]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * Block multiqueue core code | |
4 | * | |
5 | * Copyright (C) 2013-2014 Jens Axboe | |
6 | * Copyright (C) 2013-2014 Christoph Hellwig | |
7 | */ | |
8 | #include <linux/kernel.h> | |
9 | #include <linux/module.h> | |
10 | #include <linux/backing-dev.h> | |
11 | #include <linux/bio.h> | |
12 | #include <linux/blkdev.h> | |
13 | #include <linux/kmemleak.h> | |
14 | #include <linux/mm.h> | |
15 | #include <linux/init.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/smp.h> | |
19 | #include <linux/llist.h> | |
20 | #include <linux/list_sort.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/cache.h> | |
23 | #include <linux/sched/sysctl.h> | |
24 | #include <linux/sched/topology.h> | |
25 | #include <linux/sched/signal.h> | |
26 | #include <linux/delay.h> | |
27 | #include <linux/crash_dump.h> | |
28 | #include <linux/prefetch.h> | |
29 | #include <linux/blk-crypto.h> | |
30 | ||
31 | #include <trace/events/block.h> | |
32 | ||
33 | #include <linux/blk-mq.h> | |
34 | #include <linux/t10-pi.h> | |
35 | #include "blk.h" | |
36 | #include "blk-mq.h" | |
37 | #include "blk-mq-debugfs.h" | |
38 | #include "blk-mq-tag.h" | |
39 | #include "blk-pm.h" | |
40 | #include "blk-stat.h" | |
41 | #include "blk-mq-sched.h" | |
42 | #include "blk-rq-qos.h" | |
43 | ||
44 | static DEFINE_PER_CPU(struct list_head, blk_cpu_done); | |
45 | ||
46 | static void blk_mq_poll_stats_start(struct request_queue *q); | |
47 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb); | |
48 | ||
49 | static int blk_mq_poll_stats_bkt(const struct request *rq) | |
50 | { | |
51 | int ddir, sectors, bucket; | |
52 | ||
53 | ddir = rq_data_dir(rq); | |
54 | sectors = blk_rq_stats_sectors(rq); | |
55 | ||
56 | bucket = ddir + 2 * ilog2(sectors); | |
57 | ||
58 | if (bucket < 0) | |
59 | return -1; | |
60 | else if (bucket >= BLK_MQ_POLL_STATS_BKTS) | |
61 | return ddir + BLK_MQ_POLL_STATS_BKTS - 2; | |
62 | ||
63 | return bucket; | |
64 | } | |
65 | ||
66 | /* | |
67 | * Check if any of the ctx, dispatch list or elevator | |
68 | * have pending work in this hardware queue. | |
69 | */ | |
70 | static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) | |
71 | { | |
72 | return !list_empty_careful(&hctx->dispatch) || | |
73 | sbitmap_any_bit_set(&hctx->ctx_map) || | |
74 | blk_mq_sched_has_work(hctx); | |
75 | } | |
76 | ||
77 | /* | |
78 | * Mark this ctx as having pending work in this hardware queue | |
79 | */ | |
80 | static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, | |
81 | struct blk_mq_ctx *ctx) | |
82 | { | |
83 | const int bit = ctx->index_hw[hctx->type]; | |
84 | ||
85 | if (!sbitmap_test_bit(&hctx->ctx_map, bit)) | |
86 | sbitmap_set_bit(&hctx->ctx_map, bit); | |
87 | } | |
88 | ||
89 | static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, | |
90 | struct blk_mq_ctx *ctx) | |
91 | { | |
92 | const int bit = ctx->index_hw[hctx->type]; | |
93 | ||
94 | sbitmap_clear_bit(&hctx->ctx_map, bit); | |
95 | } | |
96 | ||
97 | struct mq_inflight { | |
98 | struct hd_struct *part; | |
99 | unsigned int inflight[2]; | |
100 | }; | |
101 | ||
102 | static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx, | |
103 | struct request *rq, void *priv, | |
104 | bool reserved) | |
105 | { | |
106 | struct mq_inflight *mi = priv; | |
107 | ||
108 | if (rq->part == mi->part) | |
109 | mi->inflight[rq_data_dir(rq)]++; | |
110 | ||
111 | return true; | |
112 | } | |
113 | ||
114 | unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part) | |
115 | { | |
116 | struct mq_inflight mi = { .part = part }; | |
117 | ||
118 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); | |
119 | ||
120 | return mi.inflight[0] + mi.inflight[1]; | |
121 | } | |
122 | ||
123 | void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part, | |
124 | unsigned int inflight[2]) | |
125 | { | |
126 | struct mq_inflight mi = { .part = part }; | |
127 | ||
128 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); | |
129 | inflight[0] = mi.inflight[0]; | |
130 | inflight[1] = mi.inflight[1]; | |
131 | } | |
132 | ||
133 | void blk_freeze_queue_start(struct request_queue *q) | |
134 | { | |
135 | mutex_lock(&q->mq_freeze_lock); | |
136 | if (++q->mq_freeze_depth == 1) { | |
137 | percpu_ref_kill(&q->q_usage_counter); | |
138 | mutex_unlock(&q->mq_freeze_lock); | |
139 | if (queue_is_mq(q)) | |
140 | blk_mq_run_hw_queues(q, false); | |
141 | } else { | |
142 | mutex_unlock(&q->mq_freeze_lock); | |
143 | } | |
144 | } | |
145 | EXPORT_SYMBOL_GPL(blk_freeze_queue_start); | |
146 | ||
147 | void blk_mq_freeze_queue_wait(struct request_queue *q) | |
148 | { | |
149 | wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); | |
150 | } | |
151 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); | |
152 | ||
153 | int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, | |
154 | unsigned long timeout) | |
155 | { | |
156 | return wait_event_timeout(q->mq_freeze_wq, | |
157 | percpu_ref_is_zero(&q->q_usage_counter), | |
158 | timeout); | |
159 | } | |
160 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); | |
161 | ||
162 | /* | |
163 | * Guarantee no request is in use, so we can change any data structure of | |
164 | * the queue afterward. | |
165 | */ | |
166 | void blk_freeze_queue(struct request_queue *q) | |
167 | { | |
168 | /* | |
169 | * In the !blk_mq case we are only calling this to kill the | |
170 | * q_usage_counter, otherwise this increases the freeze depth | |
171 | * and waits for it to return to zero. For this reason there is | |
172 | * no blk_unfreeze_queue(), and blk_freeze_queue() is not | |
173 | * exported to drivers as the only user for unfreeze is blk_mq. | |
174 | */ | |
175 | blk_freeze_queue_start(q); | |
176 | blk_mq_freeze_queue_wait(q); | |
177 | } | |
178 | ||
179 | void blk_mq_freeze_queue(struct request_queue *q) | |
180 | { | |
181 | /* | |
182 | * ...just an alias to keep freeze and unfreeze actions balanced | |
183 | * in the blk_mq_* namespace | |
184 | */ | |
185 | blk_freeze_queue(q); | |
186 | } | |
187 | EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); | |
188 | ||
189 | void blk_mq_unfreeze_queue(struct request_queue *q) | |
190 | { | |
191 | mutex_lock(&q->mq_freeze_lock); | |
192 | q->mq_freeze_depth--; | |
193 | WARN_ON_ONCE(q->mq_freeze_depth < 0); | |
194 | if (!q->mq_freeze_depth) { | |
195 | percpu_ref_resurrect(&q->q_usage_counter); | |
196 | wake_up_all(&q->mq_freeze_wq); | |
197 | } | |
198 | mutex_unlock(&q->mq_freeze_lock); | |
199 | } | |
200 | EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); | |
201 | ||
202 | /* | |
203 | * FIXME: replace the scsi_internal_device_*block_nowait() calls in the | |
204 | * mpt3sas driver such that this function can be removed. | |
205 | */ | |
206 | void blk_mq_quiesce_queue_nowait(struct request_queue *q) | |
207 | { | |
208 | blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); | |
209 | } | |
210 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); | |
211 | ||
212 | /** | |
213 | * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished | |
214 | * @q: request queue. | |
215 | * | |
216 | * Note: this function does not prevent that the struct request end_io() | |
217 | * callback function is invoked. Once this function is returned, we make | |
218 | * sure no dispatch can happen until the queue is unquiesced via | |
219 | * blk_mq_unquiesce_queue(). | |
220 | */ | |
221 | void blk_mq_quiesce_queue(struct request_queue *q) | |
222 | { | |
223 | struct blk_mq_hw_ctx *hctx; | |
224 | unsigned int i; | |
225 | bool rcu = false; | |
226 | ||
227 | blk_mq_quiesce_queue_nowait(q); | |
228 | ||
229 | queue_for_each_hw_ctx(q, hctx, i) { | |
230 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
231 | synchronize_srcu(hctx->srcu); | |
232 | else | |
233 | rcu = true; | |
234 | } | |
235 | if (rcu) | |
236 | synchronize_rcu(); | |
237 | } | |
238 | EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); | |
239 | ||
240 | /* | |
241 | * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() | |
242 | * @q: request queue. | |
243 | * | |
244 | * This function recovers queue into the state before quiescing | |
245 | * which is done by blk_mq_quiesce_queue. | |
246 | */ | |
247 | void blk_mq_unquiesce_queue(struct request_queue *q) | |
248 | { | |
249 | blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); | |
250 | ||
251 | /* dispatch requests which are inserted during quiescing */ | |
252 | blk_mq_run_hw_queues(q, true); | |
253 | } | |
254 | EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); | |
255 | ||
256 | void blk_mq_wake_waiters(struct request_queue *q) | |
257 | { | |
258 | struct blk_mq_hw_ctx *hctx; | |
259 | unsigned int i; | |
260 | ||
261 | queue_for_each_hw_ctx(q, hctx, i) | |
262 | if (blk_mq_hw_queue_mapped(hctx)) | |
263 | blk_mq_tag_wakeup_all(hctx->tags, true); | |
264 | } | |
265 | ||
266 | /* | |
267 | * Only need start/end time stamping if we have iostat or | |
268 | * blk stats enabled, or using an IO scheduler. | |
269 | */ | |
270 | static inline bool blk_mq_need_time_stamp(struct request *rq) | |
271 | { | |
272 | return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator; | |
273 | } | |
274 | ||
275 | static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, | |
276 | unsigned int tag, u64 alloc_time_ns) | |
277 | { | |
278 | struct blk_mq_tags *tags = blk_mq_tags_from_data(data); | |
279 | struct request *rq = tags->static_rqs[tag]; | |
280 | ||
281 | if (data->q->elevator) { | |
282 | rq->tag = BLK_MQ_NO_TAG; | |
283 | rq->internal_tag = tag; | |
284 | } else { | |
285 | rq->tag = tag; | |
286 | rq->internal_tag = BLK_MQ_NO_TAG; | |
287 | } | |
288 | ||
289 | /* csd/requeue_work/fifo_time is initialized before use */ | |
290 | rq->q = data->q; | |
291 | rq->mq_ctx = data->ctx; | |
292 | rq->mq_hctx = data->hctx; | |
293 | rq->rq_flags = 0; | |
294 | rq->cmd_flags = data->cmd_flags; | |
295 | if (data->flags & BLK_MQ_REQ_PREEMPT) | |
296 | rq->rq_flags |= RQF_PREEMPT; | |
297 | if (blk_queue_io_stat(data->q)) | |
298 | rq->rq_flags |= RQF_IO_STAT; | |
299 | INIT_LIST_HEAD(&rq->queuelist); | |
300 | INIT_HLIST_NODE(&rq->hash); | |
301 | RB_CLEAR_NODE(&rq->rb_node); | |
302 | rq->rq_disk = NULL; | |
303 | rq->part = NULL; | |
304 | #ifdef CONFIG_BLK_RQ_ALLOC_TIME | |
305 | rq->alloc_time_ns = alloc_time_ns; | |
306 | #endif | |
307 | if (blk_mq_need_time_stamp(rq)) | |
308 | rq->start_time_ns = ktime_get_ns(); | |
309 | else | |
310 | rq->start_time_ns = 0; | |
311 | rq->io_start_time_ns = 0; | |
312 | rq->stats_sectors = 0; | |
313 | rq->nr_phys_segments = 0; | |
314 | #if defined(CONFIG_BLK_DEV_INTEGRITY) | |
315 | rq->nr_integrity_segments = 0; | |
316 | #endif | |
317 | blk_crypto_rq_set_defaults(rq); | |
318 | /* tag was already set */ | |
319 | WRITE_ONCE(rq->deadline, 0); | |
320 | ||
321 | rq->timeout = 0; | |
322 | ||
323 | rq->end_io = NULL; | |
324 | rq->end_io_data = NULL; | |
325 | ||
326 | data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++; | |
327 | refcount_set(&rq->ref, 1); | |
328 | ||
329 | if (!op_is_flush(data->cmd_flags)) { | |
330 | struct elevator_queue *e = data->q->elevator; | |
331 | ||
332 | rq->elv.icq = NULL; | |
333 | if (e && e->type->ops.prepare_request) { | |
334 | if (e->type->icq_cache) | |
335 | blk_mq_sched_assign_ioc(rq); | |
336 | ||
337 | e->type->ops.prepare_request(rq); | |
338 | rq->rq_flags |= RQF_ELVPRIV; | |
339 | } | |
340 | } | |
341 | ||
342 | data->hctx->queued++; | |
343 | return rq; | |
344 | } | |
345 | ||
346 | static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data) | |
347 | { | |
348 | struct request_queue *q = data->q; | |
349 | struct elevator_queue *e = q->elevator; | |
350 | u64 alloc_time_ns = 0; | |
351 | unsigned int tag; | |
352 | ||
353 | /* alloc_time includes depth and tag waits */ | |
354 | if (blk_queue_rq_alloc_time(q)) | |
355 | alloc_time_ns = ktime_get_ns(); | |
356 | ||
357 | if (data->cmd_flags & REQ_NOWAIT) | |
358 | data->flags |= BLK_MQ_REQ_NOWAIT; | |
359 | ||
360 | if (e) { | |
361 | /* | |
362 | * Flush requests are special and go directly to the | |
363 | * dispatch list. Don't include reserved tags in the | |
364 | * limiting, as it isn't useful. | |
365 | */ | |
366 | if (!op_is_flush(data->cmd_flags) && | |
367 | e->type->ops.limit_depth && | |
368 | !(data->flags & BLK_MQ_REQ_RESERVED)) | |
369 | e->type->ops.limit_depth(data->cmd_flags, data); | |
370 | } | |
371 | ||
372 | retry: | |
373 | data->ctx = blk_mq_get_ctx(q); | |
374 | data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); | |
375 | if (!e) | |
376 | blk_mq_tag_busy(data->hctx); | |
377 | ||
378 | /* | |
379 | * Waiting allocations only fail because of an inactive hctx. In that | |
380 | * case just retry the hctx assignment and tag allocation as CPU hotplug | |
381 | * should have migrated us to an online CPU by now. | |
382 | */ | |
383 | tag = blk_mq_get_tag(data); | |
384 | if (tag == BLK_MQ_NO_TAG) { | |
385 | if (data->flags & BLK_MQ_REQ_NOWAIT) | |
386 | return NULL; | |
387 | ||
388 | /* | |
389 | * Give up the CPU and sleep for a random short time to ensure | |
390 | * that thread using a realtime scheduling class are migrated | |
391 | * off the CPU, and thus off the hctx that is going away. | |
392 | */ | |
393 | msleep(3); | |
394 | goto retry; | |
395 | } | |
396 | return blk_mq_rq_ctx_init(data, tag, alloc_time_ns); | |
397 | } | |
398 | ||
399 | struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, | |
400 | blk_mq_req_flags_t flags) | |
401 | { | |
402 | struct blk_mq_alloc_data data = { | |
403 | .q = q, | |
404 | .flags = flags, | |
405 | .cmd_flags = op, | |
406 | }; | |
407 | struct request *rq; | |
408 | int ret; | |
409 | ||
410 | ret = blk_queue_enter(q, flags); | |
411 | if (ret) | |
412 | return ERR_PTR(ret); | |
413 | ||
414 | rq = __blk_mq_alloc_request(&data); | |
415 | if (!rq) | |
416 | goto out_queue_exit; | |
417 | rq->__data_len = 0; | |
418 | rq->__sector = (sector_t) -1; | |
419 | rq->bio = rq->biotail = NULL; | |
420 | return rq; | |
421 | out_queue_exit: | |
422 | blk_queue_exit(q); | |
423 | return ERR_PTR(-EWOULDBLOCK); | |
424 | } | |
425 | EXPORT_SYMBOL(blk_mq_alloc_request); | |
426 | ||
427 | struct request *blk_mq_alloc_request_hctx(struct request_queue *q, | |
428 | unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx) | |
429 | { | |
430 | struct blk_mq_alloc_data data = { | |
431 | .q = q, | |
432 | .flags = flags, | |
433 | .cmd_flags = op, | |
434 | }; | |
435 | u64 alloc_time_ns = 0; | |
436 | unsigned int cpu; | |
437 | unsigned int tag; | |
438 | int ret; | |
439 | ||
440 | /* alloc_time includes depth and tag waits */ | |
441 | if (blk_queue_rq_alloc_time(q)) | |
442 | alloc_time_ns = ktime_get_ns(); | |
443 | ||
444 | /* | |
445 | * If the tag allocator sleeps we could get an allocation for a | |
446 | * different hardware context. No need to complicate the low level | |
447 | * allocator for this for the rare use case of a command tied to | |
448 | * a specific queue. | |
449 | */ | |
450 | if (WARN_ON_ONCE(!(flags & (BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_RESERVED)))) | |
451 | return ERR_PTR(-EINVAL); | |
452 | ||
453 | if (hctx_idx >= q->nr_hw_queues) | |
454 | return ERR_PTR(-EIO); | |
455 | ||
456 | ret = blk_queue_enter(q, flags); | |
457 | if (ret) | |
458 | return ERR_PTR(ret); | |
459 | ||
460 | /* | |
461 | * Check if the hardware context is actually mapped to anything. | |
462 | * If not tell the caller that it should skip this queue. | |
463 | */ | |
464 | ret = -EXDEV; | |
465 | data.hctx = q->queue_hw_ctx[hctx_idx]; | |
466 | if (!blk_mq_hw_queue_mapped(data.hctx)) | |
467 | goto out_queue_exit; | |
468 | cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); | |
469 | data.ctx = __blk_mq_get_ctx(q, cpu); | |
470 | ||
471 | if (!q->elevator) | |
472 | blk_mq_tag_busy(data.hctx); | |
473 | ||
474 | ret = -EWOULDBLOCK; | |
475 | tag = blk_mq_get_tag(&data); | |
476 | if (tag == BLK_MQ_NO_TAG) | |
477 | goto out_queue_exit; | |
478 | return blk_mq_rq_ctx_init(&data, tag, alloc_time_ns); | |
479 | ||
480 | out_queue_exit: | |
481 | blk_queue_exit(q); | |
482 | return ERR_PTR(ret); | |
483 | } | |
484 | EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); | |
485 | ||
486 | static void __blk_mq_free_request(struct request *rq) | |
487 | { | |
488 | struct request_queue *q = rq->q; | |
489 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
490 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
491 | const int sched_tag = rq->internal_tag; | |
492 | ||
493 | blk_crypto_free_request(rq); | |
494 | blk_pm_mark_last_busy(rq); | |
495 | rq->mq_hctx = NULL; | |
496 | if (rq->tag != BLK_MQ_NO_TAG) | |
497 | blk_mq_put_tag(hctx->tags, ctx, rq->tag); | |
498 | if (sched_tag != BLK_MQ_NO_TAG) | |
