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" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f JA |
39 | |
40 | static DEFINE_MUTEX(all_q_mutex); | |
41 | static LIST_HEAD(all_q_list); | |
42 | ||
34dbad5d OS |
43 | static void blk_mq_poll_stats_start(struct request_queue *q); |
44 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
2719aa21 | 45 | static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync); |
34dbad5d | 46 | |
720b8ccc SB |
47 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
48 | { | |
49 | int ddir, bytes, bucket; | |
50 | ||
99c749a4 | 51 | ddir = rq_data_dir(rq); |
720b8ccc SB |
52 | bytes = blk_rq_bytes(rq); |
53 | ||
54 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
55 | ||
56 | if (bucket < 0) | |
57 | return -1; | |
58 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
59 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
60 | ||
61 | return bucket; | |
62 | } | |
63 | ||
320ae51f JA |
64 | /* |
65 | * Check if any of the ctx's have pending work in this hardware queue | |
66 | */ | |
50e1dab8 | 67 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 68 | { |
bd166ef1 JA |
69 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
70 | !list_empty_careful(&hctx->dispatch) || | |
71 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
72 | } |
73 | ||
320ae51f JA |
74 | /* |
75 | * Mark this ctx as having pending work in this hardware queue | |
76 | */ | |
77 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
78 | struct blk_mq_ctx *ctx) | |
79 | { | |
88459642 OS |
80 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
81 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
82 | } |
83 | ||
84 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
85 | struct blk_mq_ctx *ctx) | |
86 | { | |
88459642 | 87 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
88 | } |
89 | ||
1671d522 | 90 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 91 | { |
4ecd4fef | 92 | int freeze_depth; |
cddd5d17 | 93 | |
4ecd4fef CH |
94 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
95 | if (freeze_depth == 1) { | |
3ef28e83 | 96 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 97 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 98 | } |
f3af020b | 99 | } |
1671d522 | 100 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 101 | |
6bae363e | 102 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 103 | { |
3ef28e83 | 104 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 105 | } |
6bae363e | 106 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 107 | |
f91328c4 KB |
108 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
109 | unsigned long timeout) | |
110 | { | |
111 | return wait_event_timeout(q->mq_freeze_wq, | |
112 | percpu_ref_is_zero(&q->q_usage_counter), | |
113 | timeout); | |
114 | } | |
115 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 116 | |
f3af020b TH |
117 | /* |
118 | * Guarantee no request is in use, so we can change any data structure of | |
119 | * the queue afterward. | |
120 | */ | |
3ef28e83 | 121 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 122 | { |
3ef28e83 DW |
123 | /* |
124 | * In the !blk_mq case we are only calling this to kill the | |
125 | * q_usage_counter, otherwise this increases the freeze depth | |
126 | * and waits for it to return to zero. For this reason there is | |
127 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
128 | * exported to drivers as the only user for unfreeze is blk_mq. | |
129 | */ | |
1671d522 | 130 | blk_freeze_queue_start(q); |
f3af020b TH |
131 | blk_mq_freeze_queue_wait(q); |
132 | } | |
3ef28e83 DW |
133 | |
134 | void blk_mq_freeze_queue(struct request_queue *q) | |
135 | { | |
136 | /* | |
137 | * ...just an alias to keep freeze and unfreeze actions balanced | |
138 | * in the blk_mq_* namespace | |
139 | */ | |
140 | blk_freeze_queue(q); | |
141 | } | |
c761d96b | 142 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 143 | |
b4c6a028 | 144 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 145 | { |
4ecd4fef | 146 | int freeze_depth; |
320ae51f | 147 | |
4ecd4fef CH |
148 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
149 | WARN_ON_ONCE(freeze_depth < 0); | |
150 | if (!freeze_depth) { | |
3ef28e83 | 151 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 152 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 153 | } |
320ae51f | 154 | } |
b4c6a028 | 155 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 156 | |
6a83e74d BVA |
157 | /** |
158 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
159 | * @q: request queue. | |
160 | * | |
161 | * Note: this function does not prevent that the struct request end_io() | |
162 | * callback function is invoked. Additionally, it is not prevented that | |
163 | * new queue_rq() calls occur unless the queue has been stopped first. | |
164 | */ | |
165 | void blk_mq_quiesce_queue(struct request_queue *q) | |
166 | { | |
167 | struct blk_mq_hw_ctx *hctx; | |
168 | unsigned int i; | |
169 | bool rcu = false; | |
170 | ||
2719aa21 | 171 | __blk_mq_stop_hw_queues(q, true); |
6a83e74d BVA |
172 | |
173 | queue_for_each_hw_ctx(q, hctx, i) { | |
174 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
175 | synchronize_srcu(&hctx->queue_rq_srcu); | |
176 | else | |
177 | rcu = true; | |
178 | } | |
179 | if (rcu) | |
180 | synchronize_rcu(); | |
181 | } | |
182 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
183 | ||
aed3ea94 JA |
184 | void blk_mq_wake_waiters(struct request_queue *q) |
185 | { | |
186 | struct blk_mq_hw_ctx *hctx; | |
187 | unsigned int i; | |
188 | ||
189 | queue_for_each_hw_ctx(q, hctx, i) | |
190 | if (blk_mq_hw_queue_mapped(hctx)) | |
191 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
192 | |
193 | /* | |
194 | * If we are called because the queue has now been marked as | |
195 | * dying, we need to ensure that processes currently waiting on | |
196 | * the queue are notified as well. | |
197 | */ | |
198 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
199 | } |
200 | ||
320ae51f JA |
201 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
202 | { | |
203 | return blk_mq_has_free_tags(hctx->tags); | |
204 | } | |
205 | EXPORT_SYMBOL(blk_mq_can_queue); | |
206 | ||
2c3ad667 JA |
207 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
208 | struct request *rq, unsigned int op) | |
320ae51f | 209 | { |
af76e555 CH |
210 | INIT_LIST_HEAD(&rq->queuelist); |
211 | /* csd/requeue_work/fifo_time is initialized before use */ | |
212 | rq->q = q; | |
320ae51f | 213 | rq->mq_ctx = ctx; |
ef295ecf | 214 | rq->cmd_flags = op; |
e8064021 CH |
215 | if (blk_queue_io_stat(q)) |
216 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
217 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
218 | rq->cpu = -1; | |
af76e555 CH |
219 | INIT_HLIST_NODE(&rq->hash); |
220 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
221 | rq->rq_disk = NULL; |
222 | rq->part = NULL; | |
3ee32372 | 223 | rq->start_time = jiffies; |
af76e555 CH |
224 | #ifdef CONFIG_BLK_CGROUP |
225 | rq->rl = NULL; | |
0fec08b4 | 226 | set_start_time_ns(rq); |
af76e555 CH |
227 | rq->io_start_time_ns = 0; |
228 | #endif | |
229 | rq->nr_phys_segments = 0; | |
230 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
231 | rq->nr_integrity_segments = 0; | |
232 | #endif | |
af76e555 CH |
233 | rq->special = NULL; |
234 | /* tag was already set */ | |
af76e555 | 235 | rq->extra_len = 0; |
af76e555 | 236 | |
af76e555 | 237 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
238 | rq->timeout = 0; |
239 | ||
af76e555 CH |
240 | rq->end_io = NULL; |
241 | rq->end_io_data = NULL; | |
242 | rq->next_rq = NULL; | |
243 | ||
ef295ecf | 244 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 245 | } |
2c3ad667 | 246 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 247 | |
2c3ad667 JA |
248 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
249 | unsigned int op) | |
5dee8577 CH |
250 | { |
251 | struct request *rq; | |
252 | unsigned int tag; | |
253 | ||
cb96a42c | 254 | tag = blk_mq_get_tag(data); |
5dee8577 | 255 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
256 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
257 | ||
258 | rq = tags->static_rqs[tag]; | |
5dee8577 | 259 | |
bd166ef1 JA |
260 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
261 | rq->tag = -1; | |
262 | rq->internal_tag = tag; | |
263 | } else { | |
200e86b3 JA |
264 | if (blk_mq_tag_busy(data->hctx)) { |
265 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
266 | atomic_inc(&data->hctx->nr_active); | |
267 | } | |
bd166ef1 JA |
268 | rq->tag = tag; |
269 | rq->internal_tag = -1; | |
562bef42 | 270 | data->hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
271 | } |
272 | ||
ef295ecf | 273 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
274 | return rq; |
275 | } | |
276 | ||
277 | return NULL; | |
278 | } | |
2c3ad667 | 279 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 280 | |
6f3b0e8b CH |
281 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
282 | unsigned int flags) | |
320ae51f | 283 | { |
5a797e00 | 284 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 285 | struct request *rq; |
a492f075 | 286 | int ret; |
320ae51f | 287 | |
6f3b0e8b | 288 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
289 | if (ret) |
290 | return ERR_PTR(ret); | |
320ae51f | 291 | |
bd166ef1 | 292 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 293 | |
bd166ef1 JA |
294 | blk_mq_put_ctx(alloc_data.ctx); |
295 | blk_queue_exit(q); | |
296 | ||
297 | if (!rq) | |
a492f075 | 298 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
299 | |
300 | rq->__data_len = 0; | |
301 | rq->__sector = (sector_t) -1; | |
302 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
303 | return rq; |
304 | } | |
4bb659b1 | 305 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 306 | |
1f5bd336 ML |
307 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
308 | unsigned int flags, unsigned int hctx_idx) | |
309 | { | |
6d2809d5 | 310 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 311 | struct request *rq; |
6d2809d5 | 312 | unsigned int cpu; |
1f5bd336 ML |
313 | int ret; |
314 | ||
315 | /* | |
316 | * If the tag allocator sleeps we could get an allocation for a | |
317 | * different hardware context. No need to complicate the low level | |
318 | * allocator for this for the rare use case of a command tied to | |
319 | * a specific queue. | |
320 | */ | |
321 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
322 | return ERR_PTR(-EINVAL); | |
323 | ||
324 | if (hctx_idx >= q->nr_hw_queues) | |
325 | return ERR_PTR(-EIO); | |
326 | ||
327 | ret = blk_queue_enter(q, true); | |
328 | if (ret) | |
329 | return ERR_PTR(ret); | |
330 | ||
c8712c6a CH |
331 | /* |
332 | * Check if the hardware context is actually mapped to anything. | |
333 | * If not tell the caller that it should skip this queue. | |
334 | */ | |
6d2809d5 OS |
335 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
336 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
337 | blk_queue_exit(q); | |
338 | return ERR_PTR(-EXDEV); | |
c8712c6a | 339 | } |
6d2809d5 OS |
340 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
341 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 342 | |
6d2809d5 | 343 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
c8712c6a | 344 | |
c8712c6a | 345 | blk_queue_exit(q); |
6d2809d5 OS |
346 | |
347 | if (!rq) | |
348 | return ERR_PTR(-EWOULDBLOCK); | |
349 | ||
350 | return rq; | |
1f5bd336 ML |
351 | } |
352 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
353 | ||
bd166ef1 JA |
354 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
355 | struct request *rq) | |
320ae51f | 356 | { |
bd166ef1 | 357 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
358 | struct request_queue *q = rq->q; |
359 | ||
e8064021 | 360 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 361 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
362 | |
363 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 364 | rq->rq_flags = 0; |
0d2602ca | 365 | |
af76e555 | 366 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 367 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
368 | if (rq->tag != -1) |
369 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
370 | if (sched_tag != -1) | |
c05f8525 | 371 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 372 | blk_mq_sched_restart(hctx); |
3ef28e83 | 373 | blk_queue_exit(q); |
320ae51f JA |
374 | } |
375 | ||
bd166ef1 | 376 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 377 | struct request *rq) |
320ae51f JA |
378 | { |
379 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
380 | |
381 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
382 | __blk_mq_finish_request(hctx, ctx, rq); |
383 | } | |
384 | ||
385 | void blk_mq_finish_request(struct request *rq) | |
386 | { | |
387 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 388 | } |
5b727272 | 389 | EXPORT_SYMBOL_GPL(blk_mq_finish_request); |
7c7f2f2b JA |
390 | |
