]>
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> |
320ae51f JA |
25 | |
26 | #include <trace/events/block.h> | |
27 | ||
28 | #include <linux/blk-mq.h> | |
29 | #include "blk.h" | |
30 | #include "blk-mq.h" | |
31 | #include "blk-mq-tag.h" | |
32 | ||
33 | static DEFINE_MUTEX(all_q_mutex); | |
34 | static LIST_HEAD(all_q_list); | |
35 | ||
36 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx); | |
37 | ||
320ae51f JA |
38 | /* |
39 | * Check if any of the ctx's have pending work in this hardware queue | |
40 | */ | |
41 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
42 | { | |
43 | unsigned int i; | |
44 | ||
569fd0ce | 45 | for (i = 0; i < hctx->ctx_map.size; i++) |
1429d7c9 | 46 | if (hctx->ctx_map.map[i].word) |
320ae51f JA |
47 | return true; |
48 | ||
49 | return false; | |
50 | } | |
51 | ||
1429d7c9 JA |
52 | static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx, |
53 | struct blk_mq_ctx *ctx) | |
54 | { | |
55 | return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word]; | |
56 | } | |
57 | ||
58 | #define CTX_TO_BIT(hctx, ctx) \ | |
59 | ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1)) | |
60 | ||
320ae51f JA |
61 | /* |
62 | * Mark this ctx as having pending work in this hardware queue | |
63 | */ | |
64 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
65 | struct blk_mq_ctx *ctx) | |
66 | { | |
1429d7c9 JA |
67 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); |
68 | ||
69 | if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word)) | |
70 | set_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
71 | } | |
72 | ||
73 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
74 | struct blk_mq_ctx *ctx) | |
75 | { | |
76 | struct blk_align_bitmap *bm = get_bm(hctx, ctx); | |
77 | ||
78 | clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word); | |
320ae51f JA |
79 | } |
80 | ||
b4c6a028 | 81 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 82 | { |
4ecd4fef | 83 | int freeze_depth; |
cddd5d17 | 84 | |
4ecd4fef CH |
85 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
86 | if (freeze_depth == 1) { | |
3ef28e83 | 87 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 88 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 89 | } |
f3af020b | 90 | } |
b4c6a028 | 91 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
92 | |
93 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
94 | { | |
3ef28e83 | 95 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
96 | } |
97 | ||
f3af020b TH |
98 | /* |
99 | * Guarantee no request is in use, so we can change any data structure of | |
100 | * the queue afterward. | |
101 | */ | |
3ef28e83 | 102 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 103 | { |
3ef28e83 DW |
104 | /* |
105 | * In the !blk_mq case we are only calling this to kill the | |
106 | * q_usage_counter, otherwise this increases the freeze depth | |
107 | * and waits for it to return to zero. For this reason there is | |
108 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
109 | * exported to drivers as the only user for unfreeze is blk_mq. | |
110 | */ | |
f3af020b TH |
111 | blk_mq_freeze_queue_start(q); |
112 | blk_mq_freeze_queue_wait(q); | |
113 | } | |
3ef28e83 DW |
114 | |
115 | void blk_mq_freeze_queue(struct request_queue *q) | |
116 | { | |
117 | /* | |
118 | * ...just an alias to keep freeze and unfreeze actions balanced | |
119 | * in the blk_mq_* namespace | |
120 | */ | |
121 | blk_freeze_queue(q); | |
122 | } | |
c761d96b | 123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 124 | |
b4c6a028 | 125 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 126 | { |
4ecd4fef | 127 | int freeze_depth; |
320ae51f | 128 | |
4ecd4fef CH |
129 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
130 | WARN_ON_ONCE(freeze_depth < 0); | |
131 | if (!freeze_depth) { | |
3ef28e83 | 132 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 133 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 134 | } |
320ae51f | 135 | } |
b4c6a028 | 136 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 137 | |
aed3ea94 JA |
138 | void blk_mq_wake_waiters(struct request_queue *q) |
139 | { | |
140 | struct blk_mq_hw_ctx *hctx; | |
141 | unsigned int i; | |
142 | ||
143 | queue_for_each_hw_ctx(q, hctx, i) | |
144 | if (blk_mq_hw_queue_mapped(hctx)) | |
145 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
146 | |
147 | /* | |
148 | * If we are called because the queue has now been marked as | |
149 | * dying, we need to ensure that processes currently waiting on | |
150 | * the queue are notified as well. | |
151 | */ | |
152 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
153 | } |
154 | ||
320ae51f JA |
155 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
156 | { | |
157 | return blk_mq_has_free_tags(hctx->tags); | |
158 | } | |
159 | EXPORT_SYMBOL(blk_mq_can_queue); | |
160 | ||
94eddfbe | 161 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
cc6e3b10 MC |
162 | struct request *rq, int op, |
163 | unsigned int op_flags) | |
320ae51f | 164 | { |
94eddfbe | 165 | if (blk_queue_io_stat(q)) |
cc6e3b10 | 166 | op_flags |= REQ_IO_STAT; |
94eddfbe | 167 | |
af76e555 CH |
168 | INIT_LIST_HEAD(&rq->queuelist); |
169 | /* csd/requeue_work/fifo_time is initialized before use */ | |
170 | rq->q = q; | |
320ae51f | 171 | rq->mq_ctx = ctx; |
cc6e3b10 | 172 | req_set_op_attrs(rq, op, op_flags); |
af76e555 CH |
173 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
174 | rq->cpu = -1; | |
af76e555 CH |
175 | INIT_HLIST_NODE(&rq->hash); |
176 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
177 | rq->rq_disk = NULL; |
178 | rq->part = NULL; | |
3ee32372 | 179 | rq->start_time = jiffies; |
af76e555 CH |
180 | #ifdef CONFIG_BLK_CGROUP |
181 | rq->rl = NULL; | |
0fec08b4 | 182 | set_start_time_ns(rq); |
af76e555 CH |
183 | rq->io_start_time_ns = 0; |
184 | #endif | |
185 | rq->nr_phys_segments = 0; | |
186 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
187 | rq->nr_integrity_segments = 0; | |
188 | #endif | |
af76e555 CH |
189 | rq->special = NULL; |
190 | /* tag was already set */ | |
191 | rq->errors = 0; | |
af76e555 | 192 | |
6f4a1626 TB |
193 | rq->cmd = rq->__cmd; |
194 | ||
af76e555 CH |
195 | rq->extra_len = 0; |
196 | rq->sense_len = 0; | |
197 | rq->resid_len = 0; | |
198 | rq->sense = NULL; | |
199 | ||
af76e555 | 200 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
201 | rq->timeout = 0; |
202 | ||
af76e555 CH |
203 | rq->end_io = NULL; |
204 | rq->end_io_data = NULL; | |
205 | rq->next_rq = NULL; | |
206 | ||
d9d8c5c4 | 207 | ctx->rq_dispatched[rw_is_sync(op, op_flags)]++; |
320ae51f JA |
208 | } |
209 | ||
5dee8577 | 210 | static struct request * |
cc6e3b10 | 211 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags) |
5dee8577 CH |
212 | { |
213 | struct request *rq; | |
214 | unsigned int tag; | |
215 | ||
cb96a42c | 216 | tag = blk_mq_get_tag(data); |
5dee8577 | 217 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 218 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 219 | |
cb96a42c | 220 | if (blk_mq_tag_busy(data->hctx)) { |
5dee8577 | 221 | rq->cmd_flags = REQ_MQ_INFLIGHT; |
cb96a42c | 222 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
223 | } |
224 | ||
225 | rq->tag = tag; | |
cc6e3b10 | 226 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags); |
5dee8577 CH |
227 | return rq; |
228 | } | |
229 | ||
230 | return NULL; | |
231 | } | |
232 | ||
6f3b0e8b CH |
233 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
234 | unsigned int flags) | |
320ae51f | 235 | { |
d852564f CH |
236 | struct blk_mq_ctx *ctx; |
237 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 238 | struct request *rq; |
cb96a42c | 239 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 240 | int ret; |
320ae51f | 241 | |
6f3b0e8b | 242 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
243 | if (ret) |
244 | return ERR_PTR(ret); | |
320ae51f | 245 | |
d852564f CH |
246 | ctx = blk_mq_get_ctx(q); |
247 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
6f3b0e8b | 248 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
d852564f | 249 | |
cc6e3b10 | 250 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
6f3b0e8b | 251 | if (!rq && !(flags & BLK_MQ_REQ_NOWAIT)) { |
d852564f CH |
252 | __blk_mq_run_hw_queue(hctx); |
253 | blk_mq_put_ctx(ctx); | |
254 | ||
255 | ctx = blk_mq_get_ctx(q); | |
256 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
6f3b0e8b | 257 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
cc6e3b10 | 258 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); |
cb96a42c | 259 | ctx = alloc_data.ctx; |
d852564f CH |
260 | } |
261 | blk_mq_put_ctx(ctx); | |
c76541a9 | 262 | if (!rq) { |
3ef28e83 | 263 | blk_queue_exit(q); |
a492f075 | 264 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 265 | } |
0c4de0f3 CH |
266 | |
267 | rq->__data_len = 0; | |
268 | rq->__sector = (sector_t) -1; | |
269 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
270 | return rq; |
271 | } | |
4bb659b1 | 272 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 273 | |
1f5bd336 ML |
274 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
275 | unsigned int flags, unsigned int hctx_idx) | |
276 | { | |
277 | struct blk_mq_hw_ctx *hctx; | |
278 | struct blk_mq_ctx *ctx; | |
279 | struct request *rq; | |
280 | struct blk_mq_alloc_data alloc_data; | |
281 | int ret; | |
282 | ||
283 | /* | |
284 | * If the tag allocator sleeps we could get an allocation for a | |
285 | * different hardware context. No need to complicate the low level | |
286 | * allocator for this for the rare use case of a command tied to | |
287 | * a specific queue. | |
288 | */ | |
289 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
290 | return ERR_PTR(-EINVAL); | |
291 | ||
292 | if (hctx_idx >= q->nr_hw_queues) | |
293 | return ERR_PTR(-EIO); | |
294 | ||
295 | ret = blk_queue_enter(q, true); | |
296 | if (ret) | |
297 | return ERR_PTR(ret); | |
298 | ||
299 | hctx = q->queue_hw_ctx[hctx_idx]; | |
300 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); | |
301 | ||
302 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
303 | rq = __blk_mq_alloc_request(&alloc_data, rw, 0); | |
304 | if (!rq) { | |
305 | blk_queue_exit(q); | |
306 | return ERR_PTR(-EWOULDBLOCK); | |
307 | } | |
308 | ||
309 | return rq; | |
310 | } | |
311 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
312 | ||
320ae51f JA |
313 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
314 | struct blk_mq_ctx *ctx, struct request *rq) | |
315 | { | |
316 | const int tag = rq->tag; | |
317 | struct request_queue *q = rq->q; | |
318 | ||
0d2602ca JA |
319 | if (rq->cmd_flags & REQ_MQ_INFLIGHT) |
320 | atomic_dec(&hctx->nr_active); | |
683d0e12 | 321 | rq->cmd_flags = 0; |
0d2602ca | 322 | |
