]>
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 | |
aed3ea94 JA |
118 | void blk_mq_wake_waiters(struct request_queue *q) |
119 | { | |
120 | struct blk_mq_hw_ctx *hctx; | |
121 | unsigned int i; | |
122 | ||
123 | queue_for_each_hw_ctx(q, hctx, i) | |
124 | if (blk_mq_hw_queue_mapped(hctx)) | |
125 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
126 | |
127 | /* | |
128 | * If we are called because the queue has now been marked as | |
129 | * dying, we need to ensure that processes currently waiting on | |
130 | * the queue are notified as well. | |
131 | */ | |
132 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
133 | } |
134 | ||
320ae51f JA |
135 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
136 | { | |
137 | return blk_mq_has_free_tags(hctx->tags); | |
138 | } | |
139 | EXPORT_SYMBOL(blk_mq_can_queue); | |
140 | ||
94eddfbe | 141 | static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
ef295ecf | 142 | struct request *rq, unsigned int op) |
320ae51f | 143 | { |
af76e555 CH |
144 | INIT_LIST_HEAD(&rq->queuelist); |
145 | /* csd/requeue_work/fifo_time is initialized before use */ | |
146 | rq->q = q; | |
320ae51f | 147 | rq->mq_ctx = ctx; |
ef295ecf | 148 | rq->cmd_flags = op; |
e8064021 CH |
149 | if (blk_queue_io_stat(q)) |
150 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
151 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
152 | rq->cpu = -1; | |
af76e555 CH |
153 | INIT_HLIST_NODE(&rq->hash); |
154 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
155 | rq->rq_disk = NULL; |
156 | rq->part = NULL; | |
3ee32372 | 157 | rq->start_time = jiffies; |
af76e555 CH |
158 | #ifdef CONFIG_BLK_CGROUP |
159 | rq->rl = NULL; | |
0fec08b4 | 160 | set_start_time_ns(rq); |
af76e555 CH |
161 | rq->io_start_time_ns = 0; |
162 | #endif | |
163 | rq->nr_phys_segments = 0; | |
164 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
165 | rq->nr_integrity_segments = 0; | |
166 | #endif | |
af76e555 CH |
167 | rq->special = NULL; |
168 | /* tag was already set */ | |
169 | rq->errors = 0; | |
af76e555 | 170 | |
6f4a1626 TB |
171 | rq->cmd = rq->__cmd; |
172 | ||
af76e555 CH |
173 | rq->extra_len = 0; |
174 | rq->sense_len = 0; | |
175 | rq->resid_len = 0; | |
176 | rq->sense = NULL; | |
177 | ||
af76e555 | 178 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
179 | rq->timeout = 0; |
180 | ||
af76e555 CH |
181 | rq->end_io = NULL; |
182 | rq->end_io_data = NULL; | |
183 | rq->next_rq = NULL; | |
184 | ||
ef295ecf | 185 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f JA |
186 | } |
187 | ||
5dee8577 | 188 | static struct request * |
ef295ecf | 189 | __blk_mq_alloc_request(struct blk_mq_alloc_data *data, unsigned int op) |
5dee8577 CH |
190 | { |
191 | struct request *rq; | |
192 | unsigned int tag; | |
193 | ||
cb96a42c | 194 | tag = blk_mq_get_tag(data); |
5dee8577 | 195 | if (tag != BLK_MQ_TAG_FAIL) { |
cb96a42c | 196 | rq = data->hctx->tags->rqs[tag]; |
5dee8577 | 197 | |
cb96a42c | 198 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 199 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 200 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
201 | } |
202 | ||
203 | rq->tag = tag; | |
ef295ecf | 204 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
205 | return rq; |
206 | } | |
207 | ||
208 | return NULL; | |
209 | } | |
210 | ||
6f3b0e8b CH |
211 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
212 | unsigned int flags) | |
320ae51f | 213 | { |
d852564f CH |
214 | struct blk_mq_ctx *ctx; |
215 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 216 | struct request *rq; |
cb96a42c | 217 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 218 | int ret; |
320ae51f | 219 | |
6f3b0e8b | 220 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
221 | if (ret) |
222 | return ERR_PTR(ret); | |
320ae51f | 223 | |
d852564f | 224 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 225 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 226 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
ef295ecf | 227 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f | 228 | blk_mq_put_ctx(ctx); |
841bac2c | 229 | |
c76541a9 | 230 | if (!rq) { |
3ef28e83 | 231 | blk_queue_exit(q); |
a492f075 | 232 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 233 | } |
0c4de0f3 CH |
234 | |
235 | rq->__data_len = 0; | |
236 | rq->__sector = (sector_t) -1; | |
237 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
238 | return rq; |
239 | } | |
4bb659b1 | 240 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 241 | |
1f5bd336 ML |
242 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
243 | unsigned int flags, unsigned int hctx_idx) | |
244 | { | |
245 | struct blk_mq_hw_ctx *hctx; | |
246 | struct blk_mq_ctx *ctx; | |
247 | struct request *rq; | |
248 | struct blk_mq_alloc_data alloc_data; | |
249 | int ret; | |
250 | ||
251 | /* | |
252 | * If the tag allocator sleeps we could get an allocation for a | |
253 | * different hardware context. No need to complicate the low level | |
254 | * allocator for this for the rare use case of a command tied to | |
255 | * a specific queue. | |
256 | */ | |
257 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
258 | return ERR_PTR(-EINVAL); | |
259 | ||
260 | if (hctx_idx >= q->nr_hw_queues) | |
261 | return ERR_PTR(-EIO); | |
262 | ||
263 | ret = blk_queue_enter(q, true); | |
264 | if (ret) | |
265 | return ERR_PTR(ret); | |
266 | ||
c8712c6a CH |
267 | /* |
268 | * Check if the hardware context is actually mapped to anything. | |
269 | * If not tell the caller that it should skip this queue. | |
270 | */ | |
1f5bd336 | 271 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
272 | if (!blk_mq_hw_queue_mapped(hctx)) { |
273 | ret = -EXDEV; | |
274 | goto out_queue_exit; | |
275 | } | |
1f5bd336 ML |
276 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
277 | ||
278 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 279 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 280 | if (!rq) { |
c8712c6a CH |
281 | ret = -EWOULDBLOCK; |
282 | goto out_queue_exit; | |
1f5bd336 ML |
283 | } |
284 | ||
285 | return rq; | |
c8712c6a CH |
286 | |
287 | out_queue_exit: | |
288 | blk_queue_exit(q); | |
289 | return ERR_PTR(ret); | |
1f5bd336 ML |
290 | } |
291 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
292 | ||
320ae51f JA |
293 | static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, |
294 | struct blk_mq_ctx *ctx, struct request *rq) | |
295 | { | |
296 | const int tag = rq->tag; | |
297 | struct request_queue *q = rq->q; | |
298 | ||
e8064021 | 299 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 300 | atomic_dec(&hctx->nr_active); |
e8064021 | 301 | rq->rq_flags = 0; |
0d2602ca | 302 | |
af76e555 | 303 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
40aabb67 | 304 | blk_mq_put_tag(hctx, ctx, tag); |
3ef28e83 | 305 | blk_queue_exit(q); |
320ae51f JA |
306 | } |
307 | ||
7c7f2f2b | 308 | void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq) |
320ae51f JA |
309 | { |
310 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
311 | |
312 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 313 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b JA |
314 | |
315 | } | |
316 | EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request); | |
317 | ||
318 | void blk_mq_free_request(struct request *rq) | |
319 | { | |
7d7e0f90 | 320 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 321 | } |
1a3b595a | 322 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 323 | |
c8a446ad | 324 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 325 | { |
0d11e6ac ML |
326 | blk_account_io_done(rq); |
327 | ||
91b63639 | 328 | if (rq->end_io) { |
320ae51f | 329 | rq->end_io(rq, error); |
91b63639 CH |
330 | } else { |
331 | if (unlikely(blk_bidi_rq(rq))) | |
332 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 333 | blk_mq_free_request(rq); |
91b63639 | 334 | } |
320ae51f | 335 | } |
c8a446ad | 336 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 337 | |
c8a446ad | 338 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
339 | { |
340 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
341 | BUG(); | |
c8a446ad | 342 | __blk_mq_end_request(rq, error); |
63151a44 | 343 | } |
c8a446ad | 344 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 345 | |
30a91cb4 | 346 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 347 | { |
3d6efbf6 | 348 | struct request *rq = data; |
320ae51f | 349 | |
30a91cb4 | 350 | rq->q->softirq_done_fn(rq); |
320ae51f | 351 | } |
320ae51f | 352 | |
ed851860 | 353 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
354 | { |
355 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 356 | bool shared = false; |
320ae51f JA |
357 | int cpu; |
358 | ||
38535201 | 359 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
360 | rq->q->softirq_done_fn(rq); |
361 | return; | |
362 | } | |
320ae51f JA |
363 | |
364 | cpu = get_cpu(); | |
38535201 CH |
365 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
366 | shared = cpus_share_cache(cpu, ctx->cpu); | |
367 | ||
368 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 369 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
370 | rq->csd.info = rq; |
371 | rq->csd.flags = 0; | |
c46fff2a | 372 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 373 | } else { |
30a91cb4 | 374 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 375 | } |
320ae51f JA |
376 | put_cpu(); |
377 | } | |
30a91cb4 | 378 | |
1fa8cc52 | 379 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
380 | { |
381 | struct request_queue *q = rq->q; | |
382 | ||
383 | if (!q->softirq_done_fn) | |
c8a446ad | 384 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
385 | else |
386 | blk_mq_ipi_complete_request(rq); | |
387 | } | |
388 | ||
30a91cb4 CH |
389 | /** |
390 | * blk_mq_complete_request - end I/O on a request | |
391 | * @rq: the request being processed | |
392 | * | |
393 | * Description: | |
394 | * Ends all I/O on a request. It does not handle partial completions. | |
