]>
Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f | 39 | |
34dbad5d OS |
40 | static void blk_mq_poll_stats_start(struct request_queue *q); |
41 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
42 | ||
720b8ccc SB |
43 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
44 | { | |
45 | int ddir, bytes, bucket; | |
46 | ||
99c749a4 | 47 | ddir = rq_data_dir(rq); |
720b8ccc SB |
48 | bytes = blk_rq_bytes(rq); |
49 | ||
50 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
51 | ||
52 | if (bucket < 0) | |
53 | return -1; | |
54 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
55 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
56 | ||
57 | return bucket; | |
58 | } | |
59 | ||
320ae51f JA |
60 | /* |
61 | * Check if any of the ctx's have pending work in this hardware queue | |
62 | */ | |
50e1dab8 | 63 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 64 | { |
bd166ef1 JA |
65 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
66 | !list_empty_careful(&hctx->dispatch) || | |
67 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
68 | } |
69 | ||
320ae51f JA |
70 | /* |
71 | * Mark this ctx as having pending work in this hardware queue | |
72 | */ | |
73 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
74 | struct blk_mq_ctx *ctx) | |
75 | { | |
88459642 OS |
76 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
77 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
78 | } |
79 | ||
80 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
88459642 | 83 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
84 | } |
85 | ||
f299b7c7 JA |
86 | struct mq_inflight { |
87 | struct hd_struct *part; | |
88 | unsigned int *inflight; | |
89 | }; | |
90 | ||
91 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
92 | struct request *rq, void *priv, | |
93 | bool reserved) | |
94 | { | |
95 | struct mq_inflight *mi = priv; | |
96 | ||
97 | if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) && | |
98 | !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
99 | /* | |
100 | * Count as inflight if it either matches the partition we | |
101 | * asked for, or if it's the root | |
102 | */ | |
103 | if (rq->part == mi->part || mi->part->partno) | |
104 | mi->inflight[0]++; | |
105 | } | |
106 | } | |
107 | ||
108 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
109 | unsigned int inflight[2]) | |
110 | { | |
111 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
112 | ||
113 | inflight[0] = 0; | |
114 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); | |
115 | } | |
116 | ||
1671d522 | 117 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 118 | { |
4ecd4fef | 119 | int freeze_depth; |
cddd5d17 | 120 | |
4ecd4fef CH |
121 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
122 | if (freeze_depth == 1) { | |
3ef28e83 | 123 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 124 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 125 | } |
f3af020b | 126 | } |
1671d522 | 127 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 128 | |
6bae363e | 129 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 130 | { |
3ef28e83 | 131 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 132 | } |
6bae363e | 133 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 134 | |
f91328c4 KB |
135 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
136 | unsigned long timeout) | |
137 | { | |
138 | return wait_event_timeout(q->mq_freeze_wq, | |
139 | percpu_ref_is_zero(&q->q_usage_counter), | |
140 | timeout); | |
141 | } | |
142 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 143 | |
f3af020b TH |
144 | /* |
145 | * Guarantee no request is in use, so we can change any data structure of | |
146 | * the queue afterward. | |
147 | */ | |
3ef28e83 | 148 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 149 | { |
3ef28e83 DW |
150 | /* |
151 | * In the !blk_mq case we are only calling this to kill the | |
152 | * q_usage_counter, otherwise this increases the freeze depth | |
153 | * and waits for it to return to zero. For this reason there is | |
154 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
155 | * exported to drivers as the only user for unfreeze is blk_mq. | |
156 | */ | |
1671d522 | 157 | blk_freeze_queue_start(q); |
f3af020b TH |
158 | blk_mq_freeze_queue_wait(q); |
159 | } | |
3ef28e83 DW |
160 | |
161 | void blk_mq_freeze_queue(struct request_queue *q) | |
162 | { | |
163 | /* | |
164 | * ...just an alias to keep freeze and unfreeze actions balanced | |
165 | * in the blk_mq_* namespace | |
166 | */ | |
167 | blk_freeze_queue(q); | |
168 | } | |
c761d96b | 169 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 170 | |
b4c6a028 | 171 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 172 | { |
4ecd4fef | 173 | int freeze_depth; |
320ae51f | 174 | |
4ecd4fef CH |
175 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
176 | WARN_ON_ONCE(freeze_depth < 0); | |
177 | if (!freeze_depth) { | |
3ef28e83 | 178 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 179 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 180 | } |
320ae51f | 181 | } |
b4c6a028 | 182 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 183 | |
852ec809 BVA |
184 | /* |
185 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
186 | * mpt3sas driver such that this function can be removed. | |
187 | */ | |
188 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
189 | { | |
190 | unsigned long flags; | |
191 | ||
192 | spin_lock_irqsave(q->queue_lock, flags); | |
193 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
194 | spin_unlock_irqrestore(q->queue_lock, flags); | |
195 | } | |
196 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
197 | ||
6a83e74d | 198 | /** |
69e07c4a | 199 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
200 | * @q: request queue. |
201 | * | |
202 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
203 | * callback function is invoked. Once this function is returned, we make |
204 | * sure no dispatch can happen until the queue is unquiesced via | |
205 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
206 | */ |
207 | void blk_mq_quiesce_queue(struct request_queue *q) | |
208 | { | |
209 | struct blk_mq_hw_ctx *hctx; | |
210 | unsigned int i; | |
211 | bool rcu = false; | |
212 | ||
1d9e9bc6 | 213 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 214 | |
6a83e74d BVA |
215 | queue_for_each_hw_ctx(q, hctx, i) { |
216 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 217 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
218 | else |
219 | rcu = true; | |
220 | } | |
221 | if (rcu) | |
222 | synchronize_rcu(); | |
223 | } | |
224 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
225 | ||
e4e73913 ML |
226 | /* |
227 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
228 | * @q: request queue. | |
229 | * | |
230 | * This function recovers queue into the state before quiescing | |
231 | * which is done by blk_mq_quiesce_queue. | |
232 | */ | |
233 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
234 | { | |
852ec809 BVA |
235 | unsigned long flags; |
236 | ||
237 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 238 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 239 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 240 | |
1d9e9bc6 ML |
241 | /* dispatch requests which are inserted during quiescing */ |
242 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
243 | } |
244 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
245 | ||
aed3ea94 JA |
246 | void blk_mq_wake_waiters(struct request_queue *q) |
247 | { | |
248 | struct blk_mq_hw_ctx *hctx; | |
249 | unsigned int i; | |
250 | ||
251 | queue_for_each_hw_ctx(q, hctx, i) | |
252 | if (blk_mq_hw_queue_mapped(hctx)) | |
253 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
254 | |
255 | /* | |
256 | * If we are called because the queue has now been marked as | |
257 | * dying, we need to ensure that processes currently waiting on | |
258 | * the queue are notified as well. | |
259 | */ | |
260 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
261 | } |
262 | ||
320ae51f JA |
263 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
264 | { | |
265 | return blk_mq_has_free_tags(hctx->tags); | |
266 | } | |
267 | EXPORT_SYMBOL(blk_mq_can_queue); | |
268 | ||
e4cdf1a1 CH |
269 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
270 | unsigned int tag, unsigned int op) | |
320ae51f | 271 | { |
e4cdf1a1 CH |
272 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
273 | struct request *rq = tags->static_rqs[tag]; | |
274 | ||
c3a148d2 BVA |
275 | rq->rq_flags = 0; |
276 | ||
e4cdf1a1 CH |
277 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
278 | rq->tag = -1; | |
279 | rq->internal_tag = tag; | |
280 | } else { | |
281 | if (blk_mq_tag_busy(data->hctx)) { | |
282 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
283 | atomic_inc(&data->hctx->nr_active); | |
284 | } | |
285 | rq->tag = tag; | |
286 | rq->internal_tag = -1; | |
287 | data->hctx->tags->rqs[rq->tag] = rq; | |
288 | } | |
289 | ||
af76e555 CH |
290 | INIT_LIST_HEAD(&rq->queuelist); |
291 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
292 | rq->q = data->q; |
293 | rq->mq_ctx = data->ctx; | |
ef295ecf | 294 | rq->cmd_flags = op; |
e4cdf1a1 | 295 | if (blk_queue_io_stat(data->q)) |
e8064021 | 296 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
297 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
298 | rq->cpu = -1; | |
af76e555 CH |
299 | INIT_HLIST_NODE(&rq->hash); |
300 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
301 | rq->rq_disk = NULL; |
302 | rq->part = NULL; | |
3ee32372 | 303 | rq->start_time = jiffies; |
af76e555 CH |
304 | #ifdef CONFIG_BLK_CGROUP |
305 | rq->rl = NULL; | |
0fec08b4 | 306 | set_start_time_ns(rq); |
af76e555 CH |
307 | rq->io_start_time_ns = 0; |
308 | #endif | |
309 | rq->nr_phys_segments = 0; | |
310 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
311 | rq->nr_integrity_segments = 0; | |
312 | #endif | |
af76e555 CH |
313 | rq->special = NULL; |
314 | /* tag was already set */ | |
af76e555 | 315 | rq->extra_len = 0; |
af76e555 | 316 | |
af76e555 | 317 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
318 | rq->timeout = 0; |
319 | ||
af76e555 CH |
320 | rq->end_io = NULL; |
321 | rq->end_io_data = NULL; | |
322 | rq->next_rq = NULL; | |
323 | ||
e4cdf1a1 CH |
324 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
325 | return rq; | |
5dee8577 CH |
326 | } |
327 | ||
d2c0d383 CH |
328 | static struct request *blk_mq_get_request(struct request_queue *q, |
329 | struct bio *bio, unsigned int op, | |
330 | struct blk_mq_alloc_data *data) | |
331 | { | |
332 | struct elevator_queue *e = q->elevator; | |
333 | struct request *rq; | |
e4cdf1a1 | 334 | unsigned int tag; |
d2c0d383 CH |
335 | |
336 | blk_queue_enter_live(q); | |
337 | data->q = q; | |
338 | if (likely(!data->ctx)) | |
339 | data->ctx = blk_mq_get_ctx(q); | |
340 | if (likely(!data->hctx)) | |
341 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
342 | if (op & REQ_NOWAIT) |
343 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
344 | |
345 | if (e) { | |
346 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
347 | ||
348 | /* | |
349 | * Flush requests are special and go directly to the | |
350 | * dispatch list. | |
351 | */ | |
5bbf4e5a CH |
352 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
353 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
354 | } |
355 | ||
e4cdf1a1 CH |
356 | tag = blk_mq_get_tag(data); |
357 | if (tag == BLK_MQ_TAG_FAIL) { | |
037cebb8 CH |
358 | blk_queue_exit(q); |
359 | return NULL; | |
d2c0d383 CH |
360 | } |
361 | ||
e4cdf1a1 | 362 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
363 | if (!op_is_flush(op)) { |
364 | rq->elv.icq = NULL; | |
5bbf4e5a | 365 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
366 | if (e->type->icq_cache && rq_ioc(bio)) |
367 | blk_mq_sched_assign_ioc(rq, bio); | |
368 | ||
5bbf4e5a CH |
369 | e->type->ops.mq.prepare_request(rq, bio); |
370 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 371 | } |
037cebb8 CH |
372 | } |
373 | data->hctx->queued++; | |
374 | return rq; | |
d2c0d383 CH |
375 | } |
376 | ||
cd6ce148 | 377 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
6f3b0e8b | 378 | unsigned int flags) |
320ae51f | 379 | { |
5a797e00 | 380 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 381 | struct request *rq; |
a492f075 | 382 | int ret; |
320ae51f | 383 | |
6f3b0e8b | 384 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
385 | if (ret) |
386 | return ERR_PTR(ret); | |
320ae51f | 387 | |
cd6ce148 | 388 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
841bac2c | 389 | |
bd166ef1 JA |
390 | blk_mq_put_ctx(alloc_data.ctx); |
391 | blk_queue_exit(q); | |
392 | ||
393 | if (!rq) | |
a492f075 | 394 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
395 | |
396 | rq->__data_len = 0; | |
397 | rq->__sector = (sector_t) -1; | |
