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/blk-integrity.h>
14 #include <linux/kmemleak.h>
15 #include <linux/mm.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <linux/workqueue.h>
19 #include <linux/smp.h>
20 #include <linux/interrupt.h>
21 #include <linux/llist.h>
22 #include <linux/cpu.h>
23 #include <linux/cache.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 #include <linux/part_stat.h>
31 #include <linux/sched/isolation.h>
32
33 #include <trace/events/block.h>
34
35 #include <linux/t10-pi.h>
36 #include "blk.h"
37 #include "blk-mq.h"
38 #include "blk-mq-debugfs.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 llist_head, blk_cpu_done);
45 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd);
46 static DEFINE_MUTEX(blk_mq_cpuhp_lock);
47
48 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
49 static void blk_mq_request_bypass_insert(struct request *rq,
50 blk_insert_t flags);
51 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
52 struct list_head *list);
53 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
54 struct io_comp_batch *iob, unsigned int flags);
55
56 /*
57 * Check if any of the ctx, dispatch list or elevator
58 * have pending work in this hardware queue.
59 */
blk_mq_hctx_has_pending(struct blk_mq_hw_ctx * hctx)60 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
61 {
62 return !list_empty_careful(&hctx->dispatch) ||
63 sbitmap_any_bit_set(&hctx->ctx_map) ||
64 blk_mq_sched_has_work(hctx);
65 }
66
67 /*
68 * Mark this ctx as having pending work in this hardware queue
69 */
blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx * hctx,struct blk_mq_ctx * ctx)70 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
71 struct blk_mq_ctx *ctx)
72 {
73 const int bit = ctx->index_hw[hctx->type];
74
75 if (!sbitmap_test_bit(&hctx->ctx_map, bit))
76 sbitmap_set_bit(&hctx->ctx_map, bit);
77 }
78
blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx * hctx,struct blk_mq_ctx * ctx)79 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
80 struct blk_mq_ctx *ctx)
81 {
82 const int bit = ctx->index_hw[hctx->type];
83
84 sbitmap_clear_bit(&hctx->ctx_map, bit);
85 }
86
87 struct mq_inflight {
88 struct block_device *part;
89 unsigned int inflight[2];
90 };
91
blk_mq_check_inflight(struct request * rq,void * priv)92 static bool blk_mq_check_inflight(struct request *rq, void *priv)
93 {
94 struct mq_inflight *mi = priv;
95
96 if (rq->rq_flags & RQF_IO_STAT &&
97 (!bdev_is_partition(mi->part) || rq->part == mi->part) &&
98 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
99 mi->inflight[rq_data_dir(rq)]++;
100
101 return true;
102 }
103
blk_mq_in_flight(struct request_queue * q,struct block_device * part)104 unsigned int blk_mq_in_flight(struct request_queue *q,
105 struct block_device *part)
106 {
107 struct mq_inflight mi = { .part = part };
108
109 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
110
111 return mi.inflight[0] + mi.inflight[1];
112 }
113
blk_mq_in_flight_rw(struct request_queue * q,struct block_device * part,unsigned int inflight[2])114 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
115 unsigned int inflight[2])
116 {
117 struct mq_inflight mi = { .part = part };
118
119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
120 inflight[0] = mi.inflight[0];
121 inflight[1] = mi.inflight[1];
122 }
123
124 #ifdef CONFIG_LOCKDEP
blk_freeze_set_owner(struct request_queue * q,struct task_struct * owner)125 static bool blk_freeze_set_owner(struct request_queue *q,
126 struct task_struct *owner)
127 {
128 if (!owner)
129 return false;
130
131 if (!q->mq_freeze_depth) {
132 q->mq_freeze_owner = owner;
133 q->mq_freeze_owner_depth = 1;
134 return true;
135 }
136
137 if (owner == q->mq_freeze_owner)
138 q->mq_freeze_owner_depth += 1;
139 return false;
140 }
141
142 /* verify the last unfreeze in owner context */
blk_unfreeze_check_owner(struct request_queue * q)143 static bool blk_unfreeze_check_owner(struct request_queue *q)
144 {
145 if (!q->mq_freeze_owner)
146 return false;
147 if (q->mq_freeze_owner != current)
148 return false;
149 if (--q->mq_freeze_owner_depth == 0) {
150 q->mq_freeze_owner = NULL;
151 return true;
152 }
153 return false;
154 }
155
156 #else
157
blk_freeze_set_owner(struct request_queue * q,struct task_struct * owner)158 static bool blk_freeze_set_owner(struct request_queue *q,
159 struct task_struct *owner)
160 {
161 return false;
162 }
163
blk_unfreeze_check_owner(struct request_queue * q)164 static bool blk_unfreeze_check_owner(struct request_queue *q)
165 {
166 return false;
167 }
168 #endif
169
__blk_freeze_queue_start(struct request_queue * q,struct task_struct * owner)170 bool __blk_freeze_queue_start(struct request_queue *q,
171 struct task_struct *owner)
172 {
173 bool freeze;
174
175 mutex_lock(&q->mq_freeze_lock);
176 freeze = blk_freeze_set_owner(q, owner);
177 if (++q->mq_freeze_depth == 1) {
178 percpu_ref_kill(&q->q_usage_counter);
179 mutex_unlock(&q->mq_freeze_lock);
180 if (queue_is_mq(q))
181 blk_mq_run_hw_queues(q, false);
182 } else {
183 mutex_unlock(&q->mq_freeze_lock);
184 }
185
186 return freeze;
187 }
188
blk_freeze_queue_start(struct request_queue * q)189 void blk_freeze_queue_start(struct request_queue *q)
190 {
191 if (__blk_freeze_queue_start(q, current))
192 blk_freeze_acquire_lock(q, false, false);
193 }
194 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
195
blk_mq_freeze_queue_wait(struct request_queue * q)196 void blk_mq_freeze_queue_wait(struct request_queue *q)
197 {
198 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
199 }
200 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
201
blk_mq_freeze_queue_wait_timeout(struct request_queue * q,unsigned long timeout)202 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
203 unsigned long timeout)
204 {
205 return wait_event_timeout(q->mq_freeze_wq,
206 percpu_ref_is_zero(&q->q_usage_counter),
207 timeout);
208 }
209 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
210
blk_mq_freeze_queue(struct request_queue * q)211 void blk_mq_freeze_queue(struct request_queue *q)
212 {
213 blk_freeze_queue_start(q);
214 blk_mq_freeze_queue_wait(q);
215 }
216 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
217
__blk_mq_unfreeze_queue(struct request_queue * q,bool force_atomic)218 bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
219 {
220 bool unfreeze;
221
222 mutex_lock(&q->mq_freeze_lock);
223 if (force_atomic)
224 q->q_usage_counter.data->force_atomic = true;
225 q->mq_freeze_depth--;
226 WARN_ON_ONCE(q->mq_freeze_depth < 0);
227 if (!q->mq_freeze_depth) {
228 percpu_ref_resurrect(&q->q_usage_counter);
229 wake_up_all(&q->mq_freeze_wq);
230 }
231 unfreeze = blk_unfreeze_check_owner(q);
232 mutex_unlock(&q->mq_freeze_lock);
233
234 return unfreeze;
235 }
236
blk_mq_unfreeze_queue(struct request_queue * q)237 void blk_mq_unfreeze_queue(struct request_queue *q)
238 {
239 if (__blk_mq_unfreeze_queue(q, false))
240 blk_unfreeze_release_lock(q, false, false);
241 }
242 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
243
244 /*
245 * non_owner variant of blk_freeze_queue_start
246 *
247 * Unlike blk_freeze_queue_start, the queue doesn't need to be unfrozen
248 * by the same task. This is fragile and should not be used if at all
249 * possible.
250 */
blk_freeze_queue_start_non_owner(struct request_queue * q)251 void blk_freeze_queue_start_non_owner(struct request_queue *q)
252 {
253 __blk_freeze_queue_start(q, NULL);
254 }
255 EXPORT_SYMBOL_GPL(blk_freeze_queue_start_non_owner);
256
257 /* non_owner variant of blk_mq_unfreeze_queue */
blk_mq_unfreeze_queue_non_owner(struct request_queue * q)258 void blk_mq_unfreeze_queue_non_owner(struct request_queue *q)
259 {
260 __blk_mq_unfreeze_queue(q, false);
261 }
262 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue_non_owner);
263
264 /*
265 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
266 * mpt3sas driver such that this function can be removed.
267 */
blk_mq_quiesce_queue_nowait(struct request_queue * q)268 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
269 {
270 unsigned long flags;
271
272 spin_lock_irqsave(&q->queue_lock, flags);
273 if (!q->quiesce_depth++)
274 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
275 spin_unlock_irqrestore(&q->queue_lock, flags);
276 }
277 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
278
279 /**
280 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
281 * @set: tag_set to wait on
282 *
283 * Note: it is driver's responsibility for making sure that quiesce has
284 * been started on or more of the request_queues of the tag_set. This
285 * function only waits for the quiesce on those request_queues that had
286 * the quiesce flag set using blk_mq_quiesce_queue_nowait.
287 */
blk_mq_wait_quiesce_done(struct blk_mq_tag_set * set)288 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
289 {
290 if (set->flags & BLK_MQ_F_BLOCKING)
291 synchronize_srcu(set->srcu);
292 else
293 synchronize_rcu();
294 }
295 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
296
297 /**
298 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
299 * @q: request queue.
300 *
301 * Note: this function does not prevent that the struct request end_io()
302 * callback function is invoked. Once this function is returned, we make
303 * sure no dispatch can happen until the queue is unquiesced via
304 * blk_mq_unquiesce_queue().
305 */
blk_mq_quiesce_queue(struct request_queue * q)306 void blk_mq_quiesce_queue(struct request_queue *q)
307 {
308 blk_mq_quiesce_queue_nowait(q);
309 /* nothing to wait for non-mq queues */
310 if (queue_is_mq(q))
311 blk_mq_wait_quiesce_done(q->tag_set);
312 }
313 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
314
315 /*
316 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
317 * @q: request queue.
318 *
319 * This function recovers queue into the state before quiescing
320 * which is done by blk_mq_quiesce_queue.
321 */
blk_mq_unquiesce_queue(struct request_queue * q)322 void blk_mq_unquiesce_queue(struct request_queue *q)
323 {
324 unsigned long flags;
325 bool run_queue = false;
326
327 spin_lock_irqsave(&q->queue_lock, flags);
328 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
329 ;
330 } else if (!--q->quiesce_depth) {
331 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
332 run_queue = true;
333 }
334 spin_unlock_irqrestore(&q->queue_lock, flags);
335
336 /* dispatch requests which are inserted during quiescing */
337 if (run_queue)
338 blk_mq_run_hw_queues(q, true);
339 }
340 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
341
blk_mq_quiesce_tagset(struct blk_mq_tag_set * set)342 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
343 {
344 struct request_queue *q;
345
346 mutex_lock(&set->tag_list_lock);
347 list_for_each_entry(q, &set->tag_list, tag_set_list) {
348 if (!blk_queue_skip_tagset_quiesce(q))
349 blk_mq_quiesce_queue_nowait(q);
350 }
351 mutex_unlock(&set->tag_list_lock);
352
353 blk_mq_wait_quiesce_done(set);
354 }
355 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
356
blk_mq_unquiesce_tagset(struct blk_mq_tag_set * set)357 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
358 {
359 struct request_queue *q;
360
361 mutex_lock(&set->tag_list_lock);
362 list_for_each_entry(q, &set->tag_list, tag_set_list) {
363 if (!blk_queue_skip_tagset_quiesce(q))
364 blk_mq_unquiesce_queue(q);
365 }
366 mutex_unlock(&set->tag_list_lock);
367 }
368 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
369
blk_mq_wake_waiters(struct request_queue * q)370 void blk_mq_wake_waiters(struct request_queue *q)
371 {
372 struct blk_mq_hw_ctx *hctx;
373 unsigned long i;
374
375 queue_for_each_hw_ctx(q, hctx, i)
376 if (blk_mq_hw_queue_mapped(hctx))
377 blk_mq_tag_wakeup_all(hctx->tags, true);
378 }
379
blk_rq_init(struct request_queue * q,struct request * rq)380 void blk_rq_init(struct request_queue *q, struct request *rq)
381 {
382 memset(rq, 0, sizeof(*rq));
383
384 INIT_LIST_HEAD(&rq->queuelist);
385 rq->q = q;
386 rq->__sector = (sector_t) -1;
387 INIT_HLIST_NODE(&rq->hash);
388 RB_CLEAR_NODE(&rq->rb_node);
389 rq->tag = BLK_MQ_NO_TAG;
390 rq->internal_tag = BLK_MQ_NO_TAG;
391 rq->start_time_ns = blk_time_get_ns();
392 blk_crypto_rq_set_defaults(rq);
393 }
394 EXPORT_SYMBOL(blk_rq_init);
395
396 /* Set start and alloc time when the allocated request is actually used */
blk_mq_rq_time_init(struct request * rq,u64 alloc_time_ns)397 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
398 {
399 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
400 if (blk_queue_rq_alloc_time(rq->q))
401 rq->alloc_time_ns = alloc_time_ns;
402 else
403 rq->alloc_time_ns = 0;
404 #endif
405 }
406
blk_mq_rq_ctx_init(struct blk_mq_alloc_data * data,struct blk_mq_tags * tags,unsigned int tag)407 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
408 struct blk_mq_tags *tags, unsigned int tag)
409 {
410 struct blk_mq_ctx *ctx = data->ctx;
411 struct blk_mq_hw_ctx *hctx = data->hctx;
412 struct request_queue *q = data->q;
413 struct request *rq = tags->static_rqs[tag];
414
415 rq->q = q;
416 rq->mq_ctx = ctx;
417 rq->mq_hctx = hctx;
418 rq->cmd_flags = data->cmd_flags;
419
420 if (data->flags & BLK_MQ_REQ_PM)
421 data->rq_flags |= RQF_PM;
422 rq->rq_flags = data->rq_flags;
423
424 if (data->rq_flags & RQF_SCHED_TAGS) {
425 rq->tag = BLK_MQ_NO_TAG;
426 rq->internal_tag = tag;
427 } else {
428 rq->tag = tag;
429 rq->internal_tag = BLK_MQ_NO_TAG;
430 }
431 rq->timeout = 0;
432
433 rq->part = NULL;
434 rq->io_start_time_ns = 0;
435 rq->stats_sectors = 0;
436 rq->nr_phys_segments = 0;
437 rq->nr_integrity_segments = 0;
438 rq->end_io = NULL;
439 rq->end_io_data = NULL;
440
441 blk_crypto_rq_set_defaults(rq);
442 INIT_LIST_HEAD(&rq->queuelist);
443 /* tag was already set */
444 WRITE_ONCE(rq->deadline, 0);
445 req_ref_set(rq, 1);
446
447 if (rq->rq_flags & RQF_USE_SCHED) {
448 struct elevator_queue *e = data->q->elevator;
449
450 INIT_HLIST_NODE(&rq->hash);
451 RB_CLEAR_NODE(&rq->rb_node);
452
453 if (e->type->ops.prepare_request)
454 e->type->ops.prepare_request(rq);
455 }
456
457 return rq;
458 }
459
460 static inline struct request *
__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data * data)461 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
462 {
463 unsigned int tag, tag_offset;
464 struct blk_mq_tags *tags;
465 struct request *rq;
466 unsigned long tag_mask;
467 int i, nr = 0;
468
469 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
470 if (unlikely(!tag_mask))
471 return NULL;
472
473 tags = blk_mq_tags_from_data(data);
474 for (i = 0; tag_mask; i++) {
475 if (!(tag_mask & (1UL << i)))
476 continue;
477 tag = tag_offset + i;
478 prefetch(tags->static_rqs[tag]);
479 tag_mask &= ~(1UL << i);
480 rq = blk_mq_rq_ctx_init(data, tags, tag);
481 rq_list_add_head(data->cached_rqs, rq);
482 nr++;
483 }
484 if (!(data->rq_flags & RQF_SCHED_TAGS))
485 blk_mq_add_active_requests(data->hctx, nr);
486 /* caller already holds a reference, add for remainder */
487 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
488 data->nr_tags -= nr;
489
490 return rq_list_pop(data->cached_rqs);
491 }
492
__blk_mq_alloc_requests(struct blk_mq_alloc_data * data)493 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
494 {
495 struct request_queue *q = data->q;
496 u64 alloc_time_ns = 0;
497 struct request *rq;
498 unsigned int tag;
499
500 /* alloc_time includes depth and tag waits */
501 if (blk_queue_rq_alloc_time(q))
502 alloc_time_ns = blk_time_get_ns();
503
504 if (data->cmd_flags & REQ_NOWAIT)
505 data->flags |= BLK_MQ_REQ_NOWAIT;
506
507 retry:
508 data->ctx = blk_mq_get_ctx(q);
509 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
510
511 if (q->elevator) {
512 /*
513 * All requests use scheduler tags when an I/O scheduler is
514 * enabled for the queue.
515 */
516 data->rq_flags |= RQF_SCHED_TAGS;
517
518 /*
519 * Flush/passthrough requests are special and go directly to the
520 * dispatch list.
521 */
522 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
523 !blk_op_is_passthrough(data->cmd_flags)) {
524 struct elevator_mq_ops *ops = &q->elevator->type->ops;
525
526 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
527
528 data->rq_flags |= RQF_USE_SCHED;
529 if (ops->limit_depth)
530 ops->limit_depth(data->cmd_flags, data);
531 }
532 } else {
533 blk_mq_tag_busy(data->hctx);
534 }
535
536 if (data->flags & BLK_MQ_REQ_RESERVED)
537 data->rq_flags |= RQF_RESV;
538
539 /*
540 * Try batched alloc if we want more than 1 tag.
541 */
542 if (data->nr_tags > 1) {
543 rq = __blk_mq_alloc_requests_batch(data);
544 if (rq) {
545 blk_mq_rq_time_init(rq, alloc_time_ns);
546 return rq;
547 }
548 data->nr_tags = 1;
549 }
550
551 /*
552 * Waiting allocations only fail because of an inactive hctx. In that
553 * case just retry the hctx assignment and tag allocation as CPU hotplug
554 * should have migrated us to an online CPU by now.
555 */
556 tag = blk_mq_get_tag(data);
557 if (tag == BLK_MQ_NO_TAG) {
558 if (data->flags & BLK_MQ_REQ_NOWAIT)
559 return NULL;
560 /*
561 * Give up the CPU and sleep for a random short time to
562 * ensure that thread using a realtime scheduling class
563 * are migrated off the CPU, and thus off the hctx that
564 * is going away.
565 */
566 msleep(3);
567 goto retry;
568 }
569
570 if (!(data->rq_flags & RQF_SCHED_TAGS))
571 blk_mq_inc_active_requests(data->hctx);
572 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
573 blk_mq_rq_time_init(rq, alloc_time_ns);
574 return rq;
575 }
576
blk_mq_rq_cache_fill(struct request_queue * q,struct blk_plug * plug,blk_opf_t opf,blk_mq_req_flags_t flags)577 static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
578 struct blk_plug *plug,
579 blk_opf_t opf,
580 blk_mq_req_flags_t flags)
581 {
582 struct blk_mq_alloc_data data = {
583 .q = q,
584 .flags = flags,
585 .cmd_flags = opf,
586 .nr_tags = plug->nr_ios,
587 .cached_rqs = &plug->cached_rqs,
588 };
589 struct request *rq;
590
591 if (blk_queue_enter(q, flags))
592 return NULL;
593
594 plug->nr_ios = 1;
595
596 rq = __blk_mq_alloc_requests(&data);
597 if (unlikely(!rq))
598 blk_queue_exit(q);
599 return rq;
600 }
601
blk_mq_alloc_cached_request(struct request_queue * q,blk_opf_t opf,blk_mq_req_flags_t flags)602 static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
603 blk_opf_t opf,
604 blk_mq_req_flags_t flags)
605 {
606 struct blk_plug *plug = current->plug;
607 struct request *rq;
608
609 if (!plug)
610 return NULL;
611
612 if (rq_list_empty(&plug->cached_rqs)) {
613 if (plug->nr_ios == 1)
614 return NULL;
615 rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
616 if (!rq)
617 return NULL;
618 } else {
619 rq = rq_list_peek(&plug->cached_rqs);
620 if (!rq || rq->q != q)
621 return NULL;
622
623 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
624 return NULL;
625 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
626 return NULL;
627
628 rq_list_pop(&plug->cached_rqs);
629 blk_mq_rq_time_init(rq, blk_time_get_ns());
630 }
631
632 rq->cmd_flags = opf;
633 INIT_LIST_HEAD(&rq->queuelist);
634 return rq;
635 }
636
blk_mq_alloc_request(struct request_queue * q,blk_opf_t opf,blk_mq_req_flags_t flags)637 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
638 blk_mq_req_flags_t flags)
639 {
640 struct request *rq;
641
642 rq = blk_mq_alloc_cached_request(q, opf, flags);
643 if (!rq) {
644 struct blk_mq_alloc_data data = {
645 .q = q,
646 .flags = flags,
647 .cmd_flags = opf,
648 .nr_tags = 1,
649 };
650 int ret;
651
652 ret = blk_queue_enter(q, flags);
653 if (ret)
654 return ERR_PTR(ret);
655
656 rq = __blk_mq_alloc_requests(&data);
657 if (!rq)
658 goto out_queue_exit;
659 }
660 rq->__data_len = 0;
661 rq->__sector = (sector_t) -1;
662 rq->bio = rq->biotail = NULL;
663 return rq;
664 out_queue_exit:
665 blk_queue_exit(q);
666 return ERR_PTR(-EWOULDBLOCK);
667 }
668 EXPORT_SYMBOL(blk_mq_alloc_request);
669
blk_mq_alloc_request_hctx(struct request_queue * q,blk_opf_t opf,blk_mq_req_flags_t flags,unsigned int hctx_idx)670 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
671 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
672 {
673 struct blk_mq_alloc_data data = {
674 .q = q,
675 .flags = flags,
676 .cmd_flags = opf,
677 .nr_tags = 1,
678 };
679 u64 alloc_time_ns = 0;
680 struct request *rq;
681 unsigned int cpu;
682 unsigned int tag;
683 int ret;
684
685 /* alloc_time includes depth and tag waits */
686 if (blk_queue_rq_alloc_time(q))
687 alloc_time_ns = blk_time_get_ns();
688
689 /*
690 * If the tag allocator sleeps we could get an allocation for a
691 * different hardware context. No need to complicate the low level
692 * allocator for this for the rare use case of a command tied to
693 * a specific queue.
