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 | ||
320ae51f JA |
37 | /* |
38 | * Check if any of the ctx's have pending work in this hardware queue | |
39 | */ | |
40 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
41 | { | |
88459642 | 42 | return sbitmap_any_bit_set(&hctx->ctx_map); |
1429d7c9 JA |
43 | } |
44 | ||
320ae51f JA |
45 | /* |
46 | * Mark this ctx as having pending work in this hardware queue | |
47 | */ | |
48 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
49 | struct blk_mq_ctx *ctx) | |
50 | { | |
88459642 OS |
51 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
52 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
53 | } |
54 | ||
55 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
56 | struct blk_mq_ctx *ctx) | |
57 | { | |
88459642 | 58 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
59 | } |
60 | ||
b4c6a028 | 61 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 62 | { |
4ecd4fef | 63 | int freeze_depth; |
cddd5d17 | 64 | |
4ecd4fef CH |
65 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
66 | if (freeze_depth == 1) { | |
3ef28e83 | 67 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 68 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 69 | } |
f3af020b | 70 | } |
b4c6a028 | 71 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
72 | |
73 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
74 | { | |
3ef28e83 | 75 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
76 | } |
77 | ||
f3af020b TH |
78 | /* |
79 | * Guarantee no request is in use, so we can change any data structure of | |
80 | * the queue afterward. | |
81 | */ | |
3ef28e83 | 82 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 83 | { |
3ef28e83 DW |
84 | /* |
85 | * In the !blk_mq case we are only calling this to kill the | |
86 | * q_usage_counter, otherwise this increases the freeze depth | |
87 | * and waits for it to return to zero. For this reason there is | |
88 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
89 | * exported to drivers as the only user for unfreeze is blk_mq. | |
90 | */ | |
f3af020b TH |
91 | blk_mq_freeze_queue_start(q); |
92 | blk_mq_freeze_queue_wait(q); | |
93 | } | |
3ef28e83 DW |
94 | |
95 | void blk_mq_freeze_queue(struct request_queue *q) | |
96 | { | |
97 | /* | |
98 | * ...just an alias to keep freeze and unfreeze actions balanced | |
99 | * in the blk_mq_* namespace | |
100 | */ | |
101 | blk_freeze_queue(q); | |
102 | } | |
c761d96b | 103 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 104 | |
b4c6a028 | 105 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 106 | { |
4ecd4fef | 107 | int freeze_depth; |
320ae51f | 108 | |
4ecd4fef CH |
109 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
110 | WARN_ON_ONCE(freeze_depth < 0); | |
111 | if (!freeze_depth) { | |
3ef28e83 | 112 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 113 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 114 | } |
320ae51f | 115 | } |
b4c6a028 | 116 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 117 | |
6a83e74d BVA |
118 | /** |
119 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
120 | * @q: request queue. | |
121 | * | |
122 | * Note: this function does not prevent that the struct request end_io() | |
123 | * callback function is invoked. Additionally, it is not prevented that | |
124 | * new queue_rq() calls occur unless the queue has been stopped first. | |
125 | */ | |
126 | void blk_mq_quiesce_queue(struct request_queue *q) | |
127 | { | |
128 | struct blk_mq_hw_ctx *hctx; | |
129 | unsigned int i; | |
130 | bool rcu = false; | |
131 | ||
132 | blk_mq_stop_hw_queues(q); | |
133 | ||
134 | queue_for_each_hw_ctx(q, hctx, i) { | |
135 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
136 | synchronize_srcu(&hctx->queue_rq_srcu); | |
137 | else | |
138 | rcu = true; | |
139 | } | |
140 | if (rcu) | |
141 | synchronize_rcu(); | |
142 | } | |
143 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
144 | ||
aed3ea94 JA |
145 | void blk_mq_wake_waiters(struct request_queue *q) |
146 | { | |
147 | struct blk_mq_hw_ctx *hctx; | |
148 | unsigned int i; | |
149 | ||
150 | queue_for_each_hw_ctx(q, hctx, i) | |
151 | if (blk_mq_hw_queue_mapped(hctx)) | |
152 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
153 | |
154 | /* | |
155 | * If we are called because the queue has now been marked as | |
156 | * dying, we need to ensure that processes currently waiting on | |
157 | * the queue are notified as well. | |
158 | */ | |
159 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
160 | } |
161 | ||
320ae51f JA |
162 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
163 | { | |
164 | return blk_mq_has_free_tags(hctx->tags); | |
165 | } | |
166 | EXPORT_SYMBOL(blk_mq_can_queue); | |
167 | ||
94eddfbe | 168 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
ef295ecf | 169 | struct request *rq, unsigned int op) |
320ae51f | 170 | { |
af76e555 CH |
171 | INIT_LIST_HEAD(&rq->queuelist); |
172 | /* csd/requeue_work/fifo_time is initialized before use */ | |
173 | rq->q = q; | |
320ae51f | 174 | rq->mq_ctx = ctx; |
ef295ecf | 175 | rq->cmd_flags = op; |
e8064021 CH |
176 | if (blk_queue_io_stat(q)) |
177 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
178 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
179 | rq->cpu = -1; | |
af76e555 CH |
180 | INIT_HLIST_NODE(&rq->hash); |
181 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
182 | rq->rq_disk = NULL; |
183 | rq->part = NULL; | |
3ee32372 | 184 | rq->start_time = jiffies; |
af76e555 CH |
185 | #ifdef CONFIG_BLK_CGROUP |
186 | rq->rl = NULL; | |
0fec08b4 | 187 | set_start_time_ns(rq); |
af76e555 CH |
188 | rq->io_start_time_ns = 0; |
189 | #endif | |
190 | rq->nr_phys_segments = 0; | |
191 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
192 | rq->nr_integrity_segments = 0; | |
193 | #endif | |
af76e555 CH |
194 | rq->special = NULL; |
195 | /* tag was already set */ | |
196 | rq->errors = 0; | |
af76e555 | 197 | |
6f4a1626 TB |
198 | rq->cmd = rq->__cmd; |
199 | ||
af76e555 CH |
200 | rq->extra_len = 0; |
201 | rq->sense_len = 0; | |
202 | rq->resid_len = 0; | |
203 | rq->sense = NULL; | |
204 | ||
af76e555 | 205 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
206 | rq->timeout = 0; |
207 | ||
af76e555 CH |
208 | rq->end_io = NULL; |
209 | rq->end_io_data = NULL; | |
210 | rq->next_rq = NULL; | |
211 | ||
ef295ecf | 212 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f JA |
213 | } |
214 | ||
5dee8577 | 215 | static struct request * |
ef295ecf | 216 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, unsigned int op) |
5dee8577 CH |
217 | { |
218 | struct request *rq; | |
219 | unsigned int tag; | |
220 | ||
cb96a42c | 221 | tag = blk_mq_get_tag(data); |
5dee8577 | 222 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 223 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 224 | |
cb96a42c | 225 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 226 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 227 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
228 | } |
229 | ||
230 | rq->tag = tag; | |
ef295ecf | 231 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
232 | return rq; |
233 | } | |
234 | ||
235 | return NULL; | |
236 | } | |
237 | ||
6f3b0e8b CH |
238 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
239 | unsigned int flags) | |
320ae51f | 240 | { |
d852564f CH |
241 | struct blk_mq_ctx *ctx; |
242 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 243 | struct request *rq; |
cb96a42c | 244 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 245 | int ret; |
320ae51f | 246 | |
6f3b0e8b | 247 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
248 | if (ret) |
249 | return ERR_PTR(ret); | |
320ae51f | 250 | |
d852564f | 251 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 252 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 253 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
ef295ecf | 254 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f | 255 | blk_mq_put_ctx(ctx); |
841bac2c | 256 | |
c76541a9 | 257 | if (!rq) { |
3ef28e83 | 258 | blk_queue_exit(q); |
a492f075 | 259 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 260 | } |
0c4de0f3 CH |
261 | |
262 | rq->__data_len = 0; | |
263 | rq->__sector = (sector_t) -1; | |
264 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
265 | return rq; |
266 | } | |
4bb659b1 | 267 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 268 | |
1f5bd336 ML |
269 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
270 | unsigned int flags, unsigned int hctx_idx) | |
271 | { | |
272 | struct blk_mq_hw_ctx *hctx; | |
273 | struct blk_mq_ctx *ctx; | |
274 | struct request *rq; | |
275 | struct blk_mq_alloc_data alloc_data; | |
276 | int ret; | |
277 | ||
278 | /* | |
279 | * If the tag allocator sleeps we could get an allocation for a | |
280 | * different hardware context. No need to complicate the low level | |
281 | * allocator for this for the rare use case of a command tied to | |
282 | * a specific queue. | |
283 | */ | |
284 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
285 | return ERR_PTR(-EINVAL); | |
286 | ||
287 | if (hctx_idx >= q->nr_hw_queues) | |
288 | return ERR_PTR(-EIO); | |
289 | ||
290 | ret = blk_queue_enter(q, true); | |
291 | if (ret) | |
292 | return ERR_PTR(ret); | |
293 | ||
c8712c6a CH |
294 | /* |
295 | * Check if the hardware context is actually mapped to anything. | |
296 | * If not tell the caller that it should skip this queue. | |
297 | */ | |
1f5bd336 | 298 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
299 | if (!blk_mq_hw_queue_mapped(hctx)) { |
300 | ret = -EXDEV; | |
301 | goto out_queue_exit; | |
302 | } | |
1f5bd336 ML |
