
1129 | if (hctx->queue->nr_hw_queues == 1) |
1130 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1131 | |
1132 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1133 | int next_cpu; |
506e931f JA |
1134 | |
1135 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1136 | if (next_cpu >= nr_cpu_ids) | |
1137 | next_cpu = cpumask_first(hctx->cpumask); | |
1138 | ||
1139 | hctx->next_cpu = next_cpu; | |
1140 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1141 | } | |
1142 | ||
b657d7e6 | 1143 | return hctx->next_cpu; |
506e931f JA |
1144 | } |
1145 | ||
320ae51f JA |
1146 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1147 | { | |
5d1b25c1 BVA |
1148 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1149 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1150 | return; |
1151 | ||
1b792f2f | 1152 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1153 | int cpu = get_cpu(); |
1154 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1155 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1156 | put_cpu(); |
398205b8 PB |
1157 | return; |
1158 | } | |
e4043dcf | 1159 | |
2a90d4aa | 1160 | put_cpu(); |
e4043dcf | 1161 | } |
398205b8 | 1162 | |
27489a3c | 1163 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1164 | } |
1165 | ||
b94ec296 | 1166 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1167 | { |
1168 | struct blk_mq_hw_ctx *hctx; | |
1169 | int i; | |
1170 | ||
1171 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1172 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1173 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1174 | continue; |
1175 | ||
b94ec296 | 1176 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1177 | } |
1178 | } | |
b94ec296 | 1179 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1180 | |
fd001443 BVA |
1181 | /** |
1182 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1183 | * @q: request queue. | |
1184 | * | |
1185 | * The caller is responsible for serializing this function against | |
1186 | * blk_mq_{start,stop}_hw_queue(). | |
1187 | */ | |
1188 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1189 | { | |
1190 | struct blk_mq_hw_ctx *hctx; | |
1191 | int i; | |
1192 | ||
1193 | queue_for_each_hw_ctx(q, hctx, i) | |
1194 | if (blk_mq_hctx_stopped(hctx)) | |
1195 | return true; | |
1196 | ||
1197 | return false; | |
1198 | } | |
1199 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1200 | ||
320ae51f JA |
1201 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1202 | { | |
27489a3c | 1203 | cancel_work(&hctx->run_work); |
70f4db63 | 1204 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1205 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1206 | } | |
1207 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1208 | ||
280d45f6 CH |
1209 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1210 | { | |
1211 | struct blk_mq_hw_ctx *hctx; | |
1212 | int i; | |
1213 | ||
1214 | queue_for_each_hw_ctx(q, hctx, i) | |
1215 | blk_mq_stop_hw_queue(hctx); | |
1216 | } | |
1217 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1218 | ||
320ae51f JA |
1219 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1220 | { | |
1221 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1222 | |
0ffbce80 | 1223 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1224 | } |
1225 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1226 | ||
2f268556 CH |
1227 | void blk_mq_start_hw_queues(struct request_queue *q) |
1228 | { | |
1229 | struct blk_mq_hw_ctx *hctx; | |
1230 | int i; | |
1231 | ||
1232 | queue_for_each_hw_ctx(q, hctx, i) | |
1233 | blk_mq_start_hw_queue(hctx); | |
1234 | } | |
1235 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1236 | ||
ae911c5e JA |
1237 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1238 | { | |
1239 | if (!blk_mq_hctx_stopped(hctx)) | |
1240 | return; | |
1241 | ||
1242 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1243 | blk_mq_run_hw_queue(hctx, async); | |
1244 | } | |
1245 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1246 | ||
1b4a3258 | 1247 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1248 | { |
1249 | struct blk_mq_hw_ctx *hctx; | |
1250 | int i; | |
1251 | ||
ae911c5e JA |
1252 | queue_for_each_hw_ctx(q, hctx, i) |
1253 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1254 | } |
1255 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1256 | ||
70f4db63 | 1257 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1258 | { |
1259 | struct blk_mq_hw_ctx *hctx; | |
1260 | ||
27489a3c | 1261 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1262 | |
320ae51f JA |
1263 | __blk_mq_run_hw_queue(hctx); |
1264 | } | |
1265 | ||
70f4db63 CH |
1266 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1267 | { | |
1268 | struct blk_mq_hw_ctx *hctx; | |
1269 | ||
1270 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1271 | ||
1272 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1273 | __blk_mq_run_hw_queue(hctx); | |
1274 | } | |
1275 | ||
1276 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1277 | { | |
19c66e59 ML |
1278 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1279 | return; | |
70f4db63 | 1280 | |
7e79dadc | 1281 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1282 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1283 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1284 | } |
1285 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1286 | ||
cfd0c552 | 1287 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1288 | struct request *rq, |
1289 | bool at_head) | |
320ae51f | 1290 | { |
e57690fe JA |
1291 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1292 | ||
01b983c9 JA |
1293 | trace_block_rq_insert(hctx->queue, rq); |
1294 | ||
72a0a36e CH |
1295 | if (at_head) |
1296 | list_add(&rq->queuelist, &ctx->rq_list); | |
1297 | else | |
1298 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1299 | } |
4bb659b1 | 1300 | |
2c3ad667 JA |
1301 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1302 | bool at_head) | |
cfd0c552 ML |
1303 | { |
1304 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1305 | ||
e57690fe | 1306 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1307 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1308 | } |
1309 | ||
bd166ef1 JA |
1310 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1311 | struct list_head *list) | |
320ae51f JA |
1312 | |
1313 | { | |
320ae51f JA |
1314 | /* |
1315 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1316 | * offline now | |
1317 | */ | |
1318 | spin_lock(&ctx->lock); | |
1319 | while (!list_empty(list)) { | |
1320 | struct request *rq; | |
1321 | ||
1322 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1323 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1324 | list_del_init(&rq->queuelist); |
e57690fe | 1325 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1326 | } |
cfd0c552 | 1327 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1328 | spin_unlock(&ctx->lock); |
320ae51f JA |
1329 | } |
1330 | ||
1331 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1332 | { | |
1333 | struct request *rqa = container_of(a, struct request, queuelist); | |
1334 | struct request *rqb = container_of(b, struct request, queuelist); | |
1335 | ||
1336 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1337 | (rqa->mq_ctx == rqb->mq_ctx && | |
1338 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1339 | } | |
1340 | ||
1341 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1342 | { | |
1343 | struct blk_mq_ctx *this_ctx; | |
1344 | struct request_queue *this_q; | |
1345 | struct request *rq; | |
1346 | LIST_HEAD(list); | |
1347 | LIST_HEAD(ctx_list); | |
1348 | unsigned int depth; | |
1349 | ||
1350 | list_splice_init(&plug->mq_list, &list); | |
1351 | ||
1352 | list_sort(NULL, &list, plug_ctx_cmp); | |
1353 | ||
1354 | this_q = NULL; | |
1355 | this_ctx = NULL; | |
1356 | depth = 0; | |
1357 | ||
1358 | while (!list_empty(&list)) { | |
1359 | rq = list_entry_rq(list.next); | |
