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 | ||
52d7f1b5 | 504 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
505 | } |
506 | ||
507 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head) | |
508 | { | |
509 | struct request_queue *q = rq->q; | |
510 | unsigned long flags; | |
511 | ||
512 | /* | |
513 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
514 | * request head insertation from the workqueue. | |
515 | */ | |
e8064021 | 516 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
517 | |
518 | spin_lock_irqsave(&q->requeue_lock, flags); | |
519 | if (at_head) { | |
e8064021 | 520 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
521 | list_add(&rq->queuelist, &q->requeue_list); |
522 | } else { | |
523 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
524 | } | |
525 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
526 | } | |
527 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
528 | ||
529 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
530 | { | |
2849450a | 531 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
532 | } |
533 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
534 | ||
2849450a MS |
535 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
536 | unsigned long msecs) | |
537 | { | |
538 | kblockd_schedule_delayed_work(&q->requeue_work, | |
539 | msecs_to_jiffies(msecs)); | |
540 | } | |
541 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
542 | ||
1885b24d JA |
543 | void blk_mq_abort_requeue_list(struct request_queue *q) |
544 | { | |
545 | unsigned long flags; | |
546 | LIST_HEAD(rq_list); | |
547 | ||
548 | spin_lock_irqsave(&q->requeue_lock, flags); | |
549 | list_splice_init(&q->requeue_list, &rq_list); | |
550 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
551 | ||
552 | while (!list_empty(&rq_list)) { | |
553 | struct request *rq; | |
554 | ||
555 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
556 | list_del_init(&rq->queuelist); | |
557 | rq->errors = -EIO; | |
558 | blk_mq_end_request(rq, rq->errors); | |
559 | } | |
560 | } | |
561 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
562 | ||
0e62f51f JA |
563 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
564 | { | |
88c7b2b7 JA |
565 | if (tag < tags->nr_tags) { |
566 | prefetch(tags->rqs[tag]); | |
4ee86bab | 567 | return tags->rqs[tag]; |
88c7b2b7 | 568 | } |
4ee86bab HR |
569 | |
570 | return NULL; | |
24d2f903 CH |
571 | } |
572 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
573 | ||
320ae51f | 574 | struct blk_mq_timeout_data { |
46f92d42 CH |
575 | unsigned long next; |
576 | unsigned int next_set; | |
320ae51f JA |
577 | }; |
578 | ||
90415837 | 579 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 580 | { |
46f92d42 CH |
581 | struct blk_mq_ops *ops = req->q->mq_ops; |
582 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; | |
87ee7b11 JA |
583 | |
584 | /* | |
585 | * We know that complete is set at this point. If STARTED isn't set | |
586 | * anymore, then the request isn't active and the "timeout" should | |
587 | * just be ignored. This can happen due to the bitflag ordering. | |
588 | * Timeout first checks if STARTED is set, and if it is, assumes | |
589 | * the request is active. But if we race with completion, then | |
590 | * we both flags will get cleared. So check here again, and ignore | |
591 | * a timeout event with a request that isn't active. | |
592 | */ | |
46f92d42 CH |
593 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
594 | return; | |
87ee7b11 | 595 | |
46f92d42 | 596 | if (ops->timeout) |
0152fb6b | 597 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
598 | |
599 | switch (ret) { | |
600 | case BLK_EH_HANDLED: | |
601 | __blk_mq_complete_request(req); | |
602 | break; | |
603 | case BLK_EH_RESET_TIMER: | |
604 | blk_add_timer(req); | |
605 | blk_clear_rq_complete(req); | |
606 | break; | |
607 | case BLK_EH_NOT_HANDLED: | |
608 | break; | |
609 | default: | |
610 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
611 | break; | |
612 | } | |
87ee7b11 | 613 | } |
5b3f25fc | 614 | |
81481eb4 CH |
615 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
616 | struct request *rq, void *priv, bool reserved) | |
617 | { | |
618 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 619 | |
eb130dbf KB |
620 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
621 | /* | |
622 | * If a request wasn't started before the queue was | |
623 | * marked dying, kill it here or it'll go unnoticed. | |
624 | */ | |
a59e0f57 KB |
625 | if (unlikely(blk_queue_dying(rq->q))) { |
626 | rq->errors = -EIO; | |
627 | blk_mq_end_request(rq, rq->errors); | |
628 | } | |
46f92d42 | 629 | return; |
eb130dbf | 630 | } |
87ee7b11 | 631 | |
46f92d42 CH |
632 | if (time_after_eq(jiffies, rq->deadline)) { |
633 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 634 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
635 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
636 | data->next = rq->deadline; | |
637 | data->next_set = 1; | |
638 | } | |
87ee7b11 JA |
639 | } |
640 | ||
287922eb | 641 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 642 | { |
287922eb CH |
643 | struct request_queue *q = |
644 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
645 | struct blk_mq_timeout_data data = { |
646 | .next = 0, | |
647 | .next_set = 0, | |
648 | }; | |
81481eb4 | 649 | int i; |
320ae51f | 650 | |
71f79fb3 GKB |
651 | /* A deadlock might occur if a request is stuck requiring a |
652 | * timeout at the same time a queue freeze is waiting | |
653 | * completion, since the timeout code would not be able to | |
654 | * acquire the queue reference here. | |
655 | * | |
656 | * That's why we don't use blk_queue_enter here; instead, we use | |
657 | * percpu_ref_tryget directly, because we need to be able to | |
658 | * obtain a reference even in the short window between the queue | |
659 | * starting to freeze, by dropping the first reference in | |
660 | * blk_mq_freeze_queue_start, and the moment the last request is | |
661 | * consumed, marked by the instant q_usage_counter reaches | |
662 | * zero. | |
663 | */ | |
664 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
665 | return; |
666 | ||
0bf6cd5b | 667 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 668 | |
81481eb4 CH |
669 | if (data.next_set) { |
670 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
671 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 672 | } else { |
0bf6cd5b CH |
673 | struct blk_mq_hw_ctx *hctx; |
674 | ||
f054b56c ML |
675 | queue_for_each_hw_ctx(q, hctx, i) { |
676 | /* the hctx may be unmapped, so check it here */ | |
677 | if (blk_mq_hw_queue_mapped(hctx)) | |
678 | blk_mq_tag_idle(hctx); | |
679 | } | |
0d2602ca | 680 | } |
287922eb | 681 | blk_queue_exit(q); |
320ae51f JA |
682 | } |
683 | ||
684 | /* | |
685 | * Reverse check our software queue for entries that we could potentially | |
686 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
687 | * too much time checking for merges. | |
688 | */ | |
689 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
690 | struct blk_mq_ctx *ctx, struct bio *bio) | |
691 | { | |
692 | struct request *rq; | |
693 | int checked = 8; | |
694 | ||
695 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
696 | int el_ret; | |
697 | ||
698 | if (!checked--) | |
699 | break; | |
700 | ||
701 | if (!blk_rq_merge_ok(rq, bio)) | |
702 | continue; | |
703 | ||
704 | el_ret = blk_try_merge(rq, bio); | |
705 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
706 | if (bio_attempt_back_merge(q, rq, bio)) { | |
707 | ctx->rq_merged++; | |
708 | return true; | |
709 | } | |
710 | break; | |
711 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
712 | if (bio_attempt_front_merge(q, rq, bio)) { | |
713 | ctx->rq_merged++; | |
714 | return true; | |
715 | } | |
716 | break; | |
717 | } | |
718 | } | |
719 | ||
720 | return false; | |
721 | } | |
722 | ||
88459642 OS |
723 | struct flush_busy_ctx_data { |
724 | struct blk_mq_hw_ctx *hctx; | |
725 | struct list_head *list; | |
726 | }; | |
727 | ||
728 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
729 | { | |
730 | struct flush_busy_ctx_data *flush_data = data; | |
731 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
732 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
733 | ||
734 | sbitmap_clear_bit(sb, bitnr); | |
