Commit | Line | Data |
---|---|---|
75bb4625 JA |
1 | /* |
2 | * Block multiqueue core code | |
3 | * | |
4 | * Copyright (C) 2013-2014 Jens Axboe | |
5 | * Copyright (C) 2013-2014 Christoph Hellwig | |
6 | */ | |
320ae51f JA |
7 | #include <linux/kernel.h> |
8 | #include <linux/module.h> | |
9 | #include <linux/backing-dev.h> | |
10 | #include <linux/bio.h> | |
11 | #include <linux/blkdev.h> | |
f75782e4 | 12 | #include <linux/kmemleak.h> |
320ae51f JA |
13 | #include <linux/mm.h> |
14 | #include <linux/init.h> | |
15 | #include <linux/slab.h> | |
16 | #include <linux/workqueue.h> | |
17 | #include <linux/smp.h> | |
18 | #include <linux/llist.h> | |
19 | #include <linux/list_sort.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/cache.h> | |
22 | #include <linux/sched/sysctl.h> | |
105ab3d8 | 23 | #include <linux/sched/topology.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
320ae51f | 25 | #include <linux/delay.h> |
aedcd72f | 26 | #include <linux/crash_dump.h> |
88c7b2b7 | 27 | #include <linux/prefetch.h> |
320ae51f JA |
28 | |
29 | #include <trace/events/block.h> | |
30 | ||
31 | #include <linux/blk-mq.h> | |
32 | #include "blk.h" | |
33 | #include "blk-mq.h" | |
9c1051aa | 34 | #include "blk-mq-debugfs.h" |
320ae51f | 35 | #include "blk-mq-tag.h" |
cf43e6be | 36 | #include "blk-stat.h" |
87760e5e | 37 | #include "blk-wbt.h" |
bd166ef1 | 38 | #include "blk-mq-sched.h" |
320ae51f | 39 | |
34dbad5d OS |
40 | static void blk_mq_poll_stats_start(struct request_queue *q); |
41 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
42 | ||
720b8ccc SB |
43 | static int blk_mq_poll_stats_bkt(const struct request *rq) |
44 | { | |
45 | int ddir, bytes, bucket; | |
46 | ||
99c749a4 | 47 | ddir = rq_data_dir(rq); |
720b8ccc SB |
48 | bytes = blk_rq_bytes(rq); |
49 | ||
50 | bucket = ddir + 2*(ilog2(bytes) - 9); | |
51 | ||
52 | if (bucket < 0) | |
53 | return -1; | |
54 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
55 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
56 | ||
57 | return bucket; | |
58 | } | |
59 | ||
320ae51f JA |
60 | /* |
61 | * Check if any of the ctx's have pending work in this hardware queue | |
62 | */ | |
50e1dab8 | 63 | bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) |
320ae51f | 64 | { |
bd166ef1 JA |
65 | return sbitmap_any_bit_set(&hctx->ctx_map) || |
66 | !list_empty_careful(&hctx->dispatch) || | |
67 | blk_mq_sched_has_work(hctx); | |
1429d7c9 JA |
68 | } |
69 | ||
320ae51f JA |
70 | /* |
71 | * Mark this ctx as having pending work in this hardware queue | |
72 | */ | |
73 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
74 | struct blk_mq_ctx *ctx) | |
75 | { | |
88459642 OS |
76 | if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw)) |
77 | sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw); | |
1429d7c9 JA |
78 | } |
79 | ||
80 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
88459642 | 83 | sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw); |
320ae51f JA |
84 | } |
85 | ||
f299b7c7 JA |
86 | struct mq_inflight { |
87 | struct hd_struct *part; | |
88 | unsigned int *inflight; | |
89 | }; | |
90 | ||
91 | static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
92 | struct request *rq, void *priv, | |
93 | bool reserved) | |
94 | { | |
95 | struct mq_inflight *mi = priv; | |
96 | ||
97 | if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) && | |
98 | !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
99 | /* | |
b8d62b3a JA |
100 | * index[0] counts the specific partition that was asked |
101 | * for. index[1] counts the ones that are active on the | |
102 | * whole device, so increment that if mi->part is indeed | |
103 | * a partition, and not a whole device. | |
f299b7c7 | 104 | */ |
b8d62b3a | 105 | if (rq->part == mi->part) |
f299b7c7 | 106 | mi->inflight[0]++; |
b8d62b3a JA |
107 | if (mi->part->partno) |
108 | mi->inflight[1]++; | |
f299b7c7 JA |
109 | } |
110 | } | |
111 | ||
112 | void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part, | |
113 | unsigned int inflight[2]) | |
114 | { | |
115 | struct mq_inflight mi = { .part = part, .inflight = inflight, }; | |
116 | ||
b8d62b3a | 117 | inflight[0] = inflight[1] = 0; |
f299b7c7 JA |
118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); |
119 | } | |
120 | ||
1671d522 | 121 | void blk_freeze_queue_start(struct request_queue *q) |
43a5e4e2 | 122 | { |
4ecd4fef | 123 | int freeze_depth; |
cddd5d17 | 124 | |
4ecd4fef CH |
125 | freeze_depth = atomic_inc_return(&q->mq_freeze_depth); |
126 | if (freeze_depth == 1) { | |
3ef28e83 | 127 | percpu_ref_kill(&q->q_usage_counter); |
b94ec296 | 128 | blk_mq_run_hw_queues(q, false); |
cddd5d17 | 129 | } |
f3af020b | 130 | } |
1671d522 | 131 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); |
f3af020b | 132 | |
6bae363e | 133 | void blk_mq_freeze_queue_wait(struct request_queue *q) |
f3af020b | 134 | { |
3ef28e83 | 135 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); |
43a5e4e2 | 136 | } |
6bae363e | 137 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); |
43a5e4e2 | 138 | |
f91328c4 KB |
139 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, |
140 | unsigned long timeout) | |
141 | { | |
142 | return wait_event_timeout(q->mq_freeze_wq, | |
143 | percpu_ref_is_zero(&q->q_usage_counter), | |
144 | timeout); | |
145 | } | |
146 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
43a5e4e2 | 147 | |
f3af020b TH |
148 | /* |
149 | * Guarantee no request is in use, so we can change any data structure of | |
150 | * the queue afterward. | |
151 | */ | |
3ef28e83 | 152 | void blk_freeze_queue(struct request_queue *q) |
f3af020b | 153 | { |
3ef28e83 DW |
154 | /* |
155 | * In the !blk_mq case we are only calling this to kill the | |
156 | * q_usage_counter, otherwise this increases the freeze depth | |
157 | * and waits for it to return to zero. For this reason there is | |
158 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
159 | * exported to drivers as the only user for unfreeze is blk_mq. | |
160 | */ | |
1671d522 | 161 | blk_freeze_queue_start(q); |
f3af020b TH |
162 | blk_mq_freeze_queue_wait(q); |
163 | } | |
3ef28e83 DW |
164 | |
165 | void blk_mq_freeze_queue(struct request_queue *q) | |
166 | { | |
167 | /* | |
168 | * ...just an alias to keep freeze and unfreeze actions balanced | |
169 | * in the blk_mq_* namespace | |
170 | */ | |
171 | blk_freeze_queue(q); | |
172 | } | |
c761d96b | 173 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); |
f3af020b | 174 | |
b4c6a028 | 175 | void blk_mq_unfreeze_queue(struct request_queue *q) |
320ae51f | 176 | { |
4ecd4fef | 177 | int freeze_depth; |
320ae51f | 178 | |
4ecd4fef CH |
179 | freeze_depth = atomic_dec_return(&q->mq_freeze_depth); |
180 | WARN_ON_ONCE(freeze_depth < 0); | |
181 | if (!freeze_depth) { | |
3ef28e83 | 182 | percpu_ref_reinit(&q->q_usage_counter); |
320ae51f | 183 | wake_up_all(&q->mq_freeze_wq); |
add703fd | 184 | } |
320ae51f | 185 | } |
b4c6a028 | 186 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); |
320ae51f | 187 | |
852ec809 BVA |
188 | /* |
189 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
190 | * mpt3sas driver such that this function can be removed. | |
191 | */ | |
192 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
193 | { | |
194 | unsigned long flags; | |
195 | ||
196 | spin_lock_irqsave(q->queue_lock, flags); | |
197 | queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
198 | spin_unlock_irqrestore(q->queue_lock, flags); | |
199 | } | |
200 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
201 | ||
6a83e74d | 202 | /** |
69e07c4a | 203 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished |
6a83e74d BVA |
204 | * @q: request queue. |
205 | * | |
206 | * Note: this function does not prevent that the struct request end_io() | |
69e07c4a ML |
207 | * callback function is invoked. Once this function is returned, we make |
208 | * sure no dispatch can happen until the queue is unquiesced via | |
209 | * blk_mq_unquiesce_queue(). | |
6a83e74d BVA |
210 | */ |
211 | void blk_mq_quiesce_queue(struct request_queue *q) | |
212 | { | |
213 | struct blk_mq_hw_ctx *hctx; | |
214 | unsigned int i; | |
215 | bool rcu = false; | |
216 | ||
1d9e9bc6 | 217 | blk_mq_quiesce_queue_nowait(q); |
f4560ffe | 218 | |
6a83e74d BVA |
219 | queue_for_each_hw_ctx(q, hctx, i) { |
220 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
07319678 | 221 | synchronize_srcu(hctx->queue_rq_srcu); |
6a83e74d BVA |
222 | else |
223 | rcu = true; | |
224 | } | |
225 | if (rcu) | |
226 | synchronize_rcu(); | |
227 | } | |
228 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
229 | ||
e4e73913 ML |
230 | /* |
231 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
232 | * @q: request queue. | |
233 | * | |
234 | * This function recovers queue into the state before quiescing | |
235 | * which is done by blk_mq_quiesce_queue. | |
236 | */ | |
237 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
238 | { | |
852ec809 BVA |
239 | unsigned long flags; |
240 | ||
241 | spin_lock_irqsave(q->queue_lock, flags); | |
f4560ffe | 242 | queue_flag_clear(QUEUE_FLAG_QUIESCED, q); |
852ec809 | 243 | spin_unlock_irqrestore(q->queue_lock, flags); |
f4560ffe | 244 | |
1d9e9bc6 ML |
245 | /* dispatch requests which are inserted during quiescing */ |
246 | blk_mq_run_hw_queues(q, true); | |
e4e73913 ML |
247 | } |
248 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
249 | ||
aed3ea94 JA |
250 | void blk_mq_wake_waiters(struct request_queue *q) |
251 | { | |
252 | struct blk_mq_hw_ctx *hctx; | |
253 | unsigned int i; | |
254 | ||
255 | queue_for_each_hw_ctx(q, hctx, i) | |
256 | if (blk_mq_hw_queue_mapped(hctx)) | |
257 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
3fd5940c KB |
258 | |
259 | /* | |
260 | * If we are called because the queue has now been marked as | |
261 | * dying, we need to ensure that processes currently waiting on | |
262 | * the queue are notified as well. | |
263 | */ | |
264 | wake_up_all(&q->mq_freeze_wq); | |
aed3ea94 JA |
265 | } |
266 | ||
320ae51f JA |
267 | bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx) |
268 | { | |
269 | return blk_mq_has_free_tags(hctx->tags); | |
270 | } | |
271 | EXPORT_SYMBOL(blk_mq_can_queue); | |
272 | ||
e4cdf1a1 CH |
273 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, |
274 | unsigned int tag, unsigned int op) | |
320ae51f | 275 | { |
e4cdf1a1 CH |
276 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); |
277 | struct request *rq = tags->static_rqs[tag]; | |
278 | ||
c3a148d2 BVA |
279 | rq->rq_flags = 0; |
280 | ||
e4cdf1a1 CH |
281 | if (data->flags & BLK_MQ_REQ_INTERNAL) { |
282 | rq->tag = -1; | |
283 | rq->internal_tag = tag; | |
284 | } else { | |
285 | if (blk_mq_tag_busy(data->hctx)) { | |
286 | rq->rq_flags = RQF_MQ_INFLIGHT; | |
287 | atomic_inc(&data->hctx->nr_active); | |
288 | } | |
289 | rq->tag = tag; | |
290 | rq->internal_tag = -1; | |
291 | data->hctx->tags->rqs[rq->tag] = rq; | |
292 | } | |
293 | ||
af76e555 CH |
294 | INIT_LIST_HEAD(&rq->queuelist); |
295 | /* csd/requeue_work/fifo_time is initialized before use */ | |
e4cdf1a1 CH |
296 | rq->q = data->q; |
297 | rq->mq_ctx = data->ctx; | |
ef295ecf | 298 | rq->cmd_flags = op; |
e4cdf1a1 | 299 | if (blk_queue_io_stat(data->q)) |
e8064021 | 300 | rq->rq_flags |= RQF_IO_STAT; |
af76e555 CH |
301 | /* do not touch atomic flags, it needs atomic ops against the timer */ |
302 | rq->cpu = -1; | |
af76e555 CH |
303 | INIT_HLIST_NODE(&rq->hash); |
