]>
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 | |
85acb3ba SL |
484 | if (blk_rq_rl(rq)) |
485 | blk_put_rl(blk_rq_rl(rq)); | |
486 | ||
af76e555 | 487 | clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags); |
06426adf | 488 | clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); |
bd166ef1 JA |
489 | if (rq->tag != -1) |
490 | blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag); | |
491 | if (sched_tag != -1) | |
c05f8525 | 492 | blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag); |
6d8c6c0f | 493 | blk_mq_sched_restart(hctx); |
3ef28e83 | 494 | blk_queue_exit(q); |
320ae51f | 495 | } |
1a3b595a | 496 | EXPORT_SYMBOL_GPL(blk_mq_free_request); |
320ae51f | 497 | |
2a842aca | 498 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) |
320ae51f | 499 | { |
0d11e6ac ML |
500 | blk_account_io_done(rq); |
501 | ||
91b63639 | 502 | if (rq->end_io) { |
87760e5e | 503 | wbt_done(rq->q->rq_wb, &rq->issue_stat); |
320ae51f | 504 | rq->end_io(rq, error); |
91b63639 CH |
505 | } else { |
506 | if (unlikely(blk_bidi_rq(rq))) | |
507 | blk_mq_free_request(rq->next_rq); | |
320ae51f | 508 | blk_mq_free_request(rq); |
91b63639 | 509 | } |
320ae51f | 510 | } |
c8a446ad | 511 | EXPORT_SYMBOL(__blk_mq_end_request); |
63151a44 | 512 | |
2a842aca | 513 | void blk_mq_end_request(struct request *rq, blk_status_t error) |
63151a44 CH |
514 | { |
515 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
516 | BUG(); | |
c8a446ad | 517 | __blk_mq_end_request(rq, error); |
63151a44 | 518 | } |
c8a446ad | 519 | EXPORT_SYMBOL(blk_mq_end_request); |
320ae51f | 520 | |
30a91cb4 | 521 | static void __blk_mq_complete_request_remote(void *data) |
320ae51f | 522 | { |
3d6efbf6 | 523 | struct request *rq = data; |
320ae51f | 524 | |
30a91cb4 | 525 | rq->q->softirq_done_fn(rq); |
320ae51f | 526 | } |
320ae51f | 527 | |
453f8341 | 528 | static void __blk_mq_complete_request(struct request *rq) |
320ae51f JA |
529 | { |
530 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
38535201 | 531 | bool shared = false; |
320ae51f JA |
532 | int cpu; |
533 | ||
453f8341 CH |
534 | if (rq->internal_tag != -1) |
535 | blk_mq_sched_completed_request(rq); | |
536 | if (rq->rq_flags & RQF_STATS) { | |
537 | blk_mq_poll_stats_start(rq->q); | |
538 | blk_stat_add(rq); | |
539 | } | |
540 | ||
38535201 | 541 | if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) { |
30a91cb4 CH |
542 | rq->q->softirq_done_fn(rq); |
543 | return; | |
544 | } | |
320ae51f JA |
545 | |
546 | cpu = get_cpu(); | |
38535201 CH |
547 | if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags)) |
548 | shared = cpus_share_cache(cpu, ctx->cpu); | |
549 | ||
550 | if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) { | |
30a91cb4 | 551 | rq->csd.func = __blk_mq_complete_request_remote; |
3d6efbf6 CH |
552 | rq->csd.info = rq; |
553 | rq->csd.flags = 0; | |
c46fff2a | 554 | smp_call_function_single_async(ctx->cpu, &rq->csd); |
3d6efbf6 | 555 | } else { |
30a91cb4 | 556 | rq->q->softirq_done_fn(rq); |
3d6efbf6 | 557 | } |
320ae51f JA |
558 | put_cpu(); |
559 | } | |
30a91cb4 CH |
560 | |
561 | /** | |
562 | * blk_mq_complete_request - end I/O on a request | |
563 | * @rq: the request being processed | |
564 | * | |
565 | * Description: | |
566 | * Ends all I/O on a request. It does not handle partial completions. | |
567 | * The actual completion happens out-of-order, through a IPI handler. | |
568 | **/ | |
08e0029a | 569 | void blk_mq_complete_request(struct request *rq) |
30a91cb4 | 570 | { |
95f09684 JA |
571 | struct request_queue *q = rq->q; |
572 | ||
573 | if (unlikely(blk_should_fake_timeout(q))) | |
30a91cb4 | 574 | return; |
08e0029a | 575 | if (!blk_mark_rq_complete(rq)) |
ed851860 | 576 | __blk_mq_complete_request(rq); |
30a91cb4 CH |
577 | } |
578 | EXPORT_SYMBOL(blk_mq_complete_request); | |
320ae51f | 579 | |
973c0191 KB |
580 | int blk_mq_request_started(struct request *rq) |
581 | { | |
582 | return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
583 | } | |
584 | EXPORT_SYMBOL_GPL(blk_mq_request_started); | |
585 | ||
e2490073 | 586 | void blk_mq_start_request(struct request *rq) |
320ae51f JA |
587 | { |
588 | struct request_queue *q = rq->q; | |
589 | ||
bd166ef1 JA |
590 | blk_mq_sched_started_request(rq); |
591 | ||
320ae51f JA |
592 | trace_block_rq_issue(q, rq); |
593 | ||
cf43e6be | 594 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { |
88eeca49 | 595 | blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq)); |
cf43e6be | 596 | rq->rq_flags |= RQF_STATS; |
87760e5e | 597 | wbt_issue(q->rq_wb, &rq->issue_stat); |
cf43e6be JA |
598 | } |
599 | ||
2b8393b4 | 600 | blk_add_timer(rq); |
87ee7b11 | 601 | |
a7af0af3 | 602 | WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)); |
538b7534 | 603 | |
87ee7b11 JA |
604 | /* |
605 | * Mark us as started and clear complete. Complete might have been | |
606 | * set if requeue raced with timeout, which then marked it as | |
607 | * complete. So be sure to clear complete again when we start | |
608 | * the request, otherwise we'll ignore the completion event. | |
a7af0af3 PZ |
609 | * |
610 | * Ensure that ->deadline is visible before we set STARTED, such that | |
611 | * blk_mq_check_expired() is guaranteed to observe our ->deadline when | |
612 | * it observes STARTED. | |
87ee7b11 | 613 | */ |
a7af0af3 PZ |
614 | smp_wmb(); |
615 | set_bit(REQ_ATOM_STARTED, &rq->atomic_flags); | |
616 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) { | |
617 | /* | |
618 | * Coherence order guarantees these consecutive stores to a | |
619 | * single variable propagate in the specified order. Thus the | |
620 | * clear_bit() is ordered _after_ the set bit. See | |
621 | * blk_mq_check_expired(). | |
622 | * | |
623 | * (the bits must be part of the same byte for this to be | |
624 | * true). | |
625 | */ | |
4b570521 | 626 | clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags); |
a7af0af3 | 627 | } |
49f5baa5 CH |
628 | |
629 | if (q->dma_drain_size && blk_rq_bytes(rq)) { | |
630 | /* | |
631 | * Make sure space for the drain appears. We know we can do | |
632 | * this because max_hw_segments has been adjusted to be one | |
633 | * fewer than the device can handle. | |
634 | */ | |
635 | rq->nr_phys_segments++; | |
636 | } | |
320ae51f | 637 | } |
e2490073 | 638 | EXPORT_SYMBOL(blk_mq_start_request); |
320ae51f | 639 | |
d9d149a3 ML |
640 | /* |
641 | * When we reach here because queue is busy, REQ_ATOM_COMPLETE | |
48b99c9d | 642 | * flag isn't set yet, so there may be race with timeout handler, |
d9d149a3 ML |
643 | * but given rq->deadline is just set in .queue_rq() under |
644 | * this situation, the race won't be possible in reality because | |
645 | * rq->timeout should be set as big enough to cover the window | |
646 | * between blk_mq_start_request() called from .queue_rq() and | |
647 | * clearing REQ_ATOM_STARTED here. | |
648 | */ | |
ed0791b2 | 649 | static void __blk_mq_requeue_request(struct request *rq) |
320ae51f JA |
650 | { |
651 | struct request_queue *q = rq->q; | |
652 | ||
653 | trace_block_rq_requeue(q, rq); | |
87760e5e | 654 | wbt_requeue(q->rq_wb, &rq->issue_stat); |
bd166ef1 | 655 | blk_mq_sched_requeue_request(rq); |
49f5baa5 | 656 | |
e2490073 CH |
657 | if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) { |
658 | if (q->dma_drain_size && blk_rq_bytes(rq)) | |
659 | rq->nr_phys_segments--; | |
660 | } | |
320ae51f JA |
661 | } |
662 | ||
2b053aca | 663 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) |
ed0791b2 | 664 | { |
ed0791b2 | 665 | __blk_mq_requeue_request(rq); |
ed0791b2 | 666 | |
ed0791b2 | 667 | BUG_ON(blk_queued_rq(rq)); |
2b053aca | 668 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); |
ed0791b2 CH |
669 | } |
670 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
671 | ||
6fca6a61 CH |
672 | static void blk_mq_requeue_work(struct work_struct *work) |
673 | { | |
674 | struct request_queue *q = | |
2849450a | 675 | container_of(work, struct request_queue, requeue_work.work); |
6fca6a61 CH |
676 | LIST_HEAD(rq_list); |
677 | struct request *rq, *next; | |
6fca6a61 | 678 | |
18e9781d | 679 | spin_lock_irq(&q->requeue_lock); |
6fca6a61 | 680 | list_splice_init(&q->requeue_list, &rq_list); |
18e9781d | 681 | spin_unlock_irq(&q->requeue_lock); |
6fca6a61 CH |
682 | |
683 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
e8064021 | 684 | if (!(rq->rq_flags & RQF_SOFTBARRIER)) |
6fca6a61 CH |
685 | continue; |
686 | ||
e8064021 | 687 | rq->rq_flags &= ~RQF_SOFTBARRIER; |
6fca6a61 | 688 | list_del_init(&rq->queuelist); |
bd6737f1 | 689 | blk_mq_sched_insert_request(rq, true, false, false, true); |
6fca6a61 CH |
690 | } |
691 | ||
692 | while (!list_empty(&rq_list)) { | |
693 | rq = list_entry(rq_list.next, struct request, queuelist); | |
694 | list_del_init(&rq->queuelist); | |
bd6737f1 | 695 | blk_mq_sched_insert_request(rq, false, false, false, true); |
6fca6a61 CH |
696 | } |
697 | ||
52d7f1b5 | 698 | blk_mq_run_hw_queues(q, false); |
6fca6a61 CH |
699 | } |
700 | ||
2b053aca BVA |
701 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, |
702 | bool kick_requeue_list) | |
6fca6a61 CH |
703 | { |
704 | struct request_queue *q = rq->q; | |
705 | unsigned long flags; | |
706 | ||
707 | /* | |
708 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
709 | * request head insertation from the workqueue. | |
710 | */ | |
e8064021 | 711 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); |
6fca6a61 CH |
712 | |
713 | spin_lock_irqsave(&q->requeue_lock, flags); | |
714 | if (at_head) { | |
e8064021 | 715 | rq->rq_flags |= RQF_SOFTBARRIER; |
6fca6a61 CH |
716 | list_add(&rq->queuelist, &q->requeue_list); |
717 | } else { | |
718 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
719 | } | |
720 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
2b053aca BVA |
721 | |
722 | if (kick_requeue_list) | |
723 | blk_mq_kick_requeue_list(q); | |
6fca6a61 CH |
724 | } |
725 | EXPORT_SYMBOL(blk_mq_add_to_requeue_list); | |
726 | ||
727 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
728 | { | |
2849450a | 729 | kblockd_schedule_delayed_work(&q->requeue_work, 0); |
6fca6a61 CH |
730 | } |
731 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
732 | ||
2849450a MS |
733 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, |
734 | unsigned long msecs) | |
735 | { | |
d4acf365 BVA |
736 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, |
737 | msecs_to_jiffies(msecs)); | |
2849450a MS |
738 | } |
739 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
740 | ||
0e62f51f JA |
741 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) |
742 | { | |
88c7b2b7 JA |
743 | if (tag < tags->nr_tags) { |
744 | prefetch(tags->rqs[tag]); | |
4ee86bab | 745 | return tags->rqs[tag]; |
88c7b2b7 | 746 | } |
4ee86bab HR |
747 | |
748 | return NULL; | |
24d2f903 CH |
749 | } |
750 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
751 | ||
320ae51f | 752 | struct blk_mq_timeout_data { |
46f92d42 CH |
753 | unsigned long next; |
754 | unsigned int next_set; | |
320ae51f JA |
755 | }; |
756 | ||
90415837 | 757 | void blk_mq_rq_timed_out(struct request *req, bool reserved) |
320ae51f | 758 | { |
f8a5b122 | 759 | const struct blk_mq_ops *ops = req->q->mq_ops; |
46f92d42 | 760 | enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER; |
87ee7b11 JA |
761 | |
762 | /* | |
763 | * We know that complete is set at this point. If STARTED isn't set | |
764 | * anymore, then the request isn't active and the "timeout" should | |
765 | * just be ignored. This can happen due to the bitflag ordering. | |
766 | * Timeout first checks if STARTED is set, and if it is, assumes | |
767 | * the request is active. But if we race with completion, then | |