499 | blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); | |
500 | blk_mq_sched_restart(hctx); | |
501 | blk_queue_exit(q); | |
502 | } | |
503 | ||
504 | void blk_mq_free_request(struct request *rq) | |
505 | { | |
506 | struct request_queue *q = rq->q; | |
507 | struct elevator_queue *e = q->elevator; | |
508 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
509 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
510 | ||
511 | if (rq->rq_flags & RQF_ELVPRIV) { | |
512 | if (e && e->type->ops.finish_request) | |
513 | e->type->ops.finish_request(rq); | |
514 | if (rq->elv.icq) { | |
515 | put_io_context(rq->elv.icq->ioc); | |
516 | rq->elv.icq = NULL; | |
517 | } | |
518 | } | |
519 | ||
520 | ctx->rq_completed[rq_is_sync(rq)]++; | |
521 | if (rq->rq_flags & RQF_MQ_INFLIGHT) | |
522 | __blk_mq_dec_active_requests(hctx); | |
523 | ||
524 | if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) | |
525 | laptop_io_completion(q->backing_dev_info); | |
526 | ||
527 | rq_qos_done(q, rq); | |
528 | ||
529 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
530 | if (refcount_dec_and_test(&rq->ref)) | |
531 | __blk_mq_free_request(rq); | |
532 | } | |
533 | EXPORT_SYMBOL_GPL(blk_mq_free_request); | |
534 | ||
535 | inline void __blk_mq_end_request(struct request *rq, blk_status_t error) | |
536 | { | |
537 | u64 now = 0; | |
538 | ||
539 | if (blk_mq_need_time_stamp(rq)) | |
540 | now = ktime_get_ns(); | |
541 | ||
542 | if (rq->rq_flags & RQF_STATS) { | |
543 | blk_mq_poll_stats_start(rq->q); | |
544 | blk_stat_add(rq, now); | |
545 | } | |
546 | ||
547 | blk_mq_sched_completed_request(rq, now); | |
548 | ||
549 | blk_account_io_done(rq, now); | |
550 | ||
551 | if (rq->end_io) { | |
552 | rq_qos_done(rq->q, rq); | |
553 | rq->end_io(rq, error); | |
554 | } else { | |
555 | blk_mq_free_request(rq); | |
556 | } | |
557 | } | |
558 | EXPORT_SYMBOL(__blk_mq_end_request); | |
559 | ||
560 | void blk_mq_end_request(struct request *rq, blk_status_t error) | |
561 | { | |
562 | if (blk_update_request(rq, error, blk_rq_bytes(rq))) | |
563 | BUG(); | |
564 | __blk_mq_end_request(rq, error); | |
565 | } | |
566 | EXPORT_SYMBOL(blk_mq_end_request); | |
567 | ||
568 | /* | |
569 | * Softirq action handler - move entries to local list and loop over them | |
570 | * while passing them to the queue registered handler. | |
571 | */ | |
572 | static __latent_entropy void blk_done_softirq(struct softirq_action *h) | |
573 | { | |
574 | struct list_head *cpu_list, local_list; | |
575 | ||
576 | local_irq_disable(); | |
577 | cpu_list = this_cpu_ptr(&blk_cpu_done); | |
578 | list_replace_init(cpu_list, &local_list); | |
579 | local_irq_enable(); | |
580 | ||
581 | while (!list_empty(&local_list)) { | |
582 | struct request *rq; | |
583 | ||
584 | rq = list_entry(local_list.next, struct request, ipi_list); | |
585 | list_del_init(&rq->ipi_list); | |
586 | rq->q->mq_ops->complete(rq); | |
587 | } | |
588 | } | |
589 | ||
590 | static void blk_mq_trigger_softirq(struct request *rq) | |
591 | { | |
592 | struct list_head *list; | |
593 | unsigned long flags; | |
594 | ||
595 | local_irq_save(flags); | |
596 | list = this_cpu_ptr(&blk_cpu_done); | |
597 | list_add_tail(&rq->ipi_list, list); | |
598 | ||
599 | /* | |
600 | * If the list only contains our just added request, signal a raise of | |
601 | * the softirq. If there are already entries there, someone already | |
602 | * raised the irq but it hasn't run yet. | |
603 | */ | |
604 | if (list->next == &rq->ipi_list) | |
605 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
606 | local_irq_restore(flags); | |
607 | } | |
608 | ||
609 | static int blk_softirq_cpu_dead(unsigned int cpu) | |
610 | { | |
611 | /* | |
612 | * If a CPU goes away, splice its entries to the current CPU | |
613 | * and trigger a run of the softirq | |
614 | */ | |
615 | local_irq_disable(); | |
616 | list_splice_init(&per_cpu(blk_cpu_done, cpu), | |
617 | this_cpu_ptr(&blk_cpu_done)); | |
618 | raise_softirq_irqoff(BLOCK_SOFTIRQ); | |
619 | local_irq_enable(); | |
620 | ||
621 | return 0; | |
622 | } | |
623 | ||
624 | ||
625 | static void __blk_mq_complete_request_remote(void *data) | |
626 | { | |
627 | struct request *rq = data; | |
628 | ||
629 | /* | |
630 | * For most of single queue controllers, there is only one irq vector | |
631 | * for handling I/O completion, and the only irq's affinity is set | |
632 | * to all possible CPUs. On most of ARCHs, this affinity means the irq | |
633 | * is handled on one specific CPU. | |
634 | * | |
635 | * So complete I/O requests in softirq context in case of single queue | |
636 | * devices to avoid degrading I/O performance due to irqsoff latency. | |
637 | */ | |
638 | if (rq->q->nr_hw_queues == 1) | |
639 | blk_mq_trigger_softirq(rq); | |
640 | else | |
641 | rq->q->mq_ops->complete(rq); | |
642 | } | |
643 | ||
644 | static inline bool blk_mq_complete_need_ipi(struct request *rq) | |
645 | { | |
646 | int cpu = raw_smp_processor_id(); | |
647 | ||
648 | if (!IS_ENABLED(CONFIG_SMP) || | |
649 | !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) | |
650 | return false; | |
651 | ||
652 | /* same CPU or cache domain? Complete locally */ | |
653 | if (cpu == rq->mq_ctx->cpu || | |
654 | (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && | |
655 | cpus_share_cache(cpu, rq->mq_ctx->cpu))) | |
656 | return false; | |
657 | ||
658 | /* don't try to IPI to an offline CPU */ | |
659 | return cpu_online(rq->mq_ctx->cpu); | |
660 | } | |
661 | ||
662 | bool blk_mq_complete_request_remote(struct request *rq) | |
663 | { | |
664 | WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); | |
665 | ||
666 | /* | |
667 | * For a polled request, always complete locallly, it's pointless | |
668 | * to redirect the completion. | |
669 | */ | |
670 | if (rq->cmd_flags & REQ_HIPRI) | |
671 | return false; | |
672 | ||
673 | if (blk_mq_complete_need_ipi(rq)) { | |
674 | rq->csd.func = __blk_mq_complete_request_remote; | |
675 | rq->csd.info = rq; | |
676 | rq->csd.flags = 0; | |
677 | smp_call_function_single_async(rq->mq_ctx->cpu, &rq->csd); | |
678 | } else { | |
679 | if (rq->q->nr_hw_queues > 1) | |
680 | return false; | |
681 | blk_mq_trigger_softirq(rq); | |
682 | } | |
683 | ||
684 | return true; | |
685 | } | |
686 | EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); | |
687 | ||
688 | /** | |
689 | * blk_mq_complete_request - end I/O on a request | |
690 | * @rq: the request being processed | |
691 | * | |
692 | * Description: | |
693 | * Complete a request by scheduling the ->complete_rq operation. | |
694 | **/ | |
695 | void blk_mq_complete_request(struct request *rq) | |
696 | { | |
697 | if (!blk_mq_complete_request_remote(rq)) | |
698 | rq->q->mq_ops->complete(rq); | |
699 | } | |
700 | EXPORT_SYMBOL(blk_mq_complete_request); | |
701 | ||
702 | static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx) | |
703 | __releases(hctx->srcu) | |
704 | { | |
705 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) | |
706 | rcu_read_unlock(); | |
707 | else | |
708 | srcu_read_unlock(hctx->srcu, srcu_idx); | |
709 | } | |
710 | ||
711 | static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx) | |
712 | __acquires(hctx->srcu) | |
713 | { | |
714 | if (!(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
715 | /* shut up gcc false positive */ | |
716 | *srcu_idx = 0; | |
717 | rcu_read_lock(); | |
718 | } else | |
719 | *srcu_idx = srcu_read_lock(hctx->srcu); | |
720 | } | |
721 | ||
722 | /** | |
723 | * blk_mq_start_request - Start processing a request | |
724 | * @rq: Pointer to request to be started | |
725 | * | |
726 | * Function used by device drivers to notify the block layer that a request | |
727 | * is going to be processed now, so blk layer can do proper initializations | |
728 | * such as starting the timeout timer. | |
729 | */ | |
730 | void blk_mq_start_request(struct request *rq) | |
731 | { | |
732 | struct request_queue *q = rq->q; | |
733 | ||
734 | trace_block_rq_issue(q, rq); | |
735 | ||
736 | if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) { | |
737 | rq->io_start_time_ns = ktime_get_ns(); | |
738 | rq->stats_sectors = blk_rq_sectors(rq); | |
739 | rq->rq_flags |= RQF_STATS; | |
740 | rq_qos_issue(q, rq); | |
741 | } | |
742 | ||
743 | WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); | |
744 | ||
745 | blk_add_timer(rq); | |
746 | WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); | |
747 | ||
748 | #ifdef CONFIG_BLK_DEV_INTEGRITY | |
749 | if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) | |
750 | q->integrity.profile->prepare_fn(rq); | |
751 | #endif | |
752 | } | |
753 | EXPORT_SYMBOL(blk_mq_start_request); | |
754 | ||
755 | static void __blk_mq_requeue_request(struct request *rq) | |
756 | { | |
757 | struct request_queue *q = rq->q; | |
758 | ||
759 | blk_mq_put_driver_tag(rq); | |
760 | ||
761 | trace_block_rq_requeue(q, rq); | |
762 | rq_qos_requeue(q, rq); | |
763 | ||
764 | if (blk_mq_request_started(rq)) { | |
765 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
766 | rq->rq_flags &= ~RQF_TIMED_OUT; | |
767 | } | |
768 | } | |
769 | ||
770 | void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) | |
771 | { | |
772 | __blk_mq_requeue_request(rq); | |
773 | ||
774 | /* this request will be re-inserted to io scheduler queue */ | |
775 | blk_mq_sched_requeue_request(rq); | |
776 | ||
777 | BUG_ON(!list_empty(&rq->queuelist)); | |
778 | blk_mq_add_to_requeue_list(rq, true, kick_requeue_list); | |
779 | } | |
780 | EXPORT_SYMBOL(blk_mq_requeue_request); | |
781 | ||
782 | static void blk_mq_requeue_work(struct work_struct *work) | |
783 | { | |
784 | struct request_queue *q = | |
785 | container_of(work, struct request_queue, requeue_work.work); | |
786 | LIST_HEAD(rq_list); | |
787 | struct request *rq, *next; | |
788 | ||
789 | spin_lock_irq(&q->requeue_lock); | |
790 | list_splice_init(&q->requeue_list, &rq_list); | |
791 | spin_unlock_irq(&q->requeue_lock); | |
792 | ||
793 | list_for_each_entry_safe(rq, next, &rq_list, queuelist) { | |
794 | if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP))) | |
795 | continue; | |
796 | ||
797 | rq->rq_flags &= ~RQF_SOFTBARRIER; | |
798 | list_del_init(&rq->queuelist); | |
799 | /* | |
800 | * If RQF_DONTPREP, rq has contained some driver specific | |
801 | * data, so insert it to hctx dispatch list to avoid any | |
802 | * merge. | |
803 | */ | |
804 | if (rq->rq_flags & RQF_DONTPREP) | |
805 | blk_mq_request_bypass_insert(rq, false, false); | |
806 | else | |
807 | blk_mq_sched_insert_request(rq, true, false, false); | |
808 | } | |
809 | ||
810 | while (!list_empty(&rq_list)) { | |
811 | rq = list_entry(rq_list.next, struct request, queuelist); | |
812 | list_del_init(&rq->queuelist); | |
813 | blk_mq_sched_insert_request(rq, false, false, false); | |
814 | } | |
815 | ||
816 | blk_mq_run_hw_queues(q, false); | |
817 | } | |
818 | ||
819 | void blk_mq_add_to_requeue_list(struct request *rq, bool at_head, | |
820 | bool kick_requeue_list) | |
821 | { | |
822 | struct request_queue *q = rq->q; | |
823 | unsigned long flags; | |
824 | ||
825 | /* | |
826 | * We abuse this flag that is otherwise used by the I/O scheduler to | |
827 | * request head insertion from the workqueue. | |
828 | */ | |
829 | BUG_ON(rq->rq_flags & RQF_SOFTBARRIER); | |
830 | ||
831 | spin_lock_irqsave(&q->requeue_lock, flags); | |
832 | if (at_head) { | |
833 | rq->rq_flags |= RQF_SOFTBARRIER; | |
834 | list_add(&rq->queuelist, &q->requeue_list); | |
835 | } else { | |
836 | list_add_tail(&rq->queuelist, &q->requeue_list); | |
837 | } | |
838 | spin_unlock_irqrestore(&q->requeue_lock, flags); | |
839 | ||
840 | if (kick_requeue_list) | |
841 | blk_mq_kick_requeue_list(q); | |
842 | } | |
843 | ||
844 | void blk_mq_kick_requeue_list(struct request_queue *q) | |
845 | { | |
846 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); | |
847 | } | |
848 | EXPORT_SYMBOL(blk_mq_kick_requeue_list); | |
849 | ||
850 | void blk_mq_delay_kick_requeue_list(struct request_queue *q, | |
851 | unsigned long msecs) | |
852 | { | |
853 | kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, | |
854 | msecs_to_jiffies(msecs)); | |
855 | } | |
856 | EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); | |
857 | ||
858 | struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag) | |
859 | { | |
860 | if (tag < tags->nr_tags) { | |
861 | prefetch(tags->rqs[tag]); | |
862 | return tags->rqs[tag]; | |
863 | } | |
864 | ||
865 | return NULL; | |
866 | } | |
867 | EXPORT_SYMBOL(blk_mq_tag_to_rq); | |
868 | ||
869 | static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
870 | void *priv, bool reserved) | |
871 | { | |
872 | /* | |
873 | * If we find a request that isn't idle and the queue matches, | |
874 | * we know the queue is busy. Return false to stop the iteration. | |
875 | */ | |
876 | if (blk_mq_request_started(rq) && rq->q == hctx->queue) { | |
877 | bool *busy = priv; | |
878 | ||
879 | *busy = true; | |
880 | return false; | |
881 | } | |
882 | ||
883 | return true; | |
884 | } | |
885 | ||
886 | bool blk_mq_queue_inflight(struct request_queue *q) | |
887 | { | |
888 | bool busy = false; | |
889 | ||
890 | blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); | |
891 | return busy; | |
892 | } | |
893 | EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); | |
894 | ||
895 | static void blk_mq_rq_timed_out(struct request *req, bool reserved) | |
896 | { | |
897 | req->rq_flags |= RQF_TIMED_OUT; | |
898 | if (req->q->mq_ops->timeout) { | |
899 | enum blk_eh_timer_return ret; | |
900 | ||
901 | ret = req->q->mq_ops->timeout(req, reserved); | |
902 | if (ret == BLK_EH_DONE) | |
903 | return; | |
904 | WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); | |
905 | } | |
906 | ||
907 | blk_add_timer(req); | |
908 | } | |
909 | ||
910 | static bool blk_mq_req_expired(struct request *rq, unsigned long *next) | |
911 | { | |
912 | unsigned long deadline; | |
913 | ||
914 | if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) | |
915 | return false; | |
916 | if (rq->rq_flags & RQF_TIMED_OUT) | |
917 | return false; | |
918 | ||
919 | deadline = READ_ONCE(rq->deadline); | |
920 | if (time_after_eq(jiffies, deadline)) | |
921 | return true; | |
922 | ||
923 | if (*next == 0) | |
924 | *next = deadline; | |
925 | else if (time_after(*next, deadline)) | |
926 | *next = deadline; | |
927 | return false; | |
928 | } | |
929 | ||
930 | static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx, | |
931 | struct request *rq, void *priv, bool reserved) | |
932 | { | |
933 | unsigned long *next = priv; | |
934 | ||
935 | /* | |
936 | * Just do a quick check if it is expired before locking the request in | |
937 | * so we're not unnecessarilly synchronizing across CPUs. | |
938 | */ | |
939 | if (!blk_mq_req_expired(rq, next)) | |
940 | return true; | |
941 | ||
942 | /* | |
943 | * We have reason to believe the request may be expired. Take a | |
944 | * reference on the request to lock this request lifetime into its | |
945 | * currently allocated context to prevent it from being reallocated in | |
946 | * the event the completion by-passes this timeout handler. | |
947 | * | |
948 | * If the reference was already released, then the driver beat the | |
949 | * timeout handler to posting a natural completion. | |
950 | */ | |
951 | if (!refcount_inc_not_zero(&rq->ref)) | |
952 | return true; | |
953 | ||
954 | /* | |
955 | * The request is now locked and cannot be reallocated underneath the | |
956 | * timeout handler's processing. Re-verify this exact request is truly | |
957 | * expired; if it is not expired, then the request was completed and | |
958 | * reallocated as a new request. | |
959 | */ | |
960 | if (blk_mq_req_expired(rq, next)) | |
961 | blk_mq_rq_timed_out(rq, reserved); | |
962 | ||
963 | if (is_flush_rq(rq, hctx)) | |
964 | rq->end_io(rq, 0); | |
965 | else if (refcount_dec_and_test(&rq->ref)) | |
966 | __blk_mq_free_request(rq); | |
967 | ||
968 | return true; | |
969 | } | |
970 | ||
971 | static void blk_mq_timeout_work(struct work_struct *work) | |
972 | { | |
973 | struct request_queue *q = | |
974 | container_of(work, struct request_queue, timeout_work); | |
975 | unsigned long next = 0; | |
976 | struct blk_mq_hw_ctx *hctx; | |
977 | int i; | |
978 | ||
979 | /* A deadlock might occur if a request is stuck requiring a | |
980 | * timeout at the same time a queue freeze is waiting | |
981 | * completion, since the timeout code would not be able to | |
982 | * acquire the queue reference here. | |
983 | * | |
984 | * That's why we don't use blk_queue_enter here; instead, we use | |
985 | * percpu_ref_tryget directly, because we need to be able to | |
986 | * obtain a reference even in the short window between the queue | |