694 */
695 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
696 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
697 return ERR_PTR(-EINVAL);
698
699 if (hctx_idx >= q->nr_hw_queues)
700 return ERR_PTR(-EIO);
701
702 ret = blk_queue_enter(q, flags);
703 if (ret)
704 return ERR_PTR(ret);
705
706 /*
707 * Check if the hardware context is actually mapped to anything.
708 * If not tell the caller that it should skip this queue.
709 */
710 ret = -EXDEV;
711 data.hctx = xa_load(&q->hctx_table, hctx_idx);
712 if (!blk_mq_hw_queue_mapped(data.hctx))
713 goto out_queue_exit;
714 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
715 if (cpu >= nr_cpu_ids)
716 goto out_queue_exit;
717 data.ctx = __blk_mq_get_ctx(q, cpu);
718
719 if (q->elevator)
720 data.rq_flags |= RQF_SCHED_TAGS;
721 else
722 blk_mq_tag_busy(data.hctx);
723
724 if (flags & BLK_MQ_REQ_RESERVED)
725 data.rq_flags |= RQF_RESV;
726
727 ret = -EWOULDBLOCK;
728 tag = blk_mq_get_tag(&data);
729 if (tag == BLK_MQ_NO_TAG)
730 goto out_queue_exit;
731 if (!(data.rq_flags & RQF_SCHED_TAGS))
732 blk_mq_inc_active_requests(data.hctx);
733 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
734 blk_mq_rq_time_init(rq, alloc_time_ns);
735 rq->__data_len = 0;
736 rq->__sector = (sector_t) -1;
737 rq->bio = rq->biotail = NULL;
738 return rq;
739
740 out_queue_exit:
741 blk_queue_exit(q);
742 return ERR_PTR(ret);
743 }
744 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
745
blk_mq_finish_request(struct request * rq)746 static void blk_mq_finish_request(struct request *rq)
747 {
748 struct request_queue *q = rq->q;
749
750 blk_zone_finish_request(rq);
751
752 if (rq->rq_flags & RQF_USE_SCHED) {
753 q->elevator->type->ops.finish_request(rq);
754 /*
755 * For postflush request that may need to be
756 * completed twice, we should clear this flag
757 * to avoid double finish_request() on the rq.
758 */
759 rq->rq_flags &= ~RQF_USE_SCHED;
760 }
761 }
762
__blk_mq_free_request(struct request * rq)763 static void __blk_mq_free_request(struct request *rq)
764 {
765 struct request_queue *q = rq->q;
766 struct blk_mq_ctx *ctx = rq->mq_ctx;
767 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
768 const int sched_tag = rq->internal_tag;
769
770 blk_crypto_free_request(rq);
771 blk_pm_mark_last_busy(rq);
772 rq->mq_hctx = NULL;
773
774 if (rq->tag != BLK_MQ_NO_TAG) {
775 blk_mq_dec_active_requests(hctx);
776 blk_mq_put_tag(hctx->tags, ctx, rq->tag);
777 }
778 if (sched_tag != BLK_MQ_NO_TAG)
779 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
780 blk_mq_sched_restart(hctx);
781 blk_queue_exit(q);
782 }
783
blk_mq_free_request(struct request * rq)784 void blk_mq_free_request(struct request *rq)
785 {
786 struct request_queue *q = rq->q;
787
788 blk_mq_finish_request(rq);
789
790 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
791 laptop_io_completion(q->disk->bdi);
792
793 rq_qos_done(q, rq);
794
795 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
796 if (req_ref_put_and_test(rq))
797 __blk_mq_free_request(rq);
798 }
799 EXPORT_SYMBOL_GPL(blk_mq_free_request);
800
blk_mq_free_plug_rqs(struct blk_plug * plug)801 void blk_mq_free_plug_rqs(struct blk_plug *plug)
802 {
803 struct request *rq;
804
805 while ((rq = rq_list_pop(&plug->cached_rqs)) != NULL)
806 blk_mq_free_request(rq);
807 }
808
blk_dump_rq_flags(struct request * rq,char * msg)809 void blk_dump_rq_flags(struct request *rq, char *msg)
810 {
811 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
812 rq->q->disk ? rq->q->disk->disk_name : "?",
813 (__force unsigned long long) rq->cmd_flags);
814
815 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
816 (unsigned long long)blk_rq_pos(rq),
817 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
818 printk(KERN_INFO " bio %p, biotail %p, len %u\n",
819 rq->bio, rq->biotail, blk_rq_bytes(rq));
820 }
821 EXPORT_SYMBOL(blk_dump_rq_flags);
822
blk_account_io_completion(struct request * req,unsigned int bytes)823 static void blk_account_io_completion(struct request *req, unsigned int bytes)
824 {
825 if (req->rq_flags & RQF_IO_STAT) {
826 const int sgrp = op_stat_group(req_op(req));
827
828 part_stat_lock();
829 part_stat_add(req->part, sectors[sgrp], bytes >> 9);
830 part_stat_unlock();
831 }
832 }
833
blk_print_req_error(struct request * req,blk_status_t status)834 static void blk_print_req_error(struct request *req, blk_status_t status)
835 {
836 printk_ratelimited(KERN_ERR
837 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
838 "phys_seg %u prio class %u\n",
839 blk_status_to_str(status),
840 req->q->disk ? req->q->disk->disk_name : "?",
841 blk_rq_pos(req), (__force u32)req_op(req),
842 blk_op_str(req_op(req)),
843 (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
844 req->nr_phys_segments,
845 IOPRIO_PRIO_CLASS(req_get_ioprio(req)));
846 }
847
848 /*
849 * Fully end IO on a request. Does not support partial completions, or
850 * errors.
851 */
blk_complete_request(struct request * req)852 static void blk_complete_request(struct request *req)
853 {
854 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
855 int total_bytes = blk_rq_bytes(req);
856 struct bio *bio = req->bio;
857
858 trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
859
860 if (!bio)
861 return;
862
863 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
864 blk_integrity_complete(req, total_bytes);
865
866 /*
867 * Upper layers may call blk_crypto_evict_key() anytime after the last
868 * bio_endio(). Therefore, the keyslot must be released before that.
869 */
870 blk_crypto_rq_put_keyslot(req);
871
872 blk_account_io_completion(req, total_bytes);
873
874 do {
875 struct bio *next = bio->bi_next;
876
877 /* Completion has already been traced */
878 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
879
880 blk_zone_update_request_bio(req, bio);
881
882 if (!is_flush)
883 bio_endio(bio);
884 bio = next;
885 } while (bio);
886
887 /*
888 * Reset counters so that the request stacking driver
889 * can find how many bytes remain in the request
890 * later.
891 */
892 if (!req->end_io) {
893 req->bio = NULL;
894 req->__data_len = 0;
895 }
896 }
897
898 /**
899 * blk_update_request - Complete multiple bytes without completing the request
900 * @req: the request being processed
901 * @error: block status code
902 * @nr_bytes: number of bytes to complete for @req
903 *
904 * Description:
905 * Ends I/O on a number of bytes attached to @req, but doesn't complete
906 * the request structure even if @req doesn't have leftover.
907 * If @req has leftover, sets it up for the next range of segments.
908 *
909 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
910 * %false return from this function.
911 *
912 * Note:
913 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
914 * except in the consistency check at the end of this function.
915 *
916 * Return:
917 * %false - this request doesn't have any more data
918 * %true - this request has more data
919 **/
blk_update_request(struct request * req,blk_status_t error,unsigned int nr_bytes)920 bool blk_update_request(struct request *req, blk_status_t error,
921 unsigned int nr_bytes)
922 {
923 bool is_flush = req->rq_flags & RQF_FLUSH_SEQ;
924 bool quiet = req->rq_flags & RQF_QUIET;
925 int total_bytes;
926
927 trace_block_rq_complete(req, error, nr_bytes);
928
929 if (!req->bio)
930 return false;
931
932 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
933 error == BLK_STS_OK)
934 blk_integrity_complete(req, nr_bytes);
935
936 /*
937 * Upper layers may call blk_crypto_evict_key() anytime after the last
938 * bio_endio(). Therefore, the keyslot must be released before that.
939 */
940 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
941 __blk_crypto_rq_put_keyslot(req);
942
943 if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) &&
944 !test_bit(GD_DEAD, &req->q->disk->state)) {
945 blk_print_req_error(req, error);
946 trace_block_rq_error(req, error, nr_bytes);
947 }
948
949 blk_account_io_completion(req, nr_bytes);
950
951 total_bytes = 0;
952 while (req->bio) {
953 struct bio *bio = req->bio;
954 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
955
956 if (unlikely(error))
957 bio->bi_status = error;
958
959 if (bio_bytes == bio->bi_iter.bi_size) {
960 req->bio = bio->bi_next;
961 } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) {
962 /*
963 * Partial zone append completions cannot be supported
964 * as the BIO fragments may end up not being written
965 * sequentially.
966 */
967 bio->bi_status = BLK_STS_IOERR;
968 }
969
970 /* Completion has already been traced */
971 bio_clear_flag(bio, BIO_TRACE_COMPLETION);
972 if (unlikely(quiet))
973 bio_set_flag(bio, BIO_QUIET);
974
975 bio_advance(bio, bio_bytes);
976
977 /* Don't actually finish bio if it's part of flush sequence */
978 if (!bio->bi_iter.bi_size) {
979 blk_zone_update_request_bio(req, bio);
980 if (!is_flush)
981 bio_endio(bio);
982 }
983
984 total_bytes += bio_bytes;
985 nr_bytes -= bio_bytes;
986
987 if (!nr_bytes)
988 break;
989 }
990
991 /*
992 * completely done
993 */
994 if (!req->bio) {
995 /*
996 * Reset counters so that the request stacking driver
997 * can find how many bytes remain in the request
998 * later.
999 */
1000 req->__data_len = 0;
1001 return false;
1002 }
1003
1004 req->__data_len -= total_bytes;
1005
1006 /* update sector only for requests with clear definition of sector */
1007 if (!blk_rq_is_passthrough(req))
1008 req->__sector += total_bytes >> 9;
1009
1010 /* mixed attributes always follow the first bio */
1011 if (req->rq_flags & RQF_MIXED_MERGE) {
1012 req->cmd_flags &= ~REQ_FAILFAST_MASK;
1013 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
1014 }
1015
1016 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
1017 /*
1018 * If total number of sectors is less than the first segment
1019 * size, something has gone terribly wrong.
1020 */
1021 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
1022 blk_dump_rq_flags(req, "request botched");
1023 req->__data_len = blk_rq_cur_bytes(req);
1024 }
1025
1026 /* recalculate the number of segments */
1027 req->nr_phys_segments = blk_recalc_rq_segments(req);
1028 }
1029
1030 return true;
1031 }
1032 EXPORT_SYMBOL_GPL(blk_update_request);
1033
blk_account_io_done(struct request * req,u64 now)1034 static inline void blk_account_io_done(struct request *req, u64 now)
1035 {
1036 trace_block_io_done(req);
1037
1038 /*
1039 * Account IO completion. flush_rq isn't accounted as a
1040 * normal IO on queueing nor completion. Accounting the
1041 * containing request is enough.
1042 */
1043 if ((req->rq_flags & (RQF_IO_STAT|RQF_FLUSH_SEQ)) == RQF_IO_STAT) {
1044 const int sgrp = op_stat_group(req_op(req));
1045
1046 part_stat_lock();
1047 update_io_ticks(req->part, jiffies, true);
1048 part_stat_inc(req->part, ios[sgrp]);
1049 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
1050 part_stat_local_dec(req->part,
1051 in_flight[op_is_write(req_op(req))]);
1052 part_stat_unlock();
1053 }
1054 }
1055
blk_rq_passthrough_stats(struct request * req)1056 static inline bool blk_rq_passthrough_stats(struct request *req)
1057 {
1058 struct bio *bio = req->bio;
1059
1060 if (!blk_queue_passthrough_stat(req->q))
1061 return false;
1062
1063 /* Requests without a bio do not transfer data. */
1064 if (!bio)
1065 return false;
1066
1067 /*
1068 * Stats are accumulated in the bdev, so must have one attached to a
1069 * bio to track stats. Most drivers do not set the bdev for passthrough
1070 * requests, but nvme is one that will set it.
1071 */
1072 if (!bio->bi_bdev)
1073 return false;
1074
1075 /*
1076 * We don't know what a passthrough command does, but we know the
1077 * payload size and data direction. Ensuring the size is aligned to the
1078 * block size filters out most commands with payloads that don't
1079 * represent sector access.
1080 */
1081 if (blk_rq_bytes(req) & (bdev_logical_block_size(bio->bi_bdev) - 1))
1082 return false;
1083 return true;
1084 }
1085
blk_account_io_start(struct request * req)1086 static inline void blk_account_io_start(struct request *req)
1087 {
1088 trace_block_io_start(req);
1089
1090 if (!blk_queue_io_stat(req->q))
1091 return;
1092 if (blk_rq_is_passthrough(req) && !blk_rq_passthrough_stats(req))
1093 return;
1094
1095 req->rq_flags |= RQF_IO_STAT;
1096 req->start_time_ns = blk_time_get_ns();
1097
1098 /*
1099 * All non-passthrough requests are created from a bio with one
1100 * exception: when a flush command that is part of a flush sequence
1101 * generated by the state machine in blk-flush.c is cloned onto the
1102 * lower device by dm-multipath we can get here without a bio.
1103 */
1104 if (req->bio)
1105 req->part = req->bio->bi_bdev;
1106 else
1107 req->part = req->q->disk->part0;
1108
1109 part_stat_lock();
1110 update_io_ticks(req->part, jiffies, false);
1111 part_stat_local_inc(req->part, in_flight[op_is_write(req_op(req))]);
1112 part_stat_unlock();
1113 }
1114
__blk_mq_end_request_acct(struct request * rq,u64 now)1115 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1116 {
1117 if (rq->rq_flags & RQF_STATS)
1118 blk_stat_add(rq, now);
1119
1120 blk_mq_sched_completed_request(rq, now);
1121 blk_account_io_done(rq, now);
1122 }
1123
__blk_mq_end_request(struct request * rq,blk_status_t error)1124 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1125 {
1126 if (blk_mq_need_time_stamp(rq))
1127 __blk_mq_end_request_acct(rq, blk_time_get_ns());
1128
1129 blk_mq_finish_request(rq);
1130
1131 if (rq->end_io) {
1132 rq_qos_done(rq->q, rq);
1133 if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1134 blk_mq_free_request(rq);
1135 } else {
1136 blk_mq_free_request(rq);
1137 }
1138 }
1139 EXPORT_SYMBOL(__blk_mq_end_request);
1140
blk_mq_end_request(struct request * rq,blk_status_t error)1141 void blk_mq_end_request(struct request *rq, blk_status_t error)
1142 {
1143 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1144 BUG();
1145 __blk_mq_end_request(rq, error);
1146 }
1147 EXPORT_SYMBOL(blk_mq_end_request);
1148
1149 #define TAG_COMP_BATCH 32
1150
blk_mq_flush_tag_batch(struct blk_mq_hw_ctx * hctx,int * tag_array,int nr_tags)1151 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1152 int *tag_array, int nr_tags)
1153 {
1154 struct request_queue *q = hctx->queue;
1155
1156 blk_mq_sub_active_requests(hctx, nr_tags);
1157
1158 blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1159 percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1160 }
1161
blk_mq_end_request_batch(struct io_comp_batch * iob)1162 void blk_mq_end_request_batch(struct io_comp_batch *iob)
1163 {
1164 int tags[TAG_COMP_BATCH], nr_tags = 0;
1165 struct blk_mq_hw_ctx *cur_hctx = NULL;
1166 struct request *rq;
1167 u64 now = 0;
1168
1169 if (iob->need_ts)
1170 now = blk_time_get_ns();
1171
1172 while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1173 prefetch(rq->bio);
1174 prefetch(rq->rq_next);
1175
1176 blk_complete_request(rq);
1177 if (iob->need_ts)
1178 __blk_mq_end_request_acct(rq, now);
1179
1180 blk_mq_finish_request(rq);
1181
1182 rq_qos_done(rq->q, rq);
1183
1184 /*
1185 * If end_io handler returns NONE, then it still has
1186 * ownership of the request.
1187 */
1188 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1189 continue;
1190
1191 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1192 if (!req_ref_put_and_test(rq))
1193 continue;
1194
1195 blk_crypto_free_request(rq);
1196 blk_pm_mark_last_busy(rq);
1197
1198 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1199 if (cur_hctx)
1200 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1201 nr_tags = 0;
1202 cur_hctx = rq->mq_hctx;
1203 }
1204 tags[nr_tags++] = rq->tag;
1205 }
1206
1207 if (nr_tags)
1208 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1209 }
1210 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1211
blk_complete_reqs(struct llist_head * list)1212 static void blk_complete_reqs(struct llist_head *list)
1213 {
1214 struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1215 struct request *rq, *next;
1216
1217 llist_for_each_entry_safe(rq, next, entry, ipi_list)
1218 rq->q->mq_ops->complete(rq);
1219 }
1220
blk_done_softirq(void)1221 static __latent_entropy void blk_done_softirq(void)
1222 {
1223 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1224 }
1225
blk_softirq_cpu_dead(unsigned int cpu)1226 static int blk_softirq_cpu_dead(unsigned int cpu)
1227 {
1228 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1229 return 0;
1230 }
1231
__blk_mq_complete_request_remote(void * data)1232 static void __blk_mq_complete_request_remote(void *data)
1233 {
1234 __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1235 }
1236
blk_mq_complete_need_ipi(struct request * rq)1237 static inline bool blk_mq_complete_need_ipi(struct request *rq)
1238 {
1239 int cpu = raw_smp_processor_id();
1240
1241 if (!IS_ENABLED(CONFIG_SMP) ||
1242 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1243 return false;
1244 /*
1245 * With force threaded interrupts enabled, raising softirq from an SMP
1246 * function call will always result in waking the ksoftirqd thread.
1247 * This is probably worse than completing the request on a different
1248 * cache domain.
1249 */
1250 if (force_irqthreads())
1251 return false;
1252
1253 /* same CPU or cache domain and capacity? Complete locally */
1254 if (cpu == rq->mq_ctx->cpu ||
1255 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1256 cpus_share_cache(cpu, rq->mq_ctx->cpu) &&
1257 cpus_equal_capacity(cpu, rq->mq_ctx->cpu)))
1258 return false;
1259
1260 /* don't try to IPI to an offline CPU */
1261 return cpu_online(rq->mq_ctx->cpu);
1262 }
1263
blk_mq_complete_send_ipi(struct request * rq)1264 static void blk_mq_complete_send_ipi(struct request *rq)
1265 {
1266 unsigned int cpu;
1267
1268 cpu = rq->mq_ctx->cpu;
1269 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu)))
1270 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu));
1271 }
1272
blk_mq_raise_softirq(struct request * rq)1273 static void blk_mq_raise_softirq(struct request *rq)
1274 {
1275 struct llist_head *list;
1276
1277 preempt_disable();
1278 list = this_cpu_ptr(&blk_cpu_done);
1279 if (llist_add(&rq->ipi_list, list))
1280 raise_softirq(BLOCK_SOFTIRQ);
1281 preempt_enable();
1282 }
1283
blk_mq_complete_request_remote(struct request * rq)1284 bool blk_mq_complete_request_remote(struct request *rq)
1285 {
1286 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1287
1288 /*
1289 * For request which hctx has only one ctx mapping,
1290 * or a polled request, always complete locally,
1291 * it's pointless to redirect the completion.
1292 */
1293 if ((rq->mq_hctx->nr_ctx == 1 &&
1294 rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1295 rq->cmd_flags & REQ_POLLED)
1296 return false;
1297
1298 if (blk_mq_complete_need_ipi(rq)) {
1299 blk_mq_complete_send_ipi(rq);
1300 return true;
1301 }
1302
1303 if (rq->q->nr_hw_queues == 1) {
1304 blk_mq_raise_softirq(rq);
1305 return true;
1306 }
1307 return false;
1308 }
1309 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1310
1311 /**
1312 * blk_mq_complete_request - end I/O on a request
1313 * @rq: the request being processed
1314 *
1315 * Description:
1316 * Complete a request by scheduling the ->complete_rq operation.
1317 **/
blk_mq_complete_request(struct request * rq)1318 void blk_mq_complete_request(struct request *rq)
1319 {
1320 if (!blk_mq_complete_request_remote(rq))
1321 rq->q->mq_ops->complete(rq);
1322 }
1323 EXPORT_SYMBOL(blk_mq_complete_request);
1324
1325 /**
1326 * blk_mq_start_request - Start processing a request
1327 * @rq: Pointer to request to be started
1328 *
1329 * Function used by device drivers to notify the block layer that a request
1330 * is going to be processed now, so blk layer can do proper initializations
1331 * such as starting the timeout timer.