735 | spin_lock(&ctx->lock); | |
736 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
737 | spin_unlock(&ctx->lock); | |
738 | return true; | |
739 | } | |
740 | ||
1429d7c9 JA |
741 | /* |
742 | * Process software queues that have been marked busy, splicing them | |
743 | * to the for-dispatch | |
744 | */ | |
745 | static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
746 | { | |
88459642 OS |
747 | struct flush_busy_ctx_data data = { |
748 | .hctx = hctx, | |
749 | .list = list, | |
750 | }; | |
1429d7c9 | 751 | |
88459642 | 752 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 753 | } |
1429d7c9 | 754 | |
703fd1c0 JA |
755 | static inline unsigned int queued_to_index(unsigned int queued) |
756 | { | |
757 | if (!queued) | |
758 | return 0; | |
1429d7c9 | 759 | |
703fd1c0 | 760 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
761 | } |
762 | ||
320ae51f JA |
763 | /* |
764 | * Run this hardware queue, pulling any software queues mapped to it in. | |
765 | * Note that this function currently has various problems around ordering | |
766 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
767 | * items on the hctx->dispatch list. Ignore that for now. | |
768 | */ | |
769 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
770 | { | |
771 | struct request_queue *q = hctx->queue; | |
320ae51f JA |
772 | struct request *rq; |
773 | LIST_HEAD(rq_list); | |
74c45052 JA |
774 | LIST_HEAD(driver_list); |
775 | struct list_head *dptr; | |
1429d7c9 | 776 | int queued; |
320ae51f | 777 | |
5d1b25c1 | 778 | if (unlikely(blk_mq_hctx_stopped(hctx))) |
320ae51f JA |
779 | return; |
780 | ||
0e87e58b JA |
781 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
782 | cpu_online(hctx->next_cpu)); | |
783 | ||
320ae51f JA |
784 | hctx->run++; |
785 | ||
786 | /* | |
787 | * Touch any software queue that has pending entries. | |
788 | */ | |
1429d7c9 | 789 | flush_busy_ctxs(hctx, &rq_list); |
320ae51f JA |
790 | |
791 | /* | |
792 | * If we have previous entries on our dispatch list, grab them | |
793 | * and stuff them at the front for more fair dispatch. | |
794 | */ | |
795 | if (!list_empty_careful(&hctx->dispatch)) { | |
796 | spin_lock(&hctx->lock); | |
797 | if (!list_empty(&hctx->dispatch)) | |
798 | list_splice_init(&hctx->dispatch, &rq_list); | |
799 | spin_unlock(&hctx->lock); | |
800 | } | |
801 | ||
74c45052 JA |
802 | /* |
803 | * Start off with dptr being NULL, so we start the first request | |
804 | * immediately, even if we have more pending. | |
805 | */ | |
806 | dptr = NULL; | |
807 | ||
320ae51f JA |
808 | /* |
809 | * Now process all the entries, sending them to the driver. | |
810 | */ | |
1429d7c9 | 811 | queued = 0; |
320ae51f | 812 | while (!list_empty(&rq_list)) { |
74c45052 | 813 | struct blk_mq_queue_data bd; |
320ae51f JA |
814 | int ret; |
815 | ||
816 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
817 | list_del_init(&rq->queuelist); | |
320ae51f | 818 | |
74c45052 JA |
819 | bd.rq = rq; |
820 | bd.list = dptr; | |
821 | bd.last = list_empty(&rq_list); | |
822 | ||
823 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
824 | switch (ret) { |
825 | case BLK_MQ_RQ_QUEUE_OK: | |
826 | queued++; | |
52b9c330 | 827 | break; |
320ae51f | 828 | case BLK_MQ_RQ_QUEUE_BUSY: |
320ae51f | 829 | list_add(&rq->queuelist, &rq_list); |
ed0791b2 | 830 | __blk_mq_requeue_request(rq); |
320ae51f JA |
831 | break; |
832 | default: | |
833 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 834 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 835 | rq->errors = -EIO; |
c8a446ad | 836 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
837 | break; |
838 | } | |
839 | ||
840 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
841 | break; | |
74c45052 JA |
842 | |
843 | /* | |
844 | * We've done the first request. If we have more than 1 | |
845 | * left in the list, set dptr to defer issue. | |
846 | */ | |
847 | if (!dptr && rq_list.next != rq_list.prev) | |
848 | dptr = &driver_list; | |
320ae51f JA |
849 | } |
850 | ||
703fd1c0 | 851 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
852 | |
853 | /* | |
854 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
855 | * that is where we will continue on next queue run. | |
856 | */ | |
857 | if (!list_empty(&rq_list)) { | |
858 | spin_lock(&hctx->lock); | |
859 | list_splice(&rq_list, &hctx->dispatch); | |
860 | spin_unlock(&hctx->lock); | |
9ba52e58 SL |
861 | /* |
862 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
863 | * it's possible the queue is stopped and restarted again | |
864 | * before this. Queue restart will dispatch requests. And since | |
865 | * requests in rq_list aren't added into hctx->dispatch yet, | |
866 | * the requests in rq_list might get lost. | |
867 | * | |
868 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
869 | **/ | |
870 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f JA |
871 | } |
872 | } | |
873 | ||
506e931f JA |
874 | /* |
875 | * It'd be great if the workqueue API had a way to pass | |
876 | * in a mask and had some smarts for more clever placement. | |
877 | * For now we just round-robin here, switching for every | |
878 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
879 | */ | |
880 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
881 | { | |
b657d7e6 CH |
882 | if (hctx->queue->nr_hw_queues == 1) |
883 | return WORK_CPU_UNBOUND; | |
506e931f JA |
884 | |
885 | if (--hctx->next_cpu_batch <= 0) { | |
b657d7e6 | 886 | int cpu = hctx->next_cpu, next_cpu; |
506e931f JA |
887 | |
888 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
889 | if (next_cpu >= nr_cpu_ids) | |
890 | next_cpu = cpumask_first(hctx->cpumask); | |
891 | ||
892 | hctx->next_cpu = next_cpu; | |
893 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
b657d7e6 CH |
894 | |
895 | return cpu; | |
506e931f JA |
896 | } |
897 | ||
b657d7e6 | 898 | return hctx->next_cpu; |
506e931f JA |
899 | } |
900 | ||
320ae51f JA |
901 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
902 | { | |
5d1b25c1 BVA |
903 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
904 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
905 | return; |
906 | ||
1b792f2f | 907 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
908 | int cpu = get_cpu(); |
909 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 910 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 911 | put_cpu(); |
398205b8 PB |
912 | return; |
913 | } | |
e4043dcf | 914 | |
2a90d4aa | 915 | put_cpu(); |
e4043dcf | 916 | } |
398205b8 | 917 | |
27489a3c | 918 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
919 | } |
920 | ||
b94ec296 | 921 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
922 | { |
923 | struct blk_mq_hw_ctx *hctx; | |
924 | int i; | |
925 | ||
926 | queue_for_each_hw_ctx(q, hctx, i) { | |
927 | if ((!blk_mq_hctx_has_pending(hctx) && | |
928 | list_empty_careful(&hctx->dispatch)) || | |
5d1b25c1 | 929 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
930 | continue; |
931 | ||
b94ec296 | 932 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
933 | } |
934 | } | |
b94ec296 | 935 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 936 | |
fd001443 BVA |
937 | /** |
938 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
939 | * @q: request queue. | |
940 | * | |
941 | * The caller is responsible for serializing this function against | |
942 | * blk_mq_{start,stop}_hw_queue(). | |
943 | */ | |
944 | bool blk_mq_queue_stopped(struct request_queue *q) | |
945 | { | |
946 | struct blk_mq_hw_ctx *hctx; | |
947 | int i; | |
948 | ||
949 | queue_for_each_hw_ctx(q, hctx, i) | |
950 | if (blk_mq_hctx_stopped(hctx)) | |
951 | return true; | |
952 | ||
953 | return false; | |
954 | } | |
955 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
956 | ||
320ae51f JA |
957 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
958 | { | |
27489a3c | 959 | cancel_work(&hctx->run_work); |
70f4db63 | 960 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
961 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
962 | } | |
963 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
964 | ||
280d45f6 CH |
965 | void blk_mq_stop_hw_queues(struct request_queue *q) |
966 | { | |
967 | struct blk_mq_hw_ctx *hctx; | |
968 | int i; | |
969 | ||
970 | queue_for_each_hw_ctx(q, hctx, i) | |