1360 | list_del_init(&rq->queuelist); | |
1361 | BUG_ON(!rq->q); | |
1362 | if (rq->mq_ctx != this_ctx) { | |
1363 | if (this_ctx) { | |
bd166ef1 JA |
1364 | trace_block_unplug(this_q, depth, from_schedule); |
1365 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1366 | &ctx_list, | |
1367 | from_schedule); | |
320ae51f JA |
1368 | } |
1369 | ||
1370 | this_ctx = rq->mq_ctx; | |
1371 | this_q = rq->q; | |
1372 | depth = 0; | |
1373 | } | |
1374 | ||
1375 | depth++; | |
1376 | list_add_tail(&rq->queuelist, &ctx_list); | |
1377 | } | |
1378 | ||
1379 | /* | |
1380 | * If 'this_ctx' is set, we know we have entries to complete | |
1381 | * on 'ctx_list'. Do those. | |
1382 | */ | |
1383 | if (this_ctx) { | |
bd166ef1 JA |
1384 | trace_block_unplug(this_q, depth, from_schedule); |
1385 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1386 | from_schedule); | |
320ae51f JA |
1387 | } |
1388 | } | |
1389 | ||
1390 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1391 | { | |
1392 | init_request_from_bio(rq, bio); | |
4b570521 | 1393 | |
6e85eaf3 | 1394 | blk_account_io_start(rq, true); |
320ae51f JA |
1395 | } |
1396 | ||
274a5843 JA |
1397 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1398 | { | |
1399 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1400 | !blk_queue_nomerges(hctx->queue); | |
1401 | } | |
1402 | ||
07068d5b JA |
1403 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1404 | struct blk_mq_ctx *ctx, | |
1405 | struct request *rq, struct bio *bio) | |
320ae51f | 1406 | { |
e18378a6 | 1407 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1408 | blk_mq_bio_to_request(rq, bio); |
1409 | spin_lock(&ctx->lock); | |
1410 | insert_rq: | |
1411 | __blk_mq_insert_request(hctx, rq, false); | |
1412 | spin_unlock(&ctx->lock); | |
1413 | return false; | |
1414 | } else { | |
274a5843 JA |
1415 | struct request_queue *q = hctx->queue; |
1416 | ||
07068d5b JA |
1417 | spin_lock(&ctx->lock); |
1418 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1419 | blk_mq_bio_to_request(rq, bio); | |
1420 | goto insert_rq; | |
1421 | } | |
320ae51f | 1422 | |
07068d5b | 1423 | spin_unlock(&ctx->lock); |
bd166ef1 | 1424 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1425 | return true; |
14ec77f3 | 1426 | } |
07068d5b | 1427 | } |
14ec77f3 | 1428 | |
fd2d3326 JA |
1429 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1430 | { | |
bd166ef1 JA |
1431 | if (rq->tag != -1) |
1432 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1433 | ||
1434 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1435 | } |
1436 | ||
066a4a73 | 1437 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1438 | { |
f984df1f | 1439 | struct request_queue *q = rq->q; |
f984df1f SL |
1440 | struct blk_mq_queue_data bd = { |
1441 | .rq = rq, | |
1442 | .list = NULL, | |
1443 | .last = 1 | |
1444 | }; | |
bd166ef1 JA |
1445 | struct blk_mq_hw_ctx *hctx; |
1446 | blk_qc_t new_cookie; | |
1447 | int ret; | |
f984df1f | 1448 | |
bd166ef1 | 1449 | if (q->elevator) |
2253efc8 BVA |
1450 | goto insert; |
1451 | ||
bd166ef1 JA |
1452 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1453 | goto insert; | |
1454 | ||
1455 | new_cookie = request_to_qc_t(hctx, rq); | |
1456 | ||
f984df1f SL |
1457 | /* |
1458 | * For OK queue, we are done. For error, kill it. Any other | |
1459 | * error (busy), just add it to our list as we previously | |
1460 | * would have done | |
1461 | */ | |
1462 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1463 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1464 | *cookie = new_cookie; | |
2253efc8 | 1465 | return; |
7b371636 | 1466 | } |
f984df1f | 1467 | |
7b371636 JA |
1468 | __blk_mq_requeue_request(rq); |
1469 | ||
1470 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1471 | *cookie = BLK_QC_T_NONE; | |
1472 | rq->errors = -EIO; | |
1473 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1474 | return; |
f984df1f | 1475 | } |
7b371636 | 1476 | |
2253efc8 | 1477 | insert: |
bd6737f1 | 1478 | blk_mq_sched_insert_request(rq, false, true, true, false); |
f984df1f SL |
1479 | } |
1480 | ||
07068d5b JA |
1481 | /* |
1482 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1483 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1484 | * hardware for SYNC IO. | |
1485 | */ | |
dece1635 | 1486 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1487 | { |
ef295ecf | 1488 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1489 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1490 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1491 | struct request *rq; |
6a83e74d | 1492 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1493 | struct blk_plug *plug; |
5b3f341f | 1494 | struct request *same_queue_rq = NULL; |
7b371636 | 1495 | blk_qc_t cookie; |
87760e5e | 1496 | unsigned int wb_acct; |
07068d5b JA |
1497 | |
1498 | blk_queue_bounce(q, &bio); | |
1499 | ||
1500 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1501 | bio_io_error(bio); |
dece1635 | 1502 | return BLK_QC_T_NONE; |
07068d5b JA |
1503 | } |
1504 | ||
54efd50b KO |
1505 | blk_queue_split(q, &bio, q->bio_split); |
1506 | ||
87c279e6 OS |
1507 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1508 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1509 | return BLK_QC_T_NONE; | |
f984df1f | 1510 | |
bd166ef1 JA |
1511 | if (blk_mq_sched_bio_merge(q, bio)) |
1512 | return BLK_QC_T_NONE; | |
1513 | ||
87760e5e JA |
1514 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1515 | ||
bd166ef1 JA |
1516 | trace_block_getrq(q, bio, bio->bi_opf); |
1517 | ||
1518 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1519 | if (unlikely(!rq)) { |
1520 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1521 | return BLK_QC_T_NONE; |
87760e5e JA |
1522 | } |
1523 | ||
1524 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1525 | |
fd2d3326 | 1526 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1527 | |
1528 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1529 | if (q->elevator) |
1530 | goto elv_insert; | |
07068d5b JA |
1531 | blk_mq_bio_to_request(rq, bio); |
1532 | blk_insert_flush(rq); | |
0c2a6fe4 | 1533 | goto run_queue; |
07068d5b JA |
1534 | } |
1535 | ||
f984df1f | 1536 | plug = current->plug; |
e167dfb5 JA |
1537 | /* |
1538 | * If the driver supports defer issued based on 'last', then | |
1539 | * queue it up like normal since we can potentially save some | |
1540 | * CPU this way. | |
1541 | */ | |
f984df1f SL |
1542 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1543 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1544 | struct request *old_rq = NULL; | |
07068d5b JA |
1545 | |
1546 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1547 | |
1548 | /* | |
6a83e74d | 1549 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1550 | * Otherwise the existing request in the plug list will be |
1551 | * issued. So the plug list will have one request at most | |
07068d5b | 1552 | */ |
f984df1f | 1553 | if (plug) { |
5b3f341f SL |
1554 | /* |
1555 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1556 | * happens, same_queue_rq is invalid and plug list is |
1557 | * empty | |
1558 | */ | |
5b3f341f SL |
1559 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1560 | old_rq = same_queue_rq; | |
f984df1f | 1561 | list_del_init(&old_rq->queuelist); |
07068d5b | 1562 | } |
f984df1f SL |
1563 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1564 | } else /* is_sync */ | |
1565 | old_rq = rq; | |
1566 | blk_mq_put_ctx(data.ctx); | |
1567 | if (!old_rq) | |
7b371636 | 1568 | goto done; |
6a83e74d BVA |
1569 | |
1570 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1571 | rcu_read_lock(); | |
066a4a73 | 1572 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1573 | rcu_read_unlock(); |
1574 | } else { | |
1575 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1576 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1577 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1578 | } | |
7b371636 | 1579 | goto done; |