1332 */
blk_mq_start_request(struct request * rq)1333 void blk_mq_start_request(struct request *rq)
1334 {
1335 struct request_queue *q = rq->q;
1336
1337 trace_block_rq_issue(rq);
1338
1339 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) &&
1340 !blk_rq_is_passthrough(rq)) {
1341 rq->io_start_time_ns = blk_time_get_ns();
1342 rq->stats_sectors = blk_rq_sectors(rq);
1343 rq->rq_flags |= RQF_STATS;
1344 rq_qos_issue(q, rq);
1345 }
1346
1347 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1348
1349 blk_add_timer(rq);
1350 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1351 rq->mq_hctx->tags->rqs[rq->tag] = rq;
1352
1353 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1354 blk_integrity_prepare(rq);
1355
1356 if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1357 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1358 }
1359 EXPORT_SYMBOL(blk_mq_start_request);
1360
1361 /*
1362 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1363 * queues. This is important for md arrays to benefit from merging
1364 * requests.
1365 */
blk_plug_max_rq_count(struct blk_plug * plug)1366 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1367 {
1368 if (plug->multiple_queues)
1369 return BLK_MAX_REQUEST_COUNT * 2;
1370 return BLK_MAX_REQUEST_COUNT;
1371 }
1372
blk_add_rq_to_plug(struct blk_plug * plug,struct request * rq)1373 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1374 {
1375 struct request *last = rq_list_peek(&plug->mq_list);
1376
1377 if (!plug->rq_count) {
1378 trace_block_plug(rq->q);
1379 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1380 (!blk_queue_nomerges(rq->q) &&
1381 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1382 blk_mq_flush_plug_list(plug, false);
1383 last = NULL;
1384 trace_block_plug(rq->q);
1385 }
1386
1387 if (!plug->multiple_queues && last && last->q != rq->q)
1388 plug->multiple_queues = true;
1389 /*
1390 * Any request allocated from sched tags can't be issued to
1391 * ->queue_rqs() directly
1392 */
1393 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1394 plug->has_elevator = true;
1395 rq_list_add_tail(&plug->mq_list, rq);
1396 plug->rq_count++;
1397 }
1398
1399 /**
1400 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1401 * @rq: request to insert
1402 * @at_head: insert request at head or tail of queue
1403 *
1404 * Description:
1405 * Insert a fully prepared request at the back of the I/O scheduler queue
1406 * for execution. Don't wait for completion.
1407 *
1408 * Note:
1409 * This function will invoke @done directly if the queue is dead.
1410 */
blk_execute_rq_nowait(struct request * rq,bool at_head)1411 void blk_execute_rq_nowait(struct request *rq, bool at_head)
1412 {
1413 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1414
1415 WARN_ON(irqs_disabled());
1416 WARN_ON(!blk_rq_is_passthrough(rq));
1417
1418 blk_account_io_start(rq);
1419
1420 if (current->plug && !at_head) {
1421 blk_add_rq_to_plug(current->plug, rq);
1422 return;
1423 }
1424
1425 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1426 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
1427 }
1428 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1429
1430 struct blk_rq_wait {
1431 struct completion done;
1432 blk_status_t ret;
1433 };
1434
blk_end_sync_rq(struct request * rq,blk_status_t ret)1435 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1436 {
1437 struct blk_rq_wait *wait = rq->end_io_data;
1438
1439 wait->ret = ret;
1440 complete(&wait->done);
1441 return RQ_END_IO_NONE;
1442 }
1443
blk_rq_is_poll(struct request * rq)1444 bool blk_rq_is_poll(struct request *rq)
1445 {
1446 if (!rq->mq_hctx)
1447 return false;
1448 if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1449 return false;
1450 return true;
1451 }
1452 EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1453
blk_rq_poll_completion(struct request * rq,struct completion * wait)1454 static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1455 {
1456 do {
1457 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1458 cond_resched();
1459 } while (!completion_done(wait));
1460 }
1461
1462 /**
1463 * blk_execute_rq - insert a request into queue for execution
1464 * @rq: request to insert
1465 * @at_head: insert request at head or tail of queue
1466 *
1467 * Description:
1468 * Insert a fully prepared request at the back of the I/O scheduler queue
1469 * for execution and wait for completion.
1470 * Return: The blk_status_t result provided to blk_mq_end_request().
1471 */
blk_execute_rq(struct request * rq,bool at_head)1472 blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1473 {
1474 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1475 struct blk_rq_wait wait = {
1476 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1477 };
1478
1479 WARN_ON(irqs_disabled());
1480 WARN_ON(!blk_rq_is_passthrough(rq));
1481
1482 rq->end_io_data = &wait;
1483 rq->end_io = blk_end_sync_rq;
1484
1485 blk_account_io_start(rq);
1486 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1487 blk_mq_run_hw_queue(hctx, false);
1488
1489 if (blk_rq_is_poll(rq))
1490 blk_rq_poll_completion(rq, &wait.done);
1491 else
1492 blk_wait_io(&wait.done);
1493
1494 return wait.ret;
1495 }
1496 EXPORT_SYMBOL(blk_execute_rq);
1497
__blk_mq_requeue_request(struct request * rq)1498 static void __blk_mq_requeue_request(struct request *rq)
1499 {
1500 struct request_queue *q = rq->q;
1501
1502 blk_mq_put_driver_tag(rq);
1503
1504 trace_block_rq_requeue(rq);
1505 rq_qos_requeue(q, rq);
1506
1507 if (blk_mq_request_started(rq)) {
1508 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1509 rq->rq_flags &= ~RQF_TIMED_OUT;
1510 }
1511 }
1512
blk_mq_requeue_request(struct request * rq,bool kick_requeue_list)1513 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1514 {
1515 struct request_queue *q = rq->q;
1516 unsigned long flags;
1517
1518 __blk_mq_requeue_request(rq);
1519
1520 /* this request will be re-inserted to io scheduler queue */
1521 blk_mq_sched_requeue_request(rq);
1522
1523 spin_lock_irqsave(&q->requeue_lock, flags);
1524 list_add_tail(&rq->queuelist, &q->requeue_list);
1525 spin_unlock_irqrestore(&q->requeue_lock, flags);
1526
1527 if (kick_requeue_list)
1528 blk_mq_kick_requeue_list(q);
1529 }
1530 EXPORT_SYMBOL(blk_mq_requeue_request);
1531
blk_mq_requeue_work(struct work_struct * work)1532 static void blk_mq_requeue_work(struct work_struct *work)
1533 {
1534 struct request_queue *q =
1535 container_of(work, struct request_queue, requeue_work.work);
1536 LIST_HEAD(rq_list);
1537 LIST_HEAD(flush_list);
1538 struct request *rq;
1539
1540 spin_lock_irq(&q->requeue_lock);
1541 list_splice_init(&q->requeue_list, &rq_list);
1542 list_splice_init(&q->flush_list, &flush_list);
1543 spin_unlock_irq(&q->requeue_lock);
1544
1545 while (!list_empty(&rq_list)) {
1546 rq = list_entry(rq_list.next, struct request, queuelist);
1547 list_del_init(&rq->queuelist);
1548 /*
1549 * If RQF_DONTPREP is set, the request has been started by the
1550 * driver already and might have driver-specific data allocated
1551 * already. Insert it into the hctx dispatch list to avoid
1552 * block layer merges for the request.
1553 */
1554 if (rq->rq_flags & RQF_DONTPREP)
1555 blk_mq_request_bypass_insert(rq, 0);
1556 else
1557 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1558 }
1559
1560 while (!list_empty(&flush_list)) {
1561 rq = list_entry(flush_list.next, struct request, queuelist);
1562 list_del_init(&rq->queuelist);
1563 blk_mq_insert_request(rq, 0);
1564 }
1565
1566 blk_mq_run_hw_queues(q, false);
1567 }
1568
blk_mq_kick_requeue_list(struct request_queue * q)1569 void blk_mq_kick_requeue_list(struct request_queue *q)
1570 {
1571 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1572 }
1573 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1574
blk_mq_delay_kick_requeue_list(struct request_queue * q,unsigned long msecs)1575 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1576 unsigned long msecs)
1577 {
1578 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1579 msecs_to_jiffies(msecs));
1580 }
1581 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1582
blk_is_flush_data_rq(struct request * rq)1583 static bool blk_is_flush_data_rq(struct request *rq)
1584 {
1585 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq);
1586 }
1587
blk_mq_rq_inflight(struct request * rq,void * priv)1588 static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1589 {
1590 /*
1591 * If we find a request that isn't idle we know the queue is busy
1592 * as it's checked in the iter.
1593 * Return false to stop the iteration.
1594 *
1595 * In case of queue quiesce, if one flush data request is completed,
1596 * don't count it as inflight given the flush sequence is suspended,
1597 * and the original flush data request is invisible to driver, just
1598 * like other pending requests because of quiesce
1599 */
1600 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) &&
1601 blk_is_flush_data_rq(rq) &&
1602 blk_mq_request_completed(rq))) {
1603 bool *busy = priv;
1604
1605 *busy = true;
1606 return false;
1607 }
1608
1609 return true;
1610 }
1611
blk_mq_queue_inflight(struct request_queue * q)1612 bool blk_mq_queue_inflight(struct request_queue *q)
1613 {
1614 bool busy = false;
1615
1616 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1617 return busy;
1618 }
1619 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1620
blk_mq_rq_timed_out(struct request * req)1621 static void blk_mq_rq_timed_out(struct request *req)
1622 {
1623 req->rq_flags |= RQF_TIMED_OUT;
1624 if (req->q->mq_ops->timeout) {
1625 enum blk_eh_timer_return ret;
1626
1627 ret = req->q->mq_ops->timeout(req);
1628 if (ret == BLK_EH_DONE)
1629 return;
1630 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1631 }
1632
1633 blk_add_timer(req);
1634 }
1635
1636 struct blk_expired_data {
1637 bool has_timedout_rq;
1638 unsigned long next;
1639 unsigned long timeout_start;
1640 };
1641
blk_mq_req_expired(struct request * rq,struct blk_expired_data * expired)1642 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1643 {
1644 unsigned long deadline;
1645
1646 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1647 return false;
1648 if (rq->rq_flags & RQF_TIMED_OUT)
1649 return false;
1650
1651 deadline = READ_ONCE(rq->deadline);
1652 if (time_after_eq(expired->timeout_start, deadline))
1653 return true;
1654
1655 if (expired->next == 0)
1656 expired->next = deadline;
1657 else if (time_after(expired->next, deadline))
1658 expired->next = deadline;
1659 return false;
1660 }
1661
blk_mq_put_rq_ref(struct request * rq)1662 void blk_mq_put_rq_ref(struct request *rq)
1663 {
1664 if (is_flush_rq(rq)) {
1665 if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1666 blk_mq_free_request(rq);
1667 } else if (req_ref_put_and_test(rq)) {
1668 __blk_mq_free_request(rq);
1669 }
1670 }
1671
blk_mq_check_expired(struct request * rq,void * priv)1672 static bool blk_mq_check_expired(struct request *rq, void *priv)
1673 {
1674 struct blk_expired_data *expired = priv;
1675
1676 /*
1677 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1678 * be reallocated underneath the timeout handler's processing, then
1679 * the expire check is reliable. If the request is not expired, then
1680 * it was completed and reallocated as a new request after returning
1681 * from blk_mq_check_expired().
1682 */
1683 if (blk_mq_req_expired(rq, expired)) {
1684 expired->has_timedout_rq = true;
1685 return false;
1686 }
1687 return true;
1688 }
1689
blk_mq_handle_expired(struct request * rq,void * priv)1690 static bool blk_mq_handle_expired(struct request *rq, void *priv)
1691 {
1692 struct blk_expired_data *expired = priv;
1693
1694 if (blk_mq_req_expired(rq, expired))
1695 blk_mq_rq_timed_out(rq);
1696 return true;
1697 }
1698
blk_mq_timeout_work(struct work_struct * work)1699 static void blk_mq_timeout_work(struct work_struct *work)
1700 {
1701 struct request_queue *q =
1702 container_of(work, struct request_queue, timeout_work);
1703 struct blk_expired_data expired = {
1704 .timeout_start = jiffies,
1705 };
1706 struct blk_mq_hw_ctx *hctx;
1707 unsigned long i;
1708
1709 /* A deadlock might occur if a request is stuck requiring a
1710 * timeout at the same time a queue freeze is waiting
1711 * completion, since the timeout code would not be able to
1712 * acquire the queue reference here.
1713 *
1714 * That's why we don't use blk_queue_enter here; instead, we use
1715 * percpu_ref_tryget directly, because we need to be able to
1716 * obtain a reference even in the short window between the queue
1717 * starting to freeze, by dropping the first reference in
1718 * blk_freeze_queue_start, and the moment the last request is
1719 * consumed, marked by the instant q_usage_counter reaches
1720 * zero.
1721 */
1722 if (!percpu_ref_tryget(&q->q_usage_counter))
1723 return;
1724
1725 /* check if there is any timed-out request */
1726 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1727 if (expired.has_timedout_rq) {
1728 /*
1729 * Before walking tags, we must ensure any submit started
1730 * before the current time has finished. Since the submit
1731 * uses srcu or rcu, wait for a synchronization point to
1732 * ensure all running submits have finished
1733 */
1734 blk_mq_wait_quiesce_done(q->tag_set);
1735
1736 expired.next = 0;
1737 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1738 }
1739
1740 if (expired.next != 0) {
1741 mod_timer(&q->timeout, expired.next);
1742 } else {
1743 /*
1744 * Request timeouts are handled as a forward rolling timer. If
1745 * we end up here it means that no requests are pending and
1746 * also that no request has been pending for a while. Mark
1747 * each hctx as idle.
1748 */
1749 queue_for_each_hw_ctx(q, hctx, i) {
1750 /* the hctx may be unmapped, so check it here */
1751 if (blk_mq_hw_queue_mapped(hctx))
1752 blk_mq_tag_idle(hctx);
1753 }
1754 }
1755 blk_queue_exit(q);
1756 }
1757
1758 struct flush_busy_ctx_data {
1759 struct blk_mq_hw_ctx *hctx;
1760 struct list_head *list;
1761 };
1762
flush_busy_ctx(struct sbitmap * sb,unsigned int bitnr,void * data)1763 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1764 {
1765 struct flush_busy_ctx_data *flush_data = data;
1766 struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1767 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1768 enum hctx_type type = hctx->type;
1769
1770 spin_lock(&ctx->lock);
1771 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1772 sbitmap_clear_bit(sb, bitnr);
1773 spin_unlock(&ctx->lock);
1774 return true;
1775 }
1776
1777 /*
1778 * Process software queues that have been marked busy, splicing them
1779 * to the for-dispatch
1780 */
blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx * hctx,struct list_head * list)1781 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1782 {
1783 struct flush_busy_ctx_data data = {
1784 .hctx = hctx,
1785 .list = list,
1786 };
1787
1788 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1789 }
1790
1791 struct dispatch_rq_data {
1792 struct blk_mq_hw_ctx *hctx;
1793 struct request *rq;
1794 };
1795
dispatch_rq_from_ctx(struct sbitmap * sb,unsigned int bitnr,void * data)1796 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1797 void *data)
1798 {
1799 struct dispatch_rq_data *dispatch_data = data;
1800 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1801 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1802 enum hctx_type type = hctx->type;
1803
1804 spin_lock(&ctx->lock);
1805 if (!list_empty(&ctx->rq_lists[type])) {
1806 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1807 list_del_init(&dispatch_data->rq->queuelist);
1808 if (list_empty(&ctx->rq_lists[type]))
1809 sbitmap_clear_bit(sb, bitnr);
1810 }
1811 spin_unlock(&ctx->lock);
1812
1813 return !dispatch_data->rq;
1814 }
1815
blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx * hctx,struct blk_mq_ctx * start)1816 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1817 struct blk_mq_ctx *start)
1818 {
1819 unsigned off = start ? start->index_hw[hctx->type] : 0;
1820 struct dispatch_rq_data data = {
1821 .hctx = hctx,
1822 .rq = NULL,
1823 };
1824
1825 __sbitmap_for_each_set(&hctx->ctx_map, off,
1826 dispatch_rq_from_ctx, &data);
1827
1828 return data.rq;
1829 }
1830
__blk_mq_alloc_driver_tag(struct request * rq)1831 bool __blk_mq_alloc_driver_tag(struct request *rq)
1832 {
1833 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1834 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1835 int tag;
1836
1837 blk_mq_tag_busy(rq->mq_hctx);
1838
1839 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1840 bt = &rq->mq_hctx->tags->breserved_tags;
1841 tag_offset = 0;
1842 } else {
1843 if (!hctx_may_queue(rq->mq_hctx, bt))
1844 return false;
1845 }
1846
1847 tag = __sbitmap_queue_get(bt);
1848 if (tag == BLK_MQ_NO_TAG)
1849 return false;
1850
1851 rq->tag = tag + tag_offset;
1852 blk_mq_inc_active_requests(rq->mq_hctx);
1853 return true;
1854 }
1855
blk_mq_dispatch_wake(wait_queue_entry_t * wait,unsigned mode,int flags,void * key)1856 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1857 int flags, void *key)
1858 {
1859 struct blk_mq_hw_ctx *hctx;
1860
1861 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1862
1863 spin_lock(&hctx->dispatch_wait_lock);
1864 if (!list_empty(&wait->entry)) {
1865 struct sbitmap_queue *sbq;
1866
1867 list_del_init(&wait->entry);
1868 sbq = &hctx->tags->bitmap_tags;
1869 atomic_dec(&sbq->ws_active);
1870 }
1871 spin_unlock(&hctx->dispatch_wait_lock);
1872
1873 blk_mq_run_hw_queue(hctx, true);
1874 return 1;
1875 }
1876
1877 /*
1878 * Mark us waiting for a tag. For shared tags, this involves hooking us into
1879 * the tag wakeups. For non-shared tags, we can simply mark us needing a
1880 * restart. For both cases, take care to check the condition again after
1881 * marking us as waiting.
1882 */
blk_mq_mark_tag_wait(struct blk_mq_hw_ctx * hctx,struct request * rq)1883 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1884 struct request *rq)
1885 {
1886 struct sbitmap_queue *sbq;
1887 struct wait_queue_head *wq;
1888 wait_queue_entry_t *wait;
1889 bool ret;
1890
1891 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1892 !(blk_mq_is_shared_tags(hctx->flags))) {
1893 blk_mq_sched_mark_restart_hctx(hctx);
1894
1895 /*
1896 * It's possible that a tag was freed in the window between the
1897 * allocation failure and adding the hardware queue to the wait
1898 * queue.
1899 *
1900 * Don't clear RESTART here, someone else could have set it.
1901 * At most this will cost an extra queue run.
1902 */
1903 return blk_mq_get_driver_tag(rq);
1904 }
1905
1906 wait = &hctx->dispatch_wait;
1907 if (!list_empty_careful(&wait->entry))
1908 return false;
1909
1910 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1911 sbq = &hctx->tags->breserved_tags;
1912 else
1913 sbq = &hctx->tags->bitmap_tags;
1914 wq = &bt_wait_ptr(sbq, hctx)->wait;
1915
1916 spin_lock_irq(&wq->lock);
1917 spin_lock(&hctx->dispatch_wait_lock);
1918 if (!list_empty(&wait->entry)) {
1919 spin_unlock(&hctx->dispatch_wait_lock);
1920 spin_unlock_irq(&wq->lock);
1921 return false;
1922 }
1923
1924 atomic_inc(&sbq->ws_active);
1925 wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1926 __add_wait_queue(wq, wait);
1927
1928 /*
1929 * Add one explicit barrier since blk_mq_get_driver_tag() may
1930 * not imply barrier in case of failure.
1931 *
1932 * Order adding us to wait queue and allocating driver tag.
1933 *
1934 * The pair is the one implied in sbitmap_queue_wake_up() which
1935 * orders clearing sbitmap tag bits and waitqueue_active() in
1936 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless
1937 *
1938 * Otherwise, re-order of adding wait queue and getting driver tag
1939 * may cause __sbitmap_queue_wake_up() to wake up nothing because
1940 * the waitqueue_active() may not observe us in wait queue.
1941 */
1942 smp_mb();
1943
1944 /*
1945 * It's possible that a tag was freed in the window between the
1946 * allocation failure and adding the hardware queue to the wait
1947 * queue.
1948 */
1949 ret = blk_mq_get_driver_tag(rq);
1950 if (!ret) {
1951 spin_unlock(&hctx->dispatch_wait_lock);
1952 spin_unlock_irq(&wq->lock);
1953 return false;
1954 }
1955
1956 /*
1957 * We got a tag, remove ourselves from the wait queue to ensure
1958 * someone else gets the wakeup.
