1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Block multiqueue core code 4 * 5 * Copyright (C) 2013-2014 Jens Axboe 6 * Copyright (C) 2013-2014 Christoph Hellwig 7 */ 8 #include <linux/kernel.h> 9 #include <linux/module.h> 10 #include <linux/backing-dev.h> 11 #include <linux/bio.h> 12 #include <linux/blkdev.h> 13 #include <linux/blk-integrity.h> 14 #include <linux/kmemleak.h> 15 #include <linux/mm.h> 16 #include <linux/init.h> 17 #include <linux/slab.h> 18 #include <linux/workqueue.h> 19 #include <linux/smp.h> 20 #include <linux/interrupt.h> 21 #include <linux/llist.h> 22 #include <linux/cpu.h> 23 #include <linux/cache.h> 24 #include <linux/sched/sysctl.h> 25 #include <linux/sched/topology.h> 26 #include <linux/sched/signal.h> 27 #include <linux/delay.h> 28 #include <linux/crash_dump.h> 29 #include <linux/prefetch.h> 30 #include <linux/blk-crypto.h> 31 #include <linux/part_stat.h> 32 33 #include <trace/events/block.h> 34 35 #include <linux/t10-pi.h> 36 #include "blk.h" 37 #include "blk-mq.h" 38 #include "blk-mq-debugfs.h" 39 #include "blk-pm.h" 40 #include "blk-stat.h" 41 #include "blk-mq-sched.h" 42 #include "blk-rq-qos.h" 43 #include "blk-ioprio.h" 44 45 static DEFINE_PER_CPU(struct llist_head, blk_cpu_done); 46 static DEFINE_PER_CPU(call_single_data_t, blk_cpu_csd); 47 48 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags); 49 static void blk_mq_request_bypass_insert(struct request *rq, 50 blk_insert_t flags); 51 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, 52 struct list_head *list); 53 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, 54 struct io_comp_batch *iob, unsigned int flags); 55 56 /* 57 * Check if any of the ctx, dispatch list or elevator 58 * have pending work in this hardware queue. 59 */ 60 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx) 61 { 62 return !list_empty_careful(&hctx->dispatch) || 63 sbitmap_any_bit_set(&hctx->ctx_map) || 64 blk_mq_sched_has_work(hctx); 65 } 66 67 /* 68 * Mark this ctx as having pending work in this hardware queue 69 */ 70 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx, 71 struct blk_mq_ctx *ctx) 72 { 73 const int bit = ctx->index_hw[hctx->type]; 74 75 if (!sbitmap_test_bit(&hctx->ctx_map, bit)) 76 sbitmap_set_bit(&hctx->ctx_map, bit); 77 } 78 79 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx, 80 struct blk_mq_ctx *ctx) 81 { 82 const int bit = ctx->index_hw[hctx->type]; 83 84 sbitmap_clear_bit(&hctx->ctx_map, bit); 85 } 86 87 struct mq_inflight { 88 struct block_device *part; 89 unsigned int inflight[2]; 90 }; 91 92 static bool blk_mq_check_inflight(struct request *rq, void *priv) 93 { 94 struct mq_inflight *mi = priv; 95 96 if (rq->part && blk_do_io_stat(rq) && 97 (!mi->part->bd_partno || rq->part == mi->part) && 98 blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) 99 mi->inflight[rq_data_dir(rq)]++; 100 101 return true; 102 } 103 104 unsigned int blk_mq_in_flight(struct request_queue *q, 105 struct block_device *part) 106 { 107 struct mq_inflight mi = { .part = part }; 108 109 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); 110 111 return mi.inflight[0] + mi.inflight[1]; 112 } 113 114 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, 115 unsigned int inflight[2]) 116 { 117 struct mq_inflight mi = { .part = part }; 118 119 blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi); 120 inflight[0] = mi.inflight[0]; 121 inflight[1] = mi.inflight[1]; 122 } 123 124 void blk_freeze_queue_start(struct request_queue *q) 125 { 126 mutex_lock(&q->mq_freeze_lock); 127 if (++q->mq_freeze_depth == 1) { 128 percpu_ref_kill(&q->q_usage_counter); 129 mutex_unlock(&q->mq_freeze_lock); 130 if (queue_is_mq(q)) 131 blk_mq_run_hw_queues(q, false); 132 } else { 133 mutex_unlock(&q->mq_freeze_lock); 134 } 135 } 136 EXPORT_SYMBOL_GPL(blk_freeze_queue_start); 137 138 void blk_mq_freeze_queue_wait(struct request_queue *q) 139 { 140 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter)); 141 } 142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); 143 144 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, 145 unsigned long timeout) 146 { 147 return wait_event_timeout(q->mq_freeze_wq, 148 percpu_ref_is_zero(&q->q_usage_counter), 149 timeout); 150 } 151 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout); 152 153 /* 154 * Guarantee no request is in use, so we can change any data structure of 155 * the queue afterward. 156 */ 157 void blk_freeze_queue(struct request_queue *q) 158 { 159 /* 160 * In the !blk_mq case we are only calling this to kill the 161 * q_usage_counter, otherwise this increases the freeze depth 162 * and waits for it to return to zero. For this reason there is 163 * no blk_unfreeze_queue(), and blk_freeze_queue() is not 164 * exported to drivers as the only user for unfreeze is blk_mq. 165 */ 166 blk_freeze_queue_start(q); 167 blk_mq_freeze_queue_wait(q); 168 } 169 170 void blk_mq_freeze_queue(struct request_queue *q) 171 { 172 /* 173 * ...just an alias to keep freeze and unfreeze actions balanced 174 * in the blk_mq_* namespace 175 */ 176 blk_freeze_queue(q); 177 } 178 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue); 179 180 void __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic) 181 { 182 mutex_lock(&q->mq_freeze_lock); 183 if (force_atomic) 184 q->q_usage_counter.data->force_atomic = true; 185 q->mq_freeze_depth--; 186 WARN_ON_ONCE(q->mq_freeze_depth < 0); 187 if (!q->mq_freeze_depth) { 188 percpu_ref_resurrect(&q->q_usage_counter); 189 wake_up_all(&q->mq_freeze_wq); 190 } 191 mutex_unlock(&q->mq_freeze_lock); 192 } 193 194 void blk_mq_unfreeze_queue(struct request_queue *q) 195 { 196 __blk_mq_unfreeze_queue(q, false); 197 } 198 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue); 199 200 /* 201 * FIXME: replace the scsi_internal_device_*block_nowait() calls in the 202 * mpt3sas driver such that this function can be removed. 203 */ 204 void blk_mq_quiesce_queue_nowait(struct request_queue *q) 205 { 206 unsigned long flags; 207 208 spin_lock_irqsave(&q->queue_lock, flags); 209 if (!q->quiesce_depth++) 210 blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q); 211 spin_unlock_irqrestore(&q->queue_lock, flags); 212 } 213 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait); 214 215 /** 216 * blk_mq_wait_quiesce_done() - wait until in-progress quiesce is done 217 * @set: tag_set to wait on 218 * 219 * Note: it is driver's responsibility for making sure that quiesce has 220 * been started on or more of the request_queues of the tag_set. This 221 * function only waits for the quiesce on those request_queues that had 222 * the quiesce flag set using blk_mq_quiesce_queue_nowait. 223 */ 224 void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set) 225 { 226 if (set->flags & BLK_MQ_F_BLOCKING) 227 synchronize_srcu(set->srcu); 228 else 229 synchronize_rcu(); 230 } 231 EXPORT_SYMBOL_GPL(blk_mq_wait_quiesce_done); 232 233 /** 234 * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished 235 * @q: request queue. 236 * 237 * Note: this function does not prevent that the struct request end_io() 238 * callback function is invoked. Once this function is returned, we make 239 * sure no dispatch can happen until the queue is unquiesced via 240 * blk_mq_unquiesce_queue(). 241 */ 242 void blk_mq_quiesce_queue(struct request_queue *q) 243 { 244 blk_mq_quiesce_queue_nowait(q); 245 /* nothing to wait for non-mq queues */ 246 if (queue_is_mq(q)) 247 blk_mq_wait_quiesce_done(q->tag_set); 248 } 249 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue); 250 251 /* 252 * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue() 253 * @q: request queue. 254 * 255 * This function recovers queue into the state before quiescing 256 * which is done by blk_mq_quiesce_queue. 257 */ 258 void blk_mq_unquiesce_queue(struct request_queue *q) 259 { 260 unsigned long flags; 261 bool run_queue = false; 262 263 spin_lock_irqsave(&q->queue_lock, flags); 264 if (WARN_ON_ONCE(q->quiesce_depth <= 0)) { 265 ; 266 } else if (!