]>
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 JA |
39 | |
40 | static DEFINE_MUTEX(all_q_mutex); | |
41 | static LIST_HEAD(all_q_list); | |
42 | ||
34dbad5d OS |
43 | static void blk_mq_poll_stats_start(struct request_queue *q); |
44 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
45 | ||
720b8ccc SB |
46 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
47 | { | |
48 | int ddir, bytes, bucket; | |
49 | ||
99c749a4 | 50 | ddir = rq_data_dir(rq); |
720b8ccc SB |
51 | bytes = blk_rq_bytes(rq); |
52 | ||
53 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
54 | ||
55 | if (bucket < 0) | |
56 | return -1; | |
57 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
58 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
59 | ||
60 | return bucket; | |
61 | } | |
62 | ||
320ae51f JA |
63 | /* |
64 | * Check if any of the ctx's have pending work in this hardware queue | |
65 | */ | |
50e1dab8 | 66 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 67 | { |
bd166ef1 JA |
68 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
69 | !list_empty_careful(&hctx->dispatch) || | |
70 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
71 | } |
72 | ||
320ae51f JA |
73 | /* |
74 | * Mark this ctx as having pending work in this hardware queue | |
75 | */ | |
76 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
77 | struct blk_mq_ctx *ctx) | |
78 | { | |
88459642 OS |
79 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
80 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
81 | } |
82 | ||
83 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
84 | struct blk_mq_ctx *ctx) | |
85 | { | |
88459642 | 86 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
87 | } |
88 | ||
1671d522 | 89 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 90 | { |
4ecd4fef | 91 | int freeze_depth; |
cddd5d17 | 92 | |
4ecd4fef CH |
93 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
94 | if (freeze_depth == 1) { | |
3ef28e83 | 95 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 96 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 97 | } |
f3af020b | 98 | } |
1671d522 | 99 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 100 | |
6bae363e | 101 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 102 | { |
3ef28e83 | 103 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 104 | } |
6bae363e | 105 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 106 | |
f91328c4 KB |
107 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
108 | unsigned long timeout) | |
109 | { | |
110 | return wait_event_timeout(q->mq_freeze_wq, | |
111 | percpu_ref_is_zero(&q->q_usage_counter), | |
112 | timeout); | |
113 | } | |
114 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 115 | |
f3af020b TH |
116 | /* |
117 | * Guarantee no request is in use, so we can change any data structure of | |
118 | * the queue afterward. | |
119 | */ | |
3ef28e83 | 120 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 121 | { |
3ef28e83 DW |
122 | /* |
123 | * In the !blk_mq case we are only calling this to kill the | |
124 | * q_usage_counter, otherwise this increases the freeze depth | |
125 | * and waits for it to return to zero. For this reason there is | |
126 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
127 | * exported to drivers as the only user for unfreeze is blk_mq. | |
128 | */ | |
1671d522 | 129 | blk_freeze_queue_start(q); |
f3af020b TH |
130 | blk_mq_freeze_queue_wait(q); |
131 | } | |
3ef28e83 DW |
132 | |
133 | void blk_mq_freeze_queue(struct request_queue *q) | |
134 | { | |
135 | /* | |
136 | * ...just an alias to keep freeze and unfreeze actions balanced | |
137 | * in the blk_mq_* namespace | |
138 | */ | |
139 | blk_freeze_queue(q); | |
140 | } | |
c761d96b | 141 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 142 | |
b4c6a028 | 143 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 144 | { |
4ecd4fef | 145 | int freeze_depth; |
320ae51f | 146 | |
4ecd4fef CH |
147 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
148 | WARN_ON_ONCE(freeze_depth < 0); | |
149 | if (!freeze_depth) { | |
3ef28e83 | 150 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 151 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 152 | } |
320ae51f | 153 | } |
b4c6a028 | 154 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 155 | |
6a83e74d | 156 | /** |
69e07c4a | 157 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
158 | * @q: request queue. |
159 | * | |
160 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
161 | * callback function is invoked. Once this function is returned, we make |
162 | * sure no dispatch can happen until the queue is unquiesced via | |
163 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
164 | */ |
165 | void blk_mq_quiesce_queue(struct request_queue *q) | |
166 | { | |
167 | struct blk_mq_hw_ctx *hctx; | |
168 | unsigned int i; | |
169 | bool rcu = false; | |
170 | ||
1d9e9bc6 | 171 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 172 | |
6a83e74d BVA |
173 | queue_for_each_hw_ctx(q, hctx, i) { |
174 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 175 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
176 | else |
177 | rcu = true; | |
178 | } | |
179 | if (rcu) | |
180 | synchronize_rcu(); | |
181 | } | |
182 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
183 | ||
e4e73913 ML |
184 | /* |
185 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
186 | * @q: request queue. | |
187 | * | |
188 | * This function recovers queue into the state before quiescing | |
189 | * which is done by blk_mq_quiesce_queue. | |
190 | */ | |
191 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
192 | { | |
f4560ffe ML |
193 | spin_lock_irq(q->queue_lock); |
194 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
195 | spin_unlock_irq(q->queue_lock); | |
196 | ||
1d9e9bc6 ML |
197 | /* dispatch requests which are inserted during quiescing */ |
198 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
199 | } |
200 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
201 | ||
aed3ea94 JA |
202 | void blk_mq_wake_waiters(struct request_queue *q) |
203 | { | |
204 | struct blk_mq_hw_ctx *hctx; | |
205 | unsigned int i; | |
206 | ||
207 | queue_for_each_hw_ctx(q, hctx, i) | |
208 | if (blk_mq_hw_queue_mapped(hctx)) | |
209 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
210 | |
211 | /* | |
212 | * If we are called because the queue has now been marked as | |
213 | * dying, we need to ensure that processes currently waiting on | |
214 | * the queue are notified as well. | |
215 | */ | |
216 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
217 | } |
218 | ||
320ae51f JA |
219 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
220 | { | |
221 | return blk_mq_has_free_tags(hctx->tags); | |
222 | } | |
223 | EXPORT_SYMBOL(blk_mq_can_queue); | |
224 | ||
e4cdf1a1 CH |
225 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
226 | unsigned int tag, unsigned int op) | |
320ae51f | 227 | { |
e4cdf1a1 CH |
228 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
229 | struct request *rq = tags->static_rqs[tag]; | |
230 | ||
c3a148d2 BVA |
231 | rq->rq_flags = 0; |
232 | ||
e4cdf1a1 CH |
233 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
234 | rq->tag = -1; | |
235 | rq->internal_tag = tag; | |
236 | } else { | |
237 | if (blk_mq_tag_busy(data->hctx)) { | |
238 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
239 | atomic_inc(&data->hctx->nr_active); | |
240 | } | |
241 | rq->tag = tag; | |
242 | rq->internal_tag = -1; | |
243 | data->hctx->tags->rqs[rq->tag] = rq; | |
244 | } | |
245 | ||
af76e555 CH |
246 | INIT_LIST_HEAD(&rq->queuelist); |
247 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
248 | rq->q = data->q; |
249 | rq->mq_ctx = data->ctx; | |
ef295ecf | 250 | rq->cmd_flags = op; |
e4cdf1a1 | 251 | if (blk_queue_io_stat(data->q)) |
e8064021 | 252 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
253 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
254 | rq->cpu = -1; | |
af76e555 CH |
255 | INIT_HLIST_NODE(&rq->hash); |
256 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
257 | rq->rq_disk = NULL; |
258 | rq->part = NULL; | |
3ee32372 | 259 | rq->start_time = jiffies; |
af76e555 CH |
260 | #ifdef CONFIG_BLK_CGROUP |
261 | rq->rl = NULL; | |
0fec08b4 | 262 | set_start_time_ns(rq); |
af76e555 CH |
263 | rq->io_start_time_ns = 0; |
264 | #endif | |
265 | rq->nr_phys_segments = 0; | |
266 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
267 | rq->nr_integrity_segments = 0; | |
268 | #endif | |
af76e555 CH |
269 | rq->special = NULL; |
270 | /* tag was already set */ | |
af76e555 | 271 | rq->extra_len = 0; |
af76e555 | 272 | |
af76e555 | 273 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
274 | rq->timeout = 0; |
275 | ||
af76e555 CH |
276 | rq->end_io = NULL; |
277 | rq->end_io_data = NULL; | |
278 | rq->next_rq = NULL; | |
279 | ||
e4cdf1a1 CH |
280 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
281 | return rq; | |
5dee8577 CH |
282 | } |
283 | ||
d2c0d383 CH |
284 | static struct request *blk_mq_get_request(struct request_queue *q, |
285 | struct bio *bio, unsigned int op, | |
286 | struct blk_mq_alloc_data *data) | |
287 | { | |
288 | struct elevator_queue *e = q->elevator; | |
289 | struct request *rq; | |
e4cdf1a1 | 290 | unsigned int tag; |
d2c0d383 CH |
291 | |
292 | blk_queue_enter_live(q); | |
293 | data->q = q; | |
294 | if (likely(!data->ctx)) | |
295 | data->ctx = blk_mq_get_ctx(q); | |
296 | if (likely(!data->hctx)) | |
297 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
298 | if (op & REQ_NOWAIT) |
299 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
300 | |
301 | if (e) { | |
302 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
303 | ||
304 | /* | |
305 | * Flush requests are special and go directly to the | |
306 | * dispatch list. | |
307 | */ | |
5bbf4e5a CH |
308 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
309 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
310 | } |
311 | ||
e4cdf1a1 CH |
312 | tag = blk_mq_get_tag(data); |
313 | if (tag == BLK_MQ_TAG_FAIL) { | |
037cebb8 CH |
314 | blk_queue_exit(q); |
315 | return NULL; | |
d2c0d383 CH |
316 | } |
317 | ||
e4cdf1a1 | 318 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
319 | if (!op_is_flush(op)) { |
320 | rq->elv.icq = NULL; | |
5bbf4e5a | 321 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
322 | if (e->type->icq_cache && rq_ioc(bio)) |
323 | blk_mq_sched_assign_ioc(rq, bio); | |
324 | ||
5bbf4e5a CH |
325 | e->type->ops.mq.prepare_request(rq, bio); |
326 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 327 | } |
037cebb8 CH |
328 | } |
329 | data->hctx->queued++; | |
330 | return rq; | |
d2c0d383 CH |
331 | } |
332 | ||
cd6ce148 | 333 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
6f3b0e8b | 334 | unsigned int flags) |
320ae51f | 335 | { |
5a797e00 | 336 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 337 | struct request *rq; |
a492f075 | 338 | int ret; |
320ae51f | 339 | |
6f3b0e8b | 340 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
341 | if (ret) |
342 | return ERR_PTR(ret); | |
320ae51f | 343 | |
cd6ce148 | 344 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
841bac2c | 345 | |
bd166ef1 JA |
346 | blk_mq_put_ctx(alloc_data.ctx); |
347 | blk_queue_exit(q); | |
348 | ||
349 | if (!rq) | |
a492f075 | 350 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 CH |
351 | |
352 | rq->__data_len = 0; | |
353 | rq->__sector = (sector_t) -1; | |
354 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
355 | return rq; |
356 | } | |
4bb659b1 | 357 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 358 | |
cd6ce148 BVA |
359 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
360 | unsigned int op, unsigned int flags, unsigned int hctx_idx) | |
1f5bd336 | 361 | { |
6d2809d5 | 362 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 363 | struct request *rq; |
6d2809d5 | 364 | unsigned int cpu; |
1f5bd336 ML |
365 | int ret; |
366 | ||
367 | /* | |
368 | * If the tag allocator sleeps we could get an allocation for a | |
369 | * different hardware context. No need to complicate the low level | |
370 | * allocator for this for the rare use case of a command tied to | |
371 | * a specific queue. | |
372 | */ | |
373 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
374 | return ERR_PTR(-EINVAL); | |
375 | ||
376 | if (hctx_idx >= q->nr_hw_queues) | |
377 | return ERR_PTR(-EIO); | |
378 | ||
379 | ret = blk_queue_enter(q, true); | |
380 | if (ret) | |
381 | return ERR_PTR(ret); | |
382 | ||
c8712c6a CH |
383 | /* |
384 | * Check if the hardware context is actually mapped to anything. | |
385 | * If not tell the caller that it should skip this queue. | |
386 | */ | |
6d2809d5 OS |
387 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
388 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
389 | blk_queue_exit(q); | |
390 | return ERR_PTR(-EXDEV); | |
c8712c6a | 391 | } |
6d2809d5 OS |
392 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
393 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 394 | |
cd6ce148 | 395 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