48b99c9d | 768 | * both flags will get cleared. So check here again, and ignore |
87ee7b11 JA |
769 | * a timeout event with a request that isn't active. |
770 | */ | |
46f92d42 CH |
771 | if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags)) |
772 | return; | |
87ee7b11 | 773 | |
46f92d42 | 774 | if (ops->timeout) |
0152fb6b | 775 | ret = ops->timeout(req, reserved); |
46f92d42 CH |
776 | |
777 | switch (ret) { | |
778 | case BLK_EH_HANDLED: | |
779 | __blk_mq_complete_request(req); | |
780 | break; | |
781 | case BLK_EH_RESET_TIMER: | |
782 | blk_add_timer(req); | |
783 | blk_clear_rq_complete(req); | |
784 | break; | |
785 | case BLK_EH_NOT_HANDLED: | |
786 | break; | |
787 | default: | |
788 | printk(KERN_ERR "block: bad eh return: %d\n", ret); | |
789 | break; | |
790 | } | |
87ee7b11 | 791 | } |
5b3f25fc | 792 | |
81481eb4 CH |
793 | static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, |
794 | struct request *rq, void *priv, bool reserved) | |
795 | { | |
796 | struct blk_mq_timeout_data *data = priv; | |
a7af0af3 | 797 | unsigned long deadline; |
87ee7b11 | 798 | |
95a49603 | 799 | if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) |
46f92d42 | 800 | return; |
87ee7b11 | 801 | |
a7af0af3 PZ |
802 | /* |
803 | * Ensures that if we see STARTED we must also see our | |
804 | * up-to-date deadline, see blk_mq_start_request(). | |
805 | */ | |
806 | smp_rmb(); | |
807 | ||
808 | deadline = READ_ONCE(rq->deadline); | |
809 | ||
d9d149a3 ML |
810 | /* |
811 | * The rq being checked may have been freed and reallocated | |
812 | * out already here, we avoid this race by checking rq->deadline | |
813 | * and REQ_ATOM_COMPLETE flag together: | |
814 | * | |
815 | * - if rq->deadline is observed as new value because of | |
816 | * reusing, the rq won't be timed out because of timing. | |
817 | * - if rq->deadline is observed as previous value, | |
818 | * REQ_ATOM_COMPLETE flag won't be cleared in reuse path | |
819 | * because we put a barrier between setting rq->deadline | |
820 | * and clearing the flag in blk_mq_start_request(), so | |
821 | * this rq won't be timed out too. | |
822 | */ | |
a7af0af3 PZ |
823 | if (time_after_eq(jiffies, deadline)) { |
824 | if (!blk_mark_rq_complete(rq)) { | |
825 | /* | |
826 | * Again coherence order ensures that consecutive reads | |
827 | * from the same variable must be in that order. This | |
828 | * ensures that if we see COMPLETE clear, we must then | |
829 | * see STARTED set and we'll ignore this timeout. | |
830 | * | |
831 | * (There's also the MB implied by the test_and_clear()) | |
832 | */ | |
0152fb6b | 833 | blk_mq_rq_timed_out(rq, reserved); |
a7af0af3 PZ |
834 | } |
835 | } else if (!data->next_set || time_after(data->next, deadline)) { | |
836 | data->next = deadline; | |
46f92d42 CH |
837 | data->next_set = 1; |
838 | } | |
87ee7b11 JA |
839 | } |
840 | ||
287922eb | 841 | static void blk_mq_timeout_work(struct work_struct *work) |
320ae51f | 842 | { |
287922eb CH |
843 | struct request_queue *q = |
844 | container_of(work, struct request_queue, timeout_work); | |
81481eb4 CH |
845 | struct blk_mq_timeout_data data = { |
846 | .next = 0, | |
847 | .next_set = 0, | |
848 | }; | |
81481eb4 | 849 | int i; |
320ae51f | 850 | |
71f79fb3 GKB |
851 | /* A deadlock might occur if a request is stuck requiring a |
852 | * timeout at the same time a queue freeze is waiting | |
853 | * completion, since the timeout code would not be able to | |
854 | * acquire the queue reference here. | |
855 | * | |
856 | * That's why we don't use blk_queue_enter here; instead, we use | |
857 | * percpu_ref_tryget directly, because we need to be able to | |
858 | * obtain a reference even in the short window between the queue | |
859 | * starting to freeze, by dropping the first reference in | |
1671d522 | 860 | * blk_freeze_queue_start, and the moment the last request is |
71f79fb3 GKB |
861 | * consumed, marked by the instant q_usage_counter reaches |
862 | * zero. | |
863 | */ | |
864 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
287922eb CH |
865 | return; |
866 | ||
0bf6cd5b | 867 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data); |
320ae51f | 868 | |
81481eb4 CH |
869 | if (data.next_set) { |
870 | data.next = blk_rq_timeout(round_jiffies_up(data.next)); | |
871 | mod_timer(&q->timeout, data.next); | |
0d2602ca | 872 | } else { |
0bf6cd5b CH |
873 | struct blk_mq_hw_ctx *hctx; |
874 | ||
f054b56c ML |
875 | queue_for_each_hw_ctx(q, hctx, i) { |
876 | /* the hctx may be unmapped, so check it here */ | |
877 | if (blk_mq_hw_queue_mapped(hctx)) | |
878 | blk_mq_tag_idle(hctx); | |
879 | } | |
0d2602ca | 880 | } |
287922eb | 881 | blk_queue_exit(q); |
320ae51f JA |
882 | } |
883 | ||
88459642 OS |
884 | struct flush_busy_ctx_data { |
885 | struct blk_mq_hw_ctx *hctx; | |
886 | struct list_head *list; | |
887 | }; | |
888 | ||
889 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
890 | { | |
891 | struct flush_busy_ctx_data *flush_data = data; | |
892 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
893 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
894 | ||
895 | sbitmap_clear_bit(sb, bitnr); | |
896 | spin_lock(&ctx->lock); | |
897 | list_splice_tail_init(&ctx->rq_list, flush_data->list); | |
898 | spin_unlock(&ctx->lock); | |
899 | return true; | |
900 | } | |
901 | ||
1429d7c9 JA |
902 | /* |
903 | * Process software queues that have been marked busy, splicing them | |
904 | * to the for-dispatch | |
905 | */ | |
2c3ad667 | 906 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) |
1429d7c9 | 907 | { |
88459642 OS |
908 | struct flush_busy_ctx_data data = { |
909 | .hctx = hctx, | |
910 | .list = list, | |
911 | }; | |
1429d7c9 | 912 | |
88459642 | 913 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); |
1429d7c9 | 914 | } |
2c3ad667 | 915 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); |
1429d7c9 | 916 | |
703fd1c0 JA |
917 | static inline unsigned int queued_to_index(unsigned int queued) |
918 | { | |
919 | if (!queued) | |
920 | return 0; | |
1429d7c9 | 921 | |
703fd1c0 | 922 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); |
1429d7c9 JA |
923 | } |
924 | ||
bd6737f1 JA |
925 | bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx, |
926 | bool wait) | |
bd166ef1 JA |
927 | { |
928 | struct blk_mq_alloc_data data = { | |
929 | .q = rq->q, | |
bd166ef1 JA |
930 | .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), |
931 | .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT, | |
932 | }; | |
933 | ||
5feeacdd JA |
934 | might_sleep_if(wait); |
935 | ||
81380ca1 OS |
936 | if (rq->tag != -1) |
937 | goto done; | |
bd166ef1 | 938 | |
415b806d SG |
939 | if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag)) |
940 | data.flags |= BLK_MQ_REQ_RESERVED; | |
941 | ||
bd166ef1 JA |
942 | rq->tag = blk_mq_get_tag(&data); |
943 | if (rq->tag >= 0) { | |
200e86b3 JA |
944 | if (blk_mq_tag_busy(data.hctx)) { |
945 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
946 | atomic_inc(&data.hctx->nr_active); | |
947 | } | |
bd166ef1 | 948 | data.hctx->tags->rqs[rq->tag] = rq; |
bd166ef1 JA |
949 | } |
950 | ||
81380ca1 OS |
951 | done: |
952 | if (hctx) | |
953 | *hctx = data.hctx; | |
954 | return rq->tag != -1; | |
bd166ef1 JA |
955 | } |
956 | ||
113285b4 JA |
957 | static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, |
958 | struct request *rq) | |
99cf1dc5 | 959 | { |
99cf1dc5 JA |
960 | blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag); |
961 | rq->tag = -1; | |
962 | ||
963 | if (rq->rq_flags & RQF_MQ_INFLIGHT) { | |
964 | rq->rq_flags &= ~RQF_MQ_INFLIGHT; | |
965 | atomic_dec(&hctx->nr_active); | |
966 | } | |
967 | } | |
968 | ||
113285b4 JA |
969 | static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx, |
970 | struct request *rq) | |
971 | { | |
972 | if (rq->tag == -1 || rq->internal_tag == -1) | |
973 | return; | |
974 | ||
975 | __blk_mq_put_driver_tag(hctx, rq); | |
976 | } | |
977 | ||
978 | static void blk_mq_put_driver_tag(struct request *rq) | |
979 | { | |
980 | struct blk_mq_hw_ctx *hctx; | |
981 | ||
982 | if (rq->tag == -1 || rq->internal_tag == -1) | |
983 | return; | |
984 | ||
985 | hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu); | |
986 | __blk_mq_put_driver_tag(hctx, rq); | |
987 | } | |
988 | ||
bd166ef1 JA |
989 | /* |
990 | * If we fail getting a driver tag because all the driver tags are already | |
991 | * assigned and on the dispatch list, BUT the first entry does not have a | |
992 | * tag, then we could deadlock. For that case, move entries with assigned | |
993 | * driver tags to the front, leaving the set of tagged requests in the | |
994 | * same order, and the untagged set in the same order. | |
995 | */ | |
996 | static bool reorder_tags_to_front(struct list_head *list) | |
997 | { | |
998 | struct request *rq, *tmp, *first = NULL; | |
999 | ||
1000 | list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) { | |
1001 | if (rq == first) | |
1002 | break; | |
1003 | if (rq->tag != -1) { | |
1004 | list_move(&rq->queuelist, list); | |
1005 | if (!first) | |
1006 | first = rq; | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | return first != NULL; | |
1011 | } | |
1012 | ||
ac6424b9 | 1013 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags, |
da55f2cc OS |
1014 | void *key) |
1015 | { | |
1016 | struct blk_mq_hw_ctx *hctx; | |
1017 | ||
1018 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1019 | ||
2055da97 | 1020 | list_del(&wait->entry); |
da55f2cc OS |
1021 | clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state); |
1022 | blk_mq_run_hw_queue(hctx, true); | |
1023 | return 1; | |
1024 | } | |
1025 | ||
1026 | static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx) | |
1027 | { | |
1028 | struct sbq_wait_state *ws; | |
1029 | ||
1030 | /* | |
1031 | * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait. | |
1032 | * The thread which wins the race to grab this bit adds the hardware | |
1033 | * queue to the wait queue. | |
1034 | */ | |
1035 | if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) || | |
1036 | test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
1037 | return false; | |
1038 | ||
1039 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
1040 | ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx); | |
1041 | ||
1042 | /* | |
1043 | * As soon as this returns, it's no longer safe to fiddle with | |
1044 | * hctx->dispatch_wait, since a completion can wake up the wait queue | |
1045 | * and unlock the bit. | |
1046 | */ | |
1047 | add_wait_queue(&ws->wait, &hctx->dispatch_wait); | |
1048 | return true; | |
1049 | } | |
1050 | ||
de148297 ML |
1051 | bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list, |
1052 | bool got_budget) | |
320ae51f | 1053 | { |
81380ca1 | 1054 | struct blk_mq_hw_ctx *hctx; |
320ae51f | 1055 | struct request *rq; |
fc17b653 | 1056 | int errors, queued; |
320ae51f | 1057 | |
81380ca1 OS |
1058 | if (list_empty(list)) |
1059 | return false; | |
1060 | ||
de148297 ML |
1061 | WARN_ON(!list_is_singular(list) && got_budget); |
1062 | ||
320ae51f JA |
1063 | /* |
1064 | * Now process all the entries, sending them to the driver. | |
1065 | */ | |
93efe981 | 1066 | errors = queued = 0; |
81380ca1 | 1067 | do { |
74c45052 | 1068 | struct blk_mq_queue_data bd; |
fc17b653 | 1069 | blk_status_t ret; |
320ae51f | 1070 | |
f04c3df3 | 1071 | rq = list_first_entry(list, struct request, queuelist); |
bd166ef1 JA |
1072 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1073 | if (!queued && reorder_tags_to_front(list)) | |
1074 | continue; | |
3c782d67 JA |
1075 | |
1076 | /* | |
da55f2cc OS |
1077 | * The initial allocation attempt failed, so we need to |