304 | RB_CLEAR_NODE(&rq->rb_node); | |
af76e555 CH |
305 | rq->rq_disk = NULL; |
306 | rq->part = NULL; | |
3ee32372 | 307 | rq->start_time = jiffies; |
af76e555 CH |
308 | #ifdef CONFIG_BLK_CGROUP |
309 | rq->rl = NULL; | |
0fec08b4 | 310 | set_start_time_ns(rq); |
af76e555 CH |
311 | rq->io_start_time_ns = 0; |
312 | #endif | |
313 | rq->nr_phys_segments = 0; | |
314 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
315 | rq->nr_integrity_segments = 0; | |
316 | #endif | |
af76e555 CH |
317 | rq->special = NULL; |
318 | /* tag was already set */ | |
af76e555 | 319 | rq->extra_len = 0; |
af76e555 | 320 | |
af76e555 | 321 | INIT_LIST_HEAD(&rq->timeout_list); |
f6be4fb4 JA |
322 | rq->timeout = 0; |
323 | ||
af76e555 CH |
324 | rq->end_io = NULL; |
325 | rq->end_io_data = NULL; | |
326 | rq->next_rq = NULL; | |
327 | ||
e4cdf1a1 CH |
328 | data->ctx->rq_dispatched[op_is_sync(op)]++; |
329 | return rq; | |
5dee8577 CH |
330 | } |
331 | ||
d2c0d383 CH |
332 | static struct request *blk_mq_get_request(struct request_queue *q, |
333 | struct bio *bio, unsigned int op, | |
334 | struct blk_mq_alloc_data *data) | |
335 | { | |
336 | struct elevator_queue *e = q->elevator; | |
337 | struct request *rq; | |
e4cdf1a1 | 338 | unsigned int tag; |
1ad43c00 | 339 | struct blk_mq_ctx *local_ctx = NULL; |
d2c0d383 CH |
340 | |
341 | blk_queue_enter_live(q); | |
342 | data->q = q; | |
343 | if (likely(!data->ctx)) | |
1ad43c00 | 344 | data->ctx = local_ctx = blk_mq_get_ctx(q); |
d2c0d383 CH |
345 | if (likely(!data->hctx)) |
346 | data->hctx = blk_mq_map_queue(q, data->ctx->cpu); | |
03a07c92 GR |
347 | if (op & REQ_NOWAIT) |
348 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
d2c0d383 CH |
349 | |
350 | if (e) { | |
351 | data->flags |= BLK_MQ_REQ_INTERNAL; | |
352 | ||
353 | /* | |
354 | * Flush requests are special and go directly to the | |
355 | * dispatch list. | |
356 | */ | |
5bbf4e5a CH |
357 | if (!op_is_flush(op) && e->type->ops.mq.limit_depth) |
358 | e->type->ops.mq.limit_depth(op, data); | |
d2c0d383 CH |
359 | } |
360 | ||
e4cdf1a1 CH |
361 | tag = blk_mq_get_tag(data); |
362 | if (tag == BLK_MQ_TAG_FAIL) { | |
1ad43c00 ML |
363 | if (local_ctx) { |
364 | blk_mq_put_ctx(local_ctx); | |
365 | data->ctx = NULL; | |
366 | } | |
037cebb8 CH |
367 | blk_queue_exit(q); |
368 | return NULL; | |
d2c0d383 CH |
369 | } |
370 | ||
e4cdf1a1 | 371 | rq = blk_mq_rq_ctx_init(data, tag, op); |
037cebb8 CH |
372 | if (!op_is_flush(op)) { |
373 | rq->elv.icq = NULL; | |
5bbf4e5a | 374 | if (e && e->type->ops.mq.prepare_request) { |
44e8c2bf CH |
375 | if (e->type->icq_cache && rq_ioc(bio)) |
376 | blk_mq_sched_assign_ioc(rq, bio); | |
377 | ||
5bbf4e5a CH |
378 | e->type->ops.mq.prepare_request(rq, bio); |
379 | rq->rq_flags |= RQF_ELVPRIV; | |
44e8c2bf | 380 | } |
037cebb8 CH |
381 | } |
382 | data->hctx->queued++; | |
383 | return rq; | |
d2c0d383 CH |
384 | } |
385 | ||
cd6ce148 | 386 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, |
6f3b0e8b | 387 | unsigned int flags) |
320ae51f | 388 | { |
5a797e00 | 389 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
bd166ef1 | 390 | struct request *rq; |
a492f075 | 391 | int ret; |
320ae51f | 392 | |
6f3b0e8b | 393 | ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT); |
a492f075 JL |
394 | if (ret) |
395 | return ERR_PTR(ret); | |
320ae51f | 396 | |
cd6ce148 | 397 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 398 | blk_queue_exit(q); |
841bac2c | 399 | |
bd166ef1 | 400 | if (!rq) |
a492f075 | 401 | return ERR_PTR(-EWOULDBLOCK); |
0c4de0f3 | 402 | |
1ad43c00 | 403 | blk_mq_put_ctx(alloc_data.ctx); |
1ad43c00 | 404 | |
0c4de0f3 CH |
405 | rq->__data_len = 0; |
406 | rq->__sector = (sector_t) -1; | |
407 | rq->bio = rq->biotail = NULL; | |
320ae51f JA |
408 | return rq; |
409 | } | |
4bb659b1 | 410 | EXPORT_SYMBOL(blk_mq_alloc_request); |
320ae51f | 411 | |
cd6ce148 BVA |
412 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, |
413 | unsigned int op, unsigned int flags, unsigned int hctx_idx) | |
1f5bd336 | 414 | { |
6d2809d5 | 415 | struct blk_mq_alloc_data alloc_data = { .flags = flags }; |
1f5bd336 | 416 | struct request *rq; |
6d2809d5 | 417 | unsigned int cpu; |
1f5bd336 ML |
418 | int ret; |
419 | ||
420 | /* | |
421 | * If the tag allocator sleeps we could get an allocation for a | |
422 | * different hardware context. No need to complicate the low level | |
423 | * allocator for this for the rare use case of a command tied to | |
424 | * a specific queue. | |
425 | */ | |
426 | if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT))) | |
427 | return ERR_PTR(-EINVAL); | |
428 | ||
429 | if (hctx_idx >= q->nr_hw_queues) | |
430 | return ERR_PTR(-EIO); | |
431 | ||
432 | ret = blk_queue_enter(q, true); | |
433 | if (ret) | |
434 | return ERR_PTR(ret); | |
435 | ||
c8712c6a CH |
436 | /* |
437 | * Check if the hardware context is actually mapped to anything. | |
438 | * If not tell the caller that it should skip this queue. | |
439 | */ | |
6d2809d5 OS |
440 | alloc_data.hctx = q->queue_hw_ctx[hctx_idx]; |
441 | if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) { | |
442 | blk_queue_exit(q); | |
443 | return ERR_PTR(-EXDEV); | |
c8712c6a | 444 | } |
6d2809d5 OS |
445 | cpu = cpumask_first(alloc_data.hctx->cpumask); |
446 | alloc_data.ctx = __blk_mq_get_ctx(q, cpu); | |
1f5bd336 | 447 | |
cd6ce148 | 448 | rq = blk_mq_get_request(q, NULL, op, &alloc_data); |
3280d66a | 449 | blk_queue_exit(q); |
c8712c6a | 450 | |
6d2809d5 OS |
451 | if (!rq) |
452 | return ERR_PTR(-EWOULDBLOCK); | |
453 | ||
454 | return rq; | |
1f5bd336 ML |
455 | } |
456 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
457 | ||
6af54051 | 458 | void blk_mq_free_request(struct request *rq) |
320ae51f | 459 | { |
320ae51f | 460 | struct request_queue *q = rq->q; |
6af54051 CH |
461 | struct elevator_queue *e = q->elevator; |
462 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
463 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu); | |
464 | const int sched_tag = rq->internal_tag; | |
465 | ||
5bbf4e5a | 466 | if (rq->rq_flags & RQF_ELVPRIV) { |
6af54051 CH |
467 | if (e && e->type->ops.mq.finish_request) |
468 | e->type->ops.mq.finish_request(rq); | |
469 | if (rq->elv.icq) { | |
470 | put_io_context(rq->elv.icq->ioc); | |
471 | rq->elv.icq = NULL; | |
472 | } | |
473 | } | |
320ae51f | 474 | |
6af54051 | 475 | ctx->rq_completed[rq_is_sync(rq)]++; |
e8064021 | 476 | if (rq->rq_flags & RQF_MQ_INFLIGHT) |
0d2602ca | 477 | atomic_dec(&hctx->nr_active); |
87760e5e | 478 | |
7beb2f84 JA |
479 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) |
480 | laptop_io_completion(q->backing_dev_info); | |
481 | ||
87760e5e | 482 | wbt_done(q->rq_wb, &rq->issue_stat); |
0d2602ca | 483 | |
af76e555 | 484 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 485 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
486 | if (rq->tag != -1) |
487 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
488 | if (sched_tag != -1) | |
c05f8525 | 489 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 490 | blk_mq_sched_restart(hctx); |
3ef28e83 | 491 | blk_queue_exit(q); |
320ae51f | 492 | } |
1a3b595a | 493 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 494 | |
2a842aca | 495 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 496 | { |
0d11e6ac ML |
497 | blk_account_io_done(rq); |
498 | ||
91b63639 | 499 | if (rq->end_io) { |
87760e5e | 500 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 501 | rq->end_io(rq, error); |
91b63639 CH |
502 | } else { |
503 | if (unlikely(blk_bidi_rq(rq))) | |
504 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 505 | blk_mq_free_request(rq); |
91b63639 | 506 | } |
320ae51f | 507 | } |
c8a446ad | 508 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 509 | |
2a842aca | 510 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
511 | { |
512 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
513 | BUG(); | |
c8a446ad | 514 | __blk_mq_end_request(rq, error); |
63151a44 | 515 | } |
c8a446ad | 516 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 517 | |
30a91cb4 | 518 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 519 | { |
3d6efbf6 | 520 | struct request *rq = data; |
320ae51f | 521 | |
30a91cb4 | 522 | rq->q->softirq_done_fn(rq); |
320ae51f | 523 | } |
320ae51f | 524 | |
453f8341 | 525 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
526 | { |
527 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 528 | bool shared = false; |
320ae51f JA |
529 | int cpu; |
530 | ||
453f8341 CH |
531 | if (rq->internal_tag != -1) |
532 | blk_mq_sched_completed_request(rq); | |
533 | if (rq->rq_flags & RQF_STATS) { | |
534 | blk_mq_poll_stats_start(rq->q); | |
535 | blk_stat_add(rq); | |
536 | } | |
537 | ||
38535201 | 538 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
539 | rq->q->softirq_done_fn(rq); |
540 | return; | |
541 | } | |
320ae51f JA |
542 | |
543 | cpu = get_cpu(); | |
38535201 CH |
544 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
545 | shared = cpus_share_cache(cpu, ctx->cpu); | |
546 | ||
547 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 548 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
549 | rq->csd.info = rq; |
550 | rq->csd.flags = 0; | |
c46fff2a | 551 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 552 | } else { |
30a91cb4 | 553 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 554 | } |
320ae51f JA |
555 | put_cpu(); |
556 | } | |
30a91cb4 CH |
557 | |
558 | /** | |
559 | * blk_mq_complete_request - end I/O on a request | |
560 | * @rq: the request being processed | |
561 | * | |
562 | * Description: | |
563 | * Ends all I/O on a request. It does not handle partial completions. | |
564 | * The actual completion happens out-of-order, through a IPI handler. | |
565 | **/ | |
08e0029a | 566 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 567 | { |
95f09684 JA |
568 | struct request_queue *q = rq->q; |
569 | ||
570 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 571 | return; |
08e0029a | 572 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 573 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
574 | } |
575 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 576 | |
973c0191 KB |
577 | int blk_mq_request_started(struct request *rq) |
578 | { | |
579 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
580 | } | |
581 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
582 | ||
e2490073 | 583 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
584 | { |
585 | struct request_queue *q = rq->q; | |
586 | ||
bd166ef1 JA |
587 | blk_mq_sched_started_request(rq); |
588 | ||
320ae51f JA |
589 | trace_block_rq_issue(q, rq); |
590 | ||
cf43e6be | 591 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 592 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 593 | rq->rq_flags |= RQF_STATS; |
87760e5e | 594 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
595 | } |
596 | ||
2b8393b4 | 597 | blk_add_timer(rq); |
87ee7b11 | 598 | |
538b7534 JA |
599 | /* |
600 | * Ensure that ->deadline is visible before set the started | |
601 | * flag and clear the completed flag. | |
602 | */ | |
603 | smp_mb__before_atomic(); | |
604 | ||
87ee7b11 JA |
605 | /* |
606 | * Mark us as started and clear complete. Complete might have been | |
607 | * set if requeue raced with timeout, which then marked it as | |
608 | * complete. So be sure to clear complete again when we start | |
609 | * the request, otherwise we'll ignore the completion event. | |