398 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
399 | return rq; |
400 | } | |
4bb659b1 | 401 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 402 | |
cd6ce148 BVA |
403 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
404 | unsigned int op, unsigned int flags, unsigned int hctx_idx) | |
1f5bd336 | 405 | { |
6d2809d5 | 406 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 407 | struct request *rq; |
6d2809d5 | 408 | unsigned int cpu; |
1f5bd336 ML |
409 | int ret; |
410 | ||
411 | /* | |
412 | * If the tag allocator sleeps we could get an allocation for a | |
413 | * different hardware context. No need to complicate the low level | |
414 | * allocator for this for the rare use case of a command tied to | |
415 | * a specific queue. | |
416 | */ | |
417 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
418 | return ERR_PTR(-EINVAL); | |
419 | ||
420 | if (hctx_idx >= q->nr_hw_queues) | |
421 | return ERR_PTR(-EIO); | |
422 | ||
423 | ret = blk_queue_enter(q, true); | |
424 | if (ret) | |
425 | return ERR_PTR(ret); | |
426 | ||
c8712c6a CH |
427 | /* |
428 | * Check if the hardware context is actually mapped to anything. | |
429 | * If not tell the caller that it should skip this queue. | |
430 | */ | |
6d2809d5 OS |
431 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
432 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
433 | blk_queue_exit(q); | |
434 | return ERR_PTR(-EXDEV); | |
c8712c6a | 435 | } |
6d2809d5 OS |
436 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
437 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 438 | |
cd6ce148 | 439 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
c8712c6a | 440 | |
c8712c6a | 441 | blk_queue_exit(q); |
6d2809d5 OS |
442 | |
443 | if (!rq) | |
444 | return ERR_PTR(-EWOULDBLOCK); | |
445 | ||
446 | return rq; | |
1f5bd336 ML |
447 | } |
448 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
449 | ||
6af54051 | 450 | void blk_mq_free_request(struct request *rq) |
320ae51f | 451 | { |
320ae51f | 452 | struct request_queue *q = rq->q; |
6af54051 CH |
453 | struct elevator_queue *e = q->elevator; |
454 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
455 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
456 | const int sched_tag = rq->internal_tag; | |
457 | ||
5bbf4e5a | 458 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
459 | if (e && e->type->ops.mq.finish_request) |
460 | e->type->ops.mq.finish_request(rq); | |
461 | if (rq->elv.icq) { | |
462 | put_io_context(rq->elv.icq->ioc); | |
463 | rq->elv.icq = NULL; | |
464 | } | |
465 | } | |
320ae51f | 466 | |
6af54051 | 467 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 468 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 469 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
470 | |
471 | wbt_done(q->rq_wb, &rq->issue_stat); | |
0d2602ca | 472 | |
af76e555 | 473 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 474 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
475 | if (rq->tag != -1) |
476 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
477 | if (sched_tag != -1) | |
c05f8525 | 478 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 479 | blk_mq_sched_restart(hctx); |
3ef28e83 | 480 | blk_queue_exit(q); |
320ae51f | 481 | } |
1a3b595a | 482 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 483 | |
2a842aca | 484 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 485 | { |
0d11e6ac ML |
486 | blk_account_io_done(rq); |
487 | ||
91b63639 | 488 | if (rq->end_io) { |
87760e5e | 489 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 490 | rq->end_io(rq, error); |
91b63639 CH |
491 | } else { |
492 | if (unlikely(blk_bidi_rq(rq))) | |
493 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 494 | blk_mq_free_request(rq); |
91b63639 | 495 | } |
320ae51f | 496 | } |
c8a446ad | 497 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 498 | |
2a842aca | 499 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
500 | { |
501 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
502 | BUG(); | |
c8a446ad | 503 | __blk_mq_end_request(rq, error); |
63151a44 | 504 | } |
c8a446ad | 505 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 506 | |
30a91cb4 | 507 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 508 | { |
3d6efbf6 | 509 | struct request *rq = data; |
320ae51f | 510 | |
30a91cb4 | 511 | rq->q->softirq_done_fn(rq); |
320ae51f | 512 | } |
320ae51f | 513 | |
453f8341 | 514 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
515 | { |
516 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 517 | bool shared = false; |
320ae51f JA |
518 | int cpu; |
519 | ||
453f8341 CH |
520 | if (rq->internal_tag != -1) |
521 | blk_mq_sched_completed_request(rq); | |
522 | if (rq->rq_flags & RQF_STATS) { | |
523 | blk_mq_poll_stats_start(rq->q); | |
524 | blk_stat_add(rq); | |
525 | } | |
526 | ||
38535201 | 527 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
528 | rq->q->softirq_done_fn(rq); |
529 | return; | |
530 | } | |
320ae51f JA |
531 | |
532 | cpu = get_cpu(); | |
38535201 CH |
533 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
534 | shared = cpus_share_cache(cpu, ctx->cpu); | |
535 | ||
536 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 537 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
538 | rq->csd.info = rq; |
539 | rq->csd.flags = 0; | |
c46fff2a | 540 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 541 | } else { |
30a91cb4 | 542 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 543 | } |
320ae51f JA |
544 | put_cpu(); |
545 | } | |
30a91cb4 CH |
546 | |
547 | /** | |
548 | * blk_mq_complete_request - end I/O on a request | |
549 | * @rq: the request being processed | |
550 | * | |
551 | * Description: | |
552 | * Ends all I/O on a request. It does not handle partial completions. | |
553 | * The actual completion happens out-of-order, through a IPI handler. | |
554 | **/ | |
08e0029a | 555 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 556 | { |
95f09684 JA |
557 | struct request_queue *q = rq->q; |
558 | ||
559 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 560 | return; |
08e0029a | 561 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 562 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
563 | } |
564 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 565 | |
973c0191 KB |
566 | int blk_mq_request_started(struct request *rq) |
567 | { | |
568 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
569 | } | |
570 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
571 | ||
e2490073 | 572 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
573 | { |
574 | struct request_queue *q = rq->q; | |
575 | ||
bd166ef1 JA |
576 | blk_mq_sched_started_request(rq); |
577 | ||
320ae51f JA |
578 | trace_block_rq_issue(q, rq); |
579 | ||
cf43e6be | 580 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 581 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 582 | rq->rq_flags |= RQF_STATS; |
87760e5e | 583 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
584 | } |
585 | ||
2b8393b4 | 586 | blk_add_timer(rq); |
87ee7b11 | 587 | |
538b7534 JA |
588 | /* |
589 | * Ensure that ->deadline is visible before set the started | |
590 | * flag and clear the completed flag. | |
591 | */ | |
592 | smp_mb__before_atomic(); | |
593 | ||
87ee7b11 JA |
594 | /* |
595 | * Mark us as started and clear complete. Complete might have been | |
596 | * set if requeue raced with timeout, which then marked it as | |
597 | * complete. So be sure to clear complete again when we start | |
598 | * the request, otherwise we'll ignore the completion event. | |
599 | */ | |
4b570521 JA |
600 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
601 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
602 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
603 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
604 | |
605 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
606 | /* | |
607 | * Make sure space for the drain appears. We know we can do | |
608 | * this because max_hw_segments has been adjusted to be one | |
609 | * fewer than the device can handle. | |
610 | */ | |
611 | rq->nr_phys_segments++; | |
612 | } | |
320ae51f | 613 | } |
e2490073 | 614 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 615 | |
d9d149a3 ML |
616 | /* |
617 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 618 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
619 | * but given rq->deadline is just set in .queue_rq() under |
620 | * this situation, the race won't be possible in reality because | |
621 | * rq->timeout should be set as big enough to cover the window | |
622 | * between blk_mq_start_request() called from .queue_rq() and | |
623 | * clearing REQ_ATOM_STARTED here. | |
624 | */ | |
ed0791b2 | 625 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
626 | { |
627 | struct request_queue *q = rq->q; | |
628 | ||
629 | trace_block_rq_requeue(q, rq); | |
87760e5e | 630 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 631 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 632 | |
e2490073 CH |
633 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
634 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
635 | rq->nr_phys_segments--; | |
636 | } | |
320ae51f JA |
637 | } |
638 | ||
2b053aca | 639 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 640 | { |
ed0791b2 | 641 | __blk_mq_requeue_request(rq); |
ed0791b2 | 642 | |
ed0791b2 | 643 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 644 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
645 | } |
646 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
647 | ||
6fca6a61 CH |
648 | static void blk_mq_requeue_work(struct work_struct *work) |
649 | { | |
650 | struct request_queue *q = | |
2849450a | 651 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
652 | LIST_HEAD(rq_list); |
653 | struct request *rq, *next; | |
6fca6a61 | 654 | |
18e9781d | 655 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 656 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 657 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
658 | |
659 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 660 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
661 | continue; |
662 | ||
e8064021 | 663 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 664 | list_del_init(&rq->queuelist); |
bd6737f1 | 665 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
666 | } |
667 | ||
668 | while (!list_empty(&rq_list)) { | |
669 | rq = list_entry(rq_list.next, struct request, queuelist); | |
670 | list_del_init(&rq->queuelist); | |
bd6737f1 | 671 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
672 | } |
673 | ||
52d7f1b5 | 674 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
675 | } |
676 | ||
2b053aca BVA |
677 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
678 | bool kick_requeue_list) | |
6fca6a61 CH |
679 | { |
680 | struct request_queue *q = rq->q; | |
681 | unsigned long flags; | |
682 | ||
683 | /* | |
684 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
685 | * request head insertation from the workqueue. | |
686 | */ | |
e8064021 | 687 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
688 | |
689 | spin_lock_irqsave(&q->requeue_lock, flags); | |
690 | if (at_head) { | |
e8064021 | 691 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
692 | list_add(&rq->queuelist, &q->requeue_list); |
693 | } else { | |
694 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
695 | } | |
696 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
697 | |
698 | if (kick_requeue_list) | |
699 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
700 | } |
701 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
702 | ||
703 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
704 | { | |
2849450a | 705 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
706 | } |
707 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
708 | ||
2849450a MS |
709 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
710 | unsigned long msecs) | |
711 | { | |
712 | kblockd_schedule_delayed_work(&q->requeue_work, | |
713 | msecs_to_jiffies(msecs)); | |
714 | } | |
715 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
716 | ||
0e62f51f JA |
717 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
718 | { | |
88c7b2b7 JA |
719 | if (tag < tags->nr_tags) { |
720 | prefetch(tags->rqs[tag]); | |
4ee86bab | 721 | return tags->rqs[tag]; |
88c7b2b7 | 722 | } |
4ee86bab HR |
723 | |
724 | return NULL; | |
24d2f903 CH |
725 | } |
726 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
727 | ||
320ae51f | 728 | struct blk_mq_timeout_data { |
46f92d42 CH |
729 | unsigned long next; |
730 | unsigned int next_set; | |
320ae51f JA |
731 | }; |
732 | ||
90415837 | 733 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 734 | { |
f8a5b122 | 735 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 736 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
