]>
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" | |
cf43e6be | 33 | #include "blk-stat.h" |
87760e5e | 34 | #include "blk-wbt.h" |
320ae51f JA |
35 | |
36 | static DEFINE_MUTEX(all_q_mutex); | |
37 | static LIST_HEAD(all_q_list); | |
38 | ||
320ae51f JA |
39 | /* |
40 | * Check if any of the ctx's have pending work in this hardware queue | |
41 | */ | |
42 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
43 | { | |
88459642 | 44 | return sbitmap_any_bit_set(&hctx->ctx_map); |
1429d7c9 JA |
45 | } |
46 | ||
320ae51f JA |
47 | /* |
48 | * Mark this ctx as having pending work in this hardware queue | |
49 | */ | |
50 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
51 | struct blk_mq_ctx *ctx) | |
52 | { | |
88459642 OS |
53 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
54 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
55 | } |
56 | ||
57 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
58 | struct blk_mq_ctx *ctx) | |
59 | { | |
88459642 | 60 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
61 | } |
62 | ||
b4c6a028 | 63 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 64 | { |
4ecd4fef | 65 | int freeze_depth; |
cddd5d17 | 66 | |
4ecd4fef CH |
67 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
68 | if (freeze_depth == 1) { | |
3ef28e83 | 69 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 70 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 71 | } |
f3af020b | 72 | } |
b4c6a028 | 73 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b TH |
74 | |
75 | static void blk_mq_freeze_queue_wait(struct request_queue *q) | |
76 | { | |
3ef28e83 | 77 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 ML |
78 | } |
79 | ||
f3af020b TH |
80 | /* |
81 | * Guarantee no request is in use, so we can change any data structure of | |
82 | * the queue afterward. | |
83 | */ | |
3ef28e83 | 84 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 85 | { |
3ef28e83 DW |
86 | /* |
87 | * In the !blk_mq case we are only calling this to kill the | |
88 | * q_usage_counter, otherwise this increases the freeze depth | |
89 | * and waits for it to return to zero. For this reason there is | |
90 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
91 | * exported to drivers as the only user for unfreeze is blk_mq. | |
92 | */ | |
f3af020b TH |
93 | blk_mq_freeze_queue_start(q); |
94 | blk_mq_freeze_queue_wait(q); | |
95 | } | |
3ef28e83 DW |
96 | |
97 | void blk_mq_freeze_queue(struct request_queue *q) | |
98 | { | |
99 | /* | |
100 | * ...just an alias to keep freeze and unfreeze actions balanced | |
101 | * in the blk_mq_* namespace | |
102 | */ | |
103 | blk_freeze_queue(q); | |
104 | } | |
c761d96b | 105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 106 | |
b4c6a028 | 107 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 108 | { |
4ecd4fef | 109 | int freeze_depth; |
320ae51f | 110 | |
4ecd4fef CH |
111 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
112 | WARN_ON_ONCE(freeze_depth < 0); | |
113 | if (!freeze_depth) { | |
3ef28e83 | 114 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 115 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 116 | } |
320ae51f | 117 | } |
b4c6a028 | 118 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 119 | |
6a83e74d BVA |
120 | /** |
121 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
122 | * @q: request queue. | |
123 | * | |
124 | * Note: this function does not prevent that the struct request end_io() | |
125 | * callback function is invoked. Additionally, it is not prevented that | |
126 | * new queue_rq() calls occur unless the queue has been stopped first. | |
127 | */ | |
128 | void blk_mq_quiesce_queue(struct request_queue *q) | |
129 | { | |
130 | struct blk_mq_hw_ctx *hctx; | |
131 | unsigned int i; | |
132 | bool rcu = false; | |
133 | ||
134 | blk_mq_stop_hw_queues(q); | |
135 | ||
136 | queue_for_each_hw_ctx(q, hctx, i) { | |
137 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
138 | synchronize_srcu(&hctx->queue_rq_srcu); | |
139 | else | |
140 | rcu = true; | |
141 | } | |
142 | if (rcu) | |
143 | synchronize_rcu(); | |
144 | } | |
145 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
146 | ||
aed3ea94 JA |
147 | void blk_mq_wake_waiters(struct request_queue *q) |
148 | { | |
149 | struct blk_mq_hw_ctx *hctx; | |
150 | unsigned int i; | |
151 | ||
152 | queue_for_each_hw_ctx(q, hctx, i) | |
153 | if (blk_mq_hw_queue_mapped(hctx)) | |
154 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
155 | |
156 | /* | |
157 | * If we are called because the queue has now been marked as | |
158 | * dying, we need to ensure that processes currently waiting on | |
159 | * the queue are notified as well. | |
160 | */ | |
161 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
162 | } |
163 | ||
320ae51f JA |
164 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
165 | { | |
166 | return blk_mq_has_free_tags(hctx->tags); | |
167 | } | |
168 | EXPORT_SYMBOL(blk_mq_can_queue); | |
169 | ||
2c3ad667 JA |
170 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
171 | struct request *rq, unsigned int op) | |
320ae51f | 172 | { |
af76e555 CH |
173 | INIT_LIST_HEAD(&rq->queuelist); |
174 | /* csd/requeue_work/fifo_time is initialized before use */ | |
175 | rq->q = q; | |
320ae51f | 176 | rq->mq_ctx = ctx; |
ef295ecf | 177 | rq->cmd_flags = op; |
e8064021 CH |
178 | if (blk_queue_io_stat(q)) |
179 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
180 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
181 | rq->cpu = -1; | |
af76e555 CH |
182 | INIT_HLIST_NODE(&rq->hash); |
183 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
184 | rq->rq_disk = NULL; |
185 | rq->part = NULL; | |
3ee32372 | 186 | rq->start_time = jiffies; |
af76e555 CH |
187 | #ifdef CONFIG_BLK_CGROUP |
188 | rq->rl = NULL; | |
0fec08b4 | 189 | set_start_time_ns(rq); |
af76e555 CH |
190 | rq->io_start_time_ns = 0; |
191 | #endif | |
192 | rq->nr_phys_segments = 0; | |
193 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
194 | rq->nr_integrity_segments = 0; | |
195 | #endif | |
af76e555 CH |
196 | rq->special = NULL; |
197 | /* tag was already set */ | |
198 | rq->errors = 0; | |
af76e555 | 199 | |
6f4a1626 TB |
200 | rq->cmd = rq->__cmd; |
201 | ||
af76e555 CH |
202 | rq->extra_len = 0; |
203 | rq->sense_len = 0; | |
204 | rq->resid_len = 0; | |
205 | rq->sense = NULL; | |
206 | ||
af76e555 | 207 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
208 | rq->timeout = 0; |
209 | ||
af76e555 CH |
210 | rq->end_io = NULL; |
211 | rq->end_io_data = NULL; | |
212 | rq->next_rq = NULL; | |
213 | ||
ef295ecf | 214 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 215 | } |
2c3ad667 | 216 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 217 | |
2c3ad667 JA |
218 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
219 | unsigned int op) | |
5dee8577 CH |
220 | { |
221 | struct request *rq; | |
222 | unsigned int tag; | |
223 | ||
cb96a42c | 224 | tag = blk_mq_get_tag(data); |
5dee8577 | 225 | if (tag != BLK_MQ_TAG_FAIL) { |
2af8cbe3 | 226 | rq = data->hctx->tags->static_rqs[tag]; |
5dee8577 | 227 | |
cb96a42c | 228 | if (blk_mq_tag_busy(data->hctx)) { |
e8064021 | 229 | rq->rq_flags = RQF_MQ_INFLIGHT; |
cb96a42c | 230 | atomic_inc(&data->hctx->nr_active); |
5dee8577 CH |
231 | } |
232 | ||
233 | rq->tag = tag; | |
2af8cbe3 | 234 | data->hctx->tags->rqs[tag] = rq; |
ef295ecf | 235 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
236 | return rq; |
237 | } | |
238 | ||
239 | return NULL; | |
240 | } | |
2c3ad667 | 241 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 242 | |
6f3b0e8b CH |
243 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
244 | unsigned int flags) | |
320ae51f | 245 | { |
d852564f CH |
246 | struct blk_mq_ctx *ctx; |
247 | struct blk_mq_hw_ctx *hctx; | |
320ae51f | 248 | struct request *rq; |
cb96a42c | 249 | struct blk_mq_alloc_data alloc_data; |
a492f075 | 250 | int ret; |
320ae51f | 251 | |
6f3b0e8b | 252 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
253 | if (ret) |
254 | return ERR_PTR(ret); | |
320ae51f | 255 | |
d852564f | 256 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 257 | hctx = blk_mq_map_queue(q, ctx->cpu); |
6f3b0e8b | 258 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); |
ef295ecf | 259 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
d852564f | 260 | blk_mq_put_ctx(ctx); |
841bac2c | 261 | |
c76541a9 | 262 | if (!rq) { |
3ef28e83 | 263 | blk_queue_exit(q); |
a492f075 | 264 | return ERR_PTR(-EWOULDBLOCK); |
c76541a9 | 265 | } |
0c4de0f3 CH |
266 | |
267 | rq->__data_len = 0; | |
268 | rq->__sector = (sector_t) -1; | |
269 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
270 | return rq; |
271 | } | |
4bb659b1 | 272 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 273 | |
1f5bd336 ML |
274 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
275 | unsigned int flags, unsigned int hctx_idx) | |
276 | { | |
277 | struct blk_mq_hw_ctx *hctx; | |
278 | struct blk_mq_ctx *ctx; | |
279 | struct request *rq; | |
280 | struct blk_mq_alloc_data alloc_data; | |
281 | int ret; | |
282 | ||
283 | /* | |
284 | * If the tag allocator sleeps we could get an allocation for a | |
285 | * different hardware context. No need to complicate the low level | |
286 | * allocator for this for the rare use case of a command tied to | |
287 | * a specific queue. | |
288 | */ | |
289 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
290 | return ERR_PTR(-EINVAL); | |
291 | ||
292 | if (hctx_idx >= q->nr_hw_queues) | |
293 | return ERR_PTR(-EIO); | |
294 | ||
295 | ret = blk_queue_enter(q, true); | |
296 | if (ret) | |
297 | return ERR_PTR(ret); | |
298 | ||
c8712c6a CH |
299 | /* |
300 | * Check if the hardware context is actually mapped to anything. | |
301 | * If not tell the caller that it should skip this queue. | |
302 | */ | |
1f5bd336 | 303 | hctx = q->queue_hw_ctx[hctx_idx]; |
c8712c6a CH |
304 | if (!blk_mq_hw_queue_mapped(hctx)) { |
305 | ret = -EXDEV; | |
306 | goto out_queue_exit; | |
307 | } | |
1f5bd336 ML |
308 | ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask)); |
309 | ||
310 | blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx); | |
ef295ecf | 311 | rq = __blk_mq_alloc_request(&alloc_data, rw); |
1f5bd336 | 312 | if (!rq) { |
c8712c6a CH |
313 | ret = -EWOULDBLOCK; |
314 | goto out_queue_exit; | |
1f5bd336 ML |
315 | } |
316 | ||
317 | return rq; | |
c8712c6a CH |
318 | |
319 | out_queue_exit: | |
320 | blk_queue_exit(q); | |
321 | return ERR_PTR(ret); | |
1f5bd336 ML |
322 | } |
323 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
324 | ||
2c3ad667 JA |
325 | void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
326 | struct request *rq) | |
320ae51f JA |
327 | { |
328 | const int tag = rq->tag; | |
329 | struct request_queue *q = rq->q; | |
330 | ||
e8064021 | 331 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 332 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
333 | |
334 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 335 | rq->rq_flags = 0; |
0d2602ca | 336 | |
af76e555 | 337 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 338 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
4941115b | 339 | blk_mq_put_tag(hctx, hctx->tags, ctx, tag); |
3ef28e83 | 340 | blk_queue_exit(q); |
320ae51f JA |
341 | } |
342 | ||
16a3c2a7 JA |
343 | static void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, |
344 | struct request *rq) | |
320ae51f JA |
345 | { |
346 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
347 | |
348 | ctx->rq_completed[rq_is_sync(rq)]++; | |
320ae51f | 349 | __blk_mq_free_request(hctx, ctx, rq); |
7c7f2f2b | 350 | } |
7c7f2f2b JA |
351 | |
352 | void blk_mq_free_request(struct request *rq) | |
353 | { | |
7d7e0f90 | 354 | blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); |
320ae51f | 355 | } |
1a3b595a | 356 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 357 | |
c8a446ad | 358 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 359 | { |
0d11e6ac ML |
360 | blk_account_io_done(rq); |
361 | ||
91b63639 | 362 | if (rq->end_io) { |