610 | */ | |
4b570521 JA |
611 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
612 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
613 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
614 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); | |
49f5baa5 CH |
615 | |
616 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
617 | /* | |
618 | * Make sure space for the drain appears. We know we can do | |
619 | * this because max_hw_segments has been adjusted to be one | |
620 | * fewer than the device can handle. | |
621 | */ | |
622 | rq->nr_phys_segments++; | |
623 | } | |
320ae51f | 624 | } |
e2490073 | 625 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 626 | |
d9d149a3 ML |
627 | /* |
628 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 629 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
630 | * but given rq->deadline is just set in .queue_rq() under |
631 | * this situation, the race won't be possible in reality because | |
632 | * rq->timeout should be set as big enough to cover the window | |
633 | * between blk_mq_start_request() called from .queue_rq() and | |
634 | * clearing REQ_ATOM_STARTED here. | |
635 | */ | |
ed0791b2 | 636 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
637 | { |
638 | struct request_queue *q = rq->q; | |
639 | ||
640 | trace_block_rq_requeue(q, rq); | |
87760e5e | 641 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 642 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 643 | |
e2490073 CH |
644 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
645 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
646 | rq->nr_phys_segments--; | |
647 | } | |
320ae51f JA |
648 | } |
649 | ||
2b053aca | 650 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 651 | { |
ed0791b2 | 652 | __blk_mq_requeue_request(rq); |
ed0791b2 | 653 | |
ed0791b2 | 654 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 655 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
656 | } |
657 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
658 | ||
6fca6a61 CH |
659 | static void blk_mq_requeue_work(struct work_struct *work) |
660 | { | |
661 | struct request_queue *q = | |
2849450a | 662 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
663 | LIST_HEAD(rq_list); |
664 | struct request *rq, *next; | |
6fca6a61 | 665 | |
18e9781d | 666 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 667 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 668 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
669 | |
670 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 671 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
672 | continue; |
673 | ||
e8064021 | 674 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 675 | list_del_init(&rq->queuelist); |
bd6737f1 | 676 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
677 | } |
678 | ||
679 | while (!list_empty(&rq_list)) { | |
680 | rq = list_entry(rq_list.next, struct request, queuelist); | |
681 | list_del_init(&rq->queuelist); | |
bd6737f1 | 682 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
683 | } |
684 | ||
52d7f1b5 | 685 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
686 | } |
687 | ||
2b053aca BVA |
688 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
689 | bool kick_requeue_list) | |
6fca6a61 CH |
690 | { |
691 | struct request_queue *q = rq->q; | |
692 | unsigned long flags; | |
693 | ||
694 | /* | |
695 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
696 | * request head insertation from the workqueue. | |
697 | */ | |
e8064021 | 698 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
699 | |
700 | spin_lock_irqsave(&q->requeue_lock, flags); | |
701 | if (at_head) { | |
e8064021 | 702 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
703 | list_add(&rq->queuelist, &q->requeue_list); |
704 | } else { | |
705 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
706 | } | |
707 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
708 | |
709 | if (kick_requeue_list) | |
710 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
711 | } |
712 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
713 | ||
714 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
715 | { | |
2849450a | 716 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
717 | } |
718 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
719 | ||
2849450a MS |
720 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
721 | unsigned long msecs) | |
722 | { | |
d4acf365 BVA |
723 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
724 | msecs_to_jiffies(msecs)); | |
2849450a MS |
725 | } |
726 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
727 | ||
0e62f51f JA |
728 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
729 | { | |
88c7b2b7 JA |
730 | if (tag < tags->nr_tags) { |
731 | prefetch(tags->rqs[tag]); | |
4ee86bab | 732 | return tags->rqs[tag]; |
88c7b2b7 | 733 | } |
4ee86bab HR |
734 | |
735 | return NULL; | |
24d2f903 CH |
736 | } |
737 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
738 | ||
320ae51f | 739 | struct blk_mq_timeout_data { |
46f92d42 CH |
740 | unsigned long next; |
741 | unsigned int next_set; | |
320ae51f JA |
742 | }; |
743 | ||
90415837 | 744 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 745 | { |
f8a5b122 | 746 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 747 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
748 | |
749 | /* | |
750 | * We know that complete is set at this point. If STARTED isn't set | |
751 | * anymore, then the request isn't active and the "timeout" should | |
752 | * just be ignored. This can happen due to the bitflag ordering. | |
753 | * Timeout first checks if STARTED is set, and if it is, assumes | |
754 | * the request is active. But if we race with completion, then | |
48b99c9d | 755 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
756 | * a timeout event with a request that isn't active. |
757 | */ | |
46f92d42 CH |
758 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
759 | return; | |
87ee7b11 | 760 | |
46f92d42 | 761 | if (ops->timeout) |
0152fb6b | 762 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
763 | |
764 | switch (ret) { | |
765 | case BLK_EH_HANDLED: | |
766 | __blk_mq_complete_request(req); | |
767 | break; | |
768 | case BLK_EH_RESET_TIMER: | |
769 | blk_add_timer(req); | |
770 | blk_clear_rq_complete(req); | |
771 | break; | |
772 | case BLK_EH_NOT_HANDLED: | |
773 | break; | |
774 | default: | |
775 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
776 | break; | |
777 | } | |
87ee7b11 | 778 | } |
5b3f25fc | 779 | |
81481eb4 CH |
780 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
781 | struct request *rq, void *priv, bool reserved) | |
782 | { | |
783 | struct blk_mq_timeout_data *data = priv; | |
87ee7b11 | 784 | |
95a49603 | 785 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 786 | return; |
87ee7b11 | 787 | |
d9d149a3 ML |
788 | /* |
789 | * The rq being checked may have been freed and reallocated | |
790 | * out already here, we avoid this race by checking rq->deadline | |
791 | * and REQ_ATOM_COMPLETE flag together: | |
792 | * | |
793 | * - if rq->deadline is observed as new value because of | |
794 | * reusing, the rq won't be timed out because of timing. | |
795 | * - if rq->deadline is observed as previous value, | |
796 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
797 | * because we put a barrier between setting rq->deadline | |
798 | * and clearing the flag in blk_mq_start_request(), so | |
799 | * this rq won't be timed out too. | |
800 | */ | |
46f92d42 CH |
801 | if (time_after_eq(jiffies, rq->deadline)) { |
802 | if (!blk_mark_rq_complete(rq)) | |
0152fb6b | 803 | blk_mq_rq_timed_out(rq, reserved); |
46f92d42 CH |
804 | } else if (!data->next_set || time_after(data->next, rq->deadline)) { |
805 | data->next = rq->deadline; | |
806 | data->next_set = 1; | |
807 | } | |
87ee7b11 JA |
808 | } |
809 | ||
287922eb | 810 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 811 | { |
287922eb CH |
812 | struct request_queue *q = |
813 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
814 | struct blk_mq_timeout_data data = { |
815 | .next = 0, | |
816 | .next_set = 0, | |
817 | }; | |
81481eb4 | 818 | int i; |
320ae51f | 819 | |
71f79fb3 GKB |
820 | /* A deadlock might occur if a request is stuck requiring a |
821 | * timeout at the same time a queue freeze is waiting | |
822 | * completion, since the timeout code would not be able to | |
823 | * acquire the queue reference here. | |
824 | * | |
825 | * That's why we don't use blk_queue_enter here; instead, we use | |
826 | * percpu_ref_tryget directly, because we need to be able to | |
827 | * obtain a reference even in the short window between the queue | |
828 | * starting to freeze, by dropping the first reference in | |
1671d522 | 829 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
830 | * consumed, marked by the instant q_usage_counter reaches |
831 | * zero. | |
832 | */ | |
833 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
834 | return; |
835 | ||
0bf6cd5b | 836 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 837 | |
81481eb4 CH |
838 | if (data.next_set) { |
839 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
840 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 841 | } else { |
0bf6cd5b CH |
842 | struct blk_mq_hw_ctx *hctx; |
843 | ||
f054b56c ML |
844 | queue_for_each_hw_ctx(q, hctx, i) { |
845 | /* the hctx may be unmapped, so check it here */ | |
846 | if (blk_mq_hw_queue_mapped(hctx)) | |
847 | blk_mq_tag_idle(hctx); | |
848 | } | |
0d2602ca | 849 | } |
287922eb | 850 | blk_queue_exit(q); |
320ae51f JA |
851 | } |
852 | ||
88459642 OS |
853 | struct flush_busy_ctx_data { |
854 | struct blk_mq_hw_ctx *hctx; | |
855 | struct list_head *list; | |
856 | }; | |
857 | ||
858 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
859 | { | |
860 | struct flush_busy_ctx_data *flush_data = data; | |
861 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
862 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
863 | ||
864 | sbitmap_clear_bit(sb, bitnr); | |
865 | spin_lock(&ctx->lock); | |
866 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
867 | spin_unlock(&ctx->lock); | |
868 | return true; | |
869 | } | |
870 | ||
1429d7c9 JA |
871 | /* |
872 | * Process software queues that have been marked busy, splicing them | |
873 | * to the for-dispatch | |
874 | */ | |
2c3ad667 | 875 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 876 | { |
88459642 OS |
877 | struct flush_busy_ctx_data data = { |
878 | .hctx = hctx, | |
879 | .list = list, | |
880 | }; | |
1429d7c9 | 881 | |
88459642 | 882 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 883 | } |
2c3ad667 | 884 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 885 | |
703fd1c0 JA |
886 | static inline unsigned int queued_to_index(unsigned int queued) |
887 | { | |
888 | if (!queued) | |
889 | return 0; | |
1429d7c9 | 890 | |
703fd1c0 | 891 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
892 | } |
893 | ||
bd6737f1 JA |
894 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
895 | bool wait) | |
bd166ef1 JA |
896 | { |
897 | struct blk_mq_alloc_data data = { | |
898 | .q = rq->q, | |
bd166ef1 JA |
899 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
900 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
901 | }; | |
902 | ||
5feeacdd JA |
903 | might_sleep_if(wait); |
904 | ||
81380ca1 OS |