af76e555 | 323 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
0d2602ca | 324 | blk_mq_put_tag(hctx, tag, &ctx->last_tag); |
3ef28e83 | 325 | blk_queue_exit(q); |
320ae51f JA |
326 | } |
327 | ||
7c7f2f2b | 328 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
329 | { |
330 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
331 | |
332 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 333 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
334 | |
335 | } | |
336 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
337 | ||
338 | void blk_mq_free_request(struct request *rq) | |
339 | { | |
340 | struct blk_mq_hw_ctx *hctx; | |
341 | struct request_queue *q = rq->q; | |
342 | ||
343 | hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu); | |
344 | blk_mq_free_hctx_request(hctx, rq); | |
320ae51f | 345 | } |
1a3b595a | 346 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 347 | |
c8a446ad | 348 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 349 | { |
0d11e6ac ML |
350 | blk_account_io_done(rq); |
351 | ||
91b63639 | 352 | if (rq->end_io) { |
320ae51f | 353 | rq->end_io(rq, error); |
91b63639 CH |
354 | } else { |
355 | if (unlikely(blk_bidi_rq(rq))) | |
356 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 357 | blk_mq_free_request(rq); |
91b63639 | 358 | } |
320ae51f | 359 | } |
c8a446ad | 360 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 361 | |
c8a446ad | 362 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
363 | { |
364 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
365 | BUG(); | |
c8a446ad | 366 | __blk_mq_end_request(rq, error); |
63151a44 | 367 | } |
c8a446ad | 368 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 369 | |
30a91cb4 | 370 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 371 | { |
3d6efbf6 | 372 | struct request *rq = data; |
320ae51f | 373 | |
30a91cb4 | 374 | rq->q->softirq_done_fn(rq); |
320ae51f | 375 | } |
320ae51f | 376 | |
ed851860 | 377 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
378 | { |
379 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 380 | bool shared = false; |
320ae51f JA |
381 | int cpu; |
382 | ||
38535201 | 383 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
384 | rq->q->softirq_done_fn(rq); |
385 | return; | |
386 | } | |
320ae51f JA |
387 | |
388 | cpu = get_cpu(); | |
38535201 CH |
389 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
390 | shared = cpus_share_cache(cpu, ctx->cpu); | |
391 | ||
392 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 393 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
394 | rq->csd.info = rq; |
395 | rq->csd.flags = 0; | |
c46fff2a | 396 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 397 | } else { |
30a91cb4 | 398 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 399 | } |
320ae51f JA |
400 | put_cpu(); |
401 | } | |
30a91cb4 | 402 | |
1fa8cc52 | 403 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
404 | { |
405 | struct request_queue *q = rq->q; | |
406 | ||
407 | if (!q->softirq_done_fn) | |
c8a446ad | 408 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
409 | else |
410 | blk_mq_ipi_complete_request(rq); | |
411 | } | |
412 | ||
30a91cb4 CH |
413 | /** |
414 | * blk_mq_complete_request - end I/O on a request | |
415 | * @rq: the request being processed | |
416 | * | |
417 | * Description: | |
418 | * Ends all I/O on a request. It does not handle partial completions. | |
419 | * The actual completion happens out-of-order, through a IPI handler. | |
420 | **/ | |
f4829a9b | 421 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 422 | { |
95f09684 JA |
423 | struct request_queue *q = rq->q; |
424 | ||
425 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 426 | return; |
f4829a9b CH |
427 | if (!blk_mark_rq_complete(rq)) { |
428 | rq->errors = error; | |
ed851860 | 429 | __blk_mq_complete_request(rq); |
f4829a9b | 430 | } |
30a91cb4 CH |
431 | } |
432 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 433 | |
973c0191 KB |
434 | int blk_mq_request_started(struct request *rq) |
435 | { | |
436 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
437 | } | |
438 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
439 | ||
e2490073 | 440 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
441 | { |
442 | struct request_queue *q = rq->q; | |
443 | ||
444 | trace_block_rq_issue(q, rq); | |
445 | ||
742ee69b | 446 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
447 | if (unlikely(blk_bidi_rq(rq))) |
448 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 449 | |
2b8393b4 | 450 | blk_add_timer(rq); |
87ee7b11 | 451 | |
538b7534 JA |
452 | /* |
453 | * Ensure that ->deadline is visible before set the started | |
454 | * flag and clear the completed flag. | |
455 | */ | |
456 | smp_mb__before_atomic(); | |
457 | ||
87ee7b11 JA |
458 | /* |
459 | * Mark us as started and clear complete. Complete might have been | |
460 | * set if requeue raced with timeout, which then marked it as | |
461 | * complete. So be sure to clear complete again when we start | |
462 | * the request, otherwise we'll ignore the completion event. | |
463 | */ | |
4b570521 JA |
464 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
465 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
466 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
467 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
468 | |
469 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
470 | /* | |
471 | * Make sure space for the drain appears. We know we can do | |
472 | * this because max_hw_segments has been adjusted to be one | |
473 | * fewer than the device can handle. | |
474 | */ | |
475 | rq->nr_phys_segments++; | |
476 | } | |
320ae51f | 477 | } |
e2490073 | 478 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 479 | |
ed0791b2 | 480 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
481 | { |
482 | struct request_queue *q = rq->q; | |
483 | ||
484 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 485 | |
e2490073 CH |
486 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
487 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
488 | rq->nr_phys_segments--; | |
489 | } | |
320ae51f JA |
490 | } |
491 | ||
ed0791b2 CH |
492 | void blk_mq_requeue_request(struct request *rq) |
493 | { | |
ed0791b2 | 494 | __blk_mq_requeue_request(rq); |
ed0791b2 | 495 | |
ed0791b2 | 496 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 497 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
498 | } |
499 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
500 | ||
6fca6a61 CH |
501 | static void blk_mq_requeue_work(struct work_struct *work) |
502 | { | |
503 | struct request_queue *q = | |
504 | container_of(work, struct request_queue, requeue_work); | |
505 | LIST_HEAD(rq_list); | |
506 | struct request *rq, *next; | |
507 | unsigned long flags; | |
508 | ||
509 | spin_lock_irqsave(&q->requeue_lock, flags); | |
510 | list_splice_init(&q->requeue_list, &rq_list); | |
511 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
512 | ||
513 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
514 | if (!(rq->cmd_flags & REQ_SOFTBARRIER)) | |
515 | continue; | |
516 | ||
517 | rq->cmd_flags &= ~REQ_SOFTBARRIER; | |
518 | list_del_init(&rq->queuelist); | |
519 | blk_mq_insert_request(rq, true, false, false); | |
520 | } | |
521 | ||
522 | while (!list_empty(&rq_list)) { | |
523 | rq = list_entry(rq_list.next, struct request, queuelist); | |
524 | list_del_init(&rq->queuelist); | |
525 | blk_mq_insert_request(rq, false, false, false); | |
526 | } | |
527 | ||
8b957415 JA |
528 | /* |
529 | * Use the start variant of queue running here, so that running | |
530 | * the requeue work will kick stopped queues. | |
531 | */ | |
532 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
533 | } |
534 | ||
535 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
536 | { | |
537 | struct request_queue *q = rq->q; | |
538 | unsigned long flags; | |
539 | ||
540 | /* | |
541 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
542 | * request head insertation from the workqueue. | |
543 | */ | |
544 | BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER); | |
545 | ||
546 | spin_lock_irqsave(&q->requeue_lock, flags); | |
547 | if (at_head) { | |
548 | rq->cmd_flags |= REQ_SOFTBARRIER; | |
549 | list_add(&rq->queuelist, &q->requeue_list); | |
550 | } else { | |
551 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
552 | } | |
553 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
554 | } | |
555 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
556 | ||
c68ed59f KB |
557 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
558 | { | |
559 | cancel_work_sync(&q->requeue_work); | |
560 | } | |
561 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
562 | ||
6fca6a61 CH |
563 | void blk_mq_kick_requeue_list(struct request_queue *q) |
564 | { | |
565 | kblockd_schedule_work(&q->requeue_work); | |
566 | } | |
567 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
568 | ||
1885b24d JA |
569 | void blk_mq_abort_requeue_list(struct request_queue *q) |
570 | { | |
571 | unsigned long flags; | |
572 | LIST_HEAD(rq_list); | |
573 | ||
574 | spin_lock_irqsave(&q->requeue_lock, flags); | |
575 | list_splice_init(&q->requeue_list, &rq_list); | |
576 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
577 | ||
578 | while (!list_empty(&rq_list)) { | |
579 | struct request *rq; | |
580 | ||
581 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
582 | list_del_init(&rq->queuelist); | |
583 | rq->errors = -EIO; | |
584 | blk_mq_end_request(rq, rq->errors); | |
585 | } | |
586 | } | |
587 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
588 | ||
0e62f51f JA |
589 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
590 | { | |
4ee86bab HR |
591 | if (tag < tags->nr_tags) |
592 | return tags->rqs[tag]; | |
593 | ||
594 | return NULL; | |
24d2f903 CH |
595 | } |
596 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
597 | ||
320ae51f | 598 | struct blk_mq_timeout_data { |
46f92d42 CH |
599 | unsigned long next; |
600 | unsigned int next_set; | |
320ae51f JA |
601 | }; |
602 | ||
90415837 | 603 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 604 | { |
46f92d42 CH |
605 | struct blk_mq_ops *ops = req->q->mq_ops; |
606 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
607 | |
608 | /* | |
609 | * We know that complete is set at this point. If STARTED isn't set | |
610 | * anymore, then the request isn't active and the "timeout" should | |
611 | * just be ignored. This can happen due to the bitflag ordering. | |
612 | * Timeout first checks if STARTED is set, and if it is, assumes | |
613 | * the request is active. But if we race with completion, then | |
614 | * we both flags will get cleared. So check here again, and ignore | |
615 | * a timeout event with a request that isn't active. | |