c8712c6a | 396 | |
c8712c6a | 397 | blk_queue_exit(q); |
6d2809d5 OS |
398 | |
399 | if (!rq) | |
400 | return ERR_PTR(-EWOULDBLOCK); | |
401 | ||
402 | return rq; | |
1f5bd336 ML |
403 | } |
404 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
405 | ||
6af54051 | 406 | void blk_mq_free_request(struct request *rq) |
320ae51f | 407 | { |
320ae51f | 408 | struct request_queue *q = rq->q; |
6af54051 CH |
409 | struct elevator_queue *e = q->elevator; |
410 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
411 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
412 | const int sched_tag = rq->internal_tag; | |
413 | ||
5bbf4e5a | 414 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
415 | if (e && e->type->ops.mq.finish_request) |
416 | e->type->ops.mq.finish_request(rq); | |
417 | if (rq->elv.icq) { | |
418 | put_io_context(rq->elv.icq->ioc); | |
419 | rq->elv.icq = NULL; | |
420 | } | |
421 | } | |
320ae51f | 422 | |
6af54051 | 423 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 424 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 425 | atomic_dec(&hctx->nr_active); |
87760e5e JA |
426 | |
427 | wbt_done(q->rq_wb, &rq->issue_stat); | |
0d2602ca | 428 | |
af76e555 | 429 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 430 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
431 | if (rq->tag != -1) |
432 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
433 | if (sched_tag != -1) | |
c05f8525 | 434 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 435 | blk_mq_sched_restart(hctx); |
3ef28e83 | 436 | blk_queue_exit(q); |
320ae51f | 437 | } |
1a3b595a | 438 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 439 | |
2a842aca | 440 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 441 | { |
0d11e6ac ML |
442 | blk_account_io_done(rq); |
443 | ||
91b63639 | 444 | if (rq->end_io) { |
87760e5e | 445 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 446 | rq->end_io(rq, error); |
91b63639 CH |
447 | } else { |
448 | if (unlikely(blk_bidi_rq(rq))) | |
449 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 450 | blk_mq_free_request(rq); |
91b63639 | 451 | } |
320ae51f | 452 | } |
c8a446ad | 453 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 454 | |
2a842aca | 455 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
456 | { |
457 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
458 | BUG(); | |
c8a446ad | 459 | __blk_mq_end_request(rq, error); |
63151a44 | 460 | } |
c8a446ad | 461 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 462 | |
30a91cb4 | 463 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 464 | { |
3d6efbf6 | 465 | struct request *rq = data; |
320ae51f | 466 | |
30a91cb4 | 467 | rq->q->softirq_done_fn(rq); |
320ae51f | 468 | } |
320ae51f | 469 | |
453f8341 | 470 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
471 | { |
472 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 473 | bool shared = false; |
320ae51f JA |
474 | int cpu; |
475 | ||
453f8341 CH |
476 | if (rq->internal_tag != -1) |
477 | blk_mq_sched_completed_request(rq); | |
478 | if (rq->rq_flags & RQF_STATS) { | |
479 | blk_mq_poll_stats_start(rq->q); | |
480 | blk_stat_add(rq); | |
481 | } | |
482 | ||
38535201 | 483 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
484 | rq->q->softirq_done_fn(rq); |
485 | return; | |
486 | } | |
320ae51f JA |
487 | |
488 | cpu = get_cpu(); | |
38535201 CH |
489 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
490 | shared = cpus_share_cache(cpu, ctx->cpu); | |
491 | ||
492 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 493 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
494 | rq->csd.info = rq; |
495 | rq->csd.flags = 0; | |
c46fff2a | 496 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 497 | } else { |
30a91cb4 | 498 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 499 | } |
320ae51f JA |
500 | put_cpu(); |
501 | } | |
30a91cb4 CH |
502 | |
503 | /** | |
504 | * blk_mq_complete_request - end I/O on a request | |
505 | * @rq: the request being processed | |
506 | * | |
507 | * Description: | |
508 | * Ends all I/O on a request. It does not handle partial completions. | |
509 | * The actual completion happens out-of-order, through a IPI handler. | |
510 | **/ | |
08e0029a | 511 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 512 | { |
95f09684 JA |
513 | struct request_queue *q = rq->q; |
514 | ||
515 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 516 | return; |
08e0029a | 517 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 518 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
519 | } |
520 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 521 | |
973c0191 KB |
522 | int blk_mq_request_started(struct request *rq) |
523 | { | |
524 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
525 | } | |
526 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
527 | ||
e2490073 | 528 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
529 | { |
530 | struct request_queue *q = rq->q; | |
531 | ||
bd166ef1 JA |
532 | blk_mq_sched_started_request(rq); |
533 | ||
320ae51f JA |
534 | trace_block_rq_issue(q, rq); |
535 | ||
cf43e6be | 536 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 537 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 538 | rq->rq_flags |= RQF_STATS; |
87760e5e | 539 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
540 | } |
541 | ||
2b8393b4 | 542 | blk_add_timer(rq); |
87ee7b11 | 543 | |
538b7534 JA |
544 | /* |
545 | * Ensure that ->deadline is visible before set the started | |
546 | * flag and clear the completed flag. | |
547 | */ | |
548 | smp_mb__before_atomic(); | |
549 | ||
87ee7b11 JA |
550 | /* |
551 | * Mark us as started and clear complete. Complete might have been | |
552 | * set if requeue raced with timeout, which then marked it as | |
553 | * complete. So be sure to clear complete again when we start | |
554 | * the request, otherwise we'll ignore the completion event. | |
555 | */ | |
4b570521 JA |
556 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
557 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
558 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
559 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
560 | |
561 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
562 | /* | |
563 | * Make sure space for the drain appears. We know we can do | |
564 | * this because max_hw_segments has been adjusted to be one | |
565 | * fewer than the device can handle. | |
566 | */ | |
567 | rq->nr_phys_segments++; | |
568 | } | |
320ae51f | 569 | } |
e2490073 | 570 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 571 | |
d9d149a3 ML |
572 | /* |
573 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 574 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
575 | * but given rq->deadline is just set in .queue_rq() under |
576 | * this situation, the race won't be possible in reality because | |
577 | * rq->timeout should be set as big enough to cover the window | |
578 | * between blk_mq_start_request() called from .queue_rq() and | |
579 | * clearing REQ_ATOM_STARTED here. | |
580 | */ | |
ed0791b2 | 581 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
582 | { |
583 | struct request_queue *q = rq->q; | |
584 | ||
585 | trace_block_rq_requeue(q, rq); | |
87760e5e | 586 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 587 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 588 | |
e2490073 CH |
589 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
590 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
591 | rq->nr_phys_segments--; | |
592 | } | |
320ae51f JA |
593 | } |
594 | ||
2b053aca | 595 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 596 | { |
ed0791b2 | 597 | __blk_mq_requeue_request(rq); |
ed0791b2 | 598 | |
ed0791b2 | 599 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 600 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
601 | } |
602 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
603 | ||
6fca6a61 CH |
604 | static void blk_mq_requeue_work(struct work_struct *work) |
605 | { | |
606 | struct request_queue *q = | |
2849450a | 607 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
608 | LIST_HEAD(rq_list); |
609 | struct request *rq, *next; | |
610 | unsigned long flags; | |
611 | ||
612 | spin_lock_irqsave(&q->requeue_lock, flags); | |
613 | list_splice_init(&q->requeue_list, &rq_list); | |
614 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
615 | ||
616 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 617 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
618 | continue; |
619 | ||
e8064021 | 620 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 621 | list_del_init(&rq->queuelist); |
bd6737f1 | 622 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
623 | } |
624 | ||
625 | while (!list_empty(&rq_list)) { | |
626 | rq = list_entry(rq_list.next, struct request, queuelist); | |
627 | list_del_init(&rq->queuelist); | |
bd6737f1 | 628 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
629 | } |
630 | ||
52d7f1b5 | 631 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
632 | } |
633 | ||
2b053aca BVA |
634 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
635 | bool kick_requeue_list) | |
6fca6a61 CH |
636 | { |
637 | struct request_queue *q = rq->q; | |
638 | unsigned long flags; | |
639 | ||
640 | /* | |
641 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
642 | * request head insertation from the workqueue. | |
643 | */ | |
e8064021 | 644 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
645 | |
646 | spin_lock_irqsave(&q->requeue_lock, flags); | |
647 | if (at_head) { | |
e8064021 | 648 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
649 | list_add(&rq->queuelist, &q->requeue_list); |
650 | } else { | |
651 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
652 | } | |
653 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
654 | |
655 | if (kick_requeue_list) | |
656 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
657 | } |
658 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
659 | ||
660 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
661 | { | |
2849450a | 662 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
663 | } |
664 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
665 | ||
2849450a MS |
666 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
667 | unsigned long msecs) | |
668 | { | |
669 | kblockd_schedule_delayed_work(&q->requeue_work, | |
670 | msecs_to_jiffies(msecs)); | |
671 | } | |
672 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
673 | ||
0e62f51f JA |
674 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
675 | { | |
88c7b2b7 JA |
676 | if (tag < tags->nr_tags) { |
677 | prefetch(tags->rqs[tag]); | |
4ee86bab | 678 | return tags->rqs[tag]; |
88c7b2b7 | 679 | } |
4ee86bab HR |
680 | |
681 | return NULL; | |
24d2f903 CH |
682 | } |
683 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
684 | ||
320ae51f | 685 | struct blk_mq_timeout_data { |
46f92d42 CH |
686 | unsigned long next; |
687 | unsigned int next_set; | |
320ae51f JA |
688 | }; |
689 | ||
90415837 | 690 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 691 | { |
f8a5b122 | 692 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 693 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
694 | |
695 | /* | |
696 | * We know that complete is set at this point. If STARTED isn't set | |
697 | * anymore, then the request isn't active and the "timeout" should | |
698 | * just be ignored. This can happen due to the bitflag ordering. | |
699 | * Timeout first checks if STARTED is set, and if it is, assumes | |
700 | * the request is active. But if we race with completion, then | |
48b99c9d | 701 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
702 | * a timeout event with a request that isn't active. |
703 | */ | |
46f92d42 CH |
704 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
705 | return; | |
87ee7b11 | 706 | |
46f92d42 | 707 | if (ops->timeout) |
0152fb6b | 708 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
709 | |
710 | switch (ret) { | |
711 | case BLK_EH_HANDLED: | |
712 | __blk_mq_complete_request(req); | |
713 | break; | |
714 | case BLK_EH_RESET_TIMER: | |
715 | blk_add_timer(req); | |
716 | blk_clear_rq_complete(req); | |
717 | break; | |
718 | case BLK_EH_NOT_HANDLED: | |
719 | break; | |
720 | default: | |
721 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
722 | break; | |
723 | } | |
87ee7b11 | 724 | } |
5b3f25fc | 725 | |
81481eb4 CH |
726 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
727 | struct request *rq, void *priv, bool reserved) | |
728 | { | |
729 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 730 | |
95a49603 | 731 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 732 | return; |
87ee7b11 | 733 | |
d9d149a3 ML |
734 | /* |
735 | * The rq being checked may have been freed and reallocated | |
736 | * out already here, we avoid this race by checking rq->deadline | |
737 | * and REQ_ATOM_COMPLETE flag together: | |
738 | * | |
739 | * - if rq->deadline is observed as new value because of | |