905 | if (rq->tag != -1) |
906 | goto done; | |
bd166ef1 | 907 | |
415b806d SG |
908 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
909 | data.flags |= BLK_MQ_REQ_RESERVED; | |
910 | ||
bd166ef1 JA |
911 | rq->tag = blk_mq_get_tag(&data); |
912 | if (rq->tag >= 0) { | |
200e86b3 JA |
913 | if (blk_mq_tag_busy(data.hctx)) { |
914 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
915 | atomic_inc(&data.hctx->nr_active); | |
916 | } | |
bd166ef1 | 917 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
918 | } |
919 | ||
81380ca1 OS |
920 | done: |
921 | if (hctx) | |
922 | *hctx = data.hctx; | |
923 | return rq->tag != -1; | |
bd166ef1 JA |
924 | } |
925 | ||
113285b4 JA |
926 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
927 | struct request *rq) | |
99cf1dc5 | 928 | { |
99cf1dc5 JA |
929 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
930 | rq->tag = -1; | |
931 | ||
932 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
933 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
934 | atomic_dec(&hctx->nr_active); | |
935 | } | |
936 | } | |
937 | ||
113285b4 JA |
938 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
939 | struct request *rq) | |
940 | { | |
941 | if (rq->tag == -1 || rq->internal_tag == -1) | |
942 | return; | |
943 | ||
944 | __blk_mq_put_driver_tag(hctx, rq); | |
945 | } | |
946 | ||
947 | static void blk_mq_put_driver_tag(struct request *rq) | |
948 | { | |
949 | struct blk_mq_hw_ctx *hctx; | |
950 | ||
951 | if (rq->tag == -1 || rq->internal_tag == -1) | |
952 | return; | |
953 | ||
954 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
955 | __blk_mq_put_driver_tag(hctx, rq); | |
956 | } | |
957 | ||
bd166ef1 JA |
958 | /* |
959 | * If we fail getting a driver tag because all the driver tags are already | |
960 | * assigned and on the dispatch list, BUT the first entry does not have a | |
961 | * tag, then we could deadlock. For that case, move entries with assigned | |
962 | * driver tags to the front, leaving the set of tagged requests in the | |
963 | * same order, and the untagged set in the same order. | |
964 | */ | |
965 | static bool reorder_tags_to_front(struct list_head *list) | |
966 | { | |
967 | struct request *rq, *tmp, *first = NULL; | |
968 | ||
969 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
970 | if (rq == first) | |
971 | break; | |
972 | if (rq->tag != -1) { | |
973 | list_move(&rq->queuelist, list); | |
974 | if (!first) | |
975 | first = rq; | |
976 | } | |
977 | } | |
978 | ||
979 | return first != NULL; | |
980 | } | |
981 | ||
ac6424b9 | 982 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags, |
da55f2cc OS |
983 | void *key) |
984 | { | |
985 | struct blk_mq_hw_ctx *hctx; | |
986 | ||
987 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
988 | ||
2055da97 | 989 | list_del(&wait->entry); |
da55f2cc OS |
990 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); |
991 | blk_mq_run_hw_queue(hctx, true); | |
992 | return 1; | |
993 | } | |
994 | ||
995 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
996 | { | |
997 | struct sbq_wait_state *ws; | |
998 | ||
999 | /* | |
1000 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
1001 | * The thread which wins the race to grab this bit adds the hardware | |
1002 | * queue to the wait queue. | |
1003 | */ | |
1004 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
1005 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
1006 | return false; | |
1007 | ||
1008 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
1009 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
1010 | ||
1011 | /* | |
1012 | * As soon as this returns, it's no longer safe to fiddle with | |
1013 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
1014 | * and unlock the bit. | |
1015 | */ | |
1016 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
1017 | return true; | |
1018 | } | |
1019 | ||
81380ca1 | 1020 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list) |
320ae51f | 1021 | { |
81380ca1 | 1022 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1023 | struct request *rq; |
fc17b653 | 1024 | int errors, queued; |
320ae51f | 1025 | |
81380ca1 OS |
1026 | if (list_empty(list)) |
1027 | return false; | |
1028 | ||
320ae51f JA |
1029 | /* |
1030 | * Now process all the entries, sending them to the driver. | |
1031 | */ | |
93efe981 | 1032 | errors = queued = 0; |
81380ca1 | 1033 | do { |
74c45052 | 1034 | struct blk_mq_queue_data bd; |
fc17b653 | 1035 | blk_status_t ret; |
320ae51f | 1036 | |
f04c3df3 | 1037 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1038 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1039 | if (!queued && reorder_tags_to_front(list)) | |
1040 | continue; | |
3c782d67 JA |
1041 | |
1042 | /* | |
da55f2cc OS |
1043 | * The initial allocation attempt failed, so we need to |
1044 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1045 | */ |
807b1041 OS |
1046 | if (!blk_mq_dispatch_wait_add(hctx)) |
1047 | break; | |
1048 | ||
1049 | /* | |
1050 | * It's possible that a tag was freed in the window | |
1051 | * between the allocation failure and adding the | |
1052 | * hardware queue to the wait queue. | |
1053 | */ | |
1054 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) | |
3c782d67 | 1055 | break; |
bd166ef1 | 1056 | } |
da55f2cc | 1057 | |
320ae51f | 1058 | list_del_init(&rq->queuelist); |
320ae51f | 1059 | |
74c45052 | 1060 | bd.rq = rq; |
113285b4 JA |
1061 | |
1062 | /* | |
1063 | * Flag last if we have no more requests, or if we have more | |
1064 | * but can't assign a driver tag to it. | |
1065 | */ | |
1066 | if (list_empty(list)) | |
1067 | bd.last = true; | |
1068 | else { | |
1069 | struct request *nxt; | |
1070 | ||
1071 | nxt = list_first_entry(list, struct request, queuelist); | |
1072 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1073 | } | |
74c45052 JA |
1074 | |
1075 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1076 | if (ret == BLK_STS_RESOURCE) { |
113285b4 | 1077 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1078 | list_add(&rq->queuelist, list); |
ed0791b2 | 1079 | __blk_mq_requeue_request(rq); |
320ae51f | 1080 | break; |
fc17b653 CH |
1081 | } |
1082 | ||
1083 | if (unlikely(ret != BLK_STS_OK)) { | |
93efe981 | 1084 | errors++; |
2a842aca | 1085 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1086 | continue; |
320ae51f JA |
1087 | } |
1088 | ||
fc17b653 | 1089 | queued++; |
81380ca1 | 1090 | } while (!list_empty(list)); |
320ae51f | 1091 | |
703fd1c0 | 1092 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1093 | |
1094 | /* | |
1095 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1096 | * that is where we will continue on next queue run. | |
1097 | */ | |
f04c3df3 | 1098 | if (!list_empty(list)) { |
113285b4 | 1099 | /* |
710c785f BVA |
1100 | * If an I/O scheduler has been configured and we got a driver |
1101 | * tag for the next request already, free it again. | |
113285b4 JA |
1102 | */ |
1103 | rq = list_first_entry(list, struct request, queuelist); | |
1104 | blk_mq_put_driver_tag(rq); | |
1105 | ||
320ae51f | 1106 | spin_lock(&hctx->lock); |
c13660a0 | 1107 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1108 | spin_unlock(&hctx->lock); |
f04c3df3 | 1109 | |
9ba52e58 | 1110 | /* |
710c785f BVA |
1111 | * If SCHED_RESTART was set by the caller of this function and |
1112 | * it is no longer set that means that it was cleared by another | |
1113 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1114 | * |
710c785f BVA |
1115 | * If TAG_WAITING is set that means that an I/O scheduler has |
1116 | * been configured and another thread is waiting for a driver | |
1117 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1118 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1119 | * |
710c785f BVA |
1120 | * If no I/O scheduler has been configured it is possible that |
1121 | * the hardware queue got stopped and restarted before requests | |
1122 | * were pushed back onto the dispatch list. Rerun the queue to | |
1123 | * avoid starvation. Notes: | |
1124 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1125 | * been stopped before rerunning a queue. | |
1126 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1127 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1128 | * and dm-rq. |
bd166ef1 | 1129 | */ |
da55f2cc OS |
1130 | if (!blk_mq_sched_needs_restart(hctx) && |
1131 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1132 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1133 | } |
f04c3df3 | 1134 | |
93efe981 | 1135 | return (queued + errors) != 0; |
f04c3df3 JA |
1136 | } |
1137 | ||
6a83e74d BVA |
1138 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1139 | { | |
1140 | int srcu_idx; | |
1141 | ||
b7a71e66 JA |
1142 | /* |
1143 | * We should be running this queue from one of the CPUs that | |
1144 | * are mapped to it. | |
1145 | */ | |
6a83e74d BVA |
1146 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1147 | cpu_online(hctx->next_cpu)); | |
1148 | ||
b7a71e66 JA |
1149 | /* |
1150 | * We can't run the queue inline with ints disabled. Ensure that | |
1151 | * we catch bad users of this early. | |
1152 | */ | |
1153 | WARN_ON_ONCE(in_interrupt()); | |
1154 | ||
6a83e74d BVA |
1155 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1156 | rcu_read_lock(); | |
bd166ef1 | 1157 | blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1158 | rcu_read_unlock(); |
1159 | } else { | |
bf4907c0 JA |
1160 | might_sleep(); |
1161 | ||
07319678 | 1162 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
bd166ef1 | 1163 | blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1164 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d BVA |
1165 | } |
1166 | } | |
1167 | ||
506e931f JA |
1168 | /* |
1169 | * It'd be great if the workqueue API had a way to pass | |
1170 | * in a mask and had some smarts for more clever placement. | |
1171 | * For now we just round-robin here, switching for every | |
1172 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1173 | */ | |
1174 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1175 | { | |
b657d7e6 CH |
1176 | if (hctx->queue->nr_hw_queues == 1) |
1177 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1178 | |
1179 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1180 | int next_cpu; |
506e931f JA |
1181 | |
1182 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1183 | if (next_cpu >= nr_cpu_ids) | |
1184 | next_cpu = cpumask_first(hctx->cpumask); | |
1185 | ||
1186 | hctx->next_cpu = next_cpu; | |
1187 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1188 | } | |
1189 | ||
b657d7e6 | 1190 | return hctx->next_cpu; |
506e931f JA |
1191 | } |
1192 | ||
7587a5ae BVA |
1193 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1194 | unsigned long msecs) | |
320ae51f | 1195 | { |
5435c023 BVA |
1196 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1197 | return; | |
1198 | ||
1199 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1200 | return; |
1201 | ||
1b792f2f | 1202 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1203 | int cpu = get_cpu(); |
1204 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1205 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1206 | put_cpu(); |
398205b8 PB |
1207 | return; |
1208 | } | |
e4043dcf | 1209 | |
2a90d4aa | 1210 | put_cpu(); |
e4043dcf | 1211 | } |
398205b8 | 1212 | |
9f993737 JA |
1213 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1214 | &hctx->run_work, | |
