]>
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 | ||
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
720b8ccc SB |
45 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
46 | { | |
47 | int ddir, bytes, bucket; | |
48 | ||
49 | ddir = blk_stat_rq_ddir(rq); | |
50 | bytes = blk_rq_bytes(rq); | |
51 | ||
52 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
53 | ||
54 | if (bucket < 0) | |
55 | return -1; | |
56 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
57 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
58 | ||
59 | return bucket; | |
60 | } | |
61 | ||
320ae51f JA |
62 | /* |
63 | * Check if any of the ctx's have pending work in this hardware queue | |
64 | */ | |
50e1dab8 | 65 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 66 | { |
bd166ef1 JA |
67 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
68 | !list_empty_careful(&hctx->dispatch) || | |
69 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
70 | } |
71 | ||
320ae51f JA |
72 | /* |
73 | * Mark this ctx as having pending work in this hardware queue | |
74 | */ | |
75 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
76 | struct blk_mq_ctx *ctx) | |
77 | { | |
88459642 OS |
78 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
79 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
80 | } |
81 | ||
82 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
83 | struct blk_mq_ctx *ctx) | |
84 | { | |
88459642 | 85 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
86 | } |
87 | ||
1671d522 | 88 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 89 | { |
4ecd4fef | 90 | int freeze_depth; |
cddd5d17 | 91 | |
4ecd4fef CH |
92 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
93 | if (freeze_depth == 1) { | |
3ef28e83 | 94 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 95 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 96 | } |
f3af020b | 97 | } |
1671d522 | 98 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 99 | |
6bae363e | 100 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 101 | { |
3ef28e83 | 102 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 103 | } |
6bae363e | 104 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 105 | |
f91328c4 KB |
106 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
107 | unsigned long timeout) | |
108 | { | |
109 | return wait_event_timeout(q->mq_freeze_wq, | |
110 | percpu_ref_is_zero(&q->q_usage_counter), | |
111 | timeout); | |
112 | } | |
113 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 114 | |
f3af020b TH |
115 | /* |
116 | * Guarantee no request is in use, so we can change any data structure of | |
117 | * the queue afterward. | |
118 | */ | |
3ef28e83 | 119 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 120 | { |
3ef28e83 DW |
121 | /* |
122 | * In the !blk_mq case we are only calling this to kill the | |
123 | * q_usage_counter, otherwise this increases the freeze depth | |
124 | * and waits for it to return to zero. For this reason there is | |
125 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
126 | * exported to drivers as the only user for unfreeze is blk_mq. | |
127 | */ | |
1671d522 | 128 | blk_freeze_queue_start(q); |
f3af020b TH |
129 | blk_mq_freeze_queue_wait(q); |
130 | } | |
3ef28e83 DW |
131 | |
132 | void blk_mq_freeze_queue(struct request_queue *q) | |
133 | { | |
134 | /* | |
135 | * ...just an alias to keep freeze and unfreeze actions balanced | |
136 | * in the blk_mq_* namespace | |
137 | */ | |
138 | blk_freeze_queue(q); | |
139 | } | |
c761d96b | 140 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 141 | |
b4c6a028 | 142 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 143 | { |
4ecd4fef | 144 | int freeze_depth; |
320ae51f | 145 | |
4ecd4fef CH |
146 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
147 | WARN_ON_ONCE(freeze_depth < 0); | |
148 | if (!freeze_depth) { | |
3ef28e83 | 149 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 150 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 151 | } |
320ae51f | 152 | } |
b4c6a028 | 153 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 154 | |
6a83e74d BVA |
155 | /** |
156 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
157 | * @q: request queue. | |
158 | * | |
159 | * Note: this function does not prevent that the struct request end_io() | |
160 | * callback function is invoked. Additionally, it is not prevented that | |
161 | * new queue_rq() calls occur unless the queue has been stopped first. | |
162 | */ | |
163 | void blk_mq_quiesce_queue(struct request_queue *q) | |
164 | { | |
165 | struct blk_mq_hw_ctx *hctx; | |
166 | unsigned int i; | |
167 | bool rcu = false; | |
168 | ||
169 | blk_mq_stop_hw_queues(q); | |
170 | ||
171 | queue_for_each_hw_ctx(q, hctx, i) { | |
172 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
173 | synchronize_srcu(&hctx->queue_rq_srcu); | |
174 | else | |
175 | rcu = true; | |
176 | } | |
177 | if (rcu) | |
178 | synchronize_rcu(); | |
179 | } | |
180 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
181 | ||
aed3ea94 JA |
182 | void blk_mq_wake_waiters(struct request_queue *q) |
183 | { | |
184 | struct blk_mq_hw_ctx *hctx; | |
185 | unsigned int i; | |
186 | ||
187 | queue_for_each_hw_ctx(q, hctx, i) | |
188 | if (blk_mq_hw_queue_mapped(hctx)) | |
189 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
190 | |
191 | /* | |
192 | * If we are called because the queue has now been marked as | |
193 | * dying, we need to ensure that processes currently waiting on | |
194 | * the queue are notified as well. | |
195 | */ | |
196 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
197 | } |
198 | ||
320ae51f JA |
199 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
200 | { | |
201 | return blk_mq_has_free_tags(hctx->tags); | |
202 | } | |
203 | EXPORT_SYMBOL(blk_mq_can_queue); | |
204 | ||
2c3ad667 JA |
205 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
206 | struct request *rq, unsigned int op) | |
320ae51f | 207 | { |
af76e555 CH |
208 | INIT_LIST_HEAD(&rq->queuelist); |
209 | /* csd/requeue_work/fifo_time is initialized before use */ | |
210 | rq->q = q; | |
320ae51f | 211 | rq->mq_ctx = ctx; |
ef295ecf | 212 | rq->cmd_flags = op; |
e8064021 CH |
213 | if (blk_queue_io_stat(q)) |
214 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
215 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
216 | rq->cpu = -1; | |
af76e555 CH |
217 | INIT_HLIST_NODE(&rq->hash); |
218 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
219 | rq->rq_disk = NULL; |
220 | rq->part = NULL; | |
3ee32372 | 221 | rq->start_time = jiffies; |
af76e555 CH |
222 | #ifdef CONFIG_BLK_CGROUP |
223 | rq->rl = NULL; | |
0fec08b4 | 224 | set_start_time_ns(rq); |
af76e555 CH |
225 | rq->io_start_time_ns = 0; |
226 | #endif | |
227 | rq->nr_phys_segments = 0; | |
228 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
229 | rq->nr_integrity_segments = 0; | |
230 | #endif | |
af76e555 CH |
231 | rq->special = NULL; |
232 | /* tag was already set */ | |
af76e555 | 233 | rq->extra_len = 0; |
af76e555 | 234 | |
af76e555 | 235 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
236 | rq->timeout = 0; |
237 | ||
af76e555 CH |
238 | rq->end_io = NULL; |
239 | rq->end_io_data = NULL; | |
240 | rq->next_rq = NULL; | |
241 | ||
ef295ecf | 242 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 243 | } |
2c3ad667 | 244 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 245 | |
2c3ad667 JA |
246 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
247 | unsigned int op) | |
5dee8577 CH |
248 | { |
249 | struct request *rq; | |
250 | unsigned int tag; | |
251 | ||
cb96a42c | 252 | tag = blk_mq_get_tag(data); |
5dee8577 | 253 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
254 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
255 | ||
256 | rq = tags->static_rqs[tag]; | |
5dee8577 | 257 | |
bd166ef1 JA |
258 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
259 | rq->tag = -1; | |
260 | rq->internal_tag = tag; | |
261 | } else { | |
200e86b3 JA |
262 | if (blk_mq_tag_busy(data->hctx)) { |
263 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
264 | atomic_inc(&data->hctx->nr_active); | |
265 | } | |
bd166ef1 JA |
266 | rq->tag = tag; |
267 | rq->internal_tag = -1; | |
562bef42 | 268 | data->hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
269 | } |
270 | ||
ef295ecf | 271 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
272 | return rq; |
273 | } | |
274 | ||
275 | return NULL; | |
276 | } | |
2c3ad667 | 277 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 278 | |
6f3b0e8b CH |
279 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
280 | unsigned int flags) | |
320ae51f | 281 | { |
5a797e00 | 282 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 283 | struct request *rq; |
a492f075 | 284 | int ret; |
320ae51f | 285 | |
6f3b0e8b | 286 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
287 | if (ret) |
288 | return ERR_PTR(ret); | |
320ae51f | 289 | |
bd166ef1 | 290 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 291 | |
bd166ef1 JA |
292 | blk_mq_put_ctx(alloc_data.ctx); |
293 | blk_queue_exit(q); | |
294 | ||
295 | if (!rq) | |
a492f075 | 296 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
297 | |
298 | rq->__data_len = 0; | |
299 | rq->__sector = (sector_t) -1; | |
300 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
301 | return rq; |
302 | } | |
4bb659b1 | 303 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 304 | |
1f5bd336 ML |
305 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
306 | unsigned int flags, unsigned int hctx_idx) | |
307 | { | |
6d2809d5 | 308 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 309 | struct request *rq; |
6d2809d5 | 310 | unsigned int cpu; |
1f5bd336 ML |
311 | int ret; |
312 | ||
313 | /* | |
314 | * If the tag allocator sleeps we could get an allocation for a | |
315 | * different hardware context. No need to complicate the low level | |
316 | * allocator for this for the rare use case of a command tied to | |
317 | * a specific queue. | |
318 | */ | |
319 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
320 | return ERR_PTR(-EINVAL); | |
321 | ||
322 | if (hctx_idx >= q->nr_hw_queues) | |
323 | return ERR_PTR(-EIO); | |
324 | ||
325 | ret = blk_queue_enter(q, true); | |
326 | if (ret) | |
327 | return ERR_PTR(ret); | |
328 | ||
c8712c6a CH |
329 | /* |
330 | * Check if the hardware context is actually mapped to anything. | |
331 | * If not tell the caller that it should skip this queue. | |
332 | */ | |
6d2809d5 OS |
333 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
334 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
335 | blk_queue_exit(q); | |
336 | return ERR_PTR(-EXDEV); | |
c8712c6a | 337 | } |
6d2809d5 OS |
338 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
339 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 340 | |
6d2809d5 | 341 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
c8712c6a | 342 | |
c8712c6a | 343 | blk_queue_exit(q); |
6d2809d5 OS |
344 | |
345 | if (!rq) | |
346 | return ERR_PTR(-EWOULDBLOCK); | |
347 | ||
348 | return rq; | |
1f5bd336 ML |
349 | } |
350 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
351 | ||
bd166ef1 JA |
352 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
353 | struct request *rq) | |
320ae51f | 354 | { |
bd166ef1 | 355 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
356 | struct request_queue *q = rq->q; |
357 | ||
e8064021 | 358 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 359 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
360 | |
361 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 362 | rq->rq_flags = 0; |
0d2602ca | 363 | |
af76e555 | 364 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 365 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
366 | if (rq->tag != -1) |
367 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
368 | if (sched_tag != -1) | |
c05f8525 | 369 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 370 | blk_mq_sched_restart(hctx); |
3ef28e83 | 371 | blk_queue_exit(q); |
320ae51f JA |
372 | } |
373 | ||
bd166ef1 | 374 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 375 | struct request *rq) |
320ae51f JA |
376 | { |
377 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
378 | |
379 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
380 | __blk_mq_finish_request(hctx, ctx, rq); |
381 | } | |
382 | ||
383 | void blk_mq_finish_request(struct request *rq) | |
384 | { | |
385 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 386 | } |
5b727272 | 387 | EXPORT_SYMBOL_GPL(blk_mq_finish_request); |
7c7f2f2b JA |
388 | |
389 | void blk_mq_free_request(struct request *rq) | |
390 | { | |
bd166ef1 | 391 | blk_mq_sched_put_request(rq); |
320ae51f | 392 | } |
1a3b595a | 393 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 394 | |