391 | void blk_mq_free_request(struct request *rq) | |
392 | { | |
bd166ef1 | 393 | blk_mq_sched_put_request(rq); |
320ae51f | 394 | } |
1a3b595a | 395 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 396 | |
c8a446ad | 397 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 398 | { |
0d11e6ac ML |
399 | blk_account_io_done(rq); |
400 | ||
91b63639 | 401 | if (rq->end_io) { |
87760e5e | 402 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 403 | rq->end_io(rq, error); |
91b63639 CH |
404 | } else { |
405 | if (unlikely(blk_bidi_rq(rq))) | |
406 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 407 | blk_mq_free_request(rq); |
91b63639 | 408 | } |
320ae51f | 409 | } |
c8a446ad | 410 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 411 | |
c8a446ad | 412 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
413 | { |
414 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
415 | BUG(); | |
c8a446ad | 416 | __blk_mq_end_request(rq, error); |
63151a44 | 417 | } |
c8a446ad | 418 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 419 | |
30a91cb4 | 420 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 421 | { |
3d6efbf6 | 422 | struct request *rq = data; |
320ae51f | 423 | |
30a91cb4 | 424 | rq->q->softirq_done_fn(rq); |
320ae51f | 425 | } |
320ae51f | 426 | |
453f8341 | 427 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
428 | { |
429 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 430 | bool shared = false; |
320ae51f JA |
431 | int cpu; |
432 | ||
453f8341 CH |
433 | if (rq->internal_tag != -1) |
434 | blk_mq_sched_completed_request(rq); | |
435 | if (rq->rq_flags & RQF_STATS) { | |
436 | blk_mq_poll_stats_start(rq->q); | |
437 | blk_stat_add(rq); | |
438 | } | |
439 | ||
38535201 | 440 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
441 | rq->q->softirq_done_fn(rq); |
442 | return; | |
443 | } | |
320ae51f JA |
444 | |
445 | cpu = get_cpu(); | |
38535201 CH |
446 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
447 | shared = cpus_share_cache(cpu, ctx->cpu); | |
448 | ||
449 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 450 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
451 | rq->csd.info = rq; |
452 | rq->csd.flags = 0; | |
c46fff2a | 453 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 454 | } else { |
30a91cb4 | 455 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 456 | } |
320ae51f JA |
457 | put_cpu(); |
458 | } | |
30a91cb4 CH |
459 | |
460 | /** | |
461 | * blk_mq_complete_request - end I/O on a request | |
462 | * @rq: the request being processed | |
463 | * | |
464 | * Description: | |
465 | * Ends all I/O on a request. It does not handle partial completions. | |
466 | * The actual completion happens out-of-order, through a IPI handler. | |
467 | **/ | |
08e0029a | 468 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 469 | { |
95f09684 JA |
470 | struct request_queue *q = rq->q; |
471 | ||
472 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 473 | return; |
08e0029a | 474 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 475 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
476 | } |
477 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 478 | |
973c0191 KB |
479 | int blk_mq_request_started(struct request *rq) |
480 | { | |
481 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
482 | } | |
483 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
484 | ||
e2490073 | 485 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
486 | { |
487 | struct request_queue *q = rq->q; | |
488 | ||
bd166ef1 JA |
489 | blk_mq_sched_started_request(rq); |
490 | ||
320ae51f JA |
491 | trace_block_rq_issue(q, rq); |
492 | ||
cf43e6be | 493 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 494 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 495 | rq->rq_flags |= RQF_STATS; |
87760e5e | 496 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
497 | } |
498 | ||
2b8393b4 | 499 | blk_add_timer(rq); |
87ee7b11 | 500 | |
538b7534 JA |
501 | /* |
502 | * Ensure that ->deadline is visible before set the started | |
503 | * flag and clear the completed flag. | |
504 | */ | |
505 | smp_mb__before_atomic(); | |
506 | ||
87ee7b11 JA |
507 | /* |
508 | * Mark us as started and clear complete. Complete might have been | |
509 | * set if requeue raced with timeout, which then marked it as | |
510 | * complete. So be sure to clear complete again when we start | |
511 | * the request, otherwise we'll ignore the completion event. | |
512 | */ | |
4b570521 JA |
513 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
514 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
515 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
516 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
517 | |
518 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
519 | /* | |
520 | * Make sure space for the drain appears. We know we can do | |
521 | * this because max_hw_segments has been adjusted to be one | |
522 | * fewer than the device can handle. | |
523 | */ | |
524 | rq->nr_phys_segments++; | |
525 | } | |
320ae51f | 526 | } |
e2490073 | 527 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 528 | |
d9d149a3 ML |
529 | /* |
530 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 531 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
532 | * but given rq->deadline is just set in .queue_rq() under |
533 | * this situation, the race won't be possible in reality because | |
534 | * rq->timeout should be set as big enough to cover the window | |
535 | * between blk_mq_start_request() called from .queue_rq() and | |
536 | * clearing REQ_ATOM_STARTED here. | |
537 | */ | |
ed0791b2 | 538 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
539 | { |
540 | struct request_queue *q = rq->q; | |
541 | ||
542 | trace_block_rq_requeue(q, rq); | |
87760e5e | 543 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 544 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 545 | |
e2490073 CH |
546 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
547 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
548 | rq->nr_phys_segments--; | |
549 | } | |
320ae51f JA |
550 | } |
551 | ||
2b053aca | 552 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 553 | { |
ed0791b2 | 554 | __blk_mq_requeue_request(rq); |
ed0791b2 | 555 | |
ed0791b2 | 556 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 557 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
558 | } |
559 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
560 | ||
6fca6a61 CH |
561 | static void blk_mq_requeue_work(struct work_struct *work) |
562 | { | |
563 | struct request_queue *q = | |
2849450a | 564 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
565 | LIST_HEAD(rq_list); |
566 | struct request *rq, *next; | |
567 | unsigned long flags; | |
568 | ||
569 | spin_lock_irqsave(&q->requeue_lock, flags); | |
570 | list_splice_init(&q->requeue_list, &rq_list); | |
571 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
572 | ||
573 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 574 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
575 | continue; |
576 | ||
e8064021 | 577 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 578 | list_del_init(&rq->queuelist); |
bd6737f1 | 579 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
580 | } |
581 | ||
582 | while (!list_empty(&rq_list)) { | |
583 | rq = list_entry(rq_list.next, struct request, queuelist); | |
584 | list_del_init(&rq->queuelist); | |
bd6737f1 | 585 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
586 | } |
587 | ||
52d7f1b5 | 588 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
589 | } |
590 | ||
2b053aca BVA |
591 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
592 | bool kick_requeue_list) | |
6fca6a61 CH |
593 | { |
594 | struct request_queue *q = rq->q; | |
595 | unsigned long flags; | |
596 | ||
597 | /* | |
598 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
599 | * request head insertation from the workqueue. | |
600 | */ | |
e8064021 | 601 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
602 | |
603 | spin_lock_irqsave(&q->requeue_lock, flags); | |
604 | if (at_head) { | |
e8064021 | 605 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
606 | list_add(&rq->queuelist, &q->requeue_list); |
607 | } else { | |
608 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
609 | } | |
610 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
611 | |
612 | if (kick_requeue_list) | |
613 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
614 | } |
615 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
616 | ||
617 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
618 | { | |
2849450a | 619 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
620 | } |
621 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
622 | ||
2849450a MS |
623 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
624 | unsigned long msecs) | |
625 | { | |
626 | kblockd_schedule_delayed_work(&q->requeue_work, | |
627 | msecs_to_jiffies(msecs)); | |
628 | } | |
629 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
630 | ||
1885b24d JA |
631 | void blk_mq_abort_requeue_list(struct request_queue *q) |
632 | { | |
633 | unsigned long flags; | |
634 | LIST_HEAD(rq_list); | |
635 | ||
636 | spin_lock_irqsave(&q->requeue_lock, flags); | |
637 | list_splice_init(&q->requeue_list, &rq_list); | |
638 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
639 | ||
640 | while (!list_empty(&rq_list)) { | |
641 | struct request *rq; | |
642 | ||
643 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
644 | list_del_init(&rq->queuelist); | |
caf7df12 | 645 | blk_mq_end_request(rq, -EIO); |
1885b24d JA |
646 | } |
647 | } | |
648 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
649 | ||
0e62f51f JA |
650 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
651 | { | |
88c7b2b7 JA |
652 | if (tag < tags->nr_tags) { |
653 | prefetch(tags->rqs[tag]); | |
4ee86bab | 654 | return tags->rqs[tag]; |
88c7b2b7 | 655 | } |
4ee86bab HR |
656 | |
657 | return NULL; | |
24d2f903 CH |
658 | } |
659 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
660 | ||
320ae51f | 661 | struct blk_mq_timeout_data { |
46f92d42 CH |
662 | unsigned long next; |
663 | unsigned int next_set; | |
320ae51f JA |
664 | }; |
665 | ||
90415837 | 666 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 667 | { |
f8a5b122 | 668 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 669 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
670 | |
671 | /* | |
672 | * We know that complete is set at this point. If STARTED isn't set | |
673 | * anymore, then the request isn't active and the "timeout" should | |
674 | * just be ignored. This can happen due to the bitflag ordering. | |
675 | * Timeout first checks if STARTED is set, and if it is, assumes | |
676 | * the request is active. But if we race with completion, then | |
48b99c9d | 677 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
678 | * a timeout event with a request that isn't active. |
679 | */ | |
46f92d42 CH |
680 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
681 | return; | |
87ee7b11 | 682 | |
46f92d42 | 683 | if (ops->timeout) |
0152fb6b | 684 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
685 | |
686 | switch (ret) { | |
687 | case BLK_EH_HANDLED: | |
688 | __blk_mq_complete_request(req); | |
689 | break; | |
690 | case BLK_EH_RESET_TIMER: | |
691 | blk_add_timer(req); | |
692 | blk_clear_rq_complete(req); | |
693 | break; | |
694 | case BLK_EH_NOT_HANDLED: | |
695 | break; | |
696 | default: | |
697 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
698 | break; | |
699 | } | |
87ee7b11 | 700 | } |
5b3f25fc | 701 | |
81481eb4 CH |
702 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
703 | struct request *rq, void *priv, bool reserved) | |
704 | { | |
705 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 706 | |
95a49603 | 707 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 708 | return; |
87ee7b11 | 709 | |
d9d149a3 ML |
710 | /* |
711 | * The rq being checked may have been freed and reallocated | |
712 | * out already here, we avoid this race by checking rq->deadline | |
713 | * and REQ_ATOM_COMPLETE flag together: | |
714 | * | |
715 | * - if rq->deadline is observed as new value because of | |
716 | * reusing, the rq won't be timed out because of timing. | |
717 | * - if rq->deadline is observed as previous value, | |
718 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
719 | * because we put a barrier between setting rq->deadline | |
720 | * and clearing the flag in blk_mq_start_request(), so | |
721 | * this rq won't be timed out too. | |
722 | */ | |
46f92d42 CH |
723 | if (time_after_eq(jiffies, rq->deadline)) { |
724 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 725 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
726 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
727 | data->next = rq->deadline; | |
728 | data->next_set = 1; | |
729 | } | |
87ee7b11 JA |
730 | } |
731 | ||
287922eb | 732 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 733 | { |
