Line data Source code
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/t10-pi.h>
36 : #include "blk.h"
37 : #include "blk-mq.h"
38 : #include "blk-mq-debugfs.h"
39 : #include "blk-pm.h"
40 : #include "blk-stat.h"
41 : #include "blk-mq-sched.h"
42 : #include "blk-rq-qos.h"
43 : #include "blk-ioprio.h"
44 :
45 : static DEFINE_PER_CPU(struct llist_head, blk_cpu_done);
46 :
47 : static void blk_mq_insert_request(struct request *rq, blk_insert_t flags);
48 : static void blk_mq_request_bypass_insert(struct request *rq,
49 : blk_insert_t flags);
50 : static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
51 : struct list_head *list);
52 : static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
53 : struct io_comp_batch *iob, unsigned int flags);
54 :
55 : /*
56 : * Check if any of the ctx, dispatch list or elevator
57 : * have pending work in this hardware queue.
58 : */
59 0 : static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
60 : {
61 0 : return !list_empty_careful(&hctx->dispatch) ||
62 0 : sbitmap_any_bit_set(&hctx->ctx_map) ||
63 0 : blk_mq_sched_has_work(hctx);
64 : }
65 :
66 : /*
67 : * Mark this ctx as having pending work in this hardware queue
68 : */
69 0 : static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 : struct blk_mq_ctx *ctx)
71 : {
72 0 : const int bit = ctx->index_hw[hctx->type];
73 :
74 0 : if (!sbitmap_test_bit(&hctx->ctx_map, bit))
75 0 : sbitmap_set_bit(&hctx->ctx_map, bit);
76 0 : }
77 :
78 : static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
79 : struct blk_mq_ctx *ctx)
80 : {
81 0 : const int bit = ctx->index_hw[hctx->type];
82 :
83 0 : sbitmap_clear_bit(&hctx->ctx_map, bit);
84 : }
85 :
86 : struct mq_inflight {
87 : struct block_device *part;
88 : unsigned int inflight[2];
89 : };
90 :
91 0 : static bool blk_mq_check_inflight(struct request *rq, void *priv)
92 : {
93 0 : struct mq_inflight *mi = priv;
94 :
95 0 : if (rq->part && blk_do_io_stat(rq) &&
96 0 : (!mi->part->bd_partno || rq->part == mi->part) &&
97 0 : blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
98 0 : mi->inflight[rq_data_dir(rq)]++;
99 :
100 0 : return true;
101 : }
102 :
103 0 : unsigned int blk_mq_in_flight(struct request_queue *q,
104 : struct block_device *part)
105 : {
106 0 : struct mq_inflight mi = { .part = part };
107 :
108 0 : blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
109 :
110 0 : return mi.inflight[0] + mi.inflight[1];
111 : }
112 :
113 0 : void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
114 : unsigned int inflight[2])
115 : {
116 0 : struct mq_inflight mi = { .part = part };
117 :
118 0 : blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
119 0 : inflight[0] = mi.inflight[0];
120 0 : inflight[1] = mi.inflight[1];
121 0 : }
122 :
123 0 : void blk_freeze_queue_start(struct request_queue *q)
124 : {
125 0 : mutex_lock(&q->mq_freeze_lock);
126 0 : if (++q->mq_freeze_depth == 1) {
127 0 : percpu_ref_kill(&q->q_usage_counter);
128 0 : mutex_unlock(&q->mq_freeze_lock);
129 0 : if (queue_is_mq(q))
130 0 : blk_mq_run_hw_queues(q, false);
131 : } else {
132 0 : mutex_unlock(&q->mq_freeze_lock);
133 : }
134 0 : }
135 : EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
136 :
137 0 : void blk_mq_freeze_queue_wait(struct request_queue *q)
138 : {
139 0 : wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
140 0 : }
141 : EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
142 :
143 0 : int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
144 : unsigned long timeout)
145 : {
146 0 : return wait_event_timeout(q->mq_freeze_wq,
147 : percpu_ref_is_zero(&q->q_usage_counter),
148 : timeout);
149 : }
150 : EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
151 :
152 : /*
153 : * Guarantee no request is in use, so we can change any data structure of
154 : * the queue afterward.
155 : */
156 0 : void blk_freeze_queue(struct request_queue *q)
157 : {
158 : /*
159 : * In the !blk_mq case we are only calling this to kill the
160 : * q_usage_counter, otherwise this increases the freeze depth
161 : * and waits for it to return to zero. For this reason there is
162 : * no blk_unfreeze_queue(), and blk_freeze_queue() is not
163 : * exported to drivers as the only user for unfreeze is blk_mq.
164 : */
165 0 : blk_freeze_queue_start(q);
166 0 : blk_mq_freeze_queue_wait(q);
167 0 : }
168 :
169 0 : void blk_mq_freeze_queue(struct request_queue *q)
170 : {
171 : /*
172 : * ...just an alias to keep freeze and unfreeze actions balanced
173 : * in the blk_mq_* namespace
174 : */
175 0 : blk_freeze_queue(q);
176 0 : }
177 : EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
178 :
179 0 : void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
180 : {
181 0 : mutex_lock(&q->mq_freeze_lock);
182 0 : if (force_atomic)
183 0 : q->q_usage_counter.data->force_atomic = true;
184 0 : q->mq_freeze_depth--;
185 0 : WARN_ON_ONCE(q->mq_freeze_depth < 0);
186 0 : if (!q->mq_freeze_depth) {
187 0 : percpu_ref_resurrect(&q->q_usage_counter);
188 0 : wake_up_all(&q->mq_freeze_wq);
189 : }
190 0 : mutex_unlock(&q->mq_freeze_lock);
191 0 : }
192 :
193 0 : void blk_mq_unfreeze_queue(struct request_queue *q)
194 : {
195 0 : __blk_mq_unfreeze_queue(q, false);
196 0 : }
197 : EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
198 :
199 : /*
200 : * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
201 : * mpt3sas driver such that this function can be removed.
202 : */
203 0 : void blk_mq_quiesce_queue_nowait(struct request_queue *q)
204 : {
205 : unsigned long flags;
206 :
207 0 : spin_lock_irqsave(&q->queue_lock, flags);
208 0 : if (!q->quiesce_depth++)
209 0 : blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
210 0 : spin_unlock_irqrestore(&q->queue_lock, flags);
211 0 : }
212 : EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
213 :
214 : /**
215 : * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done
216 : * @set: tag_set to wait on
217 : *
218 : * Note: it is driver's responsibility for making sure that quiesce has
219 : * been started on or more of the request_queues of the tag_set. This
220 : * function only waits for the quiesce on those request_queues that had
221 : * the quiesce flag set using blk_mq_quiesce_queue_nowait.
222 : */
223 0 : void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)
224 : {
225 0 : if (set->flags & BLK_MQ_F_BLOCKING)
226 0 : synchronize_srcu(set->srcu);
227 : else
228 0 : synchronize_rcu();
229 0 : }
230 : EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done);
231 :
232 : /**
233 : * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
234 : * @q: request queue.
235 : *
236 : * Note: this function does not prevent that the struct request end_io()
237 : * callback function is invoked. Once this function is returned, we make
238 : * sure no dispatch can happen until the queue is unquiesced via
239 : * blk_mq_unquiesce_queue().
240 : */
241 0 : void blk_mq_quiesce_queue(struct request_queue *q)
242 : {
243 0 : blk_mq_quiesce_queue_nowait(q);
244 : /* nothing to wait for non-mq queues */
245 0 : if (queue_is_mq(q))
246 0 : blk_mq_wait_quiesce_done(q->tag_set);
247 0 : }
248 : EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
249 :
250 : /*
251 : * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
252 : * @q: request queue.
253 : *
254 : * This function recovers queue into the state before quiescing
255 : * which is done by blk_mq_quiesce_queue.
256 : */
257 0 : void blk_mq_unquiesce_queue(struct request_queue *q)
258 : {
259 : unsigned long flags;
260 0 : bool run_queue = false;
261 :
262 0 : spin_lock_irqsave(&q->queue_lock, flags);
263 0 : if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
264 : ;
265 0 : } else if (!--q->quiesce_depth) {
266 0 : blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
267 0 : run_queue = true;
268 : }
269 0 : spin_unlock_irqrestore(&q->queue_lock, flags);
270 :
271 : /* dispatch requests which are inserted during quiescing */
272 0 : if (run_queue)
273 0 : blk_mq_run_hw_queues(q, true);
274 0 : }
275 : EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
276 :
277 0 : void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set)
278 : {
279 : struct request_queue *q;
280 :
281 0 : mutex_lock(&set->tag_list_lock);
282 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
283 0 : if (!blk_queue_skip_tagset_quiesce(q))
284 0 : blk_mq_quiesce_queue_nowait(q);
285 : }
286 0 : blk_mq_wait_quiesce_done(set);
287 0 : mutex_unlock(&set->tag_list_lock);
288 0 : }
289 : EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset);
290 :
291 0 : void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set)
292 : {
293 : struct request_queue *q;
294 :
295 0 : mutex_lock(&set->tag_list_lock);
296 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
297 0 : if (!blk_queue_skip_tagset_quiesce(q))
298 0 : blk_mq_unquiesce_queue(q);
299 : }
300 0 : mutex_unlock(&set->tag_list_lock);
301 0 : }
302 : EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset);
303 :
304 0 : void blk_mq_wake_waiters(struct request_queue *q)
305 : {
306 : struct blk_mq_hw_ctx *hctx;
307 : unsigned long i;
308 :
309 0 : queue_for_each_hw_ctx(q, hctx, i)
310 0 : if (blk_mq_hw_queue_mapped(hctx))
311 0 : blk_mq_tag_wakeup_all(hctx->tags, true);
312 0 : }
313 :
314 0 : void blk_rq_init(struct request_queue *q, struct request *rq)
315 : {
316 0 : memset(rq, 0, sizeof(*rq));
317 :
318 0 : INIT_LIST_HEAD(&rq->queuelist);
319 0 : rq->q = q;
320 0 : rq->__sector = (sector_t) -1;
321 0 : INIT_HLIST_NODE(&rq->hash);
322 0 : RB_CLEAR_NODE(&rq->rb_node);
323 0 : rq->tag = BLK_MQ_NO_TAG;
324 0 : rq->internal_tag = BLK_MQ_NO_TAG;
325 0 : rq->start_time_ns = ktime_get_ns();
326 0 : rq->part = NULL;
327 0 : blk_crypto_rq_set_defaults(rq);
328 0 : }
329 : EXPORT_SYMBOL(blk_rq_init);
330 :
331 : /* Set start and alloc time when the allocated request is actually used */
332 : static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns)
333 : {
334 0 : if (blk_mq_need_time_stamp(rq))
335 0 : rq->start_time_ns = ktime_get_ns();
336 : else
337 0 : rq->start_time_ns = 0;
338 :
339 : #ifdef CONFIG_BLK_RQ_ALLOC_TIME
340 : if (blk_queue_rq_alloc_time(rq->q))
341 : rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns;
342 : else
343 : rq->alloc_time_ns = 0;
344 : #endif
345 : }
346 :
347 0 : static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
348 : struct blk_mq_tags *tags, unsigned int tag)
349 : {
350 0 : struct blk_mq_ctx *ctx = data->ctx;
351 0 : struct blk_mq_hw_ctx *hctx = data->hctx;
352 0 : struct request_queue *q = data->q;
353 0 : struct request *rq = tags->static_rqs[tag];
354 :
355 0 : rq->q = q;
356 0 : rq->mq_ctx = ctx;
357 0 : rq->mq_hctx = hctx;
358 0 : rq->cmd_flags = data->cmd_flags;
359 :
360 0 : if (data->flags & BLK_MQ_REQ_PM)
361 0 : data->rq_flags |= RQF_PM;
362 0 : if (blk_queue_io_stat(q))
363 0 : data->rq_flags |= RQF_IO_STAT;
364 0 : rq->rq_flags = data->rq_flags;
365 :
366 0 : if (data->rq_flags & RQF_SCHED_TAGS) {
367 0 : rq->tag = BLK_MQ_NO_TAG;
368 0 : rq->internal_tag = tag;
369 : } else {
370 0 : rq->tag = tag;
371 0 : rq->internal_tag = BLK_MQ_NO_TAG;
372 : }
373 0 : rq->timeout = 0;
374 :
375 0 : rq->part = NULL;
376 0 : rq->io_start_time_ns = 0;
377 0 : rq->stats_sectors = 0;
378 0 : rq->nr_phys_segments = 0;
379 : #if defined(CONFIG_BLK_DEV_INTEGRITY)
380 : rq->nr_integrity_segments = 0;
381 : #endif
382 0 : rq->end_io = NULL;
383 0 : rq->end_io_data = NULL;
384 :
385 0 : blk_crypto_rq_set_defaults(rq);
386 0 : INIT_LIST_HEAD(&rq->queuelist);
387 : /* tag was already set */
388 0 : WRITE_ONCE(rq->deadline, 0);
389 0 : req_ref_set(rq, 1);
390 :
391 0 : if (rq->rq_flags & RQF_USE_SCHED) {
392 0 : struct elevator_queue *e = data->q->elevator;
393 :
394 0 : INIT_HLIST_NODE(&rq->hash);
395 0 : RB_CLEAR_NODE(&rq->rb_node);
396 :
397 0 : if (e->type->ops.prepare_request)
398 0 : e->type->ops.prepare_request(rq);
399 : }
400 :
401 0 : return rq;
402 : }
403 :
404 : static inline struct request *
405 0 : __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
406 : {
407 : unsigned int tag, tag_offset;
408 : struct blk_mq_tags *tags;
409 : struct request *rq;
410 : unsigned long tag_mask;
411 0 : int i, nr = 0;
412 :
413 0 : tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset);
414 0 : if (unlikely(!tag_mask))
415 : return NULL;
416 :
417 0 : tags = blk_mq_tags_from_data(data);
418 0 : for (i = 0; tag_mask; i++) {
419 0 : if (!(tag_mask & (1UL << i)))
420 0 : continue;
421 0 : tag = tag_offset + i;
422 0 : prefetch(tags->static_rqs[tag]);
423 0 : tag_mask &= ~(1UL << i);
424 0 : rq = blk_mq_rq_ctx_init(data, tags, tag);
425 0 : rq_list_add(data->cached_rq, rq);
426 0 : nr++;
427 : }
428 : /* caller already holds a reference, add for remainder */
429 0 : percpu_ref_get_many(&data->q->q_usage_counter, nr - 1);
430 0 : data->nr_tags -= nr;
431 :
432 0 : return rq_list_pop(data->cached_rq);
433 : }
434 :
435 0 : static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
436 : {
437 0 : struct request_queue *q = data->q;
438 0 : u64 alloc_time_ns = 0;
439 : struct request *rq;
440 : unsigned int tag;
441 :
442 : /* alloc_time includes depth and tag waits */
443 : if (blk_queue_rq_alloc_time(q))
444 : alloc_time_ns = ktime_get_ns();
445 :
446 0 : if (data->cmd_flags & REQ_NOWAIT)
447 0 : data->flags |= BLK_MQ_REQ_NOWAIT;
448 :
449 0 : if (q->elevator) {
450 : /*
451 : * All requests use scheduler tags when an I/O scheduler is
452 : * enabled for the queue.
453 : */
454 0 : data->rq_flags |= RQF_SCHED_TAGS;
455 :
456 : /*
457 : * Flush/passthrough requests are special and go directly to the
458 : * dispatch list.
459 : */
460 0 : if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
461 0 : !blk_op_is_passthrough(data->cmd_flags)) {
462 0 : struct elevator_mq_ops *ops = &q->elevator->type->ops;
463 :
464 0 : WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
465 :
466 0 : data->rq_flags |= RQF_USE_SCHED;
467 0 : if (ops->limit_depth)
468 0 : ops->limit_depth(data->cmd_flags, data);
469 : }
470 : }
471 :
472 : retry:
473 0 : data->ctx = blk_mq_get_ctx(q);
474 0 : data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx);
475 0 : if (!(data->rq_flags & RQF_SCHED_TAGS))
476 0 : blk_mq_tag_busy(data->hctx);
477 :
478 0 : if (data->flags & BLK_MQ_REQ_RESERVED)
479 0 : data->rq_flags |= RQF_RESV;
480 :
481 : /*
482 : * Try batched alloc if we want more than 1 tag.
483 : */
484 0 : if (data->nr_tags > 1) {
485 0 : rq = __blk_mq_alloc_requests_batch(data);
486 0 : if (rq) {
487 0 : blk_mq_rq_time_init(rq, alloc_time_ns);
488 : return rq;
489 : }
490 0 : data->nr_tags = 1;
491 : }
492 :
493 : /*
494 : * Waiting allocations only fail because of an inactive hctx. In that
495 : * case just retry the hctx assignment and tag allocation as CPU hotplug
496 : * should have migrated us to an online CPU by now.
497 : */
498 0 : tag = blk_mq_get_tag(data);
499 0 : if (tag == BLK_MQ_NO_TAG) {
500 0 : if (data->flags & BLK_MQ_REQ_NOWAIT)
501 : return NULL;
502 : /*
503 : * Give up the CPU and sleep for a random short time to
504 : * ensure that thread using a realtime scheduling class
505 : * are migrated off the CPU, and thus off the hctx that
506 : * is going away.
507 : */
508 0 : msleep(3);
509 0 : goto retry;
510 : }
511 :
512 0 : rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag);
513 0 : blk_mq_rq_time_init(rq, alloc_time_ns);
514 : return rq;
515 : }
516 :
517 0 : static struct request *blk_mq_rq_cache_fill(struct request_queue *q,
518 : struct blk_plug *plug,
519 : blk_opf_t opf,
520 : blk_mq_req_flags_t flags)
521 : {
522 0 : struct blk_mq_alloc_data data = {
523 : .q = q,
524 : .flags = flags,
525 : .cmd_flags = opf,
526 0 : .nr_tags = plug->nr_ios,
527 0 : .cached_rq = &plug->cached_rq,
528 : };
529 : struct request *rq;
530 :
531 0 : if (blk_queue_enter(q, flags))
532 : return NULL;
533 :
534 0 : plug->nr_ios = 1;
535 :
536 0 : rq = __blk_mq_alloc_requests(&data);
537 0 : if (unlikely(!rq))
538 0 : blk_queue_exit(q);
539 : return rq;
540 : }
541 :
542 0 : static struct request *blk_mq_alloc_cached_request(struct request_queue *q,
543 : blk_opf_t opf,
544 : blk_mq_req_flags_t flags)
545 : {
546 0 : struct blk_plug *plug = current->plug;
547 : struct request *rq;
548 :
549 0 : if (!plug)
550 : return NULL;
551 :
552 0 : if (rq_list_empty(plug->cached_rq)) {
553 0 : if (plug->nr_ios == 1)
554 : return NULL;
555 0 : rq = blk_mq_rq_cache_fill(q, plug, opf, flags);
556 0 : if (!rq)
557 : return NULL;
558 : } else {
559 0 : rq = rq_list_peek(&plug->cached_rq);
560 0 : if (!rq || rq->q != q)
561 : return NULL;
562 :
563 0 : if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type)
564 : return NULL;
565 0 : if (op_is_flush(rq->cmd_flags) != op_is_flush(opf))
566 : return NULL;
567 :
568 0 : plug->cached_rq = rq_list_next(rq);
569 0 : blk_mq_rq_time_init(rq, 0);
570 : }
571 :
572 0 : rq->cmd_flags = opf;
573 0 : INIT_LIST_HEAD(&rq->queuelist);
574 0 : return rq;
575 : }
576 :
577 0 : struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf,
578 : blk_mq_req_flags_t flags)
579 : {
580 : struct request *rq;
581 :
582 0 : rq = blk_mq_alloc_cached_request(q, opf, flags);
583 0 : if (!rq) {
584 0 : struct blk_mq_alloc_data data = {
585 : .q = q,
586 : .flags = flags,
587 : .cmd_flags = opf,
588 : .nr_tags = 1,
589 : };
590 : int ret;
591 :
592 0 : ret = blk_queue_enter(q, flags);
593 0 : if (ret)
594 0 : return ERR_PTR(ret);
595 :
596 0 : rq = __blk_mq_alloc_requests(&data);
597 0 : if (!rq)
598 : goto out_queue_exit;
599 : }
600 0 : rq->__data_len = 0;
601 0 : rq->__sector = (sector_t) -1;
602 0 : rq->bio = rq->biotail = NULL;
603 0 : return rq;
604 : out_queue_exit:
605 0 : blk_queue_exit(q);
606 0 : return ERR_PTR(-EWOULDBLOCK);
607 : }
608 : EXPORT_SYMBOL(blk_mq_alloc_request);
609 :
610 0 : struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
611 : blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx)
612 : {
613 0 : struct blk_mq_alloc_data data = {
614 : .q = q,
615 : .flags = flags,
616 : .cmd_flags = opf,
617 : .nr_tags = 1,
618 : };
619 0 : u64 alloc_time_ns = 0;
620 : struct request *rq;
621 : unsigned int cpu;
622 : unsigned int tag;
623 : int ret;
624 :
625 : /* alloc_time includes depth and tag waits */
626 : if (blk_queue_rq_alloc_time(q))
627 : alloc_time_ns = ktime_get_ns();
628 :
629 : /*
630 : * If the tag allocator sleeps we could get an allocation for a
631 : * different hardware context. No need to complicate the low level
632 : * allocator for this for the rare use case of a command tied to
633 : * a specific queue.
634 : */
635 0 : if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) ||
636 0 : WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED)))
637 : return ERR_PTR(-EINVAL);
638 :
639 0 : if (hctx_idx >= q->nr_hw_queues)
640 : return ERR_PTR(-EIO);
641 :
642 0 : ret = blk_queue_enter(q, flags);
643 0 : if (ret)
644 0 : return ERR_PTR(ret);
645 :
646 : /*
647 : * Check if the hardware context is actually mapped to anything.
648 : * If not tell the caller that it should skip this queue.