303 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
304 | ||
305 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 306 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 307 | if (!rq) { |
c8712c6a CH |
308 | ret = -EWOULDBLOCK; |
309 | goto out_queue_exit; | |
1f5bd336 ML |
310 | } |
311 | ||
312 | return rq; | |
c8712c6a CH |
313 | |
314 | out_queue_exit: | |
315 | blk_queue_exit(q); | |
316 | return ERR_PTR(ret); | |
1f5bd336 ML |
317 | } |
318 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
319 | ||
320ae51f JA |
320 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
321 | struct blk_mq_ctx *ctx, struct request *rq) | |
322 | { | |
323 | const int tag = rq->tag; | |
324 | struct request_queue *q = rq->q; | |
325 | ||
e8064021 | 326 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 327 | atomic_dec(&hctx->nr_active); |
e8064021 | 328 | rq->rq_flags = 0; |
0d2602ca | 329 | |
af76e555 | 330 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
40aabb67 | 331 | blk_mq_put_tag(hctx, ctx, tag); |
3ef28e83 | 332 | blk_queue_exit(q); |
320ae51f JA |
333 | } |
334 | ||
7c7f2f2b | 335 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
336 | { |
337 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
338 | |
339 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 340 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
341 | |
342 | } | |
343 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
344 | ||
345 | void blk_mq_free_request(struct request *rq) | |
346 | { | |
7d7e0f90 | 347 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 348 | } |
1a3b595a | 349 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 350 | |
c8a446ad | 351 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 352 | { |
0d11e6ac ML |
353 | blk_account_io_done(rq); |
354 | ||
91b63639 | 355 | if (rq->end_io) { |
320ae51f | 356 | rq->end_io(rq, error); |
91b63639 CH |
357 | } else { |
358 | if (unlikely(blk_bidi_rq(rq))) | |
359 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 360 | blk_mq_free_request(rq); |
91b63639 | 361 | } |
320ae51f | 362 | } |
c8a446ad | 363 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 364 | |
c8a446ad | 365 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
366 | { |
367 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
368 | BUG(); | |
c8a446ad | 369 | __blk_mq_end_request(rq, error); |
63151a44 | 370 | } |
c8a446ad | 371 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 372 | |
30a91cb4 | 373 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 374 | { |
3d6efbf6 | 375 | struct request *rq = data; |
320ae51f | 376 | |
30a91cb4 | 377 | rq->q->softirq_done_fn(rq); |
320ae51f | 378 | } |
320ae51f | 379 | |
ed851860 | 380 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
381 | { |
382 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 383 | bool shared = false; |
320ae51f JA |
384 | int cpu; |
385 | ||
38535201 | 386 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
387 | rq->q->softirq_done_fn(rq); |
388 | return; | |
389 | } | |
320ae51f JA |
390 | |
391 | cpu = get_cpu(); | |
38535201 CH |
392 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
393 | shared = cpus_share_cache(cpu, ctx->cpu); | |
394 | ||
395 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 396 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
397 | rq->csd.info = rq; |
398 | rq->csd.flags = 0; | |
c46fff2a | 399 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 400 | } else { |
30a91cb4 | 401 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 402 | } |
320ae51f JA |
403 | put_cpu(); |
404 | } | |
30a91cb4 | 405 | |
1fa8cc52 | 406 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
407 | { |
408 | struct request_queue *q = rq->q; | |
409 | ||
410 | if (!q->softirq_done_fn) | |
c8a446ad | 411 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
412 | else |
413 | blk_mq_ipi_complete_request(rq); | |
414 | } | |
415 | ||
30a91cb4 CH |
416 | /** |
417 | * blk_mq_complete_request - end I/O on a request | |
418 | * @rq: the request being processed | |
419 | * | |
420 | * Description: | |
421 | * Ends all I/O on a request. It does not handle partial completions. | |
422 | * The actual completion happens out-of-order, through a IPI handler. | |
423 | **/ | |
f4829a9b | 424 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 425 | { |
95f09684 JA |
426 | struct request_queue *q = rq->q; |
427 | ||
428 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 429 | return; |
f4829a9b CH |
430 | if (!blk_mark_rq_complete(rq)) { |
431 | rq->errors = error; | |
ed851860 | 432 | __blk_mq_complete_request(rq); |
f4829a9b | 433 | } |
30a91cb4 CH |
434 | } |
435 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 436 | |
973c0191 KB |
437 | int blk_mq_request_started(struct request *rq) |
438 | { | |
439 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
440 | } | |
441 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
442 | ||
e2490073 | 443 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
444 | { |
445 | struct request_queue *q = rq->q; | |
446 | ||
447 | trace_block_rq_issue(q, rq); | |
448 | ||
742ee69b | 449 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
450 | if (unlikely(blk_bidi_rq(rq))) |
451 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 452 | |
2b8393b4 | 453 | blk_add_timer(rq); |
87ee7b11 | 454 | |
538b7534 JA |
455 | /* |
456 | * Ensure that ->deadline is visible before set the started | |
457 | * flag and clear the completed flag. | |
458 | */ | |
459 | smp_mb__before_atomic(); | |
460 | ||
87ee7b11 JA |
461 | /* |
462 | * Mark us as started and clear complete. Complete might have been | |
463 | * set if requeue raced with timeout, which then marked it as | |
464 | * complete. So be sure to clear complete again when we start | |
465 | * the request, otherwise we'll ignore the completion event. | |
466 | */ | |
4b570521 JA |
467 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
468 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
469 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
470 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
471 | |
472 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
473 | /* | |
474 | * Make sure space for the drain appears. We know we can do | |
475 | * this because max_hw_segments has been adjusted to be one | |
476 | * fewer than the device can handle. | |
477 | */ | |
478 | rq->nr_phys_segments++; | |
479 | } | |
320ae51f | 480 | } |
e2490073 | 481 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 482 | |
ed0791b2 | 483 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
484 | { |
485 | struct request_queue *q = rq->q; | |
486 | ||
487 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 488 | |
e2490073 CH |
489 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
490 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
491 | rq->nr_phys_segments--; | |
492 | } | |
320ae51f JA |
493 | } |
494 | ||
2b053aca | 495 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 496 | { |
ed0791b2 | 497 | __blk_mq_requeue_request(rq); |
ed0791b2 | 498 | |
ed0791b2 | 499 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 500 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
501 | } |
502 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
503 | ||
6fca6a61 CH |
504 | static void blk_mq_requeue_work(struct work_struct *work) |
505 | { | |
506 | struct request_queue *q = | |
2849450a | 507 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
508 | LIST_HEAD(rq_list); |
509 | struct request *rq, *next; | |
510 | unsigned long flags; | |
511 | ||
512 | spin_lock_irqsave(&q->requeue_lock, flags); | |
513 | list_splice_init(&q->requeue_list, &rq_list); | |
514 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
515 | ||
516 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 517 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
518 | continue; |
519 | ||
e8064021 | 520 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 CH |
521 | list_del_init(&rq->queuelist); |
522 | blk_mq_insert_request(rq, true, false, false); | |
523 | } | |
524 | ||
525 | while (!list_empty(&rq_list)) { | |
526 | rq = list_entry(rq_list.next, struct request, queuelist); | |
527 | list_del_init(&rq->queuelist); | |
528 | blk_mq_insert_request(rq, false, false, false); | |
529 | } | |
530 | ||
52d7f1b5 | 531 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
532 | } |
533 | ||
2b053aca BVA |
534 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
535 | bool kick_requeue_list) | |
6fca6a61 CH |
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 | */ | |
e8064021 | 544 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
545 | |
546 | spin_lock_irqsave(&q->requeue_lock, flags); | |
547 | if (at_head) { | |
e8064021 | 548 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
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); | |
2b053aca BVA |
554 | |
555 | if (kick_requeue_list) | |
556 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
557 | } |
558 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
559 | ||
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 | ||
88459642 OS |
754 | struct flush_busy_ctx_data { |
755 | struct blk_mq_hw_ctx *hctx; | |
756 | struct list_head *list; | |
757 | }; | |
758 | ||
759 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
760 | { | |
761 | struct flush_busy_ctx_data *flush_data = data; | |
762 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
763 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
764 | ||
765 | sbitmap_clear_bit(sb, bitnr); | |
766 | spin_lock(&ctx->lock); | |
767 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
768 | spin_unlock(&ctx->lock); | |
769 | return true; | |
770 | } | |
771 | ||
1429d7c9 JA |