740 | * reusing, the rq won't be timed out because of timing. | |
741 | * - if rq->deadline is observed as previous value, | |
742 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
743 | * because we put a barrier between setting rq->deadline | |
744 | * and clearing the flag in blk_mq_start_request(), so | |
745 | * this rq won't be timed out too. | |
746 | */ | |
46f92d42 CH |
747 | if (time_after_eq(jiffies, rq->deadline)) { |
748 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 749 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
750 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
751 | data->next = rq->deadline; | |
752 | data->next_set = 1; | |
753 | } | |
87ee7b11 JA |
754 | } |
755 | ||
287922eb | 756 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 757 | { |
287922eb CH |
758 | struct request_queue *q = |
759 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
760 | struct blk_mq_timeout_data data = { |
761 | .next = 0, | |
762 | .next_set = 0, | |
763 | }; | |
81481eb4 | 764 | int i; |
320ae51f | 765 | |
71f79fb3 GKB |
766 | /* A deadlock might occur if a request is stuck requiring a |
767 | * timeout at the same time a queue freeze is waiting | |
768 | * completion, since the timeout code would not be able to | |
769 | * acquire the queue reference here. | |
770 | * | |
771 | * That's why we don't use blk_queue_enter here; instead, we use | |
772 | * percpu_ref_tryget directly, because we need to be able to | |
773 | * obtain a reference even in the short window between the queue | |
774 | * starting to freeze, by dropping the first reference in | |
1671d522 | 775 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
776 | * consumed, marked by the instant q_usage_counter reaches |
777 | * zero. | |
778 | */ | |
779 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
780 | return; |
781 | ||
0bf6cd5b | 782 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 783 | |
81481eb4 CH |
784 | if (data.next_set) { |
785 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
786 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 787 | } else { |
0bf6cd5b CH |
788 | struct blk_mq_hw_ctx *hctx; |
789 | ||
f054b56c ML |
790 | queue_for_each_hw_ctx(q, hctx, i) { |
791 | /* the hctx may be unmapped, so check it here */ | |
792 | if (blk_mq_hw_queue_mapped(hctx)) | |
793 | blk_mq_tag_idle(hctx); | |
794 | } | |
0d2602ca | 795 | } |
287922eb | 796 | blk_queue_exit(q); |
320ae51f JA |
797 | } |
798 | ||
88459642 OS |
799 | struct flush_busy_ctx_data { |
800 | struct blk_mq_hw_ctx *hctx; | |
801 | struct list_head *list; | |
802 | }; | |
803 | ||
804 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
805 | { | |
806 | struct flush_busy_ctx_data *flush_data = data; | |
807 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
808 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
809 | ||
810 | sbitmap_clear_bit(sb, bitnr); | |
811 | spin_lock(&ctx->lock); | |
812 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
813 | spin_unlock(&ctx->lock); | |
814 | return true; | |
815 | } | |
816 | ||
1429d7c9 JA |
817 | /* |
818 | * Process software queues that have been marked busy, splicing them | |
819 | * to the for-dispatch | |
820 | */ | |
2c3ad667 | 821 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 822 | { |
88459642 OS |
823 | struct flush_busy_ctx_data data = { |
824 | .hctx = hctx, | |
825 | .list = list, | |
826 | }; | |
1429d7c9 | 827 | |
88459642 | 828 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 829 | } |
2c3ad667 | 830 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 831 | |
703fd1c0 JA |
832 | static inline unsigned int queued_to_index(unsigned int queued) |
833 | { | |
834 | if (!queued) | |
835 | return 0; | |
1429d7c9 | 836 | |
703fd1c0 | 837 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
838 | } |
839 | ||
bd6737f1 JA |
840 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
841 | bool wait) | |
bd166ef1 JA |
842 | { |
843 | struct blk_mq_alloc_data data = { | |
844 | .q = rq->q, | |
bd166ef1 JA |
845 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
846 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
847 | }; | |
848 | ||
5feeacdd JA |
849 | might_sleep_if(wait); |
850 | ||
81380ca1 OS |
851 | if (rq->tag != -1) |
852 | goto done; | |
bd166ef1 | 853 | |
415b806d SG |
854 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
855 | data.flags |= BLK_MQ_REQ_RESERVED; | |
856 | ||
bd166ef1 JA |
857 | rq->tag = blk_mq_get_tag(&data); |
858 | if (rq->tag >= 0) { | |
200e86b3 JA |
859 | if (blk_mq_tag_busy(data.hctx)) { |
860 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
861 | atomic_inc(&data.hctx->nr_active); | |
862 | } | |
bd166ef1 | 863 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
864 | } |
865 | ||
81380ca1 OS |
866 | done: |
867 | if (hctx) | |
868 | *hctx = data.hctx; | |
869 | return rq->tag != -1; | |
bd166ef1 JA |
870 | } |
871 | ||
113285b4 JA |
872 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
873 | struct request *rq) | |
99cf1dc5 | 874 | { |
99cf1dc5 JA |
875 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
876 | rq->tag = -1; | |
877 | ||
878 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
879 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
880 | atomic_dec(&hctx->nr_active); | |
881 | } | |
882 | } | |
883 | ||
113285b4 JA |
884 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
885 | struct request *rq) | |
886 | { | |
887 | if (rq->tag == -1 || rq->internal_tag == -1) | |
888 | return; | |
889 | ||
890 | __blk_mq_put_driver_tag(hctx, rq); | |
891 | } | |
892 | ||
893 | static void blk_mq_put_driver_tag(struct request *rq) | |
894 | { | |
895 | struct blk_mq_hw_ctx *hctx; | |
896 | ||
897 | if (rq->tag == -1 || rq->internal_tag == -1) | |
898 | return; | |
899 | ||
900 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
901 | __blk_mq_put_driver_tag(hctx, rq); | |
902 | } | |
903 | ||
bd166ef1 JA |
904 | /* |
905 | * If we fail getting a driver tag because all the driver tags are already | |
906 | * assigned and on the dispatch list, BUT the first entry does not have a | |
907 | * tag, then we could deadlock. For that case, move entries with assigned | |
908 | * driver tags to the front, leaving the set of tagged requests in the | |
909 | * same order, and the untagged set in the same order. | |
910 | */ | |
911 | static bool reorder_tags_to_front(struct list_head *list) | |
912 | { | |
913 | struct request *rq, *tmp, *first = NULL; | |
914 | ||
915 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
916 | if (rq == first) | |
917 | break; | |
918 | if (rq->tag != -1) { | |
919 | list_move(&rq->queuelist, list); | |
920 | if (!first) | |
921 | first = rq; | |
922 | } | |
923 | } | |
924 | ||
925 | return first != NULL; | |
926 | } | |
927 | ||
da55f2cc OS |
928 | static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags, |
929 | void *key) | |
930 | { | |
931 | struct blk_mq_hw_ctx *hctx; | |
932 | ||
933 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
934 | ||
935 | list_del(&wait->task_list); | |
936 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); | |
937 | blk_mq_run_hw_queue(hctx, true); | |
938 | return 1; | |
939 | } | |
940 | ||
941 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
942 | { | |
943 | struct sbq_wait_state *ws; | |
944 | ||
945 | /* | |
946 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
947 | * The thread which wins the race to grab this bit adds the hardware | |
948 | * queue to the wait queue. | |
949 | */ | |
950 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
951 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
952 | return false; | |
953 | ||
954 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
955 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
956 | ||
957 | /* | |
958 | * As soon as this returns, it's no longer safe to fiddle with | |
959 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
960 | * and unlock the bit. | |
961 | */ | |
962 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
963 | return true; | |
964 | } | |
965 | ||
81380ca1 | 966 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
320ae51f | 967 | { |
81380ca1 | 968 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 969 | struct request *rq; |
fc17b653 | 970 | int errors, queued; |
320ae51f | 971 | |
81380ca1 OS |
972 | if (list_empty(list)) |
973 | return false; | |
974 | ||
320ae51f JA |
975 | /* |
976 | * Now process all the entries, sending them to the driver. | |
977 | */ | |
93efe981 | 978 | errors = queued = 0; |
81380ca1 | 979 | do { |
74c45052 | 980 | struct blk_mq_queue_data bd; |
fc17b653 | 981 | blk_status_t ret; |
320ae51f | 982 | |
f04c3df3 | 983 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
984 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
985 | if (!queued && reorder_tags_to_front(list)) | |
986 | continue; | |
3c782d67 JA |
987 | |
988 | /* | |
da55f2cc OS |
989 | * The initial allocation attempt failed, so we need to |
990 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 991 | */ |
807b1041 OS |
992 | if (!blk_mq_dispatch_wait_add(hctx)) |
993 | break; | |
994 | ||
995 | /* | |
996 | * It's possible that a tag was freed in the window | |
997 | * between the allocation failure and adding the | |
998 | * hardware queue to the wait queue. | |
999 | */ | |
1000 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
3c782d67 | 1001 | break; |
bd166ef1 | 1002 | } |
da55f2cc | 1003 | |
320ae51f | 1004 | list_del_init(&rq->queuelist); |
320ae51f | 1005 | |
74c45052 | 1006 | bd.rq = rq; |
113285b4 JA |
1007 | |
1008 | /* | |
1009 | * Flag last if we have no more requests, or if we have more | |
1010 | * but can't assign a driver tag to it. | |
1011 | */ | |
1012 | if (list_empty(list)) | |
1013 | bd.last = true; | |
1014 | else { | |
1015 | struct request *nxt; | |
1016 | ||
1017 | nxt = list_first_entry(list, struct request, queuelist); | |
1018 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1019 | } | |
74c45052 JA |
1020 | |
1021 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1022 | if (ret == BLK_STS_RESOURCE) { |
113285b4 | 1023 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1024 | list_add(&rq->queuelist, list); |
ed0791b2 | 1025 | __blk_mq_requeue_request(rq); |
320ae51f | 1026 | break; |
fc17b653 CH |
1027 | } |
1028 | ||
1029 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1030 | errors++; |
2a842aca | 1031 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1032 | continue; |
320ae51f JA |
1033 | } |
1034 | ||
fc17b653 | 1035 | queued++; |
81380ca1 | 1036 | } while (!list_empty(list)); |
320ae51f | 1037 | |
703fd1c0 | 1038 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1039 | |
1040 | /* | |
1041 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1042 | * that is where we will continue on next queue run. | |
1043 | */ | |
f04c3df3 | 1044 | if (!list_empty(list)) { |
113285b4 | 1045 | /* |
710c785f BVA |
1046 | * If an I/O scheduler has been configured and we got a driver |
1047 | * tag for the next request already, free it again. | |
113285b4 JA |
1048 | */ |
1049 | rq = list_first_entry(list, struct request, queuelist); | |
1050 | blk_mq_put_driver_tag(rq); | |
1051 | ||
320ae51f | 1052 | spin_lock(&hctx->lock); |
c13660a0 | 1053 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1054 | spin_unlock(&hctx->lock); |
f04c3df3 | 1055 | |
9ba52e58 | 1056 | /* |
710c785f BVA |
1057 | * If SCHED_RESTART was set by the caller of this function and |
1058 | * it is no longer set that means that it was cleared by another | |
1059 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1060 | * |
710c785f BVA |
1061 | * If TAG_WAITING is set that means that an I/O scheduler has |
1062 | * been configured and another thread is waiting for a driver | |
1063 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1064 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1065 | * |
710c785f BVA |
1066 | * If no I/O scheduler has been configured it is possible that |
1067 | * the hardware queue got stopped and restarted before requests | |
1068 | * were pushed back onto the dispatch list. Rerun the queue to | |
1069 | * avoid starvation. Notes: | |
1070 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1071 | * been stopped before rerunning a queue. | |
1072 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1073 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1074 | * and dm-rq. |
bd166ef1 | 1075 | */ |
da55f2cc OS |
1076 | if (!blk_mq_sched_needs_restart(hctx) && |
1077 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1078 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1079 | } |
f04c3df3 | 1080 | |
93efe981 | 1081 | return (queued + errors) != 0; |
f04c3df3 JA |
1082 | } |
1083 | ||
6a83e74d BVA |
1084 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1085 | { | |
1086 | int srcu_idx; | |
1087 | ||
1088 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1089 | cpu_online(hctx->next_cpu)); | |
1090 | ||
1091 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1092 | rcu_read_lock(); | |
bd166ef1 | 1093 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1094 | rcu_read_unlock(); |