616 | */ | |
46f92d42 CH |
617 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
618 | return; | |
87ee7b11 | 619 | |
46f92d42 | 620 | if (ops->timeout) |
0152fb6b | 621 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
622 | |
623 | switch (ret) { | |
624 | case BLK_EH_HANDLED: | |
625 | __blk_mq_complete_request(req); | |
626 | break; | |
627 | case BLK_EH_RESET_TIMER: | |
628 | blk_add_timer(req); | |
629 | blk_clear_rq_complete(req); | |
630 | break; | |
631 | case BLK_EH_NOT_HANDLED: | |
632 | break; | |
633 | default: | |
634 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
635 | break; | |
636 | } | |
87ee7b11 | 637 | } |
5b3f25fc | 638 | |
81481eb4 CH |
639 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
640 | struct request *rq, void *priv, bool reserved) | |
641 | { | |
642 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 643 | |
eb130dbf KB |
644 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
645 | /* | |
646 | * If a request wasn't started before the queue was | |
647 | * marked dying, kill it here or it'll go unnoticed. | |
648 | */ | |
a59e0f57 KB |
649 | if (unlikely(blk_queue_dying(rq->q))) { |
650 | rq->errors = -EIO; | |
651 | blk_mq_end_request(rq, rq->errors); | |
652 | } | |
46f92d42 | 653 | return; |
eb130dbf | 654 | } |
87ee7b11 | 655 | |
46f92d42 CH |
656 | if (time_after_eq(jiffies, rq->deadline)) { |
657 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 658 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
659 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
660 | data->next = rq->deadline; | |
661 | data->next_set = 1; | |
662 | } | |
87ee7b11 JA |
663 | } |
664 | ||
287922eb | 665 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 666 | { |
287922eb CH |
667 | struct request_queue *q = |
668 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
669 | struct blk_mq_timeout_data data = { |
670 | .next = 0, | |
671 | .next_set = 0, | |
672 | }; | |
81481eb4 | 673 | int i; |
320ae51f | 674 | |
71f79fb3 GKB |
675 | /* A deadlock might occur if a request is stuck requiring a |
676 | * timeout at the same time a queue freeze is waiting | |
677 | * completion, since the timeout code would not be able to | |
678 | * acquire the queue reference here. | |
679 | * | |
680 | * That's why we don't use blk_queue_enter here; instead, we use | |
681 | * percpu_ref_tryget directly, because we need to be able to | |
682 | * obtain a reference even in the short window between the queue | |
683 | * starting to freeze, by dropping the first reference in | |
684 | * blk_mq_freeze_queue_start, and the moment the last request is | |
685 | * consumed, marked by the instant q_usage_counter reaches | |
686 | * zero. | |
687 | */ | |
688 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
689 | return; |
690 | ||
0bf6cd5b | 691 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 692 | |
81481eb4 CH |
693 | if (data.next_set) { |
694 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
695 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 696 | } else { |
0bf6cd5b CH |
697 | struct blk_mq_hw_ctx *hctx; |
698 | ||
f054b56c ML |
699 | queue_for_each_hw_ctx(q, hctx, i) { |
700 | /* the hctx may be unmapped, so check it here */ | |
701 | if (blk_mq_hw_queue_mapped(hctx)) | |
702 | blk_mq_tag_idle(hctx); | |
703 | } | |
0d2602ca | 704 | } |
287922eb | 705 | blk_queue_exit(q); |
320ae51f JA |
706 | } |
707 | ||
708 | /* | |
709 | * Reverse check our software queue for entries that we could potentially | |
710 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
711 | * too much time checking for merges. | |
712 | */ | |
713 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
714 | struct blk_mq_ctx *ctx, struct bio *bio) | |
715 | { | |
716 | struct request *rq; | |
717 | int checked = 8; | |
718 | ||
719 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
720 | int el_ret; | |
721 | ||
722 | if (!checked--) | |
723 | break; | |
724 | ||
725 | if (!blk_rq_merge_ok(rq, bio)) | |
726 | continue; | |
727 | ||
728 | el_ret = blk_try_merge(rq, bio); | |
729 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
730 | if (bio_attempt_back_merge(q, rq, bio)) { | |
731 | ctx->rq_merged++; | |
732 | return true; | |
733 | } | |
734 | break; | |
735 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
736 | if (bio_attempt_front_merge(q, rq, bio)) { | |
737 | ctx->rq_merged++; | |
738 | return true; | |
739 | } | |
740 | break; | |
741 | } | |
742 | } | |
743 | ||
744 | return false; | |
745 | } | |
746 | ||
1429d7c9 JA |
747 | /* |
748 | * Process software queues that have been marked busy, splicing them | |
749 | * to the for-dispatch | |
750 | */ | |
751 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
752 | { | |
753 | struct blk_mq_ctx *ctx; | |
754 | int i; | |
755 | ||
569fd0ce | 756 | for (i = 0; i < hctx->ctx_map.size; i++) { |
1429d7c9 JA |
757 | struct blk_align_bitmap *bm = &hctx->ctx_map.map[i]; |
758 | unsigned int off, bit; | |
759 | ||
760 | if (!bm->word) | |
761 | continue; | |
762 | ||
763 | bit = 0; | |
764 | off = i * hctx->ctx_map.bits_per_word; | |
765 | do { | |
766 | bit = find_next_bit(&bm->word, bm->depth, bit); | |
767 | if (bit >= bm->depth) | |
768 | break; | |
769 | ||
770 | ctx = hctx->ctxs[bit + off]; | |
771 | clear_bit(bit, &bm->word); | |
772 | spin_lock(&ctx->lock); | |
773 | list_splice_tail_init(&ctx->rq_list, list); | |
774 | spin_unlock(&ctx->lock); | |
775 | ||
776 | bit++; | |
777 | } while (1); | |
778 | } | |
779 | } | |
780 | ||
320ae51f JA |
781 | /* |
782 | * Run this hardware queue, pulling any software queues mapped to it in. | |
783 | * Note that this function currently has various problems around ordering | |
784 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
785 | * items on the hctx->dispatch list. Ignore that for now. | |
786 | */ | |
787 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
788 | { | |
789 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
790 | struct request *rq; |
791 | LIST_HEAD(rq_list); | |
74c45052 JA |
792 | LIST_HEAD(driver_list); |
793 | struct list_head *dptr; | |
1429d7c9 | 794 | int queued; |
320ae51f | 795 | |
5d12f905 | 796 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
797 | return; |
798 | ||
0e87e58b JA |
799 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
800 | cpu_online(hctx->next_cpu)); | |
801 | ||
320ae51f JA |
802 | hctx->run++; |
803 | ||
804 | /* | |
805 | * Touch any software queue that has pending entries. | |
806 | */ | |
1429d7c9 | 807 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
808 | |
809 | /* | |
810 | * If we have previous entries on our dispatch list, grab them | |
811 | * and stuff them at the front for more fair dispatch. | |
812 | */ | |
813 | if (!list_empty_careful(&hctx->dispatch)) { | |
814 | spin_lock(&hctx->lock); | |
815 | if (!list_empty(&hctx->dispatch)) | |
816 | list_splice_init(&hctx->dispatch, &rq_list); | |
817 | spin_unlock(&hctx->lock); | |
818 | } | |
819 | ||
74c45052 JA |
820 | /* |
821 | * Start off with dptr being NULL, so we start the first request | |
822 | * immediately, even if we have more pending. | |
823 | */ | |
824 | dptr = NULL; | |
825 | ||
320ae51f JA |
826 | /* |
827 | * Now process all the entries, sending them to the driver. | |
828 | */ | |
1429d7c9 | 829 | queued = 0; |
320ae51f | 830 | while (!list_empty(&rq_list)) { |
74c45052 | 831 | struct blk_mq_queue_data bd; |
320ae51f JA |
832 | int ret; |
833 | ||
834 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
835 | list_del_init(&rq->queuelist); | |
320ae51f | 836 | |
74c45052 JA |
837 | bd.rq = rq; |
838 | bd.list = dptr; | |
839 | bd.last = list_empty(&rq_list); | |
840 | ||
841 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
842 | switch (ret) { |
843 | case BLK_MQ_RQ_QUEUE_OK: | |
844 | queued++; | |
52b9c330 | 845 | break; |
320ae51f | 846 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 847 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 848 | __blk_mq_requeue_request(rq); |
320ae51f JA |
849 | break; |
850 | default: | |
851 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 852 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 853 | rq->errors = -EIO; |
c8a446ad | 854 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
855 | break; |
856 | } | |
857 | ||
858 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
859 | break; | |
74c45052 JA |
860 | |
861 | /* | |
862 | * We've done the first request. If we have more than 1 | |
863 | * left in the list, set dptr to defer issue. | |
864 | */ | |
865 | if (!dptr && rq_list.next != rq_list.prev) | |
866 | dptr = &driver_list; | |
320ae51f JA |
867 | } |
868 | ||
869 | if (!queued) | |
870 | hctx->dispatched[0]++; | |
871 | else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1))) | |
872 | hctx->dispatched[ilog2(queued) + 1]++; | |
873 | ||
874 | /* | |
875 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
876 | * that is where we will continue on next queue run. | |
877 | */ | |
878 | if (!list_empty(&rq_list)) { | |
879 | spin_lock(&hctx->lock); | |
880 | list_splice(&rq_list, &hctx->dispatch); | |
881 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
882 | /* |
883 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
884 | * it's possible the queue is stopped and restarted again | |
885 | * before this. Queue restart will dispatch requests. And since | |
886 | * requests in rq_list aren't added into hctx->dispatch yet, | |
887 | * the requests in rq_list might get lost. | |
888 | * | |
889 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
890 | **/ | |
891 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
892 | } |
893 | } | |
894 | ||
506e931f JA |
895 | /* |
896 | * It'd be great if the workqueue API had a way to pass | |
897 | * in a mask and had some smarts for more clever placement. | |
898 | * For now we just round-robin here, switching for every | |
899 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
900 | */ | |
901 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
902 | { | |
b657d7e6 CH |
903 | if (hctx->queue->nr_hw_queues == 1) |
904 | return WORK_CPU_UNBOUND; | |
506e931f JA |
905 | |
906 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 907 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
908 | |
909 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
910 | if (next_cpu >= nr_cpu_ids) | |
911 | next_cpu = cpumask_first(hctx->cpumask); | |
912 | ||
913 | hctx->next_cpu = next_cpu; | |
914 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
915 | |
916 | return cpu; | |
506e931f JA |
917 | } |
918 | ||
b657d7e6 | 919 | return hctx->next_cpu; |
506e931f JA |
920 | } |
921 | ||
320ae51f JA |
922 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
923 | { | |
19c66e59 ML |
