]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
23 | #include <linux/delay.h> | |
aedcd72f | 24 | #include <linux/crash_dump.h> |
88c7b2b7 | 25 | #include <linux/prefetch.h> |
320ae51f JA |
26 | |
27 | #include <trace/events/block.h> | |
28 | ||
29 | #include <linux/blk-mq.h> | |
30 | #include "blk.h" | |
31 | #include "blk-mq.h" | |
32 | #include "blk-mq-tag.h" | |
cf43e6be | 33 | #include "blk-stat.h" |
87760e5e | 34 | #include "blk-wbt.h" |
bd166ef1 | 35 | #include "blk-mq-sched.h" |
320ae51f JA |
36 | |
37 | static DEFINE_MUTEX(all_q_mutex); | |
38 | static LIST_HEAD(all_q_list); | |
39 | ||
320ae51f JA |
40 | /* |
41 | * Check if any of the ctx's have pending work in this hardware queue | |
42 | */ | |
50e1dab8 | 43 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 44 | { |
bd166ef1 JA |
45 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
46 | !list_empty_careful(&hctx->dispatch) || | |
47 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
48 | } |
49 | ||
320ae51f JA |
50 | /* |
51 | * Mark this ctx as having pending work in this hardware queue | |
52 | */ | |
53 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
54 | struct blk_mq_ctx *ctx) | |
55 | { | |
88459642 OS |
56 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
57 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
58 | } |
59 | ||
60 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
61 | struct blk_mq_ctx *ctx) | |
62 | { | |
88459642 | 63 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
64 | } |
65 | ||
b4c6a028 | 66 | void blk_mq_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 67 | { |
4ecd4fef | 68 | int freeze_depth; |
cddd5d17 | 69 | |
4ecd4fef CH |
70 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
71 | if (freeze_depth == 1) { | |
3ef28e83 | 72 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 73 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 74 | } |
f3af020b | 75 | } |
b4c6a028 | 76 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start); |
f3af020b | 77 | |
6bae363e | 78 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 79 | { |
3ef28e83 | 80 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 81 | } |
6bae363e | 82 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 83 | |
f91328c4 KB |
84 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
85 | unsigned long timeout) | |
86 | { | |
87 | return wait_event_timeout(q->mq_freeze_wq, | |
88 | percpu_ref_is_zero(&q->q_usage_counter), | |
89 | timeout); | |
90 | } | |
91 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
92 | ||
f3af020b TH |
93 | /* |
94 | * Guarantee no request is in use, so we can change any data structure of | |
95 | * the queue afterward. | |
96 | */ | |
3ef28e83 | 97 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 98 | { |
3ef28e83 DW |
99 | /* |
100 | * In the !blk_mq case we are only calling this to kill the | |
101 | * q_usage_counter, otherwise this increases the freeze depth | |
102 | * and waits for it to return to zero. For this reason there is | |
103 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
104 | * exported to drivers as the only user for unfreeze is blk_mq. | |
105 | */ | |
f3af020b TH |
106 | blk_mq_freeze_queue_start(q); |
107 | blk_mq_freeze_queue_wait(q); | |
108 | } | |
3ef28e83 DW |
109 | |
110 | void blk_mq_freeze_queue(struct request_queue *q) | |
111 | { | |
112 | /* | |
113 | * ...just an alias to keep freeze and unfreeze actions balanced | |
114 | * in the blk_mq_* namespace | |
115 | */ | |
116 | blk_freeze_queue(q); | |
117 | } | |
c761d96b | 118 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 119 | |
b4c6a028 | 120 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 121 | { |
4ecd4fef | 122 | int freeze_depth; |
320ae51f | 123 | |
4ecd4fef CH |
124 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
125 | WARN_ON_ONCE(freeze_depth < 0); | |
126 | if (!freeze_depth) { | |
3ef28e83 | 127 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 128 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 129 | } |
320ae51f | 130 | } |
b4c6a028 | 131 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 132 | |
6a83e74d BVA |
133 | /** |
134 | * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished | |
135 | * @q: request queue. | |
136 | * | |
137 | * Note: this function does not prevent that the struct request end_io() | |
138 | * callback function is invoked. Additionally, it is not prevented that | |
139 | * new queue_rq() calls occur unless the queue has been stopped first. | |
140 | */ | |
141 | void blk_mq_quiesce_queue(struct request_queue *q) | |
142 | { | |
143 | struct blk_mq_hw_ctx *hctx; | |
144 | unsigned int i; | |
145 | bool rcu = false; | |
146 | ||
147 | blk_mq_stop_hw_queues(q); | |
148 | ||
149 | queue_for_each_hw_ctx(q, hctx, i) { | |
150 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
151 | synchronize_srcu(&hctx->queue_rq_srcu); | |
152 | else | |
153 | rcu = true; | |
154 | } | |
155 | if (rcu) | |
156 | synchronize_rcu(); | |
157 | } | |
158 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
159 | ||
aed3ea94 JA |
160 | void blk_mq_wake_waiters(struct request_queue *q) |
161 | { | |
162 | struct blk_mq_hw_ctx *hctx; | |
163 | unsigned int i; | |
164 | ||
165 | queue_for_each_hw_ctx(q, hctx, i) | |
166 | if (blk_mq_hw_queue_mapped(hctx)) | |
167 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
168 | |
169 | /* | |
170 | * If we are called because the queue has now been marked as | |
171 | * dying, we need to ensure that processes currently waiting on | |
172 | * the queue are notified as well. | |
173 | */ | |
174 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
175 | } |
176 | ||
320ae51f JA |
177 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
178 | { | |
179 | return blk_mq_has_free_tags(hctx->tags); | |
180 | } | |
181 | EXPORT_SYMBOL(blk_mq_can_queue); | |
182 | ||
2c3ad667 JA |
183 | void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx, |
184 | struct request *rq, unsigned int op) | |
320ae51f | 185 | { |
af76e555 CH |
186 | INIT_LIST_HEAD(&rq->queuelist); |
187 | /* csd/requeue_work/fifo_time is initialized before use */ | |
188 | rq->q = q; | |
320ae51f | 189 | rq->mq_ctx = ctx; |
ef295ecf | 190 | rq->cmd_flags = op; |
e8064021 CH |
191 | if (blk_queue_io_stat(q)) |
192 | rq->rq_flags |= RQF_IO_STAT; | |
af76e555 CH |
193 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
194 | rq->cpu = -1; | |
af76e555 CH |
195 | INIT_HLIST_NODE(&rq->hash); |
196 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
197 | rq->rq_disk = NULL; |
198 | rq->part = NULL; | |
3ee32372 | 199 | rq->start_time = jiffies; |
af76e555 CH |
200 | #ifdef CONFIG_BLK_CGROUP |
201 | rq->rl = NULL; | |
0fec08b4 | 202 | set_start_time_ns(rq); |
af76e555 CH |
203 | rq->io_start_time_ns = 0; |
204 | #endif | |
205 | rq->nr_phys_segments = 0; | |
206 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
207 | rq->nr_integrity_segments = 0; | |
208 | #endif | |
af76e555 CH |
209 | rq->special = NULL; |
210 | /* tag was already set */ | |
211 | rq->errors = 0; | |
af76e555 | 212 | rq->extra_len = 0; |
af76e555 | 213 | |
af76e555 | 214 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
215 | rq->timeout = 0; |
216 | ||
af76e555 CH |
217 | rq->end_io = NULL; |
218 | rq->end_io_data = NULL; | |
219 | rq->next_rq = NULL; | |
220 | ||
ef295ecf | 221 | ctx->rq_dispatched[op_is_sync(op)]++; |
320ae51f | 222 | } |
2c3ad667 | 223 | EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init); |
320ae51f | 224 | |
2c3ad667 JA |
225 | struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data, |
226 | unsigned int op) | |
5dee8577 CH |
227 | { |
228 | struct request *rq; | |
229 | unsigned int tag; | |
230 | ||
cb96a42c | 231 | tag = blk_mq_get_tag(data); |
5dee8577 | 232 | if (tag != BLK_MQ_TAG_FAIL) { |
bd166ef1 JA |
233 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
234 | ||
235 | rq = tags->static_rqs[tag]; | |
5dee8577 | 236 | |
bd166ef1 JA |
237 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
238 | rq->tag = -1; | |
239 | rq->internal_tag = tag; | |
240 | } else { | |
200e86b3 JA |
241 | if (blk_mq_tag_busy(data->hctx)) { |
242 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
243 | atomic_inc(&data->hctx->nr_active); | |
244 | } | |
bd166ef1 JA |
245 | rq->tag = tag; |
246 | rq->internal_tag = -1; | |
562bef42 | 247 | data->hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
248 | } |
249 | ||
ef295ecf | 250 | blk_mq_rq_ctx_init(data->q, data->ctx, rq, op); |
5dee8577 CH |
251 | return rq; |
252 | } | |
253 | ||
254 | return NULL; | |
255 | } | |
2c3ad667 | 256 | EXPORT_SYMBOL_GPL(__blk_mq_alloc_request); |
5dee8577 | 257 | |
6f3b0e8b CH |
258 | struct request *blk_mq_alloc_request(struct request_queue *q, int rw, |
259 | unsigned int flags) | |
320ae51f | 260 | { |
5a797e00 | 261 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 262 | struct request *rq; |
a492f075 | 263 | int ret; |
320ae51f | 264 | |
6f3b0e8b | 265 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
266 | if (ret) |
267 | return ERR_PTR(ret); | |
320ae51f | 268 | |
bd166ef1 | 269 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
841bac2c | 270 | |
bd166ef1 JA |
271 | blk_mq_put_ctx(alloc_data.ctx); |
272 | blk_queue_exit(q); | |
273 | ||
274 | if (!rq) | |
a492f075 | 275 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
276 | |
277 | rq->__data_len = 0; | |
278 | rq->__sector = (sector_t) -1; | |
279 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
280 | return rq; |
281 | } | |
4bb659b1 | 282 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 283 | |
1f5bd336 ML |
284 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw, |
285 | unsigned int flags, unsigned int hctx_idx) | |
286 | { | |
6d2809d5 | 287 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 288 | struct request *rq; |
6d2809d5 | 289 | unsigned int cpu; |
1f5bd336 ML |
290 | int ret; |
291 | ||
292 | /* | |
293 | * If the tag allocator sleeps we could get an allocation for a | |
294 | * different hardware context. No need to complicate the low level | |
295 | * allocator for this for the rare use case of a command tied to | |
296 | * a specific queue. | |
297 | */ | |
298 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
299 | return ERR_PTR(-EINVAL); | |
300 | ||
301 | if (hctx_idx >= q->nr_hw_queues) | |
302 | return ERR_PTR(-EIO); | |
303 | ||
304 | ret = blk_queue_enter(q, true); | |
305 | if (ret) | |
306 | return ERR_PTR(ret); | |
307 | ||
c8712c6a CH |
308 | /* |
309 | * Check if the hardware context is actually mapped to anything. | |
310 | * If not tell the caller that it should skip this queue. | |
311 | */ | |
6d2809d5 OS |
312 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
313 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
314 | blk_queue_exit(q); | |
315 | return ERR_PTR(-EXDEV); | |
1f5bd336 | 316 | } |
6d2809d5 OS |
317 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
318 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 319 | |
6d2809d5 | 320 | rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data); |
c8712c6a | 321 | |
6d2809d5 | 322 | blk_mq_put_ctx(alloc_data.ctx); |
c8712c6a | 323 | blk_queue_exit(q); |
6d2809d5 OS |
324 | |
325 | if (!rq) | |
326 | return ERR_PTR(-EWOULDBLOCK); | |
327 | ||
328 | return rq; | |
1f5bd336 ML |
329 | } |
330 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
331 | ||
bd166ef1 JA |
332 | void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
333 | struct request *rq) | |
320ae51f | 334 | { |
bd166ef1 | 335 | const int sched_tag = rq->internal_tag; |
320ae51f JA |
336 | struct request_queue *q = rq->q; |
337 | ||
e8064021 | 338 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 339 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
340 | |
341 | wbt_done(q->rq_wb, &rq->issue_stat); | |
e8064021 | 342 | rq->rq_flags = 0; |
0d2602ca | 343 | |
af76e555 | 344 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 345 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
346 | if (rq->tag != -1) |
347 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
348 | if (sched_tag != -1) | |
349 | blk_mq_sched_completed_request(hctx, rq); | |
50e1dab8 | 350 | blk_mq_sched_restart_queues(hctx); |
3ef28e83 | 351 | blk_queue_exit(q); |
320ae51f JA |
352 | } |
353 | ||
bd166ef1 | 354 | static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx, |
16a3c2a7 | 355 | struct request *rq) |
320ae51f JA |
356 | { |
357 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
320ae51f JA |
358 | |
359 | ctx->rq_completed[rq_is_sync(rq)]++; | |
bd166ef1 JA |
360 | __blk_mq_finish_request(hctx, ctx, rq); |
361 | } | |
362 | ||
363 | void blk_mq_finish_request(struct request *rq) | |
364 | { | |
365 | blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq); | |
7c7f2f2b | 366 | } |
7c7f2f2b JA |
367 | |
368 | void blk_mq_free_request(struct request *rq) | |
369 | { | |
bd166ef1 | 370 | blk_mq_sched_put_request(rq); |
320ae51f | 371 | } |
1a3b595a | 372 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 373 | |
c8a446ad | 374 | inline void __blk_mq_end_request(struct request *rq, int error) |