1095 | } else { | |
bf4907c0 JA |
1096 | might_sleep(); |
1097 | ||
07319678 | 1098 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
bd166ef1 | 1099 | blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1100 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d BVA |
1101 | } |
1102 | } | |
1103 | ||
506e931f JA |
1104 | /* |
1105 | * It'd be great if the workqueue API had a way to pass | |
1106 | * in a mask and had some smarts for more clever placement. | |
1107 | * For now we just round-robin here, switching for every | |
1108 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1109 | */ | |
1110 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1111 | { | |
b657d7e6 CH |
1112 | if (hctx->queue->nr_hw_queues == 1) |
1113 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1114 | |
1115 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1116 | int next_cpu; |
506e931f JA |
1117 | |
1118 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1119 | if (next_cpu >= nr_cpu_ids) | |
1120 | next_cpu = cpumask_first(hctx->cpumask); | |
1121 | ||
1122 | hctx->next_cpu = next_cpu; | |
1123 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1124 | } | |
1125 | ||
b657d7e6 | 1126 | return hctx->next_cpu; |
506e931f JA |
1127 | } |
1128 | ||
7587a5ae BVA |
1129 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1130 | unsigned long msecs) | |
320ae51f | 1131 | { |
5435c023 BVA |
1132 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1133 | return; | |
1134 | ||
1135 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1136 | return; |
1137 | ||
1b792f2f | 1138 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1139 | int cpu = get_cpu(); |
1140 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1141 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1142 | put_cpu(); |
398205b8 PB |
1143 | return; |
1144 | } | |
e4043dcf | 1145 | |
2a90d4aa | 1146 | put_cpu(); |
e4043dcf | 1147 | } |
398205b8 | 1148 | |
9f993737 JA |
1149 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1150 | &hctx->run_work, | |
1151 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1152 | } |
1153 | ||
1154 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1155 | { | |
1156 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1157 | } | |
1158 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1159 | ||
1160 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1161 | { | |
1162 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1163 | } |
5b727272 | 1164 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1165 | |
b94ec296 | 1166 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1167 | { |
1168 | struct blk_mq_hw_ctx *hctx; | |
1169 | int i; | |
1170 | ||
1171 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1172 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1173 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1174 | continue; |
1175 | ||
b94ec296 | 1176 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1177 | } |
1178 | } | |
b94ec296 | 1179 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1180 | |
fd001443 BVA |
1181 | /** |
1182 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1183 | * @q: request queue. | |
1184 | * | |
1185 | * The caller is responsible for serializing this function against | |
1186 | * blk_mq_{start,stop}_hw_queue(). | |
1187 | */ | |
1188 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1189 | { | |
1190 | struct blk_mq_hw_ctx *hctx; | |
1191 | int i; | |
1192 | ||
1193 | queue_for_each_hw_ctx(q, hctx, i) | |
1194 | if (blk_mq_hctx_stopped(hctx)) | |
1195 | return true; | |
1196 | ||
1197 | return false; | |
1198 | } | |
1199 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1200 | ||
39a70c76 ML |
1201 | /* |
1202 | * This function is often used for pausing .queue_rq() by driver when | |
1203 | * there isn't enough resource or some conditions aren't satisfied, and | |
1204 | * BLK_MQ_RQ_QUEUE_BUSY is usually returned. | |
1205 | * | |
1206 | * We do not guarantee that dispatch can be drained or blocked | |
1207 | * after blk_mq_stop_hw_queue() returns. Please use | |
1208 | * blk_mq_quiesce_queue() for that requirement. | |
1209 | */ | |
2719aa21 JA |
1210 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1211 | { | |
641a9ed6 | 1212 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1213 | |
641a9ed6 | 1214 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1215 | } |
641a9ed6 | 1216 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1217 | |
39a70c76 ML |
1218 | /* |
1219 | * This function is often used for pausing .queue_rq() by driver when | |
1220 | * there isn't enough resource or some conditions aren't satisfied, and | |
1221 | * BLK_MQ_RQ_QUEUE_BUSY is usually returned. | |
1222 | * | |
1223 | * We do not guarantee that dispatch can be drained or blocked | |
1224 | * after blk_mq_stop_hw_queues() returns. Please use | |
1225 | * blk_mq_quiesce_queue() for that requirement. | |
1226 | */ | |
2719aa21 JA |
1227 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1228 | { | |
641a9ed6 ML |
1229 | struct blk_mq_hw_ctx *hctx; |
1230 | int i; | |
1231 | ||
1232 | queue_for_each_hw_ctx(q, hctx, i) | |
1233 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1234 | } |
1235 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1236 | ||
320ae51f JA |
1237 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1238 | { | |
1239 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1240 | |
0ffbce80 | 1241 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1242 | } |
1243 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1244 | ||
2f268556 CH |
1245 | void blk_mq_start_hw_queues(struct request_queue *q) |
1246 | { | |
1247 | struct blk_mq_hw_ctx *hctx; | |
1248 | int i; | |
1249 | ||
1250 | queue_for_each_hw_ctx(q, hctx, i) | |
1251 | blk_mq_start_hw_queue(hctx); | |
1252 | } | |
1253 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1254 | ||
ae911c5e JA |
1255 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1256 | { | |
1257 | if (!blk_mq_hctx_stopped(hctx)) | |
1258 | return; | |
1259 | ||
1260 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1261 | blk_mq_run_hw_queue(hctx, async); | |
1262 | } | |
1263 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1264 | ||
1b4a3258 | 1265 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1266 | { |
1267 | struct blk_mq_hw_ctx *hctx; | |
1268 | int i; | |
1269 | ||
ae911c5e JA |
1270 | queue_for_each_hw_ctx(q, hctx, i) |
1271 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1272 | } |
1273 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1274 | ||
70f4db63 | 1275 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1276 | { |
1277 | struct blk_mq_hw_ctx *hctx; | |
1278 | ||
9f993737 | 1279 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1280 | |
21c6e939 JA |
1281 | /* |
1282 | * If we are stopped, don't run the queue. The exception is if | |
1283 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1284 | * the STOPPED bit and run it. | |
1285 | */ | |
1286 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1287 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1288 | return; | |
7587a5ae | 1289 | |
21c6e939 JA |
1290 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1291 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1292 | } | |
7587a5ae BVA |
1293 | |
1294 | __blk_mq_run_hw_queue(hctx); | |
1295 | } | |
1296 | ||
70f4db63 CH |
1297 | |
1298 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1299 | { | |
5435c023 | 1300 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1301 | return; |
70f4db63 | 1302 | |
21c6e939 JA |
1303 | /* |
1304 | * Stop the hw queue, then modify currently delayed work. | |
1305 | * This should prevent us from running the queue prematurely. | |
1306 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1307 | */ | |
7e79dadc | 1308 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1309 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1310 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1311 | &hctx->run_work, | |
1312 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1313 | } |
1314 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1315 | ||
cfd0c552 | 1316 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1317 | struct request *rq, |
1318 | bool at_head) | |
320ae51f | 1319 | { |
e57690fe JA |
1320 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1321 | ||
7b607814 BVA |
1322 | lockdep_assert_held(&ctx->lock); |
1323 | ||
01b983c9 JA |
1324 | trace_block_rq_insert(hctx->queue, rq); |
1325 | ||
72a0a36e CH |
1326 | if (at_head) |
1327 | list_add(&rq->queuelist, &ctx->rq_list); | |
1328 | else | |
1329 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1330 | } |
4bb659b1 | 1331 | |
2c3ad667 JA |
1332 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1333 | bool at_head) | |
cfd0c552 ML |
1334 | { |
1335 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1336 | ||
7b607814 BVA |
1337 | lockdep_assert_held(&ctx->lock); |
1338 | ||
e57690fe | 1339 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1340 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1341 | } |
1342 | ||
bd166ef1 JA |
1343 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1344 | struct list_head *list) | |
320ae51f JA |
1345 | |
1346 | { | |
320ae51f JA |
1347 | /* |
1348 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1349 | * offline now | |
1350 | */ | |
1351 | spin_lock(&ctx->lock); | |
1352 | while (!list_empty(list)) { | |
1353 | struct request *rq; | |
1354 | ||
1355 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1356 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1357 | list_del_init(&rq->queuelist); |
e57690fe | 1358 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1359 | } |
cfd0c552 | 1360 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1361 | spin_unlock(&ctx->lock); |
320ae51f JA |
1362 | } |
1363 | ||
1364 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1365 | { | |
1366 | struct request *rqa = container_of(a, struct request, queuelist); | |
1367 | struct request *rqb = container_of(b, struct request, queuelist); | |
1368 | ||
1369 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1370 | (rqa->mq_ctx == rqb->mq_ctx && | |
1371 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1372 | } | |
1373 | ||
1374 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1375 | { | |
1376 | struct blk_mq_ctx *this_ctx; | |
1377 | struct request_queue *this_q; | |
1378 | struct request *rq; | |
1379 | LIST_HEAD(list); | |
1380 | LIST_HEAD(ctx_list); | |
1381 | unsigned int depth; | |
1382 | ||
1383 | list_splice_init(&plug->mq_list, &list); | |
1384 | ||
1385 | list_sort(NULL, &list, plug_ctx_cmp); | |
1386 | ||
1387 | this_q = NULL; | |
1388 | this_ctx = NULL; | |
1389 | depth = 0; | |
1390 | ||
1391 | while (!list_empty(&list)) { | |
1392 | rq = list_entry_rq(list.next); | |
1393 | list_del_init(&rq->queuelist); | |
1394 | BUG_ON(!rq->q); | |
1395 | if (rq->mq_ctx != this_ctx) { | |
1396 | if (this_ctx) { | |
bd166ef1 JA |
1397 | trace_block_unplug(this_q, depth, from_schedule); |
1398 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1399 | &ctx_list, | |
1400 | from_schedule); | |
320ae51f JA |
1401 | } |
1402 | ||
1403 | this_ctx = rq->mq_ctx; | |
1404 | this_q = rq->q; | |
1405 | depth = 0; | |
1406 | } | |
1407 | ||
1408 | depth++; | |
1409 | list_add_tail(&rq->queuelist, &ctx_list); | |
1410 | } | |
1411 | ||
1412 | /* | |
1413 | * If 'this_ctx' is set, we know we have entries to complete | |
1414 | * on 'ctx_list'. Do those. | |
1415 | */ | |
1416 | if (this_ctx) { | |
bd166ef1 JA |
1417 | trace_block_unplug(this_q, depth, from_schedule); |
1418 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1419 | from_schedule); | |
320ae51f JA |
1420 | } |
1421 | } | |
1422 | ||
1423 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1424 | { | |
da8d7f07 | 1425 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1426 | |
6e85eaf3 | 1427 | blk_account_io_start(rq, true); |
320ae51f JA |
1428 | } |
1429 | ||
274a5843 JA |
1430 | static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx) |
1431 | { | |
1432 | return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) && | |
1433 | !blk_queue_nomerges(hctx->queue); | |
1434 | } | |
1435 | ||
ab42f35d ML |
1436 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1437 | struct blk_mq_ctx *ctx, | |
1438 | struct request *rq) | |
1439 | { | |
1440 | spin_lock(&ctx->lock); | |
1441 | __blk_mq_insert_request(hctx, rq, false); | |
1442 | spin_unlock(&ctx->lock); | |
07068d5b | 1443 | } |
14ec77f3 | 1444 | |
fd2d3326 JA |
1445 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1446 | { | |
bd166ef1 JA |
1447 | if (rq->tag != -1) |
1448 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1449 | ||
1450 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1451 | } |
1452 | ||
d964f04a ML |
1453 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1454 | struct request *rq, | |
1455 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1456 | { |
f984df1f | 1457 | struct request_queue *q = rq->q; |
f984df1f SL |
1458 | struct blk_mq_queue_data bd = { |
