]>
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> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
34 | #include "blk-mq-tag.h" | |
cf43e6be | 35 | #include "blk-stat.h" |
87760e5e | 36 | #include "blk-wbt.h" |
bd166ef1 | 37 | #include "blk-mq-sched.h" |
320ae51f JA |
38 | |
39 | static DEFINE_MUTEX(all_q_mutex); | |
40 | static LIST_HEAD(all_q_list); | |
41 | ||
34dbad5d OS |
42 | static void blk_mq_poll_stats_start(struct request_queue *q); |
43 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
44 | ||
320ae51f JA |
45 | /* |
46 | * Check if any of the ctx's have pending work in this hardware queue | |
47 | */ | |
50e1dab8 | 48 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 49 | { |
bd166ef1 JA |
50 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
51 | !list_empty_careful(&hctx->dispatch) || | |
52 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
53 | } |
54 | ||
320ae51f JA |
55 | /* |
56 | * Mark this ctx as having pending work in this hardware queue | |
57 | */ | |
58 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
59 | struct blk_mq_ctx *ctx) | |
60 | { | |
88459642 OS |
61 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
62 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
63 | } |
64 | ||
65 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
66 | struct blk_mq_ctx *ctx) | |
67 | { | |
88459642 | 68 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
69 | } |
70 | ||
b4c6a028 | 71 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 72 | { |
4ecd4fef | 73 | int freeze_depth; |
cddd5d17 | 74 | |
4ecd4fef CH |
75 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
76 | if (freeze_depth == 1) { | |
3ef28e83 | 77 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 78 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 79 | } |
f3af020b | 80 | } |
b4c6a028 | 81 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b | 82 | |
6bae363e | 83 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 84 | { |
3ef28e83 | 85 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 86 | } |
6bae363e | 87 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 88 | |
f91328c4 KB |
89 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
90 | unsigned long timeout) | |
91 | { | |
92 | return wait_event_timeout(q->mq_freeze_wq, | |
93 | percpu_ref_is_zero(&q->q_usage_counter), | |
94 | timeout); | |
95 | } | |
96 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 97 | |
f3af020b TH |
98 | /* |
99 | * Guarantee no request is in use, so we can change any data structure of | |
100 | * the queue afterward. | |
101 | */ | |
3ef28e83 | 102 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 103 | { |
3ef28e83 DW |
104 | /* |
105 | * In the !blk_mq case we are only calling this to kill the | |
106 | * q_usage_counter, otherwise this increases the freeze depth | |
107 | * and waits for it to return to zero. For this reason there is | |
108 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
109 | * exported to drivers as the only user for unfreeze is blk_mq. | |
110 | */ | |
f3af020b TH |
111 | blk_mq_freeze_queue_start(q); |
112 | blk_mq_freeze_queue_wait(q); | |
113 | } | |
3ef28e83 DW |
114 | |
115 | void blk_mq_freeze_queue(struct request_queue *q) | |
116 | { | |
117 | /* | |
118 | * ...just an alias to keep freeze and unfreeze actions balanced | |
119 | * in the blk_mq_* namespace | |
120 | */ | |
121 | blk_freeze_queue(q); | |
122 | } | |
c761d96b | 123 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 124 | |
b4c6a028 | 125 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 126 | { |
4ecd4fef | 127 | int freeze_depth; |
320ae51f | 128 | |
4ecd4fef CH |
129 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
130 | WARN_ON_ONCE(freeze_depth < 0); | |
131 | if (!freeze_depth) { | |
3ef28e83 | 132 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 133 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 134 | } |
320ae51f | 135 | } |
b4c6a028 | 136 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 137 | |
6a83e74d BVA |
138 | /** |
139 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
140 | * @q: request queue. | |
141 | * | |
142 | * Note: this function does not prevent that the struct request end_io() | |
143 | * callback function is invoked. Additionally, it is not prevented that | |
144 | * new queue_rq() calls occur unless the queue has been stopped first. | |
145 | */ | |
146 | void blk_mq_quiesce_queue(struct request_queue *q) | |
147 | { | |
148 | struct blk_mq_hw_ctx *hctx; | |
149 | unsigned int i; | |
150 | bool rcu = false; | |
151 | ||
152 | blk_mq_stop_hw_queues(q); | |
153 | ||
154 | queue_for_each_hw_ctx(q, hctx, i) { | |
155 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
156 | synchronize_srcu(&hctx->queue_rq_srcu); | |
157 | else | |
158 | rcu = true; | |
159 | } | |
160 | if (rcu) | |
161 | synchronize_rcu(); | |
162 | } | |
163 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
164 | ||
aed3ea94 JA |
165 | void blk_mq_wake_waiters(struct request_queue *q) |
166 | { | |
167 | struct blk_mq_hw_ctx *hctx; | |
168 | unsigned int i; | |
169 | ||
170 | queue_for_each_hw_ctx(q, hctx, i) | |
171 | if (blk_mq_hw_queue_mapped(hctx)) | |
172 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
173 | |
174 | /* | |
175 | * If we are called because the queue has now been marked as | |
176 | * dying, we need to ensure that processes currently waiting on | |
177 | * the queue are notified as well. | |
178 | */ | |
179 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
180 | } |
181 | ||
320ae51f JA |
182 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
183 | { | |
184 | return blk_mq_has_free_tags(hctx->tags); | |
185 | } | |
186 | EXPORT_SYMBOL(blk_mq_can_queue); | |
187 | ||
2c3ad667 JA |
188 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
189 | struct request *rq, unsigned int op) | |
320ae51f | 190 | { |
af76e555 CH |
191 | INIT_LIST_HEAD(&rq->queuelist); |
192 | /* csd/requeue_work/fifo_time is initialized before use */ | |
193 | rq->q = q; | |
320ae51f | 194 | rq->mq_ctx = ctx; |
ef295ecf | 195 | rq->cmd_flags = op; |
e8064021 CH |
196 | if (blk_queue_io_stat(q)) |
197 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
198 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
199 | rq->cpu = -1; | |
af76e555 CH |
200 | INIT_HLIST_NODE(&rq->hash); |
201 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
202 | rq->rq_disk = NULL; |
203 | rq->part = NULL; | |
3ee32372 | 204 | rq->start_time = jiffies; |
af76e555 CH |
205 | #ifdef CONFIG_BLK_CGROUP |
206 | rq->rl = NULL; | |
0fec08b4 | 207 | set_start_time_ns(rq); |
af76e555 CH |
208 | rq->io_start_time_ns = 0; |
209 | #endif | |
210 | rq->nr_phys_segments = 0; | |
211 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
212 | rq->nr_integrity_segments = 0; | |
213 | #endif | |
af76e555 CH |
214 | rq->special = NULL; |
215 | /* tag was already set */ | |
216 | rq->errors = 0; | |
af76e555 | 217 | rq->extra_len = 0; |
af76e555 | 218 | |
af76e555 | 219 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
220 | rq->timeout = 0; |
221 | ||
af76e555 CH |
222 | rq->end_io = NULL; |
223 | rq->end_io_data = NULL; | |
224 | rq->next_rq = NULL; | |
225 | ||
ef295ecf | 226 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 227 | } |
2c3ad667 | 228 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 229 | |
2c3ad667 JA |
230 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
231 | unsigned int op) | |
5dee8577 CH |
232 | { |
233 | struct request *rq; | |
234 | unsigned int tag; | |
235 | ||
cb96a42c | 236 | tag = blk_mq_get_tag(data); |
5dee8577 | 237 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
238 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
239 | ||
240 | rq = tags->static_rqs[tag]; | |
5dee8577 | 241 | |
bd166ef1 JA |
242 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
243 | rq->tag = -1; | |
244 | rq->internal_tag = tag; | |
245 | } else { | |
200e86b3 JA |
246 | if (blk_mq_tag_busy(data->hctx)) { |
247 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
248 | atomic_inc(&data->hctx->nr_active); | |
249 | } | |
bd166ef1 JA |
250 | rq->tag = tag; |
251 | rq->internal_tag = -1; | |
562bef42 | 252 | data->hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
253 | } |
254 | ||
ef295ecf | 255 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
256 | return rq; |
257 | } | |
258 | ||
259 | return NULL; | |
260 | } | |
2c3ad667 | 261 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 262 | |
6f3b0e8b CH |
263 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
264 | unsigned int flags) | |
320ae51f | 265 | { |
5a797e00 | 266 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 267 | struct request *rq; |
a492f075 | 268 | int ret; |
320ae51f | 269 | |
6f3b0e8b | 270 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
271 | if (ret) |
272 | return ERR_PTR(ret); | |
320ae51f | 273 | |
bd166ef1 | 274 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 275 | |
bd166ef1 JA |
276 | blk_mq_put_ctx(alloc_data.ctx); |
277 | blk_queue_exit(q); | |
278 | ||
279 | if (!rq) | |
a492f075 | 280 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
281 | |
282 | rq->__data_len = 0; | |
283 | rq->__sector = (sector_t) -1; | |
284 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
285 | return rq; |
286 | } | |
4bb659b1 | 287 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 288 | |
1f5bd336 ML |
289 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
290 | unsigned int flags, unsigned int hctx_idx) | |
291 | { | |
6d2809d5 | 292 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 293 | struct request *rq; |
6d2809d5 | 294 | unsigned int cpu; |
1f5bd336 ML |
295 | int ret; |
296 | ||
297 | /* | |
298 | * If the tag allocator sleeps we could get an allocation for a | |
299 | * different hardware context. No need to complicate the low level | |
300 | * allocator for this for the rare use case of a command tied to | |
301 | * a specific queue. | |
302 | */ | |
303 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
304 | return ERR_PTR(-EINVAL); | |
305 | ||
306 | if (hctx_idx >= q->nr_hw_queues) | |
307 | return ERR_PTR(-EIO); | |
308 | ||
309 | ret = blk_queue_enter(q, true); | |
310 | if (ret) | |
311 | return ERR_PTR(ret); | |
312 | ||
c8712c6a CH |
313 | /* |
314 | * Check if the hardware context is actually mapped to anything. | |
315 | * If not tell the caller that it should skip this queue. | |
316 | */ | |
6d2809d5 OS |
317 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
318 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
319 | blk_queue_exit(q); | |
320 | return ERR_PTR(-EXDEV); | |
c8712c6a | 321 | } |
6d2809d5 OS |
322 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
323 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 324 | |
6d2809d5 | 325 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
c8712c6a | 326 | |
6d2809d5 | 327 | blk_mq_put_ctx(alloc_data.ctx); |
c8712c6a | 328 | blk_queue_exit(q); |
6d2809d5 OS |
329 | |
330 | if (!rq) | |
331 | return ERR_PTR(-EWOULDBLOCK); | |
332 | ||
333 | return rq; | |
1f5bd336 ML |
334 | } |
335 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
336 | ||
bd166ef1 JA |
337 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
338 | struct request *rq) | |
320ae51f | 339 | { |
bd166ef1 | 340 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
341 | struct request_queue *q = rq->q; |
342 | ||
e8064021 | 343 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 344 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
345 | |
346 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 347 | rq->rq_flags = 0; |
0d2602ca | 348 | |
af76e555 | 349 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 350 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
351 | if (rq->tag != -1) |
352 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
353 | if (sched_tag != -1) | |
354 | blk_mq_sched_completed_request(hctx, rq); | |
50e1dab8 | 355 | blk_mq_sched_restart_queues(hctx); |
3ef28e83 | 356 | blk_queue_exit(q); |
320ae51f JA |
357 | } |
358 | ||
bd166ef1 | 359 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 360 | struct request *rq) |
320ae51f JA |
361 | { |
362 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
363 | |
364 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
365 | __blk_mq_finish_request(hctx, ctx, rq); |
366 | } | |
367 | ||
368 | void blk_mq_finish_request(struct request *rq) | |
369 | { | |
370 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 371 | } |
7c7f2f2b JA |
372 | |
373 | void blk_mq_free_request(struct request *rq) | |
374 | { | |
