FILE
blk-mq.c
File preview with syntax highlighting.
linux-6.19.14/block/blk-mq.c
1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Block multiqueue core code 4 * 5 * Copyright (C) 2013-2014 Jens Axboe 6 * Copyright (C) 2013-2014 Christoph Hellwig 7 */ 8#include <linux/kernel.h> 9#include <linux/module.h> 10#include <linux/backing-dev.h> 11#include <linux/bio.h> 12#include <linux/blkdev.h> 13#include <linux/blk-integrity.h> 14#include <linux/kmemleak.h> 15#include <linux/mm.h> 16#include <linux/init.h> 17#include <linux/slab.h> 18#include <linux/workqueue.h> 19#include <linux/smp.h> 20#include <linux/interrupt.h> 21#include <linux/llist.h> 22#include <linux/cpu.h> 23#include <linux/cache.h> 24#include <linux/sched/topology.h> 25#include <linux/sched/signal.h> 26#include <linux/suspend.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#include <linux/sched/isolation.h> 33 34#include <trace/events/block.h> 35 36#include <linux/t10-pi.h> 37#include "blk.h" 38#include "blk-mq.h" 39#include "blk-mq-debugfs.h" 40#include "blk-pm.h" 41#include "blk-stat.h" 42#include "blk-mq-sched.h" 43#include "blk-rq-qos.h" 44 45static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); 46static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd); 47static DEFINE_MUTEX(blk_mq_cpuhp_lock); 48 49static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); 50static void blk_mq_request_bypass_insert(struct request *rq, 51 blk_insert_t flags); 52static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, 53 struct list_head *list); 54static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, 55 struct io_comp_batch *iob, unsigned int flags); 56 57/* 58 * Check if any of the ctx, dispatch list or elevator 59 * have pending work in this hardware queue. 60 */ 61static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) 62{ 63 return !list_empty_careful(&hctx->dispatch) || 64 sbitmap_any_bit_set(&hctx->ctx_map) || 65 blk_mq_sched_has_work(hctx); 66} 67 68/* 69 * Mark this ctx as having pending work in this hardware queue 70 */ 71static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, 72 struct blk_mq_ctx *ctx) 73{ 74 const int bit = ctx->index_hw[hctx->type]; 75 76 if (!sbitmap_test_bit(&hctx->ctx_map, bit)) 77 sbitmap_set_bit(&hctx->ctx_map, bit); 78} 79 80static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, 81 struct blk_mq_ctx *ctx) 82{ 83 const int bit = ctx->index_hw[hctx->type]; 84 85 sbitmap_clear_bit(&hctx->ctx_map, bit); 86} 87 88struct mq_inflight { 89 struct block_device *part; 90 unsigned int inflight[2]; 91}; 92 93static bool blk_mq_check_in_driver(struct request *rq, void *priv) 94{ 95 struct mq_inflight *mi = priv; 96 97 if (rq->rq_flags & RQF_IO_STAT && 98 (!bdev_is_partition(mi->part) || rq->part == mi->part) && 99 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) 100 mi->inflight[rq_data_dir(rq)]++; 101 102 return true; 103} 104 105void blk_mq_in_driver_rw(struct block_device *part, unsigned int inflight[2]) 106{ 107 struct mq_inflight mi = { .part = part }; 108 109 blk_mq_queue_tag_busy_iter(bdev_get_queue(part), blk_mq_check_in_driver, 110 &mi); 111 inflight[READ] = mi.inflight[READ]; 112 inflight[WRITE] = mi.inflight[WRITE]; 113} 114 115#ifdef CONFIG_LOCKDEP 116static bool blk_freeze_set_owner(struct request_queue *q, 117 struct task_struct *owner) 118{ 119 if (!owner) 120 return false; 121 122 if (!q->mq_freeze_depth) { 123 q->mq_freeze_owner = owner; 124 q->mq_freeze_owner_depth = 1; 125 q->mq_freeze_disk_dead = !q->disk || 126 test_bit(GD_DEAD, &q->disk->state) || 127 !blk_queue_registered(q); 128 q->mq_freeze_queue_dying = blk_queue_dying(q); 129 return true; 130 } 131 132 if (owner == q->mq_freeze_owner) 133 q->mq_freeze_owner_depth += 1; 134 return false; 135} 136 137/* verify the last unfreeze in owner context */ 138static bool blk_unfreeze_check_owner(struct request_queue *q) 139{ 140 if (q->mq_freeze_owner != current) 141 return false; 142 if (--q->mq_freeze_owner_depth == 0) { 143 q->mq_freeze_owner = NULL; 144 return true; 145 } 146 return false; 147} 148 149#else 150 151static bool blk_freeze_set_owner(struct request_queue *q, 152 struct task_struct *owner) 153{ 154 return false; 155} 156 157static bool blk_unfreeze_check_owner(struct request_queue *q) 158{ 159 return false; 160} 161#endif 162 163bool __blk_freeze_queue_start(struct request_queue *q, 164 struct task_struct *owner) 165{ 166 bool freeze; 167 168 mutex_lock(&q->mq_freeze_lock); 169 freeze = blk_freeze_set_owner(q, owner); 170 if (++q->mq_freeze_depth == 1) { 171 percpu_ref_kill(&q->q_usage_counter); 172 mutex_unlock(&q->mq_freeze_lock); 173 if (queue_is_mq(q)) 174 blk_mq_run_hw_queues(q, false); 175 } else { 176 mutex_unlock(&q->mq_freeze_lock); 177 } 178 179 return freeze; 180} 181 182void blk_freeze_queue_start(struct request_queue *q) 183{ 184 if (__blk_freeze_queue_start(q, current)) 185 blk_freeze_acquire_lock(q); 186} 187EXPORT_SYMBOL_GPL(blk_freeze_queue_start); 188 189void blk_mq_freeze_queue_wait(struct request_queue *q) 190{ 191 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); 192} 193EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); 194 195int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, 196 unsigned long timeout) 197{ 198 return wait_event_timeout(q->mq_freeze_wq, 199 percpu_ref_is_zero(&q->q_usage_counter), 200 timeout); 201} 202EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); 203 204void blk_mq_freeze_queue_nomemsave(struct request_queue *q) 205{ 206 blk_freeze_queue_start(q); 207 blk_mq_freeze_queue_wait(q); 208} 209EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_nomemsave); 210 211bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) 212{ 213 bool unfreeze; 214 215 mutex_lock(&q->mq_freeze_lock); 216 if (force_atomic) 217 q->q_usage_counter.data->force_atomic = true; 218 q->mq_freeze_depth--; 219 WARN_ON_ONCE(q->mq_freeze_depth < 0); 220 if (!q->mq_freeze_depth) { 221 percpu_ref_resurrect(&q->q_usage_counter); 222 wake_up_all(&q->mq_freeze_wq); 223 } 224 unfreeze = blk_unfreeze_check_owner(q); 225 mutex_unlock(&q->mq_freeze_lock); 226 227 return unfreeze; 228} 229 230void blk_mq_unfreeze_queue_nomemrestore(struct request_queue *q) 231{ 232 if (__blk_mq_unfreeze_queue(q, false)) 233 blk_unfreeze_release_lock(q); 234} 235EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue_nomemrestore); 236 237/* 238 * non_owner variant of blk_freeze_queue_start 239 * 240 * Unlike blk_freeze_queue_start, the queue doesn't need to be unfrozen 241 * by the same task. This is fragile and should not be used if at all 242 * possible. 243 */ 244void blk_freeze_queue_start_non_owner(struct request_queue *q) 245{ 246 __blk_freeze_queue_start(q, NULL); 247} 248EXPORT_SYMBOL_GPL(blk_freeze_queue_start_non_owner); 249 250/* non_owner variant of blk_mq_unfreeze_queue */ 251void blk_mq_unfreeze_queue_non_owner(struct request_queue *q) 252{ 253 __blk_mq_unfreeze_queue(q, false); 254} 255EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue_non_owner); 256 257/* 258 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the 259 * mpt3sas driver such that this function can be removed. 260 */ 261void blk_mq_quiesce_queue_nowait(struct request_queue *q) 262{ 263 unsigned long flags; 264 265 spin_lock_irqsave(&q->queue_lock, flags); 266 if (!q->quiesce_depth++) 267 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); 268 spin_unlock_irqrestore(&q->queue_lock, flags); 269} 270EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); 271 272/** 273 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done 274 * @set: tag_set to wait on 275 * 276 * Note: it is driver's responsibility for making sure that quiesce has 277 * been started on or more of the request_queues of the tag_set. This 278 * function only waits for the quiesce on those request_queues that had 279 * the quiesce flag set using blk_mq_quiesce_queue_nowait. 280 */ 281void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) 282{ 283 if (set->flags & BLK_MQ_F_BLOCKING) 284 synchronize_srcu(set->srcu); 285 else 286 synchronize_rcu(); 287} 288EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); 289 290/** 291 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished 292 * @q: request queue. 293 * 294 * Note: this function does not prevent that the struct request end_io() 295 * callback function is invoked. Once this function is returned, we make 296 * sure no dispatch can happen until the queue is unquiesced via 297 * blk_mq_unquiesce_queue(). 298 */ 299void blk_mq_quiesce_queue(struct request_queue *q) 300{ 301 blk_mq_quiesce_queue_nowait(q); 302 /* nothing to wait for non-mq queues */ 303 if (queue_is_mq(q)) 304 blk_mq_wait_quiesce_done(q->tag_set); 305} 306EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); 307 308/* 309 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() 310 * @q: request queue. 311 * 312 * This function recovers queue into the state before quiescing 313 * which is done by blk_mq_quiesce_queue. 314 */ 315void blk_mq_unquiesce_queue(struct request_queue *q) 316{ 317 unsigned long flags; 318 bool run_queue = false; 319 320 spin_lock_irqsave(&q->queue_lock, flags); 321 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { 322 ; 323 } else if (!--q->quiesce_depth) { 324 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); 325 run_queue = true; 326 } 327 spin_unlock_irqrestore(&q->queue_lock, flags); 328 329 /* dispatch requests which are inserted during quiescing */ 330 if (run_queue) 331 blk_mq_run_hw_queues(q, true); 332} 333EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); 334 335void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) 336{ 337 struct request_queue *q; 338 339 rcu_read_lock(); 340 list_for_each_entry_rcu(q, &set->tag_list, tag_set_list) { 341 if (!blk_queue_skip_tagset_quiesce(q)) 342 blk_mq_quiesce_queue_nowait(q); 343 } 344 rcu_read_unlock(); 345 346 blk_mq_wait_quiesce_done(set); 347} 348EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); 349 350void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) 351{ 352 struct request_queue *q; 353 354 rcu_read_lock(); 355 list_for_each_entry_rcu(q, &set->tag_list, tag_set_list) { 356 if (!blk_queue_skip_tagset_quiesce(q)) 357 blk_mq_unquiesce_queue(q); 358 } 359 rcu_read_unlock(); 360} 361EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); 362 363void blk_mq_wake_waiters(struct request_queue *q) 364{ 365 struct blk_mq_hw_ctx *hctx; 366 unsigned long i; 367 368 queue_for_each_hw_ctx(q, hctx, i) 369 if (blk_mq_hw_queue_mapped(hctx)) 370 blk_mq_tag_wakeup_all(hctx->tags, true); 371} 372 373void blk_rq_init(struct request_queue *q, struct request *rq) 374{ 375 memset(rq, 0, sizeof(*rq)); 376 377 INIT_LIST_HEAD(&rq->queuelist); 378 rq->q = q; 379 rq->__sector = (sector_t) -1; 380 rq->phys_gap_bit = 0; 381 INIT_HLIST_NODE(&rq->hash); 382 RB_CLEAR_NODE(&rq->rb_node); 383 rq->tag = BLK_MQ_NO_TAG; 384 rq->internal_tag = BLK_MQ_NO_TAG; 385 rq->start_time_ns = blk_time_get_ns(); 386 blk_crypto_rq_set_defaults(rq); 387} 388EXPORT_SYMBOL(blk_rq_init); 389 390/* Set start and alloc time when the allocated request is actually used */ 391static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns) 392{ 393#ifdef CONFIG_BLK_RQ_ALLOC_TIME 394 if (blk_queue_rq_alloc_time(rq->q)) 395 rq->alloc_time_ns = alloc_time_ns; 396 else 397 rq->alloc_time_ns = 0; 398#endif 399} 400 401static inline void blk_mq_bio_issue_init(struct request_queue *q, 402 struct bio *bio) 403{ 404#ifdef CONFIG_BLK_CGROUP 405 if (test_bit(QUEUE_FLAG_BIO_ISSUE_TIME, &q->queue_flags)) 406 bio->issue_time_ns = blk_time_get_ns(); 407#endif 408} 409 410static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, 411 struct blk_mq_tags *tags, unsigned int tag) 412{ 413 struct blk_mq_ctx *ctx = data->ctx; 414 struct blk_mq_hw_ctx *hctx = data->hctx; 415 struct request_queue *q = data->q; 416 struct request *rq = tags->static_rqs[tag]; 417 418 rq->q = q; 419 rq->mq_ctx = ctx; 420 rq->mq_hctx = hctx; 421 rq->cmd_flags = data->cmd_flags; 422 423 if (data->flags & BLK_MQ_REQ_PM) 424 data->rq_flags |= RQF_PM; 425 rq->rq_flags = data->rq_flags; 426 427 if (data->rq_flags & RQF_SCHED_TAGS) { 428 rq->tag = BLK_MQ_NO_TAG; 429 rq->internal_tag = tag; 430 } else { 431 rq->tag = tag; 432 rq->internal_tag = BLK_MQ_NO_TAG; 433 } 434 rq->timeout = 0; 435 436 rq->part = NULL; 437 rq->io_start_time_ns = 0; 438 rq->stats_sectors = 0; 439 rq->nr_phys_segments = 0; 440 rq->nr_integrity_segments = 0; 441 rq->end_io = NULL; 442 rq->end_io_data = NULL; 443 444 blk_crypto_rq_set_defaults(rq); 445 INIT_LIST_HEAD(&rq->queuelist); 446 /* tag was already set */ 447 WRITE_ONCE(rq->deadline, 0); 448 req_ref_set(rq, 1); 449 450 if (rq->rq_flags & RQF_USE_SCHED) { 451 struct elevator_queue *e = data->q->elevator; 452 453 INIT_HLIST_NODE(&rq->hash); 454 RB_CLEAR_NODE(&rq->rb_node); 455 456 if (e->type->ops.prepare_request) 457 e->type->ops.prepare_request(rq); 458 } 459 460 return rq; 461} 462 463static inline struct request * 464__blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data) 465{ 466 unsigned int tag, tag_offset; 467 struct blk_mq_tags *tags; 468 struct request *rq; 469 unsigned long tag_mask; 470 int i, nr = 0; 471 472 do { 473 tag_mask = blk_mq_get_tags(data, data->nr_tags - nr, &tag_offset); 474 if (unlikely(!tag_mask)) { 475 if (nr == 0) 476 return NULL; 477 break; 478 } 479 tags = blk_mq_tags_from_data(data); 480 for (i = 0; tag_mask; i++) { 481 if (!(tag_mask & (1UL << i))) 482 continue; 483 tag = tag_offset + i; 484 prefetch(tags->static_rqs[tag]); 485 tag_mask &= ~(1UL << i); 486 rq = blk_mq_rq_ctx_init(data, tags, tag); 487 rq_list_add_head(data->cached_rqs, rq); 488 nr++; 489 } 490 } while (data->nr_tags > nr); 491 492 if (!(data->rq_flags & RQF_SCHED_TAGS)) 493 blk_mq_add_active_requests(data->hctx, nr); 494 /* caller already holds a reference, add for remainder */ 495 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); 496 data->nr_tags -= nr; 497 498 return rq_list_pop(data->cached_rqs); 499} 500 501static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) 502{ 503 struct request_queue *q = data->q; 504 u64 alloc_time_ns = 0; 505 struct request *rq; 506 unsigned int tag; 507 508 /* alloc_time includes depth and tag waits */ 509 if (blk_queue_rq_alloc_time(q)) 510 alloc_time_ns = blk_time_get_ns(); 511 512 if (data->cmd_flags & REQ_NOWAIT) 513 data->flags |= BLK_MQ_REQ_NOWAIT; 514 515retry: 516 data->ctx = blk_mq_get_ctx(q); 517 data->hctx = blk_mq_map_queue(data->cmd_flags, data->ctx); 518 519 if (q->elevator) { 520 /* 521 * All requests use scheduler tags when an I/O scheduler is 522 * enabled for the queue. 523 */ 524 data->rq_flags |= RQF_SCHED_TAGS; 525 526 /* 527 * Flush/passthrough requests are special and go directly to the 528 * dispatch list. 529 */ 530 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && 531 !blk_op_is_passthrough(data->cmd_flags)) { 532 struct elevator_mq_ops *ops = &q->elevator->type->ops; 533 534 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); 535 536 data->rq_flags |= RQF_USE_SCHED; 537 if (ops->limit_depth) 538 ops->limit_depth(data->cmd_flags, data); 539 } 540 } else { 541 blk_mq_tag_busy(data->hctx); 542 } 543 544 if (data->flags & BLK_MQ_REQ_RESERVED) 545 data->rq_flags |= RQF_RESV; 546 547 /* 548 * Try batched alloc if we want more than 1 tag. 549 */ 550 if (data->nr_tags > 1) { 551 rq = __blk_mq_alloc_requests_batch(data); 552 if (rq) { 553 blk_mq_rq_time_init(rq, alloc_time_ns); 554 return rq; 555 } 556 data->nr_tags = 1; 557 } 558 559 /* 560 * Waiting allocations only fail because of an inactive hctx. In that 561 * case just retry the hctx assignment and tag allocation as CPU hotplug 562 * should have migrated us to an online CPU by now. 563 */ 564 tag = blk_mq_get_tag(data); 565 if (tag == BLK_MQ_NO_TAG) { 566 if (data->flags & BLK_MQ_REQ_NOWAIT) 567 return NULL; 568 /* 569 * Give up the CPU and sleep for a random short time to 570 * ensure that thread using a realtime scheduling class 571 * are migrated off the CPU, and thus off the hctx that 572 * is going away. 573 */ 574 msleep(3); 575 goto retry; 576 } 577 578 if (!(data->rq_flags & RQF_SCHED_TAGS)) 579 blk_mq_inc_active_requests(data->hctx); 580 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag); 581 blk_mq_rq_time_init(rq, alloc_time_ns); 582 return rq; 583} 584 585static struct request *blk_mq_rq_cache_fill(struct request_queue *q, 586 struct blk_plug *plug, 587 blk_opf_t opf, 588 blk_mq_req_flags_t flags) 589{ 590 struct blk_mq_alloc_data data = { 591 .q = q, 592 .flags = flags, 593 .shallow_depth = 0, 594 .cmd_flags = opf, 595 .rq_flags = 0, 596 .nr_tags = plug->nr_ios, 597 .cached_rqs = &plug->cached_rqs, 598 .ctx = NULL, 599 .hctx = NULL 600 }; 601 struct request *rq; 602 603 if (blk_queue_enter(q, flags)) 604 return NULL; 605 606 plug->nr_ios = 1; 607 608 rq = __blk_mq_alloc_requests(&data); 609 if (unlikely(!rq)) 610 blk_queue_exit(q); 611 return rq; 612} 613 614static struct request *blk_mq_alloc_cached_request(struct request_queue *q, 615 blk_opf_t opf, 616 blk_mq_req_flags_t flags) 617{ 618 struct blk_plug *plug = current->plug; 619 struct request *rq; 620 621 if (!plug) 622 return NULL; 623 624 if (rq_list_empty(&plug->cached_rqs)) { 625 if (plug->nr_ios == 1) 626 return NULL; 627 rq = blk_mq_rq_cache_fill(q, plug, opf, flags); 628 if (!rq) 629 return NULL; 630 } else { 631 rq = rq_list_peek(&plug->cached_rqs); 632 if (!rq || rq->q != q) 633 return NULL; 634 635 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) 636 return NULL; 637 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) 638 return NULL; 639 640 rq_list_pop(&plug->cached_rqs); 641 blk_mq_rq_time_init(rq, blk_time_get_ns()); 642 } 643 644 rq->cmd_flags = opf; 645 INIT_LIST_HEAD(&rq->queuelist); 646 return rq; 647} 648 649struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, 650 blk_mq_req_flags_t flags) 651{ 652 struct request *rq; 653 654 rq = blk_mq_alloc_cached_request(q, opf, flags); 655 if (!rq) { 656 struct blk_mq_alloc_data data = { 657 .q = q, 658 .flags = flags, 659 .shallow_depth = 0, 660 .cmd_flags = opf, 661 .rq_flags = 0, 662 .nr_tags = 1, 663 .cached_rqs = NULL, 664 .ctx = NULL, 665 .hctx = NULL 666 }; 667 int ret; 668 669 ret = blk_queue_enter(q, flags); 670 if (ret) 671 return ERR_PTR(ret); 672 673 rq = __blk_mq_alloc_requests(&data); 674 if (!rq) 675 goto out_queue_exit; 676 } 677 rq->__data_len = 0; 678 rq->phys_gap_bit = 0; 679 rq->__sector = (sector_t) -1; 680 rq->bio = rq->biotail = NULL; 681 return rq; 682out_queue_exit: 683 blk_queue_exit(q); 684 return ERR_PTR(-EWOULDBLOCK); 685} 686EXPORT_SYMBOL(blk_mq_alloc_request); 687 688struct request *blk_mq_alloc_request_hctx(struct request_queue *q, 689 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) 690{ 691 struct blk_mq_alloc_data data = { 692 .q = q, 693 .flags = flags, 694 .shallow_depth = 0, 695 .cmd_flags = opf, 696 .rq_flags = 0, 697 .nr_tags = 1, 698 .cached_rqs = NULL, 699 .ctx = NULL, 700 .hctx = NULL 701 }; 702 u64 alloc_time_ns = 0; 703 struct request *rq; 704 unsigned int cpu; 705 unsigned int tag; 706 int ret; 707 708 /* alloc_time includes depth and tag waits */ 709 if (blk_queue_rq_alloc_time(q)) 710 alloc_time_ns = blk_time_get_ns(); 711 712 /* 713 * If the tag allocator sleeps we could get an allocation for a 714 * different hardware context. No need to complicate the low level 715 * allocator for this for the rare use case of a command tied to 716 * a specific queue. 