--q->quiesce_depth) { 267 blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q); 268 run_queue = true; 269 } 270 spin_unlock_irqrestore(&q->queue_lock, flags); 271 272 /* dispatch requests which are inserted during quiescing */ 273 if (run_queue) 274 blk_mq_run_hw_queues(q, true); 275 } 276 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue); 277 278 void blk_mq_quiesce_tagset(struct blk_mq_tag_set *set) 279 { 280 struct request_queue *q; 281 282 mutex_lock(&set->tag_list_lock); 283 list_for_each_entry(q, &set->tag_list, tag_set_list) { 284 if (!blk_queue_skip_tagset_quiesce(q)) 285 blk_mq_quiesce_queue_nowait(q); 286 } 287 blk_mq_wait_quiesce_done(set); 288 mutex_unlock(&set->tag_list_lock); 289 } 290 EXPORT_SYMBOL_GPL(blk_mq_quiesce_tagset); 291 292 void blk_mq_unquiesce_tagset(struct blk_mq_tag_set *set) 293 { 294 struct request_queue *q; 295 296 mutex_lock(&set->tag_list_lock); 297 list_for_each_entry(q, &set->tag_list, tag_set_list) { 298 if (!blk_queue_skip_tagset_quiesce(q)) 299 blk_mq_unquiesce_queue(q); 300 } 301 mutex_unlock(&set->tag_list_lock); 302 } 303 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_tagset); 304 305 void blk_mq_wake_waiters(struct request_queue *q) 306 { 307 struct blk_mq_hw_ctx *hctx; 308 unsigned long i; 309 310 queue_for_each_hw_ctx(q, hctx, i) 311 if (blk_mq_hw_queue_mapped(hctx)) 312 blk_mq_tag_wakeup_all(hctx->tags, true); 313 } 314 315 void blk_rq_init(struct request_queue *q, struct request *rq) 316 { 317 memset(rq, 0, sizeof(*rq)); 318 319 INIT_LIST_HEAD(&rq->queuelist); 320 rq->q = q; 321 rq->__sector = (sector_t) -1; 322 INIT_HLIST_NODE(&rq->hash); 323 RB_CLEAR_NODE(&rq->rb_node); 324 rq->tag = BLK_MQ_NO_TAG; 325 rq->internal_tag = BLK_MQ_NO_TAG; 326 rq->start_time_ns = ktime_get_ns(); 327 rq->part = NULL; 328 blk_crypto_rq_set_defaults(rq); 329 } 330 EXPORT_SYMBOL(blk_rq_init); 331 332 /* Set start and alloc time when the allocated request is actually used */ 333 static inline void blk_mq_rq_time_init(struct request *rq, u64 alloc_time_ns) 334 { 335 if (blk_mq_need_time_stamp(rq)) 336 rq->start_time_ns = ktime_get_ns(); 337 else 338 rq->start_time_ns = 0; 339 340 #ifdef CONFIG_BLK_RQ_ALLOC_TIME 341 if (blk_queue_rq_alloc_time(rq->q)) 342 rq->alloc_time_ns = alloc_time_ns ?: rq->start_time_ns; 343 else 344 rq->alloc_time_ns = 0; 345 #endif 346 } 347 348 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data, 349 struct blk_mq_tags *tags, unsigned int tag) 350 { 351 struct blk_mq_ctx *ctx = data->ctx; 352 struct blk_mq_hw_ctx *hctx = data->hctx; 353 struct request_queue *q = data->q; 354 struct request *rq = tags->static_rqs[tag]; 355 356 rq->q = q; 357 rq->mq_ctx = ctx; 358 rq->mq_hctx = hctx; 359 rq->cmd_flags = data->cmd_flags; 360 361 if (data->flags & BLK_MQ_REQ_PM) 362 data->rq_flags |= RQF_PM; 363 if (blk_queue_io_stat(q)) 364 data->rq_flags |= RQF_IO_STAT; 365 rq->rq_flags = data->rq_flags; 366 367 if (data->rq_flags & RQF_SCHED_TAGS) { 368 rq->tag = BLK_MQ_NO_TAG; 369 rq->internal_tag = tag; 370 } else { 371 rq->tag = tag; 372 rq->internal_tag = BLK_MQ_NO_TAG; 373 } 374 rq->timeout = 0; 375 376 rq->part = NULL; 377 rq->io_start_time_ns = 0; 378 rq->stats_sectors = 0; 379 rq->nr_phys_segments = 0; 380 #if defined(CONFIG_BLK_DEV_INTEGRITY) 381 rq->nr_integrity_segments = 0; 382 #endif 383 rq->end_io = NULL; 384 rq->end_io_data = NULL; 385 386 blk_crypto_rq_set_defaults(rq); 387 INIT_LIST_HEAD(&rq->queuelist); 388 /* tag was already set */ 389 WRITE_ONCE(rq->deadline, 0); 390 req_ref_set(rq, 1); 391 392 if (rq->rq_flags & RQF_USE_SCHED) { 393 struct elevator_queue *e = data->q->elevator; 394 395 INIT_HLIST_NODE(&rq->hash); 396 RB_CLEAR_NODE(&rq->rb_node); 397 398 if (e->type->ops.prepare_request) 399 e->type->ops.prepare_request(rq); 400 } 401 402 return rq; 403 } 404 405 static inline struct request * 406 __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data) 407 { 408 unsigned int tag, tag_offset; 409 struct blk_mq_tags *tags; 410 struct request *rq; 411 unsigned long tag_mask; 412 int i, nr = 0; 413 414 tag_mask = blk_mq_get_tags(data, data->nr_tags, &tag_offset); 415 if (unlikely(!tag_mask)) 416 return NULL; 417 418 tags = blk_mq_tags_from_data(data); 419 for (i = 0; tag_mask; i++) { 420 if (!(tag_mask & (1UL << i))) 421 continue; 422 tag = tag_offset + i; 423 prefetch(tags->static_rqs[tag]); 424 tag_mask &= ~(1UL << i); 425 rq = blk_mq_rq_ctx_init(data, tags, tag); 426 rq_list_add(data->cached_rq, rq); 427 nr++; 428 } 429 if (!(data->rq_flags & RQF_SCHED_TAGS)) 430 blk_mq_add_active_requests(data->hctx, nr); 431 /* caller already holds a reference, add for remainder */ 432 percpu_ref_get_many(&data->q->q_usage_counter, nr - 1); 433 data->nr_tags -= nr; 434 435 return rq_list_pop(data->cached_rq); 436 } 437 438 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) 439 { 440 struct request_queue *q = data->q; 441 u64 alloc_time_ns = 0; 442 struct request *rq; 443 unsigned int tag; 444 445 /* alloc_time includes depth and tag waits */ 446 if (blk_queue_rq_alloc_time(q)) 447 alloc_time_ns = ktime_get_ns(); 448 449 if (data->cmd_flags & REQ_NOWAIT) 450 data->flags |= BLK_MQ_REQ_NOWAIT; 451 452 if (q->elevator) { 453 /* 454 * All requests use scheduler tags when an I/O scheduler is 455 * enabled for the queue. 456 */ 457 data->rq_flags |= RQF_SCHED_TAGS; 458 459 /* 460 * Flush/passthrough requests are special and go directly to the 461 * dispatch list. 462 */ 463 if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH && 464 !blk_op_is_passthrough(data->cmd_flags)) { 465 struct elevator_mq_ops *ops = &q->elevator->type->ops; 466 467 WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED); 468 469 data->rq_flags |= RQF_USE_SCHED; 470 if (ops->limit_depth) 471 ops->limit_depth(data->cmd_flags, data); 472 } 473 } 474 475 retry: 476 data->ctx = blk_mq_get_ctx(q); 477 data->hctx = blk_mq_map_queue(q, data->cmd_flags, data->ctx); 478 if (!(data->rq_flags & RQF_SCHED_TAGS)) 479 blk_mq_tag_busy(data->hctx); 480 481 if (data->flags & BLK_MQ_REQ_RESERVED) 482 data->rq_flags |= RQF_RESV; 483 484 /* 485 * Try batched alloc if we want more than 1 tag. 486 */ 487 if (data->nr_tags > 1) { 488 rq = __blk_mq_alloc_requests_batch(data); 489 if (rq) { 490 blk_mq_rq_time_init(rq, alloc_time_ns); 491 return rq; 492 } 493 data->nr_tags = 1; 494 } 495 496 /* 497 * Waiting allocations only fail because of an inactive hctx. In that 498 * case just retry the hctx assignment and tag allocation as CPU hotplug 499 * should have migrated us to an online CPU by now. 500 */ 501 tag = blk_mq_get_tag(data); 502 if (tag == BLK_MQ_NO_TAG) { 503 if (data->flags & BLK_MQ_REQ_NOWAIT) 504 return NULL; 505 /* 506 * Give up the CPU and sleep for a random short time to 507 * ensure that thread using a realtime scheduling class 508 * are migrated off the CPU, and thus off the hctx that 509 * is going away. 510 */ 511 msleep(3); 512 goto retry; 513 } 514 515 if (!(data->rq_flags & RQF_SCHED_TAGS)) 516 blk_mq_inc_active_requests(data->hctx); 517 rq = blk_mq_rq_ctx_init(data, blk_mq_tags_from_data(data), tag); 518 blk_mq_rq_time_init(rq, alloc_time_ns); 519 return rq; 520 } 521 522 static struct request *blk_mq_rq_cache_fill(struct request_queue *q, 523 struct blk_plug *plug, 524 blk_opf_t opf, 525 blk_mq_req_flags_t flags) 526 { 527 struct blk_mq_alloc_data data = { 528 .q = q, 529 .flags = flags, 530 .cmd_flags = opf, 531 .nr_tags = plug->nr_ios, 532 .cached_rq = &plug->cached_rq, 533 }; 534 struct request *rq; 535 536 if (blk_queue_enter(q, flags)) 537 return NULL; 538 539 plug->nr_ios = 1; 540 541 rq = __blk_mq_alloc_requests(&data); 542 if (unlikely(!rq)) 543 blk_queue_exit(q); 544 return rq; 545 } 546 547 static struct request *blk_mq_alloc_cached_request(struct request_queue *q, 548 blk_opf_t opf, 549 blk_mq_req_flags_t flags) 550 { 551 struct blk_plug *plug = current->plug; 552 struct request *rq; 553 554 if (!plug) 555 return NULL; 556 557 if (rq_list_empty(plug->cached_rq)) { 558 if (plug->nr_ios == 1) 559 return NULL; 560 rq = blk_mq_rq_cache_fill(q, plug, opf, flags); 561 if (!rq) 562 return NULL; 563 } else { 564 rq = rq_list_peek(&plug->cached_rq); 565 if (!rq || rq->q != q) 566 return NULL; 567 568 if (blk_mq_get_hctx_type(opf) != rq->mq_hctx->type) 569 return NULL; 570 if (op_is_flush(rq->cmd_flags) != op_is_flush(opf)) 571 return NULL; 572 573 plug->cached_rq = rq_list_next(rq); 574 blk_mq_rq_time_init(rq, 0); 575 } 576 577 rq->cmd_flags = opf; 578 INIT_LIST_HEAD(&rq->queuelist); 579 return rq; 580 } 581 582 struct request *blk_mq_alloc_request(struct request_queue *q, blk_opf_t opf, 583 blk_mq_req_flags_t flags) 584 { 585 struct request *rq; 586 587 rq = blk_mq_alloc_cached_request(q, opf, flags); 588 if (!rq) { 589 struct blk_mq_alloc_data data = { 590 .q = q, 591 .flags = flags, 592 .cmd_flags = opf, 593 .nr_tags = 1, 594 }; 595 int ret; 596 597 ret = blk_queue_enter(q, flags); 598 if (ret) 599 return ERR_PTR(ret); 600 601 rq = __blk_mq_alloc_requests(&data); 602 if (!rq) 603 goto out_queue_exit; 604 } 605 rq->__data_len = 0; 606 rq->__sector = (sector_t) -1; 607 rq->bio = rq->biotail = NULL; 608 return rq; 609 out_queue_exit: 610 blk_queue_exit(q); 611 return ERR_PTR(-EWOULDBLOCK); 612 } 613 EXPORT_SYMBOL(blk_mq_alloc_request); 614 615 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, 616 blk_opf_t opf, blk_mq_req_flags_t flags, unsigned int hctx_idx) 617 { 618 struct blk_mq_alloc_data data = { 619 .q = q, 620 .flags = flags, 621 .cmd_flags = opf, 622 .nr_tags = 1, 623 }; 624 u64 alloc_time_ns = 0; 625 struct request *rq; 626 unsigned int cpu; 627 unsigned int tag; 628 int ret; 629 630 /* alloc_time includes depth and tag waits */ 631 if (blk_queue_rq_alloc_time(q)) 632 alloc_time_ns = ktime_get_ns(); 633 634 /* 635 * If the tag allocator sleeps we could get an allocation for a 636 * different hardware context. No need to complicate the low level 637 * allocator for this for the rare use case of a command tied to 638 * a specific queue. 639 */ 640 if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)) || 641 WARN_ON_ONCE(!(flags & BLK_MQ_REQ_RESERVED))) 642 return ERR_PTR(-EINVAL); 643 644 if (hctx_idx >= q->nr_hw_queues) 645 return ERR_PTR(-EIO); 646 647 ret = blk_queue_enter(q, flags); 648 if (ret) 649 return ERR_PTR(ret); 650 651 /* 652 * Check if the hardware context is actually mapped to anything. 653 * If not tell the caller that it should skip this queue. 