]>
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> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
33 | ||
34 | static DEFINE_MUTEX(all_q_mutex); | |
35 | static LIST_HEAD(all_q_list); | |
36 | ||
37 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
38 | ||
320ae51f JA |
39 | /* |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
44 | unsigned int i; | |
45 | ||
569fd0ce | 46 | for (i = 0; i < hctx->ctx_map.size; i++) |
1429d7c9 | 47 | if (hctx->ctx_map.map[i].word) |
320ae51f JA |
48 | return true; |
49 | ||
50 | return false; | |
51 | } | |
52 | ||
1429d7c9 JA |
53 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
54 | struct blk_mq_ctx *ctx) | |
55 | { | |
56 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
57 | } | |
58 | ||
59 | #define CTX_TO_BIT(hctx, ctx) \ | |
60 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
61 | ||
320ae51f JA |
62 | /* |
63 | * Mark this ctx as having pending work in this hardware queue | |
64 | */ | |
65 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
66 | struct blk_mq_ctx *ctx) | |
67 | { | |
1429d7c9 JA |
68 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
69 | ||
70 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
71 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
72 | } | |
73 | ||
74 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
75 | struct blk_mq_ctx *ctx) | |
76 | { | |
77 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
78 | ||
79 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
80 | } |
81 | ||
b4c6a028 | 82 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 83 | { |
4ecd4fef | 84 | int freeze_depth; |
cddd5d17 | 85 | |
4ecd4fef CH |
86 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
87 | if (freeze_depth == 1) { | |
3ef28e83 | 88 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 89 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 90 | } |
f3af020b | 91 | } |
b4c6a028 | 92 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
93 | |
94 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
95 | { | |
3ef28e83 | 96 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
97 | } |
98 | ||
f3af020b TH |
99 | /* |
100 | * Guarantee no request is in use, so we can change any data structure of | |
101 | * the queue afterward. | |
102 | */ | |
3ef28e83 | 103 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 104 | { |
3ef28e83 DW |
105 | /* |
106 | * In the !blk_mq case we are only calling this to kill the | |
107 | * q_usage_counter, otherwise this increases the freeze depth | |
108 | * and waits for it to return to zero. For this reason there is | |
109 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
110 | * exported to drivers as the only user for unfreeze is blk_mq. | |
111 | */ | |
f3af020b TH |
112 | blk_mq_freeze_queue_start(q); |
113 | blk_mq_freeze_queue_wait(q); | |
114 | } | |
3ef28e83 DW |
115 | |
116 | void blk_mq_freeze_queue(struct request_queue *q) | |
117 | { | |
118 | /* | |
119 | * ...just an alias to keep freeze and unfreeze actions balanced | |
120 | * in the blk_mq_* namespace | |
121 | */ | |
122 | blk_freeze_queue(q); | |
123 | } | |
c761d96b | 124 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 125 | |
b4c6a028 | 126 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 127 | { |
4ecd4fef | 128 | int freeze_depth; |
320ae51f | 129 | |
4ecd4fef CH |
130 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
131 | WARN_ON_ONCE(freeze_depth < 0); | |
132 | if (!freeze_depth) { | |
3ef28e83 | 133 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 134 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 135 | } |
320ae51f | 136 | } |
b4c6a028 | 137 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 138 | |
aed3ea94 JA |
139 | void blk_mq_wake_waiters(struct request_queue *q) |
140 | { | |
141 | struct blk_mq_hw_ctx *hctx; | |
142 | unsigned int i; | |
143 | ||
144 | queue_for_each_hw_ctx(q, hctx, i) | |
145 | if (blk_mq_hw_queue_mapped(hctx)) | |
146 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
147 | |
148 | /* | |
149 | * If we are called because the queue has now been marked as | |
150 | * dying, we need to ensure that processes currently waiting on | |
151 | * the queue are notified as well. | |
152 | */ | |
153 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
154 | } |
155 | ||
320ae51f JA |
156 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
157 | { | |
158 | return blk_mq_has_free_tags(hctx->tags); | |
159 | } | |
160 | EXPORT_SYMBOL(blk_mq_can_queue); | |
161 | ||
94eddfbe | 162 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
cc6e3b10 MC |
163 | struct request *rq, int op, |
164 | unsigned int op_flags) | |
320ae51f | 165 | { |
94eddfbe | 166 | if (blk_queue_io_stat(q)) |
cc6e3b10 | 167 | op_flags |= REQ_IO_STAT; |
94eddfbe | 168 | |
af76e555 CH |
169 | INIT_LIST_HEAD(&rq->queuelist); |
170 | /* csd/requeue_work/fifo_time is initialized before use */ | |
171 | rq->q = q; | |
320ae51f | 172 | rq->mq_ctx = ctx; |
cc6e3b10 | 173 | req_set_op_attrs(rq, op, op_flags); |
af76e555 CH |
174 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
175 | rq->cpu = -1; | |
af76e555 CH |
176 | INIT_HLIST_NODE(&rq->hash); |
177 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
178 | rq->rq_disk = NULL; |
179 | rq->part = NULL; | |
3ee32372 | 180 | rq->start_time = jiffies; |
af76e555 CH |
181 | #ifdef CONFIG_BLK_CGROUP |
182 | rq->rl = NULL; | |
0fec08b4 | 183 | set_start_time_ns(rq); |
af76e555 CH |
184 | rq->io_start_time_ns = 0; |
185 | #endif | |
186 | rq->nr_phys_segments = 0; | |
187 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
188 | rq->nr_integrity_segments = 0; | |
189 | #endif | |
af76e555 CH |
190 | rq->special = NULL; |
191 | /* tag was already set */ | |
192 | rq->errors = 0; | |
af76e555 | 193 | |
6f4a1626 TB |
194 | rq->cmd = rq->__cmd; |
195 | ||
af76e555 CH |
196 | rq->extra_len = 0; |
197 | rq->sense_len = 0; | |
198 | rq->resid_len = 0; | |
199 | rq->sense = NULL; | |
200 | ||
af76e555 | 201 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
202 | rq->timeout = 0; |
203 | ||
af76e555 CH |
204 | rq->end_io = NULL; |
205 | rq->end_io_data = NULL; | |
206 | rq->next_rq = NULL; | |
207 | ||
d9d8c5c4 | 208 | ctx->rq_dispatched[rw_is_sync(op, op_flags)]++; |
320ae51f JA |
209 | } |
210 | ||
5dee8577 | 211 | static struct request * |
cc6e3b10 | 212 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags) |
5dee8577 CH |
213 | { |
214 | struct request *rq; | |
215 | unsigned int tag; | |
216 | ||
cb96a42c | 217 | tag = blk_mq_get_tag(data); |
5dee8577 | 218 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 219 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 220 | |
cb96a42c | 221 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 222 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 223 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
224 | } |
225 | ||
226 | rq->tag = tag; | |
cc6e3b10 | 227 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags); |
5dee8577 CH |
228 | return rq; |
229 | } | |
230 | ||
231 | return NULL; | |
232 | } | |
233 | ||
6f3b0e8b CH |
234 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
235 | unsigned int flags) | |
320ae51f | 236 | { |
d852564f CH |
237 | struct blk_mq_ctx *ctx; |
238 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 239 | struct request *rq; |
cb96a42c | 240 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 241 | int ret; |
320ae51f | 242 | |
6f3b0e8b | 243 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
244 | if (ret) |
245 | return ERR_PTR(ret); | |
320ae51f | 246 | |
d852564f | 247 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 248 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 249 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
d852564f | 250 | |
cc6e3b10 | 251 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
6f3b0e8b | 252 | if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) { |
d852564f CH |
253 | __blk_mq_run_hw_queue(hctx); |
254 | blk_mq_put_ctx(ctx); | |
255 | ||
256 | ctx = blk_mq_get_ctx(q); | |
7d7e0f90 | 257 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 258 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
cc6e3b10 | 259 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
cb96a42c | 260 | ctx = alloc_data.ctx; |
d852564f CH |
261 | } |
262 | blk_mq_put_ctx(ctx); | |
c76541a9 | 263 | if (!rq) { |
3ef28e83 | 264 | blk_queue_exit(q); |
a492f075 | 265 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 266 | } |
0c4de0f3 CH |
267 | |
268 | rq->__data_len = 0; | |
269 | rq->__sector = (sector_t) -1; | |
270 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
271 | return rq; |
272 | } | |
4bb659b1 | 273 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 274 | |
1f5bd336 ML |
275 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
276 | unsigned int flags, unsigned int hctx_idx) | |
277 | { | |
278 | struct blk_mq_hw_ctx *hctx; | |
279 | struct blk_mq_ctx *ctx; | |
280 | struct request *rq; | |
281 | struct blk_mq_alloc_data alloc_data; | |
282 | int ret; | |
283 | ||
284 | /* | |
285 | * If the tag allocator sleeps we could get an allocation for a | |
286 | * different hardware context. No need to complicate the low level | |
287 | * allocator for this for the rare use case of a command tied to | |
288 | * a specific queue. | |
289 | */ | |
290 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
291 | return ERR_PTR(-EINVAL); | |
292 | ||
293 | if (hctx_idx >= q->nr_hw_queues) | |
294 | return ERR_PTR(-EIO); | |
295 | ||
296 | ret = blk_queue_enter(q, true); | |
297 | if (ret) | |
298 | return ERR_PTR(ret); | |
299 | ||
300 | hctx = q->queue_hw_ctx[hctx_idx]; | |
301 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); | |
302 | ||
303 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
304 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); | |
305 | if (!rq) { | |
306 | blk_queue_exit(q); | |
307 | return ERR_PTR(-EWOULDBLOCK); | |
308 | } | |
309 | ||
310 | return rq; | |
311 | } | |
312 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
313 | ||
320ae51f JA |
314 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
315 | struct blk_mq_ctx *ctx, struct request *rq) | |
316 | { | |
317 | const int tag = rq->tag; | |
318 | struct request_queue *q = rq->q; | |
319 | ||
0d2602ca JA |
320 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
321 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 322 | rq->cmd_flags = 0; |
0d2602ca | 323 | |
af76e555 | 324 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 325 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
3ef28e83 | 326 | blk_queue_exit(q); |
320ae51f JA |
327 | } |
328 | ||
7c7f2f2b | 329 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