987 | * starting to freeze, by dropping the first reference in | |
988 | * blk_freeze_queue_start, and the moment the last request is | |
989 | * consumed, marked by the instant q_usage_counter reaches | |
990 | * zero. | |
991 | */ | |
992 | if (!percpu_ref_tryget(&q->q_usage_counter)) | |
993 | return; | |
994 | ||
995 | blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next); | |
996 | ||
997 | if (next != 0) { | |
998 | mod_timer(&q->timeout, next); | |
999 | } else { | |
1000 | /* | |
1001 | * Request timeouts are handled as a forward rolling timer. If | |
1002 | * we end up here it means that no requests are pending and | |
1003 | * also that no request has been pending for a while. Mark | |
1004 | * each hctx as idle. | |
1005 | */ | |
1006 | queue_for_each_hw_ctx(q, hctx, i) { | |
1007 | /* the hctx may be unmapped, so check it here */ | |
1008 | if (blk_mq_hw_queue_mapped(hctx)) | |
1009 | blk_mq_tag_idle(hctx); | |
1010 | } | |
1011 | } | |
1012 | blk_queue_exit(q); | |
1013 | } | |
1014 | ||
1015 | struct flush_busy_ctx_data { | |
1016 | struct blk_mq_hw_ctx *hctx; | |
1017 | struct list_head *list; | |
1018 | }; | |
1019 | ||
1020 | static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) | |
1021 | { | |
1022 | struct flush_busy_ctx_data *flush_data = data; | |
1023 | struct blk_mq_hw_ctx *hctx = flush_data->hctx; | |
1024 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1025 | enum hctx_type type = hctx->type; | |
1026 | ||
1027 | spin_lock(&ctx->lock); | |
1028 | list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); | |
1029 | sbitmap_clear_bit(sb, bitnr); | |
1030 | spin_unlock(&ctx->lock); | |
1031 | return true; | |
1032 | } | |
1033 | ||
1034 | /* | |
1035 | * Process software queues that have been marked busy, splicing them | |
1036 | * to the for-dispatch | |
1037 | */ | |
1038 | void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) | |
1039 | { | |
1040 | struct flush_busy_ctx_data data = { | |
1041 | .hctx = hctx, | |
1042 | .list = list, | |
1043 | }; | |
1044 | ||
1045 | sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); | |
1046 | } | |
1047 | EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); | |
1048 | ||
1049 | struct dispatch_rq_data { | |
1050 | struct blk_mq_hw_ctx *hctx; | |
1051 | struct request *rq; | |
1052 | }; | |
1053 | ||
1054 | static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, | |
1055 | void *data) | |
1056 | { | |
1057 | struct dispatch_rq_data *dispatch_data = data; | |
1058 | struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; | |
1059 | struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; | |
1060 | enum hctx_type type = hctx->type; | |
1061 | ||
1062 | spin_lock(&ctx->lock); | |
1063 | if (!list_empty(&ctx->rq_lists[type])) { | |
1064 | dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); | |
1065 | list_del_init(&dispatch_data->rq->queuelist); | |
1066 | if (list_empty(&ctx->rq_lists[type])) | |
1067 | sbitmap_clear_bit(sb, bitnr); | |
1068 | } | |
1069 | spin_unlock(&ctx->lock); | |
1070 | ||
1071 | return !dispatch_data->rq; | |
1072 | } | |
1073 | ||
1074 | struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, | |
1075 | struct blk_mq_ctx *start) | |
1076 | { | |
1077 | unsigned off = start ? start->index_hw[hctx->type] : 0; | |
1078 | struct dispatch_rq_data data = { | |
1079 | .hctx = hctx, | |
1080 | .rq = NULL, | |
1081 | }; | |
1082 | ||
1083 | __sbitmap_for_each_set(&hctx->ctx_map, off, | |
1084 | dispatch_rq_from_ctx, &data); | |
1085 | ||
1086 | return data.rq; | |
1087 | } | |
1088 | ||
1089 | static inline unsigned int queued_to_index(unsigned int queued) | |
1090 | { | |
1091 | if (!queued) | |
1092 | return 0; | |
1093 | ||
1094 | return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1); | |
1095 | } | |
1096 | ||
1097 | static bool __blk_mq_get_driver_tag(struct request *rq) | |
1098 | { | |
1099 | struct sbitmap_queue *bt = rq->mq_hctx->tags->bitmap_tags; | |
1100 | unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; | |
1101 | int tag; | |
1102 | ||
1103 | blk_mq_tag_busy(rq->mq_hctx); | |
1104 | ||
1105 | if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { | |
1106 | bt = rq->mq_hctx->tags->breserved_tags; | |
1107 | tag_offset = 0; | |
1108 | } else { | |
1109 | if (!hctx_may_queue(rq->mq_hctx, bt)) | |
1110 | return false; | |
1111 | } | |
1112 | ||
1113 | tag = __sbitmap_queue_get(bt); | |
1114 | if (tag == BLK_MQ_NO_TAG) | |
1115 | return false; | |
1116 | ||
1117 | rq->tag = tag + tag_offset; | |
1118 | return true; | |
1119 | } | |
1120 | ||
1121 | static bool blk_mq_get_driver_tag(struct request *rq) | |
1122 | { | |
1123 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1124 | ||
1125 | if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_get_driver_tag(rq)) | |
1126 | return false; | |
1127 | ||
1128 | if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && | |
1129 | !(rq->rq_flags & RQF_MQ_INFLIGHT)) { | |
1130 | rq->rq_flags |= RQF_MQ_INFLIGHT; | |
1131 | __blk_mq_inc_active_requests(hctx); | |
1132 | } | |
1133 | hctx->tags->rqs[rq->tag] = rq; | |
1134 | return true; | |
1135 | } | |
1136 | ||
1137 | static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, | |
1138 | int flags, void *key) | |
1139 | { | |
1140 | struct blk_mq_hw_ctx *hctx; | |
1141 | ||
1142 | hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); | |
1143 | ||
1144 | spin_lock(&hctx->dispatch_wait_lock); | |
1145 | if (!list_empty(&wait->entry)) { | |
1146 | struct sbitmap_queue *sbq; | |
1147 | ||
1148 | list_del_init(&wait->entry); | |
1149 | sbq = hctx->tags->bitmap_tags; | |
1150 | atomic_dec(&sbq->ws_active); | |
1151 | } | |
1152 | spin_unlock(&hctx->dispatch_wait_lock); | |
1153 | ||
1154 | blk_mq_run_hw_queue(hctx, true); | |
1155 | return 1; | |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * Mark us waiting for a tag. For shared tags, this involves hooking us into | |
1160 | * the tag wakeups. For non-shared tags, we can simply mark us needing a | |
1161 | * restart. For both cases, take care to check the condition again after | |
1162 | * marking us as waiting. | |
1163 | */ | |
1164 | static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, | |
1165 | struct request *rq) | |
1166 | { | |
1167 | struct sbitmap_queue *sbq = hctx->tags->bitmap_tags; | |
1168 | struct wait_queue_head *wq; | |
1169 | wait_queue_entry_t *wait; | |
1170 | bool ret; | |
1171 | ||
1172 | if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
1173 | blk_mq_sched_mark_restart_hctx(hctx); | |
1174 | ||
1175 | /* | |
1176 | * It's possible that a tag was freed in the window between the | |
1177 | * allocation failure and adding the hardware queue to the wait | |
1178 | * queue. | |
1179 | * | |
1180 | * Don't clear RESTART here, someone else could have set it. | |
1181 | * At most this will cost an extra queue run. | |
1182 | */ | |
1183 | return blk_mq_get_driver_tag(rq); | |
1184 | } | |
1185 | ||
1186 | wait = &hctx->dispatch_wait; | |
1187 | if (!list_empty_careful(&wait->entry)) | |
1188 | return false; | |
1189 | ||
1190 | wq = &bt_wait_ptr(sbq, hctx)->wait; | |
1191 | ||
1192 | spin_lock_irq(&wq->lock); | |
1193 | spin_lock(&hctx->dispatch_wait_lock); | |
1194 | if (!list_empty(&wait->entry)) { | |
1195 | spin_unlock(&hctx->dispatch_wait_lock); | |
1196 | spin_unlock_irq(&wq->lock); | |
1197 | return false; | |
1198 | } | |
1199 | ||
1200 | atomic_inc(&sbq->ws_active); | |
1201 | wait->flags &= ~WQ_FLAG_EXCLUSIVE; | |
1202 | __add_wait_queue(wq, wait); | |
1203 | ||
1204 | /* | |
1205 | * It's possible that a tag was freed in the window between the | |
1206 | * allocation failure and adding the hardware queue to the wait | |
1207 | * queue. | |
1208 | */ | |
1209 | ret = blk_mq_get_driver_tag(rq); | |
1210 | if (!ret) { | |
1211 | spin_unlock(&hctx->dispatch_wait_lock); | |
1212 | spin_unlock_irq(&wq->lock); | |
1213 | return false; | |
1214 | } | |
1215 | ||
1216 | /* | |
1217 | * We got a tag, remove ourselves from the wait queue to ensure | |
1218 | * someone else gets the wakeup. | |
1219 | */ | |
1220 | list_del_init(&wait->entry); | |
1221 | atomic_dec(&sbq->ws_active); | |
1222 | spin_unlock(&hctx->dispatch_wait_lock); | |
1223 | spin_unlock_irq(&wq->lock); | |
1224 | ||
1225 | return true; | |
1226 | } | |
1227 | ||
1228 | #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 | |
1229 | #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 | |
1230 | /* | |
1231 | * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): | |
1232 | * - EWMA is one simple way to compute running average value | |
1233 | * - weight(7/8 and 1/8) is applied so that it can decrease exponentially | |
1234 | * - take 4 as factor for avoiding to get too small(0) result, and this | |
1235 | * factor doesn't matter because EWMA decreases exponentially | |
1236 | */ | |
1237 | static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) | |
1238 | { | |
1239 | unsigned int ewma; | |
1240 | ||
1241 | if (hctx->queue->elevator) | |
1242 | return; | |
1243 | ||
1244 | ewma = hctx->dispatch_busy; | |
1245 | ||
1246 | if (!ewma && !busy) | |
1247 | return; | |
1248 | ||
1249 | ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; | |
1250 | if (busy) | |
1251 | ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; | |
1252 | ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; | |
1253 | ||
1254 | hctx->dispatch_busy = ewma; | |
1255 | } | |
1256 | ||
1257 | #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ | |
1258 | ||
1259 | static void blk_mq_handle_dev_resource(struct request *rq, | |
1260 | struct list_head *list) | |
1261 | { | |
1262 | struct request *next = | |
1263 | list_first_entry_or_null(list, struct request, queuelist); | |
1264 | ||
1265 | /* | |
1266 | * If an I/O scheduler has been configured and we got a driver tag for | |
1267 | * the next request already, free it. | |
1268 | */ | |
1269 | if (next) | |
1270 | blk_mq_put_driver_tag(next); | |
1271 | ||
1272 | list_add(&rq->queuelist, list); | |
1273 | __blk_mq_requeue_request(rq); | |
1274 | } | |
1275 | ||
1276 | static void blk_mq_handle_zone_resource(struct request *rq, | |
1277 | struct list_head *zone_list) | |
1278 | { | |
1279 | /* | |
1280 | * If we end up here it is because we cannot dispatch a request to a | |
1281 | * specific zone due to LLD level zone-write locking or other zone | |
1282 | * related resource not being available. In this case, set the request | |
1283 | * aside in zone_list for retrying it later. | |
1284 | */ | |
1285 | list_add(&rq->queuelist, zone_list); | |
1286 | __blk_mq_requeue_request(rq); | |
1287 | } | |
1288 | ||
1289 | enum prep_dispatch { | |
1290 | PREP_DISPATCH_OK, | |
1291 | PREP_DISPATCH_NO_TAG, | |
1292 | PREP_DISPATCH_NO_BUDGET, | |
1293 | }; | |
1294 | ||
1295 | static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, | |
1296 | bool need_budget) | |
1297 | { | |
1298 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1299 | ||
1300 | if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) { | |
1301 | blk_mq_put_driver_tag(rq); | |
1302 | return PREP_DISPATCH_NO_BUDGET; | |
1303 | } | |
1304 | ||
1305 | if (!blk_mq_get_driver_tag(rq)) { | |
1306 | /* | |
1307 | * The initial allocation attempt failed, so we need to | |
1308 | * rerun the hardware queue when a tag is freed. The | |
1309 | * waitqueue takes care of that. If the queue is run | |
1310 | * before we add this entry back on the dispatch list, | |
1311 | * we'll re-run it below. | |
1312 | */ | |
1313 | if (!blk_mq_mark_tag_wait(hctx, rq)) { | |
1314 | /* | |
1315 | * All budgets not got from this function will be put | |
1316 | * together during handling partial dispatch | |
1317 | */ | |
1318 | if (need_budget) | |
1319 | blk_mq_put_dispatch_budget(rq->q); | |
1320 | return PREP_DISPATCH_NO_TAG; | |
1321 | } | |
1322 | } | |
1323 | ||
1324 | return PREP_DISPATCH_OK; | |
1325 | } | |
1326 | ||
1327 | /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ | |
1328 | static void blk_mq_release_budgets(struct request_queue *q, | |
1329 | unsigned int nr_budgets) | |
1330 | { | |
1331 | int i; | |
1332 | ||
1333 | for (i = 0; i < nr_budgets; i++) | |
1334 | blk_mq_put_dispatch_budget(q); | |
1335 | } | |
1336 | ||
1337 | /* | |
1338 | * Returns true if we did some work AND can potentially do more. | |
1339 | */ | |
1340 | bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, | |
1341 | unsigned int nr_budgets) | |
1342 | { | |
1343 | enum prep_dispatch prep; | |
1344 | struct request_queue *q = hctx->queue; | |
1345 | struct request *rq, *nxt; | |
1346 | int errors, queued; | |
1347 | blk_status_t ret = BLK_STS_OK; | |
1348 | LIST_HEAD(zone_list); | |
1349 | ||
1350 | if (list_empty(list)) | |
1351 | return false; | |
1352 | ||
1353 | /* | |
1354 | * Now process all the entries, sending them to the driver. | |
1355 | */ | |
1356 | errors = queued = 0; | |
1357 | do { | |
1358 | struct blk_mq_queue_data bd; | |
1359 | ||
1360 | rq = list_first_entry(list, struct request, queuelist); | |
1361 | ||
1362 | WARN_ON_ONCE(hctx != rq->mq_hctx); | |
1363 | prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); | |
1364 | if (prep != PREP_DISPATCH_OK) | |
1365 | break; | |
1366 | ||
1367 | list_del_init(&rq->queuelist); | |
1368 | ||
1369 | bd.rq = rq; | |
1370 | ||
1371 | /* | |
1372 | * Flag last if we have no more requests, or if we have more | |
1373 | * but can't assign a driver tag to it. | |
1374 | */ | |
1375 | if (list_empty(list)) | |
1376 | bd.last = true; | |
1377 | else { | |
1378 | nxt = list_first_entry(list, struct request, queuelist); | |
1379 | bd.last = !blk_mq_get_driver_tag(nxt); | |
1380 | } | |
1381 | ||
1382 | /* | |
1383 | * once the request is queued to lld, no need to cover the | |
1384 | * budget any more | |
1385 | */ | |
1386 | if (nr_budgets) | |
1387 | nr_budgets--; | |
1388 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1389 | switch (ret) { | |
1390 | case BLK_STS_OK: | |
1391 | queued++; | |
1392 | break; | |
1393 | case BLK_STS_RESOURCE: | |
1394 | case BLK_STS_DEV_RESOURCE: | |
1395 | blk_mq_handle_dev_resource(rq, list); | |
1396 | goto out; | |
1397 | case BLK_STS_ZONE_RESOURCE: | |
1398 | /* | |
1399 | * Move the request to zone_list and keep going through | |
1400 | * the dispatch list to find more requests the drive can | |
1401 | * accept. | |
1402 | */ | |
1403 | blk_mq_handle_zone_resource(rq, &zone_list); | |
1404 | break; | |
1405 | default: | |
1406 | errors++; | |
1407 | blk_mq_end_request(rq, BLK_STS_IOERR); | |
1408 | } | |
1409 | } while (!list_empty(list)); | |
1410 | out: | |
1411 | if (!list_empty(&zone_list)) | |
1412 | list_splice_tail_init(&zone_list, list); | |
1413 | ||
1414 | hctx->dispatched[queued_to_index(queued)]++; | |
1415 | ||
1416 | /* | |
1417 | * Any items that need requeuing? Stuff them into hctx->dispatch, | |
1418 | * that is where we will continue on next queue run. | |
1419 | */ | |
1420 | if (!list_empty(list)) { | |
1421 | bool needs_restart; | |
1422 | /* For non-shared tags, the RESTART check will suffice */ | |
1423 | bool no_tag = prep == PREP_DISPATCH_NO_TAG && | |
1424 | (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED); | |
1425 | bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET; | |
1426 | ||
1427 | blk_mq_release_budgets(q, nr_budgets); | |
1428 | ||
1429 | /* | |
1430 | * If we didn't flush the entire list, we could have told | |
1431 | * the driver there was more coming, but that turned out to | |
1432 | * be a lie. | |
1433 | */ | |
1434 | if (q->mq_ops->commit_rqs && queued) | |
1435 | q->mq_ops->commit_rqs(hctx); | |
1436 | ||
1437 | spin_lock(&hctx->lock); | |
1438 | list_splice_tail_init(list, &hctx->dispatch); | |
1439 | spin_unlock(&hctx->lock); | |
1440 | ||
1441 | /* | |
1442 | * Order adding requests to hctx->dispatch and checking | |
1443 | * SCHED_RESTART flag. The pair of this smp_mb() is the one | |
1444 | * in blk_mq_sched_restart(). Avoid restart code path to | |
1445 | * miss the new added requests to hctx->dispatch, meantime | |
1446 | * SCHED_RESTART is observed here. | |
1447 | */ | |
1448 | smp_mb(); | |
1449 | ||
1450 | /* | |
1451 | * If SCHED_RESTART was set by the caller of this function and | |
1452 | * it is no longer set that means that it was cleared by another | |
1453 | * thread and hence that a queue rerun is needed. | |
1454 | * | |
1455 | * If 'no_tag' is set, that means that we failed getting | |
1456 | * a driver tag with an I/O scheduler attached. If our dispatch | |
1457 | * waitqueue is no longer active, ensure that we run the queue | |
1458 | * AFTER adding our entries back to the list. | |
1459 | * | |
1460 | * If no I/O scheduler has been configured it is possible that | |
1461 | * the hardware queue got stopped and restarted before requests | |
1462 | * were pushed back onto the dispatch list. Rerun the queue to | |