737 | |
738 | /* | |
739 | * We know that complete is set at this point. If STARTED isn't set | |
740 | * anymore, then the request isn't active and the "timeout" should | |
741 | * just be ignored. This can happen due to the bitflag ordering. | |
742 | * Timeout first checks if STARTED is set, and if it is, assumes | |
743 | * the request is active. But if we race with completion, then | |
48b99c9d | 744 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
745 | * a timeout event with a request that isn't active. |
746 | */ | |
46f92d42 CH |
747 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
748 | return; | |
87ee7b11 | 749 | |
46f92d42 | 750 | if (ops->timeout) |
0152fb6b | 751 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
752 | |
753 | switch (ret) { | |
754 | case BLK_EH_HANDLED: | |
755 | __blk_mq_complete_request(req); | |
756 | break; | |
757 | case BLK_EH_RESET_TIMER: | |
758 | blk_add_timer(req); | |
759 | blk_clear_rq_complete(req); | |
760 | break; | |
761 | case BLK_EH_NOT_HANDLED: | |
762 | break; | |
763 | default: | |
764 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
765 | break; | |
766 | } | |
87ee7b11 | 767 | } |
5b3f25fc | 768 | |
81481eb4 CH |
769 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
770 | struct request *rq, void *priv, bool reserved) | |
771 | { | |
772 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 773 | |
95a49603 | 774 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 775 | return; |
87ee7b11 | 776 | |
d9d149a3 ML |
777 | /* |
778 | * The rq being checked may have been freed and reallocated | |
779 | * out already here, we avoid this race by checking rq->deadline | |
780 | * and REQ_ATOM_COMPLETE flag together: | |
781 | * | |
782 | * - if rq->deadline is observed as new value because of | |
783 | * reusing, the rq won't be timed out because of timing. | |
784 | * - if rq->deadline is observed as previous value, | |
785 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
786 | * because we put a barrier between setting rq->deadline | |
787 | * and clearing the flag in blk_mq_start_request(), so | |
788 | * this rq won't be timed out too. | |
789 | */ | |
46f92d42 CH |
790 | if (time_after_eq(jiffies, rq->deadline)) { |
791 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 792 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
793 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
794 | data->next = rq->deadline; | |
795 | data->next_set = 1; | |
796 | } | |
87ee7b11 JA |
797 | } |
798 | ||
287922eb | 799 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 800 | { |
287922eb CH |
801 | struct request_queue *q = |
802 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
803 | struct blk_mq_timeout_data data = { |
804 | .next = 0, | |
805 | .next_set = 0, | |
806 | }; | |
81481eb4 | 807 | int i; |
320ae51f | 808 | |
71f79fb3 GKB |
809 | /* A deadlock might occur if a request is stuck requiring a |
810 | * timeout at the same time a queue freeze is waiting | |
811 | * completion, since the timeout code would not be able to | |
812 | * acquire the queue reference here. | |
813 | * | |
814 | * That's why we don't use blk_queue_enter here; instead, we use | |
815 | * percpu_ref_tryget directly, because we need to be able to | |
816 | * obtain a reference even in the short window between the queue | |
817 | * starting to freeze, by dropping the first reference in | |
1671d522 | 818 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
819 | * consumed, marked by the instant q_usage_counter reaches |
820 | * zero. | |
821 | */ | |
822 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
823 | return; |
824 | ||
0bf6cd5b | 825 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 826 | |
81481eb4 CH |
827 | if (data.next_set) { |
828 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
829 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 830 | } else { |
0bf6cd5b CH |
831 | struct blk_mq_hw_ctx *hctx; |
832 | ||
f054b56c ML |
833 | queue_for_each_hw_ctx(q, hctx, i) { |
834 | /* the hctx may be unmapped, so check it here */ | |
835 | if (blk_mq_hw_queue_mapped(hctx)) | |
836 | blk_mq_tag_idle(hctx); | |
837 | } | |
0d2602ca | 838 | } |
287922eb | 839 | blk_queue_exit(q); |
320ae51f JA |
840 | } |
841 | ||
88459642 OS |
842 | struct flush_busy_ctx_data { |
843 | struct blk_mq_hw_ctx *hctx; | |
844 | struct list_head *list; | |
845 | }; | |
846 | ||
847 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
848 | { | |
849 | struct flush_busy_ctx_data *flush_data = data; | |
850 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
851 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
852 | ||
853 | sbitmap_clear_bit(sb, bitnr); | |
854 | spin_lock(&ctx->lock); | |
855 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
856 | spin_unlock(&ctx->lock); | |
857 | return true; | |
858 | } | |
859 | ||
1429d7c9 JA |
860 | /* |
861 | * Process software queues that have been marked busy, splicing them | |
862 | * to the for-dispatch | |
863 | */ | |
2c3ad667 | 864 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 865 | { |
88459642 OS |
866 | struct flush_busy_ctx_data data = { |
867 | .hctx = hctx, | |
868 | .list = list, | |
869 | }; | |
1429d7c9 | 870 | |
88459642 | 871 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 872 | } |
2c3ad667 | 873 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 874 | |
703fd1c0 JA |
875 | static inline unsigned int queued_to_index(unsigned int queued) |
876 | { | |
877 | if (!queued) | |
878 | return 0; | |
1429d7c9 | 879 | |
703fd1c0 | 880 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
881 | } |
882 | ||
bd6737f1 JA |
883 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
884 | bool wait) | |
bd166ef1 JA |
885 | { |
886 | struct blk_mq_alloc_data data = { | |
887 | .q = rq->q, | |
bd166ef1 JA |
888 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
889 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
890 | }; | |
891 | ||
5feeacdd JA |
892 | might_sleep_if(wait); |
893 | ||
81380ca1 OS |
894 | if (rq->tag != -1) |
895 | goto done; | |
bd166ef1 | 896 | |
415b806d SG |
897 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
898 | data.flags |= BLK_MQ_REQ_RESERVED; | |
899 | ||
bd166ef1 JA |
900 | rq->tag = blk_mq_get_tag(&data); |
901 | if (rq->tag >= 0) { | |
200e86b3 JA |
902 | if (blk_mq_tag_busy(data.hctx)) { |
903 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
904 | atomic_inc(&data.hctx->nr_active); | |
905 | } | |
bd166ef1 | 906 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
907 | } |
908 | ||
81380ca1 OS |
909 | done: |
910 | if (hctx) | |
911 | *hctx = data.hctx; | |
912 | return rq->tag != -1; | |
bd166ef1 JA |
913 | } |
914 | ||
113285b4 JA |
915 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
916 | struct request *rq) | |
99cf1dc5 | 917 | { |
99cf1dc5 JA |
918 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
919 | rq->tag = -1; | |
920 | ||
921 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
922 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
923 | atomic_dec(&hctx->nr_active); | |
924 | } | |
925 | } | |
926 | ||
113285b4 JA |
927 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
928 | struct request *rq) | |
929 | { | |
930 | if (rq->tag == -1 || rq->internal_tag == -1) | |
931 | return; | |
932 | ||
933 | __blk_mq_put_driver_tag(hctx, rq); | |
934 | } | |
935 | ||
936 | static void blk_mq_put_driver_tag(struct request *rq) | |
937 | { | |
938 | struct blk_mq_hw_ctx *hctx; | |
939 | ||
940 | if (rq->tag == -1 || rq->internal_tag == -1) | |
941 | return; | |
942 | ||
943 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
944 | __blk_mq_put_driver_tag(hctx, rq); | |
945 | } | |
946 | ||
bd166ef1 JA |
947 | /* |
948 | * If we fail getting a driver tag because all the driver tags are already | |
949 | * assigned and on the dispatch list, BUT the first entry does not have a | |
950 | * tag, then we could deadlock. For that case, move entries with assigned | |
951 | * driver tags to the front, leaving the set of tagged requests in the | |
952 | * same order, and the untagged set in the same order. | |
953 | */ | |
954 | static bool reorder_tags_to_front(struct list_head *list) | |
955 | { | |
956 | struct request *rq, *tmp, *first = NULL; | |
957 | ||
958 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
959 | if (rq == first) | |
960 | break; | |
961 | if (rq->tag != -1) { | |
962 | list_move(&rq->queuelist, list); | |
963 | if (!first) | |
964 | first = rq; | |
965 | } | |
966 | } | |
967 | ||
968 | return first != NULL; | |
969 | } | |
970 | ||
ac6424b9 | 971 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags, |
da55f2cc OS |
972 | void *key) |
973 | { | |
974 | struct blk_mq_hw_ctx *hctx; | |
975 | ||
976 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
977 | ||
2055da97 | 978 | list_del(&wait->entry); |
da55f2cc OS |
979 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); |
980 | blk_mq_run_hw_queue(hctx, true); | |
981 | return 1; | |
982 | } | |
983 | ||
984 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
985 | { | |
986 | struct sbq_wait_state *ws; | |
987 | ||
988 | /* | |
989 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
990 | * The thread which wins the race to grab this bit adds the hardware | |
991 | * queue to the wait queue. | |
992 | */ | |
993 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
994 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
995 | return false; | |
996 | ||
997 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
998 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
999 | ||
1000 | /* | |
1001 | * As soon as this returns, it's no longer safe to fiddle with | |
1002 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
1003 | * and unlock the bit. | |
1004 | */ | |
1005 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
1006 | return true; | |
1007 | } | |
1008 | ||
81380ca1 | 1009 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
320ae51f | 1010 | { |
81380ca1 | 1011 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1012 | struct request *rq; |
fc17b653 | 1013 | int errors, queued; |
320ae51f | 1014 | |
81380ca1 OS |
1015 | if (list_empty(list)) |
1016 | return false; | |
1017 | ||
320ae51f JA |
1018 | /* |
1019 | * Now process all the entries, sending them to the driver. | |
1020 | */ | |
93efe981 | 1021 | errors = queued = 0; |
81380ca1 | 1022 | do { |
74c45052 | 1023 | struct blk_mq_queue_data bd; |
fc17b653 | 1024 | blk_status_t ret; |
320ae51f | 1025 | |
f04c3df3 | 1026 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1027 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1028 | if (!queued && reorder_tags_to_front(list)) | |
1029 | continue; | |
3c782d67 JA |
1030 | |
1031 | /* | |
da55f2cc OS |
1032 | * The initial allocation attempt failed, so we need to |
1033 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1034 | */ |
807b1041 OS |
1035 | if (!blk_mq_dispatch_wait_add(hctx)) |
1036 | break; | |
1037 | ||
1038 | /* | |
1039 | * It's possible that a tag was freed in the window | |
1040 | * between the allocation failure and adding the | |
1041 | * hardware queue to the wait queue. | |
1042 | */ | |
1043 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
3c782d67 | 1044 | break; |
bd166ef1 | 1045 | } |
da55f2cc | 1046 | |
320ae51f | 1047 | list_del_init(&rq->queuelist); |
320ae51f | 1048 | |
74c45052 | 1049 | bd.rq = rq; |
113285b4 JA |
1050 | |
1051 | /* | |
1052 | * Flag last if we have no more requests, or if we have more | |
1053 | * but can't assign a driver tag to it. | |
1054 | */ | |
1055 | if (list_empty(list)) | |
1056 | bd.last = true; | |
1057 | else { | |
1058 | struct request *nxt; | |
1059 | ||
1060 | nxt = list_first_entry(list, struct request, queuelist); | |
1061 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1062 | } | |
74c45052 JA |
1063 | |
1064 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1065 | if (ret == BLK_STS_RESOURCE) { |
113285b4 | 1066 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1067 | list_add(&rq->queuelist, list); |
ed0791b2 | 1068 | __blk_mq_requeue_request(rq); |
320ae51f | 1069 | break; |
fc17b653 CH |
1070 | } |
1071 | ||
1072 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1073 | errors++; |
2a842aca | 1074 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1075 | continue; |
320ae51f JA |
1076 | } |
1077 | ||
fc17b653 | 1078 | queued++; |
81380ca1 | 1079 | } while (!list_empty(list)); |
320ae51f | 1080 | |
703fd1c0 | 1081 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1082 | |
1083 | /* | |
1084 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1085 | * that is where we will continue on next queue run. | |
1086 | */ | |
f04c3df3 | 1087 | if (!list_empty(list)) { |
113285b4 | 1088 | /* |
710c785f BVA |
1089 | * If an I/O scheduler has been configured and we got a driver |