1459 | .rq = rq, | |
d945a365 | 1460 | .last = true, |
f984df1f | 1461 | }; |
bd166ef1 | 1462 | blk_qc_t new_cookie; |
f06345ad | 1463 | blk_status_t ret; |
d964f04a ML |
1464 | bool run_queue = true; |
1465 | ||
f4560ffe ML |
1466 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1467 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1468 | run_queue = false; |
1469 | goto insert; | |
1470 | } | |
f984df1f | 1471 | |
bd166ef1 | 1472 | if (q->elevator) |
2253efc8 BVA |
1473 | goto insert; |
1474 | ||
d964f04a | 1475 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1476 | goto insert; |
1477 | ||
1478 | new_cookie = request_to_qc_t(hctx, rq); | |
1479 | ||
f984df1f SL |
1480 | /* |
1481 | * For OK queue, we are done. For error, kill it. Any other | |
1482 | * error (busy), just add it to our list as we previously | |
1483 | * would have done | |
1484 | */ | |
1485 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1486 | switch (ret) { |
1487 | case BLK_STS_OK: | |
7b371636 | 1488 | *cookie = new_cookie; |
2253efc8 | 1489 | return; |
fc17b653 CH |
1490 | case BLK_STS_RESOURCE: |
1491 | __blk_mq_requeue_request(rq); | |
1492 | goto insert; | |
1493 | default: | |
7b371636 | 1494 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1495 | blk_mq_end_request(rq, ret); |
2253efc8 | 1496 | return; |
f984df1f | 1497 | } |
7b371636 | 1498 | |
2253efc8 | 1499 | insert: |
d964f04a | 1500 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1501 | } |
1502 | ||
5eb6126e CH |
1503 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1504 | struct request *rq, blk_qc_t *cookie) | |
1505 | { | |
1506 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1507 | rcu_read_lock(); | |
d964f04a | 1508 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1509 | rcu_read_unlock(); |
1510 | } else { | |
bf4907c0 JA |
1511 | unsigned int srcu_idx; |
1512 | ||
1513 | might_sleep(); | |
1514 | ||
07319678 | 1515 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1516 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1517 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1518 | } |
1519 | } | |
1520 | ||
dece1635 | 1521 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1522 | { |
ef295ecf | 1523 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1524 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1525 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1526 | struct request *rq; |
5eb6126e | 1527 | unsigned int request_count = 0; |
f984df1f | 1528 | struct blk_plug *plug; |
5b3f341f | 1529 | struct request *same_queue_rq = NULL; |
7b371636 | 1530 | blk_qc_t cookie; |
87760e5e | 1531 | unsigned int wb_acct; |
07068d5b JA |
1532 | |
1533 | blk_queue_bounce(q, &bio); | |
1534 | ||
af67c31f | 1535 | blk_queue_split(q, &bio); |
f36ea50c | 1536 | |
07068d5b | 1537 | if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) { |
4246a0b6 | 1538 | bio_io_error(bio); |
dece1635 | 1539 | return BLK_QC_T_NONE; |
07068d5b JA |
1540 | } |
1541 | ||
87c279e6 OS |
1542 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1543 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1544 | return BLK_QC_T_NONE; | |
f984df1f | 1545 | |
bd166ef1 JA |
1546 | if (blk_mq_sched_bio_merge(q, bio)) |
1547 | return BLK_QC_T_NONE; | |
1548 | ||
87760e5e JA |
1549 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1550 | ||
bd166ef1 JA |
1551 | trace_block_getrq(q, bio, bio->bi_opf); |
1552 | ||
d2c0d383 | 1553 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1554 | if (unlikely(!rq)) { |
1555 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1556 | if (bio->bi_opf & REQ_NOWAIT) |
1557 | bio_wouldblock_error(bio); | |
dece1635 | 1558 | return BLK_QC_T_NONE; |
87760e5e JA |
1559 | } |
1560 | ||
1561 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1562 | |
fd2d3326 | 1563 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1564 | |
f984df1f | 1565 | plug = current->plug; |
07068d5b | 1566 | if (unlikely(is_flush_fua)) { |
f984df1f | 1567 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1568 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1569 | if (q->elevator) { |
1570 | blk_mq_sched_insert_request(rq, false, true, true, | |
1571 | true); | |
6a83e74d | 1572 | } else { |
a4d907b6 CH |
1573 | blk_insert_flush(rq); |
1574 | blk_mq_run_hw_queue(data.hctx, true); | |
6a83e74d | 1575 | } |
a4d907b6 | 1576 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1577 | struct request *last = NULL; |
1578 | ||
b00c53e8 | 1579 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1580 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1581 | |
1582 | /* | |
1583 | * @request_count may become stale because of schedule | |
1584 | * out, so check the list again. | |
1585 | */ | |
1586 | if (list_empty(&plug->mq_list)) | |
1587 | request_count = 0; | |
254d259d CH |
1588 | else if (blk_queue_nomerges(q)) |
1589 | request_count = blk_plug_queued_count(q); | |
1590 | ||
676d0607 | 1591 | if (!request_count) |
e6c4438b | 1592 | trace_block_plug(q); |
600271d9 SL |
1593 | else |
1594 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1595 | |
600271d9 SL |
1596 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1597 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1598 | blk_flush_plug_list(plug, false); |
1599 | trace_block_plug(q); | |
320ae51f | 1600 | } |
b094f89c | 1601 | |
e6c4438b | 1602 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1603 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1604 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1605 | |
07068d5b | 1606 | /* |
6a83e74d | 1607 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1608 | * Otherwise the existing request in the plug list will be |
1609 | * issued. So the plug list will have one request at most | |
2299722c CH |
1610 | * The plug list might get flushed before this. If that happens, |
1611 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1612 | */ |
2299722c CH |
1613 | if (list_empty(&plug->mq_list)) |
1614 | same_queue_rq = NULL; | |
1615 | if (same_queue_rq) | |
1616 | list_del_init(&same_queue_rq->queuelist); | |
1617 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1618 | ||
bf4907c0 JA |
1619 | blk_mq_put_ctx(data.ctx); |
1620 | ||
dad7a3be ML |
1621 | if (same_queue_rq) { |
1622 | data.hctx = blk_mq_map_queue(q, | |
1623 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1624 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1625 | &cookie); | |
dad7a3be | 1626 | } |
a4d907b6 | 1627 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1628 | blk_mq_put_ctx(data.ctx); |
2299722c | 1629 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1630 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1631 | } else if (q->elevator) { |
b00c53e8 | 1632 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1633 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1634 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1635 | } else { |
b00c53e8 | 1636 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1637 | blk_mq_bio_to_request(rq, bio); |
1638 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1639 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1640 | } |
320ae51f | 1641 | |
7b371636 | 1642 | return cookie; |
320ae51f JA |
1643 | } |
1644 | ||
cc71a6f4 JA |
1645 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1646 | unsigned int hctx_idx) | |
95363efd | 1647 | { |
e9b267d9 | 1648 | struct page *page; |
320ae51f | 1649 | |
24d2f903 | 1650 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1651 | int i; |
320ae51f | 1652 | |
24d2f903 | 1653 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1654 | struct request *rq = tags->static_rqs[i]; |
1655 | ||
1656 | if (!rq) | |
e9b267d9 | 1657 | continue; |
d6296d39 | 1658 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1659 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1660 | } |
320ae51f | 1661 | } |
320ae51f | 1662 | |
24d2f903 CH |
1663 | while (!list_empty(&tags->page_list)) { |
1664 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1665 | list_del_init(&page->lru); |
f75782e4 CM |
1666 | /* |
1667 | * Remove kmemleak object previously allocated in | |
1668 | * blk_mq_init_rq_map(). | |
1669 | */ | |
1670 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1671 | __free_pages(page, page->private); |
1672 | } | |
cc71a6f4 | 1673 | } |
320ae51f | 1674 | |
cc71a6f4 JA |
1675 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1676 | { | |
24d2f903 | 1677 | kfree(tags->rqs); |
cc71a6f4 | 1678 | tags->rqs = NULL; |
2af8cbe3 JA |
1679 | kfree(tags->static_rqs); |
1680 | tags->static_rqs = NULL; | |
320ae51f | 1681 | |
24d2f903 | 1682 | blk_mq_free_tags(tags); |
320ae51f JA |
1683 | } |
1684 | ||
cc71a6f4 JA |
1685 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1686 | unsigned int hctx_idx, | |
1687 | unsigned int nr_tags, | |
1688 | unsigned int reserved_tags) | |
320ae51f | 1689 | { |
24d2f903 | 1690 | struct blk_mq_tags *tags; |
59f082e4 | 1691 | int node; |
320ae51f | 1692 | |
59f082e4 SL |
1693 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1694 | if (node == NUMA_NO_NODE) | |
1695 | node = set->numa_node; | |
1696 | ||
1697 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1698 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1699 | if (!tags) |
1700 | return NULL; | |
320ae51f | 1701 | |
cc71a6f4 | 1702 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1703 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1704 | node); |
24d2f903 CH |
1705 | if (!tags->rqs) { |
1706 | blk_mq_free_tags(tags); | |
1707 | return NULL; | |
1708 | } | |
320ae51f | 1709 | |
2af8cbe3 JA |
1710 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1711 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1712 | node); |
2af8cbe3 JA |
1713 | if (!tags->static_rqs) { |
1714 | kfree(tags->rqs); | |
1715 | blk_mq_free_tags(tags); | |
1716 | return NULL; | |
1717 | } | |
1718 | ||
cc71a6f4 JA |
1719 | return tags; |
1720 | } | |
1721 | ||
1722 | static size_t order_to_size(unsigned int order) | |
1723 | { | |
1724 | return (size_t)PAGE_SIZE << order; | |
1725 | } | |
1726 | ||
1727 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1728 | unsigned int hctx_idx, unsigned int depth) | |
1729 | { | |
1730 | unsigned int i, j, entries_per_page, max_order = 4; | |
1731 | size_t rq_size, left; | |
59f082e4 SL |
1732 | int node; |
1733 | ||
1734 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1735 | if (node == NUMA_NO_NODE) | |
1736 | node = set->numa_node; | |
cc71a6f4 JA |
1737 | |
1738 | INIT_LIST_HEAD(&tags->page_list); | |
1739 | ||
320ae51f JA |
1740 | /* |
1741 | * rq_size is the size of the request plus driver payload, rounded | |
1742 | * to the cacheline size | |
1743 | */ | |
24d2f903 | 1744 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1745 | cache_line_size()); |
cc71a6f4 | 1746 | left = rq_size * depth; |
320ae51f | 1747 | |
cc71a6f4 | 1748 | for (i = 0; i < depth; ) { |
320ae51f JA |
1749 | int this_order = max_order; |
1750 | struct page *page; | |
1751 | int to_do; | |
1752 | void *p; | |
1753 | ||
b3a834b1 | 1754 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1755 | this_order--; |
1756 | ||
1757 | do { | |
59f082e4 | 1758 | page = alloc_pages_node(node, |
36e1f3d1 | 1759 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1760 | this_order); |
320ae51f JA |
1761 | if (page) |
1762 | break; | |
1763 | if (!this_order--) | |
1764 | break; | |
1765 | if (order_to_size(this_order) < rq_size) | |
1766 | break; | |
1767 | } while (1); | |
1768 | ||
1769 | if (!page) | |
24d2f903 | 1770 | goto fail; |
320ae51f JA |
1771 | |
1772 | page->private = this_order; | |
24d2f903 | 1773 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1774 | |
1775 | p = page_address(page); | |
f75782e4 CM |
1776 | /* |
1777 | * Allow kmemleak to scan these pages as they contain pointers | |
1778 | * to additional allocations like via ops->init_request(). | |
1779 | */ | |
36e1f3d1 | 1780 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1781 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1782 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1783 | left -= to_do * rq_size; |
1784 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1785 | struct request *rq = p; |
1786 | ||
1787 | tags->static_rqs[i] = rq; | |
24d2f903 | 1788 | if (set->ops->init_request) { |
d6296d39 | 1789 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1790 | node)) { |
2af8cbe3 | 1791 | tags->static_rqs[i] = NULL; |
24d2f903 | 1792 | goto fail; |
a5164405 | 1793 | } |
e9b267d9 CH |
1794 | } |
1795 | ||
320ae51f JA |
1796 | p += rq_size; |
1797 | i++; | |
1798 | } | |
1799 | } | |
cc71a6f4 | 1800 | return 0; |
320ae51f | 1801 | |
24d2f903 | 1802 | fail: |
cc71a6f4 JA |
1803 | blk_mq_free_rqs(set, tags, hctx_idx); |
1804 | return -ENOMEM; | |
320ae51f JA |
1805 | } |
1806 | ||
e57690fe JA |
1807 | /* |
1808 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1809 | * software queue to the hw queue dispatch list, and ensure that it | |
1810 | * gets run. | |
1811 | */ | |
9467f859 | 1812 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1813 | { |