395 | * The actual completion happens out-of-order, through a IPI handler. | |
396 | **/ | |
f4829a9b | 397 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 398 | { |
95f09684 JA |
399 | struct request_queue *q = rq->q; |
400 | ||
401 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 402 | return; |
f4829a9b CH |
403 | if (!blk_mark_rq_complete(rq)) { |
404 | rq->errors = error; | |
ed851860 | 405 | __blk_mq_complete_request(rq); |
f4829a9b | 406 | } |
30a91cb4 CH |
407 | } |
408 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 409 | |
973c0191 KB |
410 | int blk_mq_request_started(struct request *rq) |
411 | { | |
412 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
413 | } | |
414 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
415 | ||
e2490073 | 416 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
417 | { |
418 | struct request_queue *q = rq->q; | |
419 | ||
420 | trace_block_rq_issue(q, rq); | |
421 | ||
742ee69b | 422 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
423 | if (unlikely(blk_bidi_rq(rq))) |
424 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 425 | |
2b8393b4 | 426 | blk_add_timer(rq); |
87ee7b11 | 427 | |
538b7534 JA |
428 | /* |
429 | * Ensure that ->deadline is visible before set the started | |
430 | * flag and clear the completed flag. | |
431 | */ | |
432 | smp_mb__before_atomic(); | |
433 | ||
87ee7b11 JA |
434 | /* |
435 | * Mark us as started and clear complete. Complete might have been | |
436 | * set if requeue raced with timeout, which then marked it as | |
437 | * complete. So be sure to clear complete again when we start | |
438 | * the request, otherwise we'll ignore the completion event. | |
439 | */ | |
4b570521 JA |
440 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
441 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
442 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
443 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
444 | |
445 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
446 | /* | |
447 | * Make sure space for the drain appears. We know we can do | |
448 | * this because max_hw_segments has been adjusted to be one | |
449 | * fewer than the device can handle. | |
450 | */ | |
451 | rq->nr_phys_segments++; | |
452 | } | |
320ae51f | 453 | } |
e2490073 | 454 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 455 | |
ed0791b2 | 456 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
457 | { |
458 | struct request_queue *q = rq->q; | |
459 | ||
460 | trace_block_rq_requeue(q, rq); | |
49f5baa5 | 461 | |
e2490073 CH |
462 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
463 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
464 | rq->nr_phys_segments--; | |
465 | } | |
320ae51f JA |
466 | } |
467 | ||
ed0791b2 CH |
468 | void blk_mq_requeue_request(struct request *rq) |
469 | { | |
ed0791b2 | 470 | __blk_mq_requeue_request(rq); |
ed0791b2 | 471 | |
ed0791b2 | 472 | BUG_ON(blk_queued_rq(rq)); |
6fca6a61 | 473 | blk_mq_add_to_requeue_list(rq, true); |
ed0791b2 CH |
474 | } |
475 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
476 | ||
6fca6a61 CH |
477 | static void blk_mq_requeue_work(struct work_struct *work) |
478 | { | |
479 | struct request_queue *q = | |
2849450a | 480 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
481 | LIST_HEAD(rq_list); |
482 | struct request *rq, *next; | |
483 | unsigned long flags; | |
484 | ||
485 | spin_lock_irqsave(&q->requeue_lock, flags); | |
486 | list_splice_init(&q->requeue_list, &rq_list); | |
487 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
488 | ||
489 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 490 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
491 | continue; |
492 | ||
e8064021 | 493 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 CH |
494 | list_del_init(&rq->queuelist); |
495 | blk_mq_insert_request(rq, true, false, false); | |
496 | } | |
497 | ||
498 | while (!list_empty(&rq_list)) { | |
499 | rq = list_entry(rq_list.next, struct request, queuelist); | |
500 | list_del_init(&rq->queuelist); | |
501 | blk_mq_insert_request(rq, false, false, false); | |
502 | } | |
503 | ||
8b957415 JA |
504 | /* |
505 | * Use the start variant of queue running here, so that running | |
506 | * the requeue work will kick stopped queues. | |
507 | */ | |
508 | blk_mq_start_hw_queues(q); | |
6fca6a61 CH |
509 | } |
510 | ||
511 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
512 | { | |
513 | struct request_queue *q = rq->q; | |
514 | unsigned long flags; | |
515 | ||
516 | /* | |
517 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
518 | * request head insertation from the workqueue. | |
519 | */ | |
e8064021 | 520 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
521 | |
522 | spin_lock_irqsave(&q->requeue_lock, flags); | |
523 | if (at_head) { | |
e8064021 | 524 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
525 | list_add(&rq->queuelist, &q->requeue_list); |
526 | } else { | |
527 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
528 | } | |
529 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
530 | } | |
531 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
532 | ||
c68ed59f KB |
533 | void blk_mq_cancel_requeue_work(struct request_queue *q) |
534 | { | |
2849450a | 535 | cancel_delayed_work_sync(&q->requeue_work); |
c68ed59f KB |
536 | } |
537 | EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work); | |
538 | ||
6fca6a61 CH |
539 | void blk_mq_kick_requeue_list(struct request_queue *q) |
540 | { | |
2849450a | 541 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
542 | } |
543 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
544 | ||
2849450a MS |
545 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
546 | unsigned long msecs) | |
547 | { | |
548 | kblockd_schedule_delayed_work(&q->requeue_work, | |
549 | msecs_to_jiffies(msecs)); | |
550 | } | |
551 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
552 | ||
1885b24d JA |
553 | void blk_mq_abort_requeue_list(struct request_queue *q) |
554 | { | |
555 | unsigned long flags; | |
556 | LIST_HEAD(rq_list); | |
557 | ||
558 | spin_lock_irqsave(&q->requeue_lock, flags); | |
559 | list_splice_init(&q->requeue_list, &rq_list); | |
560 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
561 | ||
562 | while (!list_empty(&rq_list)) { | |
563 | struct request *rq; | |
564 | ||
565 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
566 | list_del_init(&rq->queuelist); | |
567 | rq->errors = -EIO; | |
568 | blk_mq_end_request(rq, rq->errors); | |
569 | } | |
570 | } | |
571 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
572 | ||
0e62f51f JA |
573 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
574 | { | |
88c7b2b7 JA |
575 | if (tag < tags->nr_tags) { |
576 | prefetch(tags->rqs[tag]); | |
4ee86bab | 577 | return tags->rqs[tag]; |
88c7b2b7 | 578 | } |
4ee86bab HR |
579 | |
580 | return NULL; | |
24d2f903 CH |
581 | } |
582 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
583 | ||
320ae51f | 584 | struct blk_mq_timeout_data { |
46f92d42 CH |
585 | unsigned long next; |
586 | unsigned int next_set; | |
320ae51f JA |
587 | }; |
588 | ||
90415837 | 589 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 590 | { |
46f92d42 CH |
591 | struct blk_mq_ops *ops = req->q->mq_ops; |
592 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
593 | |
594 | /* | |
595 | * We know that complete is set at this point. If STARTED isn't set | |
596 | * anymore, then the request isn't active and the "timeout" should | |
597 | * just be ignored. This can happen due to the bitflag ordering. | |
598 | * Timeout first checks if STARTED is set, and if it is, assumes | |
599 | * the request is active. But if we race with completion, then | |
600 | * we both flags will get cleared. So check here again, and ignore | |
601 | * a timeout event with a request that isn't active. | |
602 | */ | |
46f92d42 CH |
603 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
604 | return; | |
87ee7b11 | 605 | |
46f92d42 | 606 | if (ops->timeout) |
0152fb6b | 607 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
608 | |
609 | switch (ret) { | |
610 | case BLK_EH_HANDLED: | |
611 | __blk_mq_complete_request(req); | |
612 | break; | |
613 | case BLK_EH_RESET_TIMER: | |
614 | blk_add_timer(req); | |
615 | blk_clear_rq_complete(req); | |
616 | break; | |
617 | case BLK_EH_NOT_HANDLED: | |
618 | break; | |
619 | default: | |
620 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
621 | break; | |
622 | } | |
87ee7b11 | 623 | } |
5b3f25fc | 624 | |
81481eb4 CH |
625 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
626 | struct request *rq, void *priv, bool reserved) | |
627 | { | |
628 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 629 | |
eb130dbf KB |
630 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
631 | /* | |
632 | * If a request wasn't started before the queue was | |
633 | * marked dying, kill it here or it'll go unnoticed. | |
634 | */ | |
a59e0f57 KB |
635 | if (unlikely(blk_queue_dying(rq->q))) { |
636 | rq->errors = -EIO; | |
637 | blk_mq_end_request(rq, rq->errors); | |
638 | } | |
46f92d42 | 639 | return; |
eb130dbf | 640 | } |
87ee7b11 | 641 | |
46f92d42 CH |
642 | if (time_after_eq(jiffies, rq->deadline)) { |
643 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 644 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
645 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
646 | data->next = rq->deadline; | |
647 | data->next_set = 1; | |
648 | } | |
87ee7b11 JA |
649 | } |
650 | ||
287922eb | 651 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 652 | { |
287922eb CH |
653 | struct request_queue *q = |
654 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
655 | struct blk_mq_timeout_data data = { |
656 | .next = 0, | |
657 | .next_set = 0, | |
658 | }; | |
81481eb4 | 659 | int i; |
320ae51f | 660 | |
71f79fb3 GKB |
661 | /* A deadlock might occur if a request is stuck requiring a |
662 | * timeout at the same time a queue freeze is waiting | |