1090 | * tag for the next request already, free it again. | |
113285b4 JA |
1091 | */ |
1092 | rq = list_first_entry(list, struct request, queuelist); | |
1093 | blk_mq_put_driver_tag(rq); | |
1094 | ||
320ae51f | 1095 | spin_lock(&hctx->lock); |
c13660a0 | 1096 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1097 | spin_unlock(&hctx->lock); |
f04c3df3 | 1098 | |
9ba52e58 | 1099 | /* |
710c785f BVA |
1100 | * If SCHED_RESTART was set by the caller of this function and |
1101 | * it is no longer set that means that it was cleared by another | |
1102 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1103 | * |
710c785f BVA |
1104 | * If TAG_WAITING is set that means that an I/O scheduler has |
1105 | * been configured and another thread is waiting for a driver | |
1106 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1107 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1108 | * |
710c785f BVA |
1109 | * If no I/O scheduler has been configured it is possible that |
1110 | * the hardware queue got stopped and restarted before requests | |
1111 | * were pushed back onto the dispatch list. Rerun the queue to | |
1112 | * avoid starvation. Notes: | |
1113 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1114 | * been stopped before rerunning a queue. | |
1115 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1116 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1117 | * and dm-rq. |
bd166ef1 | 1118 | */ |
da55f2cc OS |
1119 | if (!blk_mq_sched_needs_restart(hctx) && |
1120 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1121 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1122 | } |
f04c3df3 | 1123 | |
93efe981 | 1124 | return (queued + errors) != 0; |
f04c3df3 JA |
1125 | } |
1126 | ||
6a83e74d BVA |
1127 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1128 | { | |
1129 | int srcu_idx; | |
1130 | ||
b7a71e66 JA |
1131 | /* |
1132 | * We should be running this queue from one of the CPUs that | |
1133 | * are mapped to it. | |
1134 | */ | |
6a83e74d BVA |
1135 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1136 | cpu_online(hctx->next_cpu)); | |
1137 | ||
b7a71e66 JA |
1138 | /* |
1139 | * We can't run the queue inline with ints disabled. Ensure that | |
1140 | * we catch bad users of this early. | |
1141 | */ | |
1142 | WARN_ON_ONCE(in_interrupt()); | |
1143 | ||
6a83e74d BVA |
1144 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1145 | rcu_read_lock(); | |
bd166ef1 | 1146 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1147 | rcu_read_unlock(); |
1148 | } else { | |
bf4907c0 JA |
1149 | might_sleep(); |
1150 | ||
07319678 | 1151 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
bd166ef1 | 1152 | blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1153 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d BVA |
1154 | } |
1155 | } | |
1156 | ||
506e931f JA |
1157 | /* |
1158 | * It'd be great if the workqueue API had a way to pass | |
1159 | * in a mask and had some smarts for more clever placement. | |
1160 | * For now we just round-robin here, switching for every | |
1161 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1162 | */ | |
1163 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1164 | { | |
b657d7e6 CH |
1165 | if (hctx->queue->nr_hw_queues == 1) |
1166 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1167 | |
1168 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1169 | int next_cpu; |
506e931f JA |
1170 | |
1171 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1172 | if (next_cpu >= nr_cpu_ids) | |
1173 | next_cpu = cpumask_first(hctx->cpumask); | |
1174 | ||
1175 | hctx->next_cpu = next_cpu; | |
1176 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1177 | } | |
1178 | ||
b657d7e6 | 1179 | return hctx->next_cpu; |
506e931f JA |
1180 | } |
1181 | ||
7587a5ae BVA |
1182 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1183 | unsigned long msecs) | |
320ae51f | 1184 | { |
5435c023 BVA |
1185 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1186 | return; | |
1187 | ||
1188 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1189 | return; |
1190 | ||
1b792f2f | 1191 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1192 | int cpu = get_cpu(); |
1193 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1194 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1195 | put_cpu(); |
398205b8 PB |
1196 | return; |
1197 | } | |
e4043dcf | 1198 | |
2a90d4aa | 1199 | put_cpu(); |
e4043dcf | 1200 | } |
398205b8 | 1201 | |
9f993737 JA |
1202 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1203 | &hctx->run_work, | |
1204 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1205 | } |
1206 | ||
1207 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1208 | { | |
1209 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1210 | } | |
1211 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1212 | ||
1213 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1214 | { | |
1215 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1216 | } |
5b727272 | 1217 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1218 | |
b94ec296 | 1219 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1220 | { |
1221 | struct blk_mq_hw_ctx *hctx; | |
1222 | int i; | |
1223 | ||
1224 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1225 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1226 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1227 | continue; |
1228 | ||
b94ec296 | 1229 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1230 | } |
1231 | } | |
b94ec296 | 1232 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1233 | |
fd001443 BVA |
1234 | /** |
1235 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1236 | * @q: request queue. | |
1237 | * | |
1238 | * The caller is responsible for serializing this function against | |
1239 | * blk_mq_{start,stop}_hw_queue(). | |
1240 | */ | |
1241 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1242 | { | |
1243 | struct blk_mq_hw_ctx *hctx; | |
1244 | int i; | |
1245 | ||
1246 | queue_for_each_hw_ctx(q, hctx, i) | |
1247 | if (blk_mq_hctx_stopped(hctx)) | |
1248 | return true; | |
1249 | ||
1250 | return false; | |
1251 | } | |
1252 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1253 | ||
39a70c76 ML |
1254 | /* |
1255 | * This function is often used for pausing .queue_rq() by driver when | |
1256 | * there isn't enough resource or some conditions aren't satisfied, and | |
1257 | * BLK_MQ_RQ_QUEUE_BUSY is usually returned. | |
1258 | * | |
1259 | * We do not guarantee that dispatch can be drained or blocked | |
1260 | * after blk_mq_stop_hw_queue() returns. Please use | |
1261 | * blk_mq_quiesce_queue() for that requirement. | |
1262 | */ | |
2719aa21 JA |
1263 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1264 | { | |
641a9ed6 | 1265 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1266 | |
641a9ed6 | 1267 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1268 | } |
641a9ed6 | 1269 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1270 | |
39a70c76 ML |
1271 | /* |
1272 | * This function is often used for pausing .queue_rq() by driver when | |
1273 | * there isn't enough resource or some conditions aren't satisfied, and | |
1274 | * BLK_MQ_RQ_QUEUE_BUSY is usually returned. | |
1275 | * | |
1276 | * We do not guarantee that dispatch can be drained or blocked | |
1277 | * after blk_mq_stop_hw_queues() returns. Please use | |
1278 | * blk_mq_quiesce_queue() for that requirement. | |
1279 | */ | |
2719aa21 JA |
1280 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1281 | { | |
641a9ed6 ML |
1282 | struct blk_mq_hw_ctx *hctx; |
1283 | int i; | |
1284 | ||
1285 | queue_for_each_hw_ctx(q, hctx, i) | |
1286 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1287 | } |
1288 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1289 | ||
320ae51f JA |
1290 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1291 | { | |
1292 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1293 | |
0ffbce80 | 1294 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1295 | } |
1296 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1297 | ||
2f268556 CH |
1298 | void blk_mq_start_hw_queues(struct request_queue *q) |
1299 | { | |
1300 | struct blk_mq_hw_ctx *hctx; | |
1301 | int i; | |
1302 | ||
1303 | queue_for_each_hw_ctx(q, hctx, i) | |
1304 | blk_mq_start_hw_queue(hctx); | |
1305 | } | |
1306 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1307 | ||
ae911c5e JA |
1308 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1309 | { | |
1310 | if (!blk_mq_hctx_stopped(hctx)) | |
1311 | return; | |
1312 | ||
1313 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1314 | blk_mq_run_hw_queue(hctx, async); | |
1315 | } | |
1316 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1317 | ||
1b4a3258 | 1318 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1319 | { |
1320 | struct blk_mq_hw_ctx *hctx; | |
1321 | int i; | |
1322 | ||
ae911c5e JA |
1323 | queue_for_each_hw_ctx(q, hctx, i) |
1324 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1325 | } |
1326 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1327 | ||
70f4db63 | 1328 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1329 | { |
1330 | struct blk_mq_hw_ctx *hctx; | |
1331 | ||
9f993737 | 1332 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1333 | |
21c6e939 JA |
1334 | /* |
1335 | * If we are stopped, don't run the queue. The exception is if | |
1336 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1337 | * the STOPPED bit and run it. | |
1338 | */ | |
1339 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1340 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1341 | return; | |
7587a5ae | 1342 | |
21c6e939 JA |
1343 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1344 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1345 | } | |
7587a5ae BVA |
1346 | |
1347 | __blk_mq_run_hw_queue(hctx); | |
1348 | } | |
1349 | ||
70f4db63 CH |
1350 | |
1351 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1352 | { | |
5435c023 | 1353 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1354 | return; |
70f4db63 | 1355 | |
21c6e939 JA |
1356 | /* |
1357 | * Stop the hw queue, then modify currently delayed work. | |
1358 | * This should prevent us from running the queue prematurely. | |
1359 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1360 | */ | |
7e79dadc | 1361 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1362 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1363 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1364 | &hctx->run_work, | |
1365 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1366 | } |
1367 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1368 | ||
cfd0c552 | 1369 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1370 | struct request *rq, |
1371 | bool at_head) | |
320ae51f | 1372 | { |
e57690fe JA |
1373 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1374 | ||
7b607814 BVA |
1375 | lockdep_assert_held(&ctx->lock); |
1376 | ||
01b983c9 JA |
1377 | trace_block_rq_insert(hctx->queue, rq); |
1378 | ||
72a0a36e CH |
1379 | if (at_head) |
1380 | list_add(&rq->queuelist, &ctx->rq_list); | |
1381 | else | |
1382 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1383 | } |
4bb659b1 | 1384 | |
2c3ad667 JA |
1385 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1386 | bool at_head) | |
cfd0c552 ML |
1387 | { |
1388 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1389 | ||
7b607814 BVA |
1390 | lockdep_assert_held(&ctx->lock); |
1391 | ||
e57690fe | 1392 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1393 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1394 | } |
1395 | ||
bd166ef1 JA |
1396 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1397 | struct list_head *list) | |
320ae51f JA |
1398 | |
1399 | { | |
320ae51f JA |
1400 | /* |
1401 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1402 | * offline now | |
1403 | */ | |
1404 | spin_lock(&ctx->lock); | |
1405 | while (!list_empty(list)) { | |
1406 | struct request *rq; | |
1407 | ||
1408 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1409 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1410 | list_del_init(&rq->queuelist); |
e57690fe | 1411 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1412 | } |
cfd0c552 | 1413 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1414 | spin_unlock(&ctx->lock); |
320ae51f JA |
1415 | } |
1416 | ||
1417 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1418 | { | |
1419 | struct request *rqa = container_of(a, struct request, queuelist); | |
1420 | struct request *rqb = container_of(b, struct request, queuelist); | |
1421 | ||
1422 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1423 | (rqa->mq_ctx == rqb->mq_ctx && | |
1424 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1425 | } | |
1426 | ||
1427 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1428 | { | |