654 */ 655 ret = -EXDEV; 656 data.hctx = xa_load(&q->hctx_table, hctx_idx); 657 if (!blk_mq_hw_queue_mapped(data.hctx)) 658 goto out_queue_exit; 659 cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask); 660 if (cpu >= nr_cpu_ids) 661 goto out_queue_exit; 662 data.ctx = __blk_mq_get_ctx(q, cpu); 663 664 if (q->elevator) 665 data.rq_flags |= RQF_SCHED_TAGS; 666 else 667 blk_mq_tag_busy(data.hctx); 668 669 if (flags & BLK_MQ_REQ_RESERVED) 670 data.rq_flags |= RQF_RESV; 671 672 ret = -EWOULDBLOCK; 673 tag = blk_mq_get_tag(&data); 674 if (tag == BLK_MQ_NO_TAG) 675 goto out_queue_exit; 676 if (!(data.rq_flags & RQF_SCHED_TAGS)) 677 blk_mq_inc_active_requests(data.hctx); 678 rq = blk_mq_rq_ctx_init(&data, blk_mq_tags_from_data(&data), tag); 679 blk_mq_rq_time_init(rq, alloc_time_ns); 680 rq->__data_len = 0; 681 rq->__sector = (sector_t) -1; 682 rq->bio = rq->biotail = NULL; 683 return rq; 684 685 out_queue_exit: 686 blk_queue_exit(q); 687 return ERR_PTR(ret); 688 } 689 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx); 690 691 static void blk_mq_finish_request(struct request *rq) 692 { 693 struct request_queue *q = rq->q; 694 695 if (rq->rq_flags & RQF_USE_SCHED) { 696 q->elevator->type->ops.finish_request(rq); 697 /* 698 * For postflush request that may need to be 699 * completed twice, we should clear this flag 700 * to avoid double finish_request() on the rq. 701 */ 702 rq->rq_flags &= ~RQF_USE_SCHED; 703 } 704 } 705 706 static void __blk_mq_free_request(struct request *rq) 707 { 708 struct request_queue *q = rq->q; 709 struct blk_mq_ctx *ctx = rq->mq_ctx; 710 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 711 const int sched_tag = rq->internal_tag; 712 713 blk_crypto_free_request(rq); 714 blk_pm_mark_last_busy(rq); 715 rq->mq_hctx = NULL; 716 717 if (rq->tag != BLK_MQ_NO_TAG) { 718 blk_mq_dec_active_requests(hctx); 719 blk_mq_put_tag(hctx->tags, ctx, rq->tag); 720 } 721 if (sched_tag != BLK_MQ_NO_TAG) 722 blk_mq_put_tag(hctx->sched_tags, ctx, sched_tag); 723 blk_mq_sched_restart(hctx); 724 blk_queue_exit(q); 725 } 726 727 void blk_mq_free_request(struct request *rq) 728 { 729 struct request_queue *q = rq->q; 730 731 blk_mq_finish_request(rq); 732 733 if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq))) 734 laptop_io_completion(q->disk->bdi); 735 736 rq_qos_done(q, rq); 737 738 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 739 if (req_ref_put_and_test(rq)) 740 __blk_mq_free_request(rq); 741 } 742 EXPORT_SYMBOL_GPL(blk_mq_free_request); 743 744 void blk_mq_free_plug_rqs(struct blk_plug *plug) 745 { 746 struct request *rq; 747 748 while ((rq = rq_list_pop(&plug->cached_rq)) != NULL) 749 blk_mq_free_request(rq); 750 } 751 752 void blk_dump_rq_flags(struct request *rq, char *msg) 753 { 754 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, 755 rq->q->disk ? rq->q->disk->disk_name : "?", 756 (__force unsigned long long) rq->cmd_flags); 757 758 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", 759 (unsigned long long)blk_rq_pos(rq), 760 blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); 761 printk(KERN_INFO " bio %p, biotail %p, len %u\n", 762 rq->bio, rq->biotail, blk_rq_bytes(rq)); 763 } 764 EXPORT_SYMBOL(blk_dump_rq_flags); 765 766 static void req_bio_endio(struct request *rq, struct bio *bio, 767 unsigned int nbytes, blk_status_t error) 768 { 769 if (unlikely(error)) { 770 bio->bi_status = error; 771 } else if (req_op(rq) == REQ_OP_ZONE_APPEND) { 772 /* 773 * Partial zone append completions cannot be supported as the 774 * BIO fragments may end up not being written sequentially. 775 */ 776 if (bio->bi_iter.bi_size != nbytes) 777 bio->bi_status = BLK_STS_IOERR; 778 else 779 bio->bi_iter.bi_sector = rq->__sector; 780 } 781 782 bio_advance(bio, nbytes); 783 784 if (unlikely(rq->rq_flags & RQF_QUIET)) 785 bio_set_flag(bio, BIO_QUIET); 786 /* don't actually finish bio if it's part of flush sequence */ 787 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) 788 bio_endio(bio); 789 } 790 791 static void blk_account_io_completion(struct request *req, unsigned int bytes) 792 { 793 if (req->part && blk_do_io_stat(req)) { 794 const int sgrp = op_stat_group(req_op(req)); 795 796 part_stat_lock(); 797 part_stat_add(req->part, sectors[sgrp], bytes >> 9); 798 part_stat_unlock(); 799 } 800 } 801 802 static void blk_print_req_error(struct request *req, blk_status_t status) 803 { 804 printk_ratelimited(KERN_ERR 805 "%s error, dev %s, sector %llu op 0x%x:(%s) flags 0x%x " 806 "phys_seg %u prio class %u\n", 807 blk_status_to_str(status), 808 req->q->disk ? req->q->disk->disk_name : "?", 809 blk_rq_pos(req), (__force u32)req_op(req), 810 blk_op_str(req_op(req)), 811 (__force u32)(req->cmd_flags & ~REQ_OP_MASK), 812 req->nr_phys_segments, 813 IOPRIO_PRIO_CLASS(req->ioprio)); 814 } 815 816 /* 817 * Fully end IO on a request. Does not support partial completions, or 818 * errors. 819 */ 820 static void blk_complete_request(struct request *req) 821 { 822 const bool is_flush = (req->rq_flags & RQF_FLUSH_SEQ) != 0; 823 int total_bytes = blk_rq_bytes(req); 824 struct bio *bio = req->bio; 825 826 trace_block_rq_complete(req, BLK_STS_OK, total_bytes); 827 828 if (!bio) 829 return; 830 831 #ifdef CONFIG_BLK_DEV_INTEGRITY 832 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ) 833 req->q->integrity.profile->complete_fn(req, total_bytes); 834 #endif 835 836 /* 837 * Upper layers may call blk_crypto_evict_key() anytime after the last 838 * bio_endio(). Therefore, the keyslot must be released before that. 839 */ 840 blk_crypto_rq_put_keyslot(req); 841 842 blk_account_io_completion(req, total_bytes); 843 844 do { 845 struct bio *next = bio->bi_next; 846 847 /* Completion has already been traced */ 848 bio_clear_flag(bio, BIO_TRACE_COMPLETION); 849 850 if (req_op(req) == REQ_OP_ZONE_APPEND) 851 bio->bi_iter.bi_sector = req->__sector; 852 853 if (!is_flush) 854 bio_endio(bio); 855 bio = next; 856 } while (bio); 857 858 /* 859 * Reset counters so that the request stacking driver 860 * can find how many bytes remain in the request 861 * later. 862 */ 863 if (!req->end_io) { 864 req->bio = NULL; 865 req->__data_len = 0; 866 } 867 } 868 869 /** 870 * blk_update_request - Complete multiple bytes without completing the request 871 * @req: the request being processed 872 * @error: block status code 873 * @nr_bytes: number of bytes to complete for @req 874 * 875 * Description: 876 * Ends I/O on a number of bytes attached to @req, but doesn't complete 877 * the request structure even if @req doesn't have leftover. 878 * If @req has leftover, sets it up for the next range of segments. 879 * 880 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees 881 * %false return from this function. 882 * 883 * Note: 884 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function 885 * except in the consistency check at the end of this function. 886 * 887 * Return: 888 * %false - this request doesn't have any more data 889 * %true - this request has more data 890 **/ 891 bool blk_update_request(struct request *req, blk_status_t error, 892 unsigned int nr_bytes) 893 { 894 int total_bytes; 895 896 trace_block_rq_complete(req, error, nr_bytes); 897 898 if (!req->bio) 899 return false; 900 901 #ifdef CONFIG_BLK_DEV_INTEGRITY 902 if (blk_integrity_rq(req) && req_op(req) == REQ_OP_READ && 903 error == BLK_STS_OK) 904 req->q->integrity.profile->complete_fn(req, nr_bytes); 905 #endif 906 907 /* 908 * Upper layers may call blk_crypto_evict_key() anytime after the last 909 * bio_endio(). Therefore, the keyslot must be released before that. 910 */ 911 if (blk_crypto_rq_has_keyslot(req) && nr_bytes >= blk_rq_bytes(req)) 912 __blk_crypto_rq_put_keyslot(req); 913 914 if (unlikely(error && !blk_rq_is_passthrough(req) && 915 !(req->rq_flags & RQF_QUIET)) && 916 !test_bit(GD_DEAD, &req->q->disk->state)) { 917 blk_print_req_error(req, error); 918 trace_block_rq_error(req, error, nr_bytes); 919 } 920 921 blk_account_io_completion(req, nr_bytes); 922 923 total_bytes = 0; 924 while (req->bio) { 925 struct bio *bio = req->bio; 926 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); 927 928 if (bio_bytes == bio->bi_iter.bi_size) 929 req->bio = bio->bi_next; 930 931 /* Completion has already been traced */ 932 bio_clear_flag(bio, BIO_TRACE_COMPLETION); 933 req_bio_endio(req, bio, bio_bytes, error); 934 935 total_bytes += bio_bytes; 936 nr_bytes -= bio_bytes; 937 938 if (!nr_bytes) 939 break; 940 } 941 942 /* 943 * completely done 944 */ 945 if (!req->bio) { 946 /* 947 * Reset counters so that the request stacking driver 948 * can find how many bytes remain in the request 949 * later. 950 */ 951 req->__data_len = 0; 952 return false; 953 } 954 955 req->__data_len -= total_bytes; 956 957 /* update sector only for requests with clear definition of sector */ 958 if (!blk_rq_is_passthrough(req)) 959 req->__sector += total_bytes >> 9; 960 961 /* mixed attributes always follow the first bio */ 962 if (req->rq_flags & RQF_MIXED_MERGE) { 963 req->cmd_flags &= ~REQ_FAILFAST_MASK; 964 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; 965 } 966 967 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { 968 /* 969 * If total number of sectors is less than the first segment 970 * size, something has gone terribly wrong. 