c8a446ad | 395 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 396 | { |
0d11e6ac ML |
397 | blk_account_io_done(rq); |
398 | ||
91b63639 | 399 | if (rq->end_io) { |
87760e5e | 400 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 401 | rq->end_io(rq, error); |
91b63639 CH |
402 | } else { |
403 | if (unlikely(blk_bidi_rq(rq))) | |
404 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 405 | blk_mq_free_request(rq); |
91b63639 | 406 | } |
320ae51f | 407 | } |
c8a446ad | 408 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 409 | |
c8a446ad | 410 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
411 | { |
412 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
413 | BUG(); | |
c8a446ad | 414 | __blk_mq_end_request(rq, error); |
63151a44 | 415 | } |
c8a446ad | 416 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 417 | |
30a91cb4 | 418 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 419 | { |
3d6efbf6 | 420 | struct request *rq = data; |
320ae51f | 421 | |
30a91cb4 | 422 | rq->q->softirq_done_fn(rq); |
320ae51f | 423 | } |
320ae51f | 424 | |
453f8341 | 425 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
426 | { |
427 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 428 | bool shared = false; |
320ae51f JA |
429 | int cpu; |
430 | ||
453f8341 CH |
431 | if (rq->internal_tag != -1) |
432 | blk_mq_sched_completed_request(rq); | |
433 | if (rq->rq_flags & RQF_STATS) { | |
434 | blk_mq_poll_stats_start(rq->q); | |
435 | blk_stat_add(rq); | |
436 | } | |
437 | ||
38535201 | 438 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
439 | rq->q->softirq_done_fn(rq); |
440 | return; | |
441 | } | |
320ae51f JA |
442 | |
443 | cpu = get_cpu(); | |
38535201 CH |
444 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
445 | shared = cpus_share_cache(cpu, ctx->cpu); | |
446 | ||
447 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 448 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
449 | rq->csd.info = rq; |
450 | rq->csd.flags = 0; | |
c46fff2a | 451 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 452 | } else { |
30a91cb4 | 453 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 454 | } |
320ae51f JA |
455 | put_cpu(); |
456 | } | |
30a91cb4 CH |
457 | |
458 | /** | |
459 | * blk_mq_complete_request - end I/O on a request | |
460 | * @rq: the request being processed | |
461 | * | |
462 | * Description: | |
463 | * Ends all I/O on a request. It does not handle partial completions. | |
464 | * The actual completion happens out-of-order, through a IPI handler. | |
465 | **/ | |
08e0029a | 466 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 467 | { |
95f09684 JA |
468 | struct request_queue *q = rq->q; |
469 | ||
470 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 471 | return; |
08e0029a | 472 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 473 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
474 | } |
475 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 476 | |
973c0191 KB |
477 | int blk_mq_request_started(struct request *rq) |
478 | { | |
479 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
480 | } | |
481 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
482 | ||
e2490073 | 483 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
484 | { |
485 | struct request_queue *q = rq->q; | |
486 | ||
bd166ef1 JA |
487 | blk_mq_sched_started_request(rq); |
488 | ||
320ae51f JA |
489 | trace_block_rq_issue(q, rq); |
490 | ||
cf43e6be | 491 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 492 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 493 | rq->rq_flags |= RQF_STATS; |
87760e5e | 494 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
495 | } |
496 | ||
2b8393b4 | 497 | blk_add_timer(rq); |
87ee7b11 | 498 | |
538b7534 JA |
499 | /* |
500 | * Ensure that ->deadline is visible before set the started | |
501 | * flag and clear the completed flag. | |
502 | */ | |
503 | smp_mb__before_atomic(); | |
504 | ||
87ee7b11 JA |
505 | /* |
506 | * Mark us as started and clear complete. Complete might have been | |
507 | * set if requeue raced with timeout, which then marked it as | |
508 | * complete. So be sure to clear complete again when we start | |
509 | * the request, otherwise we'll ignore the completion event. | |
510 | */ | |
4b570521 JA |
511 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
512 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
513 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
514 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
515 | |
516 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
517 | /* | |
518 | * Make sure space for the drain appears. We know we can do | |
519 | * this because max_hw_segments has been adjusted to be one | |
520 | * fewer than the device can handle. | |
521 | */ | |
522 | rq->nr_phys_segments++; | |
523 | } | |
320ae51f | 524 | } |
e2490073 | 525 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 526 | |
d9d149a3 ML |
527 | /* |
528 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 529 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
530 | * but given rq->deadline is just set in .queue_rq() under |
531 | * this situation, the race won't be possible in reality because | |
532 | * rq->timeout should be set as big enough to cover the window | |
533 | * between blk_mq_start_request() called from .queue_rq() and | |
534 | * clearing REQ_ATOM_STARTED here. | |
535 | */ | |
ed0791b2 | 536 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
537 | { |
538 | struct request_queue *q = rq->q; | |
539 | ||
540 | trace_block_rq_requeue(q, rq); | |
87760e5e | 541 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 542 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 543 | |
e2490073 CH |
544 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
545 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
546 | rq->nr_phys_segments--; | |
547 | } | |
320ae51f JA |
548 | } |
549 | ||
2b053aca | 550 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 551 | { |
ed0791b2 | 552 | __blk_mq_requeue_request(rq); |
ed0791b2 | 553 | |
ed0791b2 | 554 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 555 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
556 | } |
557 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
558 | ||
6fca6a61 CH |
559 | static void blk_mq_requeue_work(struct work_struct *work) |
560 | { | |
561 | struct request_queue *q = | |
2849450a | 562 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
563 | LIST_HEAD(rq_list); |
564 | struct request *rq, *next; | |
565 | unsigned long flags; | |
566 | ||
567 | spin_lock_irqsave(&q->requeue_lock, flags); | |
568 | list_splice_init(&q->requeue_list, &rq_list); | |
569 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
570 | ||
571 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 572 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
573 | continue; |
574 | ||
e8064021 | 575 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 576 | list_del_init(&rq->queuelist); |
bd6737f1 | 577 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
578 | } |
579 | ||
580 | while (!list_empty(&rq_list)) { | |
581 | rq = list_entry(rq_list.next, struct request, queuelist); | |
582 | list_del_init(&rq->queuelist); | |
bd6737f1 | 583 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
584 | } |
585 | ||
52d7f1b5 | 586 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
587 | } |
588 | ||
2b053aca BVA |
589 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
590 | bool kick_requeue_list) | |
6fca6a61 CH |
591 | { |
592 | struct request_queue *q = rq->q; | |
593 | unsigned long flags; | |
594 | ||
595 | /* | |
596 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
597 | * request head insertation from the workqueue. | |
598 | */ | |
e8064021 | 599 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
600 | |
601 | spin_lock_irqsave(&q->requeue_lock, flags); | |
602 | if (at_head) { | |
e8064021 | 603 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
604 | list_add(&rq->queuelist, &q->requeue_list); |
605 | } else { | |
606 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
607 | } | |
608 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
609 | |
610 | if (kick_requeue_list) | |
611 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
612 | } |
613 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
614 | ||
615 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
616 | { | |
2849450a | 617 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
618 | } |
619 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
620 | ||
2849450a MS |
621 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
622 | unsigned long msecs) | |
623 | { | |
624 | kblockd_schedule_delayed_work(&q->requeue_work, | |
625 | msecs_to_jiffies(msecs)); | |
626 | } | |
627 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
628 | ||
1885b24d JA |
629 | void blk_mq_abort_requeue_list(struct request_queue *q) |
630 | { | |
631 | unsigned long flags; | |
632 | LIST_HEAD(rq_list); | |
633 | ||
634 | spin_lock_irqsave(&q->requeue_lock, flags); | |
635 | list_splice_init(&q->requeue_list, &rq_list); | |
636 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
637 | ||
638 | while (!list_empty(&rq_list)) { | |
639 | struct request *rq; | |
640 | ||
641 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
642 | list_del_init(&rq->queuelist); | |
caf7df12 | 643 | blk_mq_end_request(rq, -EIO); |
1885b24d JA |
644 | } |
645 | } | |
646 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
647 | ||
0e62f51f JA |
648 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
649 | { | |
88c7b2b7 JA |
650 | if (tag < tags->nr_tags) { |
651 | prefetch(tags->rqs[tag]); | |
4ee86bab | 652 | return tags->rqs[tag]; |
88c7b2b7 | 653 | } |
4ee86bab HR |
654 | |
655 | return NULL; | |
24d2f903 CH |
656 | } |
657 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
658 | ||
320ae51f | 659 | struct blk_mq_timeout_data { |
46f92d42 CH |
660 | unsigned long next; |
661 | unsigned int next_set; | |
320ae51f JA |
662 | }; |
663 | ||
90415837 | 664 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 665 | { |
f8a5b122 | 666 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 667 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
668 | |
669 | /* | |
670 | * We know that complete is set at this point. If STARTED isn't set | |
671 | * anymore, then the request isn't active and the "timeout" should | |
672 | * just be ignored. This can happen due to the bitflag ordering. | |
673 | * Timeout first checks if STARTED is set, and if it is, assumes | |
674 | * the request is active. But if we race with completion, then | |
48b99c9d | 675 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
676 | * a timeout event with a request that isn't active. |
677 | */ | |
46f92d42 CH |
678 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
679 | return; | |
87ee7b11 | 680 | |
46f92d42 | 681 | if (ops->timeout) |
0152fb6b | 682 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
683 | |
684 | switch (ret) { | |
685 | case BLK_EH_HANDLED: | |
686 | __blk_mq_complete_request(req); | |
687 | break; | |
688 | case BLK_EH_RESET_TIMER: | |
689 | blk_add_timer(req); | |
690 | blk_clear_rq_complete(req); | |
691 | break; | |
692 | case BLK_EH_NOT_HANDLED: | |
693 | break; | |
694 | default: | |
695 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
696 | break; | |
697 | } | |
87ee7b11 | 698 | } |
5b3f25fc | 699 | |
81481eb4 CH |
700 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
701 | struct request *rq, void *priv, bool reserved) | |
702 | { | |
703 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 704 | |
a4ef8e56 | 705 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 706 | return; |
87ee7b11 | 707 | |
d9d149a3 ML |
708 | /* |
709 | * The rq being checked may have been freed and reallocated | |
710 | * out already here, we avoid this race by checking rq->deadline | |
711 | * and REQ_ATOM_COMPLETE flag together: | |
712 | * | |
713 | * - if rq->deadline is observed as new value because of | |
714 | * reusing, the rq won't be timed out because of timing. | |
715 | * - if rq->deadline is observed as previous value, | |
716 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
717 | * because we put a barrier between setting rq->deadline | |
718 | * and clearing the flag in blk_mq_start_request(), so | |
719 | * this rq won't be timed out too. | |
720 | */ | |
46f92d42 CH |
721 | if (time_after_eq(jiffies, rq->deadline)) { |
722 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 723 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
724 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
725 | data->next = rq->deadline; | |
726 | data->next_set = 1; | |
727 | } | |
87ee7b11 JA |
728 | } |
729 | ||
287922eb | 730 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 731 | { |
287922eb CH |