87760e5e | 363 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 364 | rq->end_io(rq, error); |
91b63639 CH |
365 | } else { |
366 | if (unlikely(blk_bidi_rq(rq))) | |
367 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 368 | blk_mq_free_request(rq); |
91b63639 | 369 | } |
320ae51f | 370 | } |
c8a446ad | 371 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 372 | |
c8a446ad | 373 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
374 | { |
375 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
376 | BUG(); | |
c8a446ad | 377 | __blk_mq_end_request(rq, error); |
63151a44 | 378 | } |
c8a446ad | 379 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 380 | |
30a91cb4 | 381 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 382 | { |
3d6efbf6 | 383 | struct request *rq = data; |
320ae51f | 384 | |
30a91cb4 | 385 | rq->q->softirq_done_fn(rq); |
320ae51f | 386 | } |
320ae51f | 387 | |
ed851860 | 388 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
389 | { |
390 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 391 | bool shared = false; |
320ae51f JA |
392 | int cpu; |
393 | ||
38535201 | 394 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
395 | rq->q->softirq_done_fn(rq); |
396 | return; | |
397 | } | |
320ae51f JA |
398 | |
399 | cpu = get_cpu(); | |
38535201 CH |
400 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
401 | shared = cpus_share_cache(cpu, ctx->cpu); | |
402 | ||
403 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 404 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
405 | rq->csd.info = rq; |
406 | rq->csd.flags = 0; | |
c46fff2a | 407 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 408 | } else { |
30a91cb4 | 409 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 410 | } |
320ae51f JA |
411 | put_cpu(); |
412 | } | |
30a91cb4 | 413 | |
cf43e6be JA |
414 | static void blk_mq_stat_add(struct request *rq) |
415 | { | |
416 | if (rq->rq_flags & RQF_STATS) { | |
417 | /* | |
418 | * We could rq->mq_ctx here, but there's less of a risk | |
419 | * of races if we have the completion event add the stats | |
420 | * to the local software queue. | |
421 | */ | |
422 | struct blk_mq_ctx *ctx; | |
423 | ||
424 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
425 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
426 | } | |
427 | } | |
428 | ||
1fa8cc52 | 429 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
430 | { |
431 | struct request_queue *q = rq->q; | |
432 | ||
cf43e6be JA |
433 | blk_mq_stat_add(rq); |
434 | ||
ed851860 | 435 | if (!q->softirq_done_fn) |
c8a446ad | 436 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
437 | else |
438 | blk_mq_ipi_complete_request(rq); | |
439 | } | |
440 | ||
30a91cb4 CH |
441 | /** |
442 | * blk_mq_complete_request - end I/O on a request | |
443 | * @rq: the request being processed | |
444 | * | |
445 | * Description: | |
446 | * Ends all I/O on a request. It does not handle partial completions. | |
447 | * The actual completion happens out-of-order, through a IPI handler. | |
448 | **/ | |
f4829a9b | 449 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 450 | { |
95f09684 JA |
451 | struct request_queue *q = rq->q; |
452 | ||
453 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 454 | return; |
f4829a9b CH |
455 | if (!blk_mark_rq_complete(rq)) { |
456 | rq->errors = error; | |
ed851860 | 457 | __blk_mq_complete_request(rq); |
f4829a9b | 458 | } |
30a91cb4 CH |
459 | } |
460 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 461 | |
973c0191 KB |
462 | int blk_mq_request_started(struct request *rq) |
463 | { | |
464 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
465 | } | |
466 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
467 | ||
e2490073 | 468 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
469 | { |
470 | struct request_queue *q = rq->q; | |
471 | ||
472 | trace_block_rq_issue(q, rq); | |
473 | ||
742ee69b | 474 | rq->resid_len = blk_rq_bytes(rq); |
91b63639 CH |
475 | if (unlikely(blk_bidi_rq(rq))) |
476 | rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq); | |
742ee69b | 477 | |
cf43e6be JA |
478 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
479 | blk_stat_set_issue_time(&rq->issue_stat); | |
480 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 481 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
482 | } |
483 | ||
2b8393b4 | 484 | blk_add_timer(rq); |
87ee7b11 | 485 | |
538b7534 JA |
486 | /* |
487 | * Ensure that ->deadline is visible before set the started | |
488 | * flag and clear the completed flag. | |
489 | */ | |
490 | smp_mb__before_atomic(); | |
491 | ||
87ee7b11 JA |
492 | /* |
493 | * Mark us as started and clear complete. Complete might have been | |
494 | * set if requeue raced with timeout, which then marked it as | |
495 | * complete. So be sure to clear complete again when we start | |
496 | * the request, otherwise we'll ignore the completion event. | |
497 | */ | |
4b570521 JA |
498 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
499 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
500 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
501 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
502 | |
503 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
504 | /* | |
505 | * Make sure space for the drain appears. We know we can do | |
506 | * this because max_hw_segments has been adjusted to be one | |
507 | * fewer than the device can handle. | |
508 | */ | |
509 | rq->nr_phys_segments++; | |
510 | } | |
320ae51f | 511 | } |
e2490073 | 512 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 513 | |
ed0791b2 | 514 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
515 | { |
516 | struct request_queue *q = rq->q; | |
517 | ||
518 | trace_block_rq_requeue(q, rq); | |
87760e5e | 519 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
49f5baa5 | 520 | |
e2490073 CH |
521 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
522 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
523 | rq->nr_phys_segments--; | |
524 | } | |
320ae51f JA |
525 | } |
526 | ||
2b053aca | 527 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 528 | { |
ed0791b2 | 529 | __blk_mq_requeue_request(rq); |
ed0791b2 | 530 | |
ed0791b2 | 531 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 532 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
533 | } |
534 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
535 | ||
6fca6a61 CH |
536 | static void blk_mq_requeue_work(struct work_struct *work) |
537 | { | |
538 | struct request_queue *q = | |
2849450a | 539 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
540 | LIST_HEAD(rq_list); |
541 | struct request *rq, *next; | |
542 | unsigned long flags; | |
543 | ||
544 | spin_lock_irqsave(&q->requeue_lock, flags); | |
545 | list_splice_init(&q->requeue_list, &rq_list); | |
546 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
547 | ||
548 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 549 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
550 | continue; |
551 | ||
e8064021 | 552 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 CH |
553 | list_del_init(&rq->queuelist); |
554 | blk_mq_insert_request(rq, true, false, false); | |
555 | } | |
556 | ||
557 | while (!list_empty(&rq_list)) { | |
558 | rq = list_entry(rq_list.next, struct request, queuelist); | |
559 | list_del_init(&rq->queuelist); | |
560 | blk_mq_insert_request(rq, false, false, false); | |
561 | } | |
562 | ||
52d7f1b5 | 563 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
564 | } |
565 | ||
2b053aca BVA |
566 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
567 | bool kick_requeue_list) | |
6fca6a61 CH |
568 | { |
569 | struct request_queue *q = rq->q; | |
570 | unsigned long flags; | |
571 | ||
572 | /* | |
573 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
574 | * request head insertation from the workqueue. | |
575 | */ | |
e8064021 | 576 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
577 | |
578 | spin_lock_irqsave(&q->requeue_lock, flags); | |
579 | if (at_head) { | |
e8064021 | 580 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
581 | list_add(&rq->queuelist, &q->requeue_list); |
582 | } else { | |
583 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
584 | } | |
585 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
586 | |
587 | if (kick_requeue_list) | |
588 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
589 | } |
590 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
591 | ||
592 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
593 | { | |
2849450a | 594 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
595 | } |
596 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
597 | ||
2849450a MS |
598 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
599 | unsigned long msecs) | |
600 | { | |
601 | kblockd_schedule_delayed_work(&q->requeue_work, | |
602 | msecs_to_jiffies(msecs)); | |
603 | } | |
604 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
605 | ||
1885b24d JA |
606 | void blk_mq_abort_requeue_list(struct request_queue *q) |
607 | { | |
608 | unsigned long flags; | |
609 | LIST_HEAD(rq_list); | |
610 | ||
611 | spin_lock_irqsave(&q->requeue_lock, flags); | |
612 | list_splice_init(&q->requeue_list, &rq_list); | |
613 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
614 | ||
615 | while (!list_empty(&rq_list)) { | |
616 | struct request *rq; | |
617 | ||
618 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
619 | list_del_init(&rq->queuelist); | |
620 | rq->errors = -EIO; | |
621 | blk_mq_end_request(rq, rq->errors); | |
622 | } | |
623 | } | |
624 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
625 | ||
0e62f51f JA |
626 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
627 | { | |
88c7b2b7 JA |
628 | if (tag < tags->nr_tags) { |
629 | prefetch(tags->rqs[tag]); | |
4ee86bab | 630 | return tags->rqs[tag]; |
88c7b2b7 | 631 | } |
4ee86bab HR |
632 | |
633 | return NULL; | |
24d2f903 CH |
634 | } |
635 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
636 | ||
320ae51f | 637 | struct blk_mq_timeout_data { |
46f92d42 CH |
638 | unsigned long next; |
639 | unsigned int next_set; | |
320ae51f JA |
640 | }; |
641 | ||
90415837 | 642 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 643 | { |
f8a5b122 | 644 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 645 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
646 | |
647 | /* | |
648 | * We know that complete is set at this point. If STARTED isn't set | |
649 | * anymore, then the request isn't active and the "timeout" should | |
650 | * just be ignored. This can happen due to the bitflag ordering. | |
651 | * Timeout first checks if STARTED is set, and if it is, assumes | |
652 | * the request is active. But if we race with completion, then | |
653 | * we both flags will get cleared. So check here again, and ignore | |
654 | * a timeout event with a request that isn't active. | |
655 | */ | |
46f92d42 CH |
656 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
657 | return; | |
87ee7b11 | 658 | |
46f92d42 | 659 | if (ops->timeout) |
0152fb6b | 660 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
661 | |
662 | switch (ret) { | |
663 | case BLK_EH_HANDLED: | |
664 | __blk_mq_complete_request(req); | |
665 | break; | |
666 | case BLK_EH_RESET_TIMER: | |
667 | blk_add_timer(req); | |
668 | blk_clear_rq_complete(req); | |
669 | break; | |
670 | case BLK_EH_NOT_HANDLED: | |
671 | break; | |
672 | default: | |
673 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
674 | break; | |
675 | } | |
87ee7b11 | 676 | } |
5b3f25fc | 677 | |
81481eb4 CH |
678 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
679 | struct request *rq, void *priv, bool reserved) | |
680 | { | |
681 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 682 | |
eb130dbf KB |
683 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
684 | /* | |
685 | * If a request wasn't started before the queue was | |
686 | * marked dying, kill it here or it'll go unnoticed. | |
687 | */ | |
a59e0f57 KB |
688 | if (unlikely(blk_queue_dying(rq->q))) { |