1078 | * rerun the hardware queue when a tag is freed. | |
3c782d67 | 1079 | */ |
de148297 ML |
1080 | if (!blk_mq_dispatch_wait_add(hctx)) { |
1081 | if (got_budget) | |
1082 | blk_mq_put_dispatch_budget(hctx); | |
807b1041 | 1083 | break; |
de148297 | 1084 | } |
807b1041 OS |
1085 | |
1086 | /* | |
1087 | * It's possible that a tag was freed in the window | |
1088 | * between the allocation failure and adding the | |
1089 | * hardware queue to the wait queue. | |
1090 | */ | |
de148297 ML |
1091 | if (!blk_mq_get_driver_tag(rq, &hctx, false)) { |
1092 | if (got_budget) | |
1093 | blk_mq_put_dispatch_budget(hctx); | |
1094 | break; | |
1095 | } | |
1096 | } | |
1097 | ||
1098 | if (!got_budget) { | |
1099 | ret = blk_mq_get_dispatch_budget(hctx); | |
1100 | if (ret == BLK_STS_RESOURCE) | |
3c782d67 | 1101 | break; |
de148297 ML |
1102 | if (ret != BLK_STS_OK) |
1103 | goto fail_rq; | |
bd166ef1 | 1104 | } |
da55f2cc | 1105 | |
320ae51f | 1106 | list_del_init(&rq->queuelist); |
320ae51f | 1107 | |
74c45052 | 1108 | bd.rq = rq; |
113285b4 JA |
1109 | |
1110 | /* | |
1111 | * Flag last if we have no more requests, or if we have more | |
1112 | * but can't assign a driver tag to it. | |
1113 | */ | |
1114 | if (list_empty(list)) | |
1115 | bd.last = true; | |
1116 | else { | |
1117 | struct request *nxt; | |
1118 | ||
1119 | nxt = list_first_entry(list, struct request, queuelist); | |
1120 | bd.last = !blk_mq_get_driver_tag(nxt, NULL, false); | |
1121 | } | |
74c45052 JA |
1122 | |
1123 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 | 1124 | if (ret == BLK_STS_RESOURCE) { |
113285b4 | 1125 | blk_mq_put_driver_tag_hctx(hctx, rq); |
f04c3df3 | 1126 | list_add(&rq->queuelist, list); |
ed0791b2 | 1127 | __blk_mq_requeue_request(rq); |
320ae51f | 1128 | break; |
fc17b653 CH |
1129 | } |
1130 | ||
de148297 | 1131 | fail_rq: |
fc17b653 | 1132 | if (unlikely(ret != BLK_STS_OK)) { |
93efe981 | 1133 | errors++; |
2a842aca | 1134 | blk_mq_end_request(rq, BLK_STS_IOERR); |
fc17b653 | 1135 | continue; |
320ae51f JA |
1136 | } |
1137 | ||
fc17b653 | 1138 | queued++; |
81380ca1 | 1139 | } while (!list_empty(list)); |
320ae51f | 1140 | |
703fd1c0 | 1141 | hctx->dispatched[queued_to_index(queued)]++; |
320ae51f JA |
1142 | |
1143 | /* | |
1144 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1145 | * that is where we will continue on next queue run. | |
1146 | */ | |
f04c3df3 | 1147 | if (!list_empty(list)) { |
113285b4 | 1148 | /* |
710c785f BVA |
1149 | * If an I/O scheduler has been configured and we got a driver |
1150 | * tag for the next request already, free it again. | |
113285b4 JA |
1151 | */ |
1152 | rq = list_first_entry(list, struct request, queuelist); | |
1153 | blk_mq_put_driver_tag(rq); | |
1154 | ||
320ae51f | 1155 | spin_lock(&hctx->lock); |
c13660a0 | 1156 | list_splice_init(list, &hctx->dispatch); |
320ae51f | 1157 | spin_unlock(&hctx->lock); |
f04c3df3 | 1158 | |
9ba52e58 | 1159 | /* |
710c785f BVA |
1160 | * If SCHED_RESTART was set by the caller of this function and |
1161 | * it is no longer set that means that it was cleared by another | |
1162 | * thread and hence that a queue rerun is needed. | |
9ba52e58 | 1163 | * |
710c785f BVA |
1164 | * If TAG_WAITING is set that means that an I/O scheduler has |
1165 | * been configured and another thread is waiting for a driver | |
1166 | * tag. To guarantee fairness, do not rerun this hardware queue | |
1167 | * but let the other thread grab the driver tag. | |
bd166ef1 | 1168 | * |
710c785f BVA |
1169 | * If no I/O scheduler has been configured it is possible that |
1170 | * the hardware queue got stopped and restarted before requests | |
1171 | * were pushed back onto the dispatch list. Rerun the queue to | |
1172 | * avoid starvation. Notes: | |
1173 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1174 | * been stopped before rerunning a queue. | |
1175 | * - Some but not all block drivers stop a queue before | |
fc17b653 | 1176 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq |
710c785f | 1177 | * and dm-rq. |
bd166ef1 | 1178 | */ |
da55f2cc OS |
1179 | if (!blk_mq_sched_needs_restart(hctx) && |
1180 | !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) | |
bd166ef1 | 1181 | blk_mq_run_hw_queue(hctx, true); |
320ae51f | 1182 | } |
f04c3df3 | 1183 | |
93efe981 | 1184 | return (queued + errors) != 0; |
f04c3df3 JA |
1185 | } |
1186 | ||
6a83e74d BVA |
1187 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) |
1188 | { | |
1189 | int srcu_idx; | |
de148297 | 1190 | bool run_queue; |
6a83e74d | 1191 | |
b7a71e66 JA |
1192 | /* |
1193 | * We should be running this queue from one of the CPUs that | |
1194 | * are mapped to it. | |
1195 | */ | |
6a83e74d BVA |
1196 | WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && |
1197 | cpu_online(hctx->next_cpu)); | |
1198 | ||
b7a71e66 JA |
1199 | /* |
1200 | * We can't run the queue inline with ints disabled. Ensure that | |
1201 | * we catch bad users of this early. | |
1202 | */ | |
1203 | WARN_ON_ONCE(in_interrupt()); | |
1204 | ||
6a83e74d BVA |
1205 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { |
1206 | rcu_read_lock(); | |
de148297 | 1207 | run_queue = blk_mq_sched_dispatch_requests(hctx); |
6a83e74d BVA |
1208 | rcu_read_unlock(); |
1209 | } else { | |
bf4907c0 JA |
1210 | might_sleep(); |
1211 | ||
07319678 | 1212 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
de148297 | 1213 | run_queue = blk_mq_sched_dispatch_requests(hctx); |
07319678 | 1214 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
6a83e74d | 1215 | } |
de148297 ML |
1216 | |
1217 | if (run_queue) | |
1218 | blk_mq_run_hw_queue(hctx, true); | |
6a83e74d BVA |
1219 | } |
1220 | ||
506e931f JA |
1221 | /* |
1222 | * It'd be great if the workqueue API had a way to pass | |
1223 | * in a mask and had some smarts for more clever placement. | |
1224 | * For now we just round-robin here, switching for every | |
1225 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1226 | */ | |
1227 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1228 | { | |
b657d7e6 CH |
1229 | if (hctx->queue->nr_hw_queues == 1) |
1230 | return WORK_CPU_UNBOUND; | |
506e931f JA |
1231 | |
1232 | if (--hctx->next_cpu_batch <= 0) { | |
c02ebfdd | 1233 | int next_cpu; |
506e931f JA |
1234 | |
1235 | next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask); | |
1236 | if (next_cpu >= nr_cpu_ids) | |
1237 | next_cpu = cpumask_first(hctx->cpumask); | |
1238 | ||
1239 | hctx->next_cpu = next_cpu; | |
1240 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1241 | } | |
1242 | ||
b657d7e6 | 1243 | return hctx->next_cpu; |
506e931f JA |
1244 | } |
1245 | ||
7587a5ae BVA |
1246 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, |
1247 | unsigned long msecs) | |
320ae51f | 1248 | { |
5435c023 BVA |
1249 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
1250 | return; | |
1251 | ||
1252 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
320ae51f JA |
1253 | return; |
1254 | ||
1b792f2f | 1255 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { |
2a90d4aa PB |
1256 | int cpu = get_cpu(); |
1257 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
398205b8 | 1258 | __blk_mq_run_hw_queue(hctx); |
2a90d4aa | 1259 | put_cpu(); |
398205b8 PB |
1260 | return; |
1261 | } | |
e4043dcf | 1262 | |
2a90d4aa | 1263 | put_cpu(); |
e4043dcf | 1264 | } |
398205b8 | 1265 | |
9f993737 JA |
1266 | kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx), |
1267 | &hctx->run_work, | |
1268 | msecs_to_jiffies(msecs)); | |
7587a5ae BVA |
1269 | } |
1270 | ||
1271 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1272 | { | |
1273 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1274 | } | |
1275 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1276 | ||
1277 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1278 | { | |
1279 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
320ae51f | 1280 | } |
5b727272 | 1281 | EXPORT_SYMBOL(blk_mq_run_hw_queue); |
320ae51f | 1282 | |
b94ec296 | 1283 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1284 | { |
1285 | struct blk_mq_hw_ctx *hctx; | |
1286 | int i; | |
1287 | ||
1288 | queue_for_each_hw_ctx(q, hctx, i) { | |
bd166ef1 | 1289 | if (!blk_mq_hctx_has_pending(hctx) || |
5d1b25c1 | 1290 | blk_mq_hctx_stopped(hctx)) |
320ae51f JA |
1291 | continue; |
1292 | ||
b94ec296 | 1293 | blk_mq_run_hw_queue(hctx, async); |
320ae51f JA |
1294 | } |
1295 | } | |
b94ec296 | 1296 | EXPORT_SYMBOL(blk_mq_run_hw_queues); |
320ae51f | 1297 | |
fd001443 BVA |
1298 | /** |
1299 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1300 | * @q: request queue. | |
1301 | * | |
1302 | * The caller is responsible for serializing this function against | |
1303 | * blk_mq_{start,stop}_hw_queue(). | |
1304 | */ | |
1305 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1306 | { | |
1307 | struct blk_mq_hw_ctx *hctx; | |
1308 | int i; | |
1309 | ||
1310 | queue_for_each_hw_ctx(q, hctx, i) | |
1311 | if (blk_mq_hctx_stopped(hctx)) | |
1312 | return true; | |
1313 | ||
1314 | return false; | |
1315 | } | |
1316 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1317 | ||
39a70c76 ML |
1318 | /* |
1319 | * This function is often used for pausing .queue_rq() by driver when | |
1320 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1321 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1322 | * |
1323 | * We do not guarantee that dispatch can be drained or blocked | |
1324 | * after blk_mq_stop_hw_queue() returns. Please use | |
1325 | * blk_mq_quiesce_queue() for that requirement. | |
1326 | */ | |
2719aa21 JA |
1327 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) |
1328 | { | |
641a9ed6 | 1329 | cancel_delayed_work(&hctx->run_work); |
280d45f6 | 1330 | |
641a9ed6 | 1331 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); |
2719aa21 | 1332 | } |
641a9ed6 | 1333 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); |
2719aa21 | 1334 | |
39a70c76 ML |
1335 | /* |
1336 | * This function is often used for pausing .queue_rq() by driver when | |
1337 | * there isn't enough resource or some conditions aren't satisfied, and | |
4d606219 | 1338 | * BLK_STS_RESOURCE is usually returned. |
39a70c76 ML |
1339 | * |
1340 | * We do not guarantee that dispatch can be drained or blocked | |
1341 | * after blk_mq_stop_hw_queues() returns. Please use | |
1342 | * blk_mq_quiesce_queue() for that requirement. | |
1343 | */ | |
2719aa21 JA |
1344 | void blk_mq_stop_hw_queues(struct request_queue *q) |
1345 | { | |
641a9ed6 ML |
1346 | struct blk_mq_hw_ctx *hctx; |
1347 | int i; | |
1348 | ||
1349 | queue_for_each_hw_ctx(q, hctx, i) | |
1350 | blk_mq_stop_hw_queue(hctx); | |
280d45f6 CH |
1351 | } |
1352 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1353 | ||
320ae51f JA |
1354 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) |
1355 | { | |
1356 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
e4043dcf | 1357 | |
0ffbce80 | 1358 | blk_mq_run_hw_queue(hctx, false); |
320ae51f JA |
1359 | } |
1360 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1361 | ||
2f268556 CH |
1362 | void blk_mq_start_hw_queues(struct request_queue *q) |
1363 | { | |
1364 | struct blk_mq_hw_ctx *hctx; | |
1365 | int i; | |
1366 | ||
1367 | queue_for_each_hw_ctx(q, hctx, i) | |
1368 | blk_mq_start_hw_queue(hctx); | |
1369 | } | |
1370 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1371 | ||
ae911c5e JA |
1372 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) |
1373 | { | |
1374 | if (!blk_mq_hctx_stopped(hctx)) | |
1375 | return; | |
1376 | ||
1377 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1378 | blk_mq_run_hw_queue(hctx, async); | |
1379 | } | |
1380 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1381 | ||