1215 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1216 | } |
1217 | ||
1218 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1219 | { | |
1220 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1221 | } | |
1222 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1223 | ||
1224 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1225 | { | |
1226 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1227 | } |
5b727272 | 1228 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1229 | |
b94ec296 | 1230 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1231 | { |
1232 | struct blk_mq_hw_ctx *hctx; | |
1233 | int i; | |
1234 | ||
1235 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1236 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1237 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1238 | continue; |
1239 | ||
b94ec296 | 1240 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1241 | } |
1242 | } | |
b94ec296 | 1243 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1244 | |
fd001443 BVA |
1245 | /** |
1246 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1247 | * @q: request queue. | |
1248 | * | |
1249 | * The caller is responsible for serializing this function against | |
1250 | * blk_mq_{start,stop}_hw_queue(). | |
1251 | */ | |
1252 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1253 | { | |
1254 | struct blk_mq_hw_ctx *hctx; | |
1255 | int i; | |
1256 | ||
1257 | queue_for_each_hw_ctx(q, hctx, i) | |
1258 | if (blk_mq_hctx_stopped(hctx)) | |
1259 | return true; | |
1260 | ||
1261 | return false; | |
1262 | } | |
1263 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1264 | ||
39a70c76 ML |
1265 | /* |
1266 | * This function is often used for pausing .queue_rq() by driver when | |
1267 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1268 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1269 | * |
1270 | * We do not guarantee that dispatch can be drained or blocked | |
1271 | * after blk_mq_stop_hw_queue() returns. Please use | |
1272 | * blk_mq_quiesce_queue() for that requirement. | |
1273 | */ | |
2719aa21 JA |
1274 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1275 | { | |
641a9ed6 | 1276 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1277 | |
641a9ed6 | 1278 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1279 | } |
641a9ed6 | 1280 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1281 | |
39a70c76 ML |
1282 | /* |
1283 | * This function is often used for pausing .queue_rq() by driver when | |
1284 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1285 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1286 | * |
1287 | * We do not guarantee that dispatch can be drained or blocked | |
1288 | * after blk_mq_stop_hw_queues() returns. Please use | |
1289 | * blk_mq_quiesce_queue() for that requirement. | |
1290 | */ | |
2719aa21 JA |
1291 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1292 | { | |
641a9ed6 ML |
1293 | struct blk_mq_hw_ctx *hctx; |
1294 | int i; | |
1295 | ||
1296 | queue_for_each_hw_ctx(q, hctx, i) | |
1297 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1298 | } |
1299 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1300 | ||
320ae51f JA |
1301 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1302 | { | |
1303 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1304 | |
0ffbce80 | 1305 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1306 | } |
1307 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1308 | ||
2f268556 CH |
1309 | void blk_mq_start_hw_queues(struct request_queue *q) |
1310 | { | |
1311 | struct blk_mq_hw_ctx *hctx; | |
1312 | int i; | |
1313 | ||
1314 | queue_for_each_hw_ctx(q, hctx, i) | |
1315 | blk_mq_start_hw_queue(hctx); | |
1316 | } | |
1317 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1318 | ||
ae911c5e JA |
1319 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1320 | { | |
1321 | if (!blk_mq_hctx_stopped(hctx)) | |
1322 | return; | |
1323 | ||
1324 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1325 | blk_mq_run_hw_queue(hctx, async); | |
1326 | } | |
1327 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1328 | ||
1b4a3258 | 1329 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1330 | { |
1331 | struct blk_mq_hw_ctx *hctx; | |
1332 | int i; | |
1333 | ||
ae911c5e JA |
1334 | queue_for_each_hw_ctx(q, hctx, i) |
1335 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1336 | } |
1337 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1338 | ||
70f4db63 | 1339 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1340 | { |
1341 | struct blk_mq_hw_ctx *hctx; | |
1342 | ||
9f993737 | 1343 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1344 | |
21c6e939 JA |
1345 | /* |
1346 | * If we are stopped, don't run the queue. The exception is if | |
1347 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1348 | * the STOPPED bit and run it. | |
1349 | */ | |
1350 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1351 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1352 | return; | |
7587a5ae | 1353 | |
21c6e939 JA |
1354 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1355 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1356 | } | |
7587a5ae BVA |
1357 | |
1358 | __blk_mq_run_hw_queue(hctx); | |
1359 | } | |
1360 | ||
70f4db63 CH |
1361 | |
1362 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1363 | { | |
5435c023 | 1364 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1365 | return; |
70f4db63 | 1366 | |
21c6e939 JA |
1367 | /* |
1368 | * Stop the hw queue, then modify currently delayed work. | |
1369 | * This should prevent us from running the queue prematurely. | |
1370 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1371 | */ | |
7e79dadc | 1372 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1373 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1374 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1375 | &hctx->run_work, | |
1376 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1377 | } |
1378 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1379 | ||
cfd0c552 | 1380 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1381 | struct request *rq, |
1382 | bool at_head) | |
320ae51f | 1383 | { |
e57690fe JA |
1384 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1385 | ||
7b607814 BVA |
1386 | lockdep_assert_held(&ctx->lock); |
1387 | ||
01b983c9 JA |
1388 | trace_block_rq_insert(hctx->queue, rq); |
1389 | ||
72a0a36e CH |
1390 | if (at_head) |
1391 | list_add(&rq->queuelist, &ctx->rq_list); | |
1392 | else | |
1393 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1394 | } |
4bb659b1 | 1395 | |
2c3ad667 JA |
1396 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1397 | bool at_head) | |
cfd0c552 ML |
1398 | { |
1399 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1400 | ||
7b607814 BVA |
1401 | lockdep_assert_held(&ctx->lock); |
1402 | ||
e57690fe | 1403 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1404 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1405 | } |
1406 | ||
157f377b JA |
1407 | /* |
1408 | * Should only be used carefully, when the caller knows we want to | |
1409 | * bypass a potential IO scheduler on the target device. | |
1410 | */ | |
1411 | void blk_mq_request_bypass_insert(struct request *rq) | |
1412 | { | |
1413 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1414 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1415 | ||
1416 | spin_lock(&hctx->lock); | |
1417 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1418 | spin_unlock(&hctx->lock); | |
1419 | ||
1420 | blk_mq_run_hw_queue(hctx, false); | |
1421 | } | |
1422 | ||
bd166ef1 JA |
1423 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1424 | struct list_head *list) | |
320ae51f JA |
1425 | |
1426 | { | |
320ae51f JA |
1427 | /* |
1428 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1429 | * offline now | |
1430 | */ | |
1431 | spin_lock(&ctx->lock); | |
1432 | while (!list_empty(list)) { | |
1433 | struct request *rq; | |
1434 | ||
1435 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1436 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1437 | list_del_init(&rq->queuelist); |
e57690fe | 1438 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1439 | } |
cfd0c552 | 1440 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1441 | spin_unlock(&ctx->lock); |
320ae51f JA |
1442 | } |
1443 | ||
1444 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1445 | { | |
1446 | struct request *rqa = container_of(a, struct request, queuelist); | |
1447 | struct request *rqb = container_of(b, struct request, queuelist); | |
1448 | ||
1449 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1450 | (rqa->mq_ctx == rqb->mq_ctx && | |
1451 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1452 | } | |
1453 | ||
1454 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1455 | { | |
1456 | struct blk_mq_ctx *this_ctx; | |
1457 | struct request_queue *this_q; | |
1458 | struct request *rq; | |
1459 | LIST_HEAD(list); | |
1460 | LIST_HEAD(ctx_list); | |
1461 | unsigned int depth; | |
1462 | ||
1463 | list_splice_init(&plug->mq_list, &list); | |
1464 | ||
1465 | list_sort(NULL, &list, plug_ctx_cmp); | |
1466 | ||
1467 | this_q = NULL; | |
1468 | this_ctx = NULL; | |
1469 | depth = 0; | |
1470 | ||
1471 | while (!list_empty(&list)) { | |
1472 | rq = list_entry_rq(list.next); | |
1473 | list_del_init(&rq->queuelist); | |
1474 | BUG_ON(!rq->q); | |
1475 | if (rq->mq_ctx != this_ctx) { | |
1476 | if (this_ctx) { | |
bd166ef1 JA |
1477 | trace_block_unplug(this_q, depth, from_schedule); |
1478 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1479 | &ctx_list, | |
1480 | from_schedule); | |
320ae51f JA |
1481 | } |
1482 | ||
1483 | this_ctx = rq->mq_ctx; | |
1484 | this_q = rq->q; | |
1485 | depth = 0; | |
1486 | } | |
1487 | ||
1488 | depth++; | |
1489 | list_add_tail(&rq->queuelist, &ctx_list); | |
1490 | } | |
1491 | ||
1492 | /* | |
1493 | * If 'this_ctx' is set, we know we have entries to complete | |
1494 | * on 'ctx_list'. Do those. | |
1495 | */ | |
1496 | if (this_ctx) { | |
bd166ef1 JA |
1497 | trace_block_unplug(this_q, depth, from_schedule); |
1498 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1499 | from_schedule); | |
320ae51f JA |
1500 | } |
1501 | } | |
1502 | ||
1503 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1504 | { | |
da8d7f07 | 1505 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1506 | |
6e85eaf3 | 1507 | blk_account_io_start(rq, true); |
320ae51f JA |
1508 | } |
1509 | ||
ab42f35d ML |
1510 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1511 | struct blk_mq_ctx *ctx, | |
1512 | struct request *rq) | |
1513 | { | |
1514 | spin_lock(&ctx->lock); | |
1515 | __blk_mq_insert_request(hctx, rq, false); | |
1516 | spin_unlock(&ctx->lock); | |
07068d5b | 1517 | } |
14ec77f3 | 1518 | |
fd2d3326 JA |
1519 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1520 | { | |
bd166ef1 JA |
1521 | if (rq->tag != -1) |
1522 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1523 | ||
1524 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1525 | } |
1526 | ||
d964f04a ML |
1527 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1528 | struct request *rq, | |
1529 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1530 | { |
f984df1f | 1531 | struct request_queue *q = rq->q; |
f984df1f SL |
1532 | struct blk_mq_queue_data bd = { |
1533 | .rq = rq, | |
d945a365 | 1534 | .last = true, |
f984df1f | 1535 | }; |
bd166ef1 | 1536 | blk_qc_t new_cookie; |