732 | struct request_queue *q = |
733 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
734 | struct blk_mq_timeout_data data = { |
735 | .next = 0, | |
736 | .next_set = 0, | |
737 | }; | |
81481eb4 | 738 | int i; |
320ae51f | 739 | |
71f79fb3 GKB |
740 | /* A deadlock might occur if a request is stuck requiring a |
741 | * timeout at the same time a queue freeze is waiting | |
742 | * completion, since the timeout code would not be able to | |
743 | * acquire the queue reference here. | |
744 | * | |
745 | * That's why we don't use blk_queue_enter here; instead, we use | |
746 | * percpu_ref_tryget directly, because we need to be able to | |
747 | * obtain a reference even in the short window between the queue | |
748 | * starting to freeze, by dropping the first reference in | |
1671d522 | 749 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
750 | * consumed, marked by the instant q_usage_counter reaches |
751 | * zero. | |
752 | */ | |
753 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
754 | return; |
755 | ||
0bf6cd5b | 756 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 757 | |
81481eb4 CH |
758 | if (data.next_set) { |
759 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
760 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 761 | } else { |
0bf6cd5b CH |
762 | struct blk_mq_hw_ctx *hctx; |
763 | ||
f054b56c ML |
764 | queue_for_each_hw_ctx(q, hctx, i) { |
765 | /* the hctx may be unmapped, so check it here */ | |
766 | if (blk_mq_hw_queue_mapped(hctx)) | |
767 | blk_mq_tag_idle(hctx); | |
768 | } | |
0d2602ca | 769 | } |
287922eb | 770 | blk_queue_exit(q); |
320ae51f JA |
771 | } |
772 | ||
773 | /* | |
774 | * Reverse check our software queue for entries that we could potentially | |
775 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
776 | * too much time checking for merges. | |
777 | */ | |
778 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
779 | struct blk_mq_ctx *ctx, struct bio *bio) | |
780 | { | |
781 | struct request *rq; | |
782 | int checked = 8; | |
783 | ||
784 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 785 | bool merged = false; |
320ae51f JA |
786 | |
787 | if (!checked--) | |
788 | break; | |
789 | ||
790 | if (!blk_rq_merge_ok(rq, bio)) | |
791 | continue; | |
792 | ||
34fe7c05 CH |
793 | switch (blk_try_merge(rq, bio)) { |
794 | case ELEVATOR_BACK_MERGE: | |
795 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
796 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 797 | break; |
34fe7c05 CH |
798 | case ELEVATOR_FRONT_MERGE: |
799 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
800 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 801 | break; |
1e739730 CH |
802 | case ELEVATOR_DISCARD_MERGE: |
803 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 804 | break; |
34fe7c05 CH |
805 | default: |
806 | continue; | |
320ae51f | 807 | } |
34fe7c05 CH |
808 | |
809 | if (merged) | |
810 | ctx->rq_merged++; | |
811 | return merged; | |
320ae51f JA |
812 | } |
813 | ||
814 | return false; | |
815 | } | |
816 | ||
88459642 OS |
817 | struct flush_busy_ctx_data { |
818 | struct blk_mq_hw_ctx *hctx; | |
819 | struct list_head *list; | |
820 | }; | |
821 | ||
822 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
823 | { | |
824 | struct flush_busy_ctx_data *flush_data = data; | |
825 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
826 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
827 | ||
828 | sbitmap_clear_bit(sb, bitnr); | |
829 | spin_lock(&ctx->lock); | |
830 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
831 | spin_unlock(&ctx->lock); | |
832 | return true; | |
833 | } | |
834 | ||
1429d7c9 JA |
835 | /* |
836 | * Process software queues that have been marked busy, splicing them | |
837 | * to the for-dispatch | |
838 | */ | |
2c3ad667 | 839 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 840 | { |
88459642 OS |
841 | struct flush_busy_ctx_data data = { |
842 | .hctx = hctx, | |
843 | .list = list, | |
844 | }; | |
1429d7c9 | 845 | |
88459642 | 846 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 847 | } |
2c3ad667 | 848 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 849 | |
703fd1c0 JA |
850 | static inline unsigned int queued_to_index(unsigned int queued) |
851 | { | |
852 | if (!queued) | |
853 | return 0; | |
1429d7c9 | 854 | |
703fd1c0 | 855 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
856 | } |
857 | ||
bd6737f1 JA |
858 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
859 | bool wait) | |
bd166ef1 JA |
860 | { |
861 | struct blk_mq_alloc_data data = { | |
862 | .q = rq->q, | |
bd166ef1 JA |
863 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
864 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
865 | }; | |
866 | ||
5feeacdd JA |
867 | might_sleep_if(wait); |
868 | ||
81380ca1 OS |
869 | if (rq->tag != -1) |
870 | goto done; | |
bd166ef1 | 871 | |
415b806d SG |
872 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
873 | data.flags |= BLK_MQ_REQ_RESERVED; | |
874 | ||
bd166ef1 JA |
875 | rq->tag = blk_mq_get_tag(&data); |
876 | if (rq->tag >= 0) { | |
200e86b3 JA |
877 | if (blk_mq_tag_busy(data.hctx)) { |
878 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
879 | atomic_inc(&data.hctx->nr_active); | |
880 | } | |
bd166ef1 | 881 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
882 | } |
883 | ||
81380ca1 OS |
884 | done: |
885 | if (hctx) | |
886 | *hctx = data.hctx; | |
887 | return rq->tag != -1; | |
bd166ef1 JA |
888 | } |
889 | ||
113285b4 JA |
890 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
891 | struct request *rq) | |
99cf1dc5 | 892 | { |
99cf1dc5 JA |
893 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
894 | rq->tag = -1; | |
895 | ||
896 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
897 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
898 | atomic_dec(&hctx->nr_active); | |
899 | } | |
900 | } | |
901 | ||
113285b4 JA |
902 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
903 | struct request *rq) | |
904 | { | |
905 | if (rq->tag == -1 || rq->internal_tag == -1) | |
906 | return; | |
907 | ||
908 | __blk_mq_put_driver_tag(hctx, rq); | |
909 | } | |
910 | ||
911 | static void blk_mq_put_driver_tag(struct request *rq) | |
912 | { | |
913 | struct blk_mq_hw_ctx *hctx; | |
914 | ||
915 | if (rq->tag == -1 || rq->internal_tag == -1) | |
916 | return; | |
917 | ||
918 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
919 | __blk_mq_put_driver_tag(hctx, rq); | |
920 | } | |
921 | ||
bd166ef1 JA |
922 | /* |
923 | * If we fail getting a driver tag because all the driver tags are already | |
924 | * assigned and on the dispatch list, BUT the first entry does not have a | |
925 | * tag, then we could deadlock. For that case, move entries with assigned | |
926 | * driver tags to the front, leaving the set of tagged requests in the | |
927 | * same order, and the untagged set in the same order. | |
928 | */ | |
929 | static bool reorder_tags_to_front(struct list_head *list) | |
930 | { | |
931 | struct request *rq, *tmp, *first = NULL; | |
932 | ||
933 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
934 | if (rq == first) | |
935 | break; | |
936 | if (rq->tag != -1) { | |
937 | list_move(&rq->queuelist, list); | |
938 | if (!first) | |
939 | first = rq; | |
940 | } | |
941 | } | |
942 | ||
943 | return first != NULL; | |
944 | } | |
945 | ||
da55f2cc OS |
946 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
947 | void *key) | |
948 | { | |
949 | struct blk_mq_hw_ctx *hctx; | |
950 | ||
951 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
952 | ||
953 | list_del(&wait->task_list); | |
954 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
955 | blk_mq_run_hw_queue(hctx, true); | |
956 | return 1; | |
957 | } | |
958 | ||
959 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
960 | { | |
961 | struct sbq_wait_state *ws; | |
962 | ||
963 | /* | |
964 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
965 | * The thread which wins the race to grab this bit adds the hardware | |
966 | * queue to the wait queue. | |
967 | */ | |
968 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
969 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
970 | return false; | |
971 | ||
972 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
973 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
974 | ||
975 | /* | |
976 | * As soon as this returns, it's no longer safe to fiddle with | |
977 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
978 | * and unlock the bit. | |
979 | */ | |
980 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
981 | return true; | |
982 | } | |
983 | ||
81380ca1 | 984 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
320ae51f | 985 | { |
81380ca1 | 986 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 987 | struct request *rq; |
3e8a7069 | 988 | int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 989 | |
81380ca1 OS |
990 | if (list_empty(list)) |
991 | return false; | |
992 | ||
320ae51f JA |
993 | /* |
994 | * Now process all the entries, sending them to the driver. | |
995 | */ | |
3e8a7069 | 996 | errors = queued = 0; |
81380ca1 | 997 | do { |
74c45052 | 998 | struct blk_mq_queue_data bd; |
320ae51f | 999 | |
f04c3df3 | 1000 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1001 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1002 | if (!queued && reorder_tags_to_front(list)) | |
1003 | continue; | |
3c782d67 JA |
1004 | |
1005 | /* | |
da55f2cc OS |
1006 | * The initial allocation attempt failed, so we need to |
1007 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1008 | */ |
807b1041 OS |
1009 | if (!blk_mq_dispatch_wait_add(hctx)) |
1010 | break; | |
1011 | ||
1012 | /* | |
1013 | * It's possible that a tag was freed in the window | |
1014 | * between the allocation failure and adding the | |
1015 | * hardware queue to the wait queue. | |
1016 | */ | |
1017 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
3c782d67 | 1018 | break; |
bd166ef1 | 1019 | } |
da55f2cc | 1020 | |
320ae51f | 1021 | list_del_init(&rq->queuelist); |
320ae51f | 1022 | |
74c45052 | 1023 | bd.rq = rq; |
113285b4 JA |
1024 | |
1025 | /* | |
1026 | * Flag last if we have no more requests, or if we have more | |
1027 | * but can't assign a driver tag to it. | |
1028 | */ | |
1029 | if (list_empty(list)) | |
1030 | bd.last = true; | |
1031 | else { | |
1032 | struct request *nxt; | |
1033 | ||
1034 | nxt = list_first_entry(list, struct request, queuelist); | |
1035 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1036 | } | |
74c45052 JA |
1037 | |
1038 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
1039 | switch (ret) { |
1040 | case BLK_MQ_RQ_QUEUE_OK: | |
1041 | queued++; | |
52b9c330 | 1042 | break; |
320ae51f | 1043 | case BLK_MQ_RQ_QUEUE_BUSY: |
113285b4 | 1044 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1045 | list_add(&rq->queuelist, list); |
ed0791b2 | 1046 | __blk_mq_requeue_request(rq); |
320ae51f JA |
1047 | break; |
1048 | default: | |
1049 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 1050 | case BLK_MQ_RQ_QUEUE_ERROR: |
3e8a7069 | 1051 | errors++; |
caf7df12 | 1052 | blk_mq_end_request(rq, -EIO); |
320ae51f JA |
1053 | break; |
1054 | } | |
1055 | ||
1056 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1057 | break; | |
81380ca1 | 1058 | } while (!list_empty(list)); |
320ae51f | 1059 | |
703fd1c0 | 1060 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1061 | |
1062 | /* | |
1063 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1064 | * that is where we will continue on next queue run. | |
1065 | */ | |
f04c3df3 | 1066 | if (!list_empty(list)) { |
113285b4 | 1067 | /* |
710c785f BVA |
1068 | * If an I/O scheduler has been configured and we got a driver |
1069 | * tag for the next request already, free it again. | |
113285b4 JA |
1070 | */ |
1071 | rq = list_first_entry(list, struct request, queuelist); | |
1072 | blk_mq_put_driver_tag(rq); | |
1073 | ||
320ae51f | 1074 | spin_lock(&hctx->lock); |
c13660a0 | 1075 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1076 | spin_unlock(&hctx->lock); |
f04c3df3 | 1077 | |
9ba52e58 | 1078 | /* |
710c785f BVA |
1079 | * If SCHED_RESTART was set by the caller of this function and |
1080 | * it is no longer set that means that it was cleared by another | |
1081 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1082 | * |
710c785f BVA |
1083 | * If TAG_WAITING is set that means that an I/O scheduler has |
1084 | * been configured and another thread is waiting for a driver | |
1085 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1086 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1087 | * |
710c785f BVA |
1088 | * If no I/O scheduler has been configured it is possible that |
1089 | * the hardware queue got stopped and restarted before requests | |
1090 | * were pushed back onto the dispatch list. Rerun the queue to | |