971 */ 972 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { 973 blk_dump_rq_flags(req, "request botched"); 974 req->__data_len = blk_rq_cur_bytes(req); 975 } 976 977 /* recalculate the number of segments */ 978 req->nr_phys_segments = blk_recalc_rq_segments(req); 979 } 980 981 return true; 982 } 983 EXPORT_SYMBOL_GPL(blk_update_request); 984 985 static inline void blk_account_io_done(struct request *req, u64 now) 986 { 987 trace_block_io_done(req); 988 989 /* 990 * Account IO completion. flush_rq isn't accounted as a 991 * normal IO on queueing nor completion. Accounting the 992 * containing request is enough. 993 */ 994 if (blk_do_io_stat(req) && req->part && 995 !(req->rq_flags & RQF_FLUSH_SEQ)) { 996 const int sgrp = op_stat_group(req_op(req)); 997 998 part_stat_lock(); 999 update_io_ticks(req->part, jiffies, true); 1000 part_stat_inc(req->part, ios[sgrp]); 1001 part_stat_add(req->part, nsecs[sgrp], now - req->start_time_ns); 1002 part_stat_unlock(); 1003 } 1004 } 1005 1006 static inline void blk_account_io_start(struct request *req) 1007 { 1008 trace_block_io_start(req); 1009 1010 if (blk_do_io_stat(req)) { 1011 /* 1012 * All non-passthrough requests are created from a bio with one 1013 * exception: when a flush command that is part of a flush sequence 1014 * generated by the state machine in blk-flush.c is cloned onto the 1015 * lower device by dm-multipath we can get here without a bio. 1016 */ 1017 if (req->bio) 1018 req->part = req->bio->bi_bdev; 1019 else 1020 req->part = req->q->disk->part0; 1021 1022 part_stat_lock(); 1023 update_io_ticks(req->part, jiffies, false); 1024 part_stat_unlock(); 1025 } 1026 } 1027 1028 static inline void __blk_mq_end_request_acct(struct request *rq, u64 now) 1029 { 1030 if (rq->rq_flags & RQF_STATS) 1031 blk_stat_add(rq, now); 1032 1033 blk_mq_sched_completed_request(rq, now); 1034 blk_account_io_done(rq, now); 1035 } 1036 1037 inline void __blk_mq_end_request(struct request *rq, blk_status_t error) 1038 { 1039 if (blk_mq_need_time_stamp(rq)) 1040 __blk_mq_end_request_acct(rq, ktime_get_ns()); 1041 1042 blk_mq_finish_request(rq); 1043 1044 if (rq->end_io) { 1045 rq_qos_done(rq->q, rq); 1046 if (rq->end_io(rq, error) == RQ_END_IO_FREE) 1047 blk_mq_free_request(rq); 1048 } else { 1049 blk_mq_free_request(rq); 1050 } 1051 } 1052 EXPORT_SYMBOL(__blk_mq_end_request); 1053 1054 void blk_mq_end_request(struct request *rq, blk_status_t error) 1055 { 1056 if (blk_update_request(rq, error, blk_rq_bytes(rq))) 1057 BUG(); 1058 __blk_mq_end_request(rq, error); 1059 } 1060 EXPORT_SYMBOL(blk_mq_end_request); 1061 1062 #define TAG_COMP_BATCH 32 1063 1064 static inline void blk_mq_flush_tag_batch(struct blk_mq_hw_ctx *hctx, 1065 int *tag_array, int nr_tags) 1066 { 1067 struct request_queue *q = hctx->queue; 1068 1069 blk_mq_sub_active_requests(hctx, nr_tags); 1070 1071 blk_mq_put_tags(hctx->tags, tag_array, nr_tags); 1072 percpu_ref_put_many(&q->q_usage_counter, nr_tags); 1073 } 1074 1075 void blk_mq_end_request_batch(struct io_comp_batch *iob) 1076 { 1077 int tags[TAG_COMP_BATCH], nr_tags = 0; 1078 struct blk_mq_hw_ctx *cur_hctx = NULL; 1079 struct request *rq; 1080 u64 now = 0; 1081 1082 if (iob->need_ts) 1083 now = ktime_get_ns(); 1084 1085 while ((rq = rq_list_pop(&iob->req_list)) != NULL) { 1086 prefetch(rq->bio); 1087 prefetch(rq->rq_next); 1088 1089 blk_complete_request(rq); 1090 if (iob->need_ts) 1091 __blk_mq_end_request_acct(rq, now); 1092 1093 blk_mq_finish_request(rq); 1094 1095 rq_qos_done(rq->q, rq); 1096 1097 /* 1098 * If end_io handler returns NONE, then it still has 1099 * ownership of the request. 1100 */ 1101 if (rq->end_io && rq->end_io(rq, 0) == RQ_END_IO_NONE) 1102 continue; 1103 1104 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 1105 if (!req_ref_put_and_test(rq)) 1106 continue; 1107 1108 blk_crypto_free_request(rq); 1109 blk_pm_mark_last_busy(rq); 1110 1111 if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) { 1112 if (cur_hctx) 1113 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); 1114 nr_tags = 0; 1115 cur_hctx = rq->mq_hctx; 1116 } 1117 tags[nr_tags++] = rq->tag; 1118 } 1119 1120 if (nr_tags) 1121 blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags); 1122 } 1123 EXPORT_SYMBOL_GPL(blk_mq_end_request_batch); 1124 1125 static void blk_complete_reqs(struct llist_head *list) 1126 { 1127 struct llist_node *entry = llist_reverse_order(llist_del_all(list)); 1128 struct request *rq, *next; 1129 1130 llist_for_each_entry_safe(rq, next, entry, ipi_list) 1131 rq->q->mq_ops->complete(rq); 1132 } 1133 1134 static __latent_entropy void blk_done_softirq(struct softirq_action *h) 1135 { 1136 blk_complete_reqs(this_cpu_ptr(&blk_cpu_done)); 1137 } 1138 1139 static int blk_softirq_cpu_dead(unsigned int cpu) 1140 { 1141 blk_complete_reqs(&per_cpu(blk_cpu_done, cpu)); 1142 return 0; 1143 } 1144 1145 static void __blk_mq_complete_request_remote(void *data) 1146 { 1147 __raise_softirq_irqoff(BLOCK_SOFTIRQ); 1148 } 1149 1150 static inline bool blk_mq_complete_need_ipi(struct request *rq) 1151 { 1152 int cpu = raw_smp_processor_id(); 1153 1154 if (!IS_ENABLED(CONFIG_SMP) || 1155 !test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) 1156 return false; 1157 /* 1158 * With force threaded interrupts enabled, raising softirq from an SMP 1159 * function call will always result in waking the ksoftirqd thread. 1160 * This is probably worse than completing the request on a different 1161 * cache domain. 1162 */ 1163 if (force_irqthreads()) 1164 return false; 1165 1166 /* same CPU or cache domain? Complete locally */ 1167 if (cpu == rq->mq_ctx->cpu || 1168 (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags) && 1169 cpus_share_cache(cpu, rq->mq_ctx->cpu))) 1170 return false; 1171 1172 /* don't try to IPI to an offline CPU */ 1173 return cpu_online(rq->mq_ctx->cpu); 1174 } 1175 1176 static void blk_mq_complete_send_ipi(struct request *rq) 1177 { 1178 unsigned int cpu; 1179 1180 cpu = rq->mq_ctx->cpu; 1181 if (llist_add(&rq->ipi_list, &per_cpu(blk_cpu_done, cpu))) 1182 smp_call_function_single_async(cpu, &per_cpu(blk_cpu_csd, cpu)); 1183 } 1184 1185 static void blk_mq_raise_softirq(struct request *rq) 1186 { 1187 struct llist_head *list; 1188 1189 preempt_disable(); 1190 list = this_cpu_ptr(&blk_cpu_done); 1191 if (llist_add(&rq->ipi_list, list)) 1192 raise_softirq(BLOCK_SOFTIRQ); 1193 preempt_enable(); 1194 } 1195 1196 bool blk_mq_complete_request_remote(struct request *rq) 1197 { 1198 WRITE_ONCE(rq->state, MQ_RQ_COMPLETE); 1199 1200 /* 1201 * For request which hctx has only one ctx mapping, 1202 * or a polled request, always complete locally, 1203 * it's pointless to redirect the completion. 1204 */ 1205 if ((rq->mq_hctx->nr_ctx == 1 && 1206 rq->mq_ctx->cpu == raw_smp_processor_id()) || 1207 rq->cmd_flags & REQ_POLLED) 1208 return false; 1209 1210 if (blk_mq_complete_need_ipi(rq)) { 1211 blk_mq_complete_send_ipi(rq); 1212 return true; 1213 } 1214 1215 if (rq->q->nr_hw_queues == 1) { 1216 blk_mq_raise_softirq(rq); 1217 return true; 1218 } 1219 return false; 1220 } 1221 EXPORT_SYMBOL_GPL(blk_mq_complete_request_remote); 1222 1223 /** 1224 * blk_mq_complete_request - end I/O on a request 1225 * @rq: the request being processed 1226 * 1227 * Description: 1228 * Complete a request by scheduling the ->complete_rq operation. 1229 **/ 1230 void blk_mq_complete_request(struct request *rq) 1231 { 1232 if (!blk_mq_complete_request_remote(rq)) 1233 rq->q->mq_ops->complete(rq); 1234 } 1235 EXPORT_SYMBOL(blk_mq_complete_request); 1236 1237 /** 1238 * blk_mq_start_request - Start processing a request 1239 * @rq: Pointer to request to be started 1240 * 1241 * Function used by device drivers to notify the block layer that a request 1242 * is going to be processed now, so blk layer can do proper initializations 1243 * such as starting the timeout timer. 1244 */ 1245 void blk_mq_start_request(struct request *rq) 1246 { 1247 struct request_queue *q = rq->q; 1248 1249 trace_block_rq_issue(rq); 1250 1251 if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags) && 1252 !blk_rq_is_passthrough(rq)) { 1253 rq->io_start_time_ns = ktime_get_ns(); 1254 rq->stats_sectors = blk_rq_sectors(rq); 1255 rq->rq_flags |= RQF_STATS; 1256 rq_qos_issue(q, rq); 1257 } 1258 1259 WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE); 1260 1261 blk_add_timer(rq); 1262 WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT); 1263 rq->mq_hctx->tags->rqs[rq->tag] = rq; 1264 1265 #ifdef CONFIG_BLK_DEV_INTEGRITY 1266 if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE) 1267 q->integrity.profile->prepare_fn(rq); 1268 #endif 1269 if (rq->bio && rq->bio->bi_opf & REQ_POLLED) 1270 WRITE_ONCE(rq->bio->bi_cookie, rq->mq_hctx->queue_num); 1271 } 1272 EXPORT_SYMBOL(blk_mq_start_request); 1273 1274 /* 1275 * Allow 2x BLK_MAX_REQUEST_COUNT requests on plug queue for multiple 1276 * queues. This is important for md arrays to benefit from merging 1277 * requests. 