689 | rq->errors = -EIO; | |
690 | blk_mq_end_request(rq, rq->errors); | |
691 | } | |
46f92d42 | 692 | return; |
eb130dbf | 693 | } |
87ee7b11 | 694 | |
46f92d42 CH |
695 | if (time_after_eq(jiffies, rq->deadline)) { |
696 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 697 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
698 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
699 | data->next = rq->deadline; | |
700 | data->next_set = 1; | |
701 | } | |
87ee7b11 JA |
702 | } |
703 | ||
287922eb | 704 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 705 | { |
287922eb CH |
706 | struct request_queue *q = |
707 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
708 | struct blk_mq_timeout_data data = { |
709 | .next = 0, | |
710 | .next_set = 0, | |
711 | }; | |
81481eb4 | 712 | int i; |
320ae51f | 713 | |
71f79fb3 GKB |
714 | /* A deadlock might occur if a request is stuck requiring a |
715 | * timeout at the same time a queue freeze is waiting | |
716 | * completion, since the timeout code would not be able to | |
717 | * acquire the queue reference here. | |
718 | * | |
719 | * That's why we don't use blk_queue_enter here; instead, we use | |
720 | * percpu_ref_tryget directly, because we need to be able to | |
721 | * obtain a reference even in the short window between the queue | |
722 | * starting to freeze, by dropping the first reference in | |
723 | * blk_mq_freeze_queue_start, and the moment the last request is | |
724 | * consumed, marked by the instant q_usage_counter reaches | |
725 | * zero. | |
726 | */ | |
727 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
728 | return; |
729 | ||
0bf6cd5b | 730 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 731 | |
81481eb4 CH |
732 | if (data.next_set) { |
733 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
734 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 735 | } else { |
0bf6cd5b CH |
736 | struct blk_mq_hw_ctx *hctx; |
737 | ||
f054b56c ML |
738 | queue_for_each_hw_ctx(q, hctx, i) { |
739 | /* the hctx may be unmapped, so check it here */ | |
740 | if (blk_mq_hw_queue_mapped(hctx)) | |
741 | blk_mq_tag_idle(hctx); | |
742 | } | |
0d2602ca | 743 | } |
287922eb | 744 | blk_queue_exit(q); |
320ae51f JA |
745 | } |
746 | ||
747 | /* | |
748 | * Reverse check our software queue for entries that we could potentially | |
749 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
750 | * too much time checking for merges. | |
751 | */ | |
752 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
753 | struct blk_mq_ctx *ctx, struct bio *bio) | |
754 | { | |
755 | struct request *rq; | |
756 | int checked = 8; | |
757 | ||
758 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
759 | int el_ret; | |
760 | ||
761 | if (!checked--) | |
762 | break; | |
763 | ||
764 | if (!blk_rq_merge_ok(rq, bio)) | |
765 | continue; | |
766 | ||
767 | el_ret = blk_try_merge(rq, bio); | |
768 | if (el_ret == ELEVATOR_BACK_MERGE) { | |
769 | if (bio_attempt_back_merge(q, rq, bio)) { | |
770 | ctx->rq_merged++; | |
771 | return true; | |
772 | } | |
773 | break; | |
774 | } else if (el_ret == ELEVATOR_FRONT_MERGE) { | |
775 | if (bio_attempt_front_merge(q, rq, bio)) { | |
776 | ctx->rq_merged++; | |
777 | return true; | |
778 | } | |
779 | break; | |
780 | } | |
781 | } | |
782 | ||
783 | return false; | |
784 | } | |
785 | ||
88459642 OS |
786 | struct flush_busy_ctx_data { |
787 | struct blk_mq_hw_ctx *hctx; | |
788 | struct list_head *list; | |
789 | }; | |
790 | ||
791 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
792 | { | |
793 | struct flush_busy_ctx_data *flush_data = data; | |
794 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
795 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
796 | ||
797 | sbitmap_clear_bit(sb, bitnr); | |
798 | spin_lock(&ctx->lock); | |
799 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
800 | spin_unlock(&ctx->lock); | |
801 | return true; | |
802 | } | |
803 | ||
1429d7c9 JA |
804 | /* |
805 | * Process software queues that have been marked busy, splicing them | |
806 | * to the for-dispatch | |
807 | */ | |
2c3ad667 | 808 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 809 | { |
88459642 OS |
810 | struct flush_busy_ctx_data data = { |
811 | .hctx = hctx, | |
812 | .list = list, | |
813 | }; | |
1429d7c9 | 814 | |
88459642 | 815 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 816 | } |
2c3ad667 | 817 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 818 | |
703fd1c0 JA |
819 | static inline unsigned int queued_to_index(unsigned int queued) |
820 | { | |
821 | if (!queued) | |
822 | return 0; | |
1429d7c9 | 823 | |
703fd1c0 | 824 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
825 | } |
826 | ||
f04c3df3 | 827 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
828 | { |
829 | struct request_queue *q = hctx->queue; | |
320ae51f | 830 | struct request *rq; |
74c45052 JA |
831 | LIST_HEAD(driver_list); |
832 | struct list_head *dptr; | |
f04c3df3 | 833 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 834 | |
74c45052 JA |
835 | /* |
836 | * Start off with dptr being NULL, so we start the first request | |
837 | * immediately, even if we have more pending. | |
838 | */ | |
839 | dptr = NULL; | |
840 | ||
320ae51f JA |
841 | /* |
842 | * Now process all the entries, sending them to the driver. | |
843 | */ | |
1429d7c9 | 844 | queued = 0; |
f04c3df3 | 845 | while (!list_empty(list)) { |
74c45052 | 846 | struct blk_mq_queue_data bd; |
320ae51f | 847 | |
f04c3df3 | 848 | rq = list_first_entry(list, struct request, queuelist); |
320ae51f | 849 | list_del_init(&rq->queuelist); |
320ae51f | 850 | |
74c45052 JA |
851 | bd.rq = rq; |
852 | bd.list = dptr; | |
f04c3df3 | 853 | bd.last = list_empty(list); |
74c45052 JA |
854 | |
855 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
856 | switch (ret) { |
857 | case BLK_MQ_RQ_QUEUE_OK: | |
858 | queued++; | |
52b9c330 | 859 | break; |
320ae51f | 860 | case BLK_MQ_RQ_QUEUE_BUSY: |
f04c3df3 | 861 | list_add(&rq->queuelist, list); |
ed0791b2 | 862 | __blk_mq_requeue_request(rq); |
320ae51f JA |
863 | break; |
864 | default: | |
865 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 866 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 867 | rq->errors = -EIO; |
c8a446ad | 868 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
869 | break; |
870 | } | |
871 | ||
872 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
873 | break; | |
74c45052 JA |
874 | |
875 | /* | |
876 | * We've done the first request. If we have more than 1 | |
877 | * left in the list, set dptr to defer issue. | |
878 | */ | |
f04c3df3 | 879 | if (!dptr && list->next != list->prev) |
74c45052 | 880 | dptr = &driver_list; |
320ae51f JA |
881 | } |
882 | ||
703fd1c0 | 883 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
884 | |
885 | /* | |
886 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
887 | * that is where we will continue on next queue run. | |
888 | */ | |
f04c3df3 | 889 | if (!list_empty(list)) { |
320ae51f | 890 | spin_lock(&hctx->lock); |
f04c3df3 | 891 | list_splice(list, &hctx->dispatch); |
320ae51f | 892 | spin_unlock(&hctx->lock); |
f04c3df3 | 893 | |
9ba52e58 SL |
894 | /* |
895 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
896 | * it's possible the queue is stopped and restarted again | |
897 | * before this. Queue restart will dispatch requests. And since | |
898 | * requests in rq_list aren't added into hctx->dispatch yet, | |
899 | * the requests in rq_list might get lost. | |
900 | * | |
901 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
902 | **/ | |
903 | blk_mq_run_hw_queue(hctx, true); | |
320ae51f | 904 | } |
f04c3df3 JA |
905 | |
906 | return ret != BLK_MQ_RQ_QUEUE_BUSY; | |
907 | } | |
908 | ||
909 | /* | |
910 | * Run this hardware queue, pulling any software queues mapped to it in. | |
911 | * Note that this function currently has various problems around ordering | |
912 | * of IO. In particular, we'd like FIFO behaviour on handling existing | |
913 | * items on the hctx->dispatch list. Ignore that for now. | |
914 | */ | |
915 | static void blk_mq_process_rq_list(struct blk_mq_hw_ctx *hctx) | |
916 | { | |
917 | LIST_HEAD(rq_list); | |
918 | LIST_HEAD(driver_list); | |
919 | ||
920 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
921 | return; | |
922 | ||
923 | hctx->run++; | |
924 | ||
925 | /* | |
926 | * Touch any software queue that has pending entries. | |
927 | */ | |
2c3ad667 | 928 | blk_mq_flush_busy_ctxs(hctx, &rq_list); |
f04c3df3 JA |
929 | |
930 | /* | |
931 | * If we have previous entries on our dispatch list, grab them | |
932 | * and stuff them at the front for more fair dispatch. | |
933 | */ | |
934 | if (!list_empty_careful(&hctx->dispatch)) { | |
935 | spin_lock(&hctx->lock); | |
936 | if (!list_empty(&hctx->dispatch)) | |
937 | list_splice_init(&hctx->dispatch, &rq_list); | |
938 | spin_unlock(&hctx->lock); | |
939 | } | |
940 | ||
941 | blk_mq_dispatch_rq_list(hctx, &rq_list); | |
320ae51f JA |
942 | } |
943 | ||
6a83e74d BVA |
944 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
945 | { | |
946 | int srcu_idx; | |
947 | ||
948 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
949 | cpu_online(hctx->next_cpu)); | |
950 | ||
951 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
952 | rcu_read_lock(); | |
953 | blk_mq_process_rq_list(hctx); | |
954 | rcu_read_unlock(); | |
955 | } else { | |
956 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
957 | blk_mq_process_rq_list(hctx); | |
958 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
959 | } | |
960 | } | |
961 | ||
506e931f JA |
962 | /* |
963 | * It'd be great if the workqueue API had a way to pass | |
964 | * in a mask and had some smarts for more clever placement. | |
965 | * For now we just round-robin here, switching for every | |
966 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
967 | */ | |
968 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
969 | { | |
b657d7e6 CH |
970 | if (hctx->queue->nr_hw_queues == 1) |
971 | return WORK_CPU_UNBOUND; | |
506e931f JA |
972 | |
973 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 974 | int next_cpu; |
506e931f JA |
975 | |
976 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
977 | if (next_cpu >= nr_cpu_ids) | |
978 | next_cpu = cpumask_first(hctx->cpumask); | |
979 | ||
980 | hctx->next_cpu = next_cpu; | |
981 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
982 | } | |
983 | ||
b657d7e6 | 984 | return hctx->next_cpu; |
506e931f JA |
985 | } |
986 | ||
320ae51f JA |
987 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
988 | { | |
5d1b25c1 BVA |
989 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
990 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
991 | return; |
992 | ||
1b792f2f | 993 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
994 | int cpu = get_cpu(); |
995 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 996 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 997 | put_cpu(); |
398205b8 PB |
998 | return; |
999 | } | |
e4043dcf | 1000 | |
2a90d4aa | 1001 | put_cpu(); |
e4043dcf | 1002 | } |
398205b8 | 1003 | |
27489a3c | 1004 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1005 | } |
1006 | ||
b94ec296 | 1007 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1008 | { |
1009 | struct blk_mq_hw_ctx *hctx; | |
1010 | int i; | |
1011 | ||
1012 | queue_for_each_hw_ctx(q, hctx, i) { | |
1013 | if ((!blk_mq_hctx_has_pending(hctx) && | |
1014 | list_empty_careful(&hctx->dispatch)) || | |
5d1b25c1 | 1015 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1016 | continue; |
1017 | ||
b94ec296 | 1018 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1019 | } |
1020 | } | |
b94ec296 | 1021 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1022 | |
fd001443 BVA |
1023 | /** |
1024 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1025 | * @q: request queue. | |
1026 | * | |
1027 | * The caller is responsible for serializing this function against | |
1028 | * blk_mq_{start,stop}_hw_queue(). | |
1029 | */ | |