320ae51f | 375 | { |
0d11e6ac ML |
376 | blk_account_io_done(rq); |
377 | ||
91b63639 | 378 | if (rq->end_io) { |
87760e5e | 379 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 380 | rq->end_io(rq, error); |
91b63639 CH |
381 | } else { |
382 | if (unlikely(blk_bidi_rq(rq))) | |
383 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 384 | blk_mq_free_request(rq); |
91b63639 | 385 | } |
320ae51f | 386 | } |
c8a446ad | 387 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 388 | |
c8a446ad | 389 | void blk_mq_end_request(struct request *rq, int error) |
63151a44 CH |
390 | { |
391 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
392 | BUG(); | |
c8a446ad | 393 | __blk_mq_end_request(rq, error); |
63151a44 | 394 | } |
c8a446ad | 395 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 396 | |
30a91cb4 | 397 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 398 | { |
3d6efbf6 | 399 | struct request *rq = data; |
320ae51f | 400 | |
30a91cb4 | 401 | rq->q->softirq_done_fn(rq); |
320ae51f | 402 | } |
320ae51f | 403 | |
ed851860 | 404 | static void blk_mq_ipi_complete_request(struct request *rq) |
320ae51f JA |
405 | { |
406 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 407 | bool shared = false; |
320ae51f JA |
408 | int cpu; |
409 | ||
38535201 | 410 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
411 | rq->q->softirq_done_fn(rq); |
412 | return; | |
413 | } | |
320ae51f JA |
414 | |
415 | cpu = get_cpu(); | |
38535201 CH |
416 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
417 | shared = cpus_share_cache(cpu, ctx->cpu); | |
418 | ||
419 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 420 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
421 | rq->csd.info = rq; |
422 | rq->csd.flags = 0; | |
c46fff2a | 423 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 424 | } else { |
30a91cb4 | 425 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 426 | } |
320ae51f JA |
427 | put_cpu(); |
428 | } | |
30a91cb4 | 429 | |
cf43e6be JA |
430 | static void blk_mq_stat_add(struct request *rq) |
431 | { | |
432 | if (rq->rq_flags & RQF_STATS) { | |
433 | /* | |
434 | * We could rq->mq_ctx here, but there's less of a risk | |
435 | * of races if we have the completion event add the stats | |
436 | * to the local software queue. | |
437 | */ | |
438 | struct blk_mq_ctx *ctx; | |
439 | ||
440 | ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id()); | |
441 | blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq); | |
442 | } | |
443 | } | |
444 | ||
1fa8cc52 | 445 | static void __blk_mq_complete_request(struct request *rq) |
ed851860 JA |
446 | { |
447 | struct request_queue *q = rq->q; | |
448 | ||
cf43e6be JA |
449 | blk_mq_stat_add(rq); |
450 | ||
ed851860 | 451 | if (!q->softirq_done_fn) |
c8a446ad | 452 | blk_mq_end_request(rq, rq->errors); |
ed851860 JA |
453 | else |
454 | blk_mq_ipi_complete_request(rq); | |
455 | } | |
456 | ||
30a91cb4 CH |
457 | /** |
458 | * blk_mq_complete_request - end I/O on a request | |
459 | * @rq: the request being processed | |
460 | * | |
461 | * Description: | |
462 | * Ends all I/O on a request. It does not handle partial completions. | |
463 | * The actual completion happens out-of-order, through a IPI handler. | |
464 | **/ | |
f4829a9b | 465 | void blk_mq_complete_request(struct request *rq, int error) |
30a91cb4 | 466 | { |
95f09684 JA |
467 | struct request_queue *q = rq->q; |
468 | ||
469 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 470 | return; |
f4829a9b CH |
471 | if (!blk_mark_rq_complete(rq)) { |
472 | rq->errors = error; | |
ed851860 | 473 | __blk_mq_complete_request(rq); |
f4829a9b | 474 | } |
30a91cb4 CH |
475 | } |
476 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 477 | |
973c0191 KB |
478 | int blk_mq_request_started(struct request *rq) |
479 | { | |
480 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
481 | } | |
482 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
483 | ||
e2490073 | 484 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
485 | { |
486 | struct request_queue *q = rq->q; | |
487 | ||
bd166ef1 JA |
488 | blk_mq_sched_started_request(rq); |
489 | ||
320ae51f JA |
490 | trace_block_rq_issue(q, rq); |
491 | ||
cf43e6be JA |
492 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
493 | blk_stat_set_issue_time(&rq->issue_stat); | |
494 | rq->rq_flags |= RQF_STATS; | |
87760e5e | 495 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
496 | } |
497 | ||
2b8393b4 | 498 | blk_add_timer(rq); |
87ee7b11 | 499 | |
538b7534 JA |
500 | /* |
501 | * Ensure that ->deadline is visible before set the started | |
502 | * flag and clear the completed flag. | |
503 | */ | |
504 | smp_mb__before_atomic(); | |
505 | ||
87ee7b11 JA |
506 | /* |
507 | * Mark us as started and clear complete. Complete might have been | |
508 | * set if requeue raced with timeout, which then marked it as | |
509 | * complete. So be sure to clear complete again when we start | |
510 | * the request, otherwise we'll ignore the completion event. | |
511 | */ | |
4b570521 JA |
512 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
513 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
514 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
515 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
516 | |
517 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
518 | /* | |
519 | * Make sure space for the drain appears. We know we can do | |
520 | * this because max_hw_segments has been adjusted to be one | |
521 | * fewer than the device can handle. | |
522 | */ | |
523 | rq->nr_phys_segments++; | |
524 | } | |
320ae51f | 525 | } |
e2490073 | 526 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 527 | |
ed0791b2 | 528 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
529 | { |
530 | struct request_queue *q = rq->q; | |
531 | ||
532 | trace_block_rq_requeue(q, rq); | |
87760e5e | 533 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 534 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 535 | |
e2490073 CH |
536 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
537 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
538 | rq->nr_phys_segments--; | |
539 | } | |
320ae51f JA |
540 | } |
541 | ||
2b053aca | 542 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 543 | { |
ed0791b2 | 544 | __blk_mq_requeue_request(rq); |
ed0791b2 | 545 | |
ed0791b2 | 546 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 547 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
548 | } |
549 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
550 | ||
6fca6a61 CH |
551 | static void blk_mq_requeue_work(struct work_struct *work) |
552 | { | |
553 | struct request_queue *q = | |
2849450a | 554 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
555 | LIST_HEAD(rq_list); |
556 | struct request *rq, *next; | |
557 | unsigned long flags; | |
558 | ||
559 | spin_lock_irqsave(&q->requeue_lock, flags); | |
560 | list_splice_init(&q->requeue_list, &rq_list); | |
561 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
562 | ||
563 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 564 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
565 | continue; |
566 | ||
e8064021 | 567 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 568 | list_del_init(&rq->queuelist); |
bd6737f1 | 569 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
570 | } |
571 | ||
572 | while (!list_empty(&rq_list)) { | |
573 | rq = list_entry(rq_list.next, struct request, queuelist); | |
574 | list_del_init(&rq->queuelist); | |
bd6737f1 | 575 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
576 | } |
577 | ||
52d7f1b5 | 578 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
579 | } |
580 | ||
2b053aca BVA |
581 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
582 | bool kick_requeue_list) | |
6fca6a61 CH |
583 | { |
584 | struct request_queue *q = rq->q; | |
585 | unsigned long flags; | |
586 | ||
587 | /* | |
588 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
589 | * request head insertation from the workqueue. | |
590 | */ | |
e8064021 | 591 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
592 | |
593 | spin_lock_irqsave(&q->requeue_lock, flags); | |
594 | if (at_head) { | |
e8064021 | 595 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
596 | list_add(&rq->queuelist, &q->requeue_list); |
597 | } else { | |
598 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
599 | } | |
600 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
601 | |
602 | if (kick_requeue_list) | |
603 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
604 | } |
605 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
606 | ||
607 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
608 | { | |
2849450a | 609 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
610 | } |
611 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
612 | ||
2849450a MS |
613 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
614 | unsigned long msecs) | |
615 | { | |
616 | kblockd_schedule_delayed_work(&q->requeue_work, | |
617 | msecs_to_jiffies(msecs)); | |
618 | } | |
619 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
620 | ||
1885b24d JA |
621 | void blk_mq_abort_requeue_list(struct request_queue *q) |
622 | { | |
623 | unsigned long flags; | |
624 | LIST_HEAD(rq_list); | |
625 | ||
626 | spin_lock_irqsave(&q->requeue_lock, flags); | |
627 | list_splice_init(&q->requeue_list, &rq_list); | |
628 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
629 | ||
630 | while (!list_empty(&rq_list)) { | |
631 | struct request *rq; | |
632 | ||
633 | rq = list_first_entry(&rq_list, struct request, queuelist); | |
634 | list_del_init(&rq->queuelist); | |
635 | rq->errors = -EIO; | |
636 | blk_mq_end_request(rq, rq->errors); | |
637 | } | |
638 | } | |
639 | EXPORT_SYMBOL(blk_mq_abort_requeue_list); | |
640 | ||
0e62f51f JA |
641 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
642 | { | |
88c7b2b7 JA |
643 | if (tag < tags->nr_tags) { |
644 | prefetch(tags->rqs[tag]); | |
4ee86bab | 645 | return tags->rqs[tag]; |
88c7b2b7 | 646 | } |
4ee86bab HR |
647 | |
648 | return NULL; | |
24d2f903 CH |
649 | } |
650 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
651 | ||
320ae51f | 652 | struct blk_mq_timeout_data { |
46f92d42 CH |
653 | unsigned long next; |
654 | unsigned int next_set; | |
320ae51f JA |
655 | }; |
656 | ||
90415837 | 657 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 658 | { |
f8a5b122 | 659 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 660 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
661 | |
662 | /* | |
663 | * We know that complete is set at this point. If STARTED isn't set | |
664 | * anymore, then the request isn't active and the "timeout" should | |
665 | * just be ignored. This can happen due to the bitflag ordering. | |
666 | * Timeout first checks if STARTED is set, and if it is, assumes | |
667 | * the request is active. But if we race with completion, then | |
668 | * we both flags will get cleared. So check here again, and ignore | |
669 | * a timeout event with a request that isn't active. | |
670 | */ | |
46f92d42 CH |
671 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
672 | return; | |
87ee7b11 | 673 | |
46f92d42 | 674 | if (ops->timeout) |
0152fb6b | 675 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
676 | |
677 | switch (ret) { | |
678 | case BLK_EH_HANDLED: | |
679 | __blk_mq_complete_request(req); | |
680 | break; | |
681 | case BLK_EH_RESET_TIMER: | |
682 | blk_add_timer(req); | |
683 | blk_clear_rq_complete(req); | |
684 | break; | |
685 | case BLK_EH_NOT_HANDLED: | |
686 | break; | |
687 | default: | |
688 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
689 | break; | |
690 | } | |
87ee7b11 | 691 | } |
5b3f25fc | 692 | |
81481eb4 CH |
693 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
694 | struct request *rq, void *priv, bool reserved) | |
695 | { | |
696 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 697 | |
eb130dbf KB |
698 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
699 | /* | |
700 | * If a request wasn't started before the queue was | |
701 | * marked dying, kill it here or it'll go unnoticed. | |
702 | */ | |
a59e0f57 KB |
703 | if (unlikely(blk_queue_dying(rq->q))) { |
704 | rq->errors = -EIO; | |
705 | blk_mq_end_request(rq, rq->errors); | |
706 | } | |
46f92d42 | 707 | return; |
eb130dbf | 708 | } |
87ee7b11 | 709 | |
46f92d42 CH |
710 | if (time_after_eq(jiffies, rq->deadline)) { |
711 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 712 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
713 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
714 | data->next = rq->deadline; | |
715 | data->next_set = 1; | |
716 | } | |
87ee7b11 JA |
717 | } |
718 | ||
287922eb | 719 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 720 | { |
287922eb CH |
721 | struct request_queue *q = |
722 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
723 | struct blk_mq_timeout_data data = { |
724 | .next = 0, | |
725 | .next_set = 0, | |
726 | }; | |
81481eb4 | 727 | int i; |
320ae51f | 728 | |
71f79fb3 GKB |
729 | /* A deadlock might occur if a request is stuck requiring a |