717 */ 718 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || 719 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) 720 return ERR_PTR(-EINVAL); 721 722 if (hctx_idx >= q->nr_hw_queues) 723 return ERR_PTR(-EIO); 724 725 ret = blk_queue_enter(q, flags); 726 if (ret) 727 return ERR_PTR(ret); 728 729 /* 730 * Check if the hardware context is actually mapped to anything. 731 * If not tell the caller that it should skip this queue. 732 */ 733 ret = -EXDEV; 734 data.hctx = q->queue_hw_ctx[hctx_idx]; 735 if (!blk_mq_hw_queue_mapped(data.hctx)) 736 goto out_queue_exit; 737 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); 738 if (cpu >= nr_cpu_ids) 739 goto out_queue_exit; 740 data.ctx = __blk_mq_get_ctx(q, cpu); 741 742 if (q->elevator) 743 data.rq_flags |= RQF_SCHED_TAGS; 744 else 745 blk_mq_tag_busy(data.hctx); 746 747 if (flags & BLK_MQ_REQ_RESERVED) 748 data.rq_flags |= RQF_RESV; 749 750 ret = -EWOULDBLOCK; 751 tag = blk_mq_get_tag(&data); 752 if (tag == BLK_MQ_NO_TAG) 753 goto out_queue_exit; 754 if (!(data.rq_flags & RQF_SCHED_TAGS)) 755 blk_mq_inc_active_requests(data.hctx); 756 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag); 757 blk_mq_rq_time_init(rq, alloc_time_ns); 758 rq->__data_len = 0; 759 rq->phys_gap_bit = 0; 760 rq->__sector = (sector_t) -1; 761 rq->bio = rq->biotail = NULL; 762 return rq; 763 764out_queue_exit: 765 blk_queue_exit(q); 766 return ERR_PTR(ret); 767} 768EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); 769 770static void blk_mq_finish_request(struct request *rq) 771{ 772 struct request_queue *q = rq->q; 773 774 blk_zone_finish_request(rq); 775 776 if (rq->rq_flags & RQF_USE_SCHED) { 777 q->elevator->type->ops.finish_request(rq); 778 /* 779 * For postflush request that may need to be 780 * completed twice, we should clear this flag 781 * to avoid double finish_request() on the rq. 782 */ 783 rq->rq_flags &= ~RQF_USE_SCHED; 784 } 785} 786 787static void __blk_mq_free_request(struct request *rq) 788{ 789 struct request_queue *q = rq->q; 790 struct blk_mq_ctx *ctx = rq->mq_ctx; 791 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 792 const int sched_tag = rq->internal_tag; 793 794 blk_crypto_free_request(rq); 795 blk_pm_mark_last_busy(rq); 796 rq->mq_hctx = NULL; 797 798 if (rq->tag != BLK_MQ_NO_TAG) { 799 blk_mq_dec_active_requests(hctx); 800 blk_mq_put_tag(hctx->tags, ctx, rq->tag); 801 } 802 if (sched_tag != BLK_MQ_NO_TAG) 803 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); 804 blk_mq_sched_restart(hctx); 805 blk_queue_exit(q); 806} 807 808void blk_mq_free_request(struct request *rq) 809{ 810 struct request_queue *q = rq->q; 811 812 blk_mq_finish_request(rq); 813 814 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) 815 laptop_io_completion(q->disk->bdi); 816 817 rq_qos_done(q, rq); 818 819 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 820 if (req_ref_put_and_test(rq)) 821 __blk_mq_free_request(rq); 822} 823EXPORT_SYMBOL_GPL(blk_mq_free_request); 824 825void blk_mq_free_plug_rqs(struct blk_plug *plug) 826{ 827 struct request *rq; 828 829 while ((rq = rq_list_pop(&plug->cached_rqs)) != NULL) 830 blk_mq_free_request(rq); 831} 832 833void blk_dump_rq_flags(struct request *rq, char *msg) 834{ 835 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, 836 rq->q->disk ? rq->q->disk->disk_name : "?", 837 (__force unsigned long long) rq->cmd_flags); 838 839 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", 840 (unsigned long long)blk_rq_pos(rq), 841 blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); 842 printk(KERN_INFO " bio %p, biotail %p, len %u\n", 843 rq->bio, rq->biotail, blk_rq_bytes(rq)); 844} 845EXPORT_SYMBOL(blk_dump_rq_flags); 846 847static void blk_account_io_completion(struct request *req, unsigned int bytes) 848{ 849 if (req->rq_flags & RQF_IO_STAT) { 850 const int sgrp = op_stat_group(req_op(req)); 851 852 part_stat_lock(); 853 part_stat_add(req->part, sectors[sgrp], bytes >> 9); 854 part_stat_unlock(); 855 } 856} 857 858static void blk_print_req_error(struct request *req, blk_status_t status) 859{ 860 printk_ratelimited(KERN_ERR 861 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " 862 "phys_seg %u prio class %u\n", 863 blk_status_to_str(status), 864 req->q->disk ? req->q->disk->disk_name : "?", 865 blk_rq_pos(req), (__force u32)req_op(req), 866 blk_op_str(req_op(req)), 867 (__force u32)(req->cmd_flags & ~REQ_OP_MASK), 868 req->nr_phys_segments, 869 IOPRIO_PRIO_CLASS(req_get_ioprio(req))); 870} 871 872/* 873 * Fully end IO on a request. Does not support partial completions, or 874 * errors. 875 */ 876static void blk_complete_request(struct request *req) 877{ 878 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; 879 int total_bytes = blk_rq_bytes(req); 880 struct bio *bio = req->bio; 881 882 trace_block_rq_complete(req, BLK_STS_OK, total_bytes); 883 884 if (!bio) 885 return; 886 887 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) 888 blk_integrity_complete(req, total_bytes); 889 890 /* 891 * Upper layers may call blk_crypto_evict_key() anytime after the last 892 * bio_endio(). Therefore, the keyslot must be released before that. 893 */ 894 blk_crypto_rq_put_keyslot(req); 895 896 blk_account_io_completion(req, total_bytes); 897 898 do { 899 struct bio *next = bio->bi_next; 900 901 /* Completion has already been traced */ 902 bio_clear_flag(bio, BIO_TRACE_COMPLETION); 903 904 if (blk_req_bio_is_zone_append(req, bio)) 905 blk_zone_append_update_request_bio(req, bio); 906 907 if (!is_flush) 908 bio_endio(bio); 909 bio = next; 910 } while (bio); 911 912 /* 913 * Reset counters so that the request stacking driver 914 * can find how many bytes remain in the request 915 * later. 916 */ 917 if (!req->end_io) { 918 req->bio = NULL; 919 req->__data_len = 0; 920 } 921} 922 923/** 924 * blk_update_request - Complete multiple bytes without completing the request 925 * @req: the request being processed 926 * @error: block status code 927 * @nr_bytes: number of bytes to complete for @req 928 * 929 * Description: 930 * Ends I/O on a number of bytes attached to @req, but doesn't complete 931 * the request structure even if @req doesn't have leftover. 932 * If @req has leftover, sets it up for the next range of segments. 933 * 934 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees 935 * %false return from this function. 936 * 937 * Note: 938 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function 939 * except in the consistency check at the end of this function. 940 * 941 * Return: 942 * %false - this request doesn't have any more data 943 * %true - this request has more data 944 **/ 945bool blk_update_request(struct request *req, blk_status_t error, 946 unsigned int nr_bytes) 947{ 948 bool is_flush = req->rq_flags & RQF_FLUSH_SEQ; 949 bool quiet = req->rq_flags & RQF_QUIET; 950 int total_bytes; 951 952 trace_block_rq_complete(req, error, nr_bytes); 953 954 if (!req->bio) 955 return false; 956 957 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && 958 error == BLK_STS_OK) 959 blk_integrity_complete(req, nr_bytes); 960 961 /* 962 * Upper layers may call blk_crypto_evict_key() anytime after the last 963 * bio_endio(). Therefore, the keyslot must be released before that. 964 */ 965 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) 966 __blk_crypto_rq_put_keyslot(req); 967 968 if (unlikely(error && !blk_rq_is_passthrough(req) && !quiet) && 969 !test_bit(GD_DEAD, &req->q->disk->state)) { 970 blk_print_req_error(req, error); 971 trace_block_rq_error(req, error, nr_bytes); 972 } 973 974 blk_account_io_completion(req, nr_bytes); 975 976 total_bytes = 0; 977 while (req->bio) { 978 struct bio *bio = req->bio; 979 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); 980 981 if (unlikely(error)) 982 bio->bi_status = error; 983 984 if (bio_bytes == bio->bi_iter.bi_size) { 985 req->bio = bio->bi_next; 986 } else if (bio_is_zone_append(bio) && error == BLK_STS_OK) { 987 /* 988 * Partial zone append completions cannot be supported 989 * as the BIO fragments may end up not being written 990 * sequentially. 991 */ 992 bio->bi_status = BLK_STS_IOERR; 993 } 994 995 /* Completion has already been traced */ 996 bio_clear_flag(bio, BIO_TRACE_COMPLETION); 997 if (unlikely(quiet)) 998 bio_set_flag(bio, BIO_QUIET); 999 1000 bio_advance(bio, bio_bytes); 1001 1002 /* Don't actually finish bio if it's part of flush sequence */ 1003 if (!bio->bi_iter.bi_size) { 1004 if (blk_req_bio_is_zone_append(req, bio)) 1005 blk_zone_append_update_request_bio(req, bio); 1006 if (!is_flush) 1007 bio_endio(bio); 1008 } 1009 1010 total_bytes += bio_bytes; 1011 nr_bytes -= bio_bytes; 1012 1013 if (!nr_bytes) 1014 break; 1015 } 1016 1017 /* 1018 * completely done 1019 */ 1020 if (!req->bio) { 1021 /* 1022 * Reset counters so that the request stacking driver 1023 * can find how many bytes remain in the request 1024 * later. 1025 */ 1026 req->__data_len = 0; 1027 return false; 1028 } 1029 1030 req->__data_len -= total_bytes; 1031 1032 /* update sector only for requests with clear definition of sector */ 1033 if (!blk_rq_is_passthrough(req)) 1034 req->__sector += total_bytes >> 9; 1035 1036 /* mixed attributes always follow the first bio */ 1037 if (req->rq_flags & RQF_MIXED_MERGE) { 1038 req->cmd_flags &= ~REQ_FAILFAST_MASK; 1039 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; 1040 } 1041 1042 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { 1043 /* 1044 * If total number of sectors is less than the first segment 1045 * size, something has gone terribly wrong. 1046 */ 1047 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { 1048 blk_dump_rq_flags(req, "request botched"); 1049 req->__data_len = blk_rq_cur_bytes(req); 1050 } 1051 1052 /* recalculate the number of segments */ 1053 req->nr_phys_segments = blk_recalc_rq_segments(req); 1054 } 1055 1056 return true; 1057} 1058EXPORT_SYMBOL_GPL(blk_update_request); 1059 1060static inline void blk_account_io_done(struct request *req, u64 now) 1061{ 1062 trace_block_io_done(req); 1063 1064 /* 1065 * Account IO completion. flush_rq isn't accounted as a 1066 * normal IO on queueing nor completion. Accounting the 1067 * containing request is enough. 1068 */ 1069 if ((req->rq_flags & (RQF_IO_STAT|RQF_FLUSH_SEQ)) == RQF_IO_STAT) { 1070 const int sgrp = op_stat_group(req_op(req)); 1071 1072 part_stat_lock(); 1073 update_io_ticks(req->part, jiffies, true); 1074 part_stat_inc(req->part, ios[sgrp]); 1075 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); 1076 part_stat_local_dec(req->part, 1077 in_flight[op_is_write(req_op(req))]); 1078 part_stat_unlock(); 1079 } 1080} 1081 1082static inline bool blk_rq_passthrough_stats(struct request *req) 1083{ 1084 struct bio *bio = req->bio; 1085 1086 if (!blk_queue_passthrough_stat(req->q)) 1087 return false; 1088 1089 /* Requests without a bio do not transfer data. */ 1090 if (!bio) 1091 return false; 1092 1093 /* 1094 * Stats are accumulated in the bdev, so must have one attached to a 1095 * bio to track stats. Most drivers do not set the bdev for passthrough 1096 * requests, but nvme is one that will set it. 1097 */ 1098 if (!bio->bi_bdev) 1099 return false; 1100 1101 /* 1102 * We don't know what a passthrough command does, but we know the 1103 * payload size and data direction. Ensuring the size is aligned to the 1104 * block size filters out most commands with payloads that don't 1105 * represent sector access. 1106 */ 1107 if (blk_rq_bytes(req) & (bdev_logical_block_size(bio->bi_bdev) - 1)) 1108 return false; 1109 return true; 1110} 1111 1112static inline void blk_account_io_start(struct request *req) 1113{ 1114 trace_block_io_start(req); 1115 1116 if (!blk_queue_io_stat(req->q)) 1117 return; 1118 if (blk_rq_is_passthrough(req) && !blk_rq_passthrough_stats(req)) 1119 return; 1120 1121 req->rq_flags |= RQF_IO_STAT; 1122 req->start_time_ns = blk_time_get_ns(); 1123 1124 /* 1125 * All non-passthrough requests are created from a bio with one 1126 * exception: when a flush command that is part of a flush sequence 1127 * generated by the state machine in blk-flush.c is cloned onto the 1128 * lower device by dm-multipath we can get here without a bio. 1129 */ 1130 if (req->bio) 1131 req->part = req->bio->bi_bdev; 1132 else 1133 req->part = req->q->disk->part0; 1134 1135 part_stat_lock(); 1136 update_io_ticks(req->part, jiffies, false); 1137 part_stat_local_inc(req->part, in_flight[op_is_write(req_op(req))]); 1138 part_stat_unlock(); 1139} 1140 1141static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) 1142{ 1143 if (rq->rq_flags & RQF_STATS) 1144 blk_stat_add(rq, now); 1145 1146 blk_mq_sched_completed_request(rq, now); 1147 blk_account_io_done(rq, now); 1148} 1149 1150inline void __blk_mq_end_request(struct request *rq, blk_status_t error) 1151{ 1152 if (blk_mq_need_time_stamp(rq)) 1153 __blk_mq_end_request_acct(rq, blk_time_get_ns()); 1154 1155 blk_mq_finish_request(rq); 1156 1157 if (rq->end_io) { 1158 rq_qos_done(rq->q, rq); 1159 if (rq->end_io(rq, error) == RQ_END_IO_FREE) 1160 blk_mq_free_request(rq); 1161 } else { 1162 blk_mq_free_request(rq); 1163 } 1164} 1165EXPORT_SYMBOL(__blk_mq_end_request); 1166 1167void blk_mq_end_request(struct request *rq, blk_status_t error) 1168{ 1169 if (blk_update_request(rq, error, blk_rq_bytes(rq))) 1170 BUG(); 1171 __blk_mq_end_request(rq, error); 1172} 1173EXPORT_SYMBOL(blk_mq_end_request); 1174 1175#define TAG_COMP_BATCH 32 1176 1177static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, 1178 int *tag_array, int nr_tags) 1179{ 1180 struct request_queue *q = hctx->queue; 1181 1182 blk_mq_sub_active_requests(hctx, nr_tags); 1183 1184 blk_mq_put_tags(hctx->tags, tag_array, nr_tags); 1185 percpu_ref_put_many(&q->q_usage_counter, nr_tags); 1186} 1187 1188void blk_mq_end_request_batch(struct io_comp_batch *iob) 1189{ 1190 int tags[TAG_COMP_BATCH], nr_tags = 0; 1191 struct blk_mq_hw_ctx *cur_hctx = NULL; 1192 struct request *rq; 1193 u64 now = 0; 1194 1195 if (iob->need_ts) 1196 now = blk_time_get_ns(); 1197 1198 while ((rq = rq_list_pop(&iob->req_list)) != NULL) { 1199 prefetch(rq->bio); 1200 prefetch(rq->rq_next); 1201 1202 blk_complete_request(rq); 1203 if (iob->need_ts) 1204 __blk_mq_end_request_acct(rq, now); 1205 1206 blk_mq_finish_request(rq); 1207 1208 rq_qos_done(rq->q, rq); 1209 1210 /* 1211 * If end_io handler returns NONE, then it still has 1212 * ownership of the request. 1213 */ 1214 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) 1215 continue; 1216 1217 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 1218 if (!req_ref_put_and_test(rq)) 1219 continue; 1220 1221 blk_crypto_free_request(rq); 1222 blk_pm_mark_last_busy(rq); 1223 1224 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { 1225 if (cur_hctx) 1226 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); 1227 nr_tags = 0; 1228 cur_hctx = rq->mq_hctx; 1229 } 1230 tags[nr_tags++] = rq->tag; 1231 } 1232 1233 if (nr_tags) 1234 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); 1235} 1236EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); 1237 1238static void blk_complete_reqs(struct llist_head *list) 1239{ 1240 struct llist_node *entry = llist_reverse_order(llist_del_all(list)); 1241 struct request *rq, *next; 1242 1243 llist_for_each_entry_safe(rq, next, entry, ipi_list) 1244 rq->q->mq_ops->complete(rq); 1245} 1246 1247static __latent_entropy void blk_done_softirq(void) 1248{ 1249 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); 1250} 1251 1252static int blk_softirq_cpu_dead(unsigned int cpu) 1253{ 1254 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); 1255 return 0; 1256} 1257 1258static void __blk_mq_complete_request_remote(void *data) 1259{ 1260 __raise_softirq_irqoff(BLOCK_SOFTIRQ); 1261} 1262 1263static inline bool blk_mq_complete_need_ipi(struct request *rq) 1264{ 1265 int cpu = raw_smp_processor_id(); 1266 1267 if (!IS_ENABLED(CONFIG_SMP) || 1268 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) 1269 return false; 1270 /* 1271 * With force threaded interrupts enabled, raising softirq from an SMP 1272 * function call will always result in waking the ksoftirqd thread. 1273 * This is probably worse than completing the request on a different 1274 * cache domain. 1275 */ 1276 if (force_irqthreads()) 1277 return false; 1278 1279 /* same CPU or cache domain and capacity? Complete locally */ 1280 if (cpu == rq->mq_ctx->cpu || 1281 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && 1282 cpus_share_cache(cpu, rq->mq_ctx->cpu) && 1283 cpus_equal_capacity(cpu, rq->mq_ctx->cpu))) 1284 return false; 1285 1286 /* don't try to IPI to an offline CPU */ 1287 return cpu_online(rq->mq_ctx->cpu); 1288} 1289 1290static void blk_mq_complete_send_ipi(struct request *rq) 1291{ 1292 unsigned int cpu; 1293 1294 cpu = rq->mq_ctx->cpu; 1295 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) 1296 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); 1297} 1298 1299static void blk_mq_raise_softirq(struct request *rq) 1300{ 1301 struct llist_head *list; 1302 1303 preempt_disable(); 1304 list = this_cpu_ptr(&blk_cpu_done); 1305 if (llist_add(&rq->ipi_list, list)) 1306 raise_softirq(BLOCK_SOFTIRQ); 1307 preempt_enable(); 1308} 1309 1310bool blk_mq_complete_request_remote(struct request *rq) 1311{ 1312 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); 1313 1314 /* 1315 * For request which hctx has only one ctx mapping, 1316 * or a polled request, always complete locally, 1317 * it's pointless to redirect the completion. 1318 */ 1319 if ((rq->mq_hctx->nr_ctx == 1 && 1320 rq->mq_ctx->cpu == raw_smp_processor_id()) || 1321 rq->cmd_flags & REQ_POLLED) 1322 return false; 1323 1324 if (blk_mq_complete_need_ipi(rq)) { 1325 blk_mq_complete_send_ipi(rq); 1326 return true; 1327 } 1328 1329 if (rq->q->nr_hw_queues == 1) { 1330 blk_mq_raise_softirq(rq); 1331 return true; 1332 } 1333 return false; 1334} 1335EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); 1336 1337/** 1338 * blk_mq_complete_request - end I/O on a request 1339 * @rq: the request being processed 1340 * 1341 * Description: 1342 * Complete a request by scheduling the ->complete_rq operation. 1343 **/ 1344void blk_mq_complete_request(struct request *rq) 1345{ 1346 if (!blk_mq_complete_request_remote(rq)) 1347 rq->q->mq_ops->complete(rq); 1348} 1349EXPORT_SYMBOL(blk_mq_complete_request); 1350 1351/** 1352 * blk_mq_start_request - Start processing a request 1353 * @rq: Pointer to request to be started 1354 * 1355 * Function used by device drivers to notify the block layer that a request 1356 * is going to be processed now, so blk layer can do proper initializations 1357 * such as starting the timeout timer. 