1959 */
1960 list_del_init(&wait->entry);
1961 atomic_dec(&sbq->ws_active);
1962 spin_unlock(&hctx->dispatch_wait_lock);
1963 spin_unlock_irq(&wq->lock);
1964
1965 return true;
1966 }
1967
1968 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1969 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1970 /*
1971 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1972 * - EWMA is one simple way to compute running average value
1973 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1974 * - take 4 as factor for avoiding to get too small(0) result, and this
1975 * factor doesn't matter because EWMA decreases exponentially
1976 */
blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx * hctx,bool busy)1977 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1978 {
1979 unsigned int ewma;
1980
1981 ewma = hctx->dispatch_busy;
1982
1983 if (!ewma && !busy)
1984 return;
1985
1986 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1987 if (busy)
1988 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1989 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1990
1991 hctx->dispatch_busy = ewma;
1992 }
1993
1994 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1995
blk_mq_handle_dev_resource(struct request * rq,struct list_head * list)1996 static void blk_mq_handle_dev_resource(struct request *rq,
1997 struct list_head *list)
1998 {
1999 list_add(&rq->queuelist, list);
2000 __blk_mq_requeue_request(rq);
2001 }
2002
2003 enum prep_dispatch {
2004 PREP_DISPATCH_OK,
2005 PREP_DISPATCH_NO_TAG,
2006 PREP_DISPATCH_NO_BUDGET,
2007 };
2008
blk_mq_prep_dispatch_rq(struct request * rq,bool need_budget)2009 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
2010 bool need_budget)
2011 {
2012 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2013 int budget_token = -1;
2014
2015 if (need_budget) {
2016 budget_token = blk_mq_get_dispatch_budget(rq->q);
2017 if (budget_token < 0) {
2018 blk_mq_put_driver_tag(rq);
2019 return PREP_DISPATCH_NO_BUDGET;
2020 }
2021 blk_mq_set_rq_budget_token(rq, budget_token);
2022 }
2023
2024 if (!blk_mq_get_driver_tag(rq)) {
2025 /*
2026 * The initial allocation attempt failed, so we need to
2027 * rerun the hardware queue when a tag is freed. The
2028 * waitqueue takes care of that. If the queue is run
2029 * before we add this entry back on the dispatch list,
2030 * we'll re-run it below.
2031 */
2032 if (!blk_mq_mark_tag_wait(hctx, rq)) {
2033 /*
2034 * All budgets not got from this function will be put
2035 * together during handling partial dispatch
2036 */
2037 if (need_budget)
2038 blk_mq_put_dispatch_budget(rq->q, budget_token);
2039 return PREP_DISPATCH_NO_TAG;
2040 }
2041 }
2042
2043 return PREP_DISPATCH_OK;
2044 }
2045
2046 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
blk_mq_release_budgets(struct request_queue * q,struct list_head * list)2047 static void blk_mq_release_budgets(struct request_queue *q,
2048 struct list_head *list)
2049 {
2050 struct request *rq;
2051
2052 list_for_each_entry(rq, list, queuelist) {
2053 int budget_token = blk_mq_get_rq_budget_token(rq);
2054
2055 if (budget_token >= 0)
2056 blk_mq_put_dispatch_budget(q, budget_token);
2057 }
2058 }
2059
2060 /*
2061 * blk_mq_commit_rqs will notify driver using bd->last that there is no
2062 * more requests. (See comment in struct blk_mq_ops for commit_rqs for
2063 * details)
2064 * Attention, we should explicitly call this in unusual cases:
2065 * 1) did not queue everything initially scheduled to queue
2066 * 2) the last attempt to queue a request failed
2067 */
blk_mq_commit_rqs(struct blk_mq_hw_ctx * hctx,int queued,bool from_schedule)2068 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
2069 bool from_schedule)
2070 {
2071 if (hctx->queue->mq_ops->commit_rqs && queued) {
2072 trace_block_unplug(hctx->queue, queued, !from_schedule);
2073 hctx->queue->mq_ops->commit_rqs(hctx);
2074 }
2075 }
2076
2077 /*
2078 * Returns true if we did some work AND can potentially do more.
2079 */
blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx * hctx,struct list_head * list,unsigned int nr_budgets)2080 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
2081 unsigned int nr_budgets)
2082 {
2083 enum prep_dispatch prep;
2084 struct request_queue *q = hctx->queue;
2085 struct request *rq;
2086 int queued;
2087 blk_status_t ret = BLK_STS_OK;
2088 bool needs_resource = false;
2089
2090 if (list_empty(list))
2091 return false;
2092
2093 /*
2094 * Now process all the entries, sending them to the driver.
2095 */
2096 queued = 0;
2097 do {
2098 struct blk_mq_queue_data bd;
2099
2100 rq = list_first_entry(list, struct request, queuelist);
2101
2102 WARN_ON_ONCE(hctx != rq->mq_hctx);
2103 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2104 if (prep != PREP_DISPATCH_OK)
2105 break;
2106
2107 list_del_init(&rq->queuelist);
2108
2109 bd.rq = rq;
2110 bd.last = list_empty(list);
2111
2112 /*
2113 * once the request is queued to lld, no need to cover the
2114 * budget any more
2115 */
2116 if (nr_budgets)
2117 nr_budgets--;
2118 ret = q->mq_ops->queue_rq(hctx, &bd);
2119 switch (ret) {
2120 case BLK_STS_OK:
2121 queued++;
2122 break;
2123 case BLK_STS_RESOURCE:
2124 needs_resource = true;
2125 fallthrough;
2126 case BLK_STS_DEV_RESOURCE:
2127 blk_mq_handle_dev_resource(rq, list);
2128 goto out;
2129 default:
2130 blk_mq_end_request(rq, ret);
2131 }
2132 } while (!list_empty(list));
2133 out:
2134 /* If we didn't flush the entire list, we could have told the driver
2135 * there was more coming, but that turned out to be a lie.
2136 */
2137 if (!list_empty(list) || ret != BLK_STS_OK)
2138 blk_mq_commit_rqs(hctx, queued, false);
2139
2140 /*
2141 * Any items that need requeuing? Stuff them into hctx->dispatch,
2142 * that is where we will continue on next queue run.
2143 */
2144 if (!list_empty(list)) {
2145 bool needs_restart;
2146 /* For non-shared tags, the RESTART check will suffice */
2147 bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2148 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2149 blk_mq_is_shared_tags(hctx->flags));
2150
2151 if (nr_budgets)
2152 blk_mq_release_budgets(q, list);
2153
2154 spin_lock(&hctx->lock);
2155 list_splice_tail_init(list, &hctx->dispatch);
2156 spin_unlock(&hctx->lock);
2157
2158 /*
2159 * Order adding requests to hctx->dispatch and checking
2160 * SCHED_RESTART flag. The pair of this smp_mb() is the one
2161 * in blk_mq_sched_restart(). Avoid restart code path to
2162 * miss the new added requests to hctx->dispatch, meantime
2163 * SCHED_RESTART is observed here.
2164 */
2165 smp_mb();
2166
2167 /*
2168 * If SCHED_RESTART was set by the caller of this function and
2169 * it is no longer set that means that it was cleared by another
2170 * thread and hence that a queue rerun is needed.
2171 *
2172 * If 'no_tag' is set, that means that we failed getting
2173 * a driver tag with an I/O scheduler attached. If our dispatch
2174 * waitqueue is no longer active, ensure that we run the queue
2175 * AFTER adding our entries back to the list.
2176 *
2177 * If no I/O scheduler has been configured it is possible that
2178 * the hardware queue got stopped and restarted before requests
2179 * were pushed back onto the dispatch list. Rerun the queue to
2180 * avoid starvation. Notes:
2181 * - blk_mq_run_hw_queue() checks whether or not a queue has
2182 * been stopped before rerunning a queue.
2183 * - Some but not all block drivers stop a queue before
2184 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2185 * and dm-rq.
2186 *
2187 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2188 * bit is set, run queue after a delay to avoid IO stalls
2189 * that could otherwise occur if the queue is idle. We'll do
2190 * similar if we couldn't get budget or couldn't lock a zone
2191 * and SCHED_RESTART is set.
2192 */
2193 needs_restart = blk_mq_sched_needs_restart(hctx);
2194 if (prep == PREP_DISPATCH_NO_BUDGET)
2195 needs_resource = true;
2196 if (!needs_restart ||
2197 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2198 blk_mq_run_hw_queue(hctx, true);
2199 else if (needs_resource)
2200 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2201
2202 blk_mq_update_dispatch_busy(hctx, true);
2203 return false;
2204 }
2205
2206 blk_mq_update_dispatch_busy(hctx, false);
2207 return true;
2208 }
2209
blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx * hctx)2210 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2211 {
2212 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2213
2214 if (cpu >= nr_cpu_ids)
2215 cpu = cpumask_first(hctx->cpumask);
2216 return cpu;
2217 }
2218
2219 /*
2220 * ->next_cpu is always calculated from hctx->cpumask, so simply use
2221 * it for speeding up the check
2222 */
blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx * hctx)2223 static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx)
2224 {
2225 return hctx->next_cpu >= nr_cpu_ids;
2226 }
2227
2228 /*
2229 * It'd be great if the workqueue API had a way to pass
2230 * in a mask and had some smarts for more clever placement.
2231 * For now we just round-robin here, switching for every
2232 * BLK_MQ_CPU_WORK_BATCH queued items.
2233 */
blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx * hctx)2234 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2235 {
2236 bool tried = false;
2237 int next_cpu = hctx->next_cpu;
2238
2239 /* Switch to unbound if no allowable CPUs in this hctx */
2240 if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx))
2241 return WORK_CPU_UNBOUND;
2242
2243 if (--hctx->next_cpu_batch <= 0) {
2244 select_cpu:
2245 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2246 cpu_online_mask);
2247 if (next_cpu >= nr_cpu_ids)
2248 next_cpu = blk_mq_first_mapped_cpu(hctx);
2249 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2250 }
2251
2252 /*
2253 * Do unbound schedule if we can't find a online CPU for this hctx,
2254 * and it should only happen in the path of handling CPU DEAD.
2255 */
2256 if (!cpu_online(next_cpu)) {
2257 if (!tried) {
2258 tried = true;
2259 goto select_cpu;
2260 }
2261
2262 /*
2263 * Make sure to re-select CPU next time once after CPUs
2264 * in hctx->cpumask become online again.
2265 */
2266 hctx->next_cpu = next_cpu;
2267 hctx->next_cpu_batch = 1;
2268 return WORK_CPU_UNBOUND;
2269 }
2270
2271 hctx->next_cpu = next_cpu;
2272 return next_cpu;
2273 }
2274
2275 /**
2276 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2277 * @hctx: Pointer to the hardware queue to run.
2278 * @msecs: Milliseconds of delay to wait before running the queue.
2279 *
2280 * Run a hardware queue asynchronously with a delay of @msecs.
2281 */
blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx * hctx,unsigned long msecs)2282 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2283 {
2284 if (unlikely(blk_mq_hctx_stopped(hctx)))
2285 return;
2286 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2287 msecs_to_jiffies(msecs));
2288 }
2289 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2290
blk_mq_hw_queue_need_run(struct blk_mq_hw_ctx * hctx)2291 static inline bool blk_mq_hw_queue_need_run(struct blk_mq_hw_ctx *hctx)
2292 {
2293 bool need_run;
2294
2295 /*
2296 * When queue is quiesced, we may be switching io scheduler, or
2297 * updating nr_hw_queues, or other things, and we can't run queue
2298 * any more, even blk_mq_hctx_has_pending() can't be called safely.
2299 *
2300 * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2301 * quiesced.
2302 */
2303 __blk_mq_run_dispatch_ops(hctx->queue, false,
2304 need_run = !blk_queue_quiesced(hctx->queue) &&
2305 blk_mq_hctx_has_pending(hctx));
2306 return need_run;
2307 }
2308
2309 /**
2310 * blk_mq_run_hw_queue - Start to run a hardware queue.
2311 * @hctx: Pointer to the hardware queue to run.
2312 * @async: If we want to run the queue asynchronously.
2313 *
2314 * Check if the request queue is not in a quiesced state and if there are
2315 * pending requests to be sent. If this is true, run the queue to send requests
2316 * to hardware.
2317 */
blk_mq_run_hw_queue(struct blk_mq_hw_ctx * hctx,bool async)2318 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2319 {
2320 bool need_run;
2321
2322 /*
2323 * We can't run the queue inline with interrupts disabled.
2324 */
2325 WARN_ON_ONCE(!async && in_interrupt());
2326
2327 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
2328
2329 need_run = blk_mq_hw_queue_need_run(hctx);
2330 if (!need_run) {
2331 unsigned long flags;
2332
2333 /*
2334 * Synchronize with blk_mq_unquiesce_queue(), because we check
2335 * if hw queue is quiesced locklessly above, we need the use
2336 * ->queue_lock to make sure we see the up-to-date status to
2337 * not miss rerunning the hw queue.
2338 */
2339 spin_lock_irqsave(&hctx->queue->queue_lock, flags);
2340 need_run = blk_mq_hw_queue_need_run(hctx);
2341 spin_unlock_irqrestore(&hctx->queue->queue_lock, flags);
2342
2343 if (!need_run)
2344 return;
2345 }
2346
2347 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2348 blk_mq_delay_run_hw_queue(hctx, 0);
2349 return;
2350 }
2351
2352 blk_mq_run_dispatch_ops(hctx->queue,
2353 blk_mq_sched_dispatch_requests(hctx));
2354 }
2355 EXPORT_SYMBOL(blk_mq_run_hw_queue);
2356
2357 /*
2358 * Return prefered queue to dispatch from (if any) for non-mq aware IO
2359 * scheduler.
2360 */
blk_mq_get_sq_hctx(struct request_queue * q)2361 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2362 {
2363 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2364 /*
2365 * If the IO scheduler does not respect hardware queues when
2366 * dispatching, we just don't bother with multiple HW queues and
2367 * dispatch from hctx for the current CPU since running multiple queues
2368 * just causes lock contention inside the scheduler and pointless cache
2369 * bouncing.
2370 */
2371 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2372
2373 if (!blk_mq_hctx_stopped(hctx))
2374 return hctx;
2375 return NULL;
2376 }
2377
2378 /**
2379 * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2380 * @q: Pointer to the request queue to run.
2381 * @async: If we want to run the queue asynchronously.
2382 */
blk_mq_run_hw_queues(struct request_queue * q,bool async)2383 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2384 {
2385 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2386 unsigned long i;
2387
2388 sq_hctx = NULL;
2389 if (blk_queue_sq_sched(q))
2390 sq_hctx = blk_mq_get_sq_hctx(q);
2391 queue_for_each_hw_ctx(q, hctx, i) {
2392 if (blk_mq_hctx_stopped(hctx))
2393 continue;
2394 /*
2395 * Dispatch from this hctx either if there's no hctx preferred
2396 * by IO scheduler or if it has requests that bypass the
2397 * scheduler.
2398 */
2399 if (!sq_hctx || sq_hctx == hctx ||
2400 !list_empty_careful(&hctx->dispatch))
2401 blk_mq_run_hw_queue(hctx, async);
2402 }
2403 }
2404 EXPORT_SYMBOL(blk_mq_run_hw_queues);
2405
2406 /**
2407 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2408 * @q: Pointer to the request queue to run.
2409 * @msecs: Milliseconds of delay to wait before running the queues.
2410 */
blk_mq_delay_run_hw_queues(struct request_queue * q,unsigned long msecs)2411 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2412 {
2413 struct blk_mq_hw_ctx *hctx, *sq_hctx;
2414 unsigned long i;
2415
2416 sq_hctx = NULL;
2417 if (blk_queue_sq_sched(q))
2418 sq_hctx = blk_mq_get_sq_hctx(q);
2419 queue_for_each_hw_ctx(q, hctx, i) {
2420 if (blk_mq_hctx_stopped(hctx))
2421 continue;
2422 /*
2423 * If there is already a run_work pending, leave the
2424 * pending delay untouched. Otherwise, a hctx can stall
2425 * if another hctx is re-delaying the other's work
2426 * before the work executes.
2427 */
2428 if (delayed_work_pending(&hctx->run_work))
2429 continue;
2430 /*
2431 * Dispatch from this hctx either if there's no hctx preferred
2432 * by IO scheduler or if it has requests that bypass the
2433 * scheduler.
2434 */
2435 if (!sq_hctx || sq_hctx == hctx ||
2436 !list_empty_careful(&hctx->dispatch))
2437 blk_mq_delay_run_hw_queue(hctx, msecs);
2438 }
2439 }
2440 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2441
2442 /*
2443 * This function is often used for pausing .queue_rq() by driver when
2444 * there isn't enough resource or some conditions aren't satisfied, and
2445 * BLK_STS_RESOURCE is usually returned.
2446 *
2447 * We do not guarantee that dispatch can be drained or blocked
2448 * after blk_mq_stop_hw_queue() returns. Please use
2449 * blk_mq_quiesce_queue() for that requirement.
2450 */
blk_mq_stop_hw_queue(struct blk_mq_hw_ctx * hctx)2451 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2452 {
2453 cancel_delayed_work(&hctx->run_work);
2454
2455 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2456 }
2457 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2458
2459 /*
2460 * This function is often used for pausing .queue_rq() by driver when
2461 * there isn't enough resource or some conditions aren't satisfied, and
2462 * BLK_STS_RESOURCE is usually returned.
2463 *
2464 * We do not guarantee that dispatch can be drained or blocked
2465 * after blk_mq_stop_hw_queues() returns. Please use
2466 * blk_mq_quiesce_queue() for that requirement.
2467 */
blk_mq_stop_hw_queues(struct request_queue * q)2468 void blk_mq_stop_hw_queues(struct request_queue *q)
2469 {
2470 struct blk_mq_hw_ctx *hctx;
2471 unsigned long i;
2472
2473 queue_for_each_hw_ctx(q, hctx, i)
2474 blk_mq_stop_hw_queue(hctx);
2475 }
2476 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2477
blk_mq_start_hw_queue(struct blk_mq_hw_ctx * hctx)2478 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2479 {
2480 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2481
2482 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING);
2483 }
2484 EXPORT_SYMBOL(blk_mq_start_hw_queue);
2485
blk_mq_start_hw_queues(struct request_queue * q)2486 void blk_mq_start_hw_queues(struct request_queue *q)
2487 {
2488 struct blk_mq_hw_ctx *hctx;
2489 unsigned long i;
2490
2491 queue_for_each_hw_ctx(q, hctx, i)
2492 blk_mq_start_hw_queue(hctx);
2493 }
2494 EXPORT_SYMBOL(blk_mq_start_hw_queues);
2495
blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx * hctx,bool async)2496 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2497 {
2498 if (!blk_mq_hctx_stopped(hctx))
2499 return;
2500
2501 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2502 /*
2503 * Pairs with the smp_mb() in blk_mq_hctx_stopped() to order the
2504 * clearing of BLK_MQ_S_STOPPED above and the checking of dispatch
2505 * list in the subsequent routine.
2506 */
2507 smp_mb__after_atomic();
2508 blk_mq_run_hw_queue(hctx, async);
2509 }
2510 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2511
blk_mq_start_stopped_hw_queues(struct request_queue * q,bool async)2512 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2513 {
2514 struct blk_mq_hw_ctx *hctx;
2515 unsigned long i;
2516
2517 queue_for_each_hw_ctx(q, hctx, i)
2518 blk_mq_start_stopped_hw_queue(hctx, async ||
2519 (hctx->flags & BLK_MQ_F_BLOCKING));
2520 }
2521 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2522
blk_mq_run_work_fn(struct work_struct * work)2523 static void blk_mq_run_work_fn(struct work_struct *work)
2524 {
2525 struct blk_mq_hw_ctx *hctx =
2526 container_of(work, struct blk_mq_hw_ctx, run_work.work);
2527
2528 blk_mq_run_dispatch_ops(hctx->queue,
2529 blk_mq_sched_dispatch_requests(hctx));
2530 }
2531
2532 /**
2533 * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2534 * @rq: Pointer to request to be inserted.
2535 * @flags: BLK_MQ_INSERT_*
2536 *
2537 * Should only be used carefully, when the caller knows we want to
2538 * bypass a potential IO scheduler on the target device.
2539 */
blk_mq_request_bypass_insert(struct request * rq,blk_insert_t flags)2540 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2541 {
2542 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2543
2544 spin_lock(&hctx->lock);
2545 if (flags & BLK_MQ_INSERT_AT_HEAD)
2546 list_add(&rq->queuelist, &hctx->dispatch);
2547 else
2548 list_add_tail(&rq->queuelist, &hctx->dispatch);
2549 spin_unlock(&hctx->lock);
2550 }
2551
blk_mq_insert_requests(struct blk_mq_hw_ctx * hctx,struct blk_mq_ctx * ctx,struct list_head * list,bool run_queue_async)2552 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2553 struct blk_mq_ctx *ctx, struct list_head *list,
2554 bool run_queue_async)
2555 {
2556 struct request *rq;
2557 enum hctx_type type = hctx->type;
2558
2559 /*
2560 * Try to issue requests directly if the hw queue isn't busy to save an
2561 * extra enqueue & dequeue to the sw queue.
2562 */
2563 if (!hctx->dispatch_busy && !run_queue_async) {
2564 blk_mq_run_dispatch_ops(hctx->queue,
2565 blk_mq_try_issue_list_directly(hctx, list));
2566 if (list_empty(list))
2567 goto out;
2568 }
2569
2570 /*
2571 * preemption doesn't flush plug list, so it's possible ctx->cpu is
2572 * offline now
2573 */
2574 list_for_each_entry(rq, list, queuelist) {
2575 BUG_ON(rq->mq_ctx != ctx);
2576 trace_block_rq_insert(rq);
2577 if (rq->cmd_flags & REQ_NOWAIT)
2578 run_queue_async = true;
2579 }
2580
2581 spin_lock(&ctx->lock);
2582 list_splice_tail_init(list, &ctx->rq_lists[type]);
2583 blk_mq_hctx_mark_pending(hctx, ctx);
2584 spin_unlock(&ctx->lock);
2585 out:
2586 blk_mq_run_hw_queue(hctx, run_queue_async);
2587 }
2588
blk_mq_insert_request(struct request * rq,blk_insert_t flags)2589 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2590 {
2591 struct request_queue *q = rq->q;
2592 struct blk_mq_ctx *ctx = rq->mq_ctx;
2593 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2594
2595 if (blk_rq_is_passthrough(rq)) {
2596 /*
2597 * Passthrough request have to be added to hctx->dispatch
2598 * directly. The device may be in a situation where it can't
2599 * handle FS request, and always returns BLK_STS_RESOURCE for
2600 * them, which gets them added to hctx->dispatch.