1463 | * avoid starvation. Notes: | |
1464 | * - blk_mq_run_hw_queue() checks whether or not a queue has | |
1465 | * been stopped before rerunning a queue. | |
1466 | * - Some but not all block drivers stop a queue before | |
1467 | * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq | |
1468 | * and dm-rq. | |
1469 | * | |
1470 | * If driver returns BLK_STS_RESOURCE and SCHED_RESTART | |
1471 | * bit is set, run queue after a delay to avoid IO stalls | |
1472 | * that could otherwise occur if the queue is idle. We'll do | |
1473 | * similar if we couldn't get budget and SCHED_RESTART is set. | |
1474 | */ | |
1475 | needs_restart = blk_mq_sched_needs_restart(hctx); | |
1476 | if (!needs_restart || | |
1477 | (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) | |
1478 | blk_mq_run_hw_queue(hctx, true); | |
1479 | else if (needs_restart && (ret == BLK_STS_RESOURCE || | |
1480 | no_budget_avail)) | |
1481 | blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); | |
1482 | ||
1483 | blk_mq_update_dispatch_busy(hctx, true); | |
1484 | return false; | |
1485 | } else | |
1486 | blk_mq_update_dispatch_busy(hctx, false); | |
1487 | ||
1488 | return (queued + errors) != 0; | |
1489 | } | |
1490 | ||
1491 | /** | |
1492 | * __blk_mq_run_hw_queue - Run a hardware queue. | |
1493 | * @hctx: Pointer to the hardware queue to run. | |
1494 | * | |
1495 | * Send pending requests to the hardware. | |
1496 | */ | |
1497 | static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1498 | { | |
1499 | int srcu_idx; | |
1500 | ||
1501 | /* | |
1502 | * We should be running this queue from one of the CPUs that | |
1503 | * are mapped to it. | |
1504 | * | |
1505 | * There are at least two related races now between setting | |
1506 | * hctx->next_cpu from blk_mq_hctx_next_cpu() and running | |
1507 | * __blk_mq_run_hw_queue(): | |
1508 | * | |
1509 | * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(), | |
1510 | * but later it becomes online, then this warning is harmless | |
1511 | * at all | |
1512 | * | |
1513 | * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(), | |
1514 | * but later it becomes offline, then the warning can't be | |
1515 | * triggered, and we depend on blk-mq timeout handler to | |
1516 | * handle dispatched requests to this hctx | |
1517 | */ | |
1518 | if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) && | |
1519 | cpu_online(hctx->next_cpu)) { | |
1520 | printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n", | |
1521 | raw_smp_processor_id(), | |
1522 | cpumask_empty(hctx->cpumask) ? "inactive": "active"); | |
1523 | dump_stack(); | |
1524 | } | |
1525 | ||
1526 | /* | |
1527 | * We can't run the queue inline with ints disabled. Ensure that | |
1528 | * we catch bad users of this early. | |
1529 | */ | |
1530 | WARN_ON_ONCE(in_interrupt()); | |
1531 | ||
1532 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
1533 | ||
1534 | hctx_lock(hctx, &srcu_idx); | |
1535 | blk_mq_sched_dispatch_requests(hctx); | |
1536 | hctx_unlock(hctx, srcu_idx); | |
1537 | } | |
1538 | ||
1539 | static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) | |
1540 | { | |
1541 | int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); | |
1542 | ||
1543 | if (cpu >= nr_cpu_ids) | |
1544 | cpu = cpumask_first(hctx->cpumask); | |
1545 | return cpu; | |
1546 | } | |
1547 | ||
1548 | /* | |
1549 | * It'd be great if the workqueue API had a way to pass | |
1550 | * in a mask and had some smarts for more clever placement. | |
1551 | * For now we just round-robin here, switching for every | |
1552 | * BLK_MQ_CPU_WORK_BATCH queued items. | |
1553 | */ | |
1554 | static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) | |
1555 | { | |
1556 | bool tried = false; | |
1557 | int next_cpu = hctx->next_cpu; | |
1558 | ||
1559 | if (hctx->queue->nr_hw_queues == 1) | |
1560 | return WORK_CPU_UNBOUND; | |
1561 | ||
1562 | if (--hctx->next_cpu_batch <= 0) { | |
1563 | select_cpu: | |
1564 | next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, | |
1565 | cpu_online_mask); | |
1566 | if (next_cpu >= nr_cpu_ids) | |
1567 | next_cpu = blk_mq_first_mapped_cpu(hctx); | |
1568 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
1569 | } | |
1570 | ||
1571 | /* | |
1572 | * Do unbound schedule if we can't find a online CPU for this hctx, | |
1573 | * and it should only happen in the path of handling CPU DEAD. | |
1574 | */ | |
1575 | if (!cpu_online(next_cpu)) { | |
1576 | if (!tried) { | |
1577 | tried = true; | |
1578 | goto select_cpu; | |
1579 | } | |
1580 | ||
1581 | /* | |
1582 | * Make sure to re-select CPU next time once after CPUs | |
1583 | * in hctx->cpumask become online again. | |
1584 | */ | |
1585 | hctx->next_cpu = next_cpu; | |
1586 | hctx->next_cpu_batch = 1; | |
1587 | return WORK_CPU_UNBOUND; | |
1588 | } | |
1589 | ||
1590 | hctx->next_cpu = next_cpu; | |
1591 | return next_cpu; | |
1592 | } | |
1593 | ||
1594 | /** | |
1595 | * __blk_mq_delay_run_hw_queue - Run (or schedule to run) a hardware queue. | |
1596 | * @hctx: Pointer to the hardware queue to run. | |
1597 | * @async: If we want to run the queue asynchronously. | |
1598 | * @msecs: Microseconds of delay to wait before running the queue. | |
1599 | * | |
1600 | * If !@async, try to run the queue now. Else, run the queue asynchronously and | |
1601 | * with a delay of @msecs. | |
1602 | */ | |
1603 | static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async, | |
1604 | unsigned long msecs) | |
1605 | { | |
1606 | if (unlikely(blk_mq_hctx_stopped(hctx))) | |
1607 | return; | |
1608 | ||
1609 | if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) { | |
1610 | int cpu = get_cpu(); | |
1611 | if (cpumask_test_cpu(cpu, hctx->cpumask)) { | |
1612 | __blk_mq_run_hw_queue(hctx); | |
1613 | put_cpu(); | |
1614 | return; | |
1615 | } | |
1616 | ||
1617 | put_cpu(); | |
1618 | } | |
1619 | ||
1620 | kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, | |
1621 | msecs_to_jiffies(msecs)); | |
1622 | } | |
1623 | ||
1624 | /** | |
1625 | * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. | |
1626 | * @hctx: Pointer to the hardware queue to run. | |
1627 | * @msecs: Microseconds of delay to wait before running the queue. | |
1628 | * | |
1629 | * Run a hardware queue asynchronously with a delay of @msecs. | |
1630 | */ | |
1631 | void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) | |
1632 | { | |
1633 | __blk_mq_delay_run_hw_queue(hctx, true, msecs); | |
1634 | } | |
1635 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); | |
1636 | ||
1637 | /** | |
1638 | * blk_mq_run_hw_queue - Start to run a hardware queue. | |
1639 | * @hctx: Pointer to the hardware queue to run. | |
1640 | * @async: If we want to run the queue asynchronously. | |
1641 | * | |
1642 | * Check if the request queue is not in a quiesced state and if there are | |
1643 | * pending requests to be sent. If this is true, run the queue to send requests | |
1644 | * to hardware. | |
1645 | */ | |
1646 | void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1647 | { | |
1648 | int srcu_idx; | |
1649 | bool need_run; | |
1650 | ||
1651 | /* | |
1652 | * When queue is quiesced, we may be switching io scheduler, or | |
1653 | * updating nr_hw_queues, or other things, and we can't run queue | |
1654 | * any more, even __blk_mq_hctx_has_pending() can't be called safely. | |
1655 | * | |
1656 | * And queue will be rerun in blk_mq_unquiesce_queue() if it is | |
1657 | * quiesced. | |
1658 | */ | |
1659 | hctx_lock(hctx, &srcu_idx); | |
1660 | need_run = !blk_queue_quiesced(hctx->queue) && | |
1661 | blk_mq_hctx_has_pending(hctx); | |
1662 | hctx_unlock(hctx, srcu_idx); | |
1663 | ||
1664 | if (need_run) | |
1665 | __blk_mq_delay_run_hw_queue(hctx, async, 0); | |
1666 | } | |
1667 | EXPORT_SYMBOL(blk_mq_run_hw_queue); | |
1668 | ||
1669 | /** | |
1670 | * blk_mq_run_hw_queue - Run all hardware queues in a request queue. | |
1671 | * @q: Pointer to the request queue to run. | |
1672 | * @async: If we want to run the queue asynchronously. | |
1673 | */ | |
1674 | void blk_mq_run_hw_queues(struct request_queue *q, bool async) | |
1675 | { | |
1676 | struct blk_mq_hw_ctx *hctx; | |
1677 | int i; | |
1678 | ||
1679 | queue_for_each_hw_ctx(q, hctx, i) { | |
1680 | if (blk_mq_hctx_stopped(hctx)) | |
1681 | continue; | |
1682 | ||
1683 | blk_mq_run_hw_queue(hctx, async); | |
1684 | } | |
1685 | } | |
1686 | EXPORT_SYMBOL(blk_mq_run_hw_queues); | |
1687 | ||
1688 | /** | |
1689 | * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. | |
1690 | * @q: Pointer to the request queue to run. | |
1691 | * @msecs: Microseconds of delay to wait before running the queues. | |
1692 | */ | |
1693 | void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) | |
1694 | { | |
1695 | struct blk_mq_hw_ctx *hctx; | |
1696 | int i; | |
1697 | ||
1698 | queue_for_each_hw_ctx(q, hctx, i) { | |
1699 | if (blk_mq_hctx_stopped(hctx)) | |
1700 | continue; | |
1701 | ||
1702 | blk_mq_delay_run_hw_queue(hctx, msecs); | |
1703 | } | |
1704 | } | |
1705 | EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); | |
1706 | ||
1707 | /** | |
1708 | * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped | |
1709 | * @q: request queue. | |
1710 | * | |
1711 | * The caller is responsible for serializing this function against | |
1712 | * blk_mq_{start,stop}_hw_queue(). | |
1713 | */ | |
1714 | bool blk_mq_queue_stopped(struct request_queue *q) | |
1715 | { | |
1716 | struct blk_mq_hw_ctx *hctx; | |
1717 | int i; | |
1718 | ||
1719 | queue_for_each_hw_ctx(q, hctx, i) | |
1720 | if (blk_mq_hctx_stopped(hctx)) | |
1721 | return true; | |
1722 | ||
1723 | return false; | |
1724 | } | |
1725 | EXPORT_SYMBOL(blk_mq_queue_stopped); | |
1726 | ||
1727 | /* | |
1728 | * This function is often used for pausing .queue_rq() by driver when | |
1729 | * there isn't enough resource or some conditions aren't satisfied, and | |
1730 | * BLK_STS_RESOURCE is usually returned. | |
1731 | * | |
1732 | * We do not guarantee that dispatch can be drained or blocked | |
1733 | * after blk_mq_stop_hw_queue() returns. Please use | |
1734 | * blk_mq_quiesce_queue() for that requirement. | |
1735 | */ | |
1736 | void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1737 | { | |
1738 | cancel_delayed_work(&hctx->run_work); | |
1739 | ||
1740 | set_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1741 | } | |
1742 | EXPORT_SYMBOL(blk_mq_stop_hw_queue); | |
1743 | ||
1744 | /* | |
1745 | * This function is often used for pausing .queue_rq() by driver when | |
1746 | * there isn't enough resource or some conditions aren't satisfied, and | |
1747 | * BLK_STS_RESOURCE is usually returned. | |
1748 | * | |
1749 | * We do not guarantee that dispatch can be drained or blocked | |
1750 | * after blk_mq_stop_hw_queues() returns. Please use | |
1751 | * blk_mq_quiesce_queue() for that requirement. | |
1752 | */ | |
1753 | void blk_mq_stop_hw_queues(struct request_queue *q) | |
1754 | { | |
1755 | struct blk_mq_hw_ctx *hctx; | |
1756 | int i; | |
1757 | ||
1758 | queue_for_each_hw_ctx(q, hctx, i) | |
1759 | blk_mq_stop_hw_queue(hctx); | |
1760 | } | |
1761 | EXPORT_SYMBOL(blk_mq_stop_hw_queues); | |
1762 | ||
1763 | void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) | |
1764 | { | |
1765 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1766 | ||
1767 | blk_mq_run_hw_queue(hctx, false); | |
1768 | } | |
1769 | EXPORT_SYMBOL(blk_mq_start_hw_queue); | |
1770 | ||
1771 | void blk_mq_start_hw_queues(struct request_queue *q) | |
1772 | { | |
1773 | struct blk_mq_hw_ctx *hctx; | |
1774 | int i; | |
1775 | ||
1776 | queue_for_each_hw_ctx(q, hctx, i) | |
1777 | blk_mq_start_hw_queue(hctx); | |
1778 | } | |
1779 | EXPORT_SYMBOL(blk_mq_start_hw_queues); | |
1780 | ||
1781 | void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) | |
1782 | { | |
1783 | if (!blk_mq_hctx_stopped(hctx)) | |
1784 | return; | |
1785 | ||
1786 | clear_bit(BLK_MQ_S_STOPPED, &hctx->state); | |
1787 | blk_mq_run_hw_queue(hctx, async); | |
1788 | } | |
1789 | EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); | |
1790 | ||
1791 | void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) | |
1792 | { | |
1793 | struct blk_mq_hw_ctx *hctx; | |
1794 | int i; | |
1795 | ||
1796 | queue_for_each_hw_ctx(q, hctx, i) | |
1797 | blk_mq_start_stopped_hw_queue(hctx, async); | |
1798 | } | |
1799 | EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); | |
1800 | ||
1801 | static void blk_mq_run_work_fn(struct work_struct *work) | |
1802 | { | |
1803 | struct blk_mq_hw_ctx *hctx; | |
1804 | ||
1805 | hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work); | |
1806 | ||
1807 | /* | |
1808 | * If we are stopped, don't run the queue. | |
1809 | */ | |
1810 | if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) | |
1811 | return; | |
1812 | ||
1813 | __blk_mq_run_hw_queue(hctx); | |
1814 | } | |
1815 | ||
1816 | static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx, | |
1817 | struct request *rq, | |
1818 | bool at_head) | |
1819 | { | |
1820 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1821 | enum hctx_type type = hctx->type; | |
1822 | ||
1823 | lockdep_assert_held(&ctx->lock); | |
1824 | ||
1825 | trace_block_rq_insert(hctx->queue, rq); | |
1826 | ||
1827 | if (at_head) | |
1828 | list_add(&rq->queuelist, &ctx->rq_lists[type]); | |
1829 | else | |
1830 | list_add_tail(&rq->queuelist, &ctx->rq_lists[type]); | |
1831 | } | |
1832 | ||
1833 | void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq, | |
1834 | bool at_head) | |
1835 | { | |
1836 | struct blk_mq_ctx *ctx = rq->mq_ctx; | |
1837 | ||
1838 | lockdep_assert_held(&ctx->lock); | |
1839 | ||
1840 | __blk_mq_insert_req_list(hctx, rq, at_head); | |
1841 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1842 | } | |
1843 | ||
1844 | /** | |
1845 | * blk_mq_request_bypass_insert - Insert a request at dispatch list. | |
1846 | * @rq: Pointer to request to be inserted. | |
1847 | * @at_head: true if the request should be inserted at the head of the list. | |
1848 | * @run_queue: If we should run the hardware queue after inserting the request. | |
1849 | * | |
1850 | * Should only be used carefully, when the caller knows we want to | |
1851 | * bypass a potential IO scheduler on the target device. | |
1852 | */ | |
1853 | void blk_mq_request_bypass_insert(struct request *rq, bool at_head, | |
1854 | bool run_queue) | |
1855 | { | |
1856 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
1857 | ||
1858 | spin_lock(&hctx->lock); | |
1859 | if (at_head) | |
1860 | list_add(&rq->queuelist, &hctx->dispatch); | |
1861 | else | |
1862 | list_add_tail(&rq->queuelist, &hctx->dispatch); | |
1863 | spin_unlock(&hctx->lock); | |
1864 | ||
1865 | if (run_queue) | |
1866 | blk_mq_run_hw_queue(hctx, false); | |
1867 | } | |
1868 | ||
1869 | void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx, | |
1870 | struct list_head *list) | |
1871 | ||
1872 | { | |
1873 | struct request *rq; | |
1874 | enum hctx_type type = hctx->type; | |
1875 | ||
1876 | /* | |
1877 | * preemption doesn't flush plug list, so it's possible ctx->cpu is | |
1878 | * offline now | |
1879 | */ | |
1880 | list_for_each_entry(rq, list, queuelist) { | |
1881 | BUG_ON(rq->mq_ctx != ctx); | |
1882 | trace_block_rq_insert(hctx->queue, rq); | |
1883 | } | |
1884 | ||
1885 | spin_lock(&ctx->lock); | |
1886 | list_splice_tail_init(list, &ctx->rq_lists[type]); | |
1887 | blk_mq_hctx_mark_pending(hctx, ctx); | |
1888 | spin_unlock(&ctx->lock); | |
1889 | } | |
1890 | ||
1891 | static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) | |
1892 | { | |
1893 | struct request *rqa = container_of(a, struct request, queuelist); | |
1894 | struct request *rqb = container_of(b, struct request, queuelist); | |
1895 | ||
1896 | if (rqa->mq_ctx != rqb->mq_ctx) | |
1897 | return rqa->mq_ctx > rqb->mq_ctx; | |
1898 | if (rqa->mq_hctx != rqb->mq_hctx) | |
1899 | return rqa->mq_hctx > rqb->mq_hctx; | |
1900 | ||
1901 | return blk_rq_pos(rqa) > blk_rq_pos(rqb); | |
1902 | } | |
1903 | ||
1904 | void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) | |
1905 | { | |
1906 | LIST_HEAD(list); | |
1907 | ||
1908 | if (list_empty(&plug->mq_list)) | |
1909 | return; | |
1910 | list_splice_init(&plug->mq_list, &list); | |
1911 | ||
1912 | if (plug->rq_count > 2 && plug->multiple_queues) | |
1913 | list_sort(NULL, &list, plug_rq_cmp); | |
1914 | ||
1915 | plug->rq_count = 0; | |
1916 | ||
1917 | do { | |
1918 | struct list_head rq_list; | |
1919 | struct request *rq, *head_rq = list_entry_rq(list.next); | |
1920 | struct list_head *pos = &head_rq->queuelist; /* skip first */ | |
1921 | struct blk_mq_hw_ctx *this_hctx = head_rq->mq_hctx; | |
1922 | struct blk_mq_ctx *this_ctx = head_rq->mq_ctx; | |
1923 | unsigned int depth = 1; | |