1091 | * avoid starvation. Notes: | |
1092 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1093 | * been stopped before rerunning a queue. | |
1094 | * - Some but not all block drivers stop a queue before | |
1095 | * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq | |
1096 | * and dm-rq. | |
bd166ef1 | 1097 | */ |
da55f2cc OS |
1098 | if (!blk_mq_sched_needs_restart(hctx) && |
1099 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1100 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1101 | } |
f04c3df3 | 1102 | |
3e8a7069 | 1103 | return (queued + errors) != 0; |
f04c3df3 JA |
1104 | } |
1105 | ||
6a83e74d BVA |
1106 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1107 | { | |
1108 | int srcu_idx; | |
1109 | ||
1110 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1111 | cpu_online(hctx->next_cpu)); | |
1112 | ||
1113 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1114 | rcu_read_lock(); | |
bd166ef1 | 1115 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1116 | rcu_read_unlock(); |
1117 | } else { | |
bf4907c0 JA |
1118 | might_sleep(); |
1119 | ||
6a83e74d | 1120 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); |
bd166ef1 | 1121 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1122 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1123 | } | |
1124 | } | |
1125 | ||
506e931f JA |
1126 | /* |
1127 | * It'd be great if the workqueue API had a way to pass | |
1128 | * in a mask and had some smarts for more clever placement. | |
1129 | * For now we just round-robin here, switching for every | |
1130 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1131 | */ | |
1132 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1133 | { | |
b657d7e6 CH |
1134 | if (hctx->queue->nr_hw_queues == 1) |
1135 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1136 | |
1137 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1138 | int next_cpu; |
506e931f JA |
1139 | |
1140 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1141 | if (next_cpu >= nr_cpu_ids) | |
1142 | next_cpu = cpumask_first(hctx->cpumask); | |
1143 | ||
1144 | hctx->next_cpu = next_cpu; | |
1145 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1146 | } | |
1147 | ||
b657d7e6 | 1148 | return hctx->next_cpu; |
506e931f JA |
1149 | } |
1150 | ||
7587a5ae BVA |
1151 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1152 | unsigned long msecs) | |
320ae51f | 1153 | { |
5d1b25c1 BVA |
1154 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1155 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1156 | return; |
1157 | ||
1b792f2f | 1158 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1159 | int cpu = get_cpu(); |
1160 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1161 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1162 | put_cpu(); |
398205b8 PB |
1163 | return; |
1164 | } | |
e4043dcf | 1165 | |
2a90d4aa | 1166 | put_cpu(); |
e4043dcf | 1167 | } |
398205b8 | 1168 | |
7587a5ae BVA |
1169 | if (msecs == 0) |
1170 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), | |
1171 | &hctx->run_work); | |
1172 | else | |
1173 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1174 | &hctx->delayed_run_work, | |
1175 | msecs_to_jiffies(msecs)); | |
1176 | } | |
1177 | ||
1178 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1179 | { | |
1180 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1181 | } | |
1182 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1183 | ||
1184 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1185 | { | |
1186 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1187 | } |
5b727272 | 1188 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1189 | |
b94ec296 | 1190 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1191 | { |
1192 | struct blk_mq_hw_ctx *hctx; | |
1193 | int i; | |
1194 | ||
1195 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1196 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1197 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1198 | continue; |
1199 | ||
b94ec296 | 1200 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1201 | } |
1202 | } | |
b94ec296 | 1203 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1204 | |
fd001443 BVA |
1205 | /** |
1206 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1207 | * @q: request queue. | |
1208 | * | |
1209 | * The caller is responsible for serializing this function against | |
1210 | * blk_mq_{start,stop}_hw_queue(). | |
1211 | */ | |
1212 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1213 | { | |
1214 | struct blk_mq_hw_ctx *hctx; | |
1215 | int i; | |
1216 | ||
1217 | queue_for_each_hw_ctx(q, hctx, i) | |
1218 | if (blk_mq_hctx_stopped(hctx)) | |
1219 | return true; | |
1220 | ||
1221 | return false; | |
1222 | } | |
1223 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1224 | ||
320ae51f JA |
1225 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1226 | { | |
27489a3c | 1227 | cancel_work(&hctx->run_work); |
70f4db63 | 1228 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1229 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1230 | } | |
1231 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1232 | ||
280d45f6 CH |
1233 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1234 | { | |
1235 | struct blk_mq_hw_ctx *hctx; | |
1236 | int i; | |
1237 | ||
1238 | queue_for_each_hw_ctx(q, hctx, i) | |
1239 | blk_mq_stop_hw_queue(hctx); | |
1240 | } | |
1241 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1242 | ||
320ae51f JA |
1243 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1244 | { | |
1245 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1246 | |
0ffbce80 | 1247 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1248 | } |
1249 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1250 | ||
2f268556 CH |
1251 | void blk_mq_start_hw_queues(struct request_queue *q) |
1252 | { | |
1253 | struct blk_mq_hw_ctx *hctx; | |
1254 | int i; | |
1255 | ||
1256 | queue_for_each_hw_ctx(q, hctx, i) | |
1257 | blk_mq_start_hw_queue(hctx); | |
1258 | } | |
1259 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1260 | ||
ae911c5e JA |
1261 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1262 | { | |
1263 | if (!blk_mq_hctx_stopped(hctx)) | |
1264 | return; | |
1265 | ||
1266 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1267 | blk_mq_run_hw_queue(hctx, async); | |
1268 | } | |
1269 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1270 | ||
1b4a3258 | 1271 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1272 | { |
1273 | struct blk_mq_hw_ctx *hctx; | |
1274 | int i; | |
1275 | ||
ae911c5e JA |
1276 | queue_for_each_hw_ctx(q, hctx, i) |
1277 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1278 | } |
1279 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1280 | ||
70f4db63 | 1281 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1282 | { |
1283 | struct blk_mq_hw_ctx *hctx; | |
1284 | ||
27489a3c | 1285 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1286 | |
320ae51f JA |
1287 | __blk_mq_run_hw_queue(hctx); |
1288 | } | |
1289 | ||
7587a5ae BVA |
1290 | static void blk_mq_delayed_run_work_fn(struct work_struct *work) |
1291 | { | |
1292 | struct blk_mq_hw_ctx *hctx; | |
1293 | ||
1294 | hctx = container_of(work, struct blk_mq_hw_ctx, delayed_run_work.work); | |
1295 | ||
1296 | __blk_mq_run_hw_queue(hctx); | |
1297 | } | |
1298 | ||
70f4db63 CH |
1299 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1300 | { | |
1301 | struct blk_mq_hw_ctx *hctx; | |
1302 | ||
1303 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1304 | ||
1305 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1306 | __blk_mq_run_hw_queue(hctx); | |
1307 | } | |
1308 | ||
1309 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1310 | { | |
19c66e59 ML |
1311 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1312 | return; | |
70f4db63 | 1313 | |
7e79dadc | 1314 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1315 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1316 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1317 | } |
1318 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1319 | ||
cfd0c552 | 1320 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1321 | struct request *rq, |
1322 | bool at_head) | |
320ae51f | 1323 | { |
e57690fe JA |
1324 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1325 | ||
01b983c9 JA |
1326 | trace_block_rq_insert(hctx->queue, rq); |
1327 | ||
72a0a36e CH |
1328 | if (at_head) |
1329 | list_add(&rq->queuelist, &ctx->rq_list); | |
1330 | else | |
1331 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1332 | } |
4bb659b1 | 1333 | |
2c3ad667 JA |
1334 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1335 | bool at_head) | |
cfd0c552 ML |
1336 | { |
1337 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1338 | ||
e57690fe | 1339 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1340 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1341 | } |
1342 | ||
bd166ef1 JA |
1343 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1344 | struct list_head *list) | |
320ae51f JA |
1345 | |
1346 | { | |
320ae51f JA |
1347 | /* |
1348 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1349 | * offline now | |
1350 | */ | |
1351 | spin_lock(&ctx->lock); | |
1352 | while (!list_empty(list)) { | |
1353 | struct request *rq; | |
1354 | ||
1355 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1356 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1357 | list_del_init(&rq->queuelist); |
e57690fe | 1358 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1359 | } |
cfd0c552 | 1360 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1361 | spin_unlock(&ctx->lock); |
320ae51f JA |
1362 | } |
1363 | ||
1364 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1365 | { | |
1366 | struct request *rqa = container_of(a, struct request, queuelist); | |
1367 | struct request *rqb = container_of(b, struct request, queuelist); | |
1368 | ||
1369 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1370 | (rqa->mq_ctx == rqb->mq_ctx && | |
1371 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1372 | } | |
1373 | ||
1374 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1375 | { | |
1376 | struct blk_mq_ctx *this_ctx; | |
1377 | struct request_queue *this_q; | |
1378 | struct request *rq; | |
1379 | LIST_HEAD(list); | |
1380 | LIST_HEAD(ctx_list); | |
1381 | unsigned int depth; | |
1382 | ||
1383 | list_splice_init(&plug->mq_list, &list); | |
1384 | ||
1385 | list_sort(NULL, &list, plug_ctx_cmp); | |
1386 | ||
1387 | this_q = NULL; | |
1388 | this_ctx = NULL; | |
1389 | depth = 0; | |
1390 | ||
1391 | while (!list_empty(&list)) { | |
1392 | rq = list_entry_rq(list.next); | |
1393 | list_del_init(&rq->queuelist); | |
1394 | BUG_ON(!rq->q); | |
1395 | if (rq->mq_ctx != this_ctx) { | |
1396 | if (this_ctx) { | |
bd166ef1 JA |
1397 | trace_block_unplug(this_q, depth, from_schedule); |
1398 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1399 | &ctx_list, | |
1400 | from_schedule); | |
320ae51f JA |
1401 | } |
1402 | ||
1403 | this_ctx = rq->mq_ctx; | |
1404 | this_q = rq->q; | |
1405 | depth = 0; | |
1406 | } | |
1407 | ||
1408 | depth++; | |
1409 | list_add_tail(&rq->queuelist, &ctx_list); | |
1410 | } | |
1411 | ||
1412 | /* | |
1413 | * If 'this_ctx' is set, we know we have entries to complete | |
1414 | * on 'ctx_list'. Do those. | |
1415 | */ | |
1416 | if (this_ctx) { | |
bd166ef1 JA |
1417 | trace_block_unplug(this_q, depth, from_schedule); |
1418 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1419 | from_schedule); | |
320ae51f JA |
1420 | } |
1421 | } | |
1422 | ||
1423 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1424 | { | |
da8d7f07 | 1425 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1426 | |
6e85eaf3 | 1427 | blk_account_io_start(rq, true); |
320ae51f JA |
1428 | } |
1429 | ||
274a5843 JA |
1430 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1431 | { | |
1432 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1433 | !blk_queue_nomerges(hctx->queue); | |
1434 | } | |
1435 | ||
07068d5b JA |
1436 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1437 | struct blk_mq_ctx *ctx, | |
1438 | struct request *rq, struct bio *bio) | |
320ae51f | 1439 | { |
e18378a6 | 1440 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1441 | blk_mq_bio_to_request(rq, bio); |
1442 | spin_lock(&ctx->lock); | |
1443 | insert_rq: | |
1444 | __blk_mq_insert_request(hctx, rq, false); | |
1445 | spin_unlock(&ctx->lock); | |
1446 | return false; | |
1447 | } else { | |
274a5843 JA |
1448 | struct request_queue *q = hctx->queue; |
1449 | ||
07068d5b JA |
1450 | spin_lock(&ctx->lock); |
1451 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1452 | blk_mq_bio_to_request(rq, bio); | |
1453 | goto insert_rq; | |
1454 | } | |
320ae51f | 1455 | |
07068d5b | 1456 | spin_unlock(&ctx->lock); |
