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