1429 | struct blk_mq_ctx *this_ctx; | |
1430 | struct request_queue *this_q; | |
1431 | struct request *rq; | |
1432 | LIST_HEAD(list); | |
1433 | LIST_HEAD(ctx_list); | |
1434 | unsigned int depth; | |
1435 | ||
1436 | list_splice_init(&plug->mq_list, &list); | |
1437 | ||
1438 | list_sort(NULL, &list, plug_ctx_cmp); | |
1439 | ||
1440 | this_q = NULL; | |
1441 | this_ctx = NULL; | |
1442 | depth = 0; | |
1443 | ||
1444 | while (!list_empty(&list)) { | |
1445 | rq = list_entry_rq(list.next); | |
1446 | list_del_init(&rq->queuelist); | |
1447 | BUG_ON(!rq->q); | |
1448 | if (rq->mq_ctx != this_ctx) { | |
1449 | if (this_ctx) { | |
bd166ef1 JA |
1450 | trace_block_unplug(this_q, depth, from_schedule); |
1451 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1452 | &ctx_list, | |
1453 | from_schedule); | |
320ae51f JA |
1454 | } |
1455 | ||
1456 | this_ctx = rq->mq_ctx; | |
1457 | this_q = rq->q; | |
1458 | depth = 0; | |
1459 | } | |
1460 | ||
1461 | depth++; | |
1462 | list_add_tail(&rq->queuelist, &ctx_list); | |
1463 | } | |
1464 | ||
1465 | /* | |
1466 | * If 'this_ctx' is set, we know we have entries to complete | |
1467 | * on 'ctx_list'. Do those. | |
1468 | */ | |
1469 | if (this_ctx) { | |
bd166ef1 JA |
1470 | trace_block_unplug(this_q, depth, from_schedule); |
1471 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1472 | from_schedule); | |
320ae51f JA |
1473 | } |
1474 | } | |
1475 | ||
1476 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1477 | { | |
da8d7f07 | 1478 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1479 | |
6e85eaf3 | 1480 | blk_account_io_start(rq, true); |
320ae51f JA |
1481 | } |
1482 | ||
274a5843 JA |
1483 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1484 | { | |
1485 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1486 | !blk_queue_nomerges(hctx->queue); | |
1487 | } | |
1488 | ||
ab42f35d ML |
1489 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1490 | struct blk_mq_ctx *ctx, | |
1491 | struct request *rq) | |
1492 | { | |
1493 | spin_lock(&ctx->lock); | |
1494 | __blk_mq_insert_request(hctx, rq, false); | |
1495 | spin_unlock(&ctx->lock); | |
07068d5b | 1496 | } |
14ec77f3 | 1497 | |
fd2d3326 JA |
1498 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1499 | { | |
bd166ef1 JA |
1500 | if (rq->tag != -1) |
1501 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1502 | ||
1503 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1504 | } |
1505 | ||
d964f04a ML |
1506 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1507 | struct request *rq, | |
1508 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1509 | { |
f984df1f | 1510 | struct request_queue *q = rq->q; |
f984df1f SL |
1511 | struct blk_mq_queue_data bd = { |
1512 | .rq = rq, | |
d945a365 | 1513 | .last = true, |
f984df1f | 1514 | }; |
bd166ef1 | 1515 | blk_qc_t new_cookie; |
f06345ad | 1516 | blk_status_t ret; |
d964f04a ML |
1517 | bool run_queue = true; |
1518 | ||
f4560ffe ML |
1519 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1520 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1521 | run_queue = false; |
1522 | goto insert; | |
1523 | } | |
f984df1f | 1524 | |
bd166ef1 | 1525 | if (q->elevator) |
2253efc8 BVA |
1526 | goto insert; |
1527 | ||
d964f04a | 1528 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1529 | goto insert; |
1530 | ||
1531 | new_cookie = request_to_qc_t(hctx, rq); | |
1532 | ||
f984df1f SL |
1533 | /* |
1534 | * For OK queue, we are done. For error, kill it. Any other | |
1535 | * error (busy), just add it to our list as we previously | |
1536 | * would have done | |
1537 | */ | |
1538 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1539 | switch (ret) { |
1540 | case BLK_STS_OK: | |
7b371636 | 1541 | *cookie = new_cookie; |
2253efc8 | 1542 | return; |
fc17b653 CH |
1543 | case BLK_STS_RESOURCE: |
1544 | __blk_mq_requeue_request(rq); | |
1545 | goto insert; | |
1546 | default: | |
7b371636 | 1547 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1548 | blk_mq_end_request(rq, ret); |
2253efc8 | 1549 | return; |
f984df1f | 1550 | } |
7b371636 | 1551 | |
2253efc8 | 1552 | insert: |
d964f04a | 1553 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1554 | } |
1555 | ||
5eb6126e CH |
1556 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1557 | struct request *rq, blk_qc_t *cookie) | |
1558 | { | |
1559 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1560 | rcu_read_lock(); | |
d964f04a | 1561 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1562 | rcu_read_unlock(); |
1563 | } else { | |
bf4907c0 JA |
1564 | unsigned int srcu_idx; |
1565 | ||
1566 | might_sleep(); | |
1567 | ||
07319678 | 1568 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1569 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1570 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1571 | } |
1572 | } | |
1573 | ||
dece1635 | 1574 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1575 | { |
ef295ecf | 1576 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1577 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1578 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1579 | struct request *rq; |
5eb6126e | 1580 | unsigned int request_count = 0; |
f984df1f | 1581 | struct blk_plug *plug; |
5b3f341f | 1582 | struct request *same_queue_rq = NULL; |
7b371636 | 1583 | blk_qc_t cookie; |
87760e5e | 1584 | unsigned int wb_acct; |
07068d5b JA |
1585 | |
1586 | blk_queue_bounce(q, &bio); | |
1587 | ||
af67c31f | 1588 | blk_queue_split(q, &bio); |
f36ea50c | 1589 | |
e23947bd | 1590 | if (!bio_integrity_prep(bio)) |
dece1635 | 1591 | return BLK_QC_T_NONE; |
07068d5b | 1592 | |
87c279e6 OS |
1593 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1594 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1595 | return BLK_QC_T_NONE; | |
f984df1f | 1596 | |
bd166ef1 JA |
1597 | if (blk_mq_sched_bio_merge(q, bio)) |
1598 | return BLK_QC_T_NONE; | |
1599 | ||
87760e5e JA |
1600 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1601 | ||
bd166ef1 JA |
1602 | trace_block_getrq(q, bio, bio->bi_opf); |
1603 | ||
d2c0d383 | 1604 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1605 | if (unlikely(!rq)) { |
1606 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1607 | if (bio->bi_opf & REQ_NOWAIT) |
1608 | bio_wouldblock_error(bio); | |
dece1635 | 1609 | return BLK_QC_T_NONE; |
87760e5e JA |
1610 | } |
1611 | ||
1612 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1613 | |
fd2d3326 | 1614 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1615 | |
f984df1f | 1616 | plug = current->plug; |
07068d5b | 1617 | if (unlikely(is_flush_fua)) { |
f984df1f | 1618 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1619 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1620 | if (q->elevator) { |
1621 | blk_mq_sched_insert_request(rq, false, true, true, | |
1622 | true); | |
6a83e74d | 1623 | } else { |
a4d907b6 CH |
1624 | blk_insert_flush(rq); |
1625 | blk_mq_run_hw_queue(data.hctx, true); | |
6a83e74d | 1626 | } |
a4d907b6 | 1627 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1628 | struct request *last = NULL; |
1629 | ||
b00c53e8 | 1630 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1631 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1632 | |
1633 | /* | |
1634 | * @request_count may become stale because of schedule | |
1635 | * out, so check the list again. | |
1636 | */ | |
1637 | if (list_empty(&plug->mq_list)) | |
1638 | request_count = 0; | |
254d259d CH |
1639 | else if (blk_queue_nomerges(q)) |
1640 | request_count = blk_plug_queued_count(q); | |
1641 | ||
676d0607 | 1642 | if (!request_count) |
e6c4438b | 1643 | trace_block_plug(q); |
600271d9 SL |
1644 | else |
1645 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1646 | |
600271d9 SL |
1647 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1648 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1649 | blk_flush_plug_list(plug, false); |
1650 | trace_block_plug(q); | |
320ae51f | 1651 | } |
b094f89c | 1652 | |
e6c4438b | 1653 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1654 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1655 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1656 | |
07068d5b | 1657 | /* |
6a83e74d | 1658 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1659 | * Otherwise the existing request in the plug list will be |
1660 | * issued. So the plug list will have one request at most | |
2299722c CH |
1661 | * The plug list might get flushed before this. If that happens, |
1662 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1663 | */ |
2299722c CH |
1664 | if (list_empty(&plug->mq_list)) |
1665 | same_queue_rq = NULL; | |
1666 | if (same_queue_rq) | |
1667 | list_del_init(&same_queue_rq->queuelist); | |
1668 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1669 | ||
bf4907c0 JA |
1670 | blk_mq_put_ctx(data.ctx); |
1671 | ||
dad7a3be ML |
1672 | if (same_queue_rq) { |
1673 | data.hctx = blk_mq_map_queue(q, | |
1674 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1675 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1676 | &cookie); | |
dad7a3be | 1677 | } |
a4d907b6 | 1678 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1679 | blk_mq_put_ctx(data.ctx); |
2299722c | 1680 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1681 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1682 | } else if (q->elevator) { |
b00c53e8 | 1683 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1684 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1685 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1686 | } else { |
b00c53e8 | 1687 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1688 | blk_mq_bio_to_request(rq, bio); |
1689 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1690 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1691 | } |
320ae51f | 1692 | |
7b371636 | 1693 | return cookie; |
320ae51f JA |
1694 | } |
1695 | ||
cc71a6f4 JA |
1696 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1697 | unsigned int hctx_idx) | |
95363efd | 1698 | { |
e9b267d9 | 1699 | struct page *page; |
320ae51f | 1700 | |
24d2f903 | 1701 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1702 | int i; |
320ae51f | 1703 | |
24d2f903 | 1704 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1705 | struct request *rq = tags->static_rqs[i]; |
1706 | ||
1707 | if (!rq) | |
e9b267d9 | 1708 | continue; |
d6296d39 | 1709 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1710 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1711 | } |
320ae51f | 1712 | } |
320ae51f | 1713 | |
24d2f903 CH |
1714 | while (!list_empty(&tags->page_list)) { |
1715 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1716 | list_del_init(&page->lru); |
f75782e4 CM |
1717 | /* |
1718 | * Remove kmemleak object previously allocated in | |
1719 | * blk_mq_init_rq_map(). | |
1720 | */ | |
1721 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1722 | __free_pages(page, page->private); |
1723 | } | |
cc71a6f4 | 1724 | } |
320ae51f | 1725 | |
cc71a6f4 JA |
1726 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1727 | { | |
24d2f903 | 1728 | kfree(tags->rqs); |
cc71a6f4 | 1729 | tags->rqs = NULL; |
2af8cbe3 JA |
1730 | kfree(tags->static_rqs); |
1731 | tags->static_rqs = NULL; | |
320ae51f | 1732 | |
24d2f903 | 1733 | blk_mq_free_tags(tags); |
320ae51f JA |
1734 | } |
1735 | ||
cc71a6f4 JA |
1736 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1737 | unsigned int hctx_idx, | |
1738 | unsigned int nr_tags, | |
1739 | unsigned int reserved_tags) | |
320ae51f | 1740 | { |
24d2f903 | 1741 | struct blk_mq_tags *tags; |
59f082e4 | 1742 | int node; |
320ae51f | 1743 | |
59f082e4 SL |
1744 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1745 | if (node == NUMA_NO_NODE) | |
1746 | node = set->numa_node; | |
1747 | ||
1748 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1749 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1750 | if (!tags) |
1751 | return NULL; | |
320ae51f | 1752 | |
cc71a6f4 | 1753 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1754 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1755 | node); |
24d2f903 CH |
1756 | if (!tags->rqs) { |
1757 | blk_mq_free_tags(tags); | |
1758 | return NULL; | |
1759 | } | |
320ae51f | 1760 | |
2af8cbe3 JA |
1761 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1762 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1763 | node); |
2af8cbe3 JA |
1764 | if (!tags->static_rqs) { |
1765 | kfree(tags->rqs); | |
1766 | blk_mq_free_tags(tags); | |
1767 | return NULL; | |
1768 | } | |
1769 | ||
cc71a6f4 JA |
1770 | return tags; |
1771 | } | |
1772 | ||
1773 | static size_t order_to_size(unsigned int order) | |
1774 | { | |
1775 | return (size_t)PAGE_SIZE << order; | |
1776 | } | |
1777 | ||
1778 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1779 | unsigned int hctx_idx, unsigned int depth) | |
1780 | { | |
1781 | unsigned int i, j, entries_per_page, max_order = 4; | |