1030 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1031 | { | |
1032 | struct blk_mq_hw_ctx *hctx; | |
1033 | int i; | |
1034 | ||
1035 | queue_for_each_hw_ctx(q, hctx, i) | |
1036 | if (blk_mq_hctx_stopped(hctx)) | |
1037 | return true; | |
1038 | ||
1039 | return false; | |
1040 | } | |
1041 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1042 | ||
320ae51f JA |
1043 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1044 | { | |
27489a3c | 1045 | cancel_work(&hctx->run_work); |
70f4db63 | 1046 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1047 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1048 | } | |
1049 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1050 | ||
280d45f6 CH |
1051 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1052 | { | |
1053 | struct blk_mq_hw_ctx *hctx; | |
1054 | int i; | |
1055 | ||
1056 | queue_for_each_hw_ctx(q, hctx, i) | |
1057 | blk_mq_stop_hw_queue(hctx); | |
1058 | } | |
1059 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1060 | ||
320ae51f JA |
1061 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1062 | { | |
1063 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1064 | |
0ffbce80 | 1065 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1066 | } |
1067 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1068 | ||
2f268556 CH |
1069 | void blk_mq_start_hw_queues(struct request_queue *q) |
1070 | { | |
1071 | struct blk_mq_hw_ctx *hctx; | |
1072 | int i; | |
1073 | ||
1074 | queue_for_each_hw_ctx(q, hctx, i) | |
1075 | blk_mq_start_hw_queue(hctx); | |
1076 | } | |
1077 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1078 | ||
ae911c5e JA |
1079 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1080 | { | |
1081 | if (!blk_mq_hctx_stopped(hctx)) | |
1082 | return; | |
1083 | ||
1084 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1085 | blk_mq_run_hw_queue(hctx, async); | |
1086 | } | |
1087 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1088 | ||
1b4a3258 | 1089 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1090 | { |
1091 | struct blk_mq_hw_ctx *hctx; | |
1092 | int i; | |
1093 | ||
ae911c5e JA |
1094 | queue_for_each_hw_ctx(q, hctx, i) |
1095 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1096 | } |
1097 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1098 | ||
70f4db63 | 1099 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1100 | { |
1101 | struct blk_mq_hw_ctx *hctx; | |
1102 | ||
27489a3c | 1103 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1104 | |
320ae51f JA |
1105 | __blk_mq_run_hw_queue(hctx); |
1106 | } | |
1107 | ||
70f4db63 CH |
1108 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1109 | { | |
1110 | struct blk_mq_hw_ctx *hctx; | |
1111 | ||
1112 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1113 | ||
1114 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1115 | __blk_mq_run_hw_queue(hctx); | |
1116 | } | |
1117 | ||
1118 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1119 | { | |
19c66e59 ML |
1120 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1121 | return; | |
70f4db63 | 1122 | |
b657d7e6 CH |
1123 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1124 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1125 | } |
1126 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1127 | ||
cfd0c552 | 1128 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1129 | struct request *rq, |
1130 | bool at_head) | |
320ae51f | 1131 | { |
e57690fe JA |
1132 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1133 | ||
01b983c9 JA |
1134 | trace_block_rq_insert(hctx->queue, rq); |
1135 | ||
72a0a36e CH |
1136 | if (at_head) |
1137 | list_add(&rq->queuelist, &ctx->rq_list); | |
1138 | else | |
1139 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1140 | } |
4bb659b1 | 1141 | |
2c3ad667 JA |
1142 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1143 | bool at_head) | |
cfd0c552 ML |
1144 | { |
1145 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1146 | ||
e57690fe | 1147 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1148 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1149 | } |
1150 | ||
eeabc850 | 1151 | void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue, |
e57690fe | 1152 | bool async) |
320ae51f | 1153 | { |
e57690fe | 1154 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
eeabc850 | 1155 | struct request_queue *q = rq->q; |
7d7e0f90 | 1156 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1157 | |
a57a178a CH |
1158 | spin_lock(&ctx->lock); |
1159 | __blk_mq_insert_request(hctx, rq, at_head); | |
1160 | spin_unlock(&ctx->lock); | |
320ae51f | 1161 | |
320ae51f JA |
1162 | if (run_queue) |
1163 | blk_mq_run_hw_queue(hctx, async); | |
1164 | } | |
1165 | ||
1166 | static void blk_mq_insert_requests(struct request_queue *q, | |
1167 | struct blk_mq_ctx *ctx, | |
1168 | struct list_head *list, | |
1169 | int depth, | |
1170 | bool from_schedule) | |
1171 | ||
1172 | { | |
7d7e0f90 | 1173 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f JA |
1174 | |
1175 | trace_block_unplug(q, depth, !from_schedule); | |
1176 | ||
320ae51f JA |
1177 | /* |
1178 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1179 | * offline now | |
1180 | */ | |
1181 | spin_lock(&ctx->lock); | |
1182 | while (!list_empty(list)) { | |
1183 | struct request *rq; | |
1184 | ||
1185 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1186 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1187 | list_del_init(&rq->queuelist); |
e57690fe | 1188 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1189 | } |
cfd0c552 | 1190 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1191 | spin_unlock(&ctx->lock); |
1192 | ||
320ae51f JA |
1193 | blk_mq_run_hw_queue(hctx, from_schedule); |
1194 | } | |
1195 | ||
1196 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1197 | { | |
1198 | struct request *rqa = container_of(a, struct request, queuelist); | |
1199 | struct request *rqb = container_of(b, struct request, queuelist); | |
1200 | ||
1201 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1202 | (rqa->mq_ctx == rqb->mq_ctx && | |
1203 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1204 | } | |
1205 | ||
1206 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1207 | { | |
1208 | struct blk_mq_ctx *this_ctx; | |
1209 | struct request_queue *this_q; | |
1210 | struct request *rq; | |
1211 | LIST_HEAD(list); | |
1212 | LIST_HEAD(ctx_list); | |
1213 | unsigned int depth; | |
1214 | ||
1215 | list_splice_init(&plug->mq_list, &list); | |
1216 | ||
1217 | list_sort(NULL, &list, plug_ctx_cmp); | |
1218 | ||
1219 | this_q = NULL; | |
1220 | this_ctx = NULL; | |
1221 | depth = 0; | |
1222 | ||
1223 | while (!list_empty(&list)) { | |
1224 | rq = list_entry_rq(list.next); | |
1225 | list_del_init(&rq->queuelist); | |
1226 | BUG_ON(!rq->q); | |
1227 | if (rq->mq_ctx != this_ctx) { | |
1228 | if (this_ctx) { | |
1229 | blk_mq_insert_requests(this_q, this_ctx, | |
1230 | &ctx_list, depth, | |
1231 | from_schedule); | |
1232 | } | |
1233 | ||
1234 | this_ctx = rq->mq_ctx; | |
1235 | this_q = rq->q; | |
1236 | depth = 0; | |
1237 | } | |
1238 | ||
1239 | depth++; | |
1240 | list_add_tail(&rq->queuelist, &ctx_list); | |
1241 | } | |
1242 | ||
1243 | /* | |
1244 | * If 'this_ctx' is set, we know we have entries to complete | |
1245 | * on 'ctx_list'. Do those. | |
1246 | */ | |
1247 | if (this_ctx) { | |
1248 | blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth, | |
1249 | from_schedule); | |
1250 | } | |
1251 | } | |
1252 | ||
1253 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1254 | { | |
1255 | init_request_from_bio(rq, bio); | |
4b570521 | 1256 | |
6e85eaf3 | 1257 | blk_account_io_start(rq, true); |
320ae51f JA |
1258 | } |
1259 | ||
274a5843 JA |
1260 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1261 | { | |
1262 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1263 | !blk_queue_nomerges(hctx->queue); | |
1264 | } | |
1265 | ||
07068d5b JA |
1266 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1267 | struct blk_mq_ctx *ctx, | |
1268 | struct request *rq, struct bio *bio) | |
320ae51f | 1269 | { |
e18378a6 | 1270 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1271 | blk_mq_bio_to_request(rq, bio); |
1272 | spin_lock(&ctx->lock); | |
1273 | insert_rq: | |
1274 | __blk_mq_insert_request(hctx, rq, false); | |
1275 | spin_unlock(&ctx->lock); | |
1276 | return false; | |
1277 | } else { | |
274a5843 JA |
1278 | struct request_queue *q = hctx->queue; |
1279 | ||
07068d5b JA |
1280 | spin_lock(&ctx->lock); |
1281 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1282 | blk_mq_bio_to_request(rq, bio); | |
1283 | goto insert_rq; | |
1284 | } | |
320ae51f | 1285 | |
07068d5b JA |
1286 | spin_unlock(&ctx->lock); |
1287 | __blk_mq_free_request(hctx, ctx, rq); | |
1288 | return true; | |
14ec77f3 | 1289 | } |
07068d5b | 1290 | } |
14ec77f3 | 1291 | |
07068d5b JA |
1292 | static struct request *blk_mq_map_request(struct request_queue *q, |
1293 | struct bio *bio, | |
2552e3f8 | 1294 | struct blk_mq_alloc_data *data) |
07068d5b JA |
1295 | { |
1296 | struct blk_mq_hw_ctx *hctx; | |
1297 | struct blk_mq_ctx *ctx; | |
1298 | struct request *rq; | |
320ae51f | 1299 | |
3ef28e83 | 1300 | blk_queue_enter_live(q); |
320ae51f | 1301 | ctx = blk_mq_get_ctx(q); |
7d7e0f90 | 1302 | hctx = blk_mq_map_queue(q, ctx->cpu); |
320ae51f | 1303 | |
ef295ecf | 1304 | trace_block_getrq(q, bio, bio->bi_opf); |
2552e3f8 | 1305 | blk_mq_set_alloc_data(data, q, 0, ctx, hctx); |
ef295ecf | 1306 | rq = __blk_mq_alloc_request(data, bio->bi_opf); |
320ae51f | 1307 | |
7dd2fb68 | 1308 | data->hctx->queued++; |
07068d5b JA |
1309 | return rq; |
1310 | } | |
1311 | ||
fd2d3326 JA |
1312 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1313 | { | |
1314 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1315 | } | |
1316 | ||
066a4a73 | 1317 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f SL |
1318 | { |
1319 | int ret; | |
1320 | struct request_queue *q = rq->q; | |
066a4a73 | 1321 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu); |
f984df1f SL |
1322 | struct blk_mq_queue_data bd = { |
1323 | .rq = rq, | |
1324 | .list = NULL, | |
1325 | .last = 1 | |
1326 | }; | |
fd2d3326 | 1327 | blk_qc_t new_cookie = request_to_qc_t(hctx, rq); |
f984df1f | 1328 | |
2253efc8 BVA |
1329 | if (blk_mq_hctx_stopped(hctx)) |
1330 | goto insert; | |
1331 | ||
f984df1f SL |
1332 | /* |
1333 | * For OK queue, we are done. For error, kill it. Any other | |
1334 | * error (busy), just add it to our list as we previously | |
1335 | * would have done | |
1336 | */ | |
1337 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1338 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1339 | *cookie = new_cookie; | |
2253efc8 | 1340 | return; |
7b371636 | 1341 | } |
f984df1f | 1342 | |
7b371636 JA |
1343 | __blk_mq_requeue_request(rq); |
1344 | ||
1345 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1346 | *cookie = BLK_QC_T_NONE; | |
1347 | rq->errors = -EIO; | |
1348 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1349 | return; |
f984df1f | 1350 | } |
7b371636 | 1351 | |
2253efc8 BVA |
1352 | insert: |
1353 | blk_mq_insert_request(rq, false, true, true); | |
f984df1f SL |
1354 | } |
1355 | ||
07068d5b JA |
1356 | /* |
1357 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1358 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1359 | * hardware for SYNC IO. | |
1360 | */ | |
dece1635 | 1361 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1362 | { |
ef295ecf | 1363 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1364 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
2552e3f8 | 1365 | struct blk_mq_alloc_data data; |
07068d5b | 1366 | struct request *rq; |
6a83e74d | 1367 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1368 | struct blk_plug *plug; |
5b3f341f | 1369 | struct request *same_queue_rq = NULL; |
7b371636 | 1370 | blk_qc_t cookie; |
87760e5e | 1371 | unsigned int wb_acct; |
07068d5b JA |
1372 | |
1373 | blk_queue_bounce(q, &bio); | |
1374 | ||
1375 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1376 | bio_io_error(bio); |