649 : */
650 0 : ret = -EXDEV;
651 0 : data.hctx = xa_load(&q->hctx_table, hctx_idx);
652 0 : if (!blk_mq_hw_queue_mapped(data.hctx))
653 : goto out_queue_exit;
654 0 : cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
655 0 : if (cpu >= nr_cpu_ids)
656 : goto out_queue_exit;
657 0 : data.ctx = __blk_mq_get_ctx(q, cpu);
658 :
659 0 : if (q->elevator)
660 0 : data.rq_flags |= RQF_SCHED_TAGS;
661 : else
662 0 : blk_mq_tag_busy(data.hctx);
663 :
664 0 : if (flags & BLK_MQ_REQ_RESERVED)
665 0 : data.rq_flags |= RQF_RESV;
666 :
667 0 : ret = -EWOULDBLOCK;
668 0 : tag = blk_mq_get_tag(&data);
669 0 : if (tag == BLK_MQ_NO_TAG)
670 : goto out_queue_exit;
671 0 : rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag);
672 0 : blk_mq_rq_time_init(rq, alloc_time_ns);
673 0 : rq->__data_len = 0;
674 0 : rq->__sector = (sector_t) -1;
675 0 : rq->bio = rq->biotail = NULL;
676 0 : return rq;
677 :
678 : out_queue_exit:
679 0 : blk_queue_exit(q);
680 0 : return ERR_PTR(ret);
681 : }
682 : EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
683 :
684 0 : static void __blk_mq_free_request(struct request *rq)
685 : {
686 0 : struct request_queue *q = rq->q;
687 0 : struct blk_mq_ctx *ctx = rq->mq_ctx;
688 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
689 0 : const int sched_tag = rq->internal_tag;
690 :
691 0 : blk_crypto_free_request(rq);
692 0 : blk_pm_mark_last_busy(rq);
693 0 : rq->mq_hctx = NULL;
694 :
695 0 : if (rq->rq_flags & RQF_MQ_INFLIGHT)
696 : __blk_mq_dec_active_requests(hctx);
697 :
698 0 : if (rq->tag != BLK_MQ_NO_TAG)
699 0 : blk_mq_put_tag(hctx->tags, ctx, rq->tag);
700 0 : if (sched_tag != BLK_MQ_NO_TAG)
701 0 : blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag);
702 0 : blk_mq_sched_restart(hctx);
703 0 : blk_queue_exit(q);
704 0 : }
705 :
706 0 : void blk_mq_free_request(struct request *rq)
707 : {
708 0 : struct request_queue *q = rq->q;
709 :
710 0 : if ((rq->rq_flags & RQF_USE_SCHED) &&
711 0 : q->elevator->type->ops.finish_request)
712 0 : q->elevator->type->ops.finish_request(rq);
713 :
714 0 : if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
715 0 : laptop_io_completion(q->disk->bdi);
716 :
717 0 : rq_qos_done(q, rq);
718 :
719 0 : WRITE_ONCE(rq->state, MQ_RQ_IDLE);
720 0 : if (req_ref_put_and_test(rq))
721 0 : __blk_mq_free_request(rq);
722 0 : }
723 : EXPORT_SYMBOL_GPL(blk_mq_free_request);
724 :
725 0 : void blk_mq_free_plug_rqs(struct blk_plug *plug)
726 : {
727 : struct request *rq;
728 :
729 0 : while ((rq = rq_list_pop(&plug->cached_rq)) != NULL)
730 0 : blk_mq_free_request(rq);
731 0 : }
732 :
733 0 : void blk_dump_rq_flags(struct request *rq, char *msg)
734 : {
735 0 : printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg,
736 : rq->q->disk ? rq->q->disk->disk_name : "?",
737 : (__force unsigned long long) rq->cmd_flags);
738 :
739 0 : printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
740 : (unsigned long long)blk_rq_pos(rq),
741 : blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
742 0 : printk(KERN_INFO " bio %p, biotail %p, len %u\n",
743 : rq->bio, rq->biotail, blk_rq_bytes(rq));
744 0 : }
745 : EXPORT_SYMBOL(blk_dump_rq_flags);
746 :
747 0 : static void req_bio_endio(struct request *rq, struct bio *bio,
748 : unsigned int nbytes, blk_status_t error)
749 : {
750 0 : if (unlikely(error)) {
751 0 : bio->bi_status = error;
752 0 : } else if (req_op(rq) == REQ_OP_ZONE_APPEND) {
753 : /*
754 : * Partial zone append completions cannot be supported as the
755 : * BIO fragments may end up not being written sequentially.
756 : */
757 0 : if (bio->bi_iter.bi_size != nbytes)
758 0 : bio->bi_status = BLK_STS_IOERR;
759 : else
760 0 : bio->bi_iter.bi_sector = rq->__sector;
761 : }
762 :
763 0 : bio_advance(bio, nbytes);
764 :
765 0 : if (unlikely(rq->rq_flags & RQF_QUIET))
766 : bio_set_flag(bio, BIO_QUIET);
767 : /* don't actually finish bio if it's part of flush sequence */
768 0 : if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ))
769 0 : bio_endio(bio);
770 0 : }
771 :
772 0 : static void blk_account_io_completion(struct request *req, unsigned int bytes)
773 : {
774 0 : if (req->part && blk_do_io_stat(req)) {
775 0 : const int sgrp = op_stat_group(req_op(req));
776 :
777 0 : part_stat_lock();
778 0 : part_stat_add(req->part, sectors[sgrp], bytes >> 9);
779 0 : part_stat_unlock();
780 : }
781 0 : }
782 :
783 0 : static void blk_print_req_error(struct request *req, blk_status_t status)
784 : {
785 0 : printk_ratelimited(KERN_ERR
786 : "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x "
787 : "phys_seg %u prio class %u\n",
788 : blk_status_to_str(status),
789 : req->q->disk ? req->q->disk->disk_name : "?",
790 : blk_rq_pos(req), (__force u32)req_op(req),
791 : blk_op_str(req_op(req)),
792 : (__force u32)(req->cmd_flags & ~REQ_OP_MASK),
793 : req->nr_phys_segments,
794 : IOPRIO_PRIO_CLASS(req->ioprio));
795 0 : }
796 :
797 : /*
798 : * Fully end IO on a request. Does not support partial completions, or
799 : * errors.
800 : */
801 0 : static void blk_complete_request(struct request *req)
802 : {
803 0 : const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0;
804 0 : int total_bytes = blk_rq_bytes(req);
805 0 : struct bio *bio = req->bio;
806 :
807 0 : trace_block_rq_complete(req, BLK_STS_OK, total_bytes);
808 :
809 0 : if (!bio)
810 : return;
811 :
812 : #ifdef CONFIG_BLK_DEV_INTEGRITY
813 : if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ)
814 : req->q->integrity.profile->complete_fn(req, total_bytes);
815 : #endif
816 :
817 : /*
818 : * Upper layers may call blk_crypto_evict_key() anytime after the last
819 : * bio_endio(). Therefore, the keyslot must be released before that.
820 : */
821 0 : blk_crypto_rq_put_keyslot(req);
822 :
823 0 : blk_account_io_completion(req, total_bytes);
824 :
825 : do {
826 0 : struct bio *next = bio->bi_next;
827 :
828 : /* Completion has already been traced */
829 0 : bio_clear_flag(bio, BIO_TRACE_COMPLETION);
830 :
831 0 : if (req_op(req) == REQ_OP_ZONE_APPEND)
832 0 : bio->bi_iter.bi_sector = req->__sector;
833 :
834 0 : if (!is_flush)
835 0 : bio_endio(bio);
836 0 : bio = next;
837 0 : } while (bio);
838 :
839 : /*
840 : * Reset counters so that the request stacking driver
841 : * can find how many bytes remain in the request
842 : * later.
843 : */
844 0 : if (!req->end_io) {
845 0 : req->bio = NULL;
846 0 : req->__data_len = 0;
847 : }
848 : }
849 :
850 : /**
851 : * blk_update_request - Complete multiple bytes without completing the request
852 : * @req: the request being processed
853 : * @error: block status code
854 : * @nr_bytes: number of bytes to complete for @req
855 : *
856 : * Description:
857 : * Ends I/O on a number of bytes attached to @req, but doesn't complete
858 : * the request structure even if @req doesn't have leftover.
859 : * If @req has leftover, sets it up for the next range of segments.
860 : *
861 : * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
862 : * %false return from this function.
863 : *
864 : * Note:
865 : * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function
866 : * except in the consistency check at the end of this function.
867 : *
868 : * Return:
869 : * %false - this request doesn't have any more data
870 : * %true - this request has more data
871 : **/
872 0 : bool blk_update_request(struct request *req, blk_status_t error,
873 : unsigned int nr_bytes)
874 : {
875 : int total_bytes;
876 :
877 0 : trace_block_rq_complete(req, error, nr_bytes);
878 :
879 0 : if (!req->bio)
880 : return false;
881 :
882 : #ifdef CONFIG_BLK_DEV_INTEGRITY
883 : if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ &&
884 : error == BLK_STS_OK)
885 : req->q->integrity.profile->complete_fn(req, nr_bytes);
886 : #endif
887 :
888 : /*
889 : * Upper layers may call blk_crypto_evict_key() anytime after the last
890 : * bio_endio(). Therefore, the keyslot must be released before that.
891 : */
892 0 : if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req))
893 : __blk_crypto_rq_put_keyslot(req);
894 :
895 0 : if (unlikely(error && !blk_rq_is_passthrough(req) &&
896 0 : !(req->rq_flags & RQF_QUIET)) &&
897 0 : !test_bit(GD_DEAD, &req->q->disk->state)) {
898 0 : blk_print_req_error(req, error);
899 0 : trace_block_rq_error(req, error, nr_bytes);
900 : }
901 :
902 0 : blk_account_io_completion(req, nr_bytes);
903 :
904 0 : total_bytes = 0;
905 0 : while (req->bio) {
906 0 : struct bio *bio = req->bio;
907 0 : unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes);
908 :
909 0 : if (bio_bytes == bio->bi_iter.bi_size)
910 0 : req->bio = bio->bi_next;
911 :
912 : /* Completion has already been traced */
913 0 : bio_clear_flag(bio, BIO_TRACE_COMPLETION);
914 0 : req_bio_endio(req, bio, bio_bytes, error);
915 :
916 0 : total_bytes += bio_bytes;
917 0 : nr_bytes -= bio_bytes;
918 :
919 0 : if (!nr_bytes)
920 : break;
921 : }
922 :
923 : /*
924 : * completely done
925 : */
926 0 : if (!req->bio) {
927 : /*
928 : * Reset counters so that the request stacking driver
929 : * can find how many bytes remain in the request
930 : * later.
931 : */
932 0 : req->__data_len = 0;
933 0 : return false;
934 : }
935 :
936 0 : req->__data_len -= total_bytes;
937 :
938 : /* update sector only for requests with clear definition of sector */
939 0 : if (!blk_rq_is_passthrough(req))
940 0 : req->__sector += total_bytes >> 9;
941 :
942 : /* mixed attributes always follow the first bio */
943 0 : if (req->rq_flags & RQF_MIXED_MERGE) {
944 0 : req->cmd_flags &= ~REQ_FAILFAST_MASK;
945 0 : req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK;
946 : }
947 :
948 0 : if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) {
949 : /*
950 : * If total number of sectors is less than the first segment
951 : * size, something has gone terribly wrong.
952 : */
953 0 : if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
954 0 : blk_dump_rq_flags(req, "request botched");
955 0 : req->__data_len = blk_rq_cur_bytes(req);
956 : }
957 :
958 : /* recalculate the number of segments */
959 0 : req->nr_phys_segments = blk_recalc_rq_segments(req);
960 : }
961 :
962 : return true;
963 : }
964 : EXPORT_SYMBOL_GPL(blk_update_request);
965 :
966 0 : static inline void blk_account_io_done(struct request *req, u64 now)
967 : {
968 0 : trace_block_io_done(req);
969 :
970 : /*
971 : * Account IO completion. flush_rq isn't accounted as a
972 : * normal IO on queueing nor completion. Accounting the
973 : * containing request is enough.
974 : */
975 0 : if (blk_do_io_stat(req) && req->part &&
976 0 : !(req->rq_flags & RQF_FLUSH_SEQ)) {
977 0 : const int sgrp = op_stat_group(req_op(req));
978 :
979 0 : part_stat_lock();
980 0 : update_io_ticks(req->part, jiffies, true);
981 0 : part_stat_inc(req->part, ios[sgrp]);
982 0 : part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns);
983 0 : part_stat_unlock();
984 : }
985 0 : }
986 :
987 0 : static inline void blk_account_io_start(struct request *req)
988 : {
989 0 : trace_block_io_start(req);
990 :
991 0 : if (blk_do_io_stat(req)) {
992 : /*
993 : * All non-passthrough requests are created from a bio with one
994 : * exception: when a flush command that is part of a flush sequence
995 : * generated by the state machine in blk-flush.c is cloned onto the
996 : * lower device by dm-multipath we can get here without a bio.
997 : */
998 0 : if (req->bio)
999 0 : req->part = req->bio->bi_bdev;
1000 : else
1001 0 : req->part = req->q->disk->part0;
1002 :
1003 0 : part_stat_lock();
1004 0 : update_io_ticks(req->part, jiffies, false);
1005 0 : part_stat_unlock();
1006 : }
1007 0 : }
1008 :
1009 0 : static inline void __blk_mq_end_request_acct(struct request *rq, u64 now)
1010 : {
1011 0 : if (rq->rq_flags & RQF_STATS)
1012 0 : blk_stat_add(rq, now);
1013 :
1014 0 : blk_mq_sched_completed_request(rq, now);
1015 0 : blk_account_io_done(rq, now);
1016 0 : }
1017 :
1018 0 : inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
1019 : {
1020 0 : if (blk_mq_need_time_stamp(rq))
1021 0 : __blk_mq_end_request_acct(rq, ktime_get_ns());
1022 :
1023 0 : if (rq->end_io) {
1024 0 : rq_qos_done(rq->q, rq);
1025 0 : if (rq->end_io(rq, error) == RQ_END_IO_FREE)
1026 0 : blk_mq_free_request(rq);
1027 : } else {
1028 0 : blk_mq_free_request(rq);
1029 : }
1030 0 : }
1031 : EXPORT_SYMBOL(__blk_mq_end_request);
1032 :
1033 0 : void blk_mq_end_request(struct request *rq, blk_status_t error)
1034 : {
1035 0 : if (blk_update_request(rq, error, blk_rq_bytes(rq)))
1036 0 : BUG();
1037 0 : __blk_mq_end_request(rq, error);
1038 0 : }
1039 : EXPORT_SYMBOL(blk_mq_end_request);
1040 :
1041 : #define TAG_COMP_BATCH 32
1042 :
1043 0 : static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx,
1044 : int *tag_array, int nr_tags)
1045 : {
1046 0 : struct request_queue *q = hctx->queue;
1047 :
1048 : /*
1049 : * All requests should have been marked as RQF_MQ_INFLIGHT, so
1050 : * update hctx->nr_active in batch
1051 : */
1052 0 : if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
1053 : __blk_mq_sub_active_requests(hctx, nr_tags);
1054 :
1055 0 : blk_mq_put_tags(hctx->tags, tag_array, nr_tags);
1056 0 : percpu_ref_put_many(&q->q_usage_counter, nr_tags);
1057 0 : }
1058 :
1059 0 : void blk_mq_end_request_batch(struct io_comp_batch *iob)
1060 : {
1061 0 : int tags[TAG_COMP_BATCH], nr_tags = 0;
1062 0 : struct blk_mq_hw_ctx *cur_hctx = NULL;
1063 : struct request *rq;
1064 0 : u64 now = 0;
1065 :
1066 0 : if (iob->need_ts)
1067 0 : now = ktime_get_ns();
1068 :
1069 0 : while ((rq = rq_list_pop(&iob->req_list)) != NULL) {
1070 0 : prefetch(rq->bio);
1071 0 : prefetch(rq->rq_next);
1072 :
1073 0 : blk_complete_request(rq);
1074 0 : if (iob->need_ts)
1075 0 : __blk_mq_end_request_acct(rq, now);
1076 :
1077 0 : rq_qos_done(rq->q, rq);
1078 :
1079 : /*
1080 : * If end_io handler returns NONE, then it still has
1081 : * ownership of the request.
1082 : */
1083 0 : if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE)
1084 0 : continue;
1085 :
1086 0 : WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1087 0 : if (!req_ref_put_and_test(rq))
1088 0 : continue;
1089 :
1090 0 : blk_crypto_free_request(rq);
1091 0 : blk_pm_mark_last_busy(rq);
1092 :
1093 0 : if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
1094 0 : if (cur_hctx)
1095 0 : blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1096 0 : nr_tags = 0;
1097 0 : cur_hctx = rq->mq_hctx;
1098 : }
1099 0 : tags[nr_tags++] = rq->tag;
1100 : }
1101 :
1102 0 : if (nr_tags)
1103 0 : blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
1104 0 : }
1105 : EXPORT_SYMBOL_GPL(blk_mq_end_request_batch);
1106 :
1107 0 : static void blk_complete_reqs(struct llist_head *list)
1108 : {
1109 0 : struct llist_node *entry = llist_reverse_order(llist_del_all(list));
1110 : struct request *rq, *next;
1111 :
1112 0 : llist_for_each_entry_safe(rq, next, entry, ipi_list)
1113 0 : rq->q->mq_ops->complete(rq);
1114 0 : }
1115 :
1116 0 : static __latent_entropy void blk_done_softirq(struct softirq_action *h)
1117 : {
1118 0 : blk_complete_reqs(this_cpu_ptr(&blk_cpu_done));
1119 0 : }
1120 :
1121 0 : static int blk_softirq_cpu_dead(unsigned int cpu)
1122 : {
1123 0 : blk_complete_reqs(&per_cpu(blk_cpu_done, cpu));
1124 0 : return 0;
1125 : }
1126 :
1127 : static void __blk_mq_complete_request_remote(void *data)
1128 : {
1129 : __raise_softirq_irqoff(BLOCK_SOFTIRQ);
1130 : }
1131 :
1132 : static inline bool blk_mq_complete_need_ipi(struct request *rq)
1133 : {
1134 0 : int cpu = raw_smp_processor_id();
1135 :
1136 : if (!IS_ENABLED(CONFIG_SMP) ||
1137 : !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags))
1138 : return false;
1139 : /*
1140 : * With force threaded interrupts enabled, raising softirq from an SMP
1141 : * function call will always result in waking the ksoftirqd thread.
1142 : * This is probably worse than completing the request on a different
1143 : * cache domain.
1144 : */
1145 : if (force_irqthreads())
1146 : return false;
1147 :
1148 : /* same CPU or cache domain? Complete locally */
1149 : if (cpu == rq->mq_ctx->cpu ||
1150 : (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) &&
1151 : cpus_share_cache(cpu, rq->mq_ctx->cpu)))
1152 : return false;
1153 :
1154 : /* don't try to IPI to an offline CPU */
1155 : return cpu_online(rq->mq_ctx->cpu);
1156 : }
1157 :
1158 : static void blk_mq_complete_send_ipi(struct request *rq)
1159 : {
1160 : struct llist_head *list;
1161 : unsigned int cpu;
1162 :
1163 : cpu = rq->mq_ctx->cpu;
1164 : list = &per_cpu(blk_cpu_done, cpu);
1165 : if (llist_add(&rq->ipi_list, list)) {
1166 : INIT_CSD(&rq->csd, __blk_mq_complete_request_remote, rq);
1167 : smp_call_function_single_async(cpu, &rq->csd);
1168 : }
1169 : }
1170 :
1171 0 : static void blk_mq_raise_softirq(struct request *rq)
1172 : {
1173 : struct llist_head *list;
1174 :
1175 0 : preempt_disable();
1176 0 : list = this_cpu_ptr(&blk_cpu_done);
1177 0 : if (llist_add(&rq->ipi_list, list))
1178 0 : raise_softirq(BLOCK_SOFTIRQ);
1179 0 : preempt_enable();
1180 0 : }
1181 :
1182 0 : bool blk_mq_complete_request_remote(struct request *rq)
1183 : {
1184 0 : WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
1185 :
1186 : /*
1187 : * For request which hctx has only one ctx mapping,
1188 : * or a polled request, always complete locally,
1189 : * it's pointless to redirect the completion.
1190 : */
1191 0 : if ((rq->mq_hctx->nr_ctx == 1 &&
1192 0 : rq->mq_ctx->cpu == raw_smp_processor_id()) ||
1193 0 : rq->cmd_flags & REQ_POLLED)
1194 : return false;
1195 :
1196 0 : if (blk_mq_complete_need_ipi(rq)) {
1197 : blk_mq_complete_send_ipi(rq);
1198 : return true;
1199 : }
1200 :
1201 0 : if (rq->q->nr_hw_queues == 1) {
1202 0 : blk_mq_raise_softirq(rq);
1203 0 : return true;
1204 : }
1205 : return false;
1206 : }
1207 : EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote);
1208 :
1209 : /**
1210 : * blk_mq_complete_request - end I/O on a request
1211 : * @rq: the request being processed
1212 : *
1213 : * Description:
1214 : * Complete a request by scheduling the ->complete_rq operation.
1215 : **/
1216 0 : void blk_mq_complete_request(struct request *rq)
1217 : {
1218 0 : if (!blk_mq_complete_request_remote(rq))
1219 0 : rq->q->mq_ops->complete(rq);
1220 0 : }
1221 : EXPORT_SYMBOL(blk_mq_complete_request);
1222 :
1223 : /**
1224 : * blk_mq_start_request - Start processing a request
1225 : * @rq: Pointer to request to be started
1226 : *
1227 : * Function used by device drivers to notify the block layer that a request
1228 : * is going to be processed now, so blk layer can do proper initializations
1229 : * such as starting the timeout timer.