730 | * timeout at the same time a queue freeze is waiting | |
731 | * completion, since the timeout code would not be able to | |
732 | * acquire the queue reference here. | |
733 | * | |
734 | * That's why we don't use blk_queue_enter here; instead, we use | |
735 | * percpu_ref_tryget directly, because we need to be able to | |
736 | * obtain a reference even in the short window between the queue | |
737 | * starting to freeze, by dropping the first reference in | |
738 | * blk_mq_freeze_queue_start, and the moment the last request is | |
739 | * consumed, marked by the instant q_usage_counter reaches | |
740 | * zero. | |
741 | */ | |
742 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
743 | return; |
744 | ||
0bf6cd5b | 745 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 746 | |
81481eb4 CH |
747 | if (data.next_set) { |
748 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
749 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 750 | } else { |
0bf6cd5b CH |
751 | struct blk_mq_hw_ctx *hctx; |
752 | ||
f054b56c ML |
753 | queue_for_each_hw_ctx(q, hctx, i) { |
754 | /* the hctx may be unmapped, so check it here */ | |
755 | if (blk_mq_hw_queue_mapped(hctx)) | |
756 | blk_mq_tag_idle(hctx); | |
757 | } | |
0d2602ca | 758 | } |
287922eb | 759 | blk_queue_exit(q); |
320ae51f JA |
760 | } |
761 | ||
762 | /* | |
763 | * Reverse check our software queue for entries that we could potentially | |
764 | * merge with. Currently includes a hand-wavy stop count of 8, to not spend | |
765 | * too much time checking for merges. | |
766 | */ | |
767 | static bool blk_mq_attempt_merge(struct request_queue *q, | |
768 | struct blk_mq_ctx *ctx, struct bio *bio) | |
769 | { | |
770 | struct request *rq; | |
771 | int checked = 8; | |
772 | ||
773 | list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) { | |
34fe7c05 | 774 | bool merged = false; |
320ae51f JA |
775 | |
776 | if (!checked--) | |
777 | break; | |
778 | ||
779 | if (!blk_rq_merge_ok(rq, bio)) | |
780 | continue; | |
781 | ||
34fe7c05 CH |
782 | switch (blk_try_merge(rq, bio)) { |
783 | case ELEVATOR_BACK_MERGE: | |
784 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
785 | merged = bio_attempt_back_merge(q, rq, bio); | |
bd166ef1 | 786 | break; |
34fe7c05 CH |
787 | case ELEVATOR_FRONT_MERGE: |
788 | if (blk_mq_sched_allow_merge(q, rq, bio)) | |
789 | merged = bio_attempt_front_merge(q, rq, bio); | |
320ae51f | 790 | break; |
1e739730 CH |
791 | case ELEVATOR_DISCARD_MERGE: |
792 | merged = bio_attempt_discard_merge(q, rq, bio); | |
320ae51f | 793 | break; |
34fe7c05 CH |
794 | default: |
795 | continue; | |
320ae51f | 796 | } |
34fe7c05 CH |
797 | |
798 | if (merged) | |
799 | ctx->rq_merged++; | |
800 | return merged; | |
320ae51f JA |
801 | } |
802 | ||
803 | return false; | |
804 | } | |
805 | ||
88459642 OS |
806 | struct flush_busy_ctx_data { |
807 | struct blk_mq_hw_ctx *hctx; | |
808 | struct list_head *list; | |
809 | }; | |
810 | ||
811 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
812 | { | |
813 | struct flush_busy_ctx_data *flush_data = data; | |
814 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
815 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
816 | ||
817 | sbitmap_clear_bit(sb, bitnr); | |
818 | spin_lock(&ctx->lock); | |
819 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
820 | spin_unlock(&ctx->lock); | |
821 | return true; | |
822 | } | |
823 | ||
1429d7c9 JA |
824 | /* |
825 | * Process software queues that have been marked busy, splicing them | |
826 | * to the for-dispatch | |
827 | */ | |
2c3ad667 | 828 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 829 | { |
88459642 OS |
830 | struct flush_busy_ctx_data data = { |
831 | .hctx = hctx, | |
832 | .list = list, | |
833 | }; | |
1429d7c9 | 834 | |
88459642 | 835 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 836 | } |
2c3ad667 | 837 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 838 | |
703fd1c0 JA |
839 | static inline unsigned int queued_to_index(unsigned int queued) |
840 | { | |
841 | if (!queued) | |
842 | return 0; | |
1429d7c9 | 843 | |
703fd1c0 | 844 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
845 | } |
846 | ||
bd6737f1 JA |
847 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
848 | bool wait) | |
bd166ef1 JA |
849 | { |
850 | struct blk_mq_alloc_data data = { | |
851 | .q = rq->q, | |
bd166ef1 JA |
852 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
853 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
854 | }; | |
855 | ||
bd166ef1 JA |
856 | if (rq->tag != -1) { |
857 | done: | |
858 | if (hctx) | |
859 | *hctx = data.hctx; | |
860 | return true; | |
861 | } | |
862 | ||
415b806d SG |
863 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
864 | data.flags |= BLK_MQ_REQ_RESERVED; | |
865 | ||
bd166ef1 JA |
866 | rq->tag = blk_mq_get_tag(&data); |
867 | if (rq->tag >= 0) { | |
200e86b3 JA |
868 | if (blk_mq_tag_busy(data.hctx)) { |
869 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
870 | atomic_inc(&data.hctx->nr_active); | |
871 | } | |
bd166ef1 JA |
872 | data.hctx->tags->rqs[rq->tag] = rq; |
873 | goto done; | |
874 | } | |
875 | ||
876 | return false; | |
877 | } | |
878 | ||
113285b4 JA |
879 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
880 | struct request *rq) | |
99cf1dc5 | 881 | { |
99cf1dc5 JA |
882 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
883 | rq->tag = -1; | |
884 | ||
885 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
886 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
887 | atomic_dec(&hctx->nr_active); | |
888 | } | |
889 | } | |
890 | ||
113285b4 JA |
891 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
892 | struct request *rq) | |
893 | { | |
894 | if (rq->tag == -1 || rq->internal_tag == -1) | |
895 | return; | |
896 | ||
897 | __blk_mq_put_driver_tag(hctx, rq); | |
898 | } | |
899 | ||
900 | static void blk_mq_put_driver_tag(struct request *rq) | |
901 | { | |
902 | struct blk_mq_hw_ctx *hctx; | |
903 | ||
904 | if (rq->tag == -1 || rq->internal_tag == -1) | |
905 | return; | |
906 | ||
907 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
908 | __blk_mq_put_driver_tag(hctx, rq); | |
909 | } | |
910 | ||
bd166ef1 JA |
911 | /* |
912 | * If we fail getting a driver tag because all the driver tags are already | |
913 | * assigned and on the dispatch list, BUT the first entry does not have a | |
914 | * tag, then we could deadlock. For that case, move entries with assigned | |
915 | * driver tags to the front, leaving the set of tagged requests in the | |
916 | * same order, and the untagged set in the same order. | |
917 | */ | |
918 | static bool reorder_tags_to_front(struct list_head *list) | |
919 | { | |
920 | struct request *rq, *tmp, *first = NULL; | |
921 | ||
922 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
923 | if (rq == first) | |
924 | break; | |
925 | if (rq->tag != -1) { | |
926 | list_move(&rq->queuelist, list); | |
927 | if (!first) | |
928 | first = rq; | |
929 | } | |
930 | } | |
931 | ||
932 | return first != NULL; | |
933 | } | |
934 | ||
da55f2cc OS |
935 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
936 | void *key) | |
937 | { | |
938 | struct blk_mq_hw_ctx *hctx; | |
939 | ||
940 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
941 | ||
942 | list_del(&wait->task_list); | |
943 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
944 | blk_mq_run_hw_queue(hctx, true); | |
945 | return 1; | |
946 | } | |
947 | ||
948 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
949 | { | |
950 | struct sbq_wait_state *ws; | |
951 | ||
952 | /* | |
953 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
954 | * The thread which wins the race to grab this bit adds the hardware | |
955 | * queue to the wait queue. | |
956 | */ | |
957 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
958 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
959 | return false; | |
960 | ||
961 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
962 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
963 | ||
964 | /* | |
965 | * As soon as this returns, it's no longer safe to fiddle with | |
966 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
967 | * and unlock the bit. | |
968 | */ | |
969 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
970 | return true; | |
971 | } | |
972 | ||
f04c3df3 | 973 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
320ae51f JA |
974 | { |
975 | struct request_queue *q = hctx->queue; | |
320ae51f | 976 | struct request *rq; |
74c45052 JA |
977 | LIST_HEAD(driver_list); |
978 | struct list_head *dptr; | |
f04c3df3 | 979 | int queued, ret = BLK_MQ_RQ_QUEUE_OK; |
320ae51f | 980 | |
74c45052 JA |
981 | /* |
982 | * Start off with dptr being NULL, so we start the first request | |
983 | * immediately, even if we have more pending. | |
984 | */ | |
985 | dptr = NULL; | |
986 | ||
320ae51f JA |
987 | /* |
988 | * Now process all the entries, sending them to the driver. | |
989 | */ | |
1429d7c9 | 990 | queued = 0; |
f04c3df3 | 991 | while (!list_empty(list)) { |
74c45052 | 992 | struct blk_mq_queue_data bd; |
320ae51f | 993 | |
f04c3df3 | 994 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
995 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
996 | if (!queued && reorder_tags_to_front(list)) | |
997 | continue; | |
3c782d67 JA |
998 | |
999 | /* | |
da55f2cc OS |
1000 | * The initial allocation attempt failed, so we need to |
1001 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1002 | */ |
da55f2cc OS |
1003 | if (blk_mq_dispatch_wait_add(hctx)) { |
1004 | /* | |
1005 | * It's possible that a tag was freed in the | |
1006 | * window between the allocation failure and | |
1007 | * adding the hardware queue to the wait queue. | |
1008 | */ | |
1009 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
1010 | break; | |
1011 | } else { | |
3c782d67 | 1012 | break; |
da55f2cc | 1013 | } |
bd166ef1 | 1014 | } |
da55f2cc | 1015 | |
320ae51f | 1016 | list_del_init(&rq->queuelist); |
320ae51f | 1017 | |
74c45052 JA |
1018 | bd.rq = rq; |
1019 | bd.list = dptr; | |
113285b4 JA |
1020 | |
1021 | /* | |
1022 | * Flag last if we have no more requests, or if we have more | |
1023 | * but can't assign a driver tag to it. | |
1024 | */ | |
1025 | if (list_empty(list)) | |
1026 | bd.last = true; | |
1027 | else { | |
1028 | struct request *nxt; | |
1029 | ||
1030 | nxt = list_first_entry(list, struct request, queuelist); | |
1031 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1032 | } | |
74c45052 JA |
1033 | |
1034 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
320ae51f JA |
1035 | switch (ret) { |
1036 | case BLK_MQ_RQ_QUEUE_OK: | |
1037 | queued++; | |
52b9c330 | 1038 | break; |
320ae51f | 1039 | case BLK_MQ_RQ_QUEUE_BUSY: |
113285b4 | 1040 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1041 | list_add(&rq->queuelist, list); |
ed0791b2 | 1042 | __blk_mq_requeue_request(rq); |
320ae51f JA |
1043 | break; |
1044 | default: | |
1045 | pr_err("blk-mq: bad return on queue: %d\n", ret); | |
320ae51f | 1046 | case BLK_MQ_RQ_QUEUE_ERROR: |
1e93b8c2 | 1047 | rq->errors = -EIO; |
c8a446ad | 1048 | blk_mq_end_request(rq, rq->errors); |
320ae51f JA |
1049 | break; |
1050 | } | |
1051 | ||
1052 | if (ret == BLK_MQ_RQ_QUEUE_BUSY) | |
1053 | break; | |
74c45052 JA |
1054 | |
1055 | /* | |
1056 | * We've done the first request. If we have more than 1 | |
1057 | * left in the list, set dptr to defer issue. | |
1058 | */ | |
f04c3df3 | 1059 | if (!dptr && list->next != list->prev) |
74c45052 | 1060 | dptr = &driver_list; |
320ae51f JA |
1061 | } |
1062 | ||
703fd1c0 | 1063 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1064 | |
1065 | /* | |
1066 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1067 | * that is where we will continue on next queue run. | |
1068 | */ | |
f04c3df3 | 1069 | if (!list_empty(list)) { |
113285b4 JA |
1070 | /* |
1071 | * If we got a driver tag for the next request already, | |
1072 | * free it again. | |
1073 | */ | |
1074 | rq = list_first_entry(list, struct request, queuelist); | |
1075 | blk_mq_put_driver_tag(rq); | |
1076 | ||
320ae51f | 1077 | spin_lock(&hctx->lock); |
c13660a0 | 1078 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1079 | spin_unlock(&hctx->lock); |
f04c3df3 | 1080 | |
9ba52e58 SL |
1081 | /* |
1082 | * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but | |
1083 | * it's possible the queue is stopped and restarted again | |
1084 | * before this. Queue restart will dispatch requests. And since | |
1085 | * requests in rq_list aren't added into hctx->dispatch yet, | |
1086 | * the requests in rq_list might get lost. | |
1087 | * | |
1088 | * blk_mq_run_hw_queue() already checks the STOPPED bit | |
bd166ef1 | 1089 | * |
da55f2cc OS |
1090 | * If RESTART or TAG_WAITING is set, then let completion restart |
1091 | * the queue instead of potentially looping here. | |
bd166ef1 | 1092 | */ |
da55f2cc OS |
1093 | if (!blk_mq_sched_needs_restart(hctx) && |
1094 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1095 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1096 | } |