2601 *
2602 * If a passthrough request is required to unblock the queues,
2603 * and it is added to the scheduler queue, there is no chance to
2604 * dispatch it given we prioritize requests in hctx->dispatch.
2605 */
2606 blk_mq_request_bypass_insert(rq, flags);
2607 } else if (req_op(rq) == REQ_OP_FLUSH) {
2608 /*
2609 * Firstly normal IO request is inserted to scheduler queue or
2610 * sw queue, meantime we add flush request to dispatch queue(
2611 * hctx->dispatch) directly and there is at most one in-flight
2612 * flush request for each hw queue, so it doesn't matter to add
2613 * flush request to tail or front of the dispatch queue.
2614 *
2615 * Secondly in case of NCQ, flush request belongs to non-NCQ
2616 * command, and queueing it will fail when there is any
2617 * in-flight normal IO request(NCQ command). When adding flush
2618 * rq to the front of hctx->dispatch, it is easier to introduce
2619 * extra time to flush rq's latency because of S_SCHED_RESTART
2620 * compared with adding to the tail of dispatch queue, then
2621 * chance of flush merge is increased, and less flush requests
2622 * will be issued to controller. It is observed that ~10% time
2623 * is saved in blktests block/004 on disk attached to AHCI/NCQ
2624 * drive when adding flush rq to the front of hctx->dispatch.
2625 *
2626 * Simply queue flush rq to the front of hctx->dispatch so that
2627 * intensive flush workloads can benefit in case of NCQ HW.
2628 */
2629 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2630 } else if (q->elevator) {
2631 LIST_HEAD(list);
2632
2633 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2634
2635 list_add(&rq->queuelist, &list);
2636 q->elevator->type->ops.insert_requests(hctx, &list, flags);
2637 } else {
2638 trace_block_rq_insert(rq);
2639
2640 spin_lock(&ctx->lock);
2641 if (flags & BLK_MQ_INSERT_AT_HEAD)
2642 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2643 else
2644 list_add_tail(&rq->queuelist,
2645 &ctx->rq_lists[hctx->type]);
2646 blk_mq_hctx_mark_pending(hctx, ctx);
2647 spin_unlock(&ctx->lock);
2648 }
2649 }
2650
blk_mq_bio_to_request(struct request * rq,struct bio * bio,unsigned int nr_segs)2651 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2652 unsigned int nr_segs)
2653 {
2654 int err;
2655
2656 if (bio->bi_opf & REQ_RAHEAD)
2657 rq->cmd_flags |= REQ_FAILFAST_MASK;
2658
2659 rq->__sector = bio->bi_iter.bi_sector;
2660 blk_rq_bio_prep(rq, bio, nr_segs);
2661 if (bio_integrity(bio))
2662 rq->nr_integrity_segments = blk_rq_count_integrity_sg(rq->q,
2663 bio);
2664
2665 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2666 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2667 WARN_ON_ONCE(err);
2668
2669 blk_account_io_start(rq);
2670 }
2671
__blk_mq_issue_directly(struct blk_mq_hw_ctx * hctx,struct request * rq,bool last)2672 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2673 struct request *rq, bool last)
2674 {
2675 struct request_queue *q = rq->q;
2676 struct blk_mq_queue_data bd = {
2677 .rq = rq,
2678 .last = last,
2679 };
2680 blk_status_t ret;
2681
2682 /*
2683 * For OK queue, we are done. For error, caller may kill it.
2684 * Any other error (busy), just add it to our list as we
2685 * previously would have done.
2686 */
2687 ret = q->mq_ops->queue_rq(hctx, &bd);
2688 switch (ret) {
2689 case BLK_STS_OK:
2690 blk_mq_update_dispatch_busy(hctx, false);
2691 break;
2692 case BLK_STS_RESOURCE:
2693 case BLK_STS_DEV_RESOURCE:
2694 blk_mq_update_dispatch_busy(hctx, true);
2695 __blk_mq_requeue_request(rq);
2696 break;
2697 default:
2698 blk_mq_update_dispatch_busy(hctx, false);
2699 break;
2700 }
2701
2702 return ret;
2703 }
2704
blk_mq_get_budget_and_tag(struct request * rq)2705 static bool blk_mq_get_budget_and_tag(struct request *rq)
2706 {
2707 int budget_token;
2708
2709 budget_token = blk_mq_get_dispatch_budget(rq->q);
2710 if (budget_token < 0)
2711 return false;
2712 blk_mq_set_rq_budget_token(rq, budget_token);
2713 if (!blk_mq_get_driver_tag(rq)) {
2714 blk_mq_put_dispatch_budget(rq->q, budget_token);
2715 return false;
2716 }
2717 return true;
2718 }
2719
2720 /**
2721 * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2722 * @hctx: Pointer of the associated hardware queue.
2723 * @rq: Pointer to request to be sent.
2724 *
2725 * If the device has enough resources to accept a new request now, send the
2726 * request directly to device driver. Else, insert at hctx->dispatch queue, so
2727 * we can try send it another time in the future. Requests inserted at this
2728 * queue have higher priority.
2729 */
blk_mq_try_issue_directly(struct blk_mq_hw_ctx * hctx,struct request * rq)2730 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2731 struct request *rq)
2732 {
2733 blk_status_t ret;
2734
2735 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2736 blk_mq_insert_request(rq, 0);
2737 blk_mq_run_hw_queue(hctx, false);
2738 return;
2739 }
2740
2741 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2742 blk_mq_insert_request(rq, 0);
2743 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT);
2744 return;
2745 }
2746
2747 ret = __blk_mq_issue_directly(hctx, rq, true);
2748 switch (ret) {
2749 case BLK_STS_OK:
2750 break;
2751 case BLK_STS_RESOURCE:
2752 case BLK_STS_DEV_RESOURCE:
2753 blk_mq_request_bypass_insert(rq, 0);
2754 blk_mq_run_hw_queue(hctx, false);
2755 break;
2756 default:
2757 blk_mq_end_request(rq, ret);
2758 break;
2759 }
2760 }
2761
blk_mq_request_issue_directly(struct request * rq,bool last)2762 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2763 {
2764 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2765
2766 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2767 blk_mq_insert_request(rq, 0);
2768 blk_mq_run_hw_queue(hctx, false);
2769 return BLK_STS_OK;
2770 }
2771
2772 if (!blk_mq_get_budget_and_tag(rq))
2773 return BLK_STS_RESOURCE;
2774 return __blk_mq_issue_directly(hctx, rq, last);
2775 }
2776
blk_mq_plug_issue_direct(struct blk_plug * plug)2777 static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2778 {
2779 struct blk_mq_hw_ctx *hctx = NULL;
2780 struct request *rq;
2781 int queued = 0;
2782 blk_status_t ret = BLK_STS_OK;
2783
2784 while ((rq = rq_list_pop(&plug->mq_list))) {
2785 bool last = rq_list_empty(&plug->mq_list);
2786
2787 if (hctx != rq->mq_hctx) {
2788 if (hctx) {
2789 blk_mq_commit_rqs(hctx, queued, false);
2790 queued = 0;
2791 }
2792 hctx = rq->mq_hctx;
2793 }
2794
2795 ret = blk_mq_request_issue_directly(rq, last);
2796 switch (ret) {
2797 case BLK_STS_OK:
2798 queued++;
2799 break;
2800 case BLK_STS_RESOURCE:
2801 case BLK_STS_DEV_RESOURCE:
2802 blk_mq_request_bypass_insert(rq, 0);
2803 blk_mq_run_hw_queue(hctx, false);
2804 goto out;
2805 default:
2806 blk_mq_end_request(rq, ret);
2807 break;
2808 }
2809 }
2810
2811 out:
2812 if (ret != BLK_STS_OK)
2813 blk_mq_commit_rqs(hctx, queued, false);
2814 }
2815
__blk_mq_flush_plug_list(struct request_queue * q,struct blk_plug * plug)2816 static void __blk_mq_flush_plug_list(struct request_queue *q,
2817 struct blk_plug *plug)
2818 {
2819 if (blk_queue_quiesced(q))
2820 return;
2821 q->mq_ops->queue_rqs(&plug->mq_list);
2822 }
2823
blk_mq_dispatch_plug_list(struct blk_plug * plug,bool from_sched)2824 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2825 {
2826 struct blk_mq_hw_ctx *this_hctx = NULL;
2827 struct blk_mq_ctx *this_ctx = NULL;
2828 struct rq_list requeue_list = {};
2829 unsigned int depth = 0;
2830 bool is_passthrough = false;
2831 LIST_HEAD(list);
2832
2833 do {
2834 struct request *rq = rq_list_pop(&plug->mq_list);
2835
2836 if (!this_hctx) {
2837 this_hctx = rq->mq_hctx;
2838 this_ctx = rq->mq_ctx;
2839 is_passthrough = blk_rq_is_passthrough(rq);
2840 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2841 is_passthrough != blk_rq_is_passthrough(rq)) {
2842 rq_list_add_tail(&requeue_list, rq);
2843 continue;
2844 }
2845 list_add_tail(&rq->queuelist, &list);
2846 depth++;
2847 } while (!rq_list_empty(&plug->mq_list));
2848
2849 plug->mq_list = requeue_list;
2850 trace_block_unplug(this_hctx->queue, depth, !from_sched);
2851
2852 percpu_ref_get(&this_hctx->queue->q_usage_counter);
2853 /* passthrough requests should never be issued to the I/O scheduler */
2854 if (is_passthrough) {
2855 spin_lock(&this_hctx->lock);
2856 list_splice_tail_init(&list, &this_hctx->dispatch);
2857 spin_unlock(&this_hctx->lock);
2858 blk_mq_run_hw_queue(this_hctx, from_sched);
2859 } else if (this_hctx->queue->elevator) {
2860 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2861 &list, 0);
2862 blk_mq_run_hw_queue(this_hctx, from_sched);
2863 } else {
2864 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2865 }
2866 percpu_ref_put(&this_hctx->queue->q_usage_counter);
2867 }
2868
blk_mq_flush_plug_list(struct blk_plug * plug,bool from_schedule)2869 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2870 {
2871 struct request *rq;
2872 unsigned int depth;
2873
2874 /*
2875 * We may have been called recursively midway through handling
2876 * plug->mq_list via a schedule() in the driver's queue_rq() callback.
2877 * To avoid mq_list changing under our feet, clear rq_count early and
2878 * bail out specifically if rq_count is 0 rather than checking
2879 * whether the mq_list is empty.
2880 */
2881 if (plug->rq_count == 0)
2882 return;
2883 depth = plug->rq_count;
2884 plug->rq_count = 0;
2885
2886 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2887 struct request_queue *q;
2888
2889 rq = rq_list_peek(&plug->mq_list);
2890 q = rq->q;
2891 trace_block_unplug(q, depth, true);
2892
2893 /*
2894 * Peek first request and see if we have a ->queue_rqs() hook.
2895 * If we do, we can dispatch the whole plug list in one go. We
2896 * already know at this point that all requests belong to the
2897 * same queue, caller must ensure that's the case.
2898 */
2899 if (q->mq_ops->queue_rqs) {
2900 blk_mq_run_dispatch_ops(q,
2901 __blk_mq_flush_plug_list(q, plug));
2902 if (rq_list_empty(&plug->mq_list))
2903 return;
2904 }
2905
2906 blk_mq_run_dispatch_ops(q,
2907 blk_mq_plug_issue_direct(plug));
2908 if (rq_list_empty(&plug->mq_list))
2909 return;
2910 }
2911
2912 do {
2913 blk_mq_dispatch_plug_list(plug, from_schedule);
2914 } while (!rq_list_empty(&plug->mq_list));
2915 }
2916
blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx * hctx,struct list_head * list)2917 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2918 struct list_head *list)
2919 {
2920 int queued = 0;
2921 blk_status_t ret = BLK_STS_OK;
2922
2923 while (!list_empty(list)) {
2924 struct request *rq = list_first_entry(list, struct request,
2925 queuelist);
2926
2927 list_del_init(&rq->queuelist);
2928 ret = blk_mq_request_issue_directly(rq, list_empty(list));
2929 switch (ret) {
2930 case BLK_STS_OK:
2931 queued++;
2932 break;
2933 case BLK_STS_RESOURCE:
2934 case BLK_STS_DEV_RESOURCE:
2935 blk_mq_request_bypass_insert(rq, 0);
2936 if (list_empty(list))
2937 blk_mq_run_hw_queue(hctx, false);
2938 goto out;
2939 default:
2940 blk_mq_end_request(rq, ret);
2941 break;
2942 }
2943 }
2944
2945 out:
2946 if (ret != BLK_STS_OK)
2947 blk_mq_commit_rqs(hctx, queued, false);
2948 }
2949
blk_mq_attempt_bio_merge(struct request_queue * q,struct bio * bio,unsigned int nr_segs)2950 static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2951 struct bio *bio, unsigned int nr_segs)
2952 {
2953 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2954 if (blk_attempt_plug_merge(q, bio, nr_segs))
2955 return true;
2956 if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2957 return true;
2958 }
2959 return false;
2960 }
2961
blk_mq_get_new_requests(struct request_queue * q,struct blk_plug * plug,struct bio * bio,unsigned int nsegs)2962 static struct request *blk_mq_get_new_requests(struct request_queue *q,
2963 struct blk_plug *plug,
2964 struct bio *bio,
2965 unsigned int nsegs)
2966 {
2967 struct blk_mq_alloc_data data = {
2968 .q = q,
2969 .nr_tags = 1,
2970 .cmd_flags = bio->bi_opf,
2971 };
2972 struct request *rq;
2973
2974 rq_qos_throttle(q, bio);
2975
2976 if (plug) {
2977 data.nr_tags = plug->nr_ios;
2978 plug->nr_ios = 1;
2979 data.cached_rqs = &plug->cached_rqs;
2980 }
2981
2982 rq = __blk_mq_alloc_requests(&data);
2983 if (rq)
2984 return rq;
2985 rq_qos_cleanup(q, bio);
2986 if (bio->bi_opf & REQ_NOWAIT)
2987 bio_wouldblock_error(bio);
2988 return NULL;
2989 }
2990
2991 /*
2992 * Check if there is a suitable cached request and return it.
2993 */
blk_mq_peek_cached_request(struct blk_plug * plug,struct request_queue * q,blk_opf_t opf)2994 static struct request *blk_mq_peek_cached_request(struct blk_plug *plug,
2995 struct request_queue *q, blk_opf_t opf)
2996 {
2997 enum hctx_type type = blk_mq_get_hctx_type(opf);
2998 struct request *rq;
2999
3000 if (!plug)
3001 return NULL;
3002 rq = rq_list_peek(&plug->cached_rqs);
3003 if (!rq || rq->q != q)
3004 return NULL;
3005 if (type != rq->mq_hctx->type &&
3006 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT))
3007 return NULL;
3008 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
3009 return NULL;
3010 return rq;
3011 }
3012
blk_mq_use_cached_rq(struct request * rq,struct blk_plug * plug,struct bio * bio)3013 static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug,
3014 struct bio *bio)
3015 {
3016 if (rq_list_pop(&plug->cached_rqs) != rq)
3017 WARN_ON_ONCE(1);
3018
3019 /*
3020 * If any qos ->throttle() end up blocking, we will have flushed the
3021 * plug and hence killed the cached_rq list as well. Pop this entry
3022 * before we throttle.
3023 */
3024 rq_qos_throttle(rq->q, bio);
3025
3026 blk_mq_rq_time_init(rq, blk_time_get_ns());
3027 rq->cmd_flags = bio->bi_opf;
3028 INIT_LIST_HEAD(&rq->queuelist);
3029 }
3030
bio_unaligned(const struct bio * bio,struct request_queue * q)3031 static bool bio_unaligned(const struct bio *bio, struct request_queue *q)
3032 {
3033 unsigned int bs_mask = queue_logical_block_size(q) - 1;
3034
3035 /* .bi_sector of any zero sized bio need to be initialized */
3036 if ((bio->bi_iter.bi_size & bs_mask) ||
3037 ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask))
3038 return true;
3039 return false;
3040 }
3041
3042 /**
3043 * blk_mq_submit_bio - Create and send a request to block device.
3044 * @bio: Bio pointer.
3045 *
3046 * Builds up a request structure from @q and @bio and send to the device. The
3047 * request may not be queued directly to hardware if:
3048 * * This request can be merged with another one
3049 * * We want to place request at plug queue for possible future merging
3050 * * There is an IO scheduler active at this queue
3051 *
3052 * It will not queue the request if there is an error with the bio, or at the
3053 * request creation.
3054 */
blk_mq_submit_bio(struct bio * bio)3055 void blk_mq_submit_bio(struct bio *bio)
3056 {
3057 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
3058 struct blk_plug *plug = current->plug;
3059 const int is_sync = op_is_sync(bio->bi_opf);
3060 struct blk_mq_hw_ctx *hctx;
3061 unsigned int nr_segs;
3062 struct request *rq;
3063 blk_status_t ret;
3064
3065 /*
3066 * If the plug has a cached request for this queue, try to use it.
3067 */
3068 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf);
3069
3070 /*
3071 * A BIO that was released from a zone write plug has already been
3072 * through the preparation in this function, already holds a reference
3073 * on the queue usage counter, and is the only write BIO in-flight for
3074 * the target zone. Go straight to preparing a request for it.
3075 */
3076 if (bio_zone_write_plugging(bio)) {
3077 nr_segs = bio->__bi_nr_segments;
3078 if (rq)
3079 blk_queue_exit(q);
3080 goto new_request;
3081 }
3082
3083 bio = blk_queue_bounce(bio, q);
3084
3085 /*
3086 * The cached request already holds a q_usage_counter reference and we
3087 * don't have to acquire a new one if we use it.
3088 */
3089 if (!rq) {
3090 if (unlikely(bio_queue_enter(bio)))
3091 return;
3092 }
3093
3094 /*
3095 * Device reconfiguration may change logical block size, so alignment
3096 * check has to be done with queue usage counter held
3097 */
3098 if (unlikely(bio_unaligned(bio, q))) {
3099 bio_io_error(bio);
3100 goto queue_exit;
3101 }
3102
3103 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
3104 if (!bio)
3105 goto queue_exit;
3106
3107 if (!bio_integrity_prep(bio))
3108 goto queue_exit;
3109
3110 if (blk_mq_attempt_bio_merge(q, bio, nr_segs))
3111 goto queue_exit;
3112
3113 if (blk_queue_is_zoned(q) && blk_zone_plug_bio(bio, nr_segs))
3114 goto queue_exit;
3115
3116 new_request:
3117 if (!rq) {
3118 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
3119 if (unlikely(!rq))
3120 goto queue_exit;
3121 } else {
3122 blk_mq_use_cached_rq(rq, plug, bio);
3123 }
3124
3125 trace_block_getrq(bio);
3126
3127 rq_qos_track(q, rq, bio);
3128
3129 blk_mq_bio_to_request(rq, bio, nr_segs);
3130
3131 ret = blk_crypto_rq_get_keyslot(rq);
3132 if (ret != BLK_STS_OK) {
3133 bio->bi_status = ret;
3134 bio_endio(bio);
3135 blk_mq_free_request(rq);
3136 return;
3137 }
3138
3139 if (bio_zone_write_plugging(bio))
3140 blk_zone_write_plug_init_request(rq);
3141
3142 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
3143 return;
3144
3145 if (plug) {
3146 blk_add_rq_to_plug(plug, rq);
3147 return;
3148 }
3149
3150 hctx = rq->mq_hctx;
3151 if ((rq->rq_flags & RQF_USE_SCHED) ||
3152 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
3153 blk_mq_insert_request(rq, 0);
3154 blk_mq_run_hw_queue(hctx, true);
3155 } else {
3156 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
3157 }
3158 return;
3159
3160 queue_exit:
3161 /*
3162 * Don't drop the queue reference if we were trying to use a cached
3163 * request and thus didn't acquire one.
3164 */
3165 if (!rq)
3166 blk_queue_exit(q);
3167 }
3168
3169 #ifdef CONFIG_BLK_MQ_STACKING
3170 /**
3171 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3172 * @rq: the request being queued
3173 */
blk_insert_cloned_request(struct request * rq)3174 blk_status_t blk_insert_cloned_request(struct request *rq)
3175 {
3176 struct request_queue *q = rq->q;
3177 unsigned int max_sectors = blk_queue_get_max_sectors(rq);
3178 unsigned int max_segments = blk_rq_get_max_segments(rq);
3179 blk_status_t ret;
3180
3181 if (blk_rq_sectors(rq) > max_sectors) {
3182 /*
3183 * SCSI device does not have a good way to return if
3184 * Write Same/Zero is actually supported. If a device rejects
3185 * a non-read/write command (discard, write same,etc.) the
3186 * low-level device driver will set the relevant queue limit to
3187 * 0 to prevent blk-lib from issuing more of the offending
3188 * operations. Commands queued prior to the queue limit being
3189 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3190 * errors being propagated to upper layers.
3191 */
3192 if (max_sectors == 0)
3193 return BLK_STS_NOTSUPP;
3194
3195 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3196 __func__, blk_rq_sectors(rq), max_sectors);
3197 return BLK_STS_IOERR;
3198 }
3199
3200 /*
3201 * The queue settings related to segment counting may differ from the
3202 * original queue.
3203 */
3204 rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3205 if (rq->nr_phys_segments > max_segments) {
3206 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3207 __func__, rq->nr_phys_segments, max_segments);
3208 return BLK_STS_IOERR;
3209 }
3210
3211 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3212 return BLK_STS_IOERR;
3213
3214 ret = blk_crypto_rq_get_keyslot(rq);
3215 if (ret != BLK_STS_OK)
3216 return ret;
3217
3218 blk_account_io_start(rq);
3219
3220 /*
3221 * Since we have a scheduler attached on the top device,
3222 * bypass a potential scheduler on the bottom device for
3223 * insert.