663 | * completion, since the timeout code would not be able to | |
664 | * acquire the queue reference here. | |
665 | * | |
666 | * That's why we don't use blk_queue_enter here; instead, we use | |
667 | * percpu_ref_tryget directly, because we need to be able to | |
668 | * obtain a reference even in the short window between the queue | |
669 | * starting to freeze, by dropping the first reference in | |
670 | * blk_mq_freeze_queue_start, and the moment the last request is | |
671 | * consumed, marked by the instant q_usage_counter reaches | |
672 | * zero. | |
673 | */ | |
674 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
675 | return; |
676 | ||
0bf6cd5b | 677 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 678 | |
81481eb4 CH |
679 | if (data.next_set) { |
680 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
681 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 682 | } else { |
0bf6cd5b CH |
683 | struct blk_mq_hw_ctx *hctx; |
684 | ||
f054b56c ML |
685 | queue_for_each_hw_ctx(q, hctx, i) { |
686 | /* the hctx may be unmapped, so check it here */ | |
687 | if (blk_mq_hw_queue_mapped(hctx)) | |
688 | blk_mq_tag_idle(hctx); | |
689 | } | |
0d2602ca | 690 | } |
287922eb | 691 | blk_queue_exit(q); |
320ae51f JA |
692 | } |
693 | ||
694 | /* | |
695 | * Reverse check our software queue for entries that we could potentially | |
696 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
697 | * too much time checking for merges. | |
698 | */ | |
699 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
700 | struct blk_mq_ctx *ctx, struct bio *bio) | |
701 | { | |
702 | struct request *rq; | |
703 | int checked = 8; | |
704 | ||
705 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
706 | int el_ret; | |
707 | ||
708 | if (!checked--) | |
709 | break; | |
710 | ||
711 | if (!blk_rq_merge_ok(rq, bio)) | |
712 | continue; | |
713 | ||
714 | el_ret = blk_try_merge(rq, bio); | |
715 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
716 | if (bio_attempt_back_merge(q, rq, bio)) { | |
717 | ctx->rq_merged++; | |
718 | return true; | |
719 | } | |
720 | break; | |
721 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
722 | if (bio_attempt_front_merge(q, rq, bio)) { | |
723 | ctx->rq_merged++; | |
724 | return true; | |
725 | } | |
726 | break; | |
727 | } | |
728 | } | |
729 | ||
730 | return false; | |
731 | } | |
732 | ||
88459642 OS |
733 | struct flush_busy_ctx_data { |
734 | struct blk_mq_hw_ctx *hctx; | |
735 | struct list_head *list; | |
736 | }; | |
737 | ||
738 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
739 | { | |
740 | struct flush_busy_ctx_data *flush_data = data; | |
741 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
742 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
743 | ||
744 | sbitmap_clear_bit(sb, bitnr); | |
745 | spin_lock(&ctx->lock); | |
746 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
747 | spin_unlock(&ctx->lock); | |
748 | return true; | |
749 | } | |
750 | ||
1429d7c9 JA |
751 | /* |
752 | * Process software queues that have been marked busy, splicing them | |
753 | * to the for-dispatch | |
754 | */ | |
755 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
756 | { | |
88459642 OS |
757 | struct flush_busy_ctx_data data = { |
758 | .hctx = hctx, | |
759 | .list = list, | |
760 | }; | |
1429d7c9 | 761 | |
88459642 | 762 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 763 | } |
1429d7c9 | 764 | |
703fd1c0 JA |
765 | static inline unsigned int queued_to_index(unsigned int queued) |
766 | { | |
767 | if (!queued) | |
768 | return 0; | |
1429d7c9 | 769 | |
703fd1c0 | 770 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
771 | } |
772 | ||
320ae51f JA |
773 | /* |
774 | * Run this hardware queue, pulling any software queues mapped to it in. | |
775 | * Note that this function currently has various problems around ordering | |
776 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
777 | * items on the hctx->dispatch list. Ignore that for now. | |
778 | */ | |
779 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
780 | { | |
781 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
782 | struct request *rq; |
783 | LIST_HEAD(rq_list); | |
74c45052 JA |
784 | LIST_HEAD(driver_list); |
785 | struct list_head *dptr; | |
1429d7c9 | 786 | int queued; |
320ae51f | 787 | |
5d12f905 | 788 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state))) |
320ae51f JA |
789 | return; |
790 | ||
0e87e58b JA |
791 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
792 | cpu_online(hctx->next_cpu)); | |
793 | ||
320ae51f JA |
794 | hctx->run++; |
795 | ||
796 | /* | |
797 | * Touch any software queue that has pending entries. | |
798 | */ | |
1429d7c9 | 799 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
800 | |
801 | /* | |
802 | * If we have previous entries on our dispatch list, grab them | |
803 | * and stuff them at the front for more fair dispatch. | |
804 | */ | |
805 | if (!list_empty_careful(&hctx->dispatch)) { | |
806 | spin_lock(&hctx->lock); | |
807 | if (!list_empty(&hctx->dispatch)) | |
808 | list_splice_init(&hctx->dispatch, &rq_list); | |
809 | spin_unlock(&hctx->lock); | |
810 | } | |
811 | ||
74c45052 JA |
812 | /* |
813 | * Start off with dptr being NULL, so we start the first request | |
814 | * immediately, even if we have more pending. | |
815 | */ | |
816 | dptr = NULL; | |
817 | ||
320ae51f JA |
818 | /* |
819 | * Now process all the entries, sending them to the driver. | |
820 | */ | |
1429d7c9 | 821 | queued = 0; |
320ae51f | 822 | while (!list_empty(&rq_list)) { |
74c45052 | 823 | struct blk_mq_queue_data bd; |
320ae51f JA |
824 | int ret; |
825 | ||
826 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
827 | list_del_init(&rq->queuelist); | |
320ae51f | 828 | |
74c45052 JA |
829 | bd.rq = rq; |
830 | bd.list = dptr; | |
831 | bd.last = list_empty(&rq_list); | |
832 | ||
833 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
834 | switch (ret) { |
835 | case BLK_MQ_RQ_QUEUE_OK: | |
836 | queued++; | |
52b9c330 | 837 | break; |
320ae51f | 838 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 839 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 840 | __blk_mq_requeue_request(rq); |
320ae51f JA |
841 | break; |
842 | default: | |
843 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 844 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 845 | rq->errors = -EIO; |
c8a446ad | 846 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
847 | break; |
848 | } | |
849 | ||
850 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
851 | break; | |
74c45052 JA |
852 | |
853 | /* | |
854 | * We've done the first request. If we have more than 1 | |
855 | * left in the list, set dptr to defer issue. | |
856 | */ | |
857 | if (!dptr && rq_list.next != rq_list.prev) | |
858 | dptr = &driver_list; | |
320ae51f JA |
859 | } |
860 | ||
703fd1c0 | 861 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
862 | |
863 | /* | |
864 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
865 | * that is where we will continue on next queue run. | |
866 | */ | |
867 | if (!list_empty(&rq_list)) { | |
868 | spin_lock(&hctx->lock); | |
869 | list_splice(&rq_list, &hctx->dispatch); | |
870 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
871 | /* |
872 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
873 | * it's possible the queue is stopped and restarted again | |
874 | * before this. Queue restart will dispatch requests. And since | |
875 | * requests in rq_list aren't added into hctx->dispatch yet, | |
876 | * the requests in rq_list might get lost. | |
877 | * | |
878 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
879 | **/ | |
880 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
881 | } |
882 | } | |
883 | ||
506e931f JA |
884 | /* |
885 | * It'd be great if the workqueue API had a way to pass | |
886 | * in a mask and had some smarts for more clever placement. | |
887 | * For now we just round-robin here, switching for every | |
888 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
889 | */ | |
890 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
891 | { | |
b657d7e6 CH |
892 | if (hctx->queue->nr_hw_queues == 1) |
893 | return WORK_CPU_UNBOUND; | |
506e931f JA |
894 | |
895 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 896 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
897 | |
898 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
899 | if (next_cpu >= nr_cpu_ids) | |
900 | next_cpu = cpumask_first(hctx->cpumask); | |
901 | ||
902 | hctx->next_cpu = next_cpu; | |
903 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
904 | |
905 | return cpu; | |
506e931f JA |
906 | } |
907 | ||
b657d7e6 | 908 | return hctx->next_cpu; |
506e931f JA |
909 | } |
910 | ||
320ae51f JA |
911 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
912 | { | |
19c66e59 ML |
913 | if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) || |
914 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
915 | return; |
916 | ||
1b792f2f | 917 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
918 | int cpu = get_cpu(); |
919 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 920 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 921 | put_cpu(); |
398205b8 PB |
922 | return; |
923 | } | |
e4043dcf | 924 | |
2a90d4aa | 925 | put_cpu(); |
e4043dcf | 926 | } |
398205b8 | 927 | |
27489a3c | 928 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
929 | } |
930 | ||
b94ec296 | 931 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
932 | { |
933 | struct blk_mq_hw_ctx *hctx; | |
934 | int i; | |
935 | ||
936 | queue_for_each_hw_ctx(q, hctx, i) { | |
937 | if ((!blk_mq_hctx_has_pending(hctx) && | |
938 | list_empty_careful(&hctx->dispatch)) || | |
5d12f905 | 939 | test_bit(BLK_MQ_S_STOPPED, &hctx->state)) |
320ae51f JA |
940 | continue; |
941 | ||
b94ec296 | 942 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
943 | } |
944 | } | |
b94ec296 | 945 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f JA |
946 | |