971 | blk_mq_stop_hw_queue(hctx); | |
972 | } | |
973 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
974 | ||
320ae51f JA |
975 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
976 | { | |
977 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 978 | |
0ffbce80 | 979 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
980 | } |
981 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
982 | ||
2f268556 CH |
983 | void blk_mq_start_hw_queues(struct request_queue *q) |
984 | { | |
985 | struct blk_mq_hw_ctx *hctx; | |
986 | int i; | |
987 | ||
988 | queue_for_each_hw_ctx(q, hctx, i) | |
989 | blk_mq_start_hw_queue(hctx); | |
990 | } | |
991 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
992 | ||
1b4a3258 | 993 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
994 | { |
995 | struct blk_mq_hw_ctx *hctx; | |
996 | int i; | |
997 | ||
998 | queue_for_each_hw_ctx(q, hctx, i) { | |
5d1b25c1 | 999 | if (!blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1000 | continue; |
1001 | ||
1002 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1b4a3258 | 1003 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1004 | } |
1005 | } | |
1006 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1007 | ||
70f4db63 | 1008 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1009 | { |
1010 | struct blk_mq_hw_ctx *hctx; | |
1011 | ||
27489a3c | 1012 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1013 | |
320ae51f JA |
1014 | __blk_mq_run_hw_queue(hctx); |
1015 | } | |
1016 | ||
70f4db63 CH |
1017 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1018 | { | |
1019 | struct blk_mq_hw_ctx *hctx; | |
1020 | ||
1021 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1022 | ||
1023 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1024 | __blk_mq_run_hw_queue(hctx); | |
1025 | } | |
1026 | ||
1027 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1028 | { | |
19c66e59 ML |
1029 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1030 | return; | |
70f4db63 | 1031 | |
b657d7e6 CH |
1032 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1033 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1034 | } |
1035 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1036 | ||
cfd0c552 | 1037 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1038 | struct request *rq, |
1039 | bool at_head) | |
320ae51f | 1040 | { |
e57690fe JA |
1041 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1042 | ||
01b983c9 JA |
1043 | trace_block_rq_insert(hctx->queue, rq); |
1044 | ||
72a0a36e CH |
1045 | if (at_head) |
1046 | list_add(&rq->queuelist, &ctx->rq_list); | |
1047 | else | |
1048 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1049 | } |
4bb659b1 | 1050 | |
cfd0c552 ML |
1051 | static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, |
1052 | struct request *rq, bool at_head) | |
1053 | { | |
1054 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1055 | ||
e57690fe | 1056 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1057 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1058 | } |
1059 | ||
eeabc850 | 1060 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1061 | bool async) |
320ae51f | 1062 | { |
e57690fe | 1063 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1064 | struct request_queue *q = rq->q; |
7d7e0f90 | 1065 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1066 | |
a57a178a CH |
1067 | spin_lock(&ctx->lock); |
1068 | __blk_mq_insert_request(hctx, rq, at_head); | |
1069 | spin_unlock(&ctx->lock); | |
320ae51f | 1070 | |
320ae51f JA |
1071 | if (run_queue) |
1072 | blk_mq_run_hw_queue(hctx, async); | |
1073 | } | |
1074 | ||
1075 | static void blk_mq_insert_requests(struct request_queue *q, | |
1076 | struct blk_mq_ctx *ctx, | |
1077 | struct list_head *list, | |
1078 | int depth, | |
1079 | bool from_schedule) | |
1080 | ||
1081 | { | |
7d7e0f90 | 1082 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1083 | |
1084 | trace_block_unplug(q, depth, !from_schedule); | |
1085 | ||
320ae51f JA |
1086 | /* |
1087 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1088 | * offline now | |
1089 | */ | |
1090 | spin_lock(&ctx->lock); | |
1091 | while (!list_empty(list)) { | |
1092 | struct request *rq; | |
1093 | ||
1094 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1095 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1096 | list_del_init(&rq->queuelist); |
e57690fe | 1097 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1098 | } |
cfd0c552 | 1099 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1100 | spin_unlock(&ctx->lock); |
1101 | ||
320ae51f JA |
1102 | blk_mq_run_hw_queue(hctx, from_schedule); |
1103 | } | |
1104 | ||
1105 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1106 | { | |
1107 | struct request *rqa = container_of(a, struct request, queuelist); | |
1108 | struct request *rqb = container_of(b, struct request, queuelist); | |
1109 | ||
1110 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1111 | (rqa->mq_ctx == rqb->mq_ctx && | |
1112 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1113 | } | |
1114 | ||
1115 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1116 | { | |
1117 | struct blk_mq_ctx *this_ctx; | |
1118 | struct request_queue *this_q; | |
1119 | struct request *rq; | |
1120 | LIST_HEAD(list); | |
1121 | LIST_HEAD(ctx_list); | |
1122 | unsigned int depth; | |
1123 | ||
1124 | list_splice_init(&plug->mq_list, &list); | |
1125 | ||
1126 | list_sort(NULL, &list, plug_ctx_cmp); | |
1127 | ||
1128 | this_q = NULL; | |
1129 | this_ctx = NULL; | |
1130 | depth = 0; | |
1131 | ||
1132 | while (!list_empty(&list)) { | |
1133 | rq = list_entry_rq(list.next); | |
1134 | list_del_init(&rq->queuelist); | |
1135 | BUG_ON(!rq->q); | |
1136 | if (rq->mq_ctx != this_ctx) { | |
1137 | if (this_ctx) { | |
1138 | blk_mq_insert_requests(this_q, this_ctx, | |
1139 | &ctx_list, depth, | |
1140 | from_schedule); | |
1141 | } | |
1142 | ||
1143 | this_ctx = rq->mq_ctx; | |
1144 | this_q = rq->q; | |
1145 | depth = 0; | |
1146 | } | |
1147 | ||
1148 | depth++; | |
1149 | list_add_tail(&rq->queuelist, &ctx_list); | |
1150 | } | |
1151 | ||
1152 | /* | |
1153 | * If 'this_ctx' is set, we know we have entries to complete | |
1154 | * on 'ctx_list'. Do those. | |
1155 | */ | |
1156 | if (this_ctx) { | |
1157 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1158 | from_schedule); | |
1159 | } | |
1160 | } | |
1161 | ||
1162 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1163 | { | |
1164 | init_request_from_bio(rq, bio); | |
4b570521 | 1165 | |
a21f2a3e | 1166 | blk_account_io_start(rq, 1); |
320ae51f JA |
1167 | } |
1168 | ||
274a5843 JA |
1169 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1170 | { | |
1171 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1172 | !blk_queue_nomerges(hctx->queue); | |
1173 | } | |
1174 | ||
07068d5b JA |
1175 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1176 | struct blk_mq_ctx *ctx, | |
1177 | struct request *rq, struct bio *bio) | |
320ae51f | 1178 | { |
e18378a6 | 1179 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1180 | blk_mq_bio_to_request(rq, bio); |
1181 | spin_lock(&ctx->lock); | |
1182 | insert_rq: | |
1183 | __blk_mq_insert_request(hctx, rq, false); | |
1184 | spin_unlock(&ctx->lock); | |
1185 | return false; | |
1186 | } else { | |
274a5843 JA |
1187 | struct request_queue *q = hctx->queue; |
1188 | ||
07068d5b JA |
1189 | spin_lock(&ctx->lock); |
1190 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1191 | blk_mq_bio_to_request(rq, bio); | |
1192 | goto insert_rq; | |
1193 | } | |
320ae51f | 1194 | |
07068d5b JA |
1195 | spin_unlock(&ctx->lock); |
1196 | __blk_mq_free_request(hctx, ctx, rq); | |
1197 | return true; | |
14ec77f3 | 1198 | } |
07068d5b | 1199 | } |
14ec77f3 | 1200 | |
07068d5b JA |
1201 | static struct request *blk_mq_map_request(struct request_queue *q, |
1202 | struct bio *bio, | |
2552e3f8 | 1203 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1204 | { |
1205 | struct blk_mq_hw_ctx *hctx; | |
1206 | struct blk_mq_ctx *ctx; | |
1207 | struct request *rq; | |
320ae51f | 1208 | |
3ef28e83 | 1209 | blk_queue_enter_live(q); |
320ae51f | 1210 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1211 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1212 | |