772 | /* |
773 | * Process software queues that have been marked busy, splicing them | |
774 | * to the for-dispatch | |
775 | */ | |
776 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
777 | { | |
88459642 OS |
778 | struct flush_busy_ctx_data data = { |
779 | .hctx = hctx, | |
780 | .list = list, | |
781 | }; | |
1429d7c9 | 782 | |
88459642 | 783 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 784 | } |
1429d7c9 | 785 | |
703fd1c0 JA |
786 | static inline unsigned int queued_to_index(unsigned int queued) |
787 | { | |
788 | if (!queued) | |
789 | return 0; | |
1429d7c9 | 790 | |
703fd1c0 | 791 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
792 | } |
793 | ||
320ae51f JA |
794 | /* |
795 | * Run this hardware queue, pulling any software queues mapped to it in. | |
796 | * Note that this function currently has various problems around ordering | |
797 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
798 | * items on the hctx->dispatch list. Ignore that for now. | |
799 | */ | |
6a83e74d | 800 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) |
320ae51f JA |
801 | { |
802 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
803 | struct request *rq; |
804 | LIST_HEAD(rq_list); | |
74c45052 JA |
805 | LIST_HEAD(driver_list); |
806 | struct list_head *dptr; | |
1429d7c9 | 807 | int queued; |
320ae51f | 808 | |
5d1b25c1 | 809 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
810 | return; |
811 | ||
812 | hctx->run++; | |
813 | ||
814 | /* | |
815 | * Touch any software queue that has pending entries. | |
816 | */ | |
1429d7c9 | 817 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
818 | |
819 | /* | |
820 | * If we have previous entries on our dispatch list, grab them | |
821 | * and stuff them at the front for more fair dispatch. | |
822 | */ | |
823 | if (!list_empty_careful(&hctx->dispatch)) { | |
824 | spin_lock(&hctx->lock); | |
825 | if (!list_empty(&hctx->dispatch)) | |
826 | list_splice_init(&hctx->dispatch, &rq_list); | |
827 | spin_unlock(&hctx->lock); | |
828 | } | |
829 | ||
74c45052 JA |
830 | /* |
831 | * Start off with dptr being NULL, so we start the first request | |
832 | * immediately, even if we have more pending. | |
833 | */ | |
834 | dptr = NULL; | |
835 | ||
320ae51f JA |
836 | /* |
837 | * Now process all the entries, sending them to the driver. | |
838 | */ | |
1429d7c9 | 839 | queued = 0; |
320ae51f | 840 | while (!list_empty(&rq_list)) { |
74c45052 | 841 | struct blk_mq_queue_data bd; |
320ae51f JA |
842 | int ret; |
843 | ||
844 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
845 | list_del_init(&rq->queuelist); | |
320ae51f | 846 | |
74c45052 JA |
847 | bd.rq = rq; |
848 | bd.list = dptr; | |
849 | bd.last = list_empty(&rq_list); | |
850 | ||
851 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
852 | switch (ret) { |
853 | case BLK_MQ_RQ_QUEUE_OK: | |
854 | queued++; | |
52b9c330 | 855 | break; |
320ae51f | 856 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 857 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 858 | __blk_mq_requeue_request(rq); |
320ae51f JA |
859 | break; |
860 | default: | |
861 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 862 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 863 | rq->errors = -EIO; |
c8a446ad | 864 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
865 | break; |
866 | } | |
867 | ||
868 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
869 | break; | |
74c45052 JA |
870 | |
871 | /* | |
872 | * We've done the first request. If we have more than 1 | |
873 | * left in the list, set dptr to defer issue. | |
874 | */ | |
875 | if (!dptr && rq_list.next != rq_list.prev) | |
876 | dptr = &driver_list; | |
320ae51f JA |
877 | } |
878 | ||
703fd1c0 | 879 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
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 | ||
6a83e74d BVA |
902 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
903 | { | |
904 | int srcu_idx; | |
905 | ||
906 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
907 | cpu_online(hctx->next_cpu)); | |
908 | ||
909 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
910 | rcu_read_lock(); | |
911 | blk_mq_process_rq_list(hctx); | |
912 | rcu_read_unlock(); | |
913 | } else { | |
914 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
915 | blk_mq_process_rq_list(hctx); | |
916 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
917 | } | |
918 | } | |
919 | ||
506e931f JA |
920 | /* |
921 | * It'd be great if the workqueue API had a way to pass | |
922 | * in a mask and had some smarts for more clever placement. | |
923 | * For now we just round-robin here, switching for every | |
924 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
925 | */ | |
926 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
927 | { | |
b657d7e6 CH |
928 | if (hctx->queue->nr_hw_queues == 1) |
929 | return WORK_CPU_UNBOUND; | |
506e931f JA |
930 | |
931 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 932 | int next_cpu; |
506e931f JA |
933 | |
934 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
935 | if (next_cpu >= nr_cpu_ids) | |
936 | next_cpu = cpumask_first(hctx->cpumask); | |
937 | ||
938 | hctx->next_cpu = next_cpu; | |
939 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
940 | } | |
941 | ||
b657d7e6 | 942 | return hctx->next_cpu; |
506e931f JA |
943 | } |
944 | ||
320ae51f JA |
945 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
946 | { | |
5d1b25c1 BVA |
947 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
948 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
949 | return; |
950 | ||
1b792f2f | 951 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
952 | int cpu = get_cpu(); |
953 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 954 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 955 | put_cpu(); |
398205b8 PB |
956 | return; |
957 | } | |
e4043dcf | 958 | |
2a90d4aa | 959 | put_cpu(); |
e4043dcf | 960 | } |
398205b8 | 961 | |
27489a3c | 962 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
963 | } |
964 | ||
b94ec296 | 965 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
966 | { |
967 | struct blk_mq_hw_ctx *hctx; | |
968 | int i; | |
969 | ||
970 | queue_for_each_hw_ctx(q, hctx, i) { | |
971 | if ((!blk_mq_hctx_has_pending(hctx) && | |
972 | list_empty_careful(&hctx->dispatch)) || | |
5d1b25c1 | 973 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
974 | continue; |
975 | ||
b94ec296 | 976 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
977 | } |
978 | } | |
b94ec296 | 979 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 980 | |
fd001443 BVA |
981 | /** |
982 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
983 | * @q: request queue. | |
984 | * | |
985 | * The caller is responsible for serializing this function against | |
986 | * blk_mq_{start,stop}_hw_queue(). | |
987 | */ | |
988 | bool blk_mq_queue_stopped(struct request_queue *q) | |
989 | { | |
990 | struct blk_mq_hw_ctx *hctx; | |
991 | int i; | |
992 | ||
993 | queue_for_each_hw_ctx(q, hctx, i) | |
994 | if (blk_mq_hctx_stopped(hctx)) | |
995 | return true; | |
996 | ||
997 | return false; | |
998 | } | |
999 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1000 | ||
320ae51f JA |
1001 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1002 | { | |
27489a3c | 1003 | cancel_work(&hctx->run_work); |
70f4db63 | 1004 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1005 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1006 | } | |
1007 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1008 | ||
280d45f6 CH |
1009 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1010 | { | |
1011 | struct blk_mq_hw_ctx *hctx; | |
1012 | int i; | |
1013 | ||
1014 | queue_for_each_hw_ctx(q, hctx, i) | |
1015 | blk_mq_stop_hw_queue(hctx); | |
1016 | } | |
1017 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1018 | ||
320ae51f JA |
1019 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1020 | { | |
1021 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1022 | |
0ffbce80 | 1023 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1024 | } |
1025 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1026 | ||
2f268556 CH |
1027 | void blk_mq_start_hw_queues(struct request_queue *q) |
1028 | { | |
1029 | struct blk_mq_hw_ctx *hctx; | |
1030 | int i; | |
1031 | ||
1032 | queue_for_each_hw_ctx(q, hctx, i) | |
1033 | blk_mq_start_hw_queue(hctx); | |
1034 | } | |
1035 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1036 | ||
1b4a3258 | 1037 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1038 | { |
1039 | struct blk_mq_hw_ctx *hctx; | |
1040 | int i; | |
1041 | ||
1042 | queue_for_each_hw_ctx(q, hctx, i) { | |
5d1b25c1 | 1043 | if (!blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1044 | continue; |
1045 | ||
1046 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1047 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1048 | } |
1049 | } | |
1050 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1051 | ||
70f4db63 | 1052 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1053 | { |
1054 | struct blk_mq_hw_ctx *hctx; | |
1055 | ||
27489a3c | 1056 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1057 | |
320ae51f JA |
1058 | __blk_mq_run_hw_queue(hctx); |
1059 | } | |
1060 | ||
70f4db63 CH |
1061 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1062 | { | |
1063 | struct blk_mq_hw_ctx *hctx; | |
1064 | ||
1065 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1066 | ||
1067 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1068 | __blk_mq_run_hw_queue(hctx); | |
1069 | } | |
1070 | ||
1071 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1072 | { | |
19c66e59 ML |