947 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
948 | { | |
27489a3c | 949 | cancel_work(&hctx->run_work); |
70f4db63 | 950 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
951 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
952 | } | |
953 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
954 | ||
280d45f6 CH |
955 | void blk_mq_stop_hw_queues(struct request_queue *q) |
956 | { | |
957 | struct blk_mq_hw_ctx *hctx; | |
958 | int i; | |
959 | ||
960 | queue_for_each_hw_ctx(q, hctx, i) | |
961 | blk_mq_stop_hw_queue(hctx); | |
962 | } | |
963 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
964 | ||
320ae51f JA |
965 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
966 | { | |
967 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 968 | |
0ffbce80 | 969 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
970 | } |
971 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
972 | ||
2f268556 CH |
973 | void blk_mq_start_hw_queues(struct request_queue *q) |
974 | { | |
975 | struct blk_mq_hw_ctx *hctx; | |
976 | int i; | |
977 | ||
978 | queue_for_each_hw_ctx(q, hctx, i) | |
979 | blk_mq_start_hw_queue(hctx); | |
980 | } | |
981 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
982 | ||
1b4a3258 | 983 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
984 | { |
985 | struct blk_mq_hw_ctx *hctx; | |
986 | int i; | |
987 | ||
988 | queue_for_each_hw_ctx(q, hctx, i) { | |
989 | if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
990 | continue; | |
991 | ||
992 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 993 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
994 | } |
995 | } | |
996 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
997 | ||
70f4db63 | 998 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
999 | { |
1000 | struct blk_mq_hw_ctx *hctx; | |
1001 | ||
27489a3c | 1002 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1003 | |
320ae51f JA |
1004 | __blk_mq_run_hw_queue(hctx); |
1005 | } | |
1006 | ||
70f4db63 CH |
1007 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1008 | { | |
1009 | struct blk_mq_hw_ctx *hctx; | |
1010 | ||
1011 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1012 | ||
1013 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1014 | __blk_mq_run_hw_queue(hctx); | |
1015 | } | |
1016 | ||
1017 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1018 | { | |
19c66e59 ML |
1019 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1020 | return; | |
70f4db63 | 1021 | |
b657d7e6 CH |
1022 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1023 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1024 | } |
1025 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1026 | ||
cfd0c552 | 1027 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1028 | struct request *rq, |
1029 | bool at_head) | |
320ae51f | 1030 | { |
e57690fe JA |
1031 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1032 | ||
01b983c9 JA |
1033 | trace_block_rq_insert(hctx->queue, rq); |
1034 | ||
72a0a36e CH |
1035 | if (at_head) |
1036 | list_add(&rq->queuelist, &ctx->rq_list); | |
1037 | else | |
1038 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1039 | } |
4bb659b1 | 1040 | |
cfd0c552 ML |
1041 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1042 | struct request *rq, bool at_head) | |
1043 | { | |
1044 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1045 | ||
e57690fe | 1046 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1047 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1048 | } |
1049 | ||
eeabc850 | 1050 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1051 | bool async) |
320ae51f | 1052 | { |
e57690fe | 1053 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1054 | struct request_queue *q = rq->q; |
7d7e0f90 | 1055 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1056 | |
a57a178a CH |
1057 | spin_lock(&ctx->lock); |
1058 | __blk_mq_insert_request(hctx, rq, at_head); | |
1059 | spin_unlock(&ctx->lock); | |
320ae51f | 1060 | |
320ae51f JA |
1061 | if (run_queue) |
1062 | blk_mq_run_hw_queue(hctx, async); | |
1063 | } | |
1064 | ||
1065 | static void blk_mq_insert_requests(struct request_queue *q, | |
1066 | struct blk_mq_ctx *ctx, | |
1067 | struct list_head *list, | |
1068 | int depth, | |
1069 | bool from_schedule) | |
1070 | ||
1071 | { | |
7d7e0f90 | 1072 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1073 | |
1074 | trace_block_unplug(q, depth, !from_schedule); | |
1075 | ||
320ae51f JA |
1076 | /* |
1077 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1078 | * offline now | |
1079 | */ | |
1080 | spin_lock(&ctx->lock); | |
1081 | while (!list_empty(list)) { | |
1082 | struct request *rq; | |
1083 | ||
1084 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1085 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1086 | list_del_init(&rq->queuelist); |
e57690fe | 1087 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1088 | } |
cfd0c552 | 1089 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1090 | spin_unlock(&ctx->lock); |
1091 | ||
320ae51f JA |
1092 | blk_mq_run_hw_queue(hctx, from_schedule); |
1093 | } | |
1094 | ||
1095 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1096 | { | |
1097 | struct request *rqa = container_of(a, struct request, queuelist); | |
1098 | struct request *rqb = container_of(b, struct request, queuelist); | |
1099 | ||
1100 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1101 | (rqa->mq_ctx == rqb->mq_ctx && | |
1102 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1103 | } | |
1104 | ||
1105 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1106 | { | |
1107 | struct blk_mq_ctx *this_ctx; | |
1108 | struct request_queue *this_q; | |
1109 | struct request *rq; | |
1110 | LIST_HEAD(list); | |
1111 | LIST_HEAD(ctx_list); | |
1112 | unsigned int depth; | |
1113 | ||
1114 | list_splice_init(&plug->mq_list, &list); | |
1115 | ||
1116 | list_sort(NULL, &list, plug_ctx_cmp); | |
1117 | ||
1118 | this_q = NULL; | |
1119 | this_ctx = NULL; | |
1120 | depth = 0; | |
1121 | ||
1122 | while (!list_empty(&list)) { | |
1123 | rq = list_entry_rq(list.next); | |
1124 | list_del_init(&rq->queuelist); | |
1125 | BUG_ON(!rq->q); | |
1126 | if (rq->mq_ctx != this_ctx) { | |
1127 | if (this_ctx) { | |
1128 | blk_mq_insert_requests(this_q, this_ctx, | |
1129 | &ctx_list, depth, | |
1130 | from_schedule); | |
1131 | } | |
1132 | ||
1133 | this_ctx = rq->mq_ctx; | |
1134 | this_q = rq->q; | |
1135 | depth = 0; | |
1136 | } | |
1137 | ||
1138 | depth++; | |
1139 | list_add_tail(&rq->queuelist, &ctx_list); | |
1140 | } | |
1141 | ||
1142 | /* | |
1143 | * If 'this_ctx' is set, we know we have entries to complete | |
1144 | * on 'ctx_list'. Do those. | |
1145 | */ | |
1146 | if (this_ctx) { | |
1147 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1148 | from_schedule); | |
1149 | } | |
1150 | } | |
1151 | ||
1152 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1153 | { | |
1154 | init_request_from_bio(rq, bio); | |
4b570521 | 1155 | |
a21f2a3e | 1156 | blk_account_io_start(rq, 1); |
320ae51f JA |
1157 | } |
1158 | ||
274a5843 JA |
1159 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1160 | { | |
1161 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1162 | !blk_queue_nomerges(hctx->queue); | |
1163 | } | |
1164 | ||
07068d5b JA |
1165 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1166 | struct blk_mq_ctx *ctx, | |
1167 | struct request *rq, struct bio *bio) | |
320ae51f | 1168 | { |
e18378a6 | 1169 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1170 | blk_mq_bio_to_request(rq, bio); |
1171 | spin_lock(&ctx->lock); | |
1172 | insert_rq: | |
1173 | __blk_mq_insert_request(hctx, rq, false); | |
1174 | spin_unlock(&ctx->lock); | |
1175 | return false; | |
1176 | } else { | |
274a5843 JA |
1177 | struct request_queue *q = hctx->queue; |
1178 | ||
07068d5b JA |
1179 | spin_lock(&ctx->lock); |
1180 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1181 | blk_mq_bio_to_request(rq, bio); | |
1182 | goto insert_rq; | |
1183 | } | |
320ae51f | 1184 | |
07068d5b JA |
1185 | spin_unlock(&ctx->lock); |
1186 | __blk_mq_free_request(hctx, ctx, rq); | |
1187 | return true; | |
14ec77f3 | 1188 | } |
07068d5b | 1189 | } |
14ec77f3 | 1190 | |
07068d5b JA |
1191 | static struct request *blk_mq_map_request(struct request_queue *q, |
1192 | struct bio *bio, | |
2552e3f8 | 1193 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1194 | { |
1195 | struct blk_mq_hw_ctx *hctx; | |
1196 | struct blk_mq_ctx *ctx; | |
1197 | struct request *rq; | |
320ae51f | 1198 | |
3ef28e83 | 1199 | blk_queue_enter_live(q); |
320ae51f | 1200 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1201 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1202 | |
ef295ecf | 1203 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1204 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1205 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1206 | |
7dd2fb68 | 1207 | data->hctx->queued++; |
07068d5b JA |
1208 | return rq; |
1209 | } | |
1210 | ||
7b371636 | 1211 | static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1212 | { |
1213 | int ret; | |
1214 | struct request_queue *q = rq->q; | |
7d7e0f90 | 1215 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
f984df1f SL |
1216 | struct blk_mq_queue_data bd = { |
1217 | .rq = rq, | |
1218 | .list = NULL, | |
1219 | .last = 1 | |
1220 | }; | |
7b371636 | 1221 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f SL |
1222 | |
1223 | /* | |
1224 | * For OK queue, we are done. For error, kill it. Any other | |
1225 | * error (busy), just add it to our list as we previously | |
1226 | * would have done | |
1227 | */ | |
1228 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1229 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1230 | *cookie = new_cookie; | |
f984df1f | 1231 | return 0; |
7b371636 | 1232 | } |
f984df1f | 1233 | |
7b371636 JA |
1234 | __blk_mq_requeue_request(rq); |
1235 | ||
1236 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1237 | *cookie = BLK_QC_T_NONE; | |