1b4a3258 | 1382 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) |
320ae51f JA |
1383 | { |
1384 | struct blk_mq_hw_ctx *hctx; | |
1385 | int i; | |
1386 | ||
ae911c5e JA |
1387 | queue_for_each_hw_ctx(q, hctx, i) |
1388 | blk_mq_start_stopped_hw_queue(hctx, async); | |
320ae51f JA |
1389 | } |
1390 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1391 | ||
70f4db63 | 1392 | static void blk_mq_run_work_fn(struct work_struct *work) |
320ae51f JA |
1393 | { |
1394 | struct blk_mq_hw_ctx *hctx; | |
1395 | ||
9f993737 | 1396 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); |
320ae51f | 1397 | |
21c6e939 JA |
1398 | /* |
1399 | * If we are stopped, don't run the queue. The exception is if | |
1400 | * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear | |
1401 | * the STOPPED bit and run it. | |
1402 | */ | |
1403 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) { | |
1404 | if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state)) | |
1405 | return; | |
7587a5ae | 1406 | |
21c6e939 JA |
1407 | clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1408 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1409 | } | |
7587a5ae BVA |
1410 | |
1411 | __blk_mq_run_hw_queue(hctx); | |
1412 | } | |
1413 | ||
70f4db63 CH |
1414 | |
1415 | void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1416 | { | |
5435c023 | 1417 | if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx))) |
19c66e59 | 1418 | return; |
70f4db63 | 1419 | |
21c6e939 JA |
1420 | /* |
1421 | * Stop the hw queue, then modify currently delayed work. | |
1422 | * This should prevent us from running the queue prematurely. | |
1423 | * Mark the queue as auto-clearing STOPPED when it runs. | |
1424 | */ | |
7e79dadc | 1425 | blk_mq_stop_hw_queue(hctx); |
21c6e939 JA |
1426 | set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state); |
1427 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), | |
1428 | &hctx->run_work, | |
1429 | msecs_to_jiffies(msecs)); | |
70f4db63 CH |
1430 | } |
1431 | EXPORT_SYMBOL(blk_mq_delay_queue); | |
1432 | ||
cfd0c552 | 1433 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, |
cfd0c552 ML |
1434 | struct request *rq, |
1435 | bool at_head) | |
320ae51f | 1436 | { |
e57690fe JA |
1437 | struct blk_mq_ctx *ctx = rq->mq_ctx; |
1438 | ||
7b607814 BVA |
1439 | lockdep_assert_held(&ctx->lock); |
1440 | ||
01b983c9 JA |
1441 | trace_block_rq_insert(hctx->queue, rq); |
1442 | ||
72a0a36e CH |
1443 | if (at_head) |
1444 | list_add(&rq->queuelist, &ctx->rq_list); | |
1445 | else | |
1446 | list_add_tail(&rq->queuelist, &ctx->rq_list); | |
cfd0c552 | 1447 | } |
4bb659b1 | 1448 | |
2c3ad667 JA |
1449 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, |
1450 | bool at_head) | |
cfd0c552 ML |
1451 | { |
1452 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1453 | ||
7b607814 BVA |
1454 | lockdep_assert_held(&ctx->lock); |
1455 | ||
e57690fe | 1456 | __blk_mq_insert_req_list(hctx, rq, at_head); |
320ae51f | 1457 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f JA |
1458 | } |
1459 | ||
157f377b JA |
1460 | /* |
1461 | * Should only be used carefully, when the caller knows we want to | |
1462 | * bypass a potential IO scheduler on the target device. | |
1463 | */ | |
1464 | void blk_mq_request_bypass_insert(struct request *rq) | |
1465 | { | |
1466 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1467 | struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu); | |
1468 | ||
1469 | spin_lock(&hctx->lock); | |
1470 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1471 | spin_unlock(&hctx->lock); | |
1472 | ||
1473 | blk_mq_run_hw_queue(hctx, false); | |
1474 | } | |
1475 | ||
bd166ef1 JA |
1476 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, |
1477 | struct list_head *list) | |
320ae51f JA |
1478 | |
1479 | { | |
320ae51f JA |
1480 | /* |
1481 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1482 | * offline now | |
1483 | */ | |
1484 | spin_lock(&ctx->lock); | |
1485 | while (!list_empty(list)) { | |
1486 | struct request *rq; | |
1487 | ||
1488 | rq = list_first_entry(list, struct request, queuelist); | |
e57690fe | 1489 | BUG_ON(rq->mq_ctx != ctx); |
320ae51f | 1490 | list_del_init(&rq->queuelist); |
e57690fe | 1491 | __blk_mq_insert_req_list(hctx, rq, false); |
320ae51f | 1492 | } |
cfd0c552 | 1493 | blk_mq_hctx_mark_pending(hctx, ctx); |
320ae51f | 1494 | spin_unlock(&ctx->lock); |
320ae51f JA |
1495 | } |
1496 | ||
1497 | static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1498 | { | |
1499 | struct request *rqa = container_of(a, struct request, queuelist); | |
1500 | struct request *rqb = container_of(b, struct request, queuelist); | |
1501 | ||
1502 | return !(rqa->mq_ctx < rqb->mq_ctx || | |
1503 | (rqa->mq_ctx == rqb->mq_ctx && | |
1504 | blk_rq_pos(rqa) < blk_rq_pos(rqb))); | |
1505 | } | |
1506 | ||
1507 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1508 | { | |
1509 | struct blk_mq_ctx *this_ctx; | |
1510 | struct request_queue *this_q; | |
1511 | struct request *rq; | |
1512 | LIST_HEAD(list); | |
1513 | LIST_HEAD(ctx_list); | |
1514 | unsigned int depth; | |
1515 | ||
1516 | list_splice_init(&plug->mq_list, &list); | |
1517 | ||
1518 | list_sort(NULL, &list, plug_ctx_cmp); | |
1519 | ||
1520 | this_q = NULL; | |
1521 | this_ctx = NULL; | |
1522 | depth = 0; | |
1523 | ||
1524 | while (!list_empty(&list)) { | |
1525 | rq = list_entry_rq(list.next); | |
1526 | list_del_init(&rq->queuelist); | |
1527 | BUG_ON(!rq->q); | |
1528 | if (rq->mq_ctx != this_ctx) { | |
1529 | if (this_ctx) { | |
bd166ef1 JA |
1530 | trace_block_unplug(this_q, depth, from_schedule); |
1531 | blk_mq_sched_insert_requests(this_q, this_ctx, | |
1532 | &ctx_list, | |
1533 | from_schedule); | |
320ae51f JA |
1534 | } |
1535 | ||
1536 | this_ctx = rq->mq_ctx; | |
1537 | this_q = rq->q; | |
1538 | depth = 0; | |
1539 | } | |
1540 | ||
1541 | depth++; | |
1542 | list_add_tail(&rq->queuelist, &ctx_list); | |
1543 | } | |
1544 | ||
1545 | /* | |
1546 | * If 'this_ctx' is set, we know we have entries to complete | |
1547 | * on 'ctx_list'. Do those. | |
1548 | */ | |
1549 | if (this_ctx) { | |
bd166ef1 JA |
1550 | trace_block_unplug(this_q, depth, from_schedule); |
1551 | blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list, | |
1552 | from_schedule); | |
320ae51f JA |
1553 | } |
1554 | } | |
1555 | ||
1556 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio) | |
1557 | { | |
da8d7f07 | 1558 | blk_init_request_from_bio(rq, bio); |
4b570521 | 1559 | |
85acb3ba SL |
1560 | blk_rq_set_rl(rq, blk_get_rl(rq->q, bio)); |
1561 | ||
6e85eaf3 | 1562 | blk_account_io_start(rq, true); |
320ae51f JA |
1563 | } |
1564 | ||
ab42f35d ML |
1565 | static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx, |
1566 | struct blk_mq_ctx *ctx, | |
1567 | struct request *rq) | |
1568 | { | |
1569 | spin_lock(&ctx->lock); | |
1570 | __blk_mq_insert_request(hctx, rq, false); | |
1571 | spin_unlock(&ctx->lock); | |
07068d5b | 1572 | } |
14ec77f3 | 1573 | |
fd2d3326 JA |
1574 | static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq) |
1575 | { | |
bd166ef1 JA |
1576 | if (rq->tag != -1) |
1577 | return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false); | |
1578 | ||
1579 | return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true); | |
fd2d3326 JA |
1580 | } |
1581 | ||
d964f04a ML |
1582 | static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1583 | struct request *rq, | |
1584 | blk_qc_t *cookie, bool may_sleep) | |
f984df1f | 1585 | { |
f984df1f | 1586 | struct request_queue *q = rq->q; |
f984df1f SL |
1587 | struct blk_mq_queue_data bd = { |
1588 | .rq = rq, | |
d945a365 | 1589 | .last = true, |
f984df1f | 1590 | }; |
bd166ef1 | 1591 | blk_qc_t new_cookie; |
f06345ad | 1592 | blk_status_t ret; |
d964f04a ML |
1593 | bool run_queue = true; |
1594 | ||
f4560ffe ML |
1595 | /* RCU or SRCU read lock is needed before checking quiesced flag */ |
1596 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
d964f04a ML |
1597 | run_queue = false; |
1598 | goto insert; | |
1599 | } | |
f984df1f | 1600 | |
bd166ef1 | 1601 | if (q->elevator) |
2253efc8 BVA |
1602 | goto insert; |
1603 | ||
d964f04a | 1604 | if (!blk_mq_get_driver_tag(rq, NULL, false)) |
bd166ef1 JA |
1605 | goto insert; |
1606 | ||
de148297 ML |
1607 | ret = blk_mq_get_dispatch_budget(hctx); |
1608 | if (ret == BLK_STS_RESOURCE) { | |
1609 | blk_mq_put_driver_tag(rq); | |
1610 | goto insert; | |
1611 | } else if (ret != BLK_STS_OK) | |
1612 | goto fail_rq; | |
1613 | ||
bd166ef1 JA |
1614 | new_cookie = request_to_qc_t(hctx, rq); |
1615 | ||
f984df1f SL |
1616 | /* |
1617 | * For OK queue, we are done. For error, kill it. Any other | |
1618 | * error (busy), just add it to our list as we previously | |
1619 | * would have done | |
1620 | */ | |
1621 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
fc17b653 CH |
1622 | switch (ret) { |
1623 | case BLK_STS_OK: | |
7b371636 | 1624 | *cookie = new_cookie; |
2253efc8 | 1625 | return; |
fc17b653 CH |
1626 | case BLK_STS_RESOURCE: |
1627 | __blk_mq_requeue_request(rq); | |
1628 | goto insert; | |
1629 | default: | |
de148297 | 1630 | fail_rq: |
7b371636 | 1631 | *cookie = BLK_QC_T_NONE; |
fc17b653 | 1632 | blk_mq_end_request(rq, ret); |
2253efc8 | 1633 | return; |
f984df1f | 1634 | } |
7b371636 | 1635 | |
2253efc8 | 1636 | insert: |
d964f04a | 1637 | blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep); |
f984df1f SL |
1638 | } |
1639 | ||
5eb6126e CH |
1640 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, |
1641 | struct request *rq, blk_qc_t *cookie) | |
1642 | { | |
1643 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1644 | rcu_read_lock(); | |
d964f04a | 1645 | __blk_mq_try_issue_directly(hctx, rq, cookie, false); |
5eb6126e CH |
1646 | rcu_read_unlock(); |
1647 | } else { | |
bf4907c0 JA |
1648 | unsigned int srcu_idx; |
1649 | ||
1650 | might_sleep(); | |
1651 | ||
07319678 | 1652 | srcu_idx = srcu_read_lock(hctx->queue_rq_srcu); |
d964f04a | 1653 | __blk_mq_try_issue_directly(hctx, rq, cookie, true); |
07319678 | 1654 | srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx); |
5eb6126e CH |
1655 | } |
1656 | } | |
1657 | ||
dece1635 | 1658 | static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio) |
07068d5b | 1659 | { |
ef295ecf | 1660 | const int is_sync = op_is_sync(bio->bi_opf); |
f73f44eb | 1661 | const int is_flush_fua = op_is_flush(bio->bi_opf); |
5a797e00 | 1662 | struct blk_mq_alloc_data data = { .flags = 0 }; |
07068d5b | 1663 | struct request *rq; |
5eb6126e | 1664 | unsigned int request_count = 0; |
f984df1f | 1665 | struct blk_plug *plug; |
5b3f341f | 1666 | struct request *same_queue_rq = NULL; |
7b371636 | 1667 | blk_qc_t cookie; |
87760e5e | 1668 | unsigned int wb_acct; |
07068d5b JA |
1669 | |
1670 | blk_queue_bounce(q, &bio); | |
1671 | ||
af67c31f | 1672 | blk_queue_split(q, &bio); |
f36ea50c | 1673 | |
e23947bd | 1674 | if (!bio_integrity_prep(bio)) |
dece1635 | 1675 | return BLK_QC_T_NONE; |
07068d5b | 1676 | |
87c279e6 OS |
1677 | if (!is_flush_fua && !blk_queue_nomerges(q) && |
1678 | blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq)) | |
1679 | return BLK_QC_T_NONE; | |
f984df1f | 1680 | |
bd166ef1 JA |
1681 | if (blk_mq_sched_bio_merge(q, bio)) |
1682 | return BLK_QC_T_NONE; | |
1683 | ||
87760e5e JA |
1684 | wb_acct = wbt_wait(q->rq_wb, bio, NULL); |
1685 | ||
bd166ef1 JA |
1686 | trace_block_getrq(q, bio, bio->bi_opf); |
1687 | ||
d2c0d383 | 1688 | rq = blk_mq_get_request(q, bio, bio->bi_opf, &data); |
87760e5e JA |
1689 | if (unlikely(!rq)) { |
1690 | __wbt_done(q->rq_wb, wb_acct); | |
03a07c92 GR |
1691 | if (bio->bi_opf & REQ_NOWAIT) |
1692 | bio_wouldblock_error(bio); | |
dece1635 | 1693 | return BLK_QC_T_NONE; |
87760e5e JA |