924 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
925 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
926 | return; |
927 | ||
398205b8 | 928 | if (!async) { |
2a90d4aa PB |
929 | int cpu = get_cpu(); |
930 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 931 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 932 | put_cpu(); |
398205b8 PB |
933 | return; |
934 | } | |
e4043dcf | 935 | |
2a90d4aa | 936 | put_cpu(); |
e4043dcf | 937 | } |
398205b8 | 938 | |
b657d7e6 CH |
939 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
940 | &hctx->run_work, 0); | |
320ae51f JA |
941 | } |
942 | ||
b94ec296 | 943 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
944 | { |
945 | struct blk_mq_hw_ctx *hctx; | |
946 | int i; | |
947 | ||
948 | queue_for_each_hw_ctx(q, hctx, i) { | |
949 | if ((!blk_mq_hctx_has_pending(hctx) && | |
950 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 951 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
952 | continue; |
953 | ||
b94ec296 | 954 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
955 | } |
956 | } | |
b94ec296 | 957 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f JA |
958 | |
959 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
960 | { | |
70f4db63 CH |
961 | cancel_delayed_work(&hctx->run_work); |
962 | cancel_delayed_work(&hctx->delay_work); | |
320ae51f JA |
963 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
964 | } | |
965 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
966 | ||
280d45f6 CH |
967 | void blk_mq_stop_hw_queues(struct request_queue *q) |
968 | { | |
969 | struct blk_mq_hw_ctx *hctx; | |
970 | int i; | |
971 | ||
972 | queue_for_each_hw_ctx(q, hctx, i) | |
973 | blk_mq_stop_hw_queue(hctx); | |
974 | } | |
975 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
976 | ||
320ae51f JA |
977 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
978 | { | |
979 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 980 | |
0ffbce80 | 981 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
982 | } |
983 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
984 | ||
2f268556 CH |
985 | void blk_mq_start_hw_queues(struct request_queue *q) |
986 | { | |
987 | struct blk_mq_hw_ctx *hctx; | |
988 | int i; | |
989 | ||
990 | queue_for_each_hw_ctx(q, hctx, i) | |
991 | blk_mq_start_hw_queue(hctx); | |
992 | } | |
993 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
994 | ||
1b4a3258 | 995 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
996 | { |
997 | struct blk_mq_hw_ctx *hctx; | |
998 | int i; | |
999 | ||
1000 | queue_for_each_hw_ctx(q, hctx, i) { | |
1001 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1002 | continue; | |
1003 | ||
1004 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1005 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1006 | } |
1007 | } | |
1008 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1009 | ||
70f4db63 | 1010 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1011 | { |
1012 | struct blk_mq_hw_ctx *hctx; | |
1013 | ||
70f4db63 | 1014 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
e4043dcf | 1015 | |
320ae51f JA |
1016 | __blk_mq_run_hw_queue(hctx); |
1017 | } | |
1018 | ||
70f4db63 CH |
1019 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1020 | { | |
1021 | struct blk_mq_hw_ctx *hctx; | |
1022 | ||
1023 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1024 | ||
1025 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1026 | __blk_mq_run_hw_queue(hctx); | |
1027 | } | |
1028 | ||
1029 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1030 | { | |
19c66e59 ML |
1031 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1032 | return; | |
70f4db63 | 1033 | |
b657d7e6 CH |
1034 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1035 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1036 | } |
1037 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1038 | ||
cfd0c552 | 1039 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1040 | struct request *rq, |
1041 | bool at_head) | |
320ae51f | 1042 | { |
e57690fe JA |
1043 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1044 | ||
01b983c9 JA |
1045 | trace_block_rq_insert(hctx->queue, rq); |
1046 | ||
72a0a36e CH |
1047 | if (at_head) |
1048 | list_add(&rq->queuelist, &ctx->rq_list); | |
1049 | else | |
1050 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1051 | } |
4bb659b1 | 1052 | |
cfd0c552 ML |
1053 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1054 | struct request *rq, bool at_head) | |
1055 | { | |
1056 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1057 | ||
e57690fe | 1058 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1059 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1060 | } |
1061 | ||
eeabc850 | 1062 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1063 | bool async) |
320ae51f | 1064 | { |
e57690fe | 1065 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1066 | struct request_queue *q = rq->q; |
320ae51f | 1067 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1068 | |
320ae51f JA |
1069 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1070 | ||
a57a178a CH |
1071 | spin_lock(&ctx->lock); |
1072 | __blk_mq_insert_request(hctx, rq, at_head); | |
1073 | spin_unlock(&ctx->lock); | |
320ae51f | 1074 | |
320ae51f JA |
1075 | if (run_queue) |
1076 | blk_mq_run_hw_queue(hctx, async); | |
1077 | } | |
1078 | ||
1079 | static void blk_mq_insert_requests(struct request_queue *q, | |
1080 | struct blk_mq_ctx *ctx, | |
1081 | struct list_head *list, | |
1082 | int depth, | |
1083 | bool from_schedule) | |
1084 | ||
1085 | { | |
1086 | struct blk_mq_hw_ctx *hctx; | |
320ae51f JA |
1087 | |
1088 | trace_block_unplug(q, depth, !from_schedule); | |
1089 | ||
320ae51f JA |
1090 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
1091 | ||
1092 | /* | |
1093 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1094 | * offline now | |
1095 | */ | |
1096 | spin_lock(&ctx->lock); | |
1097 | while (!list_empty(list)) { | |
1098 | struct request *rq; | |
1099 | ||
1100 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1101 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1102 | list_del_init(&rq->queuelist); |
e57690fe | 1103 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1104 | } |
cfd0c552 | 1105 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1106 | spin_unlock(&ctx->lock); |
1107 | ||
320ae51f JA |
1108 | blk_mq_run_hw_queue(hctx, from_schedule); |
1109 | } | |
1110 | ||
1111 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1112 | { | |
1113 | struct request *rqa = container_of(a, struct request, queuelist); | |
1114 | struct request *rqb = container_of(b, struct request, queuelist); | |
1115 | ||
1116 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1117 | (rqa->mq_ctx == rqb->mq_ctx && | |
1118 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1119 | } | |
1120 | ||
1121 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1122 | { | |
1123 | struct blk_mq_ctx *this_ctx; | |
1124 | struct request_queue *this_q; | |
1125 | struct request *rq; | |
1126 | LIST_HEAD(list); | |
1127 | LIST_HEAD(ctx_list); | |
1128 | unsigned int depth; | |
1129 | ||
1130 | list_splice_init(&plug->mq_list, &list); | |
1131 | ||
1132 | list_sort(NULL, &list, plug_ctx_cmp); | |
1133 | ||
1134 | this_q = NULL; | |
1135 | this_ctx = NULL; | |
1136 | depth = 0; | |
1137 | ||
1138 | while (!list_empty(&list)) { | |
1139 | rq = list_entry_rq(list.next); | |
1140 | list_del_init(&rq->queuelist); | |
1141 | BUG_ON(!rq->q); | |
1142 | if (rq->mq_ctx != this_ctx) { | |
1143 | if (this_ctx) { | |
1144 | blk_mq_insert_requests(this_q, this_ctx, | |
1145 | &ctx_list, depth, | |
1146 | from_schedule); | |
1147 | } | |
1148 | ||
1149 | this_ctx = rq->mq_ctx; | |
1150 | this_q = rq->q; | |
1151 | depth = 0; | |
1152 | } | |
1153 | ||
1154 | depth++; | |
1155 | list_add_tail(&rq->queuelist, &ctx_list); | |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * If 'this_ctx' is set, we know we have entries to complete | |
1160 | * on 'ctx_list'. Do those. | |
1161 | */ | |
1162 | if (this_ctx) { | |
1163 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1164 | from_schedule); | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1169 | { | |
1170 | init_request_from_bio(rq, bio); | |
4b570521 | 1171 | |
a21f2a3e | 1172 | blk_account_io_start(rq, 1); |
320ae51f JA |
1173 | } |
1174 | ||
274a5843 JA |
1175 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1176 | { | |
1177 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1178 | !blk_queue_nomerges(hctx->queue); | |
1179 | } | |
1180 | ||
07068d5b JA |
1181 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1182 | struct blk_mq_ctx *ctx, | |
1183 | struct request *rq, struct bio *bio) | |
320ae51f | 1184 | { |
e18378a6 | 1185 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1186 | blk_mq_bio_to_request(rq, bio); |
1187 | spin_lock(&ctx->lock); | |
1188 | insert_rq: | |
1189 | __blk_mq_insert_request(hctx, rq, false); | |
1190 | spin_unlock(&ctx->lock); | |
1191 | return false; | |
1192 | } else { | |
274a5843 JA |
1193 | struct request_queue *q = hctx->queue; |
1194 | ||
07068d5b JA |
1195 | spin_lock(&ctx->lock); |
1196 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1197 | blk_mq_bio_to_request(rq, bio); | |
1198 | goto insert_rq; | |
1199 | } | |
320ae51f | 1200 | |
07068d5b JA |
1201 | spin_unlock(&ctx->lock); |
1202 | __blk_mq_free_request(hctx, ctx, rq); | |
1203 | return true; | |
14ec77f3 | 1204 | } |
07068d5b | 1205 | } |
14ec77f3 | 1206 | |
07068d5b JA |
1207 | struct blk_map_ctx { |
1208 | struct blk_mq_hw_ctx *hctx; | |
1209 | struct blk_mq_ctx *ctx; | |
1210 | }; | |
1211 | ||
1212 | static struct request *blk_mq_map_request(struct request_queue *q, | |
1213 | struct bio *bio, | |
1214 | struct blk_map_ctx *data) | |
1215 | { | |
1216 | struct blk_mq_hw_ctx *hctx; | |
1217 | struct blk_mq_ctx *ctx; | |
1218 | struct request *rq; | |
cc6e3b10 MC |
1219 | int op = bio_data_dir(bio); |
1220 | int op_flags = 0; | |
cb96a42c | 1221 | struct blk_mq_alloc_data alloc_data; |
320ae51f | 1222 | |
3ef28e83 | 1223 | blk_queue_enter_live(q); |
320ae51f JA |
1224 | ctx = blk_mq_get_ctx(q); |
1225 | hctx = q->mq_ops->map_queue(q, ctx->cpu); | |
1226 | ||
1eff9d32 | 1227 | if (rw_is_sync(bio_op(bio), bio->bi_opf)) |
cc6e3b10 | 1228 | op_flags |= REQ_SYNC; |
07068d5b | 1229 | |
cc6e3b10 | 1230 | trace_block_getrq(q, bio, op); |
6f3b0e8b | 1231 | blk_mq_set_alloc_data(&alloc_data, q, BLK_MQ_REQ_NOWAIT, ctx, hctx); |
cc6e3b10 | 1232 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
5dee8577 | 1233 | if (unlikely(!rq)) { |
793597a6 | 1234 | __blk_mq_run_hw_queue(hctx); |
320ae51f | 1235 | blk_mq_put_ctx(ctx); |
cc6e3b10 | 1236 | trace_block_sleeprq(q, bio, op); |
793597a6 CH |
1237 | |