1238 | rq->errors = -EIO; | |
1239 | blk_mq_end_request(rq, rq->errors); | |
1240 | return 0; | |
f984df1f | 1241 | } |
7b371636 JA |
1242 | |
1243 | return -1; | |
f984df1f SL |
1244 | } |
1245 | ||
07068d5b JA |
1246 | /* |
1247 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1248 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1249 | * hardware for SYNC IO. | |
1250 | */ | |
dece1635 | 1251 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1252 | { |
ef295ecf | 1253 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1254 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1255 | struct blk_mq_alloc_data data; |
07068d5b | 1256 | struct request *rq; |
f984df1f SL |
1257 | unsigned int request_count = 0; |
1258 | struct blk_plug *plug; | |
5b3f341f | 1259 | struct request *same_queue_rq = NULL; |
7b371636 | 1260 | blk_qc_t cookie; |
07068d5b JA |
1261 | |
1262 | blk_queue_bounce(q, &bio); | |
1263 | ||
1264 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1265 | bio_io_error(bio); |
dece1635 | 1266 | return BLK_QC_T_NONE; |
07068d5b JA |
1267 | } |
1268 | ||
54efd50b KO |
1269 | blk_queue_split(q, &bio, q->bio_split); |
1270 | ||
87c279e6 OS |
1271 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1272 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1273 | return BLK_QC_T_NONE; | |
f984df1f | 1274 | |
07068d5b JA |
1275 | rq = blk_mq_map_request(q, bio, &data); |
1276 | if (unlikely(!rq)) | |
dece1635 | 1277 | return BLK_QC_T_NONE; |
07068d5b | 1278 | |
7b371636 | 1279 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1280 | |
1281 | if (unlikely(is_flush_fua)) { | |
1282 | blk_mq_bio_to_request(rq, bio); | |
1283 | blk_insert_flush(rq); | |
1284 | goto run_queue; | |
1285 | } | |
1286 | ||
f984df1f | 1287 | plug = current->plug; |
e167dfb5 JA |
1288 | /* |
1289 | * If the driver supports defer issued based on 'last', then | |
1290 | * queue it up like normal since we can potentially save some | |
1291 | * CPU this way. | |
1292 | */ | |
f984df1f SL |
1293 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1294 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1295 | struct request *old_rq = NULL; | |
07068d5b JA |
1296 | |
1297 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1298 | |
1299 | /* | |
b094f89c | 1300 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1301 | * Otherwise the existing request in the plug list will be |
1302 | * issued. So the plug list will have one request at most | |
07068d5b | 1303 | */ |
f984df1f | 1304 | if (plug) { |
5b3f341f SL |
1305 | /* |
1306 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1307 | * happens, same_queue_rq is invalid and plug list is |
1308 | * empty | |
1309 | */ | |
5b3f341f SL |
1310 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1311 | old_rq = same_queue_rq; | |
f984df1f | 1312 | list_del_init(&old_rq->queuelist); |
07068d5b | 1313 | } |
f984df1f SL |
1314 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1315 | } else /* is_sync */ | |
1316 | old_rq = rq; | |
1317 | blk_mq_put_ctx(data.ctx); | |
1318 | if (!old_rq) | |
7b371636 JA |
1319 | goto done; |
1320 | if (!blk_mq_direct_issue_request(old_rq, &cookie)) | |
1321 | goto done; | |
f984df1f | 1322 | blk_mq_insert_request(old_rq, false, true, true); |
7b371636 | 1323 | goto done; |
07068d5b JA |
1324 | } |
1325 | ||
1326 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1327 | /* | |
1328 | * For a SYNC request, send it to the hardware immediately. For | |
1329 | * an ASYNC request, just ensure that we run it later on. The | |
1330 | * latter allows for merging opportunities and more efficient | |
1331 | * dispatching. | |
1332 | */ | |
1333 | run_queue: | |
1334 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1335 | } | |
07068d5b | 1336 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1337 | done: |
1338 | return cookie; | |
07068d5b JA |
1339 | } |
1340 | ||
1341 | /* | |
1342 | * Single hardware queue variant. This will attempt to use any per-process | |
1343 | * plug for merging and IO deferral. | |
1344 | */ | |
dece1635 | 1345 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1346 | { |
ef295ecf | 1347 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1348 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1349 | struct blk_plug *plug; |
1350 | unsigned int request_count = 0; | |
2552e3f8 | 1351 | struct blk_mq_alloc_data data; |
07068d5b | 1352 | struct request *rq; |
7b371636 | 1353 | blk_qc_t cookie; |
07068d5b | 1354 | |
07068d5b JA |
1355 | blk_queue_bounce(q, &bio); |
1356 | ||
1357 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1358 | bio_io_error(bio); |
dece1635 | 1359 | return BLK_QC_T_NONE; |
07068d5b JA |
1360 | } |
1361 | ||
54efd50b KO |
1362 | blk_queue_split(q, &bio, q->bio_split); |
1363 | ||
87c279e6 OS |
1364 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1365 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1366 | return BLK_QC_T_NONE; | |
1367 | } else | |
1368 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1369 | |
1370 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1371 | if (unlikely(!rq)) |
dece1635 | 1372 | return BLK_QC_T_NONE; |
320ae51f | 1373 | |
7b371636 | 1374 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1375 | |
1376 | if (unlikely(is_flush_fua)) { | |
1377 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1378 | blk_insert_flush(rq); |
1379 | goto run_queue; | |
1380 | } | |
1381 | ||
1382 | /* | |
1383 | * A task plug currently exists. Since this is completely lockless, | |
1384 | * utilize that to temporarily store requests until the task is | |
1385 | * either done or scheduled away. | |
1386 | */ | |
e6c4438b JM |
1387 | plug = current->plug; |
1388 | if (plug) { | |
1389 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1390 | if (!request_count) |
e6c4438b | 1391 | trace_block_plug(q); |
b094f89c JA |
1392 | |
1393 | blk_mq_put_ctx(data.ctx); | |
1394 | ||
1395 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1396 | blk_flush_plug_list(plug, false); |
1397 | trace_block_plug(q); | |
320ae51f | 1398 | } |
b094f89c | 1399 | |
e6c4438b | 1400 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1401 | return cookie; |
320ae51f JA |
1402 | } |
1403 | ||
07068d5b JA |
1404 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1405 | /* | |
1406 | * For a SYNC request, send it to the hardware immediately. For | |
1407 | * an ASYNC request, just ensure that we run it later on. The | |
1408 | * latter allows for merging opportunities and more efficient | |
1409 | * dispatching. | |
1410 | */ | |
1411 | run_queue: | |
1412 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1413 | } |
1414 | ||
07068d5b | 1415 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1416 | return cookie; |
320ae51f JA |
1417 | } |
1418 | ||
24d2f903 CH |
1419 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1420 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1421 | { |
e9b267d9 | 1422 | struct page *page; |
320ae51f | 1423 | |
24d2f903 | 1424 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1425 | int i; |
320ae51f | 1426 | |
24d2f903 CH |
1427 | for (i = 0; i < tags->nr_tags; i++) { |
1428 | if (!tags->rqs[i]) | |
e9b267d9 | 1429 | continue; |
24d2f903 CH |
1430 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1431 | hctx_idx, i); | |
a5164405 | 1432 | tags->rqs[i] = NULL; |
e9b267d9 | 1433 | } |
320ae51f | 1434 | } |
320ae51f | 1435 | |
24d2f903 CH |
1436 | while (!list_empty(&tags->page_list)) { |
1437 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1438 | list_del_init(&page->lru); |
f75782e4 CM |
1439 | /* |
1440 | * Remove kmemleak object previously allocated in | |
1441 | * blk_mq_init_rq_map(). | |
1442 | */ | |
1443 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1444 | __free_pages(page, page->private); |
1445 | } | |
1446 | ||
24d2f903 | 1447 | kfree(tags->rqs); |
320ae51f | 1448 | |
24d2f903 | 1449 | blk_mq_free_tags(tags); |
320ae51f JA |
1450 | } |
1451 | ||
1452 | static size_t order_to_size(unsigned int order) | |
1453 | { | |
4ca08500 | 1454 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1455 | } |
1456 | ||
24d2f903 CH |
1457 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1458 | unsigned int hctx_idx) | |
320ae51f | 1459 | { |
24d2f903 | 1460 | struct blk_mq_tags *tags; |
320ae51f JA |
1461 | unsigned int i, j, entries_per_page, max_order = 4; |
1462 | size_t rq_size, left; | |
1463 | ||
24d2f903 | 1464 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1465 | set->numa_node, |
1466 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1467 | if (!tags) |
1468 | return NULL; | |
320ae51f | 1469 | |
24d2f903 CH |
1470 | INIT_LIST_HEAD(&tags->page_list); |
1471 | ||
a5164405 JA |
1472 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1473 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1474 | set->numa_node); | |
24d2f903 CH |
1475 | if (!tags->rqs) { |
1476 | blk_mq_free_tags(tags); | |
1477 | return NULL; | |
1478 | } | |
320ae51f JA |
1479 | |
1480 | /* | |
1481 | * rq_size is the size of the request plus driver payload, rounded | |
1482 | * to the cacheline size | |
1483 | */ | |
24d2f903 | 1484 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1485 | cache_line_size()); |
24d2f903 | 1486 | left = rq_size * set->queue_depth; |
320ae51f | 1487 | |
24d2f903 | 1488 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1489 | int this_order = max_order; |
1490 | struct page *page; | |
1491 | int to_do; | |
1492 | void *p; | |
1493 | ||
b3a834b1 | 1494 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1495 | this_order--; |
1496 | ||
1497 | do { | |
a5164405 | 1498 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1499 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1500 | this_order); |
320ae51f JA |
1501 | if (page) |
1502 | break; | |
1503 | if (!this_order--) | |
1504 | break; | |
1505 | if (order_to_size(this_order) < rq_size) | |
1506 | break; | |
1507 | } while (1); | |
1508 | ||
1509 | if (!page) | |
24d2f903 | 1510 | goto fail; |
320ae51f JA |