bd166ef1 | 375 | blk_mq_sched_put_request(rq); |
320ae51f | 376 | } |
1a3b595a | 377 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 378 | |
c8a446ad | 379 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 380 | { |
0d11e6ac ML |
381 | blk_account_io_done(rq); |
382 | ||
91b63639 | 383 | if (rq->end_io) { |
87760e5e | 384 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 385 | rq->end_io(rq, error); |
91b63639 CH |
386 | } else { |
387 | if (unlikely(blk_bidi_rq(rq))) | |
388 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 389 | blk_mq_free_request(rq); |
91b63639 | 390 | } |
320ae51f | 391 | } |
c8a446ad | 392 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 393 | |
c8a446ad | 394 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
395 | { |
396 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
397 | BUG(); | |
c8a446ad | 398 | __blk_mq_end_request(rq, error); |
63151a44 | 399 | } |
c8a446ad | 400 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 401 | |
30a91cb4 | 402 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 403 | { |
3d6efbf6 | 404 | struct request *rq = data; |
320ae51f | 405 | |
30a91cb4 | 406 | rq->q->softirq_done_fn(rq); |
320ae51f | 407 | } |
320ae51f | 408 | |
ed851860 | 409 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
410 | { |
411 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 412 | bool shared = false; |
320ae51f JA |
413 | int cpu; |
414 | ||
38535201 | 415 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
416 | rq->q->softirq_done_fn(rq); |
417 | return; | |
418 | } | |
320ae51f JA |
419 | |
420 | cpu = get_cpu(); | |
38535201 CH |
421 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
422 | shared = cpus_share_cache(cpu, ctx->cpu); | |
423 | ||
424 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 425 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
426 | rq->csd.info = rq; |
427 | rq->csd.flags = 0; | |
c46fff2a | 428 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 429 | } else { |
30a91cb4 | 430 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 431 | } |
320ae51f JA |
432 | put_cpu(); |
433 | } | |
30a91cb4 | 434 | |
cf43e6be JA |
435 | static void blk_mq_stat_add(struct request *rq) |
436 | { | |
437 | if (rq->rq_flags & RQF_STATS) { | |
34dbad5d OS |
438 | blk_mq_poll_stats_start(rq->q); |
439 | blk_stat_add(rq); | |
cf43e6be JA |
440 | } |
441 | } | |
442 | ||
1fa8cc52 | 443 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
444 | { |
445 | struct request_queue *q = rq->q; | |
446 | ||
cf43e6be JA |
447 | blk_mq_stat_add(rq); |
448 | ||
ed851860 | 449 | if (!q->softirq_done_fn) |
c8a446ad | 450 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
451 | else |
452 | blk_mq_ipi_complete_request(rq); | |
453 | } | |
454 | ||
30a91cb4 CH |
455 | /** |
456 | * blk_mq_complete_request - end I/O on a request | |
457 | * @rq: the request being processed | |
458 | * | |
459 | * Description: | |
460 | * Ends all I/O on a request. It does not handle partial completions. | |
461 | * The actual completion happens out-of-order, through a IPI handler. | |
462 | **/ | |
f4829a9b | 463 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 464 | { |
95f09684 JA |
465 | struct request_queue *q = rq->q; |
466 | ||
467 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 468 | return; |
f4829a9b CH |
469 | if (!blk_mark_rq_complete(rq)) { |
470 | rq->errors = error; | |
ed851860 | 471 | __blk_mq_complete_request(rq); |
f4829a9b | 472 | } |
30a91cb4 CH |
473 | } |
474 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 475 | |
973c0191 KB |
476 | int blk_mq_request_started(struct request *rq) |
477 | { | |
478 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
479 | } | |
480 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
481 | ||
e2490073 | 482 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
483 | { |
484 | struct request_queue *q = rq->q; | |
485 | ||
bd166ef1 JA |
486 | blk_mq_sched_started_request(rq); |
487 | ||
320ae51f JA |
488 | trace_block_rq_issue(q, rq); |
489 | ||
cf43e6be | 490 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 491 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 492 | rq->rq_flags |= RQF_STATS; |
87760e5e | 493 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
494 | } |
495 | ||
2b8393b4 | 496 | blk_add_timer(rq); |
87ee7b11 | 497 | |
538b7534 JA |
498 | /* |
499 | * Ensure that ->deadline is visible before set the started | |
500 | * flag and clear the completed flag. | |
501 | */ | |
502 | smp_mb__before_atomic(); | |
503 | ||
87ee7b11 JA |
504 | /* |
505 | * Mark us as started and clear complete. Complete might have been | |
506 | * set if requeue raced with timeout, which then marked it as | |
507 | * complete. So be sure to clear complete again when we start | |
508 | * the request, otherwise we'll ignore the completion event. | |
509 | */ | |
4b570521 JA |
510 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
511 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
512 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
513 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
514 | |
515 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
516 | /* | |
517 | * Make sure space for the drain appears. We know we can do | |
518 | * this because max_hw_segments has been adjusted to be one | |
519 | * fewer than the device can handle. | |
520 | */ | |
521 | rq->nr_phys_segments++; | |
522 | } | |
320ae51f | 523 | } |
e2490073 | 524 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 525 | |
d9d149a3 ML |
526 | /* |
527 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
528 | * flag isn't set yet, so there may be race with timeout hanlder, | |
529 | * but given rq->deadline is just set in .queue_rq() under | |
530 | * this situation, the race won't be possible in reality because | |
531 | * rq->timeout should be set as big enough to cover the window | |
532 | * between blk_mq_start_request() called from .queue_rq() and | |
533 | * clearing REQ_ATOM_STARTED here. | |
534 | */ | |
ed0791b2 | 535 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
536 | { |
537 | struct request_queue *q = rq->q; | |
538 | ||
539 | trace_block_rq_requeue(q, rq); | |
87760e5e | 540 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 541 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 542 | |
e2490073 CH |
543 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
544 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
545 | rq->nr_phys_segments--; | |
546 | } | |
320ae51f JA |
547 | } |
548 | ||
2b053aca | 549 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 550 | { |
ed0791b2 | 551 | __blk_mq_requeue_request(rq); |
ed0791b2 | 552 | |
ed0791b2 | 553 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 554 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
555 | } |
556 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
557 | ||
6fca6a61 CH |
558 | static void blk_mq_requeue_work(struct work_struct *work) |
559 | { | |
560 | struct request_queue *q = | |
2849450a | 561 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
562 | LIST_HEAD(rq_list); |
563 | struct request *rq, *next; | |
564 | unsigned long flags; | |
565 | ||
566 | spin_lock_irqsave(&q->requeue_lock, flags); | |
567 | list_splice_init(&q->requeue_list, &rq_list); | |
568 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
569 | ||
570 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 571 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
572 | continue; |
573 | ||
e8064021 | 574 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 575 | list_del_init(&rq->queuelist); |
bd6737f1 | 576 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
577 | } |
578 | ||
579 | while (!list_empty(&rq_list)) { | |
580 | rq = list_entry(rq_list.next, struct request, queuelist); | |
581 | list_del_init(&rq->queuelist); | |
bd6737f1 | 582 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
583 | } |
584 | ||
52d7f1b5 | 585 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
586 | } |
587 | ||
2b053aca BVA |
588 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
589 | bool kick_requeue_list) | |
6fca6a61 CH |
590 | { |
591 | struct request_queue *q = rq->q; | |
592 | unsigned long flags; | |
593 | ||
594 | /* | |
595 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
596 | * request head insertation from the workqueue. | |
597 | */ | |
e8064021 | 598 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
599 | |
600 | spin_lock_irqsave(&q->requeue_lock, flags); | |
601 | if (at_head) { | |
e8064021 | 602 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
603 | list_add(&rq->queuelist, &q->requeue_list); |
604 | } else { | |
605 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
606 | } | |
607 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
608 | |
609 | if (kick_requeue_list) | |
610 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
611 | } |
612 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
613 | ||
614 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
615 | { | |
2849450a | 616 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
617 | } |
618 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
619 | ||
2849450a MS |
620 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
621 | unsigned long msecs) | |
622 | { | |
623 | kblockd_schedule_delayed_work(&q->requeue_work, | |
624 | msecs_to_jiffies(msecs)); | |
625 | } | |
626 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
627 | ||
1885b24d JA |
628 | void blk_mq_abort_requeue_list(struct request_queue *q) |
629 | { | |
630 | unsigned long flags; | |
631 | LIST_HEAD(rq_list); | |
632 | ||
633 | spin_lock_irqsave(&q->requeue_lock, flags); | |
634 | list_splice_init(&q->requeue_list, &rq_list); | |
635 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
636 | ||
637 | while (!list_empty(&rq_list)) { | |
638 | struct request *rq; | |
639 | ||
640 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
641 | list_del_init(&rq->queuelist); | |
642 | rq->errors = -EIO; | |
643 | blk_mq_end_request(rq, rq->errors); | |
644 | } | |
645 | } | |
646 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
647 | ||
0e62f51f JA |
648 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
649 | { | |
88c7b2b7 JA |
650 | if (tag < tags->nr_tags) { |
651 | prefetch(tags->rqs[tag]); | |
4ee86bab | 652 | return tags->rqs[tag]; |
88c7b2b7 | 653 | } |
4ee86bab HR |
654 | |
655 | return NULL; | |
24d2f903 CH |
656 | } |
657 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
658 | ||
320ae51f | 659 | struct blk_mq_timeout_data { |
46f92d42 CH |
660 | unsigned long next; |
661 | unsigned int next_set; | |
320ae51f JA |
662 | }; |
663 | ||
90415837 | 664 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 665 | { |
f8a5b122 | 666 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 667 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
668 | |
669 | /* | |
670 | * We know that complete is set at this point. If STARTED isn't set | |
671 | * anymore, then the request isn't active and the "timeout" should | |
672 | * just be ignored. This can happen due to the bitflag ordering. | |
673 | * Timeout first checks if STARTED is set, and if it is, assumes | |
674 | * the request is active. But if we race with completion, then | |
675 | * we both flags will get cleared. So check here again, and ignore | |
676 | * a timeout event with a request that isn't active. | |
677 | */ | |
46f92d42 CH |
678 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
679 | return; | |
87ee7b11 | 680 | |
46f92d42 | 681 | if (ops->timeout) |
0152fb6b | 682 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
683 | |
684 | switch (ret) { | |
685 | case BLK_EH_HANDLED: | |
686 | __blk_mq_complete_request(req); | |
687 | break; | |
688 | case BLK_EH_RESET_TIMER: | |
689 | blk_add_timer(req); | |
690 | blk_clear_rq_complete(req); | |
691 | break; | |
692 | case BLK_EH_NOT_HANDLED: | |
693 | break; | |
694 | default: | |
695 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
696 | break; | |
697 | } | |
87ee7b11 | 698 | } |
5b3f25fc | 699 | |
81481eb4 CH |
700 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
701 | struct request *rq, void *priv, bool reserved) | |
702 | { | |
703 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 704 | |
a4ef8e56 | 705 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 706 | return; |
87ee7b11 | 707 | |
d9d149a3 ML |
708 | /* |
709 | * The rq being checked may have been freed and reallocated | |
710 | * out already here, we avoid this race by checking rq->deadline | |
711 | * and REQ_ATOM_COMPLETE flag together: | |