1238 | ctx = blk_mq_get_ctx(q); | |
320ae51f | 1239 | hctx = q->mq_ops->map_queue(q, ctx->cpu); |
6f3b0e8b | 1240 | blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx); |
cc6e3b10 | 1241 | rq = __blk_mq_alloc_request(&alloc_data, op, op_flags); |
cb96a42c ML |
1242 | ctx = alloc_data.ctx; |
1243 | hctx = alloc_data.hctx; | |
320ae51f JA |
1244 | } |
1245 | ||
1246 | hctx->queued++; | |
07068d5b JA |
1247 | data->hctx = hctx; |
1248 | data->ctx = ctx; | |
1249 | return rq; | |
1250 | } | |
1251 | ||
7b371636 | 1252 | static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1253 | { |
1254 | int ret; | |
1255 | struct request_queue *q = rq->q; | |
1256 | struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, | |
1257 | rq->mq_ctx->cpu); | |
1258 | struct blk_mq_queue_data bd = { | |
1259 | .rq = rq, | |
1260 | .list = NULL, | |
1261 | .last = 1 | |
1262 | }; | |
7b371636 | 1263 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f SL |
1264 | |
1265 | /* | |
1266 | * For OK queue, we are done. For error, kill it. Any other | |
1267 | * error (busy), just add it to our list as we previously | |
1268 | * would have done | |
1269 | */ | |
1270 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1271 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1272 | *cookie = new_cookie; | |
f984df1f | 1273 | return 0; |
7b371636 | 1274 | } |
f984df1f | 1275 | |
7b371636 JA |
1276 | __blk_mq_requeue_request(rq); |
1277 | ||
1278 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1279 | *cookie = BLK_QC_T_NONE; | |
1280 | rq->errors = -EIO; | |
1281 | blk_mq_end_request(rq, rq->errors); | |
1282 | return 0; | |
f984df1f | 1283 | } |
7b371636 JA |
1284 | |
1285 | return -1; | |
f984df1f SL |
1286 | } |
1287 | ||
07068d5b JA |
1288 | /* |
1289 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1290 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1291 | * hardware for SYNC IO. | |
1292 | */ | |
dece1635 | 1293 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1294 | { |
1eff9d32 JA |
1295 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1296 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
07068d5b JA |
1297 | struct blk_map_ctx data; |
1298 | struct request *rq; | |
f984df1f SL |
1299 | unsigned int request_count = 0; |
1300 | struct blk_plug *plug; | |
5b3f341f | 1301 | struct request *same_queue_rq = NULL; |
7b371636 | 1302 | blk_qc_t cookie; |
07068d5b JA |
1303 | |
1304 | blk_queue_bounce(q, &bio); | |
1305 | ||
1306 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1307 | bio_io_error(bio); |
dece1635 | 1308 | return BLK_QC_T_NONE; |
07068d5b JA |
1309 | } |
1310 | ||
54efd50b KO |
1311 | blk_queue_split(q, &bio, q->bio_split); |
1312 | ||
87c279e6 OS |
1313 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1314 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1315 | return BLK_QC_T_NONE; | |
f984df1f | 1316 | |
07068d5b JA |
1317 | rq = blk_mq_map_request(q, bio, &data); |
1318 | if (unlikely(!rq)) | |
dece1635 | 1319 | return BLK_QC_T_NONE; |
07068d5b | 1320 | |
7b371636 | 1321 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1322 | |
1323 | if (unlikely(is_flush_fua)) { | |
1324 | blk_mq_bio_to_request(rq, bio); | |
1325 | blk_insert_flush(rq); | |
1326 | goto run_queue; | |
1327 | } | |
1328 | ||
f984df1f | 1329 | plug = current->plug; |
e167dfb5 JA |
1330 | /* |
1331 | * If the driver supports defer issued based on 'last', then | |
1332 | * queue it up like normal since we can potentially save some | |
1333 | * CPU this way. | |
1334 | */ | |
f984df1f SL |
1335 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1336 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1337 | struct request *old_rq = NULL; | |
07068d5b JA |
1338 | |
1339 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1340 | |
1341 | /* | |
b094f89c | 1342 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1343 | * Otherwise the existing request in the plug list will be |
1344 | * issued. So the plug list will have one request at most | |
07068d5b | 1345 | */ |
f984df1f | 1346 | if (plug) { |
5b3f341f SL |
1347 | /* |
1348 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1349 | * happens, same_queue_rq is invalid and plug list is |
1350 | * empty | |
1351 | */ | |
5b3f341f SL |
1352 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1353 | old_rq = same_queue_rq; | |
f984df1f | 1354 | list_del_init(&old_rq->queuelist); |
07068d5b | 1355 | } |
f984df1f SL |
1356 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1357 | } else /* is_sync */ | |
1358 | old_rq = rq; | |
1359 | blk_mq_put_ctx(data.ctx); | |
1360 | if (!old_rq) | |
7b371636 JA |
1361 | goto done; |
1362 | if (!blk_mq_direct_issue_request(old_rq, &cookie)) | |
1363 | goto done; | |
f984df1f | 1364 | blk_mq_insert_request(old_rq, false, true, true); |
7b371636 | 1365 | goto done; |
07068d5b JA |
1366 | } |
1367 | ||
1368 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1369 | /* | |
1370 | * For a SYNC request, send it to the hardware immediately. For | |
1371 | * an ASYNC request, just ensure that we run it later on. The | |
1372 | * latter allows for merging opportunities and more efficient | |
1373 | * dispatching. | |
1374 | */ | |
1375 | run_queue: | |
1376 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1377 | } | |
07068d5b | 1378 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1379 | done: |
1380 | return cookie; | |
07068d5b JA |
1381 | } |
1382 | ||
1383 | /* | |
1384 | * Single hardware queue variant. This will attempt to use any per-process | |
1385 | * plug for merging and IO deferral. | |
1386 | */ | |
dece1635 | 1387 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1388 | { |
1eff9d32 JA |
1389 | const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf); |
1390 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); | |
e6c4438b JM |
1391 | struct blk_plug *plug; |
1392 | unsigned int request_count = 0; | |
07068d5b JA |
1393 | struct blk_map_ctx data; |
1394 | struct request *rq; | |
7b371636 | 1395 | blk_qc_t cookie; |
07068d5b | 1396 | |
07068d5b JA |
1397 | blk_queue_bounce(q, &bio); |
1398 | ||
1399 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1400 | bio_io_error(bio); |
dece1635 | 1401 | return BLK_QC_T_NONE; |
07068d5b JA |
1402 | } |
1403 | ||
54efd50b KO |
1404 | blk_queue_split(q, &bio, q->bio_split); |
1405 | ||
87c279e6 OS |
1406 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1407 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1408 | return BLK_QC_T_NONE; | |
1409 | } else | |
1410 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1411 | |
1412 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1413 | if (unlikely(!rq)) |
dece1635 | 1414 | return BLK_QC_T_NONE; |
320ae51f | 1415 | |
7b371636 | 1416 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1417 | |
1418 | if (unlikely(is_flush_fua)) { | |
1419 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1420 | blk_insert_flush(rq); |
1421 | goto run_queue; | |
1422 | } | |
1423 | ||
1424 | /* | |
1425 | * A task plug currently exists. Since this is completely lockless, | |
1426 | * utilize that to temporarily store requests until the task is | |
1427 | * either done or scheduled away. | |
1428 | */ | |
e6c4438b JM |
1429 | plug = current->plug; |
1430 | if (plug) { | |
1431 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1432 | if (!request_count) |
e6c4438b | 1433 | trace_block_plug(q); |
b094f89c JA |
1434 | |
1435 | blk_mq_put_ctx(data.ctx); | |
1436 | ||
1437 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1438 | blk_flush_plug_list(plug, false); |
1439 | trace_block_plug(q); | |
320ae51f | 1440 | } |
b094f89c | 1441 | |
e6c4438b | 1442 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1443 | return cookie; |
320ae51f JA |
1444 | } |
1445 | ||
07068d5b JA |
1446 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1447 | /* | |
1448 | * For a SYNC request, send it to the hardware immediately. For | |
1449 | * an ASYNC request, just ensure that we run it later on. The | |
1450 | * latter allows for merging opportunities and more efficient | |
1451 | * dispatching. | |
1452 | */ | |
1453 | run_queue: | |
1454 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1455 | } |
1456 | ||
07068d5b | 1457 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1458 | return cookie; |
320ae51f JA |
1459 | } |
1460 | ||
1461 | /* | |
1462 | * Default mapping to a software queue, since we use one per CPU. | |
1463 | */ | |
1464 | struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu) | |
1465 | { | |
1466 | return q->queue_hw_ctx[q->mq_map[cpu]]; | |
1467 | } | |
1468 | EXPORT_SYMBOL(blk_mq_map_queue); | |
1469 | ||
24d2f903 CH |
1470 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1471 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1472 | { |
e9b267d9 | 1473 | struct page *page; |
320ae51f | 1474 | |
24d2f903 | 1475 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1476 | int i; |
320ae51f | 1477 | |
24d2f903 CH |
1478 | for (i = 0; i < tags->nr_tags; i++) { |
1479 | if (!tags->rqs[i]) | |
e9b267d9 | 1480 | continue; |
24d2f903 CH |
1481 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1482 | hctx_idx, i); | |
a5164405 | 1483 | tags->rqs[i] = NULL; |
e9b267d9 | 1484 | } |
320ae51f | 1485 | } |
320ae51f | 1486 | |
24d2f903 CH |
1487 | while (!list_empty(&tags->page_list)) { |
1488 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1489 | list_del_init(&page->lru); |
f75782e4 CM |
1490 | /* |
1491 | * Remove kmemleak object previously allocated in | |
1492 | * blk_mq_init_rq_map(). | |
1493 | */ | |
1494 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1495 | __free_pages(page, page->private); |
1496 | } | |
1497 | ||
24d2f903 | 1498 | kfree(tags->rqs); |
320ae51f | 1499 | |
24d2f903 | 1500 | blk_mq_free_tags(tags); |
320ae51f JA |
1501 | } |
1502 | ||
1503 | static size_t order_to_size(unsigned int order) | |
1504 | { | |
4ca08500 | 1505 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1506 | } |
1507 | ||
24d2f903 CH |
1508 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1509 | unsigned int hctx_idx) | |
320ae51f | 1510 | { |
24d2f903 | 1511 | struct blk_mq_tags *tags; |
320ae51f JA |
1512 | unsigned int i, j, entries_per_page, max_order = 4; |
1513 | size_t rq_size, left; | |
1514 | ||
24d2f903 | 1515 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1516 | set->numa_node, |
1517 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1518 | if (!tags) |
1519 | return NULL; | |
320ae51f | 1520 | |
24d2f903 CH |
1521 | INIT_LIST_HEAD(&tags->page_list); |
1522 | ||
a5164405 JA |
1523 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1524 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1525 | set->numa_node); | |
24d2f903 CH |
1526 | if (!tags->rqs) { |
1527 | blk_mq_free_tags(tags); | |