07068d5b JA |
1580 | } |
1581 | ||
bd166ef1 | 1582 | if (q->elevator) { |
0c2a6fe4 | 1583 | elv_insert: |
bd166ef1 JA |
1584 | blk_mq_put_ctx(data.ctx); |
1585 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1586 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1587 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1588 | goto done; |
1589 | } | |
07068d5b JA |
1590 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1591 | /* | |
1592 | * For a SYNC request, send it to the hardware immediately. For | |
1593 | * an ASYNC request, just ensure that we run it later on. The | |
1594 | * latter allows for merging opportunities and more efficient | |
1595 | * dispatching. | |
1596 | */ | |
0c2a6fe4 | 1597 | run_queue: |
07068d5b JA |
1598 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
1599 | } | |
07068d5b | 1600 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1601 | done: |
1602 | return cookie; | |
07068d5b JA |
1603 | } |
1604 | ||
1605 | /* | |
1606 | * Single hardware queue variant. This will attempt to use any per-process | |
1607 | * plug for merging and IO deferral. | |
1608 | */ | |
dece1635 | 1609 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1610 | { |
ef295ecf | 1611 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1612 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6c4438b JM |
1613 | struct blk_plug *plug; |
1614 | unsigned int request_count = 0; | |
5a797e00 | 1615 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1616 | struct request *rq; |
7b371636 | 1617 | blk_qc_t cookie; |
87760e5e | 1618 | unsigned int wb_acct; |
07068d5b | 1619 | |
07068d5b JA |
1620 | blk_queue_bounce(q, &bio); |
1621 | ||
1622 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1623 | bio_io_error(bio); |
dece1635 | 1624 | return BLK_QC_T_NONE; |
07068d5b JA |
1625 | } |
1626 | ||
54efd50b KO |
1627 | blk_queue_split(q, &bio, q->bio_split); |
1628 | ||
87c279e6 OS |
1629 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1630 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1631 | return BLK_QC_T_NONE; | |
1632 | } else | |
1633 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1634 | |
bd166ef1 JA |
1635 | if (blk_mq_sched_bio_merge(q, bio)) |
1636 | return BLK_QC_T_NONE; | |
1637 | ||
87760e5e JA |
1638 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1639 | ||
bd166ef1 JA |
1640 | trace_block_getrq(q, bio, bio->bi_opf); |
1641 | ||
1642 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1643 | if (unlikely(!rq)) { |
1644 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1645 | return BLK_QC_T_NONE; |
87760e5e JA |
1646 | } |
1647 | ||
1648 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1649 | |
fd2d3326 | 1650 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1651 | |
1652 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1653 | if (q->elevator) |
1654 | goto elv_insert; | |
320ae51f | 1655 | blk_mq_bio_to_request(rq, bio); |
320ae51f | 1656 | blk_insert_flush(rq); |
0c2a6fe4 | 1657 | goto run_queue; |
320ae51f JA |
1658 | } |
1659 | ||
1660 | /* | |
1661 | * A task plug currently exists. Since this is completely lockless, | |
1662 | * utilize that to temporarily store requests until the task is | |
1663 | * either done or scheduled away. | |
1664 | */ | |
e6c4438b JM |
1665 | plug = current->plug; |
1666 | if (plug) { | |
600271d9 SL |
1667 | struct request *last = NULL; |
1668 | ||
e6c4438b | 1669 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1670 | |
1671 | /* | |
1672 | * @request_count may become stale because of schedule | |
1673 | * out, so check the list again. | |
1674 | */ | |
1675 | if (list_empty(&plug->mq_list)) | |
1676 | request_count = 0; | |
676d0607 | 1677 | if (!request_count) |
e6c4438b | 1678 | trace_block_plug(q); |
600271d9 SL |
1679 | else |
1680 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1681 | |
1682 | blk_mq_put_ctx(data.ctx); | |
1683 | ||
600271d9 SL |
1684 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1685 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1686 | blk_flush_plug_list(plug, false); |
1687 | trace_block_plug(q); | |
320ae51f | 1688 | } |
b094f89c | 1689 | |
e6c4438b | 1690 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1691 | return cookie; |
320ae51f JA |
1692 | } |
1693 | ||
bd166ef1 | 1694 | if (q->elevator) { |
0c2a6fe4 | 1695 | elv_insert: |
bd166ef1 JA |
1696 | blk_mq_put_ctx(data.ctx); |
1697 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1698 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1699 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1700 | goto done; |
1701 | } | |
07068d5b JA |
1702 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1703 | /* | |
1704 | * For a SYNC request, send it to the hardware immediately. For | |
1705 | * an ASYNC request, just ensure that we run it later on. The | |
1706 | * latter allows for merging opportunities and more efficient | |
1707 | * dispatching. | |
1708 | */ | |
0c2a6fe4 | 1709 | run_queue: |
07068d5b | 1710 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
320ae51f JA |
1711 | } |
1712 | ||
07068d5b | 1713 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1714 | done: |
7b371636 | 1715 | return cookie; |
320ae51f JA |
1716 | } |
1717 | ||
cc71a6f4 JA |
1718 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1719 | unsigned int hctx_idx) | |
95363efd | 1720 | { |
e9b267d9 | 1721 | struct page *page; |
320ae51f | 1722 | |
24d2f903 | 1723 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1724 | int i; |
320ae51f | 1725 | |
24d2f903 | 1726 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1727 | struct request *rq = tags->static_rqs[i]; |
1728 | ||
1729 | if (!rq) | |
e9b267d9 | 1730 | continue; |
2af8cbe3 | 1731 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1732 | hctx_idx, i); |
2af8cbe3 | 1733 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1734 | } |
320ae51f | 1735 | } |
320ae51f | 1736 | |
24d2f903 CH |
1737 | while (!list_empty(&tags->page_list)) { |
1738 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1739 | list_del_init(&page->lru); |
f75782e4 CM |
1740 | /* |
1741 | * Remove kmemleak object previously allocated in | |
1742 | * blk_mq_init_rq_map(). | |
1743 | */ | |
1744 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1745 | __free_pages(page, page->private); |
1746 | } | |
cc71a6f4 | 1747 | } |
320ae51f | 1748 | |
cc71a6f4 JA |
1749 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1750 | { | |
24d2f903 | 1751 | kfree(tags->rqs); |
cc71a6f4 | 1752 | tags->rqs = NULL; |
2af8cbe3 JA |
1753 | kfree(tags->static_rqs); |
1754 | tags->static_rqs = NULL; | |
320ae51f | 1755 | |
24d2f903 | 1756 | blk_mq_free_tags(tags); |
320ae51f JA |
1757 | } |
1758 | ||
cc71a6f4 JA |
1759 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1760 | unsigned int hctx_idx, | |
1761 | unsigned int nr_tags, | |
1762 | unsigned int reserved_tags) | |
320ae51f | 1763 | { |
24d2f903 | 1764 | struct blk_mq_tags *tags; |
59f082e4 | 1765 | int node; |
320ae51f | 1766 | |
59f082e4 SL |
1767 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1768 | if (node == NUMA_NO_NODE) | |
1769 | node = set->numa_node; | |
1770 | ||
1771 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1772 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1773 | if (!tags) |
1774 | return NULL; | |
320ae51f | 1775 | |
cc71a6f4 | 1776 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1777 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1778 | node); |
24d2f903 CH |
1779 | if (!tags->rqs) { |
1780 | blk_mq_free_tags(tags); | |
1781 | return NULL; | |
1782 | } | |
320ae51f | 1783 | |
2af8cbe3 JA |
1784 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1785 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1786 | node); |
2af8cbe3 JA |
1787 | if (!tags->static_rqs) { |
1788 | kfree(tags->rqs); | |
1789 | blk_mq_free_tags(tags); | |
1790 | return NULL; | |
1791 | } | |
1792 | ||
cc71a6f4 JA |
1793 | return tags; |
1794 | } | |