9467f859 | 1814 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1815 | struct blk_mq_ctx *ctx; |
1816 | LIST_HEAD(tmp); | |
1817 | ||
9467f859 | 1818 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1819 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1820 | |
1821 | spin_lock(&ctx->lock); | |
1822 | if (!list_empty(&ctx->rq_list)) { | |
1823 | list_splice_init(&ctx->rq_list, &tmp); | |
1824 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1825 | } | |
1826 | spin_unlock(&ctx->lock); | |
1827 | ||
1828 | if (list_empty(&tmp)) | |
9467f859 | 1829 | return 0; |
484b4061 | 1830 | |
e57690fe JA |
1831 | spin_lock(&hctx->lock); |
1832 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1833 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1834 | |
1835 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1836 | return 0; |
484b4061 JA |
1837 | } |
1838 | ||
9467f859 | 1839 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1840 | { |
9467f859 TG |
1841 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1842 | &hctx->cpuhp_dead); | |
484b4061 JA |
1843 | } |
1844 | ||
c3b4afca | 1845 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1846 | static void blk_mq_exit_hctx(struct request_queue *q, |
1847 | struct blk_mq_tag_set *set, | |
1848 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1849 | { | |
9c1051aa OS |
1850 | blk_mq_debugfs_unregister_hctx(hctx); |
1851 | ||
08e98fc6 ML |
1852 | blk_mq_tag_idle(hctx); |
1853 | ||
f70ced09 | 1854 | if (set->ops->exit_request) |
d6296d39 | 1855 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1856 | |
93252632 OS |
1857 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1858 | ||
08e98fc6 ML |
1859 | if (set->ops->exit_hctx) |
1860 | set->ops->exit_hctx(hctx, hctx_idx); | |
1861 | ||
6a83e74d | 1862 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1863 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1864 | |
9467f859 | 1865 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1866 | blk_free_flush_queue(hctx->fq); |
88459642 | 1867 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1868 | } |
1869 | ||
624dbe47 ML |
1870 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1871 | struct blk_mq_tag_set *set, int nr_queue) | |
1872 | { | |
1873 | struct blk_mq_hw_ctx *hctx; | |
1874 | unsigned int i; | |
1875 | ||
1876 | queue_for_each_hw_ctx(q, hctx, i) { | |
1877 | if (i == nr_queue) | |
1878 | break; | |
08e98fc6 | 1879 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1880 | } |
624dbe47 ML |
1881 | } |
1882 | ||
08e98fc6 ML |
1883 | static int blk_mq_init_hctx(struct request_queue *q, |
1884 | struct blk_mq_tag_set *set, | |
1885 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1886 | { |
08e98fc6 ML |
1887 | int node; |
1888 | ||
1889 | node = hctx->numa_node; | |
1890 | if (node == NUMA_NO_NODE) | |
1891 | node = hctx->numa_node = set->numa_node; | |
1892 | ||
9f993737 | 1893 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1894 | spin_lock_init(&hctx->lock); |
1895 | INIT_LIST_HEAD(&hctx->dispatch); | |
1896 | hctx->queue = q; | |
1897 | hctx->queue_num = hctx_idx; | |
2404e607 | 1898 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1899 | |
9467f859 | 1900 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1901 | |
1902 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1903 | |
1904 | /* | |
08e98fc6 ML |
1905 | * Allocate space for all possible cpus to avoid allocation at |
1906 | * runtime | |
320ae51f | 1907 | */ |
08e98fc6 ML |
1908 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1909 | GFP_KERNEL, node); | |
1910 | if (!hctx->ctxs) | |
1911 | goto unregister_cpu_notifier; | |
320ae51f | 1912 | |
88459642 OS |
1913 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1914 | node)) | |
08e98fc6 | 1915 | goto free_ctxs; |
320ae51f | 1916 | |
08e98fc6 | 1917 | hctx->nr_ctx = 0; |
320ae51f | 1918 | |
08e98fc6 ML |
1919 | if (set->ops->init_hctx && |
1920 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1921 | goto free_bitmap; | |
320ae51f | 1922 | |
93252632 OS |
1923 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
1924 | goto exit_hctx; | |
1925 | ||
f70ced09 ML |
1926 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1927 | if (!hctx->fq) | |
93252632 | 1928 | goto sched_exit_hctx; |
320ae51f | 1929 | |
f70ced09 | 1930 | if (set->ops->init_request && |
d6296d39 CH |
1931 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
1932 | node)) | |
f70ced09 | 1933 | goto free_fq; |
320ae51f | 1934 | |
6a83e74d | 1935 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1936 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1937 | |
9c1051aa OS |
1938 | blk_mq_debugfs_register_hctx(q, hctx); |
1939 | ||
08e98fc6 | 1940 | return 0; |
320ae51f | 1941 | |
f70ced09 ML |
1942 | free_fq: |
1943 | kfree(hctx->fq); | |
93252632 OS |
1944 | sched_exit_hctx: |
1945 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
1946 | exit_hctx: |
1947 | if (set->ops->exit_hctx) | |
1948 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 1949 | free_bitmap: |
88459642 | 1950 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1951 | free_ctxs: |
1952 | kfree(hctx->ctxs); | |
1953 | unregister_cpu_notifier: | |
9467f859 | 1954 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
1955 | return -1; |
1956 | } | |
320ae51f | 1957 | |
320ae51f JA |
1958 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
1959 | unsigned int nr_hw_queues) | |
1960 | { | |
1961 | unsigned int i; | |
1962 | ||
1963 | for_each_possible_cpu(i) { | |
1964 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
1965 | struct blk_mq_hw_ctx *hctx; | |
1966 | ||
320ae51f JA |
1967 | __ctx->cpu = i; |
1968 | spin_lock_init(&__ctx->lock); | |
1969 | INIT_LIST_HEAD(&__ctx->rq_list); | |
1970 | __ctx->queue = q; | |
1971 | ||
1972 | /* If the cpu isn't online, the cpu is mapped to first hctx */ | |
320ae51f JA |
1973 | if (!cpu_online(i)) |
1974 | continue; | |
1975 | ||
7d7e0f90 | 1976 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 1977 | |
320ae51f JA |
1978 | /* |
1979 | * Set local node, IFF we have more than one hw queue. If | |
1980 | * not, we remain on the home node of the device | |
1981 | */ | |
1982 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 1983 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
1984 | } |
1985 | } | |
1986 | ||
cc71a6f4 JA |
1987 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
1988 | { | |
1989 | int ret = 0; | |
1990 | ||
1991 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
1992 | set->queue_depth, set->reserved_tags); | |
1993 | if (!set->tags[hctx_idx]) | |
1994 | return false; | |
1995 | ||
1996 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
1997 | set->queue_depth); | |
1998 | if (!ret) | |
1999 | return true; | |
2000 | ||
2001 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2002 | set->tags[hctx_idx] = NULL; | |
2003 | return false; | |
2004 | } | |
2005 | ||
2006 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2007 | unsigned int hctx_idx) | |
2008 | { | |
bd166ef1 JA |
2009 | if (set->tags[hctx_idx]) { |
2010 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2011 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2012 | set->tags[hctx_idx] = NULL; | |
2013 | } | |
cc71a6f4 JA |
2014 | } |
2015 | ||
5778322e AM |
2016 | static void blk_mq_map_swqueue(struct request_queue *q, |
2017 | const struct cpumask *online_mask) | |
320ae51f | 2018 | { |
d1b1cea1 | 2019 | unsigned int i, hctx_idx; |
320ae51f JA |
2020 | struct blk_mq_hw_ctx *hctx; |
2021 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2022 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2023 | |
60de074b AM |
2024 | /* |
2025 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2026 | */ | |
2027 | mutex_lock(&q->sysfs_lock); | |
2028 | ||
320ae51f | 2029 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2030 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2031 | hctx->nr_ctx = 0; |
2032 | } | |
2033 | ||
2034 | /* | |
2035 | * Map software to hardware queues | |
2036 | */ | |
897bb0c7 | 2037 | for_each_possible_cpu(i) { |
320ae51f | 2038 | /* If the cpu isn't online, the cpu is mapped to first hctx */ |
5778322e | 2039 | if (!cpumask_test_cpu(i, online_mask)) |
e4043dcf JA |
2040 | continue; |
2041 | ||
d1b1cea1 GKB |
2042 | hctx_idx = q->mq_map[i]; |
2043 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2044 | if (!set->tags[hctx_idx] && |
2045 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2046 | /* |
2047 | * If tags initialization fail for some hctx, | |
2048 | * that hctx won't be brought online. In this | |
2049 | * case, remap the current ctx to hctx[0] which | |
2050 | * is guaranteed to always have tags allocated | |
2051 | */ | |
cc71a6f4 | 2052 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2053 | } |
2054 | ||
897bb0c7 | 2055 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2056 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2057 | |
e4043dcf | 2058 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2059 | ctx->index_hw = hctx->nr_ctx; |
2060 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2061 | } | |
506e931f | 2062 | |
60de074b AM |
2063 | mutex_unlock(&q->sysfs_lock); |
2064 | ||
506e931f | 2065 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2066 | /* |
a68aafa5 JA |
2067 | * If no software queues are mapped to this hardware queue, |
2068 | * disable it and free the request entries. | |
484b4061 JA |
2069 | */ |
2070 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2071 | /* Never unmap queue 0. We need it as a |
2072 | * fallback in case of a new remap fails | |
2073 | * allocation | |
2074 | */ | |
cc71a6f4 JA |
2075 | if (i && set->tags[i]) |
2076 | blk_mq_free_map_and_requests(set, i); | |
2077 | ||
2a34c087 | 2078 | hctx->tags = NULL; |
484b4061 JA |
2079 | continue; |
2080 | } | |
2081 | ||
2a34c087 ML |
2082 | hctx->tags = set->tags[i]; |
2083 | WARN_ON(!hctx->tags); | |
2084 | ||
889fa31f CY |
2085 | /* |
2086 | * Set the map size to the number of mapped software queues. | |
2087 | * This is more accurate and more efficient than looping | |
2088 | * over all possibly mapped software queues. | |
2089 | */ | |
88459642 | 2090 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2091 | |
484b4061 JA |
2092 | /* |
2093 | * Initialize batch roundrobin counts | |
2094 | */ | |
506e931f JA |
2095 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2096 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2097 | } | |
320ae51f JA |
2098 | } |
2099 | ||
2404e607 | 2100 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2101 | { |
2102 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2103 | int i; |
2104 | ||
2404e607 JM |
2105 | queue_for_each_hw_ctx(q, hctx, i) { |
2106 | if (shared) | |
2107 | hctx->flags |= BLK_MQ_F_TAG_SHARED; | |
2108 | else | |
2109 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2110 | } | |
2111 | } | |
2112 | ||
2113 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared) | |
2114 | { | |
2115 | struct request_queue *q; | |
0d2602ca | 2116 | |
705cda97 BVA |
2117 | lockdep_assert_held(&set->tag_list_lock); |
2118 | ||
0d2602ca JA |
2119 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2120 | blk_mq_freeze_queue(q); | |
2404e607 | 2121 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2122 | blk_mq_unfreeze_queue(q); |
2123 | } | |
2124 | } | |
2125 | ||
2126 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2127 | { | |
2128 | struct blk_mq_tag_set *set = q->tag_set; | |
2129 | ||
0d2602ca | 2130 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2131 | list_del_rcu(&q->tag_set_list); |
2132 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2133 | if (list_is_singular(&set->tag_list)) { |
2134 | /* just transitioned to unshared */ | |
2135 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2136 | /* update existing queue */ | |
2137 | blk_mq_update_tag_set_depth(set, false); | |
2138 | } | |
0d2602ca | 2139 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2140 | |
2141 | synchronize_rcu(); | |
0d2602ca JA |
2142 | } |
2143 | ||
2144 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2145 | struct request_queue *q) | |
2146 | { | |
2147 | q->tag_set = set; | |
2148 | ||
2149 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2150 | |
2151 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2152 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2153 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2154 | /* update existing queue */ | |
2155 | blk_mq_update_tag_set_depth(set, true); | |
2156 | } | |
2157 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2158 | queue_set_hctx_shared(q, true); | |
705cda97 | 2159 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2160 | |
0d2602ca JA |
2161 | mutex_unlock(&set->tag_list_lock); |
2162 | } | |
2163 | ||
e09aae7e ML |
2164 | /* |
2165 | * It is the actual release handler for mq, but we do it from | |
2166 | * request queue's release handler for avoiding use-after-free | |
2167 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2168 | * but we can't group ctx/kctx kobj without it. | |
2169 | */ | |