287922eb CH |
734 | struct request_queue *q = |
735 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
736 | struct blk_mq_timeout_data data = { |
737 | .next = 0, | |
738 | .next_set = 0, | |
739 | }; | |
81481eb4 | 740 | int i; |
320ae51f | 741 | |
71f79fb3 GKB |
742 | /* A deadlock might occur if a request is stuck requiring a |
743 | * timeout at the same time a queue freeze is waiting | |
744 | * completion, since the timeout code would not be able to | |
745 | * acquire the queue reference here. | |
746 | * | |
747 | * That's why we don't use blk_queue_enter here; instead, we use | |
748 | * percpu_ref_tryget directly, because we need to be able to | |
749 | * obtain a reference even in the short window between the queue | |
750 | * starting to freeze, by dropping the first reference in | |
1671d522 | 751 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
752 | * consumed, marked by the instant q_usage_counter reaches |
753 | * zero. | |
754 | */ | |
755 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
756 | return; |
757 | ||
0bf6cd5b | 758 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 759 | |
81481eb4 CH |
760 | if (data.next_set) { |
761 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
762 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 763 | } else { |
0bf6cd5b CH |
764 | struct blk_mq_hw_ctx *hctx; |
765 | ||
f054b56c ML |
766 | queue_for_each_hw_ctx(q, hctx, i) { |
767 | /* the hctx may be unmapped, so check it here */ | |
768 | if (blk_mq_hw_queue_mapped(hctx)) | |
769 | blk_mq_tag_idle(hctx); | |
770 | } | |
0d2602ca | 771 | } |
287922eb | 772 | blk_queue_exit(q); |
320ae51f JA |
773 | } |
774 | ||
775 | /* | |
776 | * Reverse check our software queue for entries that we could potentially | |
777 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
778 | * too much time checking for merges. | |
779 | */ | |
780 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
781 | struct blk_mq_ctx *ctx, struct bio *bio) | |
782 | { | |
783 | struct request *rq; | |
784 | int checked = 8; | |
785 | ||
786 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 787 | bool merged = false; |
320ae51f JA |
788 | |
789 | if (!checked--) | |
790 | break; | |
791 | ||
792 | if (!blk_rq_merge_ok(rq, bio)) | |
793 | continue; | |
794 | ||
34fe7c05 CH |
795 | switch (blk_try_merge(rq, bio)) { |
796 | case ELEVATOR_BACK_MERGE: | |
797 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
798 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 799 | break; |
34fe7c05 CH |
800 | case ELEVATOR_FRONT_MERGE: |
801 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
802 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 803 | break; |
1e739730 CH |
804 | case ELEVATOR_DISCARD_MERGE: |
805 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 806 | break; |
34fe7c05 CH |
807 | default: |
808 | continue; | |
320ae51f | 809 | } |
34fe7c05 CH |
810 | |
811 | if (merged) | |
812 | ctx->rq_merged++; | |
813 | return merged; | |
320ae51f JA |
814 | } |
815 | ||
816 | return false; | |
817 | } | |
818 | ||
88459642 OS |
819 | struct flush_busy_ctx_data { |
820 | struct blk_mq_hw_ctx *hctx; | |
821 | struct list_head *list; | |
822 | }; | |
823 | ||
824 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
825 | { | |
826 | struct flush_busy_ctx_data *flush_data = data; | |
827 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
828 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
829 | ||
830 | sbitmap_clear_bit(sb, bitnr); | |
831 | spin_lock(&ctx->lock); | |
832 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
833 | spin_unlock(&ctx->lock); | |
834 | return true; | |
835 | } | |
836 | ||
1429d7c9 JA |
837 | /* |
838 | * Process software queues that have been marked busy, splicing them | |
839 | * to the for-dispatch | |
840 | */ | |
2c3ad667 | 841 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 842 | { |
88459642 OS |
843 | struct flush_busy_ctx_data data = { |
844 | .hctx = hctx, | |
845 | .list = list, | |
846 | }; | |
1429d7c9 | 847 | |
88459642 | 848 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 849 | } |
2c3ad667 | 850 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 851 | |
703fd1c0 JA |
852 | static inline unsigned int queued_to_index(unsigned int queued) |
853 | { | |
854 | if (!queued) | |
855 | return 0; | |
1429d7c9 | 856 | |
703fd1c0 | 857 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
858 | } |
859 | ||
bd6737f1 JA |
860 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
861 | bool wait) | |
bd166ef1 JA |
862 | { |
863 | struct blk_mq_alloc_data data = { | |
864 | .q = rq->q, | |
bd166ef1 JA |
865 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
866 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
867 | }; | |
868 | ||
5feeacdd JA |
869 | might_sleep_if(wait); |
870 | ||
81380ca1 OS |
871 | if (rq->tag != -1) |
872 | goto done; | |
bd166ef1 | 873 | |
415b806d SG |
874 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
875 | data.flags |= BLK_MQ_REQ_RESERVED; | |
876 | ||
bd166ef1 JA |
877 | rq->tag = blk_mq_get_tag(&data); |
878 | if (rq->tag >= 0) { | |
200e86b3 JA |
879 | if (blk_mq_tag_busy(data.hctx)) { |
880 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
881 | atomic_inc(&data.hctx->nr_active); | |
882 | } | |
bd166ef1 | 883 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
884 | } |
885 | ||
81380ca1 OS |
886 | done: |
887 | if (hctx) | |
888 | *hctx = data.hctx; | |
889 | return rq->tag != -1; | |
bd166ef1 JA |
890 | } |
891 | ||
113285b4 JA |
892 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
893 | struct request *rq) | |
99cf1dc5 | 894 | { |
99cf1dc5 JA |
895 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
896 | rq->tag = -1; | |
897 | ||
898 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
899 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
900 | atomic_dec(&hctx->nr_active); | |
901 | } | |
902 | } | |
903 | ||
113285b4 JA |
904 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
905 | struct request *rq) | |
906 | { | |
907 | if (rq->tag == -1 || rq->internal_tag == -1) | |
908 | return; | |
909 | ||
910 | __blk_mq_put_driver_tag(hctx, rq); | |
911 | } | |
912 | ||
913 | static void blk_mq_put_driver_tag(struct request *rq) | |
914 | { | |
915 | struct blk_mq_hw_ctx *hctx; | |
916 | ||
917 | if (rq->tag == -1 || rq->internal_tag == -1) | |
918 | return; | |
919 | ||
920 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
921 | __blk_mq_put_driver_tag(hctx, rq); | |
922 | } | |
923 | ||
bd166ef1 JA |
924 | /* |
925 | * If we fail getting a driver tag because all the driver tags are already | |
926 | * assigned and on the dispatch list, BUT the first entry does not have a | |
927 | * tag, then we could deadlock. For that case, move entries with assigned | |
928 | * driver tags to the front, leaving the set of tagged requests in the | |
929 | * same order, and the untagged set in the same order. | |
930 | */ | |
931 | static bool reorder_tags_to_front(struct list_head *list) | |
932 | { | |
933 | struct request *rq, *tmp, *first = NULL; | |
934 | ||
935 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
936 | if (rq == first) | |
937 | break; | |
938 | if (rq->tag != -1) { | |
939 | list_move(&rq->queuelist, list); | |
940 | if (!first) | |
941 | first = rq; | |
942 | } | |
943 | } | |
944 | ||
945 | return first != NULL; | |
946 | } | |
947 | ||
da55f2cc OS |
948 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
949 | void *key) | |
950 | { | |
951 | struct blk_mq_hw_ctx *hctx; | |
952 | ||
953 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
954 | ||
955 | list_del(&wait->task_list); | |
956 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
957 | blk_mq_run_hw_queue(hctx, true); | |
958 | return 1; | |
959 | } | |
960 | ||
961 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
962 | { | |
963 | struct sbq_wait_state *ws; | |
964 | ||
965 | /* | |
966 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
967 | * The thread which wins the race to grab this bit adds the hardware | |
968 | * queue to the wait queue. | |
969 | */ | |
970 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
971 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
972 | return false; | |
973 | ||
974 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
975 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
976 | ||
977 | /* | |
978 | * As soon as this returns, it's no longer safe to fiddle with | |
979 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
980 | * and unlock the bit. | |
981 | */ | |
982 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
983 | return true; | |
984 | } | |
985 | ||
81380ca1 | 986 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
320ae51f | 987 | { |
81380ca1 | 988 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 989 | struct request *rq; |
93efe981 | 990 | int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 991 | |
81380ca1 OS |
992 | if (list_empty(list)) |
993 | return false; | |
994 | ||
320ae51f JA |
995 | /* |
996 | * Now process all the entries, sending them to the driver. | |
997 | */ | |
93efe981 | 998 | errors = queued = 0; |
81380ca1 | 999 | do { |
74c45052 | 1000 | struct blk_mq_queue_data bd; |
320ae51f | 1001 | |
f04c3df3 | 1002 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1003 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1004 | if (!queued && reorder_tags_to_front(list)) | |
1005 | continue; | |
3c782d67 JA |
1006 | |
1007 | /* | |
da55f2cc OS |
1008 | * The initial allocation attempt failed, so we need to |
1009 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1010 | */ |
807b1041 OS |
1011 | if (!blk_mq_dispatch_wait_add(hctx)) |
1012 | break; | |
1013 | ||
1014 | /* | |
1015 | * It's possible that a tag was freed in the window | |
1016 | * between the allocation failure and adding the | |
1017 | * hardware queue to the wait queue. | |
1018 | */ | |
1019 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
3c782d67 | 1020 | break; |
bd166ef1 | 1021 | } |
da55f2cc | 1022 | |
320ae51f | 1023 | list_del_init(&rq->queuelist); |
320ae51f | 1024 | |
74c45052 | 1025 | bd.rq = rq; |
113285b4 JA |
1026 | |
1027 | /* | |
1028 | * Flag last if we have no more requests, or if we have more | |
1029 | * but can't assign a driver tag to it. | |
1030 | */ | |
1031 | if (list_empty(list)) | |
1032 | bd.last = true; | |
1033 | else { | |
1034 | struct request *nxt; | |
1035 | ||
1036 | nxt = list_first_entry(list, struct request, queuelist); | |
1037 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1038 | } | |
74c45052 JA |
1039 | |
1040 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
1041 | switch (ret) { |
1042 | case BLK_MQ_RQ_QUEUE_OK: | |
1043 | queued++; | |
52b9c330 | 1044 | break; |
320ae51f | 1045 | case BLK_MQ_RQ_QUEUE_BUSY: |
113285b4 | 1046 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1047 | list_add(&rq->queuelist, list); |
ed0791b2 | 1048 | __blk_mq_requeue_request(rq); |
320ae51f JA |
1049 | break; |
1050 | default: | |
1051 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 1052 | case BLK_MQ_RQ_QUEUE_ERROR: |
93efe981 | 1053 | errors++; |
caf7df12 | 1054 | blk_mq_end_request(rq, -EIO); |
320ae51f JA |
1055 | break; |
1056 | } | |
1057 | ||
1058 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1059 | break; | |
81380ca1 | 1060 | } while (!list_empty(list)); |
320ae51f | 1061 | |
703fd1c0 | 1062 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1063 | |
1064 | /* | |
1065 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1066 | * that is where we will continue on next queue run. | |
1067 | */ | |
f04c3df3 | 1068 | if (!list_empty(list)) { |
113285b4 | 1069 | /* |
710c785f BVA |
1070 | * If an I/O scheduler has been configured and we got a driver |
1071 | * tag for the next request already, free it again. | |
113285b4 JA |
1072 | */ |
1073 | rq = list_first_entry(list, struct request, queuelist); | |
1074 | blk_mq_put_driver_tag(rq); | |
1075 | ||
320ae51f | 1076 | spin_lock(&hctx->lock); |
c13660a0 | 1077 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1078 | spin_unlock(&hctx->lock); |
f04c3df3 | 1079 | |
9ba52e58 | 1080 | /* |
710c785f BVA |
1081 | * If SCHED_RESTART was set by the caller of this function and |
1082 | * it is no longer set that means that it was cleared by another | |
1083 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1084 | * |
710c785f BVA |
1085 | * If TAG_WAITING is set that means that an I/O scheduler has |
1086 | * been configured and another thread is waiting for a driver | |
1087 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1088 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1089 | * |
710c785f BVA |
1090 | * If no I/O scheduler has been configured it is possible that |
1091 | * the hardware queue got stopped and restarted before requests | |
1092 | * were pushed back onto the dispatch list. Rerun the queue to | |
1093 | * avoid starvation. Notes: | |