1230 : */
1231 0 : void blk_mq_start_request(struct request *rq)
1232 : {
1233 0 : struct request_queue *q = rq->q;
1234 :
1235 0 : trace_block_rq_issue(rq);
1236 :
1237 0 : if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
1238 0 : rq->io_start_time_ns = ktime_get_ns();
1239 0 : rq->stats_sectors = blk_rq_sectors(rq);
1240 0 : rq->rq_flags |= RQF_STATS;
1241 0 : rq_qos_issue(q, rq);
1242 : }
1243 :
1244 0 : WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
1245 :
1246 0 : blk_add_timer(rq);
1247 0 : WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
1248 :
1249 : #ifdef CONFIG_BLK_DEV_INTEGRITY
1250 : if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
1251 : q->integrity.profile->prepare_fn(rq);
1252 : #endif
1253 0 : if (rq->bio && rq->bio->bi_opf & REQ_POLLED)
1254 0 : WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num);
1255 0 : }
1256 : EXPORT_SYMBOL(blk_mq_start_request);
1257 :
1258 : /*
1259 : * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple
1260 : * queues. This is important for md arrays to benefit from merging
1261 : * requests.
1262 : */
1263 : static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug)
1264 : {
1265 0 : if (plug->multiple_queues)
1266 : return BLK_MAX_REQUEST_COUNT * 2;
1267 : return BLK_MAX_REQUEST_COUNT;
1268 : }
1269 :
1270 0 : static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1271 : {
1272 0 : struct request *last = rq_list_peek(&plug->mq_list);
1273 :
1274 0 : if (!plug->rq_count) {
1275 : trace_block_plug(rq->q);
1276 0 : } else if (plug->rq_count >= blk_plug_max_rq_count(plug) ||
1277 0 : (!blk_queue_nomerges(rq->q) &&
1278 0 : blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1279 0 : blk_mq_flush_plug_list(plug, false);
1280 0 : last = NULL;
1281 0 : trace_block_plug(rq->q);
1282 : }
1283 :
1284 0 : if (!plug->multiple_queues && last && last->q != rq->q)
1285 0 : plug->multiple_queues = true;
1286 : /*
1287 : * Any request allocated from sched tags can't be issued to
1288 : * ->queue_rqs() directly
1289 : */
1290 0 : if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS))
1291 0 : plug->has_elevator = true;
1292 0 : rq->rq_next = NULL;
1293 0 : rq_list_add(&plug->mq_list, rq);
1294 0 : plug->rq_count++;
1295 0 : }
1296 :
1297 : /**
1298 : * blk_execute_rq_nowait - insert a request to I/O scheduler for execution
1299 : * @rq: request to insert
1300 : * @at_head: insert request at head or tail of queue
1301 : *
1302 : * Description:
1303 : * Insert a fully prepared request at the back of the I/O scheduler queue
1304 : * for execution. Don't wait for completion.
1305 : *
1306 : * Note:
1307 : * This function will invoke @done directly if the queue is dead.
1308 : */
1309 0 : void blk_execute_rq_nowait(struct request *rq, bool at_head)
1310 : {
1311 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1312 :
1313 0 : WARN_ON(irqs_disabled());
1314 0 : WARN_ON(!blk_rq_is_passthrough(rq));
1315 :
1316 0 : blk_account_io_start(rq);
1317 :
1318 : /*
1319 : * As plugging can be enabled for passthrough requests on a zoned
1320 : * device, directly accessing the plug instead of using blk_mq_plug()
1321 : * should not have any consequences.
1322 : */
1323 0 : if (current->plug && !at_head) {
1324 0 : blk_add_rq_to_plug(current->plug, rq);
1325 0 : return;
1326 : }
1327 :
1328 0 : blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1329 0 : blk_mq_run_hw_queue(hctx, false);
1330 : }
1331 : EXPORT_SYMBOL_GPL(blk_execute_rq_nowait);
1332 :
1333 : struct blk_rq_wait {
1334 : struct completion done;
1335 : blk_status_t ret;
1336 : };
1337 :
1338 0 : static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret)
1339 : {
1340 0 : struct blk_rq_wait *wait = rq->end_io_data;
1341 :
1342 0 : wait->ret = ret;
1343 0 : complete(&wait->done);
1344 0 : return RQ_END_IO_NONE;
1345 : }
1346 :
1347 0 : bool blk_rq_is_poll(struct request *rq)
1348 : {
1349 0 : if (!rq->mq_hctx)
1350 : return false;
1351 0 : if (rq->mq_hctx->type != HCTX_TYPE_POLL)
1352 : return false;
1353 0 : return true;
1354 : }
1355 : EXPORT_SYMBOL_GPL(blk_rq_is_poll);
1356 :
1357 0 : static void blk_rq_poll_completion(struct request *rq, struct completion *wait)
1358 : {
1359 : do {
1360 0 : blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0);
1361 0 : cond_resched();
1362 0 : } while (!completion_done(wait));
1363 0 : }
1364 :
1365 : /**
1366 : * blk_execute_rq - insert a request into queue for execution
1367 : * @rq: request to insert
1368 : * @at_head: insert request at head or tail of queue
1369 : *
1370 : * Description:
1371 : * Insert a fully prepared request at the back of the I/O scheduler queue
1372 : * for execution and wait for completion.
1373 : * Return: The blk_status_t result provided to blk_mq_end_request().
1374 : */
1375 0 : blk_status_t blk_execute_rq(struct request *rq, bool at_head)
1376 : {
1377 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1378 0 : struct blk_rq_wait wait = {
1379 0 : .done = COMPLETION_INITIALIZER_ONSTACK(wait.done),
1380 : };
1381 :
1382 0 : WARN_ON(irqs_disabled());
1383 0 : WARN_ON(!blk_rq_is_passthrough(rq));
1384 :
1385 0 : rq->end_io_data = &wait;
1386 0 : rq->end_io = blk_end_sync_rq;
1387 :
1388 0 : blk_account_io_start(rq);
1389 0 : blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0);
1390 0 : blk_mq_run_hw_queue(hctx, false);
1391 :
1392 0 : if (blk_rq_is_poll(rq)) {
1393 0 : blk_rq_poll_completion(rq, &wait.done);
1394 : } else {
1395 : /*
1396 : * Prevent hang_check timer from firing at us during very long
1397 : * I/O
1398 : */
1399 0 : unsigned long hang_check = sysctl_hung_task_timeout_secs;
1400 :
1401 : if (hang_check)
1402 : while (!wait_for_completion_io_timeout(&wait.done,
1403 : hang_check * (HZ/2)))
1404 : ;
1405 : else
1406 0 : wait_for_completion_io(&wait.done);
1407 : }
1408 :
1409 0 : return wait.ret;
1410 : }
1411 : EXPORT_SYMBOL(blk_execute_rq);
1412 :
1413 0 : static void __blk_mq_requeue_request(struct request *rq)
1414 : {
1415 0 : struct request_queue *q = rq->q;
1416 :
1417 0 : blk_mq_put_driver_tag(rq);
1418 :
1419 0 : trace_block_rq_requeue(rq);
1420 0 : rq_qos_requeue(q, rq);
1421 :
1422 0 : if (blk_mq_request_started(rq)) {
1423 0 : WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1424 0 : rq->rq_flags &= ~RQF_TIMED_OUT;
1425 : }
1426 0 : }
1427 :
1428 0 : void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
1429 : {
1430 0 : struct request_queue *q = rq->q;
1431 : unsigned long flags;
1432 :
1433 0 : __blk_mq_requeue_request(rq);
1434 :
1435 : /* this request will be re-inserted to io scheduler queue */
1436 0 : blk_mq_sched_requeue_request(rq);
1437 :
1438 0 : spin_lock_irqsave(&q->requeue_lock, flags);
1439 0 : list_add_tail(&rq->queuelist, &q->requeue_list);
1440 0 : spin_unlock_irqrestore(&q->requeue_lock, flags);
1441 :
1442 0 : if (kick_requeue_list)
1443 : blk_mq_kick_requeue_list(q);
1444 0 : }
1445 : EXPORT_SYMBOL(blk_mq_requeue_request);
1446 :
1447 0 : static void blk_mq_requeue_work(struct work_struct *work)
1448 : {
1449 0 : struct request_queue *q =
1450 0 : container_of(work, struct request_queue, requeue_work.work);
1451 0 : LIST_HEAD(rq_list);
1452 0 : LIST_HEAD(flush_list);
1453 : struct request *rq;
1454 :
1455 0 : spin_lock_irq(&q->requeue_lock);
1456 0 : list_splice_init(&q->requeue_list, &rq_list);
1457 0 : list_splice_init(&q->flush_list, &flush_list);
1458 0 : spin_unlock_irq(&q->requeue_lock);
1459 :
1460 0 : while (!list_empty(&rq_list)) {
1461 0 : rq = list_entry(rq_list.next, struct request, queuelist);
1462 : /*
1463 : * If RQF_DONTPREP ist set, the request has been started by the
1464 : * driver already and might have driver-specific data allocated
1465 : * already. Insert it into the hctx dispatch list to avoid
1466 : * block layer merges for the request.
1467 : */
1468 0 : if (rq->rq_flags & RQF_DONTPREP) {
1469 0 : list_del_init(&rq->queuelist);
1470 : blk_mq_request_bypass_insert(rq, 0);
1471 : } else {
1472 0 : list_del_init(&rq->queuelist);
1473 0 : blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD);
1474 : }
1475 : }
1476 :
1477 0 : while (!list_empty(&flush_list)) {
1478 0 : rq = list_entry(flush_list.next, struct request, queuelist);
1479 0 : list_del_init(&rq->queuelist);
1480 0 : blk_mq_insert_request(rq, 0);
1481 : }
1482 :
1483 0 : blk_mq_run_hw_queues(q, false);
1484 0 : }
1485 :
1486 0 : void blk_mq_kick_requeue_list(struct request_queue *q)
1487 : {
1488 0 : kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
1489 0 : }
1490 : EXPORT_SYMBOL(blk_mq_kick_requeue_list);
1491 :
1492 0 : void blk_mq_delay_kick_requeue_list(struct request_queue *q,
1493 : unsigned long msecs)
1494 : {
1495 0 : kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
1496 : msecs_to_jiffies(msecs));
1497 0 : }
1498 : EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
1499 :
1500 0 : static bool blk_mq_rq_inflight(struct request *rq, void *priv)
1501 : {
1502 : /*
1503 : * If we find a request that isn't idle we know the queue is busy
1504 : * as it's checked in the iter.
1505 : * Return false to stop the iteration.
1506 : */
1507 0 : if (blk_mq_request_started(rq)) {
1508 0 : bool *busy = priv;
1509 :
1510 0 : *busy = true;
1511 0 : return false;
1512 : }
1513 :
1514 : return true;
1515 : }
1516 :
1517 0 : bool blk_mq_queue_inflight(struct request_queue *q)
1518 : {
1519 0 : bool busy = false;
1520 :
1521 0 : blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
1522 0 : return busy;
1523 : }
1524 : EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
1525 :
1526 0 : static void blk_mq_rq_timed_out(struct request *req)
1527 : {
1528 0 : req->rq_flags |= RQF_TIMED_OUT;
1529 0 : if (req->q->mq_ops->timeout) {
1530 : enum blk_eh_timer_return ret;
1531 :
1532 0 : ret = req->q->mq_ops->timeout(req);
1533 0 : if (ret == BLK_EH_DONE)
1534 : return;
1535 0 : WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
1536 : }
1537 :
1538 0 : blk_add_timer(req);
1539 : }
1540 :
1541 : struct blk_expired_data {
1542 : bool has_timedout_rq;
1543 : unsigned long next;
1544 : unsigned long timeout_start;
1545 : };
1546 :
1547 : static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired)
1548 : {
1549 : unsigned long deadline;
1550 :
1551 0 : if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
1552 : return false;
1553 0 : if (rq->rq_flags & RQF_TIMED_OUT)
1554 : return false;
1555 :
1556 0 : deadline = READ_ONCE(rq->deadline);
1557 0 : if (time_after_eq(expired->timeout_start, deadline))
1558 : return true;
1559 :
1560 0 : if (expired->next == 0)
1561 0 : expired->next = deadline;
1562 0 : else if (time_after(expired->next, deadline))
1563 0 : expired->next = deadline;
1564 : return false;
1565 : }
1566 :
1567 0 : void blk_mq_put_rq_ref(struct request *rq)
1568 : {
1569 0 : if (is_flush_rq(rq)) {
1570 0 : if (rq->end_io(rq, 0) == RQ_END_IO_FREE)
1571 0 : blk_mq_free_request(rq);
1572 0 : } else if (req_ref_put_and_test(rq)) {
1573 0 : __blk_mq_free_request(rq);
1574 : }
1575 0 : }
1576 :
1577 0 : static bool blk_mq_check_expired(struct request *rq, void *priv)
1578 : {
1579 0 : struct blk_expired_data *expired = priv;
1580 :
1581 : /*
1582 : * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot
1583 : * be reallocated underneath the timeout handler's processing, then
1584 : * the expire check is reliable. If the request is not expired, then
1585 : * it was completed and reallocated as a new request after returning
1586 : * from blk_mq_check_expired().
1587 : */
1588 0 : if (blk_mq_req_expired(rq, expired)) {
1589 0 : expired->has_timedout_rq = true;
1590 0 : return false;
1591 : }
1592 : return true;
1593 : }
1594 :
1595 0 : static bool blk_mq_handle_expired(struct request *rq, void *priv)
1596 : {
1597 0 : struct blk_expired_data *expired = priv;
1598 :
1599 0 : if (blk_mq_req_expired(rq, expired))
1600 0 : blk_mq_rq_timed_out(rq);
1601 0 : return true;
1602 : }
1603 :
1604 0 : static void blk_mq_timeout_work(struct work_struct *work)
1605 : {
1606 0 : struct request_queue *q =
1607 0 : container_of(work, struct request_queue, timeout_work);
1608 0 : struct blk_expired_data expired = {
1609 : .timeout_start = jiffies,
1610 : };
1611 : struct blk_mq_hw_ctx *hctx;
1612 : unsigned long i;
1613 :
1614 : /* A deadlock might occur if a request is stuck requiring a
1615 : * timeout at the same time a queue freeze is waiting
1616 : * completion, since the timeout code would not be able to
1617 : * acquire the queue reference here.
1618 : *
1619 : * That's why we don't use blk_queue_enter here; instead, we use
1620 : * percpu_ref_tryget directly, because we need to be able to
1621 : * obtain a reference even in the short window between the queue
1622 : * starting to freeze, by dropping the first reference in
1623 : * blk_freeze_queue_start, and the moment the last request is
1624 : * consumed, marked by the instant q_usage_counter reaches
1625 : * zero.
1626 : */
1627 0 : if (!percpu_ref_tryget(&q->q_usage_counter))
1628 0 : return;
1629 :
1630 : /* check if there is any timed-out request */
1631 0 : blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
1632 0 : if (expired.has_timedout_rq) {
1633 : /*
1634 : * Before walking tags, we must ensure any submit started
1635 : * before the current time has finished. Since the submit
1636 : * uses srcu or rcu, wait for a synchronization point to
1637 : * ensure all running submits have finished
1638 : */
1639 0 : blk_mq_wait_quiesce_done(q->tag_set);
1640 :
1641 0 : expired.next = 0;
1642 0 : blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired);
1643 : }
1644 :
1645 0 : if (expired.next != 0) {
1646 0 : mod_timer(&q->timeout, expired.next);
1647 : } else {
1648 : /*
1649 : * Request timeouts are handled as a forward rolling timer. If
1650 : * we end up here it means that no requests are pending and
1651 : * also that no request has been pending for a while. Mark
1652 : * each hctx as idle.
1653 : */
1654 0 : queue_for_each_hw_ctx(q, hctx, i) {
1655 : /* the hctx may be unmapped, so check it here */
1656 0 : if (blk_mq_hw_queue_mapped(hctx))
1657 : blk_mq_tag_idle(hctx);
1658 : }
1659 : }
1660 0 : blk_queue_exit(q);
1661 : }
1662 :
1663 : struct flush_busy_ctx_data {
1664 : struct blk_mq_hw_ctx *hctx;
1665 : struct list_head *list;
1666 : };
1667 :
1668 0 : static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
1669 : {
1670 0 : struct flush_busy_ctx_data *flush_data = data;
1671 0 : struct blk_mq_hw_ctx *hctx = flush_data->hctx;
1672 0 : struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1673 0 : enum hctx_type type = hctx->type;
1674 :
1675 0 : spin_lock(&ctx->lock);
1676 0 : list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
1677 0 : sbitmap_clear_bit(sb, bitnr);
1678 0 : spin_unlock(&ctx->lock);
1679 0 : return true;
1680 : }
1681 :
1682 : /*
1683 : * Process software queues that have been marked busy, splicing them
1684 : * to the for-dispatch
1685 : */
1686 0 : void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
1687 : {
1688 0 : struct flush_busy_ctx_data data = {
1689 : .hctx = hctx,
1690 : .list = list,
1691 : };
1692 :
1693 0 : sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1694 0 : }
1695 : EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1696 :
1697 : struct dispatch_rq_data {
1698 : struct blk_mq_hw_ctx *hctx;
1699 : struct request *rq;
1700 : };
1701 :
1702 0 : static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1703 : void *data)
1704 : {
1705 0 : struct dispatch_rq_data *dispatch_data = data;
1706 0 : struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1707 0 : struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1708 0 : enum hctx_type type = hctx->type;
1709 :
1710 0 : spin_lock(&ctx->lock);
1711 0 : if (!list_empty(&ctx->rq_lists[type])) {
1712 0 : dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1713 0 : list_del_init(&dispatch_data->rq->queuelist);
1714 0 : if (list_empty(&ctx->rq_lists[type]))
1715 : sbitmap_clear_bit(sb, bitnr);
1716 : }
1717 0 : spin_unlock(&ctx->lock);
1718 :
1719 0 : return !dispatch_data->rq;
1720 : }
1721 :
1722 0 : struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1723 : struct blk_mq_ctx *start)
1724 : {
1725 0 : unsigned off = start ? start->index_hw[hctx->type] : 0;
1726 0 : struct dispatch_rq_data data = {
1727 : .hctx = hctx,
1728 : .rq = NULL,
1729 : };
1730 :
1731 0 : __sbitmap_for_each_set(&hctx->ctx_map, off,
1732 : dispatch_rq_from_ctx, &data);
1733 :
1734 0 : return data.rq;
1735 : }
1736 :
1737 0 : static bool __blk_mq_alloc_driver_tag(struct request *rq)
1738 : {
1739 0 : struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags;
1740 0 : unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags;
1741 : int tag;
1742 :
1743 0 : blk_mq_tag_busy(rq->mq_hctx);
1744 :
1745 0 : if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) {
1746 0 : bt = &rq->mq_hctx->tags->breserved_tags;
1747 0 : tag_offset = 0;
1748 : } else {
1749 0 : if (!hctx_may_queue(rq->mq_hctx, bt))
1750 : return false;
1751 : }
1752 :
1753 0 : tag = __sbitmap_queue_get(bt);
1754 0 : if (tag == BLK_MQ_NO_TAG)
1755 : return false;
1756 :
1757 0 : rq->tag = tag + tag_offset;
1758 0 : return true;
1759 : }
1760 :
1761 0 : bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq)
1762 : {
1763 0 : if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
1764 : return false;
1765 :
1766 0 : if ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1767 0 : !(rq->rq_flags & RQF_MQ_INFLIGHT)) {
1768 0 : rq->rq_flags |= RQF_MQ_INFLIGHT;
1769 : __blk_mq_inc_active_requests(hctx);
1770 : }
1771 0 : hctx->tags->rqs[rq->tag] = rq;
1772 0 : return true;
1773 : }
1774 :
1775 0 : static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1776 : int flags, void *key)
1777 : {
1778 : struct blk_mq_hw_ctx *hctx;
1779 :
1780 0 : hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1781 :
1782 0 : spin_lock(&hctx->dispatch_wait_lock);
1783 0 : if (!list_empty(&wait->entry)) {
1784 : struct sbitmap_queue *sbq;
1785 :
1786 0 : list_del_init(&wait->entry);
1787 0 : sbq = &hctx->tags->bitmap_tags;
1788 0 : atomic_dec(&sbq->ws_active);
1789 : }
1790 0 : spin_unlock(&hctx->dispatch_wait_lock);
1791 :
1792 0 : blk_mq_run_hw_queue(hctx, true);
1793 0 : return 1;
1794 : }
1795 :
1796 : /*
1797 : * Mark us waiting for a tag. For shared tags, this involves hooking us into
1798 : * the tag wakeups. For non-shared tags, we can simply mark us needing a
1799 : * restart. For both cases, take care to check the condition again after
1800 : * marking us as waiting.
1801 : */
1802 0 : static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1803 : struct request *rq)
1804 : {
1805 : struct sbitmap_queue *sbq;
1806 : struct wait_queue_head *wq;
1807 : wait_queue_entry_t *wait;
1808 : bool ret;
1809 :
1810 0 : if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) &&
1811 0 : !(blk_mq_is_shared_tags(hctx->flags))) {
1812 0 : blk_mq_sched_mark_restart_hctx(hctx);
1813 :
1814 : /*
1815 : * It's possible that a tag was freed in the window between the
1816 : * allocation failure and adding the hardware queue to the wait
1817 : * queue.
1818 : *
1819 : * Don't clear RESTART here, someone else could have set it.
1820 : * At most this will cost an extra queue run.
1821 : */
1822 0 : return blk_mq_get_driver_tag(rq);
1823 : }
1824 :
1825 0 : wait = &hctx->dispatch_wait;
1826 0 : if (!list_empty_careful(&wait->entry))
1827 : return false;
1828 :
1829 0 : if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag))
1830 0 : sbq = &hctx->tags->breserved_tags;
1831 : else
1832 0 : sbq = &hctx->tags->bitmap_tags;
1833 0 : wq = &bt_wait_ptr(sbq, hctx)->wait;
1834 :
1835 0 : spin_lock_irq(&wq->lock);
1836 0 : spin_lock(&hctx->dispatch_wait_lock);
1837 0 : if (!list_empty(&wait->entry)) {
1838 0 : spin_unlock(&hctx->dispatch_wait_lock);
1839 0 : spin_unlock_irq(&wq->lock);
1840 0 : return false;
1841 : }
1842 :
1843 0 : atomic_inc(&sbq->ws_active);
1844 0 : wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1845 0 : __add_wait_queue(wq, wait);
1846 :
1847 : /*
1848 : * It's possible that a tag was freed in the window between the
1849 : * allocation failure and adding the hardware queue to the wait
1850 : * queue.