1511 | |
1512 | page->private = this_order; | |
24d2f903 | 1513 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1514 | |
1515 | p = page_address(page); | |
f75782e4 CM |
1516 | /* |
1517 | * Allow kmemleak to scan these pages as they contain pointers | |
1518 | * to additional allocations like via ops->init_request(). | |
1519 | */ | |
1520 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1521 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1522 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1523 | left -= to_do * rq_size; |
1524 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1525 | tags->rqs[i] = p; |
1526 | if (set->ops->init_request) { | |
1527 | if (set->ops->init_request(set->driver_data, | |
1528 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1529 | set->numa_node)) { |
1530 | tags->rqs[i] = NULL; | |
24d2f903 | 1531 | goto fail; |
a5164405 | 1532 | } |
e9b267d9 CH |
1533 | } |
1534 | ||
320ae51f JA |
1535 | p += rq_size; |
1536 | i++; | |
1537 | } | |
1538 | } | |
24d2f903 | 1539 | return tags; |
320ae51f | 1540 | |
24d2f903 | 1541 | fail: |
24d2f903 CH |
1542 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1543 | return NULL; | |
320ae51f JA |
1544 | } |
1545 | ||
e57690fe JA |
1546 | /* |
1547 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1548 | * software queue to the hw queue dispatch list, and ensure that it | |
1549 | * gets run. | |
1550 | */ | |
9467f859 | 1551 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1552 | { |
9467f859 | 1553 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1554 | struct blk_mq_ctx *ctx; |
1555 | LIST_HEAD(tmp); | |
1556 | ||
9467f859 | 1557 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1558 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1559 | |
1560 | spin_lock(&ctx->lock); | |
1561 | if (!list_empty(&ctx->rq_list)) { | |
1562 | list_splice_init(&ctx->rq_list, &tmp); | |
1563 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1564 | } | |
1565 | spin_unlock(&ctx->lock); | |
1566 | ||
1567 | if (list_empty(&tmp)) | |
9467f859 | 1568 | return 0; |
484b4061 | 1569 | |
e57690fe JA |
1570 | spin_lock(&hctx->lock); |
1571 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1572 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1573 | |
1574 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1575 | return 0; |
484b4061 JA |
1576 | } |
1577 | ||
9467f859 | 1578 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1579 | { |
9467f859 TG |
1580 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1581 | &hctx->cpuhp_dead); | |
484b4061 JA |
1582 | } |
1583 | ||
c3b4afca | 1584 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1585 | static void blk_mq_exit_hctx(struct request_queue *q, |
1586 | struct blk_mq_tag_set *set, | |
1587 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1588 | { | |
f70ced09 ML |
1589 | unsigned flush_start_tag = set->queue_depth; |
1590 | ||
08e98fc6 ML |
1591 | blk_mq_tag_idle(hctx); |
1592 | ||
f70ced09 ML |
1593 | if (set->ops->exit_request) |
1594 | set->ops->exit_request(set->driver_data, | |
1595 | hctx->fq->flush_rq, hctx_idx, | |
1596 | flush_start_tag + hctx_idx); | |
1597 | ||
08e98fc6 ML |
1598 | if (set->ops->exit_hctx) |
1599 | set->ops->exit_hctx(hctx, hctx_idx); | |
1600 | ||
9467f859 | 1601 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1602 | blk_free_flush_queue(hctx->fq); |
88459642 | 1603 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1604 | } |
1605 | ||
624dbe47 ML |
1606 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1607 | struct blk_mq_tag_set *set, int nr_queue) | |
1608 | { | |
1609 | struct blk_mq_hw_ctx *hctx; | |
1610 | unsigned int i; | |
1611 | ||
1612 | queue_for_each_hw_ctx(q, hctx, i) { | |
1613 | if (i == nr_queue) | |
1614 | break; | |
08e98fc6 | 1615 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1616 | } |
624dbe47 ML |
1617 | } |
1618 | ||
1619 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1620 | struct blk_mq_tag_set *set) | |
1621 | { | |
1622 | struct blk_mq_hw_ctx *hctx; | |
1623 | unsigned int i; | |
1624 | ||
e09aae7e | 1625 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1626 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1627 | } |
1628 | ||
08e98fc6 ML |
1629 | static int blk_mq_init_hctx(struct request_queue *q, |
1630 | struct blk_mq_tag_set *set, | |
1631 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1632 | { |
08e98fc6 | 1633 | int node; |
f70ced09 | 1634 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1635 | |
1636 | node = hctx->numa_node; | |
1637 | if (node == NUMA_NO_NODE) | |
1638 | node = hctx->numa_node = set->numa_node; | |
1639 | ||
27489a3c | 1640 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1641 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1642 | spin_lock_init(&hctx->lock); | |
1643 | INIT_LIST_HEAD(&hctx->dispatch); | |
1644 | hctx->queue = q; | |
1645 | hctx->queue_num = hctx_idx; | |
2404e607 | 1646 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1647 | |
9467f859 | 1648 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1649 | |
1650 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1651 | |
1652 | /* | |
08e98fc6 ML |
1653 | * Allocate space for all possible cpus to avoid allocation at |
1654 | * runtime | |
320ae51f | 1655 | */ |
08e98fc6 ML |
1656 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1657 | GFP_KERNEL, node); | |
1658 | if (!hctx->ctxs) | |
1659 | goto unregister_cpu_notifier; | |
320ae51f | 1660 | |
88459642 OS |
1661 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1662 | node)) | |
08e98fc6 | 1663 | goto free_ctxs; |
320ae51f | 1664 | |
08e98fc6 | 1665 | hctx->nr_ctx = 0; |
320ae51f | 1666 | |
08e98fc6 ML |
1667 | if (set->ops->init_hctx && |
1668 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1669 | goto free_bitmap; | |
320ae51f | 1670 | |
f70ced09 ML |
1671 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1672 | if (!hctx->fq) | |
1673 | goto exit_hctx; | |
320ae51f | 1674 | |
f70ced09 ML |
1675 | if (set->ops->init_request && |
1676 | set->ops->init_request(set->driver_data, | |
1677 | hctx->fq->flush_rq, hctx_idx, | |
1678 | flush_start_tag + hctx_idx, node)) | |
1679 | goto free_fq; | |
320ae51f | 1680 | |
08e98fc6 | 1681 | return 0; |
320ae51f | 1682 | |
f70ced09 ML |
1683 | free_fq: |
1684 | kfree(hctx->fq); | |
1685 | exit_hctx: | |
1686 | if (set->ops->exit_hctx) | |
1687 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1688 | free_bitmap: |
88459642 | 1689 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1690 | free_ctxs: |
1691 | kfree(hctx->ctxs); | |
1692 | unregister_cpu_notifier: | |
9467f859 | 1693 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1694 | return -1; |
1695 | } | |
320ae51f | 1696 | |
320ae51f JA |
1697 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1698 | unsigned int nr_hw_queues) | |
1699 | { | |
1700 | unsigned int i; | |
1701 | ||
1702 | for_each_possible_cpu(i) { | |
1703 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1704 | struct blk_mq_hw_ctx *hctx; | |
1705 | ||
1706 | memset(__ctx, 0, sizeof(*__ctx)); | |
1707 | __ctx->cpu = i; | |
1708 | spin_lock_init(&__ctx->lock); | |
1709 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1710 | __ctx->queue = q; | |
1711 | ||
1712 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1713 | if (!cpu_online(i)) |
1714 | continue; | |
1715 | ||
7d7e0f90 | 1716 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1717 | |
320ae51f JA |
1718 | /* |
1719 | * Set local node, IFF we have more than one hw queue. If | |
1720 | * not, we remain on the home node of the device | |
1721 | */ | |
1722 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1723 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1724 | } |
1725 | } | |
1726 | ||
5778322e AM |
1727 | static void blk_mq_map_swqueue(struct request_queue *q, |
1728 | const struct cpumask *online_mask) | |
320ae51f JA |
1729 | { |
1730 | unsigned int i; | |
1731 | struct blk_mq_hw_ctx *hctx; | |
1732 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1733 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1734 | |
60de074b AM |
1735 | /* |
1736 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1737 | */ | |
1738 | mutex_lock(&q->sysfs_lock); | |
1739 | ||
320ae51f | 1740 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1741 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1742 | hctx->nr_ctx = 0; |
1743 | } | |
1744 | ||
1745 | /* | |
1746 | * Map software to hardware queues | |
1747 | */ | |
897bb0c7 | 1748 | for_each_possible_cpu(i) { |
320ae51f | 1749 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1750 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1751 | continue; |
1752 | ||
897bb0c7 | 1753 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1754 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1755 | |
e4043dcf | 1756 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1757 | ctx->index_hw = hctx->nr_ctx; |
1758 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1759 | } | |
506e931f | 1760 | |
60de074b AM |
1761 | mutex_unlock(&q->sysfs_lock); |
1762 | ||
506e931f | 1763 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1764 | /* |
a68aafa5 JA |
1765 | * If no software queues are mapped to this hardware queue, |
1766 | * disable it and free the request entries. | |
484b4061 JA |
1767 | */ |
1768 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1769 | if (set->tags[i]) { |
1770 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1771 | set->tags[i] = NULL; | |
484b4061 | 1772 | } |
2a34c087 | 1773 | hctx->tags = NULL; |
484b4061 JA |
1774 | continue; |
1775 | } | |
1776 | ||
2a34c087 ML |
1777 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1778 | if (!set->tags[i]) | |
1779 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1780 | hctx->tags = set->tags[i]; | |
1781 | WARN_ON(!hctx->tags); | |
1782 | ||
889fa31f CY |
1783 | /* |
1784 | * Set the map size to the number of mapped software queues. | |
1785 | * This is more accurate and more efficient than looping | |