ef295ecf | 1213 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1214 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1215 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1216 | |
7dd2fb68 | 1217 | data->hctx->queued++; |
07068d5b JA |
1218 | return rq; |
1219 | } | |
1220 | ||
2253efc8 BVA |
1221 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1222 | struct request *rq, blk_qc_t *cookie) | |
f984df1f SL |
1223 | { |
1224 | int ret; | |
1225 | struct request_queue *q = rq->q; | |
f984df1f SL |
1226 | struct blk_mq_queue_data bd = { |
1227 | .rq = rq, | |
1228 | .list = NULL, | |
1229 | .last = 1 | |
1230 | }; | |
7b371636 | 1231 | blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num); |
f984df1f | 1232 | |
2253efc8 BVA |
1233 | if (blk_mq_hctx_stopped(hctx)) |
1234 | goto insert; | |
1235 | ||
f984df1f SL |
1236 | /* |
1237 | * For OK queue, we are done. For error, kill it. Any other | |
1238 | * error (busy), just add it to our list as we previously | |
1239 | * would have done | |
1240 | */ | |
1241 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1242 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1243 | *cookie = new_cookie; | |
2253efc8 | 1244 | return; |
7b371636 | 1245 | } |
f984df1f | 1246 | |
7b371636 JA |
1247 | __blk_mq_requeue_request(rq); |
1248 | ||
1249 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1250 | *cookie = BLK_QC_T_NONE; | |
1251 | rq->errors = -EIO; | |
1252 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1253 | return; |
f984df1f | 1254 | } |
7b371636 | 1255 | |
2253efc8 BVA |
1256 | insert: |
1257 | blk_mq_insert_request(rq, false, true, true); | |
f984df1f SL |
1258 | } |
1259 | ||
07068d5b JA |
1260 | /* |
1261 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1262 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1263 | * hardware for SYNC IO. | |
1264 | */ | |
dece1635 | 1265 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1266 | { |
ef295ecf | 1267 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1268 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1269 | struct blk_mq_alloc_data data; |
07068d5b | 1270 | struct request *rq; |
f984df1f SL |
1271 | unsigned int request_count = 0; |
1272 | struct blk_plug *plug; | |
5b3f341f | 1273 | struct request *same_queue_rq = NULL; |
7b371636 | 1274 | blk_qc_t cookie; |
07068d5b JA |
1275 | |
1276 | blk_queue_bounce(q, &bio); | |
1277 | ||
1278 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1279 | bio_io_error(bio); |
dece1635 | 1280 | return BLK_QC_T_NONE; |
07068d5b JA |
1281 | } |
1282 | ||
54efd50b KO |
1283 | blk_queue_split(q, &bio, q->bio_split); |
1284 | ||
87c279e6 OS |
1285 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1286 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1287 | return BLK_QC_T_NONE; | |
f984df1f | 1288 | |
07068d5b JA |
1289 | rq = blk_mq_map_request(q, bio, &data); |
1290 | if (unlikely(!rq)) | |
dece1635 | 1291 | return BLK_QC_T_NONE; |
07068d5b | 1292 | |
7b371636 | 1293 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
07068d5b JA |
1294 | |
1295 | if (unlikely(is_flush_fua)) { | |
1296 | blk_mq_bio_to_request(rq, bio); | |
1297 | blk_insert_flush(rq); | |
1298 | goto run_queue; | |
1299 | } | |
1300 | ||
f984df1f | 1301 | plug = current->plug; |
e167dfb5 JA |
1302 | /* |
1303 | * If the driver supports defer issued based on 'last', then | |
1304 | * queue it up like normal since we can potentially save some | |
1305 | * CPU this way. | |
1306 | */ | |
f984df1f SL |
1307 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1308 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1309 | struct request *old_rq = NULL; | |
07068d5b JA |
1310 | |
1311 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1312 | |
1313 | /* | |
b094f89c | 1314 | * We do limited pluging. If the bio can be merged, do that. |
f984df1f SL |
1315 | * Otherwise the existing request in the plug list will be |
1316 | * issued. So the plug list will have one request at most | |
07068d5b | 1317 | */ |
f984df1f | 1318 | if (plug) { |
5b3f341f SL |
1319 | /* |
1320 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1321 | * happens, same_queue_rq is invalid and plug list is |
1322 | * empty | |
1323 | */ | |
5b3f341f SL |
1324 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1325 | old_rq = same_queue_rq; | |
f984df1f | 1326 | list_del_init(&old_rq->queuelist); |
07068d5b | 1327 | } |
f984df1f SL |
1328 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1329 | } else /* is_sync */ | |
1330 | old_rq = rq; | |
1331 | blk_mq_put_ctx(data.ctx); | |
1332 | if (!old_rq) | |
7b371636 | 1333 | goto done; |
2253efc8 | 1334 | blk_mq_try_issue_directly(data.hctx, old_rq, &cookie); |
7b371636 | 1335 | goto done; |
07068d5b JA |
1336 | } |
1337 | ||
1338 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1339 | /* | |
1340 | * For a SYNC request, send it to the hardware immediately. For | |
1341 | * an ASYNC request, just ensure that we run it later on. The | |
1342 | * latter allows for merging opportunities and more efficient | |
1343 | * dispatching. | |
1344 | */ | |
1345 | run_queue: | |
1346 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1347 | } | |
07068d5b | 1348 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1349 | done: |
1350 | return cookie; | |
07068d5b JA |
1351 | } |
1352 | ||
1353 | /* | |
1354 | * Single hardware queue variant. This will attempt to use any per-process | |
1355 | * plug for merging and IO deferral. | |
1356 | */ | |
dece1635 | 1357 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1358 | { |
ef295ecf | 1359 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1360 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1361 | struct blk_plug *plug; |
1362 | unsigned int request_count = 0; | |
2552e3f8 | 1363 | struct blk_mq_alloc_data data; |
07068d5b | 1364 | struct request *rq; |
7b371636 | 1365 | blk_qc_t cookie; |
07068d5b | 1366 | |
07068d5b JA |
1367 | blk_queue_bounce(q, &bio); |
1368 | ||
1369 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1370 | bio_io_error(bio); |
dece1635 | 1371 | return BLK_QC_T_NONE; |
07068d5b JA |
1372 | } |
1373 | ||
54efd50b KO |
1374 | blk_queue_split(q, &bio, q->bio_split); |
1375 | ||
87c279e6 OS |
1376 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1377 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1378 | return BLK_QC_T_NONE; | |
1379 | } else | |
1380 | request_count = blk_plug_queued_count(q); | |
07068d5b JA |
1381 | |
1382 | rq = blk_mq_map_request(q, bio, &data); | |
ff87bcec | 1383 | if (unlikely(!rq)) |
dece1635 | 1384 | return BLK_QC_T_NONE; |
320ae51f | 1385 | |
7b371636 | 1386 | cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num); |
320ae51f JA |
1387 | |
1388 | if (unlikely(is_flush_fua)) { | |
1389 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1390 | blk_insert_flush(rq); |
1391 | goto run_queue; | |
1392 | } | |
1393 | ||
1394 | /* | |
1395 | * A task plug currently exists. Since this is completely lockless, | |
1396 | * utilize that to temporarily store requests until the task is | |
1397 | * either done or scheduled away. | |
1398 | */ | |
e6c4438b JM |
1399 | plug = current->plug; |
1400 | if (plug) { | |
1401 | blk_mq_bio_to_request(rq, bio); | |
676d0607 | 1402 | if (!request_count) |
e6c4438b | 1403 | trace_block_plug(q); |
b094f89c JA |
1404 | |
1405 | blk_mq_put_ctx(data.ctx); | |
1406 | ||
1407 | if (request_count >= BLK_MAX_REQUEST_COUNT) { | |
e6c4438b JM |
1408 | blk_flush_plug_list(plug, false); |
1409 | trace_block_plug(q); | |
320ae51f | 1410 | } |
b094f89c | 1411 | |
e6c4438b | 1412 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1413 | return cookie; |
320ae51f JA |
1414 | } |
1415 | ||
07068d5b JA |
1416 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1417 | /* | |
1418 | * For a SYNC request, send it to the hardware immediately. For | |
1419 | * an ASYNC request, just ensure that we run it later on. The | |
1420 | * latter allows for merging opportunities and more efficient | |
1421 | * dispatching. | |
1422 | */ | |
1423 | run_queue: | |
1424 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1425 | } |
1426 | ||
07068d5b | 1427 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1428 | return cookie; |
320ae51f JA |
1429 | } |
1430 | ||
24d2f903 CH |
1431 | static void blk_mq_free_rq_map(struct blk_mq_tag_set *set, |
1432 | struct blk_mq_tags *tags, unsigned int hctx_idx) | |