1694 | } |
1695 | ||
1696 | wbt_track(&rq->issue_stat, wb_acct); | |
07068d5b | 1697 | |
fd2d3326 | 1698 | cookie = request_to_qc_t(data.hctx, rq); |
07068d5b | 1699 | |
f984df1f | 1700 | plug = current->plug; |
07068d5b | 1701 | if (unlikely(is_flush_fua)) { |
f984df1f | 1702 | blk_mq_put_ctx(data.ctx); |
07068d5b | 1703 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 CH |
1704 | if (q->elevator) { |
1705 | blk_mq_sched_insert_request(rq, false, true, true, | |
1706 | true); | |
6a83e74d | 1707 | } else { |
a4d907b6 CH |
1708 | blk_insert_flush(rq); |
1709 | blk_mq_run_hw_queue(data.hctx, true); | |
6a83e74d | 1710 | } |
a4d907b6 | 1711 | } else if (plug && q->nr_hw_queues == 1) { |
600271d9 SL |
1712 | struct request *last = NULL; |
1713 | ||
b00c53e8 | 1714 | blk_mq_put_ctx(data.ctx); |
e6c4438b | 1715 | blk_mq_bio_to_request(rq, bio); |
0a6219a9 ML |
1716 | |
1717 | /* | |
1718 | * @request_count may become stale because of schedule | |
1719 | * out, so check the list again. | |
1720 | */ | |
1721 | if (list_empty(&plug->mq_list)) | |
1722 | request_count = 0; | |
254d259d CH |
1723 | else if (blk_queue_nomerges(q)) |
1724 | request_count = blk_plug_queued_count(q); | |
1725 | ||
676d0607 | 1726 | if (!request_count) |
e6c4438b | 1727 | trace_block_plug(q); |
600271d9 SL |
1728 | else |
1729 | last = list_entry_rq(plug->mq_list.prev); | |
b094f89c | 1730 | |
600271d9 SL |
1731 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && |
1732 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
e6c4438b JM |
1733 | blk_flush_plug_list(plug, false); |
1734 | trace_block_plug(q); | |
320ae51f | 1735 | } |
b094f89c | 1736 | |
e6c4438b | 1737 | list_add_tail(&rq->queuelist, &plug->mq_list); |
2299722c | 1738 | } else if (plug && !blk_queue_nomerges(q)) { |
bd166ef1 | 1739 | blk_mq_bio_to_request(rq, bio); |
07068d5b | 1740 | |
07068d5b | 1741 | /* |
6a83e74d | 1742 | * We do limited plugging. If the bio can be merged, do that. |
f984df1f SL |
1743 | * Otherwise the existing request in the plug list will be |
1744 | * issued. So the plug list will have one request at most | |
2299722c CH |
1745 | * The plug list might get flushed before this. If that happens, |
1746 | * the plug list is empty, and same_queue_rq is invalid. | |
07068d5b | 1747 | */ |
2299722c CH |
1748 | if (list_empty(&plug->mq_list)) |
1749 | same_queue_rq = NULL; | |
1750 | if (same_queue_rq) | |
1751 | list_del_init(&same_queue_rq->queuelist); | |
1752 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
1753 | ||
bf4907c0 JA |
1754 | blk_mq_put_ctx(data.ctx); |
1755 | ||
dad7a3be ML |
1756 | if (same_queue_rq) { |
1757 | data.hctx = blk_mq_map_queue(q, | |
1758 | same_queue_rq->mq_ctx->cpu); | |
2299722c CH |
1759 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, |
1760 | &cookie); | |
dad7a3be | 1761 | } |
a4d907b6 | 1762 | } else if (q->nr_hw_queues > 1 && is_sync) { |
bf4907c0 | 1763 | blk_mq_put_ctx(data.ctx); |
2299722c | 1764 | blk_mq_bio_to_request(rq, bio); |
2299722c | 1765 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); |
a4d907b6 | 1766 | } else if (q->elevator) { |
b00c53e8 | 1767 | blk_mq_put_ctx(data.ctx); |
bd166ef1 | 1768 | blk_mq_bio_to_request(rq, bio); |
a4d907b6 | 1769 | blk_mq_sched_insert_request(rq, false, true, true, true); |
ab42f35d | 1770 | } else { |
b00c53e8 | 1771 | blk_mq_put_ctx(data.ctx); |
ab42f35d ML |
1772 | blk_mq_bio_to_request(rq, bio); |
1773 | blk_mq_queue_io(data.hctx, data.ctx, rq); | |
a4d907b6 | 1774 | blk_mq_run_hw_queue(data.hctx, true); |
ab42f35d | 1775 | } |
320ae51f | 1776 | |
7b371636 | 1777 | return cookie; |
320ae51f JA |
1778 | } |
1779 | ||
cc71a6f4 JA |
1780 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, |
1781 | unsigned int hctx_idx) | |
95363efd | 1782 | { |
e9b267d9 | 1783 | struct page *page; |
320ae51f | 1784 | |
24d2f903 | 1785 | if (tags->rqs && set->ops->exit_request) { |
e9b267d9 | 1786 | int i; |
320ae51f | 1787 | |
24d2f903 | 1788 | for (i = 0; i < tags->nr_tags; i++) { |
2af8cbe3 JA |
1789 | struct request *rq = tags->static_rqs[i]; |
1790 | ||
1791 | if (!rq) | |
e9b267d9 | 1792 | continue; |
d6296d39 | 1793 | set->ops->exit_request(set, rq, hctx_idx); |
2af8cbe3 | 1794 | tags->static_rqs[i] = NULL; |
e9b267d9 | 1795 | } |
320ae51f | 1796 | } |
320ae51f | 1797 | |
24d2f903 CH |
1798 | while (!list_empty(&tags->page_list)) { |
1799 | page = list_first_entry(&tags->page_list, struct page, lru); | |
6753471c | 1800 | list_del_init(&page->lru); |
f75782e4 CM |
1801 | /* |
1802 | * Remove kmemleak object previously allocated in | |
1803 | * blk_mq_init_rq_map(). | |
1804 | */ | |
1805 | kmemleak_free(page_address(page)); | |
320ae51f JA |
1806 | __free_pages(page, page->private); |
1807 | } | |
cc71a6f4 | 1808 | } |
320ae51f | 1809 | |
cc71a6f4 JA |
1810 | void blk_mq_free_rq_map(struct blk_mq_tags *tags) |
1811 | { | |
24d2f903 | 1812 | kfree(tags->rqs); |
cc71a6f4 | 1813 | tags->rqs = NULL; |
2af8cbe3 JA |
1814 | kfree(tags->static_rqs); |
1815 | tags->static_rqs = NULL; | |
320ae51f | 1816 | |
24d2f903 | 1817 | blk_mq_free_tags(tags); |
320ae51f JA |
1818 | } |
1819 | ||
cc71a6f4 JA |
1820 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, |
1821 | unsigned int hctx_idx, | |
1822 | unsigned int nr_tags, | |
1823 | unsigned int reserved_tags) | |
320ae51f | 1824 | { |
24d2f903 | 1825 | struct blk_mq_tags *tags; |
59f082e4 | 1826 | int node; |
320ae51f | 1827 | |
59f082e4 SL |
1828 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); |
1829 | if (node == NUMA_NO_NODE) | |
1830 | node = set->numa_node; | |
1831 | ||
1832 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, | |
24391c0d | 1833 | BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); |
24d2f903 CH |
1834 | if (!tags) |
1835 | return NULL; | |
320ae51f | 1836 | |
cc71a6f4 | 1837 | tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
36e1f3d1 | 1838 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, |
59f082e4 | 1839 | node); |
24d2f903 CH |
1840 | if (!tags->rqs) { |
1841 | blk_mq_free_tags(tags); | |
1842 | return NULL; | |
1843 | } | |
320ae51f | 1844 | |
2af8cbe3 JA |
1845 | tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *), |
1846 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
59f082e4 | 1847 | node); |
2af8cbe3 JA |
1848 | if (!tags->static_rqs) { |
1849 | kfree(tags->rqs); | |
1850 | blk_mq_free_tags(tags); | |
1851 | return NULL; | |
1852 | } | |
1853 | ||
cc71a6f4 JA |
1854 | return tags; |
1855 | } | |
1856 | ||
1857 | static size_t order_to_size(unsigned int order) | |
1858 | { | |
1859 | return (size_t)PAGE_SIZE << order; | |
1860 | } | |
1861 | ||
1862 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
1863 | unsigned int hctx_idx, unsigned int depth) | |
1864 | { | |
1865 | unsigned int i, j, entries_per_page, max_order = 4; | |
1866 | size_t rq_size, left; | |
59f082e4 SL |
1867 | int node; |
1868 | ||
1869 | node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx); | |
1870 | if (node == NUMA_NO_NODE) | |
1871 | node = set->numa_node; | |
cc71a6f4 JA |
1872 | |
1873 | INIT_LIST_HEAD(&tags->page_list); | |
1874 | ||
320ae51f JA |
1875 | /* |
1876 | * rq_size is the size of the request plus driver payload, rounded | |
1877 | * to the cacheline size | |
1878 | */ | |
24d2f903 | 1879 | rq_size = round_up(sizeof(struct request) + set->cmd_size, |
320ae51f | 1880 | cache_line_size()); |
cc71a6f4 | 1881 | left = rq_size * depth; |
320ae51f | 1882 | |
cc71a6f4 | 1883 | for (i = 0; i < depth; ) { |
320ae51f JA |
1884 | int this_order = max_order; |
1885 | struct page *page; | |
1886 | int to_do; | |
1887 | void *p; | |
1888 | ||
b3a834b1 | 1889 | while (this_order && left < order_to_size(this_order - 1)) |
320ae51f JA |
1890 | this_order--; |
1891 | ||
1892 | do { | |
59f082e4 | 1893 | page = alloc_pages_node(node, |
36e1f3d1 | 1894 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, |
a5164405 | 1895 | this_order); |
320ae51f JA |
1896 | if (page) |
1897 | break; | |
1898 | if (!this_order--) | |
1899 | break; | |
1900 | if (order_to_size(this_order) < rq_size) | |
1901 | break; | |
1902 | } while (1); | |
1903 | ||
1904 | if (!page) | |
24d2f903 | 1905 | goto fail; |
320ae51f JA |
1906 | |
1907 | page->private = this_order; | |
24d2f903 | 1908 | list_add_tail(&page->lru, &tags->page_list); |
320ae51f JA |
1909 | |
1910 | p = page_address(page); | |
f75782e4 CM |
1911 | /* |
1912 | * Allow kmemleak to scan these pages as they contain pointers | |
1913 | * to additional allocations like via ops->init_request(). | |
1914 | */ | |
36e1f3d1 | 1915 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); |
320ae51f | 1916 | entries_per_page = order_to_size(this_order) / rq_size; |
cc71a6f4 | 1917 | to_do = min(entries_per_page, depth - i); |
320ae51f JA |
1918 | left -= to_do * rq_size; |
1919 | for (j = 0; j < to_do; j++) { | |
2af8cbe3 JA |
1920 | struct request *rq = p; |
1921 | ||
1922 | tags->static_rqs[i] = rq; | |
24d2f903 | 1923 | if (set->ops->init_request) { |
d6296d39 | 1924 | if (set->ops->init_request(set, rq, hctx_idx, |
59f082e4 | 1925 | node)) { |
2af8cbe3 | 1926 | tags->static_rqs[i] = NULL; |
24d2f903 | 1927 | goto fail; |
a5164405 | 1928 | } |
e9b267d9 CH |
1929 | } |
1930 | ||
320ae51f JA |
1931 | p += rq_size; |
1932 | i++; | |
1933 | } | |
1934 | } | |
cc71a6f4 | 1935 | return 0; |
320ae51f | 1936 | |
24d2f903 | 1937 | fail: |
cc71a6f4 JA |
1938 | blk_mq_free_rqs(set, tags, hctx_idx); |
1939 | return -ENOMEM; | |
320ae51f JA |
1940 | } |
1941 | ||
e57690fe JA |
1942 | /* |
1943 | * 'cpu' is going away. splice any existing rq_list entries from this | |
1944 | * software queue to the hw queue dispatch list, and ensure that it | |
1945 | * gets run. | |
1946 | */ | |
9467f859 | 1947 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) |
484b4061 | 1948 | { |
9467f859 | 1949 | struct blk_mq_hw_ctx *hctx; |
484b4061 JA |
1950 | struct blk_mq_ctx *ctx; |
1951 | LIST_HEAD(tmp); | |
1952 | ||
9467f859 | 1953 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); |
e57690fe | 1954 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); |
484b4061 JA |
1955 | |
1956 | spin_lock(&ctx->lock); | |
1957 | if (!list_empty(&ctx->rq_list)) { | |
1958 | list_splice_init(&ctx->rq_list, &tmp); | |
1959 | blk_mq_hctx_clear_pending(hctx, ctx); | |
1960 | } | |
1961 | spin_unlock(&ctx->lock); | |
1962 | ||
1963 | if (list_empty(&tmp)) | |
9467f859 | 1964 | return 0; |
484b4061 | 1965 | |
e57690fe JA |
1966 | spin_lock(&hctx->lock); |
1967 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
1968 | spin_unlock(&hctx->lock); | |
484b4061 JA |
1969 | |
1970 | blk_mq_run_hw_queue(hctx, true); | |
9467f859 | 1971 | return 0; |
484b4061 JA |
1972 | } |
1973 | ||
9467f859 | 1974 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) |
484b4061 | 1975 | { |
9467f859 TG |
1976 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, |
1977 | &hctx->cpuhp_dead); | |
484b4061 JA |
1978 | } |
1979 | ||
c3b4afca | 1980 | /* hctx->ctxs will be freed in queue's release handler */ |
08e98fc6 ML |
1981 | static void blk_mq_exit_hctx(struct request_queue *q, |
1982 | struct blk_mq_tag_set *set, | |
1983 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
1984 | { | |
9c1051aa OS |
1985 | blk_mq_debugfs_unregister_hctx(hctx); |
1986 | ||
08e98fc6 ML |
1987 | blk_mq_tag_idle(hctx); |
1988 | ||
f70ced09 | 1989 | if (set->ops->exit_request) |
d6296d39 | 1990 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); |
f70ced09 | 1991 | |
93252632 OS |
1992 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); |
1993 | ||
08e98fc6 ML |
1994 | if (set->ops->exit_hctx) |
1995 | set->ops->exit_hctx(hctx, hctx_idx); | |
1996 | ||