2170 | void blk_mq_release(struct request_queue *q) | |
2171 | { | |
2172 | struct blk_mq_hw_ctx *hctx; | |
2173 | unsigned int i; | |
2174 | ||
2175 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2176 | queue_for_each_hw_ctx(q, hctx, i) { |
2177 | if (!hctx) | |
2178 | continue; | |
6c8b232e | 2179 | kobject_put(&hctx->kobj); |
c3b4afca | 2180 | } |
e09aae7e | 2181 | |
a723bab3 AM |
2182 | q->mq_map = NULL; |
2183 | ||
e09aae7e ML |
2184 | kfree(q->queue_hw_ctx); |
2185 | ||
7ea5fe31 ML |
2186 | /* |
2187 | * release .mq_kobj and sw queue's kobject now because | |
2188 | * both share lifetime with request queue. | |
2189 | */ | |
2190 | blk_mq_sysfs_deinit(q); | |
2191 | ||
e09aae7e ML |
2192 | free_percpu(q->queue_ctx); |
2193 | } | |
2194 | ||
24d2f903 | 2195 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2196 | { |
2197 | struct request_queue *uninit_q, *q; | |
2198 | ||
2199 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2200 | if (!uninit_q) | |
2201 | return ERR_PTR(-ENOMEM); | |
2202 | ||
2203 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2204 | if (IS_ERR(q)) | |
2205 | blk_cleanup_queue(uninit_q); | |
2206 | ||
2207 | return q; | |
2208 | } | |
2209 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2210 | ||
07319678 BVA |
2211 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2212 | { | |
2213 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2214 | ||
2215 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2216 | __alignof__(struct blk_mq_hw_ctx)) != | |
2217 | sizeof(struct blk_mq_hw_ctx)); | |
2218 | ||
2219 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2220 | hw_ctx_size += sizeof(struct srcu_struct); | |
2221 | ||
2222 | return hw_ctx_size; | |
2223 | } | |
2224 | ||
868f2f0b KB |
2225 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2226 | struct request_queue *q) | |
320ae51f | 2227 | { |
868f2f0b KB |
2228 | int i, j; |
2229 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2230 | |
868f2f0b | 2231 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2232 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2233 | int node; |
f14bbe77 | 2234 | |
868f2f0b KB |
2235 | if (hctxs[i]) |
2236 | continue; | |
2237 | ||
2238 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2239 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2240 | GFP_KERNEL, node); |
320ae51f | 2241 | if (!hctxs[i]) |
868f2f0b | 2242 | break; |
320ae51f | 2243 | |
a86073e4 | 2244 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2245 | node)) { |
2246 | kfree(hctxs[i]); | |
2247 | hctxs[i] = NULL; | |
2248 | break; | |
2249 | } | |
e4043dcf | 2250 | |
0d2602ca | 2251 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2252 | hctxs[i]->numa_node = node; |
320ae51f | 2253 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2254 | |
2255 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2256 | free_cpumask_var(hctxs[i]->cpumask); | |
2257 | kfree(hctxs[i]); | |
2258 | hctxs[i] = NULL; | |
2259 | break; | |
2260 | } | |
2261 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2262 | } |
868f2f0b KB |
2263 | for (j = i; j < q->nr_hw_queues; j++) { |
2264 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2265 | ||
2266 | if (hctx) { | |
cc71a6f4 JA |
2267 | if (hctx->tags) |
2268 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2269 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2270 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2271 | hctxs[j] = NULL; |
2272 | ||
2273 | } | |
2274 | } | |
2275 | q->nr_hw_queues = i; | |
2276 | blk_mq_sysfs_register(q); | |
2277 | } | |
2278 | ||
2279 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2280 | struct request_queue *q) | |
2281 | { | |
66841672 ML |
2282 | /* mark the queue as mq asap */ |
2283 | q->mq_ops = set->ops; | |
2284 | ||
34dbad5d | 2285 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2286 | blk_mq_poll_stats_bkt, |
2287 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2288 | if (!q->poll_cb) |
2289 | goto err_exit; | |
2290 | ||
868f2f0b KB |
2291 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2292 | if (!q->queue_ctx) | |
c7de5726 | 2293 | goto err_exit; |
868f2f0b | 2294 | |
737f98cf ML |
2295 | /* init q->mq_kobj and sw queues' kobjects */ |
2296 | blk_mq_sysfs_init(q); | |
2297 | ||
868f2f0b KB |
2298 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2299 | GFP_KERNEL, set->numa_node); | |
2300 | if (!q->queue_hw_ctx) | |
2301 | goto err_percpu; | |
2302 | ||
bdd17e75 | 2303 | q->mq_map = set->mq_map; |
868f2f0b KB |
2304 | |
2305 | blk_mq_realloc_hw_ctxs(set, q); | |
2306 | if (!q->nr_hw_queues) | |
2307 | goto err_hctxs; | |
320ae51f | 2308 | |
287922eb | 2309 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2310 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2311 | |
2312 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2313 | |
94eddfbe | 2314 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2315 | |
05f1dd53 JA |
2316 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2317 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2318 | ||
1be036e9 CH |
2319 | q->sg_reserved_size = INT_MAX; |
2320 | ||
2849450a | 2321 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2322 | INIT_LIST_HEAD(&q->requeue_list); |
2323 | spin_lock_init(&q->requeue_lock); | |
2324 | ||
254d259d | 2325 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2326 | |
eba71768 JA |
2327 | /* |
2328 | * Do this after blk_queue_make_request() overrides it... | |
2329 | */ | |
2330 | q->nr_requests = set->queue_depth; | |
2331 | ||
64f1c21e JA |
2332 | /* |
2333 | * Default to classic polling | |
2334 | */ | |
2335 | q->poll_nsec = -1; | |
2336 | ||
24d2f903 CH |
2337 | if (set->ops->complete) |
2338 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2339 | |
24d2f903 | 2340 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
320ae51f | 2341 | |
eabe0659 | 2342 | get_online_cpus(); |
51d638b1 | 2343 | mutex_lock(&all_q_mutex); |
320ae51f | 2344 | |
4593fdbe | 2345 | list_add_tail(&q->all_q_node, &all_q_list); |
0d2602ca | 2346 | blk_mq_add_queue_tag_set(set, q); |
5778322e | 2347 | blk_mq_map_swqueue(q, cpu_online_mask); |
484b4061 | 2348 | |
eabe0659 | 2349 | mutex_unlock(&all_q_mutex); |
51d638b1 | 2350 | put_online_cpus(); |
4593fdbe | 2351 | |
d3484991 JA |
2352 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2353 | int ret; | |
2354 | ||
2355 | ret = blk_mq_sched_init(q); | |
2356 | if (ret) | |
2357 | return ERR_PTR(ret); | |
2358 | } | |
2359 | ||
320ae51f | 2360 | return q; |
18741986 | 2361 | |
320ae51f | 2362 | err_hctxs: |
868f2f0b | 2363 | kfree(q->queue_hw_ctx); |
320ae51f | 2364 | err_percpu: |
868f2f0b | 2365 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2366 | err_exit: |
2367 | q->mq_ops = NULL; | |
320ae51f JA |
2368 | return ERR_PTR(-ENOMEM); |
2369 | } | |
b62c21b7 | 2370 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2371 | |
2372 | void blk_mq_free_queue(struct request_queue *q) | |
2373 | { | |
624dbe47 | 2374 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2375 | |
0e626368 AM |
2376 | mutex_lock(&all_q_mutex); |
2377 | list_del_init(&q->all_q_node); | |
2378 | mutex_unlock(&all_q_mutex); | |
2379 | ||
0d2602ca JA |
2380 | blk_mq_del_queue_tag_set(q); |
2381 | ||
624dbe47 | 2382 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2383 | } |
320ae51f JA |
2384 | |
2385 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
5778322e AM |
2386 | static void blk_mq_queue_reinit(struct request_queue *q, |
2387 | const struct cpumask *online_mask) | |
320ae51f | 2388 | { |
4ecd4fef | 2389 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2390 | |
9c1051aa | 2391 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2392 | blk_mq_sysfs_unregister(q); |
2393 | ||
320ae51f JA |
2394 | /* |
2395 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2396 | * we should change hctx numa_node according to new topology (this | |
2397 | * involves free and re-allocate memory, worthy doing?) | |
2398 | */ | |
2399 | ||
5778322e | 2400 | blk_mq_map_swqueue(q, online_mask); |
320ae51f | 2401 | |
67aec14c | 2402 | blk_mq_sysfs_register(q); |
9c1051aa | 2403 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2404 | } |
2405 | ||
65d5291e SAS |
2406 | /* |
2407 | * New online cpumask which is going to be set in this hotplug event. | |
2408 | * Declare this cpumasks as global as cpu-hotplug operation is invoked | |
2409 | * one-by-one and dynamically allocating this could result in a failure. | |
2410 | */ | |
2411 | static struct cpumask cpuhp_online_new; | |
2412 | ||
2413 | static void blk_mq_queue_reinit_work(void) | |
320ae51f JA |
2414 | { |
2415 | struct request_queue *q; | |
320ae51f JA |
2416 | |
2417 | mutex_lock(&all_q_mutex); | |
f3af020b TH |
2418 | /* |
2419 | * We need to freeze and reinit all existing queues. Freezing | |
2420 | * involves synchronous wait for an RCU grace period and doing it | |
2421 | * one by one may take a long time. Start freezing all queues in | |
2422 | * one swoop and then wait for the completions so that freezing can | |
2423 | * take place in parallel. | |
2424 | */ | |
2425 | list_for_each_entry(q, &all_q_list, all_q_node) | |
1671d522 | 2426 | blk_freeze_queue_start(q); |
415d3dab | 2427 | list_for_each_entry(q, &all_q_list, all_q_node) |
f3af020b TH |
2428 | blk_mq_freeze_queue_wait(q); |
2429 | ||
320ae51f | 2430 | list_for_each_entry(q, &all_q_list, all_q_node) |
65d5291e | 2431 | blk_mq_queue_reinit(q, &cpuhp_online_new); |
f3af020b TH |
2432 | |
2433 | list_for_each_entry(q, &all_q_list, all_q_node) | |
2434 | blk_mq_unfreeze_queue(q); | |
2435 | ||
320ae51f | 2436 | mutex_unlock(&all_q_mutex); |
65d5291e SAS |
2437 | } |
2438 | ||
2439 | static int blk_mq_queue_reinit_dead(unsigned int cpu) | |
2440 | { | |
97a32864 | 2441 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); |
65d5291e SAS |
2442 | blk_mq_queue_reinit_work(); |
2443 | return 0; | |
2444 | } | |
2445 | ||
2446 | /* | |
2447 | * Before hotadded cpu starts handling requests, new mappings must be | |
2448 | * established. Otherwise, these requests in hw queue might never be | |
2449 | * dispatched. | |
2450 | * | |
2451 | * For example, there is a single hw queue (hctx) and two CPU queues (ctx0 | |
2452 | * for CPU0, and ctx1 for CPU1). | |
2453 | * | |
2454 | * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list | |
2455 | * and set bit0 in pending bitmap as ctx1->index_hw is still zero. | |
2456 | * | |
2c3ad667 JA |
2457 | * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set |
2458 | * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list. | |
2459 | * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is | |
2460 | * ignored. | |
65d5291e SAS |
2461 | */ |
2462 | static int blk_mq_queue_reinit_prepare(unsigned int cpu) | |
2463 | { | |
2464 | cpumask_copy(&cpuhp_online_new, cpu_online_mask); | |
2465 | cpumask_set_cpu(cpu, &cpuhp_online_new); | |
2466 | blk_mq_queue_reinit_work(); | |
2467 | return 0; | |
320ae51f JA |
2468 | } |
2469 | ||
a5164405 JA |
2470 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2471 | { | |
2472 | int i; | |
2473 | ||
cc71a6f4 JA |
2474 | for (i = 0; i < set->nr_hw_queues; i++) |
2475 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2476 | goto out_unwind; |
a5164405 JA |
2477 | |
2478 | return 0; | |
2479 | ||
2480 | out_unwind: | |
2481 | while (--i >= 0) | |
cc71a6f4 | 2482 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2483 | |
a5164405 JA |
2484 | return -ENOMEM; |
2485 | } | |
2486 | ||
2487 | /* | |
2488 | * Allocate the request maps associated with this tag_set. Note that this | |
2489 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2490 | * will be updated to reflect the allocated depth. | |
2491 | */ | |
2492 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2493 | { | |
2494 | unsigned int depth; | |
2495 | int err; | |
2496 | ||
2497 | depth = set->queue_depth; | |
2498 | do { | |
2499 | err = __blk_mq_alloc_rq_maps(set); | |
2500 | if (!err) | |
2501 | break; | |
2502 | ||
2503 | set->queue_depth >>= 1; | |
2504 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2505 | err = -ENOMEM; | |
2506 | break; | |
2507 | } | |
2508 | } while (set->queue_depth); | |
2509 | ||
2510 | if (!set->queue_depth || err) { | |
2511 | pr_err("blk-mq: failed to allocate request map\n"); | |
2512 | return -ENOMEM; | |
2513 | } | |
2514 | ||
2515 | if (depth != set->queue_depth) | |
2516 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2517 | depth, set->queue_depth); | |
2518 | ||
2519 | return 0; | |
2520 | } | |
2521 | ||
ebe8bddb OS |
2522 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2523 | { | |
2524 | if (set->ops->map_queues) | |
2525 | return set->ops->map_queues(set); | |
2526 | else | |
2527 | return blk_mq_map_queues(set); | |
2528 | } | |
2529 | ||
a4391c64 JA |
2530 | /* |
2531 | * Alloc a tag set to be associated with one or more request queues. | |
2532 | * May fail with EINVAL for various error conditions. May adjust the | |
2533 | * requested depth down, if if it too large. In that case, the set | |
2534 | * value will be stored in set->queue_depth. | |
2535 | */ | |
24d2f903 CH |