dece1635 | 1377 | return BLK_QC_T_NONE; |
07068d5b JA |
1378 | } |
1379 | ||
54efd50b KO |
1380 | blk_queue_split(q, &bio, q->bio_split); |
1381 | ||
87c279e6 OS |
1382 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1383 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1384 | return BLK_QC_T_NONE; | |
f984df1f | 1385 | |
87760e5e JA |
1386 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1387 | ||
07068d5b | 1388 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1389 | if (unlikely(!rq)) { |
1390 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1391 | return BLK_QC_T_NONE; |
87760e5e JA |
1392 | } |
1393 | ||
1394 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1395 | |
fd2d3326 | 1396 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1397 | |
1398 | if (unlikely(is_flush_fua)) { | |
1399 | blk_mq_bio_to_request(rq, bio); | |
1400 | blk_insert_flush(rq); | |
1401 | goto run_queue; | |
1402 | } | |
1403 | ||
f984df1f | 1404 | plug = current->plug; |
e167dfb5 JA |
1405 | /* |
1406 | * If the driver supports defer issued based on 'last', then | |
1407 | * queue it up like normal since we can potentially save some | |
1408 | * CPU this way. | |
1409 | */ | |
f984df1f SL |
1410 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1411 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1412 | struct request *old_rq = NULL; | |
07068d5b JA |
1413 | |
1414 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1415 | |
1416 | /* | |
6a83e74d | 1417 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1418 | * Otherwise the existing request in the plug list will be |
1419 | * issued. So the plug list will have one request at most | |
07068d5b | 1420 | */ |
f984df1f | 1421 | if (plug) { |
5b3f341f SL |
1422 | /* |
1423 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1424 | * happens, same_queue_rq is invalid and plug list is |
1425 | * empty | |
1426 | */ | |
5b3f341f SL |
1427 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1428 | old_rq = same_queue_rq; | |
f984df1f | 1429 | list_del_init(&old_rq->queuelist); |
07068d5b | 1430 | } |
f984df1f SL |
1431 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1432 | } else /* is_sync */ | |
1433 | old_rq = rq; | |
1434 | blk_mq_put_ctx(data.ctx); | |
1435 | if (!old_rq) | |
7b371636 | 1436 | goto done; |
6a83e74d BVA |
1437 | |
1438 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1439 | rcu_read_lock(); | |
066a4a73 | 1440 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1441 | rcu_read_unlock(); |
1442 | } else { | |
1443 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1444 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1445 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1446 | } | |
7b371636 | 1447 | goto done; |
07068d5b JA |
1448 | } |
1449 | ||
1450 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1451 | /* | |
1452 | * For a SYNC request, send it to the hardware immediately. For | |
1453 | * an ASYNC request, just ensure that we run it later on. The | |
1454 | * latter allows for merging opportunities and more efficient | |
1455 | * dispatching. | |
1456 | */ | |
1457 | run_queue: | |
1458 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
1459 | } | |
07068d5b | 1460 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1461 | done: |
1462 | return cookie; | |
07068d5b JA |
1463 | } |
1464 | ||
1465 | /* | |
1466 | * Single hardware queue variant. This will attempt to use any per-process | |
1467 | * plug for merging and IO deferral. | |
1468 | */ | |
dece1635 | 1469 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1470 | { |
ef295ecf | 1471 | const int is_sync = op_is_sync(bio->bi_opf); |
1eff9d32 | 1472 | const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA); |
e6c4438b JM |
1473 | struct blk_plug *plug; |
1474 | unsigned int request_count = 0; | |
2552e3f8 | 1475 | struct blk_mq_alloc_data data; |
07068d5b | 1476 | struct request *rq; |
7b371636 | 1477 | blk_qc_t cookie; |
87760e5e | 1478 | unsigned int wb_acct; |
07068d5b | 1479 | |
07068d5b JA |
1480 | blk_queue_bounce(q, &bio); |
1481 | ||
1482 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1483 | bio_io_error(bio); |
dece1635 | 1484 | return BLK_QC_T_NONE; |
07068d5b JA |
1485 | } |
1486 | ||
54efd50b KO |
1487 | blk_queue_split(q, &bio, q->bio_split); |
1488 | ||
87c279e6 OS |
1489 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1490 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1491 | return BLK_QC_T_NONE; | |
1492 | } else | |
1493 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1494 | |
87760e5e JA |
1495 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1496 | ||
07068d5b | 1497 | rq = blk_mq_map_request(q, bio, &data); |
87760e5e JA |
1498 | if (unlikely(!rq)) { |
1499 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1500 | return BLK_QC_T_NONE; |
87760e5e JA |
1501 | } |
1502 | ||
1503 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1504 | |
fd2d3326 | 1505 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1506 | |
1507 | if (unlikely(is_flush_fua)) { | |
1508 | blk_mq_bio_to_request(rq, bio); | |
320ae51f JA |
1509 | blk_insert_flush(rq); |
1510 | goto run_queue; | |
1511 | } | |
1512 | ||
1513 | /* | |
1514 | * A task plug currently exists. Since this is completely lockless, | |
1515 | * utilize that to temporarily store requests until the task is | |
1516 | * either done or scheduled away. | |
1517 | */ | |
e6c4438b JM |
1518 | plug = current->plug; |
1519 | if (plug) { | |
600271d9 SL |
1520 | struct request *last = NULL; |
1521 | ||
e6c4438b | 1522 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1523 | |
1524 | /* | |
1525 | * @request_count may become stale because of schedule | |
1526 | * out, so check the list again. | |
1527 | */ | |
1528 | if (list_empty(&plug->mq_list)) | |
1529 | request_count = 0; | |
676d0607 | 1530 | if (!request_count) |
e6c4438b | 1531 | trace_block_plug(q); |
600271d9 SL |
1532 | else |
1533 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1534 | |
1535 | blk_mq_put_ctx(data.ctx); | |
1536 | ||
600271d9 SL |
1537 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1538 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1539 | blk_flush_plug_list(plug, false); |
1540 | trace_block_plug(q); | |
320ae51f | 1541 | } |
b094f89c | 1542 | |
e6c4438b | 1543 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1544 | return cookie; |
320ae51f JA |
1545 | } |
1546 | ||
07068d5b JA |
1547 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1548 | /* | |
1549 | * For a SYNC request, send it to the hardware immediately. For | |
1550 | * an ASYNC request, just ensure that we run it later on. The | |
1551 | * latter allows for merging opportunities and more efficient | |
1552 | * dispatching. | |
1553 | */ | |
1554 | run_queue: | |
1555 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); | |
320ae51f JA |
1556 | } |
1557 | ||
07068d5b | 1558 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1559 | return cookie; |
320ae51f JA |
1560 | } |
1561 | ||
cc71a6f4 JA |
1562 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1563 | unsigned int hctx_idx) | |
95363efd | 1564 | { |
e9b267d9 | 1565 | struct page *page; |
320ae51f | 1566 | |
24d2f903 | 1567 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1568 | int i; |
320ae51f | 1569 | |
24d2f903 | 1570 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1571 | struct request *rq = tags->static_rqs[i]; |
1572 | ||
1573 | if (!rq) | |
e9b267d9 | 1574 | continue; |
2af8cbe3 | 1575 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1576 | hctx_idx, i); |
2af8cbe3 | 1577 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1578 | } |
320ae51f | 1579 | } |
320ae51f | 1580 | |
24d2f903 CH |
1581 | while (!list_empty(&tags->page_list)) { |
1582 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1583 | list_del_init(&page->lru); |
f75782e4 CM |
1584 | /* |
1585 | * Remove kmemleak object previously allocated in | |
1586 | * blk_mq_init_rq_map(). | |
1587 | */ | |
1588 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1589 | __free_pages(page, page->private); |
1590 | } | |
cc71a6f4 | 1591 | } |
320ae51f | 1592 | |
cc71a6f4 JA |
1593 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1594 | { | |
24d2f903 | 1595 | kfree(tags->rqs); |
cc71a6f4 | 1596 | tags->rqs = NULL; |
2af8cbe3 JA |
1597 | kfree(tags->static_rqs); |
1598 | tags->static_rqs = NULL; | |
320ae51f | 1599 | |
24d2f903 | 1600 | blk_mq_free_tags(tags); |
320ae51f JA |
1601 | } |
1602 | ||
cc71a6f4 JA |
1603 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1604 | unsigned int hctx_idx, | |
1605 | unsigned int nr_tags, | |
1606 | unsigned int reserved_tags) | |
320ae51f | 1607 | { |
24d2f903 | 1608 | struct blk_mq_tags *tags; |
320ae51f | 1609 | |
cc71a6f4 | 1610 | tags = blk_mq_init_tags(nr_tags, reserved_tags, |
24391c0d SL |
1611 | set->numa_node, |
1612 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); | |
24d2f903 CH |
1613 | if (!tags) |
1614 | return NULL; | |
320ae51f | 1615 | |
cc71a6f4 | 1616 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1617 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
a5164405 | 1618 | set->numa_node); |
24d2f903 CH |
1619 | if (!tags->rqs) { |
1620 | blk_mq_free_tags(tags); | |
1621 | return NULL; | |
1622 | } | |
320ae51f | 1623 | |
2af8cbe3 JA |
1624 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1625 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
1626 | set->numa_node); | |
1627 | if (!tags->static_rqs) { | |
1628 | kfree(tags->rqs); | |
1629 | blk_mq_free_tags(tags); | |
1630 | return NULL; | |
1631 | } | |
1632 | ||
cc71a6f4 JA |
1633 | return tags; |
1634 | } | |
1635 | ||
1636 | static size_t order_to_size(unsigned int order) | |
1637 | { | |
1638 | return (size_t)PAGE_SIZE << order; | |
1639 | } | |
1640 | ||
1641 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1642 | unsigned int hctx_idx, unsigned int depth) | |
1643 | { | |
1644 | unsigned int i, j, entries_per_page, max_order = 4; | |
1645 | size_t rq_size, left; | |
1646 | ||
1647 | INIT_LIST_HEAD(&tags->page_list); | |
1648 | ||
320ae51f JA |
1649 | /* |
1650 | * rq_size is the size of the request plus driver payload, rounded | |
1651 | * to the cacheline size | |
1652 | */ | |
24d2f903 | 1653 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1654 | cache_line_size()); |
cc71a6f4 | 1655 | left = rq_size * depth; |
320ae51f | 1656 | |
cc71a6f4 | 1657 | for (i = 0; i < depth; ) { |
320ae51f JA |
1658 | int this_order = max_order; |
1659 | struct page *page; | |
1660 | int to_do; | |
1661 | void *p; | |
1662 | ||
b3a834b1 | 1663 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1664 | this_order--; |
1665 | ||
1666 | do { | |
a5164405 | 1667 | page = alloc_pages_node(set->numa_node, |
36e1f3d1 | 1668 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1669 | this_order); |
320ae51f JA |
1670 | if (page) |
1671 | break; | |
1672 | if (!this_order--) | |
1673 | break; | |
1674 | if (order_to_size(this_order) < rq_size) | |
1675 | break; | |
1676 | } while (1); | |
1677 | ||
1678 | if (!page) | |
24d2f903 | 1679 | goto fail; |
320ae51f JA |
1680 | |
1681 | page->private = this_order; | |
24d2f903 | 1682 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1683 | |
1684 | p = page_address(page); | |
f75782e4 CM |
1685 | /* |
1686 | * Allow kmemleak to scan these pages as they contain pointers | |
1687 | * to additional allocations like via ops->init_request(). | |
1688 | */ | |
36e1f3d1 | 1689 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1690 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1691 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1692 | left -= to_do * rq_size; |
1693 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1694 | struct request *rq = p; |
1695 | ||
1696 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1697 | if (set->ops->init_request) { |
1698 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1699 | rq, hctx_idx, i, |
a5164405 | 1700 | set->numa_node)) { |
2af8cbe3 | 1701 | tags->static_rqs[i] = NULL; |
24d2f903 | 1702 | goto fail; |