1073 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1074 | return; | |
70f4db63 | 1075 | |
b657d7e6 CH |
1076 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1077 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1078 | } |
1079 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1080 | ||
cfd0c552 | 1081 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1082 | struct request *rq, |
1083 | bool at_head) | |
320ae51f | 1084 | { |
e57690fe JA |
1085 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1086 | ||
01b983c9 JA |
1087 | trace_block_rq_insert(hctx->queue, rq); |
1088 | ||
72a0a36e CH |
1089 | if (at_head) |
1090 | list_add(&rq->queuelist, &ctx->rq_list); | |
1091 | else | |
1092 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1093 | } |
4bb659b1 | 1094 | |
cfd0c552 ML |
1095 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1096 | struct request *rq, bool at_head) | |
1097 | { | |
1098 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1099 | ||
e57690fe | 1100 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1101 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1102 | } |
1103 | ||
eeabc850 | 1104 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1105 | bool async) |
320ae51f | 1106 | { |
e57690fe | 1107 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1108 | struct request_queue *q = rq->q; |
7d7e0f90 | 1109 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1110 | |
a57a178a CH |
1111 | spin_lock(&ctx->lock); |
1112 | __blk_mq_insert_request(hctx, rq, at_head); | |
1113 | spin_unlock(&ctx->lock); | |
320ae51f | 1114 | |
320ae51f JA |
1115 | if (run_queue) |
1116 | blk_mq_run_hw_queue(hctx, async); | |
1117 | } | |
1118 | ||
1119 | static void blk_mq_insert_requests(struct request_queue *q, | |
1120 | struct blk_mq_ctx *ctx, | |
1121 | struct list_head *list, | |
1122 | int depth, | |
1123 | bool from_schedule) | |
1124 | ||
1125 | { | |
7d7e0f90 | 1126 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1127 | |
1128 | trace_block_unplug(q, depth, !from_schedule); | |
1129 | ||
320ae51f JA |
1130 | /* |
1131 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1132 | * offline now | |
1133 | */ | |
1134 | spin_lock(&ctx->lock); | |
1135 | while (!list_empty(list)) { | |
1136 | struct request *rq; | |
1137 | ||
1138 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1139 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1140 | list_del_init(&rq->queuelist); |
e57690fe | 1141 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1142 | } |
cfd0c552 | 1143 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1144 | spin_unlock(&ctx->lock); |
1145 | ||
320ae51f JA |
1146 | blk_mq_run_hw_queue(hctx, from_schedule); |
1147 | } | |
1148 | ||
1149 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1150 | { | |
1151 | struct request *rqa = container_of(a, struct request, queuelist); | |
1152 | struct request *rqb = container_of(b, struct request, queuelist); | |
1153 | ||
1154 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1155 | (rqa->mq_ctx == rqb->mq_ctx && | |
1156 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1157 | } | |
1158 | ||
1159 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1160 | { | |
1161 | struct blk_mq_ctx *this_ctx; | |
1162 | struct request_queue *this_q; | |
1163 | struct request *rq; | |
1164 | LIST_HEAD(list); | |
1165 | LIST_HEAD(ctx_list); | |
1166 | unsigned int depth; | |
1167 | ||
1168 | list_splice_init(&plug->mq_list, &list); | |
1169 | ||
1170 | list_sort(NULL, &list, plug_ctx_cmp); | |
1171 | ||
1172 | this_q = NULL; | |
1173 | this_ctx = NULL; | |
1174 | depth = 0; | |
1175 | ||
1176 | while (!list_empty(&list)) { | |
1177 | rq = list_entry_rq(list.next); | |
1178 | list_del_init(&rq->queuelist); | |
1179 | BUG_ON(!rq->q); | |
1180 | if (rq->mq_ctx != this_ctx) { | |
1181 | if (this_ctx) { | |
1182 | blk_mq_insert_requests(this_q, this_ctx, | |
1183 | &ctx_list, depth, | |
1184 | from_schedule); | |
1185 | } | |
1186 | ||
1187 | this_ctx = rq->mq_ctx; | |
1188 | this_q = rq->q; | |
1189 | depth = 0; | |
1190 | } | |
1191 | ||
1192 | depth++; | |
1193 | list_add_tail(&rq->queuelist, &ctx_list); | |
1194 | } | |
1195 | ||
1196 | /* | |
1197 | * If 'this_ctx' is set, we know we have entries to complete | |
1198 | * on 'ctx_list'. Do those. | |
1199 | */ | |
1200 | if (this_ctx) { | |
1201 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1202 | from_schedule); | |
1203 | } | |
1204 | } | |
1205 | ||
1206 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1207 | { | |
1208 | init_request_from_bio(rq, bio); | |
4b570521 | 1209 | |
a21f2a3e | 1210 | blk_account_io_start(rq, 1); |
320ae51f JA |
1211 | } |
1212 | ||
274a5843 JA |
1213 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1214 | { | |
1215 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1216 | !blk_queue_nomerges(hctx->queue); | |
1217 | } | |
1218 | ||
07068d5b JA |
1219 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1220 | struct blk_mq_ctx *ctx, | |
1221 | struct request *rq, struct bio *bio) | |
320ae51f | 1222 | { |
e18378a6 | 1223 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1224 | blk_mq_bio_to_request(rq, bio); |
1225 | spin_lock(&ctx->lock); | |
1226 | insert_rq: | |
1227 | __blk_mq_insert_request(hctx, rq, false); | |
1228 | spin_unlock(&ctx->lock); | |
1229 | return false; | |
1230 | } else { | |
274a5843 JA |
1231 | struct request_queue *q = hctx->queue; |
1232 | ||
07068d5b JA |
1233 | spin_lock(&ctx->lock); |
1234 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1235 | blk_mq_bio_to_request(rq, bio); | |
1236 | goto insert_rq; | |
1237 | } | |
320ae51f | 1238 | |
07068d5b JA |
1239 | spin_unlock(&ctx->lock); |
1240 | __blk_mq_free_request(hctx, ctx, rq); | |
1241 | return true; | |
14ec77f3 | 1242 | } |
07068d5b | 1243 | } |
14ec77f3 | 1244 | |
07068d5b JA |
1245 | static struct request *blk_mq_map_request(struct request_queue *q, |
1246 | struct bio *bio, | |
2552e3f8 | 1247 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1248 | { |
1249 | struct blk_mq_hw_ctx *hctx; | |
1250 | struct blk_mq_ctx *ctx; | |
1251 | struct request *rq; | |
320ae51f | 1252 | |
3ef28e83 | 1253 | blk_queue_enter_live(q); |
320ae51f | 1254 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1255 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1256 | |
ef295ecf | 1257 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1258 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1259 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1260 | |
7dd2fb68 | 1261 | data->hctx->queued++; |
07068d5b JA |
1262 | return rq; |
1263 | } | |
1264 | ||
2253efc8 BVA |
1265 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1266 | struct request *rq, blk_qc_t *cookie) | |
f984df1f SL |
1267 | { |
1268 | int ret; | |
1269 | struct request_queue *q = rq->q; | |
f984df1f SL |
1270 | struct blk_mq_queue_data bd = { |
1271 | .rq = rq, | |
1272 | .list = NULL, | |
1273 | .last = 1 | |
1274 | }; | |
7b371636 | 1275 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f | 1276 | |
2253efc8 BVA |
1277 | if (blk_mq_hctx_stopped(hctx)) |
1278 | goto insert; | |
1279 | ||
f984df1f SL |
1280 | /* |
1281 | * For OK queue, we are done. For error, kill it. Any other | |
1282 | * error (busy), just add it to our list as we previously | |
1283 | * would have done | |
1284 | */ | |
1285 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1286 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1287 | *cookie = new_cookie; | |
2253efc8 | 1288 | return; |
7b371636 | 1289 | } |
f984df1f | 1290 | |
7b371636 JA |
1291 | __blk_mq_requeue_request(rq); |
1292 | ||
1293 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1294 | *cookie = BLK_QC_T_NONE; | |
1295 | rq->errors = -EIO; | |
1296 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1297 | return; |
f984df1f | 1298 | } |
7b371636 | 1299 | |
2253efc8 BVA |
1300 | insert: |
1301 | blk_mq_insert_request(rq, false, true, true); | |
f984df1f SL |
1302 | } |
1303 | ||
07068d5b JA |
1304 | /* |
1305 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1306 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1307 | * hardware for SYNC IO. | |
1308 | */ | |
dece1635 | 1309 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1310 | { |
ef295ecf | 1311 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1312 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1313 | struct blk_mq_alloc_data data; |
07068d5b | 1314 | struct request *rq; |
6a83e74d | 1315 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1316 | struct blk_plug *plug; |
5b3f341f | 1317 | struct request *same_queue_rq = NULL; |
7b371636 | 1318 | blk_qc_t cookie; |
07068d5b JA |
1319 | |
1320 | blk_queue_bounce(q, &bio); | |
1321 | ||
1322 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1323 | bio_io_error(bio); |
dece1635 | 1324 | return BLK_QC_T_NONE; |
07068d5b JA |
1325 | } |
1326 | ||
54efd50b KO |
1327 | blk_queue_split(q, &bio, q->bio_split); |
1328 | ||
87c279e6 OS |
1329 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1330 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1331 | return BLK_QC_T_NONE; | |
f984df1f | 1332 | |
07068d5b JA |
1333 | rq = blk_mq_map_request(q, bio, &data); |
1334 | if (unlikely(!rq)) | |
dece1635 | 1335 | return BLK_QC_T_NONE; |
07068d5b | 1336 | |
7b371636 | 1337 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1338 | |
1339 | if (unlikely(is_flush_fua)) { | |
1340 | blk_mq_bio_to_request(rq, bio); | |
1341 | blk_insert_flush(rq); | |
1342 | goto run_queue; | |
1343 | } | |
1344 | ||
f984df1f | 1345 | plug = current->plug; |
e167dfb5 JA |
1346 | /* |
1347 | * If the driver supports defer issued based on 'last', then | |
1348 | * queue it up like normal since we can potentially save some | |