1358 */ 1359void blk_mq_start_request(struct request *rq) 1360{ 1361 struct request_queue *q = rq->q; 1362 1363 trace_block_rq_issue(rq); 1364 1365 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && 1366 !blk_rq_is_passthrough(rq)) { 1367 rq->io_start_time_ns = blk_time_get_ns(); 1368 rq->stats_sectors = blk_rq_sectors(rq); 1369 rq->rq_flags |= RQF_STATS; 1370 rq_qos_issue(q, rq); 1371 } 1372 1373 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); 1374 1375 blk_add_timer(rq); 1376 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); 1377 rq->mq_hctx->tags->rqs[rq->tag] = rq; 1378 1379 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) 1380 blk_integrity_prepare(rq); 1381 1382 if (rq->bio && rq->bio->bi_opf & REQ_POLLED) 1383 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); 1384} 1385EXPORT_SYMBOL(blk_mq_start_request); 1386 1387/* 1388 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple 1389 * queues. This is important for md arrays to benefit from merging 1390 * requests. 1391 */ 1392static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) 1393{ 1394 if (plug->multiple_queues) 1395 return BLK_MAX_REQUEST_COUNT * 2; 1396 return BLK_MAX_REQUEST_COUNT; 1397} 1398 1399static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) 1400{ 1401 struct request *last = rq_list_peek(&plug->mq_list); 1402 1403 if (!plug->rq_count) { 1404 trace_block_plug(rq->q); 1405 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || 1406 (!blk_queue_nomerges(rq->q) && 1407 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { 1408 blk_mq_flush_plug_list(plug, false); 1409 last = NULL; 1410 trace_block_plug(rq->q); 1411 } 1412 1413 if (!plug->multiple_queues && last && last->q != rq->q) 1414 plug->multiple_queues = true; 1415 /* 1416 * Any request allocated from sched tags can't be issued to 1417 * ->queue_rqs() directly 1418 */ 1419 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) 1420 plug->has_elevator = true; 1421 rq_list_add_tail(&plug->mq_list, rq); 1422 plug->rq_count++; 1423} 1424 1425/** 1426 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution 1427 * @rq: request to insert 1428 * @at_head: insert request at head or tail of queue 1429 * 1430 * Description: 1431 * Insert a fully prepared request at the back of the I/O scheduler queue 1432 * for execution. Don't wait for completion. 1433 * 1434 * Note: 1435 * This function will invoke @done directly if the queue is dead. 1436 */ 1437void blk_execute_rq_nowait(struct request *rq, bool at_head) 1438{ 1439 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 1440 1441 WARN_ON(irqs_disabled()); 1442 WARN_ON(!blk_rq_is_passthrough(rq)); 1443 1444 blk_account_io_start(rq); 1445 1446 if (current->plug && !at_head) { 1447 blk_add_rq_to_plug(current->plug, rq); 1448 return; 1449 } 1450 1451 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); 1452 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); 1453} 1454EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); 1455 1456struct blk_rq_wait { 1457 struct completion done; 1458 blk_status_t ret; 1459}; 1460 1461static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) 1462{ 1463 struct blk_rq_wait *wait = rq->end_io_data; 1464 1465 wait->ret = ret; 1466 complete(&wait->done); 1467 return RQ_END_IO_NONE; 1468} 1469 1470bool blk_rq_is_poll(struct request *rq) 1471{ 1472 if (!rq->mq_hctx) 1473 return false; 1474 if (rq->mq_hctx->type != HCTX_TYPE_POLL) 1475 return false; 1476 return true; 1477} 1478EXPORT_SYMBOL_GPL(blk_rq_is_poll); 1479 1480static void blk_rq_poll_completion(struct request *rq, struct completion *wait) 1481{ 1482 do { 1483 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, BLK_POLL_ONESHOT); 1484 cond_resched(); 1485 } while (!completion_done(wait)); 1486} 1487 1488/** 1489 * blk_execute_rq - insert a request into queue for execution 1490 * @rq: request to insert 1491 * @at_head: insert request at head or tail of queue 1492 * 1493 * Description: 1494 * Insert a fully prepared request at the back of the I/O scheduler queue 1495 * for execution and wait for completion. 1496 * Return: The blk_status_t result provided to blk_mq_end_request(). 1497 */ 1498blk_status_t blk_execute_rq(struct request *rq, bool at_head) 1499{ 1500 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 1501 struct blk_rq_wait wait = { 1502 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), 1503 }; 1504 1505 WARN_ON(irqs_disabled()); 1506 WARN_ON(!blk_rq_is_passthrough(rq)); 1507 1508 rq->end_io_data = &wait; 1509 rq->end_io = blk_end_sync_rq; 1510 1511 blk_account_io_start(rq); 1512 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); 1513 blk_mq_run_hw_queue(hctx, false); 1514 1515 if (blk_rq_is_poll(rq)) 1516 blk_rq_poll_completion(rq, &wait.done); 1517 else 1518 blk_wait_io(&wait.done); 1519 1520 return wait.ret; 1521} 1522EXPORT_SYMBOL(blk_execute_rq); 1523 1524static void __blk_mq_requeue_request(struct request *rq) 1525{ 1526 struct request_queue *q = rq->q; 1527 1528 blk_mq_put_driver_tag(rq); 1529 1530 trace_block_rq_requeue(rq); 1531 rq_qos_requeue(q, rq); 1532 1533 if (blk_mq_request_started(rq)) { 1534 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 1535 rq->rq_flags &= ~RQF_TIMED_OUT; 1536 } 1537} 1538 1539void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) 1540{ 1541 struct request_queue *q = rq->q; 1542 unsigned long flags; 1543 1544 __blk_mq_requeue_request(rq); 1545 1546 /* this request will be re-inserted to io scheduler queue */ 1547 blk_mq_sched_requeue_request(rq); 1548 1549 spin_lock_irqsave(&q->requeue_lock, flags); 1550 list_add_tail(&rq->queuelist, &q->requeue_list); 1551 spin_unlock_irqrestore(&q->requeue_lock, flags); 1552 1553 if (kick_requeue_list) 1554 blk_mq_kick_requeue_list(q); 1555} 1556EXPORT_SYMBOL(blk_mq_requeue_request); 1557 1558static void blk_mq_requeue_work(struct work_struct *work) 1559{ 1560 struct request_queue *q = 1561 container_of(work, struct request_queue, requeue_work.work); 1562 LIST_HEAD(rq_list); 1563 LIST_HEAD(flush_list); 1564 struct request *rq; 1565 1566 spin_lock_irq(&q->requeue_lock); 1567 list_splice_init(&q->requeue_list, &rq_list); 1568 list_splice_init(&q->flush_list, &flush_list); 1569 spin_unlock_irq(&q->requeue_lock); 1570 1571 while (!list_empty(&rq_list)) { 1572 rq = list_entry(rq_list.next, struct request, queuelist); 1573 list_del_init(&rq->queuelist); 1574 /* 1575 * If RQF_DONTPREP is set, the request has been started by the 1576 * driver already and might have driver-specific data allocated 1577 * already. Insert it into the hctx dispatch list to avoid 1578 * block layer merges for the request. 1579 */ 1580 if (rq->rq_flags & RQF_DONTPREP) 1581 blk_mq_request_bypass_insert(rq, 0); 1582 else 1583 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); 1584 } 1585 1586 while (!list_empty(&flush_list)) { 1587 rq = list_entry(flush_list.next, struct request, queuelist); 1588 list_del_init(&rq->queuelist); 1589 blk_mq_insert_request(rq, 0); 1590 } 1591 1592 blk_mq_run_hw_queues(q, false); 1593} 1594 1595void blk_mq_kick_requeue_list(struct request_queue *q) 1596{ 1597 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); 1598} 1599EXPORT_SYMBOL(blk_mq_kick_requeue_list); 1600 1601void blk_mq_delay_kick_requeue_list(struct request_queue *q, 1602 unsigned long msecs) 1603{ 1604 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 1605 msecs_to_jiffies(msecs)); 1606} 1607EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); 1608 1609static bool blk_is_flush_data_rq(struct request *rq) 1610{ 1611 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); 1612} 1613 1614static bool blk_mq_rq_inflight(struct request *rq, void *priv) 1615{ 1616 /* 1617 * If we find a request that isn't idle we know the queue is busy 1618 * as it's checked in the iter. 1619 * Return false to stop the iteration. 1620 * 1621 * In case of queue quiesce, if one flush data request is completed, 1622 * don't count it as inflight given the flush sequence is suspended, 1623 * and the original flush data request is invisible to driver, just 1624 * like other pending requests because of quiesce 1625 */ 1626 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && 1627 blk_is_flush_data_rq(rq) && 1628 blk_mq_request_completed(rq))) { 1629 bool *busy = priv; 1630 1631 *busy = true; 1632 return false; 1633 } 1634 1635 return true; 1636} 1637 1638bool blk_mq_queue_inflight(struct request_queue *q) 1639{ 1640 bool busy = false; 1641 1642 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); 1643 return busy; 1644} 1645EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); 1646 1647static void blk_mq_rq_timed_out(struct request *req) 1648{ 1649 req->rq_flags |= RQF_TIMED_OUT; 1650 if (req->q->mq_ops->timeout) { 1651 enum blk_eh_timer_return ret; 1652 1653 ret = req->q->mq_ops->timeout(req); 1654 if (ret == BLK_EH_DONE) 1655 return; 1656 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); 1657 } 1658 1659 blk_add_timer(req); 1660} 1661 1662struct blk_expired_data { 1663 bool has_timedout_rq; 1664 unsigned long next; 1665 unsigned long timeout_start; 1666}; 1667 1668static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) 1669{ 1670 unsigned long deadline; 1671 1672 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) 1673 return false; 1674 if (rq->rq_flags & RQF_TIMED_OUT) 1675 return false; 1676 1677 deadline = READ_ONCE(rq->deadline); 1678 if (time_after_eq(expired->timeout_start, deadline)) 1679 return true; 1680 1681 if (expired->next == 0) 1682 expired->next = deadline; 1683 else if (time_after(expired->next, deadline)) 1684 expired->next = deadline; 1685 return false; 1686} 1687 1688void blk_mq_put_rq_ref(struct request *rq) 1689{ 1690 if (is_flush_rq(rq)) { 1691 if (rq->end_io(rq, 0) == RQ_END_IO_FREE) 1692 blk_mq_free_request(rq); 1693 } else if (req_ref_put_and_test(rq)) { 1694 __blk_mq_free_request(rq); 1695 } 1696} 1697 1698static bool blk_mq_check_expired(struct request *rq, void *priv) 1699{ 1700 struct blk_expired_data *expired = priv; 1701 1702 /* 1703 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot 1704 * be reallocated underneath the timeout handler's processing, then 1705 * the expire check is reliable. If the request is not expired, then 1706 * it was completed and reallocated as a new request after returning 1707 * from blk_mq_check_expired(). 1708 */ 1709 if (blk_mq_req_expired(rq, expired)) { 1710 expired->has_timedout_rq = true; 1711 return false; 1712 } 1713 return true; 1714} 1715 1716static bool blk_mq_handle_expired(struct request *rq, void *priv) 1717{ 1718 struct blk_expired_data *expired = priv; 1719 1720 if (blk_mq_req_expired(rq, expired)) 1721 blk_mq_rq_timed_out(rq); 1722 return true; 1723} 1724 1725static void blk_mq_timeout_work(struct work_struct *work) 1726{ 1727 struct request_queue *q = 1728 container_of(work, struct request_queue, timeout_work); 1729 struct blk_expired_data expired = { 1730 .timeout_start = jiffies, 1731 }; 1732 struct blk_mq_hw_ctx *hctx; 1733 unsigned long i; 1734 1735 /* A deadlock might occur if a request is stuck requiring a 1736 * timeout at the same time a queue freeze is waiting 1737 * completion, since the timeout code would not be able to 1738 * acquire the queue reference here. 1739 * 1740 * That's why we don't use blk_queue_enter here; instead, we use 1741 * percpu_ref_tryget directly, because we need to be able to 1742 * obtain a reference even in the short window between the queue 1743 * starting to freeze, by dropping the first reference in 1744 * blk_freeze_queue_start, and the moment the last request is 1745 * consumed, marked by the instant q_usage_counter reaches 1746 * zero. 1747 */ 1748 if (!percpu_ref_tryget(&q->q_usage_counter)) 1749 return; 1750 1751 /* check if there is any timed-out request */ 1752 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); 1753 if (expired.has_timedout_rq) { 1754 /* 1755 * Before walking tags, we must ensure any submit started 1756 * before the current time has finished. Since the submit 1757 * uses srcu or rcu, wait for a synchronization point to 1758 * ensure all running submits have finished 1759 */ 1760 blk_mq_wait_quiesce_done(q->tag_set); 1761 1762 expired.next = 0; 1763 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); 1764 } 1765 1766 if (expired.next != 0) { 1767 mod_timer(&q->timeout, expired.next); 1768 } else { 1769 /* 1770 * Request timeouts are handled as a forward rolling timer. If 1771 * we end up here it means that no requests are pending and 1772 * also that no request has been pending for a while. Mark 1773 * each hctx as idle. 1774 */ 1775 queue_for_each_hw_ctx(q, hctx, i) { 1776 /* the hctx may be unmapped, so check it here */ 1777 if (blk_mq_hw_queue_mapped(hctx)) 1778 blk_mq_tag_idle(hctx); 1779 } 1780 } 1781 blk_queue_exit(q); 1782} 1783 1784struct flush_busy_ctx_data { 1785 struct blk_mq_hw_ctx *hctx; 1786 struct list_head *list; 1787}; 1788 1789static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) 1790{ 1791 struct flush_busy_ctx_data *flush_data = data; 1792 struct blk_mq_hw_ctx *hctx = flush_data->hctx; 1793 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; 1794 enum hctx_type type = hctx->type; 1795 1796 spin_lock(&ctx->lock); 1797 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); 1798 sbitmap_clear_bit(sb, bitnr); 1799 spin_unlock(&ctx->lock); 1800 return true; 1801} 1802 1803/* 1804 * Process software queues that have been marked busy, splicing them 1805 * to the for-dispatch 1806 */ 1807void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) 1808{ 1809 struct flush_busy_ctx_data data = { 1810 .hctx = hctx, 1811 .list = list, 1812 }; 1813 1814 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); 1815} 1816 1817struct dispatch_rq_data { 1818 struct blk_mq_hw_ctx *hctx; 1819 struct request *rq; 1820}; 1821 1822static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, 1823 void *data) 1824{ 1825 struct dispatch_rq_data *dispatch_data = data; 1826 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; 1827 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; 1828 enum hctx_type type = hctx->type; 1829 1830 spin_lock(&ctx->lock); 1831 if (!list_empty(&ctx->rq_lists[type])) { 1832 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); 1833 list_del_init(&dispatch_data->rq->queuelist); 1834 if (list_empty(&ctx->rq_lists[type])) 1835 sbitmap_clear_bit(sb, bitnr); 1836 } 1837 spin_unlock(&ctx->lock); 1838 1839 return !dispatch_data->rq; 1840} 1841 1842struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, 1843 struct blk_mq_ctx *start) 1844{ 1845 unsigned off = start ? start->index_hw[hctx->type] : 0; 1846 struct dispatch_rq_data data = { 1847 .hctx = hctx, 1848 .rq = NULL, 1849 }; 1850 1851 __sbitmap_for_each_set(&hctx->ctx_map, off, 1852 dispatch_rq_from_ctx, &data); 1853 1854 return data.rq; 1855} 1856 1857bool __blk_mq_alloc_driver_tag(struct request *rq) 1858{ 1859 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; 1860 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; 1861 int tag; 1862 1863 blk_mq_tag_busy(rq->mq_hctx); 1864 1865 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { 1866 bt = &rq->mq_hctx->tags->breserved_tags; 1867 tag_offset = 0; 1868 } else { 1869 if (!hctx_may_queue(rq->mq_hctx, bt)) 1870 return false; 1871 } 1872 1873 tag = __sbitmap_queue_get(bt); 1874 if (tag == BLK_MQ_NO_TAG) 1875 return false; 1876 1877 rq->tag = tag + tag_offset; 1878 blk_mq_inc_active_requests(rq->mq_hctx); 1879 return true; 1880} 1881 1882static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, 1883 int flags, void *key) 1884{ 1885 struct blk_mq_hw_ctx *hctx; 1886 1887 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); 1888 1889 spin_lock(&hctx->dispatch_wait_lock); 1890 if (!list_empty(&wait->entry)) { 1891 struct sbitmap_queue *sbq; 1892 1893 list_del_init(&wait->entry); 1894 sbq = &hctx->tags->bitmap_tags; 1895 atomic_dec(&sbq->ws_active); 1896 } 1897 spin_unlock(&hctx->dispatch_wait_lock); 1898 1899 blk_mq_run_hw_queue(hctx, true); 1900 return 1; 1901} 1902 1903/* 1904 * Mark us waiting for a tag. For shared tags, this involves hooking us into 1905 * the tag wakeups. For non-shared tags, we can simply mark us needing a 1906 * restart. For both cases, take care to check the condition again after 1907 * marking us as waiting. 1908 */ 1909static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, 1910 struct request *rq) 1911{ 1912 struct sbitmap_queue *sbq; 1913 struct wait_queue_head *wq; 1914 wait_queue_entry_t *wait; 1915 bool ret; 1916 1917 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && 1918 !(blk_mq_is_shared_tags(hctx->flags))) { 1919 blk_mq_sched_mark_restart_hctx(hctx); 1920 1921 /* 1922 * It's possible that a tag was freed in the window between the 1923 * allocation failure and adding the hardware queue to the wait 1924 * queue. 1925 * 1926 * Don't clear RESTART here, someone else could have set it. 1927 * At most this will cost an extra queue run. 1928 */ 1929 return blk_mq_get_driver_tag(rq); 1930 } 1931 1932 wait = &hctx->dispatch_wait; 1933 if (!list_empty_careful(&wait->entry)) 1934 return false; 1935 1936 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) 1937 sbq = &hctx->tags->breserved_tags; 1938 else 1939 sbq = &hctx->tags->bitmap_tags; 1940 wq = &bt_wait_ptr(sbq, hctx)->wait; 1941 1942 spin_lock_irq(&wq->lock); 1943 spin_lock(&hctx->dispatch_wait_lock); 1944 if (!list_empty(&wait->entry)) { 1945 spin_unlock(&hctx->dispatch_wait_lock); 1946 spin_unlock_irq(&wq->lock); 1947 return false; 1948 } 1949 1950 atomic_inc(&sbq->ws_active); 1951 wait->flags &= ~WQ_FLAG_EXCLUSIVE; 1952 __add_wait_queue(wq, wait); 1953 1954 /* 1955 * Add one explicit barrier since blk_mq_get_driver_tag() may 1956 * not imply barrier in case of failure. 1957 * 1958 * Order adding us to wait queue and allocating driver tag. 1959 * 1960 * The pair is the one implied in sbitmap_queue_wake_up() which 1961 * orders clearing sbitmap tag bits and waitqueue_active() in 1962 * __sbitmap_queue_wake_up(), since waitqueue_active() is lockless 1963 * 1964 * Otherwise, re-order of adding wait queue and getting driver tag 1965 * may cause __sbitmap_queue_wake_up() to wake up nothing because 1966 * the waitqueue_active() may not observe us in wait queue. 1967 */ 1968 smp_mb(); 1969 1970 /* 1971 * It's possible that a tag was freed in the window between the 1972 * allocation failure and adding the hardware queue to the wait 1973 * queue. 1974 */ 1975 ret = blk_mq_get_driver_tag(rq); 1976 if (!ret) { 1977 spin_unlock(&hctx->dispatch_wait_lock); 1978 spin_unlock_irq(&wq->lock); 1979 return false; 1980 } 1981 1982 /* 1983 * We got a tag, remove ourselves from the wait queue to ensure 1984 * someone else gets the wakeup. 