1094 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1095 | * been stopped before rerunning a queue. | |
1096 | * - Some but not all block drivers stop a queue before | |
1097 | * returning BLK_MQ_RQ_QUEUE_BUSY. Two exceptions are scsi-mq | |
1098 | * and dm-rq. | |
bd166ef1 | 1099 | */ |
da55f2cc OS |
1100 | if (!blk_mq_sched_needs_restart(hctx) && |
1101 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1102 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1103 | } |
f04c3df3 | 1104 | |
93efe981 | 1105 | return (queued + errors) != 0; |
f04c3df3 JA |
1106 | } |
1107 | ||
6a83e74d BVA |
1108 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1109 | { | |
1110 | int srcu_idx; | |
1111 | ||
1112 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1113 | cpu_online(hctx->next_cpu)); | |
1114 | ||
1115 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1116 | rcu_read_lock(); | |
bd166ef1 | 1117 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1118 | rcu_read_unlock(); |
1119 | } else { | |
bf4907c0 JA |
1120 | might_sleep(); |
1121 | ||
6a83e74d | 1122 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); |
bd166ef1 | 1123 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1124 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1125 | } | |
1126 | } | |
1127 | ||
506e931f JA |
1128 | /* |
1129 | * It'd be great if the workqueue API had a way to pass | |
1130 | * in a mask and had some smarts for more clever placement. | |
1131 | * For now we just round-robin here, switching for every | |
1132 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1133 | */ | |
1134 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1135 | { | |
b657d7e6 CH |
1136 | if (hctx->queue->nr_hw_queues == 1) |
1137 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1138 | |
1139 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1140 | int next_cpu; |
506e931f JA |
1141 | |
1142 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1143 | if (next_cpu >= nr_cpu_ids) | |
1144 | next_cpu = cpumask_first(hctx->cpumask); | |
1145 | ||
1146 | hctx->next_cpu = next_cpu; | |
1147 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1148 | } | |
1149 | ||
b657d7e6 | 1150 | return hctx->next_cpu; |
506e931f JA |
1151 | } |
1152 | ||
7587a5ae BVA |
1153 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1154 | unsigned long msecs) | |
320ae51f | 1155 | { |
5d1b25c1 BVA |
1156 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1157 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1158 | return; |
1159 | ||
1b792f2f | 1160 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1161 | int cpu = get_cpu(); |
1162 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1163 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1164 | put_cpu(); |
398205b8 PB |
1165 | return; |
1166 | } | |
e4043dcf | 1167 | |
2a90d4aa | 1168 | put_cpu(); |
e4043dcf | 1169 | } |
398205b8 | 1170 | |
9f993737 JA |
1171 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1172 | &hctx->run_work, | |
1173 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1174 | } |
1175 | ||
1176 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1177 | { | |
1178 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1179 | } | |
1180 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1181 | ||
1182 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1183 | { | |
1184 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1185 | } |
5b727272 | 1186 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1187 | |
b94ec296 | 1188 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1189 | { |
1190 | struct blk_mq_hw_ctx *hctx; | |
1191 | int i; | |
1192 | ||
1193 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1194 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1195 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1196 | continue; |
1197 | ||
b94ec296 | 1198 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1199 | } |
1200 | } | |
b94ec296 | 1201 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1202 | |
fd001443 BVA |
1203 | /** |
1204 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1205 | * @q: request queue. | |
1206 | * | |
1207 | * The caller is responsible for serializing this function against | |
1208 | * blk_mq_{start,stop}_hw_queue(). | |
1209 | */ | |
1210 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1211 | { | |
1212 | struct blk_mq_hw_ctx *hctx; | |
1213 | int i; | |
1214 | ||
1215 | queue_for_each_hw_ctx(q, hctx, i) | |
1216 | if (blk_mq_hctx_stopped(hctx)) | |
1217 | return true; | |
1218 | ||
1219 | return false; | |
1220 | } | |
1221 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1222 | ||
2719aa21 | 1223 | static void __blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx, bool sync) |
320ae51f | 1224 | { |
2719aa21 JA |
1225 | if (sync) |
1226 | cancel_delayed_work_sync(&hctx->run_work); | |
1227 | else | |
1228 | cancel_delayed_work(&hctx->run_work); | |
1229 | ||
320ae51f JA |
1230 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1231 | } | |
2719aa21 JA |
1232 | |
1233 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1234 | { | |
1235 | __blk_mq_stop_hw_queue(hctx, false); | |
1236 | } | |
320ae51f JA |
1237 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
1238 | ||
ebd76857 | 1239 | static void __blk_mq_stop_hw_queues(struct request_queue *q, bool sync) |
280d45f6 CH |
1240 | { |
1241 | struct blk_mq_hw_ctx *hctx; | |
1242 | int i; | |
1243 | ||
1244 | queue_for_each_hw_ctx(q, hctx, i) | |
2719aa21 JA |
1245 | __blk_mq_stop_hw_queue(hctx, sync); |
1246 | } | |
1247 | ||
1248 | void blk_mq_stop_hw_queues(struct request_queue *q) | |
1249 | { | |
1250 | __blk_mq_stop_hw_queues(q, false); | |
280d45f6 CH |
1251 | } |
1252 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1253 | ||
320ae51f JA |
1254 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1255 | { | |
1256 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1257 | |
0ffbce80 | 1258 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1259 | } |
1260 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1261 | ||
2f268556 CH |
1262 | void blk_mq_start_hw_queues(struct request_queue *q) |
1263 | { | |
1264 | struct blk_mq_hw_ctx *hctx; | |
1265 | int i; | |
1266 | ||
1267 | queue_for_each_hw_ctx(q, hctx, i) | |
1268 | blk_mq_start_hw_queue(hctx); | |
1269 | } | |
1270 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1271 | ||
ae911c5e JA |
1272 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1273 | { | |
1274 | if (!blk_mq_hctx_stopped(hctx)) | |
1275 | return; | |
1276 | ||
1277 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1278 | blk_mq_run_hw_queue(hctx, async); | |
1279 | } | |
1280 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1281 | ||
1b4a3258 | 1282 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1283 | { |
1284 | struct blk_mq_hw_ctx *hctx; | |
1285 | int i; | |
1286 | ||
ae911c5e JA |
1287 | queue_for_each_hw_ctx(q, hctx, i) |
1288 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1289 | } |
1290 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1291 | ||
70f4db63 | 1292 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1293 | { |
1294 | struct blk_mq_hw_ctx *hctx; | |
1295 | ||
9f993737 | 1296 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1297 | |
21c6e939 JA |
1298 | /* |
1299 | * If we are stopped, don't run the queue. The exception is if | |
1300 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1301 | * the STOPPED bit and run it. | |
1302 | */ | |
1303 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1304 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1305 | return; | |
7587a5ae | 1306 | |
21c6e939 JA |
1307 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1308 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1309 | } | |
7587a5ae BVA |
1310 | |
1311 | __blk_mq_run_hw_queue(hctx); | |
1312 | } | |
1313 | ||
70f4db63 CH |
1314 | |
1315 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1316 | { | |
19c66e59 ML |
1317 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1318 | return; | |
70f4db63 | 1319 | |
21c6e939 JA |
1320 | /* |
1321 | * Stop the hw queue, then modify currently delayed work. | |
1322 | * This should prevent us from running the queue prematurely. | |
1323 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1324 | */ | |
7e79dadc | 1325 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1326 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1327 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1328 | &hctx->run_work, | |
1329 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1330 | } |
1331 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1332 | ||
cfd0c552 | 1333 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1334 | struct request *rq, |
1335 | bool at_head) | |
320ae51f | 1336 | { |
e57690fe JA |
1337 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1338 | ||
01b983c9 JA |
1339 | trace_block_rq_insert(hctx->queue, rq); |
1340 | ||
72a0a36e CH |
1341 | if (at_head) |
1342 | list_add(&rq->queuelist, &ctx->rq_list); | |
1343 | else | |
1344 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1345 | } |
4bb659b1 | 1346 | |
2c3ad667 JA |
1347 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1348 | bool at_head) | |
cfd0c552 ML |
1349 | { |
1350 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1351 | ||
e57690fe | 1352 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1353 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1354 | } |
1355 | ||
bd166ef1 JA |
1356 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1357 | struct list_head *list) | |
320ae51f JA |
1358 | |
1359 | { | |
320ae51f JA |
1360 | /* |
1361 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1362 | * offline now | |
1363 | */ | |
1364 | spin_lock(&ctx->lock); | |
1365 | while (!list_empty(list)) { | |
1366 | struct request *rq; | |
1367 | ||
1368 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1369 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1370 | list_del_init(&rq->queuelist); |
e57690fe | 1371 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1372 | } |
cfd0c552 | 1373 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1374 | spin_unlock(&ctx->lock); |
320ae51f JA |
1375 | } |
1376 | ||
1377 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1378 | { | |
1379 | struct request *rqa = container_of(a, struct request, queuelist); | |
1380 | struct request *rqb = container_of(b, struct request, queuelist); | |
1381 | ||
1382 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1383 | (rqa->mq_ctx == rqb->mq_ctx && | |
1384 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1385 | } | |
1386 | ||
1387 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1388 | { | |
1389 | struct blk_mq_ctx *this_ctx; | |
1390 | struct request_queue *this_q; | |
1391 | struct request *rq; | |
1392 | LIST_HEAD(list); | |
1393 | LIST_HEAD(ctx_list); | |
1394 | unsigned int depth; | |
1395 | ||
1396 | list_splice_init(&plug->mq_list, &list); | |
1397 | ||
1398 | list_sort(NULL, &list, plug_ctx_cmp); | |
1399 | ||
1400 | this_q = NULL; | |
1401 | this_ctx = NULL; | |
1402 | depth = 0; | |
1403 | ||
1404 | while (!list_empty(&list)) { | |
1405 | rq = list_entry_rq(list.next); | |
1406 | list_del_init(&rq->queuelist); | |
1407 | BUG_ON(!rq->q); | |
1408 | if (rq->mq_ctx != this_ctx) { | |
1409 | if (this_ctx) { | |
bd166ef1 JA |
1410 | trace_block_unplug(this_q, depth, from_schedule); |
1411 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1412 | &ctx_list, | |
1413 | from_schedule); | |
320ae51f JA |
1414 | } |
1415 | ||
1416 | this_ctx = rq->mq_ctx; | |
1417 | this_q = rq->q; | |
1418 | depth = 0; | |
1419 | } | |
1420 | ||
1421 | depth++; | |
1422 | list_add_tail(&rq->queuelist, &ctx_list); | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * If 'this_ctx' is set, we know we have entries to complete | |
1427 | * on 'ctx_list'. Do those. | |
1428 | */ | |
1429 | if (this_ctx) { | |
bd166ef1 JA |
1430 | trace_block_unplug(this_q, depth, from_schedule); |
1431 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1432 | from_schedule); | |
320ae51f JA |
1433 | } |
1434 | } | |
1435 | ||
1436 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1437 | { | |
da8d7f07 | 1438 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1439 | |
6e85eaf3 | 1440 | blk_account_io_start(rq, true); |
320ae51f JA |
1441 | } |
1442 | ||
274a5843 JA |
1443 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1444 | { | |
1445 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1446 | !blk_queue_nomerges(hctx->queue); | |
1447 | } | |
1448 | ||
07068d5b JA |
1449 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1450 | struct blk_mq_ctx *ctx, | |
1451 | struct request *rq, struct bio *bio) | |
320ae51f | 1452 | { |
e18378a6 | 1453 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1454 | blk_mq_bio_to_request(rq, bio); |
1455 | spin_lock(&ctx->lock); | |
1456 | insert_rq: | |
1457 | __blk_mq_insert_request(hctx, rq, false); | |
1458 | spin_unlock(&ctx->lock); | |