6a83e74d | 1997 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 1998 | cleanup_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 1999 | |
9467f859 | 2000 | blk_mq_remove_cpuhp(hctx); |
f70ced09 | 2001 | blk_free_flush_queue(hctx->fq); |
88459642 | 2002 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2003 | } |
2004 | ||
624dbe47 ML |
2005 | static void blk_mq_exit_hw_queues(struct request_queue *q, |
2006 | struct blk_mq_tag_set *set, int nr_queue) | |
2007 | { | |
2008 | struct blk_mq_hw_ctx *hctx; | |
2009 | unsigned int i; | |
2010 | ||
2011 | queue_for_each_hw_ctx(q, hctx, i) { | |
2012 | if (i == nr_queue) | |
2013 | break; | |
08e98fc6 | 2014 | blk_mq_exit_hctx(q, set, hctx, i); |
624dbe47 | 2015 | } |
624dbe47 ML |
2016 | } |
2017 | ||
08e98fc6 ML |
2018 | static int blk_mq_init_hctx(struct request_queue *q, |
2019 | struct blk_mq_tag_set *set, | |
2020 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
320ae51f | 2021 | { |
08e98fc6 ML |
2022 | int node; |
2023 | ||
2024 | node = hctx->numa_node; | |
2025 | if (node == NUMA_NO_NODE) | |
2026 | node = hctx->numa_node = set->numa_node; | |
2027 | ||
9f993737 | 2028 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); |
08e98fc6 ML |
2029 | spin_lock_init(&hctx->lock); |
2030 | INIT_LIST_HEAD(&hctx->dispatch); | |
2031 | hctx->queue = q; | |
2404e607 | 2032 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED; |
08e98fc6 | 2033 | |
9467f859 | 2034 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); |
08e98fc6 ML |
2035 | |
2036 | hctx->tags = set->tags[hctx_idx]; | |
320ae51f JA |
2037 | |
2038 | /* | |
08e98fc6 ML |
2039 | * Allocate space for all possible cpus to avoid allocation at |
2040 | * runtime | |
320ae51f | 2041 | */ |
08e98fc6 ML |
2042 | hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *), |
2043 | GFP_KERNEL, node); | |
2044 | if (!hctx->ctxs) | |
2045 | goto unregister_cpu_notifier; | |
320ae51f | 2046 | |
88459642 OS |
2047 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL, |
2048 | node)) | |
08e98fc6 | 2049 | goto free_ctxs; |
320ae51f | 2050 | |
08e98fc6 | 2051 | hctx->nr_ctx = 0; |
320ae51f | 2052 | |
08e98fc6 ML |
2053 | if (set->ops->init_hctx && |
2054 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2055 | goto free_bitmap; | |
320ae51f | 2056 | |
93252632 OS |
2057 | if (blk_mq_sched_init_hctx(q, hctx, hctx_idx)) |
2058 | goto exit_hctx; | |
2059 | ||
f70ced09 ML |
2060 | hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size); |
2061 | if (!hctx->fq) | |
93252632 | 2062 | goto sched_exit_hctx; |
320ae51f | 2063 | |
f70ced09 | 2064 | if (set->ops->init_request && |
d6296d39 CH |
2065 | set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx, |
2066 | node)) | |
f70ced09 | 2067 | goto free_fq; |
320ae51f | 2068 | |
6a83e74d | 2069 | if (hctx->flags & BLK_MQ_F_BLOCKING) |
07319678 | 2070 | init_srcu_struct(hctx->queue_rq_srcu); |
6a83e74d | 2071 | |
9c1051aa OS |
2072 | blk_mq_debugfs_register_hctx(q, hctx); |
2073 | ||
08e98fc6 | 2074 | return 0; |
320ae51f | 2075 | |
f70ced09 ML |
2076 | free_fq: |
2077 | kfree(hctx->fq); | |
93252632 OS |
2078 | sched_exit_hctx: |
2079 | blk_mq_sched_exit_hctx(q, hctx, hctx_idx); | |
f70ced09 ML |
2080 | exit_hctx: |
2081 | if (set->ops->exit_hctx) | |
2082 | set->ops->exit_hctx(hctx, hctx_idx); | |
08e98fc6 | 2083 | free_bitmap: |
88459642 | 2084 | sbitmap_free(&hctx->ctx_map); |
08e98fc6 ML |
2085 | free_ctxs: |
2086 | kfree(hctx->ctxs); | |
2087 | unregister_cpu_notifier: | |
9467f859 | 2088 | blk_mq_remove_cpuhp(hctx); |
08e98fc6 ML |
2089 | return -1; |
2090 | } | |
320ae51f | 2091 | |
320ae51f JA |
2092 | static void blk_mq_init_cpu_queues(struct request_queue *q, |
2093 | unsigned int nr_hw_queues) | |
2094 | { | |
2095 | unsigned int i; | |
2096 | ||
2097 | for_each_possible_cpu(i) { | |
2098 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2099 | struct blk_mq_hw_ctx *hctx; | |
2100 | ||
320ae51f JA |
2101 | __ctx->cpu = i; |
2102 | spin_lock_init(&__ctx->lock); | |
2103 | INIT_LIST_HEAD(&__ctx->rq_list); | |
2104 | __ctx->queue = q; | |
2105 | ||
4b855ad3 CH |
2106 | /* If the cpu isn't present, the cpu is mapped to first hctx */ |
2107 | if (!cpu_present(i)) | |
320ae51f JA |
2108 | continue; |
2109 | ||
7d7e0f90 | 2110 | hctx = blk_mq_map_queue(q, i); |
e4043dcf | 2111 | |
320ae51f JA |
2112 | /* |
2113 | * Set local node, IFF we have more than one hw queue. If | |
2114 | * not, we remain on the home node of the device | |
2115 | */ | |
2116 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
bffed457 | 2117 | hctx->numa_node = local_memory_node(cpu_to_node(i)); |
320ae51f JA |
2118 | } |
2119 | } | |
2120 | ||
cc71a6f4 JA |
2121 | static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx) |
2122 | { | |
2123 | int ret = 0; | |
2124 | ||
2125 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2126 | set->queue_depth, set->reserved_tags); | |
2127 | if (!set->tags[hctx_idx]) | |
2128 | return false; | |
2129 | ||
2130 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2131 | set->queue_depth); | |
2132 | if (!ret) | |
2133 | return true; | |
2134 | ||
2135 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2136 | set->tags[hctx_idx] = NULL; | |
2137 | return false; | |
2138 | } | |
2139 | ||
2140 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2141 | unsigned int hctx_idx) | |
2142 | { | |
bd166ef1 JA |
2143 | if (set->tags[hctx_idx]) { |
2144 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2145 | blk_mq_free_rq_map(set->tags[hctx_idx]); | |
2146 | set->tags[hctx_idx] = NULL; | |
2147 | } | |
cc71a6f4 JA |
2148 | } |
2149 | ||
4b855ad3 | 2150 | static void blk_mq_map_swqueue(struct request_queue *q) |
320ae51f | 2151 | { |
d1b1cea1 | 2152 | unsigned int i, hctx_idx; |
320ae51f JA |
2153 | struct blk_mq_hw_ctx *hctx; |
2154 | struct blk_mq_ctx *ctx; | |
2a34c087 | 2155 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2156 | |
60de074b AM |
2157 | /* |
2158 | * Avoid others reading imcomplete hctx->cpumask through sysfs | |
2159 | */ | |
2160 | mutex_lock(&q->sysfs_lock); | |
2161 | ||
320ae51f | 2162 | queue_for_each_hw_ctx(q, hctx, i) { |
e4043dcf | 2163 | cpumask_clear(hctx->cpumask); |
320ae51f JA |
2164 | hctx->nr_ctx = 0; |
2165 | } | |
2166 | ||
2167 | /* | |
4b855ad3 CH |
2168 | * Map software to hardware queues. |
2169 | * | |
2170 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
320ae51f | 2171 | */ |
4b855ad3 | 2172 | for_each_present_cpu(i) { |
d1b1cea1 GKB |
2173 | hctx_idx = q->mq_map[i]; |
2174 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
cc71a6f4 JA |
2175 | if (!set->tags[hctx_idx] && |
2176 | !__blk_mq_alloc_rq_map(set, hctx_idx)) { | |
d1b1cea1 GKB |
2177 | /* |
2178 | * If tags initialization fail for some hctx, | |
2179 | * that hctx won't be brought online. In this | |
2180 | * case, remap the current ctx to hctx[0] which | |
2181 | * is guaranteed to always have tags allocated | |
2182 | */ | |
cc71a6f4 | 2183 | q->mq_map[i] = 0; |
d1b1cea1 GKB |
2184 | } |
2185 | ||
897bb0c7 | 2186 | ctx = per_cpu_ptr(q->queue_ctx, i); |
7d7e0f90 | 2187 | hctx = blk_mq_map_queue(q, i); |
868f2f0b | 2188 | |
e4043dcf | 2189 | cpumask_set_cpu(i, hctx->cpumask); |
320ae51f JA |
2190 | ctx->index_hw = hctx->nr_ctx; |
2191 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2192 | } | |
506e931f | 2193 | |
60de074b AM |
2194 | mutex_unlock(&q->sysfs_lock); |
2195 | ||
506e931f | 2196 | queue_for_each_hw_ctx(q, hctx, i) { |
484b4061 | 2197 | /* |
a68aafa5 JA |
2198 | * If no software queues are mapped to this hardware queue, |
2199 | * disable it and free the request entries. | |
484b4061 JA |
2200 | */ |
2201 | if (!hctx->nr_ctx) { | |
d1b1cea1 GKB |
2202 | /* Never unmap queue 0. We need it as a |
2203 | * fallback in case of a new remap fails | |
2204 | * allocation | |
2205 | */ | |
cc71a6f4 JA |
2206 | if (i && set->tags[i]) |
2207 | blk_mq_free_map_and_requests(set, i); | |
2208 | ||
2a34c087 | 2209 | hctx->tags = NULL; |
484b4061 JA |
2210 | continue; |
2211 | } | |
2212 | ||
2a34c087 ML |
2213 | hctx->tags = set->tags[i]; |
2214 | WARN_ON(!hctx->tags); | |
2215 | ||
889fa31f CY |
2216 | /* |
2217 | * Set the map size to the number of mapped software queues. | |
2218 | * This is more accurate and more efficient than looping | |
2219 | * over all possibly mapped software queues. | |
2220 | */ | |
88459642 | 2221 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); |
889fa31f | 2222 | |
484b4061 JA |
2223 | /* |
2224 | * Initialize batch roundrobin counts | |
2225 | */ | |
506e931f JA |
2226 | hctx->next_cpu = cpumask_first(hctx->cpumask); |
2227 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2228 | } | |
320ae51f JA |
2229 | } |
2230 | ||
8e8320c9 JA |
2231 | /* |
2232 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2233 | * the queue isn't live yet. | |
2234 | */ | |
2404e607 | 2235 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) |
0d2602ca JA |
2236 | { |
2237 | struct blk_mq_hw_ctx *hctx; | |
0d2602ca JA |
2238 | int i; |
2239 | ||
2404e607 | 2240 | queue_for_each_hw_ctx(q, hctx, i) { |
8e8320c9 JA |
2241 | if (shared) { |
2242 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2243 | atomic_inc(&q->shared_hctx_restart); | |
2404e607 | 2244 | hctx->flags |= BLK_MQ_F_TAG_SHARED; |
8e8320c9 JA |
2245 | } else { |
2246 | if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) | |
2247 | atomic_dec(&q->shared_hctx_restart); | |
2404e607 | 2248 | hctx->flags &= ~BLK_MQ_F_TAG_SHARED; |
8e8320c9 | 2249 | } |
2404e607 JM |
2250 | } |
2251 | } | |
2252 | ||
8e8320c9 JA |
2253 | static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, |
2254 | bool shared) | |
2404e607 JM |
2255 | { |
2256 | struct request_queue *q; | |
0d2602ca | 2257 | |
705cda97 BVA |
2258 | lockdep_assert_held(&set->tag_list_lock); |
2259 | ||
0d2602ca JA |
2260 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2261 | blk_mq_freeze_queue(q); | |
2404e607 | 2262 | queue_set_hctx_shared(q, shared); |
0d2602ca JA |
2263 | blk_mq_unfreeze_queue(q); |
2264 | } | |
2265 | } | |
2266 | ||
2267 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2268 | { | |
2269 | struct blk_mq_tag_set *set = q->tag_set; | |
2270 | ||
0d2602ca | 2271 | mutex_lock(&set->tag_list_lock); |
705cda97 BVA |
2272 | list_del_rcu(&q->tag_set_list); |
2273 | INIT_LIST_HEAD(&q->tag_set_list); | |
2404e607 JM |
2274 | if (list_is_singular(&set->tag_list)) { |
2275 | /* just transitioned to unshared */ | |
2276 | set->flags &= ~BLK_MQ_F_TAG_SHARED; | |
2277 | /* update existing queue */ | |
2278 | blk_mq_update_tag_set_depth(set, false); | |
2279 | } | |
0d2602ca | 2280 | mutex_unlock(&set->tag_list_lock); |
705cda97 BVA |
2281 | |
2282 | synchronize_rcu(); | |
0d2602ca JA |
2283 | } |
2284 | ||
2285 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2286 | struct request_queue *q) | |
2287 | { | |
2288 | q->tag_set = set; | |
2289 | ||
2290 | mutex_lock(&set->tag_list_lock); | |
2404e607 JM |
2291 | |
2292 | /* Check to see if we're transitioning to shared (from 1 to 2 queues). */ | |
2293 | if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) { | |
2294 | set->flags |= BLK_MQ_F_TAG_SHARED; | |
2295 | /* update existing queue */ | |
2296 | blk_mq_update_tag_set_depth(set, true); | |
2297 | } | |
2298 | if (set->flags & BLK_MQ_F_TAG_SHARED) | |
2299 | queue_set_hctx_shared(q, true); | |
705cda97 | 2300 | list_add_tail_rcu(&q->tag_set_list, &set->tag_list); |
2404e607 | 2301 | |
0d2602ca JA |
2302 | mutex_unlock(&set->tag_list_lock); |
2303 | } | |
2304 | ||
e09aae7e ML |
2305 | /* |
2306 | * It is the actual release handler for mq, but we do it from | |
2307 | * request queue's release handler for avoiding use-after-free | |
2308 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2309 | * but we can't group ctx/kctx kobj without it. | |