f04c3df3 | 1097 | |
2aa0f21d | 1098 | return queued != 0; |
f04c3df3 JA |
1099 | } |
1100 | ||
6a83e74d BVA |
1101 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1102 | { | |
1103 | int srcu_idx; | |
1104 | ||
1105 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1106 | cpu_online(hctx->next_cpu)); | |
1107 | ||
1108 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1109 | rcu_read_lock(); | |
bd166ef1 | 1110 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1111 | rcu_read_unlock(); |
1112 | } else { | |
1113 | srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu); | |
bd166ef1 | 1114 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1115 | srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx); |
1116 | } | |
1117 | } | |
1118 | ||
506e931f JA |
1119 | /* |
1120 | * It'd be great if the workqueue API had a way to pass | |
1121 | * in a mask and had some smarts for more clever placement. | |
1122 | * For now we just round-robin here, switching for every | |
1123 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1124 | */ | |
1125 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1126 | { | |
b657d7e6 CH |
1127 | if (hctx->queue->nr_hw_queues == 1) |
1128 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1129 | |
1130 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1131 | int next_cpu; |
506e931f JA |
1132 | |
1133 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1134 | if (next_cpu >= nr_cpu_ids) | |
1135 | next_cpu = cpumask_first(hctx->cpumask); | |
1136 | ||
1137 | hctx->next_cpu = next_cpu; | |
1138 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1139 | } | |
1140 | ||
b657d7e6 | 1141 | return hctx->next_cpu; |
506e931f JA |
1142 | } |
1143 | ||
320ae51f JA |
1144 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1145 | { | |
5d1b25c1 BVA |
1146 | if (unlikely(blk_mq_hctx_stopped(hctx) || |
1147 | !blk_mq_hw_queue_mapped(hctx))) | |
320ae51f JA |
1148 | return; |
1149 | ||
1b792f2f | 1150 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1151 | int cpu = get_cpu(); |
1152 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1153 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1154 | put_cpu(); |
398205b8 PB |
1155 | return; |
1156 | } | |
e4043dcf | 1157 | |
2a90d4aa | 1158 | put_cpu(); |
e4043dcf | 1159 | } |
398205b8 | 1160 | |
27489a3c | 1161 | kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work); |
320ae51f JA |
1162 | } |
1163 | ||
b94ec296 | 1164 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1165 | { |
1166 | struct blk_mq_hw_ctx *hctx; | |
1167 | int i; | |
1168 | ||
1169 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1170 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1171 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1172 | continue; |
1173 | ||
b94ec296 | 1174 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1175 | } |
1176 | } | |
b94ec296 | 1177 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1178 | |
fd001443 BVA |
1179 | /** |
1180 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1181 | * @q: request queue. | |
1182 | * | |
1183 | * The caller is responsible for serializing this function against | |
1184 | * blk_mq_{start,stop}_hw_queue(). | |
1185 | */ | |
1186 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1187 | { | |
1188 | struct blk_mq_hw_ctx *hctx; | |
1189 | int i; | |
1190 | ||
1191 | queue_for_each_hw_ctx(q, hctx, i) | |
1192 | if (blk_mq_hctx_stopped(hctx)) | |
1193 | return true; | |
1194 | ||
1195 | return false; | |
1196 | } | |
1197 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1198 | ||
320ae51f JA |
1199 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1200 | { | |
27489a3c | 1201 | cancel_work(&hctx->run_work); |
70f4db63 | 1202 | cancel_delayed_work(&hctx->delay_work); |
320ae51f JA |
1203 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
1204 | } | |
1205 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1206 | ||
280d45f6 CH |
1207 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1208 | { | |
1209 | struct blk_mq_hw_ctx *hctx; | |
1210 | int i; | |
1211 | ||
1212 | queue_for_each_hw_ctx(q, hctx, i) | |
1213 | blk_mq_stop_hw_queue(hctx); | |
1214 | } | |
1215 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1216 | ||
320ae51f JA |
1217 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1218 | { | |
1219 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1220 | |
0ffbce80 | 1221 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1222 | } |
1223 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1224 | ||
2f268556 CH |
1225 | void blk_mq_start_hw_queues(struct request_queue *q) |
1226 | { | |
1227 | struct blk_mq_hw_ctx *hctx; | |
1228 | int i; | |
1229 | ||
1230 | queue_for_each_hw_ctx(q, hctx, i) | |
1231 | blk_mq_start_hw_queue(hctx); | |
1232 | } | |
1233 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1234 | ||
ae911c5e JA |
1235 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1236 | { | |
1237 | if (!blk_mq_hctx_stopped(hctx)) | |
1238 | return; | |
1239 | ||
1240 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1241 | blk_mq_run_hw_queue(hctx, async); | |
1242 | } | |
1243 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1244 | ||
1b4a3258 | 1245 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1246 | { |
1247 | struct blk_mq_hw_ctx *hctx; | |
1248 | int i; | |
1249 | ||
ae911c5e JA |
1250 | queue_for_each_hw_ctx(q, hctx, i) |
1251 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1252 | } |
1253 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1254 | ||
70f4db63 | 1255 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1256 | { |
1257 | struct blk_mq_hw_ctx *hctx; | |
1258 | ||
27489a3c | 1259 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work); |
e4043dcf | 1260 | |
320ae51f JA |
1261 | __blk_mq_run_hw_queue(hctx); |
1262 | } | |
1263 | ||
70f4db63 CH |
1264 | static void blk_mq_delay_work_fn(struct work_struct *work) |
1265 | { | |
1266 | struct blk_mq_hw_ctx *hctx; | |
1267 | ||
1268 | hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work); | |
1269 | ||
1270 | if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1271 | __blk_mq_run_hw_queue(hctx); | |
1272 | } | |
1273 | ||
1274 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1275 | { | |
19c66e59 ML |
1276 | if (unlikely(!blk_mq_hw_queue_mapped(hctx))) |
1277 | return; | |
70f4db63 | 1278 | |
7e79dadc | 1279 | blk_mq_stop_hw_queue(hctx); |
b657d7e6 CH |
1280 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1281 | &hctx->delay_work, msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1282 | } |
1283 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1284 | ||
cfd0c552 | 1285 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1286 | struct request *rq, |
1287 | bool at_head) | |
320ae51f | 1288 | { |
e57690fe JA |
1289 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1290 | ||
01b983c9 JA |
1291 | trace_block_rq_insert(hctx->queue, rq); |
1292 | ||
72a0a36e CH |
1293 | if (at_head) |
1294 | list_add(&rq->queuelist, &ctx->rq_list); | |
1295 | else | |
1296 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1297 | } |
4bb659b1 | 1298 | |
2c3ad667 JA |
1299 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1300 | bool at_head) | |
cfd0c552 ML |
1301 | { |
1302 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1303 | ||
e57690fe | 1304 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1305 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1306 | } |
1307 | ||
bd166ef1 JA |
1308 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1309 | struct list_head *list) | |
320ae51f JA |
1310 | |
1311 | { | |
320ae51f JA |
1312 | /* |
1313 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1314 | * offline now | |
1315 | */ | |
1316 | spin_lock(&ctx->lock); | |
1317 | while (!list_empty(list)) { | |
1318 | struct request *rq; | |
1319 | ||
1320 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1321 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1322 | list_del_init(&rq->queuelist); |
e57690fe | 1323 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1324 | } |
cfd0c552 | 1325 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1326 | spin_unlock(&ctx->lock); |
320ae51f JA |
1327 | } |
1328 | ||
1329 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1330 | { | |
1331 | struct request *rqa = container_of(a, struct request, queuelist); | |
1332 | struct request *rqb = container_of(b, struct request, queuelist); | |
1333 | ||
1334 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1335 | (rqa->mq_ctx == rqb->mq_ctx && | |
1336 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1337 | } | |
1338 | ||
1339 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1340 | { | |
1341 | struct blk_mq_ctx *this_ctx; | |
1342 | struct request_queue *this_q; | |
1343 | struct request *rq; | |
1344 | LIST_HEAD(list); | |
1345 | LIST_HEAD(ctx_list); | |
1346 | unsigned int depth; | |
1347 | ||
1348 | list_splice_init(&plug->mq_list, &list); | |
1349 | ||
1350 | list_sort(NULL, &list, plug_ctx_cmp); | |
1351 | ||
1352 | this_q = NULL; | |
1353 | this_ctx = NULL; | |
1354 | depth = 0; | |
1355 | ||
1356 | while (!list_empty(&list)) { | |
1357 | rq = list_entry_rq(list.next); | |
1358 | list_del_init(&rq->queuelist); | |
1359 | BUG_ON(!rq->q); | |
1360 | if (rq->mq_ctx != this_ctx) { | |
1361 | if (this_ctx) { | |
bd166ef1 JA |
1362 | trace_block_unplug(this_q, depth, from_schedule); |
1363 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1364 | &ctx_list, | |
1365 | from_schedule); | |
320ae51f JA |
1366 | } |
1367 | ||
1368 | this_ctx = rq->mq_ctx; | |
1369 | this_q = rq->q; | |
1370 | depth = 0; | |
1371 | } | |
1372 | ||
1373 | depth++; | |
1374 | list_add_tail(&rq->queuelist, &ctx_list); | |
1375 | } | |
1376 | ||
1377 | /* | |
1378 | * If 'this_ctx' is set, we know we have entries to complete | |
1379 | * on 'ctx_list'. Do those. | |
1380 | */ | |
1381 | if (this_ctx) { | |
bd166ef1 JA |
1382 | trace_block_unplug(this_q, depth, from_schedule); |
1383 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1384 | from_schedule); | |
320ae51f JA |
1385 | } |
1386 | } | |
1387 | ||
1388 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1389 | { | |
1390 | init_request_from_bio(rq, bio); | |
4b570521 | 1391 | |
6e85eaf3 | 1392 | blk_account_io_start(rq, true); |
320ae51f JA |
1393 | } |
1394 | ||
274a5843 JA |
1395 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1396 | { | |
1397 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1398 | !blk_queue_nomerges(hctx->queue); | |
1399 | } | |
1400 | ||
07068d5b JA |
1401 | static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx, |
1402 | struct blk_mq_ctx *ctx, | |
1403 | struct request *rq, struct bio *bio) | |
320ae51f | 1404 | { |
e18378a6 | 1405 | if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) { |
07068d5b JA |
1406 | blk_mq_bio_to_request(rq, bio); |
1407 | spin_lock(&ctx->lock); | |
1408 | insert_rq: | |
1409 | __blk_mq_insert_request(hctx, rq, false); | |
1410 | spin_unlock(&ctx->lock); | |
1411 | return false; | |
1412 | } else { | |
274a5843 JA |
1413 | struct request_queue *q = hctx->queue; |
1414 | ||
07068d5b JA |
1415 | spin_lock(&ctx->lock); |
1416 | if (!blk_mq_attempt_merge(q, ctx, bio)) { | |
1417 | blk_mq_bio_to_request(rq, bio); | |
1418 | goto insert_rq; | |
1419 | } | |
320ae51f | 1420 | |
07068d5b | 1421 | spin_unlock(&ctx->lock); |
bd166ef1 | 1422 | __blk_mq_finish_request(hctx, ctx, rq); |
07068d5b | 1423 | return true; |
14ec77f3 | 1424 | } |
07068d5b | 1425 | } |
14ec77f3 | 1426 | |
fd2d3326 JA |
1427 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1428 | { | |
bd166ef1 JA |
1429 | if (rq->tag != -1) |
1430 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1431 | ||
1432 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1433 | } |
1434 | ||
066a4a73 | 1435 | static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie) |
f984df1f | 1436 | { |
f984df1f | 1437 | struct request_queue *q = rq->q; |
f984df1f SL |
1438 | struct blk_mq_queue_data bd = { |
1439 | .rq = rq, | |
1440 | .list = NULL, | |
1441 | .last = 1 | |
1442 | }; | |
bd166ef1 JA |
1443 | struct blk_mq_hw_ctx *hctx; |
1444 | blk_qc_t new_cookie; | |
1445 | int ret; | |
f984df1f | 1446 | |
bd166ef1 | 1447 | if (q->elevator) |
2253efc8 BVA |
1448 | goto insert; |
1449 | ||
bd166ef1 JA |
1450 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) |
1451 | goto insert; | |
1452 | ||
1453 | new_cookie = request_to_qc_t(hctx, rq); | |
1454 | ||
f984df1f SL |
1455 | /* |
1456 | * For OK queue, we are done. For error, kill it. Any other | |
1457 | * error (busy), just add it to our list as we previously | |
1458 | * would have done | |
1459 | */ | |
1460 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
7b371636 JA |
1461 | if (ret == BLK_MQ_RQ_QUEUE_OK) { |
1462 | *cookie = new_cookie; | |
2253efc8 | 1463 | return; |
7b371636 | 1464 | } |
f984df1f | 1465 | |
7b371636 JA |
1466 | __blk_mq_requeue_request(rq); |
1467 | ||
1468 | if (ret == BLK_MQ_RQ_QUEUE_ERROR) { | |
1469 | *cookie = BLK_QC_T_NONE; | |
1470 | rq->errors = -EIO; | |
1471 | blk_mq_end_request(rq, rq->errors); | |