1459 | return false; | |
1460 | } else { | |
274a5843 JA |
1461 | struct request_queue *q = hctx->queue; |
1462 | ||
07068d5b JA |
1463 | spin_lock(&ctx->lock); |
1464 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1465 | blk_mq_bio_to_request(rq, bio); | |
1466 | goto insert_rq; | |
1467 | } | |
320ae51f | 1468 | |
07068d5b | 1469 | spin_unlock(&ctx->lock); |
bd166ef1 | 1470 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1471 | return true; |
14ec77f3 | 1472 | } |
07068d5b | 1473 | } |
14ec77f3 | 1474 | |
fd2d3326 JA |
1475 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1476 | { | |
bd166ef1 JA |
1477 | if (rq->tag != -1) |
1478 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1479 | ||
1480 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1481 | } |
1482 | ||
5eb6126e | 1483 | static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie, |
9c621104 | 1484 | bool may_sleep) |
f984df1f | 1485 | { |
f984df1f | 1486 | struct request_queue *q = rq->q; |
f984df1f SL |
1487 | struct blk_mq_queue_data bd = { |
1488 | .rq = rq, | |
d945a365 | 1489 | .last = true, |
f984df1f | 1490 | }; |
bd166ef1 JA |
1491 | struct blk_mq_hw_ctx *hctx; |
1492 | blk_qc_t new_cookie; | |
1493 | int ret; | |
f984df1f | 1494 | |
bd166ef1 | 1495 | if (q->elevator) |
2253efc8 BVA |
1496 | goto insert; |
1497 | ||
bd166ef1 JA |
1498 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1499 | goto insert; | |
1500 | ||
1501 | new_cookie = request_to_qc_t(hctx, rq); | |
1502 | ||
f984df1f SL |
1503 | /* |
1504 | * For OK queue, we are done. For error, kill it. Any other | |
1505 | * error (busy), just add it to our list as we previously | |
1506 | * would have done | |
1507 | */ | |
1508 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1509 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1510 | *cookie = new_cookie; | |
2253efc8 | 1511 | return; |
7b371636 | 1512 | } |
f984df1f | 1513 | |
7b371636 JA |
1514 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { |
1515 | *cookie = BLK_QC_T_NONE; | |
caf7df12 | 1516 | blk_mq_end_request(rq, -EIO); |
2253efc8 | 1517 | return; |
f984df1f | 1518 | } |
7b371636 | 1519 | |
b58e1769 | 1520 | __blk_mq_requeue_request(rq); |
2253efc8 | 1521 | insert: |
9c621104 | 1522 | blk_mq_sched_insert_request(rq, false, true, false, may_sleep); |
f984df1f SL |
1523 | } |
1524 | ||
5eb6126e CH |
1525 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1526 | struct request *rq, blk_qc_t *cookie) | |
1527 | { | |
1528 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1529 | rcu_read_lock(); | |
1530 | __blk_mq_try_issue_directly(rq, cookie, false); | |
1531 | rcu_read_unlock(); | |
1532 | } else { | |
bf4907c0 JA |
1533 | unsigned int srcu_idx; |
1534 | ||
1535 | might_sleep(); | |
1536 | ||
1537 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
5eb6126e CH |
1538 | __blk_mq_try_issue_directly(rq, cookie, true); |
1539 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
1540 | } | |
1541 | } | |
1542 | ||
dece1635 | 1543 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1544 | { |
ef295ecf | 1545 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1546 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1547 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1548 | struct request *rq; |
5eb6126e | 1549 | unsigned int request_count = 0; |
f984df1f | 1550 | struct blk_plug *plug; |
5b3f341f | 1551 | struct request *same_queue_rq = NULL; |
7b371636 | 1552 | blk_qc_t cookie; |
87760e5e | 1553 | unsigned int wb_acct; |
07068d5b JA |
1554 | |
1555 | blk_queue_bounce(q, &bio); | |
1556 | ||
f36ea50c WX |
1557 | blk_queue_split(q, &bio, q->bio_split); |
1558 | ||
07068d5b | 1559 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
4246a0b6 | 1560 | bio_io_error(bio); |
dece1635 | 1561 | return BLK_QC_T_NONE; |
07068d5b JA |
1562 | } |
1563 | ||
87c279e6 OS |
1564 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1565 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1566 | return BLK_QC_T_NONE; | |
f984df1f | 1567 | |
bd166ef1 JA |
1568 | if (blk_mq_sched_bio_merge(q, bio)) |
1569 | return BLK_QC_T_NONE; | |
1570 | ||
87760e5e JA |
1571 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1572 | ||
bd166ef1 JA |
1573 | trace_block_getrq(q, bio, bio->bi_opf); |
1574 | ||
1575 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1576 | if (unlikely(!rq)) { |
1577 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1578 | return BLK_QC_T_NONE; |
87760e5e JA |
1579 | } |
1580 | ||
1581 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1582 | |
fd2d3326 | 1583 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1584 | |
f984df1f | 1585 | plug = current->plug; |
07068d5b | 1586 | if (unlikely(is_flush_fua)) { |
f984df1f | 1587 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1588 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1589 | if (q->elevator) { |
1590 | blk_mq_sched_insert_request(rq, false, true, true, | |
1591 | true); | |
6a83e74d | 1592 | } else { |
a4d907b6 CH |
1593 | blk_insert_flush(rq); |
1594 | blk_mq_run_hw_queue(data.hctx, true); | |
6a83e74d | 1595 | } |
a4d907b6 | 1596 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1597 | struct request *last = NULL; |
1598 | ||
b00c53e8 | 1599 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1600 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1601 | |
1602 | /* | |
1603 | * @request_count may become stale because of schedule | |
1604 | * out, so check the list again. | |
1605 | */ | |
1606 | if (list_empty(&plug->mq_list)) | |
1607 | request_count = 0; | |
254d259d CH |
1608 | else if (blk_queue_nomerges(q)) |
1609 | request_count = blk_plug_queued_count(q); | |
1610 | ||
676d0607 | 1611 | if (!request_count) |
e6c4438b | 1612 | trace_block_plug(q); |
600271d9 SL |
1613 | else |
1614 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1615 | |
600271d9 SL |
1616 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1617 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1618 | blk_flush_plug_list(plug, false); |
1619 | trace_block_plug(q); | |
320ae51f | 1620 | } |
b094f89c | 1621 | |
e6c4438b | 1622 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1623 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1624 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1625 | |
07068d5b | 1626 | /* |
6a83e74d | 1627 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1628 | * Otherwise the existing request in the plug list will be |
1629 | * issued. So the plug list will have one request at most | |
2299722c CH |
1630 | * The plug list might get flushed before this. If that happens, |
1631 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1632 | */ |
2299722c CH |
1633 | if (list_empty(&plug->mq_list)) |
1634 | same_queue_rq = NULL; | |
1635 | if (same_queue_rq) | |
1636 | list_del_init(&same_queue_rq->queuelist); | |
1637 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1638 | ||
bf4907c0 JA |
1639 | blk_mq_put_ctx(data.ctx); |
1640 | ||
2299722c CH |
1641 | if (same_queue_rq) |
1642 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1643 | &cookie); | |
a4d907b6 | 1644 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1645 | blk_mq_put_ctx(data.ctx); |
2299722c | 1646 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1647 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1648 | } else if (q->elevator) { |
b00c53e8 | 1649 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1650 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1651 | blk_mq_sched_insert_request(rq, false, true, true, true); |
b00c53e8 JA |
1652 | } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1653 | blk_mq_put_ctx(data.ctx); | |
a4d907b6 | 1654 | blk_mq_run_hw_queue(data.hctx, true); |
abc25a69 BVA |
1655 | } else |
1656 | blk_mq_put_ctx(data.ctx); | |
320ae51f | 1657 | |
7b371636 | 1658 | return cookie; |
320ae51f JA |
1659 | } |
1660 | ||
cc71a6f4 JA |
1661 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1662 | unsigned int hctx_idx) | |
95363efd | 1663 | { |
e9b267d9 | 1664 | struct page *page; |
320ae51f | 1665 | |
24d2f903 | 1666 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1667 | int i; |
320ae51f | 1668 | |
24d2f903 | 1669 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1670 | struct request *rq = tags->static_rqs[i]; |
1671 | ||
1672 | if (!rq) | |
e9b267d9 | 1673 | continue; |
d6296d39 | 1674 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1675 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1676 | } |
320ae51f | 1677 | } |
320ae51f | 1678 | |
24d2f903 CH |
1679 | while (!list_empty(&tags->page_list)) { |
1680 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1681 | list_del_init(&page->lru); |
f75782e4 CM |
1682 | /* |
1683 | * Remove kmemleak object previously allocated in | |
1684 | * blk_mq_init_rq_map(). | |
1685 | */ | |
1686 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1687 | __free_pages(page, page->private); |
1688 | } | |
cc71a6f4 | 1689 | } |
320ae51f | 1690 | |
cc71a6f4 JA |
1691 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1692 | { | |
24d2f903 | 1693 | kfree(tags->rqs); |
cc71a6f4 | 1694 | tags->rqs = NULL; |
2af8cbe3 JA |
1695 | kfree(tags->static_rqs); |
1696 | tags->static_rqs = NULL; | |
320ae51f | 1697 | |
24d2f903 | 1698 | blk_mq_free_tags(tags); |
320ae51f JA |
1699 | } |
1700 | ||
cc71a6f4 JA |
1701 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1702 | unsigned int hctx_idx, | |
1703 | unsigned int nr_tags, | |
1704 | unsigned int reserved_tags) | |
320ae51f | 1705 | { |
24d2f903 | 1706 | struct blk_mq_tags *tags; |
59f082e4 | 1707 | int node; |
320ae51f | 1708 | |
59f082e4 SL |
1709 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1710 | if (node == NUMA_NO_NODE) | |
1711 | node = set->numa_node; | |
1712 | ||
1713 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1714 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1715 | if (!tags) |
1716 | return NULL; | |
320ae51f | 1717 | |
cc71a6f4 | 1718 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1719 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1720 | node); |
24d2f903 CH |
1721 | if (!tags->rqs) { |
1722 | blk_mq_free_tags(tags); | |
1723 | return NULL; | |
1724 | } | |
320ae51f | 1725 | |
2af8cbe3 JA |
1726 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1727 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1728 | node); |
2af8cbe3 JA |
1729 | if (!tags->static_rqs) { |
1730 | kfree(tags->rqs); | |
1731 | blk_mq_free_tags(tags); | |
1732 | return NULL; | |
1733 | } | |
1734 | ||
cc71a6f4 JA |
1735 | return tags; |
1736 | } | |
1737 | ||
1738 | static size_t order_to_size(unsigned int order) | |
1739 | { | |
1740 | return (size_t)PAGE_SIZE << order; | |
1741 | } | |
1742 | ||
1743 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1744 | unsigned int hctx_idx, unsigned int depth) | |
1745 | { | |
1746 | unsigned int i, j, entries_per_page, max_order = 4; | |
1747 | size_t rq_size, left; | |
59f082e4 SL |
1748 | int node; |
1749 | ||
1750 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1751 | if (node == NUMA_NO_NODE) | |
1752 | node = set->numa_node; | |
cc71a6f4 JA |
1753 | |
1754 | INIT_LIST_HEAD(&tags->page_list); | |
1755 | ||
320ae51f JA |
1756 | /* |
1757 | * rq_size is the size of the request plus driver payload, rounded | |
1758 | * to the cacheline size | |
1759 | */ | |
24d2f903 | 1760 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1761 | cache_line_size()); |
cc71a6f4 | 1762 | left = rq_size * depth; |
320ae51f | 1763 | |
cc71a6f4 | 1764 | for (i = 0; i < depth; ) { |
320ae51f JA |
1765 | int this_order = max_order; |
1766 | struct page *page; | |
1767 | int to_do; | |
1768 | void *p; | |
1769 | ||
b3a834b1 | 1770 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1771 | this_order--; |
1772 | ||
1773 | do { | |
59f082e4 | 1774 | page = alloc_pages_node(node, |
36e1f3d1 | 1775 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1776 | this_order); |
320ae51f JA |
1777 | if (page) |
1778 | break; | |
1779 | if (!this_order--) | |
1780 | break; | |
1781 | if (order_to_size(this_order) < rq_size) | |
1782 | break; | |
1783 | } while (1); | |
1784 | ||
1785 | if (!page) | |
24d2f903 | 1786 | goto fail; |
320ae51f JA |
1787 | |
1788 | page->private = this_order; | |
24d2f903 | 1789 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1790 | |
1791 | p = page_address(page); | |
f75782e4 CM |
1792 | /* |
1793 | * Allow kmemleak to scan these pages as they contain pointers | |
1794 | * to additional allocations like via ops->init_request(). | |
1795 | */ | |
36e1f3d1 | 1796 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1797 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1798 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1799 | left -= to_do * rq_size; |
1800 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1801 | struct request *rq = p; |
1802 | ||
1803 | tags->static_rqs[i] = rq; | |
24d2f903 | 1804 | if (set->ops->init_request) { |
d6296d39 | 1805 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1806 | node)) { |