f06345ad | 1537 | blk_status_t ret; |
d964f04a ML |
1538 | bool run_queue = true; |
1539 | ||
f4560ffe ML |
1540 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1541 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1542 | run_queue = false; |
1543 | goto insert; | |
1544 | } | |
f984df1f | 1545 | |
bd166ef1 | 1546 | if (q->elevator) |
2253efc8 BVA |
1547 | goto insert; |
1548 | ||
d964f04a | 1549 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1550 | goto insert; |
1551 | ||
1552 | new_cookie = request_to_qc_t(hctx, rq); | |
1553 | ||
f984df1f SL |
1554 | /* |
1555 | * For OK queue, we are done. For error, kill it. Any other | |
1556 | * error (busy), just add it to our list as we previously | |
1557 | * would have done | |
1558 | */ | |
1559 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1560 | switch (ret) { |
1561 | case BLK_STS_OK: | |
7b371636 | 1562 | *cookie = new_cookie; |
2253efc8 | 1563 | return; |
fc17b653 CH |
1564 | case BLK_STS_RESOURCE: |
1565 | __blk_mq_requeue_request(rq); | |
1566 | goto insert; | |
1567 | default: | |
7b371636 | 1568 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1569 | blk_mq_end_request(rq, ret); |
2253efc8 | 1570 | return; |
f984df1f | 1571 | } |
7b371636 | 1572 | |
2253efc8 | 1573 | insert: |
d964f04a | 1574 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1575 | } |
1576 | ||
5eb6126e CH |
1577 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1578 | struct request *rq, blk_qc_t *cookie) | |
1579 | { | |
1580 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1581 | rcu_read_lock(); | |
d964f04a | 1582 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1583 | rcu_read_unlock(); |
1584 | } else { | |
bf4907c0 JA |
1585 | unsigned int srcu_idx; |
1586 | ||
1587 | might_sleep(); | |
1588 | ||
07319678 | 1589 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1590 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1591 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1592 | } |
1593 | } | |
1594 | ||
dece1635 | 1595 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1596 | { |
ef295ecf | 1597 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1598 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1599 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1600 | struct request *rq; |
5eb6126e | 1601 | unsigned int request_count = 0; |
f984df1f | 1602 | struct blk_plug *plug; |
5b3f341f | 1603 | struct request *same_queue_rq = NULL; |
7b371636 | 1604 | blk_qc_t cookie; |
87760e5e | 1605 | unsigned int wb_acct; |
07068d5b JA |
1606 | |
1607 | blk_queue_bounce(q, &bio); | |
1608 | ||
af67c31f | 1609 | blk_queue_split(q, &bio); |
f36ea50c | 1610 | |
e23947bd | 1611 | if (!bio_integrity_prep(bio)) |
dece1635 | 1612 | return BLK_QC_T_NONE; |
07068d5b | 1613 | |
87c279e6 OS |
1614 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1615 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1616 | return BLK_QC_T_NONE; | |
f984df1f | 1617 | |
bd166ef1 JA |
1618 | if (blk_mq_sched_bio_merge(q, bio)) |
1619 | return BLK_QC_T_NONE; | |
1620 | ||
87760e5e JA |
1621 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1622 | ||
bd166ef1 JA |
1623 | trace_block_getrq(q, bio, bio->bi_opf); |
1624 | ||
d2c0d383 | 1625 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1626 | if (unlikely(!rq)) { |
1627 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1628 | if (bio->bi_opf & REQ_NOWAIT) |
1629 | bio_wouldblock_error(bio); | |
dece1635 | 1630 | return BLK_QC_T_NONE; |
87760e5e JA |
1631 | } |
1632 | ||
1633 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1634 | |
fd2d3326 | 1635 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1636 | |
f984df1f | 1637 | plug = current->plug; |
07068d5b | 1638 | if (unlikely(is_flush_fua)) { |
f984df1f | 1639 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1640 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1641 | if (q->elevator) { |
1642 | blk_mq_sched_insert_request(rq, false, true, true, | |
1643 | true); | |
6a83e74d | 1644 | } else { |
a4d907b6 CH |
1645 | blk_insert_flush(rq); |
1646 | blk_mq_run_hw_queue(data.hctx, true); | |
6a83e74d | 1647 | } |
a4d907b6 | 1648 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1649 | struct request *last = NULL; |
1650 | ||
b00c53e8 | 1651 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1652 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1653 | |
1654 | /* | |
1655 | * @request_count may become stale because of schedule | |
1656 | * out, so check the list again. | |
1657 | */ | |
1658 | if (list_empty(&plug->mq_list)) | |
1659 | request_count = 0; | |
254d259d CH |
1660 | else if (blk_queue_nomerges(q)) |
1661 | request_count = blk_plug_queued_count(q); | |
1662 | ||
676d0607 | 1663 | if (!request_count) |
e6c4438b | 1664 | trace_block_plug(q); |
600271d9 SL |
1665 | else |
1666 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1667 | |
600271d9 SL |
1668 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1669 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1670 | blk_flush_plug_list(plug, false); |
1671 | trace_block_plug(q); | |
320ae51f | 1672 | } |
b094f89c | 1673 | |
e6c4438b | 1674 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1675 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1676 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1677 | |
07068d5b | 1678 | /* |
6a83e74d | 1679 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1680 | * Otherwise the existing request in the plug list will be |
1681 | * issued. So the plug list will have one request at most | |
2299722c CH |
1682 | * The plug list might get flushed before this. If that happens, |
1683 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1684 | */ |
2299722c CH |
1685 | if (list_empty(&plug->mq_list)) |
1686 | same_queue_rq = NULL; | |
1687 | if (same_queue_rq) | |
1688 | list_del_init(&same_queue_rq->queuelist); | |
1689 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1690 | ||
bf4907c0 JA |
1691 | blk_mq_put_ctx(data.ctx); |
1692 | ||
dad7a3be ML |
1693 | if (same_queue_rq) { |
1694 | data.hctx = blk_mq_map_queue(q, | |
1695 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1696 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1697 | &cookie); | |
dad7a3be | 1698 | } |
a4d907b6 | 1699 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1700 | blk_mq_put_ctx(data.ctx); |
2299722c | 1701 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1702 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1703 | } else if (q->elevator) { |
b00c53e8 | 1704 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1705 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1706 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1707 | } else { |
b00c53e8 | 1708 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1709 | blk_mq_bio_to_request(rq, bio); |
1710 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1711 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1712 | } |
320ae51f | 1713 | |
7b371636 | 1714 | return cookie; |
320ae51f JA |
1715 | } |
1716 | ||
cc71a6f4 JA |
1717 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1718 | unsigned int hctx_idx) | |
95363efd | 1719 | { |
e9b267d9 | 1720 | struct page *page; |
320ae51f | 1721 | |
24d2f903 | 1722 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1723 | int i; |
320ae51f | 1724 | |
24d2f903 | 1725 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1726 | struct request *rq = tags->static_rqs[i]; |
1727 | ||
1728 | if (!rq) | |
e9b267d9 | 1729 | continue; |
d6296d39 | 1730 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1731 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1732 | } |
320ae51f | 1733 | } |
320ae51f | 1734 | |
24d2f903 CH |
1735 | while (!list_empty(&tags->page_list)) { |
1736 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1737 | list_del_init(&page->lru); |
f75782e4 CM |
1738 | /* |
1739 | * Remove kmemleak object previously allocated in | |
1740 | * blk_mq_init_rq_map(). | |
1741 | */ | |
1742 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1743 | __free_pages(page, page->private); |
1744 | } | |
cc71a6f4 | 1745 | } |
320ae51f | 1746 | |
cc71a6f4 JA |
1747 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1748 | { | |
24d2f903 | 1749 | kfree(tags->rqs); |
cc71a6f4 | 1750 | tags->rqs = NULL; |
2af8cbe3 JA |
1751 | kfree(tags->static_rqs); |
1752 | tags->static_rqs = NULL; | |
320ae51f | 1753 | |
24d2f903 | 1754 | blk_mq_free_tags(tags); |
320ae51f JA |
1755 | } |
1756 | ||
cc71a6f4 JA |
1757 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1758 | unsigned int hctx_idx, | |
1759 | unsigned int nr_tags, | |
1760 | unsigned int reserved_tags) | |
320ae51f | 1761 | { |
24d2f903 | 1762 | struct blk_mq_tags *tags; |
59f082e4 | 1763 | int node; |
320ae51f | 1764 | |
59f082e4 SL |
1765 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1766 | if (node == NUMA_NO_NODE) | |
1767 | node = set->numa_node; | |
1768 | ||
1769 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1770 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1771 | if (!tags) |
1772 | return NULL; | |
320ae51f | 1773 | |
cc71a6f4 | 1774 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1775 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1776 | node); |
24d2f903 CH |
1777 | if (!tags->rqs) { |
1778 | blk_mq_free_tags(tags); | |
1779 | return NULL; | |
1780 | } | |
320ae51f | 1781 | |
2af8cbe3 JA |
1782 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1783 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1784 | node); |
2af8cbe3 JA |
1785 | if (!tags->static_rqs) { |
1786 | kfree(tags->rqs); | |
1787 | blk_mq_free_tags(tags); | |
1788 | return NULL; | |
1789 | } | |
1790 | ||
cc71a6f4 JA |
1791 | return tags; |
1792 | } | |
1793 | ||
1794 | static size_t order_to_size(unsigned int order) | |
1795 | { | |
1796 | return (size_t)PAGE_SIZE << order; | |
1797 | } | |
1798 | ||
1799 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1800 | unsigned int hctx_idx, unsigned int depth) | |
1801 | { | |
1802 | unsigned int i, j, entries_per_page, max_order = 4; | |
1803 | size_t rq_size, left; | |
59f082e4 SL |
1804 | int node; |
1805 | ||
1806 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1807 | if (node == NUMA_NO_NODE) | |
1808 | node = set->numa_node; | |
cc71a6f4 JA |
1809 | |
1810 | INIT_LIST_HEAD(&tags->page_list); | |
1811 | ||
320ae51f JA |
1812 | /* |
1813 | * rq_size is the size of the request plus driver payload, rounded | |
1814 | * to the cacheline size | |
1815 | */ | |
24d2f903 | 1816 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1817 | cache_line_size()); |
cc71a6f4 | 1818 | left = rq_size * depth; |
320ae51f | 1819 | |
cc71a6f4 | 1820 | for (i = 0; i < depth; ) { |
320ae51f JA |
1821 | int this_order = max_order; |
1822 | struct page *page; | |
1823 | int to_do; | |
1824 | void *p; | |
1825 | ||
b3a834b1 | 1826 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1827 | this_order--; |
1828 | ||
1829 | do { | |
59f082e4 | 1830 | page = alloc_pages_node(node, |
36e1f3d1 | 1831 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1832 | this_order); |
320ae51f JA |
1833 | if (page) |
1834 | break; | |
1835 | if (!this_order--) | |
1836 | break; | |
1837 | if (order_to_size(this_order) < rq_size) | |
1838 | break; | |
1839 | } while (1); | |
1840 | ||
1841 | if (!page) | |
24d2f903 | 1842 | goto fail; |
320ae51f JA |
1843 | |
1844 | page->private = this_order; | |