1924 | ||
1925 | list_for_each_continue(pos, &list) { | |
1926 | rq = list_entry_rq(pos); | |
1927 | BUG_ON(!rq->q); | |
1928 | if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) | |
1929 | break; | |
1930 | depth++; | |
1931 | } | |
1932 | ||
1933 | list_cut_before(&rq_list, &list, pos); | |
1934 | trace_block_unplug(head_rq->q, depth, !from_schedule); | |
1935 | blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list, | |
1936 | from_schedule); | |
1937 | } while(!list_empty(&list)); | |
1938 | } | |
1939 | ||
1940 | static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, | |
1941 | unsigned int nr_segs) | |
1942 | { | |
1943 | if (bio->bi_opf & REQ_RAHEAD) | |
1944 | rq->cmd_flags |= REQ_FAILFAST_MASK; | |
1945 | ||
1946 | rq->__sector = bio->bi_iter.bi_sector; | |
1947 | rq->write_hint = bio->bi_write_hint; | |
1948 | blk_rq_bio_prep(rq, bio, nr_segs); | |
1949 | blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); | |
1950 | ||
1951 | blk_account_io_start(rq); | |
1952 | } | |
1953 | ||
1954 | static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1955 | struct request *rq, | |
1956 | blk_qc_t *cookie, bool last) | |
1957 | { | |
1958 | struct request_queue *q = rq->q; | |
1959 | struct blk_mq_queue_data bd = { | |
1960 | .rq = rq, | |
1961 | .last = last, | |
1962 | }; | |
1963 | blk_qc_t new_cookie; | |
1964 | blk_status_t ret; | |
1965 | ||
1966 | new_cookie = request_to_qc_t(hctx, rq); | |
1967 | ||
1968 | /* | |
1969 | * For OK queue, we are done. For error, caller may kill it. | |
1970 | * Any other error (busy), just add it to our list as we | |
1971 | * previously would have done. | |
1972 | */ | |
1973 | ret = q->mq_ops->queue_rq(hctx, &bd); | |
1974 | switch (ret) { | |
1975 | case BLK_STS_OK: | |
1976 | blk_mq_update_dispatch_busy(hctx, false); | |
1977 | *cookie = new_cookie; | |
1978 | break; | |
1979 | case BLK_STS_RESOURCE: | |
1980 | case BLK_STS_DEV_RESOURCE: | |
1981 | blk_mq_update_dispatch_busy(hctx, true); | |
1982 | __blk_mq_requeue_request(rq); | |
1983 | break; | |
1984 | default: | |
1985 | blk_mq_update_dispatch_busy(hctx, false); | |
1986 | *cookie = BLK_QC_T_NONE; | |
1987 | break; | |
1988 | } | |
1989 | ||
1990 | return ret; | |
1991 | } | |
1992 | ||
1993 | static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
1994 | struct request *rq, | |
1995 | blk_qc_t *cookie, | |
1996 | bool bypass_insert, bool last) | |
1997 | { | |
1998 | struct request_queue *q = rq->q; | |
1999 | bool run_queue = true; | |
2000 | ||
2001 | /* | |
2002 | * RCU or SRCU read lock is needed before checking quiesced flag. | |
2003 | * | |
2004 | * When queue is stopped or quiesced, ignore 'bypass_insert' from | |
2005 | * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller, | |
2006 | * and avoid driver to try to dispatch again. | |
2007 | */ | |
2008 | if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) { | |
2009 | run_queue = false; | |
2010 | bypass_insert = false; | |
2011 | goto insert; | |
2012 | } | |
2013 | ||
2014 | if (q->elevator && !bypass_insert) | |
2015 | goto insert; | |
2016 | ||
2017 | if (!blk_mq_get_dispatch_budget(q)) | |
2018 | goto insert; | |
2019 | ||
2020 | if (!blk_mq_get_driver_tag(rq)) { | |
2021 | blk_mq_put_dispatch_budget(q); | |
2022 | goto insert; | |
2023 | } | |
2024 | ||
2025 | return __blk_mq_issue_directly(hctx, rq, cookie, last); | |
2026 | insert: | |
2027 | if (bypass_insert) | |
2028 | return BLK_STS_RESOURCE; | |
2029 | ||
2030 | blk_mq_sched_insert_request(rq, false, run_queue, false); | |
2031 | ||
2032 | return BLK_STS_OK; | |
2033 | } | |
2034 | ||
2035 | /** | |
2036 | * blk_mq_try_issue_directly - Try to send a request directly to device driver. | |
2037 | * @hctx: Pointer of the associated hardware queue. | |
2038 | * @rq: Pointer to request to be sent. | |
2039 | * @cookie: Request queue cookie. | |
2040 | * | |
2041 | * If the device has enough resources to accept a new request now, send the | |
2042 | * request directly to device driver. Else, insert at hctx->dispatch queue, so | |
2043 | * we can try send it another time in the future. Requests inserted at this | |
2044 | * queue have higher priority. | |
2045 | */ | |
2046 | static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, | |
2047 | struct request *rq, blk_qc_t *cookie) | |
2048 | { | |
2049 | blk_status_t ret; | |
2050 | int srcu_idx; | |
2051 | ||
2052 | might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING); | |
2053 | ||
2054 | hctx_lock(hctx, &srcu_idx); | |
2055 | ||
2056 | ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true); | |
2057 | if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) | |
2058 | blk_mq_request_bypass_insert(rq, false, true); | |
2059 | else if (ret != BLK_STS_OK) | |
2060 | blk_mq_end_request(rq, ret); | |
2061 | ||
2062 | hctx_unlock(hctx, srcu_idx); | |
2063 | } | |
2064 | ||
2065 | blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) | |
2066 | { | |
2067 | blk_status_t ret; | |
2068 | int srcu_idx; | |
2069 | blk_qc_t unused_cookie; | |
2070 | struct blk_mq_hw_ctx *hctx = rq->mq_hctx; | |
2071 | ||
2072 | hctx_lock(hctx, &srcu_idx); | |
2073 | ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last); | |
2074 | hctx_unlock(hctx, srcu_idx); | |
2075 | ||
2076 | return ret; | |
2077 | } | |
2078 | ||
2079 | void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, | |
2080 | struct list_head *list) | |
2081 | { | |
2082 | int queued = 0; | |
2083 | ||
2084 | while (!list_empty(list)) { | |
2085 | blk_status_t ret; | |
2086 | struct request *rq = list_first_entry(list, struct request, | |
2087 | queuelist); | |
2088 | ||
2089 | list_del_init(&rq->queuelist); | |
2090 | ret = blk_mq_request_issue_directly(rq, list_empty(list)); | |
2091 | if (ret != BLK_STS_OK) { | |
2092 | if (ret == BLK_STS_RESOURCE || | |
2093 | ret == BLK_STS_DEV_RESOURCE) { | |
2094 | blk_mq_request_bypass_insert(rq, false, | |
2095 | list_empty(list)); | |
2096 | break; | |
2097 | } | |
2098 | blk_mq_end_request(rq, ret); | |
2099 | } else | |
2100 | queued++; | |
2101 | } | |
2102 | ||
2103 | /* | |
2104 | * If we didn't flush the entire list, we could have told | |
2105 | * the driver there was more coming, but that turned out to | |
2106 | * be a lie. | |
2107 | */ | |
2108 | if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs && queued) | |
2109 | hctx->queue->mq_ops->commit_rqs(hctx); | |
2110 | } | |
2111 | ||
2112 | static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) | |
2113 | { | |
2114 | list_add_tail(&rq->queuelist, &plug->mq_list); | |
2115 | plug->rq_count++; | |
2116 | if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) { | |
2117 | struct request *tmp; | |
2118 | ||
2119 | tmp = list_first_entry(&plug->mq_list, struct request, | |
2120 | queuelist); | |
2121 | if (tmp->q != rq->q) | |
2122 | plug->multiple_queues = true; | |
2123 | } | |
2124 | } | |
2125 | ||
2126 | /** | |
2127 | * blk_mq_submit_bio - Create and send a request to block device. | |
2128 | * @bio: Bio pointer. | |
2129 | * | |
2130 | * Builds up a request structure from @q and @bio and send to the device. The | |
2131 | * request may not be queued directly to hardware if: | |
2132 | * * This request can be merged with another one | |
2133 | * * We want to place request at plug queue for possible future merging | |
2134 | * * There is an IO scheduler active at this queue | |
2135 | * | |
2136 | * It will not queue the request if there is an error with the bio, or at the | |
2137 | * request creation. | |
2138 | * | |
2139 | * Returns: Request queue cookie. | |
2140 | */ | |
2141 | blk_qc_t blk_mq_submit_bio(struct bio *bio) | |
2142 | { | |
2143 | struct request_queue *q = bio->bi_disk->queue; | |
2144 | const int is_sync = op_is_sync(bio->bi_opf); | |
2145 | const int is_flush_fua = op_is_flush(bio->bi_opf); | |
2146 | struct blk_mq_alloc_data data = { | |
2147 | .q = q, | |
2148 | }; | |
2149 | struct request *rq; | |
2150 | struct blk_plug *plug; | |
2151 | struct request *same_queue_rq = NULL; | |
2152 | unsigned int nr_segs; | |
2153 | blk_qc_t cookie; | |
2154 | blk_status_t ret; | |
2155 | ||
2156 | blk_queue_bounce(q, &bio); | |
2157 | __blk_queue_split(&bio, &nr_segs); | |
2158 | ||
2159 | if (!bio_integrity_prep(bio)) | |
2160 | goto queue_exit; | |
2161 | ||
2162 | if (!is_flush_fua && !blk_queue_nomerges(q) && | |
2163 | blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq)) | |
2164 | goto queue_exit; | |
2165 | ||
2166 | if (blk_mq_sched_bio_merge(q, bio, nr_segs)) | |
2167 | goto queue_exit; | |
2168 | ||
2169 | rq_qos_throttle(q, bio); | |
2170 | ||
2171 | data.cmd_flags = bio->bi_opf; | |
2172 | rq = __blk_mq_alloc_request(&data); | |
2173 | if (unlikely(!rq)) { | |
2174 | rq_qos_cleanup(q, bio); | |
2175 | if (bio->bi_opf & REQ_NOWAIT) | |
2176 | bio_wouldblock_error(bio); | |
2177 | goto queue_exit; | |
2178 | } | |
2179 | ||
2180 | trace_block_getrq(q, bio, bio->bi_opf); | |
2181 | ||
2182 | rq_qos_track(q, rq, bio); | |
2183 | ||
2184 | cookie = request_to_qc_t(data.hctx, rq); | |
2185 | ||
2186 | blk_mq_bio_to_request(rq, bio, nr_segs); | |
2187 | ||
2188 | ret = blk_crypto_init_request(rq); | |
2189 | if (ret != BLK_STS_OK) { | |
2190 | bio->bi_status = ret; | |
2191 | bio_endio(bio); | |
2192 | blk_mq_free_request(rq); | |
2193 | return BLK_QC_T_NONE; | |
2194 | } | |
2195 | ||
2196 | plug = blk_mq_plug(q, bio); | |
2197 | if (unlikely(is_flush_fua)) { | |
2198 | /* Bypass scheduler for flush requests */ | |
2199 | blk_insert_flush(rq); | |
2200 | blk_mq_run_hw_queue(data.hctx, true); | |
2201 | } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs || | |
2202 | !blk_queue_nonrot(q))) { | |
2203 | /* | |
2204 | * Use plugging if we have a ->commit_rqs() hook as well, as | |
2205 | * we know the driver uses bd->last in a smart fashion. | |
2206 | * | |
2207 | * Use normal plugging if this disk is slow HDD, as sequential | |
2208 | * IO may benefit a lot from plug merging. | |
2209 | */ | |
2210 | unsigned int request_count = plug->rq_count; | |
2211 | struct request *last = NULL; | |
2212 | ||
2213 | if (!request_count) | |
2214 | trace_block_plug(q); | |
2215 | else | |
2216 | last = list_entry_rq(plug->mq_list.prev); | |
2217 | ||
2218 | if (request_count >= BLK_MAX_REQUEST_COUNT || (last && | |
2219 | blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { | |
2220 | blk_flush_plug_list(plug, false); | |
2221 | trace_block_plug(q); | |
2222 | } | |
2223 | ||
2224 | blk_add_rq_to_plug(plug, rq); | |
2225 | } else if (q->elevator) { | |
2226 | /* Insert the request at the IO scheduler queue */ | |
2227 | blk_mq_sched_insert_request(rq, false, true, true); | |
2228 | } else if (plug && !blk_queue_nomerges(q)) { | |
2229 | /* | |
2230 | * We do limited plugging. If the bio can be merged, do that. | |
2231 | * Otherwise the existing request in the plug list will be | |
2232 | * issued. So the plug list will have one request at most | |
2233 | * The plug list might get flushed before this. If that happens, | |
2234 | * the plug list is empty, and same_queue_rq is invalid. | |
2235 | */ | |
2236 | if (list_empty(&plug->mq_list)) | |
2237 | same_queue_rq = NULL; | |
2238 | if (same_queue_rq) { | |
2239 | list_del_init(&same_queue_rq->queuelist); | |
2240 | plug->rq_count--; | |
2241 | } | |
2242 | blk_add_rq_to_plug(plug, rq); | |
2243 | trace_block_plug(q); | |
2244 | ||
2245 | if (same_queue_rq) { | |
2246 | data.hctx = same_queue_rq->mq_hctx; | |
2247 | trace_block_unplug(q, 1, true); | |
2248 | blk_mq_try_issue_directly(data.hctx, same_queue_rq, | |
2249 | &cookie); | |
2250 | } | |
2251 | } else if ((q->nr_hw_queues > 1 && is_sync) || | |
2252 | !data.hctx->dispatch_busy) { | |
2253 | /* | |
2254 | * There is no scheduler and we can try to send directly | |
2255 | * to the hardware. | |
2256 | */ | |
2257 | blk_mq_try_issue_directly(data.hctx, rq, &cookie); | |
2258 | } else { | |
2259 | /* Default case. */ | |
2260 | blk_mq_sched_insert_request(rq, false, true, true); | |
2261 | } | |
2262 | ||
2263 | return cookie; | |
2264 | queue_exit: | |
2265 | blk_queue_exit(q); | |
2266 | return BLK_QC_T_NONE; | |
2267 | } | |
2268 | EXPORT_SYMBOL_GPL(blk_mq_submit_bio); /* only for request based dm */ | |
2269 | ||
2270 | void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
2271 | unsigned int hctx_idx) | |
2272 | { | |
2273 | struct page *page; | |
2274 | ||
2275 | if (tags->rqs && set->ops->exit_request) { | |
2276 | int i; | |
2277 | ||
2278 | for (i = 0; i < tags->nr_tags; i++) { | |
2279 | struct request *rq = tags->static_rqs[i]; | |
2280 | ||
2281 | if (!rq) | |
2282 | continue; | |
2283 | set->ops->exit_request(set, rq, hctx_idx); | |
2284 | tags->static_rqs[i] = NULL; | |
2285 | } | |
2286 | } | |
2287 | ||
2288 | while (!list_empty(&tags->page_list)) { | |
2289 | page = list_first_entry(&tags->page_list, struct page, lru); | |
2290 | list_del_init(&page->lru); | |
2291 | /* | |
2292 | * Remove kmemleak object previously allocated in | |
2293 | * blk_mq_alloc_rqs(). | |
2294 | */ | |
2295 | kmemleak_free(page_address(page)); | |
2296 | __free_pages(page, page->private); | |
2297 | } | |
2298 | } | |
2299 | ||
2300 | void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags) | |
2301 | { | |
2302 | kfree(tags->rqs); | |
2303 | tags->rqs = NULL; | |
2304 | kfree(tags->static_rqs); | |
2305 | tags->static_rqs = NULL; | |
2306 | ||
2307 | blk_mq_free_tags(tags, flags); | |
2308 | } | |
2309 | ||
2310 | struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, | |
2311 | unsigned int hctx_idx, | |
2312 | unsigned int nr_tags, | |
2313 | unsigned int reserved_tags, | |
2314 | unsigned int flags) | |
2315 | { | |
2316 | struct blk_mq_tags *tags; | |
2317 | int node; | |
2318 | ||
2319 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); | |
2320 | if (node == NUMA_NO_NODE) | |
2321 | node = set->numa_node; | |
2322 | ||
2323 | tags = blk_mq_init_tags(nr_tags, reserved_tags, node, flags); | |
2324 | if (!tags) | |
2325 | return NULL; | |
2326 | ||
2327 | tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), | |
2328 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2329 | node); | |
2330 | if (!tags->rqs) { | |
2331 | blk_mq_free_tags(tags, flags); | |
2332 | return NULL; | |
2333 | } | |
2334 | ||
2335 | tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), | |
2336 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, | |
2337 | node); | |
2338 | if (!tags->static_rqs) { | |
2339 | kfree(tags->rqs); | |
2340 | blk_mq_free_tags(tags, flags); | |
2341 | return NULL; | |
2342 | } | |
2343 | ||
2344 | return tags; | |
2345 | } | |
2346 | ||
2347 | static size_t order_to_size(unsigned int order) | |
2348 | { | |
2349 | return (size_t)PAGE_SIZE << order; | |
2350 | } | |
2351 | ||
2352 | static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, | |
2353 | unsigned int hctx_idx, int node) | |
2354 | { | |
2355 | int ret; | |
2356 | ||
2357 | if (set->ops->init_request) { | |
2358 | ret = set->ops->init_request(set, rq, hctx_idx, node); | |
2359 | if (ret) | |
2360 | return ret; | |
2361 | } | |
2362 | ||
2363 | WRITE_ONCE(rq->state, MQ_RQ_IDLE); | |
2364 | return 0; | |
2365 | } | |
2366 | ||
2367 | int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, | |
2368 | unsigned int hctx_idx, unsigned int depth) | |
2369 | { | |
2370 | unsigned int i, j, entries_per_page, max_order = 4; | |
2371 | size_t rq_size, left; | |
2372 | int node; | |
2373 | ||
2374 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx); | |
2375 | if (node == NUMA_NO_NODE) | |
2376 | node = set->numa_node; | |
2377 | ||
2378 | INIT_LIST_HEAD(&tags->page_list); | |
2379 | ||
2380 | /* | |
2381 | * rq_size is the size of the request plus driver payload, rounded | |
2382 | * to the cacheline size | |
2383 | */ | |
2384 | rq_size = round_up(sizeof(struct request) + set->cmd_size, | |
2385 | cache_line_size()); | |
2386 | left = rq_size * depth; | |
2387 | ||
2388 | for (i = 0; i < depth; ) { | |
2389 | int this_order = max_order; | |
2390 | struct page *page; | |
2391 | int to_do; | |
2392 | void *p; | |
2393 | ||
2394 | while (this_order && left < order_to_size(this_order - 1)) | |
2395 | this_order--; | |
2396 | ||
2397 | do { | |
2398 | page = alloc_pages_node(node, | |
2399 | GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, | |
2400 | this_order); | |
2401 | if (page) | |
2402 | break; | |
2403 | if (!this_order--) | |
2404 | break; | |
2405 | if (order_to_size(this_order) < rq_size) | |
2406 | break; | |
2407 | } while (1); | |
2408 | ||
2409 | if (!page) | |
2410 | goto fail; | |
2411 | ||
2412 | page->private = this_order; | |
2413 | list_add_tail(&page->lru, &tags->page_list); | |
2414 | ||
2415 | p = page_address(page); | |
2416 | /* | |
2417 | * Allow kmemleak to scan these pages as they contain pointers | |