1851 : */
1852 0 : ret = blk_mq_get_driver_tag(rq);
1853 0 : if (!ret) {
1854 0 : spin_unlock(&hctx->dispatch_wait_lock);
1855 0 : spin_unlock_irq(&wq->lock);
1856 0 : return false;
1857 : }
1858 :
1859 : /*
1860 : * We got a tag, remove ourselves from the wait queue to ensure
1861 : * someone else gets the wakeup.
1862 : */
1863 0 : list_del_init(&wait->entry);
1864 0 : atomic_dec(&sbq->ws_active);
1865 0 : spin_unlock(&hctx->dispatch_wait_lock);
1866 0 : spin_unlock_irq(&wq->lock);
1867 :
1868 0 : return true;
1869 : }
1870 :
1871 : #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8
1872 : #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4
1873 : /*
1874 : * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1875 : * - EWMA is one simple way to compute running average value
1876 : * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1877 : * - take 4 as factor for avoiding to get too small(0) result, and this
1878 : * factor doesn't matter because EWMA decreases exponentially
1879 : */
1880 : static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1881 : {
1882 : unsigned int ewma;
1883 :
1884 0 : ewma = hctx->dispatch_busy;
1885 :
1886 0 : if (!ewma && !busy)
1887 : return;
1888 :
1889 0 : ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1890 : if (busy)
1891 0 : ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1892 0 : ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1893 :
1894 0 : hctx->dispatch_busy = ewma;
1895 : }
1896 :
1897 : #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */
1898 :
1899 : static void blk_mq_handle_dev_resource(struct request *rq,
1900 : struct list_head *list)
1901 : {
1902 0 : list_add(&rq->queuelist, list);
1903 0 : __blk_mq_requeue_request(rq);
1904 : }
1905 :
1906 : static void blk_mq_handle_zone_resource(struct request *rq,
1907 : struct list_head *zone_list)
1908 : {
1909 : /*
1910 : * If we end up here it is because we cannot dispatch a request to a
1911 : * specific zone due to LLD level zone-write locking or other zone
1912 : * related resource not being available. In this case, set the request
1913 : * aside in zone_list for retrying it later.
1914 : */
1915 0 : list_add(&rq->queuelist, zone_list);
1916 0 : __blk_mq_requeue_request(rq);
1917 : }
1918 :
1919 : enum prep_dispatch {
1920 : PREP_DISPATCH_OK,
1921 : PREP_DISPATCH_NO_TAG,
1922 : PREP_DISPATCH_NO_BUDGET,
1923 : };
1924 :
1925 0 : static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
1926 : bool need_budget)
1927 : {
1928 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1929 0 : int budget_token = -1;
1930 :
1931 0 : if (need_budget) {
1932 0 : budget_token = blk_mq_get_dispatch_budget(rq->q);
1933 0 : if (budget_token < 0) {
1934 : blk_mq_put_driver_tag(rq);
1935 : return PREP_DISPATCH_NO_BUDGET;
1936 : }
1937 : blk_mq_set_rq_budget_token(rq, budget_token);
1938 : }
1939 :
1940 0 : if (!blk_mq_get_driver_tag(rq)) {
1941 : /*
1942 : * The initial allocation attempt failed, so we need to
1943 : * rerun the hardware queue when a tag is freed. The
1944 : * waitqueue takes care of that. If the queue is run
1945 : * before we add this entry back on the dispatch list,
1946 : * we'll re-run it below.
1947 : */
1948 0 : if (!blk_mq_mark_tag_wait(hctx, rq)) {
1949 : /*
1950 : * All budgets not got from this function will be put
1951 : * together during handling partial dispatch
1952 : */
1953 0 : if (need_budget)
1954 0 : blk_mq_put_dispatch_budget(rq->q, budget_token);
1955 : return PREP_DISPATCH_NO_TAG;
1956 : }
1957 : }
1958 :
1959 : return PREP_DISPATCH_OK;
1960 : }
1961 :
1962 : /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
1963 0 : static void blk_mq_release_budgets(struct request_queue *q,
1964 : struct list_head *list)
1965 : {
1966 : struct request *rq;
1967 :
1968 0 : list_for_each_entry(rq, list, queuelist) {
1969 0 : int budget_token = blk_mq_get_rq_budget_token(rq);
1970 :
1971 0 : if (budget_token >= 0)
1972 : blk_mq_put_dispatch_budget(q, budget_token);
1973 : }
1974 0 : }
1975 :
1976 : /*
1977 : * blk_mq_commit_rqs will notify driver using bd->last that there is no
1978 : * more requests. (See comment in struct blk_mq_ops for commit_rqs for
1979 : * details)
1980 : * Attention, we should explicitly call this in unusual cases:
1981 : * 1) did not queue everything initially scheduled to queue
1982 : * 2) the last attempt to queue a request failed
1983 : */
1984 : static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued,
1985 : bool from_schedule)
1986 : {
1987 0 : if (hctx->queue->mq_ops->commit_rqs && queued) {
1988 0 : trace_block_unplug(hctx->queue, queued, !from_schedule);
1989 0 : hctx->queue->mq_ops->commit_rqs(hctx);
1990 : }
1991 : }
1992 :
1993 : /*
1994 : * Returns true if we did some work AND can potentially do more.
1995 : */
1996 0 : bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
1997 : unsigned int nr_budgets)
1998 : {
1999 : enum prep_dispatch prep;
2000 0 : struct request_queue *q = hctx->queue;
2001 : struct request *rq;
2002 : int queued;
2003 0 : blk_status_t ret = BLK_STS_OK;
2004 0 : LIST_HEAD(zone_list);
2005 0 : bool needs_resource = false;
2006 :
2007 0 : if (list_empty(list))
2008 : return false;
2009 :
2010 : /*
2011 : * Now process all the entries, sending them to the driver.
2012 : */
2013 : queued = 0;
2014 : do {
2015 : struct blk_mq_queue_data bd;
2016 :
2017 0 : rq = list_first_entry(list, struct request, queuelist);
2018 :
2019 0 : WARN_ON_ONCE(hctx != rq->mq_hctx);
2020 0 : prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets);
2021 0 : if (prep != PREP_DISPATCH_OK)
2022 : break;
2023 :
2024 0 : list_del_init(&rq->queuelist);
2025 :
2026 0 : bd.rq = rq;
2027 0 : bd.last = list_empty(list);
2028 :
2029 : /*
2030 : * once the request is queued to lld, no need to cover the
2031 : * budget any more
2032 : */
2033 0 : if (nr_budgets)
2034 0 : nr_budgets--;
2035 0 : ret = q->mq_ops->queue_rq(hctx, &bd);
2036 0 : switch (ret) {
2037 : case BLK_STS_OK:
2038 0 : queued++;
2039 0 : break;
2040 : case BLK_STS_RESOURCE:
2041 0 : needs_resource = true;
2042 : fallthrough;
2043 : case BLK_STS_DEV_RESOURCE:
2044 0 : blk_mq_handle_dev_resource(rq, list);
2045 0 : goto out;
2046 : case BLK_STS_ZONE_RESOURCE:
2047 : /*
2048 : * Move the request to zone_list and keep going through
2049 : * the dispatch list to find more requests the drive can
2050 : * accept.
2051 : */
2052 0 : blk_mq_handle_zone_resource(rq, &zone_list);
2053 0 : needs_resource = true;
2054 0 : break;
2055 : default:
2056 0 : blk_mq_end_request(rq, ret);
2057 : }
2058 0 : } while (!list_empty(list));
2059 : out:
2060 0 : if (!list_empty(&zone_list))
2061 : list_splice_tail_init(&zone_list, list);
2062 :
2063 : /* If we didn't flush the entire list, we could have told the driver
2064 : * there was more coming, but that turned out to be a lie.
2065 : */
2066 0 : if (!list_empty(list) || ret != BLK_STS_OK)
2067 : blk_mq_commit_rqs(hctx, queued, false);
2068 :
2069 : /*
2070 : * Any items that need requeuing? Stuff them into hctx->dispatch,
2071 : * that is where we will continue on next queue run.
2072 : */
2073 0 : if (!list_empty(list)) {
2074 : bool needs_restart;
2075 : /* For non-shared tags, the RESTART check will suffice */
2076 0 : bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
2077 0 : ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) ||
2078 0 : blk_mq_is_shared_tags(hctx->flags));
2079 :
2080 0 : if (nr_budgets)
2081 0 : blk_mq_release_budgets(q, list);
2082 :
2083 0 : spin_lock(&hctx->lock);
2084 0 : list_splice_tail_init(list, &hctx->dispatch);
2085 0 : spin_unlock(&hctx->lock);
2086 :
2087 : /*
2088 : * Order adding requests to hctx->dispatch and checking
2089 : * SCHED_RESTART flag. The pair of this smp_mb() is the one
2090 : * in blk_mq_sched_restart(). Avoid restart code path to
2091 : * miss the new added requests to hctx->dispatch, meantime
2092 : * SCHED_RESTART is observed here.
2093 : */
2094 0 : smp_mb();
2095 :
2096 : /*
2097 : * If SCHED_RESTART was set by the caller of this function and
2098 : * it is no longer set that means that it was cleared by another
2099 : * thread and hence that a queue rerun is needed.
2100 : *
2101 : * If 'no_tag' is set, that means that we failed getting
2102 : * a driver tag with an I/O scheduler attached. If our dispatch
2103 : * waitqueue is no longer active, ensure that we run the queue
2104 : * AFTER adding our entries back to the list.
2105 : *
2106 : * If no I/O scheduler has been configured it is possible that
2107 : * the hardware queue got stopped and restarted before requests
2108 : * were pushed back onto the dispatch list. Rerun the queue to
2109 : * avoid starvation. Notes:
2110 : * - blk_mq_run_hw_queue() checks whether or not a queue has
2111 : * been stopped before rerunning a queue.
2112 : * - Some but not all block drivers stop a queue before
2113 : * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
2114 : * and dm-rq.
2115 : *
2116 : * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
2117 : * bit is set, run queue after a delay to avoid IO stalls
2118 : * that could otherwise occur if the queue is idle. We'll do
2119 : * similar if we couldn't get budget or couldn't lock a zone
2120 : * and SCHED_RESTART is set.
2121 : */
2122 0 : needs_restart = blk_mq_sched_needs_restart(hctx);
2123 0 : if (prep == PREP_DISPATCH_NO_BUDGET)
2124 0 : needs_resource = true;
2125 0 : if (!needs_restart ||
2126 0 : (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
2127 0 : blk_mq_run_hw_queue(hctx, true);
2128 0 : else if (needs_resource)
2129 0 : blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
2130 :
2131 0 : blk_mq_update_dispatch_busy(hctx, true);
2132 0 : return false;
2133 : }
2134 :
2135 : blk_mq_update_dispatch_busy(hctx, false);
2136 : return true;
2137 : }
2138 :
2139 0 : static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
2140 : {
2141 0 : int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
2142 :
2143 0 : if (cpu >= nr_cpu_ids)
2144 0 : cpu = cpumask_first(hctx->cpumask);
2145 0 : return cpu;
2146 : }
2147 :
2148 : /*
2149 : * It'd be great if the workqueue API had a way to pass
2150 : * in a mask and had some smarts for more clever placement.
2151 : * For now we just round-robin here, switching for every
2152 : * BLK_MQ_CPU_WORK_BATCH queued items.
2153 : */
2154 0 : static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
2155 : {
2156 0 : bool tried = false;
2157 0 : int next_cpu = hctx->next_cpu;
2158 :
2159 0 : if (hctx->queue->nr_hw_queues == 1)
2160 : return WORK_CPU_UNBOUND;
2161 :
2162 0 : if (--hctx->next_cpu_batch <= 0) {
2163 : select_cpu:
2164 0 : next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
2165 : cpu_online_mask);
2166 0 : if (next_cpu >= nr_cpu_ids)
2167 0 : next_cpu = blk_mq_first_mapped_cpu(hctx);
2168 0 : hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2169 : }
2170 :
2171 : /*
2172 : * Do unbound schedule if we can't find a online CPU for this hctx,
2173 : * and it should only happen in the path of handling CPU DEAD.
2174 : */
2175 0 : if (!cpu_online(next_cpu)) {
2176 0 : if (!tried) {
2177 : tried = true;
2178 : goto select_cpu;
2179 : }
2180 :
2181 : /*
2182 : * Make sure to re-select CPU next time once after CPUs
2183 : * in hctx->cpumask become online again.
2184 : */
2185 0 : hctx->next_cpu = next_cpu;
2186 0 : hctx->next_cpu_batch = 1;
2187 0 : return WORK_CPU_UNBOUND;
2188 : }
2189 :
2190 0 : hctx->next_cpu = next_cpu;
2191 0 : return next_cpu;
2192 : }
2193 :
2194 : /**
2195 : * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously.
2196 : * @hctx: Pointer to the hardware queue to run.
2197 : * @msecs: Milliseconds of delay to wait before running the queue.
2198 : *
2199 : * Run a hardware queue asynchronously with a delay of @msecs.
2200 : */
2201 0 : void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
2202 : {
2203 0 : if (unlikely(blk_mq_hctx_stopped(hctx)))
2204 : return;
2205 0 : kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
2206 : msecs_to_jiffies(msecs));
2207 : }
2208 : EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
2209 :
2210 : /**
2211 : * blk_mq_run_hw_queue - Start to run a hardware queue.
2212 : * @hctx: Pointer to the hardware queue to run.
2213 : * @async: If we want to run the queue asynchronously.
2214 : *
2215 : * Check if the request queue is not in a quiesced state and if there are
2216 : * pending requests to be sent. If this is true, run the queue to send requests
2217 : * to hardware.
2218 : */
2219 0 : void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2220 : {
2221 : bool need_run;
2222 :
2223 : /*
2224 : * We can't run the queue inline with interrupts disabled.
2225 : */
2226 0 : WARN_ON_ONCE(!async && in_interrupt());
2227 :
2228 : /*
2229 : * When queue is quiesced, we may be switching io scheduler, or
2230 : * updating nr_hw_queues, or other things, and we can't run queue
2231 : * any more, even __blk_mq_hctx_has_pending() can't be called safely.
2232 : *
2233 : * And queue will be rerun in blk_mq_unquiesce_queue() if it is
2234 : * quiesced.
2235 : */
2236 0 : __blk_mq_run_dispatch_ops(hctx->queue, false,
2237 : need_run = !blk_queue_quiesced(hctx->queue) &&
2238 : blk_mq_hctx_has_pending(hctx));
2239 :
2240 0 : if (!need_run)
2241 : return;
2242 :
2243 0 : if (async || (hctx->flags & BLK_MQ_F_BLOCKING) ||
2244 0 : !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) {
2245 0 : blk_mq_delay_run_hw_queue(hctx, 0);
2246 0 : return;
2247 : }
2248 :
2249 0 : blk_mq_run_dispatch_ops(hctx->queue,
2250 : blk_mq_sched_dispatch_requests(hctx));
2251 : }
2252 : EXPORT_SYMBOL(blk_mq_run_hw_queue);
2253 :
2254 : /*
2255 : * Return prefered queue to dispatch from (if any) for non-mq aware IO
2256 : * scheduler.
2257 : */
2258 : static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q)
2259 : {
2260 0 : struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
2261 : /*
2262 : * If the IO scheduler does not respect hardware queues when
2263 : * dispatching, we just don't bother with multiple HW queues and
2264 : * dispatch from hctx for the current CPU since running multiple queues
2265 : * just causes lock contention inside the scheduler and pointless cache
2266 : * bouncing.
2267 : */
2268 0 : struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT];
2269 :
2270 0 : if (!blk_mq_hctx_stopped(hctx))
2271 : return hctx;
2272 : return NULL;
2273 : }
2274 :
2275 : /**
2276 : * blk_mq_run_hw_queues - Run all hardware queues in a request queue.
2277 : * @q: Pointer to the request queue to run.
2278 : * @async: If we want to run the queue asynchronously.
2279 : */
2280 0 : void blk_mq_run_hw_queues(struct request_queue *q, bool async)
2281 : {
2282 : struct blk_mq_hw_ctx *hctx, *sq_hctx;
2283 : unsigned long i;
2284 :
2285 0 : sq_hctx = NULL;
2286 0 : if (blk_queue_sq_sched(q))
2287 : sq_hctx = blk_mq_get_sq_hctx(q);
2288 0 : queue_for_each_hw_ctx(q, hctx, i) {
2289 0 : if (blk_mq_hctx_stopped(hctx))
2290 0 : continue;
2291 : /*
2292 : * Dispatch from this hctx either if there's no hctx preferred
2293 : * by IO scheduler or if it has requests that bypass the
2294 : * scheduler.
2295 : */
2296 0 : if (!sq_hctx || sq_hctx == hctx ||
2297 0 : !list_empty_careful(&hctx->dispatch))
2298 0 : blk_mq_run_hw_queue(hctx, async);
2299 : }
2300 0 : }
2301 : EXPORT_SYMBOL(blk_mq_run_hw_queues);
2302 :
2303 : /**
2304 : * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously.
2305 : * @q: Pointer to the request queue to run.
2306 : * @msecs: Milliseconds of delay to wait before running the queues.
2307 : */
2308 0 : void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)
2309 : {
2310 : struct blk_mq_hw_ctx *hctx, *sq_hctx;
2311 : unsigned long i;
2312 :
2313 0 : sq_hctx = NULL;
2314 0 : if (blk_queue_sq_sched(q))
2315 : sq_hctx = blk_mq_get_sq_hctx(q);
2316 0 : queue_for_each_hw_ctx(q, hctx, i) {
2317 0 : if (blk_mq_hctx_stopped(hctx))
2318 0 : continue;
2319 : /*
2320 : * If there is already a run_work pending, leave the
2321 : * pending delay untouched. Otherwise, a hctx can stall
2322 : * if another hctx is re-delaying the other's work
2323 : * before the work executes.
2324 : */
2325 0 : if (delayed_work_pending(&hctx->run_work))
2326 0 : continue;
2327 : /*
2328 : * Dispatch from this hctx either if there's no hctx preferred
2329 : * by IO scheduler or if it has requests that bypass the
2330 : * scheduler.
2331 : */
2332 0 : if (!sq_hctx || sq_hctx == hctx ||
2333 0 : !list_empty_careful(&hctx->dispatch))
2334 0 : blk_mq_delay_run_hw_queue(hctx, msecs);
2335 : }
2336 0 : }
2337 : EXPORT_SYMBOL(blk_mq_delay_run_hw_queues);
2338 :
2339 : /*
2340 : * This function is often used for pausing .queue_rq() by driver when
2341 : * there isn't enough resource or some conditions aren't satisfied, and
2342 : * BLK_STS_RESOURCE is usually returned.
2343 : *
2344 : * We do not guarantee that dispatch can be drained or blocked
2345 : * after blk_mq_stop_hw_queue() returns. Please use
2346 : * blk_mq_quiesce_queue() for that requirement.
2347 : */
2348 0 : void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
2349 : {
2350 0 : cancel_delayed_work(&hctx->run_work);
2351 :
2352 0 : set_bit(BLK_MQ_S_STOPPED, &hctx->state);
2353 0 : }
2354 : EXPORT_SYMBOL(blk_mq_stop_hw_queue);
2355 :
2356 : /*
2357 : * This function is often used for pausing .queue_rq() by driver when
2358 : * there isn't enough resource or some conditions aren't satisfied, and
2359 : * BLK_STS_RESOURCE is usually returned.
2360 : *
2361 : * We do not guarantee that dispatch can be drained or blocked
2362 : * after blk_mq_stop_hw_queues() returns. Please use
2363 : * blk_mq_quiesce_queue() for that requirement.
2364 : */
2365 0 : void blk_mq_stop_hw_queues(struct request_queue *q)
2366 : {
2367 : struct blk_mq_hw_ctx *hctx;
2368 : unsigned long i;
2369 :
2370 0 : queue_for_each_hw_ctx(q, hctx, i)
2371 0 : blk_mq_stop_hw_queue(hctx);
2372 0 : }
2373 : EXPORT_SYMBOL(blk_mq_stop_hw_queues);
2374 :
2375 0 : void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
2376 : {
2377 0 : clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2378 :
2379 0 : blk_mq_run_hw_queue(hctx, false);
2380 0 : }
2381 : EXPORT_SYMBOL(blk_mq_start_hw_queue);
2382 :
2383 0 : void blk_mq_start_hw_queues(struct request_queue *q)
2384 : {
2385 : struct blk_mq_hw_ctx *hctx;
2386 : unsigned long i;
2387 :
2388 0 : queue_for_each_hw_ctx(q, hctx, i)
2389 0 : blk_mq_start_hw_queue(hctx);
2390 0 : }
2391 : EXPORT_SYMBOL(blk_mq_start_hw_queues);
2392 :
2393 0 : void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
2394 : {
2395 0 : if (!blk_mq_hctx_stopped(hctx))
2396 : return;
2397 :
2398 0 : clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
2399 0 : blk_mq_run_hw_queue(hctx, async);
2400 : }
2401 : EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
2402 :
2403 0 : void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
2404 : {
2405 : struct blk_mq_hw_ctx *hctx;
2406 : unsigned long i;
2407 :
2408 0 : queue_for_each_hw_ctx(q, hctx, i)
2409 0 : blk_mq_start_stopped_hw_queue(hctx, async);
2410 0 : }
2411 : EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
2412 :
2413 0 : static void blk_mq_run_work_fn(struct work_struct *work)
2414 : {
2415 0 : struct blk_mq_hw_ctx *hctx =
2416 0 : container_of(work, struct blk_mq_hw_ctx, run_work.work);
2417 :
2418 0 : blk_mq_run_dispatch_ops(hctx->queue,
2419 : blk_mq_sched_dispatch_requests(hctx));
2420 0 : }
2421 :
2422 : /**
2423 : * blk_mq_request_bypass_insert - Insert a request at dispatch list.
2424 : * @rq: Pointer to request to be inserted.
2425 : * @flags: BLK_MQ_INSERT_*
2426 : *
2427 : * Should only be used carefully, when the caller knows we want to
2428 : * bypass a potential IO scheduler on the target device.