1349 | * CPU this way. | |
1350 | */ | |
f984df1f SL |
1351 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1352 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1353 | struct request *old_rq = NULL; | |
07068d5b JA |
1354 | |
1355 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1356 | |
1357 | /* | |
6a83e74d | 1358 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1359 | * Otherwise the existing request in the plug list will be |
1360 | * issued. So the plug list will have one request at most | |
07068d5b | 1361 | */ |
f984df1f | 1362 | if (plug) { |
5b3f341f SL |
1363 | /* |
1364 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1365 | * happens, same_queue_rq is invalid and plug list is |
1366 | * empty | |
1367 | */ | |
5b3f341f SL |
1368 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1369 | old_rq = same_queue_rq; | |
f984df1f | 1370 | list_del_init(&old_rq->queuelist); |
07068d5b | 1371 | } |
f984df1f SL |
1372 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1373 | } else /* is_sync */ | |
1374 | old_rq = rq; | |
1375 | blk_mq_put_ctx(data.ctx); | |
1376 | if (!old_rq) | |
7b371636 | 1377 | goto done; |
6a83e74d BVA |
1378 | |
1379 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1380 | rcu_read_lock(); | |
1381 | blk_mq_try_issue_directly(data.hctx, old_rq, &cookie); | |
1382 | rcu_read_unlock(); | |
1383 | } else { | |
1384 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
1385 | blk_mq_try_issue_directly(data.hctx, old_rq, &cookie); | |
1386 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); | |
1387 | } | |
7b371636 | 1388 | goto done; |
07068d5b JA |
1389 | } |
1390 | ||
1391 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1392 | /* | |
1393 | * For a SYNC request, send it to the hardware immediately. For | |
1394 | * an ASYNC request, just ensure that we run it later on. The | |
1395 | * latter allows for merging opportunities and more efficient | |
1396 | * dispatching. | |
1397 | */ | |
1398 | run_queue: | |
1399 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1400 | } | |
07068d5b | 1401 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1402 | done: |
1403 | return cookie; | |
07068d5b JA |
1404 | } |
1405 | ||
1406 | /* | |
1407 | * Single hardware queue variant. This will attempt to use any per-process | |
1408 | * plug for merging and IO deferral. | |
1409 | */ | |
dece1635 | 1410 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1411 | { |
ef295ecf | 1412 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1413 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1414 | struct blk_plug *plug; |
1415 | unsigned int request_count = 0; | |
2552e3f8 | 1416 | struct blk_mq_alloc_data data; |
07068d5b | 1417 | struct request *rq; |
7b371636 | 1418 | blk_qc_t cookie; |
07068d5b | 1419 | |
07068d5b JA |
1420 | blk_queue_bounce(q, &bio); |
1421 | ||
1422 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1423 | bio_io_error(bio); |
dece1635 | 1424 | return BLK_QC_T_NONE; |
07068d5b JA |
1425 | } |
1426 | ||
54efd50b KO |
1427 | blk_queue_split(q, &bio, q->bio_split); |
1428 | ||
87c279e6 OS |
1429 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1430 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1431 | return BLK_QC_T_NONE; | |
1432 | } else | |
1433 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1434 | |
1435 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1436 | if (unlikely(!rq)) |
dece1635 | 1437 | return BLK_QC_T_NONE; |
320ae51f | 1438 | |
7b371636 | 1439 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1440 | |
1441 | if (unlikely(is_flush_fua)) { | |
1442 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1443 | blk_insert_flush(rq); |
1444 | goto run_queue; | |
1445 | } | |
1446 | ||
1447 | /* | |
1448 | * A task plug currently exists. Since this is completely lockless, | |
1449 | * utilize that to temporarily store requests until the task is | |
1450 | * either done or scheduled away. | |
1451 | */ | |
e6c4438b JM |
1452 | plug = current->plug; |
1453 | if (plug) { | |
600271d9 SL |
1454 | struct request *last = NULL; |
1455 | ||
e6c4438b | 1456 | blk_mq_bio_to_request(rq, bio); |
676d0607 | 1457 | if (!request_count) |
e6c4438b | 1458 | trace_block_plug(q); |
600271d9 SL |
1459 | else |
1460 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1461 | |
1462 | blk_mq_put_ctx(data.ctx); | |
1463 | ||
600271d9 SL |
1464 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1465 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1466 | blk_flush_plug_list(plug, false); |
1467 | trace_block_plug(q); | |
320ae51f | 1468 | } |
b094f89c | 1469 | |
e6c4438b | 1470 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1471 | return cookie; |
320ae51f JA |
1472 | } |
1473 | ||
07068d5b JA |
1474 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1475 | /* | |
1476 | * For a SYNC request, send it to the hardware immediately. For | |
1477 | * an ASYNC request, just ensure that we run it later on. The | |
1478 | * latter allows for merging opportunities and more efficient | |
1479 | * dispatching. | |
1480 | */ | |
1481 | run_queue: | |
1482 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1483 | } |
1484 | ||
07068d5b | 1485 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1486 | return cookie; |
320ae51f JA |
1487 | } |
1488 | ||
24d2f903 CH |
1489 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1490 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1491 | { |
e9b267d9 | 1492 | struct page *page; |
320ae51f | 1493 | |
24d2f903 | 1494 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1495 | int i; |
320ae51f | 1496 | |
24d2f903 CH |
1497 | for (i = 0; i < tags->nr_tags; i++) { |
1498 | if (!tags->rqs[i]) | |
e9b267d9 | 1499 | continue; |
24d2f903 CH |
1500 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1501 | hctx_idx, i); | |
a5164405 | 1502 | tags->rqs[i] = NULL; |
e9b267d9 | 1503 | } |
320ae51f | 1504 | } |
320ae51f | 1505 | |
24d2f903 CH |
1506 | while (!list_empty(&tags->page_list)) { |
1507 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1508 | list_del_init(&page->lru); |
f75782e4 CM |
1509 | /* |
1510 | * Remove kmemleak object previously allocated in | |
1511 | * blk_mq_init_rq_map(). | |
1512 | */ | |
1513 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1514 | __free_pages(page, page->private); |
1515 | } | |
1516 | ||
24d2f903 | 1517 | kfree(tags->rqs); |
320ae51f | 1518 | |
24d2f903 | 1519 | blk_mq_free_tags(tags); |
320ae51f JA |
1520 | } |
1521 | ||
1522 | static size_t order_to_size(unsigned int order) | |
1523 | { | |
4ca08500 | 1524 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1525 | } |
1526 | ||
24d2f903 CH |
1527 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1528 | unsigned int hctx_idx) | |
320ae51f | 1529 | { |
24d2f903 | 1530 | struct blk_mq_tags *tags; |
320ae51f JA |
1531 | unsigned int i, j, entries_per_page, max_order = 4; |
1532 | size_t rq_size, left; | |
1533 | ||
24d2f903 | 1534 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1535 | set->numa_node, |
1536 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1537 | if (!tags) |
1538 | return NULL; | |
320ae51f | 1539 | |
24d2f903 CH |
1540 | INIT_LIST_HEAD(&tags->page_list); |
1541 | ||
a5164405 JA |
1542 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1543 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1544 | set->numa_node); | |
24d2f903 CH |
1545 | if (!tags->rqs) { |
1546 | blk_mq_free_tags(tags); | |
1547 | return NULL; | |
1548 | } | |
320ae51f JA |
1549 | |
1550 | /* | |
1551 | * rq_size is the size of the request plus driver payload, rounded | |
1552 | * to the cacheline size | |
1553 | */ | |
24d2f903 | 1554 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1555 | cache_line_size()); |
24d2f903 | 1556 | left = rq_size * set->queue_depth; |
320ae51f | 1557 | |
24d2f903 | 1558 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1559 | int this_order = max_order; |
1560 | struct page *page; | |
1561 | int to_do; | |
1562 | void *p; | |
1563 | ||
b3a834b1 | 1564 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1565 | this_order--; |
1566 | ||
1567 | do { | |
a5164405 | 1568 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1569 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1570 | this_order); |
320ae51f JA |
1571 | if (page) |
1572 | break; | |
1573 | if (!this_order--) | |
1574 | break; | |
1575 | if (order_to_size(this_order) < rq_size) | |
1576 | break; | |
1577 | } while (1); | |
1578 | ||
1579 | if (!page) | |
24d2f903 | 1580 | goto fail; |
320ae51f JA |
1581 | |
1582 | page->private = this_order; | |
24d2f903 | 1583 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1584 | |
1585 | p = page_address(page); | |
f75782e4 CM |
1586 | /* |
1587 | * Allow kmemleak to scan these pages as they contain pointers | |
1588 | * to additional allocations like via ops->init_request(). | |
1589 | */ | |
1590 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1591 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1592 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1593 | left -= to_do * rq_size; |
1594 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1595 | tags->rqs[i] = p; |
1596 | if (set->ops->init_request) { | |
1597 | if (set->ops->init_request(set->driver_data, | |
1598 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1599 | set->numa_node)) { |
1600 | tags->rqs[i] = NULL; | |
24d2f903 | 1601 | goto fail; |
a5164405 | 1602 | } |
e9b267d9 CH |
1603 | } |
1604 | ||
320ae51f JA |
1605 | p += rq_size; |
1606 | i++; | |
1607 | } | |
1608 | } | |
24d2f903 | 1609 | return tags; |
320ae51f | 1610 | |
24d2f903 | 1611 | fail: |
24d2f903 CH |
1612 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1613 | return NULL; | |