1782 | size_t rq_size, left; | |
59f082e4 SL |
1783 | int node; |
1784 | ||
1785 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1786 | if (node == NUMA_NO_NODE) | |
1787 | node = set->numa_node; | |
cc71a6f4 JA |
1788 | |
1789 | INIT_LIST_HEAD(&tags->page_list); | |
1790 | ||
320ae51f JA |
1791 | /* |
1792 | * rq_size is the size of the request plus driver payload, rounded | |
1793 | * to the cacheline size | |
1794 | */ | |
24d2f903 | 1795 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1796 | cache_line_size()); |
cc71a6f4 | 1797 | left = rq_size * depth; |
320ae51f | 1798 | |
cc71a6f4 | 1799 | for (i = 0; i < depth; ) { |
320ae51f JA |
1800 | int this_order = max_order; |
1801 | struct page *page; | |
1802 | int to_do; | |
1803 | void *p; | |
1804 | ||
b3a834b1 | 1805 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1806 | this_order--; |
1807 | ||
1808 | do { | |
59f082e4 | 1809 | page = alloc_pages_node(node, |
36e1f3d1 | 1810 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1811 | this_order); |
320ae51f JA |
1812 | if (page) |
1813 | break; | |
1814 | if (!this_order--) | |
1815 | break; | |
1816 | if (order_to_size(this_order) < rq_size) | |
1817 | break; | |
1818 | } while (1); | |
1819 | ||
1820 | if (!page) | |
24d2f903 | 1821 | goto fail; |
320ae51f JA |
1822 | |
1823 | page->private = this_order; | |
24d2f903 | 1824 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1825 | |
1826 | p = page_address(page); | |
f75782e4 CM |
1827 | /* |
1828 | * Allow kmemleak to scan these pages as they contain pointers | |
1829 | * to additional allocations like via ops->init_request(). | |
1830 | */ | |
36e1f3d1 | 1831 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1832 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1833 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1834 | left -= to_do * rq_size; |
1835 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1836 | struct request *rq = p; |
1837 | ||
1838 | tags->static_rqs[i] = rq; | |
24d2f903 | 1839 | if (set->ops->init_request) { |
d6296d39 | 1840 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1841 | node)) { |
2af8cbe3 | 1842 | tags->static_rqs[i] = NULL; |
24d2f903 | 1843 | goto fail; |
a5164405 | 1844 | } |
e9b267d9 CH |
1845 | } |
1846 | ||
320ae51f JA |
1847 | p += rq_size; |
1848 | i++; | |
1849 | } | |
1850 | } | |
cc71a6f4 | 1851 | return 0; |
320ae51f | 1852 | |
24d2f903 | 1853 | fail: |
cc71a6f4 JA |
1854 | blk_mq_free_rqs(set, tags, hctx_idx); |
1855 | return -ENOMEM; | |
320ae51f JA |
1856 | } |
1857 | ||
e57690fe JA |
1858 | /* |
1859 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1860 | * software queue to the hw queue dispatch list, and ensure that it | |
1861 | * gets run. | |
1862 | */ | |
9467f859 | 1863 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1864 | { |
9467f859 | 1865 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1866 | struct blk_mq_ctx *ctx; |
1867 | LIST_HEAD(tmp); | |
1868 | ||
9467f859 | 1869 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1870 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1871 | |
1872 | spin_lock(&ctx->lock); | |
1873 | if (!list_empty(&ctx->rq_list)) { | |
1874 | list_splice_init(&ctx->rq_list, &tmp); | |
1875 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1876 | } | |
1877 | spin_unlock(&ctx->lock); | |
1878 | ||
1879 | if (list_empty(&tmp)) | |
9467f859 | 1880 | return 0; |
484b4061 | 1881 | |
e57690fe JA |
1882 | spin_lock(&hctx->lock); |
1883 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1884 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1885 | |
1886 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1887 | return 0; |
484b4061 JA |
1888 | } |
1889 | ||
9467f859 | 1890 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1891 | { |
9467f859 TG |
1892 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1893 | &hctx->cpuhp_dead); | |
484b4061 JA |
1894 | } |
1895 | ||
c3b4afca | 1896 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1897 | static void blk_mq_exit_hctx(struct request_queue *q, |
1898 | struct blk_mq_tag_set *set, | |
1899 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1900 | { | |
9c1051aa OS |
1901 | blk_mq_debugfs_unregister_hctx(hctx); |
1902 | ||
08e98fc6 ML |
1903 | blk_mq_tag_idle(hctx); |
1904 | ||
f70ced09 | 1905 | if (set->ops->exit_request) |
d6296d39 | 1906 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1907 | |
93252632 OS |
1908 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1909 | ||
08e98fc6 ML |
1910 | if (set->ops->exit_hctx) |
1911 | set->ops->exit_hctx(hctx, hctx_idx); | |
1912 | ||
6a83e74d | 1913 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1914 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1915 | |
9467f859 | 1916 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1917 | blk_free_flush_queue(hctx->fq); |
88459642 | 1918 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1919 | } |
1920 | ||
624dbe47 ML |
1921 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1922 | struct blk_mq_tag_set *set, int nr_queue) | |
1923 | { | |
1924 | struct blk_mq_hw_ctx *hctx; | |
1925 | unsigned int i; | |
1926 | ||
1927 | queue_for_each_hw_ctx(q, hctx, i) { | |
1928 | if (i == nr_queue) | |
1929 | break; | |
08e98fc6 | 1930 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1931 | } |
624dbe47 ML |
1932 | } |
1933 | ||
08e98fc6 ML |
1934 | static int blk_mq_init_hctx(struct request_queue *q, |
1935 | struct blk_mq_tag_set *set, | |
1936 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1937 | { |
08e98fc6 ML |
1938 | int node; |
1939 | ||
1940 | node = hctx->numa_node; | |
1941 | if (node == NUMA_NO_NODE) | |
1942 | node = hctx->numa_node = set->numa_node; | |
1943 | ||
9f993737 | 1944 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1945 | spin_lock_init(&hctx->lock); |
1946 | INIT_LIST_HEAD(&hctx->dispatch); | |
1947 | hctx->queue = q; | |
2404e607 | 1948 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1949 | |
9467f859 | 1950 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1951 | |
1952 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1953 | |
1954 | /* | |
08e98fc6 ML |
1955 | * Allocate space for all possible cpus to avoid allocation at |
1956 | * runtime | |
320ae51f | 1957 | */ |
08e98fc6 ML |
1958 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1959 | GFP_KERNEL, node); | |
1960 | if (!hctx->ctxs) | |
1961 | goto unregister_cpu_notifier; | |
320ae51f | 1962 | |
88459642 OS |
1963 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1964 | node)) | |
08e98fc6 | 1965 | goto free_ctxs; |
320ae51f | 1966 | |
08e98fc6 | 1967 | hctx->nr_ctx = 0; |
320ae51f | 1968 | |
08e98fc6 ML |
1969 | if (set->ops->init_hctx && |
1970 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1971 | goto free_bitmap; | |
320ae51f | 1972 | |
93252632 OS |
1973 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
1974 | goto exit_hctx; | |
1975 | ||
f70ced09 ML |
1976 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1977 | if (!hctx->fq) | |
93252632 | 1978 | goto sched_exit_hctx; |
320ae51f | 1979 | |
f70ced09 | 1980 | if (set->ops->init_request && |
d6296d39 CH |
1981 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
1982 | node)) | |
f70ced09 | 1983 | goto free_fq; |
320ae51f | 1984 | |
6a83e74d | 1985 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1986 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1987 | |
9c1051aa OS |
1988 | blk_mq_debugfs_register_hctx(q, hctx); |
1989 | ||
08e98fc6 | 1990 | return 0; |
320ae51f | 1991 | |
f70ced09 ML |
1992 | free_fq: |
1993 | kfree(hctx->fq); | |
93252632 OS |
1994 | sched_exit_hctx: |
1995 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
1996 | exit_hctx: |
1997 | if (set->ops->exit_hctx) | |
1998 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1999 | free_bitmap: |
88459642 | 2000 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2001 | free_ctxs: |
2002 | kfree(hctx->ctxs); | |
2003 | unregister_cpu_notifier: | |
9467f859 | 2004 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2005 | return -1; |
2006 | } | |
320ae51f | 2007 | |
320ae51f JA |
2008 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2009 | unsigned int nr_hw_queues) | |
2010 | { | |
2011 | unsigned int i; | |
2012 | ||
2013 | for_each_possible_cpu(i) { | |
2014 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2015 | struct blk_mq_hw_ctx *hctx; | |
2016 | ||
320ae51f JA |
2017 | __ctx->cpu = i; |
2018 | spin_lock_init(&__ctx->lock); | |
2019 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2020 | __ctx->queue = q; | |
2021 | ||
4b855ad3 CH |
2022 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2023 | if (!cpu_present(i)) | |
320ae51f JA |
2024 | continue; |
2025 | ||
7d7e0f90 | 2026 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2027 | |
320ae51f JA |
2028 | /* |
2029 | * Set local node, IFF we have more than one hw queue. If | |
2030 | * not, we remain on the home node of the device | |
2031 | */ | |
2032 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2033 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2034 | } |
2035 | } | |
2036 | ||
cc71a6f4 JA |
2037 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2038 | { | |
2039 | int ret = 0; | |
2040 | ||
2041 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2042 | set->queue_depth, set->reserved_tags); | |
2043 | if (!set->tags[hctx_idx]) | |
2044 | return false; | |
2045 | ||
2046 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2047 | set->queue_depth); | |
2048 | if (!ret) | |
2049 | return true; | |
2050 | ||
2051 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2052 | set->tags[hctx_idx] = NULL; | |
2053 | return false; | |
2054 | } | |
2055 | ||
2056 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2057 | unsigned int hctx_idx) | |
2058 | { | |
bd166ef1 JA |
2059 | if (set->tags[hctx_idx]) { |
2060 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2061 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2062 | set->tags[hctx_idx] = NULL; | |
2063 | } | |
cc71a6f4 JA |
2064 | } |
2065 | ||
4b855ad3 | 2066 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2067 | { |
d1b1cea1 | 2068 | unsigned int i, hctx_idx; |
320ae51f JA |
2069 | struct blk_mq_hw_ctx *hctx; |
2070 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2071 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2072 | |
60de074b AM |
2073 | /* |
2074 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2075 | */ | |
2076 | mutex_lock(&q->sysfs_lock); | |
2077 | ||
320ae51f | 2078 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2079 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2080 | hctx->nr_ctx = 0; |
2081 | } | |
2082 | ||
2083 | /* | |
4b855ad3 CH |
2084 | * Map software to hardware queues. |
2085 | * | |
2086 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2087 | */ |
4b855ad3 | 2088 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2089 | hctx_idx = q->mq_map[i]; |
2090 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2091 | if (!set->tags[hctx_idx] && |
2092 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2093 | /* |
2094 | * If tags initialization fail for some hctx, | |
2095 | * that hctx won't be brought online. In this | |
2096 | * case, remap the current ctx to hctx[0] which | |
2097 | * is guaranteed to always have tags allocated | |
2098 | */ | |
cc71a6f4 | 2099 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2100 | } |
2101 | ||
897bb0c7 | 2102 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2103 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2104 | |
e4043dcf | 2105 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2106 | ctx->index_hw = hctx->nr_ctx; |
2107 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2108 | } | |
506e931f | 2109 | |
60de074b AM |
2110 | mutex_unlock(&q->sysfs_lock); |
2111 | ||
506e931f | 2112 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2113 | /* |
a68aafa5 JA |
2114 | * If no software queues are mapped to this hardware queue, |
2115 | * disable it and free the request entries. | |
484b4061 JA |
2116 | */ |
2117 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2118 | /* Never unmap queue 0. We need it as a |
2119 | * fallback in case of a new remap fails | |
2120 | * allocation | |
2121 | */ | |
cc71a6f4 JA |
2122 | if (i && set->tags[i]) |
2123 | blk_mq_free_map_and_requests(set, i); | |
2124 | ||
2a34c087 | 2125 | hctx->tags = NULL; |
484b4061 JA |
2126 | continue; |
2127 | } | |
2128 | ||
2a34c087 ML |
2129 | hctx->tags = set->tags[i]; |
2130 | WARN_ON(!hctx->tags); | |
2131 | ||
889fa31f CY |
2132 | /* |
2133 | * Set the map size to the number of mapped software queues. | |
2134 | * This is more accurate and more efficient than looping | |
2135 | * over all possibly mapped software queues. | |
2136 | */ | |
88459642 | 2137 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2138 | |
484b4061 JA |
2139 | /* |