95363efd | 1433 | { |
e9b267d9 | 1434 | struct page *page; |
320ae51f | 1435 | |
24d2f903 | 1436 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1437 | int i; |
320ae51f | 1438 | |
24d2f903 CH |
1439 | for (i = 0; i < tags->nr_tags; i++) { |
1440 | if (!tags->rqs[i]) | |
e9b267d9 | 1441 | continue; |
24d2f903 CH |
1442 | set->ops->exit_request(set->driver_data, tags->rqs[i], |
1443 | hctx_idx, i); | |
a5164405 | 1444 | tags->rqs[i] = NULL; |
e9b267d9 | 1445 | } |
320ae51f | 1446 | } |
320ae51f | 1447 | |
24d2f903 CH |
1448 | while (!list_empty(&tags->page_list)) { |
1449 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1450 | list_del_init(&page->lru); |
f75782e4 CM |
1451 | /* |
1452 | * Remove kmemleak object previously allocated in | |
1453 | * blk_mq_init_rq_map(). | |
1454 | */ | |
1455 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1456 | __free_pages(page, page->private); |
1457 | } | |
1458 | ||
24d2f903 | 1459 | kfree(tags->rqs); |
320ae51f | 1460 | |
24d2f903 | 1461 | blk_mq_free_tags(tags); |
320ae51f JA |
1462 | } |
1463 | ||
1464 | static size_t order_to_size(unsigned int order) | |
1465 | { | |
4ca08500 | 1466 | return (size_t)PAGE_SIZE << order; |
320ae51f JA |
1467 | } |
1468 | ||
24d2f903 CH |
1469 | static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set, |
1470 | unsigned int hctx_idx) | |
320ae51f | 1471 | { |
24d2f903 | 1472 | struct blk_mq_tags *tags; |
320ae51f JA |
1473 | unsigned int i, j, entries_per_page, max_order = 4; |
1474 | size_t rq_size, left; | |
1475 | ||
24d2f903 | 1476 | tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags, |
24391c0d SL |
1477 | set->numa_node, |
1478 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1479 | if (!tags) |
1480 | return NULL; | |
320ae51f | 1481 | |
24d2f903 CH |
1482 | INIT_LIST_HEAD(&tags->page_list); |
1483 | ||
a5164405 JA |
1484 | tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *), |
1485 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY, | |
1486 | set->numa_node); | |
24d2f903 CH |
1487 | if (!tags->rqs) { |
1488 | blk_mq_free_tags(tags); | |
1489 | return NULL; | |
1490 | } | |
320ae51f JA |
1491 | |
1492 | /* | |
1493 | * rq_size is the size of the request plus driver payload, rounded | |
1494 | * to the cacheline size | |
1495 | */ | |
24d2f903 | 1496 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1497 | cache_line_size()); |
24d2f903 | 1498 | left = rq_size * set->queue_depth; |
320ae51f | 1499 | |
24d2f903 | 1500 | for (i = 0; i < set->queue_depth; ) { |
320ae51f JA |
1501 | int this_order = max_order; |
1502 | struct page *page; | |
1503 | int to_do; | |
1504 | void *p; | |
1505 | ||
b3a834b1 | 1506 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1507 | this_order--; |
1508 | ||
1509 | do { | |
a5164405 | 1510 | page = alloc_pages_node(set->numa_node, |
ac211175 | 1511 | GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1512 | this_order); |
320ae51f JA |
1513 | if (page) |
1514 | break; | |
1515 | if (!this_order--) | |
1516 | break; | |
1517 | if (order_to_size(this_order) < rq_size) | |
1518 | break; | |
1519 | } while (1); | |
1520 | ||
1521 | if (!page) | |
24d2f903 | 1522 | goto fail; |
320ae51f JA |
1523 | |
1524 | page->private = this_order; | |
24d2f903 | 1525 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1526 | |
1527 | p = page_address(page); | |
f75782e4 CM |
1528 | /* |
1529 | * Allow kmemleak to scan these pages as they contain pointers | |
1530 | * to additional allocations like via ops->init_request(). | |
1531 | */ | |
1532 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL); | |
320ae51f | 1533 | entries_per_page = order_to_size(this_order) / rq_size; |
24d2f903 | 1534 | to_do = min(entries_per_page, set->queue_depth - i); |
320ae51f JA |
1535 | left -= to_do * rq_size; |
1536 | for (j = 0; j < to_do; j++) { | |
24d2f903 CH |
1537 | tags->rqs[i] = p; |
1538 | if (set->ops->init_request) { | |
1539 | if (set->ops->init_request(set->driver_data, | |
1540 | tags->rqs[i], hctx_idx, i, | |
a5164405 JA |
1541 | set->numa_node)) { |
1542 | tags->rqs[i] = NULL; | |
24d2f903 | 1543 | goto fail; |
a5164405 | 1544 | } |
e9b267d9 CH |
1545 | } |
1546 | ||
320ae51f JA |
1547 | p += rq_size; |
1548 | i++; | |
1549 | } | |
1550 | } | |
24d2f903 | 1551 | return tags; |
320ae51f | 1552 | |
24d2f903 | 1553 | fail: |
24d2f903 CH |
1554 | blk_mq_free_rq_map(set, tags, hctx_idx); |
1555 | return NULL; | |
320ae51f JA |
1556 | } |
1557 | ||
e57690fe JA |
1558 | /* |
1559 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1560 | * software queue to the hw queue dispatch list, and ensure that it | |
1561 | * gets run. | |
1562 | */ | |
9467f859 | 1563 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1564 | { |
9467f859 | 1565 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1566 | struct blk_mq_ctx *ctx; |
1567 | LIST_HEAD(tmp); | |
1568 | ||
9467f859 | 1569 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1570 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1571 | |
1572 | spin_lock(&ctx->lock); | |
1573 | if (!list_empty(&ctx->rq_list)) { | |
1574 | list_splice_init(&ctx->rq_list, &tmp); | |
1575 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1576 | } | |
1577 | spin_unlock(&ctx->lock); | |
1578 | ||
1579 | if (list_empty(&tmp)) | |
9467f859 | 1580 | return 0; |
484b4061 | 1581 | |
e57690fe JA |
1582 | spin_lock(&hctx->lock); |
1583 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1584 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1585 | |
1586 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1587 | return 0; |
484b4061 JA |
1588 | } |
1589 | ||
9467f859 | 1590 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1591 | { |
9467f859 TG |
1592 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1593 | &hctx->cpuhp_dead); | |
484b4061 JA |
1594 | } |
1595 | ||
c3b4afca | 1596 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1597 | static void blk_mq_exit_hctx(struct request_queue *q, |
1598 | struct blk_mq_tag_set *set, | |
1599 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1600 | { | |
f70ced09 ML |
1601 | unsigned flush_start_tag = set->queue_depth; |
1602 | ||
08e98fc6 ML |
1603 | blk_mq_tag_idle(hctx); |
1604 | ||
f70ced09 ML |
1605 | if (set->ops->exit_request) |
1606 | set->ops->exit_request(set->driver_data, | |
1607 | hctx->fq->flush_rq, hctx_idx, | |
1608 | flush_start_tag + hctx_idx); | |
1609 | ||
08e98fc6 ML |
1610 | if (set->ops->exit_hctx) |
1611 | set->ops->exit_hctx(hctx, hctx_idx); | |
1612 | ||
9467f859 | 1613 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1614 | blk_free_flush_queue(hctx->fq); |
88459642 | 1615 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1616 | } |
1617 | ||
624dbe47 ML |
1618 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1619 | struct blk_mq_tag_set *set, int nr_queue) | |
1620 | { | |
1621 | struct blk_mq_hw_ctx *hctx; | |
1622 | unsigned int i; | |
1623 | ||
1624 | queue_for_each_hw_ctx(q, hctx, i) { | |
1625 | if (i == nr_queue) | |
1626 | break; | |
08e98fc6 | 1627 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1628 | } |
624dbe47 ML |
1629 | } |
1630 | ||
1631 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1632 | struct blk_mq_tag_set *set) | |
1633 | { | |
1634 | struct blk_mq_hw_ctx *hctx; | |
1635 | unsigned int i; | |
1636 | ||
e09aae7e | 1637 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1638 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1639 | } |
1640 | ||
08e98fc6 ML |
1641 | static int blk_mq_init_hctx(struct request_queue *q, |
1642 | struct blk_mq_tag_set *set, | |
1643 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1644 | { |
08e98fc6 | 1645 | int node; |
f70ced09 | 1646 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1647 | |
1648 | node = hctx->numa_node; | |
1649 | if (node == NUMA_NO_NODE) | |
1650 | node = hctx->numa_node = set->numa_node; | |
1651 | ||
27489a3c | 1652 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1653 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1654 | spin_lock_init(&hctx->lock); | |
1655 | INIT_LIST_HEAD(&hctx->dispatch); | |
1656 | hctx->queue = q; | |
1657 | hctx->queue_num = hctx_idx; | |
2404e607 | 1658 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1659 | |
9467f859 | 1660 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1661 | |