bd166ef1 | 1457 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1458 | return true; |
14ec77f3 | 1459 | } |
07068d5b | 1460 | } |
14ec77f3 | 1461 | |
fd2d3326 JA |
1462 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1463 | { | |
bd166ef1 JA |
1464 | if (rq->tag != -1) |
1465 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1466 | ||
1467 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1468 | } |
1469 | ||
5eb6126e | 1470 | static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie, |
9c621104 | 1471 | bool may_sleep) |
f984df1f | 1472 | { |
f984df1f | 1473 | struct request_queue *q = rq->q; |
f984df1f SL |
1474 | struct blk_mq_queue_data bd = { |
1475 | .rq = rq, | |
d945a365 | 1476 | .last = true, |
f984df1f | 1477 | }; |
bd166ef1 JA |
1478 | struct blk_mq_hw_ctx *hctx; |
1479 | blk_qc_t new_cookie; | |
1480 | int ret; | |
f984df1f | 1481 | |
bd166ef1 | 1482 | if (q->elevator) |
2253efc8 BVA |
1483 | goto insert; |
1484 | ||
bd166ef1 JA |
1485 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1486 | goto insert; | |
1487 | ||
1488 | new_cookie = request_to_qc_t(hctx, rq); | |
1489 | ||
f984df1f SL |
1490 | /* |
1491 | * For OK queue, we are done. For error, kill it. Any other | |
1492 | * error (busy), just add it to our list as we previously | |
1493 | * would have done | |
1494 | */ | |
1495 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1496 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1497 | *cookie = new_cookie; | |
2253efc8 | 1498 | return; |
7b371636 | 1499 | } |
f984df1f | 1500 | |
7b371636 JA |
1501 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
1502 | *cookie = BLK_QC_T_NONE; | |
caf7df12 | 1503 | blk_mq_end_request(rq, -EIO); |
2253efc8 | 1504 | return; |
f984df1f | 1505 | } |
7b371636 | 1506 | |
b58e1769 | 1507 | __blk_mq_requeue_request(rq); |
2253efc8 | 1508 | insert: |
9c621104 | 1509 | blk_mq_sched_insert_request(rq, false, true, false, may_sleep); |
f984df1f SL |
1510 | } |
1511 | ||
5eb6126e CH |
1512 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1513 | struct request *rq, blk_qc_t *cookie) | |
1514 | { | |
1515 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1516 | rcu_read_lock(); | |
1517 | __blk_mq_try_issue_directly(rq, cookie, false); | |
1518 | rcu_read_unlock(); | |
1519 | } else { | |
bf4907c0 JA |
1520 | unsigned int srcu_idx; |
1521 | ||
1522 | might_sleep(); | |
1523 | ||
1524 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
5eb6126e CH |
1525 | __blk_mq_try_issue_directly(rq, cookie, true); |
1526 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
1527 | } | |
1528 | } | |
1529 | ||
dece1635 | 1530 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1531 | { |
ef295ecf | 1532 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1533 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1534 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1535 | struct request *rq; |
5eb6126e | 1536 | unsigned int request_count = 0; |
f984df1f | 1537 | struct blk_plug *plug; |
5b3f341f | 1538 | struct request *same_queue_rq = NULL; |
7b371636 | 1539 | blk_qc_t cookie; |
87760e5e | 1540 | unsigned int wb_acct; |
07068d5b JA |
1541 | |
1542 | blk_queue_bounce(q, &bio); | |
1543 | ||
1544 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1545 | bio_io_error(bio); |
dece1635 | 1546 | return BLK_QC_T_NONE; |
07068d5b JA |
1547 | } |
1548 | ||
54efd50b KO |
1549 | blk_queue_split(q, &bio, q->bio_split); |
1550 | ||
87c279e6 OS |
1551 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1552 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1553 | return BLK_QC_T_NONE; | |
f984df1f | 1554 | |
bd166ef1 JA |
1555 | if (blk_mq_sched_bio_merge(q, bio)) |
1556 | return BLK_QC_T_NONE; | |
1557 | ||
87760e5e JA |
1558 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1559 | ||
bd166ef1 JA |
1560 | trace_block_getrq(q, bio, bio->bi_opf); |
1561 | ||
1562 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1563 | if (unlikely(!rq)) { |
1564 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1565 | return BLK_QC_T_NONE; |
87760e5e JA |
1566 | } |
1567 | ||
1568 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1569 | |
fd2d3326 | 1570 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1571 | |
a4d907b6 | 1572 | plug = current->plug; |
07068d5b | 1573 | if (unlikely(is_flush_fua)) { |
b00c53e8 | 1574 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1575 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1576 | if (q->elevator) { |
1577 | blk_mq_sched_insert_request(rq, false, true, true, | |
1578 | true); | |
1579 | } else { | |
1580 | blk_insert_flush(rq); | |
1581 | blk_mq_run_hw_queue(data.hctx, true); | |
1582 | } | |
1583 | } else if (plug && q->nr_hw_queues == 1) { | |
254d259d CH |
1584 | struct request *last = NULL; |
1585 | ||
b00c53e8 | 1586 | blk_mq_put_ctx(data.ctx); |
254d259d CH |
1587 | blk_mq_bio_to_request(rq, bio); |
1588 | ||
1589 | /* | |
1590 | * @request_count may become stale because of schedule | |
1591 | * out, so check the list again. | |
1592 | */ | |
1593 | if (list_empty(&plug->mq_list)) | |
1594 | request_count = 0; | |
1595 | else if (blk_queue_nomerges(q)) | |
1596 | request_count = blk_plug_queued_count(q); | |
1597 | ||
1598 | if (!request_count) | |
1599 | trace_block_plug(q); | |
1600 | else | |
1601 | last = list_entry_rq(plug->mq_list.prev); | |
1602 | ||
254d259d CH |
1603 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1604 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1605 | blk_flush_plug_list(plug, false); | |
1606 | trace_block_plug(q); | |
1607 | } | |
1608 | ||
1609 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2299722c | 1610 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 1611 | blk_mq_bio_to_request(rq, bio); |
07068d5b JA |
1612 | |
1613 | /* | |
6a83e74d | 1614 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1615 | * Otherwise the existing request in the plug list will be |
1616 | * issued. So the plug list will have one request at most | |
2299722c CH |
1617 | * The plug list might get flushed before this. If that happens, |
1618 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1619 | */ |
2299722c CH |
1620 | if (list_empty(&plug->mq_list)) |
1621 | same_queue_rq = NULL; | |
1622 | if (same_queue_rq) | |
1623 | list_del_init(&same_queue_rq->queuelist); | |
1624 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1625 | ||
bf4907c0 JA |
1626 | blk_mq_put_ctx(data.ctx); |
1627 | ||
2299722c CH |
1628 | if (same_queue_rq) |
1629 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1630 | &cookie); | |
a4d907b6 | 1631 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1632 | blk_mq_put_ctx(data.ctx); |
2299722c | 1633 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1634 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1635 | } else if (q->elevator) { |
b00c53e8 | 1636 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1637 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1638 | blk_mq_sched_insert_request(rq, false, true, true, true); |
b00c53e8 JA |
1639 | } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1640 | blk_mq_put_ctx(data.ctx); | |
a4d907b6 | 1641 | blk_mq_run_hw_queue(data.hctx, true); |
b00c53e8 | 1642 | } |
a4d907b6 | 1643 | |
7b371636 | 1644 | return cookie; |
07068d5b JA |
1645 | } |
1646 | ||
cc71a6f4 JA |
1647 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1648 | unsigned int hctx_idx) | |
95363efd | 1649 | { |
e9b267d9 | 1650 | struct page *page; |
320ae51f | 1651 | |
24d2f903 | 1652 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1653 | int i; |
320ae51f | 1654 | |
24d2f903 | 1655 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1656 | struct request *rq = tags->static_rqs[i]; |
1657 | ||
1658 | if (!rq) | |
e9b267d9 | 1659 | continue; |
2af8cbe3 | 1660 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1661 | hctx_idx, i); |
2af8cbe3 | 1662 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1663 | } |
320ae51f | 1664 | } |
320ae51f | 1665 | |
24d2f903 CH |
1666 | while (!list_empty(&tags->page_list)) { |
1667 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1668 | list_del_init(&page->lru); |
f75782e4 CM |
1669 | /* |
1670 | * Remove kmemleak object previously allocated in | |
1671 | * blk_mq_init_rq_map(). | |
1672 | */ | |
1673 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1674 | __free_pages(page, page->private); |
1675 | } | |
cc71a6f4 | 1676 | } |
320ae51f | 1677 | |
cc71a6f4 JA |
1678 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1679 | { | |
24d2f903 | 1680 | kfree(tags->rqs); |
cc71a6f4 | 1681 | tags->rqs = NULL; |
2af8cbe3 JA |
1682 | kfree(tags->static_rqs); |
1683 | tags->static_rqs = NULL; | |
320ae51f | 1684 | |
24d2f903 | 1685 | blk_mq_free_tags(tags); |
320ae51f JA |
1686 | } |
1687 | ||
cc71a6f4 JA |
1688 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1689 | unsigned int hctx_idx, | |
1690 | unsigned int nr_tags, | |
1691 | unsigned int reserved_tags) | |
320ae51f | 1692 | { |
24d2f903 | 1693 | struct blk_mq_tags *tags; |
59f082e4 | 1694 | int node; |
320ae51f | 1695 | |
59f082e4 SL |
1696 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1697 | if (node == NUMA_NO_NODE) | |
1698 | node = set->numa_node; | |
1699 | ||
1700 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1701 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1702 | if (!tags) |
1703 | return NULL; | |
320ae51f | 1704 | |
cc71a6f4 | 1705 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1706 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1707 | node); |
24d2f903 CH |
1708 | if (!tags->rqs) { |
1709 | blk_mq_free_tags(tags); | |
1710 | return NULL; | |
1711 | } | |
320ae51f | 1712 | |
2af8cbe3 JA |
1713 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1714 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1715 | node); |
2af8cbe3 JA |
1716 | if (!tags->static_rqs) { |
1717 | kfree(tags->rqs); | |
1718 | blk_mq_free_tags(tags); | |
1719 | return NULL; | |
1720 | } | |
1721 | ||
cc71a6f4 JA |
1722 | return tags; |
1723 | } | |
1724 | ||
1725 | static size_t order_to_size(unsigned int order) | |
1726 | { | |
1727 | return (size_t)PAGE_SIZE << order; | |
1728 | } | |
1729 | ||
1730 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1731 | unsigned int hctx_idx, unsigned int depth) | |
1732 | { | |
1733 | unsigned int i, j, entries_per_page, max_order = 4; | |
1734 | size_t rq_size, left; | |
59f082e4 SL |
1735 | int node; |
1736 | ||
1737 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1738 | if (node == NUMA_NO_NODE) | |
1739 | node = set->numa_node; | |
cc71a6f4 JA |
1740 | |
1741 | INIT_LIST_HEAD(&tags->page_list); | |
1742 | ||
320ae51f JA |
1743 | /* |
1744 | * rq_size is the size of the request plus driver payload, rounded | |
1745 | * to the cacheline size | |
1746 | */ | |
24d2f903 | 1747 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1748 | cache_line_size()); |
cc71a6f4 | 1749 | left = rq_size * depth; |
320ae51f | 1750 | |
cc71a6f4 | 1751 | for (i = 0; i < depth; ) { |
320ae51f JA |
1752 | int this_order = max_order; |
1753 | struct page *page; | |
1754 | int to_do; | |
1755 | void *p; | |
1756 | ||
b3a834b1 | 1757 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1758 | this_order--; |
1759 | ||
1760 | do { | |
59f082e4 | 1761 | page = alloc_pages_node(node, |
36e1f3d1 | 1762 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1763 | this_order); |
320ae51f JA |
1764 | if (page) |
1765 | break; | |
1766 | if (!this_order--) | |
1767 | break; | |
1768 | if (order_to_size(this_order) < rq_size) | |
1769 | break; | |
1770 | } while (1); | |
1771 | ||
1772 | if (!page) | |
24d2f903 | 1773 | goto fail; |
320ae51f JA |
1774 | |
1775 | page->private = this_order; | |
24d2f903 | 1776 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1777 | |
1778 | p = page_address(page); | |
f75782e4 CM |
1779 | /* |
1780 | * Allow kmemleak to scan these pages as they contain pointers | |
1781 | * to additional allocations like via ops->init_request(). | |
1782 | */ | |
36e1f3d1 | 1783 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1784 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1785 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1786 | left -= to_do * rq_size; |
1787 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1788 | struct request *rq = p; |
1789 | ||
1790 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1791 | if (set->ops->init_request) { |
1792 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1793 | rq, hctx_idx, i, |
59f082e4 | 1794 | node)) { |
2af8cbe3 | 1795 | tags->static_rqs[i] = NULL; |
24d2f903 | 1796 | goto fail; |
a5164405 | 1797 | } |
e9b267d9 CH |
1798 | } |
1799 | ||
320ae51f JA |
1800 | p += rq_size; |
1801 | i++; | |
1802 | } | |
1803 | } | |
cc71a6f4 | 1804 | return 0; |
320ae51f | 1805 | |
24d2f903 | 1806 | fail: |
cc71a6f4 JA |