1278 */ 1279 static inline unsigned short blk_plug_max_rq_count(struct blk_plug *plug) 1280 { 1281 if (plug->multiple_queues) 1282 return BLK_MAX_REQUEST_COUNT * 2; 1283 return BLK_MAX_REQUEST_COUNT; 1284 } 1285 1286 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq) 1287 { 1288 struct request *last = rq_list_peek(&plug->mq_list); 1289 1290 if (!plug->rq_count) { 1291 trace_block_plug(rq->q); 1292 } else if (plug->rq_count >= blk_plug_max_rq_count(plug) || 1293 (!blk_queue_nomerges(rq->q) && 1294 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) { 1295 blk_mq_flush_plug_list(plug, false); 1296 last = NULL; 1297 trace_block_plug(rq->q); 1298 } 1299 1300 if (!plug->multiple_queues && last && last->q != rq->q) 1301 plug->multiple_queues = true; 1302 /* 1303 * Any request allocated from sched tags can't be issued to 1304 * ->queue_rqs() directly 1305 */ 1306 if (!plug->has_elevator && (rq->rq_flags & RQF_SCHED_TAGS)) 1307 plug->has_elevator = true; 1308 rq->rq_next = NULL; 1309 rq_list_add(&plug->mq_list, rq); 1310 plug->rq_count++; 1311 } 1312 1313 /** 1314 * blk_execute_rq_nowait - insert a request to I/O scheduler for execution 1315 * @rq: request to insert 1316 * @at_head: insert request at head or tail of queue 1317 * 1318 * Description: 1319 * Insert a fully prepared request at the back of the I/O scheduler queue 1320 * for execution. Don't wait for completion. 1321 * 1322 * Note: 1323 * This function will invoke @done directly if the queue is dead. 1324 */ 1325 void blk_execute_rq_nowait(struct request *rq, bool at_head) 1326 { 1327 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 1328 1329 WARN_ON(irqs_disabled()); 1330 WARN_ON(!blk_rq_is_passthrough(rq)); 1331 1332 blk_account_io_start(rq); 1333 1334 /* 1335 * As plugging can be enabled for passthrough requests on a zoned 1336 * device, directly accessing the plug instead of using blk_mq_plug() 1337 * should not have any consequences. 1338 */ 1339 if (current->plug && !at_head) { 1340 blk_add_rq_to_plug(current->plug, rq); 1341 return; 1342 } 1343 1344 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); 1345 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); 1346 } 1347 EXPORT_SYMBOL_GPL(blk_execute_rq_nowait); 1348 1349 struct blk_rq_wait { 1350 struct completion done; 1351 blk_status_t ret; 1352 }; 1353 1354 static enum rq_end_io_ret blk_end_sync_rq(struct request *rq, blk_status_t ret) 1355 { 1356 struct blk_rq_wait *wait = rq->end_io_data; 1357 1358 wait->ret = ret; 1359 complete(&wait->done); 1360 return RQ_END_IO_NONE; 1361 } 1362 1363 bool blk_rq_is_poll(struct request *rq) 1364 { 1365 if (!rq->mq_hctx) 1366 return false; 1367 if (rq->mq_hctx->type != HCTX_TYPE_POLL) 1368 return false; 1369 return true; 1370 } 1371 EXPORT_SYMBOL_GPL(blk_rq_is_poll); 1372 1373 static void blk_rq_poll_completion(struct request *rq, struct completion *wait) 1374 { 1375 do { 1376 blk_hctx_poll(rq->q, rq->mq_hctx, NULL, 0); 1377 cond_resched(); 1378 } while (!completion_done(wait)); 1379 } 1380 1381 /** 1382 * blk_execute_rq - insert a request into queue for execution 1383 * @rq: request to insert 1384 * @at_head: insert request at head or tail of queue 1385 * 1386 * Description: 1387 * Insert a fully prepared request at the back of the I/O scheduler queue 1388 * for execution and wait for completion. 1389 * Return: The blk_status_t result provided to blk_mq_end_request(). 1390 */ 1391 blk_status_t blk_execute_rq(struct request *rq, bool at_head) 1392 { 1393 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 1394 struct blk_rq_wait wait = { 1395 .done = COMPLETION_INITIALIZER_ONSTACK(wait.done), 1396 }; 1397 1398 WARN_ON(irqs_disabled()); 1399 WARN_ON(!blk_rq_is_passthrough(rq)); 1400 1401 rq->end_io_data = &wait; 1402 rq->end_io = blk_end_sync_rq; 1403 1404 blk_account_io_start(rq); 1405 blk_mq_insert_request(rq, at_head ? BLK_MQ_INSERT_AT_HEAD : 0); 1406 blk_mq_run_hw_queue(hctx, false); 1407 1408 if (blk_rq_is_poll(rq)) { 1409 blk_rq_poll_completion(rq, &wait.done); 1410 } else { 1411 /* 1412 * Prevent hang_check timer from firing at us during very long 1413 * I/O 1414 */ 1415 unsigned long hang_check = sysctl_hung_task_timeout_secs; 1416 1417 if (hang_check) 1418 while (!wait_for_completion_io_timeout(&wait.done, 1419 hang_check * (HZ/2))) 1420 ; 1421 else 1422 wait_for_completion_io(&wait.done); 1423 } 1424 1425 return wait.ret; 1426 } 1427 EXPORT_SYMBOL(blk_execute_rq); 1428 1429 static void __blk_mq_requeue_request(struct request *rq) 1430 { 1431 struct request_queue *q = rq->q; 1432 1433 blk_mq_put_driver_tag(rq); 1434 1435 trace_block_rq_requeue(rq); 1436 rq_qos_requeue(q, rq); 1437 1438 if (blk_mq_request_started(rq)) { 1439 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 1440 rq->rq_flags &= ~RQF_TIMED_OUT; 1441 } 1442 } 1443 1444 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list) 1445 { 1446 struct request_queue *q = rq->q; 1447 unsigned long flags; 1448 1449 __blk_mq_requeue_request(rq); 1450 1451 /* this request will be re-inserted to io scheduler queue */ 1452 blk_mq_sched_requeue_request(rq); 1453 1454 spin_lock_irqsave(&q->requeue_lock, flags); 1455 list_add_tail(&rq->queuelist, &q->requeue_list); 1456 spin_unlock_irqrestore(&q->requeue_lock, flags); 1457 1458 if (kick_requeue_list) 1459 blk_mq_kick_requeue_list(q); 1460 } 1461 EXPORT_SYMBOL(blk_mq_requeue_request); 1462 1463 static void blk_mq_requeue_work(struct work_struct *work) 1464 { 1465 struct request_queue *q = 1466 container_of(work, struct request_queue, requeue_work.work); 1467 LIST_HEAD(rq_list); 1468 LIST_HEAD(flush_list); 1469 struct request *rq; 1470 1471 spin_lock_irq(&q->requeue_lock); 1472 list_splice_init(&q->requeue_list, &rq_list); 1473 list_splice_init(&q->flush_list, &flush_list); 1474 spin_unlock_irq(&q->requeue_lock); 1475 1476 while (!list_empty(&rq_list)) { 1477 rq = list_entry(rq_list.next, struct request, queuelist); 1478 /* 1479 * If RQF_DONTPREP ist set, the request has been started by the 1480 * driver already and might have driver-specific data allocated 1481 * already. Insert it into the hctx dispatch list to avoid 1482 * block layer merges for the request. 1483 */ 1484 if (rq->rq_flags & RQF_DONTPREP) { 1485 list_del_init(&rq->queuelist); 1486 blk_mq_request_bypass_insert(rq, 0); 1487 } else { 1488 list_del_init(&rq->queuelist); 1489 blk_mq_insert_request(rq, BLK_MQ_INSERT_AT_HEAD); 1490 } 1491 } 1492 1493 while (!list_empty(&flush_list)) { 1494 rq = list_entry(flush_list.next, struct request, queuelist); 1495 list_del_init(&rq->queuelist); 1496 blk_mq_insert_request(rq, 0); 1497 } 1498 1499 blk_mq_run_hw_queues(q, false); 1500 } 1501 1502 void blk_mq_kick_requeue_list(struct request_queue *q) 1503 { 1504 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0); 1505 } 1506 EXPORT_SYMBOL(blk_mq_kick_requeue_list); 1507 1508 void blk_mq_delay_kick_requeue_list(struct request_queue *q, 1509 unsigned long msecs) 1510 { 1511 kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 1512 msecs_to_jiffies(msecs)); 1513 } 1514 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list); 1515 1516 static bool blk_is_flush_data_rq(struct request *rq) 1517 { 1518 return (rq->rq_flags & RQF_FLUSH_SEQ) && !is_flush_rq(rq); 1519 } 1520 1521 static bool blk_mq_rq_inflight(struct request *rq, void *priv) 1522 { 1523 /* 1524 * If we find a request that isn't idle we know the queue is busy 1525 * as it's checked in the iter. 1526 * Return false to stop the iteration. 1527 * 1528 * In case of queue quiesce, if one flush data request is completed, 1529 * don't count it as inflight given the flush sequence is suspended, 1530 * and the original flush data request is invisible to driver, just 1531 * like other pending requests because of quiesce 1532 */ 1533 if (blk_mq_request_started(rq) && !(blk_queue_quiesced(rq->q) && 1534 blk_is_flush_data_rq(rq) && 1535 blk_mq_request_completed(rq))) { 1536 bool *busy = priv; 1537 1538 *busy = true; 1539 return false; 1540 } 1541 1542 return true; 1543 } 1544 1545 bool blk_mq_queue_inflight(struct request_queue *q) 1546 { 1547 bool busy = false; 1548 1549 blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy); 1550 return busy; 1551 } 1552 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight); 1553 1554 static void blk_mq_rq_timed_out(struct request *req) 1555 { 1556 req->rq_flags |= RQF_TIMED_OUT; 1557 if (req->q->mq_ops->timeout) { 1558 enum blk_eh_timer_return ret; 1559 1560 ret = req->q->mq_ops->timeout(req); 1561 if (ret == BLK_EH_DONE) 1562 return; 1563 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER); 1564 } 1565 1566 blk_add_timer(req); 1567 } 1568 1569 struct blk_expired_data { 1570 bool has_timedout_rq; 1571 unsigned long next; 1572 unsigned long timeout_start; 1573 }; 1574 1575 static bool blk_mq_req_expired(struct request *rq, struct blk_expired_data *expired) 1576 { 1577 unsigned long deadline; 1578 1579 if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT) 1580 return false; 1581 if (rq->rq_flags & RQF_TIMED_OUT) 1582 return false; 1583 1584 deadline = READ_ONCE(rq->deadline); 1585 if (time_after_eq(expired->timeout_start, deadline)) 1586 return true; 1587 1588 if (expired->next == 0) 1589 expired->next = deadline; 1590 else if (time_after(expired->next, deadline)) 1591 expired->next = deadline; 1592 return false; 1593 } 1594 1595 void blk_mq_put_rq_ref(struct request *rq) 1596 { 1597 if (is_flush_rq(rq)) { 1598 if (rq->end_io(rq, 0) == RQ_END_IO_FREE) 1599 blk_mq_free_request(rq); 1600 } else if (req_ref_put_and_test(rq)) { 1601 __blk_mq_free_request(rq); 1602 } 1603 } 1604 1605 static bool blk_mq_check_expired(struct request *rq, void *priv) 1606 { 1607 struct blk_expired_data *expired = priv; 1608 1609 /* 1610 * blk_mq_queue_tag_busy_iter() has locked the request, so it cannot 1611 * be reallocated underneath the timeout handler's processing, then 1612 * the expire check is reliable. If the request is not expired, then 1613 * it was completed and reallocated as a new request after returning 1614 * from blk_mq_check_expired(). 