3224 */
3225 blk_mq_run_dispatch_ops(q,
3226 ret = blk_mq_request_issue_directly(rq, true));
3227 if (ret)
3228 blk_account_io_done(rq, blk_time_get_ns());
3229 return ret;
3230 }
3231 EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3232
3233 /**
3234 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3235 * @rq: the clone request to be cleaned up
3236 *
3237 * Description:
3238 * Free all bios in @rq for a cloned request.
3239 */
blk_rq_unprep_clone(struct request * rq)3240 void blk_rq_unprep_clone(struct request *rq)
3241 {
3242 struct bio *bio;
3243
3244 while ((bio = rq->bio) != NULL) {
3245 rq->bio = bio->bi_next;
3246
3247 bio_put(bio);
3248 }
3249 }
3250 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3251
3252 /**
3253 * blk_rq_prep_clone - Helper function to setup clone request
3254 * @rq: the request to be setup
3255 * @rq_src: original request to be cloned
3256 * @bs: bio_set that bios for clone are allocated from
3257 * @gfp_mask: memory allocation mask for bio
3258 * @bio_ctr: setup function to be called for each clone bio.
3259 * Returns %0 for success, non %0 for failure.
3260 * @data: private data to be passed to @bio_ctr
3261 *
3262 * Description:
3263 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3264 * Also, pages which the original bios are pointing to are not copied
3265 * and the cloned bios just point same pages.
3266 * So cloned bios must be completed before original bios, which means
3267 * the caller must complete @rq before @rq_src.
3268 */
blk_rq_prep_clone(struct request * rq,struct request * rq_src,struct bio_set * bs,gfp_t gfp_mask,int (* bio_ctr)(struct bio *,struct bio *,void *),void * data)3269 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3270 struct bio_set *bs, gfp_t gfp_mask,
3271 int (*bio_ctr)(struct bio *, struct bio *, void *),
3272 void *data)
3273 {
3274 struct bio *bio_src;
3275
3276 if (!bs)
3277 bs = &fs_bio_set;
3278
3279 __rq_for_each_bio(bio_src, rq_src) {
3280 struct bio *bio = bio_alloc_clone(rq->q->disk->part0, bio_src,
3281 gfp_mask, bs);
3282 if (!bio)
3283 goto free_and_out;
3284
3285 if (bio_ctr && bio_ctr(bio, bio_src, data)) {
3286 bio_put(bio);
3287 goto free_and_out;
3288 }
3289
3290 if (rq->bio) {
3291 rq->biotail->bi_next = bio;
3292 rq->biotail = bio;
3293 } else {
3294 rq->bio = rq->biotail = bio;
3295 }
3296 }
3297
3298 /* Copy attributes of the original request to the clone request. */
3299 rq->__sector = blk_rq_pos(rq_src);
3300 rq->__data_len = blk_rq_bytes(rq_src);
3301 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3302 rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3303 rq->special_vec = rq_src->special_vec;
3304 }
3305 rq->nr_phys_segments = rq_src->nr_phys_segments;
3306
3307 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3308 goto free_and_out;
3309
3310 return 0;
3311
3312 free_and_out:
3313 blk_rq_unprep_clone(rq);
3314
3315 return -ENOMEM;
3316 }
3317 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3318 #endif /* CONFIG_BLK_MQ_STACKING */
3319
3320 /*
3321 * Steal bios from a request and add them to a bio list.
3322 * The request must not have been partially completed before.
3323 */
blk_steal_bios(struct bio_list * list,struct request * rq)3324 void blk_steal_bios(struct bio_list *list, struct request *rq)
3325 {
3326 if (rq->bio) {
3327 if (list->tail)
3328 list->tail->bi_next = rq->bio;
3329 else
3330 list->head = rq->bio;
3331 list->tail = rq->biotail;
3332
3333 rq->bio = NULL;
3334 rq->biotail = NULL;
3335 }
3336
3337 rq->__data_len = 0;
3338 }
3339 EXPORT_SYMBOL_GPL(blk_steal_bios);
3340
order_to_size(unsigned int order)3341 static size_t order_to_size(unsigned int order)
3342 {
3343 return (size_t)PAGE_SIZE << order;
3344 }
3345
3346 /* called before freeing request pool in @tags */
blk_mq_clear_rq_mapping(struct blk_mq_tags * drv_tags,struct blk_mq_tags * tags)3347 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3348 struct blk_mq_tags *tags)
3349 {
3350 struct page *page;
3351 unsigned long flags;
3352
3353 /*
3354 * There is no need to clear mapping if driver tags is not initialized
3355 * or the mapping belongs to the driver tags.
3356 */
3357 if (!drv_tags || drv_tags == tags)
3358 return;
3359
3360 list_for_each_entry(page, &tags->page_list, lru) {
3361 unsigned long start = (unsigned long)page_address(page);
3362 unsigned long end = start + order_to_size(page->private);
3363 int i;
3364
3365 for (i = 0; i < drv_tags->nr_tags; i++) {
3366 struct request *rq = drv_tags->rqs[i];
3367 unsigned long rq_addr = (unsigned long)rq;
3368
3369 if (rq_addr >= start && rq_addr < end) {
3370 WARN_ON_ONCE(req_ref_read(rq) != 0);
3371 cmpxchg(&drv_tags->rqs[i], rq, NULL);
3372 }
3373 }
3374 }
3375
3376 /*
3377 * Wait until all pending iteration is done.
3378 *
3379 * Request reference is cleared and it is guaranteed to be observed
3380 * after the ->lock is released.
3381 */
3382 spin_lock_irqsave(&drv_tags->lock, flags);
3383 spin_unlock_irqrestore(&drv_tags->lock, flags);
3384 }
3385
blk_mq_free_rqs(struct blk_mq_tag_set * set,struct blk_mq_tags * tags,unsigned int hctx_idx)3386 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3387 unsigned int hctx_idx)
3388 {
3389 struct blk_mq_tags *drv_tags;
3390 struct page *page;
3391
3392 if (list_empty(&tags->page_list))
3393 return;
3394
3395 if (blk_mq_is_shared_tags(set->flags))
3396 drv_tags = set->shared_tags;
3397 else
3398 drv_tags = set->tags[hctx_idx];
3399
3400 if (tags->static_rqs && set->ops->exit_request) {
3401 int i;
3402
3403 for (i = 0; i < tags->nr_tags; i++) {
3404 struct request *rq = tags->static_rqs[i];
3405
3406 if (!rq)
3407 continue;
3408 set->ops->exit_request(set, rq, hctx_idx);
3409 tags->static_rqs[i] = NULL;
3410 }
3411 }
3412
3413 blk_mq_clear_rq_mapping(drv_tags, tags);
3414
3415 while (!list_empty(&tags->page_list)) {
3416 page = list_first_entry(&tags->page_list, struct page, lru);
3417 list_del_init(&page->lru);
3418 /*
3419 * Remove kmemleak object previously allocated in
3420 * blk_mq_alloc_rqs().
3421 */
3422 kmemleak_free(page_address(page));
3423 __free_pages(page, page->private);
3424 }
3425 }
3426
blk_mq_free_rq_map(struct blk_mq_tags * tags)3427 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3428 {
3429 kfree(tags->rqs);
3430 tags->rqs = NULL;
3431 kfree(tags->static_rqs);
3432 tags->static_rqs = NULL;
3433
3434 blk_mq_free_tags(tags);
3435 }
3436
hctx_idx_to_type(struct blk_mq_tag_set * set,unsigned int hctx_idx)3437 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3438 unsigned int hctx_idx)
3439 {
3440 int i;
3441
3442 for (i = 0; i < set->nr_maps; i++) {
3443 unsigned int start = set->map[i].queue_offset;
3444 unsigned int end = start + set->map[i].nr_queues;
3445
3446 if (hctx_idx >= start && hctx_idx < end)
3447 break;
3448 }
3449
3450 if (i >= set->nr_maps)
3451 i = HCTX_TYPE_DEFAULT;
3452
3453 return i;
3454 }
3455
blk_mq_get_hctx_node(struct blk_mq_tag_set * set,unsigned int hctx_idx)3456 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3457 unsigned int hctx_idx)
3458 {
3459 enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3460
3461 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3462 }
3463
blk_mq_alloc_rq_map(struct blk_mq_tag_set * set,unsigned int hctx_idx,unsigned int nr_tags,unsigned int reserved_tags)3464 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3465 unsigned int hctx_idx,
3466 unsigned int nr_tags,
3467 unsigned int reserved_tags)
3468 {
3469 int node = blk_mq_get_hctx_node(set, hctx_idx);
3470 struct blk_mq_tags *tags;
3471
3472 if (node == NUMA_NO_NODE)
3473 node = set->numa_node;
3474
3475 tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3476 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3477 if (!tags)
3478 return NULL;
3479
3480 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3481 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3482 node);
3483 if (!tags->rqs)
3484 goto err_free_tags;
3485
3486 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3487 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3488 node);
3489 if (!tags->static_rqs)
3490 goto err_free_rqs;
3491
3492 return tags;
3493
3494 err_free_rqs:
3495 kfree(tags->rqs);
3496 err_free_tags:
3497 blk_mq_free_tags(tags);
3498 return NULL;
3499 }
3500
blk_mq_init_request(struct blk_mq_tag_set * set,struct request * rq,unsigned int hctx_idx,int node)3501 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3502 unsigned int hctx_idx, int node)
3503 {
3504 int ret;
3505
3506 if (set->ops->init_request) {
3507 ret = set->ops->init_request(set, rq, hctx_idx, node);
3508 if (ret)
3509 return ret;
3510 }
3511
3512 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3513 return 0;
3514 }
3515
blk_mq_alloc_rqs(struct blk_mq_tag_set * set,struct blk_mq_tags * tags,unsigned int hctx_idx,unsigned int depth)3516 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3517 struct blk_mq_tags *tags,
3518 unsigned int hctx_idx, unsigned int depth)
3519 {
3520 unsigned int i, j, entries_per_page, max_order = 4;
3521 int node = blk_mq_get_hctx_node(set, hctx_idx);
3522 size_t rq_size, left;
3523
3524 if (node == NUMA_NO_NODE)
3525 node = set->numa_node;
3526
3527 INIT_LIST_HEAD(&tags->page_list);
3528
3529 /*
3530 * rq_size is the size of the request plus driver payload, rounded
3531 * to the cacheline size
3532 */
3533 rq_size = round_up(sizeof(struct request) + set->cmd_size,
3534 cache_line_size());
3535 left = rq_size * depth;
3536
3537 for (i = 0; i < depth; ) {
3538 int this_order = max_order;
3539 struct page *page;
3540 int to_do;
3541 void *p;
3542
3543 while (this_order && left < order_to_size(this_order - 1))
3544 this_order--;
3545
3546 do {
3547 page = alloc_pages_node(node,
3548 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3549 this_order);
3550 if (page)
3551 break;
3552 if (!this_order--)
3553 break;
3554 if (order_to_size(this_order) < rq_size)
3555 break;
3556 } while (1);
3557
3558 if (!page)
3559 goto fail;
3560
3561 page->private = this_order;
3562 list_add_tail(&page->lru, &tags->page_list);
3563
3564 p = page_address(page);
3565 /*
3566 * Allow kmemleak to scan these pages as they contain pointers
3567 * to additional allocations like via ops->init_request().
3568 */
3569 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3570 entries_per_page = order_to_size(this_order) / rq_size;
3571 to_do = min(entries_per_page, depth - i);
3572 left -= to_do * rq_size;
3573 for (j = 0; j < to_do; j++) {
3574 struct request *rq = p;
3575
3576 tags->static_rqs[i] = rq;
3577 if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3578 tags->static_rqs[i] = NULL;
3579 goto fail;
3580 }
3581
3582 p += rq_size;
3583 i++;
3584 }
3585 }
3586 return 0;
3587
3588 fail:
3589 blk_mq_free_rqs(set, tags, hctx_idx);
3590 return -ENOMEM;
3591 }
3592
3593 struct rq_iter_data {
3594 struct blk_mq_hw_ctx *hctx;
3595 bool has_rq;
3596 };
3597
blk_mq_has_request(struct request * rq,void * data)3598 static bool blk_mq_has_request(struct request *rq, void *data)
3599 {
3600 struct rq_iter_data *iter_data = data;
3601
3602 if (rq->mq_hctx != iter_data->hctx)
3603 return true;
3604 iter_data->has_rq = true;
3605 return false;
3606 }
3607
blk_mq_hctx_has_requests(struct blk_mq_hw_ctx * hctx)3608 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3609 {
3610 struct blk_mq_tags *tags = hctx->sched_tags ?
3611 hctx->sched_tags : hctx->tags;
3612 struct rq_iter_data data = {
3613 .hctx = hctx,
3614 };
3615
3616 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3617 return data.has_rq;
3618 }
3619
blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx * hctx,unsigned int this_cpu)3620 static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx,
3621 unsigned int this_cpu)
3622 {
3623 enum hctx_type type = hctx->type;
3624 int cpu;
3625
3626 /*
3627 * hctx->cpumask has to rule out isolated CPUs, but userspace still
3628 * might submit IOs on these isolated CPUs, so use the queue map to
3629 * check if all CPUs mapped to this hctx are offline
3630 */
3631 for_each_online_cpu(cpu) {
3632 struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue,
3633 type, cpu);
3634
3635 if (h != hctx)
3636 continue;
3637
3638 /* this hctx has at least one online CPU */
3639 if (this_cpu != cpu)
3640 return true;
3641 }
3642
3643 return false;
3644 }
3645
blk_mq_hctx_notify_offline(unsigned int cpu,struct hlist_node * node)3646 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3647 {
3648 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3649 struct blk_mq_hw_ctx, cpuhp_online);
3650
3651 if (blk_mq_hctx_has_online_cpu(hctx, cpu))
3652 return 0;
3653
3654 /*
3655 * Prevent new request from being allocated on the current hctx.
3656 *
3657 * The smp_mb__after_atomic() Pairs with the implied barrier in
3658 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3659 * seen once we return from the tag allocator.
3660 */
3661 set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3662 smp_mb__after_atomic();
3663
3664 /*
3665 * Try to grab a reference to the queue and wait for any outstanding
3666 * requests. If we could not grab a reference the queue has been
3667 * frozen and there are no requests.
3668 */
3669 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3670 while (blk_mq_hctx_has_requests(hctx))
3671 msleep(5);
3672 percpu_ref_put(&hctx->queue->q_usage_counter);
3673 }
3674
3675 return 0;
3676 }
3677
3678 /*
3679 * Check if one CPU is mapped to the specified hctx
3680 *
3681 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed
3682 * to be used for scheduling kworker only. For other usage, please call this
3683 * helper for checking if one CPU belongs to the specified hctx
3684 */
blk_mq_cpu_mapped_to_hctx(unsigned int cpu,const struct blk_mq_hw_ctx * hctx)3685 static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu,
3686 const struct blk_mq_hw_ctx *hctx)
3687 {
3688 struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue,
3689 hctx->type, cpu);
3690
3691 return mapped_hctx == hctx;
3692 }
3693
blk_mq_hctx_notify_online(unsigned int cpu,struct hlist_node * node)3694 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3695 {
3696 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3697 struct blk_mq_hw_ctx, cpuhp_online);
3698
3699 if (blk_mq_cpu_mapped_to_hctx(cpu, hctx))
3700 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3701 return 0;
3702 }
3703
3704 /*
3705 * 'cpu' is going away. splice any existing rq_list entries from this
3706 * software queue to the hw queue dispatch list, and ensure that it
3707 * gets run.
3708 */
blk_mq_hctx_notify_dead(unsigned int cpu,struct hlist_node * node)3709 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3710 {
3711 struct blk_mq_hw_ctx *hctx;
3712 struct blk_mq_ctx *ctx;
3713 LIST_HEAD(tmp);
3714 enum hctx_type type;
3715
3716 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3717 if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx))
3718 return 0;
3719
3720 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3721 type = hctx->type;
3722
3723 spin_lock(&ctx->lock);
3724 if (!list_empty(&ctx->rq_lists[type])) {
3725 list_splice_init(&ctx->rq_lists[type], &tmp);
3726 blk_mq_hctx_clear_pending(hctx, ctx);
3727 }
3728 spin_unlock(&ctx->lock);
3729
3730 if (list_empty(&tmp))
3731 return 0;
3732
3733 spin_lock(&hctx->lock);
3734 list_splice_tail_init(&tmp, &hctx->dispatch);
3735 spin_unlock(&hctx->lock);
3736
3737 blk_mq_run_hw_queue(hctx, true);
3738 return 0;
3739 }
3740
__blk_mq_remove_cpuhp(struct blk_mq_hw_ctx * hctx)3741 static void __blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3742 {
3743 lockdep_assert_held(&blk_mq_cpuhp_lock);
3744
3745 if (!(hctx->flags & BLK_MQ_F_STACKING) &&
3746 !hlist_unhashed(&hctx->cpuhp_online)) {
3747 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3748 &hctx->cpuhp_online);
3749 INIT_HLIST_NODE(&hctx->cpuhp_online);
3750 }
3751
3752 if (!hlist_unhashed(&hctx->cpuhp_dead)) {
3753 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3754 &hctx->cpuhp_dead);
3755 INIT_HLIST_NODE(&hctx->cpuhp_dead);
3756 }
3757 }
3758
blk_mq_remove_cpuhp(struct blk_mq_hw_ctx * hctx)3759 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3760 {
3761 mutex_lock(&blk_mq_cpuhp_lock);
3762 __blk_mq_remove_cpuhp(hctx);
3763 mutex_unlock(&blk_mq_cpuhp_lock);
3764 }
3765
__blk_mq_add_cpuhp(struct blk_mq_hw_ctx * hctx)3766 static void __blk_mq_add_cpuhp(struct blk_mq_hw_ctx *hctx)
3767 {
3768 lockdep_assert_held(&blk_mq_cpuhp_lock);
3769
3770 if (!(hctx->flags & BLK_MQ_F_STACKING) &&
3771 hlist_unhashed(&hctx->cpuhp_online))
3772 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3773 &hctx->cpuhp_online);
3774
3775 if (hlist_unhashed(&hctx->cpuhp_dead))
3776 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3777 &hctx->cpuhp_dead);
3778 }
3779
__blk_mq_remove_cpuhp_list(struct list_head * head)3780 static void __blk_mq_remove_cpuhp_list(struct list_head *head)
3781 {
3782 struct blk_mq_hw_ctx *hctx;
3783
3784 lockdep_assert_held(&blk_mq_cpuhp_lock);
3785
3786 list_for_each_entry(hctx, head, hctx_list)
3787 __blk_mq_remove_cpuhp(hctx);
3788 }
3789
3790 /*
3791 * Unregister cpuhp callbacks from exited hw queues
3792 *
3793 * Safe to call if this `request_queue` is live
3794 */
blk_mq_remove_hw_queues_cpuhp(struct request_queue * q)3795 static void blk_mq_remove_hw_queues_cpuhp(struct request_queue *q)
3796 {
3797 LIST_HEAD(hctx_list);
3798
3799 spin_lock(&q->unused_hctx_lock);
3800 list_splice_init(&q->unused_hctx_list, &hctx_list);
3801 spin_unlock(&q->unused_hctx_lock);
3802
3803 mutex_lock(&blk_mq_cpuhp_lock);
3804 __blk_mq_remove_cpuhp_list(&hctx_list);
3805 mutex_unlock(&blk_mq_cpuhp_lock);
3806
3807 spin_lock(&q->unused_hctx_lock);
3808 list_splice(&hctx_list, &q->unused_hctx_list);
3809 spin_unlock(&q->unused_hctx_lock);
3810 }
3811
3812 /*
3813 * Register cpuhp callbacks from all hw queues
3814 *
3815 * Safe to call if this `request_queue` is live
3816 */
blk_mq_add_hw_queues_cpuhp(struct request_queue * q)3817 static void blk_mq_add_hw_queues_cpuhp(struct request_queue *q)
3818 {
3819 struct blk_mq_hw_ctx *hctx;
3820 unsigned long i;
3821
3822 mutex_lock(&blk_mq_cpuhp_lock);
3823 queue_for_each_hw_ctx(q, hctx, i)
3824 __blk_mq_add_cpuhp(hctx);
3825 mutex_unlock(&blk_mq_cpuhp_lock);
3826 }
3827
3828 /*
3829 * Before freeing hw queue, clearing the flush request reference in
3830 * tags->rqs[] for avoiding potential UAF.
3831 */
blk_mq_clear_flush_rq_mapping(struct blk_mq_tags * tags,unsigned int queue_depth,struct request * flush_rq)3832 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3833 unsigned int queue_depth, struct request *flush_rq)
3834 {
3835 int i;
3836 unsigned long flags;
3837
3838 /* The hw queue may not be mapped yet */
3839 if (!tags)
3840 return;
3841
3842 WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3843
3844 for (i = 0; i < queue_depth; i++)
3845 cmpxchg(&tags->rqs[i], flush_rq, NULL);
3846
3847 /*
3848 * Wait until all pending iteration is done.
3849 *
3850 * Request reference is cleared and it is guaranteed to be observed
3851 * after the ->lock is released.