1528 | return NULL; | |
1529 | } | |
320ae51f JA |
1530 | |
1531 | /* | |
1532 | * rq_size is the size of the request plus driver payload, rounded | |
1533 | * to the cacheline size | |
1534 | */ | |
24d2f903 | 1535 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1536 | cache_line_size()); |
24d2f903 | 1537 | left = rq_size * set->queue_depth; |
320ae51f | 1538 | |
24d2f903 | 1539 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1540 | int this_order = max_order; |
1541 | struct page *page; | |
1542 | int to_do; | |
1543 | void *p; | |
1544 | ||
b3a834b1 | 1545 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1546 | this_order--; |
1547 | ||
1548 | do { | |
a5164405 | 1549 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1550 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1551 | this_order); |
320ae51f JA |
1552 | if (page) |
1553 | break; | |
1554 | if (!this_order--) | |
1555 | break; | |
1556 | if (order_to_size(this_order) < rq_size) | |
1557 | break; | |
1558 | } while (1); | |
1559 | ||
1560 | if (!page) | |
24d2f903 | 1561 | goto fail; |
320ae51f JA |
1562 | |
1563 | page->private = this_order; | |
24d2f903 | 1564 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1565 | |
1566 | p = page_address(page); | |
f75782e4 CM |
1567 | /* |
1568 | * Allow kmemleak to scan these pages as they contain pointers | |
1569 | * to additional allocations like via ops->init_request(). | |
1570 | */ | |
1571 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1572 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1573 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1574 | left -= to_do * rq_size; |
1575 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1576 | tags->rqs[i] = p; |
1577 | if (set->ops->init_request) { | |
1578 | if (set->ops->init_request(set->driver_data, | |
1579 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1580 | set->numa_node)) { |
1581 | tags->rqs[i] = NULL; | |
24d2f903 | 1582 | goto fail; |
a5164405 | 1583 | } |
e9b267d9 CH |
1584 | } |
1585 | ||
320ae51f JA |
1586 | p += rq_size; |
1587 | i++; | |
1588 | } | |
1589 | } | |
24d2f903 | 1590 | return tags; |
320ae51f | 1591 | |
24d2f903 | 1592 | fail: |
24d2f903 CH |
1593 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1594 | return NULL; | |
320ae51f JA |
1595 | } |
1596 | ||
1429d7c9 JA |
1597 | static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap) |
1598 | { | |
1599 | kfree(bitmap->map); | |
1600 | } | |
1601 | ||
1602 | static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node) | |
1603 | { | |
1604 | unsigned int bpw = 8, total, num_maps, i; | |
1605 | ||
1606 | bitmap->bits_per_word = bpw; | |
1607 | ||
1608 | num_maps = ALIGN(nr_cpu_ids, bpw) / bpw; | |
1609 | bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap), | |
1610 | GFP_KERNEL, node); | |
1611 | if (!bitmap->map) | |
1612 | return -ENOMEM; | |
1613 | ||
1429d7c9 JA |
1614 | total = nr_cpu_ids; |
1615 | for (i = 0; i < num_maps; i++) { | |
1616 | bitmap->map[i].depth = min(total, bitmap->bits_per_word); | |
1617 | total -= bitmap->map[i].depth; | |
1618 | } | |
1619 | ||
1620 | return 0; | |
1621 | } | |
1622 | ||
e57690fe JA |
1623 | /* |
1624 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1625 | * software queue to the hw queue dispatch list, and ensure that it | |
1626 | * gets run. | |
1627 | */ | |
484b4061 JA |
1628 | static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu) |
1629 | { | |
484b4061 JA |
1630 | struct blk_mq_ctx *ctx; |
1631 | LIST_HEAD(tmp); | |
1632 | ||
e57690fe | 1633 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1634 | |
1635 | spin_lock(&ctx->lock); | |
1636 | if (!list_empty(&ctx->rq_list)) { | |
1637 | list_splice_init(&ctx->rq_list, &tmp); | |
1638 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1639 | } | |
1640 | spin_unlock(&ctx->lock); | |
1641 | ||
1642 | if (list_empty(&tmp)) | |
1643 | return NOTIFY_OK; | |
1644 | ||
e57690fe JA |
1645 | spin_lock(&hctx->lock); |
1646 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1647 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1648 | |
1649 | blk_mq_run_hw_queue(hctx, true); | |
484b4061 JA |
1650 | return NOTIFY_OK; |
1651 | } | |
1652 | ||
484b4061 JA |
1653 | static int blk_mq_hctx_notify(void *data, unsigned long action, |
1654 | unsigned int cpu) | |
1655 | { | |
1656 | struct blk_mq_hw_ctx *hctx = data; | |
1657 | ||
1658 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) | |
1659 | return blk_mq_hctx_cpu_offline(hctx, cpu); | |
2a34c087 ML |
1660 | |
1661 | /* | |
1662 | * In case of CPU online, tags may be reallocated | |
1663 | * in blk_mq_map_swqueue() after mapping is updated. | |
1664 | */ | |
484b4061 JA |
1665 | |
1666 | return NOTIFY_OK; | |
1667 | } | |
1668 | ||
c3b4afca | 1669 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1670 | static void blk_mq_exit_hctx(struct request_queue *q, |
1671 | struct blk_mq_tag_set *set, | |
1672 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1673 | { | |
f70ced09 ML |
1674 | unsigned flush_start_tag = set->queue_depth; |
1675 | ||
08e98fc6 ML |
1676 | blk_mq_tag_idle(hctx); |
1677 | ||
f70ced09 ML |
1678 | if (set->ops->exit_request) |
1679 | set->ops->exit_request(set->driver_data, | |
1680 | hctx->fq->flush_rq, hctx_idx, | |
1681 | flush_start_tag + hctx_idx); | |
1682 | ||
08e98fc6 ML |
1683 | if (set->ops->exit_hctx) |
1684 | set->ops->exit_hctx(hctx, hctx_idx); | |
1685 | ||
1686 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
f70ced09 | 1687 | blk_free_flush_queue(hctx->fq); |
08e98fc6 ML |
1688 | blk_mq_free_bitmap(&hctx->ctx_map); |
1689 | } | |
1690 | ||
624dbe47 ML |
1691 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1692 | struct blk_mq_tag_set *set, int nr_queue) | |
1693 | { | |
1694 | struct blk_mq_hw_ctx *hctx; | |
1695 | unsigned int i; | |
1696 | ||
1697 | queue_for_each_hw_ctx(q, hctx, i) { | |
1698 | if (i == nr_queue) | |
1699 | break; | |
08e98fc6 | 1700 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1701 | } |
624dbe47 ML |
1702 | } |
1703 | ||
1704 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1705 | struct blk_mq_tag_set *set) | |
1706 | { | |
1707 | struct blk_mq_hw_ctx *hctx; | |
1708 | unsigned int i; | |
1709 | ||
e09aae7e | 1710 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1711 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1712 | } |
1713 | ||
08e98fc6 ML |
1714 | static int blk_mq_init_hctx(struct request_queue *q, |
1715 | struct blk_mq_tag_set *set, | |
1716 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1717 | { |
08e98fc6 | 1718 | int node; |
f70ced09 | 1719 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1720 | |
1721 | node = hctx->numa_node; | |
1722 | if (node == NUMA_NO_NODE) | |
1723 | node = hctx->numa_node = set->numa_node; | |
1724 | ||
1725 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
1726 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); | |
1727 | spin_lock_init(&hctx->lock); | |
1728 | INIT_LIST_HEAD(&hctx->dispatch); | |
1729 | hctx->queue = q; | |
1730 | hctx->queue_num = hctx_idx; | |
2404e607 | 1731 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 ML |
1732 | |
1733 | blk_mq_init_cpu_notifier(&hctx->cpu_notifier, | |
1734 | blk_mq_hctx_notify, hctx); | |
1735 | blk_mq_register_cpu_notifier(&hctx->cpu_notifier); | |
1736 | ||
1737 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1738 | |
1739 | /* | |
08e98fc6 ML |
1740 | * Allocate space for all possible cpus to avoid allocation at |
1741 | * runtime | |
320ae51f | 1742 | */ |
08e98fc6 ML |
1743 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1744 | GFP_KERNEL, node); | |
1745 | if (!hctx->ctxs) | |
1746 | goto unregister_cpu_notifier; | |
320ae51f | 1747 | |
08e98fc6 ML |
1748 | if (blk_mq_alloc_bitmap(&hctx->ctx_map, node)) |
1749 | goto free_ctxs; | |
320ae51f | 1750 | |
08e98fc6 | 1751 | hctx->nr_ctx = 0; |
320ae51f | 1752 | |
08e98fc6 ML |
1753 | if (set->ops->init_hctx && |
1754 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1755 | goto free_bitmap; | |
320ae51f | 1756 | |
f70ced09 ML |
1757 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1758 | if (!hctx->fq) | |
1759 | goto exit_hctx; | |
320ae51f | 1760 | |
f70ced09 ML |
1761 | if (set->ops->init_request && |
1762 | set->ops->init_request(set->driver_data, | |
1763 | hctx->fq->flush_rq, hctx_idx, | |
1764 | flush_start_tag + hctx_idx, node)) | |
1765 | goto free_fq; | |
320ae51f | 1766 | |
08e98fc6 | 1767 | return 0; |
320ae51f | 1768 | |
f70ced09 ML |
1769 | free_fq: |
1770 | kfree(hctx->fq); | |
1771 | exit_hctx: | |
1772 | if (set->ops->exit_hctx) | |
1773 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 ML |
1774 | free_bitmap: |
1775 | blk_mq_free_bitmap(&hctx->ctx_map); | |
1776 | free_ctxs: | |
1777 | kfree(hctx->ctxs); | |
1778 | unregister_cpu_notifier: | |
1779 | blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier); | |
320ae51f | 1780 | |
08e98fc6 ML |
1781 | return -1; |
1782 | } | |
320ae51f | 1783 | |
320ae51f JA |
1784 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1785 | unsigned int nr_hw_queues) | |
1786 | { | |
1787 | unsigned int i; | |
1788 | ||
1789 | for_each_possible_cpu(i) { | |
1790 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1791 | struct blk_mq_hw_ctx *hctx; | |
1792 | ||
1793 | memset(__ctx, 0, sizeof(*__ctx)); | |
1794 | __ctx->cpu = i; | |
1795 | spin_lock_init(&__ctx->lock); | |
1796 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1797 | __ctx->queue = q; | |
1798 | ||
1799 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1800 | if (!cpu_online(i)) |
1801 | continue; | |
1802 | ||
e4043dcf | 1803 | hctx = q->mq_ops->map_queue(q, i); |
e4043dcf | 1804 | |
320ae51f JA |
1805 | /* |
1806 | * Set local node, IFF we have more than one hw queue. If | |
1807 | * not, we remain on the home node of the device | |
1808 | */ | |
1809 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1810 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1811 | } |
1812 | } | |
1813 | ||
5778322e AM |
1814 | static void blk_mq_map_swqueue(struct request_queue *q, |
1815 | const struct cpumask *online_mask) | |
320ae51f JA |
1816 | { |
1817 | unsigned int i; | |
1818 | struct blk_mq_hw_ctx *hctx; | |
1819 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1820 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1821 | |
60de074b AM |
1822 | /* |
1823 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1824 | */ | |
1825 | mutex_lock(&q->sysfs_lock); | |
1826 | ||
320ae51f | 1827 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1828 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1829 | hctx->nr_ctx = 0; |
1830 | } | |
1831 | ||
1832 | /* | |
1833 | * Map software to hardware queues | |
1834 | */ | |
897bb0c7 | 1835 | for_each_possible_cpu(i) { |