2429 : */
2430 : static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)
2431 : {
2432 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2433 :
2434 0 : spin_lock(&hctx->lock);
2435 0 : if (flags & BLK_MQ_INSERT_AT_HEAD)
2436 0 : list_add(&rq->queuelist, &hctx->dispatch);
2437 : else
2438 0 : list_add_tail(&rq->queuelist, &hctx->dispatch);
2439 0 : spin_unlock(&hctx->lock);
2440 : }
2441 :
2442 0 : static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx,
2443 : struct blk_mq_ctx *ctx, struct list_head *list,
2444 : bool run_queue_async)
2445 : {
2446 : struct request *rq;
2447 0 : enum hctx_type type = hctx->type;
2448 :
2449 : /*
2450 : * Try to issue requests directly if the hw queue isn't busy to save an
2451 : * extra enqueue & dequeue to the sw queue.
2452 : */
2453 0 : if (!hctx->dispatch_busy && !run_queue_async) {
2454 0 : blk_mq_run_dispatch_ops(hctx->queue,
2455 : blk_mq_try_issue_list_directly(hctx, list));
2456 0 : if (list_empty(list))
2457 : goto out;
2458 : }
2459 :
2460 : /*
2461 : * preemption doesn't flush plug list, so it's possible ctx->cpu is
2462 : * offline now
2463 : */
2464 0 : list_for_each_entry(rq, list, queuelist) {
2465 0 : BUG_ON(rq->mq_ctx != ctx);
2466 0 : trace_block_rq_insert(rq);
2467 : }
2468 :
2469 0 : spin_lock(&ctx->lock);
2470 0 : list_splice_tail_init(list, &ctx->rq_lists[type]);
2471 0 : blk_mq_hctx_mark_pending(hctx, ctx);
2472 0 : spin_unlock(&ctx->lock);
2473 : out:
2474 0 : blk_mq_run_hw_queue(hctx, run_queue_async);
2475 0 : }
2476 :
2477 0 : static void blk_mq_insert_request(struct request *rq, blk_insert_t flags)
2478 : {
2479 0 : struct request_queue *q = rq->q;
2480 0 : struct blk_mq_ctx *ctx = rq->mq_ctx;
2481 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2482 :
2483 0 : if (blk_rq_is_passthrough(rq)) {
2484 : /*
2485 : * Passthrough request have to be added to hctx->dispatch
2486 : * directly. The device may be in a situation where it can't
2487 : * handle FS request, and always returns BLK_STS_RESOURCE for
2488 : * them, which gets them added to hctx->dispatch.
2489 : *
2490 : * If a passthrough request is required to unblock the queues,
2491 : * and it is added to the scheduler queue, there is no chance to
2492 : * dispatch it given we prioritize requests in hctx->dispatch.
2493 : */
2494 : blk_mq_request_bypass_insert(rq, flags);
2495 0 : } else if (req_op(rq) == REQ_OP_FLUSH) {
2496 : /*
2497 : * Firstly normal IO request is inserted to scheduler queue or
2498 : * sw queue, meantime we add flush request to dispatch queue(
2499 : * hctx->dispatch) directly and there is at most one in-flight
2500 : * flush request for each hw queue, so it doesn't matter to add
2501 : * flush request to tail or front of the dispatch queue.
2502 : *
2503 : * Secondly in case of NCQ, flush request belongs to non-NCQ
2504 : * command, and queueing it will fail when there is any
2505 : * in-flight normal IO request(NCQ command). When adding flush
2506 : * rq to the front of hctx->dispatch, it is easier to introduce
2507 : * extra time to flush rq's latency because of S_SCHED_RESTART
2508 : * compared with adding to the tail of dispatch queue, then
2509 : * chance of flush merge is increased, and less flush requests
2510 : * will be issued to controller. It is observed that ~10% time
2511 : * is saved in blktests block/004 on disk attached to AHCI/NCQ
2512 : * drive when adding flush rq to the front of hctx->dispatch.
2513 : *
2514 : * Simply queue flush rq to the front of hctx->dispatch so that
2515 : * intensive flush workloads can benefit in case of NCQ HW.
2516 : */
2517 : blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD);
2518 0 : } else if (q->elevator) {
2519 0 : LIST_HEAD(list);
2520 :
2521 0 : WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG);
2522 :
2523 0 : list_add(&rq->queuelist, &list);
2524 0 : q->elevator->type->ops.insert_requests(hctx, &list, flags);
2525 : } else {
2526 0 : trace_block_rq_insert(rq);
2527 :
2528 0 : spin_lock(&ctx->lock);
2529 0 : if (flags & BLK_MQ_INSERT_AT_HEAD)
2530 0 : list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]);
2531 : else
2532 0 : list_add_tail(&rq->queuelist,
2533 0 : &ctx->rq_lists[hctx->type]);
2534 0 : blk_mq_hctx_mark_pending(hctx, ctx);
2535 0 : spin_unlock(&ctx->lock);
2536 : }
2537 0 : }
2538 :
2539 0 : static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
2540 : unsigned int nr_segs)
2541 : {
2542 : int err;
2543 :
2544 0 : if (bio->bi_opf & REQ_RAHEAD)
2545 0 : rq->cmd_flags |= REQ_FAILFAST_MASK;
2546 :
2547 0 : rq->__sector = bio->bi_iter.bi_sector;
2548 0 : blk_rq_bio_prep(rq, bio, nr_segs);
2549 :
2550 : /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */
2551 0 : err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO);
2552 0 : WARN_ON_ONCE(err);
2553 :
2554 0 : blk_account_io_start(rq);
2555 0 : }
2556 :
2557 0 : static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
2558 : struct request *rq, bool last)
2559 : {
2560 0 : struct request_queue *q = rq->q;
2561 0 : struct blk_mq_queue_data bd = {
2562 : .rq = rq,
2563 : .last = last,
2564 : };
2565 : blk_status_t ret;
2566 :
2567 : /*
2568 : * For OK queue, we are done. For error, caller may kill it.
2569 : * Any other error (busy), just add it to our list as we
2570 : * previously would have done.
2571 : */
2572 0 : ret = q->mq_ops->queue_rq(hctx, &bd);
2573 0 : switch (ret) {
2574 : case BLK_STS_OK:
2575 : blk_mq_update_dispatch_busy(hctx, false);
2576 : break;
2577 : case BLK_STS_RESOURCE:
2578 : case BLK_STS_DEV_RESOURCE:
2579 0 : blk_mq_update_dispatch_busy(hctx, true);
2580 0 : __blk_mq_requeue_request(rq);
2581 0 : break;
2582 : default:
2583 : blk_mq_update_dispatch_busy(hctx, false);
2584 : break;
2585 : }
2586 :
2587 0 : return ret;
2588 : }
2589 :
2590 0 : static bool blk_mq_get_budget_and_tag(struct request *rq)
2591 : {
2592 : int budget_token;
2593 :
2594 0 : budget_token = blk_mq_get_dispatch_budget(rq->q);
2595 0 : if (budget_token < 0)
2596 : return false;
2597 0 : blk_mq_set_rq_budget_token(rq, budget_token);
2598 0 : if (!blk_mq_get_driver_tag(rq)) {
2599 0 : blk_mq_put_dispatch_budget(rq->q, budget_token);
2600 : return false;
2601 : }
2602 : return true;
2603 : }
2604 :
2605 : /**
2606 : * blk_mq_try_issue_directly - Try to send a request directly to device driver.
2607 : * @hctx: Pointer of the associated hardware queue.
2608 : * @rq: Pointer to request to be sent.
2609 : *
2610 : * If the device has enough resources to accept a new request now, send the
2611 : * request directly to device driver. Else, insert at hctx->dispatch queue, so
2612 : * we can try send it another time in the future. Requests inserted at this
2613 : * queue have higher priority.
2614 : */
2615 0 : static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
2616 : struct request *rq)
2617 : {
2618 : blk_status_t ret;
2619 :
2620 0 : if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2621 0 : blk_mq_insert_request(rq, 0);
2622 0 : return;
2623 : }
2624 :
2625 0 : if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) {
2626 0 : blk_mq_insert_request(rq, 0);
2627 0 : blk_mq_run_hw_queue(hctx, false);
2628 0 : return;
2629 : }
2630 :
2631 0 : ret = __blk_mq_issue_directly(hctx, rq, true);
2632 0 : switch (ret) {
2633 : case BLK_STS_OK:
2634 : break;
2635 : case BLK_STS_RESOURCE:
2636 : case BLK_STS_DEV_RESOURCE:
2637 0 : blk_mq_request_bypass_insert(rq, 0);
2638 0 : blk_mq_run_hw_queue(hctx, false);
2639 0 : break;
2640 : default:
2641 0 : blk_mq_end_request(rq, ret);
2642 0 : break;
2643 : }
2644 : }
2645 :
2646 0 : static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
2647 : {
2648 0 : struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
2649 :
2650 0 : if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) {
2651 0 : blk_mq_insert_request(rq, 0);
2652 0 : return BLK_STS_OK;
2653 : }
2654 :
2655 0 : if (!blk_mq_get_budget_and_tag(rq))
2656 : return BLK_STS_RESOURCE;
2657 0 : return __blk_mq_issue_directly(hctx, rq, last);
2658 : }
2659 :
2660 0 : static void blk_mq_plug_issue_direct(struct blk_plug *plug)
2661 : {
2662 0 : struct blk_mq_hw_ctx *hctx = NULL;
2663 : struct request *rq;
2664 0 : int queued = 0;
2665 0 : blk_status_t ret = BLK_STS_OK;
2666 :
2667 0 : while ((rq = rq_list_pop(&plug->mq_list))) {
2668 0 : bool last = rq_list_empty(plug->mq_list);
2669 :
2670 0 : if (hctx != rq->mq_hctx) {
2671 0 : if (hctx) {
2672 : blk_mq_commit_rqs(hctx, queued, false);
2673 : queued = 0;
2674 : }
2675 0 : hctx = rq->mq_hctx;
2676 : }
2677 :
2678 0 : ret = blk_mq_request_issue_directly(rq, last);
2679 0 : switch (ret) {
2680 : case BLK_STS_OK:
2681 0 : queued++;
2682 0 : break;
2683 : case BLK_STS_RESOURCE:
2684 : case BLK_STS_DEV_RESOURCE:
2685 0 : blk_mq_request_bypass_insert(rq, 0);
2686 0 : blk_mq_run_hw_queue(hctx, false);
2687 0 : goto out;
2688 : default:
2689 0 : blk_mq_end_request(rq, ret);
2690 0 : break;
2691 : }
2692 : }
2693 :
2694 : out:
2695 0 : if (ret != BLK_STS_OK)
2696 : blk_mq_commit_rqs(hctx, queued, false);
2697 0 : }
2698 :
2699 0 : static void __blk_mq_flush_plug_list(struct request_queue *q,
2700 : struct blk_plug *plug)
2701 : {
2702 0 : if (blk_queue_quiesced(q))
2703 : return;
2704 0 : q->mq_ops->queue_rqs(&plug->mq_list);
2705 : }
2706 :
2707 0 : static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched)
2708 : {
2709 0 : struct blk_mq_hw_ctx *this_hctx = NULL;
2710 0 : struct blk_mq_ctx *this_ctx = NULL;
2711 0 : struct request *requeue_list = NULL;
2712 0 : struct request **requeue_lastp = &requeue_list;
2713 0 : unsigned int depth = 0;
2714 0 : bool is_passthrough = false;
2715 0 : LIST_HEAD(list);
2716 :
2717 : do {
2718 0 : struct request *rq = rq_list_pop(&plug->mq_list);
2719 :
2720 0 : if (!this_hctx) {
2721 0 : this_hctx = rq->mq_hctx;
2722 0 : this_ctx = rq->mq_ctx;
2723 0 : is_passthrough = blk_rq_is_passthrough(rq);
2724 0 : } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx ||
2725 0 : is_passthrough != blk_rq_is_passthrough(rq)) {
2726 0 : rq_list_add_tail(&requeue_lastp, rq);
2727 0 : continue;
2728 : }
2729 0 : list_add(&rq->queuelist, &list);
2730 0 : depth++;
2731 0 : } while (!rq_list_empty(plug->mq_list));
2732 :
2733 0 : plug->mq_list = requeue_list;
2734 0 : trace_block_unplug(this_hctx->queue, depth, !from_sched);
2735 :
2736 0 : percpu_ref_get(&this_hctx->queue->q_usage_counter);
2737 : /* passthrough requests should never be issued to the I/O scheduler */
2738 0 : if (is_passthrough) {
2739 0 : spin_lock(&this_hctx->lock);
2740 0 : list_splice_tail_init(&list, &this_hctx->dispatch);
2741 0 : spin_unlock(&this_hctx->lock);
2742 0 : blk_mq_run_hw_queue(this_hctx, from_sched);
2743 0 : } else if (this_hctx->queue->elevator) {
2744 0 : this_hctx->queue->elevator->type->ops.insert_requests(this_hctx,
2745 : &list, 0);
2746 0 : blk_mq_run_hw_queue(this_hctx, from_sched);
2747 : } else {
2748 0 : blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched);
2749 : }
2750 0 : percpu_ref_put(&this_hctx->queue->q_usage_counter);
2751 0 : }
2752 :
2753 0 : void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2754 : {
2755 : struct request *rq;
2756 :
2757 0 : if (rq_list_empty(plug->mq_list))
2758 : return;
2759 0 : plug->rq_count = 0;
2760 :
2761 0 : if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) {
2762 : struct request_queue *q;
2763 :
2764 0 : rq = rq_list_peek(&plug->mq_list);
2765 0 : q = rq->q;
2766 :
2767 : /*
2768 : * Peek first request and see if we have a ->queue_rqs() hook.
2769 : * If we do, we can dispatch the whole plug list in one go. We
2770 : * already know at this point that all requests belong to the
2771 : * same queue, caller must ensure that's the case.
2772 : *
2773 : * Since we pass off the full list to the driver at this point,
2774 : * we do not increment the active request count for the queue.
2775 : * Bypass shared tags for now because of that.
2776 : */
2777 0 : if (q->mq_ops->queue_rqs &&
2778 0 : !(rq->mq_hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
2779 0 : blk_mq_run_dispatch_ops(q,
2780 : __blk_mq_flush_plug_list(q, plug));
2781 0 : if (rq_list_empty(plug->mq_list))
2782 : return;
2783 : }
2784 :
2785 0 : blk_mq_run_dispatch_ops(q,
2786 : blk_mq_plug_issue_direct(plug));
2787 0 : if (rq_list_empty(plug->mq_list))
2788 : return;
2789 : }
2790 :
2791 : do {
2792 0 : blk_mq_dispatch_plug_list(plug, from_schedule);
2793 0 : } while (!rq_list_empty(plug->mq_list));
2794 : }
2795 :
2796 0 : static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
2797 : struct list_head *list)
2798 : {
2799 0 : int queued = 0;
2800 0 : blk_status_t ret = BLK_STS_OK;
2801 :
2802 0 : while (!list_empty(list)) {
2803 0 : struct request *rq = list_first_entry(list, struct request,
2804 : queuelist);
2805 :
2806 0 : list_del_init(&rq->queuelist);
2807 0 : ret = blk_mq_request_issue_directly(rq, list_empty(list));
2808 0 : switch (ret) {
2809 : case BLK_STS_OK:
2810 0 : queued++;
2811 0 : break;
2812 : case BLK_STS_RESOURCE:
2813 : case BLK_STS_DEV_RESOURCE:
2814 0 : blk_mq_request_bypass_insert(rq, 0);
2815 0 : if (list_empty(list))
2816 0 : blk_mq_run_hw_queue(hctx, false);
2817 : goto out;
2818 : default:
2819 0 : blk_mq_end_request(rq, ret);
2820 0 : break;
2821 : }
2822 : }
2823 :
2824 : out:
2825 0 : if (ret != BLK_STS_OK)
2826 : blk_mq_commit_rqs(hctx, queued, false);
2827 0 : }
2828 :
2829 0 : static bool blk_mq_attempt_bio_merge(struct request_queue *q,
2830 : struct bio *bio, unsigned int nr_segs)
2831 : {
2832 0 : if (!blk_queue_nomerges(q) && bio_mergeable(bio)) {
2833 0 : if (blk_attempt_plug_merge(q, bio, nr_segs))
2834 : return true;
2835 0 : if (blk_mq_sched_bio_merge(q, bio, nr_segs))
2836 : return true;
2837 : }
2838 : return false;
2839 : }
2840 :
2841 0 : static struct request *blk_mq_get_new_requests(struct request_queue *q,
2842 : struct blk_plug *plug,
2843 : struct bio *bio,
2844 : unsigned int nsegs)
2845 : {
2846 0 : struct blk_mq_alloc_data data = {
2847 : .q = q,
2848 : .nr_tags = 1,
2849 0 : .cmd_flags = bio->bi_opf,
2850 : };
2851 : struct request *rq;
2852 :
2853 0 : if (unlikely(bio_queue_enter(bio)))
2854 : return NULL;
2855 :
2856 0 : if (blk_mq_attempt_bio_merge(q, bio, nsegs))
2857 : goto queue_exit;
2858 :
2859 0 : rq_qos_throttle(q, bio);
2860 :
2861 0 : if (plug) {
2862 0 : data.nr_tags = plug->nr_ios;
2863 0 : plug->nr_ios = 1;
2864 0 : data.cached_rq = &plug->cached_rq;
2865 : }
2866 :
2867 0 : rq = __blk_mq_alloc_requests(&data);
2868 0 : if (rq)
2869 : return rq;
2870 0 : rq_qos_cleanup(q, bio);
2871 0 : if (bio->bi_opf & REQ_NOWAIT)
2872 : bio_wouldblock_error(bio);
2873 : queue_exit:
2874 0 : blk_queue_exit(q);
2875 0 : return NULL;
2876 : }
2877 :
2878 0 : static inline struct request *blk_mq_get_cached_request(struct request_queue *q,
2879 : struct blk_plug *plug, struct bio **bio, unsigned int nsegs)
2880 : {
2881 : struct request *rq;
2882 : enum hctx_type type, hctx_type;
2883 :
2884 0 : if (!plug)
2885 : return NULL;
2886 0 : rq = rq_list_peek(&plug->cached_rq);
2887 0 : if (!rq || rq->q != q)
2888 : return NULL;
2889 :
2890 0 : if (blk_mq_attempt_bio_merge(q, *bio, nsegs)) {
2891 0 : *bio = NULL;
2892 0 : return NULL;
2893 : }
2894 :
2895 0 : type = blk_mq_get_hctx_type((*bio)->bi_opf);
2896 0 : hctx_type = rq->mq_hctx->type;
2897 0 : if (type != hctx_type &&
2898 0 : !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT))
2899 : return NULL;
2900 0 : if (op_is_flush(rq->cmd_flags) != op_is_flush((*bio)->bi_opf))
2901 : return NULL;
2902 :
2903 : /*
2904 : * If any qos ->throttle() end up blocking, we will have flushed the
2905 : * plug and hence killed the cached_rq list as well. Pop this entry
2906 : * before we throttle.
2907 : */
2908 0 : plug->cached_rq = rq_list_next(rq);
2909 0 : rq_qos_throttle(q, *bio);
2910 :
2911 0 : blk_mq_rq_time_init(rq, 0);
2912 0 : rq->cmd_flags = (*bio)->bi_opf;
2913 0 : INIT_LIST_HEAD(&rq->queuelist);
2914 0 : return rq;
2915 : }
2916 :
2917 : static void bio_set_ioprio(struct bio *bio)
2918 : {
2919 : /* Nobody set ioprio so far? Initialize it based on task's nice value */
2920 0 : if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
2921 0 : bio->bi_ioprio = get_current_ioprio();
2922 0 : blkcg_set_ioprio(bio);
2923 : }
2924 :
2925 : /**
2926 : * blk_mq_submit_bio - Create and send a request to block device.
2927 : * @bio: Bio pointer.
2928 : *
2929 : * Builds up a request structure from @q and @bio and send to the device. The
2930 : * request may not be queued directly to hardware if:
2931 : * * This request can be merged with another one
2932 : * * We want to place request at plug queue for possible future merging
2933 : * * There is an IO scheduler active at this queue
2934 : *
2935 : * It will not queue the request if there is an error with the bio, or at the
2936 : * request creation.
2937 : */
2938 0 : void blk_mq_submit_bio(struct bio *bio)
2939 : {
2940 0 : struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2941 0 : struct blk_plug *plug = blk_mq_plug(bio);
2942 0 : const int is_sync = op_is_sync(bio->bi_opf);
2943 : struct blk_mq_hw_ctx *hctx;
2944 : struct request *rq;
2945 0 : unsigned int nr_segs = 1;
2946 : blk_status_t ret;
2947 :
2948 0 : bio = blk_queue_bounce(bio, q);
2949 0 : if (bio_may_exceed_limits(bio, &q->limits)) {
2950 0 : bio = __bio_split_to_limits(bio, &q->limits, &nr_segs);
2951 0 : if (!bio)
2952 0 : return;
2953 : }
2954 :
2955 0 : if (!bio_integrity_prep(bio))
2956 : return;
2957 :
2958 0 : bio_set_ioprio(bio);
2959 :
2960 0 : rq = blk_mq_get_cached_request(q, plug, &bio, nr_segs);
2961 0 : if (!rq) {
2962 0 : if (!bio)
2963 : return;
2964 0 : rq = blk_mq_get_new_requests(q, plug, bio, nr_segs);
2965 0 : if (unlikely(!rq))
2966 : return;
2967 : }
2968 :
2969 0 : trace_block_getrq(bio);
2970 :
2971 0 : rq_qos_track(q, rq, bio);
2972 :
2973 0 : blk_mq_bio_to_request(rq, bio, nr_segs);
2974 :
2975 0 : ret = blk_crypto_rq_get_keyslot(rq);
2976 : if (ret != BLK_STS_OK) {
2977 : bio->bi_status = ret;
2978 : bio_endio(bio);
2979 : blk_mq_free_request(rq);
2980 : return;
2981 : }
2982 :
2983 0 : if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq))
2984 : return;
2985 :
2986 0 : if (plug) {
2987 0 : blk_add_rq_to_plug(plug, rq);
2988 0 : return;
2989 : }
2990 :
2991 0 : hctx = rq->mq_hctx;
2992 0 : if ((rq->rq_flags & RQF_USE_SCHED) ||
2993 0 : (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) {
2994 0 : blk_mq_insert_request(rq, 0);
2995 0 : blk_mq_run_hw_queue(hctx, true);
2996 : } else {
2997 0 : blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq));
2998 : }
2999 : }
3000 :
3001 : #ifdef CONFIG_BLK_MQ_STACKING
3002 : /**
3003 : * blk_insert_cloned_request - Helper for stacking drivers to submit a request
3004 : * @rq: the request being queued
3005 : */
3006 : blk_status_t blk_insert_cloned_request(struct request *rq)
3007 : {
3008 : struct request_queue *q = rq->q;
3009 : unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq));
3010 : unsigned int max_segments = blk_rq_get_max_segments(rq);
3011 : blk_status_t ret;
3012 :
3013 : if (blk_rq_sectors(rq) > max_sectors) {
3014 : /*
3015 : * SCSI device does not have a good way to return if
3016 : * Write Same/Zero is actually supported. If a device rejects
3017 : * a non-read/write command (discard, write same,etc.) the
3018 : * low-level device driver will set the relevant queue limit to
3019 : * 0 to prevent blk-lib from issuing more of the offending
3020 : * operations. Commands queued prior to the queue limit being
3021 : * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O
3022 : * errors being propagated to upper layers.