330 | { |
331 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
332 | |
333 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 334 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
335 | |
336 | } | |
337 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
338 | ||
339 | void blk_mq_free_request(struct request *rq) | |
340 | { | |
7d7e0f90 | 341 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 342 | } |
1a3b595a | 343 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 344 | |
c8a446ad | 345 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 346 | { |
0d11e6ac ML |
347 | blk_account_io_done(rq); |
348 | ||
91b63639 | 349 | if (rq->end_io) { |
320ae51f | 350 | rq->end_io(rq, error); |
91b63639 CH |
351 | } else { |
352 | if (unlikely(blk_bidi_rq(rq))) | |
353 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 354 | blk_mq_free_request(rq); |
91b63639 | 355 | } |
320ae51f | 356 | } |
c8a446ad | 357 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 358 | |
c8a446ad | 359 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
360 | { |
361 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
362 | BUG(); | |
c8a446ad | 363 | __blk_mq_end_request(rq, error); |
63151a44 | 364 | } |
c8a446ad | 365 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 366 | |
30a91cb4 | 367 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 368 | { |
3d6efbf6 | 369 | struct request *rq = data; |
320ae51f | 370 | |
30a91cb4 | 371 | rq->q->softirq_done_fn(rq); |
320ae51f | 372 | } |
320ae51f | 373 | |
ed851860 | 374 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
375 | { |
376 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 377 | bool shared = false; |
320ae51f JA |
378 | int cpu; |
379 | ||
38535201 | 380 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
381 | rq->q->softirq_done_fn(rq); |
382 | return; | |
383 | } | |
320ae51f JA |
384 | |
385 | cpu = get_cpu(); | |
38535201 CH |
386 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
387 | shared = cpus_share_cache(cpu, ctx->cpu); | |
388 | ||
389 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 390 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
391 | rq->csd.info = rq; |
392 | rq->csd.flags = 0; | |
c46fff2a | 393 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 394 | } else { |
30a91cb4 | 395 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 396 | } |
320ae51f JA |
397 | put_cpu(); |
398 | } | |
30a91cb4 | 399 | |
1fa8cc52 | 400 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
401 | { |
402 | struct request_queue *q = rq->q; | |
403 | ||
404 | if (!q->softirq_done_fn) | |
c8a446ad | 405 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
406 | else |
407 | blk_mq_ipi_complete_request(rq); | |
408 | } | |
409 | ||
30a91cb4 CH |
410 | /** |
411 | * blk_mq_complete_request - end I/O on a request | |
412 | * @rq: the request being processed | |
413 | * | |
414 | * Description: | |
415 | * Ends all I/O on a request. It does not handle partial completions. | |
416 | * The actual completion happens out-of-order, through a IPI handler. | |
417 | **/ | |
f4829a9b | 418 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 419 | { |
95f09684 JA |
420 | struct request_queue *q = rq->q; |
421 | ||
422 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 423 | return; |
f4829a9b CH |
424 | if (!blk_mark_rq_complete(rq)) { |
425 | rq->errors = error; | |
ed851860 | 426 | __blk_mq_complete_request(rq); |
f4829a9b | 427 | } |
30a91cb4 CH |
428 | } |
429 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 430 | |
973c0191 KB |
431 | int blk_mq_request_started(struct request *rq) |
432 | { | |
433 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
434 | } | |
435 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
436 | ||
e2490073 | 437 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
438 | { |
439 | struct request_queue *q = rq->q; | |
440 | ||
441 | trace_block_rq_issue(q, rq); | |
442 | ||
742ee69b | 443 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
444 | if (unlikely(blk_bidi_rq(rq))) |
445 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 446 | |
2b8393b4 | 447 | blk_add_timer(rq); |
87ee7b11 | 448 | |
538b7534 JA |
449 | /* |
450 | * Ensure that ->deadline is visible before set the started | |
451 | * flag and clear the completed flag. | |
452 | */ | |
453 | smp_mb__before_atomic(); | |
454 | ||
87ee7b11 JA |
455 | /* |
456 | * Mark us as started and clear complete. Complete might have been | |
457 | * set if requeue raced with timeout, which then marked it as | |
458 | * complete. So be sure to clear complete again when we start | |
459 | * the request, otherwise we'll ignore the completion event. | |
460 | */ | |
4b570521 JA |
461 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
462 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
463 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
464 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
465 | |
466 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
467 | /* | |
468 | * Make sure space for the drain appears. We know we can do | |
469 | * this because max_hw_segments has been adjusted to be one | |
470 | * fewer than the device can handle. | |
471 | */ | |
472 | rq->nr_phys_segments++; | |
473 | } | |
320ae51f | 474 | } |
e2490073 | 475 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 476 | |
ed0791b2 | 477 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
478 | { |
479 | struct request_queue *q = rq->q; | |
480 | ||
481 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 482 | |
e2490073 CH |
483 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
484 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
485 | rq->nr_phys_segments--; | |
486 | } | |
320ae51f JA |
487 | } |
488 | ||
ed0791b2 CH |
489 | void blk_mq_requeue_request(struct request *rq) |
490 | { | |
ed0791b2 | 491 | __blk_mq_requeue_request(rq); |
ed0791b2 | 492 | |
ed0791b2 | 493 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 494 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
495 | } |
496 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
497 | ||
6fca6a61 CH |
498 | static void blk_mq_requeue_work(struct work_struct *work) |
499 | { | |
500 | struct request_queue *q = | |
2849450a | 501 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
502 | LIST_HEAD(rq_list); |
503 | struct request *rq, *next; | |
504 | unsigned long flags; | |
505 | ||
506 | spin_lock_irqsave(&q->requeue_lock, flags); | |
507 | list_splice_init(&q->requeue_list, &rq_list); | |
508 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
509 | ||
510 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
511 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
512 | continue; | |
513 | ||
514 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
515 | list_del_init(&rq->queuelist); | |
516 | blk_mq_insert_request(rq, true, false, false); | |
517 | } | |
518 | ||
519 | while (!list_empty(&rq_list)) { | |
520 | rq = list_entry(rq_list.next, struct request, queuelist); | |
521 | list_del_init(&rq->queuelist); | |
522 | blk_mq_insert_request(rq, false, false, false); | |
523 | } | |
524 | ||
8b957415 JA |
525 | /* |
526 | * Use the start variant of queue running here, so that running | |
527 | * the requeue work will kick stopped queues. | |
528 | */ | |
529 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
530 | } |
531 | ||
532 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
533 | { | |
534 | struct request_queue *q = rq->q; | |
535 | unsigned long flags; | |
536 | ||
537 | /* | |
538 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
539 | * request head insertation from the workqueue. | |
540 | */ | |
541 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
542 | ||
543 | spin_lock_irqsave(&q->requeue_lock, flags); | |
544 | if (at_head) { | |
545 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
546 | list_add(&rq->queuelist, &q->requeue_list); | |
547 | } else { | |
548 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
549 | } | |
550 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
551 | } | |
552 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
553 | ||
c68ed59f KB |
554 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
555 | { | |
2849450a | 556 | cancel_delayed_work_sync(&q->requeue_work); |
c68ed59f KB |
557 | } |
558 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
559 | ||
6fca6a61 CH |
560 | void blk_mq_kick_requeue_list(struct request_queue *q) |
561 | { | |
2849450a | 562 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
563 | } |
564 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
565 | ||
2849450a MS |
566 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
567 | unsigned long msecs) | |
568 | { | |
569 | kblockd_schedule_delayed_work(&q->requeue_work, | |
570 | msecs_to_jiffies(msecs)); | |
571 | } | |
572 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
573 | ||
1885b24d JA |
574 | void blk_mq_abort_requeue_list(struct request_queue *q) |
575 | { | |
576 | unsigned long flags; | |
577 | LIST_HEAD(rq_list); | |
578 | ||
579 | spin_lock_irqsave(&q->requeue_lock, flags); | |
580 | list_splice_init(&q->requeue_list, &rq_list); | |
581 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
582 | ||
583 | while (!list_empty(&rq_list)) { | |
584 | struct request *rq; | |
585 | ||
586 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
587 | list_del_init(&rq->queuelist); | |
588 | rq->errors = -EIO; | |
589 | blk_mq_end_request(rq, rq->errors); | |
590 | } | |
591 | } | |
592 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
593 | ||
0e62f51f JA |
594 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
595 | { | |
88c7b2b7 JA |
596 | if (tag < tags->nr_tags) { |
597 | prefetch(tags->rqs[tag]); | |
4ee86bab | 598 | return tags->rqs[tag]; |
88c7b2b7 | 599 | } |
4ee86bab HR |
600 | |
601 | return NULL; | |
24d2f903 CH |
602 | } |
603 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
604 | ||
320ae51f | 605 | struct blk_mq_timeout_data { |
46f92d42 CH |
606 | unsigned long next; |
607 | unsigned int next_set; | |
320ae51f JA |
608 | }; |
609 | ||
90415837 | 610 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 611 | { |
46f92d42 CH |
612 | struct blk_mq_ops *ops = req->q->mq_ops; |
613 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
614 | |
615 | /* | |
616 | * We know that complete is set at this point. If STARTED isn't set | |
617 | * anymore, then the request isn't active and the "timeout" should | |
618 | * just be ignored. This can happen due to the bitflag ordering. | |
619 | * Timeout first checks if STARTED is set, and if it is, assumes | |
620 | * the request is active. But if we race with completion, then | |