712 | * | |
713 | * - if rq->deadline is observed as new value because of | |
714 | * reusing, the rq won't be timed out because of timing. | |
715 | * - if rq->deadline is observed as previous value, | |
716 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
717 | * because we put a barrier between setting rq->deadline | |
718 | * and clearing the flag in blk_mq_start_request(), so | |
719 | * this rq won't be timed out too. | |
720 | */ | |
46f92d42 CH |
721 | if (time_after_eq(jiffies, rq->deadline)) { |
722 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 723 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
724 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
725 | data->next = rq->deadline; | |
726 | data->next_set = 1; | |
727 | } | |
87ee7b11 JA |
728 | } |
729 | ||
287922eb | 730 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 731 | { |
287922eb CH |
732 | struct request_queue *q = |
733 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
734 | struct blk_mq_timeout_data data = { |
735 | .next = 0, | |
736 | .next_set = 0, | |
737 | }; | |
81481eb4 | 738 | int i; |
320ae51f | 739 | |
71f79fb3 GKB |
740 | /* A deadlock might occur if a request is stuck requiring a |
741 | * timeout at the same time a queue freeze is waiting | |
742 | * completion, since the timeout code would not be able to | |
743 | * acquire the queue reference here. | |
744 | * | |
745 | * That's why we don't use blk_queue_enter here; instead, we use | |
746 | * percpu_ref_tryget directly, because we need to be able to | |
747 | * obtain a reference even in the short window between the queue | |
748 | * starting to freeze, by dropping the first reference in | |
749 | * blk_mq_freeze_queue_start, and the moment the last request is | |
750 | * consumed, marked by the instant q_usage_counter reaches | |
751 | * zero. | |
752 | */ | |
753 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
754 | return; |
755 | ||
0bf6cd5b | 756 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 757 | |
81481eb4 CH |
758 | if (data.next_set) { |
759 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
760 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 761 | } else { |
0bf6cd5b CH |
762 | struct blk_mq_hw_ctx *hctx; |
763 | ||
f054b56c ML |
764 | queue_for_each_hw_ctx(q, hctx, i) { |
765 | /* the hctx may be unmapped, so check it here */ | |
766 | if (blk_mq_hw_queue_mapped(hctx)) | |
767 | blk_mq_tag_idle(hctx); | |
768 | } | |
0d2602ca | 769 | } |
287922eb | 770 | blk_queue_exit(q); |
320ae51f JA |
771 | } |
772 | ||
773 | /* | |
774 | * Reverse check our software queue for entries that we could potentially | |
775 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
776 | * too much time checking for merges. | |
777 | */ | |
778 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
779 | struct blk_mq_ctx *ctx, struct bio *bio) | |
780 | { | |
781 | struct request *rq; | |
782 | int checked = 8; | |
783 | ||
784 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 785 | bool merged = false; |
320ae51f JA |
786 | |
787 | if (!checked--) | |
788 | break; | |
789 | ||
790 | if (!blk_rq_merge_ok(rq, bio)) | |
791 | continue; | |
792 | ||
34fe7c05 CH |
793 | switch (blk_try_merge(rq, bio)) { |
794 | case ELEVATOR_BACK_MERGE: | |
795 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
796 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 797 | break; |
34fe7c05 CH |
798 | case ELEVATOR_FRONT_MERGE: |
799 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
800 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 801 | break; |
1e739730 CH |
802 | case ELEVATOR_DISCARD_MERGE: |
803 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 804 | break; |
34fe7c05 CH |
805 | default: |
806 | continue; | |
320ae51f | 807 | } |
34fe7c05 CH |
808 | |
809 | if (merged) | |
810 | ctx->rq_merged++; | |
811 | return merged; | |
320ae51f JA |
812 | } |
813 | ||
814 | return false; | |
815 | } | |
816 | ||
88459642 OS |
817 | struct flush_busy_ctx_data { |
818 | struct blk_mq_hw_ctx *hctx; | |
819 | struct list_head *list; | |
820 | }; | |
821 | ||
822 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
823 | { | |
824 | struct flush_busy_ctx_data *flush_data = data; | |
825 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
826 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
827 | ||
828 | sbitmap_clear_bit(sb, bitnr); | |
829 | spin_lock(&ctx->lock); | |
830 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
831 | spin_unlock(&ctx->lock); | |
832 | return true; | |
833 | } | |
834 | ||
1429d7c9 JA |
835 | /* |
836 | * Process software queues that have been marked busy, splicing them | |
837 | * to the for-dispatch | |
838 | */ | |
2c3ad667 | 839 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 840 | { |
88459642 OS |
841 | struct flush_busy_ctx_data data = { |
842 | .hctx = hctx, | |
843 | .list = list, | |
844 | }; | |
1429d7c9 | 845 | |
88459642 | 846 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 847 | } |
2c3ad667 | 848 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 849 | |
703fd1c0 JA |
850 | static inline unsigned int queued_to_index(unsigned int queued) |
851 | { | |
852 | if (!queued) | |
853 | return 0; | |
1429d7c9 | 854 | |
703fd1c0 | 855 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
856 | } |
857 | ||
bd6737f1 JA |
858 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
859 | bool wait) | |
bd166ef1 JA |
860 | { |
861 | struct blk_mq_alloc_data data = { | |
862 | .q = rq->q, | |
bd166ef1 JA |
863 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
864 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
865 | }; | |
866 | ||
bd166ef1 JA |
867 | if (rq->tag != -1) { |
868 | done: | |
869 | if (hctx) | |
870 | *hctx = data.hctx; | |
871 | return true; | |
872 | } | |
873 | ||
415b806d SG |
874 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
875 | data.flags |= BLK_MQ_REQ_RESERVED; | |
876 | ||
bd166ef1 JA |
877 | rq->tag = blk_mq_get_tag(&data); |
878 | if (rq->tag >= 0) { | |
200e86b3 JA |
879 | if (blk_mq_tag_busy(data.hctx)) { |
880 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
881 | atomic_inc(&data.hctx->nr_active); | |
882 | } | |
bd166ef1 JA |
883 | data.hctx->tags->rqs[rq->tag] = rq; |
884 | goto done; | |
885 | } | |
886 | ||
887 | return false; | |
888 | } | |
889 | ||
113285b4 JA |
890 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
891 | struct request *rq) | |
99cf1dc5 | 892 | { |
99cf1dc5 JA |
893 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
894 | rq->tag = -1; | |
895 | ||
896 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
897 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
898 | atomic_dec(&hctx->nr_active); | |
899 | } | |
900 | } | |
901 | ||
113285b4 JA |
902 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
903 | struct request *rq) | |
904 | { | |
905 | if (rq->tag == -1 || rq->internal_tag == -1) | |
906 | return; | |
907 | ||
908 | __blk_mq_put_driver_tag(hctx, rq); | |
909 | } | |
910 | ||
911 | static void blk_mq_put_driver_tag(struct request *rq) | |
912 | { | |
913 | struct blk_mq_hw_ctx *hctx; | |
914 | ||
915 | if (rq->tag == -1 || rq->internal_tag == -1) | |
916 | return; | |
917 | ||
918 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
919 | __blk_mq_put_driver_tag(hctx, rq); | |
920 | } | |
921 | ||
bd166ef1 JA |
922 | /* |
923 | * If we fail getting a driver tag because all the driver tags are already | |
924 | * assigned and on the dispatch list, BUT the first entry does not have a | |
925 | * tag, then we could deadlock. For that case, move entries with assigned | |
926 | * driver tags to the front, leaving the set of tagged requests in the | |
927 | * same order, and the untagged set in the same order. | |
928 | */ | |
929 | static bool reorder_tags_to_front(struct list_head *list) | |
930 | { | |
931 | struct request *rq, *tmp, *first = NULL; | |
932 | ||
933 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
934 | if (rq == first) | |
935 | break; | |
936 | if (rq->tag != -1) { | |
937 | list_move(&rq->queuelist, list); | |
938 | if (!first) | |
939 | first = rq; | |
940 | } | |
941 | } | |
942 | ||
943 | return first != NULL; | |
944 | } | |
945 | ||
da55f2cc OS |
946 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
947 | void *key) | |
948 | { | |
949 | struct blk_mq_hw_ctx *hctx; | |
950 | ||
951 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
952 | ||
953 | list_del(&wait->task_list); | |
954 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
955 | blk_mq_run_hw_queue(hctx, true); | |
956 | return 1; | |
957 | } | |
958 | ||
959 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
960 | { | |
961 | struct sbq_wait_state *ws; | |
962 | ||
963 | /* | |
964 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
965 | * The thread which wins the race to grab this bit adds the hardware | |
966 | * queue to the wait queue. | |
967 | */ | |
968 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
969 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
970 | return false; | |
971 | ||
972 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
973 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
974 | ||
975 | /* | |
976 | * As soon as this returns, it's no longer safe to fiddle with | |
977 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
978 | * and unlock the bit. | |
979 | */ | |
980 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
981 | return true; | |
982 | } | |
983 | ||
f04c3df3 | 984 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
985 | { |
986 | struct request_queue *q = hctx->queue; | |
320ae51f | 987 | struct request *rq; |
74c45052 JA |
988 | LIST_HEAD(driver_list); |
989 | struct list_head *dptr; | |
f04c3df3 | 990 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 991 | |
74c45052 JA |
992 | /* |
993 | * Start off with dptr being NULL, so we start the first request | |
994 | * immediately, even if we have more pending. | |
995 | */ | |
996 | dptr = NULL; | |
997 | ||
320ae51f JA |
998 | /* |
999 | * Now process all the entries, sending them to the driver. | |
1000 | */ | |
1429d7c9 | 1001 | queued = 0; |
f04c3df3 | 1002 | while (!list_empty(list)) { |
74c45052 | 1003 | struct blk_mq_queue_data bd; |
320ae51f | 1004 | |
f04c3df3 | 1005 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1006 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1007 | if (!queued && reorder_tags_to_front(list)) | |
1008 | continue; | |
3c782d67 JA |
1009 | |
1010 | /* | |
da55f2cc OS |
1011 | * The initial allocation attempt failed, so we need to |
1012 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1013 | */ |
da55f2cc OS |
1014 | if (blk_mq_dispatch_wait_add(hctx)) { |
1015 | /* | |
1016 | * It's possible that a tag was freed in the | |
1017 | * window between the allocation failure and | |
1018 | * adding the hardware queue to the wait queue. | |
1019 | */ | |
1020 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
1021 | break; | |
1022 | } else { | |
3c782d67 | 1023 | break; |
da55f2cc | 1024 | } |
bd166ef1 | 1025 | } |
da55f2cc | 1026 | |
320ae51f | 1027 | list_del_init(&rq->queuelist); |
320ae51f | 1028 | |
74c45052 JA |
1029 | bd.rq = rq; |
1030 | bd.list = dptr; | |
113285b4 JA |
1031 | |
1032 | /* | |
1033 | * Flag last if we have no more requests, or if we have more | |
1034 | * but can't assign a driver tag to it. | |
1035 | */ | |
1036 | if (list_empty(list)) | |
1037 | bd.last = true; | |
1038 | else { | |
1039 | struct request *nxt; | |
1040 | ||
1041 | nxt = list_first_entry(list, struct request, queuelist); | |
1042 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1043 | } | |
74c45052 JA |
1044 | |
1045 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
1046 | switch (ret) { |
1047 | case BLK_MQ_RQ_QUEUE_OK: | |
1048 | queued++; | |
52b9c330 | 1049 | break; |
320ae51f | 1050 | case BLK_MQ_RQ_QUEUE_BUSY: |
113285b4 | 1051 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1052 | list_add(&rq->queuelist, list); |
ed0791b2 | 1053 | __blk_mq_requeue_request(rq); |
320ae51f JA |
1054 | break; |
1055 | default: | |
1056 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 1057 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 1058 | rq->errors = -EIO; |
c8a446ad | 1059 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
1060 | break; |
1061 | } | |
1062 | ||
1063 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1064 | break; | |
74c45052 JA |
1065 | |
1066 | /* | |
1067 | * We've done the first request. If we have more than 1 | |
1068 | * left in the list, set dptr to defer issue. | |
1069 | */ | |