1615 */ 1616 if (blk_mq_req_expired(rq, expired)) { 1617 expired->has_timedout_rq = true; 1618 return false; 1619 } 1620 return true; 1621 } 1622 1623 static bool blk_mq_handle_expired(struct request *rq, void *priv) 1624 { 1625 struct blk_expired_data *expired = priv; 1626 1627 if (blk_mq_req_expired(rq, expired)) 1628 blk_mq_rq_timed_out(rq); 1629 return true; 1630 } 1631 1632 static void blk_mq_timeout_work(struct work_struct *work) 1633 { 1634 struct request_queue *q = 1635 container_of(work, struct request_queue, timeout_work); 1636 struct blk_expired_data expired = { 1637 .timeout_start = jiffies, 1638 }; 1639 struct blk_mq_hw_ctx *hctx; 1640 unsigned long i; 1641 1642 /* A deadlock might occur if a request is stuck requiring a 1643 * timeout at the same time a queue freeze is waiting 1644 * completion, since the timeout code would not be able to 1645 * acquire the queue reference here. 1646 * 1647 * That's why we don't use blk_queue_enter here; instead, we use 1648 * percpu_ref_tryget directly, because we need to be able to 1649 * obtain a reference even in the short window between the queue 1650 * starting to freeze, by dropping the first reference in 1651 * blk_freeze_queue_start, and the moment the last request is 1652 * consumed, marked by the instant q_usage_counter reaches 1653 * zero. 1654 */ 1655 if (!percpu_ref_tryget(&q->q_usage_counter)) 1656 return; 1657 1658 /* check if there is any timed-out request */ 1659 blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired); 1660 if (expired.has_timedout_rq) { 1661 /* 1662 * Before walking tags, we must ensure any submit started 1663 * before the current time has finished. Since the submit 1664 * uses srcu or rcu, wait for a synchronization point to 1665 * ensure all running submits have finished 1666 */ 1667 blk_mq_wait_quiesce_done(q->tag_set); 1668 1669 expired.next = 0; 1670 blk_mq_queue_tag_busy_iter(q, blk_mq_handle_expired, &expired); 1671 } 1672 1673 if (expired.next != 0) { 1674 mod_timer(&q->timeout, expired.next); 1675 } else { 1676 /* 1677 * Request timeouts are handled as a forward rolling timer. If 1678 * we end up here it means that no requests are pending and 1679 * also that no request has been pending for a while. Mark 1680 * each hctx as idle. 1681 */ 1682 queue_for_each_hw_ctx(q, hctx, i) { 1683 /* the hctx may be unmapped, so check it here */ 1684 if (blk_mq_hw_queue_mapped(hctx)) 1685 blk_mq_tag_idle(hctx); 1686 } 1687 } 1688 blk_queue_exit(q); 1689 } 1690 1691 struct flush_busy_ctx_data { 1692 struct blk_mq_hw_ctx *hctx; 1693 struct list_head *list; 1694 }; 1695 1696 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data) 1697 { 1698 struct flush_busy_ctx_data *flush_data = data; 1699 struct blk_mq_hw_ctx *hctx = flush_data->hctx; 1700 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; 1701 enum hctx_type type = hctx->type; 1702 1703 spin_lock(&ctx->lock); 1704 list_splice_tail_init(&ctx->rq_lists[type], flush_data->list); 1705 sbitmap_clear_bit(sb, bitnr); 1706 spin_unlock(&ctx->lock); 1707 return true; 1708 } 1709 1710 /* 1711 * Process software queues that have been marked busy, splicing them 1712 * to the for-dispatch 1713 */ 1714 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list) 1715 { 1716 struct flush_busy_ctx_data data = { 1717 .hctx = hctx, 1718 .list = list, 1719 }; 1720 1721 sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data); 1722 } 1723 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs); 1724 1725 struct dispatch_rq_data { 1726 struct blk_mq_hw_ctx *hctx; 1727 struct request *rq; 1728 }; 1729 1730 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr, 1731 void *data) 1732 { 1733 struct dispatch_rq_data *dispatch_data = data; 1734 struct blk_mq_hw_ctx *hctx = dispatch_data->hctx; 1735 struct blk_mq_ctx *ctx = hctx->ctxs[bitnr]; 1736 enum hctx_type type = hctx->type; 1737 1738 spin_lock(&ctx->lock); 1739 if (!list_empty(&ctx->rq_lists[type])) { 1740 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next); 1741 list_del_init(&dispatch_data->rq->queuelist); 1742 if (list_empty(&ctx->rq_lists[type])) 1743 sbitmap_clear_bit(sb, bitnr); 1744 } 1745 spin_unlock(&ctx->lock); 1746 1747 return !dispatch_data->rq; 1748 } 1749 1750 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, 1751 struct blk_mq_ctx *start) 1752 { 1753 unsigned off = start ? start->index_hw[hctx->type] : 0; 1754 struct dispatch_rq_data data = { 1755 .hctx = hctx, 1756 .rq = NULL, 1757 }; 1758 1759 __sbitmap_for_each_set(&hctx->ctx_map, off, 1760 dispatch_rq_from_ctx, &data); 1761 1762 return data.rq; 1763 } 1764 1765 bool __blk_mq_alloc_driver_tag(struct request *rq) 1766 { 1767 struct sbitmap_queue *bt = &rq->mq_hctx->tags->bitmap_tags; 1768 unsigned int tag_offset = rq->mq_hctx->tags->nr_reserved_tags; 1769 int tag; 1770 1771 blk_mq_tag_busy(rq->mq_hctx); 1772 1773 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) { 1774 bt = &rq->mq_hctx->tags->breserved_tags; 1775 tag_offset = 0; 1776 } else { 1777 if (!hctx_may_queue(rq->mq_hctx, bt)) 1778 return false; 1779 } 1780 1781 tag = __sbitmap_queue_get(bt); 1782 if (tag == BLK_MQ_NO_TAG) 1783 return false; 1784 1785 rq->tag = tag + tag_offset; 1786 blk_mq_inc_active_requests(rq->mq_hctx); 1787 return true; 1788 } 1789 1790 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, 1791 int flags, void *key) 1792 { 1793 struct blk_mq_hw_ctx *hctx; 1794 1795 hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait); 1796 1797 spin_lock(&hctx->dispatch_wait_lock); 1798 if (!list_empty(&wait->entry)) { 1799 struct sbitmap_queue *sbq; 1800 1801 list_del_init(&wait->entry); 1802 sbq = &hctx->tags->bitmap_tags; 1803 atomic_dec(&sbq->ws_active); 1804 } 1805 spin_unlock(&hctx->dispatch_wait_lock); 1806 1807 blk_mq_run_hw_queue(hctx, true); 1808 return 1; 1809 } 1810 1811 /* 1812 * Mark us waiting for a tag. For shared tags, this involves hooking us into 1813 * the tag wakeups. For non-shared tags, we can simply mark us needing a 1814 * restart. For both cases, take care to check the condition again after 1815 * marking us as waiting. 1816 */ 1817 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx, 1818 struct request *rq) 1819 { 1820 struct sbitmap_queue *sbq; 1821 struct wait_queue_head *wq; 1822 wait_queue_entry_t *wait; 1823 bool ret; 1824 1825 if (!(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) && 1826 !(blk_mq_is_shared_tags(hctx->flags))) { 1827 blk_mq_sched_mark_restart_hctx(hctx); 1828 1829 /* 1830 * It's possible that a tag was freed in the window between the 1831 * allocation failure and adding the hardware queue to the wait 1832 * queue. 1833 * 1834 * Don't clear RESTART here, someone else could have set it. 1835 * At most this will cost an extra queue run. 1836 */ 1837 return blk_mq_get_driver_tag(rq); 1838 } 1839 1840 wait = &hctx->dispatch_wait; 1841 if (!list_empty_careful(&wait->entry)) 1842 return false; 1843 1844 if (blk_mq_tag_is_reserved(rq->mq_hctx->sched_tags, rq->internal_tag)) 1845 sbq = &hctx->tags->breserved_tags; 1846 else 1847 sbq = &hctx->tags->bitmap_tags; 1848 wq = &bt_wait_ptr(sbq, hctx)->wait; 1849 1850 spin_lock_irq(&wq->lock); 1851 spin_lock(&hctx->dispatch_wait_lock); 1852 if (!list_empty(&wait->entry)) { 1853 spin_unlock(&hctx->dispatch_wait_lock); 1854 spin_unlock_irq(&wq->lock); 1855 return false; 1856 } 1857 1858 atomic_inc(&sbq->ws_active); 1859 wait->flags &= ~WQ_FLAG_EXCLUSIVE; 1860 __add_wait_queue(wq, wait); 1861 1862 /* 1863 * It's possible that a tag was freed in the window between the 1864 * allocation failure and adding the hardware queue to the wait 1865 * queue. 1866 */ 1867 ret = blk_mq_get_driver_tag(rq); 1868 if (!ret) { 1869 spin_unlock(&hctx->dispatch_wait_lock); 1870 spin_unlock_irq(&wq->lock); 1871 return false; 1872 } 1873 1874 /* 1875 * We got a tag, remove ourselves from the wait queue to ensure 1876 * someone else gets the wakeup. 1877 */ 1878 list_del_init(&wait->entry); 1879 atomic_dec(&sbq->ws_active); 1880 spin_unlock(&hctx->dispatch_wait_lock); 1881 spin_unlock_irq(&wq->lock); 1882 1883 return true; 1884 } 1885 1886 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT 8 1887 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR 4 1888 /* 1889 * Update dispatch busy with the Exponential Weighted Moving Average(EWMA): 1890 * - EWMA is one simple way to compute running average value 1891 * - weight(7/8 and 1/8) is applied so that it can decrease exponentially 1892 * - take 4 as factor for avoiding to get too small(0) result, and this 1893 * factor doesn't matter because EWMA decreases exponentially 1894 */ 1895 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy) 1896 { 1897 unsigned int ewma; 1898 1899 ewma = hctx->dispatch_busy; 1900 1901 if (!ewma && !busy) 1902 return; 1903 1904 ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1; 1905 if (busy) 1906 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR; 1907 ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT; 1908 1909 hctx->dispatch_busy = ewma; 1910 } 1911 1912 #define BLK_MQ_RESOURCE_DELAY 3 /* ms units */ 1913 1914 static void blk_mq_handle_dev_resource(struct request *rq, 1915 struct list_head *list) 1916 { 1917 list_add(&rq->queuelist, list); 1918 __blk_mq_requeue_request(rq); 1919 } 1920 1921 static void blk_mq_handle_zone_resource(struct request *rq, 1922 struct list_head *zone_list) 1923 { 1924 /* 1925 * If we end up here it is because we cannot dispatch a request to a 1926 * specific zone due to LLD level zone-write locking or other zone 1927 * related resource not being available. In this case, set the request 1928 * aside in zone_list for retrying it later. 