1662 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1663 | |
1664 | /* | |
08e98fc6 ML |
1665 | * Allocate space for all possible cpus to avoid allocation at |
1666 | * runtime | |
320ae51f | 1667 | */ |
08e98fc6 ML |
1668 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1669 | GFP_KERNEL, node); | |
1670 | if (!hctx->ctxs) | |
1671 | goto unregister_cpu_notifier; | |
320ae51f | 1672 | |
88459642 OS |
1673 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1674 | node)) | |
08e98fc6 | 1675 | goto free_ctxs; |
320ae51f | 1676 | |
08e98fc6 | 1677 | hctx->nr_ctx = 0; |
320ae51f | 1678 | |
08e98fc6 ML |
1679 | if (set->ops->init_hctx && |
1680 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1681 | goto free_bitmap; | |
320ae51f | 1682 | |
f70ced09 ML |
1683 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1684 | if (!hctx->fq) | |
1685 | goto exit_hctx; | |
320ae51f | 1686 | |
f70ced09 ML |
1687 | if (set->ops->init_request && |
1688 | set->ops->init_request(set->driver_data, | |
1689 | hctx->fq->flush_rq, hctx_idx, | |
1690 | flush_start_tag + hctx_idx, node)) | |
1691 | goto free_fq; | |
320ae51f | 1692 | |
08e98fc6 | 1693 | return 0; |
320ae51f | 1694 | |
f70ced09 ML |
1695 | free_fq: |
1696 | kfree(hctx->fq); | |
1697 | exit_hctx: | |
1698 | if (set->ops->exit_hctx) | |
1699 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1700 | free_bitmap: |
88459642 | 1701 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1702 | free_ctxs: |
1703 | kfree(hctx->ctxs); | |
1704 | unregister_cpu_notifier: | |
9467f859 | 1705 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1706 | return -1; |
1707 | } | |
320ae51f | 1708 | |
320ae51f JA |
1709 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1710 | unsigned int nr_hw_queues) | |
1711 | { | |
1712 | unsigned int i; | |
1713 | ||
1714 | for_each_possible_cpu(i) { | |
1715 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1716 | struct blk_mq_hw_ctx *hctx; | |
1717 | ||
1718 | memset(__ctx, 0, sizeof(*__ctx)); | |
1719 | __ctx->cpu = i; | |
1720 | spin_lock_init(&__ctx->lock); | |
1721 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1722 | __ctx->queue = q; | |
1723 | ||
1724 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1725 | if (!cpu_online(i)) |
1726 | continue; | |
1727 | ||
7d7e0f90 | 1728 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1729 | |
320ae51f JA |
1730 | /* |
1731 | * Set local node, IFF we have more than one hw queue. If | |
1732 | * not, we remain on the home node of the device | |
1733 | */ | |
1734 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1735 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1736 | } |
1737 | } | |
1738 | ||
5778322e AM |
1739 | static void blk_mq_map_swqueue(struct request_queue *q, |
1740 | const struct cpumask *online_mask) | |
320ae51f JA |
1741 | { |
1742 | unsigned int i; | |
1743 | struct blk_mq_hw_ctx *hctx; | |
1744 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1745 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1746 | |
60de074b AM |
1747 | /* |
1748 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1749 | */ | |
1750 | mutex_lock(&q->sysfs_lock); | |
1751 | ||
320ae51f | 1752 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1753 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1754 | hctx->nr_ctx = 0; |
1755 | } | |
1756 | ||
1757 | /* | |
1758 | * Map software to hardware queues | |
1759 | */ | |
897bb0c7 | 1760 | for_each_possible_cpu(i) { |
320ae51f | 1761 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1762 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1763 | continue; |
1764 | ||
897bb0c7 | 1765 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1766 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1767 | |
e4043dcf | 1768 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1769 | ctx->index_hw = hctx->nr_ctx; |
1770 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1771 | } | |
506e931f | 1772 | |
60de074b AM |
1773 | mutex_unlock(&q->sysfs_lock); |
1774 | ||
506e931f | 1775 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1776 | /* |
a68aafa5 JA |
1777 | * If no software queues are mapped to this hardware queue, |
1778 | * disable it and free the request entries. | |
484b4061 JA |
1779 | */ |
1780 | if (!hctx->nr_ctx) { | |
484b4061 JA |
1781 | if (set->tags[i]) { |
1782 | blk_mq_free_rq_map(set, set->tags[i], i); | |
1783 | set->tags[i] = NULL; | |
484b4061 | 1784 | } |
2a34c087 | 1785 | hctx->tags = NULL; |
484b4061 JA |
1786 | continue; |
1787 | } | |
1788 | ||
2a34c087 ML |
1789 | /* unmapped hw queue can be remapped after CPU topo changed */ |
1790 | if (!set->tags[i]) | |
1791 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
1792 | hctx->tags = set->tags[i]; | |
1793 | WARN_ON(!hctx->tags); | |
1794 | ||
889fa31f CY |
1795 | /* |
1796 | * Set the map size to the number of mapped software queues. | |
1797 | * This is more accurate and more efficient than looping | |
1798 | * over all possibly mapped software queues. | |
1799 | */ | |
88459642 | 1800 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 1801 | |
484b4061 JA |
1802 | /* |
1803 | * Initialize batch roundrobin counts | |
1804 | */ | |
506e931f JA |
1805 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
1806 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1807 | } | |
320ae51f JA |
1808 | } |
1809 | ||
2404e607 | 1810 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
1811 | { |
1812 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
1813 | int i; |
1814 | ||
2404e607 JM |
1815 | queue_for_each_hw_ctx(q, hctx, i) { |
1816 | if (shared) | |
1817 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
1818 | else | |
1819 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1820 | } | |
1821 | } | |
1822 | ||
1823 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
1824 | { | |
1825 | struct request_queue *q; | |
0d2602ca JA |
1826 | |
1827 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
1828 | blk_mq_freeze_queue(q); | |
2404e607 | 1829 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
1830 | blk_mq_unfreeze_queue(q); |
1831 | } | |
1832 | } | |
1833 | ||
1834 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
1835 | { | |
1836 | struct blk_mq_tag_set *set = q->tag_set; | |
1837 | ||
0d2602ca JA |
1838 | mutex_lock(&set->tag_list_lock); |
1839 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
1840 | if (list_is_singular(&set->tag_list)) { |
1841 | /* just transitioned to unshared */ | |
1842 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
1843 | /* update existing queue */ | |
1844 | blk_mq_update_tag_set_depth(set, false); | |
1845 | } | |
0d2602ca | 1846 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
1847 | } |
1848 | ||
1849 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
1850 | struct request_queue *q) | |
1851 | { | |
1852 | q->tag_set = set; | |
1853 | ||
1854 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
1855 | |
1856 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
1857 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
1858 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
1859 | /* update existing queue */ | |
1860 | blk_mq_update_tag_set_depth(set, true); | |
1861 | } | |
1862 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
1863 | queue_set_hctx_shared(q, true); | |
0d2602ca | 1864 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 1865 | |
0d2602ca JA |
1866 | mutex_unlock(&set->tag_list_lock); |
1867 | } | |
1868 | ||
e09aae7e ML |
1869 | /* |
1870 | * It is the actual release handler for mq, but we do it from | |
1871 | * request queue's release handler for avoiding use-after-free | |
1872 | * and headache because q->mq_kobj shouldn't have been introduced, | |
1873 | * but we can't group ctx/kctx kobj without it. | |
1874 | */ | |
1875 | void blk_mq_release(struct request_queue *q) | |
1876 | { | |
1877 | struct blk_mq_hw_ctx *hctx; | |
1878 | unsigned int i; | |
1879 | ||
1880 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
1881 | queue_for_each_hw_ctx(q, hctx, i) { |
1882 | if (!hctx) | |
1883 | continue; | |
1884 | kfree(hctx->ctxs); | |
e09aae7e | 1885 | kfree(hctx); |
c3b4afca | 1886 | } |
e09aae7e | 1887 | |
a723bab3 AM |
1888 | q->mq_map = NULL; |
1889 | ||
e09aae7e ML |
1890 | kfree(q->queue_hw_ctx); |
1891 | ||
1892 | /* ctx kobj stays in queue_ctx */ | |