f04c3df3 | 1070 | if (!dptr && list->next != list->prev) |
74c45052 | 1071 | dptr = &driver_list; |
320ae51f JA |
1072 | } |
1073 | ||
703fd1c0 | 1074 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1075 | |
1076 | /* | |
1077 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1078 | * that is where we will continue on next queue run. | |
1079 | */ | |
f04c3df3 | 1080 | if (!list_empty(list)) { |
113285b4 JA |
1081 | /* |
1082 | * If we got a driver tag for the next request already, | |
1083 | * free it again. | |
1084 | */ | |
1085 | rq = list_first_entry(list, struct request, queuelist); | |
1086 | blk_mq_put_driver_tag(rq); | |
1087 | ||
320ae51f | 1088 | spin_lock(&hctx->lock); |
c13660a0 | 1089 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1090 | spin_unlock(&hctx->lock); |
f04c3df3 | 1091 | |
9ba52e58 SL |
1092 | /* |
1093 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
1094 | * it's possible the queue is stopped and restarted again | |
1095 | * before this. Queue restart will dispatch requests. And since | |
1096 | * requests in rq_list aren't added into hctx->dispatch yet, | |
1097 | * the requests in rq_list might get lost. | |
1098 | * | |
1099 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 | 1100 | * |
da55f2cc OS |
1101 | * If RESTART or TAG_WAITING is set, then let completion restart |
1102 | * the queue instead of potentially looping here. | |
bd166ef1 | 1103 | */ |
da55f2cc OS |
1104 | if (!blk_mq_sched_needs_restart(hctx) && |
1105 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1106 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1107 | } |
f04c3df3 | 1108 | |
2aa0f21d | 1109 | return queued != 0; |
f04c3df3 JA |
1110 | } |
1111 | ||
6a83e74d BVA |
1112 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1113 | { | |
1114 | int srcu_idx; | |
1115 | ||
1116 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1117 | cpu_online(hctx->next_cpu)); | |
1118 | ||
1119 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1120 | rcu_read_lock(); | |
bd166ef1 | 1121 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1122 | rcu_read_unlock(); |
1123 | } else { | |
1124 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1125 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1126 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1127 | } | |
1128 | } | |
1129 | ||
506e931f JA |
1130 | /* |
1131 | * It'd be great if the workqueue API had a way to pass | |
1132 | * in a mask and had some smarts for more clever placement. | |
1133 | * For now we just round-robin here, switching for every | |
1134 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1135 | */ | |
1136 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1137 | { | |
b657d7e6 CH |
1138 | if (hctx->queue->nr_hw_queues == 1) |
1139 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1140 | |
1141 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1142 | int next_cpu; |
506e931f JA |
1143 | |
1144 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1145 | if (next_cpu >= nr_cpu_ids) | |
1146 | next_cpu = cpumask_first(hctx->cpumask); | |
1147 | ||
1148 | hctx->next_cpu = next_cpu; | |
1149 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1150 | } | |
1151 | ||
b657d7e6 | 1152 | return hctx->next_cpu; |
506e931f JA |
1153 | } |
1154 | ||
320ae51f JA |
1155 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1156 | { | |
5d1b25c1 BVA |
1157 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1158 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1159 | return; |
1160 | ||
1b792f2f | 1161 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1162 | int cpu = get_cpu(); |
1163 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1164 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1165 | put_cpu(); |
398205b8 PB |
1166 | return; |
1167 | } | |
e4043dcf | 1168 | |
2a90d4aa | 1169 | put_cpu(); |
e4043dcf | 1170 | } |
398205b8 | 1171 | |
27489a3c | 1172 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1173 | } |
1174 | ||
b94ec296 | 1175 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1176 | { |
1177 | struct blk_mq_hw_ctx *hctx; | |
1178 | int i; | |
1179 | ||
1180 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1181 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1182 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1183 | continue; |
1184 | ||
b94ec296 | 1185 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1186 | } |
1187 | } | |
b94ec296 | 1188 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1189 | |
fd001443 BVA |
1190 | /** |
1191 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1192 | * @q: request queue. | |
1193 | * | |
1194 | * The caller is responsible for serializing this function against | |
1195 | * blk_mq_{start,stop}_hw_queue(). | |
1196 | */ | |
1197 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1198 | { | |
1199 | struct blk_mq_hw_ctx *hctx; | |
1200 | int i; | |
1201 | ||
1202 | queue_for_each_hw_ctx(q, hctx, i) | |
1203 | if (blk_mq_hctx_stopped(hctx)) | |
1204 | return true; | |
1205 | ||
1206 | return false; | |
1207 | } | |
1208 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1209 | ||
320ae51f JA |
1210 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1211 | { | |
27489a3c | 1212 | cancel_work(&hctx->run_work); |
70f4db63 | 1213 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1214 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1215 | } | |
1216 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1217 | ||
280d45f6 CH |
1218 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1219 | { | |
1220 | struct blk_mq_hw_ctx *hctx; | |
1221 | int i; | |
1222 | ||
1223 | queue_for_each_hw_ctx(q, hctx, i) | |
1224 | blk_mq_stop_hw_queue(hctx); | |
1225 | } | |
1226 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1227 | ||
320ae51f JA |
1228 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1229 | { | |
1230 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1231 | |
0ffbce80 | 1232 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1233 | } |
1234 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1235 | ||
2f268556 CH |
1236 | void blk_mq_start_hw_queues(struct request_queue *q) |
1237 | { | |
1238 | struct blk_mq_hw_ctx *hctx; | |
1239 | int i; | |
1240 | ||
1241 | queue_for_each_hw_ctx(q, hctx, i) | |
1242 | blk_mq_start_hw_queue(hctx); | |
1243 | } | |
1244 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1245 | ||
ae911c5e JA |
1246 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1247 | { | |
1248 | if (!blk_mq_hctx_stopped(hctx)) | |
1249 | return; | |
1250 | ||
1251 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1252 | blk_mq_run_hw_queue(hctx, async); | |
1253 | } | |
1254 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1255 | ||
1b4a3258 | 1256 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1257 | { |
1258 | struct blk_mq_hw_ctx *hctx; | |
1259 | int i; | |
1260 | ||
ae911c5e JA |
1261 | queue_for_each_hw_ctx(q, hctx, i) |
1262 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1263 | } |
1264 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1265 | ||
70f4db63 | 1266 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1267 | { |
1268 | struct blk_mq_hw_ctx *hctx; | |
1269 | ||
27489a3c | 1270 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1271 | |
320ae51f JA |
1272 | __blk_mq_run_hw_queue(hctx); |
1273 | } | |
1274 | ||
70f4db63 CH |
1275 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1276 | { | |
1277 | struct blk_mq_hw_ctx *hctx; | |
1278 | ||
1279 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1280 | ||
1281 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1282 | __blk_mq_run_hw_queue(hctx); | |
1283 | } | |
1284 | ||
1285 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1286 | { | |
19c66e59 ML |
1287 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1288 | return; | |
70f4db63 | 1289 | |
7e79dadc | 1290 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1291 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1292 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1293 | } |
1294 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1295 | ||
cfd0c552 | 1296 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1297 | struct request *rq, |
1298 | bool at_head) | |
320ae51f | 1299 | { |
e57690fe JA |
1300 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1301 | ||
01b983c9 JA |
1302 | trace_block_rq_insert(hctx->queue, rq); |
1303 | ||
72a0a36e CH |
1304 | if (at_head) |
1305 | list_add(&rq->queuelist, &ctx->rq_list); | |
1306 | else | |
1307 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1308 | } |
4bb659b1 | 1309 | |
2c3ad667 JA |
1310 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1311 | bool at_head) | |
cfd0c552 ML |
1312 | { |
1313 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1314 | ||
e57690fe | 1315 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1316 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1317 | } |
1318 | ||
bd166ef1 JA |
1319 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1320 | struct list_head *list) | |
320ae51f JA |
1321 | |
1322 | { | |
320ae51f JA |
1323 | /* |
1324 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1325 | * offline now | |
1326 | */ | |
1327 | spin_lock(&ctx->lock); | |
1328 | while (!list_empty(list)) { | |
1329 | struct request *rq; | |
1330 | ||
1331 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1332 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1333 | list_del_init(&rq->queuelist); |
e57690fe | 1334 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1335 | } |
cfd0c552 | 1336 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1337 | spin_unlock(&ctx->lock); |
320ae51f JA |
1338 | } |
1339 | ||
1340 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1341 | { | |
1342 | struct request *rqa = container_of(a, struct request, queuelist); | |
1343 | struct request *rqb = container_of(b, struct request, queuelist); | |
1344 | ||
1345 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1346 | (rqa->mq_ctx == rqb->mq_ctx && | |
1347 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1348 | } | |
1349 | ||
1350 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1351 | { | |
1352 | struct blk_mq_ctx *this_ctx; | |
1353 | struct request_queue *this_q; | |
1354 | struct request *rq; | |
1355 | LIST_HEAD(list); | |
1356 | LIST_HEAD(ctx_list); | |
1357 | unsigned int depth; | |
1358 | ||
1359 | list_splice_init(&plug->mq_list, &list); | |
1360 | ||
1361 | list_sort(NULL, &list, plug_ctx_cmp); | |
1362 | ||
1363 | this_q = NULL; | |
1364 | this_ctx = NULL; | |
1365 | depth = 0; | |
1366 | ||
1367 | while (!list_empty(&list)) { | |
1368 | rq = list_entry_rq(list.next); | |
1369 | list_del_init(&rq->queuelist); | |
1370 | BUG_ON(!rq->q); | |
1371 | if (rq->mq_ctx != this_ctx) { | |
1372 | if (this_ctx) { | |
bd166ef1 JA |
1373 | trace_block_unplug(this_q, depth, from_schedule); |
1374 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1375 | &ctx_list, | |
1376 | from_schedule); | |
320ae51f JA |
1377 | } |
1378 | ||
1379 | this_ctx = rq->mq_ctx; | |
1380 | this_q = rq->q; | |
1381 | depth = 0; | |
1382 | } | |
1383 | ||
1384 | depth++; | |
1385 | list_add_tail(&rq->queuelist, &ctx_list); | |
1386 | } | |
1387 | ||
1388 | /* | |
1389 | * If 'this_ctx' is set, we know we have entries to complete | |
1390 | * on 'ctx_list'. Do those. | |
1391 | */ | |
1392 | if (this_ctx) { | |
bd166ef1 JA |
1393 | trace_block_unplug(this_q, depth, from_schedule); |
1394 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1395 | from_schedule); | |
320ae51f JA |
1396 | } |
1397 | } | |
1398 | ||
1399 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1400 | { | |
1401 | init_request_from_bio(rq, bio); | |
4b570521 | 1402 | |
6e85eaf3 | 1403 | blk_account_io_start(rq, true); |
320ae51f JA |
1404 | } |
1405 | ||
274a5843 JA |
1406 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1407 | { | |
1408 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1409 | !blk_queue_nomerges(hctx->queue); | |
1410 | } | |
1411 | ||
07068d5b JA |
1412 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1413 | struct blk_mq_ctx *ctx, | |
1414 | struct request *rq, struct bio *bio) | |
320ae51f | 1415 | { |
e18378a6 | 1416 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1417 | blk_mq_bio_to_request(rq, bio); |
1418 | spin_lock(&ctx->lock); | |
1419 | insert_rq: | |
1420 | __blk_mq_insert_request(hctx, rq, false); | |
1421 | spin_unlock(&ctx->lock); | |
1422 | return false; | |
1423 | } else { | |
274a5843 JA |
1424 | struct request_queue *q = hctx->queue; |
1425 | ||