3852 */
3853 spin_lock_irqsave(&tags->lock, flags);
3854 spin_unlock_irqrestore(&tags->lock, flags);
3855 }
3856
3857 /* hctx->ctxs will be freed in queue's release handler */
blk_mq_exit_hctx(struct request_queue * q,struct blk_mq_tag_set * set,struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)3858 static void blk_mq_exit_hctx(struct request_queue *q,
3859 struct blk_mq_tag_set *set,
3860 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3861 {
3862 struct request *flush_rq = hctx->fq->flush_rq;
3863
3864 if (blk_mq_hw_queue_mapped(hctx))
3865 blk_mq_tag_idle(hctx);
3866
3867 if (blk_queue_init_done(q))
3868 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3869 set->queue_depth, flush_rq);
3870 if (set->ops->exit_request)
3871 set->ops->exit_request(set, flush_rq, hctx_idx);
3872
3873 if (set->ops->exit_hctx)
3874 set->ops->exit_hctx(hctx, hctx_idx);
3875
3876 xa_erase(&q->hctx_table, hctx_idx);
3877
3878 spin_lock(&q->unused_hctx_lock);
3879 list_add(&hctx->hctx_list, &q->unused_hctx_list);
3880 spin_unlock(&q->unused_hctx_lock);
3881 }
3882
blk_mq_exit_hw_queues(struct request_queue * q,struct blk_mq_tag_set * set,int nr_queue)3883 static void blk_mq_exit_hw_queues(struct request_queue *q,
3884 struct blk_mq_tag_set *set, int nr_queue)
3885 {
3886 struct blk_mq_hw_ctx *hctx;
3887 unsigned long i;
3888
3889 queue_for_each_hw_ctx(q, hctx, i) {
3890 if (i == nr_queue)
3891 break;
3892 blk_mq_remove_cpuhp(hctx);
3893 blk_mq_exit_hctx(q, set, hctx, i);
3894 }
3895 }
3896
blk_mq_init_hctx(struct request_queue * q,struct blk_mq_tag_set * set,struct blk_mq_hw_ctx * hctx,unsigned hctx_idx)3897 static int blk_mq_init_hctx(struct request_queue *q,
3898 struct blk_mq_tag_set *set,
3899 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3900 {
3901 hctx->queue_num = hctx_idx;
3902
3903 hctx->tags = set->tags[hctx_idx];
3904
3905 if (set->ops->init_hctx &&
3906 set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3907 goto fail;
3908
3909 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3910 hctx->numa_node))
3911 goto exit_hctx;
3912
3913 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3914 goto exit_flush_rq;
3915
3916 return 0;
3917
3918 exit_flush_rq:
3919 if (set->ops->exit_request)
3920 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3921 exit_hctx:
3922 if (set->ops->exit_hctx)
3923 set->ops->exit_hctx(hctx, hctx_idx);
3924 fail:
3925 return -1;
3926 }
3927
3928 static struct blk_mq_hw_ctx *
blk_mq_alloc_hctx(struct request_queue * q,struct blk_mq_tag_set * set,int node)3929 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3930 int node)
3931 {
3932 struct blk_mq_hw_ctx *hctx;
3933 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3934
3935 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3936 if (!hctx)
3937 goto fail_alloc_hctx;
3938
3939 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3940 goto free_hctx;
3941
3942 atomic_set(&hctx->nr_active, 0);
3943 if (node == NUMA_NO_NODE)
3944 node = set->numa_node;
3945 hctx->numa_node = node;
3946
3947 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3948 spin_lock_init(&hctx->lock);
3949 INIT_LIST_HEAD(&hctx->dispatch);
3950 INIT_HLIST_NODE(&hctx->cpuhp_dead);
3951 INIT_HLIST_NODE(&hctx->cpuhp_online);
3952 hctx->queue = q;
3953 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3954
3955 INIT_LIST_HEAD(&hctx->hctx_list);
3956
3957 /*
3958 * Allocate space for all possible cpus to avoid allocation at
3959 * runtime
3960 */
3961 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3962 gfp, node);
3963 if (!hctx->ctxs)
3964 goto free_cpumask;
3965
3966 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3967 gfp, node, false, false))
3968 goto free_ctxs;
3969 hctx->nr_ctx = 0;
3970
3971 spin_lock_init(&hctx->dispatch_wait_lock);
3972 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3973 INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3974
3975 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3976 if (!hctx->fq)
3977 goto free_bitmap;
3978
3979 blk_mq_hctx_kobj_init(hctx);
3980
3981 return hctx;
3982
3983 free_bitmap:
3984 sbitmap_free(&hctx->ctx_map);
3985 free_ctxs:
3986 kfree(hctx->ctxs);
3987 free_cpumask:
3988 free_cpumask_var(hctx->cpumask);
3989 free_hctx:
3990 kfree(hctx);
3991 fail_alloc_hctx:
3992 return NULL;
3993 }
3994
blk_mq_init_cpu_queues(struct request_queue * q,unsigned int nr_hw_queues)3995 static void blk_mq_init_cpu_queues(struct request_queue *q,
3996 unsigned int nr_hw_queues)
3997 {
3998 struct blk_mq_tag_set *set = q->tag_set;
3999 unsigned int i, j;
4000
4001 for_each_possible_cpu(i) {
4002 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
4003 struct blk_mq_hw_ctx *hctx;
4004 int k;
4005
4006 __ctx->cpu = i;
4007 spin_lock_init(&__ctx->lock);
4008 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
4009 INIT_LIST_HEAD(&__ctx->rq_lists[k]);
4010
4011 __ctx->queue = q;
4012
4013 /*
4014 * Set local node, IFF we have more than one hw queue. If
4015 * not, we remain on the home node of the device
4016 */
4017 for (j = 0; j < set->nr_maps; j++) {
4018 hctx = blk_mq_map_queue_type(q, j, i);
4019 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
4020 hctx->numa_node = cpu_to_node(i);
4021 }
4022 }
4023 }
4024
blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set * set,unsigned int hctx_idx,unsigned int depth)4025 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
4026 unsigned int hctx_idx,
4027 unsigned int depth)
4028 {
4029 struct blk_mq_tags *tags;
4030 int ret;
4031
4032 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
4033 if (!tags)
4034 return NULL;
4035
4036 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
4037 if (ret) {
4038 blk_mq_free_rq_map(tags);
4039 return NULL;
4040 }
4041
4042 return tags;
4043 }
4044
__blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set * set,int hctx_idx)4045 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
4046 int hctx_idx)
4047 {
4048 if (blk_mq_is_shared_tags(set->flags)) {
4049 set->tags[hctx_idx] = set->shared_tags;
4050
4051 return true;
4052 }
4053
4054 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
4055 set->queue_depth);
4056
4057 return set->tags[hctx_idx];
4058 }
4059
blk_mq_free_map_and_rqs(struct blk_mq_tag_set * set,struct blk_mq_tags * tags,unsigned int hctx_idx)4060 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
4061 struct blk_mq_tags *tags,
4062 unsigned int hctx_idx)
4063 {
4064 if (tags) {
4065 blk_mq_free_rqs(set, tags, hctx_idx);
4066 blk_mq_free_rq_map(tags);
4067 }
4068 }
4069
__blk_mq_free_map_and_rqs(struct blk_mq_tag_set * set,unsigned int hctx_idx)4070 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
4071 unsigned int hctx_idx)
4072 {
4073 if (!blk_mq_is_shared_tags(set->flags))
4074 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
4075
4076 set->tags[hctx_idx] = NULL;
4077 }
4078
blk_mq_map_swqueue(struct request_queue * q)4079 static void blk_mq_map_swqueue(struct request_queue *q)
4080 {
4081 unsigned int j, hctx_idx;
4082 unsigned long i;
4083 struct blk_mq_hw_ctx *hctx;
4084 struct blk_mq_ctx *ctx;
4085 struct blk_mq_tag_set *set = q->tag_set;
4086
4087 queue_for_each_hw_ctx(q, hctx, i) {
4088 cpumask_clear(hctx->cpumask);
4089 hctx->nr_ctx = 0;
4090 hctx->dispatch_from = NULL;
4091 }
4092
4093 /*
4094 * Map software to hardware queues.
4095 *
4096 * If the cpu isn't present, the cpu is mapped to first hctx.
4097 */
4098 for_each_possible_cpu(i) {
4099
4100 ctx = per_cpu_ptr(q->queue_ctx, i);
4101 for (j = 0; j < set->nr_maps; j++) {
4102 if (!set->map[j].nr_queues) {
4103 ctx->hctxs[j] = blk_mq_map_queue_type(q,
4104 HCTX_TYPE_DEFAULT, i);
4105 continue;
4106 }
4107 hctx_idx = set->map[j].mq_map[i];
4108 /* unmapped hw queue can be remapped after CPU topo changed */
4109 if (!set->tags[hctx_idx] &&
4110 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
4111 /*
4112 * If tags initialization fail for some hctx,
4113 * that hctx won't be brought online. In this
4114 * case, remap the current ctx to hctx[0] which
4115 * is guaranteed to always have tags allocated
4116 */
4117 set->map[j].mq_map[i] = 0;
4118 }
4119
4120 hctx = blk_mq_map_queue_type(q, j, i);
4121 ctx->hctxs[j] = hctx;
4122 /*
4123 * If the CPU is already set in the mask, then we've
4124 * mapped this one already. This can happen if
4125 * devices share queues across queue maps.
4126 */
4127 if (cpumask_test_cpu(i, hctx->cpumask))
4128 continue;
4129
4130 cpumask_set_cpu(i, hctx->cpumask);
4131 hctx->type = j;
4132 ctx->index_hw[hctx->type] = hctx->nr_ctx;
4133 hctx->ctxs[hctx->nr_ctx++] = ctx;
4134
4135 /*
4136 * If the nr_ctx type overflows, we have exceeded the
4137 * amount of sw queues we can support.
4138 */
4139 BUG_ON(!hctx->nr_ctx);
4140 }
4141
4142 for (; j < HCTX_MAX_TYPES; j++)
4143 ctx->hctxs[j] = blk_mq_map_queue_type(q,
4144 HCTX_TYPE_DEFAULT, i);
4145 }
4146
4147 queue_for_each_hw_ctx(q, hctx, i) {
4148 int cpu;
4149
4150 /*
4151 * If no software queues are mapped to this hardware queue,
4152 * disable it and free the request entries.
4153 */
4154 if (!hctx->nr_ctx) {
4155 /* Never unmap queue 0. We need it as a
4156 * fallback in case of a new remap fails
4157 * allocation
4158 */
4159 if (i)
4160 __blk_mq_free_map_and_rqs(set, i);
4161
4162 hctx->tags = NULL;
4163 continue;
4164 }
4165
4166 hctx->tags = set->tags[i];
4167 WARN_ON(!hctx->tags);
4168
4169 /*
4170 * Set the map size to the number of mapped software queues.
4171 * This is more accurate and more efficient than looping
4172 * over all possibly mapped software queues.
4173 */
4174 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
4175
4176 /*
4177 * Rule out isolated CPUs from hctx->cpumask to avoid
4178 * running block kworker on isolated CPUs
4179 */
4180 for_each_cpu(cpu, hctx->cpumask) {
4181 if (cpu_is_isolated(cpu))
4182 cpumask_clear_cpu(cpu, hctx->cpumask);
4183 }
4184
4185 /*
4186 * Initialize batch roundrobin counts
4187 */
4188 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
4189 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
4190 }
4191 }
4192
4193 /*
4194 * Caller needs to ensure that we're either frozen/quiesced, or that
4195 * the queue isn't live yet.
4196 */
queue_set_hctx_shared(struct request_queue * q,bool shared)4197 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
4198 {
4199 struct blk_mq_hw_ctx *hctx;
4200 unsigned long i;
4201
4202 queue_for_each_hw_ctx(q, hctx, i) {
4203 if (shared) {
4204 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4205 } else {
4206 blk_mq_tag_idle(hctx);
4207 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4208 }
4209 }
4210 }
4211
blk_mq_update_tag_set_shared(struct blk_mq_tag_set * set,bool shared)4212 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
4213 bool shared)
4214 {
4215 struct request_queue *q;
4216
4217 lockdep_assert_held(&set->tag_list_lock);
4218
4219 list_for_each_entry(q, &set->tag_list, tag_set_list) {
4220 blk_mq_freeze_queue(q);
4221 queue_set_hctx_shared(q, shared);
4222 blk_mq_unfreeze_queue(q);
4223 }
4224 }
4225
blk_mq_del_queue_tag_set(struct request_queue * q)4226 static void blk_mq_del_queue_tag_set(struct request_queue *q)
4227 {
4228 struct blk_mq_tag_set *set = q->tag_set;
4229
4230 mutex_lock(&set->tag_list_lock);
4231 list_del(&q->tag_set_list);
4232 if (list_is_singular(&set->tag_list)) {
4233 /* just transitioned to unshared */
4234 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
4235 /* update existing queue */
4236 blk_mq_update_tag_set_shared(set, false);
4237 }
4238 mutex_unlock(&set->tag_list_lock);
4239 INIT_LIST_HEAD(&q->tag_set_list);
4240 }
4241
blk_mq_add_queue_tag_set(struct blk_mq_tag_set * set,struct request_queue * q)4242 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
4243 struct request_queue *q)
4244 {
4245 mutex_lock(&set->tag_list_lock);
4246
4247 /*
4248 * Check to see if we're transitioning to shared (from 1 to 2 queues).
4249 */
4250 if (!list_empty(&set->tag_list) &&
4251 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
4252 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
4253 /* update existing queue */
4254 blk_mq_update_tag_set_shared(set, true);
4255 }
4256 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
4257 queue_set_hctx_shared(q, true);
4258 list_add_tail(&q->tag_set_list, &set->tag_list);
4259
4260 mutex_unlock(&set->tag_list_lock);
4261 }
4262
4263 /* All allocations will be freed in release handler of q->mq_kobj */
blk_mq_alloc_ctxs(struct request_queue * q)4264 static int blk_mq_alloc_ctxs(struct request_queue *q)
4265 {
4266 struct blk_mq_ctxs *ctxs;
4267 int cpu;
4268
4269 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
4270 if (!ctxs)
4271 return -ENOMEM;
4272
4273 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
4274 if (!ctxs->queue_ctx)
4275 goto fail;
4276
4277 for_each_possible_cpu(cpu) {
4278 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
4279 ctx->ctxs = ctxs;
4280 }
4281
4282 q->mq_kobj = &ctxs->kobj;
4283 q->queue_ctx = ctxs->queue_ctx;
4284
4285 return 0;
4286 fail:
4287 kfree(ctxs);
4288 return -ENOMEM;
4289 }
4290
4291 /*
4292 * It is the actual release handler for mq, but we do it from
4293 * request queue's release handler for avoiding use-after-free
4294 * and headache because q->mq_kobj shouldn't have been introduced,
4295 * but we can't group ctx/kctx kobj without it.
4296 */
blk_mq_release(struct request_queue * q)4297 void blk_mq_release(struct request_queue *q)
4298 {
4299 struct blk_mq_hw_ctx *hctx, *next;
4300 unsigned long i;
4301
4302 queue_for_each_hw_ctx(q, hctx, i)
4303 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4304
4305 /* all hctx are in .unused_hctx_list now */
4306 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4307 list_del_init(&hctx->hctx_list);
4308 kobject_put(&hctx->kobj);
4309 }
4310
4311 xa_destroy(&q->hctx_table);
4312
4313 /*
4314 * release .mq_kobj and sw queue's kobject now because
4315 * both share lifetime with request queue.
4316 */
4317 blk_mq_sysfs_deinit(q);
4318 }
4319
blk_mq_can_poll(struct blk_mq_tag_set * set)4320 static bool blk_mq_can_poll(struct blk_mq_tag_set *set)
4321 {
4322 return set->nr_maps > HCTX_TYPE_POLL &&
4323 set->map[HCTX_TYPE_POLL].nr_queues;
4324 }
4325
blk_mq_alloc_queue(struct blk_mq_tag_set * set,struct queue_limits * lim,void * queuedata)4326 struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set,
4327 struct queue_limits *lim, void *queuedata)
4328 {
4329 struct queue_limits default_lim = { };
4330 struct request_queue *q;
4331 int ret;
4332
4333 if (!lim)
4334 lim = &default_lim;
4335 lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT;
4336 if (blk_mq_can_poll(set))
4337 lim->features |= BLK_FEAT_POLL;
4338
4339 q = blk_alloc_queue(lim, set->numa_node);
4340 if (IS_ERR(q))
4341 return q;
4342 q->queuedata = queuedata;
4343 ret = blk_mq_init_allocated_queue(set, q);
4344 if (ret) {
4345 blk_put_queue(q);
4346 return ERR_PTR(ret);
4347 }
4348 return q;
4349 }
4350 EXPORT_SYMBOL(blk_mq_alloc_queue);
4351
4352 /**
4353 * blk_mq_destroy_queue - shutdown a request queue
4354 * @q: request queue to shutdown
4355 *
4356 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future
4357 * requests will be failed with -ENODEV. The caller is responsible for dropping
4358 * the reference from blk_mq_alloc_queue() by calling blk_put_queue().
4359 *
4360 * Context: can sleep
4361 */
blk_mq_destroy_queue(struct request_queue * q)4362 void blk_mq_destroy_queue(struct request_queue *q)
4363 {
4364 WARN_ON_ONCE(!queue_is_mq(q));
4365 WARN_ON_ONCE(blk_queue_registered(q));
4366
4367 might_sleep();
4368
4369 blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4370 blk_queue_start_drain(q);
4371 blk_mq_freeze_queue_wait(q);
4372
4373 blk_sync_queue(q);
4374 blk_mq_cancel_work_sync(q);
4375 blk_mq_exit_queue(q);
4376 }
4377 EXPORT_SYMBOL(blk_mq_destroy_queue);
4378
__blk_mq_alloc_disk(struct blk_mq_tag_set * set,struct queue_limits * lim,void * queuedata,struct lock_class_key * lkclass)4379 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set,
4380 struct queue_limits *lim, void *queuedata,
4381 struct lock_class_key *lkclass)
4382 {
4383 struct request_queue *q;
4384 struct gendisk *disk;
4385
4386 q = blk_mq_alloc_queue(set, lim, queuedata);
4387 if (IS_ERR(q))
4388 return ERR_CAST(q);
4389
4390 disk = __alloc_disk_node(q, set->numa_node, lkclass);
4391 if (!disk) {
4392 blk_mq_destroy_queue(q);
4393 blk_put_queue(q);
4394 return ERR_PTR(-ENOMEM);
4395 }
4396 set_bit(GD_OWNS_QUEUE, &disk->state);
4397 return disk;
4398 }
4399 EXPORT_SYMBOL(__blk_mq_alloc_disk);
4400
blk_mq_alloc_disk_for_queue(struct request_queue * q,struct lock_class_key * lkclass)4401 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4402 struct lock_class_key *lkclass)
4403 {
4404 struct gendisk *disk;
4405
4406 if (!blk_get_queue(q))
4407 return NULL;
4408 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4409 if (!disk)
4410 blk_put_queue(q);
4411 return disk;
4412 }
4413 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4414
4415 /*
4416 * Only hctx removed from cpuhp list can be reused
4417 */
blk_mq_hctx_is_reusable(struct blk_mq_hw_ctx * hctx)4418 static bool blk_mq_hctx_is_reusable(struct blk_mq_hw_ctx *hctx)
4419 {
4420 return hlist_unhashed(&hctx->cpuhp_online) &&
4421 hlist_unhashed(&hctx->cpuhp_dead);
4422 }
4423
blk_mq_alloc_and_init_hctx(struct blk_mq_tag_set * set,struct request_queue * q,int hctx_idx,int node)4424 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4425 struct blk_mq_tag_set *set, struct request_queue *q,
4426 int hctx_idx, int node)
4427 {
4428 struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4429
4430 /* reuse dead hctx first */
4431 spin_lock(&q->unused_hctx_lock);
4432 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4433 if (tmp->numa_node == node && blk_mq_hctx_is_reusable(tmp)) {
4434 hctx = tmp;
4435 break;
4436 }
4437 }
4438 if (hctx)
4439 list_del_init(&hctx->hctx_list);
4440 spin_unlock(&q->unused_hctx_lock);
4441
4442 if (!hctx)
4443 hctx = blk_mq_alloc_hctx(q, set, node);
4444 if (!hctx)
4445 goto fail;
4446
4447 if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4448 goto free_hctx;
4449
4450 return hctx;
4451
4452 free_hctx:
4453 kobject_put(&hctx->kobj);
4454 fail:
4455 return NULL;
4456 }
4457
blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set * set,struct request_queue * q)4458 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4459 struct request_queue *q)
4460 {
4461 struct blk_mq_hw_ctx *hctx;
4462 unsigned long i, j;
4463
4464 /* protect against switching io scheduler */
4465 mutex_lock(&q->sysfs_lock);
4466 for (i = 0; i < set->nr_hw_queues; i++) {
4467 int old_node;
4468 int node = blk_mq_get_hctx_node(set, i);
4469 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4470
4471 if (old_hctx) {
4472 old_node = old_hctx->numa_node;
4473 blk_mq_exit_hctx(q, set, old_hctx, i);
4474 }
4475
4476 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4477 if (!old_hctx)
4478 break;
4479 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4480 node, old_node);
4481 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4482 WARN_ON_ONCE(!hctx);
4483 }
4484 }
4485 /*
4486 * Increasing nr_hw_queues fails. Free the newly allocated
4487 * hctxs and keep the previous q->nr_hw_queues.