1795 | ||
1796 | static size_t order_to_size(unsigned int order) | |
1797 | { | |
1798 | return (size_t)PAGE_SIZE << order; | |
1799 | } | |
1800 | ||
1801 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1802 | unsigned int hctx_idx, unsigned int depth) | |
1803 | { | |
1804 | unsigned int i, j, entries_per_page, max_order = 4; | |
1805 | size_t rq_size, left; | |
59f082e4 SL |
1806 | int node; |
1807 | ||
1808 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1809 | if (node == NUMA_NO_NODE) | |
1810 | node = set->numa_node; | |
cc71a6f4 JA |
1811 | |
1812 | INIT_LIST_HEAD(&tags->page_list); | |
1813 | ||
320ae51f JA |
1814 | /* |
1815 | * rq_size is the size of the request plus driver payload, rounded | |
1816 | * to the cacheline size | |
1817 | */ | |
24d2f903 | 1818 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1819 | cache_line_size()); |
cc71a6f4 | 1820 | left = rq_size * depth; |
320ae51f | 1821 | |
cc71a6f4 | 1822 | for (i = 0; i < depth; ) { |
320ae51f JA |
1823 | int this_order = max_order; |
1824 | struct page *page; | |
1825 | int to_do; | |
1826 | void *p; | |
1827 | ||
b3a834b1 | 1828 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1829 | this_order--; |
1830 | ||
1831 | do { | |
59f082e4 | 1832 | page = alloc_pages_node(node, |
36e1f3d1 | 1833 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1834 | this_order); |
320ae51f JA |
1835 | if (page) |
1836 | break; | |
1837 | if (!this_order--) | |
1838 | break; | |
1839 | if (order_to_size(this_order) < rq_size) | |
1840 | break; | |
1841 | } while (1); | |
1842 | ||
1843 | if (!page) | |
24d2f903 | 1844 | goto fail; |
320ae51f JA |
1845 | |
1846 | page->private = this_order; | |
24d2f903 | 1847 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1848 | |
1849 | p = page_address(page); | |
f75782e4 CM |
1850 | /* |
1851 | * Allow kmemleak to scan these pages as they contain pointers | |
1852 | * to additional allocations like via ops->init_request(). | |
1853 | */ | |
36e1f3d1 | 1854 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1855 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1856 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1857 | left -= to_do * rq_size; |
1858 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1859 | struct request *rq = p; |
1860 | ||
1861 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1862 | if (set->ops->init_request) { |
1863 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1864 | rq, hctx_idx, i, |
59f082e4 | 1865 | node)) { |
2af8cbe3 | 1866 | tags->static_rqs[i] = NULL; |
24d2f903 | 1867 | goto fail; |
a5164405 | 1868 | } |
e9b267d9 CH |
1869 | } |
1870 | ||
320ae51f JA |
1871 | p += rq_size; |
1872 | i++; | |
1873 | } | |
1874 | } | |
cc71a6f4 | 1875 | return 0; |
320ae51f | 1876 | |
24d2f903 | 1877 | fail: |
cc71a6f4 JA |
1878 | blk_mq_free_rqs(set, tags, hctx_idx); |
1879 | return -ENOMEM; | |
320ae51f JA |
1880 | } |
1881 | ||
e57690fe JA |
1882 | /* |
1883 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1884 | * software queue to the hw queue dispatch list, and ensure that it | |
1885 | * gets run. | |
1886 | */ | |
9467f859 | 1887 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1888 | { |
9467f859 | 1889 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1890 | struct blk_mq_ctx *ctx; |
1891 | LIST_HEAD(tmp); | |
1892 | ||
9467f859 | 1893 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1894 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1895 | |
1896 | spin_lock(&ctx->lock); | |
1897 | if (!list_empty(&ctx->rq_list)) { | |
1898 | list_splice_init(&ctx->rq_list, &tmp); | |
1899 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1900 | } | |
1901 | spin_unlock(&ctx->lock); | |
1902 | ||
1903 | if (list_empty(&tmp)) | |
9467f859 | 1904 | return 0; |
484b4061 | 1905 | |
e57690fe JA |
1906 | spin_lock(&hctx->lock); |
1907 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1908 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1909 | |
1910 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1911 | return 0; |
484b4061 JA |
1912 | } |
1913 | ||
9467f859 | 1914 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1915 | { |
9467f859 TG |
1916 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1917 | &hctx->cpuhp_dead); | |
484b4061 JA |
1918 | } |
1919 | ||
c3b4afca | 1920 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1921 | static void blk_mq_exit_hctx(struct request_queue *q, |
1922 | struct blk_mq_tag_set *set, | |
1923 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1924 | { | |
f70ced09 ML |
1925 | unsigned flush_start_tag = set->queue_depth; |
1926 | ||
08e98fc6 ML |
1927 | blk_mq_tag_idle(hctx); |
1928 | ||
f70ced09 ML |
1929 | if (set->ops->exit_request) |
1930 | set->ops->exit_request(set->driver_data, | |
1931 | hctx->fq->flush_rq, hctx_idx, | |
1932 | flush_start_tag + hctx_idx); | |
1933 | ||
08e98fc6 ML |
1934 | if (set->ops->exit_hctx) |
1935 | set->ops->exit_hctx(hctx, hctx_idx); | |
1936 | ||
6a83e74d BVA |
1937 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1938 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1939 | ||
9467f859 | 1940 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1941 | blk_free_flush_queue(hctx->fq); |
88459642 | 1942 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1943 | } |
1944 | ||
624dbe47 ML |
1945 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1946 | struct blk_mq_tag_set *set, int nr_queue) | |
1947 | { | |
1948 | struct blk_mq_hw_ctx *hctx; | |
1949 | unsigned int i; | |
1950 | ||
1951 | queue_for_each_hw_ctx(q, hctx, i) { | |
1952 | if (i == nr_queue) | |
1953 | break; | |
08e98fc6 | 1954 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1955 | } |
624dbe47 ML |
1956 | } |
1957 | ||
08e98fc6 ML |
1958 | static int blk_mq_init_hctx(struct request_queue *q, |
1959 | struct blk_mq_tag_set *set, | |
1960 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1961 | { |
08e98fc6 | 1962 | int node; |
f70ced09 | 1963 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1964 | |
1965 | node = hctx->numa_node; | |
1966 | if (node == NUMA_NO_NODE) | |
1967 | node = hctx->numa_node = set->numa_node; | |
1968 | ||
27489a3c | 1969 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1970 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1971 | spin_lock_init(&hctx->lock); | |
1972 | INIT_LIST_HEAD(&hctx->dispatch); | |
1973 | hctx->queue = q; | |
1974 | hctx->queue_num = hctx_idx; | |
2404e607 | 1975 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1976 | |
9467f859 | 1977 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1978 | |
1979 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1980 | |
1981 | /* | |
08e98fc6 ML |
1982 | * Allocate space for all possible cpus to avoid allocation at |
1983 | * runtime | |
320ae51f | 1984 | */ |
08e98fc6 ML |
1985 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1986 | GFP_KERNEL, node); | |
1987 | if (!hctx->ctxs) | |
1988 | goto unregister_cpu_notifier; | |
320ae51f | 1989 | |
88459642 OS |
1990 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1991 | node)) | |
08e98fc6 | 1992 | goto free_ctxs; |
320ae51f | 1993 | |
08e98fc6 | 1994 | hctx->nr_ctx = 0; |
320ae51f | 1995 | |
08e98fc6 ML |
1996 | if (set->ops->init_hctx && |
1997 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1998 | goto free_bitmap; | |
320ae51f | 1999 | |
f70ced09 ML |
2000 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2001 | if (!hctx->fq) | |
2002 | goto exit_hctx; | |
320ae51f | 2003 | |
f70ced09 ML |
2004 | if (set->ops->init_request && |
2005 | set->ops->init_request(set->driver_data, | |
2006 | hctx->fq->flush_rq, hctx_idx, | |
2007 | flush_start_tag + hctx_idx, node)) | |
2008 | goto free_fq; | |
320ae51f | 2009 | |
6a83e74d BVA |
2010 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
2011 | init_srcu_struct(&hctx->queue_rq_srcu); | |
2012 | ||
08e98fc6 | 2013 | return 0; |
320ae51f | 2014 | |
f70ced09 ML |
2015 | free_fq: |
2016 | kfree(hctx->fq); | |