2af8cbe3 | 1807 | tags->static_rqs[i] = NULL; |
24d2f903 | 1808 | goto fail; |
a5164405 | 1809 | } |
e9b267d9 CH |
1810 | } |
1811 | ||
320ae51f JA |
1812 | p += rq_size; |
1813 | i++; | |
1814 | } | |
1815 | } | |
cc71a6f4 | 1816 | return 0; |
320ae51f | 1817 | |
24d2f903 | 1818 | fail: |
cc71a6f4 JA |
1819 | blk_mq_free_rqs(set, tags, hctx_idx); |
1820 | return -ENOMEM; | |
320ae51f JA |
1821 | } |
1822 | ||
e57690fe JA |
1823 | /* |
1824 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1825 | * software queue to the hw queue dispatch list, and ensure that it | |
1826 | * gets run. | |
1827 | */ | |
9467f859 | 1828 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1829 | { |
9467f859 | 1830 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1831 | struct blk_mq_ctx *ctx; |
1832 | LIST_HEAD(tmp); | |
1833 | ||
9467f859 | 1834 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1835 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1836 | |
1837 | spin_lock(&ctx->lock); | |
1838 | if (!list_empty(&ctx->rq_list)) { | |
1839 | list_splice_init(&ctx->rq_list, &tmp); | |
1840 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1841 | } | |
1842 | spin_unlock(&ctx->lock); | |
1843 | ||
1844 | if (list_empty(&tmp)) | |
9467f859 | 1845 | return 0; |
484b4061 | 1846 | |
e57690fe JA |
1847 | spin_lock(&hctx->lock); |
1848 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1849 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1850 | |
1851 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1852 | return 0; |
484b4061 JA |
1853 | } |
1854 | ||
9467f859 | 1855 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1856 | { |
9467f859 TG |
1857 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1858 | &hctx->cpuhp_dead); | |
484b4061 JA |
1859 | } |
1860 | ||
c3b4afca | 1861 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1862 | static void blk_mq_exit_hctx(struct request_queue *q, |
1863 | struct blk_mq_tag_set *set, | |
1864 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1865 | { | |
9c1051aa OS |
1866 | blk_mq_debugfs_unregister_hctx(hctx); |
1867 | ||
08e98fc6 ML |
1868 | blk_mq_tag_idle(hctx); |
1869 | ||
f70ced09 | 1870 | if (set->ops->exit_request) |
d6296d39 | 1871 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1872 | |
93252632 OS |
1873 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1874 | ||
08e98fc6 ML |
1875 | if (set->ops->exit_hctx) |
1876 | set->ops->exit_hctx(hctx, hctx_idx); | |
1877 | ||
6a83e74d BVA |
1878 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1879 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1880 | ||
9467f859 | 1881 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1882 | blk_free_flush_queue(hctx->fq); |
88459642 | 1883 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1884 | } |
1885 | ||
624dbe47 ML |
1886 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1887 | struct blk_mq_tag_set *set, int nr_queue) | |
1888 | { | |
1889 | struct blk_mq_hw_ctx *hctx; | |
1890 | unsigned int i; | |
1891 | ||
1892 | queue_for_each_hw_ctx(q, hctx, i) { | |
1893 | if (i == nr_queue) | |
1894 | break; | |
08e98fc6 | 1895 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1896 | } |
624dbe47 ML |
1897 | } |
1898 | ||
08e98fc6 ML |
1899 | static int blk_mq_init_hctx(struct request_queue *q, |
1900 | struct blk_mq_tag_set *set, | |
1901 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1902 | { |
08e98fc6 ML |
1903 | int node; |
1904 | ||
1905 | node = hctx->numa_node; | |
1906 | if (node == NUMA_NO_NODE) | |
1907 | node = hctx->numa_node = set->numa_node; | |
1908 | ||
9f993737 | 1909 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1910 | spin_lock_init(&hctx->lock); |
1911 | INIT_LIST_HEAD(&hctx->dispatch); | |
1912 | hctx->queue = q; | |
1913 | hctx->queue_num = hctx_idx; | |
2404e607 | 1914 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1915 | |
9467f859 | 1916 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1917 | |
1918 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1919 | |
1920 | /* | |
08e98fc6 ML |
1921 | * Allocate space for all possible cpus to avoid allocation at |
1922 | * runtime | |
320ae51f | 1923 | */ |
08e98fc6 ML |
1924 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1925 | GFP_KERNEL, node); | |
1926 | if (!hctx->ctxs) | |
1927 | goto unregister_cpu_notifier; | |
320ae51f | 1928 | |
88459642 OS |
1929 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1930 | node)) | |
08e98fc6 | 1931 | goto free_ctxs; |
320ae51f | 1932 | |
08e98fc6 | 1933 | hctx->nr_ctx = 0; |
320ae51f | 1934 | |
08e98fc6 ML |
1935 | if (set->ops->init_hctx && |
1936 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1937 | goto free_bitmap; | |
320ae51f | 1938 | |
93252632 OS |
1939 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
1940 | goto exit_hctx; | |
1941 | ||
f70ced09 ML |
1942 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1943 | if (!hctx->fq) | |
93252632 | 1944 | goto sched_exit_hctx; |
320ae51f | 1945 | |
f70ced09 | 1946 | if (set->ops->init_request && |
d6296d39 CH |
1947 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
1948 | node)) | |
f70ced09 | 1949 | goto free_fq; |
320ae51f | 1950 | |
6a83e74d BVA |
1951 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1952 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1953 | ||
9c1051aa OS |
1954 | blk_mq_debugfs_register_hctx(q, hctx); |
1955 | ||
08e98fc6 | 1956 | return 0; |
320ae51f | 1957 | |
f70ced09 ML |
1958 | free_fq: |
1959 | kfree(hctx->fq); | |
93252632 OS |
1960 | sched_exit_hctx: |
1961 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
1962 | exit_hctx: |
1963 | if (set->ops->exit_hctx) | |
1964 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1965 | free_bitmap: |
88459642 | 1966 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1967 | free_ctxs: |
1968 | kfree(hctx->ctxs); | |
1969 | unregister_cpu_notifier: | |
9467f859 | 1970 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1971 | return -1; |
1972 | } | |
320ae51f | 1973 | |
320ae51f JA |
1974 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1975 | unsigned int nr_hw_queues) | |
1976 | { | |
1977 | unsigned int i; | |
1978 | ||
1979 | for_each_possible_cpu(i) { | |
1980 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1981 | struct blk_mq_hw_ctx *hctx; | |
1982 | ||
320ae51f JA |
1983 | __ctx->cpu = i; |
1984 | spin_lock_init(&__ctx->lock); | |
1985 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1986 | __ctx->queue = q; | |
1987 | ||
1988 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1989 | if (!cpu_online(i)) |
1990 | continue; | |
1991 | ||
7d7e0f90 | 1992 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1993 | |
320ae51f JA |
1994 | /* |
1995 | * Set local node, IFF we have more than one hw queue. If | |
1996 | * not, we remain on the home node of the device | |
1997 | */ | |
1998 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1999 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2000 | } |
2001 | } | |
2002 | ||
cc71a6f4 JA |
2003 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2004 | { | |
2005 | int ret = 0; | |
2006 | ||
2007 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2008 | set->queue_depth, set->reserved_tags); | |
2009 | if (!set->tags[hctx_idx]) | |
2010 | return false; | |
2011 | ||
2012 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2013 | set->queue_depth); | |
2014 | if (!ret) | |
2015 | return true; | |
2016 | ||
2017 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2018 | set->tags[hctx_idx] = NULL; | |
2019 | return false; | |
2020 | } | |
2021 | ||
2022 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2023 | unsigned int hctx_idx) | |
2024 | { | |
bd166ef1 JA |
2025 | if (set->tags[hctx_idx]) { |
2026 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2027 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2028 | set->tags[hctx_idx] = NULL; | |
2029 | } | |
cc71a6f4 JA |
2030 | } |
2031 | ||
5778322e AM |
2032 | static void blk_mq_map_swqueue(struct request_queue *q, |
2033 | const struct cpumask *online_mask) | |
320ae51f | 2034 | { |
d1b1cea1 | 2035 | unsigned int i, hctx_idx; |
320ae51f JA |
2036 | struct blk_mq_hw_ctx *hctx; |
2037 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2038 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2039 | |
60de074b AM |
2040 | /* |
2041 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2042 | */ | |
2043 | mutex_lock(&q->sysfs_lock); | |
2044 | ||
320ae51f | 2045 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2046 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2047 | hctx->nr_ctx = 0; |
2048 | } | |
2049 | ||
2050 | /* | |
2051 | * Map software to hardware queues | |
2052 | */ | |
897bb0c7 | 2053 | for_each_possible_cpu(i) { |
320ae51f | 2054 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2055 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2056 | continue; |
2057 | ||
d1b1cea1 GKB |
2058 | hctx_idx = q->mq_map[i]; |
2059 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2060 | if (!set->tags[hctx_idx] && |
2061 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2062 | /* |
2063 | * If tags initialization fail for some hctx, | |
2064 | * that hctx won't be brought online. In this | |
2065 | * case, remap the current ctx to hctx[0] which | |
2066 | * is guaranteed to always have tags allocated | |
2067 | */ | |
cc71a6f4 | 2068 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2069 | } |
2070 | ||
897bb0c7 | 2071 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2072 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2073 | |
e4043dcf | 2074 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2075 | ctx->index_hw = hctx->nr_ctx; |
2076 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2077 | } | |
506e931f | 2078 | |
60de074b AM |
2079 | mutex_unlock(&q->sysfs_lock); |
2080 | ||
506e931f | 2081 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2082 | /* |
a68aafa5 JA |
2083 | * If no software queues are mapped to this hardware queue, |
2084 | * disable it and free the request entries. | |
484b4061 JA |
2085 | */ |
2086 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2087 | /* Never unmap queue 0. We need it as a |
2088 | * fallback in case of a new remap fails | |
2089 | * allocation | |
2090 | */ | |
cc71a6f4 JA |
2091 | if (i && set->tags[i]) |
2092 | blk_mq_free_map_and_requests(set, i); | |
2093 | ||
2a34c087 | 2094 | hctx->tags = NULL; |
484b4061 JA |
2095 | continue; |
2096 | } | |
2097 | ||
2a34c087 ML |
2098 | hctx->tags = set->tags[i]; |
2099 | WARN_ON(!hctx->tags); | |
2100 | ||
889fa31f CY |
2101 | /* |
2102 | * Set the map size to the number of mapped software queues. | |
2103 | * This is more accurate and more efficient than looping | |
2104 | * over all possibly mapped software queues. | |
2105 | */ | |
88459642 | 2106 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2107 | |
484b4061 JA |
2108 | /* |
2109 | * Initialize batch roundrobin counts | |
2110 | */ | |
506e931f JA |
2111 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2112 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2113 | } | |
320ae51f JA |
2114 | } |
2115 | ||
2404e607 | 2116 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2117 | { |
2118 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2119 | int i; |
2120 | ||
2404e607 JM |
2121 | queue_for_each_hw_ctx(q, hctx, i) { |
2122 | if (shared) | |
2123 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2124 | else | |
2125 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2126 | } | |
2127 | } | |
2128 | ||
2129 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2130 | { | |
2131 | struct request_queue *q; | |
0d2602ca | 2132 | |
705cda97 BVA |
2133 | lockdep_assert_held(&set->tag_list_lock); |
2134 | ||
0d2602ca JA |
2135 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2136 | blk_mq_freeze_queue(q); | |
2404e607 | 2137 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2138 | blk_mq_unfreeze_queue(q); |
2139 | } | |
2140 | } | |
2141 | ||
2142 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2143 | { | |
2144 | struct blk_mq_tag_set *set = q->tag_set; | |
2145 | ||
0d2602ca | 2146 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2147 | list_del_rcu(&q->tag_set_list); |
2148 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2149 | if (list_is_singular(&set->tag_list)) { |
2150 | /* just transitioned to unshared */ | |
2151 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2152 | /* update existing queue */ | |
2153 | blk_mq_update_tag_set_depth(set, false); | |
2154 | } | |
0d2602ca | 2155 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2156 | |
2157 | synchronize_rcu(); | |
0d2602ca JA |
2158 | } |
2159 | ||
2160 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2161 | struct request_queue *q) | |
2162 | { | |
2163 | q->tag_set = set; | |
2164 | ||
2165 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2166 | |