2253efc8 | 1472 | return; |
f984df1f | 1473 | } |
7b371636 | 1474 | |
2253efc8 | 1475 | insert: |
bd6737f1 | 1476 | blk_mq_sched_insert_request(rq, false, true, true, false); |
f984df1f SL |
1477 | } |
1478 | ||
07068d5b JA |
1479 | /* |
1480 | * Multiple hardware queue variant. This will not use per-process plugs, | |
1481 | * but will attempt to bypass the hctx queueing if we can go straight to | |
1482 | * hardware for SYNC IO. | |
1483 | */ | |
dece1635 | 1484 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1485 | { |
ef295ecf | 1486 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1487 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1488 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1489 | struct request *rq; |
6a83e74d | 1490 | unsigned int request_count = 0, srcu_idx; |
f984df1f | 1491 | struct blk_plug *plug; |
5b3f341f | 1492 | struct request *same_queue_rq = NULL; |
7b371636 | 1493 | blk_qc_t cookie; |
87760e5e | 1494 | unsigned int wb_acct; |
07068d5b JA |
1495 | |
1496 | blk_queue_bounce(q, &bio); | |
1497 | ||
1498 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1499 | bio_io_error(bio); |
dece1635 | 1500 | return BLK_QC_T_NONE; |
07068d5b JA |
1501 | } |
1502 | ||
54efd50b KO |
1503 | blk_queue_split(q, &bio, q->bio_split); |
1504 | ||
87c279e6 OS |
1505 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1506 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1507 | return BLK_QC_T_NONE; | |
f984df1f | 1508 | |
bd166ef1 JA |
1509 | if (blk_mq_sched_bio_merge(q, bio)) |
1510 | return BLK_QC_T_NONE; | |
1511 | ||
87760e5e JA |
1512 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1513 | ||
bd166ef1 JA |
1514 | trace_block_getrq(q, bio, bio->bi_opf); |
1515 | ||
1516 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1517 | if (unlikely(!rq)) { |
1518 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1519 | return BLK_QC_T_NONE; |
87760e5e JA |
1520 | } |
1521 | ||
1522 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1523 | |
fd2d3326 | 1524 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b JA |
1525 | |
1526 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1527 | if (q->elevator) |
1528 | goto elv_insert; | |
07068d5b JA |
1529 | blk_mq_bio_to_request(rq, bio); |
1530 | blk_insert_flush(rq); | |
0c2a6fe4 | 1531 | goto run_queue; |
07068d5b JA |
1532 | } |
1533 | ||
f984df1f | 1534 | plug = current->plug; |
e167dfb5 JA |
1535 | /* |
1536 | * If the driver supports defer issued based on 'last', then | |
1537 | * queue it up like normal since we can potentially save some | |
1538 | * CPU this way. | |
1539 | */ | |
f984df1f SL |
1540 | if (((plug && !blk_queue_nomerges(q)) || is_sync) && |
1541 | !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) { | |
1542 | struct request *old_rq = NULL; | |
07068d5b JA |
1543 | |
1544 | blk_mq_bio_to_request(rq, bio); | |
07068d5b JA |
1545 | |
1546 | /* | |
6a83e74d | 1547 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1548 | * Otherwise the existing request in the plug list will be |
1549 | * issued. So the plug list will have one request at most | |
07068d5b | 1550 | */ |
f984df1f | 1551 | if (plug) { |
5b3f341f SL |
1552 | /* |
1553 | * The plug list might get flushed before this. If that | |
b094f89c JA |
1554 | * happens, same_queue_rq is invalid and plug list is |
1555 | * empty | |
1556 | */ | |
5b3f341f SL |
1557 | if (same_queue_rq && !list_empty(&plug->mq_list)) { |
1558 | old_rq = same_queue_rq; | |
f984df1f | 1559 | list_del_init(&old_rq->queuelist); |
07068d5b | 1560 | } |
f984df1f SL |
1561 | list_add_tail(&rq->queuelist, &plug->mq_list); |
1562 | } else /* is_sync */ | |
1563 | old_rq = rq; | |
1564 | blk_mq_put_ctx(data.ctx); | |
1565 | if (!old_rq) | |
7b371636 | 1566 | goto done; |
6a83e74d BVA |
1567 | |
1568 | if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1569 | rcu_read_lock(); | |
066a4a73 | 1570 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1571 | rcu_read_unlock(); |
1572 | } else { | |
1573 | srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu); | |
066a4a73 | 1574 | blk_mq_try_issue_directly(old_rq, &cookie); |
6a83e74d BVA |
1575 | srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx); |
1576 | } | |
7b371636 | 1577 | goto done; |
07068d5b JA |
1578 | } |
1579 | ||
bd166ef1 | 1580 | if (q->elevator) { |
0c2a6fe4 | 1581 | elv_insert: |
bd166ef1 JA |
1582 | blk_mq_put_ctx(data.ctx); |
1583 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1584 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1585 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1586 | goto done; |
1587 | } | |
07068d5b JA |
1588 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1589 | /* | |
1590 | * For a SYNC request, send it to the hardware immediately. For | |
1591 | * an ASYNC request, just ensure that we run it later on. The | |
1592 | * latter allows for merging opportunities and more efficient | |
1593 | * dispatching. | |
1594 | */ | |
0c2a6fe4 | 1595 | run_queue: |
07068d5b JA |
1596 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
1597 | } | |
07068d5b | 1598 | blk_mq_put_ctx(data.ctx); |
7b371636 JA |
1599 | done: |
1600 | return cookie; | |
07068d5b JA |
1601 | } |
1602 | ||
1603 | /* | |
1604 | * Single hardware queue variant. This will attempt to use any per-process | |
1605 | * plug for merging and IO deferral. | |
1606 | */ | |
dece1635 | 1607 | static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1608 | { |
ef295ecf | 1609 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1610 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
e6c4438b JM |
1611 | struct blk_plug *plug; |
1612 | unsigned int request_count = 0; | |
5a797e00 | 1613 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1614 | struct request *rq; |
7b371636 | 1615 | blk_qc_t cookie; |
87760e5e | 1616 | unsigned int wb_acct; |
07068d5b | 1617 | |
07068d5b JA |
1618 | blk_queue_bounce(q, &bio); |
1619 | ||
1620 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { | |
4246a0b6 | 1621 | bio_io_error(bio); |
dece1635 | 1622 | return BLK_QC_T_NONE; |
07068d5b JA |
1623 | } |
1624 | ||
54efd50b KO |
1625 | blk_queue_split(q, &bio, q->bio_split); |
1626 | ||
87c279e6 OS |
1627 | if (!is_flush_fua && !blk_queue_nomerges(q)) { |
1628 | if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) | |
1629 | return BLK_QC_T_NONE; | |
1630 | } else | |
1631 | request_count = blk_plug_queued_count(q); | |
07068d5b | 1632 | |
bd166ef1 JA |
1633 | if (blk_mq_sched_bio_merge(q, bio)) |
1634 | return BLK_QC_T_NONE; | |
1635 | ||
87760e5e JA |
1636 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1637 | ||
bd166ef1 JA |
1638 | trace_block_getrq(q, bio, bio->bi_opf); |
1639 | ||
1640 | rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data); | |
87760e5e JA |
1641 | if (unlikely(!rq)) { |
1642 | __wbt_done(q->rq_wb, wb_acct); | |
dece1635 | 1643 | return BLK_QC_T_NONE; |
87760e5e JA |
1644 | } |
1645 | ||
1646 | wbt_track(&rq->issue_stat, wb_acct); | |
320ae51f | 1647 | |
fd2d3326 | 1648 | cookie = request_to_qc_t(data.hctx, rq); |
320ae51f JA |
1649 | |
1650 | if (unlikely(is_flush_fua)) { | |
0c2a6fe4 JA |
1651 | if (q->elevator) |
1652 | goto elv_insert; | |
320ae51f | 1653 | blk_mq_bio_to_request(rq, bio); |
320ae51f | 1654 | blk_insert_flush(rq); |
0c2a6fe4 | 1655 | goto run_queue; |
320ae51f JA |
1656 | } |
1657 | ||
1658 | /* | |
1659 | * A task plug currently exists. Since this is completely lockless, | |
1660 | * utilize that to temporarily store requests until the task is | |
1661 | * either done or scheduled away. | |
1662 | */ | |
e6c4438b JM |
1663 | plug = current->plug; |
1664 | if (plug) { | |
600271d9 SL |
1665 | struct request *last = NULL; |
1666 | ||
e6c4438b | 1667 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1668 | |
1669 | /* | |
1670 | * @request_count may become stale because of schedule | |
1671 | * out, so check the list again. | |
1672 | */ | |
1673 | if (list_empty(&plug->mq_list)) | |
1674 | request_count = 0; | |
676d0607 | 1675 | if (!request_count) |
e6c4438b | 1676 | trace_block_plug(q); |
600271d9 SL |
1677 | else |
1678 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c JA |
1679 | |
1680 | blk_mq_put_ctx(data.ctx); | |
1681 | ||
600271d9 SL |
1682 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1683 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1684 | blk_flush_plug_list(plug, false); |
1685 | trace_block_plug(q); | |
320ae51f | 1686 | } |
b094f89c | 1687 | |
e6c4438b | 1688 | list_add_tail(&rq->queuelist, &plug->mq_list); |
7b371636 | 1689 | return cookie; |
320ae51f JA |
1690 | } |
1691 | ||
bd166ef1 | 1692 | if (q->elevator) { |
0c2a6fe4 | 1693 | elv_insert: |
bd166ef1 JA |
1694 | blk_mq_put_ctx(data.ctx); |
1695 | blk_mq_bio_to_request(rq, bio); | |
0abad774 | 1696 | blk_mq_sched_insert_request(rq, false, true, |
bd6737f1 | 1697 | !is_sync || is_flush_fua, true); |
bd166ef1 JA |
1698 | goto done; |
1699 | } | |
07068d5b JA |
1700 | if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) { |
1701 | /* | |
1702 | * For a SYNC request, send it to the hardware immediately. For | |
1703 | * an ASYNC request, just ensure that we run it later on. The | |
1704 | * latter allows for merging opportunities and more efficient | |
1705 | * dispatching. | |
1706 | */ | |
0c2a6fe4 | 1707 | run_queue: |
07068d5b | 1708 | blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua); |
320ae51f JA |
1709 | } |
1710 | ||
07068d5b | 1711 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1712 | done: |
7b371636 | 1713 | return cookie; |
320ae51f JA |
1714 | } |
1715 | ||
cc71a6f4 JA |
1716 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1717 | unsigned int hctx_idx) | |
95363efd | 1718 | { |
e9b267d9 | 1719 | struct page *page; |
320ae51f | 1720 | |
24d2f903 | 1721 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1722 | int i; |
320ae51f | 1723 | |
24d2f903 | 1724 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1725 | struct request *rq = tags->static_rqs[i]; |
1726 | ||
1727 | if (!rq) | |
e9b267d9 | 1728 | continue; |
2af8cbe3 | 1729 | set->ops->exit_request(set->driver_data, rq, |
24d2f903 | 1730 | hctx_idx, i); |
2af8cbe3 | 1731 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1732 | } |
320ae51f | 1733 | } |
320ae51f | 1734 | |
24d2f903 CH |
1735 | while (!list_empty(&tags->page_list)) { |
1736 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1737 | list_del_init(&page->lru); |
f75782e4 CM |
1738 | /* |
1739 | * Remove kmemleak object previously allocated in | |
1740 | * blk_mq_init_rq_map(). | |
1741 | */ | |
1742 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1743 | __free_pages(page, page->private); |
1744 | } | |
cc71a6f4 | 1745 | } |
320ae51f | 1746 | |
cc71a6f4 JA |
1747 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1748 | { | |
24d2f903 | 1749 | kfree(tags->rqs); |
cc71a6f4 | 1750 | tags->rqs = NULL; |
2af8cbe3 JA |
1751 | kfree(tags->static_rqs); |
1752 | tags->static_rqs = NULL; | |
320ae51f | 1753 | |
24d2f903 | 1754 | blk_mq_free_tags(tags); |
320ae51f JA |
1755 | } |
1756 | ||
cc71a6f4 JA |
1757 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1758 | unsigned int hctx_idx, | |
1759 | unsigned int nr_tags, | |
1760 | unsigned int reserved_tags) | |
320ae51f | 1761 | { |
24d2f903 | 1762 | struct blk_mq_tags *tags; |
59f082e4 | 1763 | int node; |
320ae51f | 1764 | |
59f082e4 SL |
1765 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1766 | if (node == NUMA_NO_NODE) | |
1767 | node = set->numa_node; | |
1768 | ||
1769 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1770 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1771 | if (!tags) |
1772 | return NULL; | |
320ae51f | 1773 | |
cc71a6f4 | 1774 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1775 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1776 | node); |
24d2f903 CH |
1777 | if (!tags->rqs) { |
1778 | blk_mq_free_tags(tags); | |
1779 | return NULL; | |
1780 | } | |
320ae51f | 1781 | |
2af8cbe3 JA |
1782 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1783 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1784 | node); |
2af8cbe3 JA |
1785 | if (!tags->static_rqs) { |
1786 | kfree(tags->rqs); | |
1787 | blk_mq_free_tags(tags); | |
1788 | return NULL; | |
1789 | } | |
1790 | ||
cc71a6f4 JA |
1791 | return tags; |
1792 | } | |
1793 | ||
1794 | static size_t order_to_size(unsigned int order) | |
1795 | { | |
1796 | return (size_t)PAGE_SIZE << order; | |
1797 | } | |
1798 | ||
1799 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1800 | unsigned int hctx_idx, unsigned int depth) | |
1801 | { | |
1802 | unsigned int i, j, entries_per_page, max_order = 4; | |
1803 | size_t rq_size, left; | |
59f082e4 SL |
1804 | int node; |
1805 | ||
1806 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1807 | if (node == NUMA_NO_NODE) | |
1808 | node = set->numa_node; | |
cc71a6f4 JA |
1809 | |
1810 | INIT_LIST_HEAD(&tags->page_list); | |
1811 | ||
320ae51f JA |
1812 | /* |
1813 | * rq_size is the size of the request plus driver payload, rounded | |
1814 | * to the cacheline size | |
1815 | */ | |