1985 */ 1986 list_del_init(&wait->entry); 1987 atomic_dec(&sbq->ws_active); 1988 spin_unlock(&hctx->dispatch_wait_lock); 1989 spin_unlock_irq(&wq->lock); 1990 1991 return true; 1992} 1993 1994#define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 1995#define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 1996/* 1997 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): 1998 * - EWMA is one simple way to compute running average value 1999 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially 2000 * - take 4 as factor for avoiding to get too small(0) result, and this 2001 * factor doesn't matter because EWMA decreases exponentially 2002 */ 2003static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) 2004{ 2005 unsigned int ewma; 2006 2007 ewma = hctx->dispatch_busy; 2008 2009 if (!ewma && !busy) 2010 return; 2011 2012 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; 2013 if (busy) 2014 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; 2015 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; 2016 2017 hctx->dispatch_busy = ewma; 2018} 2019 2020#define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ 2021 2022static void blk_mq_handle_dev_resource(struct request *rq, 2023 struct list_head *list) 2024{ 2025 list_add(&rq->queuelist, list); 2026 __blk_mq_requeue_request(rq); 2027} 2028 2029enum prep_dispatch { 2030 PREP_DISPATCH_OK, 2031 PREP_DISPATCH_NO_TAG, 2032 PREP_DISPATCH_NO_BUDGET, 2033}; 2034 2035static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, 2036 bool need_budget) 2037{ 2038 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2039 int budget_token = -1; 2040 2041 if (need_budget) { 2042 budget_token = blk_mq_get_dispatch_budget(rq->q); 2043 if (budget_token < 0) { 2044 blk_mq_put_driver_tag(rq); 2045 return PREP_DISPATCH_NO_BUDGET; 2046 } 2047 blk_mq_set_rq_budget_token(rq, budget_token); 2048 } 2049 2050 if (!blk_mq_get_driver_tag(rq)) { 2051 /* 2052 * The initial allocation attempt failed, so we need to 2053 * rerun the hardware queue when a tag is freed. The 2054 * waitqueue takes care of that. If the queue is run 2055 * before we add this entry back on the dispatch list, 2056 * we'll re-run it below. 2057 */ 2058 if (!blk_mq_mark_tag_wait(hctx, rq)) { 2059 /* 2060 * All budgets not got from this function will be put 2061 * together during handling partial dispatch 2062 */ 2063 if (need_budget) 2064 blk_mq_put_dispatch_budget(rq->q, budget_token); 2065 return PREP_DISPATCH_NO_TAG; 2066 } 2067 } 2068 2069 return PREP_DISPATCH_OK; 2070} 2071 2072/* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ 2073static void blk_mq_release_budgets(struct request_queue *q, 2074 struct list_head *list) 2075{ 2076 struct request *rq; 2077 2078 list_for_each_entry(rq, list, queuelist) { 2079 int budget_token = blk_mq_get_rq_budget_token(rq); 2080 2081 if (budget_token >= 0) 2082 blk_mq_put_dispatch_budget(q, budget_token); 2083 } 2084} 2085 2086/* 2087 * blk_mq_commit_rqs will notify driver using bd->last that there is no 2088 * more requests. (See comment in struct blk_mq_ops for commit_rqs for 2089 * details) 2090 * Attention, we should explicitly call this in unusual cases: 2091 * 1) did not queue everything initially scheduled to queue 2092 * 2) the last attempt to queue a request failed 2093 */ 2094static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, 2095 bool from_schedule) 2096{ 2097 if (hctx->queue->mq_ops->commit_rqs && queued) { 2098 trace_block_unplug(hctx->queue, queued, !from_schedule); 2099 hctx->queue->mq_ops->commit_rqs(hctx); 2100 } 2101} 2102 2103/* 2104 * Returns true if we did some work AND can potentially do more. 2105 */ 2106bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, 2107 bool get_budget) 2108{ 2109 enum prep_dispatch prep; 2110 struct request_queue *q = hctx->queue; 2111 struct request *rq; 2112 int queued; 2113 blk_status_t ret = BLK_STS_OK; 2114 bool needs_resource = false; 2115 2116 if (list_empty(list)) 2117 return false; 2118 2119 /* 2120 * Now process all the entries, sending them to the driver. 2121 */ 2122 queued = 0; 2123 do { 2124 struct blk_mq_queue_data bd; 2125 2126 rq = list_first_entry(list, struct request, queuelist); 2127 2128 WARN_ON_ONCE(hctx != rq->mq_hctx); 2129 prep = blk_mq_prep_dispatch_rq(rq, get_budget); 2130 if (prep != PREP_DISPATCH_OK) 2131 break; 2132 2133 list_del_init(&rq->queuelist); 2134 2135 bd.rq = rq; 2136 bd.last = list_empty(list); 2137 2138 ret = q->mq_ops->queue_rq(hctx, &bd); 2139 switch (ret) { 2140 case BLK_STS_OK: 2141 queued++; 2142 break; 2143 case BLK_STS_RESOURCE: 2144 needs_resource = true; 2145 fallthrough; 2146 case BLK_STS_DEV_RESOURCE: 2147 blk_mq_handle_dev_resource(rq, list); 2148 goto out; 2149 default: 2150 blk_mq_end_request(rq, ret); 2151 } 2152 } while (!list_empty(list)); 2153out: 2154 /* If we didn't flush the entire list, we could have told the driver 2155 * there was more coming, but that turned out to be a lie. 2156 */ 2157 if (!list_empty(list) || ret != BLK_STS_OK) 2158 blk_mq_commit_rqs(hctx, queued, false); 2159 2160 /* 2161 * Any items that need requeuing? Stuff them into hctx->dispatch, 2162 * that is where we will continue on next queue run. 2163 */ 2164 if (!list_empty(list)) { 2165 bool needs_restart; 2166 /* For non-shared tags, the RESTART check will suffice */ 2167 bool no_tag = prep == PREP_DISPATCH_NO_TAG && 2168 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || 2169 blk_mq_is_shared_tags(hctx->flags)); 2170 2171 /* 2172 * If the caller allocated budgets, free the budgets of the 2173 * requests that have not yet been passed to the block driver. 2174 */ 2175 if (!get_budget) 2176 blk_mq_release_budgets(q, list); 2177 2178 spin_lock(&hctx->lock); 2179 list_splice_tail_init(list, &hctx->dispatch); 2180 spin_unlock(&hctx->lock); 2181 2182 /* 2183 * Order adding requests to hctx->dispatch and checking 2184 * SCHED_RESTART flag. The pair of this smp_mb() is the one 2185 * in blk_mq_sched_restart(). Avoid restart code path to 2186 * miss the new added requests to hctx->dispatch, meantime 2187 * SCHED_RESTART is observed here. 2188 */ 2189 smp_mb(); 2190 2191 /* 2192 * If SCHED_RESTART was set by the caller of this function and 2193 * it is no longer set that means that it was cleared by another 2194 * thread and hence that a queue rerun is needed. 2195 * 2196 * If 'no_tag' is set, that means that we failed getting 2197 * a driver tag with an I/O scheduler attached. If our dispatch 2198 * waitqueue is no longer active, ensure that we run the queue 2199 * AFTER adding our entries back to the list. 2200 * 2201 * If no I/O scheduler has been configured it is possible that 2202 * the hardware queue got stopped and restarted before requests 2203 * were pushed back onto the dispatch list. Rerun the queue to 2204 * avoid starvation. Notes: 2205 * - blk_mq_run_hw_queue() checks whether or not a queue has 2206 * been stopped before rerunning a queue. 2207 * - Some but not all block drivers stop a queue before 2208 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq 2209 * and dm-rq. 2210 * 2211 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART 2212 * bit is set, run queue after a delay to avoid IO stalls 2213 * that could otherwise occur if the queue is idle. We'll do 2214 * similar if we couldn't get budget or couldn't lock a zone 2215 * and SCHED_RESTART is set. 2216 */ 2217 needs_restart = blk_mq_sched_needs_restart(hctx); 2218 if (prep == PREP_DISPATCH_NO_BUDGET) 2219 needs_resource = true; 2220 if (!needs_restart || 2221 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) 2222 blk_mq_run_hw_queue(hctx, true); 2223 else if (needs_resource) 2224 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); 2225 2226 blk_mq_update_dispatch_busy(hctx, true); 2227 return false; 2228 } 2229 2230 blk_mq_update_dispatch_busy(hctx, false); 2231 return true; 2232} 2233 2234static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) 2235{ 2236 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); 2237 2238 if (cpu >= nr_cpu_ids) 2239 cpu = cpumask_first(hctx->cpumask); 2240 return cpu; 2241} 2242 2243/* 2244 * ->next_cpu is always calculated from hctx->cpumask, so simply use 2245 * it for speeding up the check 2246 */ 2247static bool blk_mq_hctx_empty_cpumask(struct blk_mq_hw_ctx *hctx) 2248{ 2249 return hctx->next_cpu >= nr_cpu_ids; 2250} 2251 2252/* 2253 * It'd be great if the workqueue API had a way to pass 2254 * in a mask and had some smarts for more clever placement. 2255 * For now we just round-robin here, switching for every 2256 * BLK_MQ_CPU_WORK_BATCH queued items. 2257 */ 2258static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) 2259{ 2260 bool tried = false; 2261 int next_cpu = hctx->next_cpu; 2262 2263 /* Switch to unbound if no allowable CPUs in this hctx */ 2264 if (hctx->queue->nr_hw_queues == 1 || blk_mq_hctx_empty_cpumask(hctx)) 2265 return WORK_CPU_UNBOUND; 2266 2267 if (--hctx->next_cpu_batch <= 0) { 2268select_cpu: 2269 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, 2270 cpu_online_mask); 2271 if (next_cpu >= nr_cpu_ids) 2272 next_cpu = blk_mq_first_mapped_cpu(hctx); 2273 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; 2274 } 2275 2276 /* 2277 * Do unbound schedule if we can't find a online CPU for this hctx, 2278 * and it should only happen in the path of handling CPU DEAD. 2279 */ 2280 if (!cpu_online(next_cpu)) { 2281 if (!tried) { 2282 tried = true; 2283 goto select_cpu; 2284 } 2285 2286 /* 2287 * Make sure to re-select CPU next time once after CPUs 2288 * in hctx->cpumask become online again. 2289 */ 2290 hctx->next_cpu = next_cpu; 2291 hctx->next_cpu_batch = 1; 2292 return WORK_CPU_UNBOUND; 2293 } 2294 2295 hctx->next_cpu = next_cpu; 2296 return next_cpu; 2297} 2298 2299/** 2300 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. 2301 * @hctx: Pointer to the hardware queue to run. 2302 * @msecs: Milliseconds of delay to wait before running the queue. 2303 * 2304 * Run a hardware queue asynchronously with a delay of @msecs. 2305 */ 2306void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) 2307{ 2308 if (unlikely(blk_mq_hctx_stopped(hctx))) 2309 return; 2310 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, 2311 msecs_to_jiffies(msecs)); 2312} 2313EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); 2314 2315static inline bool blk_mq_hw_queue_need_run(struct blk_mq_hw_ctx *hctx) 2316{ 2317 bool need_run; 2318 2319 /* 2320 * When queue is quiesced, we may be switching io scheduler, or 2321 * updating nr_hw_queues, or other things, and we can't run queue 2322 * any more, even blk_mq_hctx_has_pending() can't be called safely. 2323 * 2324 * And queue will be rerun in blk_mq_unquiesce_queue() if it is 2325 * quiesced. 2326 */ 2327 __blk_mq_run_dispatch_ops(hctx->queue, false, 2328 need_run = !blk_queue_quiesced(hctx->queue) && 2329 blk_mq_hctx_has_pending(hctx)); 2330 return need_run; 2331} 2332 2333/** 2334 * blk_mq_run_hw_queue - Start to run a hardware queue. 2335 * @hctx: Pointer to the hardware queue to run. 2336 * @async: If we want to run the queue asynchronously. 2337 * 2338 * Check if the request queue is not in a quiesced state and if there are 2339 * pending requests to be sent. If this is true, run the queue to send requests 2340 * to hardware. 2341 */ 2342void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) 2343{ 2344 bool need_run; 2345 2346 /* 2347 * We can't run the queue inline with interrupts disabled. 2348 */ 2349 WARN_ON_ONCE(!async && in_interrupt()); 2350 2351 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); 2352 2353 need_run = blk_mq_hw_queue_need_run(hctx); 2354 if (!need_run) { 2355 unsigned long flags; 2356 2357 /* 2358 * Synchronize with blk_mq_unquiesce_queue(), because we check 2359 * if hw queue is quiesced locklessly above, we need the use 2360 * ->queue_lock to make sure we see the up-to-date status to 2361 * not miss rerunning the hw queue. 2362 */ 2363 spin_lock_irqsave(&hctx->queue->queue_lock, flags); 2364 need_run = blk_mq_hw_queue_need_run(hctx); 2365 spin_unlock_irqrestore(&hctx->queue->queue_lock, flags); 2366 2367 if (!need_run) 2368 return; 2369 } 2370 2371 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { 2372 blk_mq_delay_run_hw_queue(hctx, 0); 2373 return; 2374 } 2375 2376 blk_mq_run_dispatch_ops(hctx->queue, 2377 blk_mq_sched_dispatch_requests(hctx)); 2378} 2379EXPORT_SYMBOL(blk_mq_run_hw_queue); 2380 2381/* 2382 * Return prefered queue to dispatch from (if any) for non-mq aware IO 2383 * scheduler. 2384 */ 2385static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) 2386{ 2387 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 2388 /* 2389 * If the IO scheduler does not respect hardware queues when 2390 * dispatching, we just don't bother with multiple HW queues and 2391 * dispatch from hctx for the current CPU since running multiple queues 2392 * just causes lock contention inside the scheduler and pointless cache 2393 * bouncing. 2394 */ 2395 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; 2396 2397 if (!blk_mq_hctx_stopped(hctx)) 2398 return hctx; 2399 return NULL; 2400} 2401 2402/** 2403 * blk_mq_run_hw_queues - Run all hardware queues in a request queue. 2404 * @q: Pointer to the request queue to run. 2405 * @async: If we want to run the queue asynchronously. 2406 */ 2407void blk_mq_run_hw_queues(struct request_queue *q, bool async) 2408{ 2409 struct blk_mq_hw_ctx *hctx, *sq_hctx; 2410 unsigned long i; 2411 2412 sq_hctx = NULL; 2413 if (blk_queue_sq_sched(q)) 2414 sq_hctx = blk_mq_get_sq_hctx(q); 2415 queue_for_each_hw_ctx(q, hctx, i) { 2416 if (blk_mq_hctx_stopped(hctx)) 2417 continue; 2418 /* 2419 * Dispatch from this hctx either if there's no hctx preferred 2420 * by IO scheduler or if it has requests that bypass the 2421 * scheduler. 2422 */ 2423 if (!sq_hctx || sq_hctx == hctx || 2424 !list_empty_careful(&hctx->dispatch)) 2425 blk_mq_run_hw_queue(hctx, async); 2426 } 2427} 2428EXPORT_SYMBOL(blk_mq_run_hw_queues); 2429 2430/** 2431 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. 2432 * @q: Pointer to the request queue to run. 2433 * @msecs: Milliseconds of delay to wait before running the queues. 2434 */ 2435void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) 2436{ 2437 struct blk_mq_hw_ctx *hctx, *sq_hctx; 2438 unsigned long i; 2439 2440 sq_hctx = NULL; 2441 if (blk_queue_sq_sched(q)) 2442 sq_hctx = blk_mq_get_sq_hctx(q); 2443 queue_for_each_hw_ctx(q, hctx, i) { 2444 if (blk_mq_hctx_stopped(hctx)) 2445 continue; 2446 /* 2447 * If there is already a run_work pending, leave the 2448 * pending delay untouched. Otherwise, a hctx can stall 2449 * if another hctx is re-delaying the other's work 2450 * before the work executes. 2451 */ 2452 if (delayed_work_pending(&hctx->run_work)) 2453 continue; 2454 /* 2455 * Dispatch from this hctx either if there's no hctx preferred 2456 * by IO scheduler or if it has requests that bypass the 2457 * scheduler. 2458 */ 2459 if (!sq_hctx || sq_hctx == hctx || 2460 !list_empty_careful(&hctx->dispatch)) 2461 blk_mq_delay_run_hw_queue(hctx, msecs); 2462 } 2463} 2464EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); 2465 2466/* 2467 * This function is often used for pausing .queue_rq() by driver when 2468 * there isn't enough resource or some conditions aren't satisfied, and 2469 * BLK_STS_RESOURCE is usually returned. 2470 * 2471 * We do not guarantee that dispatch can be drained or blocked 2472 * after blk_mq_stop_hw_queue() returns. Please use 2473 * blk_mq_quiesce_queue() for that requirement. 2474 */ 2475void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) 2476{ 2477 cancel_delayed_work(&hctx->run_work); 2478 2479 set_bit(BLK_MQ_S_STOPPED, &hctx->state); 2480} 2481EXPORT_SYMBOL(blk_mq_stop_hw_queue); 2482 2483/* 2484 * This function is often used for pausing .queue_rq() by driver when 2485 * there isn't enough resource or some conditions aren't satisfied, and 2486 * BLK_STS_RESOURCE is usually returned. 2487 * 2488 * We do not guarantee that dispatch can be drained or blocked 2489 * after blk_mq_stop_hw_queues() returns. Please use 2490 * blk_mq_quiesce_queue() for that requirement. 2491 */ 2492void blk_mq_stop_hw_queues(struct request_queue *q) 2493{ 2494 struct blk_mq_hw_ctx *hctx; 2495 unsigned long i; 2496 2497 queue_for_each_hw_ctx(q, hctx, i) 2498 blk_mq_stop_hw_queue(hctx); 2499} 2500EXPORT_SYMBOL(blk_mq_stop_hw_queues); 2501 2502void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) 2503{ 2504 clear_bit(BLK_MQ_S_STOPPED, &hctx->state); 2505 2506 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); 2507} 2508EXPORT_SYMBOL(blk_mq_start_hw_queue); 2509 2510void blk_mq_start_hw_queues(struct request_queue *q) 2511{ 2512 struct blk_mq_hw_ctx *hctx; 2513 unsigned long i; 2514 2515 queue_for_each_hw_ctx(q, hctx, i) 2516 blk_mq_start_hw_queue(hctx); 2517} 2518EXPORT_SYMBOL(blk_mq_start_hw_queues); 2519 2520void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) 2521{ 2522 if (!blk_mq_hctx_stopped(hctx)) 2523 return; 2524 2525 clear_bit(BLK_MQ_S_STOPPED, &hctx->state); 2526 /* 2527 * Pairs with the smp_mb() in blk_mq_hctx_stopped() to order the 2528 * clearing of BLK_MQ_S_STOPPED above and the checking of dispatch 2529 * list in the subsequent routine. 2530 */ 2531 smp_mb__after_atomic(); 2532 blk_mq_run_hw_queue(hctx, async); 2533} 2534EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); 2535 2536void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) 2537{ 2538 struct blk_mq_hw_ctx *hctx; 2539 unsigned long i; 2540 2541 queue_for_each_hw_ctx(q, hctx, i) 2542 blk_mq_start_stopped_hw_queue(hctx, async || 2543 (hctx->flags & BLK_MQ_F_BLOCKING)); 2544} 2545EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); 2546 2547static void blk_mq_run_work_fn(struct work_struct *work) 2548{ 2549 struct blk_mq_hw_ctx *hctx = 2550 container_of(work, struct blk_mq_hw_ctx, run_work.work); 2551 2552 blk_mq_run_dispatch_ops(hctx->queue, 2553 blk_mq_sched_dispatch_requests(hctx)); 2554} 2555 2556/** 2557 * blk_mq_request_bypass_insert - Insert a request at dispatch list. 2558 * @rq: Pointer to request to be inserted. 2559 * @flags: BLK_MQ_INSERT_* 2560 * 2561 * Should only be used carefully, when the caller knows we want to 2562 * bypass a potential IO scheduler on the target device. 2563 */ 2564static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) 2565{ 2566 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2567 2568 spin_lock(&hctx->lock); 2569 if (flags & BLK_MQ_INSERT_AT_HEAD) 2570 list_add(&rq->queuelist, &hctx->dispatch); 2571 else 2572 list_add_tail(&rq->queuelist, &hctx->dispatch); 2573 spin_unlock(&hctx->lock); 2574} 2575 2576static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, 2577 struct blk_mq_ctx *ctx, struct list_head *list, 2578 bool run_queue_async) 2579{ 2580 struct request *rq; 2581 enum hctx_type type = hctx->type; 2582 2583 /* 2584 * Try to issue requests directly if the hw queue isn't busy to save an 2585 * extra enqueue & dequeue to the sw queue. 