2167 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2168 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2169 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2170 | /* update existing queue */ | |
2171 | blk_mq_update_tag_set_depth(set, true); | |
2172 | } | |
2173 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2174 | queue_set_hctx_shared(q, true); | |
705cda97 | 2175 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2176 | |
0d2602ca JA |
2177 | mutex_unlock(&set->tag_list_lock); |
2178 | } | |
2179 | ||
e09aae7e ML |
2180 | /* |
2181 | * It is the actual release handler for mq, but we do it from | |
2182 | * request queue's release handler for avoiding use-after-free | |
2183 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2184 | * but we can't group ctx/kctx kobj without it. | |
2185 | */ | |
2186 | void blk_mq_release(struct request_queue *q) | |
2187 | { | |
2188 | struct blk_mq_hw_ctx *hctx; | |
2189 | unsigned int i; | |
2190 | ||
2191 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2192 | queue_for_each_hw_ctx(q, hctx, i) { |
2193 | if (!hctx) | |
2194 | continue; | |
6c8b232e | 2195 | kobject_put(&hctx->kobj); |
c3b4afca | 2196 | } |
e09aae7e | 2197 | |
a723bab3 AM |
2198 | q->mq_map = NULL; |
2199 | ||
e09aae7e ML |
2200 | kfree(q->queue_hw_ctx); |
2201 | ||
7ea5fe31 ML |
2202 | /* |
2203 | * release .mq_kobj and sw queue's kobject now because | |
2204 | * both share lifetime with request queue. | |
2205 | */ | |
2206 | blk_mq_sysfs_deinit(q); | |
2207 | ||
e09aae7e ML |
2208 | free_percpu(q->queue_ctx); |
2209 | } | |
2210 | ||
24d2f903 | 2211 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2212 | { |
2213 | struct request_queue *uninit_q, *q; | |
2214 | ||
2215 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2216 | if (!uninit_q) | |
2217 | return ERR_PTR(-ENOMEM); | |
2218 | ||
2219 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2220 | if (IS_ERR(q)) | |
2221 | blk_cleanup_queue(uninit_q); | |
2222 | ||
2223 | return q; | |
2224 | } | |
2225 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2226 | ||
868f2f0b KB |
2227 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2228 | struct request_queue *q) | |
320ae51f | 2229 | { |
868f2f0b KB |
2230 | int i, j; |
2231 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2232 | |
868f2f0b | 2233 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2234 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2235 | int node; |
f14bbe77 | 2236 | |
868f2f0b KB |
2237 | if (hctxs[i]) |
2238 | continue; | |
2239 | ||
2240 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2241 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2242 | GFP_KERNEL, node); | |
320ae51f | 2243 | if (!hctxs[i]) |
868f2f0b | 2244 | break; |
320ae51f | 2245 | |
a86073e4 | 2246 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2247 | node)) { |
2248 | kfree(hctxs[i]); | |
2249 | hctxs[i] = NULL; | |
2250 | break; | |
2251 | } | |
e4043dcf | 2252 | |
0d2602ca | 2253 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2254 | hctxs[i]->numa_node = node; |
320ae51f | 2255 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2256 | |
2257 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2258 | free_cpumask_var(hctxs[i]->cpumask); | |
2259 | kfree(hctxs[i]); | |
2260 | hctxs[i] = NULL; | |
2261 | break; | |
2262 | } | |
2263 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2264 | } |
868f2f0b KB |
2265 | for (j = i; j < q->nr_hw_queues; j++) { |
2266 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2267 | ||
2268 | if (hctx) { | |
cc71a6f4 JA |
2269 | if (hctx->tags) |
2270 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2271 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2272 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2273 | hctxs[j] = NULL; |
2274 | ||
2275 | } | |
2276 | } | |
2277 | q->nr_hw_queues = i; | |
2278 | blk_mq_sysfs_register(q); | |
2279 | } | |
2280 | ||
2281 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2282 | struct request_queue *q) | |
2283 | { | |
66841672 ML |
2284 | /* mark the queue as mq asap */ |
2285 | q->mq_ops = set->ops; | |
2286 | ||
34dbad5d | 2287 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2288 | blk_mq_poll_stats_bkt, |
2289 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2290 | if (!q->poll_cb) |
2291 | goto err_exit; | |
2292 | ||
868f2f0b KB |
2293 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2294 | if (!q->queue_ctx) | |
c7de5726 | 2295 | goto err_exit; |
868f2f0b | 2296 | |
737f98cf ML |
2297 | /* init q->mq_kobj and sw queues' kobjects */ |
2298 | blk_mq_sysfs_init(q); | |
2299 | ||
868f2f0b KB |
2300 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2301 | GFP_KERNEL, set->numa_node); | |
2302 | if (!q->queue_hw_ctx) | |
2303 | goto err_percpu; | |
2304 | ||
bdd17e75 | 2305 | q->mq_map = set->mq_map; |
868f2f0b KB |
2306 | |
2307 | blk_mq_realloc_hw_ctxs(set, q); | |
2308 | if (!q->nr_hw_queues) | |
2309 | goto err_hctxs; | |
320ae51f | 2310 | |
287922eb | 2311 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2312 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2313 | |
2314 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2315 | |
94eddfbe | 2316 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2317 | |
05f1dd53 JA |
2318 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2319 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2320 | ||
1be036e9 CH |
2321 | q->sg_reserved_size = INT_MAX; |
2322 | ||
2849450a | 2323 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2324 | INIT_LIST_HEAD(&q->requeue_list); |
2325 | spin_lock_init(&q->requeue_lock); | |
2326 | ||
254d259d | 2327 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2328 | |
eba71768 JA |
2329 | /* |
2330 | * Do this after blk_queue_make_request() overrides it... | |
2331 | */ | |
2332 | q->nr_requests = set->queue_depth; | |
2333 | ||
64f1c21e JA |
2334 | /* |
2335 | * Default to classic polling | |
2336 | */ | |
2337 | q->poll_nsec = -1; | |
2338 | ||
24d2f903 CH |
2339 | if (set->ops->complete) |
2340 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2341 | |
24d2f903 | 2342 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2343 | |
eabe0659 | 2344 | get_online_cpus(); |
51d638b1 | 2345 | mutex_lock(&all_q_mutex); |
320ae51f | 2346 | |
4593fdbe | 2347 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2348 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2349 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2350 | |
eabe0659 | 2351 | mutex_unlock(&all_q_mutex); |
51d638b1 | 2352 | put_online_cpus(); |
4593fdbe | 2353 | |
d3484991 JA |
2354 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2355 | int ret; | |
2356 | ||
2357 | ret = blk_mq_sched_init(q); | |
2358 | if (ret) | |
2359 | return ERR_PTR(ret); | |
2360 | } | |
2361 | ||
320ae51f | 2362 | return q; |
18741986 | 2363 | |
320ae51f | 2364 | err_hctxs: |
868f2f0b | 2365 | kfree(q->queue_hw_ctx); |
320ae51f | 2366 | err_percpu: |
868f2f0b | 2367 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2368 | err_exit: |
2369 | q->mq_ops = NULL; | |
320ae51f JA |
2370 | return ERR_PTR(-ENOMEM); |
2371 | } | |
b62c21b7 | 2372 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2373 | |
2374 | void blk_mq_free_queue(struct request_queue *q) | |
2375 | { | |
624dbe47 | 2376 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2377 | |
0e626368 AM |
2378 | mutex_lock(&all_q_mutex); |
2379 | list_del_init(&q->all_q_node); | |
2380 | mutex_unlock(&all_q_mutex); | |
2381 | ||
0d2602ca JA |
2382 | blk_mq_del_queue_tag_set(q); |
2383 | ||
624dbe47 | 2384 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2385 | } |
320ae51f JA |
2386 | |
2387 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2388 | static void blk_mq_queue_reinit(struct request_queue *q, |
2389 | const struct cpumask *online_mask) | |
320ae51f | 2390 | { |
4ecd4fef | 2391 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2392 | |
9c1051aa | 2393 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2394 | blk_mq_sysfs_unregister(q); |
2395 | ||
320ae51f JA |
2396 | /* |
2397 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2398 | * we should change hctx numa_node according to new topology (this | |
2399 | * involves free and re-allocate memory, worthy doing?) | |
2400 | */ | |
2401 | ||
5778322e | 2402 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2403 | |
67aec14c | 2404 | blk_mq_sysfs_register(q); |
9c1051aa | 2405 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2406 | } |
2407 | ||
65d5291e SAS |
2408 | /* |
2409 | * New online cpumask which is going to be set in this hotplug event. | |
2410 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2411 | * one-by-one and dynamically allocating this could result in a failure. | |
2412 | */ | |
2413 | static struct cpumask cpuhp_online_new; | |
2414 | ||
2415 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2416 | { |
2417 | struct request_queue *q; | |
320ae51f JA |
2418 | |
2419 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2420 | /* |
2421 | * We need to freeze and reinit all existing queues. Freezing | |
2422 | * involves synchronous wait for an RCU grace period and doing it | |
2423 | * one by one may take a long time. Start freezing all queues in | |
2424 | * one swoop and then wait for the completions so that freezing can | |
2425 | * take place in parallel. | |
2426 | */ | |
2427 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1671d522 | 2428 | blk_freeze_queue_start(q); |
415d3dab | 2429 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2430 | blk_mq_freeze_queue_wait(q); |
2431 | ||
320ae51f | 2432 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2433 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2434 | |
2435 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2436 | blk_mq_unfreeze_queue(q); | |
2437 | ||
320ae51f | 2438 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2439 | } |
2440 | ||
2441 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2442 | { | |
97a32864 | 2443 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2444 | blk_mq_queue_reinit_work(); |
2445 | return 0; | |
2446 | } | |
2447 | ||
2448 | /* | |
2449 | * Before hotadded cpu starts handling requests, new mappings must be | |
2450 | * established. Otherwise, these requests in hw queue might never be | |
2451 | * dispatched. | |
2452 | * | |
2453 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2454 | * for CPU0, and ctx1 for CPU1). | |
2455 | * | |
2456 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2457 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2458 | * | |
2c3ad667 JA |
2459 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2460 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2461 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2462 | * ignored. | |
65d5291e SAS |
2463 | */ |
2464 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2465 | { | |
2466 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2467 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2468 | blk_mq_queue_reinit_work(); | |
2469 | return 0; | |
320ae51f JA |
2470 | } |
2471 | ||
a5164405 JA |
2472 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2473 | { | |
2474 | int i; | |
2475 | ||
cc71a6f4 JA |
2476 | for (i = 0; i < set->nr_hw_queues; i++) |
2477 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2478 | goto out_unwind; |
a5164405 JA |
2479 | |
2480 | return 0; | |
2481 | ||
2482 | out_unwind: | |
2483 | while (--i >= 0) | |
cc71a6f4 | 2484 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2485 | |
a5164405 JA |
2486 | return -ENOMEM; |
2487 | } | |
2488 | ||
2489 | /* | |
2490 | * Allocate the request maps associated with this tag_set. Note that this | |
2491 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2492 | * will be updated to reflect the allocated depth. | |
2493 | */ | |
2494 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2495 | { | |
2496 | unsigned int depth; | |
2497 | int err; | |
2498 | ||
2499 | depth = set->queue_depth; | |
2500 | do { | |
2501 | err = __blk_mq_alloc_rq_maps(set); | |
2502 | if (!err) | |
2503 | break; | |
2504 | ||
2505 | set->queue_depth >>= 1; | |
2506 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2507 | err = -ENOMEM; | |
2508 | break; | |
2509 | } | |
2510 | } while (set->queue_depth); | |
2511 | ||
2512 | if (!set->queue_depth || err) { | |
2513 | pr_err("blk-mq: failed to allocate request map\n"); | |
2514 | return -ENOMEM; | |
2515 | } | |
2516 | ||
2517 | if (depth != set->queue_depth) | |
2518 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2519 | depth, set->queue_depth); | |
2520 | ||
2521 | return 0; | |
2522 | } | |
2523 | ||
ebe8bddb OS |
2524 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2525 | { | |
2526 | if (set->ops->map_queues) | |
2527 | return set->ops->map_queues(set); | |
2528 | else | |
2529 | return blk_mq_map_queues(set); | |
2530 | } | |
2531 | ||
a4391c64 JA |
2532 | /* |
2533 | * Alloc a tag set to be associated with one or more request queues. | |