1893 | free_percpu(q->queue_ctx); | |
1894 | } | |
1895 | ||
24d2f903 | 1896 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
1897 | { |
1898 | struct request_queue *uninit_q, *q; | |
1899 | ||
1900 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
1901 | if (!uninit_q) | |
1902 | return ERR_PTR(-ENOMEM); | |
1903 | ||
1904 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
1905 | if (IS_ERR(q)) | |
1906 | blk_cleanup_queue(uninit_q); | |
1907 | ||
1908 | return q; | |
1909 | } | |
1910 | EXPORT_SYMBOL(blk_mq_init_queue); | |
1911 | ||
868f2f0b KB |
1912 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
1913 | struct request_queue *q) | |
320ae51f | 1914 | { |
868f2f0b KB |
1915 | int i, j; |
1916 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 1917 | |
868f2f0b | 1918 | blk_mq_sysfs_unregister(q); |
24d2f903 | 1919 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 1920 | int node; |
f14bbe77 | 1921 | |
868f2f0b KB |
1922 | if (hctxs[i]) |
1923 | continue; | |
1924 | ||
1925 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
1926 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
1927 | GFP_KERNEL, node); | |
320ae51f | 1928 | if (!hctxs[i]) |
868f2f0b | 1929 | break; |
320ae51f | 1930 | |
a86073e4 | 1931 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
1932 | node)) { |
1933 | kfree(hctxs[i]); | |
1934 | hctxs[i] = NULL; | |
1935 | break; | |
1936 | } | |
e4043dcf | 1937 | |
0d2602ca | 1938 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 1939 | hctxs[i]->numa_node = node; |
320ae51f | 1940 | hctxs[i]->queue_num = i; |
868f2f0b KB |
1941 | |
1942 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
1943 | free_cpumask_var(hctxs[i]->cpumask); | |
1944 | kfree(hctxs[i]); | |
1945 | hctxs[i] = NULL; | |
1946 | break; | |
1947 | } | |
1948 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 1949 | } |
868f2f0b KB |
1950 | for (j = i; j < q->nr_hw_queues; j++) { |
1951 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
1952 | ||
1953 | if (hctx) { | |
1954 | if (hctx->tags) { | |
1955 | blk_mq_free_rq_map(set, hctx->tags, j); | |
1956 | set->tags[j] = NULL; | |
1957 | } | |
1958 | blk_mq_exit_hctx(q, set, hctx, j); | |
1959 | free_cpumask_var(hctx->cpumask); | |
1960 | kobject_put(&hctx->kobj); | |
1961 | kfree(hctx->ctxs); | |
1962 | kfree(hctx); | |
1963 | hctxs[j] = NULL; | |
1964 | ||
1965 | } | |
1966 | } | |
1967 | q->nr_hw_queues = i; | |
1968 | blk_mq_sysfs_register(q); | |
1969 | } | |
1970 | ||
1971 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
1972 | struct request_queue *q) | |
1973 | { | |
66841672 ML |
1974 | /* mark the queue as mq asap */ |
1975 | q->mq_ops = set->ops; | |
1976 | ||
868f2f0b KB |
1977 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
1978 | if (!q->queue_ctx) | |
c7de5726 | 1979 | goto err_exit; |
868f2f0b KB |
1980 | |
1981 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
1982 | GFP_KERNEL, set->numa_node); | |
1983 | if (!q->queue_hw_ctx) | |
1984 | goto err_percpu; | |
1985 | ||
bdd17e75 | 1986 | q->mq_map = set->mq_map; |
868f2f0b KB |
1987 | |
1988 | blk_mq_realloc_hw_ctxs(set, q); | |
1989 | if (!q->nr_hw_queues) | |
1990 | goto err_hctxs; | |
320ae51f | 1991 | |
287922eb | 1992 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 1993 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
1994 | |
1995 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 1996 | |
94eddfbe | 1997 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 1998 | |
05f1dd53 JA |
1999 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2000 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2001 | ||
1be036e9 CH |
2002 | q->sg_reserved_size = INT_MAX; |
2003 | ||
2849450a | 2004 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2005 | INIT_LIST_HEAD(&q->requeue_list); |
2006 | spin_lock_init(&q->requeue_lock); | |
2007 | ||
07068d5b JA |
2008 | if (q->nr_hw_queues > 1) |
2009 | blk_queue_make_request(q, blk_mq_make_request); | |
2010 | else | |
2011 | blk_queue_make_request(q, blk_sq_make_request); | |
2012 | ||
eba71768 JA |
2013 | /* |
2014 | * Do this after blk_queue_make_request() overrides it... | |
2015 | */ | |
2016 | q->nr_requests = set->queue_depth; | |
2017 | ||
24d2f903 CH |
2018 | if (set->ops->complete) |
2019 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2020 | |
24d2f903 | 2021 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2022 | |
5778322e | 2023 | get_online_cpus(); |
320ae51f | 2024 | mutex_lock(&all_q_mutex); |
320ae51f | 2025 | |
4593fdbe | 2026 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2027 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2028 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2029 | |
4593fdbe | 2030 | mutex_unlock(&all_q_mutex); |
5778322e | 2031 | put_online_cpus(); |
4593fdbe | 2032 | |
320ae51f | 2033 | return q; |
18741986 | 2034 | |
320ae51f | 2035 | err_hctxs: |
868f2f0b | 2036 | kfree(q->queue_hw_ctx); |
320ae51f | 2037 | err_percpu: |
868f2f0b | 2038 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2039 | err_exit: |
2040 | q->mq_ops = NULL; | |
320ae51f JA |
2041 | return ERR_PTR(-ENOMEM); |
2042 | } | |
b62c21b7 | 2043 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2044 | |
2045 | void blk_mq_free_queue(struct request_queue *q) | |
2046 | { | |
624dbe47 | 2047 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2048 | |
0e626368 AM |
2049 | mutex_lock(&all_q_mutex); |
2050 | list_del_init(&q->all_q_node); | |
2051 | mutex_unlock(&all_q_mutex); | |
2052 | ||
0d2602ca JA |
2053 | blk_mq_del_queue_tag_set(q); |
2054 | ||
624dbe47 ML |
2055 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2056 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2057 | } |
320ae51f JA |
2058 | |
2059 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2060 | static void blk_mq_queue_reinit(struct request_queue *q, |
2061 | const struct cpumask *online_mask) | |
320ae51f | 2062 | { |
4ecd4fef | 2063 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2064 | |
67aec14c JA |
2065 | blk_mq_sysfs_unregister(q); |
2066 | ||
320ae51f JA |
2067 | /* |
2068 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2069 | * we should change hctx numa_node according to new topology (this | |
2070 | * involves free and re-allocate memory, worthy doing?) | |
2071 | */ | |
2072 | ||
5778322e | 2073 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2074 | |
67aec14c | 2075 | blk_mq_sysfs_register(q); |
320ae51f JA |
2076 | } |
2077 | ||
65d5291e SAS |
2078 | /* |
2079 | * New online cpumask which is going to be set in this hotplug event. | |
2080 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2081 | * one-by-one and dynamically allocating this could result in a failure. | |
2082 | */ | |
2083 | static struct cpumask cpuhp_online_new; | |
2084 | ||
2085 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2086 | { |
2087 | struct request_queue *q; | |
320ae51f JA |
2088 | |
2089 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2090 | /* |
2091 | * We need to freeze and reinit all existing queues. Freezing | |
2092 | * involves synchronous wait for an RCU grace period and doing it | |
2093 | * one by one may take a long time. Start freezing all queues in | |
2094 | * one swoop and then wait for the completions so that freezing can | |
2095 | * take place in parallel. | |
2096 | */ | |
2097 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2098 | blk_mq_freeze_queue_start(q); | |
f054b56c | 2099 | list_for_each_entry(q, &all_q_list, all_q_node) { |
f3af020b TH |
2100 | blk_mq_freeze_queue_wait(q); |
2101 | ||
f054b56c ML |
2102 | /* |
2103 | * timeout handler can't touch hw queue during the | |
2104 | * reinitialization | |
2105 | */ | |
2106 | del_timer_sync(&q->timeout); | |
2107 | } | |
2108 | ||
320ae51f | 2109 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2110 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2111 | |
2112 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2113 | blk_mq_unfreeze_queue(q); | |
2114 | ||
320ae51f | 2115 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2116 | } |
2117 | ||
2118 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2119 | { | |
97a32864 | 2120 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2121 | blk_mq_queue_reinit_work(); |
2122 | return 0; | |
2123 | } | |
2124 | ||
2125 | /* | |