2418 | * to additional allocations like via ops->init_request(). | |
2419 | */ | |
2420 | kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); | |
2421 | entries_per_page = order_to_size(this_order) / rq_size; | |
2422 | to_do = min(entries_per_page, depth - i); | |
2423 | left -= to_do * rq_size; | |
2424 | for (j = 0; j < to_do; j++) { | |
2425 | struct request *rq = p; | |
2426 | ||
2427 | tags->static_rqs[i] = rq; | |
2428 | if (blk_mq_init_request(set, rq, hctx_idx, node)) { | |
2429 | tags->static_rqs[i] = NULL; | |
2430 | goto fail; | |
2431 | } | |
2432 | ||
2433 | p += rq_size; | |
2434 | i++; | |
2435 | } | |
2436 | } | |
2437 | return 0; | |
2438 | ||
2439 | fail: | |
2440 | blk_mq_free_rqs(set, tags, hctx_idx); | |
2441 | return -ENOMEM; | |
2442 | } | |
2443 | ||
2444 | struct rq_iter_data { | |
2445 | struct blk_mq_hw_ctx *hctx; | |
2446 | bool has_rq; | |
2447 | }; | |
2448 | ||
2449 | static bool blk_mq_has_request(struct request *rq, void *data, bool reserved) | |
2450 | { | |
2451 | struct rq_iter_data *iter_data = data; | |
2452 | ||
2453 | if (rq->mq_hctx != iter_data->hctx) | |
2454 | return true; | |
2455 | iter_data->has_rq = true; | |
2456 | return false; | |
2457 | } | |
2458 | ||
2459 | static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) | |
2460 | { | |
2461 | struct blk_mq_tags *tags = hctx->sched_tags ? | |
2462 | hctx->sched_tags : hctx->tags; | |
2463 | struct rq_iter_data data = { | |
2464 | .hctx = hctx, | |
2465 | }; | |
2466 | ||
2467 | blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); | |
2468 | return data.has_rq; | |
2469 | } | |
2470 | ||
2471 | static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, | |
2472 | struct blk_mq_hw_ctx *hctx) | |
2473 | { | |
2474 | if (cpumask_next_and(-1, hctx->cpumask, cpu_online_mask) != cpu) | |
2475 | return false; | |
2476 | if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) | |
2477 | return false; | |
2478 | return true; | |
2479 | } | |
2480 | ||
2481 | static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) | |
2482 | { | |
2483 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2484 | struct blk_mq_hw_ctx, cpuhp_online); | |
2485 | ||
2486 | if (!cpumask_test_cpu(cpu, hctx->cpumask) || | |
2487 | !blk_mq_last_cpu_in_hctx(cpu, hctx)) | |
2488 | return 0; | |
2489 | ||
2490 | /* | |
2491 | * Prevent new request from being allocated on the current hctx. | |
2492 | * | |
2493 | * The smp_mb__after_atomic() Pairs with the implied barrier in | |
2494 | * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is | |
2495 | * seen once we return from the tag allocator. | |
2496 | */ | |
2497 | set_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2498 | smp_mb__after_atomic(); | |
2499 | ||
2500 | /* | |
2501 | * Try to grab a reference to the queue and wait for any outstanding | |
2502 | * requests. If we could not grab a reference the queue has been | |
2503 | * frozen and there are no requests. | |
2504 | */ | |
2505 | if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { | |
2506 | while (blk_mq_hctx_has_requests(hctx)) | |
2507 | msleep(5); | |
2508 | percpu_ref_put(&hctx->queue->q_usage_counter); | |
2509 | } | |
2510 | ||
2511 | return 0; | |
2512 | } | |
2513 | ||
2514 | static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) | |
2515 | { | |
2516 | struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, | |
2517 | struct blk_mq_hw_ctx, cpuhp_online); | |
2518 | ||
2519 | if (cpumask_test_cpu(cpu, hctx->cpumask)) | |
2520 | clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); | |
2521 | return 0; | |
2522 | } | |
2523 | ||
2524 | /* | |
2525 | * 'cpu' is going away. splice any existing rq_list entries from this | |
2526 | * software queue to the hw queue dispatch list, and ensure that it | |
2527 | * gets run. | |
2528 | */ | |
2529 | static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) | |
2530 | { | |
2531 | struct blk_mq_hw_ctx *hctx; | |
2532 | struct blk_mq_ctx *ctx; | |
2533 | LIST_HEAD(tmp); | |
2534 | enum hctx_type type; | |
2535 | ||
2536 | hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); | |
2537 | if (!cpumask_test_cpu(cpu, hctx->cpumask)) | |
2538 | return 0; | |
2539 | ||
2540 | ctx = __blk_mq_get_ctx(hctx->queue, cpu); | |
2541 | type = hctx->type; | |
2542 | ||
2543 | spin_lock(&ctx->lock); | |
2544 | if (!list_empty(&ctx->rq_lists[type])) { | |
2545 | list_splice_init(&ctx->rq_lists[type], &tmp); | |
2546 | blk_mq_hctx_clear_pending(hctx, ctx); | |
2547 | } | |
2548 | spin_unlock(&ctx->lock); | |
2549 | ||
2550 | if (list_empty(&tmp)) | |
2551 | return 0; | |
2552 | ||
2553 | spin_lock(&hctx->lock); | |
2554 | list_splice_tail_init(&tmp, &hctx->dispatch); | |
2555 | spin_unlock(&hctx->lock); | |
2556 | ||
2557 | blk_mq_run_hw_queue(hctx, true); | |
2558 | return 0; | |
2559 | } | |
2560 | ||
2561 | static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) | |
2562 | { | |
2563 | if (!(hctx->flags & BLK_MQ_F_STACKING)) | |
2564 | cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2565 | &hctx->cpuhp_online); | |
2566 | cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, | |
2567 | &hctx->cpuhp_dead); | |
2568 | } | |
2569 | ||
2570 | /* hctx->ctxs will be freed in queue's release handler */ | |
2571 | static void blk_mq_exit_hctx(struct request_queue *q, | |
2572 | struct blk_mq_tag_set *set, | |
2573 | struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) | |
2574 | { | |
2575 | if (blk_mq_hw_queue_mapped(hctx)) | |
2576 | blk_mq_tag_idle(hctx); | |
2577 | ||
2578 | if (set->ops->exit_request) | |
2579 | set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); | |
2580 | ||
2581 | if (set->ops->exit_hctx) | |
2582 | set->ops->exit_hctx(hctx, hctx_idx); | |
2583 | ||
2584 | blk_mq_remove_cpuhp(hctx); | |
2585 | ||
2586 | spin_lock(&q->unused_hctx_lock); | |
2587 | list_add(&hctx->hctx_list, &q->unused_hctx_list); | |
2588 | spin_unlock(&q->unused_hctx_lock); | |
2589 | } | |
2590 | ||
2591 | static void blk_mq_exit_hw_queues(struct request_queue *q, | |
2592 | struct blk_mq_tag_set *set, int nr_queue) | |
2593 | { | |
2594 | struct blk_mq_hw_ctx *hctx; | |
2595 | unsigned int i; | |
2596 | ||
2597 | queue_for_each_hw_ctx(q, hctx, i) { | |
2598 | if (i == nr_queue) | |
2599 | break; | |
2600 | blk_mq_debugfs_unregister_hctx(hctx); | |
2601 | blk_mq_exit_hctx(q, set, hctx, i); | |
2602 | } | |
2603 | } | |
2604 | ||
2605 | static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set) | |
2606 | { | |
2607 | int hw_ctx_size = sizeof(struct blk_mq_hw_ctx); | |
2608 | ||
2609 | BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu), | |
2610 | __alignof__(struct blk_mq_hw_ctx)) != | |
2611 | sizeof(struct blk_mq_hw_ctx)); | |
2612 | ||
2613 | if (tag_set->flags & BLK_MQ_F_BLOCKING) | |
2614 | hw_ctx_size += sizeof(struct srcu_struct); | |
2615 | ||
2616 | return hw_ctx_size; | |
2617 | } | |
2618 | ||
2619 | static int blk_mq_init_hctx(struct request_queue *q, | |
2620 | struct blk_mq_tag_set *set, | |
2621 | struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) | |
2622 | { | |
2623 | hctx->queue_num = hctx_idx; | |
2624 | ||
2625 | if (!(hctx->flags & BLK_MQ_F_STACKING)) | |
2626 | cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, | |
2627 | &hctx->cpuhp_online); | |
2628 | cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); | |
2629 | ||
2630 | hctx->tags = set->tags[hctx_idx]; | |
2631 | ||
2632 | if (set->ops->init_hctx && | |
2633 | set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) | |
2634 | goto unregister_cpu_notifier; | |
2635 | ||
2636 | if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, | |
2637 | hctx->numa_node)) | |
2638 | goto exit_hctx; | |
2639 | return 0; | |
2640 | ||
2641 | exit_hctx: | |
2642 | if (set->ops->exit_hctx) | |
2643 | set->ops->exit_hctx(hctx, hctx_idx); | |
2644 | unregister_cpu_notifier: | |
2645 | blk_mq_remove_cpuhp(hctx); | |
2646 | return -1; | |
2647 | } | |
2648 | ||
2649 | static struct blk_mq_hw_ctx * | |
2650 | blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, | |
2651 | int node) | |
2652 | { | |
2653 | struct blk_mq_hw_ctx *hctx; | |
2654 | gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; | |
2655 | ||
2656 | hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node); | |
2657 | if (!hctx) | |
2658 | goto fail_alloc_hctx; | |
2659 | ||
2660 | if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) | |
2661 | goto free_hctx; | |
2662 | ||
2663 | atomic_set(&hctx->nr_active, 0); | |
2664 | atomic_set(&hctx->elevator_queued, 0); | |
2665 | if (node == NUMA_NO_NODE) | |
2666 | node = set->numa_node; | |
2667 | hctx->numa_node = node; | |
2668 | ||
2669 | INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); | |
2670 | spin_lock_init(&hctx->lock); | |
2671 | INIT_LIST_HEAD(&hctx->dispatch); | |
2672 | hctx->queue = q; | |
2673 | hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; | |
2674 | ||
2675 | INIT_LIST_HEAD(&hctx->hctx_list); | |
2676 | ||
2677 | /* | |
2678 | * Allocate space for all possible cpus to avoid allocation at | |
2679 | * runtime | |
2680 | */ | |
2681 | hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), | |
2682 | gfp, node); | |
2683 | if (!hctx->ctxs) | |
2684 | goto free_cpumask; | |
2685 | ||
2686 | if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), | |
2687 | gfp, node)) | |
2688 | goto free_ctxs; | |
2689 | hctx->nr_ctx = 0; | |
2690 | ||
2691 | spin_lock_init(&hctx->dispatch_wait_lock); | |
2692 | init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); | |
2693 | INIT_LIST_HEAD(&hctx->dispatch_wait.entry); | |
2694 | ||
2695 | hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); | |
2696 | if (!hctx->fq) | |
2697 | goto free_bitmap; | |
2698 | ||
2699 | if (hctx->flags & BLK_MQ_F_BLOCKING) | |
2700 | init_srcu_struct(hctx->srcu); | |
2701 | blk_mq_hctx_kobj_init(hctx); | |
2702 | ||
2703 | return hctx; | |
2704 | ||
2705 | free_bitmap: | |
2706 | sbitmap_free(&hctx->ctx_map); | |
2707 | free_ctxs: | |
2708 | kfree(hctx->ctxs); | |
2709 | free_cpumask: | |
2710 | free_cpumask_var(hctx->cpumask); | |
2711 | free_hctx: | |
2712 | kfree(hctx); | |
2713 | fail_alloc_hctx: | |
2714 | return NULL; | |
2715 | } | |
2716 | ||
2717 | static void blk_mq_init_cpu_queues(struct request_queue *q, | |
2718 | unsigned int nr_hw_queues) | |
2719 | { | |
2720 | struct blk_mq_tag_set *set = q->tag_set; | |
2721 | unsigned int i, j; | |
2722 | ||
2723 | for_each_possible_cpu(i) { | |
2724 | struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); | |
2725 | struct blk_mq_hw_ctx *hctx; | |
2726 | int k; | |
2727 | ||
2728 | __ctx->cpu = i; | |
2729 | spin_lock_init(&__ctx->lock); | |
2730 | for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) | |
2731 | INIT_LIST_HEAD(&__ctx->rq_lists[k]); | |
2732 | ||
2733 | __ctx->queue = q; | |
2734 | ||
2735 | /* | |
2736 | * Set local node, IFF we have more than one hw queue. If | |
2737 | * not, we remain on the home node of the device | |
2738 | */ | |
2739 | for (j = 0; j < set->nr_maps; j++) { | |
2740 | hctx = blk_mq_map_queue_type(q, j, i); | |
2741 | if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) | |
2742 | hctx->numa_node = local_memory_node(cpu_to_node(i)); | |
2743 | } | |
2744 | } | |
2745 | } | |
2746 | ||
2747 | static bool __blk_mq_alloc_map_and_request(struct blk_mq_tag_set *set, | |
2748 | int hctx_idx) | |
2749 | { | |
2750 | unsigned int flags = set->flags; | |
2751 | int ret = 0; | |
2752 | ||
2753 | set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx, | |
2754 | set->queue_depth, set->reserved_tags, flags); | |
2755 | if (!set->tags[hctx_idx]) | |
2756 | return false; | |
2757 | ||
2758 | ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx, | |
2759 | set->queue_depth); | |
2760 | if (!ret) | |
2761 | return true; | |
2762 | ||
2763 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); | |
2764 | set->tags[hctx_idx] = NULL; | |
2765 | return false; | |
2766 | } | |
2767 | ||
2768 | static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set, | |
2769 | unsigned int hctx_idx) | |
2770 | { | |
2771 | unsigned int flags = set->flags; | |
2772 | ||
2773 | if (set->tags && set->tags[hctx_idx]) { | |
2774 | blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx); | |
2775 | blk_mq_free_rq_map(set->tags[hctx_idx], flags); | |
2776 | set->tags[hctx_idx] = NULL; | |
2777 | } | |
2778 | } | |
2779 | ||
2780 | static void blk_mq_map_swqueue(struct request_queue *q) | |
2781 | { | |
2782 | unsigned int i, j, hctx_idx; | |
2783 | struct blk_mq_hw_ctx *hctx; | |
2784 | struct blk_mq_ctx *ctx; | |
2785 | struct blk_mq_tag_set *set = q->tag_set; | |
2786 | ||
2787 | queue_for_each_hw_ctx(q, hctx, i) { | |
2788 | cpumask_clear(hctx->cpumask); | |
2789 | hctx->nr_ctx = 0; | |
2790 | hctx->dispatch_from = NULL; | |
2791 | } | |
2792 | ||
2793 | /* | |
2794 | * Map software to hardware queues. | |
2795 | * | |
2796 | * If the cpu isn't present, the cpu is mapped to first hctx. | |
2797 | */ | |
2798 | for_each_possible_cpu(i) { | |
2799 | ||
2800 | ctx = per_cpu_ptr(q->queue_ctx, i); | |
2801 | for (j = 0; j < set->nr_maps; j++) { | |
2802 | if (!set->map[j].nr_queues) { | |
2803 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2804 | HCTX_TYPE_DEFAULT, i); | |
2805 | continue; | |
2806 | } | |
2807 | hctx_idx = set->map[j].mq_map[i]; | |
2808 | /* unmapped hw queue can be remapped after CPU topo changed */ | |
2809 | if (!set->tags[hctx_idx] && | |
2810 | !__blk_mq_alloc_map_and_request(set, hctx_idx)) { | |
2811 | /* | |
2812 | * If tags initialization fail for some hctx, | |
2813 | * that hctx won't be brought online. In this | |
2814 | * case, remap the current ctx to hctx[0] which | |
2815 | * is guaranteed to always have tags allocated | |
2816 | */ | |
2817 | set->map[j].mq_map[i] = 0; | |
2818 | } | |
2819 | ||
2820 | hctx = blk_mq_map_queue_type(q, j, i); | |
2821 | ctx->hctxs[j] = hctx; | |
2822 | /* | |
2823 | * If the CPU is already set in the mask, then we've | |
2824 | * mapped this one already. This can happen if | |
2825 | * devices share queues across queue maps. | |
2826 | */ | |
2827 | if (cpumask_test_cpu(i, hctx->cpumask)) | |
2828 | continue; | |
2829 | ||
2830 | cpumask_set_cpu(i, hctx->cpumask); | |
2831 | hctx->type = j; | |
2832 | ctx->index_hw[hctx->type] = hctx->nr_ctx; | |
2833 | hctx->ctxs[hctx->nr_ctx++] = ctx; | |
2834 | ||
2835 | /* | |
2836 | * If the nr_ctx type overflows, we have exceeded the | |
2837 | * amount of sw queues we can support. | |
2838 | */ | |
2839 | BUG_ON(!hctx->nr_ctx); | |
2840 | } | |
2841 | ||
2842 | for (; j < HCTX_MAX_TYPES; j++) | |
2843 | ctx->hctxs[j] = blk_mq_map_queue_type(q, | |
2844 | HCTX_TYPE_DEFAULT, i); | |
2845 | } | |
2846 | ||
2847 | queue_for_each_hw_ctx(q, hctx, i) { | |
2848 | /* | |
2849 | * If no software queues are mapped to this hardware queue, | |
2850 | * disable it and free the request entries. | |
2851 | */ | |
2852 | if (!hctx->nr_ctx) { | |
2853 | /* Never unmap queue 0. We need it as a | |
2854 | * fallback in case of a new remap fails | |
2855 | * allocation | |
2856 | */ | |
2857 | if (i && set->tags[i]) | |
2858 | blk_mq_free_map_and_requests(set, i); | |
2859 | ||
2860 | hctx->tags = NULL; | |
2861 | continue; | |
2862 | } | |
2863 | ||
2864 | hctx->tags = set->tags[i]; | |
2865 | WARN_ON(!hctx->tags); | |
2866 | ||
2867 | /* | |
2868 | * Set the map size to the number of mapped software queues. | |
2869 | * This is more accurate and more efficient than looping | |
2870 | * over all possibly mapped software queues. | |
2871 | */ | |
2872 | sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); | |
2873 | ||
2874 | /* | |
2875 | * Initialize batch roundrobin counts | |
2876 | */ | |
2877 | hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); | |
2878 | hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; | |
2879 | } | |
2880 | } | |
2881 | ||
2882 | /* | |
2883 | * Caller needs to ensure that we're either frozen/quiesced, or that | |
2884 | * the queue isn't live yet. | |
2885 | */ | |
2886 | static void queue_set_hctx_shared(struct request_queue *q, bool shared) | |
2887 | { | |
2888 | struct blk_mq_hw_ctx *hctx; | |
2889 | int i; | |
2890 | ||
2891 | queue_for_each_hw_ctx(q, hctx, i) { | |
2892 | if (shared) | |
2893 | hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2894 | else | |
2895 | hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; | |
2896 | } | |
2897 | } | |
2898 | ||
2899 | static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, | |
2900 | bool shared) | |
2901 | { | |
2902 | struct request_queue *q; | |
2903 | ||
2904 | lockdep_assert_held(&set->tag_list_lock); | |
2905 | ||
2906 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
2907 | blk_mq_freeze_queue(q); | |
2908 | queue_set_hctx_shared(q, shared); | |