621 | * we both flags will get cleared. So check here again, and ignore | |
622 | * a timeout event with a request that isn't active. | |
623 | */ | |
46f92d42 CH |
624 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
625 | return; | |
87ee7b11 | 626 | |
46f92d42 | 627 | if (ops->timeout) |
0152fb6b | 628 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
629 | |
630 | switch (ret) { | |
631 | case BLK_EH_HANDLED: | |
632 | __blk_mq_complete_request(req); | |
633 | break; | |
634 | case BLK_EH_RESET_TIMER: | |
635 | blk_add_timer(req); | |
636 | blk_clear_rq_complete(req); | |
637 | break; | |
638 | case BLK_EH_NOT_HANDLED: | |
639 | break; | |
640 | default: | |
641 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
642 | break; | |
643 | } | |
87ee7b11 | 644 | } |
5b3f25fc | 645 | |
81481eb4 CH |
646 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
647 | struct request *rq, void *priv, bool reserved) | |
648 | { | |
649 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 650 | |
eb130dbf KB |
651 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
652 | /* | |
653 | * If a request wasn't started before the queue was | |
654 | * marked dying, kill it here or it'll go unnoticed. | |
655 | */ | |
a59e0f57 KB |
656 | if (unlikely(blk_queue_dying(rq->q))) { |
657 | rq->errors = -EIO; | |
658 | blk_mq_end_request(rq, rq->errors); | |
659 | } | |
46f92d42 | 660 | return; |
eb130dbf | 661 | } |
87ee7b11 | 662 | |
46f92d42 CH |
663 | if (time_after_eq(jiffies, rq->deadline)) { |
664 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 665 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
666 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
667 | data->next = rq->deadline; | |
668 | data->next_set = 1; | |
669 | } | |
87ee7b11 JA |
670 | } |
671 | ||
287922eb | 672 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 673 | { |
287922eb CH |
674 | struct request_queue *q = |
675 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
676 | struct blk_mq_timeout_data data = { |
677 | .next = 0, | |
678 | .next_set = 0, | |
679 | }; | |
81481eb4 | 680 | int i; |
320ae51f | 681 | |
71f79fb3 GKB |
682 | /* A deadlock might occur if a request is stuck requiring a |
683 | * timeout at the same time a queue freeze is waiting | |
684 | * completion, since the timeout code would not be able to | |
685 | * acquire the queue reference here. | |
686 | * | |
687 | * That's why we don't use blk_queue_enter here; instead, we use | |
688 | * percpu_ref_tryget directly, because we need to be able to | |
689 | * obtain a reference even in the short window between the queue | |
690 | * starting to freeze, by dropping the first reference in | |
691 | * blk_mq_freeze_queue_start, and the moment the last request is | |
692 | * consumed, marked by the instant q_usage_counter reaches | |
693 | * zero. | |
694 | */ | |
695 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
696 | return; |
697 | ||
0bf6cd5b | 698 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 699 | |
81481eb4 CH |
700 | if (data.next_set) { |
701 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
702 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 703 | } else { |
0bf6cd5b CH |
704 | struct blk_mq_hw_ctx *hctx; |
705 | ||
f054b56c ML |
706 | queue_for_each_hw_ctx(q, hctx, i) { |
707 | /* the hctx may be unmapped, so check it here */ | |
708 | if (blk_mq_hw_queue_mapped(hctx)) | |
709 | blk_mq_tag_idle(hctx); | |
710 | } | |
0d2602ca | 711 | } |
287922eb | 712 | blk_queue_exit(q); |
320ae51f JA |
713 | } |
714 | ||
715 | /* | |
716 | * Reverse check our software queue for entries that we could potentially | |
717 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
718 | * too much time checking for merges. | |
719 | */ | |
720 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
721 | struct blk_mq_ctx *ctx, struct bio *bio) | |
722 | { | |
723 | struct request *rq; | |
724 | int checked = 8; | |
725 | ||
726 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
727 | int el_ret; | |
728 | ||
729 | if (!checked--) | |
730 | break; | |
731 | ||
732 | if (!blk_rq_merge_ok(rq, bio)) | |
733 | continue; | |
734 | ||
735 | el_ret = blk_try_merge(rq, bio); | |
736 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
737 | if (bio_attempt_back_merge(q, rq, bio)) { | |
738 | ctx->rq_merged++; | |
739 | return true; | |
740 | } | |
741 | break; | |
742 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
743 | if (bio_attempt_front_merge(q, rq, bio)) { | |
744 | ctx->rq_merged++; | |
745 | return true; | |
746 | } | |
747 | break; | |
748 | } | |
749 | } | |
750 | ||
751 | return false; | |
752 | } | |
753 | ||
1429d7c9 JA |
754 | /* |
755 | * Process software queues that have been marked busy, splicing them | |
756 | * to the for-dispatch | |
757 | */ | |
758 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
759 | { | |
760 | struct blk_mq_ctx *ctx; | |
761 | int i; | |
762 | ||
569fd0ce | 763 | for (i = 0; i < hctx->ctx_map.size; i++) { |
1429d7c9 JA |
764 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
765 | unsigned int off, bit; | |
766 | ||
767 | if (!bm->word) | |
768 | continue; | |
769 | ||
770 | bit = 0; | |
771 | off = i * hctx->ctx_map.bits_per_word; | |
772 | do { | |
773 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
774 | if (bit >= bm->depth) | |
775 | break; | |
776 | ||
777 | ctx = hctx->ctxs[bit + off]; | |
778 | clear_bit(bit, &bm->word); | |
779 | spin_lock(&ctx->lock); | |
780 | list_splice_tail_init(&ctx->rq_list, list); | |
781 | spin_unlock(&ctx->lock); | |
782 | ||
783 | bit++; | |
784 | } while (1); | |
785 | } | |
786 | } | |
787 | ||
320ae51f JA |
788 | /* |
789 | * Run this hardware queue, pulling any software queues mapped to it in. | |
790 | * Note that this function currently has various problems around ordering | |
791 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
792 | * items on the hctx->dispatch list. Ignore that for now. | |
793 | */ | |
794 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
795 | { | |
796 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
797 | struct request *rq; |
798 | LIST_HEAD(rq_list); | |
74c45052 JA |
799 | LIST_HEAD(driver_list); |
800 | struct list_head *dptr; | |
1429d7c9 | 801 | int queued; |
320ae51f | 802 | |
5d12f905 | 803 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
804 | return; |
805 | ||
0e87e58b JA |
806 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
807 | cpu_online(hctx->next_cpu)); | |
808 | ||
320ae51f JA |
809 | hctx->run++; |
810 | ||
811 | /* | |
812 | * Touch any software queue that has pending entries. | |
813 | */ | |
1429d7c9 | 814 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
815 | |
816 | /* | |
817 | * If we have previous entries on our dispatch list, grab them | |
818 | * and stuff them at the front for more fair dispatch. | |
819 | */ | |
820 | if (!list_empty_careful(&hctx->dispatch)) { | |
821 | spin_lock(&hctx->lock); | |
822 | if (!list_empty(&hctx->dispatch)) | |
823 | list_splice_init(&hctx->dispatch, &rq_list); | |
824 | spin_unlock(&hctx->lock); | |
825 | } | |
826 | ||
74c45052 JA |
827 | /* |
828 | * Start off with dptr being NULL, so we start the first request | |
829 | * immediately, even if we have more pending. | |
830 | */ | |
831 | dptr = NULL; | |
832 | ||
320ae51f JA |
833 | /* |
834 | * Now process all the entries, sending them to the driver. | |
835 | */ | |
1429d7c9 | 836 | queued = 0; |
320ae51f | 837 | while (!list_empty(&rq_list)) { |
74c45052 | 838 | struct blk_mq_queue_data bd; |
320ae51f JA |
839 | int ret; |
840 | ||
841 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
842 | list_del_init(&rq->queuelist); | |
320ae51f | 843 | |
74c45052 JA |
844 | bd.rq = rq; |
845 | bd.list = dptr; | |
846 | bd.last = list_empty(&rq_list); | |
847 | ||
848 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
849 | switch (ret) { |
850 | case BLK_MQ_RQ_QUEUE_OK: | |
851 | queued++; | |
52b9c330 | 852 | break; |
320ae51f | 853 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 854 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 855 | __blk_mq_requeue_request(rq); |
320ae51f JA |
856 | break; |
857 | default: | |
858 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 859 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 860 | rq->errors = -EIO; |
c8a446ad | 861 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
862 | break; |
863 | } | |
864 | ||
865 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
866 | break; | |
74c45052 JA |
867 | |
868 | /* | |
869 | * We've done the first request. If we have more than 1 | |
870 | * left in the list, set dptr to defer issue. | |
871 | */ | |
872 | if (!dptr && rq_list.next != rq_list.prev) | |
873 | dptr = &driver_list; | |
320ae51f JA |
874 | } |
875 | ||
876 | if (!queued) | |
877 | hctx->dispatched[0]++; | |
878 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
879 | hctx->dispatched[ilog2(queued) + 1]++; | |
880 | ||
881 | /* | |
882 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
883 | * that is where we will continue on next queue run. | |
884 | */ | |
885 | if (!list_empty(&rq_list)) { | |
886 | spin_lock(&hctx->lock); | |
887 | list_splice(&rq_list, &hctx->dispatch); | |
888 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
889 | /* |
890 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
891 | * it's possible the queue is stopped and restarted again | |
892 | * before this. Queue restart will dispatch requests. And since | |
893 | * requests in rq_list aren't added into hctx->dispatch yet, | |
894 | * the requests in rq_list might get lost. | |
895 | * | |
896 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
897 | **/ | |
898 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
899 | } |
900 | } | |
901 | ||
506e931f JA |
902 | /* |
903 | * It'd be great if the workqueue API had a way to pass | |
904 | * in a mask and had some smarts for more clever placement. | |
905 | * For now we just round-robin here, switching for every | |
906 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
907 | */ | |
908 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
909 | { | |
b657d7e6 CH |
910 | if (hctx->queue->nr_hw_queues == 1) |
911 | return WORK_CPU_UNBOUND; | |
506e931f JA |
912 | |
913 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 914 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
915 | |
916 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
917 | if (next_cpu >= nr_cpu_ids) | |
918 | next_cpu = cpumask_first(hctx->cpumask); | |
919 | ||
920 | hctx->next_cpu = next_cpu; | |
921 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
922 | |
923 | return cpu; | |
506e931f JA |