2140 | * Initialize batch roundrobin counts | |
2141 | */ | |
506e931f JA |
2142 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2143 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2144 | } | |
320ae51f JA |
2145 | } |
2146 | ||
8e8320c9 JA |
2147 | /* |
2148 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2149 | * the queue isn't live yet. | |
2150 | */ | |
2404e607 | 2151 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2152 | { |
2153 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2154 | int i; |
2155 | ||
2404e607 | 2156 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2157 | if (shared) { |
2158 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2159 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2160 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2161 | } else { |
2162 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2163 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2164 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2165 | } |
2404e607 JM |
2166 | } |
2167 | } | |
2168 | ||
8e8320c9 JA |
2169 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2170 | bool shared) | |
2404e607 JM |
2171 | { |
2172 | struct request_queue *q; | |
0d2602ca | 2173 | |
705cda97 BVA |
2174 | lockdep_assert_held(&set->tag_list_lock); |
2175 | ||
0d2602ca JA |
2176 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2177 | blk_mq_freeze_queue(q); | |
2404e607 | 2178 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2179 | blk_mq_unfreeze_queue(q); |
2180 | } | |
2181 | } | |
2182 | ||
2183 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2184 | { | |
2185 | struct blk_mq_tag_set *set = q->tag_set; | |
2186 | ||
0d2602ca | 2187 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2188 | list_del_rcu(&q->tag_set_list); |
2189 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2190 | if (list_is_singular(&set->tag_list)) { |
2191 | /* just transitioned to unshared */ | |
2192 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2193 | /* update existing queue */ | |
2194 | blk_mq_update_tag_set_depth(set, false); | |
2195 | } | |
0d2602ca | 2196 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2197 | |
2198 | synchronize_rcu(); | |
0d2602ca JA |
2199 | } |
2200 | ||
2201 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2202 | struct request_queue *q) | |
2203 | { | |
2204 | q->tag_set = set; | |
2205 | ||
2206 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2207 | |
2208 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2209 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2210 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2211 | /* update existing queue */ | |
2212 | blk_mq_update_tag_set_depth(set, true); | |
2213 | } | |
2214 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2215 | queue_set_hctx_shared(q, true); | |
705cda97 | 2216 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2217 | |
0d2602ca JA |
2218 | mutex_unlock(&set->tag_list_lock); |
2219 | } | |
2220 | ||
e09aae7e ML |
2221 | /* |
2222 | * It is the actual release handler for mq, but we do it from | |
2223 | * request queue's release handler for avoiding use-after-free | |
2224 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2225 | * but we can't group ctx/kctx kobj without it. | |
2226 | */ | |
2227 | void blk_mq_release(struct request_queue *q) | |
2228 | { | |
2229 | struct blk_mq_hw_ctx *hctx; | |
2230 | unsigned int i; | |
2231 | ||
2232 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2233 | queue_for_each_hw_ctx(q, hctx, i) { |
2234 | if (!hctx) | |
2235 | continue; | |
6c8b232e | 2236 | kobject_put(&hctx->kobj); |
c3b4afca | 2237 | } |
e09aae7e | 2238 | |
a723bab3 AM |
2239 | q->mq_map = NULL; |
2240 | ||
e09aae7e ML |
2241 | kfree(q->queue_hw_ctx); |
2242 | ||
7ea5fe31 ML |
2243 | /* |
2244 | * release .mq_kobj and sw queue's kobject now because | |
2245 | * both share lifetime with request queue. | |
2246 | */ | |
2247 | blk_mq_sysfs_deinit(q); | |
2248 | ||
e09aae7e ML |
2249 | free_percpu(q->queue_ctx); |
2250 | } | |
2251 | ||
24d2f903 | 2252 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2253 | { |
2254 | struct request_queue *uninit_q, *q; | |
2255 | ||
2256 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2257 | if (!uninit_q) | |
2258 | return ERR_PTR(-ENOMEM); | |
2259 | ||
2260 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2261 | if (IS_ERR(q)) | |
2262 | blk_cleanup_queue(uninit_q); | |
2263 | ||
2264 | return q; | |
2265 | } | |
2266 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2267 | ||
07319678 BVA |
2268 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2269 | { | |
2270 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2271 | ||
2272 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2273 | __alignof__(struct blk_mq_hw_ctx)) != | |
2274 | sizeof(struct blk_mq_hw_ctx)); | |
2275 | ||
2276 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2277 | hw_ctx_size += sizeof(struct srcu_struct); | |
2278 | ||
2279 | return hw_ctx_size; | |
2280 | } | |
2281 | ||
868f2f0b KB |
2282 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2283 | struct request_queue *q) | |
320ae51f | 2284 | { |
868f2f0b KB |
2285 | int i, j; |
2286 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2287 | |
868f2f0b | 2288 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2289 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2290 | int node; |
f14bbe77 | 2291 | |
868f2f0b KB |
2292 | if (hctxs[i]) |
2293 | continue; | |
2294 | ||
2295 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2296 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2297 | GFP_KERNEL, node); |
320ae51f | 2298 | if (!hctxs[i]) |
868f2f0b | 2299 | break; |
320ae51f | 2300 | |
a86073e4 | 2301 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2302 | node)) { |
2303 | kfree(hctxs[i]); | |
2304 | hctxs[i] = NULL; | |
2305 | break; | |
2306 | } | |
e4043dcf | 2307 | |
0d2602ca | 2308 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2309 | hctxs[i]->numa_node = node; |
320ae51f | 2310 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2311 | |
2312 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2313 | free_cpumask_var(hctxs[i]->cpumask); | |
2314 | kfree(hctxs[i]); | |
2315 | hctxs[i] = NULL; | |
2316 | break; | |
2317 | } | |
2318 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2319 | } |
868f2f0b KB |
2320 | for (j = i; j < q->nr_hw_queues; j++) { |
2321 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2322 | ||
2323 | if (hctx) { | |
cc71a6f4 JA |
2324 | if (hctx->tags) |
2325 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2326 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2327 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2328 | hctxs[j] = NULL; |
2329 | ||
2330 | } | |
2331 | } | |
2332 | q->nr_hw_queues = i; | |
2333 | blk_mq_sysfs_register(q); | |
2334 | } | |
2335 | ||
2336 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2337 | struct request_queue *q) | |
2338 | { | |
66841672 ML |
2339 | /* mark the queue as mq asap */ |
2340 | q->mq_ops = set->ops; | |
2341 | ||
34dbad5d | 2342 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2343 | blk_mq_poll_stats_bkt, |
2344 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2345 | if (!q->poll_cb) |
2346 | goto err_exit; | |
2347 | ||
868f2f0b KB |
2348 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2349 | if (!q->queue_ctx) | |
c7de5726 | 2350 | goto err_exit; |
868f2f0b | 2351 | |
737f98cf ML |
2352 | /* init q->mq_kobj and sw queues' kobjects */ |
2353 | blk_mq_sysfs_init(q); | |
2354 | ||
868f2f0b KB |
2355 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2356 | GFP_KERNEL, set->numa_node); | |
2357 | if (!q->queue_hw_ctx) | |
2358 | goto err_percpu; | |
2359 | ||
bdd17e75 | 2360 | q->mq_map = set->mq_map; |
868f2f0b KB |
2361 | |
2362 | blk_mq_realloc_hw_ctxs(set, q); | |
2363 | if (!q->nr_hw_queues) | |
2364 | goto err_hctxs; | |
320ae51f | 2365 | |
287922eb | 2366 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2367 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2368 | |
2369 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2370 | |
94eddfbe | 2371 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2372 | |
05f1dd53 JA |
2373 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2374 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2375 | ||
1be036e9 CH |
2376 | q->sg_reserved_size = INT_MAX; |
2377 | ||
2849450a | 2378 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2379 | INIT_LIST_HEAD(&q->requeue_list); |
2380 | spin_lock_init(&q->requeue_lock); | |
2381 | ||
254d259d | 2382 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2383 | |
eba71768 JA |
2384 | /* |
2385 | * Do this after blk_queue_make_request() overrides it... | |
2386 | */ | |
2387 | q->nr_requests = set->queue_depth; | |
2388 | ||
64f1c21e JA |
2389 | /* |
2390 | * Default to classic polling | |
2391 | */ | |
2392 | q->poll_nsec = -1; | |
2393 | ||
24d2f903 CH |
2394 | if (set->ops->complete) |
2395 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2396 | |
24d2f903 | 2397 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2398 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2399 | blk_mq_map_swqueue(q); |
4593fdbe | 2400 | |
d3484991 JA |
2401 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2402 | int ret; | |
2403 | ||
2404 | ret = blk_mq_sched_init(q); | |
2405 | if (ret) | |
2406 | return ERR_PTR(ret); | |
2407 | } | |
2408 | ||
320ae51f | 2409 | return q; |
18741986 | 2410 | |
320ae51f | 2411 | err_hctxs: |
868f2f0b | 2412 | kfree(q->queue_hw_ctx); |
320ae51f | 2413 | err_percpu: |
868f2f0b | 2414 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2415 | err_exit: |
2416 | q->mq_ops = NULL; | |
320ae51f JA |
2417 | return ERR_PTR(-ENOMEM); |
2418 | } | |
b62c21b7 | 2419 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2420 | |
2421 | void blk_mq_free_queue(struct request_queue *q) | |
2422 | { | |
624dbe47 | 2423 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2424 | |
0d2602ca | 2425 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2426 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2427 | } |
320ae51f JA |
2428 | |
2429 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2430 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2431 | { |
4ecd4fef | 2432 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2433 | |
9c1051aa | 2434 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2435 | blk_mq_sysfs_unregister(q); |
2436 | ||
320ae51f JA |
2437 | /* |
2438 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2439 | * we should change hctx numa_node according to new topology (this | |
2440 | * involves free and re-allocate memory, worthy doing?) | |
2441 | */ | |
2442 | ||
4b855ad3 | 2443 | blk_mq_map_swqueue(q); |
320ae51f | 2444 | |
67aec14c | 2445 | blk_mq_sysfs_register(q); |
9c1051aa | 2446 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2447 | } |
2448 | ||
a5164405 JA |
2449 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2450 | { | |
2451 | int i; | |
2452 | ||
cc71a6f4 JA |
2453 | for (i = 0; i < set->nr_hw_queues; i++) |
2454 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2455 | goto out_unwind; |
a5164405 JA |
2456 | |
2457 | return 0; | |
2458 | ||
2459 | out_unwind: | |
2460 | while (--i >= 0) | |
cc71a6f4 | 2461 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2462 | |
a5164405 JA |
2463 | return -ENOMEM; |
2464 | } | |
2465 | ||
2466 | /* | |
2467 | * Allocate the request maps associated with this tag_set. Note that this | |
2468 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2469 | * will be updated to reflect the allocated depth. | |
2470 | */ | |
2471 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2472 | { | |
2473 | unsigned int depth; | |
2474 | int err; | |
2475 | ||
2476 | depth = set->queue_depth; | |
2477 | do { | |
2478 | err = __blk_mq_alloc_rq_maps(set); | |
2479 | if (!err) | |
2480 | break; | |
2481 | ||
2482 | set->queue_depth >>= 1; | |
2483 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2484 | err = -ENOMEM; | |
2485 | break; | |
2486 | } | |
2487 | } while (set->queue_depth); | |
2488 | ||
2489 | if (!set->queue_depth || err) { | |
2490 | pr_err("blk-mq: failed to allocate request map\n"); | |
2491 | return -ENOMEM; | |
2492 | } | |
2493 | ||
2494 | if (depth != set->queue_depth) | |
2495 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2496 | depth, set->queue_depth); | |
2497 | ||
2498 | return 0; | |
2499 | } | |
2500 | ||
ebe8bddb OS |
2501 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2502 | { | |
2503 | if (set->ops->map_queues) | |
2504 | return set->ops->map_queues(set); | |
2505 | else | |
2506 | return blk_mq_map_queues(set); | |
2507 | } | |
2508 | ||
a4391c64 JA |
2509 | /* |
2510 | * Alloc a tag set to be associated with one or more request queues. | |
2511 | * May fail with EINVAL for various error conditions. May adjust the | |