1786 | * over all possibly mapped software queues. | |
1787 | */ | |
88459642 | 1788 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 1789 | |
484b4061 JA |
1790 | /* |
1791 | * Initialize batch roundrobin counts | |
1792 | */ | |
506e931f JA |
1793 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1794 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1795 | } | |
320ae51f JA |
1796 | } |
1797 | ||
2404e607 | 1798 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1799 | { |
1800 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1801 | int i; |
1802 | ||
2404e607 JM |
1803 | queue_for_each_hw_ctx(q, hctx, i) { |
1804 | if (shared) | |
1805 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1806 | else | |
1807 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1808 | } | |
1809 | } | |
1810 | ||
1811 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1812 | { | |
1813 | struct request_queue *q; | |
0d2602ca JA |
1814 | |
1815 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1816 | blk_mq_freeze_queue(q); | |
2404e607 | 1817 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1818 | blk_mq_unfreeze_queue(q); |
1819 | } | |
1820 | } | |
1821 | ||
1822 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1823 | { | |
1824 | struct blk_mq_tag_set *set = q->tag_set; | |
1825 | ||
0d2602ca JA |
1826 | mutex_lock(&set->tag_list_lock); |
1827 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1828 | if (list_is_singular(&set->tag_list)) { |
1829 | /* just transitioned to unshared */ | |
1830 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1831 | /* update existing queue */ | |
1832 | blk_mq_update_tag_set_depth(set, false); | |
1833 | } | |
0d2602ca | 1834 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1835 | } |
1836 | ||
1837 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1838 | struct request_queue *q) | |
1839 | { | |
1840 | q->tag_set = set; | |
1841 | ||
1842 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1843 | |
1844 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1845 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1846 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1847 | /* update existing queue */ | |
1848 | blk_mq_update_tag_set_depth(set, true); | |
1849 | } | |
1850 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1851 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1852 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1853 | |
0d2602ca JA |
1854 | mutex_unlock(&set->tag_list_lock); |
1855 | } | |
1856 | ||
e09aae7e ML |
1857 | /* |
1858 | * It is the actual release handler for mq, but we do it from | |
1859 | * request queue's release handler for avoiding use-after-free | |
1860 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1861 | * but we can't group ctx/kctx kobj without it. | |
1862 | */ | |
1863 | void blk_mq_release(struct request_queue *q) | |
1864 | { | |
1865 | struct blk_mq_hw_ctx *hctx; | |
1866 | unsigned int i; | |
1867 | ||
1868 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1869 | queue_for_each_hw_ctx(q, hctx, i) { |
1870 | if (!hctx) | |
1871 | continue; | |
1872 | kfree(hctx->ctxs); | |
e09aae7e | 1873 | kfree(hctx); |
c3b4afca | 1874 | } |
e09aae7e | 1875 | |
a723bab3 AM |
1876 | q->mq_map = NULL; |
1877 | ||
e09aae7e ML |
1878 | kfree(q->queue_hw_ctx); |
1879 | ||
1880 | /* ctx kobj stays in queue_ctx */ | |
1881 | free_percpu(q->queue_ctx); | |
1882 | } | |
1883 | ||
24d2f903 | 1884 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1885 | { |
1886 | struct request_queue *uninit_q, *q; | |
1887 | ||
1888 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1889 | if (!uninit_q) | |
1890 | return ERR_PTR(-ENOMEM); | |
1891 | ||
1892 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1893 | if (IS_ERR(q)) | |
1894 | blk_cleanup_queue(uninit_q); | |
1895 | ||
1896 | return q; | |
1897 | } | |
1898 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1899 | ||
868f2f0b KB |
1900 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
1901 | struct request_queue *q) | |
320ae51f | 1902 | { |
868f2f0b KB |
1903 | int i, j; |
1904 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 1905 | |
868f2f0b | 1906 | blk_mq_sysfs_unregister(q); |
24d2f903 | 1907 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 1908 | int node; |
f14bbe77 | 1909 | |
868f2f0b KB |
1910 | if (hctxs[i]) |
1911 | continue; | |
1912 | ||
1913 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
1914 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1915 | GFP_KERNEL, node); | |
320ae51f | 1916 | if (!hctxs[i]) |
868f2f0b | 1917 | break; |
320ae51f | 1918 | |
a86073e4 | 1919 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
1920 | node)) { |
1921 | kfree(hctxs[i]); | |
1922 | hctxs[i] = NULL; | |
1923 | break; | |
1924 | } | |
e4043dcf | 1925 | |
0d2602ca | 1926 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1927 | hctxs[i]->numa_node = node; |
320ae51f | 1928 | hctxs[i]->queue_num = i; |
868f2f0b KB |
1929 | |
1930 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
1931 | free_cpumask_var(hctxs[i]->cpumask); | |
1932 | kfree(hctxs[i]); | |
1933 | hctxs[i] = NULL; | |
1934 | break; | |
1935 | } | |
1936 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 1937 | } |
868f2f0b KB |
1938 | for (j = i; j < q->nr_hw_queues; j++) { |
1939 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
1940 | ||
1941 | if (hctx) { | |
1942 | if (hctx->tags) { | |
1943 | blk_mq_free_rq_map(set, hctx->tags, j); | |
1944 | set->tags[j] = NULL; | |
1945 | } | |
1946 | blk_mq_exit_hctx(q, set, hctx, j); | |
1947 | free_cpumask_var(hctx->cpumask); | |
1948 | kobject_put(&hctx->kobj); | |
1949 | kfree(hctx->ctxs); | |
1950 | kfree(hctx); | |
1951 | hctxs[j] = NULL; | |
1952 | ||
1953 | } | |
1954 | } | |
1955 | q->nr_hw_queues = i; | |
1956 | blk_mq_sysfs_register(q); | |
1957 | } | |
1958 | ||
1959 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
1960 | struct request_queue *q) | |
1961 | { | |
66841672 ML |
1962 | /* mark the queue as mq asap */ |
1963 | q->mq_ops = set->ops; | |
1964 | ||
868f2f0b KB |
1965 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
1966 | if (!q->queue_ctx) | |
c7de5726 | 1967 | goto err_exit; |
868f2f0b KB |
1968 | |
1969 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
1970 | GFP_KERNEL, set->numa_node); | |
1971 | if (!q->queue_hw_ctx) | |
1972 | goto err_percpu; | |
1973 | ||
bdd17e75 | 1974 | q->mq_map = set->mq_map; |
868f2f0b KB |
1975 | |
1976 | blk_mq_realloc_hw_ctxs(set, q); | |
1977 | if (!q->nr_hw_queues) | |
1978 | goto err_hctxs; | |
320ae51f | 1979 | |
287922eb | 1980 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 1981 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
1982 | |
1983 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 1984 | |
94eddfbe | 1985 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1986 | |
05f1dd53 JA |
1987 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
1988 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
1989 | ||
1be036e9 CH |
1990 | q->sg_reserved_size = INT_MAX; |
1991 | ||
2849450a | 1992 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
1993 | INIT_LIST_HEAD(&q->requeue_list); |
1994 | spin_lock_init(&q->requeue_lock); | |
1995 | ||
07068d5b JA |
1996 | if (q->nr_hw_queues > 1) |
1997 | blk_queue_make_request(q, blk_mq_make_request); | |
1998 | else | |
1999 | blk_queue_make_request(q, blk_sq_make_request); | |
2000 | ||
eba71768 JA |
2001 | /* |
2002 | * Do this after blk_queue_make_request() overrides it... | |
2003 | */ | |
2004 | q->nr_requests = set->queue_depth; | |
2005 | ||
24d2f903 CH |
2006 | if (set->ops->complete) |
2007 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2008 | |
24d2f903 | 2009 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2010 | |
5778322e | 2011 | get_online_cpus(); |
320ae51f | 2012 | mutex_lock(&all_q_mutex); |
320ae51f | 2013 | |
4593fdbe | 2014 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2015 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2016 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2017 | |
4593fdbe | 2018 | mutex_unlock(&all_q_mutex); |
5778322e | 2019 | put_online_cpus(); |
4593fdbe | 2020 | |
320ae51f | 2021 | return q; |
18741986 | 2022 | |
320ae51f | 2023 | err_hctxs: |
868f2f0b | 2024 | kfree(q->queue_hw_ctx); |
320ae51f | 2025 | err_percpu: |
868f2f0b | 2026 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2027 | err_exit: |
2028 | q->mq_ops = NULL; | |
320ae51f JA |
2029 | return ERR_PTR(-ENOMEM); |
2030 | } | |
b62c21b7 | 2031 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2032 | |
2033 | void blk_mq_free_queue(struct request_queue *q) | |
2034 | { | |
624dbe47 | 2035 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2036 | |
0e626368 AM |
2037 | mutex_lock(&all_q_mutex); |
2038 | list_del_init(&q->all_q_node); | |
2039 | mutex_unlock(&all_q_mutex); | |
2040 | ||
0d2602ca JA |
2041 | blk_mq_del_queue_tag_set(q); |
2042 | ||
624dbe47 ML |
2043 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2044 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2045 | } |
320ae51f JA |
2046 | |
2047 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2048 | static void blk_mq_queue_reinit(struct request_queue *q, |
2049 | const struct cpumask *online_mask) | |
320ae51f | 2050 | { |
4ecd4fef | 2051 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2052 | |
67aec14c JA |
2053 | blk_mq_sysfs_unregister(q); |
2054 | ||
320ae51f JA |
2055 | /* |
2056 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2057 | * we should change hctx numa_node according to new topology (this | |
2058 | * involves free and re-allocate memory, worthy doing?) | |
2059 | */ | |
2060 | ||
5778322e | 2061 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2062 | |
67aec14c | 2063 | blk_mq_sysfs_register(q); |
320ae51f JA |
2064 | } |
2065 | ||
65d5291e SAS |
2066 | /* |
2067 | * New online cpumask which is going to be set in this hotplug event. | |