2586 */ 2587 if (!hctx->dispatch_busy && !run_queue_async) { 2588 blk_mq_run_dispatch_ops(hctx->queue, 2589 blk_mq_try_issue_list_directly(hctx, list)); 2590 if (list_empty(list)) 2591 goto out; 2592 } 2593 2594 /* 2595 * preemption doesn't flush plug list, so it's possible ctx->cpu is 2596 * offline now 2597 */ 2598 list_for_each_entry(rq, list, queuelist) { 2599 BUG_ON(rq->mq_ctx != ctx); 2600 trace_block_rq_insert(rq); 2601 if (rq->cmd_flags & REQ_NOWAIT) 2602 run_queue_async = true; 2603 } 2604 2605 spin_lock(&ctx->lock); 2606 list_splice_tail_init(list, &ctx->rq_lists[type]); 2607 blk_mq_hctx_mark_pending(hctx, ctx); 2608 spin_unlock(&ctx->lock); 2609out: 2610 blk_mq_run_hw_queue(hctx, run_queue_async); 2611} 2612 2613static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) 2614{ 2615 struct request_queue *q = rq->q; 2616 struct blk_mq_ctx *ctx = rq->mq_ctx; 2617 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2618 2619 if (blk_rq_is_passthrough(rq)) { 2620 /* 2621 * Passthrough request have to be added to hctx->dispatch 2622 * directly. The device may be in a situation where it can't 2623 * handle FS request, and always returns BLK_STS_RESOURCE for 2624 * them, which gets them added to hctx->dispatch. 2625 * 2626 * If a passthrough request is required to unblock the queues, 2627 * and it is added to the scheduler queue, there is no chance to 2628 * dispatch it given we prioritize requests in hctx->dispatch. 2629 */ 2630 blk_mq_request_bypass_insert(rq, flags); 2631 } else if (req_op(rq) == REQ_OP_FLUSH) { 2632 /* 2633 * Firstly normal IO request is inserted to scheduler queue or 2634 * sw queue, meantime we add flush request to dispatch queue( 2635 * hctx->dispatch) directly and there is at most one in-flight 2636 * flush request for each hw queue, so it doesn't matter to add 2637 * flush request to tail or front of the dispatch queue. 2638 * 2639 * Secondly in case of NCQ, flush request belongs to non-NCQ 2640 * command, and queueing it will fail when there is any 2641 * in-flight normal IO request(NCQ command). When adding flush 2642 * rq to the front of hctx->dispatch, it is easier to introduce 2643 * extra time to flush rq's latency because of S_SCHED_RESTART 2644 * compared with adding to the tail of dispatch queue, then 2645 * chance of flush merge is increased, and less flush requests 2646 * will be issued to controller. It is observed that ~10% time 2647 * is saved in blktests block/004 on disk attached to AHCI/NCQ 2648 * drive when adding flush rq to the front of hctx->dispatch. 2649 * 2650 * Simply queue flush rq to the front of hctx->dispatch so that 2651 * intensive flush workloads can benefit in case of NCQ HW. 2652 */ 2653 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); 2654 } else if (q->elevator) { 2655 LIST_HEAD(list); 2656 2657 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); 2658 2659 list_add(&rq->queuelist, &list); 2660 q->elevator->type->ops.insert_requests(hctx, &list, flags); 2661 } else { 2662 trace_block_rq_insert(rq); 2663 2664 spin_lock(&ctx->lock); 2665 if (flags & BLK_MQ_INSERT_AT_HEAD) 2666 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); 2667 else 2668 list_add_tail(&rq->queuelist, 2669 &ctx->rq_lists[hctx->type]); 2670 blk_mq_hctx_mark_pending(hctx, ctx); 2671 spin_unlock(&ctx->lock); 2672 } 2673} 2674 2675static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, 2676 unsigned int nr_segs) 2677{ 2678 int err; 2679 2680 if (bio->bi_opf & REQ_RAHEAD) 2681 rq->cmd_flags |= REQ_FAILFAST_MASK; 2682 2683 rq->bio = rq->biotail = bio; 2684 rq->__sector = bio->bi_iter.bi_sector; 2685 rq->__data_len = bio->bi_iter.bi_size; 2686 rq->phys_gap_bit = bio->bi_bvec_gap_bit; 2687 2688 rq->nr_phys_segments = nr_segs; 2689 if (bio_integrity(bio)) 2690 rq->nr_integrity_segments = blk_rq_count_integrity_sg(rq->q, 2691 bio); 2692 2693 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ 2694 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); 2695 WARN_ON_ONCE(err); 2696 2697 blk_account_io_start(rq); 2698} 2699 2700static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, 2701 struct request *rq, bool last) 2702{ 2703 struct request_queue *q = rq->q; 2704 struct blk_mq_queue_data bd = { 2705 .rq = rq, 2706 .last = last, 2707 }; 2708 blk_status_t ret; 2709 2710 /* 2711 * For OK queue, we are done. For error, caller may kill it. 2712 * Any other error (busy), just add it to our list as we 2713 * previously would have done. 2714 */ 2715 ret = q->mq_ops->queue_rq(hctx, &bd); 2716 switch (ret) { 2717 case BLK_STS_OK: 2718 blk_mq_update_dispatch_busy(hctx, false); 2719 break; 2720 case BLK_STS_RESOURCE: 2721 case BLK_STS_DEV_RESOURCE: 2722 blk_mq_update_dispatch_busy(hctx, true); 2723 __blk_mq_requeue_request(rq); 2724 break; 2725 default: 2726 blk_mq_update_dispatch_busy(hctx, false); 2727 break; 2728 } 2729 2730 return ret; 2731} 2732 2733static bool blk_mq_get_budget_and_tag(struct request *rq) 2734{ 2735 int budget_token; 2736 2737 budget_token = blk_mq_get_dispatch_budget(rq->q); 2738 if (budget_token < 0) 2739 return false; 2740 blk_mq_set_rq_budget_token(rq, budget_token); 2741 if (!blk_mq_get_driver_tag(rq)) { 2742 blk_mq_put_dispatch_budget(rq->q, budget_token); 2743 return false; 2744 } 2745 return true; 2746} 2747 2748/** 2749 * blk_mq_try_issue_directly - Try to send a request directly to device driver. 2750 * @hctx: Pointer of the associated hardware queue. 2751 * @rq: Pointer to request to be sent. 2752 * 2753 * If the device has enough resources to accept a new request now, send the 2754 * request directly to device driver. Else, insert at hctx->dispatch queue, so 2755 * we can try send it another time in the future. Requests inserted at this 2756 * queue have higher priority. 2757 */ 2758static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, 2759 struct request *rq) 2760{ 2761 blk_status_t ret; 2762 2763 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { 2764 blk_mq_insert_request(rq, 0); 2765 blk_mq_run_hw_queue(hctx, false); 2766 return; 2767 } 2768 2769 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { 2770 blk_mq_insert_request(rq, 0); 2771 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); 2772 return; 2773 } 2774 2775 ret = __blk_mq_issue_directly(hctx, rq, true); 2776 switch (ret) { 2777 case BLK_STS_OK: 2778 break; 2779 case BLK_STS_RESOURCE: 2780 case BLK_STS_DEV_RESOURCE: 2781 blk_mq_request_bypass_insert(rq, 0); 2782 blk_mq_run_hw_queue(hctx, false); 2783 break; 2784 default: 2785 blk_mq_end_request(rq, ret); 2786 break; 2787 } 2788} 2789 2790static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) 2791{ 2792 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2793 2794 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { 2795 blk_mq_insert_request(rq, 0); 2796 blk_mq_run_hw_queue(hctx, false); 2797 return BLK_STS_OK; 2798 } 2799 2800 if (!blk_mq_get_budget_and_tag(rq)) 2801 return BLK_STS_RESOURCE; 2802 return __blk_mq_issue_directly(hctx, rq, last); 2803} 2804 2805static void blk_mq_issue_direct(struct rq_list *rqs) 2806{ 2807 struct blk_mq_hw_ctx *hctx = NULL; 2808 struct request *rq; 2809 int queued = 0; 2810 blk_status_t ret = BLK_STS_OK; 2811 2812 while ((rq = rq_list_pop(rqs))) { 2813 bool last = rq_list_empty(rqs); 2814 2815 if (hctx != rq->mq_hctx) { 2816 if (hctx) { 2817 blk_mq_commit_rqs(hctx, queued, false); 2818 queued = 0; 2819 } 2820 hctx = rq->mq_hctx; 2821 } 2822 2823 ret = blk_mq_request_issue_directly(rq, last); 2824 switch (ret) { 2825 case BLK_STS_OK: 2826 queued++; 2827 break; 2828 case BLK_STS_RESOURCE: 2829 case BLK_STS_DEV_RESOURCE: 2830 blk_mq_request_bypass_insert(rq, 0); 2831 blk_mq_run_hw_queue(hctx, false); 2832 goto out; 2833 default: 2834 blk_mq_end_request(rq, ret); 2835 break; 2836 } 2837 } 2838 2839out: 2840 if (ret != BLK_STS_OK) 2841 blk_mq_commit_rqs(hctx, queued, false); 2842} 2843 2844static void __blk_mq_flush_list(struct request_queue *q, struct rq_list *rqs) 2845{ 2846 if (blk_queue_quiesced(q)) 2847 return; 2848 q->mq_ops->queue_rqs(rqs); 2849} 2850 2851static unsigned blk_mq_extract_queue_requests(struct rq_list *rqs, 2852 struct rq_list *queue_rqs) 2853{ 2854 struct request *rq = rq_list_pop(rqs); 2855 struct request_queue *this_q = rq->q; 2856 struct request **prev = &rqs->head; 2857 struct rq_list matched_rqs = {}; 2858 struct request *last = NULL; 2859 unsigned depth = 1; 2860 2861 rq_list_add_tail(&matched_rqs, rq); 2862 while ((rq = *prev)) { 2863 if (rq->q == this_q) { 2864 /* move rq from rqs to matched_rqs */ 2865 *prev = rq->rq_next; 2866 rq_list_add_tail(&matched_rqs, rq); 2867 depth++; 2868 } else { 2869 /* leave rq in rqs */ 2870 prev = &rq->rq_next; 2871 last = rq; 2872 } 2873 } 2874 2875 rqs->tail = last; 2876 *queue_rqs = matched_rqs; 2877 return depth; 2878} 2879 2880static void blk_mq_dispatch_queue_requests(struct rq_list *rqs, unsigned depth) 2881{ 2882 struct request_queue *q = rq_list_peek(rqs)->q; 2883 2884 trace_block_unplug(q, depth, true); 2885 2886 /* 2887 * Peek first request and see if we have a ->queue_rqs() hook. 2888 * If we do, we can dispatch the whole list in one go. 2889 * We already know at this point that all requests belong to the 2890 * same queue, caller must ensure that's the case. 2891 */ 2892 if (q->mq_ops->queue_rqs) { 2893 blk_mq_run_dispatch_ops(q, __blk_mq_flush_list(q, rqs)); 2894 if (rq_list_empty(rqs)) 2895 return; 2896 } 2897 2898 blk_mq_run_dispatch_ops(q, blk_mq_issue_direct(rqs)); 2899} 2900 2901static void blk_mq_dispatch_list(struct rq_list *rqs, bool from_sched) 2902{ 2903 struct blk_mq_hw_ctx *this_hctx = NULL; 2904 struct blk_mq_ctx *this_ctx = NULL; 2905 struct rq_list requeue_list = {}; 2906 unsigned int depth = 0; 2907 bool is_passthrough = false; 2908 LIST_HEAD(list); 2909 2910 do { 2911 struct request *rq = rq_list_pop(rqs); 2912 2913 if (!this_hctx) { 2914 this_hctx = rq->mq_hctx; 2915 this_ctx = rq->mq_ctx; 2916 is_passthrough = blk_rq_is_passthrough(rq); 2917 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || 2918 is_passthrough != blk_rq_is_passthrough(rq)) { 2919 rq_list_add_tail(&requeue_list, rq); 2920 continue; 2921 } 2922 list_add_tail(&rq->queuelist, &list); 2923 depth++; 2924 } while (!rq_list_empty(rqs)); 2925 2926 *rqs = requeue_list; 2927 trace_block_unplug(this_hctx->queue, depth, !from_sched); 2928 2929 percpu_ref_get(&this_hctx->queue->q_usage_counter); 2930 /* passthrough requests should never be issued to the I/O scheduler */ 2931 if (is_passthrough) { 2932 spin_lock(&this_hctx->lock); 2933 list_splice_tail_init(&list, &this_hctx->dispatch); 2934 spin_unlock(&this_hctx->lock); 2935 blk_mq_run_hw_queue(this_hctx, from_sched); 2936 } else if (this_hctx->queue->elevator) { 2937 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, 2938 &list, 0); 2939 blk_mq_run_hw_queue(this_hctx, from_sched); 2940 } else { 2941 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); 2942 } 2943 percpu_ref_put(&this_hctx->queue->q_usage_counter); 2944} 2945 2946static void blk_mq_dispatch_multiple_queue_requests(struct rq_list *rqs) 2947{ 2948 do { 2949 struct rq_list queue_rqs; 2950 unsigned depth; 2951 2952 depth = blk_mq_extract_queue_requests(rqs, &queue_rqs); 2953 blk_mq_dispatch_queue_requests(&queue_rqs, depth); 2954 while (!rq_list_empty(&queue_rqs)) 2955 blk_mq_dispatch_list(&queue_rqs, false); 2956 } while (!rq_list_empty(rqs)); 2957} 2958 2959void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) 2960{ 2961 unsigned int depth; 2962 2963 /* 2964 * We may have been called recursively midway through handling 2965 * plug->mq_list via a schedule() in the driver's queue_rq() callback. 2966 * To avoid mq_list changing under our feet, clear rq_count early and 2967 * bail out specifically if rq_count is 0 rather than checking 2968 * whether the mq_list is empty. 2969 */ 2970 if (plug->rq_count == 0) 2971 return; 2972 depth = plug->rq_count; 2973 plug->rq_count = 0; 2974 2975 if (!plug->has_elevator && !from_schedule) { 2976 if (plug->multiple_queues) { 2977 blk_mq_dispatch_multiple_queue_requests(&plug->mq_list); 2978 return; 2979 } 2980 2981 blk_mq_dispatch_queue_requests(&plug->mq_list, depth); 2982 if (rq_list_empty(&plug->mq_list)) 2983 return; 2984 } 2985 2986 do { 2987 blk_mq_dispatch_list(&plug->mq_list, from_schedule); 2988 } while (!rq_list_empty(&plug->mq_list)); 2989} 2990 2991static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, 2992 struct list_head *list) 2993{ 2994 int queued = 0; 2995 blk_status_t ret = BLK_STS_OK; 2996 2997 while (!list_empty(list)) { 2998 struct request *rq = list_first_entry(list, struct request, 2999 queuelist); 3000 3001 list_del_init(&rq->queuelist); 3002 ret = blk_mq_request_issue_directly(rq, list_empty(list)); 3003 switch (ret) { 3004 case BLK_STS_OK: 3005 queued++; 3006 break; 3007 case BLK_STS_RESOURCE: 3008 case BLK_STS_DEV_RESOURCE: 3009 blk_mq_request_bypass_insert(rq, 0); 3010 if (list_empty(list)) 3011 blk_mq_run_hw_queue(hctx, false); 3012 goto out; 3013 default: 3014 blk_mq_end_request(rq, ret); 3015 break; 3016 } 3017 } 3018 3019out: 3020 if (ret != BLK_STS_OK) 3021 blk_mq_commit_rqs(hctx, queued, false); 3022} 3023 3024static bool blk_mq_attempt_bio_merge(struct request_queue *q, 3025 struct bio *bio, unsigned int nr_segs) 3026{ 3027 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { 3028 if (blk_attempt_plug_merge(q, bio, nr_segs)) 3029 return true; 3030 if (blk_mq_sched_bio_merge(q, bio, nr_segs)) 3031 return true; 3032 } 3033 return false; 3034} 3035 3036static struct request *blk_mq_get_new_requests(struct request_queue *q, 3037 struct blk_plug *plug, 3038 struct bio *bio) 3039{ 3040 struct blk_mq_alloc_data data = { 3041 .q = q, 3042 .flags = 0, 3043 .shallow_depth = 0, 3044 .cmd_flags = bio->bi_opf, 3045 .rq_flags = 0, 3046 .nr_tags = 1, 3047 .cached_rqs = NULL, 3048 .ctx = NULL, 3049 .hctx = NULL 3050 }; 3051 struct request *rq; 3052 3053 rq_qos_throttle(q, bio); 3054 3055 if (plug) { 3056 data.nr_tags = plug->nr_ios; 3057 plug->nr_ios = 1; 3058 data.cached_rqs = &plug->cached_rqs; 3059 } 3060 3061 rq = __blk_mq_alloc_requests(&data); 3062 if (unlikely(!rq)) 3063 rq_qos_cleanup(q, bio); 3064 return rq; 3065} 3066 3067/* 3068 * Check if there is a suitable cached request and return it. 3069 */ 3070static struct request *blk_mq_peek_cached_request(struct blk_plug *plug, 3071 struct request_queue *q, blk_opf_t opf) 3072{ 3073 enum hctx_type type = blk_mq_get_hctx_type(opf); 3074 struct request *rq; 3075 3076 if (!plug) 3077 return NULL; 3078 rq = rq_list_peek(&plug->cached_rqs); 3079 if (!rq || rq->q != q) 3080 return NULL; 3081 if (type != rq->mq_hctx->type && 3082 (type != HCTX_TYPE_READ || rq->mq_hctx->type != HCTX_TYPE_DEFAULT)) 3083 return NULL; 3084 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) 3085 return NULL; 3086 return rq; 3087} 3088 3089static void blk_mq_use_cached_rq(struct request *rq, struct blk_plug *plug, 3090 struct bio *bio) 3091{ 3092 if (rq_list_pop(&plug->cached_rqs) != rq) 3093 WARN_ON_ONCE(1); 3094 3095 /* 3096 * If any qos ->throttle() end up blocking, we will have flushed the 3097 * plug and hence killed the cached_rq list as well. Pop this entry 3098 * before we throttle. 3099 */ 3100 rq_qos_throttle(rq->q, bio); 3101 3102 blk_mq_rq_time_init(rq, blk_time_get_ns()); 3103 rq->cmd_flags = bio->bi_opf; 3104 INIT_LIST_HEAD(&rq->queuelist); 3105} 3106 3107static bool bio_unaligned(const struct bio *bio, struct request_queue *q) 3108{ 3109 unsigned int bs_mask = queue_logical_block_size(q) - 1; 3110 3111 /* .bi_sector of any zero sized bio need to be initialized */ 3112 if ((bio->bi_iter.bi_size & bs_mask) || 3113 ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & bs_mask)) 3114 return true; 3115 return false; 3116} 3117 3118/** 3119 * blk_mq_submit_bio - Create and send a request to block device. 3120 * @bio: Bio pointer. 3121 * 3122 * Builds up a request structure from @q and @bio and send to the device. The 3123 * request may not be queued directly to hardware if: 3124 * * This request can be merged with another one 3125 * * We want to place request at plug queue for possible future merging 3126 * * There is an IO scheduler active at this queue 3127 * 3128 * It will not queue the request if there is an error with the bio, or at the 3129 * request creation. 3130 */ 3131void blk_mq_submit_bio(struct bio *bio) 3132{ 3133 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 3134 struct blk_plug *plug = current->plug; 3135 const int is_sync = op_is_sync(bio->bi_opf); 3136 struct blk_mq_hw_ctx *hctx; 3137 unsigned int nr_segs; 3138 struct request *rq; 3139 blk_status_t ret; 3140 3141 /* 3142 * If the plug has a cached request for this queue, try to use it. 3143 */ 3144 rq = blk_mq_peek_cached_request(plug, q, bio->bi_opf); 3145 3146 /* 3147 * A BIO that was released from a zone write plug has already been 3148 * through the preparation in this function, already holds a reference 3149 * on the queue usage counter, and is the only write BIO in-flight for 3150 * the target zone. Go straight to preparing a request for it. 3151 */ 3152 if (bio_zone_write_plugging(bio)) { 3153 nr_segs = bio->__bi_nr_segments; 3154 if (rq) 3155 blk_queue_exit(q); 3156 goto new_request; 3157 } 3158 3159 /* 3160 * The cached request already holds a q_usage_counter reference and we 3161 * don't have to acquire a new one if we use it. 3162 */ 3163 if (!rq) { 3164 if (unlikely(bio_queue_enter(bio))) 3165 return; 3166 } 3167 3168 /* 3169 * Device reconfiguration may change logical block size or reduce the 3170 * number of poll queues, so the checks for alignment and poll support 3171 * have to be done with queue usage counter held. 3172 */ 3173 if (unlikely(bio_unaligned(bio, q))) { 3174 bio_io_error(bio); 3175 goto queue_exit; 3176 } 3177 3178 if ((bio->bi_opf & REQ_POLLED) && !blk_mq_can_poll(q)) { 3179 bio->bi_status = BLK_STS_NOTSUPP; 3180 bio_endio(bio); 3181 goto queue_exit; 3182 } 3183 3184 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); 3185 if (!bio) 3186 goto queue_exit; 3187 3188 if (!bio_integrity_prep(bio)) 3189 goto queue_exit; 3190 3191 blk_mq_bio_issue_init(q, bio); 3192 if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) 3193 goto queue_exit; 3194 3195 if (bio_needs_zone_write_plugging(bio)) { 3196 if (blk_zone_plug_bio(bio, nr_segs)) 3197 goto queue_exit; 3198 } 3199 3200new_request: 3201 if (rq) { 3202 blk_mq_use_cached_rq(rq, plug, bio); 3203 } else { 3204 rq = blk_mq_get_new_requests(q, plug, bio); 3205 if (unlikely(!rq)) { 3206 if (bio->bi_opf & REQ_NOWAIT) 3207 bio_wouldblock_error(bio); 3208 goto queue_exit; 3209 } 3210 } 3211 3212 trace_block_getrq(bio); 3213 3214 rq_qos_track(q, rq, bio); 3215 3216 blk_mq_bio_to_request(rq, bio, nr_segs); 3217 3218 ret = blk_crypto_rq_get_keyslot(rq); 3219 if (ret != BLK_STS_OK) { 3220 bio->bi_status = ret; 3221 bio_endio(bio); 3222 blk_mq_free_request(rq); 3223 return; 3224 } 3225 3226 if (bio_zone_write_plugging(bio)) 3227 blk_zone_write_plug_init_request(rq); 3228 3229 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) 3230 return; 3231 3232 if (plug) { 3233 blk_add_rq_to_plug(plug, rq); 3234 return; 3235 } 3236 3237 hctx = rq->mq_hctx; 3238 if ((rq->rq_flags & RQF_USE_SCHED) || 3239 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { 3240 blk_mq_insert_request(rq, 0); 3241 blk_mq_run_hw_queue(hctx, true); 3242 } else { 3243 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); 3244 } 3245 return; 3246 3247queue_exit: 3248 /* 3249 * Don't drop the queue reference if we were trying to use a cached 3250 * request and thus didn't acquire one. 3251 */ 3252 if (!rq) 3253 blk_queue_exit(q); 3254} 3255 3256#ifdef CONFIG_BLK_MQ_STACKING 3257/** 3258 * blk_insert_cloned_request - Helper for stacking drivers to submit a request 3259 * @rq: the request being queued 3260 */ 3261blk_status_t blk_insert_cloned_request(struct request *rq) 3262{ 3263 struct request_queue *q = rq->q; 3264 unsigned int max_sectors = blk_queue_get_max_sectors(rq); 3265 unsigned int max_segments = blk_rq_get_max_segments(rq); 3266 blk_status_t ret; 3267 3268 if (blk_rq_sectors(rq) > max_sectors) { 3269 /* 3270 * SCSI device does not have a good way to return if 3271 * Write Same/Zero is actually supported. If a device rejects 3272 * a non-read/write command (discard, write same,etc.) the 3273 * low-level device driver will set the relevant queue limit to 3274 * 0 to prevent blk-lib from issuing more of the offending 3275 * operations. Commands queued prior to the queue limit being 3276 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O 3277 * errors being propagated to upper layers. 