24d2f903 | 1816 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1817 | cache_line_size()); |
cc71a6f4 | 1818 | left = rq_size * depth; |
320ae51f | 1819 | |
cc71a6f4 | 1820 | for (i = 0; i < depth; ) { |
320ae51f JA |
1821 | int this_order = max_order; |
1822 | struct page *page; | |
1823 | int to_do; | |
1824 | void *p; | |
1825 | ||
b3a834b1 | 1826 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1827 | this_order--; |
1828 | ||
1829 | do { | |
59f082e4 | 1830 | page = alloc_pages_node(node, |
36e1f3d1 | 1831 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1832 | this_order); |
320ae51f JA |
1833 | if (page) |
1834 | break; | |
1835 | if (!this_order--) | |
1836 | break; | |
1837 | if (order_to_size(this_order) < rq_size) | |
1838 | break; | |
1839 | } while (1); | |
1840 | ||
1841 | if (!page) | |
24d2f903 | 1842 | goto fail; |
320ae51f JA |
1843 | |
1844 | page->private = this_order; | |
24d2f903 | 1845 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1846 | |
1847 | p = page_address(page); | |
f75782e4 CM |
1848 | /* |
1849 | * Allow kmemleak to scan these pages as they contain pointers | |
1850 | * to additional allocations like via ops->init_request(). | |
1851 | */ | |
36e1f3d1 | 1852 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1853 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1854 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1855 | left -= to_do * rq_size; |
1856 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1857 | struct request *rq = p; |
1858 | ||
1859 | tags->static_rqs[i] = rq; | |
24d2f903 CH |
1860 | if (set->ops->init_request) { |
1861 | if (set->ops->init_request(set->driver_data, | |
2af8cbe3 | 1862 | rq, hctx_idx, i, |
59f082e4 | 1863 | node)) { |
2af8cbe3 | 1864 | tags->static_rqs[i] = NULL; |
24d2f903 | 1865 | goto fail; |
a5164405 | 1866 | } |
e9b267d9 CH |
1867 | } |
1868 | ||
320ae51f JA |
1869 | p += rq_size; |
1870 | i++; | |
1871 | } | |
1872 | } | |
cc71a6f4 | 1873 | return 0; |
320ae51f | 1874 | |
24d2f903 | 1875 | fail: |
cc71a6f4 JA |
1876 | blk_mq_free_rqs(set, tags, hctx_idx); |
1877 | return -ENOMEM; | |
320ae51f JA |
1878 | } |
1879 | ||
e57690fe JA |
1880 | /* |
1881 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1882 | * software queue to the hw queue dispatch list, and ensure that it | |
1883 | * gets run. | |
1884 | */ | |
9467f859 | 1885 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1886 | { |
9467f859 | 1887 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1888 | struct blk_mq_ctx *ctx; |
1889 | LIST_HEAD(tmp); | |
1890 | ||
9467f859 | 1891 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1892 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1893 | |
1894 | spin_lock(&ctx->lock); | |
1895 | if (!list_empty(&ctx->rq_list)) { | |
1896 | list_splice_init(&ctx->rq_list, &tmp); | |
1897 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1898 | } | |
1899 | spin_unlock(&ctx->lock); | |
1900 | ||
1901 | if (list_empty(&tmp)) | |
9467f859 | 1902 | return 0; |
484b4061 | 1903 | |
e57690fe JA |
1904 | spin_lock(&hctx->lock); |
1905 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1906 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1907 | |
1908 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1909 | return 0; |
484b4061 JA |
1910 | } |
1911 | ||
9467f859 | 1912 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1913 | { |
9467f859 TG |
1914 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1915 | &hctx->cpuhp_dead); | |
484b4061 JA |
1916 | } |
1917 | ||
c3b4afca | 1918 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1919 | static void blk_mq_exit_hctx(struct request_queue *q, |
1920 | struct blk_mq_tag_set *set, | |
1921 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1922 | { | |
f70ced09 ML |
1923 | unsigned flush_start_tag = set->queue_depth; |
1924 | ||
08e98fc6 ML |
1925 | blk_mq_tag_idle(hctx); |
1926 | ||
f70ced09 ML |
1927 | if (set->ops->exit_request) |
1928 | set->ops->exit_request(set->driver_data, | |
1929 | hctx->fq->flush_rq, hctx_idx, | |
1930 | flush_start_tag + hctx_idx); | |
1931 | ||
08e98fc6 ML |
1932 | if (set->ops->exit_hctx) |
1933 | set->ops->exit_hctx(hctx, hctx_idx); | |
1934 | ||
6a83e74d BVA |
1935 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
1936 | cleanup_srcu_struct(&hctx->queue_rq_srcu); | |
1937 | ||
9467f859 | 1938 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1939 | blk_free_flush_queue(hctx->fq); |
88459642 | 1940 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1941 | } |
1942 | ||
624dbe47 ML |
1943 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1944 | struct blk_mq_tag_set *set, int nr_queue) | |
1945 | { | |
1946 | struct blk_mq_hw_ctx *hctx; | |
1947 | unsigned int i; | |
1948 | ||
1949 | queue_for_each_hw_ctx(q, hctx, i) { | |
1950 | if (i == nr_queue) | |
1951 | break; | |
08e98fc6 | 1952 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1953 | } |
624dbe47 ML |
1954 | } |
1955 | ||
1956 | static void blk_mq_free_hw_queues(struct request_queue *q, | |
1957 | struct blk_mq_tag_set *set) | |
1958 | { | |
1959 | struct blk_mq_hw_ctx *hctx; | |
1960 | unsigned int i; | |
1961 | ||
e09aae7e | 1962 | queue_for_each_hw_ctx(q, hctx, i) |
624dbe47 | 1963 | free_cpumask_var(hctx->cpumask); |
624dbe47 ML |
1964 | } |
1965 | ||
08e98fc6 ML |
1966 | static int blk_mq_init_hctx(struct request_queue *q, |
1967 | struct blk_mq_tag_set *set, | |
1968 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1969 | { |
08e98fc6 | 1970 | int node; |
f70ced09 | 1971 | unsigned flush_start_tag = set->queue_depth; |
08e98fc6 ML |
1972 | |
1973 | node = hctx->numa_node; | |
1974 | if (node == NUMA_NO_NODE) | |
1975 | node = hctx->numa_node = set->numa_node; | |
1976 | ||
27489a3c | 1977 | INIT_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1978 | INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn); |
1979 | spin_lock_init(&hctx->lock); | |
1980 | INIT_LIST_HEAD(&hctx->dispatch); | |
1981 | hctx->queue = q; | |
1982 | hctx->queue_num = hctx_idx; | |
2404e607 | 1983 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1984 | |
9467f859 | 1985 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1986 | |
1987 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1988 | |
1989 | /* | |
08e98fc6 ML |
1990 | * Allocate space for all possible cpus to avoid allocation at |
1991 | * runtime | |
320ae51f | 1992 | */ |
08e98fc6 ML |
1993 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1994 | GFP_KERNEL, node); | |
1995 | if (!hctx->ctxs) | |
1996 | goto unregister_cpu_notifier; | |
320ae51f | 1997 | |
88459642 OS |
1998 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1999 | node)) | |
08e98fc6 | 2000 | goto free_ctxs; |
320ae51f | 2001 | |
08e98fc6 | 2002 | hctx->nr_ctx = 0; |
320ae51f | 2003 | |
08e98fc6 ML |
2004 | if (set->ops->init_hctx && |
2005 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2006 | goto free_bitmap; | |
320ae51f | 2007 | |
f70ced09 ML |
2008 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2009 | if (!hctx->fq) | |
2010 | goto exit_hctx; | |
320ae51f | 2011 | |
f70ced09 ML |
2012 | if (set->ops->init_request && |
2013 | set->ops->init_request(set->driver_data, | |
2014 | hctx->fq->flush_rq, hctx_idx, | |
2015 | flush_start_tag + hctx_idx, node)) | |
2016 | goto free_fq; | |
320ae51f | 2017 | |
6a83e74d BVA |
2018 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
2019 | init_srcu_struct(&hctx->queue_rq_srcu); | |
2020 | ||
08e98fc6 | 2021 | return 0; |
320ae51f | 2022 | |
f70ced09 ML |
2023 | free_fq: |
2024 | kfree(hctx->fq); | |
2025 | exit_hctx: | |
2026 | if (set->ops->exit_hctx) | |
2027 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2028 | free_bitmap: |
88459642 | 2029 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2030 | free_ctxs: |
2031 | kfree(hctx->ctxs); | |
2032 | unregister_cpu_notifier: | |
9467f859 | 2033 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2034 | return -1; |
2035 | } | |
320ae51f | 2036 | |
320ae51f JA |
2037 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2038 | unsigned int nr_hw_queues) | |
2039 | { | |
2040 | unsigned int i; | |
2041 | ||
2042 | for_each_possible_cpu(i) { | |
2043 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2044 | struct blk_mq_hw_ctx *hctx; | |
2045 | ||
2046 | memset(__ctx, 0, sizeof(*__ctx)); | |
2047 | __ctx->cpu = i; | |
2048 | spin_lock_init(&__ctx->lock); | |
2049 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2050 | __ctx->queue = q; | |
cf43e6be JA |
2051 | blk_stat_init(&__ctx->stat[BLK_STAT_READ]); |
2052 | blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]); | |
320ae51f JA |
2053 | |
2054 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
2055 | if (!cpu_online(i)) |
2056 | continue; | |
2057 | ||
7d7e0f90 | 2058 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2059 | |
320ae51f JA |
2060 | /* |
2061 | * Set local node, IFF we have more than one hw queue. If | |
2062 | * not, we remain on the home node of the device | |
2063 | */ | |
2064 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2065 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2066 | } |
2067 | } | |
2068 | ||
cc71a6f4 JA |
2069 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2070 | { | |
2071 | int ret = 0; | |
2072 | ||
2073 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2074 | set->queue_depth, set->reserved_tags); | |
2075 | if (!set->tags[hctx_idx]) | |
2076 | return false; | |
2077 | ||
2078 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2079 | set->queue_depth); | |
2080 | if (!ret) | |
2081 | return true; | |
2082 | ||
2083 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2084 | set->tags[hctx_idx] = NULL; | |
2085 | return false; | |
2086 | } | |
2087 | ||
2088 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2089 | unsigned int hctx_idx) | |
2090 | { | |
bd166ef1 JA |
2091 | if (set->tags[hctx_idx]) { |
2092 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2093 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2094 | set->tags[hctx_idx] = NULL; | |
2095 | } | |
cc71a6f4 JA |
2096 | } |
2097 | ||
5778322e AM |
2098 | static void blk_mq_map_swqueue(struct request_queue *q, |
2099 | const struct cpumask *online_mask) | |
320ae51f | 2100 | { |
d1b1cea1 | 2101 | unsigned int i, hctx_idx; |
320ae51f JA |
2102 | struct blk_mq_hw_ctx *hctx; |
2103 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2104 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2105 | |
60de074b AM |
2106 | /* |
2107 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2108 | */ | |
2109 | mutex_lock(&q->sysfs_lock); | |
2110 | ||
320ae51f | 2111 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2112 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2113 | hctx->nr_ctx = 0; |
2114 | } | |
2115 | ||
2116 | /* | |
2117 | * Map software to hardware queues | |
2118 | */ | |
897bb0c7 | 2119 | for_each_possible_cpu(i) { |
320ae51f | 2120 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2121 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2122 | continue; |
2123 | ||
d1b1cea1 GKB |
2124 | hctx_idx = q->mq_map[i]; |
2125 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2126 | if (!set->tags[hctx_idx] && |
2127 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2128 | /* |
2129 | * If tags initialization fail for some hctx, | |
2130 | * that hctx won't be brought online. In this | |
2131 | * case, remap the current ctx to hctx[0] which | |
2132 | * is guaranteed to always have tags allocated | |
2133 | */ | |
cc71a6f4 | 2134 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2135 | } |
2136 | ||
897bb0c7 | 2137 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2138 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2139 | |
e4043dcf | 2140 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2141 | ctx->index_hw = hctx->nr_ctx; |
2142 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2143 | } | |
506e931f | 2144 | |
60de074b AM |
2145 | mutex_unlock(&q->sysfs_lock); |
2146 | ||
506e931f | 2147 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2148 | /* |
a68aafa5 JA |
2149 | * If no software queues are mapped to this hardware queue, |
2150 | * disable it and free the request entries. | |
484b4061 JA |
2151 | */ |
2152 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2153 | /* Never unmap queue 0. We need it as a |
2154 | * fallback in case of a new remap fails | |
2155 | * allocation | |
2156 | */ | |
cc71a6f4 JA |
2157 | if (i && set->tags[i]) |
2158 | blk_mq_free_map_and_requests(set, i); | |
2159 | ||
2a34c087 | 2160 | hctx->tags = NULL; |
484b4061 JA |
2161 | continue; |
2162 | } | |
2163 | ||
2a34c087 ML |
2164 | hctx->tags = set->tags[i]; |
2165 | WARN_ON(!hctx->tags); | |
2166 | ||
889fa31f CY |
2167 | /* |
2168 | * Set the map size to the number of mapped software queues. | |
2169 | * This is more accurate and more efficient than looping | |
2170 | * over all possibly mapped software queues. | |
2171 | */ | |
88459642 | 2172 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2173 | |
484b4061 JA |
2174 | /* |
2175 | * Initialize batch roundrobin counts | |
2176 | */ | |
506e931f JA |