2310 | */ | |
2311 | void blk_mq_release(struct request_queue *q) | |
2312 | { | |
2313 | struct blk_mq_hw_ctx *hctx; | |
2314 | unsigned int i; | |
2315 | ||
2316 | /* hctx kobj stays in hctx */ | |
c3b4afca ML |
2317 | queue_for_each_hw_ctx(q, hctx, i) { |
2318 | if (!hctx) | |
2319 | continue; | |
6c8b232e | 2320 | kobject_put(&hctx->kobj); |
c3b4afca | 2321 | } |
e09aae7e | 2322 | |
a723bab3 AM |
2323 | q->mq_map = NULL; |
2324 | ||
e09aae7e ML |
2325 | kfree(q->queue_hw_ctx); |
2326 | ||
7ea5fe31 ML |
2327 | /* |
2328 | * release .mq_kobj and sw queue's kobject now because | |
2329 | * both share lifetime with request queue. | |
2330 | */ | |
2331 | blk_mq_sysfs_deinit(q); | |
2332 | ||
e09aae7e ML |
2333 | free_percpu(q->queue_ctx); |
2334 | } | |
2335 | ||
24d2f903 | 2336 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) |
b62c21b7 MS |
2337 | { |
2338 | struct request_queue *uninit_q, *q; | |
2339 | ||
2340 | uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node); | |
2341 | if (!uninit_q) | |
2342 | return ERR_PTR(-ENOMEM); | |
2343 | ||
2344 | q = blk_mq_init_allocated_queue(set, uninit_q); | |
2345 | if (IS_ERR(q)) | |
2346 | blk_cleanup_queue(uninit_q); | |
2347 | ||
2348 | return q; | |
2349 | } | |
2350 | EXPORT_SYMBOL(blk_mq_init_queue); | |
2351 | ||
07319678 BVA |
2352 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) |
2353 | { | |
2354 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2355 | ||
2356 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu), | |
2357 | __alignof__(struct blk_mq_hw_ctx)) != | |
2358 | sizeof(struct blk_mq_hw_ctx)); | |
2359 | ||
2360 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2361 | hw_ctx_size += sizeof(struct srcu_struct); | |
2362 | ||
2363 | return hw_ctx_size; | |
2364 | } | |
2365 | ||
868f2f0b KB |
2366 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, |
2367 | struct request_queue *q) | |
320ae51f | 2368 | { |
868f2f0b KB |
2369 | int i, j; |
2370 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
f14bbe77 | 2371 | |
868f2f0b | 2372 | blk_mq_sysfs_unregister(q); |
24d2f903 | 2373 | for (i = 0; i < set->nr_hw_queues; i++) { |
868f2f0b | 2374 | int node; |
f14bbe77 | 2375 | |
868f2f0b KB |
2376 | if (hctxs[i]) |
2377 | continue; | |
2378 | ||
2379 | node = blk_mq_hw_queue_to_node(q->mq_map, i); | |
07319678 | 2380 | hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set), |
cdef54dd | 2381 | GFP_KERNEL, node); |
320ae51f | 2382 | if (!hctxs[i]) |
868f2f0b | 2383 | break; |
320ae51f | 2384 | |
a86073e4 | 2385 | if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL, |
868f2f0b KB |
2386 | node)) { |
2387 | kfree(hctxs[i]); | |
2388 | hctxs[i] = NULL; | |
2389 | break; | |
2390 | } | |
e4043dcf | 2391 | |
0d2602ca | 2392 | atomic_set(&hctxs[i]->nr_active, 0); |
f14bbe77 | 2393 | hctxs[i]->numa_node = node; |
320ae51f | 2394 | hctxs[i]->queue_num = i; |
868f2f0b KB |
2395 | |
2396 | if (blk_mq_init_hctx(q, set, hctxs[i], i)) { | |
2397 | free_cpumask_var(hctxs[i]->cpumask); | |
2398 | kfree(hctxs[i]); | |
2399 | hctxs[i] = NULL; | |
2400 | break; | |
2401 | } | |
2402 | blk_mq_hctx_kobj_init(hctxs[i]); | |
320ae51f | 2403 | } |
868f2f0b KB |
2404 | for (j = i; j < q->nr_hw_queues; j++) { |
2405 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
2406 | ||
2407 | if (hctx) { | |
cc71a6f4 JA |
2408 | if (hctx->tags) |
2409 | blk_mq_free_map_and_requests(set, j); | |
868f2f0b | 2410 | blk_mq_exit_hctx(q, set, hctx, j); |
868f2f0b | 2411 | kobject_put(&hctx->kobj); |
868f2f0b KB |
2412 | hctxs[j] = NULL; |
2413 | ||
2414 | } | |
2415 | } | |
2416 | q->nr_hw_queues = i; | |
2417 | blk_mq_sysfs_register(q); | |
2418 | } | |
2419 | ||
2420 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
2421 | struct request_queue *q) | |
2422 | { | |
66841672 ML |
2423 | /* mark the queue as mq asap */ |
2424 | q->mq_ops = set->ops; | |
2425 | ||
34dbad5d | 2426 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, |
720b8ccc SB |
2427 | blk_mq_poll_stats_bkt, |
2428 | BLK_MQ_POLL_STATS_BKTS, q); | |
34dbad5d OS |
2429 | if (!q->poll_cb) |
2430 | goto err_exit; | |
2431 | ||
868f2f0b KB |
2432 | q->queue_ctx = alloc_percpu(struct blk_mq_ctx); |
2433 | if (!q->queue_ctx) | |
c7de5726 | 2434 | goto err_exit; |
868f2f0b | 2435 | |
737f98cf ML |
2436 | /* init q->mq_kobj and sw queues' kobjects */ |
2437 | blk_mq_sysfs_init(q); | |
2438 | ||
868f2f0b KB |
2439 | q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)), |
2440 | GFP_KERNEL, set->numa_node); | |
2441 | if (!q->queue_hw_ctx) | |
2442 | goto err_percpu; | |
2443 | ||
bdd17e75 | 2444 | q->mq_map = set->mq_map; |
868f2f0b KB |
2445 | |
2446 | blk_mq_realloc_hw_ctxs(set, q); | |
2447 | if (!q->nr_hw_queues) | |
2448 | goto err_hctxs; | |
320ae51f | 2449 | |
287922eb | 2450 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); |
e56f698b | 2451 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); |
320ae51f JA |
2452 | |
2453 | q->nr_queues = nr_cpu_ids; | |
320ae51f | 2454 | |
94eddfbe | 2455 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; |
320ae51f | 2456 | |
05f1dd53 JA |
2457 | if (!(set->flags & BLK_MQ_F_SG_MERGE)) |
2458 | q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE; | |
2459 | ||
1be036e9 CH |
2460 | q->sg_reserved_size = INT_MAX; |
2461 | ||
2849450a | 2462 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); |
6fca6a61 CH |
2463 | INIT_LIST_HEAD(&q->requeue_list); |
2464 | spin_lock_init(&q->requeue_lock); | |
2465 | ||
254d259d | 2466 | blk_queue_make_request(q, blk_mq_make_request); |
07068d5b | 2467 | |
eba71768 JA |
2468 | /* |
2469 | * Do this after blk_queue_make_request() overrides it... | |
2470 | */ | |
2471 | q->nr_requests = set->queue_depth; | |
2472 | ||
64f1c21e JA |
2473 | /* |
2474 | * Default to classic polling | |
2475 | */ | |
2476 | q->poll_nsec = -1; | |
2477 | ||
24d2f903 CH |
2478 | if (set->ops->complete) |
2479 | blk_queue_softirq_done(q, set->ops->complete); | |
30a91cb4 | 2480 | |
24d2f903 | 2481 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); |
0d2602ca | 2482 | blk_mq_add_queue_tag_set(set, q); |
4b855ad3 | 2483 | blk_mq_map_swqueue(q); |
4593fdbe | 2484 | |
d3484991 JA |
2485 | if (!(set->flags & BLK_MQ_F_NO_SCHED)) { |
2486 | int ret; | |
2487 | ||
2488 | ret = blk_mq_sched_init(q); | |
2489 | if (ret) | |
2490 | return ERR_PTR(ret); | |
2491 | } | |
2492 | ||
320ae51f | 2493 | return q; |
18741986 | 2494 | |
320ae51f | 2495 | err_hctxs: |
868f2f0b | 2496 | kfree(q->queue_hw_ctx); |
320ae51f | 2497 | err_percpu: |
868f2f0b | 2498 | free_percpu(q->queue_ctx); |
c7de5726 ML |
2499 | err_exit: |
2500 | q->mq_ops = NULL; | |
320ae51f JA |
2501 | return ERR_PTR(-ENOMEM); |
2502 | } | |
b62c21b7 | 2503 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); |
320ae51f JA |
2504 | |
2505 | void blk_mq_free_queue(struct request_queue *q) | |
2506 | { | |
624dbe47 | 2507 | struct blk_mq_tag_set *set = q->tag_set; |
320ae51f | 2508 | |
0d2602ca | 2509 | blk_mq_del_queue_tag_set(q); |
624dbe47 | 2510 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); |
320ae51f | 2511 | } |
320ae51f JA |
2512 | |
2513 | /* Basically redo blk_mq_init_queue with queue frozen */ | |
4b855ad3 | 2514 | static void blk_mq_queue_reinit(struct request_queue *q) |
320ae51f | 2515 | { |
4ecd4fef | 2516 | WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth)); |
320ae51f | 2517 | |
9c1051aa | 2518 | blk_mq_debugfs_unregister_hctxs(q); |
67aec14c JA |
2519 | blk_mq_sysfs_unregister(q); |
2520 | ||
320ae51f JA |
2521 | /* |
2522 | * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe | |
2523 | * we should change hctx numa_node according to new topology (this | |
2524 | * involves free and re-allocate memory, worthy doing?) | |
2525 | */ | |
2526 | ||
4b855ad3 | 2527 | blk_mq_map_swqueue(q); |
320ae51f | 2528 | |
67aec14c | 2529 | blk_mq_sysfs_register(q); |
9c1051aa | 2530 | blk_mq_debugfs_register_hctxs(q); |
320ae51f JA |
2531 | } |
2532 | ||
a5164405 JA |
2533 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) |
2534 | { | |
2535 | int i; | |
2536 | ||
cc71a6f4 JA |
2537 | for (i = 0; i < set->nr_hw_queues; i++) |
2538 | if (!__blk_mq_alloc_rq_map(set, i)) | |
a5164405 | 2539 | goto out_unwind; |
a5164405 JA |
2540 | |
2541 | return 0; | |
2542 | ||
2543 | out_unwind: | |
2544 | while (--i >= 0) | |
cc71a6f4 | 2545 | blk_mq_free_rq_map(set->tags[i]); |
a5164405 | 2546 | |
a5164405 JA |
2547 | return -ENOMEM; |
2548 | } | |
2549 | ||
2550 | /* | |
2551 | * Allocate the request maps associated with this tag_set. Note that this | |
2552 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
2553 | * will be updated to reflect the allocated depth. | |
2554 | */ | |
2555 | static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
2556 | { | |
2557 | unsigned int depth; | |
2558 | int err; | |
2559 | ||
2560 | depth = set->queue_depth; | |
2561 | do { | |
2562 | err = __blk_mq_alloc_rq_maps(set); | |
2563 | if (!err) | |
2564 | break; | |
2565 | ||
2566 | set->queue_depth >>= 1; | |
2567 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
2568 | err = -ENOMEM; | |
2569 | break; | |
2570 | } | |
2571 | } while (set->queue_depth); | |
2572 | ||
2573 | if (!set->queue_depth || err) { | |
2574 | pr_err("blk-mq: failed to allocate request map\n"); | |
2575 | return -ENOMEM; | |
2576 | } | |
2577 | ||
2578 | if (depth != set->queue_depth) | |
2579 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
2580 | depth, set->queue_depth); | |
2581 | ||
2582 | return 0; | |
2583 | } | |
2584 | ||
ebe8bddb OS |
2585 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) |
2586 | { | |
2587 | if (set->ops->map_queues) | |
2588 | return set->ops->map_queues(set); | |
2589 | else | |
2590 | return blk_mq_map_queues(set); | |
2591 | } | |
2592 | ||
a4391c64 JA |
2593 | /* |
2594 | * Alloc a tag set to be associated with one or more request queues. | |
2595 | * May fail with EINVAL for various error conditions. May adjust the | |
2596 | * requested depth down, if if it too large. In that case, the set | |
2597 | * value will be stored in set->queue_depth. | |
2598 | */ | |
24d2f903 CH |
2599 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) |
2600 | { | |
da695ba2 CH |
2601 | int ret; |
2602 | ||
205fb5f5 BVA |
2603 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); |
2604 | ||
24d2f903 CH |
2605 | if (!set->nr_hw_queues) |
2606 | return -EINVAL; | |
a4391c64 | 2607 | if (!set->queue_depth) |
24d2f903 CH |
2608 | return -EINVAL; |
2609 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
2610 | return -EINVAL; | |
2611 | ||
7d7e0f90 | 2612 | if (!set->ops->queue_rq) |
24d2f903 CH |
2613 | return -EINVAL; |
2614 | ||
de148297 ML |
2615 | if (!set->ops->get_budget ^ !set->ops->put_budget) |
2616 | return -EINVAL; | |
2617 | ||
a4391c64 JA |
2618 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { |
2619 | pr_info("blk-mq: reduced tag depth to %u\n", | |
2620 | BLK_MQ_MAX_DEPTH); | |
2621 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
2622 | } | |
24d2f903 | 2623 | |
6637fadf SL |
2624 | /* |
2625 | * If a crashdump is active, then we are potentially in a very | |
2626 | * memory constrained environment. Limit us to 1 queue and | |
2627 | * 64 tags to prevent using too much memory. | |
2628 | */ | |
2629 | if (is_kdump_kernel()) { | |
2630 | set->nr_hw_queues = 1; | |
2631 | set->queue_depth = min(64U, set->queue_depth); | |
2632 | } | |
868f2f0b KB |
2633 | /* |
2634 | * There is no use for more h/w queues than cpus. | |
2635 | */ | |
2636 | if (set->nr_hw_queues > nr_cpu_ids) | |
2637 | set->nr_hw_queues = nr_cpu_ids; | |
6637fadf | 2638 | |
868f2f0b | 2639 | set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *), |
24d2f903 CH |