2017 | exit_hctx: | |
2018 | if (set->ops->exit_hctx) | |
2019 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2020 | free_bitmap: |
88459642 | 2021 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2022 | free_ctxs: |
2023 | kfree(hctx->ctxs); | |
2024 | unregister_cpu_notifier: | |
9467f859 | 2025 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2026 | return -1; |
2027 | } | |
320ae51f | 2028 | |
320ae51f JA |
2029 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2030 | unsigned int nr_hw_queues) | |
2031 | { | |
2032 | unsigned int i; | |
2033 | ||
2034 | for_each_possible_cpu(i) { | |
2035 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2036 | struct blk_mq_hw_ctx *hctx; | |
2037 | ||
320ae51f JA |
2038 | __ctx->cpu = i; |
2039 | spin_lock_init(&__ctx->lock); | |
2040 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2041 | __ctx->queue = q; | |
cf43e6be JA |
2042 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
2043 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
2044 | |
2045 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
2046 | if (!cpu_online(i)) |
2047 | continue; | |
2048 | ||
7d7e0f90 | 2049 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2050 | |
320ae51f JA |
2051 | /* |
2052 | * Set local node, IFF we have more than one hw queue. If | |
2053 | * not, we remain on the home node of the device | |
2054 | */ | |
2055 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2056 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2057 | } |
2058 | } | |
2059 | ||
cc71a6f4 JA |
2060 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2061 | { | |
2062 | int ret = 0; | |
2063 | ||
2064 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2065 | set->queue_depth, set->reserved_tags); | |
2066 | if (!set->tags[hctx_idx]) | |
2067 | return false; | |
2068 | ||
2069 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2070 | set->queue_depth); | |
2071 | if (!ret) | |
2072 | return true; | |
2073 | ||
2074 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2075 | set->tags[hctx_idx] = NULL; | |
2076 | return false; | |
2077 | } | |
2078 | ||
2079 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2080 | unsigned int hctx_idx) | |
2081 | { | |
bd166ef1 JA |
2082 | if (set->tags[hctx_idx]) { |
2083 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2084 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2085 | set->tags[hctx_idx] = NULL; | |
2086 | } | |
cc71a6f4 JA |
2087 | } |
2088 | ||
5778322e AM |
2089 | static void blk_mq_map_swqueue(struct request_queue *q, |
2090 | const struct cpumask *online_mask) | |
320ae51f | 2091 | { |
d1b1cea1 | 2092 | unsigned int i, hctx_idx; |
320ae51f JA |
2093 | struct blk_mq_hw_ctx *hctx; |
2094 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2095 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2096 | |
60de074b AM |
2097 | /* |
2098 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2099 | */ | |
2100 | mutex_lock(&q->sysfs_lock); | |
2101 | ||
320ae51f | 2102 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2103 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2104 | hctx->nr_ctx = 0; |
2105 | } | |
2106 | ||
2107 | /* | |
2108 | * Map software to hardware queues | |
2109 | */ | |
897bb0c7 | 2110 | for_each_possible_cpu(i) { |
320ae51f | 2111 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2112 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2113 | continue; |
2114 | ||
d1b1cea1 GKB |
2115 | hctx_idx = q->mq_map[i]; |
2116 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2117 | if (!set->tags[hctx_idx] && |
2118 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2119 | /* |
2120 | * If tags initialization fail for some hctx, | |
2121 | * that hctx won't be brought online. In this | |
2122 | * case, remap the current ctx to hctx[0] which | |
2123 | * is guaranteed to always have tags allocated | |
2124 | */ | |
cc71a6f4 | 2125 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2126 | } |
2127 | ||
897bb0c7 | 2128 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2129 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2130 | |
e4043dcf | 2131 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2132 | ctx->index_hw = hctx->nr_ctx; |
2133 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2134 | } | |
506e931f | 2135 | |
60de074b AM |
2136 | mutex_unlock(&q->sysfs_lock); |
2137 | ||
506e931f | 2138 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2139 | /* |
a68aafa5 JA |
2140 | * If no software queues are mapped to this hardware queue, |
2141 | * disable it and free the request entries. | |
484b4061 JA |
2142 | */ |
2143 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2144 | /* Never unmap queue 0. We need it as a |
2145 | * fallback in case of a new remap fails | |
2146 | * allocation | |
2147 | */ | |
cc71a6f4 JA |
2148 | if (i && set->tags[i]) |
2149 | blk_mq_free_map_and_requests(set, i); | |
2150 | ||
2a34c087 | 2151 | hctx->tags = NULL; |
484b4061 JA |
2152 | continue; |
2153 | } | |
2154 | ||
2a34c087 ML |
2155 | hctx->tags = set->tags[i]; |
2156 | WARN_ON(!hctx->tags); | |
2157 | ||
889fa31f CY |
2158 | /* |
2159 | * Set the map size to the number of mapped software queues. | |
2160 | * This is more accurate and more efficient than looping | |
2161 | * over all possibly mapped software queues. | |
2162 | */ | |
88459642 | 2163 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2164 | |
484b4061 JA |
2165 | /* |
2166 | * Initialize batch roundrobin counts | |
2167 | */ | |
506e931f JA |
2168 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2169 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2170 | } | |
320ae51f JA |
2171 | } |
2172 | ||
2404e607 | 2173 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2174 | { |
2175 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2176 | int i; |
2177 | ||
2404e607 JM |
2178 | queue_for_each_hw_ctx(q, hctx, i) { |
2179 | if (shared) | |
2180 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2181 | else | |
2182 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2183 | } | |
2184 | } | |
2185 | ||
2186 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2187 | { | |
2188 | struct request_queue *q; | |
0d2602ca JA |
2189 | |
2190 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2191 | blk_mq_freeze_queue(q); | |
2404e607 | 2192 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2193 | blk_mq_unfreeze_queue(q); |
2194 | } | |
2195 | } | |
2196 | ||
2197 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2198 | { | |
2199 | struct blk_mq_tag_set *set = q->tag_set; | |
2200 | ||
0d2602ca JA |
2201 | mutex_lock(&set->tag_list_lock); |
2202 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2203 | if (list_is_singular(&set->tag_list)) { |
2204 | /* just transitioned to unshared */ | |
2205 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2206 | /* update existing queue */ | |
2207 | blk_mq_update_tag_set_depth(set, false); | |
2208 | } | |
0d2602ca | 2209 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2210 | } |
2211 | ||
2212 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2213 | struct request_queue *q) | |
2214 | { | |
2215 | q->tag_set = set; | |
2216 | ||
2217 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2218 | |
2219 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2220 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2221 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2222 | /* update existing queue */ | |
2223 | blk_mq_update_tag_set_depth(set, true); | |
2224 | } | |
2225 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2226 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2227 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2228 | |
0d2602ca JA |
2229 | mutex_unlock(&set->tag_list_lock); |
2230 | } | |
2231 | ||
e09aae7e ML |
2232 | /* |
2233 | * It is the actual release handler for mq, but we do it from | |
2234 | * request queue's release handler for avoiding use-after-free | |