1807 | blk_mq_free_rqs(set, tags, hctx_idx); |
1808 | return -ENOMEM; | |
320ae51f JA |
1809 | } |
1810 | ||
e57690fe JA |
1811 | /* |
1812 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1813 | * software queue to the hw queue dispatch list, and ensure that it | |
1814 | * gets run. | |
1815 | */ | |
9467f859 | 1816 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1817 | { |
9467f859 | 1818 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1819 | struct blk_mq_ctx *ctx; |
1820 | LIST_HEAD(tmp); | |
1821 | ||
9467f859 | 1822 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1823 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1824 | |
1825 | spin_lock(&ctx->lock); | |
1826 | if (!list_empty(&ctx->rq_list)) { | |
1827 | list_splice_init(&ctx->rq_list, &tmp); | |
1828 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1829 | } | |
1830 | spin_unlock(&ctx->lock); | |
1831 | ||
1832 | if (list_empty(&tmp)) | |
9467f859 | 1833 | return 0; |
484b4061 | 1834 | |
e57690fe JA |
1835 | spin_lock(&hctx->lock); |
1836 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1837 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1838 | |
1839 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1840 | return 0; |
484b4061 JA |
1841 | } |
1842 | ||
9467f859 | 1843 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1844 | { |
9467f859 TG |
1845 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1846 | &hctx->cpuhp_dead); | |
484b4061 JA |
1847 | } |
1848 | ||
c3b4afca | 1849 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1850 | static void blk_mq_exit_hctx(struct request_queue *q, |
1851 | struct blk_mq_tag_set *set, | |
1852 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1853 | { | |
f70ced09 ML |
1854 | unsigned flush_start_tag = set->queue_depth; |
1855 | ||
08e98fc6 ML |
1856 | blk_mq_tag_idle(hctx); |
1857 | ||
f70ced09 ML |
1858 | if (set->ops->exit_request) |
1859 | set->ops->exit_request(set->driver_data, | |
1860 | hctx->fq->flush_rq, hctx_idx, | |
1861 | flush_start_tag + hctx_idx); | |
1862 | ||
93252632 OS |
1863 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1864 | ||
08e98fc6 ML |
1865 | if (set->ops->exit_hctx) |
1866 | set->ops->exit_hctx(hctx, hctx_idx); | |
1867 | ||
6a83e74d BVA |
1868 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1869 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1870 | ||
9467f859 | 1871 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1872 | blk_free_flush_queue(hctx->fq); |
88459642 | 1873 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1874 | } |
1875 | ||
624dbe47 ML |
1876 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1877 | struct blk_mq_tag_set *set, int nr_queue) | |
1878 | { | |
1879 | struct blk_mq_hw_ctx *hctx; | |
1880 | unsigned int i; | |
1881 | ||
1882 | queue_for_each_hw_ctx(q, hctx, i) { | |
1883 | if (i == nr_queue) | |
1884 | break; | |
08e98fc6 | 1885 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1886 | } |
624dbe47 ML |
1887 | } |
1888 | ||
08e98fc6 ML |
1889 | static int blk_mq_init_hctx(struct request_queue *q, |
1890 | struct blk_mq_tag_set *set, | |
1891 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1892 | { |
08e98fc6 | 1893 | int node; |
f70ced09 | 1894 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1895 | |
1896 | node = hctx->numa_node; | |
1897 | if (node == NUMA_NO_NODE) | |
1898 | node = hctx->numa_node = set->numa_node; | |
1899 | ||
27489a3c | 1900 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
7587a5ae | 1901 | INIT_DELAYED_WORK(&hctx->delayed_run_work, blk_mq_delayed_run_work_fn); |
08e98fc6 ML |
1902 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1903 | spin_lock_init(&hctx->lock); | |
1904 | INIT_LIST_HEAD(&hctx->dispatch); | |
1905 | hctx->queue = q; | |
1906 | hctx->queue_num = hctx_idx; | |
2404e607 | 1907 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1908 | |
9467f859 | 1909 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1910 | |
1911 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1912 | |
1913 | /* | |
08e98fc6 ML |
1914 | * Allocate space for all possible cpus to avoid allocation at |
1915 | * runtime | |
320ae51f | 1916 | */ |
08e98fc6 ML |
1917 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1918 | GFP_KERNEL, node); | |
1919 | if (!hctx->ctxs) | |
1920 | goto unregister_cpu_notifier; | |
320ae51f | 1921 | |
88459642 OS |
1922 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1923 | node)) | |
08e98fc6 | 1924 | goto free_ctxs; |
320ae51f | 1925 | |
08e98fc6 | 1926 | hctx->nr_ctx = 0; |
320ae51f | 1927 | |
08e98fc6 ML |
1928 | if (set->ops->init_hctx && |
1929 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1930 | goto free_bitmap; | |
320ae51f | 1931 | |
93252632 OS |
1932 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
1933 | goto exit_hctx; | |
1934 | ||
f70ced09 ML |
1935 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1936 | if (!hctx->fq) | |
93252632 | 1937 | goto sched_exit_hctx; |
320ae51f | 1938 | |
f70ced09 ML |
1939 | if (set->ops->init_request && |
1940 | set->ops->init_request(set->driver_data, | |
1941 | hctx->fq->flush_rq, hctx_idx, | |
1942 | flush_start_tag + hctx_idx, node)) | |
1943 | goto free_fq; | |
320ae51f | 1944 | |
6a83e74d BVA |
1945 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1946 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1947 | ||
08e98fc6 | 1948 | return 0; |
320ae51f | 1949 | |
f70ced09 ML |
1950 | free_fq: |
1951 | kfree(hctx->fq); | |
93252632 OS |
1952 | sched_exit_hctx: |
1953 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
1954 | exit_hctx: |
1955 | if (set->ops->exit_hctx) | |
1956 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1957 | free_bitmap: |
88459642 | 1958 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1959 | free_ctxs: |
1960 | kfree(hctx->ctxs); | |
1961 | unregister_cpu_notifier: | |
9467f859 | 1962 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1963 | return -1; |
1964 | } | |
320ae51f | 1965 | |
320ae51f JA |
1966 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1967 | unsigned int nr_hw_queues) | |
1968 | { | |
1969 | unsigned int i; | |
1970 | ||
1971 | for_each_possible_cpu(i) { | |
1972 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1973 | struct blk_mq_hw_ctx *hctx; | |
1974 | ||
320ae51f JA |
1975 | __ctx->cpu = i; |
1976 | spin_lock_init(&__ctx->lock); | |
1977 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1978 | __ctx->queue = q; | |
1979 | ||
1980 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1981 | if (!cpu_online(i)) |
1982 | continue; | |
1983 | ||
7d7e0f90 | 1984 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1985 | |
320ae51f JA |
1986 | /* |
1987 | * Set local node, IFF we have more than one hw queue. If | |
1988 | * not, we remain on the home node of the device | |
1989 | */ | |
1990 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1991 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1992 | } |
1993 | } | |
1994 | ||
cc71a6f4 JA |
1995 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1996 | { | |
1997 | int ret = 0; | |
1998 | ||
1999 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2000 | set->queue_depth, set->reserved_tags); | |
2001 | if (!set->tags[hctx_idx]) | |
2002 | return false; | |
2003 | ||
2004 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2005 | set->queue_depth); | |
2006 | if (!ret) | |
2007 | return true; | |
2008 | ||
2009 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2010 | set->tags[hctx_idx] = NULL; | |
2011 | return false; | |
2012 | } | |
2013 | ||
2014 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2015 | unsigned int hctx_idx) | |
2016 | { | |
bd166ef1 JA |
2017 | if (set->tags[hctx_idx]) { |
2018 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2019 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2020 | set->tags[hctx_idx] = NULL; | |
2021 | } | |
cc71a6f4 JA |
2022 | } |
2023 | ||
5778322e AM |
2024 | static void blk_mq_map_swqueue(struct request_queue *q, |
2025 | const struct cpumask *online_mask) | |
320ae51f | 2026 | { |
d1b1cea1 | 2027 | unsigned int i, hctx_idx; |
320ae51f JA |
2028 | struct blk_mq_hw_ctx *hctx; |
2029 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2030 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2031 | |
60de074b AM |
2032 | /* |
2033 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2034 | */ | |
2035 | mutex_lock(&q->sysfs_lock); | |
2036 | ||
320ae51f | 2037 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2038 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2039 | hctx->nr_ctx = 0; |
2040 | } | |
2041 | ||
2042 | /* | |
2043 | * Map software to hardware queues | |
2044 | */ | |
897bb0c7 | 2045 | for_each_possible_cpu(i) { |
320ae51f | 2046 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2047 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2048 | continue; |
2049 | ||
d1b1cea1 GKB |
2050 | hctx_idx = q->mq_map[i]; |
2051 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2052 | if (!set->tags[hctx_idx] && |
2053 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2054 | /* |
2055 | * If tags initialization fail for some hctx, | |
2056 | * that hctx won't be brought online. In this | |
2057 | * case, remap the current ctx to hctx[0] which | |
2058 | * is guaranteed to always have tags allocated | |
2059 | */ | |
cc71a6f4 | 2060 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2061 | } |
2062 | ||
897bb0c7 | 2063 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2064 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2065 | |
e4043dcf | 2066 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2067 | ctx->index_hw = hctx->nr_ctx; |
2068 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2069 | } | |
506e931f | 2070 | |
60de074b AM |
2071 | mutex_unlock(&q->sysfs_lock); |
2072 | ||
506e931f | 2073 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2074 | /* |
a68aafa5 JA |
2075 | * If no software queues are mapped to this hardware queue, |
2076 | * disable it and free the request entries. | |
484b4061 JA |
2077 | */ |
2078 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2079 | /* Never unmap queue 0. We need it as a |
2080 | * fallback in case of a new remap fails | |
2081 | * allocation | |
2082 | */ | |
cc71a6f4 JA |
2083 | if (i && set->tags[i]) |
2084 | blk_mq_free_map_and_requests(set, i); | |
2085 | ||
2a34c087 | 2086 | hctx->tags = NULL; |
484b4061 JA |
2087 | continue; |
2088 | } | |
2089 | ||
2a34c087 ML |
2090 | hctx->tags = set->tags[i]; |
2091 | WARN_ON(!hctx->tags); | |
2092 | ||
889fa31f CY |
2093 | /* |
2094 | * Set the map size to the number of mapped software queues. | |
2095 | * This is more accurate and more efficient than looping | |
2096 | * over all possibly mapped software queues. | |
2097 | */ | |
88459642 | 2098 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2099 | |
484b4061 JA |
2100 | /* |
2101 | * Initialize batch roundrobin counts | |
2102 | */ | |
506e931f JA |
2103 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2104 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2105 | } | |
320ae51f JA |
2106 | } |
2107 | ||
2404e607 | 2108 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2109 | { |
2110 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2111 | int i; |
2112 | ||
2404e607 JM |
2113 | queue_for_each_hw_ctx(q, hctx, i) { |
2114 | if (shared) | |
2115 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2116 | else | |
2117 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2118 | } | |
2119 | } | |
2120 | ||
2121 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2122 | { | |
2123 | struct request_queue *q; | |
0d2602ca | 2124 | |
705cda97 BVA |
2125 | lockdep_assert_held(&set->tag_list_lock); |
2126 | ||
0d2602ca JA |
2127 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2128 | blk_mq_freeze_queue(q); | |
2404e607 | 2129 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2130 | blk_mq_unfreeze_queue(q); |
2131 | } | |
2132 | } | |
2133 | ||
2134 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2135 | { | |
2136 | struct blk_mq_tag_set *set = q->tag_set; | |
2137 | ||
0d2602ca | 2138 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2139 | list_del_rcu(&q->tag_set_list); |
2140 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2141 | if (list_is_singular(&set->tag_list)) { |
2142 | /* just transitioned to unshared */ | |
2143 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2144 | /* update existing queue */ | |
2145 | blk_mq_update_tag_set_depth(set, false); | |
2146 | } | |
0d2602ca | 2147 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2148 | |
2149 | synchronize_rcu(); | |
0d2602ca JA |
2150 | } |
2151 | ||
2152 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2153 | struct request_queue *q) | |
2154 | { | |
2155 | q->tag_set = set; | |
2156 | ||