2126 | * Before hotadded cpu starts handling requests, new mappings must be | |
2127 | * established. Otherwise, these requests in hw queue might never be | |
2128 | * dispatched. | |
2129 | * | |
2130 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2131 | * for CPU0, and ctx1 for CPU1). | |
2132 | * | |
2133 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2134 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2135 | * | |
2136 | * And then while running hw queue, flush_busy_ctxs() finds bit0 is set in | |
2137 | * pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2138 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list | |
2139 | * is ignored. | |
2140 | */ | |
2141 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2142 | { | |
2143 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2144 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2145 | blk_mq_queue_reinit_work(); | |
2146 | return 0; | |
320ae51f JA |
2147 | } |
2148 | ||
a5164405 JA |
2149 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2150 | { | |
2151 | int i; | |
2152 | ||
2153 | for (i = 0; i < set->nr_hw_queues; i++) { | |
2154 | set->tags[i] = blk_mq_init_rq_map(set, i); | |
2155 | if (!set->tags[i]) | |
2156 | goto out_unwind; | |
2157 | } | |
2158 | ||
2159 | return 0; | |
2160 | ||
2161 | out_unwind: | |
2162 | while (--i >= 0) | |
2163 | blk_mq_free_rq_map(set, set->tags[i], i); | |
2164 | ||
a5164405 JA |
2165 | return -ENOMEM; |
2166 | } | |
2167 | ||
2168 | /* | |
2169 | * Allocate the request maps associated with this tag_set. Note that this | |
2170 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2171 | * will be updated to reflect the allocated depth. | |
2172 | */ | |
2173 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2174 | { | |
2175 | unsigned int depth; | |
2176 | int err; | |
2177 | ||
2178 | depth = set->queue_depth; | |
2179 | do { | |
2180 | err = __blk_mq_alloc_rq_maps(set); | |
2181 | if (!err) | |
2182 | break; | |
2183 | ||
2184 | set->queue_depth >>= 1; | |
2185 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2186 | err = -ENOMEM; | |
2187 | break; | |
2188 | } | |
2189 | } while (set->queue_depth); | |
2190 | ||
2191 | if (!set->queue_depth || err) { | |
2192 | pr_err("blk-mq: failed to allocate request map\n"); | |
2193 | return -ENOMEM; | |
2194 | } | |
2195 | ||
2196 | if (depth != set->queue_depth) | |
2197 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2198 | depth, set->queue_depth); | |
2199 | ||
2200 | return 0; | |
2201 | } | |
2202 | ||
a4391c64 JA |
2203 | /* |
2204 | * Alloc a tag set to be associated with one or more request queues. | |
2205 | * May fail with EINVAL for various error conditions. May adjust the | |
2206 | * requested depth down, if if it too large. In that case, the set | |
2207 | * value will be stored in set->queue_depth. | |
2208 | */ | |
24d2f903 CH |
2209 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2210 | { | |
da695ba2 CH |
2211 | int ret; |
2212 | ||
205fb5f5 BVA |
2213 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2214 | ||
24d2f903 CH |
2215 | if (!set->nr_hw_queues) |
2216 | return -EINVAL; | |
a4391c64 | 2217 | if (!set->queue_depth) |
24d2f903 CH |
2218 | return -EINVAL; |
2219 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2220 | return -EINVAL; | |
2221 | ||
7d7e0f90 | 2222 | if (!set->ops->queue_rq) |
24d2f903 CH |
2223 | return -EINVAL; |
2224 | ||
a4391c64 JA |
2225 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2226 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2227 | BLK_MQ_MAX_DEPTH); | |
2228 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2229 | } | |
24d2f903 | 2230 | |
6637fadf SL |
2231 | /* |
2232 | * If a crashdump is active, then we are potentially in a very | |
2233 | * memory constrained environment. Limit us to 1 queue and | |
2234 | * 64 tags to prevent using too much memory. | |
2235 | */ | |
2236 | if (is_kdump_kernel()) { | |
2237 | set->nr_hw_queues = 1; | |
2238 | set->queue_depth = min(64U, set->queue_depth); | |
2239 | } | |
868f2f0b KB |
2240 | /* |
2241 | * There is no use for more h/w queues than cpus. | |
2242 | */ | |
2243 | if (set->nr_hw_queues > nr_cpu_ids) | |
2244 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2245 | |
868f2f0b | 2246 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2247 | GFP_KERNEL, set->numa_node); |
2248 | if (!set->tags) | |
a5164405 | 2249 | return -ENOMEM; |
24d2f903 | 2250 | |
da695ba2 CH |
2251 | ret = -ENOMEM; |
2252 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2253 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2254 | if (!set->mq_map) |
2255 | goto out_free_tags; | |
2256 | ||
da695ba2 CH |
2257 | if (set->ops->map_queues) |
2258 | ret = set->ops->map_queues(set); | |
2259 | else | |
2260 | ret = blk_mq_map_queues(set); | |
2261 | if (ret) | |
2262 | goto out_free_mq_map; | |
2263 | ||
2264 | ret = blk_mq_alloc_rq_maps(set); | |
2265 | if (ret) | |
bdd17e75 | 2266 | goto out_free_mq_map; |
24d2f903 | 2267 | |
0d2602ca JA |
2268 | mutex_init(&set->tag_list_lock); |
2269 | INIT_LIST_HEAD(&set->tag_list); | |
2270 | ||
24d2f903 | 2271 | return 0; |
bdd17e75 CH |
2272 | |
2273 | out_free_mq_map: | |
2274 | kfree(set->mq_map); | |
2275 | set->mq_map = NULL; | |
2276 | out_free_tags: | |
5676e7b6 RE |
2277 | kfree(set->tags); |
2278 | set->tags = NULL; | |
da695ba2 | 2279 | return ret; |
24d2f903 CH |
2280 | } |
2281 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2282 | ||
2283 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2284 | { | |
2285 | int i; | |
2286 | ||
868f2f0b | 2287 | for (i = 0; i < nr_cpu_ids; i++) { |
f42d79ab | 2288 | if (set->tags[i]) |
484b4061 JA |
2289 | blk_mq_free_rq_map(set, set->tags[i], i); |
2290 | } | |
2291 | ||
bdd17e75 CH |
2292 | kfree(set->mq_map); |
2293 | set->mq_map = NULL; | |
2294 | ||
981bd189 | 2295 | kfree(set->tags); |
5676e7b6 | 2296 | set->tags = NULL; |
24d2f903 CH |
2297 | } |
2298 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2299 | ||
e3a2b3f9 JA |
2300 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2301 | { | |
2302 | struct blk_mq_tag_set *set = q->tag_set; | |
2303 | struct blk_mq_hw_ctx *hctx; | |
2304 | int i, ret; | |
2305 | ||
2306 | if (!set || nr > set->queue_depth) | |
2307 | return -EINVAL; | |
2308 | ||
2309 | ret = 0; | |
2310 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2311 | if (!hctx->tags) |
2312 | continue; | |
e3a2b3f9 JA |
2313 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2314 | if (ret) | |
2315 | break; | |
2316 | } | |
2317 | ||
2318 | if (!ret) | |
2319 | q->nr_requests = nr; | |
2320 | ||
2321 | return ret; | |
2322 | } | |
2323 | ||
868f2f0b KB |
2324 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2325 | { | |
2326 | struct request_queue *q; | |
2327 | ||
2328 | if (nr_hw_queues > nr_cpu_ids) | |
2329 | nr_hw_queues = nr_cpu_ids; | |
2330 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2331 | return; | |
2332 | ||
2333 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2334 | blk_mq_freeze_queue(q); | |
2335 | ||
2336 | set->nr_hw_queues = nr_hw_queues; | |
2337 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2338 | blk_mq_realloc_hw_ctxs(set, q); | |
2339 | ||
2340 | if (q->nr_hw_queues > 1) | |
2341 | blk_queue_make_request(q, blk_mq_make_request); | |
2342 | else | |
2343 | blk_queue_make_request(q, blk_sq_make_request); | |
2344 | ||
2345 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2346 | } | |
2347 | ||
2348 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2349 | blk_mq_unfreeze_queue(q); | |
2350 | } | |
2351 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2352 | ||
676141e4 JA |
2353 | void blk_mq_disable_hotplug(void) |
2354 | { | |
2355 | mutex_lock(&all_q_mutex); | |
2356 | } | |
2357 | ||
2358 | void blk_mq_enable_hotplug(void) | |
2359 | { | |
2360 | mutex_unlock(&all_q_mutex); | |
2361 | } | |
2362 | ||
320ae51f JA |
2363 | static int __init blk_mq_init(void) |
2364 | { | |
9467f859 TG |
2365 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2366 | blk_mq_hctx_notify_dead); | |
320ae51f | 2367 | |
65d5291e SAS |
2368 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2369 | blk_mq_queue_reinit_prepare, | |
2370 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2371 | return 0; |
2372 | } | |
2373 | subsys_initcall(blk_mq_init); |