320ae51f JA |
1614 | } |
1615 | ||
e57690fe JA |
1616 | /* |
1617 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1618 | * software queue to the hw queue dispatch list, and ensure that it | |
1619 | * gets run. | |
1620 | */ | |
9467f859 | 1621 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1622 | { |
9467f859 | 1623 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1624 | struct blk_mq_ctx *ctx; |
1625 | LIST_HEAD(tmp); | |
1626 | ||
9467f859 | 1627 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1628 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1629 | |
1630 | spin_lock(&ctx->lock); | |
1631 | if (!list_empty(&ctx->rq_list)) { | |
1632 | list_splice_init(&ctx->rq_list, &tmp); | |
1633 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1634 | } | |
1635 | spin_unlock(&ctx->lock); | |
1636 | ||
1637 | if (list_empty(&tmp)) | |
9467f859 | 1638 | return 0; |
484b4061 | 1639 | |
e57690fe JA |
1640 | spin_lock(&hctx->lock); |
1641 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1642 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1643 | |
1644 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1645 | return 0; |
484b4061 JA |
1646 | } |
1647 | ||
9467f859 | 1648 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1649 | { |
9467f859 TG |
1650 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1651 | &hctx->cpuhp_dead); | |
484b4061 JA |
1652 | } |
1653 | ||
c3b4afca | 1654 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1655 | static void blk_mq_exit_hctx(struct request_queue *q, |
1656 | struct blk_mq_tag_set *set, | |
1657 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1658 | { | |
f70ced09 ML |
1659 | unsigned flush_start_tag = set->queue_depth; |
1660 | ||
08e98fc6 ML |
1661 | blk_mq_tag_idle(hctx); |
1662 | ||
f70ced09 ML |
1663 | if (set->ops->exit_request) |
1664 | set->ops->exit_request(set->driver_data, | |
1665 | hctx->fq->flush_rq, hctx_idx, | |
1666 | flush_start_tag + hctx_idx); | |
1667 | ||
08e98fc6 ML |
1668 | if (set->ops->exit_hctx) |
1669 | set->ops->exit_hctx(hctx, hctx_idx); | |
1670 | ||
6a83e74d BVA |
1671 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1672 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1673 | ||
9467f859 | 1674 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1675 | blk_free_flush_queue(hctx->fq); |
88459642 | 1676 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1677 | } |
1678 | ||
624dbe47 ML |
1679 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1680 | struct blk_mq_tag_set *set, int nr_queue) | |
1681 | { | |
1682 | struct blk_mq_hw_ctx *hctx; | |
1683 | unsigned int i; | |
1684 | ||
1685 | queue_for_each_hw_ctx(q, hctx, i) { | |
1686 | if (i == nr_queue) | |
1687 | break; | |
08e98fc6 | 1688 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1689 | } |
624dbe47 ML |
1690 | } |
1691 | ||
1692 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1693 | struct blk_mq_tag_set *set) | |
1694 | { | |
1695 | struct blk_mq_hw_ctx *hctx; | |
1696 | unsigned int i; | |
1697 | ||
e09aae7e | 1698 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1699 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1700 | } |
1701 | ||
08e98fc6 ML |
1702 | static int blk_mq_init_hctx(struct request_queue *q, |
1703 | struct blk_mq_tag_set *set, | |
1704 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1705 | { |
08e98fc6 | 1706 | int node; |
f70ced09 | 1707 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1708 | |
1709 | node = hctx->numa_node; | |
1710 | if (node == NUMA_NO_NODE) | |
1711 | node = hctx->numa_node = set->numa_node; | |
1712 | ||
27489a3c | 1713 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1714 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1715 | spin_lock_init(&hctx->lock); | |
1716 | INIT_LIST_HEAD(&hctx->dispatch); | |
1717 | hctx->queue = q; | |
1718 | hctx->queue_num = hctx_idx; | |
2404e607 | 1719 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1720 | |
9467f859 | 1721 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1722 | |
1723 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1724 | |
1725 | /* | |
08e98fc6 ML |
1726 | * Allocate space for all possible cpus to avoid allocation at |
1727 | * runtime | |
320ae51f | 1728 | */ |
08e98fc6 ML |
1729 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1730 | GFP_KERNEL, node); | |
1731 | if (!hctx->ctxs) | |
1732 | goto unregister_cpu_notifier; | |
320ae51f | 1733 | |
88459642 OS |
1734 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1735 | node)) | |
08e98fc6 | 1736 | goto free_ctxs; |
320ae51f | 1737 | |
08e98fc6 | 1738 | hctx->nr_ctx = 0; |
320ae51f | 1739 | |
08e98fc6 ML |
1740 | if (set->ops->init_hctx && |
1741 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1742 | goto free_bitmap; | |
320ae51f | 1743 | |
f70ced09 ML |
1744 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1745 | if (!hctx->fq) | |
1746 | goto exit_hctx; | |
320ae51f | 1747 | |
f70ced09 ML |
1748 | if (set->ops->init_request && |
1749 | set->ops->init_request(set->driver_data, | |
1750 | hctx->fq->flush_rq, hctx_idx, | |
1751 | flush_start_tag + hctx_idx, node)) | |
1752 | goto free_fq; | |
320ae51f | 1753 | |
6a83e74d BVA |
1754 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1755 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1756 | ||
08e98fc6 | 1757 | return 0; |
320ae51f | 1758 | |
f70ced09 ML |
1759 | free_fq: |
1760 | kfree(hctx->fq); | |
1761 | exit_hctx: | |
1762 | if (set->ops->exit_hctx) | |
1763 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1764 | free_bitmap: |
88459642 | 1765 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1766 | free_ctxs: |
1767 | kfree(hctx->ctxs); | |
1768 | unregister_cpu_notifier: | |
9467f859 | 1769 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1770 | return -1; |
1771 | } | |
320ae51f | 1772 | |
320ae51f JA |
1773 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1774 | unsigned int nr_hw_queues) | |
1775 | { | |
1776 | unsigned int i; | |
1777 | ||
1778 | for_each_possible_cpu(i) { | |
1779 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1780 | struct blk_mq_hw_ctx *hctx; | |
1781 | ||
1782 | memset(__ctx, 0, sizeof(*__ctx)); | |
1783 | __ctx->cpu = i; | |
1784 | spin_lock_init(&__ctx->lock); | |
1785 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1786 | __ctx->queue = q; | |
1787 | ||
1788 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1789 | if (!cpu_online(i)) |
1790 | continue; | |
1791 | ||
7d7e0f90 | 1792 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1793 | |
320ae51f JA |
1794 | /* |
1795 | * Set local node, IFF we have more than one hw queue. If | |
1796 | * not, we remain on the home node of the device | |
1797 | */ | |
1798 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1799 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1800 | } |
1801 | } | |
1802 | ||
5778322e AM |
1803 | static void blk_mq_map_swqueue(struct request_queue *q, |
1804 | const struct cpumask *online_mask) | |
320ae51f JA |
1805 | { |
1806 | unsigned int i; | |
1807 | struct blk_mq_hw_ctx *hctx; | |
1808 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1809 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1810 | |
60de074b AM |
1811 | /* |
1812 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1813 | */ | |
1814 | mutex_lock(&q->sysfs_lock); | |
1815 | ||
320ae51f | 1816 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1817 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1818 | hctx->nr_ctx = 0; |
1819 | } | |
1820 | ||
1821 | /* | |
1822 | * Map software to hardware queues | |
1823 | */ | |
897bb0c7 | 1824 | for_each_possible_cpu(i) { |
320ae51f | 1825 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1826 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1827 | continue; |
1828 | ||
897bb0c7 | 1829 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1830 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1831 | |
e4043dcf | 1832 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1833 | ctx->index_hw = hctx->nr_ctx; |
1834 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1835 | } | |
506e931f | 1836 | |
60de074b AM |
1837 | mutex_unlock(&q->sysfs_lock); |
1838 | ||
506e931f | 1839 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1840 | /* |
a68aafa5 JA |
1841 | * If no software queues are mapped to this hardware queue, |
1842 | * disable it and free the request entries. | |
484b4061 JA |
1843 | */ |
1844 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1845 | if (set->tags[i]) { |
1846 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1847 | set->tags[i] = NULL; | |
484b4061 | 1848 | } |
2a34c087 | 1849 | hctx->tags = NULL; |
484b4061 JA |
1850 | continue; |
1851 | } | |
1852 | ||
2a34c087 ML |
1853 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1854 | if (!set->tags[i]) | |
1855 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1856 | hctx->tags = set->tags[i]; | |
1857 | WARN_ON(!hctx->tags); | |
1858 | ||
889fa31f CY |
1859 | /* |
1860 | * Set the map size to the number of mapped software queues. | |
1861 | * This is more accurate and more efficient than looping | |
1862 | * over all possibly mapped software queues. | |
1863 | */ | |
88459642 | 1864 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 1865 | |
484b4061 JA |
1866 | /* |
1867 | * Initialize batch roundrobin counts | |
1868 | */ | |
506e931f JA |
1869 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1870 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1871 | } | |
320ae51f JA |
1872 | } |
1873 | ||
2404e607 | 1874 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1875 | { |
1876 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1877 | int i; |
1878 | ||