3278 */ 3279 if (max_sectors == 0) 3280 return BLK_STS_NOTSUPP; 3281 3282 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", 3283 __func__, blk_rq_sectors(rq), max_sectors); 3284 return BLK_STS_IOERR; 3285 } 3286 3287 /* 3288 * The queue settings related to segment counting may differ from the 3289 * original queue. 3290 */ 3291 rq->nr_phys_segments = blk_recalc_rq_segments(rq); 3292 if (rq->nr_phys_segments > max_segments) { 3293 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", 3294 __func__, rq->nr_phys_segments, max_segments); 3295 return BLK_STS_IOERR; 3296 } 3297 3298 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) 3299 return BLK_STS_IOERR; 3300 3301 ret = blk_crypto_rq_get_keyslot(rq); 3302 if (ret != BLK_STS_OK) 3303 return ret; 3304 3305 blk_account_io_start(rq); 3306 3307 /* 3308 * Since we have a scheduler attached on the top device, 3309 * bypass a potential scheduler on the bottom device for 3310 * insert. 3311 */ 3312 blk_mq_run_dispatch_ops(q, 3313 ret = blk_mq_request_issue_directly(rq, true)); 3314 if (ret) 3315 blk_account_io_done(rq, blk_time_get_ns()); 3316 return ret; 3317} 3318EXPORT_SYMBOL_GPL(blk_insert_cloned_request); 3319 3320/** 3321 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request 3322 * @rq: the clone request to be cleaned up 3323 * 3324 * Description: 3325 * Free all bios in @rq for a cloned request. 3326 */ 3327void blk_rq_unprep_clone(struct request *rq) 3328{ 3329 struct bio *bio; 3330 3331 while ((bio = rq->bio) != NULL) { 3332 rq->bio = bio->bi_next; 3333 3334 bio_put(bio); 3335 } 3336} 3337EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); 3338 3339/** 3340 * blk_rq_prep_clone - Helper function to setup clone request 3341 * @rq: the request to be setup 3342 * @rq_src: original request to be cloned 3343 * @bs: bio_set that bios for clone are allocated from 3344 * @gfp_mask: memory allocation mask for bio 3345 * @bio_ctr: setup function to be called for each clone bio. 3346 * Returns %0 for success, non %0 for failure. 3347 * @data: private data to be passed to @bio_ctr 3348 * 3349 * Description: 3350 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. 3351 * Also, pages which the original bios are pointing to are not copied 3352 * and the cloned bios just point same pages. 3353 * So cloned bios must be completed before original bios, which means 3354 * the caller must complete @rq before @rq_src. 3355 */ 3356int blk_rq_prep_clone(struct request *rq, struct request *rq_src, 3357 struct bio_set *bs, gfp_t gfp_mask, 3358 int (*bio_ctr)(struct bio *, struct bio *, void *), 3359 void *data) 3360{ 3361 struct bio *bio_src; 3362 3363 if (!bs) 3364 bs = &fs_bio_set; 3365 3366 __rq_for_each_bio(bio_src, rq_src) { 3367 struct bio *bio = bio_alloc_clone(rq->q->disk->part0, bio_src, 3368 gfp_mask, bs); 3369 if (!bio) 3370 goto free_and_out; 3371 3372 if (bio_ctr && bio_ctr(bio, bio_src, data)) { 3373 bio_put(bio); 3374 goto free_and_out; 3375 } 3376 3377 if (rq->bio) { 3378 rq->biotail->bi_next = bio; 3379 rq->biotail = bio; 3380 } else { 3381 rq->bio = rq->biotail = bio; 3382 } 3383 } 3384 3385 /* Copy attributes of the original request to the clone request. */ 3386 rq->__sector = blk_rq_pos(rq_src); 3387 rq->__data_len = blk_rq_bytes(rq_src); 3388 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { 3389 rq->rq_flags |= RQF_SPECIAL_PAYLOAD; 3390 rq->special_vec = rq_src->special_vec; 3391 } 3392 rq->nr_phys_segments = rq_src->nr_phys_segments; 3393 rq->nr_integrity_segments = rq_src->nr_integrity_segments; 3394 rq->phys_gap_bit = rq_src->phys_gap_bit; 3395 3396 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) 3397 goto free_and_out; 3398 3399 return 0; 3400 3401free_and_out: 3402 blk_rq_unprep_clone(rq); 3403 3404 return -ENOMEM; 3405} 3406EXPORT_SYMBOL_GPL(blk_rq_prep_clone); 3407#endif /* CONFIG_BLK_MQ_STACKING */ 3408 3409/* 3410 * Steal bios from a request and add them to a bio list. 3411 * The request must not have been partially completed before. 3412 */ 3413void blk_steal_bios(struct bio_list *list, struct request *rq) 3414{ 3415 if (rq->bio) { 3416 if (list->tail) 3417 list->tail->bi_next = rq->bio; 3418 else 3419 list->head = rq->bio; 3420 list->tail = rq->biotail; 3421 3422 rq->bio = NULL; 3423 rq->biotail = NULL; 3424 } 3425 3426 rq->__data_len = 0; 3427} 3428EXPORT_SYMBOL_GPL(blk_steal_bios); 3429 3430static size_t order_to_size(unsigned int order) 3431{ 3432 return (size_t)PAGE_SIZE << order; 3433} 3434 3435/* called before freeing request pool in @tags */ 3436static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, 3437 struct blk_mq_tags *tags) 3438{ 3439 struct page *page; 3440 3441 /* 3442 * There is no need to clear mapping if driver tags is not initialized 3443 * or the mapping belongs to the driver tags. 3444 */ 3445 if (!drv_tags || drv_tags == tags) 3446 return; 3447 3448 list_for_each_entry(page, &tags->page_list, lru) { 3449 unsigned long start = (unsigned long)page_address(page); 3450 unsigned long end = start + order_to_size(page->private); 3451 int i; 3452 3453 for (i = 0; i < drv_tags->nr_tags; i++) { 3454 struct request *rq = drv_tags->rqs[i]; 3455 unsigned long rq_addr = (unsigned long)rq; 3456 3457 if (rq_addr >= start && rq_addr < end) { 3458 WARN_ON_ONCE(req_ref_read(rq) != 0); 3459 cmpxchg(&drv_tags->rqs[i], rq, NULL); 3460 } 3461 } 3462 } 3463} 3464 3465void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, 3466 unsigned int hctx_idx) 3467{ 3468 struct blk_mq_tags *drv_tags; 3469 3470 if (list_empty(&tags->page_list)) 3471 return; 3472 3473 if (blk_mq_is_shared_tags(set->flags)) 3474 drv_tags = set->shared_tags; 3475 else 3476 drv_tags = set->tags[hctx_idx]; 3477 3478 if (tags->static_rqs && set->ops->exit_request) { 3479 int i; 3480 3481 for (i = 0; i < tags->nr_tags; i++) { 3482 struct request *rq = tags->static_rqs[i]; 3483 3484 if (!rq) 3485 continue; 3486 set->ops->exit_request(set, rq, hctx_idx); 3487 tags->static_rqs[i] = NULL; 3488 } 3489 } 3490 3491 blk_mq_clear_rq_mapping(drv_tags, tags); 3492 /* 3493 * Free request pages in SRCU callback, which is called from 3494 * blk_mq_free_tags(). 3495 */ 3496} 3497 3498void blk_mq_free_rq_map(struct blk_mq_tag_set *set, struct blk_mq_tags *tags) 3499{ 3500 kfree(tags->rqs); 3501 tags->rqs = NULL; 3502 kfree(tags->static_rqs); 3503 tags->static_rqs = NULL; 3504 3505 blk_mq_free_tags(set, tags); 3506} 3507 3508static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, 3509 unsigned int hctx_idx) 3510{ 3511 int i; 3512 3513 for (i = 0; i < set->nr_maps; i++) { 3514 unsigned int start = set->map[i].queue_offset; 3515 unsigned int end = start + set->map[i].nr_queues; 3516 3517 if (hctx_idx >= start && hctx_idx < end) 3518 break; 3519 } 3520 3521 if (i >= set->nr_maps) 3522 i = HCTX_TYPE_DEFAULT; 3523 3524 return i; 3525} 3526 3527static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, 3528 unsigned int hctx_idx) 3529{ 3530 enum hctx_type type = hctx_idx_to_type(set, hctx_idx); 3531 3532 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); 3533} 3534 3535static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, 3536 unsigned int hctx_idx, 3537 unsigned int nr_tags, 3538 unsigned int reserved_tags) 3539{ 3540 int node = blk_mq_get_hctx_node(set, hctx_idx); 3541 struct blk_mq_tags *tags; 3542 3543 if (node == NUMA_NO_NODE) 3544 node = set->numa_node; 3545 3546 tags = blk_mq_init_tags(nr_tags, reserved_tags, set->flags, node); 3547 if (!tags) 3548 return NULL; 3549 3550 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), 3551 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, 3552 node); 3553 if (!tags->rqs) 3554 goto err_free_tags; 3555 3556 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), 3557 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, 3558 node); 3559 if (!tags->static_rqs) 3560 goto err_free_rqs; 3561 3562 return tags; 3563 3564err_free_rqs: 3565 kfree(tags->rqs); 3566err_free_tags: 3567 blk_mq_free_tags(set, tags); 3568 return NULL; 3569} 3570 3571static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, 3572 unsigned int hctx_idx, int node) 3573{ 3574 int ret; 3575 3576 if (set->ops->init_request) { 3577 ret = set->ops->init_request(set, rq, hctx_idx, node); 3578 if (ret) 3579 return ret; 3580 } 3581 3582 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 3583 return 0; 3584} 3585 3586static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, 3587 struct blk_mq_tags *tags, 3588 unsigned int hctx_idx, unsigned int depth) 3589{ 3590 unsigned int i, j, entries_per_page, max_order = 4; 3591 int node = blk_mq_get_hctx_node(set, hctx_idx); 3592 size_t rq_size, left; 3593 3594 if (node == NUMA_NO_NODE) 3595 node = set->numa_node; 3596 3597 /* 3598 * rq_size is the size of the request plus driver payload, rounded 3599 * to the cacheline size 3600 */ 3601 rq_size = round_up(sizeof(struct request) + set->cmd_size, 3602 cache_line_size()); 3603 left = rq_size * depth; 3604 3605 for (i = 0; i < depth; ) { 3606 int this_order = max_order; 3607 struct page *page; 3608 int to_do; 3609 void *p; 3610 3611 while (this_order && left < order_to_size(this_order - 1)) 3612 this_order--; 3613 3614 do { 3615 page = alloc_pages_node(node, 3616 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, 3617 this_order); 3618 if (page) 3619 break; 3620 if (!this_order--) 3621 break; 3622 if (order_to_size(this_order) < rq_size) 3623 break; 3624 } while (1); 3625 3626 if (!page) 3627 goto fail; 3628 3629 page->private = this_order; 3630 list_add_tail(&page->lru, &tags->page_list); 3631 3632 p = page_address(page); 3633 /* 3634 * Allow kmemleak to scan these pages as they contain pointers 3635 * to additional allocations like via ops->init_request(). 3636 */ 3637 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); 3638 entries_per_page = order_to_size(this_order) / rq_size; 3639 to_do = min(entries_per_page, depth - i); 3640 left -= to_do * rq_size; 3641 for (j = 0; j < to_do; j++) { 3642 struct request *rq = p; 3643 3644 tags->static_rqs[i] = rq; 3645 if (blk_mq_init_request(set, rq, hctx_idx, node)) { 3646 tags->static_rqs[i] = NULL; 3647 goto fail; 3648 } 3649 3650 p += rq_size; 3651 i++; 3652 } 3653 } 3654 return 0; 3655 3656fail: 3657 blk_mq_free_rqs(set, tags, hctx_idx); 3658 return -ENOMEM; 3659} 3660 3661struct rq_iter_data { 3662 struct blk_mq_hw_ctx *hctx; 3663 bool has_rq; 3664}; 3665 3666static bool blk_mq_has_request(struct request *rq, void *data) 3667{ 3668 struct rq_iter_data *iter_data = data; 3669 3670 if (rq->mq_hctx != iter_data->hctx) 3671 return true; 3672 iter_data->has_rq = true; 3673 return false; 3674} 3675 3676static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) 3677{ 3678 struct blk_mq_tags *tags = hctx->sched_tags ? 3679 hctx->sched_tags : hctx->tags; 3680 struct rq_iter_data data = { 3681 .hctx = hctx, 3682 }; 3683 int srcu_idx; 3684 3685 srcu_idx = srcu_read_lock(&hctx->queue->tag_set->tags_srcu); 3686 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); 3687 srcu_read_unlock(&hctx->queue->tag_set->tags_srcu, srcu_idx); 3688 3689 return data.has_rq; 3690} 3691 3692static bool blk_mq_hctx_has_online_cpu(struct blk_mq_hw_ctx *hctx, 3693 unsigned int this_cpu) 3694{ 3695 enum hctx_type type = hctx->type; 3696 int cpu; 3697 3698 /* 3699 * hctx->cpumask has to rule out isolated CPUs, but userspace still 3700 * might submit IOs on these isolated CPUs, so use the queue map to 3701 * check if all CPUs mapped to this hctx are offline 3702 */ 3703 for_each_online_cpu(cpu) { 3704 struct blk_mq_hw_ctx *h = blk_mq_map_queue_type(hctx->queue, 3705 type, cpu); 3706 3707 if (h != hctx) 3708 continue; 3709 3710 /* this hctx has at least one online CPU */ 3711 if (this_cpu != cpu) 3712 return true; 3713 } 3714 3715 return false; 3716} 3717 3718static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) 3719{ 3720 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, 3721 struct blk_mq_hw_ctx, cpuhp_online); 3722 int ret = 0; 3723 3724 if (!hctx->nr_ctx || blk_mq_hctx_has_online_cpu(hctx, cpu)) 3725 return 0; 3726 3727 /* 3728 * Prevent new request from being allocated on the current hctx. 3729 * 3730 * The smp_mb__after_atomic() Pairs with the implied barrier in 3731 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is 3732 * seen once we return from the tag allocator. 3733 */ 3734 set_bit(BLK_MQ_S_INACTIVE, &hctx->state); 3735 smp_mb__after_atomic(); 3736 3737 /* 3738 * Try to grab a reference to the queue and wait for any outstanding 3739 * requests. If we could not grab a reference the queue has been 3740 * frozen and there are no requests. 3741 */ 3742 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { 3743 while (blk_mq_hctx_has_requests(hctx)) { 3744 /* 3745 * The wakeup capable IRQ handler of block device is 3746 * not called during suspend. Skip the loop by checking 3747 * pm_wakeup_pending to prevent the deadlock and improve 3748 * suspend latency. 3749 */ 3750 if (pm_wakeup_pending()) { 3751 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); 3752 ret = -EBUSY; 3753 break; 3754 } 3755 msleep(5); 3756 } 3757 percpu_ref_put(&hctx->queue->q_usage_counter); 3758 } 3759 3760 return ret; 3761} 3762 3763/* 3764 * Check if one CPU is mapped to the specified hctx 3765 * 3766 * Isolated CPUs have been ruled out from hctx->cpumask, which is supposed 3767 * to be used for scheduling kworker only. For other usage, please call this 3768 * helper for checking if one CPU belongs to the specified hctx 3769 */ 3770static bool blk_mq_cpu_mapped_to_hctx(unsigned int cpu, 3771 const struct blk_mq_hw_ctx *hctx) 3772{ 3773 struct blk_mq_hw_ctx *mapped_hctx = blk_mq_map_queue_type(hctx->queue, 3774 hctx->type, cpu); 3775 3776 return mapped_hctx == hctx; 3777} 3778 3779static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) 3780{ 3781 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, 3782 struct blk_mq_hw_ctx, cpuhp_online); 3783 3784 if (blk_mq_cpu_mapped_to_hctx(cpu, hctx)) 3785 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); 3786 return 0; 3787} 3788 3789/* 3790 * 'cpu' is going away. splice any existing rq_list entries from this 3791 * software queue to the hw queue dispatch list, and ensure that it 3792 * gets run. 3793 */ 3794static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) 3795{ 3796 struct blk_mq_hw_ctx *hctx; 3797 struct blk_mq_ctx *ctx; 3798 LIST_HEAD(tmp); 3799 enum hctx_type type; 3800 3801 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); 3802 if (!blk_mq_cpu_mapped_to_hctx(cpu, hctx)) 3803 return 0; 3804 3805 ctx = __blk_mq_get_ctx(hctx->queue, cpu); 3806 type = hctx->type; 3807 3808 spin_lock(&ctx->lock); 3809 if (!list_empty(&ctx->rq_lists[type])) { 3810 list_splice_init(&ctx->rq_lists[type], &tmp); 3811 blk_mq_hctx_clear_pending(hctx, ctx); 3812 } 3813 spin_unlock(&ctx->lock); 3814 3815 if (list_empty(&tmp)) 3816 return 0; 3817 3818 spin_lock(&hctx->lock); 3819 list_splice_tail_init(&tmp, &hctx->dispatch); 3820 spin_unlock(&hctx->lock); 3821 3822 blk_mq_run_hw_queue(hctx, true); 3823 return 0; 3824} 3825 3826static void __blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) 3827{ 3828 lockdep_assert_held(&blk_mq_cpuhp_lock); 3829 3830 if (!(hctx->flags & BLK_MQ_F_STACKING) && 3831 !hlist_unhashed(&hctx->cpuhp_online)) { 3832 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, 3833 &hctx->cpuhp_online); 3834 INIT_HLIST_NODE(&hctx->cpuhp_online); 3835 } 3836 3837 if (!hlist_unhashed(&hctx->cpuhp_dead)) { 3838 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, 3839 &hctx->cpuhp_dead); 3840 INIT_HLIST_NODE(&hctx->cpuhp_dead); 3841 } 3842} 3843 3844static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) 3845{ 3846 mutex_lock(&blk_mq_cpuhp_lock); 3847 __blk_mq_remove_cpuhp(hctx); 3848 mutex_unlock(&blk_mq_cpuhp_lock); 3849} 3850 3851static void __blk_mq_add_cpuhp(struct blk_mq_hw_ctx *hctx) 3852{ 3853 lockdep_assert_held(&blk_mq_cpuhp_lock); 3854 3855 if (!