1929 */ 1930 list_add(&rq->queuelist, zone_list); 1931 __blk_mq_requeue_request(rq); 1932 } 1933 1934 enum prep_dispatch { 1935 PREP_DISPATCH_OK, 1936 PREP_DISPATCH_NO_TAG, 1937 PREP_DISPATCH_NO_BUDGET, 1938 }; 1939 1940 static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq, 1941 bool need_budget) 1942 { 1943 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 1944 int budget_token = -1; 1945 1946 if (need_budget) { 1947 budget_token = blk_mq_get_dispatch_budget(rq->q); 1948 if (budget_token < 0) { 1949 blk_mq_put_driver_tag(rq); 1950 return PREP_DISPATCH_NO_BUDGET; 1951 } 1952 blk_mq_set_rq_budget_token(rq, budget_token); 1953 } 1954 1955 if (!blk_mq_get_driver_tag(rq)) { 1956 /* 1957 * The initial allocation attempt failed, so we need to 1958 * rerun the hardware queue when a tag is freed. The 1959 * waitqueue takes care of that. If the queue is run 1960 * before we add this entry back on the dispatch list, 1961 * we'll re-run it below. 1962 */ 1963 if (!blk_mq_mark_tag_wait(hctx, rq)) { 1964 /* 1965 * All budgets not got from this function will be put 1966 * together during handling partial dispatch 1967 */ 1968 if (need_budget) 1969 blk_mq_put_dispatch_budget(rq->q, budget_token); 1970 return PREP_DISPATCH_NO_TAG; 1971 } 1972 } 1973 1974 return PREP_DISPATCH_OK; 1975 } 1976 1977 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */ 1978 static void blk_mq_release_budgets(struct request_queue *q, 1979 struct list_head *list) 1980 { 1981 struct request *rq; 1982 1983 list_for_each_entry(rq, list, queuelist) { 1984 int budget_token = blk_mq_get_rq_budget_token(rq); 1985 1986 if (budget_token >= 0) 1987 blk_mq_put_dispatch_budget(q, budget_token); 1988 } 1989 } 1990 1991 /* 1992 * blk_mq_commit_rqs will notify driver using bd->last that there is no 1993 * more requests. (See comment in struct blk_mq_ops for commit_rqs for 1994 * details) 1995 * Attention, we should explicitly call this in unusual cases: 1996 * 1) did not queue everything initially scheduled to queue 1997 * 2) the last attempt to queue a request failed 1998 */ 1999 static void blk_mq_commit_rqs(struct blk_mq_hw_ctx *hctx, int queued, 2000 bool from_schedule) 2001 { 2002 if (hctx->queue->mq_ops->commit_rqs && queued) { 2003 trace_block_unplug(hctx->queue, queued, !from_schedule); 2004 hctx->queue->mq_ops->commit_rqs(hctx); 2005 } 2006 } 2007 2008 /* 2009 * Returns true if we did some work AND can potentially do more. 2010 */ 2011 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list, 2012 unsigned int nr_budgets) 2013 { 2014 enum prep_dispatch prep; 2015 struct request_queue *q = hctx->queue; 2016 struct request *rq; 2017 int queued; 2018 blk_status_t ret = BLK_STS_OK; 2019 LIST_HEAD(zone_list); 2020 bool needs_resource = false; 2021 2022 if (list_empty(list)) 2023 return false; 2024 2025 /* 2026 * Now process all the entries, sending them to the driver. 2027 */ 2028 queued = 0; 2029 do { 2030 struct blk_mq_queue_data bd; 2031 2032 rq = list_first_entry(list, struct request, queuelist); 2033 2034 WARN_ON_ONCE(hctx != rq->mq_hctx); 2035 prep = blk_mq_prep_dispatch_rq(rq, !nr_budgets); 2036 if (prep != PREP_DISPATCH_OK) 2037 break; 2038 2039 list_del_init(&rq->queuelist); 2040 2041 bd.rq = rq; 2042 bd.last = list_empty(list); 2043 2044 /* 2045 * once the request is queued to lld, no need to cover the 2046 * budget any more 2047 */ 2048 if (nr_budgets) 2049 nr_budgets--; 2050 ret = q->mq_ops->queue_rq(hctx, &bd); 2051 switch (ret) { 2052 case BLK_STS_OK: 2053 queued++; 2054 break; 2055 case BLK_STS_RESOURCE: 2056 needs_resource = true; 2057 fallthrough; 2058 case BLK_STS_DEV_RESOURCE: 2059 blk_mq_handle_dev_resource(rq, list); 2060 goto out; 2061 case BLK_STS_ZONE_RESOURCE: 2062 /* 2063 * Move the request to zone_list and keep going through 2064 * the dispatch list to find more requests the drive can 2065 * accept. 2066 */ 2067 blk_mq_handle_zone_resource(rq, &zone_list); 2068 needs_resource = true; 2069 break; 2070 default: 2071 blk_mq_end_request(rq, ret); 2072 } 2073 } while (!list_empty(list)); 2074 out: 2075 if (!list_empty(&zone_list)) 2076 list_splice_tail_init(&zone_list, list); 2077 2078 /* If we didn't flush the entire list, we could have told the driver 2079 * there was more coming, but that turned out to be a lie. 2080 */ 2081 if (!list_empty(list) || ret != BLK_STS_OK) 2082 blk_mq_commit_rqs(hctx, queued, false); 2083 2084 /* 2085 * Any items that need requeuing? Stuff them into hctx->dispatch, 2086 * that is where we will continue on next queue run. 2087 */ 2088 if (!list_empty(list)) { 2089 bool needs_restart; 2090 /* For non-shared tags, the RESTART check will suffice */ 2091 bool no_tag = prep == PREP_DISPATCH_NO_TAG && 2092 ((hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) || 2093 blk_mq_is_shared_tags(hctx->flags)); 2094 2095 if (nr_budgets) 2096 blk_mq_release_budgets(q, list); 2097 2098 spin_lock(&hctx->lock); 2099 list_splice_tail_init(list, &hctx->dispatch); 2100 spin_unlock(&hctx->lock); 2101 2102 /* 2103 * Order adding requests to hctx->dispatch and checking 2104 * SCHED_RESTART flag. The pair of this smp_mb() is the one 2105 * in blk_mq_sched_restart(). Avoid restart code path to 2106 * miss the new added requests to hctx->dispatch, meantime 2107 * SCHED_RESTART is observed here. 2108 */ 2109 smp_mb(); 2110 2111 /* 2112 * If SCHED_RESTART was set by the caller of this function and 2113 * it is no longer set that means that it was cleared by another 2114 * thread and hence that a queue rerun is needed. 2115 * 2116 * If 'no_tag' is set, that means that we failed getting 2117 * a driver tag with an I/O scheduler attached. If our dispatch 2118 * waitqueue is no longer active, ensure that we run the queue 2119 * AFTER adding our entries back to the list. 2120 * 2121 * If no I/O scheduler has been configured it is possible that 2122 * the hardware queue got stopped and restarted before requests 2123 * were pushed back onto the dispatch list. Rerun the queue to 2124 * avoid starvation. Notes: 2125 * - blk_mq_run_hw_queue() checks whether or not a queue has 2126 * been stopped before rerunning a queue. 2127 * - Some but not all block drivers stop a queue before 2128 * returning BLK_STS_RESOURCE. Two exceptions are scsi-mq 2129 * and dm-rq. 2130 * 2131 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART 2132 * bit is set, run queue after a delay to avoid IO stalls 2133 * that could otherwise occur if the queue is idle. We'll do 2134 * similar if we couldn't get budget or couldn't lock a zone 2135 * and SCHED_RESTART is set. 2136 */ 2137 needs_restart = blk_mq_sched_needs_restart(hctx); 2138 if (prep == PREP_DISPATCH_NO_BUDGET) 2139 needs_resource = true; 2140 if (!needs_restart || 2141 (no_tag && list_empty_careful(&hctx->dispatch_wait.entry))) 2142 blk_mq_run_hw_queue(hctx, true); 2143 else if (needs_resource) 2144 blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY); 2145 2146 blk_mq_update_dispatch_busy(hctx, true); 2147 return false; 2148 } 2149 2150 blk_mq_update_dispatch_busy(hctx, false); 2151 return true; 2152 } 2153 2154 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx) 2155 { 2156 int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask); 2157 2158 if (cpu >= nr_cpu_ids) 2159 cpu = cpumask_first(hctx->cpumask); 2160 return cpu; 2161 } 2162 2163 /* 2164 * It'd be great if the workqueue API had a way to pass 2165 * in a mask and had some smarts for more clever placement. 2166 * For now we just round-robin here, switching for every 2167 * BLK_MQ_CPU_WORK_BATCH queued items. 2168 */ 2169 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx) 2170 { 2171 bool tried = false; 2172 int next_cpu = hctx->next_cpu; 2173 2174 if (hctx->queue->nr_hw_queues == 1) 2175 return WORK_CPU_UNBOUND; 2176 2177 if (--hctx->next_cpu_batch <= 0) { 2178 select_cpu: 2179 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask, 2180 cpu_online_mask); 2181 if (next_cpu >= nr_cpu_ids) 2182 next_cpu = blk_mq_first_mapped_cpu(hctx); 2183 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; 2184 } 2185 2186 /* 2187 * Do unbound schedule if we can't find a online CPU for this hctx, 2188 * and it should only happen in the path of handling CPU DEAD. 2189 */ 2190 if (!cpu_online(next_cpu)) { 2191 if (!tried) { 2192 tried = true; 2193 goto select_cpu; 2194 } 2195 2196 /* 2197 * Make sure to re-select CPU next time once after CPUs 2198 * in hctx->cpumask become online again. 2199 */ 2200 hctx->next_cpu = next_cpu; 2201 hctx->next_cpu_batch = 1; 2202 return WORK_CPU_UNBOUND; 2203 } 2204 2205 hctx->next_cpu = next_cpu; 2206 return next_cpu; 2207 } 2208 2209 /** 2210 * blk_mq_delay_run_hw_queue - Run a hardware queue asynchronously. 2211 * @hctx: Pointer to the hardware queue to run. 2212 * @msecs: Milliseconds of delay to wait before running the queue. 2213 * 2214 * Run a hardware queue asynchronously with a delay of @msecs. 