2404e607 JM |
1879 | queue_for_each_hw_ctx(q, hctx, i) { |
1880 | if (shared) | |
1881 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1882 | else | |
1883 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1884 | } | |
1885 | } | |
1886 | ||
1887 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1888 | { | |
1889 | struct request_queue *q; | |
0d2602ca JA |
1890 | |
1891 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1892 | blk_mq_freeze_queue(q); | |
2404e607 | 1893 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1894 | blk_mq_unfreeze_queue(q); |
1895 | } | |
1896 | } | |
1897 | ||
1898 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1899 | { | |
1900 | struct blk_mq_tag_set *set = q->tag_set; | |
1901 | ||
0d2602ca JA |
1902 | mutex_lock(&set->tag_list_lock); |
1903 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1904 | if (list_is_singular(&set->tag_list)) { |
1905 | /* just transitioned to unshared */ | |
1906 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1907 | /* update existing queue */ | |
1908 | blk_mq_update_tag_set_depth(set, false); | |
1909 | } | |
0d2602ca | 1910 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1911 | } |
1912 | ||
1913 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1914 | struct request_queue *q) | |
1915 | { | |
1916 | q->tag_set = set; | |
1917 | ||
1918 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1919 | |
1920 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1921 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1922 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1923 | /* update existing queue */ | |
1924 | blk_mq_update_tag_set_depth(set, true); | |
1925 | } | |
1926 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1927 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1928 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1929 | |
0d2602ca JA |
1930 | mutex_unlock(&set->tag_list_lock); |
1931 | } | |
1932 | ||
e09aae7e ML |
1933 | /* |
1934 | * It is the actual release handler for mq, but we do it from | |
1935 | * request queue's release handler for avoiding use-after-free | |
1936 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1937 | * but we can't group ctx/kctx kobj without it. | |
1938 | */ | |
1939 | void blk_mq_release(struct request_queue *q) | |
1940 | { | |
1941 | struct blk_mq_hw_ctx *hctx; | |
1942 | unsigned int i; | |
1943 | ||
1944 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1945 | queue_for_each_hw_ctx(q, hctx, i) { |
1946 | if (!hctx) | |
1947 | continue; | |
1948 | kfree(hctx->ctxs); | |
e09aae7e | 1949 | kfree(hctx); |
c3b4afca | 1950 | } |
e09aae7e | 1951 | |
a723bab3 AM |
1952 | q->mq_map = NULL; |
1953 | ||
e09aae7e ML |
1954 | kfree(q->queue_hw_ctx); |
1955 | ||
1956 | /* ctx kobj stays in queue_ctx */ | |
1957 | free_percpu(q->queue_ctx); | |
1958 | } | |
1959 | ||
24d2f903 | 1960 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1961 | { |
1962 | struct request_queue *uninit_q, *q; | |
1963 | ||
1964 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1965 | if (!uninit_q) | |
1966 | return ERR_PTR(-ENOMEM); | |
1967 | ||
1968 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1969 | if (IS_ERR(q)) | |
1970 | blk_cleanup_queue(uninit_q); | |
1971 | ||
1972 | return q; | |
1973 | } | |
1974 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1975 | ||
868f2f0b KB |
1976 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
1977 | struct request_queue *q) | |
320ae51f | 1978 | { |
868f2f0b KB |
1979 | int i, j; |
1980 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 1981 | |
868f2f0b | 1982 | blk_mq_sysfs_unregister(q); |
24d2f903 | 1983 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 1984 | int node; |
f14bbe77 | 1985 | |
868f2f0b KB |
1986 | if (hctxs[i]) |
1987 | continue; | |
1988 | ||
1989 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
1990 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1991 | GFP_KERNEL, node); | |
320ae51f | 1992 | if (!hctxs[i]) |
868f2f0b | 1993 | break; |
320ae51f | 1994 | |
a86073e4 | 1995 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
1996 | node)) { |
1997 | kfree(hctxs[i]); | |
1998 | hctxs[i] = NULL; | |
1999 | break; | |
2000 | } | |
e4043dcf | 2001 | |
0d2602ca | 2002 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2003 | hctxs[i]->numa_node = node; |
320ae51f | 2004 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2005 | |
2006 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2007 | free_cpumask_var(hctxs[i]->cpumask); | |
2008 | kfree(hctxs[i]); | |
2009 | hctxs[i] = NULL; | |
2010 | break; | |
2011 | } | |
2012 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2013 | } |
868f2f0b KB |
2014 | for (j = i; j < q->nr_hw_queues; j++) { |
2015 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2016 | ||
2017 | if (hctx) { | |
2018 | if (hctx->tags) { | |
2019 | blk_mq_free_rq_map(set, hctx->tags, j); | |
2020 | set->tags[j] = NULL; | |
2021 | } | |
2022 | blk_mq_exit_hctx(q, set, hctx, j); | |
2023 | free_cpumask_var(hctx->cpumask); | |
2024 | kobject_put(&hctx->kobj); | |
2025 | kfree(hctx->ctxs); | |
2026 | kfree(hctx); | |
2027 | hctxs[j] = NULL; | |
2028 | ||
2029 | } | |
2030 | } | |
2031 | q->nr_hw_queues = i; | |
2032 | blk_mq_sysfs_register(q); | |
2033 | } | |
2034 | ||
2035 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2036 | struct request_queue *q) | |
2037 | { | |
66841672 ML |
2038 | /* mark the queue as mq asap */ |
2039 | q->mq_ops = set->ops; | |
2040 | ||
868f2f0b KB |
2041 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2042 | if (!q->queue_ctx) | |
c7de5726 | 2043 | goto err_exit; |
868f2f0b KB |
2044 | |
2045 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2046 | GFP_KERNEL, set->numa_node); | |
2047 | if (!q->queue_hw_ctx) | |
2048 | goto err_percpu; | |
2049 | ||
bdd17e75 | 2050 | q->mq_map = set->mq_map; |
868f2f0b KB |
2051 | |
2052 | blk_mq_realloc_hw_ctxs(set, q); | |
2053 | if (!q->nr_hw_queues) | |
2054 | goto err_hctxs; | |
320ae51f | 2055 | |
287922eb | 2056 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2057 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2058 | |
2059 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2060 | |
94eddfbe | 2061 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2062 | |
05f1dd53 JA |
2063 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2064 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2065 | ||
1be036e9 CH |
2066 | q->sg_reserved_size = INT_MAX; |
2067 | ||
2849450a | 2068 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2069 | INIT_LIST_HEAD(&q->requeue_list); |
2070 | spin_lock_init(&q->requeue_lock); | |
2071 | ||
07068d5b JA |
2072 | if (q->nr_hw_queues > 1) |
2073 | blk_queue_make_request(q, blk_mq_make_request); | |
2074 | else | |
2075 | blk_queue_make_request(q, blk_sq_make_request); | |
2076 | ||
eba71768 JA |
2077 | /* |
2078 | * Do this after blk_queue_make_request() overrides it... | |
2079 | */ | |
2080 | q->nr_requests = set->queue_depth; | |
2081 | ||
24d2f903 CH |
2082 | if (set->ops->complete) |
2083 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2084 | |
24d2f903 | 2085 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2086 | |
5778322e | 2087 | get_online_cpus(); |
320ae51f | 2088 | mutex_lock(&all_q_mutex); |
320ae51f | 2089 | |
4593fdbe | 2090 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2091 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2092 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2093 | |
4593fdbe | 2094 | mutex_unlock(&all_q_mutex); |
5778322e | 2095 | put_online_cpus(); |
4593fdbe | 2096 | |
320ae51f | 2097 | return q; |
18741986 | 2098 | |
320ae51f | 2099 | err_hctxs: |
868f2f0b | 2100 | kfree(q->queue_hw_ctx); |
320ae51f | 2101 | err_percpu: |
868f2f0b | 2102 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2103 | err_exit: |
2104 | q->mq_ops = NULL; | |
320ae51f JA |
2105 | return ERR_PTR(-ENOMEM); |
2106 | } | |
b62c21b7 | 2107 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2108 | |
2109 | void blk_mq_free_queue(struct request_queue *q) | |
2110 | { | |
624dbe47 | 2111 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2112 | |
0e626368 AM |
2113 | mutex_lock(&all_q_mutex); |
2114 | list_del_init(&q->all_q_node); | |
2115 | mutex_unlock(&all_q_mutex); | |
2116 | ||
0d2602ca JA |
2117 | blk_mq_del_queue_tag_set(q); |
2118 | ||
624dbe47 ML |
2119 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2120 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2121 | } |
320ae51f JA |
2122 | |
2123 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2124 | static void blk_mq_queue_reinit(struct request_queue *q, |
2125 | const struct cpumask *online_mask) | |
320ae51f | 2126 | { |
4ecd4fef | 2127 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2128 | |
67aec14c JA |
2129 | blk_mq_sysfs_unregister(q); |
2130 | ||
320ae51f JA |
2131 | /* |
2132 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2133 | * we should change hctx numa_node according to new topology (this | |
2134 | * involves free and re-allocate memory, worthy doing?) | |
2135 | */ | |
2136 | ||
5778322e | 2137 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2138 | |
67aec14c | 2139 | blk_mq_sysfs_register(q); |
320ae51f JA |
2140 | } |
2141 | ||
65d5291e SAS |
2142 | /* |
2143 | * New online cpumask which is going to be set in this hotplug event. | |
2144 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2145 | * one-by-one and dynamically allocating this could result in a failure. | |
2146 | */ | |
2147 | static struct cpumask cpuhp_online_new; | |
2148 | ||
2149 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2150 | { |
2151 | struct request_queue *q; | |
320ae51f JA |
2152 | |