2235 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2236 | * but we can't group ctx/kctx kobj without it. | |
2237 | */ | |
2238 | void blk_mq_release(struct request_queue *q) | |
2239 | { | |
2240 | struct blk_mq_hw_ctx *hctx; | |
2241 | unsigned int i; | |
2242 | ||
bd166ef1 JA |
2243 | blk_mq_sched_teardown(q); |
2244 | ||
e09aae7e | 2245 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2246 | queue_for_each_hw_ctx(q, hctx, i) { |
2247 | if (!hctx) | |
2248 | continue; | |
6c8b232e | 2249 | kobject_put(&hctx->kobj); |
c3b4afca | 2250 | } |
e09aae7e | 2251 | |
a723bab3 AM |
2252 | q->mq_map = NULL; |
2253 | ||
e09aae7e ML |
2254 | kfree(q->queue_hw_ctx); |
2255 | ||
7ea5fe31 ML |
2256 | /* |
2257 | * release .mq_kobj and sw queue's kobject now because | |
2258 | * both share lifetime with request queue. | |
2259 | */ | |
2260 | blk_mq_sysfs_deinit(q); | |
2261 | ||
e09aae7e ML |
2262 | free_percpu(q->queue_ctx); |
2263 | } | |
2264 | ||
24d2f903 | 2265 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2266 | { |
2267 | struct request_queue *uninit_q, *q; | |
2268 | ||
2269 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2270 | if (!uninit_q) | |
2271 | return ERR_PTR(-ENOMEM); | |
2272 | ||
2273 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2274 | if (IS_ERR(q)) | |
2275 | blk_cleanup_queue(uninit_q); | |
2276 | ||
2277 | return q; | |
2278 | } | |
2279 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2280 | ||
868f2f0b KB |
2281 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2282 | struct request_queue *q) | |
320ae51f | 2283 | { |
868f2f0b KB |
2284 | int i, j; |
2285 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2286 | |
868f2f0b | 2287 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2288 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2289 | int node; |
f14bbe77 | 2290 | |
868f2f0b KB |
2291 | if (hctxs[i]) |
2292 | continue; | |
2293 | ||
2294 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2295 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2296 | GFP_KERNEL, node); | |
320ae51f | 2297 | if (!hctxs[i]) |
868f2f0b | 2298 | break; |
320ae51f | 2299 | |
a86073e4 | 2300 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2301 | node)) { |
2302 | kfree(hctxs[i]); | |
2303 | hctxs[i] = NULL; | |
2304 | break; | |
2305 | } | |
e4043dcf | 2306 | |
0d2602ca | 2307 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2308 | hctxs[i]->numa_node = node; |
320ae51f | 2309 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2310 | |
2311 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2312 | free_cpumask_var(hctxs[i]->cpumask); | |
2313 | kfree(hctxs[i]); | |
2314 | hctxs[i] = NULL; | |
2315 | break; | |
2316 | } | |
2317 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2318 | } |
868f2f0b KB |
2319 | for (j = i; j < q->nr_hw_queues; j++) { |
2320 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2321 | ||
2322 | if (hctx) { | |
cc71a6f4 JA |
2323 | if (hctx->tags) |
2324 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2325 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2326 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2327 | hctxs[j] = NULL; |
2328 | ||
2329 | } | |
2330 | } | |
2331 | q->nr_hw_queues = i; | |
2332 | blk_mq_sysfs_register(q); | |
2333 | } | |
2334 | ||
2335 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2336 | struct request_queue *q) | |
2337 | { | |
66841672 ML |
2338 | /* mark the queue as mq asap */ |
2339 | q->mq_ops = set->ops; | |
2340 | ||
868f2f0b KB |
2341 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2342 | if (!q->queue_ctx) | |
c7de5726 | 2343 | goto err_exit; |
868f2f0b | 2344 | |
737f98cf ML |
2345 | /* init q->mq_kobj and sw queues' kobjects */ |
2346 | blk_mq_sysfs_init(q); | |
2347 | ||
868f2f0b KB |
2348 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2349 | GFP_KERNEL, set->numa_node); | |
2350 | if (!q->queue_hw_ctx) | |
2351 | goto err_percpu; | |
2352 | ||
bdd17e75 | 2353 | q->mq_map = set->mq_map; |
868f2f0b KB |
2354 | |
2355 | blk_mq_realloc_hw_ctxs(set, q); | |
2356 | if (!q->nr_hw_queues) | |
2357 | goto err_hctxs; | |
320ae51f | 2358 | |
287922eb | 2359 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2360 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2361 | |
2362 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2363 | |
94eddfbe | 2364 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2365 | |
05f1dd53 JA |
2366 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2367 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2368 | ||
1be036e9 CH |
2369 | q->sg_reserved_size = INT_MAX; |
2370 | ||
2849450a | 2371 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2372 | INIT_LIST_HEAD(&q->requeue_list); |
2373 | spin_lock_init(&q->requeue_lock); | |
2374 | ||
07068d5b JA |
2375 | if (q->nr_hw_queues > 1) |
2376 | blk_queue_make_request(q, blk_mq_make_request); | |
2377 | else | |
2378 | blk_queue_make_request(q, blk_sq_make_request); | |
2379 | ||
eba71768 JA |
2380 | /* |
2381 | * Do this after blk_queue_make_request() overrides it... | |
2382 | */ | |
2383 | q->nr_requests = set->queue_depth; | |
2384 | ||
64f1c21e JA |
2385 | /* |
2386 | * Default to classic polling | |
2387 | */ | |
2388 | q->poll_nsec = -1; | |
2389 | ||
24d2f903 CH |
2390 | if (set->ops->complete) |
2391 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2392 | |
24d2f903 | 2393 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2394 | |
5778322e | 2395 | get_online_cpus(); |
320ae51f | 2396 | mutex_lock(&all_q_mutex); |
320ae51f | 2397 | |
4593fdbe | 2398 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2399 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2400 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2401 | |
4593fdbe | 2402 | mutex_unlock(&all_q_mutex); |
5778322e | 2403 | put_online_cpus(); |
4593fdbe | 2404 | |
d3484991 JA |
2405 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2406 | int ret; | |
2407 | ||
2408 | ret = blk_mq_sched_init(q); | |
2409 | if (ret) | |
2410 | return ERR_PTR(ret); | |
2411 | } | |
2412 | ||
320ae51f | 2413 | return q; |
18741986 | 2414 | |
320ae51f | 2415 | err_hctxs: |
868f2f0b | 2416 | kfree(q->queue_hw_ctx); |
320ae51f | 2417 | err_percpu: |
868f2f0b | 2418 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2419 | err_exit: |
2420 | q->mq_ops = NULL; | |
320ae51f JA |
2421 | return ERR_PTR(-ENOMEM); |
2422 | } | |
b62c21b7 | 2423 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2424 | |
2425 | void blk_mq_free_queue(struct request_queue *q) | |
2426 | { | |
624dbe47 | 2427 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2428 | |
0e626368 AM |
2429 | mutex_lock(&all_q_mutex); |
2430 | list_del_init(&q->all_q_node); | |
2431 | mutex_unlock(&all_q_mutex); | |
2432 | ||
87760e5e JA |
2433 | wbt_exit(q); |
2434 | ||
0d2602ca JA |
2435 | blk_mq_del_queue_tag_set(q); |
2436 | ||
624dbe47 | 2437 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2438 | } |
320ae51f JA |
2439 | |
2440 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2441 | static void blk_mq_queue_reinit(struct request_queue *q, |
2442 | const struct cpumask *online_mask) | |
320ae51f | 2443 | { |
4ecd4fef | 2444 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2445 | |
67aec14c JA |
2446 | blk_mq_sysfs_unregister(q); |
2447 | ||
320ae51f JA |
2448 | /* |
2449 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2450 | * we should change hctx numa_node according to new topology (this | |
2451 | * involves free and re-allocate memory, worthy doing?) | |
2452 | */ | |
2453 | ||
5778322e | 2454 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2455 | |
67aec14c | 2456 | blk_mq_sysfs_register(q); |
320ae51f JA |
2457 | } |
2458 | ||
65d5291e SAS |
2459 | /* |
2460 | * New online cpumask which is going to be set in this hotplug event. | |
2461 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2462 | * one-by-one and dynamically allocating this could result in a failure. | |
2463 | */ | |