07068d5b JA |
1426 | spin_lock(&ctx->lock); |
1427 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1428 | blk_mq_bio_to_request(rq, bio); | |
1429 | goto insert_rq; | |
1430 | } | |
320ae51f | 1431 | |
07068d5b | 1432 | spin_unlock(&ctx->lock); |
bd166ef1 | 1433 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1434 | return true; |
14ec77f3 | 1435 | } |
07068d5b | 1436 | } |
14ec77f3 | 1437 | |
fd2d3326 JA |
1438 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1439 | { | |
bd166ef1 JA |
1440 | if (rq->tag != -1) |
1441 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1442 | ||
1443 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1444 | } |
1445 | ||
5eb6126e | 1446 | static void __blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie, |
9c621104 | 1447 | bool may_sleep) |
f984df1f | 1448 | { |
f984df1f | 1449 | struct request_queue *q = rq->q; |
f984df1f SL |
1450 | struct blk_mq_queue_data bd = { |
1451 | .rq = rq, | |
1452 | .list = NULL, | |
1453 | .last = 1 | |
1454 | }; | |
bd166ef1 JA |
1455 | struct blk_mq_hw_ctx *hctx; |
1456 | blk_qc_t new_cookie; | |
1457 | int ret; | |
f984df1f | 1458 | |
bd166ef1 | 1459 | if (q->elevator) |
2253efc8 BVA |
1460 | goto insert; |
1461 | ||
bd166ef1 JA |
1462 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1463 | goto insert; | |
1464 | ||
1465 | new_cookie = request_to_qc_t(hctx, rq); | |
1466 | ||
f984df1f SL |
1467 | /* |
1468 | * For OK queue, we are done. For error, kill it. Any other | |
1469 | * error (busy), just add it to our list as we previously | |
1470 | * would have done | |
1471 | */ | |
1472 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1473 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1474 | *cookie = new_cookie; | |
2253efc8 | 1475 | return; |
7b371636 | 1476 | } |
f984df1f | 1477 | |
7b371636 JA |
1478 | __blk_mq_requeue_request(rq); |
1479 | ||
1480 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1481 | *cookie = BLK_QC_T_NONE; | |
1482 | rq->errors = -EIO; | |
1483 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1484 | return; |
f984df1f | 1485 | } |
7b371636 | 1486 | |
2253efc8 | 1487 | insert: |
9c621104 | 1488 | blk_mq_sched_insert_request(rq, false, true, false, may_sleep); |
f984df1f SL |
1489 | } |
1490 | ||
5eb6126e CH |
1491 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1492 | struct request *rq, blk_qc_t *cookie) | |
1493 | { | |
1494 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1495 | rcu_read_lock(); | |
1496 | __blk_mq_try_issue_directly(rq, cookie, false); | |
1497 | rcu_read_unlock(); | |
1498 | } else { | |
1499 | unsigned int srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
1500 | __blk_mq_try_issue_directly(rq, cookie, true); | |
1501 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); | |
1502 | } | |
1503 | } | |
1504 | ||
dece1635 | 1505 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1506 | { |
ef295ecf | 1507 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1508 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1509 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1510 | struct request *rq; |
5eb6126e | 1511 | unsigned int request_count = 0; |
f984df1f | 1512 | struct blk_plug *plug; |
5b3f341f | 1513 | struct request *same_queue_rq = NULL; |
7b371636 | 1514 | blk_qc_t cookie; |
87760e5e | 1515 | unsigned int wb_acct; |
07068d5b JA |
1516 | |
1517 | blk_queue_bounce(q, &bio); | |
1518 | ||
1519 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1520 | bio_io_error(bio); |
dece1635 | 1521 | return BLK_QC_T_NONE; |
07068d5b JA |
1522 | } |
1523 | ||
54efd50b KO |
1524 | blk_queue_split(q, &bio, q->bio_split); |
1525 | ||
87c279e6 OS |
1526 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1527 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1528 | return BLK_QC_T_NONE; | |
f984df1f | 1529 | |
bd166ef1 JA |
1530 | if (blk_mq_sched_bio_merge(q, bio)) |
1531 | return BLK_QC_T_NONE; | |
1532 | ||
87760e5e JA |
1533 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1534 | ||
bd166ef1 JA |
1535 | trace_block_getrq(q, bio, bio->bi_opf); |
1536 | ||
1537 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1538 | if (unlikely(!rq)) { |
1539 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1540 | return BLK_QC_T_NONE; |
87760e5e JA |
1541 | } |
1542 | ||
1543 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1544 | |
fd2d3326 | 1545 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1546 | |
a4d907b6 | 1547 | plug = current->plug; |
07068d5b JA |
1548 | if (unlikely(is_flush_fua)) { |
1549 | blk_mq_bio_to_request(rq, bio); | |
a4d907b6 CH |
1550 | if (q->elevator) { |
1551 | blk_mq_sched_insert_request(rq, false, true, true, | |
1552 | true); | |
1553 | } else { | |
1554 | blk_insert_flush(rq); | |
1555 | blk_mq_run_hw_queue(data.hctx, true); | |
1556 | } | |
1557 | } else if (plug && q->nr_hw_queues == 1) { | |
254d259d CH |
1558 | struct request *last = NULL; |
1559 | ||
1560 | blk_mq_bio_to_request(rq, bio); | |
1561 | ||
1562 | /* | |
1563 | * @request_count may become stale because of schedule | |
1564 | * out, so check the list again. | |
1565 | */ | |
1566 | if (list_empty(&plug->mq_list)) | |
1567 | request_count = 0; | |
1568 | else if (blk_queue_nomerges(q)) | |
1569 | request_count = blk_plug_queued_count(q); | |
1570 | ||
1571 | if (!request_count) | |
1572 | trace_block_plug(q); | |
1573 | else | |
1574 | last = list_entry_rq(plug->mq_list.prev); | |
1575 | ||
254d259d CH |
1576 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1577 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
1578 | blk_flush_plug_list(plug, false); | |
1579 | trace_block_plug(q); | |
1580 | } | |
1581 | ||
1582 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2299722c | 1583 | } else if (plug && !blk_queue_nomerges(q)) { |
07068d5b | 1584 | blk_mq_bio_to_request(rq, bio); |
07068d5b JA |
1585 | |
1586 | /* | |
6a83e74d | 1587 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1588 | * Otherwise the existing request in the plug list will be |
1589 | * issued. So the plug list will have one request at most | |
2299722c CH |
1590 | * The plug list might get flushed before this. If that happens, |
1591 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1592 | */ |
2299722c CH |
1593 | if (list_empty(&plug->mq_list)) |
1594 | same_queue_rq = NULL; | |
1595 | if (same_queue_rq) | |
1596 | list_del_init(&same_queue_rq->queuelist); | |
1597 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1598 | ||
2299722c CH |
1599 | if (same_queue_rq) |
1600 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
1601 | &cookie); | |
a4d907b6 | 1602 | } else if (q->nr_hw_queues > 1 && is_sync) { |
2299722c | 1603 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1604 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1605 | } else if (q->elevator) { |
bd166ef1 | 1606 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1607 | blk_mq_sched_insert_request(rq, false, true, true, true); |
1608 | } else if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { | |
1609 | blk_mq_run_hw_queue(data.hctx, true); | |
07068d5b | 1610 | } |
a4d907b6 | 1611 | |
07068d5b | 1612 | blk_mq_put_ctx(data.ctx); |
7b371636 | 1613 | return cookie; |
07068d5b JA |
1614 | } |
1615 | ||
cc71a6f4 JA |
1616 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1617 | unsigned int hctx_idx) | |
95363efd | 1618 | { |
e9b267d9 | 1619 | struct page *page; |
320ae51f | 1620 | |
24d2f903 | 1621 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1622 | int i; |
320ae51f | 1623 | |
24d2f903 | 1624 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1625 | struct request *rq = tags->static_rqs[i]; |
1626 | ||
1627 | if (!rq) | |
e9b267d9 | 1628 | continue; |
2af8cbe3 | 1629 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1630 | hctx_idx, i); |
2af8cbe3 | 1631 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1632 | } |
320ae51f | 1633 | } |
320ae51f | 1634 | |
24d2f903 CH |
1635 | while (!list_empty(&tags->page_list)) { |
1636 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1637 | list_del_init(&page->lru); |
f75782e4 CM |
1638 | /* |
1639 | * Remove kmemleak object previously allocated in | |
1640 | * blk_mq_init_rq_map(). | |
1641 | */ | |
1642 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1643 | __free_pages(page, page->private); |
1644 | } | |
cc71a6f4 | 1645 | } |
320ae51f | 1646 | |
cc71a6f4 JA |
1647 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1648 | { | |
24d2f903 | 1649 | kfree(tags->rqs); |
cc71a6f4 | 1650 | tags->rqs = NULL; |
2af8cbe3 JA |
1651 | kfree(tags->static_rqs); |
1652 | tags->static_rqs = NULL; | |
320ae51f | 1653 | |
24d2f903 | 1654 | blk_mq_free_tags(tags); |
320ae51f JA |
1655 | } |
1656 | ||
cc71a6f4 JA |
1657 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1658 | unsigned int hctx_idx, | |
1659 | unsigned int nr_tags, | |
1660 | unsigned int reserved_tags) | |
320ae51f | 1661 | { |
24d2f903 | 1662 | struct blk_mq_tags *tags; |
59f082e4 | 1663 | int node; |
320ae51f | 1664 | |
59f082e4 SL |
1665 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1666 | if (node == NUMA_NO_NODE) | |
1667 | node = set->numa_node; | |
1668 | ||
1669 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1670 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1671 | if (!tags) |
1672 | return NULL; | |
320ae51f | 1673 | |
cc71a6f4 | 1674 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1675 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1676 | node); |
24d2f903 CH |
1677 | if (!tags->rqs) { |
1678 | blk_mq_free_tags(tags); | |
1679 | return NULL; | |
1680 | } | |
320ae51f | 1681 | |
2af8cbe3 JA |
1682 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1683 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1684 | node); |
2af8cbe3 JA |
1685 | if (!tags->static_rqs) { |
1686 | kfree(tags->rqs); | |
1687 | blk_mq_free_tags(tags); | |
1688 | return NULL; | |
1689 | } | |
1690 | ||
cc71a6f4 JA |
1691 | return tags; |
1692 | } | |
1693 | ||
1694 | static size_t order_to_size(unsigned int order) | |
1695 | { | |
1696 | return (size_t)PAGE_SIZE << order; | |
1697 | } | |
1698 | ||
1699 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1700 | unsigned int hctx_idx, unsigned int depth) | |
1701 | { | |
1702 | unsigned int i, j, entries_per_page, max_order = 4; | |
1703 | size_t rq_size, left; | |
59f082e4 SL |
1704 | int node; |
1705 | ||
1706 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1707 | if (node == NUMA_NO_NODE) | |
1708 | node = set->numa_node; | |
cc71a6f4 JA |
1709 | |
1710 | INIT_LIST_HEAD(&tags->page_list); | |
1711 | ||
320ae51f JA |
1712 | /* |
1713 | * rq_size is the size of the request plus driver payload, rounded | |
1714 | * to the cacheline size | |
1715 | */ | |
24d2f903 | 1716 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1717 | cache_line_size()); |
cc71a6f4 | 1718 | left = rq_size * depth; |
320ae51f | 1719 | |
cc71a6f4 | 1720 | for (i = 0; i < depth; ) { |
320ae51f JA |
1721 | int this_order = max_order; |
1722 | struct page *page; | |
1723 | int to_do; | |
1724 | void *p; | |
1725 | ||
b3a834b1 | 1726 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1727 | this_order--; |
1728 | ||
1729 | do { | |
59f082e4 | 1730 | page = alloc_pages_node(node, |
36e1f3d1 | 1731 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1732 | this_order); |
320ae51f JA |
1733 | if (page) |
1734 | break; | |
1735 | if (!this_order--) | |
1736 | break; | |
1737 | if (order_to_size(this_order) < rq_size) | |
1738 | break; | |
1739 | } while (1); | |
1740 | ||
1741 | if (!page) | |
24d2f903 | 1742 | goto fail; |
320ae51f JA |
1743 | |
1744 | page->private = this_order; | |
24d2f903 | 1745 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1746 | |
1747 | p = page_address(page); | |
f75782e4 CM |
1748 | /* |
1749 | * Allow kmemleak to scan these pages as they contain pointers | |
1750 | * to additional allocations like via ops->init_request(). | |
1751 | */ | |
36e1f3d1 | 1752 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1753 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1754 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1755 | left -= to_do * rq_size; |
1756 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1757 | struct request *rq = p; |
1758 | ||
1759 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1760 | if (set->ops->init_request) { |
1761 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1762 | rq, hctx_idx, i, |
59f082e4 | 1763 | node)) { |
2af8cbe3 | 1764 | tags->static_rqs[i] = NULL; |
24d2f903 | 1765 | goto fail; |
a5164405 | 1766 | } |
e9b267d9 CH |
1767 | } |