2153 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2154 | /* |
2155 | * We need to freeze and reinit all existing queues. Freezing | |
2156 | * involves synchronous wait for an RCU grace period and doing it | |
2157 | * one by one may take a long time. Start freezing all queues in | |
2158 | * one swoop and then wait for the completions so that freezing can | |
2159 | * take place in parallel. | |
2160 | */ | |
2161 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2162 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2163 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2164 | blk_mq_freeze_queue_wait(q); |
2165 | ||
f054b56c ML |
2166 | /* |
2167 | * timeout handler can't touch hw queue during the | |
2168 | * reinitialization | |
2169 | */ | |
2170 | del_timer_sync(&q->timeout); | |
2171 | } | |
2172 | ||
320ae51f | 2173 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2174 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2175 | |
2176 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2177 | blk_mq_unfreeze_queue(q); | |
2178 | ||
320ae51f | 2179 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2180 | } |
2181 | ||
2182 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2183 | { | |
97a32864 | 2184 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2185 | blk_mq_queue_reinit_work(); |
2186 | return 0; | |
2187 | } | |
2188 | ||
2189 | /* | |
2190 | * Before hotadded cpu starts handling requests, new mappings must be | |
2191 | * established. Otherwise, these requests in hw queue might never be | |
2192 | * dispatched. | |
2193 | * | |
2194 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2195 | * for CPU0, and ctx1 for CPU1). | |
2196 | * | |
2197 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2198 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2199 | * | |
2200 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in | |
2201 | * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2202 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list | |
2203 | * is ignored. | |
2204 | */ | |
2205 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2206 | { | |
2207 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2208 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2209 | blk_mq_queue_reinit_work(); | |
2210 | return 0; | |
320ae51f JA |
2211 | } |
2212 | ||
a5164405 JA |
2213 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2214 | { | |
2215 | int i; | |
2216 | ||
2217 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2218 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2219 | if (!set->tags[i]) | |
2220 | goto out_unwind; | |
2221 | } | |
2222 | ||
2223 | return 0; | |
2224 | ||
2225 | out_unwind: | |
2226 | while (--i >= 0) | |
2227 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2228 | ||
a5164405 JA |
2229 | return -ENOMEM; |
2230 | } | |
2231 | ||
2232 | /* | |
2233 | * Allocate the request maps associated with this tag_set. Note that this | |
2234 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2235 | * will be updated to reflect the allocated depth. | |
2236 | */ | |
2237 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2238 | { | |
2239 | unsigned int depth; | |
2240 | int err; | |
2241 | ||
2242 | depth = set->queue_depth; | |
2243 | do { | |
2244 | err = __blk_mq_alloc_rq_maps(set); | |
2245 | if (!err) | |
2246 | break; | |
2247 | ||
2248 | set->queue_depth >>= 1; | |
2249 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2250 | err = -ENOMEM; | |
2251 | break; | |
2252 | } | |
2253 | } while (set->queue_depth); | |
2254 | ||
2255 | if (!set->queue_depth || err) { | |
2256 | pr_err("blk-mq: failed to allocate request map\n"); | |
2257 | return -ENOMEM; | |
2258 | } | |
2259 | ||
2260 | if (depth != set->queue_depth) | |
2261 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2262 | depth, set->queue_depth); | |
2263 | ||
2264 | return 0; | |
2265 | } | |
2266 | ||
a4391c64 JA |
2267 | /* |
2268 | * Alloc a tag set to be associated with one or more request queues. | |
2269 | * May fail with EINVAL for various error conditions. May adjust the | |
2270 | * requested depth down, if if it too large. In that case, the set | |
2271 | * value will be stored in set->queue_depth. | |
2272 | */ | |
24d2f903 CH |
2273 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2274 | { | |
da695ba2 CH |
2275 | int ret; |
2276 | ||
205fb5f5 BVA |
2277 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2278 | ||
24d2f903 CH |
2279 | if (!set->nr_hw_queues) |
2280 | return -EINVAL; | |
a4391c64 | 2281 | if (!set->queue_depth) |
24d2f903 CH |
2282 | return -EINVAL; |
2283 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2284 | return -EINVAL; | |
2285 | ||
7d7e0f90 | 2286 | if (!set->ops->queue_rq) |
24d2f903 CH |
2287 | return -EINVAL; |
2288 | ||
a4391c64 JA |
2289 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2290 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2291 | BLK_MQ_MAX_DEPTH); | |
2292 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2293 | } | |
24d2f903 | 2294 | |
6637fadf SL |
2295 | /* |
2296 | * If a crashdump is active, then we are potentially in a very | |
2297 | * memory constrained environment. Limit us to 1 queue and | |
2298 | * 64 tags to prevent using too much memory. | |
2299 | */ | |
2300 | if (is_kdump_kernel()) { | |
2301 | set->nr_hw_queues = 1; | |
2302 | set->queue_depth = min(64U, set->queue_depth); | |
2303 | } | |
868f2f0b KB |
2304 | /* |
2305 | * There is no use for more h/w queues than cpus. | |
2306 | */ | |
2307 | if (set->nr_hw_queues > nr_cpu_ids) | |
2308 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2309 | |
868f2f0b | 2310 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2311 | GFP_KERNEL, set->numa_node); |
2312 | if (!set->tags) | |
a5164405 | 2313 | return -ENOMEM; |
24d2f903 | 2314 | |
da695ba2 CH |
2315 | ret = -ENOMEM; |
2316 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2317 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2318 | if (!set->mq_map) |
2319 | goto out_free_tags; | |
2320 | ||
da695ba2 CH |
2321 | if (set->ops->map_queues) |
2322 | ret = set->ops->map_queues(set); | |
2323 | else | |
2324 | ret = blk_mq_map_queues(set); | |
2325 | if (ret) | |
2326 | goto out_free_mq_map; | |
2327 | ||
2328 | ret = blk_mq_alloc_rq_maps(set); | |
2329 | if (ret) | |
bdd17e75 | 2330 | goto out_free_mq_map; |
24d2f903 | 2331 | |
0d2602ca JA |
2332 | mutex_init(&set->tag_list_lock); |
2333 | INIT_LIST_HEAD(&set->tag_list); | |
2334 | ||
24d2f903 | 2335 | return 0; |
bdd17e75 CH |
2336 | |
2337 | out_free_mq_map: | |
2338 | kfree(set->mq_map); | |
2339 | set->mq_map = NULL; | |
2340 | out_free_tags: | |
5676e7b6 RE |
2341 | kfree(set->tags); |
2342 | set->tags = NULL; | |
da695ba2 | 2343 | return ret; |
24d2f903 CH |
2344 | } |
2345 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2346 | ||
2347 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2348 | { | |
2349 | int i; | |
2350 | ||
868f2f0b | 2351 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2352 | if (set->tags[i]) |
484b4061 JA |
2353 | blk_mq_free_rq_map(set, set->tags[i], i); |
2354 | } | |
2355 | ||
bdd17e75 CH |
2356 | kfree(set->mq_map); |
2357 | set->mq_map = NULL; | |
2358 | ||
981bd189 | 2359 | kfree(set->tags); |
5676e7b6 | 2360 | set->tags = NULL; |
24d2f903 CH |
2361 | } |
2362 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2363 | ||
e3a2b3f9 JA |
2364 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2365 | { | |
2366 | struct blk_mq_tag_set *set = q->tag_set; | |
2367 | struct blk_mq_hw_ctx *hctx; | |
2368 | int i, ret; | |
2369 | ||
2370 | if (!set || nr > set->queue_depth) | |
2371 | return -EINVAL; | |
2372 | ||
2373 | ret = 0; | |
2374 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2375 | if (!hctx->tags) |
2376 | continue; | |
e3a2b3f9 JA |
2377 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2378 | if (ret) | |
2379 | break; | |
2380 | } | |
2381 | ||
2382 | if (!ret) | |
2383 | q->nr_requests = nr; | |
2384 | ||
2385 | return ret; | |
2386 | } | |
2387 | ||
868f2f0b KB |
2388 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2389 | { | |
2390 | struct request_queue *q; | |
2391 | ||
2392 | if (nr_hw_queues > nr_cpu_ids) | |
2393 | nr_hw_queues = nr_cpu_ids; | |
2394 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2395 | return; | |
2396 | ||
2397 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2398 | blk_mq_freeze_queue(q); | |
2399 | ||
2400 | set->nr_hw_queues = nr_hw_queues; | |
2401 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2402 | blk_mq_realloc_hw_ctxs(set, q); | |
2403 | ||
2404 | if (q->nr_hw_queues > 1) | |
2405 | blk_queue_make_request(q, blk_mq_make_request); | |
2406 | else | |
2407 | blk_queue_make_request(q, blk_sq_make_request); | |
2408 | ||
2409 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2410 | } | |
2411 | ||
2412 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2413 | blk_mq_unfreeze_queue(q); | |
2414 | } | |
2415 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2416 | ||
676141e4 JA |
2417 | void blk_mq_disable_hotplug(void) |
2418 | { | |
2419 | mutex_lock(&all_q_mutex); | |
2420 | } | |
2421 | ||
2422 | void blk_mq_enable_hotplug(void) | |
2423 | { | |
2424 | mutex_unlock(&all_q_mutex); | |
2425 | } | |
2426 | ||
320ae51f JA |
2427 | static int __init blk_mq_init(void) |
2428 | { | |
9467f859 TG |
2429 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2430 | blk_mq_hctx_notify_dead); | |
320ae51f | 2431 | |
65d5291e SAS |
2432 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2433 | blk_mq_queue_reinit_prepare, | |
2434 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2435 | return 0; |
2436 | } | |
2437 | subsys_initcall(blk_mq_init); |