2177 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2178 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2179 | } | |
320ae51f JA |
2180 | } |
2181 | ||
2404e607 | 2182 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2183 | { |
2184 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2185 | int i; |
2186 | ||
2404e607 JM |
2187 | queue_for_each_hw_ctx(q, hctx, i) { |
2188 | if (shared) | |
2189 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2190 | else | |
2191 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2192 | } | |
2193 | } | |
2194 | ||
2195 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2196 | { | |
2197 | struct request_queue *q; | |
0d2602ca JA |
2198 | |
2199 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2200 | blk_mq_freeze_queue(q); | |
2404e607 | 2201 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2202 | blk_mq_unfreeze_queue(q); |
2203 | } | |
2204 | } | |
2205 | ||
2206 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2207 | { | |
2208 | struct blk_mq_tag_set *set = q->tag_set; | |
2209 | ||
0d2602ca JA |
2210 | mutex_lock(&set->tag_list_lock); |
2211 | list_del_init(&q->tag_set_list); | |
2404e607 JM |
2212 | if (list_is_singular(&set->tag_list)) { |
2213 | /* just transitioned to unshared */ | |
2214 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2215 | /* update existing queue */ | |
2216 | blk_mq_update_tag_set_depth(set, false); | |
2217 | } | |
0d2602ca | 2218 | mutex_unlock(&set->tag_list_lock); |
0d2602ca JA |
2219 | } |
2220 | ||
2221 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2222 | struct request_queue *q) | |
2223 | { | |
2224 | q->tag_set = set; | |
2225 | ||
2226 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2227 | |
2228 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2229 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2230 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2231 | /* update existing queue */ | |
2232 | blk_mq_update_tag_set_depth(set, true); | |
2233 | } | |
2234 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2235 | queue_set_hctx_shared(q, true); | |
0d2602ca | 2236 | list_add_tail(&q->tag_set_list, &set->tag_list); |
2404e607 | 2237 | |
0d2602ca JA |
2238 | mutex_unlock(&set->tag_list_lock); |
2239 | } | |
2240 | ||
e09aae7e ML |
2241 | /* |
2242 | * It is the actual release handler for mq, but we do it from | |
2243 | * request queue's release handler for avoiding use-after-free | |
2244 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2245 | * but we can't group ctx/kctx kobj without it. | |
2246 | */ | |
2247 | void blk_mq_release(struct request_queue *q) | |
2248 | { | |
2249 | struct blk_mq_hw_ctx *hctx; | |
2250 | unsigned int i; | |
2251 | ||
bd166ef1 JA |
2252 | blk_mq_sched_teardown(q); |
2253 | ||
e09aae7e | 2254 | /* hctx kobj stays in hctx */ |
c3b4afca ML |
2255 | queue_for_each_hw_ctx(q, hctx, i) { |
2256 | if (!hctx) | |
2257 | continue; | |
2258 | kfree(hctx->ctxs); | |
e09aae7e | 2259 | kfree(hctx); |
c3b4afca | 2260 | } |
e09aae7e | 2261 | |
a723bab3 AM |
2262 | q->mq_map = NULL; |
2263 | ||
e09aae7e ML |
2264 | kfree(q->queue_hw_ctx); |
2265 | ||
2266 | /* ctx kobj stays in queue_ctx */ | |
2267 | free_percpu(q->queue_ctx); | |
2268 | } | |
2269 | ||
24d2f903 | 2270 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2271 | { |
2272 | struct request_queue *uninit_q, *q; | |
2273 | ||
2274 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2275 | if (!uninit_q) | |
2276 | return ERR_PTR(-ENOMEM); | |
2277 | ||
2278 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2279 | if (IS_ERR(q)) | |
2280 | blk_cleanup_queue(uninit_q); | |
2281 | ||
2282 | return q; | |
2283 | } | |
2284 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2285 | ||
868f2f0b KB |
2286 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2287 | struct request_queue *q) | |
320ae51f | 2288 | { |
868f2f0b KB |
2289 | int i, j; |
2290 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2291 | |
868f2f0b | 2292 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2293 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2294 | int node; |
f14bbe77 | 2295 | |
868f2f0b KB |
2296 | if (hctxs[i]) |
2297 | continue; | |
2298 | ||
2299 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
cdef54dd CH |
2300 | hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx), |
2301 | GFP_KERNEL, node); | |
320ae51f | 2302 | if (!hctxs[i]) |
868f2f0b | 2303 | break; |
320ae51f | 2304 | |
a86073e4 | 2305 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2306 | node)) { |
2307 | kfree(hctxs[i]); | |
2308 | hctxs[i] = NULL; | |
2309 | break; | |
2310 | } | |
e4043dcf | 2311 | |
0d2602ca | 2312 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2313 | hctxs[i]->numa_node = node; |
320ae51f | 2314 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2315 | |
2316 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2317 | free_cpumask_var(hctxs[i]->cpumask); | |
2318 | kfree(hctxs[i]); | |
2319 | hctxs[i] = NULL; | |
2320 | break; | |
2321 | } | |
2322 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2323 | } |
868f2f0b KB |
2324 | for (j = i; j < q->nr_hw_queues; j++) { |
2325 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2326 | ||
2327 | if (hctx) { | |
cc71a6f4 JA |
2328 | if (hctx->tags) |
2329 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b KB |
2330 | blk_mq_exit_hctx(q, set, hctx, j); |
2331 | free_cpumask_var(hctx->cpumask); | |
2332 | kobject_put(&hctx->kobj); | |
2333 | kfree(hctx->ctxs); | |
2334 | kfree(hctx); | |
2335 | hctxs[j] = NULL; | |
2336 | ||
2337 | } | |
2338 | } | |
2339 | q->nr_hw_queues = i; | |
2340 | blk_mq_sysfs_register(q); | |
2341 | } | |
2342 | ||
2343 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2344 | struct request_queue *q) | |
2345 | { | |
66841672 ML |
2346 | /* mark the queue as mq asap */ |
2347 | q->mq_ops = set->ops; | |
2348 | ||
868f2f0b KB |
2349 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2350 | if (!q->queue_ctx) | |
c7de5726 | 2351 | goto err_exit; |
868f2f0b KB |
2352 | |
2353 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), | |
2354 | GFP_KERNEL, set->numa_node); | |
2355 | if (!q->queue_hw_ctx) | |
2356 | goto err_percpu; | |
2357 | ||
bdd17e75 | 2358 | q->mq_map = set->mq_map; |
868f2f0b KB |
2359 | |
2360 | blk_mq_realloc_hw_ctxs(set, q); | |
2361 | if (!q->nr_hw_queues) | |
2362 | goto err_hctxs; | |
320ae51f | 2363 | |
287922eb | 2364 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2365 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2366 | |
2367 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2368 | |
94eddfbe | 2369 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2370 | |
05f1dd53 JA |
2371 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2372 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2373 | ||
1be036e9 CH |
2374 | q->sg_reserved_size = INT_MAX; |
2375 | ||
2849450a | 2376 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2377 | INIT_LIST_HEAD(&q->requeue_list); |
2378 | spin_lock_init(&q->requeue_lock); | |
2379 | ||
07068d5b JA |
2380 | if (q->nr_hw_queues > 1) |
2381 | blk_queue_make_request(q, blk_mq_make_request); | |
2382 | else | |
2383 | blk_queue_make_request(q, blk_sq_make_request); | |
2384 | ||
eba71768 JA |
2385 | /* |
2386 | * Do this after blk_queue_make_request() overrides it... | |
2387 | */ | |
2388 | q->nr_requests = set->queue_depth; | |
2389 | ||
64f1c21e JA |
2390 | /* |
2391 | * Default to classic polling | |
2392 | */ | |
2393 | q->poll_nsec = -1; | |
2394 | ||
24d2f903 CH |
2395 | if (set->ops->complete) |
2396 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2397 | |
24d2f903 | 2398 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2399 | |
5778322e | 2400 | get_online_cpus(); |
320ae51f | 2401 | mutex_lock(&all_q_mutex); |
320ae51f | 2402 | |
4593fdbe | 2403 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2404 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2405 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2406 | |
4593fdbe | 2407 | mutex_unlock(&all_q_mutex); |
5778322e | 2408 | put_online_cpus(); |
4593fdbe | 2409 | |
d3484991 JA |
2410 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2411 | int ret; | |
2412 | ||
2413 | ret = blk_mq_sched_init(q); | |
2414 | if (ret) | |
2415 | return ERR_PTR(ret); | |
2416 | } | |
2417 | ||
320ae51f | 2418 | return q; |
18741986 | 2419 | |
320ae51f | 2420 | err_hctxs: |
868f2f0b | 2421 | kfree(q->queue_hw_ctx); |
320ae51f | 2422 | err_percpu: |
868f2f0b | 2423 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2424 | err_exit: |
2425 | q->mq_ops = NULL; | |
320ae51f JA |
2426 | return ERR_PTR(-ENOMEM); |
2427 | } | |
b62c21b7 | 2428 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2429 | |
2430 | void blk_mq_free_queue(struct request_queue *q) | |
2431 | { | |
624dbe47 | 2432 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2433 | |
0e626368 AM |
2434 | mutex_lock(&all_q_mutex); |
2435 | list_del_init(&q->all_q_node); | |
2436 | mutex_unlock(&all_q_mutex); | |
2437 | ||
87760e5e JA |
2438 | wbt_exit(q); |
2439 | ||
0d2602ca JA |
2440 | blk_mq_del_queue_tag_set(q); |
2441 | ||
624dbe47 ML |
2442 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
2443 | blk_mq_free_hw_queues(q, set); | |
320ae51f | 2444 | } |
320ae51f JA |
2445 | |
2446 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2447 | static void blk_mq_queue_reinit(struct request_queue *q, |
2448 | const struct cpumask *online_mask) | |
320ae51f | 2449 | { |
4ecd4fef | 2450 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2451 | |
67aec14c JA |
2452 | blk_mq_sysfs_unregister(q); |
2453 | ||
320ae51f JA |
2454 | /* |
2455 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2456 | * we should change hctx numa_node according to new topology (this | |
2457 | * involves free and re-allocate memory, worthy doing?) | |
2458 | */ | |
2459 | ||
5778322e | 2460 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2461 | |
67aec14c | 2462 | blk_mq_sysfs_register(q); |
320ae51f JA |
2463 | } |
2464 | ||
65d5291e SAS |
2465 | /* |
2466 | * New online cpumask which is going to be set in this hotplug event. | |
2467 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2468 | * one-by-one and dynamically allocating this could result in a failure. | |
2469 | */ | |
2470 | static struct cpumask cpuhp_online_new; | |
2471 | ||
2472 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2473 | { |
2474 | struct request_queue *q; | |
320ae51f JA |
2475 | |
2476 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2477 | /* |
2478 | * We need to freeze and reinit all existing queues. Freezing | |
2479 | * involves synchronous wait for an RCU grace period and doing it | |
2480 | * one by one may take a long time. Start freezing all queues in | |
2481 | * one swoop and then wait for the completions so that freezing can | |
2482 | * take place in parallel. | |
2483 | */ | |
2484 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2485 | blk_mq_freeze_queue_start(q); | |
415d3dab | 2486 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2487 | blk_mq_freeze_queue_wait(q); |
2488 | ||
320ae51f | 2489 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2490 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2491 | |
2492 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2493 | blk_mq_unfreeze_queue(q); | |
2494 | ||
320ae51f | 2495 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2496 | } |
2497 | ||
2498 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2499 | { | |
97a32864 | 2500 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2501 | blk_mq_queue_reinit_work(); |
2502 | return 0; | |
2503 | } | |
2504 | ||
2505 | /* | |
2506 | * Before hotadded cpu starts handling requests, new mappings must be | |
2507 | * established. Otherwise, these requests in hw queue might never be | |
2508 | * dispatched. | |
2509 | * | |
2510 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2511 | * for CPU0, and ctx1 for CPU1). | |
2512 | * | |
2513 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2514 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2515 | * | |
2c3ad667 JA |
2516 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2517 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2518 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2519 | * ignored. | |
65d5291e SAS |
2520 | */ |
2521 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2522 | { | |
2523 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2524 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2525 | blk_mq_queue_reinit_work(); | |
2526 | return 0; | |
320ae51f JA |
2527 | } |
2528 | ||
a5164405 JA |
2529 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2530 | { | |
2531 | int i; | |
2532 | ||
cc71a6f4 JA |
2533 | for (i = 0; i < set->nr_hw_queues; i++) |
2534 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2535 | goto out_unwind; |
a5164405 JA |
2536 | |
2537 | return 0; | |
2538 | ||
2539 | out_unwind: | |
2540 | while (--i >= 0) | |
cc71a6f4 | 2541 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2542 | |
a5164405 JA |
2543 | return -ENOMEM; |
2544 | } | |
2545 | ||