1768 | ||
320ae51f JA |
1769 | p += rq_size; |
1770 | i++; | |
1771 | } | |
1772 | } | |
cc71a6f4 | 1773 | return 0; |
320ae51f | 1774 | |
24d2f903 | 1775 | fail: |
cc71a6f4 JA |
1776 | blk_mq_free_rqs(set, tags, hctx_idx); |
1777 | return -ENOMEM; | |
320ae51f JA |
1778 | } |
1779 | ||
e57690fe JA |
1780 | /* |
1781 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1782 | * software queue to the hw queue dispatch list, and ensure that it | |
1783 | * gets run. | |
1784 | */ | |
9467f859 | 1785 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1786 | { |
9467f859 | 1787 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1788 | struct blk_mq_ctx *ctx; |
1789 | LIST_HEAD(tmp); | |
1790 | ||
9467f859 | 1791 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1792 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1793 | |
1794 | spin_lock(&ctx->lock); | |
1795 | if (!list_empty(&ctx->rq_list)) { | |
1796 | list_splice_init(&ctx->rq_list, &tmp); | |
1797 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1798 | } | |
1799 | spin_unlock(&ctx->lock); | |
1800 | ||
1801 | if (list_empty(&tmp)) | |
9467f859 | 1802 | return 0; |
484b4061 | 1803 | |
e57690fe JA |
1804 | spin_lock(&hctx->lock); |
1805 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1806 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1807 | |
1808 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1809 | return 0; |
484b4061 JA |
1810 | } |
1811 | ||
9467f859 | 1812 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1813 | { |
9467f859 TG |
1814 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1815 | &hctx->cpuhp_dead); | |
484b4061 JA |
1816 | } |
1817 | ||
c3b4afca | 1818 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1819 | static void blk_mq_exit_hctx(struct request_queue *q, |
1820 | struct blk_mq_tag_set *set, | |
1821 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1822 | { | |
f70ced09 ML |
1823 | unsigned flush_start_tag = set->queue_depth; |
1824 | ||
08e98fc6 ML |
1825 | blk_mq_tag_idle(hctx); |
1826 | ||
f70ced09 ML |
1827 | if (set->ops->exit_request) |
1828 | set->ops->exit_request(set->driver_data, | |
1829 | hctx->fq->flush_rq, hctx_idx, | |
1830 | flush_start_tag + hctx_idx); | |
1831 | ||
08e98fc6 ML |
1832 | if (set->ops->exit_hctx) |
1833 | set->ops->exit_hctx(hctx, hctx_idx); | |
1834 | ||
6a83e74d BVA |
1835 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1836 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1837 | ||
9467f859 | 1838 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1839 | blk_free_flush_queue(hctx->fq); |
88459642 | 1840 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1841 | } |
1842 | ||
624dbe47 ML |
1843 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1844 | struct blk_mq_tag_set *set, int nr_queue) | |
1845 | { | |
1846 | struct blk_mq_hw_ctx *hctx; | |
1847 | unsigned int i; | |
1848 | ||
1849 | queue_for_each_hw_ctx(q, hctx, i) { | |
1850 | if (i == nr_queue) | |
1851 | break; | |
08e98fc6 | 1852 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1853 | } |
624dbe47 ML |
1854 | } |
1855 | ||
08e98fc6 ML |
1856 | static int blk_mq_init_hctx(struct request_queue *q, |
1857 | struct blk_mq_tag_set *set, | |
1858 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1859 | { |
08e98fc6 | 1860 | int node; |
f70ced09 | 1861 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1862 | |
1863 | node = hctx->numa_node; | |
1864 | if (node == NUMA_NO_NODE) | |
1865 | node = hctx->numa_node = set->numa_node; | |
1866 | ||
27489a3c | 1867 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1868 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1869 | spin_lock_init(&hctx->lock); | |
1870 | INIT_LIST_HEAD(&hctx->dispatch); | |
1871 | hctx->queue = q; | |
1872 | hctx->queue_num = hctx_idx; | |
2404e607 | 1873 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1874 | |
9467f859 | 1875 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1876 | |
1877 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1878 | |
1879 | /* | |
08e98fc6 ML |
1880 | * Allocate space for all possible cpus to avoid allocation at |
1881 | * runtime | |
320ae51f | 1882 | */ |
08e98fc6 ML |
1883 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1884 | GFP_KERNEL, node); | |
1885 | if (!hctx->ctxs) | |
1886 | goto unregister_cpu_notifier; | |
320ae51f | 1887 | |
88459642 OS |
1888 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1889 | node)) | |
08e98fc6 | 1890 | goto free_ctxs; |
320ae51f | 1891 | |
08e98fc6 | 1892 | hctx->nr_ctx = 0; |
320ae51f | 1893 | |
08e98fc6 ML |
1894 | if (set->ops->init_hctx && |
1895 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1896 | goto free_bitmap; | |
320ae51f | 1897 | |
f70ced09 ML |
1898 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1899 | if (!hctx->fq) | |
1900 | goto exit_hctx; | |
320ae51f | 1901 | |
f70ced09 ML |
1902 | if (set->ops->init_request && |
1903 | set->ops->init_request(set->driver_data, | |
1904 | hctx->fq->flush_rq, hctx_idx, | |
1905 | flush_start_tag + hctx_idx, node)) | |
1906 | goto free_fq; | |
320ae51f | 1907 | |
6a83e74d BVA |
1908 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1909 | init_srcu_struct(&hctx->queue_rq_srcu); | |
1910 | ||
08e98fc6 | 1911 | return 0; |
320ae51f | 1912 | |
f70ced09 ML |
1913 | free_fq: |
1914 | kfree(hctx->fq); | |
1915 | exit_hctx: | |
1916 | if (set->ops->exit_hctx) | |
1917 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1918 | free_bitmap: |
88459642 | 1919 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1920 | free_ctxs: |
1921 | kfree(hctx->ctxs); | |
1922 | unregister_cpu_notifier: | |
9467f859 | 1923 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1924 | return -1; |
1925 | } | |
320ae51f | 1926 | |
320ae51f JA |
1927 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1928 | unsigned int nr_hw_queues) | |
1929 | { | |
1930 | unsigned int i; | |
1931 | ||
1932 | for_each_possible_cpu(i) { | |
1933 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1934 | struct blk_mq_hw_ctx *hctx; | |
1935 | ||
320ae51f JA |
1936 | __ctx->cpu = i; |
1937 | spin_lock_init(&__ctx->lock); | |
1938 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1939 | __ctx->queue = q; | |
1940 | ||
1941 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1942 | if (!cpu_online(i)) |
1943 | continue; | |
1944 | ||
7d7e0f90 | 1945 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1946 | |
320ae51f JA |
1947 | /* |
1948 | * Set local node, IFF we have more than one hw queue. If | |
1949 | * not, we remain on the home node of the device | |
1950 | */ | |
1951 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1952 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1953 | } |
1954 | } | |
1955 | ||
cc71a6f4 JA |
1956 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1957 | { | |
1958 | int ret = 0; | |
1959 | ||
1960 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1961 | set->queue_depth, set->reserved_tags); | |
1962 | if (!set->tags[hctx_idx]) | |
1963 | return false; | |
1964 | ||
1965 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1966 | set->queue_depth); | |
1967 | if (!ret) | |
1968 | return true; | |
1969 | ||
1970 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1971 | set->tags[hctx_idx] = NULL; | |
1972 | return false; | |
1973 | } | |
1974 | ||
1975 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
1976 | unsigned int hctx_idx) | |
1977 | { | |
bd166ef1 JA |
1978 | if (set->tags[hctx_idx]) { |
1979 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
1980 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
1981 | set->tags[hctx_idx] = NULL; | |
1982 | } | |
cc71a6f4 JA |
1983 | } |
1984 | ||
5778322e AM |
1985 | static void blk_mq_map_swqueue(struct request_queue *q, |
1986 | const struct cpumask *online_mask) | |
320ae51f | 1987 | { |
d1b1cea1 | 1988 | unsigned int i, hctx_idx; |
320ae51f JA |
1989 | struct blk_mq_hw_ctx *hctx; |
1990 | struct blk_mq_ctx *ctx; | |
2a34c087 | 1991 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 1992 | |
60de074b AM |
1993 | /* |
1994 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
1995 | */ | |
1996 | mutex_lock(&q->sysfs_lock); | |
1997 | ||
320ae51f | 1998 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 1999 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2000 | hctx->nr_ctx = 0; |
2001 | } | |
2002 | ||
2003 | /* | |
2004 | * Map software to hardware queues | |
2005 | */ | |
897bb0c7 | 2006 | for_each_possible_cpu(i) { |
320ae51f | 2007 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2008 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2009 | continue; |
2010 | ||
d1b1cea1 GKB |
2011 | hctx_idx = q->mq_map[i]; |
2012 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2013 | if (!set->tags[hctx_idx] && |
2014 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2015 | /* |
2016 | * If tags initialization fail for some hctx, | |
2017 | * that hctx won't be brought online. In this | |
2018 | * case, remap the current ctx to hctx[0] which | |
2019 | * is guaranteed to always have tags allocated | |
2020 | */ | |
cc71a6f4 | 2021 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2022 | } |
2023 | ||
897bb0c7 | 2024 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2025 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2026 | |
e4043dcf | 2027 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2028 | ctx->index_hw = hctx->nr_ctx; |
2029 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2030 | } | |
506e931f | 2031 | |
60de074b AM |
2032 | mutex_unlock(&q->sysfs_lock); |
2033 | ||
506e931f | 2034 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2035 | /* |
a68aafa5 JA |
2036 | * If no software queues are mapped to this hardware queue, |
2037 | * disable it and free the request entries. | |
484b4061 JA |
2038 | */ |
2039 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2040 | /* Never unmap queue 0. We need it as a |
2041 | * fallback in case of a new remap fails | |
2042 | * allocation | |
2043 | */ | |
cc71a6f4 JA |
2044 | if (i && set->tags[i]) |
2045 | blk_mq_free_map_and_requests(set, i); | |
2046 | ||
2a34c087 | 2047 | hctx->tags = NULL; |
484b4061 JA |
2048 | continue; |
2049 | } | |
2050 | ||
2a34c087 ML |
2051 | hctx->tags = set->tags[i]; |
2052 | WARN_ON(!hctx->tags); | |
2053 | ||
889fa31f CY |
2054 | /* |
2055 | * Set the map size to the number of mapped software queues. | |
2056 | * This is more accurate and more efficient than looping | |
2057 | * over all possibly mapped software queues. | |
2058 | */ | |
88459642 | 2059 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2060 | |
484b4061 JA |
2061 | /* |
2062 | * Initialize batch roundrobin counts | |
2063 | */ | |
506e931f JA |
2064 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2065 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2066 | } | |
320ae51f JA |
2067 | } |
2068 | ||
2404e607 | 2069 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2070 | { |
2071 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2072 | int i; |
2073 | ||
2404e607 JM |
2074 | queue_for_each_hw_ctx(q, hctx, i) { |
2075 | if (shared) | |
2076 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2077 | else | |
2078 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2079 | } | |
2080 | } | |
2081 | ||
2082 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2083 | { | |
2084 | struct request_queue *q; | |
0d2602ca JA |
2085 | |
2086 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2087 | blk_mq_freeze_queue(q); | |
2404e607 | 2088 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2089 | blk_mq_unfreeze_queue(q); |
2090 | } | |
2091 | } | |
2092 | ||
2093 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2094 | { | |
2095 | struct blk_mq_tag_set *set = q->tag_set; | |
2096 | ||
0d2602ca JA |
2097 | mutex_lock(&set->tag_list_lock); |
2098 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2099 | if (list_is_singular(&set->tag_list)) { |
2100 | /* just transitioned to unshared */ | |
2101 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2102 | /* update existing queue */ | |
2103 | blk_mq_update_tag_set_depth(set, false); | |
2104 | } | |
0d2602ca | 2105 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2106 | } |
2107 | ||
2108 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2109 | struct request_queue *q) | |
2110 | { | |
2111 | q->tag_set = set; | |
2112 | ||
2113 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2114 | |
2115 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2116 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2117 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2118 | /* update existing queue */ | |