924 | } |
925 | ||
b657d7e6 | 926 | return hctx->next_cpu; |
506e931f JA |
927 | } |
928 | ||
320ae51f JA |
929 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
930 | { | |
19c66e59 ML |
931 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
932 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
933 | return; |
934 | ||
398205b8 | 935 | if (!async) { |
2a90d4aa PB |
936 | int cpu = get_cpu(); |
937 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 938 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 939 | put_cpu(); |
398205b8 PB |
940 | return; |
941 | } | |
e4043dcf | 942 | |
2a90d4aa | 943 | put_cpu(); |
e4043dcf | 944 | } |
398205b8 | 945 | |
27489a3c | 946 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
947 | } |
948 | ||
b94ec296 | 949 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
950 | { |
951 | struct blk_mq_hw_ctx *hctx; | |
952 | int i; | |
953 | ||
954 | queue_for_each_hw_ctx(q, hctx, i) { | |
955 | if ((!blk_mq_hctx_has_pending(hctx) && | |
956 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 957 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
958 | continue; |
959 | ||
b94ec296 | 960 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
961 | } |
962 | } | |
b94ec296 | 963 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f JA |
964 | |
965 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
966 | { | |
27489a3c | 967 | cancel_work(&hctx->run_work); |
70f4db63 | 968 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
969 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
970 | } | |
971 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
972 | ||
280d45f6 CH |
973 | void blk_mq_stop_hw_queues(struct request_queue *q) |
974 | { | |
975 | struct blk_mq_hw_ctx *hctx; | |
976 | int i; | |
977 | ||
978 | queue_for_each_hw_ctx(q, hctx, i) | |
979 | blk_mq_stop_hw_queue(hctx); | |
980 | } | |
981 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
982 | ||
320ae51f JA |
983 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
984 | { | |
985 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 986 | |
0ffbce80 | 987 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
988 | } |
989 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
990 | ||
2f268556 CH |
991 | void blk_mq_start_hw_queues(struct request_queue *q) |
992 | { | |
993 | struct blk_mq_hw_ctx *hctx; | |
994 | int i; | |
995 | ||
996 | queue_for_each_hw_ctx(q, hctx, i) | |
997 | blk_mq_start_hw_queue(hctx); | |
998 | } | |
999 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1000 | ||
1b4a3258 | 1001 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1002 | { |
1003 | struct blk_mq_hw_ctx *hctx; | |
1004 | int i; | |
1005 | ||
1006 | queue_for_each_hw_ctx(q, hctx, i) { | |
1007 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1008 | continue; | |
1009 | ||
1010 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1011 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1012 | } |
1013 | } | |
1014 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1015 | ||
70f4db63 | 1016 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1017 | { |
1018 | struct blk_mq_hw_ctx *hctx; | |
1019 | ||
27489a3c | 1020 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1021 | |
320ae51f JA |
1022 | __blk_mq_run_hw_queue(hctx); |
1023 | } | |
1024 | ||
70f4db63 CH |
1025 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1026 | { | |
1027 | struct blk_mq_hw_ctx *hctx; | |
1028 | ||
1029 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1030 | ||
1031 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1032 | __blk_mq_run_hw_queue(hctx); | |
1033 | } | |
1034 | ||
1035 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1036 | { | |
19c66e59 ML |
1037 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1038 | return; | |
70f4db63 | 1039 | |
b657d7e6 CH |
1040 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1041 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1042 | } |
1043 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1044 | ||
cfd0c552 | 1045 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1046 | struct request *rq, |
1047 | bool at_head) | |
320ae51f | 1048 | { |
e57690fe JA |
1049 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1050 | ||
01b983c9 JA |
1051 | trace_block_rq_insert(hctx->queue, rq); |
1052 | ||
72a0a36e CH |
1053 | if (at_head) |
1054 | list_add(&rq->queuelist, &ctx->rq_list); | |
1055 | else | |
1056 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1057 | } |
4bb659b1 | 1058 | |
cfd0c552 ML |
1059 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1060 | struct request *rq, bool at_head) | |
1061 | { | |
1062 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1063 | ||
e57690fe | 1064 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1065 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1066 | } |
1067 | ||
eeabc850 | 1068 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1069 | bool async) |
320ae51f | 1070 | { |
e57690fe | 1071 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1072 | struct request_queue *q = rq->q; |
7d7e0f90 | 1073 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1074 | |
a57a178a CH |
1075 | spin_lock(&ctx->lock); |
1076 | __blk_mq_insert_request(hctx, rq, at_head); | |
1077 | spin_unlock(&ctx->lock); | |
320ae51f | 1078 | |
320ae51f JA |
1079 | if (run_queue) |
1080 | blk_mq_run_hw_queue(hctx, async); | |
1081 | } | |
1082 | ||
1083 | static void blk_mq_insert_requests(struct request_queue *q, | |
1084 | struct blk_mq_ctx *ctx, | |
1085 | struct list_head *list, | |
1086 | int depth, | |
1087 | bool from_schedule) | |
1088 | ||
1089 | { | |
7d7e0f90 | 1090 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1091 | |
1092 | trace_block_unplug(q, depth, !from_schedule); | |
1093 | ||
320ae51f JA |
1094 | /* |
1095 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1096 | * offline now | |
1097 | */ | |
1098 | spin_lock(&ctx->lock); | |
1099 | while (!list_empty(list)) { | |
1100 | struct request *rq; | |
1101 | ||
1102 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1103 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1104 | list_del_init(&rq->queuelist); |
e57690fe | 1105 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1106 | } |
cfd0c552 | 1107 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1108 | spin_unlock(&ctx->lock); |
1109 | ||
320ae51f JA |
1110 | blk_mq_run_hw_queue(hctx, from_schedule); |
1111 | } | |
1112 | ||
1113 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1114 | { | |
1115 | struct request *rqa = container_of(a, struct request, queuelist); | |
1116 | struct request *rqb = container_of(b, struct request, queuelist); | |
1117 | ||
1118 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1119 | (rqa->mq_ctx == rqb->mq_ctx && | |
1120 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1121 | } | |
1122 | ||
1123 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1124 | { | |
1125 | struct blk_mq_ctx *this_ctx; | |
1126 | struct request_queue *this_q; | |
1127 | struct request *rq; | |
1128 | LIST_HEAD(list); | |
1129 | LIST_HEAD(ctx_list); | |
1130 | unsigned int depth; | |
1131 | ||
1132 | list_splice_init(&plug->mq_list, &list); | |
1133 | ||
1134 | list_sort(NULL, &list, plug_ctx_cmp); | |
1135 | ||
1136 | this_q = NULL; | |
1137 | this_ctx = NULL; | |
1138 | depth = 0; | |
1139 | ||
1140 | while (!list_empty(&list)) { | |
1141 | rq = list_entry_rq(list.next); | |
1142 | list_del_init(&rq->queuelist); | |
1143 | BUG_ON(!rq->q); | |
1144 | if (rq->mq_ctx != this_ctx) { | |
1145 | if (this_ctx) { | |
1146 | blk_mq_insert_requests(this_q, this_ctx, | |
1147 | &ctx_list, depth, | |
1148 | from_schedule); | |
1149 | } | |
1150 | ||
1151 | this_ctx = rq->mq_ctx; | |
1152 | this_q = rq->q; | |
1153 | depth = 0; | |
1154 | } | |
1155 | ||
1156 | depth++; | |
1157 | list_add_tail(&rq->queuelist, &ctx_list); | |
1158 | } | |
1159 | ||
1160 | /* | |
1161 | * If 'this_ctx' is set, we know we have entries to complete | |
1162 | * on 'ctx_list'. Do those. | |
1163 | */ | |
1164 | if (this_ctx) { | |
1165 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1166 | from_schedule); | |
1167 | } | |
1168 | } | |
1169 | ||
1170 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1171 | { | |
1172 | init_request_from_bio(rq, bio); | |
4b570521 | 1173 | |
a21f2a3e | 1174 | blk_account_io_start(rq, 1); |
320ae51f JA |
1175 | } |
1176 | ||
274a5843 JA |
1177 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1178 | { | |
1179 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1180 | !blk_queue_nomerges(hctx->queue); | |
1181 | } | |
1182 | ||
07068d5b JA |
1183 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1184 | struct blk_mq_ctx *ctx, | |
1185 | struct request *rq, struct bio *bio) | |
320ae51f | 1186 | { |
e18378a6 | 1187 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1188 | blk_mq_bio_to_request(rq, bio); |
1189 | spin_lock(&ctx->lock); | |
1190 | insert_rq: | |
1191 | __blk_mq_insert_request(hctx, rq, false); | |
1192 | spin_unlock(&ctx->lock); | |
1193 | return false; | |
1194 | } else { | |
274a5843 JA |
1195 | struct request_queue *q = hctx->queue; |
1196 | ||
07068d5b JA |
1197 | spin_lock(&ctx->lock); |
1198 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1199 | blk_mq_bio_to_request(rq, bio); | |
1200 | goto insert_rq; | |
1201 | } | |
320ae51f | 1202 | |
07068d5b JA |
1203 | spin_unlock(&ctx->lock); |
1204 | __blk_mq_free_request(hctx, ctx, rq); | |
1205 | return true; | |
14ec77f3 | 1206 | } |
07068d5b | 1207 | } |
14ec77f3 | 1208 | |
07068d5b JA |
1209 | struct blk_map_ctx { |
1210 | struct blk_mq_hw_ctx *hctx; | |
1211 | struct blk_mq_ctx *ctx; | |
1212 | }; | |
1213 | ||
1214 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1215 | struct bio *bio, | |
1216 | struct blk_map_ctx *data) | |
1217 | { | |
1218 | struct blk_mq_hw_ctx *hctx; | |
1219 | struct blk_mq_ctx *ctx; | |
1220 | struct request *rq; | |
cc6e3b10 MC |
1221 | int op = bio_data_dir(bio); |
1222 | int op_flags = 0; | |
cb96a42c | 1223 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1224 | |
3ef28e83 | 1225 | blk_queue_enter_live(q); |
320ae51f | 1226 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1227 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1228 | |
1eff9d32 | 1229 | if (rw_is_sync(bio_op(bio), bio->bi_opf)) |
cc6e3b10 | 1230 | op_flags |= REQ_SYNC; |
07068d5b | 1231 | |
cc6e3b10 | 1232 | trace_block_getrq(q, bio, op); |
6f3b0e8b | 1233 | blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx); |
cc6e3b10 | 1234 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
5dee8577 | 1235 | if (unlikely(!rq)) { |
793597a6 | 1236 | __blk_mq_run_hw_queue(hctx); |