24d2f903 | 1845 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1846 | |
1847 | p = page_address(page); | |
f75782e4 CM |
1848 | /* |
1849 | * Allow kmemleak to scan these pages as they contain pointers | |
1850 | * to additional allocations like via ops->init_request(). | |
1851 | */ | |
36e1f3d1 | 1852 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1853 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1854 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1855 | left -= to_do * rq_size; |
1856 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1857 | struct request *rq = p; |
1858 | ||
1859 | tags->static_rqs[i] = rq; | |
24d2f903 | 1860 | if (set->ops->init_request) { |
d6296d39 | 1861 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1862 | node)) { |
2af8cbe3 | 1863 | tags->static_rqs[i] = NULL; |
24d2f903 | 1864 | goto fail; |
a5164405 | 1865 | } |
e9b267d9 CH |
1866 | } |
1867 | ||
320ae51f JA |
1868 | p += rq_size; |
1869 | i++; | |
1870 | } | |
1871 | } | |
cc71a6f4 | 1872 | return 0; |
320ae51f | 1873 | |
24d2f903 | 1874 | fail: |
cc71a6f4 JA |
1875 | blk_mq_free_rqs(set, tags, hctx_idx); |
1876 | return -ENOMEM; | |
320ae51f JA |
1877 | } |
1878 | ||
e57690fe JA |
1879 | /* |
1880 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1881 | * software queue to the hw queue dispatch list, and ensure that it | |
1882 | * gets run. | |
1883 | */ | |
9467f859 | 1884 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1885 | { |
9467f859 | 1886 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1887 | struct blk_mq_ctx *ctx; |
1888 | LIST_HEAD(tmp); | |
1889 | ||
9467f859 | 1890 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1891 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1892 | |
1893 | spin_lock(&ctx->lock); | |
1894 | if (!list_empty(&ctx->rq_list)) { | |
1895 | list_splice_init(&ctx->rq_list, &tmp); | |
1896 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1897 | } | |
1898 | spin_unlock(&ctx->lock); | |
1899 | ||
1900 | if (list_empty(&tmp)) | |
9467f859 | 1901 | return 0; |
484b4061 | 1902 | |
e57690fe JA |
1903 | spin_lock(&hctx->lock); |
1904 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1905 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1906 | |
1907 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1908 | return 0; |
484b4061 JA |
1909 | } |
1910 | ||
9467f859 | 1911 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1912 | { |
9467f859 TG |
1913 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1914 | &hctx->cpuhp_dead); | |
484b4061 JA |
1915 | } |
1916 | ||
c3b4afca | 1917 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1918 | static void blk_mq_exit_hctx(struct request_queue *q, |
1919 | struct blk_mq_tag_set *set, | |
1920 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1921 | { | |
9c1051aa OS |
1922 | blk_mq_debugfs_unregister_hctx(hctx); |
1923 | ||
08e98fc6 ML |
1924 | blk_mq_tag_idle(hctx); |
1925 | ||
f70ced09 | 1926 | if (set->ops->exit_request) |
d6296d39 | 1927 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1928 | |
93252632 OS |
1929 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1930 | ||
08e98fc6 ML |
1931 | if (set->ops->exit_hctx) |
1932 | set->ops->exit_hctx(hctx, hctx_idx); | |
1933 | ||
6a83e74d | 1934 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1935 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1936 | |
9467f859 | 1937 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 1938 | blk_free_flush_queue(hctx->fq); |
88459642 | 1939 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
1940 | } |
1941 | ||
624dbe47 ML |
1942 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
1943 | struct blk_mq_tag_set *set, int nr_queue) | |
1944 | { | |
1945 | struct blk_mq_hw_ctx *hctx; | |
1946 | unsigned int i; | |
1947 | ||
1948 | queue_for_each_hw_ctx(q, hctx, i) { | |
1949 | if (i == nr_queue) | |
1950 | break; | |
08e98fc6 | 1951 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 1952 | } |
624dbe47 ML |
1953 | } |
1954 | ||
08e98fc6 ML |
1955 | static int blk_mq_init_hctx(struct request_queue *q, |
1956 | struct blk_mq_tag_set *set, | |
1957 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 1958 | { |
08e98fc6 ML |
1959 | int node; |
1960 | ||
1961 | node = hctx->numa_node; | |
1962 | if (node == NUMA_NO_NODE) | |
1963 | node = hctx->numa_node = set->numa_node; | |
1964 | ||
9f993737 | 1965 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
1966 | spin_lock_init(&hctx->lock); |
1967 | INIT_LIST_HEAD(&hctx->dispatch); | |
1968 | hctx->queue = q; | |
2404e607 | 1969 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 1970 | |
9467f859 | 1971 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
1972 | |
1973 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
1974 | |
1975 | /* | |
08e98fc6 ML |
1976 | * Allocate space for all possible cpus to avoid allocation at |
1977 | * runtime | |
320ae51f | 1978 | */ |
08e98fc6 ML |
1979 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
1980 | GFP_KERNEL, node); | |
1981 | if (!hctx->ctxs) | |
1982 | goto unregister_cpu_notifier; | |
320ae51f | 1983 | |
88459642 OS |
1984 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
1985 | node)) | |
08e98fc6 | 1986 | goto free_ctxs; |
320ae51f | 1987 | |
08e98fc6 | 1988 | hctx->nr_ctx = 0; |
320ae51f | 1989 | |
08e98fc6 ML |
1990 | if (set->ops->init_hctx && |
1991 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
1992 | goto free_bitmap; | |
320ae51f | 1993 | |
93252632 OS |
1994 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
1995 | goto exit_hctx; | |
1996 | ||
f70ced09 ML |
1997 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
1998 | if (!hctx->fq) | |
93252632 | 1999 | goto sched_exit_hctx; |
320ae51f | 2000 | |
f70ced09 | 2001 | if (set->ops->init_request && |
d6296d39 CH |
2002 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
2003 | node)) | |
f70ced09 | 2004 | goto free_fq; |
320ae51f | 2005 | |
6a83e74d | 2006 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2007 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2008 | |
9c1051aa OS |
2009 | blk_mq_debugfs_register_hctx(q, hctx); |
2010 | ||
08e98fc6 | 2011 | return 0; |
320ae51f | 2012 | |
f70ced09 ML |
2013 | free_fq: |
2014 | kfree(hctx->fq); | |
93252632 OS |
2015 | sched_exit_hctx: |
2016 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2017 | exit_hctx: |
2018 | if (set->ops->exit_hctx) | |
2019 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2020 | free_bitmap: |
88459642 | 2021 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2022 | free_ctxs: |
2023 | kfree(hctx->ctxs); | |
2024 | unregister_cpu_notifier: | |
9467f859 | 2025 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2026 | return -1; |
2027 | } | |
320ae51f | 2028 | |
320ae51f JA |
2029 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2030 | unsigned int nr_hw_queues) | |
2031 | { | |
2032 | unsigned int i; | |
2033 | ||
2034 | for_each_possible_cpu(i) { | |
2035 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2036 | struct blk_mq_hw_ctx *hctx; | |
2037 | ||
320ae51f JA |
2038 | __ctx->cpu = i; |
2039 | spin_lock_init(&__ctx->lock); | |
2040 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2041 | __ctx->queue = q; | |
2042 | ||
4b855ad3 CH |
2043 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2044 | if (!cpu_present(i)) | |
320ae51f JA |
2045 | continue; |
2046 | ||
7d7e0f90 | 2047 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2048 | |
320ae51f JA |
2049 | /* |
2050 | * Set local node, IFF we have more than one hw queue. If | |
2051 | * not, we remain on the home node of the device | |
2052 | */ | |
2053 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2054 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2055 | } |
2056 | } | |
2057 | ||
cc71a6f4 JA |
2058 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2059 | { | |
2060 | int ret = 0; | |
2061 | ||
2062 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2063 | set->queue_depth, set->reserved_tags); | |
2064 | if (!set->tags[hctx_idx]) | |
2065 | return false; | |
2066 | ||
2067 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2068 | set->queue_depth); | |
2069 | if (!ret) | |
2070 | return true; | |
2071 | ||
2072 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2073 | set->tags[hctx_idx] = NULL; | |
2074 | return false; | |
2075 | } | |
2076 | ||
2077 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2078 | unsigned int hctx_idx) | |
2079 | { | |
bd166ef1 JA |
2080 | if (set->tags[hctx_idx]) { |
2081 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2082 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2083 | set->tags[hctx_idx] = NULL; | |
2084 | } | |
cc71a6f4 JA |
2085 | } |
2086 | ||
4b855ad3 | 2087 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2088 | { |
d1b1cea1 | 2089 | unsigned int i, hctx_idx; |
320ae51f JA |
2090 | struct blk_mq_hw_ctx *hctx; |
2091 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2092 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2093 | |
60de074b AM |
2094 | /* |
2095 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2096 | */ | |
2097 | mutex_lock(&q->sysfs_lock); | |
2098 | ||
320ae51f | 2099 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2100 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2101 | hctx->nr_ctx = 0; |
2102 | } | |
2103 | ||
2104 | /* | |
4b855ad3 CH |
2105 | * Map software to hardware queues. |
2106 | * | |
2107 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2108 | */ |
4b855ad3 | 2109 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2110 | hctx_idx = q->mq_map[i]; |
2111 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2112 | if (!set->tags[hctx_idx] && |
2113 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2114 | /* |
2115 | * If tags initialization fail for some hctx, | |
2116 | * that hctx won't be brought online. In this | |
2117 | * case, remap the current ctx to hctx[0] which | |
2118 | * is guaranteed to always have tags allocated | |
2119 | */ | |
cc71a6f4 | 2120 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2121 | } |
2122 | ||
897bb0c7 | 2123 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2124 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2125 | |
e4043dcf | 2126 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2127 | ctx->index_hw = hctx->nr_ctx; |
2128 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2129 | } | |
506e931f | 2130 | |
60de074b AM |
2131 | mutex_unlock(&q->sysfs_lock); |
2132 | ||
506e931f | 2133 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2134 | /* |
a68aafa5 JA |
2135 | * If no software queues are mapped to this hardware queue, |
2136 | * disable it and free the request entries. | |
484b4061 JA |
2137 | */ |
2138 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2139 | /* Never unmap queue 0. We need it as a |
2140 | * fallback in case of a new remap fails | |
2141 | * allocation | |
2142 | */ | |
cc71a6f4 JA |
2143 | if (i && set->tags[i]) |
2144 | blk_mq_free_map_and_requests(set, i); | |
2145 | ||
2a34c087 | 2146 | hctx->tags = NULL; |
484b4061 JA |
2147 | continue; |
2148 | } | |
2149 | ||
2a34c087 ML |
2150 | hctx->tags = set->tags[i]; |
2151 | WARN_ON(!hctx->tags); | |
2152 | ||
889fa31f CY |
2153 | /* |
2154 | * Set the map size to the number of mapped software queues. | |
2155 | * This is more accurate and more efficient than looping | |
2156 | * over all possibly mapped software queues. | |
2157 | */ | |