(hctx->flags & BLK_MQ_F_STACKING) && 3856 hlist_unhashed(&hctx->cpuhp_online)) 3857 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, 3858 &hctx->cpuhp_online); 3859 3860 if (hlist_unhashed(&hctx->cpuhp_dead)) 3861 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, 3862 &hctx->cpuhp_dead); 3863} 3864 3865static void __blk_mq_remove_cpuhp_list(struct list_head *head) 3866{ 3867 struct blk_mq_hw_ctx *hctx; 3868 3869 lockdep_assert_held(&blk_mq_cpuhp_lock); 3870 3871 list_for_each_entry(hctx, head, hctx_list) 3872 __blk_mq_remove_cpuhp(hctx); 3873} 3874 3875/* 3876 * Unregister cpuhp callbacks from exited hw queues 3877 * 3878 * Safe to call if this `request_queue` is live 3879 */ 3880static void blk_mq_remove_hw_queues_cpuhp(struct request_queue *q) 3881{ 3882 LIST_HEAD(hctx_list); 3883 3884 spin_lock(&q->unused_hctx_lock); 3885 list_splice_init(&q->unused_hctx_list, &hctx_list); 3886 spin_unlock(&q->unused_hctx_lock); 3887 3888 mutex_lock(&blk_mq_cpuhp_lock); 3889 __blk_mq_remove_cpuhp_list(&hctx_list); 3890 mutex_unlock(&blk_mq_cpuhp_lock); 3891 3892 spin_lock(&q->unused_hctx_lock); 3893 list_splice(&hctx_list, &q->unused_hctx_list); 3894 spin_unlock(&q->unused_hctx_lock); 3895} 3896 3897/* 3898 * Register cpuhp callbacks from all hw queues 3899 * 3900 * Safe to call if this `request_queue` is live 3901 */ 3902static void blk_mq_add_hw_queues_cpuhp(struct request_queue *q) 3903{ 3904 struct blk_mq_hw_ctx *hctx; 3905 unsigned long i; 3906 3907 mutex_lock(&blk_mq_cpuhp_lock); 3908 queue_for_each_hw_ctx(q, hctx, i) 3909 __blk_mq_add_cpuhp(hctx); 3910 mutex_unlock(&blk_mq_cpuhp_lock); 3911} 3912 3913/* 3914 * Before freeing hw queue, clearing the flush request reference in 3915 * tags->rqs[] for avoiding potential UAF. 3916 */ 3917static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, 3918 unsigned int queue_depth, struct request *flush_rq) 3919{ 3920 int i; 3921 3922 /* The hw queue may not be mapped yet */ 3923 if (!tags) 3924 return; 3925 3926 WARN_ON_ONCE(req_ref_read(flush_rq) != 0); 3927 3928 for (i = 0; i < queue_depth; i++) 3929 cmpxchg(&tags->rqs[i], flush_rq, NULL); 3930} 3931 3932static void blk_free_flush_queue_callback(struct rcu_head *head) 3933{ 3934 struct blk_flush_queue *fq = 3935 container_of(head, struct blk_flush_queue, rcu_head); 3936 3937 blk_free_flush_queue(fq); 3938} 3939 3940/* hctx->ctxs will be freed in queue's release handler */ 3941static void blk_mq_exit_hctx(struct request_queue *q, 3942 struct blk_mq_tag_set *set, 3943 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) 3944{ 3945 struct request *flush_rq = hctx->fq->flush_rq; 3946 3947 if (blk_mq_hw_queue_mapped(hctx)) 3948 blk_mq_tag_idle(hctx); 3949 3950 if (blk_queue_init_done(q)) 3951 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], 3952 set->queue_depth, flush_rq); 3953 if (set->ops->exit_request) 3954 set->ops->exit_request(set, flush_rq, hctx_idx); 3955 3956 if (set->ops->exit_hctx) 3957 set->ops->exit_hctx(hctx, hctx_idx); 3958 3959 call_srcu(&set->tags_srcu, &hctx->fq->rcu_head, 3960 blk_free_flush_queue_callback); 3961 hctx->fq = NULL; 3962 3963 spin_lock(&q->unused_hctx_lock); 3964 list_add(&hctx->hctx_list, &q->unused_hctx_list); 3965 spin_unlock(&q->unused_hctx_lock); 3966} 3967 3968static void blk_mq_exit_hw_queues(struct request_queue *q, 3969 struct blk_mq_tag_set *set, int nr_queue) 3970{ 3971 struct blk_mq_hw_ctx *hctx; 3972 unsigned long i; 3973 3974 queue_for_each_hw_ctx(q, hctx, i) { 3975 if (i == nr_queue) 3976 break; 3977 blk_mq_remove_cpuhp(hctx); 3978 blk_mq_exit_hctx(q, set, hctx, i); 3979 } 3980} 3981 3982static int blk_mq_init_hctx(struct request_queue *q, 3983 struct blk_mq_tag_set *set, 3984 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) 3985{ 3986 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; 3987 3988 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); 3989 if (!hctx->fq) 3990 goto fail; 3991 3992 hctx->queue_num = hctx_idx; 3993 3994 hctx->tags = set->tags[hctx_idx]; 3995 3996 if (set->ops->init_hctx && 3997 set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) 3998 goto fail_free_fq; 3999 4000 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, 4001 hctx->numa_node)) 4002 goto exit_hctx; 4003 4004 return 0; 4005 4006 exit_hctx: 4007 if (set->ops->exit_hctx) 4008 set->ops->exit_hctx(hctx, hctx_idx); 4009 fail_free_fq: 4010 blk_free_flush_queue(hctx->fq); 4011 hctx->fq = NULL; 4012 fail: 4013 return -1; 4014} 4015 4016static struct blk_mq_hw_ctx * 4017blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, 4018 int node) 4019{ 4020 struct blk_mq_hw_ctx *hctx; 4021 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; 4022 4023 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); 4024 if (!hctx) 4025 goto fail_alloc_hctx; 4026 4027 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) 4028 goto free_hctx; 4029 4030 atomic_set(&hctx->nr_active, 0); 4031 if (node == NUMA_NO_NODE) 4032 node = set->numa_node; 4033 hctx->numa_node = node; 4034 4035 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); 4036 spin_lock_init(&hctx->lock); 4037 INIT_LIST_HEAD(&hctx->dispatch); 4038 INIT_HLIST_NODE(&hctx->cpuhp_dead); 4039 INIT_HLIST_NODE(&hctx->cpuhp_online); 4040 hctx->queue = q; 4041 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; 4042 4043 INIT_LIST_HEAD(&hctx->hctx_list); 4044 4045 /* 4046 * Allocate space for all possible cpus to avoid allocation at 4047 * runtime 4048 */ 4049 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), 4050 gfp, node); 4051 if (!hctx->ctxs) 4052 goto free_cpumask; 4053 4054 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), 4055 gfp, node, false, false)) 4056 goto free_ctxs; 4057 hctx->nr_ctx = 0; 4058 4059 spin_lock_init(&hctx->dispatch_wait_lock); 4060 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); 4061 INIT_LIST_HEAD(&hctx->dispatch_wait.entry); 4062 4063 blk_mq_hctx_kobj_init(hctx); 4064 4065 return hctx; 4066 4067 free_ctxs: 4068 kfree(hctx->ctxs); 4069 free_cpumask: 4070 free_cpumask_var(hctx->cpumask); 4071 free_hctx: 4072 kfree(hctx); 4073 fail_alloc_hctx: 4074 return NULL; 4075} 4076 4077static void blk_mq_init_cpu_queues(struct request_queue *q, 4078 unsigned int nr_hw_queues) 4079{ 4080 struct blk_mq_tag_set *set = q->tag_set; 4081 unsigned int i, j; 4082 4083 for_each_possible_cpu(i) { 4084 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); 4085 struct blk_mq_hw_ctx *hctx; 4086 int k; 4087 4088 __ctx->cpu = i; 4089 spin_lock_init(&__ctx->lock); 4090 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) 4091 INIT_LIST_HEAD(&__ctx->rq_lists[k]); 4092 4093 __ctx->queue = q; 4094 4095 /* 4096 * Set local node, IFF we have more than one hw queue. If 4097 * not, we remain on the home node of the device 4098 */ 4099 for (j = 0; j < set->nr_maps; j++) { 4100 hctx = blk_mq_map_queue_type(q, j, i); 4101 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) 4102 hctx->numa_node = cpu_to_node(i); 4103 } 4104 } 4105} 4106 4107struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, 4108 unsigned int hctx_idx, 4109 unsigned int depth) 4110{ 4111 struct blk_mq_tags *tags; 4112 int ret; 4113 4114 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); 4115 if (!tags) 4116 return NULL; 4117 4118 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); 4119 if (ret) { 4120 blk_mq_free_rq_map(set, tags); 4121 return NULL; 4122 } 4123 4124 return tags; 4125} 4126 4127static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, 4128 int hctx_idx) 4129{ 4130 if (blk_mq_is_shared_tags(set->flags)) { 4131 set->tags[hctx_idx] = set->shared_tags; 4132 4133 return true; 4134 } 4135 4136 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, 4137 set->queue_depth); 4138 4139 return set->tags[hctx_idx]; 4140} 4141 4142void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, 4143 struct blk_mq_tags *tags, 4144 unsigned int hctx_idx) 4145{ 4146 if (tags) { 4147 blk_mq_free_rqs(set, tags, hctx_idx); 4148 blk_mq_free_rq_map(set, tags); 4149 } 4150} 4151 4152static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, 4153 unsigned int hctx_idx) 4154{ 4155 if (!blk_mq_is_shared_tags(set->flags)) 4156 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); 4157 4158 set->tags[hctx_idx] = NULL; 4159} 4160 4161static void blk_mq_map_swqueue(struct request_queue *q) 4162{ 4163 unsigned int j, hctx_idx; 4164 unsigned long i; 4165 struct blk_mq_hw_ctx *hctx; 4166 struct blk_mq_ctx *ctx; 4167 struct blk_mq_tag_set *set = q->tag_set; 4168 4169 queue_for_each_hw_ctx(q, hctx, i) { 4170 cpumask_clear(hctx->cpumask); 4171 hctx->nr_ctx = 0; 4172 hctx->dispatch_from = NULL; 4173 } 4174 4175 /* 4176 * Map software to hardware queues. 4177 * 4178 * If the cpu isn't present, the cpu is mapped to first hctx. 4179 */ 4180 for_each_possible_cpu(i) { 4181 4182 ctx = per_cpu_ptr(q->queue_ctx, i); 4183 for (j = 0; j < set->nr_maps; j++) { 4184 if (!set->map[j].nr_queues) { 4185 ctx->hctxs[j] = blk_mq_map_queue_type(q, 4186 HCTX_TYPE_DEFAULT, i); 4187 continue; 4188 } 4189 hctx_idx = set->map[j].mq_map[i]; 4190 /* unmapped hw queue can be remapped after CPU topo changed */ 4191 if (!set->tags[hctx_idx] && 4192 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { 4193 /* 4194 * If tags initialization fail for some hctx, 4195 * that hctx won't be brought online. In this 4196 * case, remap the current ctx to hctx[0] which 4197 * is guaranteed to always have tags allocated 4198 */ 4199 set->map[j].mq_map[i] = 0; 4200 } 4201 4202 hctx = blk_mq_map_queue_type(q, j, i); 4203 ctx->hctxs[j] = hctx; 4204 /* 4205 * If the CPU is already set in the mask, then we've 4206 * mapped this one already. This can happen if 4207 * devices share queues across queue maps. 4208 */ 4209 if (cpumask_test_cpu(i, hctx->cpumask)) 4210 continue; 4211 4212 cpumask_set_cpu(i, hctx->cpumask); 4213 hctx->type = j; 4214 ctx->index_hw[hctx->type] = hctx->nr_ctx; 4215 hctx->ctxs[hctx->nr_ctx++] = ctx; 4216 4217 /* 4218 * If the nr_ctx type overflows, we have exceeded the 4219 * amount of sw queues we can support. 4220 */ 4221 BUG_ON(!hctx->nr_ctx); 4222 } 4223 4224 for (; j < HCTX_MAX_TYPES; j++) 4225 ctx->hctxs[j] = blk_mq_map_queue_type(q, 4226 HCTX_TYPE_DEFAULT, i); 4227 } 4228 4229 queue_for_each_hw_ctx(q, hctx, i) { 4230 int cpu; 4231 4232 /* 4233 * If no software queues are mapped to this hardware queue, 4234 * disable it and free the request entries. 4235 */ 4236 if (!hctx->nr_ctx) { 4237 /* Never unmap queue 0. We need it as a 4238 * fallback in case of a new remap fails 4239 * allocation 4240 */ 4241 if (i) 4242 __blk_mq_free_map_and_rqs(set, i); 4243 4244 hctx->tags = NULL; 4245 continue; 4246 } 4247 4248 hctx->tags = set->tags[i]; 4249 WARN_ON(!hctx->tags); 4250 4251 /* 4252 * Set the map size to the number of mapped software queues. 4253 * This is more accurate and more efficient than looping 4254 * over all possibly mapped software queues. 4255 */ 4256 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); 4257 4258 /* 4259 * Rule out isolated CPUs from hctx->cpumask to avoid 4260 * running block kworker on isolated CPUs 4261 */ 4262 for_each_cpu(cpu, hctx->cpumask) { 4263 if (cpu_is_isolated(cpu)) 4264 cpumask_clear_cpu(cpu, hctx->cpumask); 4265 } 4266 4267 /* 4268 * Initialize batch roundrobin counts 4269 */ 4270 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); 4271 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; 4272 } 4273} 4274 4275/* 4276 * Caller needs to ensure that we're either frozen/quiesced, or that 4277 * the queue isn't live yet. 4278 */ 4279static void queue_set_hctx_shared(struct request_queue *q, bool shared) 4280{ 4281 struct blk_mq_hw_ctx *hctx; 4282 unsigned long i; 4283 4284 queue_for_each_hw_ctx(q, hctx, i) { 4285 if (shared) { 4286 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; 4287 } else { 4288 blk_mq_tag_idle(hctx); 4289 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; 4290 } 4291 } 4292} 4293 4294static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, 4295 bool shared) 4296{ 4297 struct request_queue *q; 4298 unsigned int memflags; 4299 4300 lockdep_assert_held(&set->tag_list_lock); 4301 4302 list_for_each_entry(q, &set->tag_list, tag_set_list) { 4303 memflags = blk_mq_freeze_queue(q); 4304 queue_set_hctx_shared(q, shared); 4305 blk_mq_unfreeze_queue(q, memflags); 4306 } 4307} 4308 4309static void blk_mq_del_queue_tag_set(struct request_queue *q) 4310{ 4311 struct blk_mq_tag_set *set = q->tag_set; 4312 4313 mutex_lock(&set->tag_list_lock); 4314 list_del_rcu(&q->tag_set_list); 4315 if (list_is_singular(&set->tag_list)) { 4316 /* just transitioned to unshared */ 4317 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; 4318 /* update existing queue */ 4319 blk_mq_update_tag_set_shared(set, false); 4320 } 4321 mutex_unlock(&set->tag_list_lock); 4322} 4323 4324static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, 4325 struct request_queue *q) 4326{ 4327 mutex_lock(&set->tag_list_lock); 4328 4329 /* 4330 * Check to see if we're transitioning to shared (from 1 to 2 queues). 4331 */ 4332 if (!list_empty(&set->tag_list) && 4333 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { 4334 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; 4335 /* update existing queue */ 4336 blk_mq_update_tag_set_shared(set, true); 4337 } 4338 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 4339 queue_set_hctx_shared(q, true); 4340 list_add_tail_rcu(&q->tag_set_list, &set->tag_list); 4341 4342 mutex_unlock(&set->tag_list_lock); 4343} 4344 4345/* All allocations will be freed in release handler of q->mq_kobj */ 4346static int blk_mq_alloc_ctxs(struct request_queue *q) 4347{ 4348 struct blk_mq_ctxs *ctxs; 4349 int cpu; 4350 4351 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); 4352 if (!ctxs) 4353 return -ENOMEM; 4354 4355 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); 4356 if (!ctxs->queue_ctx) 4357 goto fail; 4358 4359 for_each_possible_cpu(cpu) { 4360 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); 4361 ctx->ctxs = ctxs; 4362 } 4363 4364 q->mq_kobj = &ctxs->kobj; 4365 q->queue_ctx = ctxs->queue_ctx; 4366 4367 return 0; 4368 fail: 4369 kfree(ctxs); 4370 return -ENOMEM; 4371} 4372 4373/* 4374 * It is the actual release handler for mq, but we do it from 4375 * request queue's release handler for avoiding use-after-free 4376 * and headache because q->mq_kobj shouldn't have been introduced, 4377 * but we can't group ctx/kctx kobj without it. 4378 */ 4379void blk_mq_release(struct request_queue *q) 4380{ 4381 struct blk_mq_hw_ctx *hctx, *next; 4382 unsigned long i; 4383 4384 queue_for_each_hw_ctx(q, hctx, i) 4385 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); 4386 4387 /* all hctx are in .unused_hctx_list now */ 4388 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { 4389 list_del_init(&hctx->hctx_list); 4390 kobject_put(&hctx->kobj); 4391 } 4392 4393 kfree(q->queue_hw_ctx); 4394 4395 /* 4396 * release .mq_kobj and sw queue's kobject now because 4397 * both share lifetime with request queue. 4398 */ 4399 blk_mq_sysfs_deinit(q); 4400} 4401 4402struct request_queue *blk_mq_alloc_queue(struct blk_mq_tag_set *set, 4403 struct queue_limits *lim, void *queuedata) 4404{ 4405 struct queue_limits default_lim = { }; 4406 struct request_queue *q; 4407 int ret; 4408 4409 if (!lim) 4410 lim = &default_lim; 4411 lim->features |= BLK_FEAT_IO_STAT | BLK_FEAT_NOWAIT; 4412 if (set->nr_maps > HCTX_TYPE_POLL) 4413 lim->features |= BLK_FEAT_POLL; 4414 4415 q = blk_alloc_queue(lim, set->numa_node); 4416 if (IS_ERR(q)) 4417 return q; 4418 q->queuedata = queuedata; 4419 ret = blk_mq_init_allocated_queue(set, q); 4420 if (ret) { 4421 blk_put_queue(q); 4422 return ERR_PTR(ret); 4423 } 4424 return q; 4425} 4426EXPORT_SYMBOL(blk_mq_alloc_queue); 4427 4428/** 4429 * blk_mq_destroy_queue - shutdown a request queue 4430 * @q: request queue to shutdown 4431 * 4432 * This shuts down a request queue allocated by blk_mq_alloc_queue(). All future 4433 * requests will be failed with -ENODEV. The caller is responsible for dropping 4434 * the reference from blk_mq_alloc_queue() by calling blk_put_queue(). 4435 * 4436 * Context: can sleep 4437 */ 4438void blk_mq_destroy_queue(struct request_queue *q) 4439{ 4440 WARN_ON_ONCE(!queue_is_mq(q)); 4441 WARN_ON_ONCE(blk_queue_registered(q)); 4442 4443 might_sleep(); 4444 4445 blk_queue_flag_set(QUEUE_FLAG_DYING, q); 4446 blk_queue_start_drain(q); 4447 blk_mq_freeze_queue_wait(q); 4448 4449 blk_sync_queue(q); 4450 blk_mq_cancel_work_sync(q); 4451 blk_mq_exit_queue(q); 4452} 4453EXPORT_SYMBOL(blk_mq_destroy_queue); 4454 4455struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, 4456 struct queue_limits *lim, void *queuedata, 4457 struct lock_class_key *lkclass) 4458{ 4459 struct request_queue *q; 4460 struct gendisk *disk; 4461 4462 q = blk_mq_alloc_queue(set, lim, queuedata); 4463 if (IS_ERR(q)) 4464 return ERR_CAST(q); 4465 4466 disk = __alloc_disk_node(q, set->numa_node, lkclass); 4467 if (!disk) { 4468 blk_mq_destroy_queue(q); 4469 blk_put_queue(q); 4470 return ERR_PTR(-ENOMEM); 4471 } 4472 set_bit(GD_OWNS_QUEUE, &disk->state); 4473 return disk; 4474} 4475EXPORT_SYMBOL(__blk_mq_alloc_disk); 4476 4477struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, 4478 struct lock_class_key *lkclass) 4479{ 4480 struct gendisk *disk; 4481 4482 if (!blk_get_queue(q)) 4483 return NULL; 4484 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); 4485 if (!disk) 4486 blk_put_queue(q); 4487 return disk; 4488} 4489EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); 4490 4491/* 4492 * Only hctx removed from cpuhp list can be reused 4493 */ 4494static bool blk_mq_hctx_is_reusable(struct blk_mq_hw_ctx *hctx) 4495{ 4496 return hlist_unhashed(&hctx->cpuhp_online) && 4497 hlist_unhashed(&hctx->cpuhp_dead); 4498} 4499 4500static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( 4501 struct blk_mq_tag_set *set, struct request_queue *q, 4502 int hctx_idx, int node) 4503{ 4504 struct blk_mq_hw_ctx *hctx = NULL, *tmp; 4505 4506 /* reuse dead hctx first */ 4507 spin_lock(&q->unused_hctx_lock); 4508 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { 4509 if (tmp->numa_node == node && blk_mq_hctx_is_reusable(tmp)) { 4510 hctx = tmp; 4511 break; 4512 } 4513 } 4514 if (hctx) 4515 list_del_init(&hctx->hctx_list); 4516 spin_unlock(&q->unused_hctx_lock); 4517 4518 if (!hctx) 4519 hctx = blk_mq_alloc_hctx(q, set, node); 4520 if (!hctx) 4521 goto fail; 4522 4523 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) 4524 goto free_hctx; 4525 4526 return hctx; 4527 4528 free_hctx: 4529 kobject_put(&hctx->kobj); 4530 fail: 4531 return NULL; 4532} 4533 4534static void __blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, 4535 struct request_queue *q) 4536{ 4537 int i, j, end; 4538 struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx; 4539 4540 if (q->nr_hw_queues < set->nr_hw_queues) { 4541 struct blk_mq_hw_ctx **new_hctxs; 4542 4543 new_hctxs = kcalloc_node(set->nr_hw_queues, 4544 sizeof(*new_hctxs), GFP_KERNEL, 4545 set->numa_node); 4546 if (!new_hctxs) 4547 return; 4548 if (hctxs) 4549 memcpy(new_hctxs, hctxs, q->nr_hw_queues * 4550 sizeof(*hctxs)); 4551 rcu_assign_pointer(q->queue_hw_ctx, new_hctxs); 4552 /* 4553 * Make sure reading the old queue_hw_ctx from other 4554 * context concurrently won't trigger uaf. 4555 */ 4556 kfree_rcu_mightsleep(hctxs); 4557 hctxs = new_hctxs; 4558 } 4559 4560 for (i = 0; i < set->nr_hw_queues; i++) { 4561 int old_node; 4562 int node = blk_mq_get_hctx_node(set, i); 4563 struct blk_mq_hw_ctx *old_hctx = hctxs[i]; 4564 4565 if (old_hctx) { 4566 old_node = old_hctx->numa_node; 4567 blk_mq_exit_hctx(q, set, old_hctx, i); 4568 } 4569 4570 hctxs[i] = blk_mq_alloc_and_init_hctx(set, q, i, node); 4571 if (!hctxs[i]) { 4572 if (!old_hctx) 4573 break; 4574 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", 4575 node, old_node); 4576 hctxs[i] = blk_mq_alloc_and_init_hctx(set, q, i, 4577 old_node); 4578 WARN_ON_ONCE(!hctxs[i]); 4579 } 4580 } 4581 /* 4582 * Increasing nr_hw_queues fails. Free the newly allocated 4583 * hctxs and keep the previous q->nr_hw_queues. 