3023 : */
3024 : if (max_sectors == 0)
3025 : return BLK_STS_NOTSUPP;
3026 :
3027 : printk(KERN_ERR "%s: over max size limit. (%u > %u)\n",
3028 : __func__, blk_rq_sectors(rq), max_sectors);
3029 : return BLK_STS_IOERR;
3030 : }
3031 :
3032 : /*
3033 : * The queue settings related to segment counting may differ from the
3034 : * original queue.
3035 : */
3036 : rq->nr_phys_segments = blk_recalc_rq_segments(rq);
3037 : if (rq->nr_phys_segments > max_segments) {
3038 : printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n",
3039 : __func__, rq->nr_phys_segments, max_segments);
3040 : return BLK_STS_IOERR;
3041 : }
3042 :
3043 : if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq)))
3044 : return BLK_STS_IOERR;
3045 :
3046 : ret = blk_crypto_rq_get_keyslot(rq);
3047 : if (ret != BLK_STS_OK)
3048 : return ret;
3049 :
3050 : blk_account_io_start(rq);
3051 :
3052 : /*
3053 : * Since we have a scheduler attached on the top device,
3054 : * bypass a potential scheduler on the bottom device for
3055 : * insert.
3056 : */
3057 : blk_mq_run_dispatch_ops(q,
3058 : ret = blk_mq_request_issue_directly(rq, true));
3059 : if (ret)
3060 : blk_account_io_done(rq, ktime_get_ns());
3061 : return ret;
3062 : }
3063 : EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
3064 :
3065 : /**
3066 : * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
3067 : * @rq: the clone request to be cleaned up
3068 : *
3069 : * Description:
3070 : * Free all bios in @rq for a cloned request.
3071 : */
3072 : void blk_rq_unprep_clone(struct request *rq)
3073 : {
3074 : struct bio *bio;
3075 :
3076 : while ((bio = rq->bio) != NULL) {
3077 : rq->bio = bio->bi_next;
3078 :
3079 : bio_put(bio);
3080 : }
3081 : }
3082 : EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
3083 :
3084 : /**
3085 : * blk_rq_prep_clone - Helper function to setup clone request
3086 : * @rq: the request to be setup
3087 : * @rq_src: original request to be cloned
3088 : * @bs: bio_set that bios for clone are allocated from
3089 : * @gfp_mask: memory allocation mask for bio
3090 : * @bio_ctr: setup function to be called for each clone bio.
3091 : * Returns %0 for success, non %0 for failure.
3092 : * @data: private data to be passed to @bio_ctr
3093 : *
3094 : * Description:
3095 : * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
3096 : * Also, pages which the original bios are pointing to are not copied
3097 : * and the cloned bios just point same pages.
3098 : * So cloned bios must be completed before original bios, which means
3099 : * the caller must complete @rq before @rq_src.
3100 : */
3101 : int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
3102 : struct bio_set *bs, gfp_t gfp_mask,
3103 : int (*bio_ctr)(struct bio *, struct bio *, void *),
3104 : void *data)
3105 : {
3106 : struct bio *bio, *bio_src;
3107 :
3108 : if (!bs)
3109 : bs = &fs_bio_set;
3110 :
3111 : __rq_for_each_bio(bio_src, rq_src) {
3112 : bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask,
3113 : bs);
3114 : if (!bio)
3115 : goto free_and_out;
3116 :
3117 : if (bio_ctr && bio_ctr(bio, bio_src, data))
3118 : goto free_and_out;
3119 :
3120 : if (rq->bio) {
3121 : rq->biotail->bi_next = bio;
3122 : rq->biotail = bio;
3123 : } else {
3124 : rq->bio = rq->biotail = bio;
3125 : }
3126 : bio = NULL;
3127 : }
3128 :
3129 : /* Copy attributes of the original request to the clone request. */
3130 : rq->__sector = blk_rq_pos(rq_src);
3131 : rq->__data_len = blk_rq_bytes(rq_src);
3132 : if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) {
3133 : rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
3134 : rq->special_vec = rq_src->special_vec;
3135 : }
3136 : rq->nr_phys_segments = rq_src->nr_phys_segments;
3137 : rq->ioprio = rq_src->ioprio;
3138 :
3139 : if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0)
3140 : goto free_and_out;
3141 :
3142 : return 0;
3143 :
3144 : free_and_out:
3145 : if (bio)
3146 : bio_put(bio);
3147 : blk_rq_unprep_clone(rq);
3148 :
3149 : return -ENOMEM;
3150 : }
3151 : EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
3152 : #endif /* CONFIG_BLK_MQ_STACKING */
3153 :
3154 : /*
3155 : * Steal bios from a request and add them to a bio list.
3156 : * The request must not have been partially completed before.
3157 : */
3158 0 : void blk_steal_bios(struct bio_list *list, struct request *rq)
3159 : {
3160 0 : if (rq->bio) {
3161 0 : if (list->tail)
3162 0 : list->tail->bi_next = rq->bio;
3163 : else
3164 0 : list->head = rq->bio;
3165 0 : list->tail = rq->biotail;
3166 :
3167 0 : rq->bio = NULL;
3168 0 : rq->biotail = NULL;
3169 : }
3170 :
3171 0 : rq->__data_len = 0;
3172 0 : }
3173 : EXPORT_SYMBOL_GPL(blk_steal_bios);
3174 :
3175 : static size_t order_to_size(unsigned int order)
3176 : {
3177 0 : return (size_t)PAGE_SIZE << order;
3178 : }
3179 :
3180 : /* called before freeing request pool in @tags */
3181 0 : static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
3182 : struct blk_mq_tags *tags)
3183 : {
3184 : struct page *page;
3185 : unsigned long flags;
3186 :
3187 : /*
3188 : * There is no need to clear mapping if driver tags is not initialized
3189 : * or the mapping belongs to the driver tags.
3190 : */
3191 0 : if (!drv_tags || drv_tags == tags)
3192 : return;
3193 :
3194 0 : list_for_each_entry(page, &tags->page_list, lru) {
3195 0 : unsigned long start = (unsigned long)page_address(page);
3196 0 : unsigned long end = start + order_to_size(page->private);
3197 : int i;
3198 :
3199 0 : for (i = 0; i < drv_tags->nr_tags; i++) {
3200 0 : struct request *rq = drv_tags->rqs[i];
3201 0 : unsigned long rq_addr = (unsigned long)rq;
3202 :
3203 0 : if (rq_addr >= start && rq_addr < end) {
3204 0 : WARN_ON_ONCE(req_ref_read(rq) != 0);
3205 0 : cmpxchg(&drv_tags->rqs[i], rq, NULL);
3206 : }
3207 : }
3208 : }
3209 :
3210 : /*
3211 : * Wait until all pending iteration is done.
3212 : *
3213 : * Request reference is cleared and it is guaranteed to be observed
3214 : * after the ->lock is released.
3215 : */
3216 0 : spin_lock_irqsave(&drv_tags->lock, flags);
3217 0 : spin_unlock_irqrestore(&drv_tags->lock, flags);
3218 : }
3219 :
3220 0 : void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
3221 : unsigned int hctx_idx)
3222 : {
3223 : struct blk_mq_tags *drv_tags;
3224 : struct page *page;
3225 :
3226 0 : if (list_empty(&tags->page_list))
3227 : return;
3228 :
3229 0 : if (blk_mq_is_shared_tags(set->flags))
3230 0 : drv_tags = set->shared_tags;
3231 : else
3232 0 : drv_tags = set->tags[hctx_idx];
3233 :
3234 0 : if (tags->static_rqs && set->ops->exit_request) {
3235 : int i;
3236 :
3237 0 : for (i = 0; i < tags->nr_tags; i++) {
3238 0 : struct request *rq = tags->static_rqs[i];
3239 :
3240 0 : if (!rq)
3241 0 : continue;
3242 0 : set->ops->exit_request(set, rq, hctx_idx);
3243 0 : tags->static_rqs[i] = NULL;
3244 : }
3245 : }
3246 :
3247 0 : blk_mq_clear_rq_mapping(drv_tags, tags);
3248 :
3249 0 : while (!list_empty(&tags->page_list)) {
3250 0 : page = list_first_entry(&tags->page_list, struct page, lru);
3251 0 : list_del_init(&page->lru);
3252 : /*
3253 : * Remove kmemleak object previously allocated in
3254 : * blk_mq_alloc_rqs().
3255 : */
3256 0 : kmemleak_free(page_address(page));
3257 0 : __free_pages(page, page->private);
3258 : }
3259 : }
3260 :
3261 0 : void blk_mq_free_rq_map(struct blk_mq_tags *tags)
3262 : {
3263 0 : kfree(tags->rqs);
3264 0 : tags->rqs = NULL;
3265 0 : kfree(tags->static_rqs);
3266 0 : tags->static_rqs = NULL;
3267 :
3268 0 : blk_mq_free_tags(tags);
3269 0 : }
3270 :
3271 : static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set,
3272 : unsigned int hctx_idx)
3273 : {
3274 : int i;
3275 :
3276 0 : for (i = 0; i < set->nr_maps; i++) {
3277 0 : unsigned int start = set->map[i].queue_offset;
3278 0 : unsigned int end = start + set->map[i].nr_queues;
3279 :
3280 0 : if (hctx_idx >= start && hctx_idx < end)
3281 : break;
3282 : }
3283 :
3284 0 : if (i >= set->nr_maps)
3285 0 : i = HCTX_TYPE_DEFAULT;
3286 :
3287 0 : return i;
3288 : }
3289 :
3290 0 : static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set,
3291 : unsigned int hctx_idx)
3292 : {
3293 0 : enum hctx_type type = hctx_idx_to_type(set, hctx_idx);
3294 :
3295 0 : return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx);
3296 : }
3297 :
3298 0 : static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
3299 : unsigned int hctx_idx,
3300 : unsigned int nr_tags,
3301 : unsigned int reserved_tags)
3302 : {
3303 0 : int node = blk_mq_get_hctx_node(set, hctx_idx);
3304 : struct blk_mq_tags *tags;
3305 :
3306 0 : if (node == NUMA_NO_NODE)
3307 0 : node = set->numa_node;
3308 :
3309 0 : tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
3310 0 : BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
3311 0 : if (!tags)
3312 : return NULL;
3313 :
3314 0 : tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3315 : GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3316 : node);
3317 0 : if (!tags->rqs)
3318 : goto err_free_tags;
3319 :
3320 0 : tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
3321 : GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
3322 : node);
3323 0 : if (!tags->static_rqs)
3324 : goto err_free_rqs;
3325 :
3326 : return tags;
3327 :
3328 : err_free_rqs:
3329 0 : kfree(tags->rqs);
3330 : err_free_tags:
3331 0 : blk_mq_free_tags(tags);
3332 0 : return NULL;
3333 : }
3334 :
3335 : static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
3336 : unsigned int hctx_idx, int node)
3337 : {
3338 : int ret;
3339 :
3340 0 : if (set->ops->init_request) {
3341 0 : ret = set->ops->init_request(set, rq, hctx_idx, node);
3342 0 : if (ret)
3343 : return ret;
3344 : }
3345 :
3346 0 : WRITE_ONCE(rq->state, MQ_RQ_IDLE);
3347 : return 0;
3348 : }
3349 :
3350 0 : static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set,
3351 : struct blk_mq_tags *tags,
3352 : unsigned int hctx_idx, unsigned int depth)
3353 : {
3354 0 : unsigned int i, j, entries_per_page, max_order = 4;
3355 0 : int node = blk_mq_get_hctx_node(set, hctx_idx);
3356 : size_t rq_size, left;
3357 :
3358 0 : if (node == NUMA_NO_NODE)
3359 0 : node = set->numa_node;
3360 :
3361 0 : INIT_LIST_HEAD(&tags->page_list);
3362 :
3363 : /*
3364 : * rq_size is the size of the request plus driver payload, rounded
3365 : * to the cacheline size
3366 : */
3367 0 : rq_size = round_up(sizeof(struct request) + set->cmd_size,
3368 : cache_line_size());
3369 0 : left = rq_size * depth;
3370 :
3371 0 : for (i = 0; i < depth; ) {
3372 : int this_order = max_order;
3373 : struct page *page;
3374 : int to_do;
3375 : void *p;
3376 :
3377 0 : while (this_order && left < order_to_size(this_order - 1))
3378 : this_order--;
3379 :
3380 : do {
3381 0 : page = alloc_pages_node(node,
3382 : GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
3383 : this_order);
3384 0 : if (page)
3385 : break;
3386 0 : if (!this_order--)
3387 : break;
3388 0 : if (order_to_size(this_order) < rq_size)
3389 : break;
3390 : } while (1);
3391 :
3392 0 : if (!page)
3393 : goto fail;
3394 :
3395 0 : page->private = this_order;
3396 0 : list_add_tail(&page->lru, &tags->page_list);
3397 :
3398 0 : p = page_address(page);
3399 : /*
3400 : * Allow kmemleak to scan these pages as they contain pointers
3401 : * to additional allocations like via ops->init_request().
3402 : */
3403 0 : kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
3404 0 : entries_per_page = order_to_size(this_order) / rq_size;
3405 0 : to_do = min(entries_per_page, depth - i);
3406 0 : left -= to_do * rq_size;
3407 0 : for (j = 0; j < to_do; j++) {
3408 0 : struct request *rq = p;
3409 :
3410 0 : tags->static_rqs[i] = rq;
3411 0 : if (blk_mq_init_request(set, rq, hctx_idx, node)) {
3412 0 : tags->static_rqs[i] = NULL;
3413 0 : goto fail;
3414 : }
3415 :
3416 0 : p += rq_size;
3417 0 : i++;
3418 : }
3419 : }
3420 : return 0;
3421 :
3422 : fail:
3423 0 : blk_mq_free_rqs(set, tags, hctx_idx);
3424 0 : return -ENOMEM;
3425 : }
3426 :
3427 : struct rq_iter_data {
3428 : struct blk_mq_hw_ctx *hctx;
3429 : bool has_rq;
3430 : };
3431 :
3432 0 : static bool blk_mq_has_request(struct request *rq, void *data)
3433 : {
3434 0 : struct rq_iter_data *iter_data = data;
3435 :
3436 0 : if (rq->mq_hctx != iter_data->hctx)
3437 : return true;
3438 0 : iter_data->has_rq = true;
3439 0 : return false;
3440 : }
3441 :
3442 0 : static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx)
3443 : {
3444 0 : struct blk_mq_tags *tags = hctx->sched_tags ?
3445 0 : hctx->sched_tags : hctx->tags;
3446 0 : struct rq_iter_data data = {
3447 : .hctx = hctx,
3448 : };
3449 :
3450 0 : blk_mq_all_tag_iter(tags, blk_mq_has_request, &data);
3451 0 : return data.has_rq;
3452 : }
3453 :
3454 0 : static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu,
3455 : struct blk_mq_hw_ctx *hctx)
3456 : {
3457 0 : if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu)
3458 : return false;
3459 0 : if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids)
3460 : return false;
3461 0 : return true;
3462 : }
3463 :
3464 0 : static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node)
3465 : {
3466 0 : struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3467 : struct blk_mq_hw_ctx, cpuhp_online);
3468 :
3469 0 : if (!cpumask_test_cpu(cpu, hctx->cpumask) ||
3470 0 : !blk_mq_last_cpu_in_hctx(cpu, hctx))
3471 : return 0;
3472 :
3473 : /*
3474 : * Prevent new request from being allocated on the current hctx.
3475 : *
3476 : * The smp_mb__after_atomic() Pairs with the implied barrier in
3477 : * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is
3478 : * seen once we return from the tag allocator.
3479 : */
3480 0 : set_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3481 0 : smp_mb__after_atomic();
3482 :
3483 : /*
3484 : * Try to grab a reference to the queue and wait for any outstanding
3485 : * requests. If we could not grab a reference the queue has been
3486 : * frozen and there are no requests.
3487 : */
3488 0 : if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) {
3489 0 : while (blk_mq_hctx_has_requests(hctx))
3490 0 : msleep(5);
3491 0 : percpu_ref_put(&hctx->queue->q_usage_counter);
3492 : }
3493 :
3494 : return 0;
3495 : }
3496 :
3497 0 : static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node)
3498 : {
3499 0 : struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node,
3500 : struct blk_mq_hw_ctx, cpuhp_online);
3501 :
3502 0 : if (cpumask_test_cpu(cpu, hctx->cpumask))
3503 0 : clear_bit(BLK_MQ_S_INACTIVE, &hctx->state);
3504 0 : return 0;
3505 : }
3506 :
3507 : /*
3508 : * 'cpu' is going away. splice any existing rq_list entries from this
3509 : * software queue to the hw queue dispatch list, and ensure that it
3510 : * gets run.
3511 : */
3512 0 : static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
3513 : {
3514 : struct blk_mq_hw_ctx *hctx;
3515 : struct blk_mq_ctx *ctx;
3516 0 : LIST_HEAD(tmp);
3517 : enum hctx_type type;
3518 :
3519 0 : hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
3520 0 : if (!cpumask_test_cpu(cpu, hctx->cpumask))
3521 : return 0;
3522 :
3523 0 : ctx = __blk_mq_get_ctx(hctx->queue, cpu);
3524 0 : type = hctx->type;
3525 :
3526 0 : spin_lock(&ctx->lock);
3527 0 : if (!list_empty(&ctx->rq_lists[type])) {
3528 0 : list_splice_init(&ctx->rq_lists[type], &tmp);
3529 : blk_mq_hctx_clear_pending(hctx, ctx);
3530 : }
3531 0 : spin_unlock(&ctx->lock);
3532 :
3533 0 : if (list_empty(&tmp))
3534 : return 0;
3535 :
3536 0 : spin_lock(&hctx->lock);
3537 0 : list_splice_tail_init(&tmp, &hctx->dispatch);
3538 0 : spin_unlock(&hctx->lock);
3539 :
3540 0 : blk_mq_run_hw_queue(hctx, true);
3541 0 : return 0;
3542 : }
3543 :
3544 0 : static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
3545 : {
3546 0 : if (!(hctx->flags & BLK_MQ_F_STACKING))
3547 0 : cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3548 : &hctx->cpuhp_online);
3549 0 : cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
3550 : &hctx->cpuhp_dead);
3551 0 : }
3552 :
3553 : /*
3554 : * Before freeing hw queue, clearing the flush request reference in
3555 : * tags->rqs[] for avoiding potential UAF.
3556 : */
3557 0 : static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
3558 : unsigned int queue_depth, struct request *flush_rq)
3559 : {
3560 : int i;
3561 : unsigned long flags;
3562 :
3563 : /* The hw queue may not be mapped yet */
3564 0 : if (!tags)
3565 : return;
3566 :
3567 0 : WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
3568 :
3569 0 : for (i = 0; i < queue_depth; i++)
3570 0 : cmpxchg(&tags->rqs[i], flush_rq, NULL);
3571 :
3572 : /*
3573 : * Wait until all pending iteration is done.
3574 : *
3575 : * Request reference is cleared and it is guaranteed to be observed
3576 : * after the ->lock is released.