2640 | GFP_KERNEL, set->numa_node); |
2641 | if (!set->tags) | |
a5164405 | 2642 | return -ENOMEM; |
24d2f903 | 2643 | |
da695ba2 CH |
2644 | ret = -ENOMEM; |
2645 | set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids, | |
2646 | GFP_KERNEL, set->numa_node); | |
bdd17e75 CH |
2647 | if (!set->mq_map) |
2648 | goto out_free_tags; | |
2649 | ||
ebe8bddb | 2650 | ret = blk_mq_update_queue_map(set); |
da695ba2 CH |
2651 | if (ret) |
2652 | goto out_free_mq_map; | |
2653 | ||
2654 | ret = blk_mq_alloc_rq_maps(set); | |
2655 | if (ret) | |
bdd17e75 | 2656 | goto out_free_mq_map; |
24d2f903 | 2657 | |
0d2602ca JA |
2658 | mutex_init(&set->tag_list_lock); |
2659 | INIT_LIST_HEAD(&set->tag_list); | |
2660 | ||
24d2f903 | 2661 | return 0; |
bdd17e75 CH |
2662 | |
2663 | out_free_mq_map: | |
2664 | kfree(set->mq_map); | |
2665 | set->mq_map = NULL; | |
2666 | out_free_tags: | |
5676e7b6 RE |
2667 | kfree(set->tags); |
2668 | set->tags = NULL; | |
da695ba2 | 2669 | return ret; |
24d2f903 CH |
2670 | } |
2671 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
2672 | ||
2673 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
2674 | { | |
2675 | int i; | |
2676 | ||
cc71a6f4 JA |
2677 | for (i = 0; i < nr_cpu_ids; i++) |
2678 | blk_mq_free_map_and_requests(set, i); | |
484b4061 | 2679 | |
bdd17e75 CH |
2680 | kfree(set->mq_map); |
2681 | set->mq_map = NULL; | |
2682 | ||
981bd189 | 2683 | kfree(set->tags); |
5676e7b6 | 2684 | set->tags = NULL; |
24d2f903 CH |
2685 | } |
2686 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
2687 | ||
e3a2b3f9 JA |
2688 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) |
2689 | { | |
2690 | struct blk_mq_tag_set *set = q->tag_set; | |
2691 | struct blk_mq_hw_ctx *hctx; | |
2692 | int i, ret; | |
2693 | ||
bd166ef1 | 2694 | if (!set) |
e3a2b3f9 JA |
2695 | return -EINVAL; |
2696 | ||
70f36b60 | 2697 | blk_mq_freeze_queue(q); |
70f36b60 | 2698 | |
e3a2b3f9 JA |
2699 | ret = 0; |
2700 | queue_for_each_hw_ctx(q, hctx, i) { | |
e9137d4b KB |
2701 | if (!hctx->tags) |
2702 | continue; | |
bd166ef1 JA |
2703 | /* |
2704 | * If we're using an MQ scheduler, just update the scheduler | |
2705 | * queue depth. This is similar to what the old code would do. | |
2706 | */ | |
70f36b60 JA |
2707 | if (!hctx->sched_tags) { |
2708 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, | |
2709 | min(nr, set->queue_depth), | |
2710 | false); | |
2711 | } else { | |
2712 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
2713 | nr, true); | |
2714 | } | |
e3a2b3f9 JA |
2715 | if (ret) |
2716 | break; | |
2717 | } | |
2718 | ||
2719 | if (!ret) | |
2720 | q->nr_requests = nr; | |
2721 | ||
70f36b60 | 2722 | blk_mq_unfreeze_queue(q); |
70f36b60 | 2723 | |
e3a2b3f9 JA |
2724 | return ret; |
2725 | } | |
2726 | ||
e4dc2b32 KB |
2727 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, |
2728 | int nr_hw_queues) | |
868f2f0b KB |
2729 | { |
2730 | struct request_queue *q; | |
2731 | ||
705cda97 BVA |
2732 | lockdep_assert_held(&set->tag_list_lock); |
2733 | ||
868f2f0b KB |
2734 | if (nr_hw_queues > nr_cpu_ids) |
2735 | nr_hw_queues = nr_cpu_ids; | |
2736 | if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues) | |
2737 | return; | |
2738 | ||
2739 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2740 | blk_mq_freeze_queue(q); | |
2741 | ||
2742 | set->nr_hw_queues = nr_hw_queues; | |
ebe8bddb | 2743 | blk_mq_update_queue_map(set); |
868f2f0b KB |
2744 | list_for_each_entry(q, &set->tag_list, tag_set_list) { |
2745 | blk_mq_realloc_hw_ctxs(set, q); | |
4b855ad3 | 2746 | blk_mq_queue_reinit(q); |
868f2f0b KB |
2747 | } |
2748 | ||
2749 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
2750 | blk_mq_unfreeze_queue(q); | |
2751 | } | |
e4dc2b32 KB |
2752 | |
2753 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
2754 | { | |
2755 | mutex_lock(&set->tag_list_lock); | |
2756 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
2757 | mutex_unlock(&set->tag_list_lock); | |
2758 | } | |
868f2f0b KB |
2759 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); |
2760 | ||
34dbad5d OS |
2761 | /* Enable polling stats and return whether they were already enabled. */ |
2762 | static bool blk_poll_stats_enable(struct request_queue *q) | |
2763 | { | |
2764 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2765 | test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags)) | |
2766 | return true; | |
2767 | blk_stat_add_callback(q, q->poll_cb); | |
2768 | return false; | |
2769 | } | |
2770 | ||
2771 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
2772 | { | |
2773 | /* | |
2774 | * We don't arm the callback if polling stats are not enabled or the | |
2775 | * callback is already active. | |
2776 | */ | |
2777 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
2778 | blk_stat_is_active(q->poll_cb)) | |
2779 | return; | |
2780 | ||
2781 | blk_stat_activate_msecs(q->poll_cb, 100); | |
2782 | } | |
2783 | ||
2784 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
2785 | { | |
2786 | struct request_queue *q = cb->data; | |
720b8ccc | 2787 | int bucket; |
34dbad5d | 2788 | |
720b8ccc SB |
2789 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { |
2790 | if (cb->stat[bucket].nr_samples) | |
2791 | q->poll_stat[bucket] = cb->stat[bucket]; | |
2792 | } | |
34dbad5d OS |
2793 | } |
2794 | ||
64f1c21e JA |
2795 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, |
2796 | struct blk_mq_hw_ctx *hctx, | |
2797 | struct request *rq) | |
2798 | { | |
64f1c21e | 2799 | unsigned long ret = 0; |
720b8ccc | 2800 | int bucket; |
64f1c21e JA |
2801 | |
2802 | /* | |
2803 | * If stats collection isn't on, don't sleep but turn it on for | |
2804 | * future users | |
2805 | */ | |
34dbad5d | 2806 | if (!blk_poll_stats_enable(q)) |
64f1c21e JA |
2807 | return 0; |
2808 | ||
64f1c21e JA |
2809 | /* |
2810 | * As an optimistic guess, use half of the mean service time | |
2811 | * for this type of request. We can (and should) make this smarter. | |
2812 | * For instance, if the completion latencies are tight, we can | |
2813 | * get closer than just half the mean. This is especially | |
2814 | * important on devices where the completion latencies are longer | |
720b8ccc SB |
2815 | * than ~10 usec. We do use the stats for the relevant IO size |
2816 | * if available which does lead to better estimates. | |
64f1c21e | 2817 | */ |
720b8ccc SB |
2818 | bucket = blk_mq_poll_stats_bkt(rq); |
2819 | if (bucket < 0) | |
2820 | return ret; | |
2821 | ||
2822 | if (q->poll_stat[bucket].nr_samples) | |
2823 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
64f1c21e JA |
2824 | |
2825 | return ret; | |
2826 | } | |
2827 | ||
06426adf | 2828 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, |
64f1c21e | 2829 | struct blk_mq_hw_ctx *hctx, |
06426adf JA |
2830 | struct request *rq) |
2831 | { | |
2832 | struct hrtimer_sleeper hs; | |
2833 | enum hrtimer_mode mode; | |
64f1c21e | 2834 | unsigned int nsecs; |
06426adf JA |
2835 | ktime_t kt; |
2836 | ||
64f1c21e JA |
2837 | if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags)) |
2838 | return false; | |
2839 | ||
2840 | /* | |
2841 | * poll_nsec can be: | |
2842 | * | |
2843 | * -1: don't ever hybrid sleep | |
2844 | * 0: use half of prev avg | |
2845 | * >0: use this specific value | |
2846 | */ | |
2847 | if (q->poll_nsec == -1) | |
2848 | return false; | |
2849 | else if (q->poll_nsec > 0) | |
2850 | nsecs = q->poll_nsec; | |
2851 | else | |
2852 | nsecs = blk_mq_poll_nsecs(q, hctx, rq); | |
2853 | ||
2854 | if (!nsecs) | |
06426adf JA |
2855 | return false; |
2856 | ||
2857 | set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags); | |
2858 | ||
2859 | /* | |
2860 | * This will be replaced with the stats tracking code, using | |
2861 | * 'avg_completion_time / 2' as the pre-sleep target. | |
2862 | */ | |
8b0e1953 | 2863 | kt = nsecs; |
06426adf JA |
2864 | |
2865 | mode = HRTIMER_MODE_REL; | |
2866 | hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode); | |
2867 | hrtimer_set_expires(&hs.timer, kt); | |
2868 | ||
2869 | hrtimer_init_sleeper(&hs, current); | |
2870 | do { | |
2871 | if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) | |
2872 | break; | |
2873 | set_current_state(TASK_UNINTERRUPTIBLE); | |
2874 | hrtimer_start_expires(&hs.timer, mode); | |
2875 | if (hs.task) | |
2876 | io_schedule(); | |
2877 | hrtimer_cancel(&hs.timer); | |
2878 | mode = HRTIMER_MODE_ABS; | |
2879 | } while (hs.task && !signal_pending(current)); | |
2880 | ||
2881 | __set_current_state(TASK_RUNNING); | |
2882 | destroy_hrtimer_on_stack(&hs.timer); | |
2883 | return true; | |
2884 | } | |
2885 | ||
bbd7bb70 JA |
2886 | static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq) |
2887 | { | |
2888 | struct request_queue *q = hctx->queue; | |
2889 | long state; | |
2890 | ||
06426adf JA |
2891 | /* |
2892 | * If we sleep, have the caller restart the poll loop to reset | |
2893 | * the state. Like for the other success return cases, the | |
2894 | * caller is responsible for checking if the IO completed. If | |
2895 | * the IO isn't complete, we'll get called again and will go | |
2896 | * straight to the busy poll loop. | |
2897 | */ | |
64f1c21e | 2898 | if (blk_mq_poll_hybrid_sleep(q, hctx, rq)) |
06426adf JA |
2899 | return true; |
2900 | ||
bbd7bb70 JA |
2901 | hctx->poll_considered++; |
2902 | ||
2903 | state = current->state; | |
2904 | while (!need_resched()) { | |
2905 | int ret; | |
2906 | ||
2907 | hctx->poll_invoked++; | |
2908 | ||
2909 | ret = q->mq_ops->poll(hctx, rq->tag); | |
2910 | if (ret > 0) { | |
2911 | hctx->poll_success++; | |
2912 | set_current_state(TASK_RUNNING); | |
2913 | return true; | |
2914 | } | |
2915 | ||
2916 | if (signal_pending_state(state, current)) | |
2917 | set_current_state(TASK_RUNNING); | |
2918 | ||
2919 | if (current->state == TASK_RUNNING) | |
2920 | return true; | |
2921 | if (ret < 0) | |
2922 | break; | |
2923 | cpu_relax(); | |
2924 | } | |
2925 | ||
2926 | return false; | |
2927 | } | |
2928 | ||
2929 | bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie) | |
2930 | { | |
2931 | struct blk_mq_hw_ctx *hctx; | |
2932 | struct blk_plug *plug; | |
2933 | struct request *rq; | |
2934 | ||
2935 | if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) || | |
2936 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
2937 | return false; | |
2938 | ||
2939 | plug = current->plug; | |
2940 | if (plug) | |
2941 | blk_flush_plug_list(plug, false); | |
2942 | ||
2943 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
bd166ef1 JA |
2944 | if (!blk_qc_t_is_internal(cookie)) |
2945 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3a07bb1d | 2946 | else { |
bd166ef1 | 2947 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); |
3a07bb1d JA |
2948 | /* |
2949 | * With scheduling, if the request has completed, we'll | |
2950 | * get a NULL return here, as we clear the sched tag when | |
2951 | * that happens. The request still remains valid, like always, | |
2952 | * so we should be safe with just the NULL check. | |
2953 | */ | |
2954 | if (!rq) | |
2955 | return false; | |
2956 | } | |
bbd7bb70 JA |
2957 | |
2958 | return __blk_mq_poll(hctx, rq); | |
2959 | } | |
2960 | EXPORT_SYMBOL_GPL(blk_mq_poll); | |
2961 | ||
320ae51f JA |
2962 | static int __init blk_mq_init(void) |
2963 | { | |
fc13457f JA |
2964 | /* |
2965 | * See comment in block/blk.h rq_atomic_flags enum | |
2966 | */ | |
2967 | BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) != | |
2968 | (REQ_ATOM_COMPLETE / BITS_PER_BYTE)); | |
2969 | ||
9467f859 TG |
2970 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, |
2971 | blk_mq_hctx_notify_dead); | |
320ae51f JA |
2972 | return 0; |
2973 | } | |
2974 | subsys_initcall(blk_mq_init); |