320ae51f | 1237 | blk_mq_put_ctx(ctx); |
cc6e3b10 | 1238 | trace_block_sleeprq(q, bio, op); |
793597a6 CH |
1239 | |
1240 | ctx = blk_mq_get_ctx(q); | |
7d7e0f90 | 1241 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 1242 | blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx); |
cc6e3b10 | 1243 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
cb96a42c ML |
1244 | ctx = alloc_data.ctx; |
1245 | hctx = alloc_data.hctx; | |
320ae51f JA |
1246 | } |
1247 | ||
1248 | hctx->queued++; | |
07068d5b JA |
1249 | data->hctx = hctx; |
1250 | data->ctx = ctx; | |
1251 | return rq; | |
1252 | } | |
1253 | ||
7b371636 | 1254 | static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1255 | { |
1256 | int ret; | |
1257 | struct request_queue *q = rq->q; | |
7d7e0f90 | 1258 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
f984df1f SL |
1259 | struct blk_mq_queue_data bd = { |
1260 | .rq = rq, | |
1261 | .list = NULL, | |
1262 | .last = 1 | |
1263 | }; | |
7b371636 | 1264 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f SL |
1265 | |
1266 | /* | |
1267 | * For OK queue, we are done. For error, kill it. Any other | |
1268 | * error (busy), just add it to our list as we previously | |
1269 | * would have done | |
1270 | */ | |
1271 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1272 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1273 | *cookie = new_cookie; | |
f984df1f | 1274 | return 0; |
7b371636 | 1275 | } |
f984df1f | 1276 | |
7b371636 JA |
1277 | __blk_mq_requeue_request(rq); |
1278 | ||
1279 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1280 | *cookie = BLK_QC_T_NONE; | |
1281 | rq->errors = -EIO; | |
1282 | blk_mq_end_request(rq, rq->errors); | |
1283 | return 0; | |
f984df1f | 1284 | } |
7b371636 JA |
1285 | |
1286 | return -1; | |
f984df1f SL |
1287 | } |
1288 | ||
07068d5b JA |
1289 | /* |
1290 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1291 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1292 | * hardware for SYNC IO. | |
1293 | */ | |
dece1635 | 1294 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1295 | { |
1eff9d32 JA |
1296 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1297 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
07068d5b JA |
1298 | struct blk_map_ctx data; |
1299 | struct request *rq; | |
f984df1f SL |
1300 | unsigned int request_count = 0; |
1301 | struct blk_plug *plug; | |
5b3f341f | 1302 | struct request *same_queue_rq = NULL; |
7b371636 | 1303 | blk_qc_t cookie; |
07068d5b JA |
1304 | |
1305 | blk_queue_bounce(q, &bio); | |
1306 | ||
1307 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1308 | bio_io_error(bio); |
dece1635 | 1309 | return BLK_QC_T_NONE; |
07068d5b JA |
1310 | } |
1311 | ||
54efd50b KO |
1312 | blk_queue_split(q, &bio, q->bio_split); |
1313 | ||
87c279e6 OS |
1314 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1315 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1316 | return BLK_QC_T_NONE; | |
f984df1f | 1317 | |
07068d5b JA |
1318 | rq = blk_mq_map_request(q, bio, &data); |
1319 | if (unlikely(!rq)) | |
dece1635 | 1320 | return BLK_QC_T_NONE; |
07068d5b | 1321 | |
7b371636 | 1322 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1323 | |
1324 | if (unlikely(is_flush_fua)) { | |
1325 | blk_mq_bio_to_request(rq, bio); | |
1326 | blk_insert_flush(rq); | |
1327 | goto run_queue; | |
1328 | } | |
1329 | ||
f984df1f | 1330 | plug = current->plug; |
e167dfb5 JA |
1331 | /* |
1332 | * If the driver supports defer issued based on 'last', then | |
1333 | * queue it up like normal since we can potentially save some | |
1334 | * CPU this way. | |
1335 | */ | |
f984df1f SL |
1336 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1337 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1338 | struct request *old_rq = NULL; | |
07068d5b JA |
1339 | |
1340 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1341 | |
1342 | /* | |
b094f89c | 1343 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1344 | * Otherwise the existing request in the plug list will be |
1345 | * issued. So the plug list will have one request at most | |
07068d5b | 1346 | */ |
f984df1f | 1347 | if (plug) { |
5b3f341f SL |
1348 | /* |
1349 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1350 | * happens, same_queue_rq is invalid and plug list is |
1351 | * empty | |
1352 | */ | |
5b3f341f SL |
1353 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1354 | old_rq = same_queue_rq; | |
f984df1f | 1355 | list_del_init(&old_rq->queuelist); |
07068d5b | 1356 | } |
f984df1f SL |
1357 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1358 | } else /* is_sync */ | |
1359 | old_rq = rq; | |
1360 | blk_mq_put_ctx(data.ctx); | |
1361 | if (!old_rq) | |
7b371636 JA |
1362 | goto done; |
1363 | if (!blk_mq_direct_issue_request(old_rq, &cookie)) | |
1364 | goto done; | |
f984df1f | 1365 | blk_mq_insert_request(old_rq, false, true, true); |
7b371636 | 1366 | goto done; |
07068d5b JA |
1367 | } |
1368 | ||
1369 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1370 | /* | |
1371 | * For a SYNC request, send it to the hardware immediately. For | |
1372 | * an ASYNC request, just ensure that we run it later on. The | |
1373 | * latter allows for merging opportunities and more efficient | |
1374 | * dispatching. | |
1375 | */ | |
1376 | run_queue: | |
1377 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1378 | } | |
07068d5b | 1379 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1380 | done: |
1381 | return cookie; | |
07068d5b JA |
1382 | } |
1383 | ||
1384 | /* | |
1385 | * Single hardware queue variant. This will attempt to use any per-process | |
1386 | * plug for merging and IO deferral. | |
1387 | */ | |
dece1635 | 1388 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1389 | { |
1eff9d32 JA |
1390 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1391 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
e6c4438b JM |
1392 | struct blk_plug *plug; |
1393 | unsigned int request_count = 0; | |
07068d5b JA |
1394 | struct blk_map_ctx data; |
1395 | struct request *rq; | |
7b371636 | 1396 | blk_qc_t cookie; |
07068d5b | 1397 | |
07068d5b JA |
1398 | blk_queue_bounce(q, &bio); |
1399 | ||
1400 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1401 | bio_io_error(bio); |
dece1635 | 1402 | return BLK_QC_T_NONE; |
07068d5b JA |
1403 | } |
1404 | ||
54efd50b KO |
1405 | blk_queue_split(q, &bio, q->bio_split); |
1406 | ||
87c279e6 OS |
1407 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1408 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1409 | return BLK_QC_T_NONE; | |
1410 | } else | |
1411 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1412 | |
1413 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1414 | if (unlikely(!rq)) |
dece1635 | 1415 | return BLK_QC_T_NONE; |
320ae51f | 1416 | |
7b371636 | 1417 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1418 | |
1419 | if (unlikely(is_flush_fua)) { | |
1420 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1421 | blk_insert_flush(rq); |
1422 | goto run_queue; | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * A task plug currently exists. Since this is completely lockless, | |
1427 | * utilize that to temporarily store requests until the task is | |
1428 | * either done or scheduled away. | |
1429 | */ | |
e6c4438b JM |
1430 | plug = current->plug; |
1431 | if (plug) { | |
1432 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1433 | if (!request_count) |
e6c4438b | 1434 | trace_block_plug(q); |
b094f89c JA |
1435 | |
1436 | blk_mq_put_ctx(data.ctx); | |
1437 | ||
1438 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1439 | blk_flush_plug_list(plug, false); |
1440 | trace_block_plug(q); | |
320ae51f | 1441 | } |
b094f89c | 1442 | |
e6c4438b | 1443 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1444 | return cookie; |
320ae51f JA |
1445 | } |
1446 | ||
07068d5b JA |
1447 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1448 | /* | |
1449 | * For a SYNC request, send it to the hardware immediately. For | |
1450 | * an ASYNC request, just ensure that we run it later on. The | |
1451 | * latter allows for merging opportunities and more efficient | |
1452 | * dispatching. | |
1453 | */ | |
1454 | run_queue: | |
1455 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1456 | } |
1457 | ||
07068d5b | 1458 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1459 | return cookie; |
320ae51f JA |
1460 | } |
1461 | ||
24d2f903 CH |
1462 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1463 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1464 | { |
e9b267d9 | 1465 | struct page *page; |
320ae51f | 1466 | |
24d2f903 | 1467 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1468 | int i; |
320ae51f | 1469 | |
24d2f903 CH |
1470 | for (i = 0; i < tags->nr_tags; i++) { |
1471 | if (!tags->rqs[i]) | |
e9b267d9 | 1472 | continue; |
24d2f903 CH |
1473 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1474 | hctx_idx, i); | |
a5164405 | 1475 | tags->rqs[i] = NULL; |
e9b267d9 | 1476 | } |
320ae51f | 1477 | } |
320ae51f | 1478 | |
24d2f903 CH |
1479 | while (!list_empty(&tags->page_list)) { |
1480 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1481 | list_del_init(&page->lru); |
f75782e4 CM |
1482 | /* |
1483 | * Remove kmemleak object previously allocated in | |
1484 | * blk_mq_init_rq_map(). | |
1485 | */ | |
1486 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1487 | __free_pages(page, page->private); |
1488 | } | |
1489 | ||
24d2f903 | 1490 | kfree(tags->rqs); |
320ae51f | 1491 | |
24d2f903 | 1492 | blk_mq_free_tags(tags); |
320ae51f JA |
1493 | } |
1494 | ||
1495 | static size_t order_to_size(unsigned int order) | |
1496 | { | |
4ca08500 | 1497 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1498 | } |
1499 | ||
24d2f903 CH |
1500 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1501 | unsigned int hctx_idx) | |
320ae51f | 1502 | { |
24d2f903 | 1503 | struct blk_mq_tags *tags; |
320ae51f JA |
1504 | unsigned int i, j, entries_per_page, max_order = 4; |
1505 | size_t rq_size, left; | |
1506 | ||
24d2f903 | 1507 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1508 | set->numa_node, |
1509 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1510 | if (!tags) |
1511 | return NULL; | |
320ae51f | 1512 | |
24d2f903 CH |
1513 | INIT_LIST_HEAD(&tags->page_list); |
1514 | ||
a5164405 JA |
1515 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1516 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1517 | set->numa_node); | |
24d2f903 CH |
1518 | if (!tags->rqs) { |
1519 | blk_mq_free_tags(tags); | |
1520 | return NULL; | |
1521 | } | |
320ae51f JA |
1522 | |
1523 | /* | |
1524 | * rq_size is the size of the request plus driver payload, rounded | |