2512 | * requested depth down, if if it too large. In that case, the set | |
2513 | * value will be stored in set->queue_depth. | |
2514 | */ | |
24d2f903 CH |
2515 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2516 | { | |
da695ba2 CH |
2517 | int ret; |
2518 | ||
205fb5f5 BVA |
2519 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2520 | ||
24d2f903 CH |
2521 | if (!set->nr_hw_queues) |
2522 | return -EINVAL; | |
a4391c64 | 2523 | if (!set->queue_depth) |
24d2f903 CH |
2524 | return -EINVAL; |
2525 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2526 | return -EINVAL; | |
2527 | ||
7d7e0f90 | 2528 | if (!set->ops->queue_rq) |
24d2f903 CH |
2529 | return -EINVAL; |
2530 | ||
a4391c64 JA |
2531 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2532 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2533 | BLK_MQ_MAX_DEPTH); | |
2534 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2535 | } | |
24d2f903 | 2536 | |
6637fadf SL |
2537 | /* |
2538 | * If a crashdump is active, then we are potentially in a very | |
2539 | * memory constrained environment. Limit us to 1 queue and | |
2540 | * 64 tags to prevent using too much memory. | |
2541 | */ | |
2542 | if (is_kdump_kernel()) { | |
2543 | set->nr_hw_queues = 1; | |
2544 | set->queue_depth = min(64U, set->queue_depth); | |
2545 | } | |
868f2f0b KB |
2546 | /* |
2547 | * There is no use for more h/w queues than cpus. | |
2548 | */ | |
2549 | if (set->nr_hw_queues > nr_cpu_ids) | |
2550 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2551 | |
868f2f0b | 2552 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2553 | GFP_KERNEL, set->numa_node); |
2554 | if (!set->tags) | |
a5164405 | 2555 | return -ENOMEM; |
24d2f903 | 2556 | |
da695ba2 CH |
2557 | ret = -ENOMEM; |
2558 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2559 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2560 | if (!set->mq_map) |
2561 | goto out_free_tags; | |
2562 | ||
ebe8bddb | 2563 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2564 | if (ret) |
2565 | goto out_free_mq_map; | |
2566 | ||
2567 | ret = blk_mq_alloc_rq_maps(set); | |
2568 | if (ret) | |
bdd17e75 | 2569 | goto out_free_mq_map; |
24d2f903 | 2570 | |
0d2602ca JA |
2571 | mutex_init(&set->tag_list_lock); |
2572 | INIT_LIST_HEAD(&set->tag_list); | |
2573 | ||
24d2f903 | 2574 | return 0; |
bdd17e75 CH |
2575 | |
2576 | out_free_mq_map: | |
2577 | kfree(set->mq_map); | |
2578 | set->mq_map = NULL; | |
2579 | out_free_tags: | |
5676e7b6 RE |
2580 | kfree(set->tags); |
2581 | set->tags = NULL; | |
da695ba2 | 2582 | return ret; |
24d2f903 CH |
2583 | } |
2584 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2585 | ||
2586 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2587 | { | |
2588 | int i; | |
2589 | ||
cc71a6f4 JA |
2590 | for (i = 0; i < nr_cpu_ids; i++) |
2591 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2592 | |
bdd17e75 CH |
2593 | kfree(set->mq_map); |
2594 | set->mq_map = NULL; | |
2595 | ||
981bd189 | 2596 | kfree(set->tags); |
5676e7b6 | 2597 | set->tags = NULL; |
24d2f903 CH |
2598 | } |
2599 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2600 | ||
e3a2b3f9 JA |
2601 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2602 | { | |
2603 | struct blk_mq_tag_set *set = q->tag_set; | |
2604 | struct blk_mq_hw_ctx *hctx; | |
2605 | int i, ret; | |
2606 | ||
bd166ef1 | 2607 | if (!set) |
e3a2b3f9 JA |
2608 | return -EINVAL; |
2609 | ||
70f36b60 | 2610 | blk_mq_freeze_queue(q); |
70f36b60 | 2611 | |
e3a2b3f9 JA |
2612 | ret = 0; |
2613 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2614 | if (!hctx->tags) |
2615 | continue; | |
bd166ef1 JA |
2616 | /* |
2617 | * If we're using an MQ scheduler, just update the scheduler | |
2618 | * queue depth. This is similar to what the old code would do. | |
2619 | */ | |
70f36b60 JA |
2620 | if (!hctx->sched_tags) { |
2621 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2622 | min(nr, set->queue_depth), | |
2623 | false); | |
2624 | } else { | |
2625 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2626 | nr, true); | |
2627 | } | |
e3a2b3f9 JA |
2628 | if (ret) |
2629 | break; | |
2630 | } | |
2631 | ||
2632 | if (!ret) | |
2633 | q->nr_requests = nr; | |
2634 | ||
70f36b60 | 2635 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2636 | |
e3a2b3f9 JA |
2637 | return ret; |
2638 | } | |
2639 | ||
e4dc2b32 KB |
2640 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2641 | int nr_hw_queues) | |
868f2f0b KB |
2642 | { |
2643 | struct request_queue *q; | |
2644 | ||
705cda97 BVA |
2645 | lockdep_assert_held(&set->tag_list_lock); |
2646 | ||
868f2f0b KB |
2647 | if (nr_hw_queues > nr_cpu_ids) |
2648 | nr_hw_queues = nr_cpu_ids; | |
2649 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2650 | return; | |
2651 | ||
2652 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2653 | blk_mq_freeze_queue(q); | |
2654 | ||
2655 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2656 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2657 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2658 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2659 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2660 | } |
2661 | ||
2662 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2663 | blk_mq_unfreeze_queue(q); | |
2664 | } | |
e4dc2b32 KB |
2665 | |
2666 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2667 | { | |
2668 | mutex_lock(&set->tag_list_lock); | |
2669 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2670 | mutex_unlock(&set->tag_list_lock); | |
2671 | } | |
868f2f0b KB |
2672 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2673 | ||
34dbad5d OS |
2674 | /* Enable polling stats and return whether they were already enabled. */ |
2675 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2676 | { | |
2677 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2678 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2679 | return true; | |
2680 | blk_stat_add_callback(q, q->poll_cb); | |
2681 | return false; | |
2682 | } | |
2683 | ||
2684 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2685 | { | |
2686 | /* | |
2687 | * We don't arm the callback if polling stats are not enabled or the | |
2688 | * callback is already active. | |
2689 | */ | |
2690 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2691 | blk_stat_is_active(q->poll_cb)) | |
2692 | return; | |
2693 | ||
2694 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2695 | } | |
2696 | ||
2697 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2698 | { | |
2699 | struct request_queue *q = cb->data; | |
720b8ccc | 2700 | int bucket; |
34dbad5d | 2701 | |
720b8ccc SB |
2702 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2703 | if (cb->stat[bucket].nr_samples) | |
2704 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2705 | } | |
34dbad5d OS |
2706 | } |
2707 | ||
64f1c21e JA |
2708 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2709 | struct blk_mq_hw_ctx *hctx, | |
2710 | struct request *rq) | |
2711 | { | |
64f1c21e | 2712 | unsigned long ret = 0; |
720b8ccc | 2713 | int bucket; |
64f1c21e JA |
2714 | |
2715 | /* | |
2716 | * If stats collection isn't on, don't sleep but turn it on for | |
2717 | * future users | |
2718 | */ | |
34dbad5d | 2719 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2720 | return 0; |
2721 | ||
64f1c21e JA |
2722 | /* |
2723 | * As an optimistic guess, use half of the mean service time | |
2724 | * for this type of request. We can (and should) make this smarter. | |
2725 | * For instance, if the completion latencies are tight, we can | |
2726 | * get closer than just half the mean. This is especially | |
2727 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2728 | * than ~10 usec. We do use the stats for the relevant IO size |
2729 | * if available which does lead to better estimates. | |
64f1c21e | 2730 | */ |
720b8ccc SB |
2731 | bucket = blk_mq_poll_stats_bkt(rq); |
2732 | if (bucket < 0) | |
2733 | return ret; | |
2734 | ||
2735 | if (q->poll_stat[bucket].nr_samples) | |
2736 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2737 | |
2738 | return ret; | |
2739 | } | |
2740 | ||
06426adf | 2741 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2742 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2743 | struct request *rq) |
2744 | { | |
2745 | struct hrtimer_sleeper hs; | |
2746 | enum hrtimer_mode mode; | |
64f1c21e | 2747 | unsigned int nsecs; |
06426adf JA |
2748 | ktime_t kt; |
2749 | ||
64f1c21e JA |
2750 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2751 | return false; | |
2752 | ||
2753 | /* | |
2754 | * poll_nsec can be: | |
2755 | * | |
2756 | * -1: don't ever hybrid sleep | |
2757 | * 0: use half of prev avg | |
2758 | * >0: use this specific value | |
2759 | */ | |
2760 | if (q->poll_nsec == -1) | |
2761 | return false; | |
2762 | else if (q->poll_nsec > 0) | |
2763 | nsecs = q->poll_nsec; | |
2764 | else | |
2765 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2766 | ||
2767 | if (!nsecs) | |
06426adf JA |
2768 | return false; |
2769 | ||
2770 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2771 | ||
2772 | /* | |
2773 | * This will be replaced with the stats tracking code, using | |
2774 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2775 | */ | |
8b0e1953 | 2776 | kt = nsecs; |
06426adf JA |
2777 | |
2778 | mode = HRTIMER_MODE_REL; | |
2779 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2780 | hrtimer_set_expires(&hs.timer, kt); | |
2781 | ||
2782 | hrtimer_init_sleeper(&hs, current); | |
2783 | do { | |
2784 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2785 | break; | |
2786 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2787 | hrtimer_start_expires(&hs.timer, mode); | |
2788 | if (hs.task) | |
2789 | io_schedule(); | |
2790 | hrtimer_cancel(&hs.timer); | |
2791 | mode = HRTIMER_MODE_ABS; | |
2792 | } while (hs.task && !signal_pending(current)); | |
2793 | ||
2794 | __set_current_state(TASK_RUNNING); | |
2795 | destroy_hrtimer_on_stack(&hs.timer); | |
2796 | return true; | |
2797 | } | |
2798 | ||
bbd7bb70 JA |
2799 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2800 | { | |
2801 | struct request_queue *q = hctx->queue; | |
2802 | long state; | |
2803 | ||
06426adf JA |
2804 | /* |
2805 | * If we sleep, have the caller restart the poll loop to reset | |
2806 | * the state. Like for the other success return cases, the | |
2807 | * caller is responsible for checking if the IO completed. If | |
2808 | * the IO isn't complete, we'll get called again and will go | |
2809 | * straight to the busy poll loop. | |
2810 | */ | |
64f1c21e | 2811 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2812 | return true; |
2813 | ||
bbd7bb70 JA |
2814 | hctx->poll_considered++; |
2815 | ||
2816 | state = current->state; | |
2817 | while (!need_resched()) { | |
2818 | int ret; | |
2819 | ||
2820 | hctx->poll_invoked++; | |
2821 | ||
2822 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2823 | if (ret > 0) { | |
2824 | hctx->poll_success++; | |
2825 | set_current_state(TASK_RUNNING); | |
2826 | return true; | |
2827 | } | |
2828 | ||
2829 | if (signal_pending_state(state, current)) | |
2830 | set_current_state(TASK_RUNNING); | |
2831 | ||
2832 | if (current->state == TASK_RUNNING) | |
2833 | return true; | |
2834 | if (ret < 0) | |
2835 | break; | |
2836 | cpu_relax(); | |
2837 | } | |
2838 | ||
2839 | return false; | |
2840 | } | |
2841 | ||
2842 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2843 | { | |
2844 | struct blk_mq_hw_ctx *hctx; | |
2845 | struct blk_plug *plug; | |
2846 | struct request *rq; | |
2847 | ||
2848 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2849 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2850 | return false; | |
2851 | ||
2852 | plug = current->plug; | |
2853 | if (plug) | |
2854 | blk_flush_plug_list(plug, false); | |
2855 | ||
2856 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2857 | if (!blk_qc_t_is_internal(cookie)) |
2858 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2859 | else { |
bd166ef1 | 2860 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
2861 | /* |
2862 | * With scheduling, if the request has completed, we'll | |
2863 | * get a NULL return here, as we clear the sched tag when | |
2864 | * that happens. The request still remains valid, like always, | |
2865 | * so we should be safe with just the NULL check. | |
2866 | */ | |
2867 | if (!rq) | |
2868 | return false; | |
2869 | } | |
bbd7bb70 JA |
2870 | |
2871 | return __blk_mq_poll(hctx, rq); | |
2872 | } | |
2873 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2874 | ||
320ae51f JA |
2875 | static int __init blk_mq_init(void) |
2876 | { | |
9467f859 TG |
2877 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2878 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
2879 | return 0; |
2880 | } | |
2881 | subsys_initcall(blk_mq_init); |