2464 | static struct cpumask cpuhp_online_new; | |
2465 | ||
2466 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2467 | { |
2468 | struct request_queue *q; | |
320ae51f JA |
2469 | |
2470 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2471 | /* |
2472 | * We need to freeze and reinit all existing queues. Freezing | |
2473 | * involves synchronous wait for an RCU grace period and doing it | |
2474 | * one by one may take a long time. Start freezing all queues in | |
2475 | * one swoop and then wait for the completions so that freezing can | |
2476 | * take place in parallel. | |
2477 | */ | |
2478 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2479 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2480 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2481 | blk_mq_freeze_queue_wait(q); |
2482 | ||
320ae51f | 2483 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2484 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2485 | |
2486 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2487 | blk_mq_unfreeze_queue(q); | |
2488 | ||
320ae51f | 2489 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2490 | } |
2491 | ||
2492 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2493 | { | |
97a32864 | 2494 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2495 | blk_mq_queue_reinit_work(); |
2496 | return 0; | |
2497 | } | |
2498 | ||
2499 | /* | |
2500 | * Before hotadded cpu starts handling requests, new mappings must be | |
2501 | * established. Otherwise, these requests in hw queue might never be | |
2502 | * dispatched. | |
2503 | * | |
2504 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2505 | * for CPU0, and ctx1 for CPU1). | |
2506 | * | |
2507 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2508 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2509 | * | |
2c3ad667 JA |
2510 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2511 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2512 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2513 | * ignored. | |
65d5291e SAS |
2514 | */ |
2515 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2516 | { | |
2517 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2518 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2519 | blk_mq_queue_reinit_work(); | |
2520 | return 0; | |
320ae51f JA |
2521 | } |
2522 | ||
a5164405 JA |
2523 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2524 | { | |
2525 | int i; | |
2526 | ||
cc71a6f4 JA |
2527 | for (i = 0; i < set->nr_hw_queues; i++) |
2528 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2529 | goto out_unwind; |
a5164405 JA |
2530 | |
2531 | return 0; | |
2532 | ||
2533 | out_unwind: | |
2534 | while (--i >= 0) | |
cc71a6f4 | 2535 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2536 | |
a5164405 JA |
2537 | return -ENOMEM; |
2538 | } | |
2539 | ||
2540 | /* | |
2541 | * Allocate the request maps associated with this tag_set. Note that this | |
2542 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2543 | * will be updated to reflect the allocated depth. | |
2544 | */ | |
2545 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2546 | { | |
2547 | unsigned int depth; | |
2548 | int err; | |
2549 | ||
2550 | depth = set->queue_depth; | |
2551 | do { | |
2552 | err = __blk_mq_alloc_rq_maps(set); | |
2553 | if (!err) | |
2554 | break; | |
2555 | ||
2556 | set->queue_depth >>= 1; | |
2557 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2558 | err = -ENOMEM; | |
2559 | break; | |
2560 | } | |
2561 | } while (set->queue_depth); | |
2562 | ||
2563 | if (!set->queue_depth || err) { | |
2564 | pr_err("blk-mq: failed to allocate request map\n"); | |
2565 | return -ENOMEM; | |
2566 | } | |
2567 | ||
2568 | if (depth != set->queue_depth) | |
2569 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2570 | depth, set->queue_depth); | |
2571 | ||
2572 | return 0; | |
2573 | } | |
2574 | ||
a4391c64 JA |
2575 | /* |
2576 | * Alloc a tag set to be associated with one or more request queues. | |
2577 | * May fail with EINVAL for various error conditions. May adjust the | |
2578 | * requested depth down, if if it too large. In that case, the set | |
2579 | * value will be stored in set->queue_depth. | |
2580 | */ | |
24d2f903 CH |
2581 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2582 | { | |
da695ba2 CH |
2583 | int ret; |
2584 | ||
205fb5f5 BVA |
2585 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2586 | ||
24d2f903 CH |
2587 | if (!set->nr_hw_queues) |
2588 | return -EINVAL; | |
a4391c64 | 2589 | if (!set->queue_depth) |
24d2f903 CH |
2590 | return -EINVAL; |
2591 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2592 | return -EINVAL; | |
2593 | ||
7d7e0f90 | 2594 | if (!set->ops->queue_rq) |
24d2f903 CH |
2595 | return -EINVAL; |
2596 | ||
a4391c64 JA |
2597 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2598 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2599 | BLK_MQ_MAX_DEPTH); | |
2600 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2601 | } | |
24d2f903 | 2602 | |
6637fadf SL |
2603 | /* |
2604 | * If a crashdump is active, then we are potentially in a very | |
2605 | * memory constrained environment. Limit us to 1 queue and | |
2606 | * 64 tags to prevent using too much memory. | |
2607 | */ | |
2608 | if (is_kdump_kernel()) { | |
2609 | set->nr_hw_queues = 1; | |
2610 | set->queue_depth = min(64U, set->queue_depth); | |
2611 | } | |
868f2f0b KB |
2612 | /* |
2613 | * There is no use for more h/w queues than cpus. | |
2614 | */ | |
2615 | if (set->nr_hw_queues > nr_cpu_ids) | |
2616 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2617 | |
868f2f0b | 2618 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2619 | GFP_KERNEL, set->numa_node); |
2620 | if (!set->tags) | |
a5164405 | 2621 | return -ENOMEM; |
24d2f903 | 2622 | |
da695ba2 CH |
2623 | ret = -ENOMEM; |
2624 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2625 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2626 | if (!set->mq_map) |
2627 | goto out_free_tags; | |
2628 | ||
da695ba2 CH |
2629 | if (set->ops->map_queues) |
2630 | ret = set->ops->map_queues(set); | |
2631 | else | |
2632 | ret = blk_mq_map_queues(set); | |
2633 | if (ret) | |
2634 | goto out_free_mq_map; | |
2635 | ||
2636 | ret = blk_mq_alloc_rq_maps(set); | |
2637 | if (ret) | |
bdd17e75 | 2638 | goto out_free_mq_map; |
24d2f903 | 2639 | |
0d2602ca JA |
2640 | mutex_init(&set->tag_list_lock); |
2641 | INIT_LIST_HEAD(&set->tag_list); | |
2642 | ||
24d2f903 | 2643 | return 0; |
bdd17e75 CH |
2644 | |
2645 | out_free_mq_map: | |
2646 | kfree(set->mq_map); | |
2647 | set->mq_map = NULL; | |
2648 | out_free_tags: | |
5676e7b6 RE |
2649 | kfree(set->tags); |
2650 | set->tags = NULL; | |
da695ba2 | 2651 | return ret; |
24d2f903 CH |
2652 | } |
2653 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2654 | ||
2655 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2656 | { | |
2657 | int i; | |
2658 | ||
cc71a6f4 JA |
2659 | for (i = 0; i < nr_cpu_ids; i++) |
2660 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2661 | |
bdd17e75 CH |
2662 | kfree(set->mq_map); |
2663 | set->mq_map = NULL; | |
2664 | ||
981bd189 | 2665 | kfree(set->tags); |
5676e7b6 | 2666 | set->tags = NULL; |
24d2f903 CH |
2667 | } |
2668 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2669 | ||
e3a2b3f9 JA |
2670 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2671 | { | |
2672 | struct blk_mq_tag_set *set = q->tag_set; | |
2673 | struct blk_mq_hw_ctx *hctx; | |
2674 | int i, ret; | |
2675 | ||
bd166ef1 | 2676 | if (!set) |
e3a2b3f9 JA |
2677 | return -EINVAL; |
2678 | ||
70f36b60 JA |
2679 | blk_mq_freeze_queue(q); |
2680 | blk_mq_quiesce_queue(q); | |
2681 | ||
e3a2b3f9 JA |
2682 | ret = 0; |
2683 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2684 | if (!hctx->tags) |
2685 | continue; | |
bd166ef1 JA |
2686 | /* |
2687 | * If we're using an MQ scheduler, just update the scheduler | |
2688 | * queue depth. This is similar to what the old code would do. | |