2157 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2158 | |
2159 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2160 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2161 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2162 | /* update existing queue */ | |
2163 | blk_mq_update_tag_set_depth(set, true); | |
2164 | } | |
2165 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2166 | queue_set_hctx_shared(q, true); | |
705cda97 | 2167 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2168 | |
0d2602ca JA |
2169 | mutex_unlock(&set->tag_list_lock); |
2170 | } | |
2171 | ||
e09aae7e ML |
2172 | /* |
2173 | * It is the actual release handler for mq, but we do it from | |
2174 | * request queue's release handler for avoiding use-after-free | |
2175 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2176 | * but we can't group ctx/kctx kobj without it. | |
2177 | */ | |
2178 | void blk_mq_release(struct request_queue *q) | |
2179 | { | |
2180 | struct blk_mq_hw_ctx *hctx; | |
2181 | unsigned int i; | |
2182 | ||
2183 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2184 | queue_for_each_hw_ctx(q, hctx, i) { |
2185 | if (!hctx) | |
2186 | continue; | |
6c8b232e | 2187 | kobject_put(&hctx->kobj); |
c3b4afca | 2188 | } |
e09aae7e | 2189 | |
a723bab3 AM |
2190 | q->mq_map = NULL; |
2191 | ||
e09aae7e ML |
2192 | kfree(q->queue_hw_ctx); |
2193 | ||
7ea5fe31 ML |
2194 | /* |
2195 | * release .mq_kobj and sw queue's kobject now because | |
2196 | * both share lifetime with request queue. | |
2197 | */ | |
2198 | blk_mq_sysfs_deinit(q); | |
2199 | ||
e09aae7e ML |
2200 | free_percpu(q->queue_ctx); |
2201 | } | |
2202 | ||
24d2f903 | 2203 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2204 | { |
2205 | struct request_queue *uninit_q, *q; | |
2206 | ||
2207 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2208 | if (!uninit_q) | |
2209 | return ERR_PTR(-ENOMEM); | |
2210 | ||
2211 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2212 | if (IS_ERR(q)) | |
2213 | blk_cleanup_queue(uninit_q); | |
2214 | ||
2215 | return q; | |
2216 | } | |
2217 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2218 | ||
868f2f0b KB |
2219 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2220 | struct request_queue *q) | |
320ae51f | 2221 | { |
868f2f0b KB |
2222 | int i, j; |
2223 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2224 | |
868f2f0b | 2225 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2226 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2227 | int node; |
f14bbe77 | 2228 | |
868f2f0b KB |
2229 | if (hctxs[i]) |
2230 | continue; | |
2231 | ||
2232 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2233 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2234 | GFP_KERNEL, node); | |
320ae51f | 2235 | if (!hctxs[i]) |
868f2f0b | 2236 | break; |
320ae51f | 2237 | |
a86073e4 | 2238 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2239 | node)) { |
2240 | kfree(hctxs[i]); | |
2241 | hctxs[i] = NULL; | |
2242 | break; | |
2243 | } | |
e4043dcf | 2244 | |
0d2602ca | 2245 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2246 | hctxs[i]->numa_node = node; |
320ae51f | 2247 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2248 | |
2249 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2250 | free_cpumask_var(hctxs[i]->cpumask); | |
2251 | kfree(hctxs[i]); | |
2252 | hctxs[i] = NULL; | |
2253 | break; | |
2254 | } | |
2255 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2256 | } |
868f2f0b KB |
2257 | for (j = i; j < q->nr_hw_queues; j++) { |
2258 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2259 | ||
2260 | if (hctx) { | |
cc71a6f4 JA |
2261 | if (hctx->tags) |
2262 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2263 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2264 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2265 | hctxs[j] = NULL; |
2266 | ||
2267 | } | |
2268 | } | |
2269 | q->nr_hw_queues = i; | |
2270 | blk_mq_sysfs_register(q); | |
2271 | } | |
2272 | ||
2273 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2274 | struct request_queue *q) | |
2275 | { | |
66841672 ML |
2276 | /* mark the queue as mq asap */ |
2277 | q->mq_ops = set->ops; | |
2278 | ||
34dbad5d | 2279 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2280 | blk_mq_poll_stats_bkt, |
2281 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2282 | if (!q->poll_cb) |
2283 | goto err_exit; | |
2284 | ||
868f2f0b KB |
2285 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2286 | if (!q->queue_ctx) | |
c7de5726 | 2287 | goto err_exit; |
868f2f0b | 2288 | |
737f98cf ML |
2289 | /* init q->mq_kobj and sw queues' kobjects */ |
2290 | blk_mq_sysfs_init(q); | |
2291 | ||
868f2f0b KB |
2292 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2293 | GFP_KERNEL, set->numa_node); | |
2294 | if (!q->queue_hw_ctx) | |
2295 | goto err_percpu; | |
2296 | ||
bdd17e75 | 2297 | q->mq_map = set->mq_map; |
868f2f0b KB |
2298 | |
2299 | blk_mq_realloc_hw_ctxs(set, q); | |
2300 | if (!q->nr_hw_queues) | |
2301 | goto err_hctxs; | |
320ae51f | 2302 | |
287922eb | 2303 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2304 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2305 | |
2306 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2307 | |
94eddfbe | 2308 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2309 | |
05f1dd53 JA |
2310 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2311 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2312 | ||
1be036e9 CH |
2313 | q->sg_reserved_size = INT_MAX; |
2314 | ||
2849450a | 2315 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2316 | INIT_LIST_HEAD(&q->requeue_list); |
2317 | spin_lock_init(&q->requeue_lock); | |
2318 | ||
254d259d | 2319 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2320 | |
eba71768 JA |
2321 | /* |
2322 | * Do this after blk_queue_make_request() overrides it... | |
2323 | */ | |
2324 | q->nr_requests = set->queue_depth; | |
2325 | ||
64f1c21e JA |
2326 | /* |
2327 | * Default to classic polling | |
2328 | */ | |
2329 | q->poll_nsec = -1; | |
2330 | ||
24d2f903 CH |
2331 | if (set->ops->complete) |
2332 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2333 | |
24d2f903 | 2334 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2335 | |
5778322e | 2336 | get_online_cpus(); |
320ae51f | 2337 | mutex_lock(&all_q_mutex); |
320ae51f | 2338 | |
4593fdbe | 2339 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2340 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2341 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2342 | |
4593fdbe | 2343 | mutex_unlock(&all_q_mutex); |
5778322e | 2344 | put_online_cpus(); |
4593fdbe | 2345 | |
d3484991 JA |
2346 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2347 | int ret; | |
2348 | ||
2349 | ret = blk_mq_sched_init(q); | |
2350 | if (ret) | |
2351 | return ERR_PTR(ret); | |
2352 | } | |
2353 | ||
320ae51f | 2354 | return q; |
18741986 | 2355 | |
320ae51f | 2356 | err_hctxs: |
868f2f0b | 2357 | kfree(q->queue_hw_ctx); |
320ae51f | 2358 | err_percpu: |
868f2f0b | 2359 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2360 | err_exit: |
2361 | q->mq_ops = NULL; | |
320ae51f JA |
2362 | return ERR_PTR(-ENOMEM); |
2363 | } | |
b62c21b7 | 2364 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2365 | |
2366 | void blk_mq_free_queue(struct request_queue *q) | |
2367 | { | |
624dbe47 | 2368 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2369 | |
0e626368 AM |
2370 | mutex_lock(&all_q_mutex); |
2371 | list_del_init(&q->all_q_node); | |
2372 | mutex_unlock(&all_q_mutex); | |
2373 | ||
0d2602ca JA |
2374 | blk_mq_del_queue_tag_set(q); |
2375 | ||
624dbe47 | 2376 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2377 | } |
320ae51f JA |
2378 | |
2379 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2380 | static void blk_mq_queue_reinit(struct request_queue *q, |
2381 | const struct cpumask *online_mask) | |
320ae51f | 2382 | { |
4ecd4fef | 2383 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2384 | |
67aec14c JA |
2385 | blk_mq_sysfs_unregister(q); |
2386 | ||
320ae51f JA |
2387 | /* |
2388 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2389 | * we should change hctx numa_node according to new topology (this | |
2390 | * involves free and re-allocate memory, worthy doing?) | |
2391 | */ | |
2392 | ||
5778322e | 2393 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2394 | |
67aec14c | 2395 | blk_mq_sysfs_register(q); |
320ae51f JA |
2396 | } |
2397 | ||
65d5291e SAS |
2398 | /* |
2399 | * New online cpumask which is going to be set in this hotplug event. | |
2400 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2401 | * one-by-one and dynamically allocating this could result in a failure. | |
2402 | */ | |
2403 | static struct cpumask cpuhp_online_new; | |
2404 | ||
2405 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2406 | { |
2407 | struct request_queue *q; | |
320ae51f JA |
2408 | |
2409 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2410 | /* |
2411 | * We need to freeze and reinit all existing queues. Freezing | |
2412 | * involves synchronous wait for an RCU grace period and doing it | |
2413 | * one by one may take a long time. Start freezing all queues in | |
2414 | * one swoop and then wait for the completions so that freezing can | |
2415 | * take place in parallel. | |
2416 | */ | |
2417 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1671d522 | 2418 | blk_freeze_queue_start(q); |
415d3dab | 2419 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2420 | blk_mq_freeze_queue_wait(q); |
2421 | ||
320ae51f | 2422 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2423 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2424 | |
2425 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2426 | blk_mq_unfreeze_queue(q); | |
2427 | ||
320ae51f | 2428 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2429 | } |
2430 | ||
2431 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2432 | { | |
97a32864 | 2433 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2434 | blk_mq_queue_reinit_work(); |
2435 | return 0; | |
2436 | } | |
2437 | ||
2438 | /* | |
2439 | * Before hotadded cpu starts handling requests, new mappings must be | |
2440 | * established. Otherwise, these requests in hw queue might never be | |
2441 | * dispatched. | |
2442 | * | |
2443 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2444 | * for CPU0, and ctx1 for CPU1). | |
2445 | * | |
2446 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2447 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2448 | * | |
2c3ad667 JA |
2449 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2450 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2451 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2452 | * ignored. | |
65d5291e SAS |
2453 | */ |
2454 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2455 | { | |
2456 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2457 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2458 | blk_mq_queue_reinit_work(); | |
2459 | return 0; | |
320ae51f JA |
2460 | } |
2461 | ||
a5164405 JA |
2462 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2463 | { | |
2464 | int i; | |
2465 | ||
cc71a6f4 JA |
2466 | for (i = 0; i < set->nr_hw_queues; i++) |
2467 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2468 | goto out_unwind; |
a5164405 JA |
2469 | |
2470 | return 0; | |
2471 | ||
2472 | out_unwind: | |
2473 | while (--i >= 0) | |
cc71a6f4 | 2474 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2475 | |
a5164405 JA |
2476 | return -ENOMEM; |
2477 | } | |
2478 | ||
2479 | /* | |
2480 | * Allocate the request maps associated with this tag_set. Note that this | |
2481 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2482 | * will be updated to reflect the allocated depth. | |
2483 | */ | |
2484 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2485 | { | |
2486 | unsigned int depth; | |
2487 | int err; | |
2488 | ||
2489 | depth = set->queue_depth; | |
2490 | do { | |
2491 | err = __blk_mq_alloc_rq_maps(set); | |
2492 | if (!err) | |
2493 | break; | |
2494 | ||
2495 | set->queue_depth >>= 1; | |
2496 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2497 | err = -ENOMEM; | |
2498 | break; | |
2499 | } | |
2500 | } while (set->queue_depth); | |
2501 | ||
2502 | if (!set->queue_depth || err) { | |
2503 | pr_err("blk-mq: failed to allocate request map\n"); | |
2504 | return -ENOMEM; | |
2505 | } | |
2506 | ||
2507 | if (depth != set->queue_depth) | |
2508 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2509 | depth, set->queue_depth); | |
2510 | ||
2511 | return 0; | |
2512 | } | |
2513 | ||
ebe8bddb OS |
2514 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2515 | { | |
2516 | if (set->ops->map_queues) | |
2517 | return set->ops->map_queues(set); | |
2518 | else | |
2519 | return blk_mq_map_queues(set); | |
2520 | } | |