4488 */
4489 if (i != set->nr_hw_queues) {
4490 j = q->nr_hw_queues;
4491 } else {
4492 j = i;
4493 q->nr_hw_queues = set->nr_hw_queues;
4494 }
4495
4496 xa_for_each_start(&q->hctx_table, j, hctx, j)
4497 blk_mq_exit_hctx(q, set, hctx, j);
4498 mutex_unlock(&q->sysfs_lock);
4499
4500 /* unregister cpuhp callbacks for exited hctxs */
4501 blk_mq_remove_hw_queues_cpuhp(q);
4502
4503 /* register cpuhp for new initialized hctxs */
4504 blk_mq_add_hw_queues_cpuhp(q);
4505 }
4506
blk_mq_init_allocated_queue(struct blk_mq_tag_set * set,struct request_queue * q)4507 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4508 struct request_queue *q)
4509 {
4510 /* mark the queue as mq asap */
4511 q->mq_ops = set->ops;
4512
4513 /*
4514 * ->tag_set has to be setup before initialize hctx, which cpuphp
4515 * handler needs it for checking queue mapping
4516 */
4517 q->tag_set = set;
4518
4519 if (blk_mq_alloc_ctxs(q))
4520 goto err_exit;
4521
4522 /* init q->mq_kobj and sw queues' kobjects */
4523 blk_mq_sysfs_init(q);
4524
4525 INIT_LIST_HEAD(&q->unused_hctx_list);
4526 spin_lock_init(&q->unused_hctx_lock);
4527
4528 xa_init(&q->hctx_table);
4529
4530 blk_mq_realloc_hw_ctxs(set, q);
4531 if (!q->nr_hw_queues)
4532 goto err_hctxs;
4533
4534 INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4535 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4536
4537 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4538
4539 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4540 INIT_LIST_HEAD(&q->flush_list);
4541 INIT_LIST_HEAD(&q->requeue_list);
4542 spin_lock_init(&q->requeue_lock);
4543
4544 q->nr_requests = set->queue_depth;
4545
4546 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4547 blk_mq_add_queue_tag_set(set, q);
4548 blk_mq_map_swqueue(q);
4549 return 0;
4550
4551 err_hctxs:
4552 blk_mq_release(q);
4553 err_exit:
4554 q->mq_ops = NULL;
4555 return -ENOMEM;
4556 }
4557 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4558
4559 /* tags can _not_ be used after returning from blk_mq_exit_queue */
blk_mq_exit_queue(struct request_queue * q)4560 void blk_mq_exit_queue(struct request_queue *q)
4561 {
4562 struct blk_mq_tag_set *set = q->tag_set;
4563
4564 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4565 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4566 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4567 blk_mq_del_queue_tag_set(q);
4568 }
4569
__blk_mq_alloc_rq_maps(struct blk_mq_tag_set * set)4570 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4571 {
4572 int i;
4573
4574 if (blk_mq_is_shared_tags(set->flags)) {
4575 set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4576 BLK_MQ_NO_HCTX_IDX,
4577 set->queue_depth);
4578 if (!set->shared_tags)
4579 return -ENOMEM;
4580 }
4581
4582 for (i = 0; i < set->nr_hw_queues; i++) {
4583 if (!__blk_mq_alloc_map_and_rqs(set, i))
4584 goto out_unwind;
4585 cond_resched();
4586 }
4587
4588 return 0;
4589
4590 out_unwind:
4591 while (--i >= 0)
4592 __blk_mq_free_map_and_rqs(set, i);
4593
4594 if (blk_mq_is_shared_tags(set->flags)) {
4595 blk_mq_free_map_and_rqs(set, set->shared_tags,
4596 BLK_MQ_NO_HCTX_IDX);
4597 }
4598
4599 return -ENOMEM;
4600 }
4601
4602 /*
4603 * Allocate the request maps associated with this tag_set. Note that this
4604 * may reduce the depth asked for, if memory is tight. set->queue_depth
4605 * will be updated to reflect the allocated depth.
4606 */
blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set * set)4607 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4608 {
4609 unsigned int depth;
4610 int err;
4611
4612 depth = set->queue_depth;
4613 do {
4614 err = __blk_mq_alloc_rq_maps(set);
4615 if (!err)
4616 break;
4617
4618 set->queue_depth >>= 1;
4619 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4620 err = -ENOMEM;
4621 break;
4622 }
4623 } while (set->queue_depth);
4624
4625 if (!set->queue_depth || err) {
4626 pr_err("blk-mq: failed to allocate request map\n");
4627 return -ENOMEM;
4628 }
4629
4630 if (depth != set->queue_depth)
4631 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4632 depth, set->queue_depth);
4633
4634 return 0;
4635 }
4636
blk_mq_update_queue_map(struct blk_mq_tag_set * set)4637 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4638 {
4639 /*
4640 * blk_mq_map_queues() and multiple .map_queues() implementations
4641 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4642 * number of hardware queues.
4643 */
4644 if (set->nr_maps == 1)
4645 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4646
4647 if (set->ops->map_queues) {
4648 int i;
4649
4650 /*
4651 * transport .map_queues is usually done in the following
4652 * way:
4653 *
4654 * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4655 * mask = get_cpu_mask(queue)
4656 * for_each_cpu(cpu, mask)
4657 * set->map[x].mq_map[cpu] = queue;
4658 * }
4659 *
4660 * When we need to remap, the table has to be cleared for
4661 * killing stale mapping since one CPU may not be mapped
4662 * to any hw queue.
4663 */
4664 for (i = 0; i < set->nr_maps; i++)
4665 blk_mq_clear_mq_map(&set->map[i]);
4666
4667 set->ops->map_queues(set);
4668 } else {
4669 BUG_ON(set->nr_maps > 1);
4670 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4671 }
4672 }
4673
blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set * set,int new_nr_hw_queues)4674 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4675 int new_nr_hw_queues)
4676 {
4677 struct blk_mq_tags **new_tags;
4678 int i;
4679
4680 if (set->nr_hw_queues >= new_nr_hw_queues)
4681 goto done;
4682
4683 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4684 GFP_KERNEL, set->numa_node);
4685 if (!new_tags)
4686 return -ENOMEM;
4687
4688 if (set->tags)
4689 memcpy(new_tags, set->tags, set->nr_hw_queues *
4690 sizeof(*set->tags));
4691 kfree(set->tags);
4692 set->tags = new_tags;
4693
4694 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) {
4695 if (!__blk_mq_alloc_map_and_rqs(set, i)) {
4696 while (--i >= set->nr_hw_queues)
4697 __blk_mq_free_map_and_rqs(set, i);
4698 return -ENOMEM;
4699 }
4700 cond_resched();
4701 }
4702
4703 done:
4704 set->nr_hw_queues = new_nr_hw_queues;
4705 return 0;
4706 }
4707
4708 /*
4709 * Alloc a tag set to be associated with one or more request queues.
4710 * May fail with EINVAL for various error conditions. May adjust the
4711 * requested depth down, if it's too large. In that case, the set
4712 * value will be stored in set->queue_depth.
4713 */
blk_mq_alloc_tag_set(struct blk_mq_tag_set * set)4714 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4715 {
4716 int i, ret;
4717
4718 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4719
4720 if (!set->nr_hw_queues)
4721 return -EINVAL;
4722 if (!set->queue_depth)
4723 return -EINVAL;
4724 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4725 return -EINVAL;
4726
4727 if (!set->ops->queue_rq)
4728 return -EINVAL;
4729
4730 if (!set->ops->get_budget ^ !set->ops->put_budget)
4731 return -EINVAL;
4732
4733 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4734 pr_info("blk-mq: reduced tag depth to %u\n",
4735 BLK_MQ_MAX_DEPTH);
4736 set->queue_depth = BLK_MQ_MAX_DEPTH;
4737 }
4738
4739 if (!set->nr_maps)
4740 set->nr_maps = 1;
4741 else if (set->nr_maps > HCTX_MAX_TYPES)
4742 return -EINVAL;
4743
4744 /*
4745 * If a crashdump is active, then we are potentially in a very
4746 * memory constrained environment. Limit us to 64 tags to prevent
4747 * using too much memory.
4748 */
4749 if (is_kdump_kernel())
4750 set->queue_depth = min(64U, set->queue_depth);
4751
4752 /*
4753 * There is no use for more h/w queues than cpus if we just have
4754 * a single map
4755 */
4756 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4757 set->nr_hw_queues = nr_cpu_ids;
4758
4759 if (set->flags & BLK_MQ_F_BLOCKING) {
4760 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4761 if (!set->srcu)
4762 return -ENOMEM;
4763 ret = init_srcu_struct(set->srcu);
4764 if (ret)
4765 goto out_free_srcu;
4766 }
4767
4768 ret = -ENOMEM;
4769 set->tags = kcalloc_node(set->nr_hw_queues,
4770 sizeof(struct blk_mq_tags *), GFP_KERNEL,
4771 set->numa_node);
4772 if (!set->tags)
4773 goto out_cleanup_srcu;
4774
4775 for (i = 0; i < set->nr_maps; i++) {
4776 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4777 sizeof(set->map[i].mq_map[0]),
4778 GFP_KERNEL, set->numa_node);
4779 if (!set->map[i].mq_map)
4780 goto out_free_mq_map;
4781 set->map[i].nr_queues = set->nr_hw_queues;
4782 }
4783
4784 blk_mq_update_queue_map(set);
4785
4786 ret = blk_mq_alloc_set_map_and_rqs(set);
4787 if (ret)
4788 goto out_free_mq_map;
4789
4790 mutex_init(&set->tag_list_lock);
4791 INIT_LIST_HEAD(&set->tag_list);
4792
4793 return 0;
4794
4795 out_free_mq_map:
4796 for (i = 0; i < set->nr_maps; i++) {
4797 kfree(set->map[i].mq_map);
4798 set->map[i].mq_map = NULL;
4799 }
4800 kfree(set->tags);
4801 set->tags = NULL;
4802 out_cleanup_srcu:
4803 if (set->flags & BLK_MQ_F_BLOCKING)
4804 cleanup_srcu_struct(set->srcu);
4805 out_free_srcu:
4806 if (set->flags & BLK_MQ_F_BLOCKING)
4807 kfree(set->srcu);
4808 return ret;
4809 }
4810 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4811
4812 /* allocate and initialize a tagset for a simple single-queue device */
blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set * set,const struct blk_mq_ops * ops,unsigned int queue_depth,unsigned int set_flags)4813 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4814 const struct blk_mq_ops *ops, unsigned int queue_depth,
4815 unsigned int set_flags)
4816 {
4817 memset(set, 0, sizeof(*set));
4818 set->ops = ops;
4819 set->nr_hw_queues = 1;
4820 set->nr_maps = 1;
4821 set->queue_depth = queue_depth;
4822 set->numa_node = NUMA_NO_NODE;
4823 set->flags = set_flags;
4824 return blk_mq_alloc_tag_set(set);
4825 }
4826 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4827
blk_mq_free_tag_set(struct blk_mq_tag_set * set)4828 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4829 {
4830 int i, j;
4831
4832 for (i = 0; i < set->nr_hw_queues; i++)
4833 __blk_mq_free_map_and_rqs(set, i);
4834
4835 if (blk_mq_is_shared_tags(set->flags)) {
4836 blk_mq_free_map_and_rqs(set, set->shared_tags,
4837 BLK_MQ_NO_HCTX_IDX);
4838 }
4839
4840 for (j = 0; j < set->nr_maps; j++) {
4841 kfree(set->map[j].mq_map);
4842 set->map[j].mq_map = NULL;
4843 }
4844
4845 kfree(set->tags);
4846 set->tags = NULL;
4847 if (set->flags & BLK_MQ_F_BLOCKING) {
4848 cleanup_srcu_struct(set->srcu);
4849 kfree(set->srcu);
4850 }
4851 }
4852 EXPORT_SYMBOL(blk_mq_free_tag_set);
4853
blk_mq_update_nr_requests(struct request_queue * q,unsigned int nr)4854 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4855 {
4856 struct blk_mq_tag_set *set = q->tag_set;
4857 struct blk_mq_hw_ctx *hctx;
4858 int ret;
4859 unsigned long i;
4860
4861 if (WARN_ON_ONCE(!q->mq_freeze_depth))
4862 return -EINVAL;
4863
4864 if (!set)
4865 return -EINVAL;
4866
4867 if (q->nr_requests == nr)
4868 return 0;
4869
4870 blk_mq_quiesce_queue(q);
4871
4872 ret = 0;
4873 queue_for_each_hw_ctx(q, hctx, i) {
4874 if (!hctx->tags)
4875 continue;
4876 /*
4877 * If we're using an MQ scheduler, just update the scheduler
4878 * queue depth. This is similar to what the old code would do.
4879 */
4880 if (hctx->sched_tags) {
4881 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4882 nr, true);
4883 } else {
4884 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4885 false);
4886 }
4887 if (ret)
4888 break;
4889 if (q->elevator && q->elevator->type->ops.depth_updated)
4890 q->elevator->type->ops.depth_updated(hctx);
4891 }
4892 if (!ret) {
4893 q->nr_requests = nr;
4894 if (blk_mq_is_shared_tags(set->flags)) {
4895 if (q->elevator)
4896 blk_mq_tag_update_sched_shared_tags(q);
4897 else
4898 blk_mq_tag_resize_shared_tags(set, nr);
4899 }
4900 }
4901
4902 blk_mq_unquiesce_queue(q);
4903
4904 return ret;
4905 }
4906
4907 /*
4908 * request_queue and elevator_type pair.
4909 * It is just used by __blk_mq_update_nr_hw_queues to cache
4910 * the elevator_type associated with a request_queue.
4911 */
4912 struct blk_mq_qe_pair {
4913 struct list_head node;
4914 struct request_queue *q;
4915 struct elevator_type *type;
4916 };
4917
4918 /*
4919 * Cache the elevator_type in qe pair list and switch the
4920 * io scheduler to 'none'
4921 */
blk_mq_elv_switch_none(struct list_head * head,struct request_queue * q)4922 static bool blk_mq_elv_switch_none(struct list_head *head,
4923 struct request_queue *q)
4924 {
4925 struct blk_mq_qe_pair *qe;
4926
4927 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4928 if (!qe)
4929 return false;
4930
4931 /* q->elevator needs protection from ->sysfs_lock */
4932 mutex_lock(&q->sysfs_lock);
4933
4934 /* the check has to be done with holding sysfs_lock */
4935 if (!q->elevator) {
4936 kfree(qe);
4937 goto unlock;
4938 }
4939
4940 INIT_LIST_HEAD(&qe->node);
4941 qe->q = q;
4942 qe->type = q->elevator->type;
4943 /* keep a reference to the elevator module as we'll switch back */
4944 __elevator_get(qe->type);
4945 list_add(&qe->node, head);
4946 elevator_disable(q);
4947 unlock:
4948 mutex_unlock(&q->sysfs_lock);
4949
4950 return true;
4951 }
4952
blk_lookup_qe_pair(struct list_head * head,struct request_queue * q)4953 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4954 struct request_queue *q)
4955 {
4956 struct blk_mq_qe_pair *qe;
4957
4958 list_for_each_entry(qe, head, node)
4959 if (qe->q == q)
4960 return qe;
4961
4962 return NULL;
4963 }
4964
blk_mq_elv_switch_back(struct list_head * head,struct request_queue * q)4965 static void blk_mq_elv_switch_back(struct list_head *head,
4966 struct request_queue *q)
4967 {
4968 struct blk_mq_qe_pair *qe;
4969 struct elevator_type *t;
4970
4971 qe = blk_lookup_qe_pair(head, q);
4972 if (!qe)
4973 return;
4974 t = qe->type;
4975 list_del(&qe->node);
4976 kfree(qe);
4977
4978 mutex_lock(&q->sysfs_lock);
4979 elevator_switch(q, t);
4980 /* drop the reference acquired in blk_mq_elv_switch_none */
4981 elevator_put(t);
4982 mutex_unlock(&q->sysfs_lock);
4983 }
4984
__blk_mq_update_nr_hw_queues(struct blk_mq_tag_set * set,int nr_hw_queues)4985 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4986 int nr_hw_queues)
4987 {
4988 struct request_queue *q;
4989 LIST_HEAD(head);
4990 int prev_nr_hw_queues = set->nr_hw_queues;
4991 int i;
4992
4993 lockdep_assert_held(&set->tag_list_lock);
4994
4995 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4996 nr_hw_queues = nr_cpu_ids;
4997 if (nr_hw_queues < 1)
4998 return;
4999 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
5000 return;
5001
5002 list_for_each_entry(q, &set->tag_list, tag_set_list)
5003 blk_mq_freeze_queue(q);
5004 /*
5005 * Switch IO scheduler to 'none', cleaning up the data associated
5006 * with the previous scheduler. We will switch back once we are done
5007 * updating the new sw to hw queue mappings.
5008 */
5009 list_for_each_entry(q, &set->tag_list, tag_set_list)
5010 if (!blk_mq_elv_switch_none(&head, q))
5011 goto switch_back;
5012
5013 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5014 blk_mq_debugfs_unregister_hctxs(q);
5015 blk_mq_sysfs_unregister_hctxs(q);
5016 }
5017
5018 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
5019 goto reregister;
5020
5021 fallback:
5022 blk_mq_update_queue_map(set);
5023 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5024 struct queue_limits lim;
5025
5026 blk_mq_realloc_hw_ctxs(set, q);
5027
5028 if (q->nr_hw_queues != set->nr_hw_queues) {
5029 int i = prev_nr_hw_queues;
5030
5031 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
5032 nr_hw_queues, prev_nr_hw_queues);
5033 for (; i < set->nr_hw_queues; i++)
5034 __blk_mq_free_map_and_rqs(set, i);
5035
5036 set->nr_hw_queues = prev_nr_hw_queues;
5037 goto fallback;
5038 }
5039 lim = queue_limits_start_update(q);
5040 if (blk_mq_can_poll(set))
5041 lim.features |= BLK_FEAT_POLL;
5042 else
5043 lim.features &= ~BLK_FEAT_POLL;
5044 if (queue_limits_commit_update(q, &lim) < 0)
5045 pr_warn("updating the poll flag failed\n");
5046 blk_mq_map_swqueue(q);
5047 }
5048
5049 reregister:
5050 list_for_each_entry(q, &set->tag_list, tag_set_list) {
5051 blk_mq_sysfs_register_hctxs(q);
5052 blk_mq_debugfs_register_hctxs(q);
5053 }
5054
5055 switch_back:
5056 list_for_each_entry(q, &set->tag_list, tag_set_list)
5057 blk_mq_elv_switch_back(&head, q);
5058
5059 list_for_each_entry(q, &set->tag_list, tag_set_list)
5060 blk_mq_unfreeze_queue(q);
5061
5062 /* Free the excess tags when nr_hw_queues shrink. */
5063 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
5064 __blk_mq_free_map_and_rqs(set, i);
5065 }
5066
blk_mq_update_nr_hw_queues(struct blk_mq_tag_set * set,int nr_hw_queues)5067 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
5068 {
5069 mutex_lock(&set->tag_list_lock);
5070 __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
5071 mutex_unlock(&set->tag_list_lock);
5072 }
5073 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
5074
blk_hctx_poll(struct request_queue * q,struct blk_mq_hw_ctx * hctx,struct io_comp_batch * iob,unsigned int flags)5075 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
5076 struct io_comp_batch *iob, unsigned int flags)
5077 {
5078 long state = get_current_state();
5079 int ret;
5080
5081 do {
5082 ret = q->mq_ops->poll(hctx, iob);
5083 if (ret > 0) {
5084 __set_current_state(TASK_RUNNING);
5085 return ret;
5086 }
5087
5088 if (signal_pending_state(state, current))
5089 __set_current_state(TASK_RUNNING);
5090 if (task_is_running(current))
5091 return 1;
5092
5093 if (ret < 0 || (flags & BLK_POLL_ONESHOT))
5094 break;
5095 cpu_relax();
5096 } while (!need_resched());
5097
5098 __set_current_state(TASK_RUNNING);
5099 return 0;
5100 }
5101
blk_mq_poll(struct request_queue * q,blk_qc_t cookie,struct io_comp_batch * iob,unsigned int flags)5102 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
5103 struct io_comp_batch *iob, unsigned int flags)
5104 {
5105 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
5106
5107 return blk_hctx_poll(q, hctx, iob, flags);
5108 }
5109
blk_rq_poll(struct request * rq,struct io_comp_batch * iob,unsigned int poll_flags)5110 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
5111 unsigned int poll_flags)
5112 {
5113 struct request_queue *q = rq->q;
5114 int ret;
5115
5116 if (!blk_rq_is_poll(rq))
5117 return 0;
5118 if (!percpu_ref_tryget(&q->q_usage_counter))
5119 return 0;
5120
5121 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
5122 blk_queue_exit(q);
5123
5124 return ret;
5125 }
5126 EXPORT_SYMBOL_GPL(blk_rq_poll);
5127
blk_mq_rq_cpu(struct request * rq)5128 unsigned int blk_mq_rq_cpu(struct request *rq)
5129 {
5130 return rq->mq_ctx->cpu;
5131 }
5132 EXPORT_SYMBOL(blk_mq_rq_cpu);
5133
blk_mq_cancel_work_sync(struct request_queue * q)5134 void blk_mq_cancel_work_sync(struct request_queue *q)
5135 {
5136 struct blk_mq_hw_ctx *hctx;
5137 unsigned long i;
5138
5139 cancel_delayed_work_sync(&q->requeue_work);
5140
5141 queue_for_each_hw_ctx(q, hctx, i)
5142 cancel_delayed_work_sync(&hctx->run_work);
5143 }
5144
blk_mq_init(void)5145 static int __init blk_mq_init(void)
5146 {
5147 int i;
5148
5149 for_each_possible_cpu(i)
5150 init_llist_head(&per_cpu(blk_cpu_done, i));
5151 for_each_possible_cpu(i)
5152 INIT_CSD(&per_cpu(blk_cpu_csd, i),
5153 __blk_mq_complete_request_remote, NULL);
5154 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
5155
5156 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
5157 "block/softirq:dead", NULL,
5158 blk_softirq_cpu_dead);
5159 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
5160 blk_mq_hctx_notify_dead);
5161 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
5162 blk_mq_hctx_notify_online,
5163 blk_mq_hctx_notify_offline);
5164 return 0;
5165 }
5166 subsys_initcall(blk_mq_init);
5167