2546 | /* | |
2547 | * Allocate the request maps associated with this tag_set. Note that this | |
2548 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2549 | * will be updated to reflect the allocated depth. | |
2550 | */ | |
2551 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2552 | { | |
2553 | unsigned int depth; | |
2554 | int err; | |
2555 | ||
2556 | depth = set->queue_depth; | |
2557 | do { | |
2558 | err = __blk_mq_alloc_rq_maps(set); | |
2559 | if (!err) | |
2560 | break; | |
2561 | ||
2562 | set->queue_depth >>= 1; | |
2563 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2564 | err = -ENOMEM; | |
2565 | break; | |
2566 | } | |
2567 | } while (set->queue_depth); | |
2568 | ||
2569 | if (!set->queue_depth || err) { | |
2570 | pr_err("blk-mq: failed to allocate request map\n"); | |
2571 | return -ENOMEM; | |
2572 | } | |
2573 | ||
2574 | if (depth != set->queue_depth) | |
2575 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2576 | depth, set->queue_depth); | |
2577 | ||
2578 | return 0; | |
2579 | } | |
2580 | ||
a4391c64 JA |
2581 | /* |
2582 | * Alloc a tag set to be associated with one or more request queues. | |
2583 | * May fail with EINVAL for various error conditions. May adjust the | |
2584 | * requested depth down, if if it too large. In that case, the set | |
2585 | * value will be stored in set->queue_depth. | |
2586 | */ | |
24d2f903 CH |
2587 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2588 | { | |
da695ba2 CH |
2589 | int ret; |
2590 | ||
205fb5f5 BVA |
2591 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2592 | ||
24d2f903 CH |
2593 | if (!set->nr_hw_queues) |
2594 | return -EINVAL; | |
a4391c64 | 2595 | if (!set->queue_depth) |
24d2f903 CH |
2596 | return -EINVAL; |
2597 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2598 | return -EINVAL; | |
2599 | ||
7d7e0f90 | 2600 | if (!set->ops->queue_rq) |
24d2f903 CH |
2601 | return -EINVAL; |
2602 | ||
a4391c64 JA |
2603 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2604 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2605 | BLK_MQ_MAX_DEPTH); | |
2606 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2607 | } | |
24d2f903 | 2608 | |
6637fadf SL |
2609 | /* |
2610 | * If a crashdump is active, then we are potentially in a very | |
2611 | * memory constrained environment. Limit us to 1 queue and | |
2612 | * 64 tags to prevent using too much memory. | |
2613 | */ | |
2614 | if (is_kdump_kernel()) { | |
2615 | set->nr_hw_queues = 1; | |
2616 | set->queue_depth = min(64U, set->queue_depth); | |
2617 | } | |
868f2f0b KB |
2618 | /* |
2619 | * There is no use for more h/w queues than cpus. | |
2620 | */ | |
2621 | if (set->nr_hw_queues > nr_cpu_ids) | |
2622 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2623 | |
868f2f0b | 2624 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2625 | GFP_KERNEL, set->numa_node); |
2626 | if (!set->tags) | |
a5164405 | 2627 | return -ENOMEM; |
24d2f903 | 2628 | |
da695ba2 CH |
2629 | ret = -ENOMEM; |
2630 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2631 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2632 | if (!set->mq_map) |
2633 | goto out_free_tags; | |
2634 | ||
da695ba2 CH |
2635 | if (set->ops->map_queues) |
2636 | ret = set->ops->map_queues(set); | |
2637 | else | |
2638 | ret = blk_mq_map_queues(set); | |
2639 | if (ret) | |
2640 | goto out_free_mq_map; | |
2641 | ||
2642 | ret = blk_mq_alloc_rq_maps(set); | |
2643 | if (ret) | |
bdd17e75 | 2644 | goto out_free_mq_map; |
24d2f903 | 2645 | |
0d2602ca JA |
2646 | mutex_init(&set->tag_list_lock); |
2647 | INIT_LIST_HEAD(&set->tag_list); | |
2648 | ||
24d2f903 | 2649 | return 0; |
bdd17e75 CH |
2650 | |
2651 | out_free_mq_map: | |
2652 | kfree(set->mq_map); | |
2653 | set->mq_map = NULL; | |
2654 | out_free_tags: | |
5676e7b6 RE |
2655 | kfree(set->tags); |
2656 | set->tags = NULL; | |
da695ba2 | 2657 | return ret; |
24d2f903 CH |
2658 | } |
2659 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2660 | ||
2661 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2662 | { | |
2663 | int i; | |
2664 | ||
cc71a6f4 JA |
2665 | for (i = 0; i < nr_cpu_ids; i++) |
2666 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2667 | |
bdd17e75 CH |
2668 | kfree(set->mq_map); |
2669 | set->mq_map = NULL; | |
2670 | ||
981bd189 | 2671 | kfree(set->tags); |
5676e7b6 | 2672 | set->tags = NULL; |
24d2f903 CH |
2673 | } |
2674 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2675 | ||
e3a2b3f9 JA |
2676 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2677 | { | |
2678 | struct blk_mq_tag_set *set = q->tag_set; | |
2679 | struct blk_mq_hw_ctx *hctx; | |
2680 | int i, ret; | |
2681 | ||
bd166ef1 | 2682 | if (!set) |
e3a2b3f9 JA |
2683 | return -EINVAL; |
2684 | ||
70f36b60 JA |
2685 | blk_mq_freeze_queue(q); |
2686 | blk_mq_quiesce_queue(q); | |
2687 | ||
e3a2b3f9 JA |
2688 | ret = 0; |
2689 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2690 | if (!hctx->tags) |
2691 | continue; | |
bd166ef1 JA |
2692 | /* |
2693 | * If we're using an MQ scheduler, just update the scheduler | |
2694 | * queue depth. This is similar to what the old code would do. | |
2695 | */ | |
70f36b60 JA |
2696 | if (!hctx->sched_tags) { |
2697 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2698 | min(nr, set->queue_depth), | |
2699 | false); | |
2700 | } else { | |
2701 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2702 | nr, true); | |
2703 | } | |
e3a2b3f9 JA |
2704 | if (ret) |
2705 | break; | |
2706 | } | |
2707 | ||
2708 | if (!ret) | |
2709 | q->nr_requests = nr; | |
2710 | ||
70f36b60 JA |
2711 | blk_mq_unfreeze_queue(q); |
2712 | blk_mq_start_stopped_hw_queues(q, true); | |
2713 | ||
e3a2b3f9 JA |
2714 | return ret; |
2715 | } | |
2716 | ||
868f2f0b KB |
2717 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) |
2718 | { | |
2719 | struct request_queue *q; | |
2720 | ||
2721 | if (nr_hw_queues > nr_cpu_ids) | |
2722 | nr_hw_queues = nr_cpu_ids; | |
2723 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2724 | return; | |
2725 | ||
2726 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2727 | blk_mq_freeze_queue(q); | |
2728 | ||
2729 | set->nr_hw_queues = nr_hw_queues; | |
2730 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2731 | blk_mq_realloc_hw_ctxs(set, q); | |
2732 | ||
f6f94300 JB |
2733 | /* |
2734 | * Manually set the make_request_fn as blk_queue_make_request | |
2735 | * resets a lot of the queue settings. | |
2736 | */ | |
868f2f0b | 2737 | if (q->nr_hw_queues > 1) |
f6f94300 | 2738 | q->make_request_fn = blk_mq_make_request; |
868f2f0b | 2739 | else |
f6f94300 | 2740 | q->make_request_fn = blk_sq_make_request; |
868f2f0b KB |
2741 | |
2742 | blk_mq_queue_reinit(q, cpu_online_mask); | |
2743 | } | |
2744 | ||
2745 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2746 | blk_mq_unfreeze_queue(q); | |
2747 | } | |
2748 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
2749 | ||
64f1c21e JA |
2750 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2751 | struct blk_mq_hw_ctx *hctx, | |
2752 | struct request *rq) | |
2753 | { | |
2754 | struct blk_rq_stat stat[2]; | |
2755 | unsigned long ret = 0; | |
2756 | ||
2757 | /* | |
2758 | * If stats collection isn't on, don't sleep but turn it on for | |
2759 | * future users | |
2760 | */ | |
2761 | if (!blk_stat_enable(q)) | |
2762 | return 0; | |
2763 | ||
2764 | /* | |
2765 | * We don't have to do this once per IO, should optimize this | |
2766 | * to just use the current window of stats until it changes | |
2767 | */ | |
2768 | memset(&stat, 0, sizeof(stat)); | |
2769 | blk_hctx_stat_get(hctx, stat); | |
2770 | ||
2771 | /* | |
2772 | * As an optimistic guess, use half of the mean service time | |
2773 | * for this type of request. We can (and should) make this smarter. | |
2774 | * For instance, if the completion latencies are tight, we can | |
2775 | * get closer than just half the mean. This is especially | |
2776 | * important on devices where the completion latencies are longer | |
2777 | * than ~10 usec. | |
2778 | */ | |
2779 | if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples) | |
2780 | ret = (stat[BLK_STAT_READ].mean + 1) / 2; | |
2781 | else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples) | |
2782 | ret = (stat[BLK_STAT_WRITE].mean + 1) / 2; | |
2783 | ||
2784 | return ret; | |
2785 | } | |
2786 | ||
06426adf | 2787 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2788 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2789 | struct request *rq) |
2790 | { | |
2791 | struct hrtimer_sleeper hs; | |
2792 | enum hrtimer_mode mode; | |
64f1c21e | 2793 | unsigned int nsecs; |
06426adf JA |
2794 | ktime_t kt; |
2795 | ||
64f1c21e JA |
2796 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2797 | return false; | |
2798 | ||
2799 | /* | |
2800 | * poll_nsec can be: | |
2801 | * | |
2802 | * -1: don't ever hybrid sleep | |
2803 | * 0: use half of prev avg | |
2804 | * >0: use this specific value | |
2805 | */ | |
2806 | if (q->poll_nsec == -1) | |
2807 | return false; | |
2808 | else if (q->poll_nsec > 0) | |
2809 | nsecs = q->poll_nsec; | |
2810 | else | |
2811 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2812 | ||
2813 | if (!nsecs) | |
06426adf JA |
2814 | return false; |
2815 | ||
2816 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2817 | ||
2818 | /* | |
2819 | * This will be replaced with the stats tracking code, using | |
2820 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2821 | */ | |
8b0e1953 | 2822 | kt = nsecs; |
06426adf JA |
2823 | |
2824 | mode = HRTIMER_MODE_REL; | |
2825 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2826 | hrtimer_set_expires(&hs.timer, kt); | |
2827 | ||
2828 | hrtimer_init_sleeper(&hs, current); | |
2829 | do { | |
2830 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2831 | break; | |
2832 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2833 | hrtimer_start_expires(&hs.timer, mode); | |
2834 | if (hs.task) | |
2835 | io_schedule(); | |
2836 | hrtimer_cancel(&hs.timer); | |
2837 | mode = HRTIMER_MODE_ABS; | |
2838 | } while (hs.task && !signal_pending(current)); | |
2839 | ||
2840 | __set_current_state(TASK_RUNNING); | |
2841 | destroy_hrtimer_on_stack(&hs.timer); | |
2842 | return true; | |
2843 | } | |
2844 | ||
bbd7bb70 JA |
2845 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2846 | { | |
2847 | struct request_queue *q = hctx->queue; | |
2848 | long state; | |
2849 | ||
06426adf JA |
2850 | /* |
2851 | * If we sleep, have the caller restart the poll loop to reset | |
2852 | * the state. Like for the other success return cases, the | |
2853 | * caller is responsible for checking if the IO completed. If | |
2854 | * the IO isn't complete, we'll get called again and will go | |
2855 | * straight to the busy poll loop. | |
2856 | */ | |
64f1c21e | 2857 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2858 | return true; |
2859 | ||
bbd7bb70 JA |
2860 | hctx->poll_considered++; |
2861 | ||
2862 | state = current->state; | |
2863 | while (!need_resched()) { | |
2864 | int ret; | |
2865 | ||
2866 | hctx->poll_invoked++; | |
2867 | ||
2868 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2869 | if (ret > 0) { | |
2870 | hctx->poll_success++; | |
2871 | set_current_state(TASK_RUNNING); | |
2872 | return true; | |
2873 | } | |
2874 | ||
2875 | if (signal_pending_state(state, current)) | |
2876 | set_current_state(TASK_RUNNING); | |
2877 | ||
2878 | if (current->state == TASK_RUNNING) | |
2879 | return true; | |
2880 | if (ret < 0) | |
2881 | break; | |
2882 | cpu_relax(); | |
2883 | } | |
2884 | ||
2885 | return false; | |
2886 | } | |
2887 | ||
2888 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2889 | { | |
2890 | struct blk_mq_hw_ctx *hctx; | |
2891 | struct blk_plug *plug; | |
2892 | struct request *rq; | |
2893 | ||
2894 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2895 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2896 | return false; | |
2897 | ||
2898 | plug = current->plug; | |
2899 | if (plug) | |
2900 | blk_flush_plug_list(plug, false); | |
2901 | ||
2902 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2903 | if (!blk_qc_t_is_internal(cookie)) |
2904 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
2905 | else | |
2906 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
bbd7bb70 JA |
2907 | |
2908 | return __blk_mq_poll(hctx, rq); | |
2909 | } | |
2910 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2911 | ||
676141e4 JA |
2912 | void blk_mq_disable_hotplug(void) |
2913 | { | |
2914 | mutex_lock(&all_q_mutex); | |
2915 | } | |
2916 | ||
2917 | void blk_mq_enable_hotplug(void) | |
2918 | { | |
2919 | mutex_unlock(&all_q_mutex); | |
2920 | } | |
2921 | ||
320ae51f JA |
2922 | static int __init blk_mq_init(void) |
2923 | { | |
9467f859 TG |
2924 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2925 | blk_mq_hctx_notify_dead); | |
320ae51f | 2926 | |
65d5291e SAS |
2927 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2928 | blk_mq_queue_reinit_prepare, | |
2929 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2930 | return 0; |
2931 | } | |
2932 | subsys_initcall(blk_mq_init); |