320ae51f | 1836 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1837 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1838 | continue; |
1839 | ||
897bb0c7 | 1840 | ctx = per_cpu_ptr(q->queue_ctx, i); |
320ae51f | 1841 | hctx = q->mq_ops->map_queue(q, i); |
868f2f0b | 1842 | |
e4043dcf | 1843 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1844 | ctx->index_hw = hctx->nr_ctx; |
1845 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1846 | } | |
506e931f | 1847 | |
60de074b AM |
1848 | mutex_unlock(&q->sysfs_lock); |
1849 | ||
506e931f | 1850 | queue_for_each_hw_ctx(q, hctx, i) { |
889fa31f CY |
1851 | struct blk_mq_ctxmap *map = &hctx->ctx_map; |
1852 | ||
484b4061 | 1853 | /* |
a68aafa5 JA |
1854 | * If no software queues are mapped to this hardware queue, |
1855 | * disable it and free the request entries. | |
484b4061 JA |
1856 | */ |
1857 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1858 | if (set->tags[i]) { |
1859 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1860 | set->tags[i] = NULL; | |
484b4061 | 1861 | } |
2a34c087 | 1862 | hctx->tags = NULL; |
484b4061 JA |
1863 | continue; |
1864 | } | |
1865 | ||
2a34c087 ML |
1866 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1867 | if (!set->tags[i]) | |
1868 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1869 | hctx->tags = set->tags[i]; | |
1870 | WARN_ON(!hctx->tags); | |
1871 | ||
e0e827b9 | 1872 | cpumask_copy(hctx->tags->cpumask, hctx->cpumask); |
889fa31f CY |
1873 | /* |
1874 | * Set the map size to the number of mapped software queues. | |
1875 | * This is more accurate and more efficient than looping | |
1876 | * over all possibly mapped software queues. | |
1877 | */ | |
569fd0ce | 1878 | map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word); |
889fa31f | 1879 | |
484b4061 JA |
1880 | /* |
1881 | * Initialize batch roundrobin counts | |
1882 | */ | |
506e931f JA |
1883 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1884 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1885 | } | |
320ae51f JA |
1886 | } |
1887 | ||
2404e607 | 1888 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1889 | { |
1890 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1891 | int i; |
1892 | ||
2404e607 JM |
1893 | queue_for_each_hw_ctx(q, hctx, i) { |
1894 | if (shared) | |
1895 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1896 | else | |
1897 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1898 | } | |
1899 | } | |
1900 | ||
1901 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1902 | { | |
1903 | struct request_queue *q; | |
0d2602ca JA |
1904 | |
1905 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1906 | blk_mq_freeze_queue(q); | |
2404e607 | 1907 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1908 | blk_mq_unfreeze_queue(q); |
1909 | } | |
1910 | } | |
1911 | ||
1912 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1913 | { | |
1914 | struct blk_mq_tag_set *set = q->tag_set; | |
1915 | ||
0d2602ca JA |
1916 | mutex_lock(&set->tag_list_lock); |
1917 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1918 | if (list_is_singular(&set->tag_list)) { |
1919 | /* just transitioned to unshared */ | |
1920 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1921 | /* update existing queue */ | |
1922 | blk_mq_update_tag_set_depth(set, false); | |
1923 | } | |
0d2602ca | 1924 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1925 | } |
1926 | ||
1927 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1928 | struct request_queue *q) | |
1929 | { | |
1930 | q->tag_set = set; | |
1931 | ||
1932 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1933 | |
1934 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1935 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1936 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1937 | /* update existing queue */ | |
1938 | blk_mq_update_tag_set_depth(set, true); | |
1939 | } | |
1940 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1941 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1942 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1943 | |
0d2602ca JA |
1944 | mutex_unlock(&set->tag_list_lock); |
1945 | } | |
1946 | ||
e09aae7e ML |
1947 | /* |
1948 | * It is the actual release handler for mq, but we do it from | |
1949 | * request queue's release handler for avoiding use-after-free | |
1950 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1951 | * but we can't group ctx/kctx kobj without it. | |
1952 | */ | |
1953 | void blk_mq_release(struct request_queue *q) | |
1954 | { | |
1955 | struct blk_mq_hw_ctx *hctx; | |
1956 | unsigned int i; | |
1957 | ||
1958 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1959 | queue_for_each_hw_ctx(q, hctx, i) { |
1960 | if (!hctx) | |
1961 | continue; | |
1962 | kfree(hctx->ctxs); | |
e09aae7e | 1963 | kfree(hctx); |
c3b4afca | 1964 | } |
e09aae7e | 1965 | |
a723bab3 AM |
1966 | kfree(q->mq_map); |
1967 | q->mq_map = NULL; | |
1968 | ||
e09aae7e ML |
1969 | kfree(q->queue_hw_ctx); |
1970 | ||
1971 | /* ctx kobj stays in queue_ctx */ | |
1972 | free_percpu(q->queue_ctx); | |
1973 | } | |
1974 | ||
24d2f903 | 1975 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1976 | { |
1977 | struct request_queue *uninit_q, *q; | |
1978 | ||
1979 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1980 | if (!uninit_q) | |
1981 | return ERR_PTR(-ENOMEM); | |
1982 | ||
1983 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1984 | if (IS_ERR(q)) | |
1985 | blk_cleanup_queue(uninit_q); | |
1986 | ||
1987 | return q; | |
1988 | } | |
1989 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1990 | ||
868f2f0b KB |
1991 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
1992 | struct request_queue *q) | |
320ae51f | 1993 | { |
868f2f0b KB |
1994 | int i, j; |
1995 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 1996 | |
868f2f0b | 1997 | blk_mq_sysfs_unregister(q); |
24d2f903 | 1998 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 1999 | int node; |
f14bbe77 | 2000 | |
868f2f0b KB |
2001 | if (hctxs[i]) |
2002 | continue; | |
2003 | ||
2004 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2005 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2006 | GFP_KERNEL, node); | |
320ae51f | 2007 | if (!hctxs[i]) |
868f2f0b | 2008 | break; |
320ae51f | 2009 | |
a86073e4 | 2010 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2011 | node)) { |
2012 | kfree(hctxs[i]); | |
2013 | hctxs[i] = NULL; | |
2014 | break; | |
2015 | } | |
e4043dcf | 2016 | |
0d2602ca | 2017 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2018 | hctxs[i]->numa_node = node; |
320ae51f | 2019 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2020 | |
2021 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2022 | free_cpumask_var(hctxs[i]->cpumask); | |
2023 | kfree(hctxs[i]); | |
2024 | hctxs[i] = NULL; | |
2025 | break; | |
2026 | } | |
2027 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2028 | } |
868f2f0b KB |
2029 | for (j = i; j < q->nr_hw_queues; j++) { |
2030 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2031 | ||
2032 | if (hctx) { | |
2033 | if (hctx->tags) { | |
2034 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2035 | set->tags[j] = NULL; | |
2036 | } | |
2037 | blk_mq_exit_hctx(q, set, hctx, j); | |
2038 | free_cpumask_var(hctx->cpumask); | |
2039 | kobject_put(&hctx->kobj); | |
2040 | kfree(hctx->ctxs); | |
2041 | kfree(hctx); | |
2042 | hctxs[j] = NULL; | |
2043 | ||
2044 | } | |
2045 | } | |
2046 | q->nr_hw_queues = i; | |
2047 | blk_mq_sysfs_register(q); | |
2048 | } | |
2049 | ||
2050 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2051 | struct request_queue *q) | |
2052 | { | |
66841672 ML |
2053 | /* mark the queue as mq asap */ |
2054 | q->mq_ops = set->ops; | |
2055 | ||
868f2f0b KB |
2056 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2057 | if (!q->queue_ctx) | |
c7de5726 | 2058 | goto err_exit; |
868f2f0b KB |
2059 | |
2060 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2061 | GFP_KERNEL, set->numa_node); | |
2062 | if (!q->queue_hw_ctx) | |
2063 | goto err_percpu; | |
2064 | ||
2065 | q->mq_map = blk_mq_make_queue_map(set); | |
2066 | if (!q->mq_map) | |
2067 | goto err_map; | |
2068 | ||
2069 | blk_mq_realloc_hw_ctxs(set, q); | |
2070 | if (!q->nr_hw_queues) | |
2071 | goto err_hctxs; | |
320ae51f | 2072 | |
287922eb | 2073 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2074 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2075 | |
2076 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2077 | |
94eddfbe | 2078 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2079 | |
05f1dd53 JA |
2080 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2081 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2082 | ||
1be036e9 CH |
2083 | q->sg_reserved_size = INT_MAX; |
2084 | ||
6fca6a61 CH |
2085 | INIT_WORK(&q->requeue_work, blk_mq_requeue_work); |
2086 | INIT_LIST_HEAD(&q->requeue_list); | |
2087 | spin_lock_init(&q->requeue_lock); | |
2088 | ||
07068d5b JA |
2089 | if (q->nr_hw_queues > 1) |
2090 | blk_queue_make_request(q, blk_mq_make_request); | |
2091 | else | |
2092 | blk_queue_make_request(q, blk_sq_make_request); | |
2093 | ||
eba71768 JA |
2094 | /* |
2095 | * Do this after blk_queue_make_request() overrides it... | |
2096 | */ | |
2097 | q->nr_requests = set->queue_depth; | |
2098 | ||
24d2f903 CH |
2099 | if (set->ops->complete) |
2100 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2101 | |
24d2f903 | 2102 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2103 | |
5778322e | 2104 | get_online_cpus(); |
320ae51f | 2105 | mutex_lock(&all_q_mutex); |
320ae51f | 2106 | |
4593fdbe | 2107 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2108 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2109 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2110 | |
4593fdbe | 2111 | mutex_unlock(&all_q_mutex); |
5778322e | 2112 | put_online_cpus(); |
4593fdbe | 2113 | |
320ae51f | 2114 | return q; |
18741986 | 2115 | |
320ae51f | 2116 | err_hctxs: |
868f2f0b | 2117 | kfree(q->mq_map); |
f14bbe77 | 2118 | err_map: |
868f2f0b | 2119 | kfree(q->queue_hw_ctx); |
320ae51f | 2120 | err_percpu: |
868f2f0b | 2121 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2122 | err_exit: |
2123 | q->mq_ops = NULL; | |
320ae51f JA |
2124 | return ERR_PTR(-ENOMEM); |
2125 | } | |
b62c21b7 | 2126 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2127 | |
2128 | void blk_mq_free_queue(struct request_queue *q) | |
2129 | { | |
624dbe47 | 2130 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2131 | |
0e626368 AM |
2132 | mutex_lock(&all_q_mutex); |
2133 | list_del_init(&q->all_q_node); | |
2134 | mutex_unlock(&all_q_mutex); | |
2135 | ||
0d2602ca JA |
2136 | blk_mq_del_queue_tag_set(q); |
2137 | ||
624dbe47 ML |