2119 | blk_mq_update_tag_set_depth(set, true); | |
2120 | } | |
2121 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2122 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2123 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2124 | |
0d2602ca JA |
2125 | mutex_unlock(&set->tag_list_lock); |
2126 | } | |
2127 | ||
e09aae7e ML |
2128 | /* |
2129 | * It is the actual release handler for mq, but we do it from | |
2130 | * request queue's release handler for avoiding use-after-free | |
2131 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2132 | * but we can't group ctx/kctx kobj without it. | |
2133 | */ | |
2134 | void blk_mq_release(struct request_queue *q) | |
2135 | { | |
2136 | struct blk_mq_hw_ctx *hctx; | |
2137 | unsigned int i; | |
2138 | ||
bd166ef1 JA |
2139 | blk_mq_sched_teardown(q); |
2140 | ||
e09aae7e | 2141 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2142 | queue_for_each_hw_ctx(q, hctx, i) { |
2143 | if (!hctx) | |
2144 | continue; | |
6c8b232e | 2145 | kobject_put(&hctx->kobj); |
c3b4afca | 2146 | } |
e09aae7e | 2147 | |
a723bab3 AM |
2148 | q->mq_map = NULL; |
2149 | ||
e09aae7e ML |
2150 | kfree(q->queue_hw_ctx); |
2151 | ||
7ea5fe31 ML |
2152 | /* |
2153 | * release .mq_kobj and sw queue's kobject now because | |
2154 | * both share lifetime with request queue. | |
2155 | */ | |
2156 | blk_mq_sysfs_deinit(q); | |
2157 | ||
e09aae7e ML |
2158 | free_percpu(q->queue_ctx); |
2159 | } | |
2160 | ||
24d2f903 | 2161 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2162 | { |
2163 | struct request_queue *uninit_q, *q; | |
2164 | ||
2165 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2166 | if (!uninit_q) | |
2167 | return ERR_PTR(-ENOMEM); | |
2168 | ||
2169 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2170 | if (IS_ERR(q)) | |
2171 | blk_cleanup_queue(uninit_q); | |
2172 | ||
2173 | return q; | |
2174 | } | |
2175 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2176 | ||
868f2f0b KB |
2177 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2178 | struct request_queue *q) | |
320ae51f | 2179 | { |
868f2f0b KB |
2180 | int i, j; |
2181 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2182 | |
868f2f0b | 2183 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2184 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2185 | int node; |
f14bbe77 | 2186 | |
868f2f0b KB |
2187 | if (hctxs[i]) |
2188 | continue; | |
2189 | ||
2190 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2191 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2192 | GFP_KERNEL, node); | |
320ae51f | 2193 | if (!hctxs[i]) |
868f2f0b | 2194 | break; |
320ae51f | 2195 | |
a86073e4 | 2196 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2197 | node)) { |
2198 | kfree(hctxs[i]); | |
2199 | hctxs[i] = NULL; | |
2200 | break; | |
2201 | } | |
e4043dcf | 2202 | |
0d2602ca | 2203 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2204 | hctxs[i]->numa_node = node; |
320ae51f | 2205 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2206 | |
2207 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2208 | free_cpumask_var(hctxs[i]->cpumask); | |
2209 | kfree(hctxs[i]); | |
2210 | hctxs[i] = NULL; | |
2211 | break; | |
2212 | } | |
2213 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2214 | } |
868f2f0b KB |
2215 | for (j = i; j < q->nr_hw_queues; j++) { |
2216 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2217 | ||
2218 | if (hctx) { | |
cc71a6f4 JA |
2219 | if (hctx->tags) |
2220 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2221 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2222 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2223 | hctxs[j] = NULL; |
2224 | ||
2225 | } | |
2226 | } | |
2227 | q->nr_hw_queues = i; | |
2228 | blk_mq_sysfs_register(q); | |
2229 | } | |
2230 | ||
2231 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2232 | struct request_queue *q) | |
2233 | { | |
66841672 ML |
2234 | /* mark the queue as mq asap */ |
2235 | q->mq_ops = set->ops; | |
2236 | ||
34dbad5d OS |
2237 | q->stats = blk_alloc_queue_stats(); |
2238 | if (!q->stats) | |
2239 | goto err_exit; | |
2240 | ||
2241 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, | |
2242 | blk_stat_rq_ddir, 2, q); | |
2243 | if (!q->poll_cb) | |
2244 | goto err_exit; | |
2245 | ||
868f2f0b KB |
2246 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2247 | if (!q->queue_ctx) | |
c7de5726 | 2248 | goto err_exit; |
868f2f0b | 2249 | |
737f98cf ML |
2250 | /* init q->mq_kobj and sw queues' kobjects */ |
2251 | blk_mq_sysfs_init(q); | |
2252 | ||
868f2f0b KB |
2253 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2254 | GFP_KERNEL, set->numa_node); | |
2255 | if (!q->queue_hw_ctx) | |
2256 | goto err_percpu; | |
2257 | ||
bdd17e75 | 2258 | q->mq_map = set->mq_map; |
868f2f0b KB |
2259 | |
2260 | blk_mq_realloc_hw_ctxs(set, q); | |
2261 | if (!q->nr_hw_queues) | |
2262 | goto err_hctxs; | |
320ae51f | 2263 | |
287922eb | 2264 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2265 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2266 | |
2267 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2268 | |
94eddfbe | 2269 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2270 | |
05f1dd53 JA |
2271 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2272 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2273 | ||
1be036e9 CH |
2274 | q->sg_reserved_size = INT_MAX; |
2275 | ||
2849450a | 2276 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2277 | INIT_LIST_HEAD(&q->requeue_list); |
2278 | spin_lock_init(&q->requeue_lock); | |
2279 | ||
254d259d | 2280 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2281 | |
eba71768 JA |
2282 | /* |
2283 | * Do this after blk_queue_make_request() overrides it... | |
2284 | */ | |
2285 | q->nr_requests = set->queue_depth; | |
2286 | ||
64f1c21e JA |
2287 | /* |
2288 | * Default to classic polling | |
2289 | */ | |
2290 | q->poll_nsec = -1; | |
2291 | ||
24d2f903 CH |
2292 | if (set->ops->complete) |
2293 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2294 | |
24d2f903 | 2295 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2296 | |
5778322e | 2297 | get_online_cpus(); |
320ae51f | 2298 | mutex_lock(&all_q_mutex); |
320ae51f | 2299 | |
4593fdbe | 2300 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2301 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2302 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2303 | |
4593fdbe | 2304 | mutex_unlock(&all_q_mutex); |
5778322e | 2305 | put_online_cpus(); |
4593fdbe | 2306 | |
d3484991 JA |
2307 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2308 | int ret; | |
2309 | ||
2310 | ret = blk_mq_sched_init(q); | |
2311 | if (ret) | |
2312 | return ERR_PTR(ret); | |
2313 | } | |
2314 | ||
320ae51f | 2315 | return q; |
18741986 | 2316 | |
320ae51f | 2317 | err_hctxs: |
868f2f0b | 2318 | kfree(q->queue_hw_ctx); |
320ae51f | 2319 | err_percpu: |
868f2f0b | 2320 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2321 | err_exit: |
2322 | q->mq_ops = NULL; | |
320ae51f JA |
2323 | return ERR_PTR(-ENOMEM); |
2324 | } | |
b62c21b7 | 2325 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2326 | |
2327 | void blk_mq_free_queue(struct request_queue *q) | |
2328 | { | |
624dbe47 | 2329 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2330 | |
0e626368 AM |
2331 | mutex_lock(&all_q_mutex); |
2332 | list_del_init(&q->all_q_node); | |
2333 | mutex_unlock(&all_q_mutex); | |
2334 | ||
0d2602ca JA |
2335 | blk_mq_del_queue_tag_set(q); |
2336 | ||
624dbe47 | 2337 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2338 | } |
320ae51f JA |
2339 | |
2340 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2341 | static void blk_mq_queue_reinit(struct request_queue *q, |
2342 | const struct cpumask *online_mask) | |
320ae51f | 2343 | { |
4ecd4fef | 2344 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2345 | |
67aec14c JA |
2346 | blk_mq_sysfs_unregister(q); |
2347 | ||
320ae51f JA |
2348 | /* |
2349 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2350 | * we should change hctx numa_node according to new topology (this | |
2351 | * involves free and re-allocate memory, worthy doing?) | |
2352 | */ | |
2353 | ||
5778322e | 2354 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2355 | |
67aec14c | 2356 | blk_mq_sysfs_register(q); |
320ae51f JA |
2357 | } |
2358 | ||
65d5291e SAS |
2359 | /* |
2360 | * New online cpumask which is going to be set in this hotplug event. | |
2361 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2362 | * one-by-one and dynamically allocating this could result in a failure. | |
2363 | */ | |
2364 | static struct cpumask cpuhp_online_new; | |
2365 | ||
2366 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2367 | { |
2368 | struct request_queue *q; | |
320ae51f JA |
2369 | |
2370 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2371 | /* |
2372 | * We need to freeze and reinit all existing queues. Freezing | |
2373 | * involves synchronous wait for an RCU grace period and doing it | |
2374 | * one by one may take a long time. Start freezing all queues in | |
2375 | * one swoop and then wait for the completions so that freezing can | |
2376 | * take place in parallel. | |
2377 | */ | |
2378 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2379 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2380 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2381 | blk_mq_freeze_queue_wait(q); |
2382 | ||
320ae51f | 2383 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2384 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2385 | |
2386 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2387 | blk_mq_unfreeze_queue(q); | |
2388 | ||
320ae51f | 2389 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2390 | } |
2391 | ||
2392 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2393 | { | |
97a32864 | 2394 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2395 | blk_mq_queue_reinit_work(); |
2396 | return 0; | |
2397 | } | |
2398 | ||
2399 | /* | |
2400 | * Before hotadded cpu starts handling requests, new mappings must be | |
2401 | * established. Otherwise, these requests in hw queue might never be | |
2402 | * dispatched. | |
2403 | * | |
2404 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2405 | * for CPU0, and ctx1 for CPU1). | |
2406 | * | |
2407 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2408 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2409 | * | |
2c3ad667 JA |
2410 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2411 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2412 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2413 | * ignored. | |
65d5291e SAS |
2414 | */ |
2415 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2416 | { | |
2417 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2418 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2419 | blk_mq_queue_reinit_work(); | |
2420 | return 0; | |
320ae51f JA |
2421 | } |
2422 | ||
a5164405 JA |
2423 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2424 | { | |
2425 | int i; | |
2426 | ||
cc71a6f4 JA |
2427 | for (i = 0; i < set->nr_hw_queues; i++) |
2428 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2429 | goto out_unwind; |
a5164405 JA |
2430 | |
2431 | return 0; | |
2432 | ||
2433 | out_unwind: | |
2434 | while (--i >= 0) | |
cc71a6f4 | 2435 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2436 | |
a5164405 JA |
2437 | return -ENOMEM; |
2438 | } | |
2439 | ||
2440 | /* | |
2441 | * Allocate the request maps associated with this tag_set. Note that this | |
2442 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2443 | * will be updated to reflect the allocated depth. | |
2444 | */ | |
2445 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2446 | { | |
2447 | unsigned int depth; | |
2448 | int err; | |
2449 | ||
2450 | depth = set->queue_depth; | |
2451 | do { | |
2452 | err = __blk_mq_alloc_rq_maps(set); | |
2453 | if (!err) | |
2454 | break; | |
2455 | ||
2456 | set->queue_depth >>= 1; | |
2457 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2458 | err = -ENOMEM; | |
2459 | break; | |
2460 | } | |
2461 | } while (set->queue_depth); | |
2462 | ||
2463 | if (!set->queue_depth || err) { | |
2464 | pr_err("blk-mq: failed to allocate request map\n"); | |
2465 | return -ENOMEM; | |
2466 | } | |
2467 | ||
2468 | if (depth != set->queue_depth) | |
2469 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2470 | depth, set->queue_depth); | |
2471 | ||
2472 | return 0; | |
2473 | } | |
2474 | ||
a4391c64 JA |
2475 | /* |