2909 | blk_mq_unfreeze_queue(q); | |
2910 | } | |
2911 | } | |
2912 | ||
2913 | static void blk_mq_del_queue_tag_set(struct request_queue *q) | |
2914 | { | |
2915 | struct blk_mq_tag_set *set = q->tag_set; | |
2916 | ||
2917 | mutex_lock(&set->tag_list_lock); | |
2918 | list_del(&q->tag_set_list); | |
2919 | if (list_is_singular(&set->tag_list)) { | |
2920 | /* just transitioned to unshared */ | |
2921 | set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; | |
2922 | /* update existing queue */ | |
2923 | blk_mq_update_tag_set_shared(set, false); | |
2924 | } | |
2925 | mutex_unlock(&set->tag_list_lock); | |
2926 | INIT_LIST_HEAD(&q->tag_set_list); | |
2927 | } | |
2928 | ||
2929 | static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, | |
2930 | struct request_queue *q) | |
2931 | { | |
2932 | mutex_lock(&set->tag_list_lock); | |
2933 | ||
2934 | /* | |
2935 | * Check to see if we're transitioning to shared (from 1 to 2 queues). | |
2936 | */ | |
2937 | if (!list_empty(&set->tag_list) && | |
2938 | !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { | |
2939 | set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; | |
2940 | /* update existing queue */ | |
2941 | blk_mq_update_tag_set_shared(set, true); | |
2942 | } | |
2943 | if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) | |
2944 | queue_set_hctx_shared(q, true); | |
2945 | list_add_tail(&q->tag_set_list, &set->tag_list); | |
2946 | ||
2947 | mutex_unlock(&set->tag_list_lock); | |
2948 | } | |
2949 | ||
2950 | /* All allocations will be freed in release handler of q->mq_kobj */ | |
2951 | static int blk_mq_alloc_ctxs(struct request_queue *q) | |
2952 | { | |
2953 | struct blk_mq_ctxs *ctxs; | |
2954 | int cpu; | |
2955 | ||
2956 | ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); | |
2957 | if (!ctxs) | |
2958 | return -ENOMEM; | |
2959 | ||
2960 | ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); | |
2961 | if (!ctxs->queue_ctx) | |
2962 | goto fail; | |
2963 | ||
2964 | for_each_possible_cpu(cpu) { | |
2965 | struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); | |
2966 | ctx->ctxs = ctxs; | |
2967 | } | |
2968 | ||
2969 | q->mq_kobj = &ctxs->kobj; | |
2970 | q->queue_ctx = ctxs->queue_ctx; | |
2971 | ||
2972 | return 0; | |
2973 | fail: | |
2974 | kfree(ctxs); | |
2975 | return -ENOMEM; | |
2976 | } | |
2977 | ||
2978 | /* | |
2979 | * It is the actual release handler for mq, but we do it from | |
2980 | * request queue's release handler for avoiding use-after-free | |
2981 | * and headache because q->mq_kobj shouldn't have been introduced, | |
2982 | * but we can't group ctx/kctx kobj without it. | |
2983 | */ | |
2984 | void blk_mq_release(struct request_queue *q) | |
2985 | { | |
2986 | struct blk_mq_hw_ctx *hctx, *next; | |
2987 | int i; | |
2988 | ||
2989 | queue_for_each_hw_ctx(q, hctx, i) | |
2990 | WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); | |
2991 | ||
2992 | /* all hctx are in .unused_hctx_list now */ | |
2993 | list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { | |
2994 | list_del_init(&hctx->hctx_list); | |
2995 | kobject_put(&hctx->kobj); | |
2996 | } | |
2997 | ||
2998 | kfree(q->queue_hw_ctx); | |
2999 | ||
3000 | /* | |
3001 | * release .mq_kobj and sw queue's kobject now because | |
3002 | * both share lifetime with request queue. | |
3003 | */ | |
3004 | blk_mq_sysfs_deinit(q); | |
3005 | } | |
3006 | ||
3007 | struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, | |
3008 | void *queuedata) | |
3009 | { | |
3010 | struct request_queue *uninit_q, *q; | |
3011 | ||
3012 | uninit_q = blk_alloc_queue(set->numa_node); | |
3013 | if (!uninit_q) | |
3014 | return ERR_PTR(-ENOMEM); | |
3015 | uninit_q->queuedata = queuedata; | |
3016 | ||
3017 | /* | |
3018 | * Initialize the queue without an elevator. device_add_disk() will do | |
3019 | * the initialization. | |
3020 | */ | |
3021 | q = blk_mq_init_allocated_queue(set, uninit_q, false); | |
3022 | if (IS_ERR(q)) | |
3023 | blk_cleanup_queue(uninit_q); | |
3024 | ||
3025 | return q; | |
3026 | } | |
3027 | EXPORT_SYMBOL_GPL(blk_mq_init_queue_data); | |
3028 | ||
3029 | struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) | |
3030 | { | |
3031 | return blk_mq_init_queue_data(set, NULL); | |
3032 | } | |
3033 | EXPORT_SYMBOL(blk_mq_init_queue); | |
3034 | ||
3035 | /* | |
3036 | * Helper for setting up a queue with mq ops, given queue depth, and | |
3037 | * the passed in mq ops flags. | |
3038 | */ | |
3039 | struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set, | |
3040 | const struct blk_mq_ops *ops, | |
3041 | unsigned int queue_depth, | |
3042 | unsigned int set_flags) | |
3043 | { | |
3044 | struct request_queue *q; | |
3045 | int ret; | |
3046 | ||
3047 | memset(set, 0, sizeof(*set)); | |
3048 | set->ops = ops; | |
3049 | set->nr_hw_queues = 1; | |
3050 | set->nr_maps = 1; | |
3051 | set->queue_depth = queue_depth; | |
3052 | set->numa_node = NUMA_NO_NODE; | |
3053 | set->flags = set_flags; | |
3054 | ||
3055 | ret = blk_mq_alloc_tag_set(set); | |
3056 | if (ret) | |
3057 | return ERR_PTR(ret); | |
3058 | ||
3059 | q = blk_mq_init_queue(set); | |
3060 | if (IS_ERR(q)) { | |
3061 | blk_mq_free_tag_set(set); | |
3062 | return q; | |
3063 | } | |
3064 | ||
3065 | return q; | |
3066 | } | |
3067 | EXPORT_SYMBOL(blk_mq_init_sq_queue); | |
3068 | ||
3069 | static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( | |
3070 | struct blk_mq_tag_set *set, struct request_queue *q, | |
3071 | int hctx_idx, int node) | |
3072 | { | |
3073 | struct blk_mq_hw_ctx *hctx = NULL, *tmp; | |
3074 | ||
3075 | /* reuse dead hctx first */ | |
3076 | spin_lock(&q->unused_hctx_lock); | |
3077 | list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { | |
3078 | if (tmp->numa_node == node) { | |
3079 | hctx = tmp; | |
3080 | break; | |
3081 | } | |
3082 | } | |
3083 | if (hctx) | |
3084 | list_del_init(&hctx->hctx_list); | |
3085 | spin_unlock(&q->unused_hctx_lock); | |
3086 | ||
3087 | if (!hctx) | |
3088 | hctx = blk_mq_alloc_hctx(q, set, node); | |
3089 | if (!hctx) | |
3090 | goto fail; | |
3091 | ||
3092 | if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) | |
3093 | goto free_hctx; | |
3094 | ||
3095 | return hctx; | |
3096 | ||
3097 | free_hctx: | |
3098 | kobject_put(&hctx->kobj); | |
3099 | fail: | |
3100 | return NULL; | |
3101 | } | |
3102 | ||
3103 | static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, | |
3104 | struct request_queue *q) | |
3105 | { | |
3106 | int i, j, end; | |
3107 | struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; | |
3108 | ||
3109 | if (q->nr_hw_queues < set->nr_hw_queues) { | |
3110 | struct blk_mq_hw_ctx **new_hctxs; | |
3111 | ||
3112 | new_hctxs = kcalloc_node(set->nr_hw_queues, | |
3113 | sizeof(*new_hctxs), GFP_KERNEL, | |
3114 | set->numa_node); | |
3115 | if (!new_hctxs) | |
3116 | return; | |
3117 | if (hctxs) | |
3118 | memcpy(new_hctxs, hctxs, q->nr_hw_queues * | |
3119 | sizeof(*hctxs)); | |
3120 | q->queue_hw_ctx = new_hctxs; | |
3121 | kfree(hctxs); | |
3122 | hctxs = new_hctxs; | |
3123 | } | |
3124 | ||
3125 | /* protect against switching io scheduler */ | |
3126 | mutex_lock(&q->sysfs_lock); | |
3127 | for (i = 0; i < set->nr_hw_queues; i++) { | |
3128 | int node; | |
3129 | struct blk_mq_hw_ctx *hctx; | |
3130 | ||
3131 | node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i); | |
3132 | /* | |
3133 | * If the hw queue has been mapped to another numa node, | |
3134 | * we need to realloc the hctx. If allocation fails, fallback | |
3135 | * to use the previous one. | |
3136 | */ | |
3137 | if (hctxs[i] && (hctxs[i]->numa_node == node)) | |
3138 | continue; | |
3139 | ||
3140 | hctx = blk_mq_alloc_and_init_hctx(set, q, i, node); | |
3141 | if (hctx) { | |
3142 | if (hctxs[i]) | |
3143 | blk_mq_exit_hctx(q, set, hctxs[i], i); | |
3144 | hctxs[i] = hctx; | |
3145 | } else { | |
3146 | if (hctxs[i]) | |
3147 | pr_warn("Allocate new hctx on node %d fails,\ | |
3148 | fallback to previous one on node %d\n", | |
3149 | node, hctxs[i]->numa_node); | |
3150 | else | |
3151 | break; | |
3152 | } | |
3153 | } | |
3154 | /* | |
3155 | * Increasing nr_hw_queues fails. Free the newly allocated | |
3156 | * hctxs and keep the previous q->nr_hw_queues. | |
3157 | */ | |
3158 | if (i != set->nr_hw_queues) { | |
3159 | j = q->nr_hw_queues; | |
3160 | end = i; | |
3161 | } else { | |
3162 | j = i; | |
3163 | end = q->nr_hw_queues; | |
3164 | q->nr_hw_queues = set->nr_hw_queues; | |
3165 | } | |
3166 | ||
3167 | for (; j < end; j++) { | |
3168 | struct blk_mq_hw_ctx *hctx = hctxs[j]; | |
3169 | ||
3170 | if (hctx) { | |
3171 | if (hctx->tags) | |
3172 | blk_mq_free_map_and_requests(set, j); | |
3173 | blk_mq_exit_hctx(q, set, hctx, j); | |
3174 | hctxs[j] = NULL; | |
3175 | } | |
3176 | } | |
3177 | mutex_unlock(&q->sysfs_lock); | |
3178 | } | |
3179 | ||
3180 | struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, | |
3181 | struct request_queue *q, | |
3182 | bool elevator_init) | |
3183 | { | |
3184 | /* mark the queue as mq asap */ | |
3185 | q->mq_ops = set->ops; | |
3186 | ||
3187 | q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn, | |
3188 | blk_mq_poll_stats_bkt, | |
3189 | BLK_MQ_POLL_STATS_BKTS, q); | |
3190 | if (!q->poll_cb) | |
3191 | goto err_exit; | |
3192 | ||
3193 | if (blk_mq_alloc_ctxs(q)) | |
3194 | goto err_poll; | |
3195 | ||
3196 | /* init q->mq_kobj and sw queues' kobjects */ | |
3197 | blk_mq_sysfs_init(q); | |
3198 | ||
3199 | INIT_LIST_HEAD(&q->unused_hctx_list); | |
3200 | spin_lock_init(&q->unused_hctx_lock); | |
3201 | ||
3202 | blk_mq_realloc_hw_ctxs(set, q); | |
3203 | if (!q->nr_hw_queues) | |
3204 | goto err_hctxs; | |
3205 | ||
3206 | INIT_WORK(&q->timeout_work, blk_mq_timeout_work); | |
3207 | blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); | |
3208 | ||
3209 | q->tag_set = set; | |
3210 | ||
3211 | q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; | |
3212 | if (set->nr_maps > HCTX_TYPE_POLL && | |
3213 | set->map[HCTX_TYPE_POLL].nr_queues) | |
3214 | blk_queue_flag_set(QUEUE_FLAG_POLL, q); | |
3215 | ||
3216 | q->sg_reserved_size = INT_MAX; | |
3217 | ||
3218 | INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); | |
3219 | INIT_LIST_HEAD(&q->requeue_list); | |
3220 | spin_lock_init(&q->requeue_lock); | |
3221 | ||
3222 | q->nr_requests = set->queue_depth; | |
3223 | ||
3224 | /* | |
3225 | * Default to classic polling | |
3226 | */ | |
3227 | q->poll_nsec = BLK_MQ_POLL_CLASSIC; | |
3228 | ||
3229 | blk_mq_init_cpu_queues(q, set->nr_hw_queues); | |
3230 | blk_mq_add_queue_tag_set(set, q); | |
3231 | blk_mq_map_swqueue(q); | |
3232 | ||
3233 | if (elevator_init) | |
3234 | elevator_init_mq(q); | |
3235 | ||
3236 | return q; | |
3237 | ||
3238 | err_hctxs: | |
3239 | kfree(q->queue_hw_ctx); | |
3240 | q->nr_hw_queues = 0; | |
3241 | blk_mq_sysfs_deinit(q); | |
3242 | err_poll: | |
3243 | blk_stat_free_callback(q->poll_cb); | |
3244 | q->poll_cb = NULL; | |
3245 | err_exit: | |
3246 | q->mq_ops = NULL; | |
3247 | return ERR_PTR(-ENOMEM); | |
3248 | } | |
3249 | EXPORT_SYMBOL(blk_mq_init_allocated_queue); | |
3250 | ||
3251 | /* tags can _not_ be used after returning from blk_mq_exit_queue */ | |
3252 | void blk_mq_exit_queue(struct request_queue *q) | |
3253 | { | |
3254 | struct blk_mq_tag_set *set = q->tag_set; | |
3255 | ||
3256 | blk_mq_del_queue_tag_set(q); | |
3257 | blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); | |
3258 | } | |
3259 | ||
3260 | static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) | |
3261 | { | |
3262 | int i; | |
3263 | ||
3264 | for (i = 0; i < set->nr_hw_queues; i++) | |
3265 | if (!__blk_mq_alloc_map_and_request(set, i)) | |
3266 | goto out_unwind; | |
3267 | ||
3268 | return 0; | |
3269 | ||
3270 | out_unwind: | |
3271 | while (--i >= 0) | |
3272 | blk_mq_free_map_and_requests(set, i); | |
3273 | ||
3274 | return -ENOMEM; | |
3275 | } | |
3276 | ||
3277 | /* | |
3278 | * Allocate the request maps associated with this tag_set. Note that this | |
3279 | * may reduce the depth asked for, if memory is tight. set->queue_depth | |
3280 | * will be updated to reflect the allocated depth. | |
3281 | */ | |
3282 | static int blk_mq_alloc_map_and_requests(struct blk_mq_tag_set *set) | |
3283 | { | |
3284 | unsigned int depth; | |
3285 | int err; | |
3286 | ||
3287 | depth = set->queue_depth; | |
3288 | do { | |
3289 | err = __blk_mq_alloc_rq_maps(set); | |
3290 | if (!err) | |
3291 | break; | |
3292 | ||
3293 | set->queue_depth >>= 1; | |
3294 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { | |
3295 | err = -ENOMEM; | |
3296 | break; | |
3297 | } | |
3298 | } while (set->queue_depth); | |
3299 | ||
3300 | if (!set->queue_depth || err) { | |
3301 | pr_err("blk-mq: failed to allocate request map\n"); | |
3302 | return -ENOMEM; | |
3303 | } | |
3304 | ||
3305 | if (depth != set->queue_depth) | |
3306 | pr_info("blk-mq: reduced tag depth (%u -> %u)\n", | |
3307 | depth, set->queue_depth); | |
3308 | ||
3309 | return 0; | |
3310 | } | |
3311 | ||
3312 | static int blk_mq_update_queue_map(struct blk_mq_tag_set *set) | |
3313 | { | |
3314 | /* | |
3315 | * blk_mq_map_queues() and multiple .map_queues() implementations | |
3316 | * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the | |
3317 | * number of hardware queues. | |
3318 | */ | |
3319 | if (set->nr_maps == 1) | |
3320 | set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; | |
3321 | ||
3322 | if (set->ops->map_queues && !is_kdump_kernel()) { | |
3323 | int i; | |
3324 | ||
3325 | /* | |
3326 | * transport .map_queues is usually done in the following | |
3327 | * way: | |
3328 | * | |
3329 | * for (queue = 0; queue < set->nr_hw_queues; queue++) { | |
3330 | * mask = get_cpu_mask(queue) | |
3331 | * for_each_cpu(cpu, mask) | |
3332 | * set->map[x].mq_map[cpu] = queue; | |
3333 | * } | |
3334 | * | |
3335 | * When we need to remap, the table has to be cleared for | |
3336 | * killing stale mapping since one CPU may not be mapped | |
3337 | * to any hw queue. | |
3338 | */ | |
3339 | for (i = 0; i < set->nr_maps; i++) | |
3340 | blk_mq_clear_mq_map(&set->map[i]); | |
3341 | ||
3342 | return set->ops->map_queues(set); | |
3343 | } else { | |
3344 | BUG_ON(set->nr_maps > 1); | |
3345 | return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); | |
3346 | } | |
3347 | } | |
3348 | ||
3349 | static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, | |
3350 | int cur_nr_hw_queues, int new_nr_hw_queues) | |
3351 | { | |
3352 | struct blk_mq_tags **new_tags; | |
3353 | ||
3354 | if (cur_nr_hw_queues >= new_nr_hw_queues) | |
3355 | return 0; | |
3356 | ||
3357 | new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), | |
3358 | GFP_KERNEL, set->numa_node); | |
3359 | if (!new_tags) | |
3360 | return -ENOMEM; | |
3361 | ||
3362 | if (set->tags) | |
3363 | memcpy(new_tags, set->tags, cur_nr_hw_queues * | |
3364 | sizeof(*set->tags)); | |
3365 | kfree(set->tags); | |
3366 | set->tags = new_tags; | |
3367 | set->nr_hw_queues = new_nr_hw_queues; | |
3368 | ||
3369 | return 0; | |
3370 | } | |
3371 | ||
3372 | /* | |
3373 | * Alloc a tag set to be associated with one or more request queues. | |
3374 | * May fail with EINVAL for various error conditions. May adjust the | |
3375 | * requested depth down, if it's too large. In that case, the set | |
3376 | * value will be stored in set->queue_depth. | |
3377 | */ | |
3378 | int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) | |
3379 | { | |
3380 | int i, ret; | |
3381 | ||
3382 | BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); | |
3383 | ||
3384 | if (!set->nr_hw_queues) | |
3385 | return -EINVAL; | |
3386 | if (!set->queue_depth) | |
3387 | return -EINVAL; | |
3388 | if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) | |
3389 | return -EINVAL; | |
3390 | ||
3391 | if (!set->ops->queue_rq) | |
3392 | return -EINVAL; | |
3393 | ||
3394 | if (!set->ops->get_budget ^ !set->ops->put_budget) | |
3395 | return -EINVAL; | |
3396 | ||
3397 | if (set->queue_depth > BLK_MQ_MAX_DEPTH) { | |
3398 | pr_info("blk-mq: reduced tag depth to %u\n", | |
3399 | BLK_MQ_MAX_DEPTH); | |
3400 | set->queue_depth = BLK_MQ_MAX_DEPTH; | |
3401 | } | |
3402 | ||
3403 | if (!set->nr_maps) | |
3404 | set->nr_maps = 1; | |
3405 | else if (set->nr_maps > HCTX_MAX_TYPES) | |
3406 | return -EINVAL; | |
3407 | ||
3408 | /* | |
3409 | * If a crashdump is active, then we are potentially in a very | |
3410 | * memory constrained environment. Limit us to 1 queue and | |
3411 | * 64 tags to prevent using too much memory. | |
3412 | */ | |
3413 | if (is_kdump_kernel()) { | |
3414 | set->nr_hw_queues = 1; | |
3415 | set->nr_maps = 1; | |
3416 | set->queue_depth = min(64U, set->queue_depth); | |
3417 | } | |
3418 | /* | |
3419 | * There is no use for more h/w queues than cpus if we just have | |