1525 | * to the cacheline size | |
1526 | */ | |
24d2f903 | 1527 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1528 | cache_line_size()); |
24d2f903 | 1529 | left = rq_size * set->queue_depth; |
320ae51f | 1530 | |
24d2f903 | 1531 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1532 | int this_order = max_order; |
1533 | struct page *page; | |
1534 | int to_do; | |
1535 | void *p; | |
1536 | ||
b3a834b1 | 1537 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1538 | this_order--; |
1539 | ||
1540 | do { | |
a5164405 | 1541 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1542 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1543 | this_order); |
320ae51f JA |
1544 | if (page) |
1545 | break; | |
1546 | if (!this_order--) | |
1547 | break; | |
1548 | if (order_to_size(this_order) < rq_size) | |
1549 | break; | |
1550 | } while (1); | |
1551 | ||
1552 | if (!page) | |
24d2f903 | 1553 | goto fail; |
320ae51f JA |
1554 | |
1555 | page->private = this_order; | |
24d2f903 | 1556 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1557 | |
1558 | p = page_address(page); | |
f75782e4 CM |
1559 | /* |
1560 | * Allow kmemleak to scan these pages as they contain pointers | |
1561 | * to additional allocations like via ops->init_request(). | |
1562 | */ | |
1563 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1564 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1565 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1566 | left -= to_do * rq_size; |
1567 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1568 | tags->rqs[i] = p; |
1569 | if (set->ops->init_request) { | |
1570 | if (set->ops->init_request(set->driver_data, | |
1571 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1572 | set->numa_node)) { |
1573 | tags->rqs[i] = NULL; | |
24d2f903 | 1574 | goto fail; |
a5164405 | 1575 | } |
e9b267d9 CH |
1576 | } |
1577 | ||
320ae51f JA |
1578 | p += rq_size; |
1579 | i++; | |
1580 | } | |
1581 | } | |
24d2f903 | 1582 | return tags; |
320ae51f | 1583 | |
24d2f903 | 1584 | fail: |
24d2f903 CH |
1585 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1586 | return NULL; | |
320ae51f JA |
1587 | } |
1588 | ||
1429d7c9 JA |
1589 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1590 | { | |
1591 | kfree(bitmap->map); | |
1592 | } | |
1593 | ||
1594 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1595 | { | |
1596 | unsigned int bpw = 8, total, num_maps, i; | |
1597 | ||
1598 | bitmap->bits_per_word = bpw; | |
1599 | ||
1600 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1601 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1602 | GFP_KERNEL, node); | |
1603 | if (!bitmap->map) | |
1604 | return -ENOMEM; | |
1605 | ||
1429d7c9 JA |
1606 | total = nr_cpu_ids; |
1607 | for (i = 0; i < num_maps; i++) { | |
1608 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1609 | total -= bitmap->map[i].depth; | |
1610 | } | |
1611 | ||
1612 | return 0; | |
1613 | } | |
1614 | ||
e57690fe JA |
1615 | /* |
1616 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1617 | * software queue to the hw queue dispatch list, and ensure that it | |
1618 | * gets run. | |
1619 | */ | |
484b4061 JA |
1620 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1621 | { | |
484b4061 JA |
1622 | struct blk_mq_ctx *ctx; |
1623 | LIST_HEAD(tmp); | |
1624 | ||
e57690fe | 1625 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1626 | |
1627 | spin_lock(&ctx->lock); | |
1628 | if (!list_empty(&ctx->rq_list)) { | |
1629 | list_splice_init(&ctx->rq_list, &tmp); | |
1630 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1631 | } | |
1632 | spin_unlock(&ctx->lock); | |
1633 | ||
1634 | if (list_empty(&tmp)) | |
1635 | return NOTIFY_OK; | |
1636 | ||
e57690fe JA |
1637 | spin_lock(&hctx->lock); |
1638 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1639 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1640 | |
1641 | blk_mq_run_hw_queue(hctx, true); | |
484b4061 JA |
1642 | return NOTIFY_OK; |
1643 | } | |
1644 | ||
484b4061 JA |
1645 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
1646 | unsigned int cpu) | |
1647 | { | |
1648 | struct blk_mq_hw_ctx *hctx = data; | |
1649 | ||
1650 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1651 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
2a34c087 ML |
1652 | |
1653 | /* | |
1654 | * In case of CPU online, tags may be reallocated | |
1655 | * in blk_mq_map_swqueue() after mapping is updated. | |
1656 | */ | |
484b4061 JA |
1657 | |
1658 | return NOTIFY_OK; | |
1659 | } | |
1660 | ||
c3b4afca | 1661 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1662 | static void blk_mq_exit_hctx(struct request_queue *q, |
1663 | struct blk_mq_tag_set *set, | |
1664 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1665 | { | |
f70ced09 ML |
1666 | unsigned flush_start_tag = set->queue_depth; |
1667 | ||
08e98fc6 ML |
1668 | blk_mq_tag_idle(hctx); |
1669 | ||
f70ced09 ML |
1670 | if (set->ops->exit_request) |
1671 | set->ops->exit_request(set->driver_data, | |
1672 | hctx->fq->flush_rq, hctx_idx, | |
1673 | flush_start_tag + hctx_idx); | |
1674 | ||
08e98fc6 ML |
1675 | if (set->ops->exit_hctx) |
1676 | set->ops->exit_hctx(hctx, hctx_idx); | |
1677 | ||
1678 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1679 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1680 | blk_mq_free_bitmap(&hctx->ctx_map); |
1681 | } | |
1682 | ||
624dbe47 ML |
1683 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1684 | struct blk_mq_tag_set *set, int nr_queue) | |
1685 | { | |
1686 | struct blk_mq_hw_ctx *hctx; | |
1687 | unsigned int i; | |
1688 | ||
1689 | queue_for_each_hw_ctx(q, hctx, i) { | |
1690 | if (i == nr_queue) | |
1691 | break; | |
08e98fc6 | 1692 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1693 | } |
624dbe47 ML |
1694 | } |
1695 | ||
1696 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1697 | struct blk_mq_tag_set *set) | |
1698 | { | |
1699 | struct blk_mq_hw_ctx *hctx; | |
1700 | unsigned int i; | |
1701 | ||
e09aae7e | 1702 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1703 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1704 | } |
1705 | ||
08e98fc6 ML |
1706 | static int blk_mq_init_hctx(struct request_queue *q, |
1707 | struct blk_mq_tag_set *set, | |
1708 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1709 | { |
08e98fc6 | 1710 | int node; |
f70ced09 | 1711 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1712 | |
1713 | node = hctx->numa_node; | |
1714 | if (node == NUMA_NO_NODE) | |
1715 | node = hctx->numa_node = set->numa_node; | |
1716 | ||
27489a3c | 1717 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1718 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1719 | spin_lock_init(&hctx->lock); | |
1720 | INIT_LIST_HEAD(&hctx->dispatch); | |
1721 | hctx->queue = q; | |
1722 | hctx->queue_num = hctx_idx; | |
2404e607 | 1723 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 ML |
1724 | |
1725 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1726 | blk_mq_hctx_notify, hctx); | |
1727 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1728 | ||
1729 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1730 | |
1731 | /* | |
08e98fc6 ML |
1732 | * Allocate space for all possible cpus to avoid allocation at |
1733 | * runtime | |
320ae51f | 1734 | */ |
08e98fc6 ML |
1735 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1736 | GFP_KERNEL, node); | |
1737 | if (!hctx->ctxs) | |
1738 | goto unregister_cpu_notifier; | |
320ae51f | 1739 | |
08e98fc6 ML |
1740 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1741 | goto free_ctxs; | |
320ae51f | 1742 | |
08e98fc6 | 1743 | hctx->nr_ctx = 0; |
320ae51f | 1744 | |
08e98fc6 ML |
1745 | if (set->ops->init_hctx && |
1746 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1747 | goto free_bitmap; | |
320ae51f | 1748 | |
f70ced09 ML |
1749 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1750 | if (!hctx->fq) | |
1751 | goto exit_hctx; | |
320ae51f | 1752 | |
f70ced09 ML |
1753 | if (set->ops->init_request && |
1754 | set->ops->init_request(set->driver_data, | |
1755 | hctx->fq->flush_rq, hctx_idx, | |
1756 | flush_start_tag + hctx_idx, node)) | |
1757 | goto free_fq; | |
320ae51f | 1758 | |
08e98fc6 | 1759 | return 0; |
320ae51f | 1760 | |
f70ced09 ML |
1761 | free_fq: |
1762 | kfree(hctx->fq); | |
1763 | exit_hctx: | |
1764 | if (set->ops->exit_hctx) | |
1765 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1766 | free_bitmap: |
1767 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1768 | free_ctxs: | |
1769 | kfree(hctx->ctxs); | |
1770 | unregister_cpu_notifier: | |
1771 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1772 | |
08e98fc6 ML |
1773 | return -1; |
1774 | } | |
320ae51f | 1775 | |
320ae51f JA |
1776 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1777 | unsigned int nr_hw_queues) | |
1778 | { | |
1779 | unsigned int i; | |
1780 | ||
1781 | for_each_possible_cpu(i) { | |
1782 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1783 | struct blk_mq_hw_ctx *hctx; | |
1784 | ||
1785 | memset(__ctx, 0, sizeof(*__ctx)); | |
1786 | __ctx->cpu = i; | |
1787 | spin_lock_init(&__ctx->lock); | |
1788 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1789 | __ctx->queue = q; | |
1790 | ||
1791 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1792 | if (!cpu_online(i)) |
1793 | continue; | |
1794 | ||
7d7e0f90 | 1795 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1796 | |
320ae51f JA |
1797 | /* |
1798 | * Set local node, IFF we have more than one hw queue. If | |
1799 | * not, we remain on the home node of the device | |
1800 | */ | |
1801 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1802 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1803 | } |
1804 | } | |
1805 | ||
5778322e AM |
1806 | static void blk_mq_map_swqueue(struct request_queue *q, |
1807 | const struct cpumask *online_mask) | |
320ae51f JA |
1808 | { |
1809 | unsigned int i; | |
1810 | struct blk_mq_hw_ctx *hctx; | |
1811 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1812 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1813 | |
60de074b AM |
1814 | /* |
1815 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1816 | */ | |
1817 | mutex_lock(&q->sysfs_lock); | |
1818 | ||
320ae51f | 1819 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1820 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1821 | hctx->nr_ctx = 0; |
1822 | } | |
1823 | ||
1824 | /* | |
1825 | * Map software to hardware queues | |
1826 | */ | |
897bb0c7 | 1827 | for_each_possible_cpu(i) { |
320ae51f | 1828 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1829 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1830 | continue; |
1831 | ||