a5164405 | 1703 | } |
e9b267d9 CH |
1704 | } |
1705 | ||
320ae51f JA |
1706 | p += rq_size; |
1707 | i++; | |
1708 | } | |
1709 | } | |
cc71a6f4 | 1710 | return 0; |
320ae51f | 1711 | |
24d2f903 | 1712 | fail: |
cc71a6f4 JA |
1713 | blk_mq_free_rqs(set, tags, hctx_idx); |
1714 | return -ENOMEM; | |
320ae51f JA |
1715 | } |
1716 | ||
e57690fe JA |
1717 | /* |
1718 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1719 | * software queue to the hw queue dispatch list, and ensure that it | |
1720 | * gets run. | |
1721 | */ | |
9467f859 | 1722 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1723 | { |
9467f859 | 1724 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1725 | struct blk_mq_ctx *ctx; |
1726 | LIST_HEAD(tmp); | |
1727 | ||
9467f859 | 1728 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1729 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1730 | |
1731 | spin_lock(&ctx->lock); | |
1732 | if (!list_empty(&ctx->rq_list)) { | |
1733 | list_splice_init(&ctx->rq_list, &tmp); | |
1734 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1735 | } | |
1736 | spin_unlock(&ctx->lock); | |
1737 | ||
1738 | if (list_empty(&tmp)) | |
9467f859 | 1739 | return 0; |
484b4061 | 1740 | |
e57690fe JA |
1741 | spin_lock(&hctx->lock); |
1742 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1743 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1744 | |
1745 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1746 | return 0; |
484b4061 JA |
1747 | } |
1748 | ||
9467f859 | 1749 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1750 | { |
9467f859 TG |
1751 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1752 | &hctx->cpuhp_dead); | |
484b4061 JA |
1753 | } |
1754 | ||
c3b4afca | 1755 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1756 | static void blk_mq_exit_hctx(struct request_queue *q, |
1757 | struct blk_mq_tag_set *set, | |
1758 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1759 | { | |
f70ced09 ML |
1760 | unsigned flush_start_tag = set->queue_depth; |
1761 | ||
08e98fc6 ML |
1762 | blk_mq_tag_idle(hctx); |
1763 | ||
f70ced09 ML |
1764 | if (set->ops->exit_request) |
1765 | set->ops->exit_request(set->driver_data, | |
1766 | hctx->fq->flush_rq, hctx_idx, | |
1767 | flush_start_tag + hctx_idx); | |
1768 | ||
08e98fc6 ML |
1769 | if (set->ops->exit_hctx) |
1770 | set->ops->exit_hctx(hctx, hctx_idx); | |
1771 | ||
6a83e74d BVA |
1772 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1773 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1774 | ||
9467f859 | 1775 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1776 | blk_free_flush_queue(hctx->fq); |
88459642 | 1777 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1778 | } |
1779 | ||
624dbe47 ML |
1780 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1781 | struct blk_mq_tag_set *set, int nr_queue) | |
1782 | { | |
1783 | struct blk_mq_hw_ctx *hctx; | |
1784 | unsigned int i; | |
1785 | ||
1786 | queue_for_each_hw_ctx(q, hctx, i) { | |
1787 | if (i == nr_queue) | |
1788 | break; | |
08e98fc6 | 1789 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1790 | } |
624dbe47 ML |
1791 | } |
1792 | ||
1793 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1794 | struct blk_mq_tag_set *set) | |
1795 | { | |
1796 | struct blk_mq_hw_ctx *hctx; | |
1797 | unsigned int i; | |
1798 | ||
e09aae7e | 1799 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1800 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1801 | } |
1802 | ||
08e98fc6 ML |
1803 | static int blk_mq_init_hctx(struct request_queue *q, |
1804 | struct blk_mq_tag_set *set, | |
1805 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1806 | { |
08e98fc6 | 1807 | int node; |
f70ced09 | 1808 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1809 | |
1810 | node = hctx->numa_node; | |
1811 | if (node == NUMA_NO_NODE) | |
1812 | node = hctx->numa_node = set->numa_node; | |
1813 | ||
27489a3c | 1814 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1815 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1816 | spin_lock_init(&hctx->lock); | |
1817 | INIT_LIST_HEAD(&hctx->dispatch); | |
1818 | hctx->queue = q; | |
1819 | hctx->queue_num = hctx_idx; | |
2404e607 | 1820 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1821 | |
9467f859 | 1822 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1823 | |
1824 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1825 | |
1826 | /* | |
08e98fc6 ML |
1827 | * Allocate space for all possible cpus to avoid allocation at |
1828 | * runtime | |
320ae51f | 1829 | */ |
08e98fc6 ML |
1830 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1831 | GFP_KERNEL, node); | |
1832 | if (!hctx->ctxs) | |
1833 | goto unregister_cpu_notifier; | |
320ae51f | 1834 | |
88459642 OS |
1835 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1836 | node)) | |
08e98fc6 | 1837 | goto free_ctxs; |
320ae51f | 1838 | |
08e98fc6 | 1839 | hctx->nr_ctx = 0; |
320ae51f | 1840 | |
08e98fc6 ML |
1841 | if (set->ops->init_hctx && |
1842 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1843 | goto free_bitmap; | |
320ae51f | 1844 | |
f70ced09 ML |
1845 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1846 | if (!hctx->fq) | |
1847 | goto exit_hctx; | |
320ae51f | 1848 | |
f70ced09 ML |
1849 | if (set->ops->init_request && |
1850 | set->ops->init_request(set->driver_data, | |
1851 | hctx->fq->flush_rq, hctx_idx, | |
1852 | flush_start_tag + hctx_idx, node)) | |
1853 | goto free_fq; | |
320ae51f | 1854 | |
6a83e74d BVA |
1855 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1856 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1857 | ||
08e98fc6 | 1858 | return 0; |
320ae51f | 1859 | |
f70ced09 ML |
1860 | free_fq: |
1861 | kfree(hctx->fq); | |
1862 | exit_hctx: | |
1863 | if (set->ops->exit_hctx) | |
1864 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1865 | free_bitmap: |
88459642 | 1866 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1867 | free_ctxs: |
1868 | kfree(hctx->ctxs); | |
1869 | unregister_cpu_notifier: | |
9467f859 | 1870 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1871 | return -1; |
1872 | } | |
320ae51f | 1873 | |
320ae51f JA |
1874 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1875 | unsigned int nr_hw_queues) | |
1876 | { | |
1877 | unsigned int i; | |
1878 | ||
1879 | for_each_possible_cpu(i) { | |
1880 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1881 | struct blk_mq_hw_ctx *hctx; | |
1882 | ||
1883 | memset(__ctx, 0, sizeof(*__ctx)); | |
1884 | __ctx->cpu = i; | |
1885 | spin_lock_init(&__ctx->lock); | |
1886 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1887 | __ctx->queue = q; | |
cf43e6be JA |
1888 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
1889 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
1890 | |
1891 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1892 | if (!cpu_online(i)) |
1893 | continue; | |
1894 | ||
7d7e0f90 | 1895 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1896 | |
320ae51f JA |
1897 | /* |
1898 | * Set local node, IFF we have more than one hw queue. If | |
1899 | * not, we remain on the home node of the device | |
1900 | */ | |
1901 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1902 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1903 | } |
1904 | } | |
1905 | ||
cc71a6f4 JA |
1906 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1907 | { | |
1908 | int ret = 0; | |
1909 | ||
1910 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1911 | set->queue_depth, set->reserved_tags); | |
1912 | if (!set->tags[hctx_idx]) | |
1913 | return false; | |
1914 | ||
1915 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1916 | set->queue_depth); | |
1917 | if (!ret) | |
1918 | return true; | |
1919 | ||
1920 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1921 | set->tags[hctx_idx] = NULL; | |
1922 | return false; | |
1923 | } | |
1924 | ||
1925 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1926 | unsigned int hctx_idx) | |
1927 | { | |
1928 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1929 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1930 | set->tags[hctx_idx] = NULL; | |
1931 | } | |
1932 | ||
5778322e AM |
1933 | static void blk_mq_map_swqueue(struct request_queue *q, |
1934 | const struct cpumask *online_mask) | |
320ae51f | 1935 | { |
d1b1cea1 | 1936 | unsigned int i, hctx_idx; |
320ae51f JA |
1937 | struct blk_mq_hw_ctx *hctx; |
1938 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1939 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1940 | |
60de074b AM |
1941 | /* |
1942 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1943 | */ | |
1944 | mutex_lock(&q->sysfs_lock); | |
1945 | ||
320ae51f | 1946 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1947 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
1948 | hctx->nr_ctx = 0; |
1949 | } | |
1950 | ||
1951 | /* | |
1952 | * Map software to hardware queues | |
1953 | */ | |
897bb0c7 | 1954 | for_each_possible_cpu(i) { |
320ae51f | 1955 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 1956 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
1957 | continue; |
1958 | ||
d1b1cea1 GKB |
1959 | hctx_idx = q->mq_map[i]; |
1960 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
1961 | if (!set->tags[hctx_idx] && |
1962 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
1963 | /* |
1964 | * If tags initialization fail for some hctx, | |
1965 | * that hctx won't be brought online. In this | |
1966 | * case, remap the current ctx to hctx[0] which | |
1967 | * is guaranteed to always have tags allocated | |
1968 | */ | |
cc71a6f4 | 1969 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
1970 | } |
1971 | ||
897bb0c7 | 1972 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 1973 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 1974 | |
e4043dcf | 1975 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
1976 | ctx->index_hw = hctx->nr_ctx; |
1977 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
1978 | } | |
506e931f | 1979 | |
60de074b AM |
1980 | mutex_unlock(&q->sysfs_lock); |
1981 | ||
506e931f | 1982 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 1983 | /* |
a68aafa5 JA |
1984 | * If no software queues are mapped to this hardware queue, |
1985 | * disable it and free the request entries. | |
484b4061 JA |
1986 | */ |
1987 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
1988 | /* Never unmap queue 0. We need it as a |
1989 | * fallback in case of a new remap fails | |
1990 | * allocation | |
1991 | */ | |
cc71a6f4 JA |
1992 | if (i && set->tags[i]) |
1993 | blk_mq_free_map_and_requests(set, i); | |
1994 | ||
2a34c087 | 1995 | hctx->tags = NULL; |
484b4061 JA |
1996 | continue; |
1997 | } | |
1998 | ||
2a34c087 ML |
1999 | hctx->tags = set->tags[i]; |
2000 | WARN_ON(!hctx->tags); | |
2001 | ||
889fa31f CY |
2002 | /* |
2003 | * Set the map size to the number of mapped software queues. | |
2004 | * This is more accurate and more efficient than looping | |
2005 | * over all possibly mapped software queues. | |
2006 | */ | |
88459642 | 2007 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2008 | |
484b4061 JA |
2009 | /* |
2010 | * Initialize batch roundrobin counts | |
2011 | */ | |
506e931f JA |
2012 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2013 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2014 | } | |
320ae51f JA |
2015 | } |
2016 | ||
2404e607 | 2017 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2018 | { |
2019 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2020 | int i; |
2021 | ||
2404e607 JM |
2022 | queue_for_each_hw_ctx(q, hctx, i) { |
2023 | if (shared) | |
2024 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2025 | else | |
2026 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2027 | } | |
2028 | } | |
2029 | ||
2030 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2031 | { | |
2032 | struct request_queue *q; | |
0d2602ca JA |
2033 | |
2034 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2035 | blk_mq_freeze_queue(q); | |
2404e607 | 2036 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2037 | blk_mq_unfreeze_queue(q); |