2138 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2139 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2140 | } |
320ae51f JA |
2141 | |
2142 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2143 | static void blk_mq_queue_reinit(struct request_queue *q, |
2144 | const struct cpumask *online_mask) | |
320ae51f | 2145 | { |
4ecd4fef | 2146 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2147 | |
67aec14c JA |
2148 | blk_mq_sysfs_unregister(q); |
2149 | ||
5778322e | 2150 | blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask); |
320ae51f JA |
2151 | |
2152 | /* | |
2153 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2154 | * we should change hctx numa_node according to new topology (this | |
2155 | * involves free and re-allocate memory, worthy doing?) | |
2156 | */ | |
2157 | ||
5778322e | 2158 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2159 | |
67aec14c | 2160 | blk_mq_sysfs_register(q); |
320ae51f JA |
2161 | } |
2162 | ||
f618ef7c PG |
2163 | static int blk_mq_queue_reinit_notify(struct notifier_block *nb, |
2164 | unsigned long action, void *hcpu) | |
320ae51f JA |
2165 | { |
2166 | struct request_queue *q; | |
5778322e AM |
2167 | int cpu = (unsigned long)hcpu; |
2168 | /* | |
2169 | * New online cpumask which is going to be set in this hotplug event. | |
2170 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2171 | * one-by-one and dynamically allocating this could result in a failure. | |
2172 | */ | |
2173 | static struct cpumask online_new; | |
320ae51f JA |
2174 | |
2175 | /* | |
5778322e AM |
2176 | * Before hotadded cpu starts handling requests, new mappings must |
2177 | * be established. Otherwise, these requests in hw queue might | |
2178 | * never be dispatched. | |
2179 | * | |
2180 | * For example, there is a single hw queue (hctx) and two CPU queues | |
2181 | * (ctx0 for CPU0, and ctx1 for CPU1). | |
2182 | * | |
2183 | * Now CPU1 is just onlined and a request is inserted into | |
2184 | * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is | |
2185 | * still zero. | |
2186 | * | |
2187 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is | |
2188 | * set in pending bitmap and tries to retrieve requests in | |
2189 | * hctx->ctxs[0]->rq_list. But htx->ctxs[0] is a pointer to ctx0, | |
2190 | * so the request in ctx1->rq_list is ignored. | |
320ae51f | 2191 | */ |
5778322e AM |
2192 | switch (action & ~CPU_TASKS_FROZEN) { |
2193 | case CPU_DEAD: | |
2194 | case CPU_UP_CANCELED: | |
2195 | cpumask_copy(&online_new, cpu_online_mask); | |
2196 | break; | |
2197 | case CPU_UP_PREPARE: | |
2198 | cpumask_copy(&online_new, cpu_online_mask); | |
2199 | cpumask_set_cpu(cpu, &online_new); | |
2200 | break; | |
2201 | default: | |
320ae51f | 2202 | return NOTIFY_OK; |
5778322e | 2203 | } |
320ae51f JA |
2204 | |
2205 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2206 | |
2207 | /* | |
2208 | * We need to freeze and reinit all existing queues. Freezing | |
2209 | * involves synchronous wait for an RCU grace period and doing it | |
2210 | * one by one may take a long time. Start freezing all queues in | |
2211 | * one swoop and then wait for the completions so that freezing can | |
2212 | * take place in parallel. | |
2213 | */ | |
2214 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2215 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2216 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2217 | blk_mq_freeze_queue_wait(q); |
2218 | ||
f054b56c ML |
2219 | /* |
2220 | * timeout handler can't touch hw queue during the | |
2221 | * reinitialization | |
2222 | */ | |
2223 | del_timer_sync(&q->timeout); | |
2224 | } | |
2225 | ||
320ae51f | 2226 | list_for_each_entry(q, &all_q_list, all_q_node) |
5778322e | 2227 | blk_mq_queue_reinit(q, &online_new); |
f3af020b TH |
2228 | |
2229 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2230 | blk_mq_unfreeze_queue(q); | |
2231 | ||
320ae51f JA |
2232 | mutex_unlock(&all_q_mutex); |
2233 | return NOTIFY_OK; | |
2234 | } | |
2235 | ||
a5164405 JA |
2236 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2237 | { | |
2238 | int i; | |
2239 | ||
2240 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2241 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2242 | if (!set->tags[i]) | |
2243 | goto out_unwind; | |
2244 | } | |
2245 | ||
2246 | return 0; | |
2247 | ||
2248 | out_unwind: | |
2249 | while (--i >= 0) | |
2250 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2251 | ||
a5164405 JA |
2252 | return -ENOMEM; |
2253 | } | |
2254 | ||
2255 | /* | |
2256 | * Allocate the request maps associated with this tag_set. Note that this | |
2257 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2258 | * will be updated to reflect the allocated depth. | |
2259 | */ | |
2260 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2261 | { | |
2262 | unsigned int depth; | |
2263 | int err; | |
2264 | ||
2265 | depth = set->queue_depth; | |
2266 | do { | |
2267 | err = __blk_mq_alloc_rq_maps(set); | |
2268 | if (!err) | |
2269 | break; | |
2270 | ||
2271 | set->queue_depth >>= 1; | |
2272 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2273 | err = -ENOMEM; | |
2274 | break; | |
2275 | } | |
2276 | } while (set->queue_depth); | |
2277 | ||
2278 | if (!set->queue_depth || err) { | |
2279 | pr_err("blk-mq: failed to allocate request map\n"); | |
2280 | return -ENOMEM; | |
2281 | } | |
2282 | ||
2283 | if (depth != set->queue_depth) | |
2284 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2285 | depth, set->queue_depth); | |
2286 | ||
2287 | return 0; | |
2288 | } | |
2289 | ||
f26cdc85 KB |
2290 | struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags) |
2291 | { | |
2292 | return tags->cpumask; | |
2293 | } | |
2294 | EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask); | |
2295 | ||
a4391c64 JA |
2296 | /* |
2297 | * Alloc a tag set to be associated with one or more request queues. | |
2298 | * May fail with EINVAL for various error conditions. May adjust the | |
2299 | * requested depth down, if if it too large. In that case, the set | |
2300 | * value will be stored in set->queue_depth. | |
2301 | */ | |
24d2f903 CH |
2302 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2303 | { | |
205fb5f5 BVA |
2304 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2305 | ||
24d2f903 CH |
2306 | if (!set->nr_hw_queues) |
2307 | return -EINVAL; | |
a4391c64 | 2308 | if (!set->queue_depth) |
24d2f903 CH |
2309 | return -EINVAL; |
2310 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2311 | return -EINVAL; | |
2312 | ||
f9018ac9 | 2313 | if (!set->ops->queue_rq || !set->ops->map_queue) |
24d2f903 CH |
2314 | return -EINVAL; |
2315 | ||
a4391c64 JA |
2316 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2317 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2318 | BLK_MQ_MAX_DEPTH); | |
2319 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2320 | } | |
24d2f903 | 2321 | |
6637fadf SL |
2322 | /* |
2323 | * If a crashdump is active, then we are potentially in a very | |
2324 | * memory constrained environment. Limit us to 1 queue and | |
2325 | * 64 tags to prevent using too much memory. | |
2326 | */ | |
2327 | if (is_kdump_kernel()) { | |
2328 | set->nr_hw_queues = 1; | |
2329 | set->queue_depth = min(64U, set->queue_depth); | |
2330 | } | |
868f2f0b KB |
2331 | /* |
2332 | * There is no use for more h/w queues than cpus. | |
2333 | */ | |
2334 | if (set->nr_hw_queues > nr_cpu_ids) | |
2335 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2336 | |
868f2f0b | 2337 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2338 | GFP_KERNEL, set->numa_node); |
2339 | if (!set->tags) | |
a5164405 | 2340 | return -ENOMEM; |
24d2f903 | 2341 | |
a5164405 JA |
2342 | if (blk_mq_alloc_rq_maps(set)) |
2343 | goto enomem; | |
24d2f903 | 2344 | |
0d2602ca JA |
2345 | mutex_init(&set->tag_list_lock); |
2346 | INIT_LIST_HEAD(&set->tag_list); | |
2347 | ||
24d2f903 | 2348 | return 0; |
a5164405 | 2349 | enomem: |
5676e7b6 RE |
2350 | kfree(set->tags); |
2351 | set->tags = NULL; | |
24d2f903 CH |
2352 | return -ENOMEM; |
2353 | } | |
2354 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2355 | ||
2356 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2357 | { | |
2358 | int i; | |
2359 | ||
868f2f0b | 2360 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2361 | if (set->tags[i]) |
484b4061 JA |
2362 | blk_mq_free_rq_map(set, set->tags[i], i); |
2363 | } | |
2364 | ||
981bd189 | 2365 | kfree(set->tags); |
5676e7b6 | 2366 | set->tags = NULL; |
24d2f903 CH |
2367 | } |
2368 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2369 | ||
e3a2b3f9 JA |
2370 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2371 | { | |
2372 | struct blk_mq_tag_set *set = q->tag_set; | |
2373 | struct blk_mq_hw_ctx *hctx; | |
2374 | int i, ret; | |
2375 | ||
2376 | if (!set || nr > set->queue_depth) | |
2377 | return -EINVAL; | |
2378 | ||
2379 | ret = 0; | |
2380 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2381 | if (!hctx->tags) |
2382 | continue; | |
e3a2b3f9 JA |
2383 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2384 | if (ret) | |
2385 | break; | |
2386 | } | |
2387 | ||
2388 | if (!ret) | |
2389 | q->nr_requests = nr; | |
2390 | ||
2391 | return ret; | |
2392 | } | |
2393 | ||
868f2f0b KB |
2394 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2395 | { | |
2396 | struct request_queue *q; | |
2397 | ||
2398 | if (nr_hw_queues > nr_cpu_ids) | |
2399 | nr_hw_queues = nr_cpu_ids; | |
2400 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2401 | return; | |
2402 | ||
2403 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2404 | blk_mq_freeze_queue(q); | |
2405 | ||
2406 | set->nr_hw_queues = nr_hw_queues; | |
2407 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2408 | blk_mq_realloc_hw_ctxs(set, q); | |
2409 | ||
2410 | if (q->nr_hw_queues > 1) | |
2411 | blk_queue_make_request(q, blk_mq_make_request); | |
2412 | else | |
2413 | blk_queue_make_request(q, blk_sq_make_request); | |
2414 | ||
2415 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2416 | } | |
2417 | ||
2418 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2419 | blk_mq_unfreeze_queue(q); | |
2420 | } | |
2421 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2422 | ||
676141e4 JA |
2423 | void blk_mq_disable_hotplug(void) |
2424 | { | |
2425 | mutex_lock(&all_q_mutex); | |
2426 | } | |
2427 | ||
2428 | void blk_mq_enable_hotplug(void) | |
2429 | { | |
2430 | mutex_unlock(&all_q_mutex); | |
2431 | } | |
2432 | ||
320ae51f JA |
2433 | static int __init blk_mq_init(void) |
2434 | { | |
320ae51f JA |
2435 | blk_mq_cpu_init(); |
2436 | ||
add703fd | 2437 | hotcpu_notifier(blk_mq_queue_reinit_notify, 0); |
320ae51f JA |
2438 | |
2439 | return 0; | |
2440 | } | |
2441 | subsys_initcall(blk_mq_init); |