897bb0c7 | 1832 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1833 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1834 | |
e4043dcf | 1835 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1836 | ctx->index_hw = hctx->nr_ctx; |
1837 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1838 | } | |
506e931f | 1839 | |
60de074b AM |
1840 | mutex_unlock(&q->sysfs_lock); |
1841 | ||
506e931f | 1842 | queue_for_each_hw_ctx(q, hctx, i) { |
889fa31f CY |
1843 | struct blk_mq_ctxmap *map = &hctx->ctx_map; |
1844 | ||
484b4061 | 1845 | /* |
a68aafa5 JA |
1846 | * If no software queues are mapped to this hardware queue, |
1847 | * disable it and free the request entries. | |
484b4061 JA |
1848 | */ |
1849 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1850 | if (set->tags[i]) { |
1851 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1852 | set->tags[i] = NULL; | |
484b4061 | 1853 | } |
2a34c087 | 1854 | hctx->tags = NULL; |
484b4061 JA |
1855 | continue; |
1856 | } | |
1857 | ||
2a34c087 ML |
1858 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1859 | if (!set->tags[i]) | |
1860 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1861 | hctx->tags = set->tags[i]; | |
1862 | WARN_ON(!hctx->tags); | |
1863 | ||
e0e827b9 | 1864 | cpumask_copy(hctx->tags->cpumask, hctx->cpumask); |
889fa31f CY |
1865 | /* |
1866 | * Set the map size to the number of mapped software queues. | |
1867 | * This is more accurate and more efficient than looping | |
1868 | * over all possibly mapped software queues. | |
1869 | */ | |
569fd0ce | 1870 | map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word); |
889fa31f | 1871 | |
484b4061 JA |
1872 | /* |
1873 | * Initialize batch roundrobin counts | |
1874 | */ | |
506e931f JA |
1875 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1876 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1877 | } | |
320ae51f JA |
1878 | } |
1879 | ||
2404e607 | 1880 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1881 | { |
1882 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1883 | int i; |
1884 | ||
2404e607 JM |
1885 | queue_for_each_hw_ctx(q, hctx, i) { |
1886 | if (shared) | |
1887 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1888 | else | |
1889 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1890 | } | |
1891 | } | |
1892 | ||
1893 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1894 | { | |
1895 | struct request_queue *q; | |
0d2602ca JA |
1896 | |
1897 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1898 | blk_mq_freeze_queue(q); | |
2404e607 | 1899 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1900 | blk_mq_unfreeze_queue(q); |
1901 | } | |
1902 | } | |
1903 | ||
1904 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1905 | { | |
1906 | struct blk_mq_tag_set *set = q->tag_set; | |
1907 | ||
0d2602ca JA |
1908 | mutex_lock(&set->tag_list_lock); |
1909 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1910 | if (list_is_singular(&set->tag_list)) { |
1911 | /* just transitioned to unshared */ | |
1912 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1913 | /* update existing queue */ | |
1914 | blk_mq_update_tag_set_depth(set, false); | |
1915 | } | |
0d2602ca | 1916 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1917 | } |
1918 | ||
1919 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1920 | struct request_queue *q) | |
1921 | { | |
1922 | q->tag_set = set; | |
1923 | ||
1924 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1925 | |
1926 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1927 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1928 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1929 | /* update existing queue */ | |
1930 | blk_mq_update_tag_set_depth(set, true); | |
1931 | } | |
1932 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1933 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1934 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1935 | |
0d2602ca JA |
1936 | mutex_unlock(&set->tag_list_lock); |
1937 | } | |
1938 | ||
e09aae7e ML |
1939 | /* |
1940 | * It is the actual release handler for mq, but we do it from | |
1941 | * request queue's release handler for avoiding use-after-free | |
1942 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1943 | * but we can't group ctx/kctx kobj without it. | |
1944 | */ | |
1945 | void blk_mq_release(struct request_queue *q) | |
1946 | { | |
1947 | struct blk_mq_hw_ctx *hctx; | |
1948 | unsigned int i; | |
1949 | ||
1950 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1951 | queue_for_each_hw_ctx(q, hctx, i) { |
1952 | if (!hctx) | |
1953 | continue; | |
1954 | kfree(hctx->ctxs); | |
e09aae7e | 1955 | kfree(hctx); |
c3b4afca | 1956 | } |
e09aae7e | 1957 | |
a723bab3 AM |
1958 | q->mq_map = NULL; |
1959 | ||
e09aae7e ML |
1960 | kfree(q->queue_hw_ctx); |
1961 | ||
1962 | /* ctx kobj stays in queue_ctx */ | |
1963 | free_percpu(q->queue_ctx); | |
1964 | } | |
1965 | ||
24d2f903 | 1966 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1967 | { |
1968 | struct request_queue *uninit_q, *q; | |
1969 | ||
1970 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1971 | if (!uninit_q) | |
1972 | return ERR_PTR(-ENOMEM); | |
1973 | ||
1974 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1975 | if (IS_ERR(q)) | |
1976 | blk_cleanup_queue(uninit_q); | |
1977 | ||
1978 | return q; | |
1979 | } | |
1980 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1981 | ||
868f2f0b KB |
1982 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
1983 | struct request_queue *q) | |
320ae51f | 1984 | { |
868f2f0b KB |
1985 | int i, j; |
1986 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 1987 | |
868f2f0b | 1988 | blk_mq_sysfs_unregister(q); |
24d2f903 | 1989 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 1990 | int node; |
f14bbe77 | 1991 | |
868f2f0b KB |
1992 | if (hctxs[i]) |
1993 | continue; | |
1994 | ||
1995 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
1996 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1997 | GFP_KERNEL, node); | |
320ae51f | 1998 | if (!hctxs[i]) |
868f2f0b | 1999 | break; |
320ae51f | 2000 | |
a86073e4 | 2001 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2002 | node)) { |
2003 | kfree(hctxs[i]); | |
2004 | hctxs[i] = NULL; | |
2005 | break; | |
2006 | } | |
e4043dcf | 2007 | |
0d2602ca | 2008 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2009 | hctxs[i]->numa_node = node; |
320ae51f | 2010 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2011 | |
2012 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2013 | free_cpumask_var(hctxs[i]->cpumask); | |
2014 | kfree(hctxs[i]); | |
2015 | hctxs[i] = NULL; | |
2016 | break; | |
2017 | } | |
2018 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2019 | } |
868f2f0b KB |
2020 | for (j = i; j < q->nr_hw_queues; j++) { |
2021 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2022 | ||
2023 | if (hctx) { | |
2024 | if (hctx->tags) { | |
2025 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2026 | set->tags[j] = NULL; | |
2027 | } | |
2028 | blk_mq_exit_hctx(q, set, hctx, j); | |
2029 | free_cpumask_var(hctx->cpumask); | |
2030 | kobject_put(&hctx->kobj); | |
2031 | kfree(hctx->ctxs); | |
2032 | kfree(hctx); | |
2033 | hctxs[j] = NULL; | |
2034 | ||
2035 | } | |
2036 | } | |
2037 | q->nr_hw_queues = i; | |
2038 | blk_mq_sysfs_register(q); | |
2039 | } | |
2040 | ||
2041 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2042 | struct request_queue *q) | |
2043 | { | |
66841672 ML |
2044 | /* mark the queue as mq asap */ |
2045 | q->mq_ops = set->ops; | |
2046 | ||
868f2f0b KB |
2047 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2048 | if (!q->queue_ctx) | |
c7de5726 | 2049 | goto err_exit; |
868f2f0b KB |
2050 | |
2051 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2052 | GFP_KERNEL, set->numa_node); | |
2053 | if (!q->queue_hw_ctx) | |
2054 | goto err_percpu; | |
2055 | ||
bdd17e75 | 2056 | q->mq_map = set->mq_map; |
868f2f0b KB |
2057 | |
2058 | blk_mq_realloc_hw_ctxs(set, q); | |
2059 | if (!q->nr_hw_queues) | |
2060 | goto err_hctxs; | |
320ae51f | 2061 | |
287922eb | 2062 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2063 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2064 | |
2065 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2066 | |
94eddfbe | 2067 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2068 | |
05f1dd53 JA |
2069 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2070 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2071 | ||
1be036e9 CH |
2072 | q->sg_reserved_size = INT_MAX; |
2073 | ||
2849450a | 2074 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2075 | INIT_LIST_HEAD(&q->requeue_list); |
2076 | spin_lock_init(&q->requeue_lock); | |
2077 | ||
07068d5b JA |
2078 | if (q->nr_hw_queues > 1) |
2079 | blk_queue_make_request(q, blk_mq_make_request); | |
2080 | else | |
2081 | blk_queue_make_request(q, blk_sq_make_request); | |
2082 | ||
eba71768 JA |
2083 | /* |
2084 | * Do this after blk_queue_make_request() overrides it... | |
2085 | */ | |
2086 | q->nr_requests = set->queue_depth; | |
2087 | ||
24d2f903 CH |
2088 | if (set->ops->complete) |
2089 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2090 | |
24d2f903 | 2091 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2092 | |
5778322e | 2093 | get_online_cpus(); |
320ae51f | 2094 | mutex_lock(&all_q_mutex); |
320ae51f | 2095 | |
4593fdbe | 2096 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2097 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2098 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2099 | |
4593fdbe | 2100 | mutex_unlock(&all_q_mutex); |
5778322e | 2101 | put_online_cpus(); |
4593fdbe | 2102 | |
320ae51f | 2103 | return q; |
18741986 | 2104 | |
320ae51f | 2105 | err_hctxs: |
868f2f0b | 2106 | kfree(q->queue_hw_ctx); |
320ae51f | 2107 | err_percpu: |
868f2f0b | 2108 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2109 | err_exit: |
2110 | q->mq_ops = NULL; | |
320ae51f JA |
2111 | return ERR_PTR(-ENOMEM); |
2112 | } | |
b62c21b7 | 2113 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2114 | |
2115 | void blk_mq_free_queue(struct request_queue *q) | |
2116 | { | |
624dbe47 | 2117 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2118 | |
0e626368 AM |
2119 | mutex_lock(&all_q_mutex); |
2120 | list_del_init(&q->all_q_node); | |
2121 | mutex_unlock(&all_q_mutex); | |
2122 | ||
0d2602ca JA |
2123 | blk_mq_del_queue_tag_set(q); |
2124 | ||
624dbe47 ML |
2125 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2126 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2127 | } |
320ae51f JA |
2128 | |
2129 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2130 | static void blk_mq_queue_reinit(struct request_queue *q, |