2534 | * May fail with EINVAL for various error conditions. May adjust the | |
2535 | * requested depth down, if if it too large. In that case, the set | |
2536 | * value will be stored in set->queue_depth. | |
2537 | */ | |
24d2f903 CH |
2538 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2539 | { | |
da695ba2 CH |
2540 | int ret; |
2541 | ||
205fb5f5 BVA |
2542 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2543 | ||
24d2f903 CH |
2544 | if (!set->nr_hw_queues) |
2545 | return -EINVAL; | |
a4391c64 | 2546 | if (!set->queue_depth) |
24d2f903 CH |
2547 | return -EINVAL; |
2548 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2549 | return -EINVAL; | |
2550 | ||
7d7e0f90 | 2551 | if (!set->ops->queue_rq) |
24d2f903 CH |
2552 | return -EINVAL; |
2553 | ||
a4391c64 JA |
2554 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2555 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2556 | BLK_MQ_MAX_DEPTH); | |
2557 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2558 | } | |
24d2f903 | 2559 | |
6637fadf SL |
2560 | /* |
2561 | * If a crashdump is active, then we are potentially in a very | |
2562 | * memory constrained environment. Limit us to 1 queue and | |
2563 | * 64 tags to prevent using too much memory. | |
2564 | */ | |
2565 | if (is_kdump_kernel()) { | |
2566 | set->nr_hw_queues = 1; | |
2567 | set->queue_depth = min(64U, set->queue_depth); | |
2568 | } | |
868f2f0b KB |
2569 | /* |
2570 | * There is no use for more h/w queues than cpus. | |
2571 | */ | |
2572 | if (set->nr_hw_queues > nr_cpu_ids) | |
2573 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2574 | |
868f2f0b | 2575 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2576 | GFP_KERNEL, set->numa_node); |
2577 | if (!set->tags) | |
a5164405 | 2578 | return -ENOMEM; |
24d2f903 | 2579 | |
da695ba2 CH |
2580 | ret = -ENOMEM; |
2581 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2582 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2583 | if (!set->mq_map) |
2584 | goto out_free_tags; | |
2585 | ||
ebe8bddb | 2586 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2587 | if (ret) |
2588 | goto out_free_mq_map; | |
2589 | ||
2590 | ret = blk_mq_alloc_rq_maps(set); | |
2591 | if (ret) | |
bdd17e75 | 2592 | goto out_free_mq_map; |
24d2f903 | 2593 | |
0d2602ca JA |
2594 | mutex_init(&set->tag_list_lock); |
2595 | INIT_LIST_HEAD(&set->tag_list); | |
2596 | ||
24d2f903 | 2597 | return 0; |
bdd17e75 CH |
2598 | |
2599 | out_free_mq_map: | |
2600 | kfree(set->mq_map); | |
2601 | set->mq_map = NULL; | |
2602 | out_free_tags: | |
5676e7b6 RE |
2603 | kfree(set->tags); |
2604 | set->tags = NULL; | |
da695ba2 | 2605 | return ret; |
24d2f903 CH |
2606 | } |
2607 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2608 | ||
2609 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2610 | { | |
2611 | int i; | |
2612 | ||
cc71a6f4 JA |
2613 | for (i = 0; i < nr_cpu_ids; i++) |
2614 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2615 | |
bdd17e75 CH |
2616 | kfree(set->mq_map); |
2617 | set->mq_map = NULL; | |
2618 | ||
981bd189 | 2619 | kfree(set->tags); |
5676e7b6 | 2620 | set->tags = NULL; |
24d2f903 CH |
2621 | } |
2622 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2623 | ||
e3a2b3f9 JA |
2624 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2625 | { | |
2626 | struct blk_mq_tag_set *set = q->tag_set; | |
2627 | struct blk_mq_hw_ctx *hctx; | |
2628 | int i, ret; | |
2629 | ||
bd166ef1 | 2630 | if (!set) |
e3a2b3f9 JA |
2631 | return -EINVAL; |
2632 | ||
70f36b60 | 2633 | blk_mq_freeze_queue(q); |
70f36b60 | 2634 | |
e3a2b3f9 JA |
2635 | ret = 0; |
2636 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2637 | if (!hctx->tags) |
2638 | continue; | |
bd166ef1 JA |
2639 | /* |
2640 | * If we're using an MQ scheduler, just update the scheduler | |
2641 | * queue depth. This is similar to what the old code would do. | |
2642 | */ | |
70f36b60 JA |
2643 | if (!hctx->sched_tags) { |
2644 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2645 | min(nr, set->queue_depth), | |
2646 | false); | |
2647 | } else { | |
2648 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2649 | nr, true); | |
2650 | } | |
e3a2b3f9 JA |
2651 | if (ret) |
2652 | break; | |
2653 | } | |
2654 | ||
2655 | if (!ret) | |
2656 | q->nr_requests = nr; | |
2657 | ||
70f36b60 | 2658 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2659 | |
e3a2b3f9 JA |
2660 | return ret; |
2661 | } | |
2662 | ||
868f2f0b KB |
2663 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2664 | { | |
2665 | struct request_queue *q; | |
2666 | ||
705cda97 BVA |
2667 | lockdep_assert_held(&set->tag_list_lock); |
2668 | ||
868f2f0b KB |
2669 | if (nr_hw_queues > nr_cpu_ids) |
2670 | nr_hw_queues = nr_cpu_ids; | |
2671 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2672 | return; | |
2673 | ||
2674 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2675 | blk_mq_freeze_queue(q); | |
2676 | ||
2677 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2678 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2679 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2680 | blk_mq_realloc_hw_ctxs(set, q); | |
868f2f0b KB |
2681 | blk_mq_queue_reinit(q, cpu_online_mask); |
2682 | } | |
2683 | ||
2684 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2685 | blk_mq_unfreeze_queue(q); | |
2686 | } | |
2687 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2688 | ||
34dbad5d OS |
2689 | /* Enable polling stats and return whether they were already enabled. */ |
2690 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2691 | { | |
2692 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2693 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2694 | return true; | |
2695 | blk_stat_add_callback(q, q->poll_cb); | |
2696 | return false; | |
2697 | } | |
2698 | ||
2699 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2700 | { | |
2701 | /* | |
2702 | * We don't arm the callback if polling stats are not enabled or the | |
2703 | * callback is already active. | |
2704 | */ | |
2705 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2706 | blk_stat_is_active(q->poll_cb)) | |
2707 | return; | |
2708 | ||
2709 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2710 | } | |
2711 | ||
2712 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2713 | { | |
2714 | struct request_queue *q = cb->data; | |
720b8ccc | 2715 | int bucket; |
34dbad5d | 2716 | |
720b8ccc SB |
2717 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2718 | if (cb->stat[bucket].nr_samples) | |
2719 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2720 | } | |
34dbad5d OS |
2721 | } |
2722 | ||
64f1c21e JA |
2723 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2724 | struct blk_mq_hw_ctx *hctx, | |
2725 | struct request *rq) | |
2726 | { | |
64f1c21e | 2727 | unsigned long ret = 0; |
720b8ccc | 2728 | int bucket; |
64f1c21e JA |
2729 | |
2730 | /* | |
2731 | * If stats collection isn't on, don't sleep but turn it on for | |
2732 | * future users | |
2733 | */ | |
34dbad5d | 2734 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2735 | return 0; |
2736 | ||
64f1c21e JA |
2737 | /* |
2738 | * As an optimistic guess, use half of the mean service time | |
2739 | * for this type of request. We can (and should) make this smarter. | |
2740 | * For instance, if the completion latencies are tight, we can | |
2741 | * get closer than just half the mean. This is especially | |
2742 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2743 | * than ~10 usec. We do use the stats for the relevant IO size |
2744 | * if available which does lead to better estimates. | |
64f1c21e | 2745 | */ |
720b8ccc SB |
2746 | bucket = blk_mq_poll_stats_bkt(rq); |
2747 | if (bucket < 0) | |
2748 | return ret; | |
2749 | ||
2750 | if (q->poll_stat[bucket].nr_samples) | |
2751 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2752 | |
2753 | return ret; | |
2754 | } | |
2755 | ||
06426adf | 2756 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2757 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2758 | struct request *rq) |
2759 | { | |
2760 | struct hrtimer_sleeper hs; | |
2761 | enum hrtimer_mode mode; | |
64f1c21e | 2762 | unsigned int nsecs; |
06426adf JA |
2763 | ktime_t kt; |
2764 | ||
64f1c21e JA |
2765 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2766 | return false; | |
2767 | ||
2768 | /* | |
2769 | * poll_nsec can be: | |
2770 | * | |
2771 | * -1: don't ever hybrid sleep | |
2772 | * 0: use half of prev avg | |
2773 | * >0: use this specific value | |
2774 | */ | |
2775 | if (q->poll_nsec == -1) | |
2776 | return false; | |
2777 | else if (q->poll_nsec > 0) | |
2778 | nsecs = q->poll_nsec; | |
2779 | else | |
2780 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2781 | ||
2782 | if (!nsecs) | |
06426adf JA |
2783 | return false; |
2784 | ||
2785 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2786 | ||
2787 | /* | |
2788 | * This will be replaced with the stats tracking code, using | |
2789 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2790 | */ | |
8b0e1953 | 2791 | kt = nsecs; |
06426adf JA |
2792 | |
2793 | mode = HRTIMER_MODE_REL; | |
2794 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2795 | hrtimer_set_expires(&hs.timer, kt); | |
2796 | ||
2797 | hrtimer_init_sleeper(&hs, current); | |
2798 | do { | |
2799 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2800 | break; | |
2801 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2802 | hrtimer_start_expires(&hs.timer, mode); | |
2803 | if (hs.task) | |
2804 | io_schedule(); | |
2805 | hrtimer_cancel(&hs.timer); | |
2806 | mode = HRTIMER_MODE_ABS; | |
2807 | } while (hs.task && !signal_pending(current)); | |
2808 | ||
2809 | __set_current_state(TASK_RUNNING); | |
2810 | destroy_hrtimer_on_stack(&hs.timer); | |
2811 | return true; | |
2812 | } | |
2813 | ||
bbd7bb70 JA |
2814 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2815 | { | |
2816 | struct request_queue *q = hctx->queue; | |
2817 | long state; | |
2818 | ||
06426adf JA |
2819 | /* |
2820 | * If we sleep, have the caller restart the poll loop to reset | |
2821 | * the state. Like for the other success return cases, the | |
2822 | * caller is responsible for checking if the IO completed. If | |
2823 | * the IO isn't complete, we'll get called again and will go | |
2824 | * straight to the busy poll loop. | |
2825 | */ | |
64f1c21e | 2826 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2827 | return true; |
2828 | ||
bbd7bb70 JA |
2829 | hctx->poll_considered++; |
2830 | ||
2831 | state = current->state; | |
2832 | while (!need_resched()) { | |
2833 | int ret; | |
2834 | ||
2835 | hctx->poll_invoked++; | |
2836 | ||
2837 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2838 | if (ret > 0) { | |
2839 | hctx->poll_success++; | |
2840 | set_current_state(TASK_RUNNING); | |
2841 | return true; | |
2842 | } | |
2843 | ||
2844 | if (signal_pending_state(state, current)) | |
2845 | set_current_state(TASK_RUNNING); | |
2846 | ||
2847 | if (current->state == TASK_RUNNING) | |
2848 | return true; | |
2849 | if (ret < 0) | |
2850 | break; | |
2851 | cpu_relax(); | |
2852 | } | |
2853 | ||
2854 | return false; | |
2855 | } | |
2856 | ||
2857 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2858 | { | |
2859 | struct blk_mq_hw_ctx *hctx; | |
2860 | struct blk_plug *plug; | |
2861 | struct request *rq; | |
2862 | ||
2863 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2864 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2865 | return false; | |
2866 | ||
2867 | plug = current->plug; | |
2868 | if (plug) | |
2869 | blk_flush_plug_list(plug, false); | |
2870 | ||
2871 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2872 | if (!blk_qc_t_is_internal(cookie)) |
2873 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2874 | else { |
bd166ef1 | 2875 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
2876 | /* |
2877 | * With scheduling, if the request has completed, we'll | |
2878 | * get a NULL return here, as we clear the sched tag when | |
2879 | * that happens. The request still remains valid, like always, | |
2880 | * so we should be safe with just the NULL check. | |
2881 | */ | |
2882 | if (!rq) | |
2883 | return false; | |
2884 | } | |
bbd7bb70 JA |
2885 | |
2886 | return __blk_mq_poll(hctx, rq); | |
2887 | } | |
2888 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2889 | ||
676141e4 JA |
2890 | void blk_mq_disable_hotplug(void) |
2891 | { | |
2892 | mutex_lock(&all_q_mutex); | |
2893 | } | |
2894 | ||
2895 | void blk_mq_enable_hotplug(void) | |
2896 | { | |
2897 | mutex_unlock(&all_q_mutex); | |
2898 | } | |
2899 | ||
320ae51f JA |
2900 | static int __init blk_mq_init(void) |
2901 | { | |
9467f859 TG |
2902 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2903 | blk_mq_hctx_notify_dead); | |
320ae51f | 2904 | |
65d5291e SAS |
2905 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2906 | blk_mq_queue_reinit_prepare, | |
2907 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2908 | return 0; |
2909 | } | |
2910 | subsys_initcall(blk_mq_init); |