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