2068 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2069 | * one-by-one and dynamically allocating this could result in a failure. | |
2070 | */ | |
2071 | static struct cpumask cpuhp_online_new; | |
2072 | ||
2073 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2074 | { |
2075 | struct request_queue *q; | |
320ae51f JA |
2076 | |
2077 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2078 | /* |
2079 | * We need to freeze and reinit all existing queues. Freezing | |
2080 | * involves synchronous wait for an RCU grace period and doing it | |
2081 | * one by one may take a long time. Start freezing all queues in | |
2082 | * one swoop and then wait for the completions so that freezing can | |
2083 | * take place in parallel. | |
2084 | */ | |
2085 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2086 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2087 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2088 | blk_mq_freeze_queue_wait(q); |
2089 | ||
f054b56c ML |
2090 | /* |
2091 | * timeout handler can't touch hw queue during the | |
2092 | * reinitialization | |
2093 | */ | |
2094 | del_timer_sync(&q->timeout); | |
2095 | } | |
2096 | ||
320ae51f | 2097 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2098 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2099 | |
2100 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2101 | blk_mq_unfreeze_queue(q); | |
2102 | ||
320ae51f | 2103 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2104 | } |
2105 | ||
2106 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2107 | { | |
97a32864 | 2108 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2109 | blk_mq_queue_reinit_work(); |
2110 | return 0; | |
2111 | } | |
2112 | ||
2113 | /* | |
2114 | * Before hotadded cpu starts handling requests, new mappings must be | |
2115 | * established. Otherwise, these requests in hw queue might never be | |
2116 | * dispatched. | |
2117 | * | |
2118 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2119 | * for CPU0, and ctx1 for CPU1). | |
2120 | * | |
2121 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2122 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2123 | * | |
2124 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in | |
2125 | * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2126 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list | |
2127 | * is ignored. | |
2128 | */ | |
2129 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2130 | { | |
2131 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2132 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2133 | blk_mq_queue_reinit_work(); | |
2134 | return 0; | |
320ae51f JA |
2135 | } |
2136 | ||
a5164405 JA |
2137 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2138 | { | |
2139 | int i; | |
2140 | ||
2141 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2142 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2143 | if (!set->tags[i]) | |
2144 | goto out_unwind; | |
2145 | } | |
2146 | ||
2147 | return 0; | |
2148 | ||
2149 | out_unwind: | |
2150 | while (--i >= 0) | |
2151 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2152 | ||
a5164405 JA |
2153 | return -ENOMEM; |
2154 | } | |
2155 | ||
2156 | /* | |
2157 | * Allocate the request maps associated with this tag_set. Note that this | |
2158 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2159 | * will be updated to reflect the allocated depth. | |
2160 | */ | |
2161 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2162 | { | |
2163 | unsigned int depth; | |
2164 | int err; | |
2165 | ||
2166 | depth = set->queue_depth; | |
2167 | do { | |
2168 | err = __blk_mq_alloc_rq_maps(set); | |
2169 | if (!err) | |
2170 | break; | |
2171 | ||
2172 | set->queue_depth >>= 1; | |
2173 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2174 | err = -ENOMEM; | |
2175 | break; | |
2176 | } | |
2177 | } while (set->queue_depth); | |
2178 | ||
2179 | if (!set->queue_depth || err) { | |
2180 | pr_err("blk-mq: failed to allocate request map\n"); | |
2181 | return -ENOMEM; | |
2182 | } | |
2183 | ||
2184 | if (depth != set->queue_depth) | |
2185 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2186 | depth, set->queue_depth); | |
2187 | ||
2188 | return 0; | |
2189 | } | |
2190 | ||
a4391c64 JA |
2191 | /* |
2192 | * Alloc a tag set to be associated with one or more request queues. | |
2193 | * May fail with EINVAL for various error conditions. May adjust the | |
2194 | * requested depth down, if if it too large. In that case, the set | |
2195 | * value will be stored in set->queue_depth. | |
2196 | */ | |
24d2f903 CH |
2197 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2198 | { | |
da695ba2 CH |
2199 | int ret; |
2200 | ||
205fb5f5 BVA |
2201 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2202 | ||
24d2f903 CH |
2203 | if (!set->nr_hw_queues) |
2204 | return -EINVAL; | |
a4391c64 | 2205 | if (!set->queue_depth) |
24d2f903 CH |
2206 | return -EINVAL; |
2207 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2208 | return -EINVAL; | |
2209 | ||
7d7e0f90 | 2210 | if (!set->ops->queue_rq) |
24d2f903 CH |
2211 | return -EINVAL; |
2212 | ||
a4391c64 JA |
2213 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2214 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2215 | BLK_MQ_MAX_DEPTH); | |
2216 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2217 | } | |
24d2f903 | 2218 | |
6637fadf SL |
2219 | /* |
2220 | * If a crashdump is active, then we are potentially in a very | |
2221 | * memory constrained environment. Limit us to 1 queue and | |
2222 | * 64 tags to prevent using too much memory. | |
2223 | */ | |
2224 | if (is_kdump_kernel()) { | |
2225 | set->nr_hw_queues = 1; | |
2226 | set->queue_depth = min(64U, set->queue_depth); | |
2227 | } | |
868f2f0b KB |
2228 | /* |
2229 | * There is no use for more h/w queues than cpus. | |
2230 | */ | |
2231 | if (set->nr_hw_queues > nr_cpu_ids) | |
2232 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2233 | |
868f2f0b | 2234 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2235 | GFP_KERNEL, set->numa_node); |
2236 | if (!set->tags) | |
a5164405 | 2237 | return -ENOMEM; |
24d2f903 | 2238 | |
da695ba2 CH |
2239 | ret = -ENOMEM; |
2240 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2241 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2242 | if (!set->mq_map) |
2243 | goto out_free_tags; | |
2244 | ||
da695ba2 CH |
2245 | if (set->ops->map_queues) |
2246 | ret = set->ops->map_queues(set); | |
2247 | else | |
2248 | ret = blk_mq_map_queues(set); | |
2249 | if (ret) | |
2250 | goto out_free_mq_map; | |
2251 | ||
2252 | ret = blk_mq_alloc_rq_maps(set); | |
2253 | if (ret) | |
bdd17e75 | 2254 | goto out_free_mq_map; |
24d2f903 | 2255 | |
0d2602ca JA |
2256 | mutex_init(&set->tag_list_lock); |
2257 | INIT_LIST_HEAD(&set->tag_list); | |
2258 | ||
24d2f903 | 2259 | return 0; |
bdd17e75 CH |
2260 | |
2261 | out_free_mq_map: | |
2262 | kfree(set->mq_map); | |
2263 | set->mq_map = NULL; | |
2264 | out_free_tags: | |
5676e7b6 RE |
2265 | kfree(set->tags); |
2266 | set->tags = NULL; | |
da695ba2 | 2267 | return ret; |
24d2f903 CH |
2268 | } |
2269 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2270 | ||
2271 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2272 | { | |
2273 | int i; | |
2274 | ||
868f2f0b | 2275 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2276 | if (set->tags[i]) |
484b4061 JA |
2277 | blk_mq_free_rq_map(set, set->tags[i], i); |
2278 | } | |
2279 | ||
bdd17e75 CH |
2280 | kfree(set->mq_map); |
2281 | set->mq_map = NULL; | |
2282 | ||
981bd189 | 2283 | kfree(set->tags); |
5676e7b6 | 2284 | set->tags = NULL; |
24d2f903 CH |
2285 | } |
2286 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2287 | ||
e3a2b3f9 JA |
2288 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2289 | { | |
2290 | struct blk_mq_tag_set *set = q->tag_set; | |
2291 | struct blk_mq_hw_ctx *hctx; | |
2292 | int i, ret; | |
2293 | ||
2294 | if (!set || nr > set->queue_depth) | |
2295 | return -EINVAL; | |
2296 | ||
2297 | ret = 0; | |
2298 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2299 | if (!hctx->tags) |
2300 | continue; | |
e3a2b3f9 JA |
2301 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2302 | if (ret) | |
2303 | break; | |
2304 | } | |
2305 | ||
2306 | if (!ret) | |
2307 | q->nr_requests = nr; | |
2308 | ||
2309 | return ret; | |
2310 | } | |
2311 | ||
868f2f0b KB |
2312 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2313 | { | |
2314 | struct request_queue *q; | |
2315 | ||
2316 | if (nr_hw_queues > nr_cpu_ids) | |
2317 | nr_hw_queues = nr_cpu_ids; | |
2318 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2319 | return; | |
2320 | ||
2321 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2322 | blk_mq_freeze_queue(q); | |
2323 | ||
2324 | set->nr_hw_queues = nr_hw_queues; | |
2325 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2326 | blk_mq_realloc_hw_ctxs(set, q); | |
2327 | ||
2328 | if (q->nr_hw_queues > 1) | |
2329 | blk_queue_make_request(q, blk_mq_make_request); | |
2330 | else | |
2331 | blk_queue_make_request(q, blk_sq_make_request); | |
2332 | ||
2333 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2334 | } | |
2335 | ||
2336 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2337 | blk_mq_unfreeze_queue(q); | |
2338 | } | |
2339 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2340 | ||
676141e4 JA |
2341 | void blk_mq_disable_hotplug(void) |
2342 | { | |
2343 | mutex_lock(&all_q_mutex); | |
2344 | } | |
2345 | ||
2346 | void blk_mq_enable_hotplug(void) | |
2347 | { | |
2348 | mutex_unlock(&all_q_mutex); | |
2349 | } | |
2350 | ||
320ae51f JA |
2351 | static int __init blk_mq_init(void) |
2352 | { | |
9467f859 TG |
2353 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2354 | blk_mq_hctx_notify_dead); | |
320ae51f | 2355 | |
65d5291e SAS |
2356 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2357 | blk_mq_queue_reinit_prepare, | |
2358 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2359 | return 0; |
2360 | } | |
2361 | subsys_initcall(blk_mq_init); |