3420 | * a single map | |
3421 | */ | |
3422 | if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) | |
3423 | set->nr_hw_queues = nr_cpu_ids; | |
3424 | ||
3425 | if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0) | |
3426 | return -ENOMEM; | |
3427 | ||
3428 | ret = -ENOMEM; | |
3429 | for (i = 0; i < set->nr_maps; i++) { | |
3430 | set->map[i].mq_map = kcalloc_node(nr_cpu_ids, | |
3431 | sizeof(set->map[i].mq_map[0]), | |
3432 | GFP_KERNEL, set->numa_node); | |
3433 | if (!set->map[i].mq_map) | |
3434 | goto out_free_mq_map; | |
3435 | set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; | |
3436 | } | |
3437 | ||
3438 | ret = blk_mq_update_queue_map(set); | |
3439 | if (ret) | |
3440 | goto out_free_mq_map; | |
3441 | ||
3442 | ret = blk_mq_alloc_map_and_requests(set); | |
3443 | if (ret) | |
3444 | goto out_free_mq_map; | |
3445 | ||
3446 | if (blk_mq_is_sbitmap_shared(set->flags)) { | |
3447 | atomic_set(&set->active_queues_shared_sbitmap, 0); | |
3448 | ||
3449 | if (blk_mq_init_shared_sbitmap(set, set->flags)) { | |
3450 | ret = -ENOMEM; | |
3451 | goto out_free_mq_rq_maps; | |
3452 | } | |
3453 | } | |
3454 | ||
3455 | mutex_init(&set->tag_list_lock); | |
3456 | INIT_LIST_HEAD(&set->tag_list); | |
3457 | ||
3458 | return 0; | |
3459 | ||
3460 | out_free_mq_rq_maps: | |
3461 | for (i = 0; i < set->nr_hw_queues; i++) | |
3462 | blk_mq_free_map_and_requests(set, i); | |
3463 | out_free_mq_map: | |
3464 | for (i = 0; i < set->nr_maps; i++) { | |
3465 | kfree(set->map[i].mq_map); | |
3466 | set->map[i].mq_map = NULL; | |
3467 | } | |
3468 | kfree(set->tags); | |
3469 | set->tags = NULL; | |
3470 | return ret; | |
3471 | } | |
3472 | EXPORT_SYMBOL(blk_mq_alloc_tag_set); | |
3473 | ||
3474 | void blk_mq_free_tag_set(struct blk_mq_tag_set *set) | |
3475 | { | |
3476 | int i, j; | |
3477 | ||
3478 | for (i = 0; i < set->nr_hw_queues; i++) | |
3479 | blk_mq_free_map_and_requests(set, i); | |
3480 | ||
3481 | if (blk_mq_is_sbitmap_shared(set->flags)) | |
3482 | blk_mq_exit_shared_sbitmap(set); | |
3483 | ||
3484 | for (j = 0; j < set->nr_maps; j++) { | |
3485 | kfree(set->map[j].mq_map); | |
3486 | set->map[j].mq_map = NULL; | |
3487 | } | |
3488 | ||
3489 | kfree(set->tags); | |
3490 | set->tags = NULL; | |
3491 | } | |
3492 | EXPORT_SYMBOL(blk_mq_free_tag_set); | |
3493 | ||
3494 | int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) | |
3495 | { | |
3496 | struct blk_mq_tag_set *set = q->tag_set; | |
3497 | struct blk_mq_hw_ctx *hctx; | |
3498 | int i, ret; | |
3499 | ||
3500 | if (!set) | |
3501 | return -EINVAL; | |
3502 | ||
3503 | if (q->nr_requests == nr) | |
3504 | return 0; | |
3505 | ||
3506 | blk_mq_freeze_queue(q); | |
3507 | blk_mq_quiesce_queue(q); | |
3508 | ||
3509 | ret = 0; | |
3510 | queue_for_each_hw_ctx(q, hctx, i) { | |
3511 | if (!hctx->tags) | |
3512 | continue; | |
3513 | /* | |
3514 | * If we're using an MQ scheduler, just update the scheduler | |
3515 | * queue depth. This is similar to what the old code would do. | |
3516 | */ | |
3517 | if (!hctx->sched_tags) { | |
3518 | ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, | |
3519 | false); | |
3520 | if (!ret && blk_mq_is_sbitmap_shared(set->flags)) | |
3521 | blk_mq_tag_resize_shared_sbitmap(set, nr); | |
3522 | } else { | |
3523 | ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, | |
3524 | nr, true); | |
3525 | } | |
3526 | if (ret) | |
3527 | break; | |
3528 | if (q->elevator && q->elevator->type->ops.depth_updated) | |
3529 | q->elevator->type->ops.depth_updated(hctx); | |
3530 | } | |
3531 | ||
3532 | if (!ret) | |
3533 | q->nr_requests = nr; | |
3534 | ||
3535 | blk_mq_unquiesce_queue(q); | |
3536 | blk_mq_unfreeze_queue(q); | |
3537 | ||
3538 | return ret; | |
3539 | } | |
3540 | ||
3541 | /* | |
3542 | * request_queue and elevator_type pair. | |
3543 | * It is just used by __blk_mq_update_nr_hw_queues to cache | |
3544 | * the elevator_type associated with a request_queue. | |
3545 | */ | |
3546 | struct blk_mq_qe_pair { | |
3547 | struct list_head node; | |
3548 | struct request_queue *q; | |
3549 | struct elevator_type *type; | |
3550 | }; | |
3551 | ||
3552 | /* | |
3553 | * Cache the elevator_type in qe pair list and switch the | |
3554 | * io scheduler to 'none' | |
3555 | */ | |
3556 | static bool blk_mq_elv_switch_none(struct list_head *head, | |
3557 | struct request_queue *q) | |
3558 | { | |
3559 | struct blk_mq_qe_pair *qe; | |
3560 | ||
3561 | if (!q->elevator) | |
3562 | return true; | |
3563 | ||
3564 | qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); | |
3565 | if (!qe) | |
3566 | return false; | |
3567 | ||
3568 | INIT_LIST_HEAD(&qe->node); | |
3569 | qe->q = q; | |
3570 | qe->type = q->elevator->type; | |
3571 | list_add(&qe->node, head); | |
3572 | ||
3573 | mutex_lock(&q->sysfs_lock); | |
3574 | /* | |
3575 | * After elevator_switch_mq, the previous elevator_queue will be | |
3576 | * released by elevator_release. The reference of the io scheduler | |
3577 | * module get by elevator_get will also be put. So we need to get | |
3578 | * a reference of the io scheduler module here to prevent it to be | |
3579 | * removed. | |
3580 | */ | |
3581 | __module_get(qe->type->elevator_owner); | |
3582 | elevator_switch_mq(q, NULL); | |
3583 | mutex_unlock(&q->sysfs_lock); | |
3584 | ||
3585 | return true; | |
3586 | } | |
3587 | ||
3588 | static void blk_mq_elv_switch_back(struct list_head *head, | |
3589 | struct request_queue *q) | |
3590 | { | |
3591 | struct blk_mq_qe_pair *qe; | |
3592 | struct elevator_type *t = NULL; | |
3593 | ||
3594 | list_for_each_entry(qe, head, node) | |
3595 | if (qe->q == q) { | |
3596 | t = qe->type; | |
3597 | break; | |
3598 | } | |
3599 | ||
3600 | if (!t) | |
3601 | return; | |
3602 | ||
3603 | list_del(&qe->node); | |
3604 | kfree(qe); | |
3605 | ||
3606 | mutex_lock(&q->sysfs_lock); | |
3607 | elevator_switch_mq(q, t); | |
3608 | mutex_unlock(&q->sysfs_lock); | |
3609 | } | |
3610 | ||
3611 | static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, | |
3612 | int nr_hw_queues) | |
3613 | { | |
3614 | struct request_queue *q; | |
3615 | LIST_HEAD(head); | |
3616 | int prev_nr_hw_queues; | |
3617 | ||
3618 | lockdep_assert_held(&set->tag_list_lock); | |
3619 | ||
3620 | if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) | |
3621 | nr_hw_queues = nr_cpu_ids; | |
3622 | if (nr_hw_queues < 1) | |
3623 | return; | |
3624 | if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) | |
3625 | return; | |
3626 | ||
3627 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3628 | blk_mq_freeze_queue(q); | |
3629 | /* | |
3630 | * Switch IO scheduler to 'none', cleaning up the data associated | |
3631 | * with the previous scheduler. We will switch back once we are done | |
3632 | * updating the new sw to hw queue mappings. | |
3633 | */ | |
3634 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3635 | if (!blk_mq_elv_switch_none(&head, q)) | |
3636 | goto switch_back; | |
3637 | ||
3638 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3639 | blk_mq_debugfs_unregister_hctxs(q); | |
3640 | blk_mq_sysfs_unregister(q); | |
3641 | } | |
3642 | ||
3643 | prev_nr_hw_queues = set->nr_hw_queues; | |
3644 | if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) < | |
3645 | 0) | |
3646 | goto reregister; | |
3647 | ||
3648 | set->nr_hw_queues = nr_hw_queues; | |
3649 | fallback: | |
3650 | blk_mq_update_queue_map(set); | |
3651 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3652 | blk_mq_realloc_hw_ctxs(set, q); | |
3653 | if (q->nr_hw_queues != set->nr_hw_queues) { | |
3654 | pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", | |
3655 | nr_hw_queues, prev_nr_hw_queues); | |
3656 | set->nr_hw_queues = prev_nr_hw_queues; | |
3657 | blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); | |
3658 | goto fallback; | |
3659 | } | |
3660 | blk_mq_map_swqueue(q); | |
3661 | } | |
3662 | ||
3663 | reregister: | |
3664 | list_for_each_entry(q, &set->tag_list, tag_set_list) { | |
3665 | blk_mq_sysfs_register(q); | |
3666 | blk_mq_debugfs_register_hctxs(q); | |
3667 | } | |
3668 | ||
3669 | switch_back: | |
3670 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3671 | blk_mq_elv_switch_back(&head, q); | |
3672 | ||
3673 | list_for_each_entry(q, &set->tag_list, tag_set_list) | |
3674 | blk_mq_unfreeze_queue(q); | |
3675 | } | |
3676 | ||
3677 | void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) | |
3678 | { | |
3679 | mutex_lock(&set->tag_list_lock); | |
3680 | __blk_mq_update_nr_hw_queues(set, nr_hw_queues); | |
3681 | mutex_unlock(&set->tag_list_lock); | |
3682 | } | |
3683 | EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); | |
3684 | ||
3685 | /* Enable polling stats and return whether they were already enabled. */ | |
3686 | static bool blk_poll_stats_enable(struct request_queue *q) | |
3687 | { | |
3688 | if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3689 | blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q)) | |
3690 | return true; | |
3691 | blk_stat_add_callback(q, q->poll_cb); | |
3692 | return false; | |
3693 | } | |
3694 | ||
3695 | static void blk_mq_poll_stats_start(struct request_queue *q) | |
3696 | { | |
3697 | /* | |
3698 | * We don't arm the callback if polling stats are not enabled or the | |
3699 | * callback is already active. | |
3700 | */ | |
3701 | if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) || | |
3702 | blk_stat_is_active(q->poll_cb)) | |
3703 | return; | |
3704 | ||
3705 | blk_stat_activate_msecs(q->poll_cb, 100); | |
3706 | } | |
3707 | ||
3708 | static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb) | |
3709 | { | |
3710 | struct request_queue *q = cb->data; | |
3711 | int bucket; | |
3712 | ||
3713 | for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) { | |
3714 | if (cb->stat[bucket].nr_samples) | |
3715 | q->poll_stat[bucket] = cb->stat[bucket]; | |
3716 | } | |
3717 | } | |
3718 | ||
3719 | static unsigned long blk_mq_poll_nsecs(struct request_queue *q, | |
3720 | struct request *rq) | |
3721 | { | |
3722 | unsigned long ret = 0; | |
3723 | int bucket; | |
3724 | ||
3725 | /* | |
3726 | * If stats collection isn't on, don't sleep but turn it on for | |
3727 | * future users | |
3728 | */ | |
3729 | if (!blk_poll_stats_enable(q)) | |
3730 | return 0; | |
3731 | ||
3732 | /* | |
3733 | * As an optimistic guess, use half of the mean service time | |
3734 | * for this type of request. We can (and should) make this smarter. | |
3735 | * For instance, if the completion latencies are tight, we can | |
3736 | * get closer than just half the mean. This is especially | |
3737 | * important on devices where the completion latencies are longer | |
3738 | * than ~10 usec. We do use the stats for the relevant IO size | |
3739 | * if available which does lead to better estimates. | |
3740 | */ | |
3741 | bucket = blk_mq_poll_stats_bkt(rq); | |
3742 | if (bucket < 0) | |
3743 | return ret; | |
3744 | ||
3745 | if (q->poll_stat[bucket].nr_samples) | |
3746 | ret = (q->poll_stat[bucket].mean + 1) / 2; | |
3747 | ||
3748 | return ret; | |
3749 | } | |
3750 | ||
3751 | static bool blk_mq_poll_hybrid_sleep(struct request_queue *q, | |
3752 | struct request *rq) | |
3753 | { | |
3754 | struct hrtimer_sleeper hs; | |
3755 | enum hrtimer_mode mode; | |
3756 | unsigned int nsecs; | |
3757 | ktime_t kt; | |
3758 | ||
3759 | if (rq->rq_flags & RQF_MQ_POLL_SLEPT) | |
3760 | return false; | |
3761 | ||
3762 | /* | |
3763 | * If we get here, hybrid polling is enabled. Hence poll_nsec can be: | |
3764 | * | |
3765 | * 0: use half of prev avg | |
3766 | * >0: use this specific value | |
3767 | */ | |
3768 | if (q->poll_nsec > 0) | |
3769 | nsecs = q->poll_nsec; | |
3770 | else | |
3771 | nsecs = blk_mq_poll_nsecs(q, rq); | |
3772 | ||
3773 | if (!nsecs) | |
3774 | return false; | |
3775 | ||
3776 | rq->rq_flags |= RQF_MQ_POLL_SLEPT; | |
3777 | ||
3778 | /* | |
3779 | * This will be replaced with the stats tracking code, using | |
3780 | * 'avg_completion_time / 2' as the pre-sleep target. | |
3781 | */ | |
3782 | kt = nsecs; | |
3783 | ||
3784 | mode = HRTIMER_MODE_REL; | |
3785 | hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode); | |
3786 | hrtimer_set_expires(&hs.timer, kt); | |
3787 | ||
3788 | do { | |
3789 | if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE) | |
3790 | break; | |
3791 | set_current_state(TASK_UNINTERRUPTIBLE); | |
3792 | hrtimer_sleeper_start_expires(&hs, mode); | |
3793 | if (hs.task) | |
3794 | io_schedule(); | |
3795 | hrtimer_cancel(&hs.timer); | |
3796 | mode = HRTIMER_MODE_ABS; | |
3797 | } while (hs.task && !signal_pending(current)); | |
3798 | ||
3799 | __set_current_state(TASK_RUNNING); | |
3800 | destroy_hrtimer_on_stack(&hs.timer); | |
3801 | return true; | |
3802 | } | |
3803 | ||
3804 | static bool blk_mq_poll_hybrid(struct request_queue *q, | |
3805 | struct blk_mq_hw_ctx *hctx, blk_qc_t cookie) | |
3806 | { | |
3807 | struct request *rq; | |
3808 | ||
3809 | if (q->poll_nsec == BLK_MQ_POLL_CLASSIC) | |
3810 | return false; | |
3811 | ||
3812 | if (!blk_qc_t_is_internal(cookie)) | |
3813 | rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie)); | |
3814 | else { | |
3815 | rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie)); | |
3816 | /* | |
3817 | * With scheduling, if the request has completed, we'll | |
3818 | * get a NULL return here, as we clear the sched tag when | |
3819 | * that happens. The request still remains valid, like always, | |
3820 | * so we should be safe with just the NULL check. | |
3821 | */ | |
3822 | if (!rq) | |
3823 | return false; | |
3824 | } | |
3825 | ||
3826 | return blk_mq_poll_hybrid_sleep(q, rq); | |
3827 | } | |
3828 | ||
3829 | /** | |
3830 | * blk_poll - poll for IO completions | |
3831 | * @q: the queue | |
3832 | * @cookie: cookie passed back at IO submission time | |
3833 | * @spin: whether to spin for completions | |
3834 | * | |
3835 | * Description: | |
3836 | * Poll for completions on the passed in queue. Returns number of | |
3837 | * completed entries found. If @spin is true, then blk_poll will continue | |
3838 | * looping until at least one completion is found, unless the task is | |
3839 | * otherwise marked running (or we need to reschedule). | |
3840 | */ | |
3841 | int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin) | |
3842 | { | |
3843 | struct blk_mq_hw_ctx *hctx; | |
3844 | long state; | |
3845 | ||
3846 | if (!blk_qc_t_valid(cookie) || | |
3847 | !test_bit(QUEUE_FLAG_POLL, &q->queue_flags)) | |
3848 | return 0; | |
3849 | ||
3850 | if (current->plug) | |
3851 | blk_flush_plug_list(current->plug, false); | |
3852 | ||
3853 | hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)]; | |
3854 | ||
3855 | /* | |
3856 | * If we sleep, have the caller restart the poll loop to reset | |
3857 | * the state. Like for the other success return cases, the | |
3858 | * caller is responsible for checking if the IO completed. If | |
3859 | * the IO isn't complete, we'll get called again and will go | |
3860 | * straight to the busy poll loop. | |
3861 | */ | |
3862 | if (blk_mq_poll_hybrid(q, hctx, cookie)) | |
3863 | return 1; | |
3864 | ||
3865 | hctx->poll_considered++; | |
3866 | ||
3867 | state = current->state; | |
3868 | do { | |
3869 | int ret; | |
3870 | ||
3871 | hctx->poll_invoked++; | |
3872 | ||
3873 | ret = q->mq_ops->poll(hctx); | |
3874 | if (ret > 0) { | |
3875 | hctx->poll_success++; | |
3876 | __set_current_state(TASK_RUNNING); | |
3877 | return ret; | |
3878 | } | |
3879 | ||
3880 | if (signal_pending_state(state, current)) | |
3881 | __set_current_state(TASK_RUNNING); | |
3882 | ||
3883 | if (current->state == TASK_RUNNING) | |
3884 | return 1; | |
3885 | if (ret < 0 || !spin) | |
3886 | break; | |
3887 | cpu_relax(); | |
3888 | } while (!need_resched()); | |
3889 | ||
3890 | __set_current_state(TASK_RUNNING); | |
3891 | return 0; | |
3892 | } | |
3893 | EXPORT_SYMBOL_GPL(blk_poll); | |
3894 | ||
3895 | unsigned int blk_mq_rq_cpu(struct request *rq) | |
3896 | { | |
3897 | return rq->mq_ctx->cpu; | |
3898 | } | |
3899 | EXPORT_SYMBOL(blk_mq_rq_cpu); | |
3900 | ||
3901 | static int __init blk_mq_init(void) | |
3902 | { | |
3903 | int i; | |
3904 | ||
3905 | for_each_possible_cpu(i) | |
3906 | INIT_LIST_HEAD(&per_cpu(blk_cpu_done, i)); | |
3907 | open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); | |
3908 | ||
3909 | cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, | |
3910 | "block/softirq:dead", NULL, | |
3911 | blk_softirq_cpu_dead); | |
3912 | cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, | |
3913 | blk_mq_hctx_notify_dead); | |
3914 | cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", | |
3915 | blk_mq_hctx_notify_online, | |
3916 | blk_mq_hctx_notify_offline); | |
3917 | return 0; | |
3918 | } | |
3919 | subsys_initcall(blk_mq_init); |