3577 : */
3578 0 : spin_lock_irqsave(&tags->lock, flags);
3579 0 : spin_unlock_irqrestore(&tags->lock, flags);
3580 : }
3581 :
3582 : /* hctx->ctxs will be freed in queue's release handler */
3583 0 : static void blk_mq_exit_hctx(struct request_queue *q,
3584 : struct blk_mq_tag_set *set,
3585 : struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
3586 : {
3587 0 : struct request *flush_rq = hctx->fq->flush_rq;
3588 :
3589 0 : if (blk_mq_hw_queue_mapped(hctx))
3590 : blk_mq_tag_idle(hctx);
3591 :
3592 0 : if (blk_queue_init_done(q))
3593 0 : blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
3594 : set->queue_depth, flush_rq);
3595 0 : if (set->ops->exit_request)
3596 0 : set->ops->exit_request(set, flush_rq, hctx_idx);
3597 :
3598 0 : if (set->ops->exit_hctx)
3599 0 : set->ops->exit_hctx(hctx, hctx_idx);
3600 :
3601 0 : blk_mq_remove_cpuhp(hctx);
3602 :
3603 0 : xa_erase(&q->hctx_table, hctx_idx);
3604 :
3605 0 : spin_lock(&q->unused_hctx_lock);
3606 0 : list_add(&hctx->hctx_list, &q->unused_hctx_list);
3607 0 : spin_unlock(&q->unused_hctx_lock);
3608 0 : }
3609 :
3610 0 : static void blk_mq_exit_hw_queues(struct request_queue *q,
3611 : struct blk_mq_tag_set *set, int nr_queue)
3612 : {
3613 : struct blk_mq_hw_ctx *hctx;
3614 : unsigned long i;
3615 :
3616 0 : queue_for_each_hw_ctx(q, hctx, i) {
3617 0 : if (i == nr_queue)
3618 : break;
3619 0 : blk_mq_exit_hctx(q, set, hctx, i);
3620 : }
3621 0 : }
3622 :
3623 0 : static int blk_mq_init_hctx(struct request_queue *q,
3624 : struct blk_mq_tag_set *set,
3625 : struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
3626 : {
3627 0 : hctx->queue_num = hctx_idx;
3628 :
3629 0 : if (!(hctx->flags & BLK_MQ_F_STACKING))
3630 0 : cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE,
3631 : &hctx->cpuhp_online);
3632 0 : cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
3633 :
3634 0 : hctx->tags = set->tags[hctx_idx];
3635 :
3636 0 : if (set->ops->init_hctx &&
3637 0 : set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
3638 : goto unregister_cpu_notifier;
3639 :
3640 0 : if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
3641 0 : hctx->numa_node))
3642 : goto exit_hctx;
3643 :
3644 0 : if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
3645 : goto exit_flush_rq;
3646 :
3647 : return 0;
3648 :
3649 : exit_flush_rq:
3650 0 : if (set->ops->exit_request)
3651 0 : set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
3652 : exit_hctx:
3653 0 : if (set->ops->exit_hctx)
3654 0 : set->ops->exit_hctx(hctx, hctx_idx);
3655 : unregister_cpu_notifier:
3656 0 : blk_mq_remove_cpuhp(hctx);
3657 0 : return -1;
3658 : }
3659 :
3660 : static struct blk_mq_hw_ctx *
3661 0 : blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
3662 : int node)
3663 : {
3664 : struct blk_mq_hw_ctx *hctx;
3665 0 : gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
3666 :
3667 0 : hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node);
3668 0 : if (!hctx)
3669 : goto fail_alloc_hctx;
3670 :
3671 0 : if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
3672 : goto free_hctx;
3673 :
3674 0 : atomic_set(&hctx->nr_active, 0);
3675 0 : if (node == NUMA_NO_NODE)
3676 0 : node = set->numa_node;
3677 0 : hctx->numa_node = node;
3678 :
3679 0 : INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
3680 0 : spin_lock_init(&hctx->lock);
3681 0 : INIT_LIST_HEAD(&hctx->dispatch);
3682 0 : hctx->queue = q;
3683 0 : hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED;
3684 :
3685 0 : INIT_LIST_HEAD(&hctx->hctx_list);
3686 :
3687 : /*
3688 : * Allocate space for all possible cpus to avoid allocation at
3689 : * runtime
3690 : */
3691 0 : hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
3692 : gfp, node);
3693 0 : if (!hctx->ctxs)
3694 : goto free_cpumask;
3695 :
3696 0 : if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
3697 : gfp, node, false, false))
3698 : goto free_ctxs;
3699 0 : hctx->nr_ctx = 0;
3700 :
3701 0 : spin_lock_init(&hctx->dispatch_wait_lock);
3702 0 : init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
3703 0 : INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
3704 :
3705 0 : hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp);
3706 0 : if (!hctx->fq)
3707 : goto free_bitmap;
3708 :
3709 0 : blk_mq_hctx_kobj_init(hctx);
3710 :
3711 0 : return hctx;
3712 :
3713 : free_bitmap:
3714 0 : sbitmap_free(&hctx->ctx_map);
3715 : free_ctxs:
3716 0 : kfree(hctx->ctxs);
3717 : free_cpumask:
3718 0 : free_cpumask_var(hctx->cpumask);
3719 : free_hctx:
3720 0 : kfree(hctx);
3721 : fail_alloc_hctx:
3722 : return NULL;
3723 : }
3724 :
3725 0 : static void blk_mq_init_cpu_queues(struct request_queue *q,
3726 : unsigned int nr_hw_queues)
3727 : {
3728 0 : struct blk_mq_tag_set *set = q->tag_set;
3729 : unsigned int i, j;
3730 :
3731 0 : for_each_possible_cpu(i) {
3732 0 : struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
3733 : struct blk_mq_hw_ctx *hctx;
3734 : int k;
3735 :
3736 0 : __ctx->cpu = i;
3737 0 : spin_lock_init(&__ctx->lock);
3738 0 : for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
3739 0 : INIT_LIST_HEAD(&__ctx->rq_lists[k]);
3740 :
3741 0 : __ctx->queue = q;
3742 :
3743 : /*
3744 : * Set local node, IFF we have more than one hw queue. If
3745 : * not, we remain on the home node of the device
3746 : */
3747 0 : for (j = 0; j < set->nr_maps; j++) {
3748 0 : hctx = blk_mq_map_queue_type(q, j, i);
3749 0 : if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
3750 0 : hctx->numa_node = cpu_to_node(i);
3751 : }
3752 : }
3753 0 : }
3754 :
3755 0 : struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3756 : unsigned int hctx_idx,
3757 : unsigned int depth)
3758 : {
3759 : struct blk_mq_tags *tags;
3760 : int ret;
3761 :
3762 0 : tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags);
3763 0 : if (!tags)
3764 : return NULL;
3765 :
3766 0 : ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth);
3767 0 : if (ret) {
3768 0 : blk_mq_free_rq_map(tags);
3769 0 : return NULL;
3770 : }
3771 :
3772 : return tags;
3773 : }
3774 :
3775 0 : static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
3776 : int hctx_idx)
3777 : {
3778 0 : if (blk_mq_is_shared_tags(set->flags)) {
3779 0 : set->tags[hctx_idx] = set->shared_tags;
3780 :
3781 0 : return true;
3782 : }
3783 :
3784 0 : set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx,
3785 : set->queue_depth);
3786 :
3787 0 : return set->tags[hctx_idx];
3788 : }
3789 :
3790 0 : void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3791 : struct blk_mq_tags *tags,
3792 : unsigned int hctx_idx)
3793 : {
3794 0 : if (tags) {
3795 0 : blk_mq_free_rqs(set, tags, hctx_idx);
3796 0 : blk_mq_free_rq_map(tags);
3797 : }
3798 0 : }
3799 :
3800 0 : static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
3801 : unsigned int hctx_idx)
3802 : {
3803 0 : if (!blk_mq_is_shared_tags(set->flags))
3804 0 : blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx);
3805 :
3806 0 : set->tags[hctx_idx] = NULL;
3807 0 : }
3808 :
3809 0 : static void blk_mq_map_swqueue(struct request_queue *q)
3810 : {
3811 : unsigned int j, hctx_idx;
3812 : unsigned long i;
3813 : struct blk_mq_hw_ctx *hctx;
3814 : struct blk_mq_ctx *ctx;
3815 0 : struct blk_mq_tag_set *set = q->tag_set;
3816 :
3817 0 : queue_for_each_hw_ctx(q, hctx, i) {
3818 0 : cpumask_clear(hctx->cpumask);
3819 0 : hctx->nr_ctx = 0;
3820 0 : hctx->dispatch_from = NULL;
3821 : }
3822 :
3823 : /*
3824 : * Map software to hardware queues.
3825 : *
3826 : * If the cpu isn't present, the cpu is mapped to first hctx.
3827 : */
3828 0 : for_each_possible_cpu(i) {
3829 :
3830 0 : ctx = per_cpu_ptr(q->queue_ctx, i);
3831 0 : for (j = 0; j < set->nr_maps; j++) {
3832 0 : if (!set->map[j].nr_queues) {
3833 0 : ctx->hctxs[j] = blk_mq_map_queue_type(q,
3834 : HCTX_TYPE_DEFAULT, i);
3835 0 : continue;
3836 : }
3837 0 : hctx_idx = set->map[j].mq_map[i];
3838 : /* unmapped hw queue can be remapped after CPU topo changed */
3839 0 : if (!set->tags[hctx_idx] &&
3840 0 : !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) {
3841 : /*
3842 : * If tags initialization fail for some hctx,
3843 : * that hctx won't be brought online. In this
3844 : * case, remap the current ctx to hctx[0] which
3845 : * is guaranteed to always have tags allocated
3846 : */
3847 0 : set->map[j].mq_map[i] = 0;
3848 : }
3849 :
3850 0 : hctx = blk_mq_map_queue_type(q, j, i);
3851 0 : ctx->hctxs[j] = hctx;
3852 : /*
3853 : * If the CPU is already set in the mask, then we've
3854 : * mapped this one already. This can happen if
3855 : * devices share queues across queue maps.
3856 : */
3857 0 : if (cpumask_test_cpu(i, hctx->cpumask))
3858 0 : continue;
3859 :
3860 0 : cpumask_set_cpu(i, hctx->cpumask);
3861 0 : hctx->type = j;
3862 0 : ctx->index_hw[hctx->type] = hctx->nr_ctx;
3863 0 : hctx->ctxs[hctx->nr_ctx++] = ctx;
3864 :
3865 : /*
3866 : * If the nr_ctx type overflows, we have exceeded the
3867 : * amount of sw queues we can support.
3868 : */
3869 0 : BUG_ON(!hctx->nr_ctx);
3870 : }
3871 :
3872 0 : for (; j < HCTX_MAX_TYPES; j++)
3873 0 : ctx->hctxs[j] = blk_mq_map_queue_type(q,
3874 : HCTX_TYPE_DEFAULT, i);
3875 : }
3876 :
3877 0 : queue_for_each_hw_ctx(q, hctx, i) {
3878 : /*
3879 : * If no software queues are mapped to this hardware queue,
3880 : * disable it and free the request entries.
3881 : */
3882 0 : if (!hctx->nr_ctx) {
3883 : /* Never unmap queue 0. We need it as a
3884 : * fallback in case of a new remap fails
3885 : * allocation
3886 : */
3887 0 : if (i)
3888 0 : __blk_mq_free_map_and_rqs(set, i);
3889 :
3890 0 : hctx->tags = NULL;
3891 0 : continue;
3892 : }
3893 :
3894 0 : hctx->tags = set->tags[i];
3895 0 : WARN_ON(!hctx->tags);
3896 :
3897 : /*
3898 : * Set the map size to the number of mapped software queues.
3899 : * This is more accurate and more efficient than looping
3900 : * over all possibly mapped software queues.
3901 : */
3902 0 : sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
3903 :
3904 : /*
3905 : * Initialize batch roundrobin counts
3906 : */
3907 0 : hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
3908 0 : hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
3909 : }
3910 0 : }
3911 :
3912 : /*
3913 : * Caller needs to ensure that we're either frozen/quiesced, or that
3914 : * the queue isn't live yet.
3915 : */
3916 0 : static void queue_set_hctx_shared(struct request_queue *q, bool shared)
3917 : {
3918 : struct blk_mq_hw_ctx *hctx;
3919 : unsigned long i;
3920 :
3921 0 : queue_for_each_hw_ctx(q, hctx, i) {
3922 0 : if (shared) {
3923 0 : hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3924 : } else {
3925 0 : blk_mq_tag_idle(hctx);
3926 0 : hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3927 : }
3928 : }
3929 0 : }
3930 :
3931 0 : static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set,
3932 : bool shared)
3933 : {
3934 : struct request_queue *q;
3935 :
3936 : lockdep_assert_held(&set->tag_list_lock);
3937 :
3938 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
3939 0 : blk_mq_freeze_queue(q);
3940 0 : queue_set_hctx_shared(q, shared);
3941 0 : blk_mq_unfreeze_queue(q);
3942 : }
3943 0 : }
3944 :
3945 0 : static void blk_mq_del_queue_tag_set(struct request_queue *q)
3946 : {
3947 0 : struct blk_mq_tag_set *set = q->tag_set;
3948 :
3949 0 : mutex_lock(&set->tag_list_lock);
3950 0 : list_del(&q->tag_set_list);
3951 0 : if (list_is_singular(&set->tag_list)) {
3952 : /* just transitioned to unshared */
3953 0 : set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
3954 : /* update existing queue */
3955 0 : blk_mq_update_tag_set_shared(set, false);
3956 : }
3957 0 : mutex_unlock(&set->tag_list_lock);
3958 0 : INIT_LIST_HEAD(&q->tag_set_list);
3959 0 : }
3960 :
3961 0 : static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
3962 : struct request_queue *q)
3963 : {
3964 0 : mutex_lock(&set->tag_list_lock);
3965 :
3966 : /*
3967 : * Check to see if we're transitioning to shared (from 1 to 2 queues).
3968 : */
3969 0 : if (!list_empty(&set->tag_list) &&
3970 0 : !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
3971 0 : set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
3972 : /* update existing queue */
3973 0 : blk_mq_update_tag_set_shared(set, true);
3974 : }
3975 0 : if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
3976 0 : queue_set_hctx_shared(q, true);
3977 0 : list_add_tail(&q->tag_set_list, &set->tag_list);
3978 :
3979 0 : mutex_unlock(&set->tag_list_lock);
3980 0 : }
3981 :
3982 : /* All allocations will be freed in release handler of q->mq_kobj */
3983 0 : static int blk_mq_alloc_ctxs(struct request_queue *q)
3984 : {
3985 : struct blk_mq_ctxs *ctxs;
3986 : int cpu;
3987 :
3988 0 : ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
3989 0 : if (!ctxs)
3990 : return -ENOMEM;
3991 :
3992 0 : ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
3993 0 : if (!ctxs->queue_ctx)
3994 : goto fail;
3995 :
3996 0 : for_each_possible_cpu(cpu) {
3997 0 : struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
3998 0 : ctx->ctxs = ctxs;
3999 : }
4000 :
4001 0 : q->mq_kobj = &ctxs->kobj;
4002 0 : q->queue_ctx = ctxs->queue_ctx;
4003 :
4004 : return 0;
4005 : fail:
4006 0 : kfree(ctxs);
4007 : return -ENOMEM;
4008 : }
4009 :
4010 : /*
4011 : * It is the actual release handler for mq, but we do it from
4012 : * request queue's release handler for avoiding use-after-free
4013 : * and headache because q->mq_kobj shouldn't have been introduced,
4014 : * but we can't group ctx/kctx kobj without it.
4015 : */
4016 0 : void blk_mq_release(struct request_queue *q)
4017 : {
4018 : struct blk_mq_hw_ctx *hctx, *next;
4019 : unsigned long i;
4020 :
4021 0 : queue_for_each_hw_ctx(q, hctx, i)
4022 0 : WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
4023 :
4024 : /* all hctx are in .unused_hctx_list now */
4025 0 : list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
4026 0 : list_del_init(&hctx->hctx_list);
4027 0 : kobject_put(&hctx->kobj);
4028 : }
4029 :
4030 0 : xa_destroy(&q->hctx_table);
4031 :
4032 : /*
4033 : * release .mq_kobj and sw queue's kobject now because
4034 : * both share lifetime with request queue.
4035 : */
4036 0 : blk_mq_sysfs_deinit(q);
4037 0 : }
4038 :
4039 0 : static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set,
4040 : void *queuedata)
4041 : {
4042 : struct request_queue *q;
4043 : int ret;
4044 :
4045 0 : q = blk_alloc_queue(set->numa_node);
4046 0 : if (!q)
4047 : return ERR_PTR(-ENOMEM);
4048 0 : q->queuedata = queuedata;
4049 0 : ret = blk_mq_init_allocated_queue(set, q);
4050 0 : if (ret) {
4051 0 : blk_put_queue(q);
4052 0 : return ERR_PTR(ret);
4053 : }
4054 : return q;
4055 : }
4056 :
4057 0 : struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
4058 : {
4059 0 : return blk_mq_init_queue_data(set, NULL);
4060 : }
4061 : EXPORT_SYMBOL(blk_mq_init_queue);
4062 :
4063 : /**
4064 : * blk_mq_destroy_queue - shutdown a request queue
4065 : * @q: request queue to shutdown
4066 : *
4067 : * This shuts down a request queue allocated by blk_mq_init_queue(). All future
4068 : * requests will be failed with -ENODEV. The caller is responsible for dropping
4069 : * the reference from blk_mq_init_queue() by calling blk_put_queue().
4070 : *
4071 : * Context: can sleep
4072 : */
4073 0 : void blk_mq_destroy_queue(struct request_queue *q)
4074 : {
4075 0 : WARN_ON_ONCE(!queue_is_mq(q));
4076 0 : WARN_ON_ONCE(blk_queue_registered(q));
4077 :
4078 : might_sleep();
4079 :
4080 0 : blk_queue_flag_set(QUEUE_FLAG_DYING, q);
4081 0 : blk_queue_start_drain(q);
4082 0 : blk_mq_freeze_queue_wait(q);
4083 :
4084 0 : blk_sync_queue(q);
4085 0 : blk_mq_cancel_work_sync(q);
4086 0 : blk_mq_exit_queue(q);
4087 0 : }
4088 : EXPORT_SYMBOL(blk_mq_destroy_queue);
4089 :
4090 0 : struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata,
4091 : struct lock_class_key *lkclass)
4092 : {
4093 : struct request_queue *q;
4094 : struct gendisk *disk;
4095 :
4096 0 : q = blk_mq_init_queue_data(set, queuedata);
4097 0 : if (IS_ERR(q))
4098 : return ERR_CAST(q);
4099 :
4100 0 : disk = __alloc_disk_node(q, set->numa_node, lkclass);
4101 0 : if (!disk) {
4102 0 : blk_mq_destroy_queue(q);
4103 0 : blk_put_queue(q);
4104 0 : return ERR_PTR(-ENOMEM);
4105 : }
4106 0 : set_bit(GD_OWNS_QUEUE, &disk->state);
4107 0 : return disk;
4108 : }
4109 : EXPORT_SYMBOL(__blk_mq_alloc_disk);
4110 :
4111 0 : struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q,
4112 : struct lock_class_key *lkclass)
4113 : {
4114 : struct gendisk *disk;
4115 :
4116 0 : if (!blk_get_queue(q))
4117 : return NULL;
4118 0 : disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass);
4119 0 : if (!disk)
4120 0 : blk_put_queue(q);
4121 : return disk;
4122 : }
4123 : EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue);
4124 :
4125 0 : static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
4126 : struct blk_mq_tag_set *set, struct request_queue *q,
4127 : int hctx_idx, int node)
4128 : {
4129 0 : struct blk_mq_hw_ctx *hctx = NULL, *tmp;
4130 :
4131 : /* reuse dead hctx first */
4132 0 : spin_lock(&q->unused_hctx_lock);
4133 0 : list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
4134 0 : if (tmp->numa_node == node) {
4135 : hctx = tmp;
4136 : break;
4137 : }
4138 : }
4139 0 : if (hctx)
4140 0 : list_del_init(&hctx->hctx_list);
4141 0 : spin_unlock(&q->unused_hctx_lock);
4142 :
4143 0 : if (!hctx)
4144 0 : hctx = blk_mq_alloc_hctx(q, set, node);
4145 0 : if (!hctx)
4146 : goto fail;
4147 :
4148 0 : if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
4149 : goto free_hctx;
4150 :
4151 : return hctx;
4152 :
4153 : free_hctx:
4154 0 : kobject_put(&hctx->kobj);
4155 : fail:
4156 : return NULL;
4157 : }
4158 :
4159 0 : static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
4160 : struct request_queue *q)
4161 : {
4162 : struct blk_mq_hw_ctx *hctx;
4163 : unsigned long i, j;
4164 :
4165 : /* protect against switching io scheduler */
4166 0 : mutex_lock(&q->sysfs_lock);
4167 0 : for (i = 0; i < set->nr_hw_queues; i++) {
4168 : int old_node;
4169 0 : int node = blk_mq_get_hctx_node(set, i);
4170 0 : struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
4171 :
4172 0 : if (old_hctx) {
4173 0 : old_node = old_hctx->numa_node;
4174 0 : blk_mq_exit_hctx(q, set, old_hctx, i);
4175 : }
4176 :
4177 0 : if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
4178 0 : if (!old_hctx)
4179 : break;
4180 0 : pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
4181 : node, old_node);
4182 0 : hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
4183 0 : WARN_ON_ONCE(!hctx);
4184 : }
4185 : }
4186 : /*
4187 : * Increasing nr_hw_queues fails. Free the newly allocated
4188 : * hctxs and keep the previous q->nr_hw_queues.