2131 | const struct cpumask *online_mask) | |
320ae51f | 2132 | { |
4ecd4fef | 2133 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2134 | |
67aec14c JA |
2135 | blk_mq_sysfs_unregister(q); |
2136 | ||
320ae51f JA |
2137 | /* |
2138 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2139 | * we should change hctx numa_node according to new topology (this | |
2140 | * involves free and re-allocate memory, worthy doing?) | |
2141 | */ | |
2142 | ||
5778322e | 2143 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2144 | |
67aec14c | 2145 | blk_mq_sysfs_register(q); |
320ae51f JA |
2146 | } |
2147 | ||
f618ef7c PG |
2148 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
2149 | unsigned long action, void *hcpu) | |
320ae51f JA |
2150 | { |
2151 | struct request_queue *q; | |
5778322e AM |
2152 | int cpu = (unsigned long)hcpu; |
2153 | /* | |
2154 | * New online cpumask which is going to be set in this hotplug event. | |
2155 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2156 | * one-by-one and dynamically allocating this could result in a failure. | |
2157 | */ | |
2158 | static struct cpumask online_new; | |
320ae51f JA |
2159 | |
2160 | /* | |
5778322e AM |
2161 | * Before hotadded cpu starts handling requests, new mappings must |
2162 | * be established. Otherwise, these requests in hw queue might | |
2163 | * never be dispatched. | |
2164 | * | |
2165 | * For example, there is a single hw queue (hctx) and two CPU queues | |
2166 | * (ctx0 for CPU0, and ctx1 for CPU1). | |
2167 | * | |
2168 | * Now CPU1 is just onlined and a request is inserted into | |
2169 | * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is | |
2170 | * still zero. | |
2171 | * | |
2172 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is | |
2173 | * set in pending bitmap and tries to retrieve requests in | |
2174 | * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, | |
2175 | * so the request in ctx1->rq_list is ignored. | |
320ae51f | 2176 | */ |
5778322e AM |
2177 | switch (action & ~CPU_TASKS_FROZEN) { |
2178 | case CPU_DEAD: | |
2179 | case CPU_UP_CANCELED: | |
2180 | cpumask_copy(&online_new, cpu_online_mask); | |
2181 | break; | |
2182 | case CPU_UP_PREPARE: | |
2183 | cpumask_copy(&online_new, cpu_online_mask); | |
2184 | cpumask_set_cpu(cpu, &online_new); | |
2185 | break; | |
2186 | default: | |
320ae51f | 2187 | return NOTIFY_OK; |
5778322e | 2188 | } |
320ae51f JA |
2189 | |
2190 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2191 | |
2192 | /* | |
2193 | * We need to freeze and reinit all existing queues. Freezing | |
2194 | * involves synchronous wait for an RCU grace period and doing it | |
2195 | * one by one may take a long time. Start freezing all queues in | |
2196 | * one swoop and then wait for the completions so that freezing can | |
2197 | * take place in parallel. | |
2198 | */ | |
2199 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2200 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2201 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2202 | blk_mq_freeze_queue_wait(q); |
2203 | ||
f054b56c ML |
2204 | /* |
2205 | * timeout handler can't touch hw queue during the | |
2206 | * reinitialization | |
2207 | */ | |
2208 | del_timer_sync(&q->timeout); | |
2209 | } | |
2210 | ||
320ae51f | 2211 | list_for_each_entry(q, &all_q_list, all_q_node) |
5778322e | 2212 | blk_mq_queue_reinit(q, &online_new); |
f3af020b TH |
2213 | |
2214 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2215 | blk_mq_unfreeze_queue(q); | |
2216 | ||
320ae51f JA |
2217 | mutex_unlock(&all_q_mutex); |
2218 | return NOTIFY_OK; | |
2219 | } | |
2220 | ||
a5164405 JA |
2221 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2222 | { | |
2223 | int i; | |
2224 | ||
2225 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2226 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2227 | if (!set->tags[i]) | |
2228 | goto out_unwind; | |
2229 | } | |
2230 | ||
2231 | return 0; | |
2232 | ||
2233 | out_unwind: | |
2234 | while (--i >= 0) | |
2235 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2236 | ||
a5164405 JA |
2237 | return -ENOMEM; |
2238 | } | |
2239 | ||
2240 | /* | |
2241 | * Allocate the request maps associated with this tag_set. Note that this | |
2242 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2243 | * will be updated to reflect the allocated depth. | |
2244 | */ | |
2245 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2246 | { | |
2247 | unsigned int depth; | |
2248 | int err; | |
2249 | ||
2250 | depth = set->queue_depth; | |
2251 | do { | |
2252 | err = __blk_mq_alloc_rq_maps(set); | |
2253 | if (!err) | |
2254 | break; | |
2255 | ||
2256 | set->queue_depth >>= 1; | |
2257 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2258 | err = -ENOMEM; | |
2259 | break; | |
2260 | } | |
2261 | } while (set->queue_depth); | |
2262 | ||
2263 | if (!set->queue_depth || err) { | |
2264 | pr_err("blk-mq: failed to allocate request map\n"); | |
2265 | return -ENOMEM; | |
2266 | } | |
2267 | ||
2268 | if (depth != set->queue_depth) | |
2269 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2270 | depth, set->queue_depth); | |
2271 | ||
2272 | return 0; | |
2273 | } | |
2274 | ||
f26cdc85 KB |
2275 | struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags) |
2276 | { | |
2277 | return tags->cpumask; | |
2278 | } | |
2279 | EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask); | |
2280 | ||
a4391c64 JA |
2281 | /* |
2282 | * Alloc a tag set to be associated with one or more request queues. | |
2283 | * May fail with EINVAL for various error conditions. May adjust the | |
2284 | * requested depth down, if if it too large. In that case, the set | |
2285 | * value will be stored in set->queue_depth. | |
2286 | */ | |
24d2f903 CH |
2287 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2288 | { | |
205fb5f5 BVA |
2289 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2290 | ||
24d2f903 CH |
2291 | if (!set->nr_hw_queues) |
2292 | return -EINVAL; | |
a4391c64 | 2293 | if (!set->queue_depth) |
24d2f903 CH |
2294 | return -EINVAL; |
2295 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2296 | return -EINVAL; | |
2297 | ||
7d7e0f90 | 2298 | if (!set->ops->queue_rq) |
24d2f903 CH |
2299 | return -EINVAL; |
2300 | ||
a4391c64 JA |
2301 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2302 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2303 | BLK_MQ_MAX_DEPTH); | |
2304 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2305 | } | |
24d2f903 | 2306 | |
6637fadf SL |
2307 | /* |
2308 | * If a crashdump is active, then we are potentially in a very | |
2309 | * memory constrained environment. Limit us to 1 queue and | |
2310 | * 64 tags to prevent using too much memory. | |
2311 | */ | |
2312 | if (is_kdump_kernel()) { | |
2313 | set->nr_hw_queues = 1; | |
2314 | set->queue_depth = min(64U, set->queue_depth); | |
2315 | } | |
868f2f0b KB |
2316 | /* |
2317 | * There is no use for more h/w queues than cpus. | |
2318 | */ | |
2319 | if (set->nr_hw_queues > nr_cpu_ids) | |
2320 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2321 | |
868f2f0b | 2322 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2323 | GFP_KERNEL, set->numa_node); |
2324 | if (!set->tags) | |
a5164405 | 2325 | return -ENOMEM; |
24d2f903 | 2326 | |
bdd17e75 CH |
2327 | set->mq_map = blk_mq_make_queue_map(set); |
2328 | if (!set->mq_map) | |
2329 | goto out_free_tags; | |
2330 | ||
a5164405 | 2331 | if (blk_mq_alloc_rq_maps(set)) |
bdd17e75 | 2332 | goto out_free_mq_map; |
24d2f903 | 2333 | |
0d2602ca JA |
2334 | mutex_init(&set->tag_list_lock); |
2335 | INIT_LIST_HEAD(&set->tag_list); | |
2336 | ||
24d2f903 | 2337 | return 0; |
bdd17e75 CH |
2338 | |
2339 | out_free_mq_map: | |
2340 | kfree(set->mq_map); | |
2341 | set->mq_map = NULL; | |
2342 | out_free_tags: | |
5676e7b6 RE |
2343 | kfree(set->tags); |
2344 | set->tags = NULL; | |
24d2f903 CH |
2345 | return -ENOMEM; |
2346 | } | |
2347 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2348 | ||
2349 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2350 | { | |
2351 | int i; | |
2352 | ||
868f2f0b | 2353 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2354 | if (set->tags[i]) |
484b4061 JA |
2355 | blk_mq_free_rq_map(set, set->tags[i], i); |
2356 | } | |
2357 | ||
bdd17e75 CH |
2358 | kfree(set->mq_map); |
2359 | set->mq_map = NULL; | |
2360 | ||
981bd189 | 2361 | kfree(set->tags); |
5676e7b6 | 2362 | set->tags = NULL; |
24d2f903 CH |
2363 | } |
2364 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2365 | ||
e3a2b3f9 JA |
2366 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2367 | { | |
2368 | struct blk_mq_tag_set *set = q->tag_set; | |
2369 | struct blk_mq_hw_ctx *hctx; | |
2370 | int i, ret; | |
2371 | ||
2372 | if (!set || nr > set->queue_depth) | |
2373 | return -EINVAL; | |
2374 | ||
2375 | ret = 0; | |
2376 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2377 | if (!hctx->tags) |
2378 | continue; | |
e3a2b3f9 JA |
2379 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2380 | if (ret) | |
2381 | break; | |
2382 | } | |
2383 | ||
2384 | if (!ret) | |
2385 | q->nr_requests = nr; | |
2386 | ||
2387 | return ret; | |
2388 | } | |
2389 | ||
868f2f0b KB |
2390 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2391 | { | |
2392 | struct request_queue *q; | |
2393 | ||
2394 | if (nr_hw_queues > nr_cpu_ids) | |
2395 | nr_hw_queues = nr_cpu_ids; | |
2396 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2397 | return; | |
2398 | ||
2399 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2400 | blk_mq_freeze_queue(q); | |
2401 | ||
2402 | set->nr_hw_queues = nr_hw_queues; | |
2403 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2404 | blk_mq_realloc_hw_ctxs(set, q); | |
2405 | ||
2406 | if (q->nr_hw_queues > 1) | |
2407 | blk_queue_make_request(q, blk_mq_make_request); | |
2408 | else | |
2409 | blk_queue_make_request(q, blk_sq_make_request); | |
2410 | ||
2411 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2412 | } | |
2413 | ||
2414 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2415 | blk_mq_unfreeze_queue(q); | |
2416 | } | |
2417 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2418 | ||
676141e4 JA |
2419 | void blk_mq_disable_hotplug(void) |
2420 | { | |
2421 | mutex_lock(&all_q_mutex); | |
2422 | } | |
2423 | ||
2424 | void blk_mq_enable_hotplug(void) | |
2425 | { | |
2426 | mutex_unlock(&all_q_mutex); | |
2427 | } | |
2428 | ||
320ae51f JA |
2429 | static int __init blk_mq_init(void) |
2430 | { | |
320ae51f JA |
2431 | blk_mq_cpu_init(); |
2432 | ||
add703fd | 2433 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2434 | |
2435 | return 0; | |
2436 | } | |
2437 | subsys_initcall(blk_mq_init); |