2476 | * Alloc a tag set to be associated with one or more request queues. | |
2477 | * May fail with EINVAL for various error conditions. May adjust the | |
2478 | * requested depth down, if if it too large. In that case, the set | |
2479 | * value will be stored in set->queue_depth. | |
2480 | */ | |
24d2f903 CH |
2481 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2482 | { | |
da695ba2 CH |
2483 | int ret; |
2484 | ||
205fb5f5 BVA |
2485 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2486 | ||
24d2f903 CH |
2487 | if (!set->nr_hw_queues) |
2488 | return -EINVAL; | |
a4391c64 | 2489 | if (!set->queue_depth) |
24d2f903 CH |
2490 | return -EINVAL; |
2491 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2492 | return -EINVAL; | |
2493 | ||
7d7e0f90 | 2494 | if (!set->ops->queue_rq) |
24d2f903 CH |
2495 | return -EINVAL; |
2496 | ||
a4391c64 JA |
2497 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2498 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2499 | BLK_MQ_MAX_DEPTH); | |
2500 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2501 | } | |
24d2f903 | 2502 | |
6637fadf SL |
2503 | /* |
2504 | * If a crashdump is active, then we are potentially in a very | |
2505 | * memory constrained environment. Limit us to 1 queue and | |
2506 | * 64 tags to prevent using too much memory. | |
2507 | */ | |
2508 | if (is_kdump_kernel()) { | |
2509 | set->nr_hw_queues = 1; | |
2510 | set->queue_depth = min(64U, set->queue_depth); | |
2511 | } | |
868f2f0b KB |
2512 | /* |
2513 | * There is no use for more h/w queues than cpus. | |
2514 | */ | |
2515 | if (set->nr_hw_queues > nr_cpu_ids) | |
2516 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2517 | |
868f2f0b | 2518 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2519 | GFP_KERNEL, set->numa_node); |
2520 | if (!set->tags) | |
a5164405 | 2521 | return -ENOMEM; |
24d2f903 | 2522 | |
da695ba2 CH |
2523 | ret = -ENOMEM; |
2524 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2525 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2526 | if (!set->mq_map) |
2527 | goto out_free_tags; | |
2528 | ||
da695ba2 CH |
2529 | if (set->ops->map_queues) |
2530 | ret = set->ops->map_queues(set); | |
2531 | else | |
2532 | ret = blk_mq_map_queues(set); | |
2533 | if (ret) | |
2534 | goto out_free_mq_map; | |
2535 | ||
2536 | ret = blk_mq_alloc_rq_maps(set); | |
2537 | if (ret) | |
bdd17e75 | 2538 | goto out_free_mq_map; |
24d2f903 | 2539 | |
0d2602ca JA |
2540 | mutex_init(&set->tag_list_lock); |
2541 | INIT_LIST_HEAD(&set->tag_list); | |
2542 | ||
24d2f903 | 2543 | return 0; |
bdd17e75 CH |
2544 | |
2545 | out_free_mq_map: | |
2546 | kfree(set->mq_map); | |
2547 | set->mq_map = NULL; | |
2548 | out_free_tags: | |
5676e7b6 RE |
2549 | kfree(set->tags); |
2550 | set->tags = NULL; | |
da695ba2 | 2551 | return ret; |
24d2f903 CH |
2552 | } |
2553 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2554 | ||
2555 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2556 | { | |
2557 | int i; | |
2558 | ||
cc71a6f4 JA |
2559 | for (i = 0; i < nr_cpu_ids; i++) |
2560 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2561 | |
bdd17e75 CH |
2562 | kfree(set->mq_map); |
2563 | set->mq_map = NULL; | |
2564 | ||
981bd189 | 2565 | kfree(set->tags); |
5676e7b6 | 2566 | set->tags = NULL; |
24d2f903 CH |
2567 | } |
2568 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2569 | ||
e3a2b3f9 JA |
2570 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2571 | { | |
2572 | struct blk_mq_tag_set *set = q->tag_set; | |
2573 | struct blk_mq_hw_ctx *hctx; | |
2574 | int i, ret; | |
2575 | ||
bd166ef1 | 2576 | if (!set) |
e3a2b3f9 JA |
2577 | return -EINVAL; |
2578 | ||
70f36b60 JA |
2579 | blk_mq_freeze_queue(q); |
2580 | blk_mq_quiesce_queue(q); | |
2581 | ||
e3a2b3f9 JA |
2582 | ret = 0; |
2583 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2584 | if (!hctx->tags) |
2585 | continue; | |
bd166ef1 JA |
2586 | /* |
2587 | * If we're using an MQ scheduler, just update the scheduler | |
2588 | * queue depth. This is similar to what the old code would do. | |
2589 | */ | |
70f36b60 JA |
2590 | if (!hctx->sched_tags) { |
2591 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2592 | min(nr, set->queue_depth), | |
2593 | false); | |
2594 | } else { | |
2595 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2596 | nr, true); | |
2597 | } | |
e3a2b3f9 JA |
2598 | if (ret) |
2599 | break; | |
2600 | } | |
2601 | ||
2602 | if (!ret) | |
2603 | q->nr_requests = nr; | |
2604 | ||
70f36b60 JA |
2605 | blk_mq_unfreeze_queue(q); |
2606 | blk_mq_start_stopped_hw_queues(q, true); | |
2607 | ||
e3a2b3f9 JA |
2608 | return ret; |
2609 | } | |
2610 | ||
868f2f0b KB |
2611 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2612 | { | |
2613 | struct request_queue *q; | |
2614 | ||
2615 | if (nr_hw_queues > nr_cpu_ids) | |
2616 | nr_hw_queues = nr_cpu_ids; | |
2617 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2618 | return; | |
2619 | ||
2620 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2621 | blk_mq_freeze_queue(q); | |
2622 | ||
2623 | set->nr_hw_queues = nr_hw_queues; | |
2624 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2625 | blk_mq_realloc_hw_ctxs(set, q); | |
868f2f0b KB |
2626 | blk_mq_queue_reinit(q, cpu_online_mask); |
2627 | } | |
2628 | ||
2629 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2630 | blk_mq_unfreeze_queue(q); | |
2631 | } | |
2632 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2633 | ||
34dbad5d OS |
2634 | /* Enable polling stats and return whether they were already enabled. */ |
2635 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2636 | { | |
2637 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2638 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2639 | return true; | |
2640 | blk_stat_add_callback(q, q->poll_cb); | |
2641 | return false; | |
2642 | } | |
2643 | ||
2644 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2645 | { | |
2646 | /* | |
2647 | * We don't arm the callback if polling stats are not enabled or the | |
2648 | * callback is already active. | |
2649 | */ | |
2650 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2651 | blk_stat_is_active(q->poll_cb)) | |
2652 | return; | |
2653 | ||
2654 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2655 | } | |
2656 | ||
2657 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2658 | { | |
2659 | struct request_queue *q = cb->data; | |
2660 | ||
2661 | if (cb->stat[READ].nr_samples) | |
2662 | q->poll_stat[READ] = cb->stat[READ]; | |
2663 | if (cb->stat[WRITE].nr_samples) | |
2664 | q->poll_stat[WRITE] = cb->stat[WRITE]; | |
2665 | } | |
2666 | ||
64f1c21e JA |
2667 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2668 | struct blk_mq_hw_ctx *hctx, | |
2669 | struct request *rq) | |
2670 | { | |
64f1c21e JA |
2671 | unsigned long ret = 0; |
2672 | ||
2673 | /* | |
2674 | * If stats collection isn't on, don't sleep but turn it on for | |
2675 | * future users | |
2676 | */ | |
34dbad5d | 2677 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2678 | return 0; |
2679 | ||
64f1c21e JA |
2680 | /* |
2681 | * As an optimistic guess, use half of the mean service time | |
2682 | * for this type of request. We can (and should) make this smarter. | |
2683 | * For instance, if the completion latencies are tight, we can | |
2684 | * get closer than just half the mean. This is especially | |
2685 | * important on devices where the completion latencies are longer | |
2686 | * than ~10 usec. | |
2687 | */ | |
34dbad5d OS |
2688 | if (req_op(rq) == REQ_OP_READ && q->poll_stat[READ].nr_samples) |
2689 | ret = (q->poll_stat[READ].mean + 1) / 2; | |
2690 | else if (req_op(rq) == REQ_OP_WRITE && q->poll_stat[WRITE].nr_samples) | |
2691 | ret = (q->poll_stat[WRITE].mean + 1) / 2; | |
64f1c21e JA |
2692 | |
2693 | return ret; | |
2694 | } | |
2695 | ||
06426adf | 2696 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2697 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2698 | struct request *rq) |
2699 | { | |
2700 | struct hrtimer_sleeper hs; | |
2701 | enum hrtimer_mode mode; | |
64f1c21e | 2702 | unsigned int nsecs; |
06426adf JA |
2703 | ktime_t kt; |
2704 | ||
64f1c21e JA |
2705 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2706 | return false; | |
2707 | ||
2708 | /* | |
2709 | * poll_nsec can be: | |
2710 | * | |
2711 | * -1: don't ever hybrid sleep | |
2712 | * 0: use half of prev avg | |
2713 | * >0: use this specific value | |
2714 | */ | |
2715 | if (q->poll_nsec == -1) | |
2716 | return false; | |
2717 | else if (q->poll_nsec > 0) | |
2718 | nsecs = q->poll_nsec; | |
2719 | else | |
2720 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2721 | ||
2722 | if (!nsecs) | |
06426adf JA |
2723 | return false; |
2724 | ||
2725 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2726 | ||
2727 | /* | |
2728 | * This will be replaced with the stats tracking code, using | |
2729 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2730 | */ | |
8b0e1953 | 2731 | kt = nsecs; |
06426adf JA |
2732 | |
2733 | mode = HRTIMER_MODE_REL; | |
2734 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2735 | hrtimer_set_expires(&hs.timer, kt); | |
2736 | ||
2737 | hrtimer_init_sleeper(&hs, current); | |
2738 | do { | |
2739 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2740 | break; | |
2741 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2742 | hrtimer_start_expires(&hs.timer, mode); | |
2743 | if (hs.task) | |
2744 | io_schedule(); | |
2745 | hrtimer_cancel(&hs.timer); | |
2746 | mode = HRTIMER_MODE_ABS; | |
2747 | } while (hs.task && !signal_pending(current)); | |
2748 | ||
2749 | __set_current_state(TASK_RUNNING); | |
2750 | destroy_hrtimer_on_stack(&hs.timer); | |
2751 | return true; | |
2752 | } | |
2753 | ||
bbd7bb70 JA |
2754 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2755 | { | |
2756 | struct request_queue *q = hctx->queue; | |
2757 | long state; | |
2758 | ||
06426adf JA |
2759 | /* |
2760 | * If we sleep, have the caller restart the poll loop to reset | |
2761 | * the state. Like for the other success return cases, the | |
2762 | * caller is responsible for checking if the IO completed. If | |
2763 | * the IO isn't complete, we'll get called again and will go | |
2764 | * straight to the busy poll loop. | |
2765 | */ | |
64f1c21e | 2766 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2767 | return true; |
2768 | ||
bbd7bb70 JA |
2769 | hctx->poll_considered++; |
2770 | ||
2771 | state = current->state; | |
2772 | while (!need_resched()) { | |
2773 | int ret; | |
2774 | ||
2775 | hctx->poll_invoked++; | |
2776 | ||
2777 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2778 | if (ret > 0) { | |
2779 | hctx->poll_success++; | |
2780 | set_current_state(TASK_RUNNING); | |
2781 | return true; | |
2782 | } | |
2783 | ||
2784 | if (signal_pending_state(state, current)) | |
2785 | set_current_state(TASK_RUNNING); | |
2786 | ||
2787 | if (current->state == TASK_RUNNING) | |
2788 | return true; | |
2789 | if (ret < 0) | |
2790 | break; | |
2791 | cpu_relax(); | |
2792 | } | |
2793 | ||
2794 | return false; | |
2795 | } | |
2796 | ||
2797 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2798 | { | |
2799 | struct blk_mq_hw_ctx *hctx; | |
2800 | struct blk_plug *plug; | |
2801 | struct request *rq; | |
2802 | ||
2803 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2804 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2805 | return false; | |
2806 | ||
2807 | plug = current->plug; | |
2808 | if (plug) | |
2809 | blk_flush_plug_list(plug, false); | |
2810 | ||
2811 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2812 | if (!blk_qc_t_is_internal(cookie)) |
2813 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2814 | else | |
2815 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2816 | |
2817 | return __blk_mq_poll(hctx, rq); | |
2818 | } | |
2819 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2820 | ||
676141e4 JA |
2821 | void blk_mq_disable_hotplug(void) |
2822 | { | |
2823 | mutex_lock(&all_q_mutex); | |
2824 | } | |
2825 | ||
2826 | void blk_mq_enable_hotplug(void) | |
2827 | { | |
2828 | mutex_unlock(&all_q_mutex); | |
2829 | } | |
2830 | ||
320ae51f JA |
2831 | static int __init blk_mq_init(void) |
2832 | { | |
9467f859 TG |
2833 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2834 | blk_mq_hctx_notify_dead); | |
320ae51f | 2835 | |
65d5291e SAS |
2836 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2837 | blk_mq_queue_reinit_prepare, | |
2838 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2839 | return 0; |
2840 | } | |
2841 | subsys_initcall(blk_mq_init); |