4189 : */
4190 0 : if (i != set->nr_hw_queues) {
4191 0 : j = q->nr_hw_queues;
4192 : } else {
4193 0 : j = i;
4194 0 : q->nr_hw_queues = set->nr_hw_queues;
4195 : }
4196 :
4197 0 : xa_for_each_start(&q->hctx_table, j, hctx, j)
4198 0 : blk_mq_exit_hctx(q, set, hctx, j);
4199 0 : mutex_unlock(&q->sysfs_lock);
4200 0 : }
4201 :
4202 0 : static void blk_mq_update_poll_flag(struct request_queue *q)
4203 : {
4204 0 : struct blk_mq_tag_set *set = q->tag_set;
4205 :
4206 0 : if (set->nr_maps > HCTX_TYPE_POLL &&
4207 0 : set->map[HCTX_TYPE_POLL].nr_queues)
4208 0 : blk_queue_flag_set(QUEUE_FLAG_POLL, q);
4209 : else
4210 0 : blk_queue_flag_clear(QUEUE_FLAG_POLL, q);
4211 0 : }
4212 :
4213 0 : int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
4214 : struct request_queue *q)
4215 : {
4216 : /* mark the queue as mq asap */
4217 0 : q->mq_ops = set->ops;
4218 :
4219 0 : if (blk_mq_alloc_ctxs(q))
4220 : goto err_exit;
4221 :
4222 : /* init q->mq_kobj and sw queues' kobjects */
4223 0 : blk_mq_sysfs_init(q);
4224 :
4225 0 : INIT_LIST_HEAD(&q->unused_hctx_list);
4226 0 : spin_lock_init(&q->unused_hctx_lock);
4227 :
4228 0 : xa_init(&q->hctx_table);
4229 :
4230 0 : blk_mq_realloc_hw_ctxs(set, q);
4231 0 : if (!q->nr_hw_queues)
4232 : goto err_hctxs;
4233 :
4234 0 : INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
4235 0 : blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
4236 :
4237 0 : q->tag_set = set;
4238 :
4239 0 : q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
4240 0 : blk_mq_update_poll_flag(q);
4241 :
4242 0 : INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
4243 0 : INIT_LIST_HEAD(&q->flush_list);
4244 0 : INIT_LIST_HEAD(&q->requeue_list);
4245 0 : spin_lock_init(&q->requeue_lock);
4246 :
4247 0 : q->nr_requests = set->queue_depth;
4248 :
4249 0 : blk_mq_init_cpu_queues(q, set->nr_hw_queues);
4250 0 : blk_mq_add_queue_tag_set(set, q);
4251 0 : blk_mq_map_swqueue(q);
4252 0 : return 0;
4253 :
4254 : err_hctxs:
4255 0 : blk_mq_release(q);
4256 : err_exit:
4257 0 : q->mq_ops = NULL;
4258 0 : return -ENOMEM;
4259 : }
4260 : EXPORT_SYMBOL(blk_mq_init_allocated_queue);
4261 :
4262 : /* tags can _not_ be used after returning from blk_mq_exit_queue */
4263 0 : void blk_mq_exit_queue(struct request_queue *q)
4264 : {
4265 0 : struct blk_mq_tag_set *set = q->tag_set;
4266 :
4267 : /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
4268 0 : blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
4269 : /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
4270 0 : blk_mq_del_queue_tag_set(q);
4271 0 : }
4272 :
4273 0 : static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
4274 : {
4275 : int i;
4276 :
4277 0 : if (blk_mq_is_shared_tags(set->flags)) {
4278 0 : set->shared_tags = blk_mq_alloc_map_and_rqs(set,
4279 : BLK_MQ_NO_HCTX_IDX,
4280 : set->queue_depth);
4281 0 : if (!set->shared_tags)
4282 : return -ENOMEM;
4283 : }
4284 :
4285 0 : for (i = 0; i < set->nr_hw_queues; i++) {
4286 0 : if (!__blk_mq_alloc_map_and_rqs(set, i))
4287 : goto out_unwind;
4288 0 : cond_resched();
4289 : }
4290 :
4291 : return 0;
4292 :
4293 : out_unwind:
4294 0 : while (--i >= 0)
4295 0 : __blk_mq_free_map_and_rqs(set, i);
4296 :
4297 0 : if (blk_mq_is_shared_tags(set->flags)) {
4298 0 : blk_mq_free_map_and_rqs(set, set->shared_tags,
4299 : BLK_MQ_NO_HCTX_IDX);
4300 : }
4301 :
4302 : return -ENOMEM;
4303 : }
4304 :
4305 : /*
4306 : * Allocate the request maps associated with this tag_set. Note that this
4307 : * may reduce the depth asked for, if memory is tight. set->queue_depth
4308 : * will be updated to reflect the allocated depth.
4309 : */
4310 0 : static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set)
4311 : {
4312 : unsigned int depth;
4313 : int err;
4314 :
4315 0 : depth = set->queue_depth;
4316 : do {
4317 0 : err = __blk_mq_alloc_rq_maps(set);
4318 0 : if (!err)
4319 : break;
4320 :
4321 0 : set->queue_depth >>= 1;
4322 0 : if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
4323 : err = -ENOMEM;
4324 : break;
4325 : }
4326 0 : } while (set->queue_depth);
4327 :
4328 0 : if (!set->queue_depth || err) {
4329 0 : pr_err("blk-mq: failed to allocate request map\n");
4330 0 : return -ENOMEM;
4331 : }
4332 :
4333 0 : if (depth != set->queue_depth)
4334 0 : pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
4335 : depth, set->queue_depth);
4336 :
4337 : return 0;
4338 : }
4339 :
4340 0 : static void blk_mq_update_queue_map(struct blk_mq_tag_set *set)
4341 : {
4342 : /*
4343 : * blk_mq_map_queues() and multiple .map_queues() implementations
4344 : * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
4345 : * number of hardware queues.
4346 : */
4347 0 : if (set->nr_maps == 1)
4348 0 : set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
4349 :
4350 0 : if (set->ops->map_queues && !is_kdump_kernel()) {
4351 : int i;
4352 :
4353 : /*
4354 : * transport .map_queues is usually done in the following
4355 : * way:
4356 : *
4357 : * for (queue = 0; queue < set->nr_hw_queues; queue++) {
4358 : * mask = get_cpu_mask(queue)
4359 : * for_each_cpu(cpu, mask)
4360 : * set->map[x].mq_map[cpu] = queue;
4361 : * }
4362 : *
4363 : * When we need to remap, the table has to be cleared for
4364 : * killing stale mapping since one CPU may not be mapped
4365 : * to any hw queue.
4366 : */
4367 0 : for (i = 0; i < set->nr_maps; i++)
4368 0 : blk_mq_clear_mq_map(&set->map[i]);
4369 :
4370 0 : set->ops->map_queues(set);
4371 : } else {
4372 0 : BUG_ON(set->nr_maps > 1);
4373 0 : blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4374 : }
4375 0 : }
4376 :
4377 0 : static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
4378 : int new_nr_hw_queues)
4379 : {
4380 : struct blk_mq_tags **new_tags;
4381 :
4382 0 : if (set->nr_hw_queues >= new_nr_hw_queues)
4383 : goto done;
4384 :
4385 0 : new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
4386 : GFP_KERNEL, set->numa_node);
4387 0 : if (!new_tags)
4388 : return -ENOMEM;
4389 :
4390 0 : if (set->tags)
4391 0 : memcpy(new_tags, set->tags, set->nr_hw_queues *
4392 : sizeof(*set->tags));
4393 0 : kfree(set->tags);
4394 0 : set->tags = new_tags;
4395 : done:
4396 0 : set->nr_hw_queues = new_nr_hw_queues;
4397 0 : return 0;
4398 : }
4399 :
4400 : /*
4401 : * Alloc a tag set to be associated with one or more request queues.
4402 : * May fail with EINVAL for various error conditions. May adjust the
4403 : * requested depth down, if it's too large. In that case, the set
4404 : * value will be stored in set->queue_depth.
4405 : */
4406 0 : int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
4407 : {
4408 : int i, ret;
4409 :
4410 : BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
4411 :
4412 0 : if (!set->nr_hw_queues)
4413 : return -EINVAL;
4414 0 : if (!set->queue_depth)
4415 : return -EINVAL;
4416 0 : if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
4417 : return -EINVAL;
4418 :
4419 0 : if (!set->ops->queue_rq)
4420 : return -EINVAL;
4421 :
4422 0 : if (!set->ops->get_budget ^ !set->ops->put_budget)
4423 : return -EINVAL;
4424 :
4425 0 : if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
4426 0 : pr_info("blk-mq: reduced tag depth to %u\n",
4427 : BLK_MQ_MAX_DEPTH);
4428 0 : set->queue_depth = BLK_MQ_MAX_DEPTH;
4429 : }
4430 :
4431 0 : if (!set->nr_maps)
4432 0 : set->nr_maps = 1;
4433 0 : else if (set->nr_maps > HCTX_MAX_TYPES)
4434 : return -EINVAL;
4435 :
4436 : /*
4437 : * If a crashdump is active, then we are potentially in a very
4438 : * memory constrained environment. Limit us to 1 queue and
4439 : * 64 tags to prevent using too much memory.
4440 : */
4441 : if (is_kdump_kernel()) {
4442 : set->nr_hw_queues = 1;
4443 : set->nr_maps = 1;
4444 : set->queue_depth = min(64U, set->queue_depth);
4445 : }
4446 : /*
4447 : * There is no use for more h/w queues than cpus if we just have
4448 : * a single map
4449 : */
4450 0 : if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
4451 0 : set->nr_hw_queues = nr_cpu_ids;
4452 :
4453 0 : if (set->flags & BLK_MQ_F_BLOCKING) {
4454 0 : set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL);
4455 0 : if (!set->srcu)
4456 : return -ENOMEM;
4457 0 : ret = init_srcu_struct(set->srcu);
4458 0 : if (ret)
4459 : goto out_free_srcu;
4460 : }
4461 :
4462 0 : ret = -ENOMEM;
4463 0 : set->tags = kcalloc_node(set->nr_hw_queues,
4464 : sizeof(struct blk_mq_tags *), GFP_KERNEL,
4465 : set->numa_node);
4466 0 : if (!set->tags)
4467 : goto out_cleanup_srcu;
4468 :
4469 0 : for (i = 0; i < set->nr_maps; i++) {
4470 0 : set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
4471 : sizeof(set->map[i].mq_map[0]),
4472 : GFP_KERNEL, set->numa_node);
4473 0 : if (!set->map[i].mq_map)
4474 : goto out_free_mq_map;
4475 0 : set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
4476 : }
4477 :
4478 0 : blk_mq_update_queue_map(set);
4479 :
4480 0 : ret = blk_mq_alloc_set_map_and_rqs(set);
4481 0 : if (ret)
4482 : goto out_free_mq_map;
4483 :
4484 0 : mutex_init(&set->tag_list_lock);
4485 0 : INIT_LIST_HEAD(&set->tag_list);
4486 :
4487 0 : return 0;
4488 :
4489 : out_free_mq_map:
4490 0 : for (i = 0; i < set->nr_maps; i++) {
4491 0 : kfree(set->map[i].mq_map);
4492 0 : set->map[i].mq_map = NULL;
4493 : }
4494 0 : kfree(set->tags);
4495 0 : set->tags = NULL;
4496 : out_cleanup_srcu:
4497 0 : if (set->flags & BLK_MQ_F_BLOCKING)
4498 0 : cleanup_srcu_struct(set->srcu);
4499 : out_free_srcu:
4500 0 : if (set->flags & BLK_MQ_F_BLOCKING)
4501 0 : kfree(set->srcu);
4502 : return ret;
4503 : }
4504 : EXPORT_SYMBOL(blk_mq_alloc_tag_set);
4505 :
4506 : /* allocate and initialize a tagset for a simple single-queue device */
4507 0 : int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set,
4508 : const struct blk_mq_ops *ops, unsigned int queue_depth,
4509 : unsigned int set_flags)
4510 : {
4511 0 : memset(set, 0, sizeof(*set));
4512 0 : set->ops = ops;
4513 0 : set->nr_hw_queues = 1;
4514 0 : set->nr_maps = 1;
4515 0 : set->queue_depth = queue_depth;
4516 0 : set->numa_node = NUMA_NO_NODE;
4517 0 : set->flags = set_flags;
4518 0 : return blk_mq_alloc_tag_set(set);
4519 : }
4520 : EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set);
4521 :
4522 0 : void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
4523 : {
4524 : int i, j;
4525 :
4526 0 : for (i = 0; i < set->nr_hw_queues; i++)
4527 0 : __blk_mq_free_map_and_rqs(set, i);
4528 :
4529 0 : if (blk_mq_is_shared_tags(set->flags)) {
4530 0 : blk_mq_free_map_and_rqs(set, set->shared_tags,
4531 : BLK_MQ_NO_HCTX_IDX);
4532 : }
4533 :
4534 0 : for (j = 0; j < set->nr_maps; j++) {
4535 0 : kfree(set->map[j].mq_map);
4536 0 : set->map[j].mq_map = NULL;
4537 : }
4538 :
4539 0 : kfree(set->tags);
4540 0 : set->tags = NULL;
4541 0 : if (set->flags & BLK_MQ_F_BLOCKING) {
4542 0 : cleanup_srcu_struct(set->srcu);
4543 0 : kfree(set->srcu);
4544 : }
4545 0 : }
4546 : EXPORT_SYMBOL(blk_mq_free_tag_set);
4547 :
4548 0 : int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
4549 : {
4550 0 : struct blk_mq_tag_set *set = q->tag_set;
4551 : struct blk_mq_hw_ctx *hctx;
4552 : int ret;
4553 : unsigned long i;
4554 :
4555 0 : if (!set)
4556 : return -EINVAL;
4557 :
4558 0 : if (q->nr_requests == nr)
4559 : return 0;
4560 :
4561 0 : blk_mq_freeze_queue(q);
4562 0 : blk_mq_quiesce_queue(q);
4563 :
4564 0 : ret = 0;
4565 0 : queue_for_each_hw_ctx(q, hctx, i) {
4566 0 : if (!hctx->tags)
4567 0 : continue;
4568 : /*
4569 : * If we're using an MQ scheduler, just update the scheduler
4570 : * queue depth. This is similar to what the old code would do.
4571 : */
4572 0 : if (hctx->sched_tags) {
4573 0 : ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
4574 : nr, true);
4575 : } else {
4576 0 : ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
4577 : false);
4578 : }
4579 0 : if (ret)
4580 : break;
4581 0 : if (q->elevator && q->elevator->type->ops.depth_updated)
4582 0 : q->elevator->type->ops.depth_updated(hctx);
4583 : }
4584 0 : if (!ret) {
4585 0 : q->nr_requests = nr;
4586 0 : if (blk_mq_is_shared_tags(set->flags)) {
4587 0 : if (q->elevator)
4588 0 : blk_mq_tag_update_sched_shared_tags(q);
4589 : else
4590 0 : blk_mq_tag_resize_shared_tags(set, nr);
4591 : }
4592 : }
4593 :
4594 0 : blk_mq_unquiesce_queue(q);
4595 0 : blk_mq_unfreeze_queue(q);
4596 :
4597 0 : return ret;
4598 : }
4599 :
4600 : /*
4601 : * request_queue and elevator_type pair.
4602 : * It is just used by __blk_mq_update_nr_hw_queues to cache
4603 : * the elevator_type associated with a request_queue.
4604 : */
4605 : struct blk_mq_qe_pair {
4606 : struct list_head node;
4607 : struct request_queue *q;
4608 : struct elevator_type *type;
4609 : };
4610 :
4611 : /*
4612 : * Cache the elevator_type in qe pair list and switch the
4613 : * io scheduler to 'none'
4614 : */
4615 0 : static bool blk_mq_elv_switch_none(struct list_head *head,
4616 : struct request_queue *q)
4617 : {
4618 : struct blk_mq_qe_pair *qe;
4619 :
4620 0 : qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
4621 0 : if (!qe)
4622 : return false;
4623 :
4624 : /* q->elevator needs protection from ->sysfs_lock */
4625 0 : mutex_lock(&q->sysfs_lock);
4626 :
4627 : /* the check has to be done with holding sysfs_lock */
4628 0 : if (!q->elevator) {
4629 0 : kfree(qe);
4630 0 : goto unlock;
4631 : }
4632 :
4633 0 : INIT_LIST_HEAD(&qe->node);
4634 0 : qe->q = q;
4635 0 : qe->type = q->elevator->type;
4636 : /* keep a reference to the elevator module as we'll switch back */
4637 0 : __elevator_get(qe->type);
4638 0 : list_add(&qe->node, head);
4639 0 : elevator_disable(q);
4640 : unlock:
4641 0 : mutex_unlock(&q->sysfs_lock);
4642 :
4643 0 : return true;
4644 : }
4645 :
4646 : static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head,
4647 : struct request_queue *q)
4648 : {
4649 : struct blk_mq_qe_pair *qe;
4650 :
4651 0 : list_for_each_entry(qe, head, node)
4652 0 : if (qe->q == q)
4653 : return qe;
4654 :
4655 : return NULL;
4656 : }
4657 :
4658 0 : static void blk_mq_elv_switch_back(struct list_head *head,
4659 : struct request_queue *q)
4660 : {
4661 : struct blk_mq_qe_pair *qe;
4662 : struct elevator_type *t;
4663 :
4664 0 : qe = blk_lookup_qe_pair(head, q);
4665 0 : if (!qe)
4666 : return;
4667 0 : t = qe->type;
4668 0 : list_del(&qe->node);
4669 0 : kfree(qe);
4670 :
4671 0 : mutex_lock(&q->sysfs_lock);
4672 0 : elevator_switch(q, t);
4673 : /* drop the reference acquired in blk_mq_elv_switch_none */
4674 0 : elevator_put(t);
4675 0 : mutex_unlock(&q->sysfs_lock);
4676 : }
4677 :
4678 0 : static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
4679 : int nr_hw_queues)
4680 : {
4681 : struct request_queue *q;
4682 0 : LIST_HEAD(head);
4683 : int prev_nr_hw_queues;
4684 :
4685 : lockdep_assert_held(&set->tag_list_lock);
4686 :
4687 0 : if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
4688 0 : nr_hw_queues = nr_cpu_ids;
4689 0 : if (nr_hw_queues < 1)
4690 0 : return;
4691 0 : if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
4692 : return;
4693 :
4694 0 : list_for_each_entry(q, &set->tag_list, tag_set_list)
4695 0 : blk_mq_freeze_queue(q);
4696 : /*
4697 : * Switch IO scheduler to 'none', cleaning up the data associated
4698 : * with the previous scheduler. We will switch back once we are done
4699 : * updating the new sw to hw queue mappings.
4700 : */
4701 0 : list_for_each_entry(q, &set->tag_list, tag_set_list)
4702 0 : if (!blk_mq_elv_switch_none(&head, q))
4703 : goto switch_back;
4704 :
4705 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
4706 0 : blk_mq_debugfs_unregister_hctxs(q);
4707 0 : blk_mq_sysfs_unregister_hctxs(q);
4708 : }
4709 :
4710 0 : prev_nr_hw_queues = set->nr_hw_queues;
4711 0 : if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0)
4712 : goto reregister;
4713 :
4714 : fallback:
4715 0 : blk_mq_update_queue_map(set);
4716 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
4717 0 : blk_mq_realloc_hw_ctxs(set, q);
4718 0 : blk_mq_update_poll_flag(q);
4719 0 : if (q->nr_hw_queues != set->nr_hw_queues) {
4720 0 : int i = prev_nr_hw_queues;
4721 :
4722 0 : pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
4723 : nr_hw_queues, prev_nr_hw_queues);
4724 0 : for (; i < set->nr_hw_queues; i++)
4725 0 : __blk_mq_free_map_and_rqs(set, i);
4726 :
4727 0 : set->nr_hw_queues = prev_nr_hw_queues;
4728 0 : blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
4729 0 : goto fallback;
4730 : }
4731 0 : blk_mq_map_swqueue(q);
4732 : }
4733 :
4734 : reregister:
4735 0 : list_for_each_entry(q, &set->tag_list, tag_set_list) {
4736 0 : blk_mq_sysfs_register_hctxs(q);
4737 0 : blk_mq_debugfs_register_hctxs(q);
4738 : }
4739 :
4740 : switch_back:
4741 0 : list_for_each_entry(q, &set->tag_list, tag_set_list)
4742 0 : blk_mq_elv_switch_back(&head, q);
4743 :
4744 0 : list_for_each_entry(q, &set->tag_list, tag_set_list)
4745 0 : blk_mq_unfreeze_queue(q);
4746 : }
4747 :
4748 0 : void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
4749 : {
4750 0 : mutex_lock(&set->tag_list_lock);
4751 0 : __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
4752 0 : mutex_unlock(&set->tag_list_lock);
4753 0 : }
4754 : EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
4755 :
4756 0 : static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
4757 : struct io_comp_batch *iob, unsigned int flags)
4758 : {
4759 0 : long state = get_current_state();
4760 : int ret;
4761 :
4762 : do {
4763 0 : ret = q->mq_ops->poll(hctx, iob);
4764 0 : if (ret > 0) {
4765 0 : __set_current_state(TASK_RUNNING);
4766 : return ret;
4767 : }
4768 :
4769 0 : if (signal_pending_state(state, current))
4770 0 : __set_current_state(TASK_RUNNING);
4771 0 : if (task_is_running(current))
4772 : return 1;
4773 :
4774 0 : if (ret < 0 || (flags & BLK_POLL_ONESHOT))
4775 : break;
4776 : cpu_relax();
4777 0 : } while (!need_resched());
4778 :
4779 0 : __set_current_state(TASK_RUNNING);
4780 : return 0;
4781 : }
4782 :
4783 0 : int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
4784 : struct io_comp_batch *iob, unsigned int flags)
4785 : {
4786 0 : struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie);
4787 :
4788 0 : return blk_hctx_poll(q, hctx, iob, flags);
4789 : }
4790 :
4791 0 : int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
4792 : unsigned int poll_flags)
4793 : {
4794 0 : struct request_queue *q = rq->q;
4795 : int ret;
4796 :
4797 0 : if (!blk_rq_is_poll(rq))
4798 : return 0;
4799 0 : if (!percpu_ref_tryget(&q->q_usage_counter))
4800 : return 0;
4801 :
4802 0 : ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags);
4803 0 : blk_queue_exit(q);
4804 :
4805 0 : return ret;
4806 : }
4807 : EXPORT_SYMBOL_GPL(blk_rq_poll);
4808 :
4809 0 : unsigned int blk_mq_rq_cpu(struct request *rq)
4810 : {
4811 0 : return rq->mq_ctx->cpu;
4812 : }
4813 : EXPORT_SYMBOL(blk_mq_rq_cpu);
4814 :
4815 0 : void blk_mq_cancel_work_sync(struct request_queue *q)
4816 : {
4817 : struct blk_mq_hw_ctx *hctx;
4818 : unsigned long i;
4819 :
4820 0 : cancel_delayed_work_sync(&q->requeue_work);
4821 :
4822 0 : queue_for_each_hw_ctx(q, hctx, i)
4823 0 : cancel_delayed_work_sync(&hctx->run_work);
4824 0 : }
4825 :
4826 1 : static int __init blk_mq_init(void)
4827 : {
4828 : int i;
4829 :
4830 2 : for_each_possible_cpu(i)
4831 2 : init_llist_head(&per_cpu(blk_cpu_done, i));
4832 1 : open_softirq(BLOCK_SOFTIRQ, blk_done_softirq);
4833 :
4834 1 : cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD,
4835 : "block/softirq:dead", NULL,
4836 : blk_softirq_cpu_dead);
4837 1 : cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
4838 : blk_mq_hctx_notify_dead);
4839 1 : cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online",
4840 : blk_mq_hctx_notify_online,
4841 : blk_mq_hctx_notify_offline);
4842 1 : return 0;
4843 : }
4844 : subsys_initcall(blk_mq_init);
|