4584 */ 4585 if (i != set->nr_hw_queues) { 4586 j = q->nr_hw_queues; 4587 end = i; 4588 } else { 4589 j = i; 4590 end = q->nr_hw_queues; 4591 q->nr_hw_queues = set->nr_hw_queues; 4592 } 4593 4594 for (; j < end; j++) { 4595 struct blk_mq_hw_ctx *hctx = hctxs[j]; 4596 4597 if (hctx) { 4598 blk_mq_exit_hctx(q, set, hctx, j); 4599 hctxs[j] = NULL; 4600 } 4601 } 4602} 4603 4604static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, 4605 struct request_queue *q) 4606{ 4607 __blk_mq_realloc_hw_ctxs(set, q); 4608 4609 /* unregister cpuhp callbacks for exited hctxs */ 4610 blk_mq_remove_hw_queues_cpuhp(q); 4611 4612 /* register cpuhp for new initialized hctxs */ 4613 blk_mq_add_hw_queues_cpuhp(q); 4614} 4615 4616int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, 4617 struct request_queue *q) 4618{ 4619 /* mark the queue as mq asap */ 4620 q->mq_ops = set->ops; 4621 4622 /* 4623 * ->tag_set has to be setup before initialize hctx, which cpuphp 4624 * handler needs it for checking queue mapping 4625 */ 4626 q->tag_set = set; 4627 4628 if (blk_mq_alloc_ctxs(q)) 4629 goto err_exit; 4630 4631 /* init q->mq_kobj and sw queues' kobjects */ 4632 blk_mq_sysfs_init(q); 4633 4634 INIT_LIST_HEAD(&q->unused_hctx_list); 4635 spin_lock_init(&q->unused_hctx_lock); 4636 4637 blk_mq_realloc_hw_ctxs(set, q); 4638 if (!q->nr_hw_queues) 4639 goto err_hctxs; 4640 4641 INIT_WORK(&q->timeout_work, blk_mq_timeout_work); 4642 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); 4643 4644 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; 4645 4646 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); 4647 INIT_LIST_HEAD(&q->flush_list); 4648 INIT_LIST_HEAD(&q->requeue_list); 4649 spin_lock_init(&q->requeue_lock); 4650 4651 q->nr_requests = set->queue_depth; 4652 4653 blk_mq_init_cpu_queues(q, set->nr_hw_queues); 4654 blk_mq_map_swqueue(q); 4655 blk_mq_add_queue_tag_set(set, q); 4656 return 0; 4657 4658err_hctxs: 4659 blk_mq_release(q); 4660err_exit: 4661 q->mq_ops = NULL; 4662 return -ENOMEM; 4663} 4664EXPORT_SYMBOL(blk_mq_init_allocated_queue); 4665 4666/* tags can _not_ be used after returning from blk_mq_exit_queue */ 4667void blk_mq_exit_queue(struct request_queue *q) 4668{ 4669 struct blk_mq_tag_set *set = q->tag_set; 4670 4671 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ 4672 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); 4673 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ 4674 blk_mq_del_queue_tag_set(q); 4675} 4676 4677static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) 4678{ 4679 int i; 4680 4681 if (blk_mq_is_shared_tags(set->flags)) { 4682 set->shared_tags = blk_mq_alloc_map_and_rqs(set, 4683 BLK_MQ_NO_HCTX_IDX, 4684 set->queue_depth); 4685 if (!set->shared_tags) 4686 return -ENOMEM; 4687 } 4688 4689 for (i = 0; i < set->nr_hw_queues; i++) { 4690 if (!__blk_mq_alloc_map_and_rqs(set, i)) 4691 goto out_unwind; 4692 cond_resched(); 4693 } 4694 4695 return 0; 4696 4697out_unwind: 4698 while (--i >= 0) 4699 __blk_mq_free_map_and_rqs(set, i); 4700 4701 if (blk_mq_is_shared_tags(set->flags)) { 4702 blk_mq_free_map_and_rqs(set, set->shared_tags, 4703 BLK_MQ_NO_HCTX_IDX); 4704 } 4705 4706 return -ENOMEM; 4707} 4708 4709/* 4710 * Allocate the request maps associated with this tag_set. Note that this 4711 * may reduce the depth asked for, if memory is tight. set->queue_depth 4712 * will be updated to reflect the allocated depth. 4713 */ 4714static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) 4715{ 4716 unsigned int depth; 4717 int err; 4718 4719 depth = set->queue_depth; 4720 do { 4721 err = __blk_mq_alloc_rq_maps(set); 4722 if (!err) 4723 break; 4724 4725 set->queue_depth >>= 1; 4726 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { 4727 err = -ENOMEM; 4728 break; 4729 } 4730 } while (set->queue_depth); 4731 4732 if (!set->queue_depth || err) { 4733 pr_err("blk-mq: failed to allocate request map\n"); 4734 return -ENOMEM; 4735 } 4736 4737 if (depth != set->queue_depth) 4738 pr_info("blk-mq: reduced tag depth (%u -> %u)\n", 4739 depth, set->queue_depth); 4740 4741 return 0; 4742} 4743 4744static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) 4745{ 4746 /* 4747 * blk_mq_map_queues() and multiple .map_queues() implementations 4748 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the 4749 * number of hardware queues. 4750 */ 4751 if (set->nr_maps == 1) 4752 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; 4753 4754 if (set->ops->map_queues) { 4755 int i; 4756 4757 /* 4758 * transport .map_queues is usually done in the following 4759 * way: 4760 * 4761 * for (queue = 0; queue < set->nr_hw_queues; queue++) { 4762 * mask = get_cpu_mask(queue) 4763 * for_each_cpu(cpu, mask) 4764 * set->map[x].mq_map[cpu] = queue; 4765 * } 4766 * 4767 * When we need to remap, the table has to be cleared for 4768 * killing stale mapping since one CPU may not be mapped 4769 * to any hw queue. 4770 */ 4771 for (i = 0; i < set->nr_maps; i++) 4772 blk_mq_clear_mq_map(&set->map[i]); 4773 4774 set->ops->map_queues(set); 4775 } else { 4776 BUG_ON(set->nr_maps > 1); 4777 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); 4778 } 4779} 4780 4781static struct blk_mq_tags **blk_mq_prealloc_tag_set_tags( 4782 struct blk_mq_tag_set *set, 4783 int new_nr_hw_queues) 4784{ 4785 struct blk_mq_tags **new_tags; 4786 int i; 4787 4788 if (set->nr_hw_queues >= new_nr_hw_queues) 4789 return NULL; 4790 4791 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), 4792 GFP_KERNEL, set->numa_node); 4793 if (!new_tags) 4794 return ERR_PTR(-ENOMEM); 4795 4796 if (set->tags) 4797 memcpy(new_tags, set->tags, set->nr_hw_queues * 4798 sizeof(*set->tags)); 4799 4800 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { 4801 if (blk_mq_is_shared_tags(set->flags)) { 4802 new_tags[i] = set->shared_tags; 4803 } else { 4804 new_tags[i] = blk_mq_alloc_map_and_rqs(set, i, 4805 set->queue_depth); 4806 if (!new_tags[i]) 4807 goto out_unwind; 4808 } 4809 cond_resched(); 4810 } 4811 4812 return new_tags; 4813out_unwind: 4814 while (--i >= set->nr_hw_queues) { 4815 if (!blk_mq_is_shared_tags(set->flags)) 4816 blk_mq_free_map_and_rqs(set, new_tags[i], i); 4817 } 4818 kfree(new_tags); 4819 return ERR_PTR(-ENOMEM); 4820} 4821 4822/* 4823 * Alloc a tag set to be associated with one or more request queues. 4824 * May fail with EINVAL for various error conditions. May adjust the 4825 * requested depth down, if it's too large. In that case, the set 4826 * value will be stored in set->queue_depth. 4827 */ 4828int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) 4829{ 4830 int i, ret; 4831 4832 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); 4833 4834 if (!set->nr_hw_queues) 4835 return -EINVAL; 4836 if (!set->queue_depth) 4837 return -EINVAL; 4838 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) 4839 return -EINVAL; 4840 4841 if (!set->ops->queue_rq) 4842 return -EINVAL; 4843 4844 if (!set->ops->get_budget ^ !set->ops->put_budget) 4845 return -EINVAL; 4846 4847 if (set->queue_depth > BLK_MQ_MAX_DEPTH) { 4848 pr_info("blk-mq: reduced tag depth to %u\n", 4849 BLK_MQ_MAX_DEPTH); 4850 set->queue_depth = BLK_MQ_MAX_DEPTH; 4851 } 4852 4853 if (!set->nr_maps) 4854 set->nr_maps = 1; 4855 else if (set->nr_maps > HCTX_MAX_TYPES) 4856 return -EINVAL; 4857 4858 /* 4859 * If a crashdump is active, then we are potentially in a very 4860 * memory constrained environment. Limit us to 64 tags to prevent 4861 * using too much memory. 4862 */ 4863 if (is_kdump_kernel()) 4864 set->queue_depth = min(64U, set->queue_depth); 4865 4866 /* 4867 * There is no use for more h/w queues than cpus if we just have 4868 * a single map 4869 */ 4870 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) 4871 set->nr_hw_queues = nr_cpu_ids; 4872 4873 if (set->flags & BLK_MQ_F_BLOCKING) { 4874 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); 4875 if (!set->srcu) 4876 return -ENOMEM; 4877 ret = init_srcu_struct(set->srcu); 4878 if (ret) 4879 goto out_free_srcu; 4880 } 4881 ret = init_srcu_struct(&set->tags_srcu); 4882 if (ret) 4883 goto out_cleanup_srcu; 4884 4885 init_rwsem(&set->update_nr_hwq_lock); 4886 4887 ret = -ENOMEM; 4888 set->tags = kcalloc_node(set->nr_hw_queues, 4889 sizeof(struct blk_mq_tags *), GFP_KERNEL, 4890 set->numa_node); 4891 if (!set->tags) 4892 goto out_cleanup_tags_srcu; 4893 4894 for (i = 0; i < set->nr_maps; i++) { 4895 set->map[i].mq_map = kcalloc_node(nr_cpu_ids, 4896 sizeof(set->map[i].mq_map[0]), 4897 GFP_KERNEL, set->numa_node); 4898 if (!set->map[i].mq_map) 4899 goto out_free_mq_map; 4900 set->map[i].nr_queues = set->nr_hw_queues; 4901 } 4902 4903 blk_mq_update_queue_map(set); 4904 4905 ret = blk_mq_alloc_set_map_and_rqs(set); 4906 if (ret) 4907 goto out_free_mq_map; 4908 4909 mutex_init(&set->tag_list_lock); 4910 INIT_LIST_HEAD(&set->tag_list); 4911 4912 return 0; 4913 4914out_free_mq_map: 4915 for (i = 0; i < set->nr_maps; i++) { 4916 kfree(set->map[i].mq_map); 4917 set->map[i].mq_map = NULL; 4918 } 4919 kfree(set->tags); 4920 set->tags = NULL; 4921out_cleanup_tags_srcu: 4922 cleanup_srcu_struct(&set->tags_srcu); 4923out_cleanup_srcu: 4924 if (set->flags & BLK_MQ_F_BLOCKING) 4925 cleanup_srcu_struct(set->srcu); 4926out_free_srcu: 4927 if (set->flags & BLK_MQ_F_BLOCKING) 4928 kfree(set->srcu); 4929 return ret; 4930} 4931EXPORT_SYMBOL(blk_mq_alloc_tag_set); 4932 4933/* allocate and initialize a tagset for a simple single-queue device */ 4934int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, 4935 const struct blk_mq_ops *ops, unsigned int queue_depth, 4936 unsigned int set_flags) 4937{ 4938 memset(set, 0, sizeof(*set)); 4939 set->ops = ops; 4940 set->nr_hw_queues = 1; 4941 set->nr_maps = 1; 4942 set->queue_depth = queue_depth; 4943 set->numa_node = NUMA_NO_NODE; 4944 set->flags = set_flags; 4945 return blk_mq_alloc_tag_set(set); 4946} 4947EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); 4948 4949void blk_mq_free_tag_set(struct blk_mq_tag_set *set) 4950{ 4951 int i, j; 4952 4953 for (i = 0; i < set->nr_hw_queues; i++) 4954 __blk_mq_free_map_and_rqs(set, i); 4955 4956 if (blk_mq_is_shared_tags(set->flags)) { 4957 blk_mq_free_map_and_rqs(set, set->shared_tags, 4958 BLK_MQ_NO_HCTX_IDX); 4959 } 4960 4961 for (j = 0; j < set->nr_maps; j++) { 4962 kfree(set->map[j].mq_map); 4963 set->map[j].mq_map = NULL; 4964 } 4965 4966 kfree(set->tags); 4967 set->tags = NULL; 4968 4969 srcu_barrier(&set->tags_srcu); 4970 cleanup_srcu_struct(&set->tags_srcu); 4971 if (set->flags & BLK_MQ_F_BLOCKING) { 4972 cleanup_srcu_struct(set->srcu); 4973 kfree(set->srcu); 4974 } 4975} 4976EXPORT_SYMBOL(blk_mq_free_tag_set); 4977 4978struct elevator_tags *blk_mq_update_nr_requests(struct request_queue *q, 4979 struct elevator_tags *et, 4980 unsigned int nr) 4981{ 4982 struct blk_mq_tag_set *set = q->tag_set; 4983 struct elevator_tags *old_et = NULL; 4984 struct blk_mq_hw_ctx *hctx; 4985 unsigned long i; 4986 4987 blk_mq_quiesce_queue(q); 4988 4989 if (blk_mq_is_shared_tags(set->flags)) { 4990 /* 4991 * Shared tags, for sched tags, we allocate max initially hence 4992 * tags can't grow, see blk_mq_alloc_sched_tags(). 4993 */ 4994 if (q->elevator) 4995 blk_mq_tag_update_sched_shared_tags(q, nr); 4996 else 4997 blk_mq_tag_resize_shared_tags(set, nr); 4998 } else if (!q->elevator) { 4999 /* 5000 * Non-shared hardware tags, nr is already checked from 5001 * queue_requests_store() and tags can't grow. 5002 */ 5003 queue_for_each_hw_ctx(q, hctx, i) { 5004 if (!hctx->tags) 5005 continue; 5006 sbitmap_queue_resize(&hctx->tags->bitmap_tags, 5007 nr - hctx->tags->nr_reserved_tags); 5008 } 5009 } else if (nr <= q->elevator->et->nr_requests) { 5010 /* Non-shared sched tags, and tags don't grow. */ 5011 queue_for_each_hw_ctx(q, hctx, i) { 5012 if (!hctx->sched_tags) 5013 continue; 5014 sbitmap_queue_resize(&hctx->sched_tags->bitmap_tags, 5015 nr - hctx->sched_tags->nr_reserved_tags); 5016 } 5017 } else { 5018 /* Non-shared sched tags, and tags grow */ 5019 queue_for_each_hw_ctx(q, hctx, i) 5020 hctx->sched_tags = et->tags[i]; 5021 old_et = q->elevator->et; 5022 q->elevator->et = et; 5023 } 5024 5025 q->nr_requests = nr; 5026 if (q->elevator && q->elevator->type->ops.depth_updated) 5027 q->elevator->type->ops.depth_updated(q); 5028 5029 blk_mq_unquiesce_queue(q); 5030 return old_et; 5031} 5032 5033/* 5034 * Switch back to the elevator type stored in the xarray. 5035 */ 5036static void blk_mq_elv_switch_back(struct request_queue *q, 5037 struct xarray *elv_tbl) 5038{ 5039 struct elv_change_ctx *ctx = xa_load(elv_tbl, q->id); 5040 5041 if (WARN_ON_ONCE(!ctx)) 5042 return; 5043 5044 /* The elv_update_nr_hw_queues unfreezes the queue. */ 5045 elv_update_nr_hw_queues(q, ctx); 5046 5047 /* Drop the reference acquired in blk_mq_elv_switch_none. */ 5048 if (ctx->type) 5049 elevator_put(ctx->type); 5050} 5051 5052/* 5053 * Stores elevator name and type in ctx and set current elevator to none. 5054 */ 5055static int blk_mq_elv_switch_none(struct request_queue *q, 5056 struct xarray *elv_tbl) 5057{ 5058 struct elv_change_ctx *ctx; 5059 5060 lockdep_assert_held_write(&q->tag_set->update_nr_hwq_lock); 5061 5062 /* 5063 * Accessing q->elevator without holding q->elevator_lock is safe here 5064 * because we're called from nr_hw_queue update which is protected by 5065 * set->update_nr_hwq_lock in the writer context. So, scheduler update/ 5066 * switch code (which acquires the same lock in the reader context) 5067 * can't run concurrently. 5068 */ 5069 if (q->elevator) { 5070 ctx = xa_load(elv_tbl, q->id); 5071 if (WARN_ON_ONCE(!ctx)) 5072 return -ENOENT; 5073 5074 ctx->name = q->elevator->type->elevator_name; 5075 5076 /* 5077 * Before we switch elevator to 'none', take a reference to 5078 * the elevator module so that while nr_hw_queue update is 5079 * running, no one can remove elevator module. We'd put the 5080 * reference to elevator module later when we switch back 5081 * elevator. 5082 */ 5083 __elevator_get(q->elevator->type); 5084 5085 /* 5086 * Store elevator type so that we can release the reference 5087 * taken above later. 5088 */ 5089 ctx->type = q->elevator->type; 5090 elevator_set_none(q); 5091 } 5092 return 0; 5093} 5094 5095static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, 5096 int nr_hw_queues) 5097{ 5098 struct request_queue *q; 5099 int prev_nr_hw_queues = set->nr_hw_queues; 5100 unsigned int memflags; 5101 int i; 5102 struct xarray elv_tbl; 5103 struct blk_mq_tags **new_tags; 5104 bool queues_frozen = false; 5105 5106 lockdep_assert_held(&set->tag_list_lock); 5107 5108 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) 5109 nr_hw_queues = nr_cpu_ids; 5110 if (nr_hw_queues < 1) 5111 return; 5112 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) 5113 return; 5114 5115 memflags = memalloc_noio_save(); 5116 5117 xa_init(&elv_tbl); 5118 if (blk_mq_alloc_sched_ctx_batch(&elv_tbl, set) < 0) 5119 goto out_free_ctx; 5120 5121 if (blk_mq_alloc_sched_res_batch(&elv_tbl, set, nr_hw_queues) < 0) 5122 goto out_free_ctx; 5123 5124 list_for_each_entry(q, &set->tag_list, tag_set_list) { 5125 blk_mq_debugfs_unregister_hctxs(q); 5126 blk_mq_sysfs_unregister_hctxs(q); 5127 } 5128 5129 /* 5130 * Switch IO scheduler to 'none', cleaning up the data associated 5131 * with the previous scheduler. We will switch back once we are done 5132 * updating the new sw to hw queue mappings. 5133 */ 5134 list_for_each_entry(q, &set->tag_list, tag_set_list) 5135 if (blk_mq_elv_switch_none(q, &elv_tbl)) 5136 goto switch_back; 5137 5138 new_tags = blk_mq_prealloc_tag_set_tags(set, nr_hw_queues); 5139 if (IS_ERR(new_tags)) 5140 goto switch_back; 5141 5142 list_for_each_entry(q, &set->tag_list, tag_set_list) 5143 blk_mq_freeze_queue_nomemsave(q); 5144 queues_frozen = true; 5145 if (new_tags) { 5146 kfree(set->tags); 5147 set->tags = new_tags; 5148 } 5149 set->nr_hw_queues = nr_hw_queues; 5150 5151fallback: 5152 blk_mq_update_queue_map(set); 5153 list_for_each_entry(q, &set->tag_list, tag_set_list) { 5154 __blk_mq_realloc_hw_ctxs(set, q); 5155 5156 if (q->nr_hw_queues != set->nr_hw_queues) { 5157 int i = prev_nr_hw_queues; 5158 5159 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", 5160 nr_hw_queues, prev_nr_hw_queues); 5161 for (; i < set->nr_hw_queues; i++) 5162 __blk_mq_free_map_and_rqs(set, i); 5163 5164 set->nr_hw_queues = prev_nr_hw_queues; 5165 goto fallback; 5166 } 5167 blk_mq_map_swqueue(q); 5168 } 5169switch_back: 5170 /* The blk_mq_elv_switch_back unfreezes queue for us. */ 5171 list_for_each_entry(q, &set->tag_list, tag_set_list) { 5172 /* switch_back expects queue to be frozen */ 5173 if (!queues_frozen) 5174 blk_mq_freeze_queue_nomemsave(q); 5175 blk_mq_elv_switch_back(q, &elv_tbl); 5176 } 5177 5178 list_for_each_entry(q, &set->tag_list, tag_set_list) { 5179 blk_mq_sysfs_register_hctxs(q); 5180 blk_mq_debugfs_register_hctxs(q); 5181 5182 blk_mq_remove_hw_queues_cpuhp(q); 5183 blk_mq_add_hw_queues_cpuhp(q); 5184 } 5185 5186out_free_ctx: 5187 blk_mq_free_sched_ctx_batch(&elv_tbl); 5188 xa_destroy(&elv_tbl); 5189 memalloc_noio_restore(memflags); 5190 5191 /* Free the excess tags when nr_hw_queues shrink. */ 5192 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) 5193 __blk_mq_free_map_and_rqs(set, i); 5194} 5195 5196void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) 5197{ 5198 down_write(&set->update_nr_hwq_lock); 5199 mutex_lock(&set->tag_list_lock); 5200 __blk_mq_update_nr_hw_queues(set, nr_hw_queues); 5201 mutex_unlock(&set->tag_list_lock); 5202 up_write(&set->update_nr_hwq_lock); 5203} 5204EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); 5205 5206static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, 5207 struct io_comp_batch *iob, unsigned int flags) 5208{ 5209 int ret; 5210 5211 do { 5212 ret = q->mq_ops->poll(hctx, iob); 5213 if (ret > 0) 5214 return ret; 5215 if (task_sigpending(current)) 5216 return 1; 5217 if (ret < 0 || (flags & BLK_POLL_ONESHOT)) 5218 break; 5219 cpu_relax(); 5220 } while (!need_resched()); 5221 5222 return 0; 5223} 5224 5225int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, 5226 struct io_comp_batch *iob, unsigned int flags) 5227{ 5228 if (!blk_mq_can_poll(q)) 5229 return 0; 5230 return blk_hctx_poll(q, q->queue_hw_ctx[cookie], iob, flags); 5231} 5232 5233int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, 5234 unsigned int poll_flags) 5235{ 5236 struct request_queue *q = rq->q; 5237 int ret; 5238 5239 if (!blk_rq_is_poll(rq)) 5240 return 0; 5241 if (!percpu_ref_tryget(&q->q_usage_counter)) 5242 return 0; 5243 5244 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); 5245 blk_queue_exit(q); 5246 5247 return ret; 5248} 5249EXPORT_SYMBOL_GPL(blk_rq_poll); 5250 5251unsigned int blk_mq_rq_cpu(struct request *rq) 5252{ 5253 return rq->mq_ctx->cpu; 5254} 5255EXPORT_SYMBOL(blk_mq_rq_cpu); 5256 5257void blk_mq_cancel_work_sync(struct request_queue *q) 5258{ 5259 struct blk_mq_hw_ctx *hctx; 5260 unsigned long i; 5261 5262 cancel_delayed_work_sync(&q->requeue_work); 5263 5264 queue_for_each_hw_ctx(q, hctx, i) 5265 cancel_delayed_work_sync(&hctx->run_work); 5266} 5267 5268static int __init blk_mq_init(void) 5269{ 5270 int i; 5271 5272 for_each_possible_cpu(i) 5273 init_llist_head(&per_cpu(blk_cpu_done, i)); 5274 for_each_possible_cpu(i) 5275 INIT_CSD(&per_cpu(blk_cpu_csd, i), 5276 __blk_mq_complete_request_remote, NULL); 5277 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); 5278 5279 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, 5280 "block/softirq:dead", NULL, 5281 blk_softirq_cpu_dead); 5282 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, 5283 blk_mq_hctx_notify_dead); 5284 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", 5285 blk_mq_hctx_notify_online, 5286 blk_mq_hctx_notify_offline); 5287 return 0; 5288} 5289subsys_initcall(blk_mq_init);
~/
Read-only file preview.