2038 | } | |
2039 | } | |
2040 | ||
2041 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2042 | { | |
2043 | struct blk_mq_tag_set *set = q->tag_set; | |
2044 | ||
0d2602ca JA |
2045 | mutex_lock(&set->tag_list_lock); |
2046 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2047 | if (list_is_singular(&set->tag_list)) { |
2048 | /* just transitioned to unshared */ | |
2049 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2050 | /* update existing queue */ | |
2051 | blk_mq_update_tag_set_depth(set, false); | |
2052 | } | |
0d2602ca | 2053 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2054 | } |
2055 | ||
2056 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2057 | struct request_queue *q) | |
2058 | { | |
2059 | q->tag_set = set; | |
2060 | ||
2061 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2062 | |
2063 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2064 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2065 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2066 | /* update existing queue */ | |
2067 | blk_mq_update_tag_set_depth(set, true); | |
2068 | } | |
2069 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2070 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2071 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2072 | |
0d2602ca JA |
2073 | mutex_unlock(&set->tag_list_lock); |
2074 | } | |
2075 | ||
e09aae7e ML |
2076 | /* |
2077 | * It is the actual release handler for mq, but we do it from | |
2078 | * request queue's release handler for avoiding use-after-free | |
2079 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2080 | * but we can't group ctx/kctx kobj without it. | |
2081 | */ | |
2082 | void blk_mq_release(struct request_queue *q) | |
2083 | { | |
2084 | struct blk_mq_hw_ctx *hctx; | |
2085 | unsigned int i; | |
2086 | ||
2087 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2088 | queue_for_each_hw_ctx(q, hctx, i) { |
2089 | if (!hctx) | |
2090 | continue; | |
2091 | kfree(hctx->ctxs); | |
e09aae7e | 2092 | kfree(hctx); |
c3b4afca | 2093 | } |
e09aae7e | 2094 | |
a723bab3 AM |
2095 | q->mq_map = NULL; |
2096 | ||
e09aae7e ML |
2097 | kfree(q->queue_hw_ctx); |
2098 | ||
2099 | /* ctx kobj stays in queue_ctx */ | |
2100 | free_percpu(q->queue_ctx); | |
2101 | } | |
2102 | ||
24d2f903 | 2103 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2104 | { |
2105 | struct request_queue *uninit_q, *q; | |
2106 | ||
2107 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2108 | if (!uninit_q) | |
2109 | return ERR_PTR(-ENOMEM); | |
2110 | ||
2111 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2112 | if (IS_ERR(q)) | |
2113 | blk_cleanup_queue(uninit_q); | |
2114 | ||
2115 | return q; | |
2116 | } | |
2117 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2118 | ||
868f2f0b KB |
2119 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2120 | struct request_queue *q) | |
320ae51f | 2121 | { |
868f2f0b KB |
2122 | int i, j; |
2123 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2124 | |
868f2f0b | 2125 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2126 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2127 | int node; |
f14bbe77 | 2128 | |
868f2f0b KB |
2129 | if (hctxs[i]) |
2130 | continue; | |
2131 | ||
2132 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2133 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2134 | GFP_KERNEL, node); | |
320ae51f | 2135 | if (!hctxs[i]) |
868f2f0b | 2136 | break; |
320ae51f | 2137 | |
a86073e4 | 2138 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2139 | node)) { |
2140 | kfree(hctxs[i]); | |
2141 | hctxs[i] = NULL; | |
2142 | break; | |
2143 | } | |
e4043dcf | 2144 | |
0d2602ca | 2145 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2146 | hctxs[i]->numa_node = node; |
320ae51f | 2147 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2148 | |
2149 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2150 | free_cpumask_var(hctxs[i]->cpumask); | |
2151 | kfree(hctxs[i]); | |
2152 | hctxs[i] = NULL; | |
2153 | break; | |
2154 | } | |
2155 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2156 | } |
868f2f0b KB |
2157 | for (j = i; j < q->nr_hw_queues; j++) { |
2158 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2159 | ||
2160 | if (hctx) { | |
cc71a6f4 JA |
2161 | if (hctx->tags) |
2162 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2163 | blk_mq_exit_hctx(q, set, hctx, j); |
2164 | free_cpumask_var(hctx->cpumask); | |
2165 | kobject_put(&hctx->kobj); | |
2166 | kfree(hctx->ctxs); | |
2167 | kfree(hctx); | |
2168 | hctxs[j] = NULL; | |
2169 | ||
2170 | } | |
2171 | } | |
2172 | q->nr_hw_queues = i; | |
2173 | blk_mq_sysfs_register(q); | |
2174 | } | |
2175 | ||
2176 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2177 | struct request_queue *q) | |
2178 | { | |
66841672 ML |
2179 | /* mark the queue as mq asap */ |
2180 | q->mq_ops = set->ops; | |
2181 | ||
868f2f0b KB |
2182 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2183 | if (!q->queue_ctx) | |
c7de5726 | 2184 | goto err_exit; |
868f2f0b KB |
2185 | |
2186 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2187 | GFP_KERNEL, set->numa_node); | |
2188 | if (!q->queue_hw_ctx) | |
2189 | goto err_percpu; | |
2190 | ||
bdd17e75 | 2191 | q->mq_map = set->mq_map; |
868f2f0b KB |
2192 | |
2193 | blk_mq_realloc_hw_ctxs(set, q); | |
2194 | if (!q->nr_hw_queues) | |
2195 | goto err_hctxs; | |
320ae51f | 2196 | |
287922eb | 2197 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2198 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2199 | |
2200 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2201 | |
94eddfbe | 2202 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2203 | |
05f1dd53 JA |
2204 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2205 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2206 | ||
1be036e9 CH |
2207 | q->sg_reserved_size = INT_MAX; |
2208 | ||
2849450a | 2209 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2210 | INIT_LIST_HEAD(&q->requeue_list); |
2211 | spin_lock_init(&q->requeue_lock); | |
2212 | ||
07068d5b JA |
2213 | if (q->nr_hw_queues > 1) |
2214 | blk_queue_make_request(q, blk_mq_make_request); | |
2215 | else | |
2216 | blk_queue_make_request(q, blk_sq_make_request); | |
2217 | ||
eba71768 JA |
2218 | /* |
2219 | * Do this after blk_queue_make_request() overrides it... | |
2220 | */ | |
2221 | q->nr_requests = set->queue_depth; | |
2222 | ||
64f1c21e JA |
2223 | /* |
2224 | * Default to classic polling | |
2225 | */ | |
2226 | q->poll_nsec = -1; | |
2227 | ||
24d2f903 CH |
2228 | if (set->ops->complete) |
2229 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2230 | |
24d2f903 | 2231 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2232 | |
5778322e | 2233 | get_online_cpus(); |
320ae51f | 2234 | mutex_lock(&all_q_mutex); |
320ae51f | 2235 | |
4593fdbe | 2236 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2237 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2238 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2239 | |
4593fdbe | 2240 | mutex_unlock(&all_q_mutex); |
5778322e | 2241 | put_online_cpus(); |
4593fdbe | 2242 | |
320ae51f | 2243 | return q; |
18741986 | 2244 | |
320ae51f | 2245 | err_hctxs: |
868f2f0b | 2246 | kfree(q->queue_hw_ctx); |
320ae51f | 2247 | err_percpu: |
868f2f0b | 2248 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2249 | err_exit: |
2250 | q->mq_ops = NULL; | |
320ae51f JA |
2251 | return ERR_PTR(-ENOMEM); |
2252 | } | |
b62c21b7 | 2253 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2254 | |
2255 | void blk_mq_free_queue(struct request_queue *q) | |
2256 | { | |
624dbe47 | 2257 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2258 | |
0e626368 AM |
2259 | mutex_lock(&all_q_mutex); |
2260 | list_del_init(&q->all_q_node); | |
2261 | mutex_unlock(&all_q_mutex); | |
2262 | ||
87760e5e JA |
2263 | wbt_exit(q); |
2264 | ||
0d2602ca JA |
2265 | blk_mq_del_queue_tag_set(q); |
2266 | ||
624dbe47 ML |
2267 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2268 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2269 | } |
320ae51f JA |
2270 | |
2271 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2272 | static void blk_mq_queue_reinit(struct request_queue *q, |
2273 | const struct cpumask *online_mask) | |
320ae51f | 2274 | { |
4ecd4fef | 2275 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2276 | |
67aec14c JA |
2277 | blk_mq_sysfs_unregister(q); |
2278 | ||
320ae51f JA |
2279 | /* |
2280 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2281 | * we should change hctx numa_node according to new topology (this | |
2282 | * involves free and re-allocate memory, worthy doing?) | |
2283 | */ | |
2284 | ||
5778322e | 2285 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2286 | |
67aec14c | 2287 | blk_mq_sysfs_register(q); |
320ae51f JA |
2288 | } |
2289 | ||
65d5291e SAS |
2290 | /* |
2291 | * New online cpumask which is going to be set in this hotplug event. | |
2292 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2293 | * one-by-one and dynamically allocating this could result in a failure. | |
2294 | */ | |
2295 | static struct cpumask cpuhp_online_new; | |
2296 | ||
2297 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2298 | { |
2299 | struct request_queue *q; | |
320ae51f JA |
2300 | |
2301 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2302 | /* |
2303 | * We need to freeze and reinit all existing queues. Freezing | |
2304 | * involves synchronous wait for an RCU grace period and doing it | |
2305 | * one by one may take a long time. Start freezing all queues in | |
2306 | * one swoop and then wait for the completions so that freezing can | |
2307 | * take place in parallel. | |
2308 | */ | |
2309 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2310 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2311 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2312 | blk_mq_freeze_queue_wait(q); |
2313 | ||
320ae51f | 2314 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2315 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2316 | |
2317 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2318 | blk_mq_unfreeze_queue(q); | |
2319 | ||
320ae51f | 2320 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2321 | } |
2322 | ||
2323 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2324 | { | |
97a32864 | 2325 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2326 | blk_mq_queue_reinit_work(); |
2327 | return 0; | |
2328 | } | |
2329 | ||
2330 | /* | |
2331 | * Before hotadded cpu starts handling requests, new mappings must be | |
2332 | * established. Otherwise, these requests in hw queue might never be | |
2333 | * dispatched. | |
2334 | * | |
2335 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2336 | * for CPU0, and ctx1 for CPU1). | |
2337 | * | |
2338 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2339 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2340 | * | |
2c3ad667 JA |
2341 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2342 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2343 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2344 | * ignored. | |
65d5291e SAS |
2345 | */ |
2346 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2347 | { | |
2348 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2349 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2350 | blk_mq_queue_reinit_work(); | |
2351 | return 0; | |
320ae51f JA |
2352 | } |
2353 | ||
a5164405 JA |
2354 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2355 | { | |
2356 | int i; | |
2357 | ||
cc71a6f4 JA |
2358 | for (i = 0; i < set->nr_hw_queues; i++) |
2359 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2360 | goto out_unwind; |
a5164405 JA |
2361 | |
2362 | return 0; | |
2363 | ||
2364 | out_unwind: | |
2365 | while (--i >= 0) | |
cc71a6f4 | 2366 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2367 | |
a5164405 JA |
2368 | return -ENOMEM; |
2369 | } | |
2370 | ||
2371 | /* | |
2372 | * Allocate the request maps associated with this tag_set. Note that this | |