2689 | */ | |
70f36b60 JA |
2690 | if (!hctx->sched_tags) { |
2691 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2692 | min(nr, set->queue_depth), | |
2693 | false); | |
2694 | } else { | |
2695 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2696 | nr, true); | |
2697 | } | |
e3a2b3f9 JA |
2698 | if (ret) |
2699 | break; | |
2700 | } | |
2701 | ||
2702 | if (!ret) | |
2703 | q->nr_requests = nr; | |
2704 | ||
70f36b60 JA |
2705 | blk_mq_unfreeze_queue(q); |
2706 | blk_mq_start_stopped_hw_queues(q, true); | |
2707 | ||
e3a2b3f9 JA |
2708 | return ret; |
2709 | } | |
2710 | ||
868f2f0b KB |
2711 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2712 | { | |
2713 | struct request_queue *q; | |
2714 | ||
2715 | if (nr_hw_queues > nr_cpu_ids) | |
2716 | nr_hw_queues = nr_cpu_ids; | |
2717 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2718 | return; | |
2719 | ||
2720 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2721 | blk_mq_freeze_queue(q); | |
2722 | ||
2723 | set->nr_hw_queues = nr_hw_queues; | |
2724 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2725 | blk_mq_realloc_hw_ctxs(set, q); | |
2726 | ||
f6f94300 JB |
2727 | /* |
2728 | * Manually set the make_request_fn as blk_queue_make_request | |
2729 | * resets a lot of the queue settings. | |
2730 | */ | |
868f2f0b | 2731 | if (q->nr_hw_queues > 1) |
f6f94300 | 2732 | q->make_request_fn = blk_mq_make_request; |
868f2f0b | 2733 | else |
f6f94300 | 2734 | q->make_request_fn = blk_sq_make_request; |
868f2f0b KB |
2735 | |
2736 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2737 | } | |
2738 | ||
2739 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2740 | blk_mq_unfreeze_queue(q); | |
2741 | } | |
2742 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2743 | ||
64f1c21e JA |
2744 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2745 | struct blk_mq_hw_ctx *hctx, | |
2746 | struct request *rq) | |
2747 | { | |
2748 | struct blk_rq_stat stat[2]; | |
2749 | unsigned long ret = 0; | |
2750 | ||
2751 | /* | |
2752 | * If stats collection isn't on, don't sleep but turn it on for | |
2753 | * future users | |
2754 | */ | |
2755 | if (!blk_stat_enable(q)) | |
2756 | return 0; | |
2757 | ||
2758 | /* | |
2759 | * We don't have to do this once per IO, should optimize this | |
2760 | * to just use the current window of stats until it changes | |
2761 | */ | |
2762 | memset(&stat, 0, sizeof(stat)); | |
2763 | blk_hctx_stat_get(hctx, stat); | |
2764 | ||
2765 | /* | |
2766 | * As an optimistic guess, use half of the mean service time | |
2767 | * for this type of request. We can (and should) make this smarter. | |
2768 | * For instance, if the completion latencies are tight, we can | |
2769 | * get closer than just half the mean. This is especially | |
2770 | * important on devices where the completion latencies are longer | |
2771 | * than ~10 usec. | |
2772 | */ | |
2773 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2774 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2775 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2776 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2777 | ||
2778 | return ret; | |
2779 | } | |
2780 | ||
06426adf | 2781 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2782 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2783 | struct request *rq) |
2784 | { | |
2785 | struct hrtimer_sleeper hs; | |
2786 | enum hrtimer_mode mode; | |
64f1c21e | 2787 | unsigned int nsecs; |
06426adf JA |
2788 | ktime_t kt; |
2789 | ||
64f1c21e JA |
2790 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2791 | return false; | |
2792 | ||
2793 | /* | |
2794 | * poll_nsec can be: | |
2795 | * | |
2796 | * -1: don't ever hybrid sleep | |
2797 | * 0: use half of prev avg | |
2798 | * >0: use this specific value | |
2799 | */ | |
2800 | if (q->poll_nsec == -1) | |
2801 | return false; | |
2802 | else if (q->poll_nsec > 0) | |
2803 | nsecs = q->poll_nsec; | |
2804 | else | |
2805 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2806 | ||
2807 | if (!nsecs) | |
06426adf JA |
2808 | return false; |
2809 | ||
2810 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2811 | ||
2812 | /* | |
2813 | * This will be replaced with the stats tracking code, using | |
2814 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2815 | */ | |
8b0e1953 | 2816 | kt = nsecs; |
06426adf JA |
2817 | |
2818 | mode = HRTIMER_MODE_REL; | |
2819 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2820 | hrtimer_set_expires(&hs.timer, kt); | |
2821 | ||
2822 | hrtimer_init_sleeper(&hs, current); | |
2823 | do { | |
2824 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2825 | break; | |
2826 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2827 | hrtimer_start_expires(&hs.timer, mode); | |
2828 | if (hs.task) | |
2829 | io_schedule(); | |
2830 | hrtimer_cancel(&hs.timer); | |
2831 | mode = HRTIMER_MODE_ABS; | |
2832 | } while (hs.task && !signal_pending(current)); | |
2833 | ||
2834 | __set_current_state(TASK_RUNNING); | |
2835 | destroy_hrtimer_on_stack(&hs.timer); | |
2836 | return true; | |
2837 | } | |
2838 | ||
bbd7bb70 JA |
2839 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2840 | { | |
2841 | struct request_queue *q = hctx->queue; | |
2842 | long state; | |
2843 | ||
06426adf JA |
2844 | /* |
2845 | * If we sleep, have the caller restart the poll loop to reset | |
2846 | * the state. Like for the other success return cases, the | |
2847 | * caller is responsible for checking if the IO completed. If | |
2848 | * the IO isn't complete, we'll get called again and will go | |
2849 | * straight to the busy poll loop. | |
2850 | */ | |
64f1c21e | 2851 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2852 | return true; |
2853 | ||
bbd7bb70 JA |
2854 | hctx->poll_considered++; |
2855 | ||
2856 | state = current->state; | |
2857 | while (!need_resched()) { | |
2858 | int ret; | |
2859 | ||
2860 | hctx->poll_invoked++; | |
2861 | ||
2862 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2863 | if (ret > 0) { | |
2864 | hctx->poll_success++; | |
2865 | set_current_state(TASK_RUNNING); | |
2866 | return true; | |
2867 | } | |
2868 | ||
2869 | if (signal_pending_state(state, current)) | |
2870 | set_current_state(TASK_RUNNING); | |
2871 | ||
2872 | if (current->state == TASK_RUNNING) | |
2873 | return true; | |
2874 | if (ret < 0) | |
2875 | break; | |
2876 | cpu_relax(); | |
2877 | } | |
2878 | ||
2879 | return false; | |
2880 | } | |
2881 | ||
2882 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2883 | { | |
2884 | struct blk_mq_hw_ctx *hctx; | |
2885 | struct blk_plug *plug; | |
2886 | struct request *rq; | |
2887 | ||
2888 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2889 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2890 | return false; | |
2891 | ||
2892 | plug = current->plug; | |
2893 | if (plug) | |
2894 | blk_flush_plug_list(plug, false); | |
2895 | ||
2896 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2897 | if (!blk_qc_t_is_internal(cookie)) |
2898 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2899 | else | |
2900 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2901 | |
2902 | return __blk_mq_poll(hctx, rq); | |
2903 | } | |
2904 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2905 | ||
676141e4 JA |
2906 | void blk_mq_disable_hotplug(void) |
2907 | { | |
2908 | mutex_lock(&all_q_mutex); | |
2909 | } | |
2910 | ||
2911 | void blk_mq_enable_hotplug(void) | |
2912 | { | |
2913 | mutex_unlock(&all_q_mutex); | |
2914 | } | |
2915 | ||
320ae51f JA |
2916 | static int __init blk_mq_init(void) |
2917 | { | |
9467f859 TG |
2918 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2919 | blk_mq_hctx_notify_dead); | |
320ae51f | 2920 | |
65d5291e SAS |
2921 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2922 | blk_mq_queue_reinit_prepare, | |
2923 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2924 | return 0; |
2925 | } | |
2926 | subsys_initcall(blk_mq_init); |