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