2373 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2374 | * will be updated to reflect the allocated depth. | |
2375 | */ | |
2376 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2377 | { | |
2378 | unsigned int depth; | |
2379 | int err; | |
2380 | ||
2381 | depth = set->queue_depth; | |
2382 | do { | |
2383 | err = __blk_mq_alloc_rq_maps(set); | |
2384 | if (!err) | |
2385 | break; | |
2386 | ||
2387 | set->queue_depth >>= 1; | |
2388 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2389 | err = -ENOMEM; | |
2390 | break; | |
2391 | } | |
2392 | } while (set->queue_depth); | |
2393 | ||
2394 | if (!set->queue_depth || err) { | |
2395 | pr_err("blk-mq: failed to allocate request map\n"); | |
2396 | return -ENOMEM; | |
2397 | } | |
2398 | ||
2399 | if (depth != set->queue_depth) | |
2400 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2401 | depth, set->queue_depth); | |
2402 | ||
2403 | return 0; | |
2404 | } | |
2405 | ||
a4391c64 JA |
2406 | /* |
2407 | * Alloc a tag set to be associated with one or more request queues. | |
2408 | * May fail with EINVAL for various error conditions. May adjust the | |
2409 | * requested depth down, if if it too large. In that case, the set | |
2410 | * value will be stored in set->queue_depth. | |
2411 | */ | |
24d2f903 CH |
2412 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2413 | { | |
da695ba2 CH |
2414 | int ret; |
2415 | ||
205fb5f5 BVA |
2416 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2417 | ||
24d2f903 CH |
2418 | if (!set->nr_hw_queues) |
2419 | return -EINVAL; | |
a4391c64 | 2420 | if (!set->queue_depth) |
24d2f903 CH |
2421 | return -EINVAL; |
2422 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2423 | return -EINVAL; | |
2424 | ||
7d7e0f90 | 2425 | if (!set->ops->queue_rq) |
24d2f903 CH |
2426 | return -EINVAL; |
2427 | ||
a4391c64 JA |
2428 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2429 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2430 | BLK_MQ_MAX_DEPTH); | |
2431 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2432 | } | |
24d2f903 | 2433 | |
6637fadf SL |
2434 | /* |
2435 | * If a crashdump is active, then we are potentially in a very | |
2436 | * memory constrained environment. Limit us to 1 queue and | |
2437 | * 64 tags to prevent using too much memory. | |
2438 | */ | |
2439 | if (is_kdump_kernel()) { | |
2440 | set->nr_hw_queues = 1; | |
2441 | set->queue_depth = min(64U, set->queue_depth); | |
2442 | } | |
868f2f0b KB |
2443 | /* |
2444 | * There is no use for more h/w queues than cpus. | |
2445 | */ | |
2446 | if (set->nr_hw_queues > nr_cpu_ids) | |
2447 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2448 | |
868f2f0b | 2449 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2450 | GFP_KERNEL, set->numa_node); |
2451 | if (!set->tags) | |
a5164405 | 2452 | return -ENOMEM; |
24d2f903 | 2453 | |
da695ba2 CH |
2454 | ret = -ENOMEM; |
2455 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2456 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2457 | if (!set->mq_map) |
2458 | goto out_free_tags; | |
2459 | ||
da695ba2 CH |
2460 | if (set->ops->map_queues) |
2461 | ret = set->ops->map_queues(set); | |
2462 | else | |
2463 | ret = blk_mq_map_queues(set); | |
2464 | if (ret) | |
2465 | goto out_free_mq_map; | |
2466 | ||
2467 | ret = blk_mq_alloc_rq_maps(set); | |
2468 | if (ret) | |
bdd17e75 | 2469 | goto out_free_mq_map; |
24d2f903 | 2470 | |
0d2602ca JA |
2471 | mutex_init(&set->tag_list_lock); |
2472 | INIT_LIST_HEAD(&set->tag_list); | |
2473 | ||
24d2f903 | 2474 | return 0; |
bdd17e75 CH |
2475 | |
2476 | out_free_mq_map: | |
2477 | kfree(set->mq_map); | |
2478 | set->mq_map = NULL; | |
2479 | out_free_tags: | |
5676e7b6 RE |
2480 | kfree(set->tags); |
2481 | set->tags = NULL; | |
da695ba2 | 2482 | return ret; |
24d2f903 CH |
2483 | } |
2484 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2485 | ||
2486 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2487 | { | |
2488 | int i; | |
2489 | ||
cc71a6f4 JA |
2490 | for (i = 0; i < nr_cpu_ids; i++) |
2491 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2492 | |
bdd17e75 CH |
2493 | kfree(set->mq_map); |
2494 | set->mq_map = NULL; | |
2495 | ||
981bd189 | 2496 | kfree(set->tags); |
5676e7b6 | 2497 | set->tags = NULL; |
24d2f903 CH |
2498 | } |
2499 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2500 | ||
e3a2b3f9 JA |
2501 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2502 | { | |
2503 | struct blk_mq_tag_set *set = q->tag_set; | |
2504 | struct blk_mq_hw_ctx *hctx; | |
2505 | int i, ret; | |
2506 | ||
2507 | if (!set || nr > set->queue_depth) | |
2508 | return -EINVAL; | |
2509 | ||
2510 | ret = 0; | |
2511 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2512 | if (!hctx->tags) |
2513 | continue; | |
e3a2b3f9 JA |
2514 | ret = blk_mq_tag_update_depth(hctx->tags, nr); |
2515 | if (ret) | |
2516 | break; | |
2517 | } | |
2518 | ||
2519 | if (!ret) | |
2520 | q->nr_requests = nr; | |
2521 | ||
2522 | return ret; | |
2523 | } | |
2524 | ||
868f2f0b KB |
2525 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2526 | { | |
2527 | struct request_queue *q; | |
2528 | ||
2529 | if (nr_hw_queues > nr_cpu_ids) | |
2530 | nr_hw_queues = nr_cpu_ids; | |
2531 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2532 | return; | |
2533 | ||
2534 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2535 | blk_mq_freeze_queue(q); | |
2536 | ||
2537 | set->nr_hw_queues = nr_hw_queues; | |
2538 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2539 | blk_mq_realloc_hw_ctxs(set, q); | |
2540 | ||
2541 | if (q->nr_hw_queues > 1) | |
2542 | blk_queue_make_request(q, blk_mq_make_request); | |
2543 | else | |
2544 | blk_queue_make_request(q, blk_sq_make_request); | |
2545 | ||
2546 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2547 | } | |
2548 | ||
2549 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2550 | blk_mq_unfreeze_queue(q); | |
2551 | } | |
2552 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2553 | ||
64f1c21e JA |
2554 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2555 | struct blk_mq_hw_ctx *hctx, | |
2556 | struct request *rq) | |
2557 | { | |
2558 | struct blk_rq_stat stat[2]; | |
2559 | unsigned long ret = 0; | |
2560 | ||
2561 | /* | |
2562 | * If stats collection isn't on, don't sleep but turn it on for | |
2563 | * future users | |
2564 | */ | |
2565 | if (!blk_stat_enable(q)) | |
2566 | return 0; | |
2567 | ||
2568 | /* | |
2569 | * We don't have to do this once per IO, should optimize this | |
2570 | * to just use the current window of stats until it changes | |
2571 | */ | |
2572 | memset(&stat, 0, sizeof(stat)); | |
2573 | blk_hctx_stat_get(hctx, stat); | |
2574 | ||
2575 | /* | |
2576 | * As an optimistic guess, use half of the mean service time | |
2577 | * for this type of request. We can (and should) make this smarter. | |
2578 | * For instance, if the completion latencies are tight, we can | |
2579 | * get closer than just half the mean. This is especially | |
2580 | * important on devices where the completion latencies are longer | |
2581 | * than ~10 usec. | |
2582 | */ | |
2583 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2584 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2585 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2586 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2587 | ||
2588 | return ret; | |
2589 | } | |
2590 | ||
06426adf | 2591 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2592 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2593 | struct request *rq) |
2594 | { | |
2595 | struct hrtimer_sleeper hs; | |
2596 | enum hrtimer_mode mode; | |
64f1c21e | 2597 | unsigned int nsecs; |
06426adf JA |
2598 | ktime_t kt; |
2599 | ||
64f1c21e JA |
2600 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2601 | return false; | |
2602 | ||
2603 | /* | |
2604 | * poll_nsec can be: | |
2605 | * | |
2606 | * -1: don't ever hybrid sleep | |
2607 | * 0: use half of prev avg | |
2608 | * >0: use this specific value | |
2609 | */ | |
2610 | if (q->poll_nsec == -1) | |
2611 | return false; | |
2612 | else if (q->poll_nsec > 0) | |
2613 | nsecs = q->poll_nsec; | |
2614 | else | |
2615 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2616 | ||
2617 | if (!nsecs) | |
06426adf JA |
2618 | return false; |
2619 | ||
2620 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2621 | ||
2622 | /* | |
2623 | * This will be replaced with the stats tracking code, using | |
2624 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2625 | */ | |
8b0e1953 | 2626 | kt = nsecs; |
06426adf JA |
2627 | |
2628 | mode = HRTIMER_MODE_REL; | |
2629 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2630 | hrtimer_set_expires(&hs.timer, kt); | |
2631 | ||
2632 | hrtimer_init_sleeper(&hs, current); | |
2633 | do { | |
2634 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2635 | break; | |
2636 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2637 | hrtimer_start_expires(&hs.timer, mode); | |
2638 | if (hs.task) | |
2639 | io_schedule(); | |
2640 | hrtimer_cancel(&hs.timer); | |
2641 | mode = HRTIMER_MODE_ABS; | |
2642 | } while (hs.task && !signal_pending(current)); | |
2643 | ||
2644 | __set_current_state(TASK_RUNNING); | |
2645 | destroy_hrtimer_on_stack(&hs.timer); | |
2646 | return true; | |
2647 | } | |
2648 | ||
bbd7bb70 JA |
2649 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2650 | { | |
2651 | struct request_queue *q = hctx->queue; | |
2652 | long state; | |
2653 | ||
06426adf JA |
2654 | /* |
2655 | * If we sleep, have the caller restart the poll loop to reset | |
2656 | * the state. Like for the other success return cases, the | |
2657 | * caller is responsible for checking if the IO completed. If | |
2658 | * the IO isn't complete, we'll get called again and will go | |
2659 | * straight to the busy poll loop. | |
2660 | */ | |
64f1c21e | 2661 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2662 | return true; |
2663 | ||
bbd7bb70 JA |
2664 | hctx->poll_considered++; |
2665 | ||
2666 | state = current->state; | |
2667 | while (!need_resched()) { | |
2668 | int ret; | |
2669 | ||
2670 | hctx->poll_invoked++; | |
2671 | ||
2672 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2673 | if (ret > 0) { | |
2674 | hctx->poll_success++; | |
2675 | set_current_state(TASK_RUNNING); | |
2676 | return true; | |
2677 | } | |
2678 | ||
2679 | if (signal_pending_state(state, current)) | |
2680 | set_current_state(TASK_RUNNING); | |
2681 | ||
2682 | if (current->state == TASK_RUNNING) | |
2683 | return true; | |
2684 | if (ret < 0) | |
2685 | break; | |
2686 | cpu_relax(); | |
2687 | } | |
2688 | ||
2689 | return false; | |
2690 | } | |
2691 | ||
2692 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2693 | { | |
2694 | struct blk_mq_hw_ctx *hctx; | |
2695 | struct blk_plug *plug; | |
2696 | struct request *rq; | |
2697 | ||
2698 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2699 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2700 | return false; | |
2701 | ||
2702 | plug = current->plug; | |
2703 | if (plug) | |
2704 | blk_flush_plug_list(plug, false); | |
2705 | ||
2706 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
2707 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2708 | ||
2709 | return __blk_mq_poll(hctx, rq); | |
2710 | } | |
2711 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2712 | ||
676141e4 JA |
2713 | void blk_mq_disable_hotplug(void) |
2714 | { | |
2715 | mutex_lock(&all_q_mutex); | |
2716 | } | |
2717 | ||
2718 | void blk_mq_enable_hotplug(void) | |
2719 | { | |
2720 | mutex_unlock(&all_q_mutex); | |
2721 | } | |
2722 | ||
320ae51f JA |
2723 | static int __init blk_mq_init(void) |
2724 | { | |
9467f859 TG |
2725 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2726 | blk_mq_hctx_notify_dead); | |
320ae51f | 2727 | |
65d5291e SAS |
2728 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2729 | blk_mq_queue_reinit_prepare, | |
2730 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2731 | return 0; |
2732 | } | |
2733 | subsys_initcall(blk_mq_init); |