2521 | ||
a4391c64 JA |
2522 | /* |
2523 | * Alloc a tag set to be associated with one or more request queues. | |
2524 | * May fail with EINVAL for various error conditions. May adjust the | |
2525 | * requested depth down, if if it too large. In that case, the set | |
2526 | * value will be stored in set->queue_depth. | |
2527 | */ | |
24d2f903 CH |
2528 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2529 | { | |
da695ba2 CH |
2530 | int ret; |
2531 | ||
205fb5f5 BVA |
2532 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2533 | ||
24d2f903 CH |
2534 | if (!set->nr_hw_queues) |
2535 | return -EINVAL; | |
a4391c64 | 2536 | if (!set->queue_depth) |
24d2f903 CH |
2537 | return -EINVAL; |
2538 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2539 | return -EINVAL; | |
2540 | ||
7d7e0f90 | 2541 | if (!set->ops->queue_rq) |
24d2f903 CH |
2542 | return -EINVAL; |
2543 | ||
a4391c64 JA |
2544 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2545 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2546 | BLK_MQ_MAX_DEPTH); | |
2547 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2548 | } | |
24d2f903 | 2549 | |
6637fadf SL |
2550 | /* |
2551 | * If a crashdump is active, then we are potentially in a very | |
2552 | * memory constrained environment. Limit us to 1 queue and | |
2553 | * 64 tags to prevent using too much memory. | |
2554 | */ | |
2555 | if (is_kdump_kernel()) { | |
2556 | set->nr_hw_queues = 1; | |
2557 | set->queue_depth = min(64U, set->queue_depth); | |
2558 | } | |
868f2f0b KB |
2559 | /* |
2560 | * There is no use for more h/w queues than cpus. | |
2561 | */ | |
2562 | if (set->nr_hw_queues > nr_cpu_ids) | |
2563 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2564 | |
868f2f0b | 2565 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2566 | GFP_KERNEL, set->numa_node); |
2567 | if (!set->tags) | |
a5164405 | 2568 | return -ENOMEM; |
24d2f903 | 2569 | |
da695ba2 CH |
2570 | ret = -ENOMEM; |
2571 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2572 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2573 | if (!set->mq_map) |
2574 | goto out_free_tags; | |
2575 | ||
ebe8bddb | 2576 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2577 | if (ret) |
2578 | goto out_free_mq_map; | |
2579 | ||
2580 | ret = blk_mq_alloc_rq_maps(set); | |
2581 | if (ret) | |
bdd17e75 | 2582 | goto out_free_mq_map; |
24d2f903 | 2583 | |
0d2602ca JA |
2584 | mutex_init(&set->tag_list_lock); |
2585 | INIT_LIST_HEAD(&set->tag_list); | |
2586 | ||
24d2f903 | 2587 | return 0; |
bdd17e75 CH |
2588 | |
2589 | out_free_mq_map: | |
2590 | kfree(set->mq_map); | |
2591 | set->mq_map = NULL; | |
2592 | out_free_tags: | |
5676e7b6 RE |
2593 | kfree(set->tags); |
2594 | set->tags = NULL; | |
da695ba2 | 2595 | return ret; |
24d2f903 CH |
2596 | } |
2597 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2598 | ||
2599 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2600 | { | |
2601 | int i; | |
2602 | ||
cc71a6f4 JA |
2603 | for (i = 0; i < nr_cpu_ids; i++) |
2604 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2605 | |
bdd17e75 CH |
2606 | kfree(set->mq_map); |
2607 | set->mq_map = NULL; | |
2608 | ||
981bd189 | 2609 | kfree(set->tags); |
5676e7b6 | 2610 | set->tags = NULL; |
24d2f903 CH |
2611 | } |
2612 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2613 | ||
e3a2b3f9 JA |
2614 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2615 | { | |
2616 | struct blk_mq_tag_set *set = q->tag_set; | |
2617 | struct blk_mq_hw_ctx *hctx; | |
2618 | int i, ret; | |
2619 | ||
bd166ef1 | 2620 | if (!set) |
e3a2b3f9 JA |
2621 | return -EINVAL; |
2622 | ||
70f36b60 JA |
2623 | blk_mq_freeze_queue(q); |
2624 | blk_mq_quiesce_queue(q); | |
2625 | ||
e3a2b3f9 JA |
2626 | ret = 0; |
2627 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2628 | if (!hctx->tags) |
2629 | continue; | |
bd166ef1 JA |
2630 | /* |
2631 | * If we're using an MQ scheduler, just update the scheduler | |
2632 | * queue depth. This is similar to what the old code would do. | |
2633 | */ | |
70f36b60 JA |
2634 | if (!hctx->sched_tags) { |
2635 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2636 | min(nr, set->queue_depth), | |
2637 | false); | |
2638 | } else { | |
2639 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2640 | nr, true); | |
2641 | } | |
e3a2b3f9 JA |
2642 | if (ret) |
2643 | break; | |
2644 | } | |
2645 | ||
2646 | if (!ret) | |
2647 | q->nr_requests = nr; | |
2648 | ||
70f36b60 JA |
2649 | blk_mq_unfreeze_queue(q); |
2650 | blk_mq_start_stopped_hw_queues(q, true); | |
2651 | ||
e3a2b3f9 JA |
2652 | return ret; |
2653 | } | |
2654 | ||
868f2f0b KB |
2655 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2656 | { | |
2657 | struct request_queue *q; | |
2658 | ||
705cda97 BVA |
2659 | lockdep_assert_held(&set->tag_list_lock); |
2660 | ||
868f2f0b KB |
2661 | if (nr_hw_queues > nr_cpu_ids) |
2662 | nr_hw_queues = nr_cpu_ids; | |
2663 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2664 | return; | |
2665 | ||
2666 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2667 | blk_mq_freeze_queue(q); | |
2668 | ||
2669 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2670 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2671 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2672 | blk_mq_realloc_hw_ctxs(set, q); | |
868f2f0b KB |
2673 | blk_mq_queue_reinit(q, cpu_online_mask); |
2674 | } | |
2675 | ||
2676 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2677 | blk_mq_unfreeze_queue(q); | |
2678 | } | |
2679 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2680 | ||
34dbad5d OS |
2681 | /* Enable polling stats and return whether they were already enabled. */ |
2682 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2683 | { | |
2684 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2685 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2686 | return true; | |
2687 | blk_stat_add_callback(q, q->poll_cb); | |
2688 | return false; | |
2689 | } | |
2690 | ||
2691 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2692 | { | |
2693 | /* | |
2694 | * We don't arm the callback if polling stats are not enabled or the | |
2695 | * callback is already active. | |
2696 | */ | |
2697 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2698 | blk_stat_is_active(q->poll_cb)) | |
2699 | return; | |
2700 | ||
2701 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2702 | } | |
2703 | ||
2704 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2705 | { | |
2706 | struct request_queue *q = cb->data; | |
720b8ccc | 2707 | int bucket; |
34dbad5d | 2708 | |
720b8ccc SB |
2709 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2710 | if (cb->stat[bucket].nr_samples) | |
2711 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2712 | } | |
34dbad5d OS |
2713 | } |
2714 | ||
64f1c21e JA |
2715 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2716 | struct blk_mq_hw_ctx *hctx, | |
2717 | struct request *rq) | |
2718 | { | |
64f1c21e | 2719 | unsigned long ret = 0; |
720b8ccc | 2720 | int bucket; |
64f1c21e JA |
2721 | |
2722 | /* | |
2723 | * If stats collection isn't on, don't sleep but turn it on for | |
2724 | * future users | |
2725 | */ | |
34dbad5d | 2726 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2727 | return 0; |
2728 | ||
64f1c21e JA |
2729 | /* |
2730 | * As an optimistic guess, use half of the mean service time | |
2731 | * for this type of request. We can (and should) make this smarter. | |
2732 | * For instance, if the completion latencies are tight, we can | |
2733 | * get closer than just half the mean. This is especially | |
2734 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2735 | * than ~10 usec. We do use the stats for the relevant IO size |
2736 | * if available which does lead to better estimates. | |
64f1c21e | 2737 | */ |
720b8ccc SB |
2738 | bucket = blk_mq_poll_stats_bkt(rq); |
2739 | if (bucket < 0) | |
2740 | return ret; | |
2741 | ||
2742 | if (q->poll_stat[bucket].nr_samples) | |
2743 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2744 | |
2745 | return ret; | |
2746 | } | |
2747 | ||
06426adf | 2748 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2749 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2750 | struct request *rq) |
2751 | { | |
2752 | struct hrtimer_sleeper hs; | |
2753 | enum hrtimer_mode mode; | |
64f1c21e | 2754 | unsigned int nsecs; |
06426adf JA |
2755 | ktime_t kt; |
2756 | ||
64f1c21e JA |
2757 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2758 | return false; | |
2759 | ||
2760 | /* | |
2761 | * poll_nsec can be: | |
2762 | * | |
2763 | * -1: don't ever hybrid sleep | |
2764 | * 0: use half of prev avg | |
2765 | * >0: use this specific value | |
2766 | */ | |
2767 | if (q->poll_nsec == -1) | |
2768 | return false; | |
2769 | else if (q->poll_nsec > 0) | |
2770 | nsecs = q->poll_nsec; | |
2771 | else | |
2772 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2773 | ||
2774 | if (!nsecs) | |
06426adf JA |
2775 | return false; |
2776 | ||
2777 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2778 | ||
2779 | /* | |
2780 | * This will be replaced with the stats tracking code, using | |
2781 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2782 | */ | |
8b0e1953 | 2783 | kt = nsecs; |
06426adf JA |
2784 | |
2785 | mode = HRTIMER_MODE_REL; | |
2786 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2787 | hrtimer_set_expires(&hs.timer, kt); | |
2788 | ||
2789 | hrtimer_init_sleeper(&hs, current); | |
2790 | do { | |
2791 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2792 | break; | |
2793 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2794 | hrtimer_start_expires(&hs.timer, mode); | |
2795 | if (hs.task) | |
2796 | io_schedule(); | |
2797 | hrtimer_cancel(&hs.timer); | |
2798 | mode = HRTIMER_MODE_ABS; | |
2799 | } while (hs.task && !signal_pending(current)); | |
2800 | ||
2801 | __set_current_state(TASK_RUNNING); | |
2802 | destroy_hrtimer_on_stack(&hs.timer); | |
2803 | return true; | |
2804 | } | |
2805 | ||
bbd7bb70 JA |
2806 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2807 | { | |
2808 | struct request_queue *q = hctx->queue; | |
2809 | long state; | |
2810 | ||
06426adf JA |
2811 | /* |
2812 | * If we sleep, have the caller restart the poll loop to reset | |
2813 | * the state. Like for the other success return cases, the | |
2814 | * caller is responsible for checking if the IO completed. If | |
2815 | * the IO isn't complete, we'll get called again and will go | |
2816 | * straight to the busy poll loop. | |
2817 | */ | |
64f1c21e | 2818 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2819 | return true; |
2820 | ||
bbd7bb70 JA |
2821 | hctx->poll_considered++; |
2822 | ||
2823 | state = current->state; | |
2824 | while (!need_resched()) { | |
2825 | int ret; | |
2826 | ||
2827 | hctx->poll_invoked++; | |
2828 | ||
2829 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2830 | if (ret > 0) { | |
2831 | hctx->poll_success++; | |
2832 | set_current_state(TASK_RUNNING); | |
2833 | return true; | |
2834 | } | |
2835 | ||
2836 | if (signal_pending_state(state, current)) | |
2837 | set_current_state(TASK_RUNNING); | |
2838 | ||
2839 | if (current->state == TASK_RUNNING) | |
2840 | return true; | |
2841 | if (ret < 0) | |
2842 | break; | |
2843 | cpu_relax(); | |
2844 | } | |
2845 | ||
2846 | return false; | |
2847 | } | |
2848 | ||
2849 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2850 | { | |
2851 | struct blk_mq_hw_ctx *hctx; | |
2852 | struct blk_plug *plug; | |
2853 | struct request *rq; | |
2854 | ||
2855 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2856 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2857 | return false; | |
2858 | ||
2859 | plug = current->plug; | |
2860 | if (plug) | |
2861 | blk_flush_plug_list(plug, false); | |
2862 | ||
2863 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2864 | if (!blk_qc_t_is_internal(cookie)) |
2865 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2866 | else | |
2867 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2868 | |
2869 | return __blk_mq_poll(hctx, rq); | |
2870 | } | |
2871 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2872 | ||
676141e4 JA |
2873 | void blk_mq_disable_hotplug(void) |
2874 | { | |
2875 | mutex_lock(&all_q_mutex); | |
2876 | } | |
2877 | ||
2878 | void blk_mq_enable_hotplug(void) | |
2879 | { | |
2880 | mutex_unlock(&all_q_mutex); | |
2881 | } | |
2882 | ||
320ae51f JA |
2883 | static int __init blk_mq_init(void) |
2884 | { | |
9467f859 TG |
2885 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2886 | blk_mq_hctx_notify_dead); | |
320ae51f | 2887 | |
65d5291e SAS |
2888 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2889 | blk_mq_queue_reinit_prepare, | |
2890 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2891 | return 0; |
2892 | } | |
2893 | subsys_initcall(blk_mq_init); |