2536 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2537 | { | |
da695ba2 CH |
2538 | int ret; |
2539 | ||
205fb5f5 BVA |
2540 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2541 | ||
24d2f903 CH |
2542 | if (!set->nr_hw_queues) |
2543 | return -EINVAL; | |
a4391c64 | 2544 | if (!set->queue_depth) |
24d2f903 CH |
2545 | return -EINVAL; |
2546 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2547 | return -EINVAL; | |
2548 | ||
7d7e0f90 | 2549 | if (!set->ops->queue_rq) |
24d2f903 CH |
2550 | return -EINVAL; |
2551 | ||
a4391c64 JA |
2552 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2553 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2554 | BLK_MQ_MAX_DEPTH); | |
2555 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2556 | } | |
24d2f903 | 2557 | |
6637fadf SL |
2558 | /* |
2559 | * If a crashdump is active, then we are potentially in a very | |
2560 | * memory constrained environment. Limit us to 1 queue and | |
2561 | * 64 tags to prevent using too much memory. | |
2562 | */ | |
2563 | if (is_kdump_kernel()) { | |
2564 | set->nr_hw_queues = 1; | |
2565 | set->queue_depth = min(64U, set->queue_depth); | |
2566 | } | |
868f2f0b KB |
2567 | /* |
2568 | * There is no use for more h/w queues than cpus. | |
2569 | */ | |
2570 | if (set->nr_hw_queues > nr_cpu_ids) | |
2571 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2572 | |
868f2f0b | 2573 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2574 | GFP_KERNEL, set->numa_node); |
2575 | if (!set->tags) | |
a5164405 | 2576 | return -ENOMEM; |
24d2f903 | 2577 | |
da695ba2 CH |
2578 | ret = -ENOMEM; |
2579 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2580 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2581 | if (!set->mq_map) |
2582 | goto out_free_tags; | |
2583 | ||
ebe8bddb | 2584 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2585 | if (ret) |
2586 | goto out_free_mq_map; | |
2587 | ||
2588 | ret = blk_mq_alloc_rq_maps(set); | |
2589 | if (ret) | |
bdd17e75 | 2590 | goto out_free_mq_map; |
24d2f903 | 2591 | |
0d2602ca JA |
2592 | mutex_init(&set->tag_list_lock); |
2593 | INIT_LIST_HEAD(&set->tag_list); | |
2594 | ||
24d2f903 | 2595 | return 0; |
bdd17e75 CH |
2596 | |
2597 | out_free_mq_map: | |
2598 | kfree(set->mq_map); | |
2599 | set->mq_map = NULL; | |
2600 | out_free_tags: | |
5676e7b6 RE |
2601 | kfree(set->tags); |
2602 | set->tags = NULL; | |
da695ba2 | 2603 | return ret; |
24d2f903 CH |
2604 | } |
2605 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2606 | ||
2607 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2608 | { | |
2609 | int i; | |
2610 | ||
cc71a6f4 JA |
2611 | for (i = 0; i < nr_cpu_ids; i++) |
2612 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2613 | |
bdd17e75 CH |
2614 | kfree(set->mq_map); |
2615 | set->mq_map = NULL; | |
2616 | ||
981bd189 | 2617 | kfree(set->tags); |
5676e7b6 | 2618 | set->tags = NULL; |
24d2f903 CH |
2619 | } |
2620 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2621 | ||
e3a2b3f9 JA |
2622 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2623 | { | |
2624 | struct blk_mq_tag_set *set = q->tag_set; | |
2625 | struct blk_mq_hw_ctx *hctx; | |
2626 | int i, ret; | |
2627 | ||
bd166ef1 | 2628 | if (!set) |
e3a2b3f9 JA |
2629 | return -EINVAL; |
2630 | ||
70f36b60 | 2631 | blk_mq_freeze_queue(q); |
70f36b60 | 2632 | |
e3a2b3f9 JA |
2633 | ret = 0; |
2634 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2635 | if (!hctx->tags) |
2636 | continue; | |
bd166ef1 JA |
2637 | /* |
2638 | * If we're using an MQ scheduler, just update the scheduler | |
2639 | * queue depth. This is similar to what the old code would do. | |
2640 | */ | |
70f36b60 JA |
2641 | if (!hctx->sched_tags) { |
2642 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2643 | min(nr, set->queue_depth), | |
2644 | false); | |
2645 | } else { | |
2646 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2647 | nr, true); | |
2648 | } | |
e3a2b3f9 JA |
2649 | if (ret) |
2650 | break; | |
2651 | } | |
2652 | ||
2653 | if (!ret) | |
2654 | q->nr_requests = nr; | |
2655 | ||
70f36b60 | 2656 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2657 | |
e3a2b3f9 JA |
2658 | return ret; |
2659 | } | |
2660 | ||
e4dc2b32 KB |
2661 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2662 | int nr_hw_queues) | |
868f2f0b KB |
2663 | { |
2664 | struct request_queue *q; | |
2665 | ||
705cda97 BVA |
2666 | lockdep_assert_held(&set->tag_list_lock); |
2667 | ||
868f2f0b KB |
2668 | if (nr_hw_queues > nr_cpu_ids) |
2669 | nr_hw_queues = nr_cpu_ids; | |
2670 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2671 | return; | |
2672 | ||
2673 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2674 | blk_mq_freeze_queue(q); | |
2675 | ||
2676 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2677 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2678 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2679 | blk_mq_realloc_hw_ctxs(set, q); | |
868f2f0b KB |
2680 | blk_mq_queue_reinit(q, cpu_online_mask); |
2681 | } | |
2682 | ||
2683 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2684 | blk_mq_unfreeze_queue(q); | |
2685 | } | |
e4dc2b32 KB |
2686 | |
2687 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2688 | { | |
2689 | mutex_lock(&set->tag_list_lock); | |
2690 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2691 | mutex_unlock(&set->tag_list_lock); | |
2692 | } | |
868f2f0b KB |
2693 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2694 | ||
34dbad5d OS |
2695 | /* Enable polling stats and return whether they were already enabled. */ |
2696 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2697 | { | |
2698 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2699 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2700 | return true; | |
2701 | blk_stat_add_callback(q, q->poll_cb); | |
2702 | return false; | |
2703 | } | |
2704 | ||
2705 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2706 | { | |
2707 | /* | |
2708 | * We don't arm the callback if polling stats are not enabled or the | |
2709 | * callback is already active. | |
2710 | */ | |
2711 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2712 | blk_stat_is_active(q->poll_cb)) | |
2713 | return; | |
2714 | ||
2715 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2716 | } | |
2717 | ||
2718 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2719 | { | |
2720 | struct request_queue *q = cb->data; | |
720b8ccc | 2721 | int bucket; |
34dbad5d | 2722 | |
720b8ccc SB |
2723 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2724 | if (cb->stat[bucket].nr_samples) | |
2725 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2726 | } | |
34dbad5d OS |
2727 | } |
2728 | ||
64f1c21e JA |
2729 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2730 | struct blk_mq_hw_ctx *hctx, | |
2731 | struct request *rq) | |
2732 | { | |
64f1c21e | 2733 | unsigned long ret = 0; |
720b8ccc | 2734 | int bucket; |
64f1c21e JA |
2735 | |
2736 | /* | |
2737 | * If stats collection isn't on, don't sleep but turn it on for | |
2738 | * future users | |
2739 | */ | |
34dbad5d | 2740 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2741 | return 0; |
2742 | ||
64f1c21e JA |
2743 | /* |
2744 | * As an optimistic guess, use half of the mean service time | |
2745 | * for this type of request. We can (and should) make this smarter. | |
2746 | * For instance, if the completion latencies are tight, we can | |
2747 | * get closer than just half the mean. This is especially | |
2748 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2749 | * than ~10 usec. We do use the stats for the relevant IO size |
2750 | * if available which does lead to better estimates. | |
64f1c21e | 2751 | */ |
720b8ccc SB |
2752 | bucket = blk_mq_poll_stats_bkt(rq); |
2753 | if (bucket < 0) | |
2754 | return ret; | |
2755 | ||
2756 | if (q->poll_stat[bucket].nr_samples) | |
2757 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2758 | |
2759 | return ret; | |
2760 | } | |
2761 | ||
06426adf | 2762 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2763 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2764 | struct request *rq) |
2765 | { | |
2766 | struct hrtimer_sleeper hs; | |
2767 | enum hrtimer_mode mode; | |
64f1c21e | 2768 | unsigned int nsecs; |
06426adf JA |
2769 | ktime_t kt; |
2770 | ||
64f1c21e JA |
2771 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2772 | return false; | |
2773 | ||
2774 | /* | |
2775 | * poll_nsec can be: | |
2776 | * | |
2777 | * -1: don't ever hybrid sleep | |
2778 | * 0: use half of prev avg | |
2779 | * >0: use this specific value | |
2780 | */ | |
2781 | if (q->poll_nsec == -1) | |
2782 | return false; | |
2783 | else if (q->poll_nsec > 0) | |
2784 | nsecs = q->poll_nsec; | |
2785 | else | |
2786 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2787 | ||
2788 | if (!nsecs) | |
06426adf JA |
2789 | return false; |
2790 | ||
2791 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2792 | ||
2793 | /* | |
2794 | * This will be replaced with the stats tracking code, using | |
2795 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2796 | */ | |
8b0e1953 | 2797 | kt = nsecs; |
06426adf JA |
2798 | |
2799 | mode = HRTIMER_MODE_REL; | |
2800 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2801 | hrtimer_set_expires(&hs.timer, kt); | |
2802 | ||
2803 | hrtimer_init_sleeper(&hs, current); | |
2804 | do { | |
2805 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2806 | break; | |
2807 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2808 | hrtimer_start_expires(&hs.timer, mode); | |
2809 | if (hs.task) | |
2810 | io_schedule(); | |
2811 | hrtimer_cancel(&hs.timer); | |
2812 | mode = HRTIMER_MODE_ABS; | |
2813 | } while (hs.task && !signal_pending(current)); | |
2814 | ||
2815 | __set_current_state(TASK_RUNNING); | |
2816 | destroy_hrtimer_on_stack(&hs.timer); | |
2817 | return true; | |
2818 | } | |
2819 | ||
bbd7bb70 JA |
2820 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2821 | { | |
2822 | struct request_queue *q = hctx->queue; | |
2823 | long state; | |
2824 | ||
06426adf JA |
2825 | /* |
2826 | * If we sleep, have the caller restart the poll loop to reset | |
2827 | * the state. Like for the other success return cases, the | |
2828 | * caller is responsible for checking if the IO completed. If | |
2829 | * the IO isn't complete, we'll get called again and will go | |
2830 | * straight to the busy poll loop. | |
2831 | */ | |
64f1c21e | 2832 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2833 | return true; |
2834 | ||
bbd7bb70 JA |
2835 | hctx->poll_considered++; |
2836 | ||
2837 | state = current->state; | |
2838 | while (!need_resched()) { | |
2839 | int ret; | |
2840 | ||
2841 | hctx->poll_invoked++; | |
2842 | ||
2843 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2844 | if (ret > 0) { | |
2845 | hctx->poll_success++; | |
2846 | set_current_state(TASK_RUNNING); | |
2847 | return true; | |
2848 | } | |
2849 | ||
2850 | if (signal_pending_state(state, current)) | |
2851 | set_current_state(TASK_RUNNING); | |
2852 | ||
2853 | if (current->state == TASK_RUNNING) | |
2854 | return true; | |
2855 | if (ret < 0) | |
2856 | break; | |
2857 | cpu_relax(); | |
2858 | } | |
2859 | ||
2860 | return false; | |
2861 | } | |
2862 | ||
2863 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2864 | { | |
2865 | struct blk_mq_hw_ctx *hctx; | |
2866 | struct blk_plug *plug; | |
2867 | struct request *rq; | |
2868 | ||
2869 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2870 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2871 | return false; | |
2872 | ||
2873 | plug = current->plug; | |
2874 | if (plug) | |
2875 | blk_flush_plug_list(plug, false); | |
2876 | ||
2877 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2878 | if (!blk_qc_t_is_internal(cookie)) |
2879 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2880 | else { |
bd166ef1 | 2881 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
2882 | /* |
2883 | * With scheduling, if the request has completed, we'll | |
2884 | * get a NULL return here, as we clear the sched tag when | |
2885 | * that happens. The request still remains valid, like always, | |
2886 | * so we should be safe with just the NULL check. | |
2887 | */ | |
2888 | if (!rq) | |
2889 | return false; | |
2890 | } | |
bbd7bb70 JA |
2891 | |
2892 | return __blk_mq_poll(hctx, rq); | |
2893 | } | |
2894 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2895 | ||
676141e4 JA |
2896 | void blk_mq_disable_hotplug(void) |
2897 | { | |
2898 | mutex_lock(&all_q_mutex); | |
2899 | } | |
2900 | ||
2901 | void blk_mq_enable_hotplug(void) | |
2902 | { | |
2903 | mutex_unlock(&all_q_mutex); | |
2904 | } | |
2905 | ||
320ae51f JA |
2906 | static int __init blk_mq_init(void) |
2907 | { | |
9467f859 TG |
2908 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2909 | blk_mq_hctx_notify_dead); | |
320ae51f | 2910 | |
65d5291e SAS |
2911 | cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare", |
2912 | blk_mq_queue_reinit_prepare, | |
2913 | blk_mq_queue_reinit_dead); | |
320ae51f JA |
2914 | return 0; |
2915 | } | |
2916 | subsys_initcall(blk_mq_init); |