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