88459642 | 2158 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2159 | |
484b4061 JA |
2160 | /* |
2161 | * Initialize batch roundrobin counts | |
2162 | */ | |
506e931f JA |
2163 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2164 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2165 | } | |
320ae51f JA |
2166 | } |
2167 | ||
8e8320c9 JA |
2168 | /* |
2169 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2170 | * the queue isn't live yet. | |
2171 | */ | |
2404e607 | 2172 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2173 | { |
2174 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2175 | int i; |
2176 | ||
2404e607 | 2177 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2178 | if (shared) { |
2179 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2180 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2181 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2182 | } else { |
2183 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2184 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2185 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2186 | } |
2404e607 JM |
2187 | } |
2188 | } | |
2189 | ||
8e8320c9 JA |
2190 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2191 | bool shared) | |
2404e607 JM |
2192 | { |
2193 | struct request_queue *q; | |
0d2602ca | 2194 | |
705cda97 BVA |
2195 | lockdep_assert_held(&set->tag_list_lock); |
2196 | ||
0d2602ca JA |
2197 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2198 | blk_mq_freeze_queue(q); | |
2404e607 | 2199 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2200 | blk_mq_unfreeze_queue(q); |
2201 | } | |
2202 | } | |
2203 | ||
2204 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2205 | { | |
2206 | struct blk_mq_tag_set *set = q->tag_set; | |
2207 | ||
0d2602ca | 2208 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2209 | list_del_rcu(&q->tag_set_list); |
2210 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2211 | if (list_is_singular(&set->tag_list)) { |
2212 | /* just transitioned to unshared */ | |
2213 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2214 | /* update existing queue */ | |
2215 | blk_mq_update_tag_set_depth(set, false); | |
2216 | } | |
0d2602ca | 2217 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2218 | |
2219 | synchronize_rcu(); | |
0d2602ca JA |
2220 | } |
2221 | ||
2222 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2223 | struct request_queue *q) | |
2224 | { | |
2225 | q->tag_set = set; | |
2226 | ||
2227 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2228 | |
2229 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2230 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2231 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2232 | /* update existing queue */ | |
2233 | blk_mq_update_tag_set_depth(set, true); | |
2234 | } | |
2235 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2236 | queue_set_hctx_shared(q, true); | |
705cda97 | 2237 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2238 | |
0d2602ca JA |
2239 | mutex_unlock(&set->tag_list_lock); |
2240 | } | |
2241 | ||
e09aae7e ML |
2242 | /* |
2243 | * It is the actual release handler for mq, but we do it from | |
2244 | * request queue's release handler for avoiding use-after-free | |
2245 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2246 | * but we can't group ctx/kctx kobj without it. | |
2247 | */ | |
2248 | void blk_mq_release(struct request_queue *q) | |
2249 | { | |
2250 | struct blk_mq_hw_ctx *hctx; | |
2251 | unsigned int i; | |
2252 | ||
2253 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2254 | queue_for_each_hw_ctx(q, hctx, i) { |
2255 | if (!hctx) | |
2256 | continue; | |
6c8b232e | 2257 | kobject_put(&hctx->kobj); |
c3b4afca | 2258 | } |
e09aae7e | 2259 | |
a723bab3 AM |
2260 | q->mq_map = NULL; |
2261 | ||
e09aae7e ML |
2262 | kfree(q->queue_hw_ctx); |
2263 | ||
7ea5fe31 ML |
2264 | /* |
2265 | * release .mq_kobj and sw queue's kobject now because | |
2266 | * both share lifetime with request queue. | |
2267 | */ | |
2268 | blk_mq_sysfs_deinit(q); | |
2269 | ||
e09aae7e ML |
2270 | free_percpu(q->queue_ctx); |
2271 | } | |
2272 | ||
24d2f903 | 2273 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2274 | { |
2275 | struct request_queue *uninit_q, *q; | |
2276 | ||
2277 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2278 | if (!uninit_q) | |
2279 | return ERR_PTR(-ENOMEM); | |
2280 | ||
2281 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2282 | if (IS_ERR(q)) | |
2283 | blk_cleanup_queue(uninit_q); | |
2284 | ||
2285 | return q; | |
2286 | } | |
2287 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2288 | ||
07319678 BVA |
2289 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2290 | { | |
2291 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2292 | ||
2293 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2294 | __alignof__(struct blk_mq_hw_ctx)) != | |
2295 | sizeof(struct blk_mq_hw_ctx)); | |
2296 | ||
2297 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2298 | hw_ctx_size += sizeof(struct srcu_struct); | |
2299 | ||
2300 | return hw_ctx_size; | |
2301 | } | |
2302 | ||
868f2f0b KB |
2303 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2304 | struct request_queue *q) | |
320ae51f | 2305 | { |
868f2f0b KB |
2306 | int i, j; |
2307 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2308 | |
868f2f0b | 2309 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2310 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2311 | int node; |
f14bbe77 | 2312 | |
868f2f0b KB |
2313 | if (hctxs[i]) |
2314 | continue; | |
2315 | ||
2316 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2317 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2318 | GFP_KERNEL, node); |
320ae51f | 2319 | if (!hctxs[i]) |
868f2f0b | 2320 | break; |
320ae51f | 2321 | |
a86073e4 | 2322 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2323 | node)) { |
2324 | kfree(hctxs[i]); | |
2325 | hctxs[i] = NULL; | |
2326 | break; | |
2327 | } | |
e4043dcf | 2328 | |
0d2602ca | 2329 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2330 | hctxs[i]->numa_node = node; |
320ae51f | 2331 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2332 | |
2333 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2334 | free_cpumask_var(hctxs[i]->cpumask); | |
2335 | kfree(hctxs[i]); | |
2336 | hctxs[i] = NULL; | |
2337 | break; | |
2338 | } | |
2339 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2340 | } |
868f2f0b KB |
2341 | for (j = i; j < q->nr_hw_queues; j++) { |
2342 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2343 | ||
2344 | if (hctx) { | |
cc71a6f4 JA |
2345 | if (hctx->tags) |
2346 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2347 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2348 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2349 | hctxs[j] = NULL; |
2350 | ||
2351 | } | |
2352 | } | |
2353 | q->nr_hw_queues = i; | |
2354 | blk_mq_sysfs_register(q); | |
2355 | } | |
2356 | ||
2357 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2358 | struct request_queue *q) | |
2359 | { | |
66841672 ML |
2360 | /* mark the queue as mq asap */ |
2361 | q->mq_ops = set->ops; | |
2362 | ||
34dbad5d | 2363 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2364 | blk_mq_poll_stats_bkt, |
2365 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2366 | if (!q->poll_cb) |
2367 | goto err_exit; | |
2368 | ||
868f2f0b KB |
2369 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2370 | if (!q->queue_ctx) | |
c7de5726 | 2371 | goto err_exit; |
868f2f0b | 2372 | |
737f98cf ML |
2373 | /* init q->mq_kobj and sw queues' kobjects */ |
2374 | blk_mq_sysfs_init(q); | |
2375 | ||
868f2f0b KB |
2376 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2377 | GFP_KERNEL, set->numa_node); | |
2378 | if (!q->queue_hw_ctx) | |
2379 | goto err_percpu; | |
2380 | ||
bdd17e75 | 2381 | q->mq_map = set->mq_map; |
868f2f0b KB |
2382 | |
2383 | blk_mq_realloc_hw_ctxs(set, q); | |
2384 | if (!q->nr_hw_queues) | |
2385 | goto err_hctxs; | |
320ae51f | 2386 | |
287922eb | 2387 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2388 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2389 | |
2390 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2391 | |
94eddfbe | 2392 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2393 | |
05f1dd53 JA |
2394 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2395 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2396 | ||
1be036e9 CH |
2397 | q->sg_reserved_size = INT_MAX; |
2398 | ||
2849450a | 2399 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2400 | INIT_LIST_HEAD(&q->requeue_list); |
2401 | spin_lock_init(&q->requeue_lock); | |
2402 | ||
254d259d | 2403 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2404 | |
eba71768 JA |
2405 | /* |
2406 | * Do this after blk_queue_make_request() overrides it... | |
2407 | */ | |
2408 | q->nr_requests = set->queue_depth; | |
2409 | ||
64f1c21e JA |
2410 | /* |
2411 | * Default to classic polling | |
2412 | */ | |
2413 | q->poll_nsec = -1; | |
2414 | ||
24d2f903 CH |
2415 | if (set->ops->complete) |
2416 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2417 | |
24d2f903 | 2418 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2419 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2420 | blk_mq_map_swqueue(q); |
4593fdbe | 2421 | |
d3484991 JA |
2422 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2423 | int ret; | |
2424 | ||
2425 | ret = blk_mq_sched_init(q); | |
2426 | if (ret) | |
2427 | return ERR_PTR(ret); | |
2428 | } | |
2429 | ||
320ae51f | 2430 | return q; |
18741986 | 2431 | |
320ae51f | 2432 | err_hctxs: |
868f2f0b | 2433 | kfree(q->queue_hw_ctx); |
320ae51f | 2434 | err_percpu: |
868f2f0b | 2435 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2436 | err_exit: |
2437 | q->mq_ops = NULL; | |
320ae51f JA |
2438 | return ERR_PTR(-ENOMEM); |
2439 | } | |
b62c21b7 | 2440 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2441 | |
2442 | void blk_mq_free_queue(struct request_queue *q) | |
2443 | { | |
624dbe47 | 2444 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2445 | |
0d2602ca | 2446 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2447 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2448 | } |
320ae51f JA |
2449 | |
2450 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2451 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2452 | { |
4ecd4fef | 2453 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2454 | |
9c1051aa | 2455 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2456 | blk_mq_sysfs_unregister(q); |
2457 | ||
320ae51f JA |
2458 | /* |
2459 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2460 | * we should change hctx numa_node according to new topology (this | |
2461 | * involves free and re-allocate memory, worthy doing?) | |
2462 | */ | |
2463 | ||
4b855ad3 | 2464 | blk_mq_map_swqueue(q); |
320ae51f | 2465 | |
67aec14c | 2466 | blk_mq_sysfs_register(q); |
9c1051aa | 2467 | blk_mq_debugfs_register_hctxs(q); |
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); | |
4b855ad3 | 2680 | blk_mq_queue_reinit(q); |
868f2f0b KB |
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 | ||
320ae51f JA |
2896 | static int __init blk_mq_init(void) |
2897 | { | |
9467f859 TG |
2898 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2899 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
2900 | return 0; |
2901 | } | |
2902 | subsys_initcall(blk_mq_init); |