2215 */ 2216 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs) 2217 { 2218 if (unlikely(blk_mq_hctx_stopped(hctx))) 2219 return; 2220 kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work, 2221 msecs_to_jiffies(msecs)); 2222 } 2223 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue); 2224 2225 /** 2226 * blk_mq_run_hw_queue - Start to run a hardware queue. 2227 * @hctx: Pointer to the hardware queue to run. 2228 * @async: If we want to run the queue asynchronously. 2229 * 2230 * Check if the request queue is not in a quiesced state and if there are 2231 * pending requests to be sent. If this is true, run the queue to send requests 2232 * to hardware. 2233 */ 2234 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) 2235 { 2236 bool need_run; 2237 2238 /* 2239 * We can't run the queue inline with interrupts disabled. 2240 */ 2241 WARN_ON_ONCE(!async && in_interrupt()); 2242 2243 might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING); 2244 2245 /* 2246 * When queue is quiesced, we may be switching io scheduler, or 2247 * updating nr_hw_queues, or other things, and we can't run queue 2248 * any more, even __blk_mq_hctx_has_pending() can't be called safely. 2249 * 2250 * And queue will be rerun in blk_mq_unquiesce_queue() if it is 2251 * quiesced. 2252 */ 2253 __blk_mq_run_dispatch_ops(hctx->queue, false, 2254 need_run = !blk_queue_quiesced(hctx->queue) && 2255 blk_mq_hctx_has_pending(hctx)); 2256 2257 if (!need_run) 2258 return; 2259 2260 if (async || !cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask)) { 2261 blk_mq_delay_run_hw_queue(hctx, 0); 2262 return; 2263 } 2264 2265 blk_mq_run_dispatch_ops(hctx->queue, 2266 blk_mq_sched_dispatch_requests(hctx)); 2267 } 2268 EXPORT_SYMBOL(blk_mq_run_hw_queue); 2269 2270 /* 2271 * Return prefered queue to dispatch from (if any) for non-mq aware IO 2272 * scheduler. 2273 */ 2274 static struct blk_mq_hw_ctx *blk_mq_get_sq_hctx(struct request_queue *q) 2275 { 2276 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 2277 /* 2278 * If the IO scheduler does not respect hardware queues when 2279 * dispatching, we just don't bother with multiple HW queues and 2280 * dispatch from hctx for the current CPU since running multiple queues 2281 * just causes lock contention inside the scheduler and pointless cache 2282 * bouncing. 2283 */ 2284 struct blk_mq_hw_ctx *hctx = ctx->hctxs[HCTX_TYPE_DEFAULT]; 2285 2286 if (!blk_mq_hctx_stopped(hctx)) 2287 return hctx; 2288 return NULL; 2289 } 2290 2291 /** 2292 * blk_mq_run_hw_queues - Run all hardware queues in a request queue. 2293 * @q: Pointer to the request queue to run. 2294 * @async: If we want to run the queue asynchronously. 2295 */ 2296 void blk_mq_run_hw_queues(struct request_queue *q, bool async) 2297 { 2298 struct blk_mq_hw_ctx *hctx, *sq_hctx; 2299 unsigned long i; 2300 2301 sq_hctx = NULL; 2302 if (blk_queue_sq_sched(q)) 2303 sq_hctx = blk_mq_get_sq_hctx(q); 2304 queue_for_each_hw_ctx(q, hctx, i) { 2305 if (blk_mq_hctx_stopped(hctx)) 2306 continue; 2307 /* 2308 * Dispatch from this hctx either if there's no hctx preferred 2309 * by IO scheduler or if it has requests that bypass the 2310 * scheduler. 2311 */ 2312 if (!sq_hctx || sq_hctx == hctx || 2313 !list_empty_careful(&hctx->dispatch)) 2314 blk_mq_run_hw_queue(hctx, async); 2315 } 2316 } 2317 EXPORT_SYMBOL(blk_mq_run_hw_queues); 2318 2319 /** 2320 * blk_mq_delay_run_hw_queues - Run all hardware queues asynchronously. 2321 * @q: Pointer to the request queue to run. 2322 * @msecs: Milliseconds of delay to wait before running the queues. 2323 */ 2324 void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) 2325 { 2326 struct blk_mq_hw_ctx *hctx, *sq_hctx; 2327 unsigned long i; 2328 2329 sq_hctx = NULL; 2330 if (blk_queue_sq_sched(q)) 2331 sq_hctx = blk_mq_get_sq_hctx(q); 2332 queue_for_each_hw_ctx(q, hctx, i) { 2333 if (blk_mq_hctx_stopped(hctx)) 2334 continue; 2335 /* 2336 * If there is already a run_work pending, leave the 2337 * pending delay untouched. Otherwise, a hctx can stall 2338 * if another hctx is re-delaying the other's work 2339 * before the work executes. 2340 */ 2341 if (delayed_work_pending(&hctx->run_work)) 2342 continue; 2343 /* 2344 * Dispatch from this hctx either if there's no hctx preferred 2345 * by IO scheduler or if it has requests that bypass the 2346 * scheduler. 2347 */ 2348 if (!sq_hctx || sq_hctx == hctx || 2349 !list_empty_careful(&hctx->dispatch)) 2350 blk_mq_delay_run_hw_queue(hctx, msecs); 2351 } 2352 } 2353 EXPORT_SYMBOL(blk_mq_delay_run_hw_queues); 2354 2355 /* 2356 * This function is often used for pausing .queue_rq() by driver when 2357 * there isn't enough resource or some conditions aren't satisfied, and 2358 * BLK_STS_RESOURCE is usually returned. 2359 * 2360 * We do not guarantee that dispatch can be drained or blocked 2361 * after blk_mq_stop_hw_queue() returns. Please use 2362 * blk_mq_quiesce_queue() for that requirement. 2363 */ 2364 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx) 2365 { 2366 cancel_delayed_work(&hctx->run_work); 2367 2368 set_bit(BLK_MQ_S_STOPPED, &hctx->state); 2369 } 2370 EXPORT_SYMBOL(blk_mq_stop_hw_queue); 2371 2372 /* 2373 * This function is often used for pausing .queue_rq() by driver when 2374 * there isn't enough resource or some conditions aren't satisfied, and 2375 * BLK_STS_RESOURCE is usually returned. 2376 * 2377 * We do not guarantee that dispatch can be drained or blocked 2378 * after blk_mq_stop_hw_queues() returns. Please use 2379 * blk_mq_quiesce_queue() for that requirement. 2380 */ 2381 void blk_mq_stop_hw_queues(struct request_queue *q) 2382 { 2383 struct blk_mq_hw_ctx *hctx; 2384 unsigned long i; 2385 2386 queue_for_each_hw_ctx(q, hctx, i) 2387 blk_mq_stop_hw_queue(hctx); 2388 } 2389 EXPORT_SYMBOL(blk_mq_stop_hw_queues); 2390 2391 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx) 2392 { 2393 clear_bit(BLK_MQ_S_STOPPED, &hctx->state); 2394 2395 blk_mq_run_hw_queue(hctx, hctx->flags & BLK_MQ_F_BLOCKING); 2396 } 2397 EXPORT_SYMBOL(blk_mq_start_hw_queue); 2398 2399 void blk_mq_start_hw_queues(struct request_queue *q) 2400 { 2401 struct blk_mq_hw_ctx *hctx; 2402 unsigned long i; 2403 2404 queue_for_each_hw_ctx(q, hctx, i) 2405 blk_mq_start_hw_queue(hctx); 2406 } 2407 EXPORT_SYMBOL(blk_mq_start_hw_queues); 2408 2409 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) 2410 { 2411 if (!blk_mq_hctx_stopped(hctx)) 2412 return; 2413 2414 clear_bit(BLK_MQ_S_STOPPED, &hctx->state); 2415 blk_mq_run_hw_queue(hctx, async); 2416 } 2417 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); 2418 2419 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async) 2420 { 2421 struct blk_mq_hw_ctx *hctx; 2422 unsigned long i; 2423 2424 queue_for_each_hw_ctx(q, hctx, i) 2425 blk_mq_start_stopped_hw_queue(hctx, async || 2426 (hctx->flags & BLK_MQ_F_BLOCKING)); 2427 } 2428 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues); 2429 2430 static void blk_mq_run_work_fn(struct work_struct *work) 2431 { 2432 struct blk_mq_hw_ctx *hctx = 2433 container_of(work, struct blk_mq_hw_ctx, run_work.work); 2434 2435 blk_mq_run_dispatch_ops(hctx->queue, 2436 blk_mq_sched_dispatch_requests(hctx)); 2437 } 2438 2439 /** 2440 * blk_mq_request_bypass_insert - Insert a request at dispatch list. 2441 * @rq: Pointer to request to be inserted. 2442 * @flags: BLK_MQ_INSERT_* 2443 * 2444 * Should only be used carefully, when the caller knows we want to 2445 * bypass a potential IO scheduler on the target device. 2446 */ 2447 static void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags) 2448 { 2449 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2450 2451 spin_lock(&hctx->lock); 2452 if (flags & BLK_MQ_INSERT_AT_HEAD) 2453 list_add(&rq->queuelist, &hctx->dispatch); 2454 else 2455 list_add_tail(&rq->queuelist, &hctx->dispatch); 2456 spin_unlock(&hctx->lock); 2457 } 2458 2459 static void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, 2460 struct blk_mq_ctx *ctx, struct list_head *list, 2461 bool run_queue_async) 2462 { 2463 struct request *rq; 2464 enum hctx_type type = hctx->type; 2465 2466 /* 2467 * Try to issue requests directly if the hw queue isn't busy to save an 2468 * extra enqueue & dequeue to the sw queue. 2469 */ 2470 if (!hctx->dispatch_busy && !run_queue_async) { 2471 blk_mq_run_dispatch_ops(hctx->queue, 2472 blk_mq_try_issue_list_directly(hctx, list)); 2473 if (list_empty(list)) 2474 goto out; 2475 } 2476 2477 /* 2478 * preemption doesn't flush plug list, so it's possible ctx->cpu is 2479 * offline now 2480 */ 2481 list_for_each_entry(rq, list, queuelist) { 2482 BUG_ON(rq->mq_ctx != ctx); 2483 trace_block_rq_insert(rq); 2484 if (rq->cmd_flags & REQ_NOWAIT) 2485 run_queue_async = true; 2486 } 2487 2488 spin_lock(&ctx->lock); 2489 list_splice_tail_init(list, &ctx->rq_lists[type]); 2490 blk_mq_hctx_mark_pending(hctx, ctx); 2491 spin_unlock(&ctx->lock); 2492 out: 2493 blk_mq_run_hw_queue(hctx, run_queue_async); 2494 } 2495 2496 static void blk_mq_insert_request(struct request *rq, blk_insert_t flags) 2497 { 2498 struct request_queue *q = rq->q; 2499 struct blk_mq_ctx *ctx = rq->mq_ctx; 2500 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2501 2502 if (blk_rq_is_passthrough(rq)) { 2503 /* 2504 * Passthrough request have to be added to hctx->dispatch 2505 * directly. The device may be in a situation where it can't 2506 * handle FS request, and always returns BLK_STS_RESOURCE for 2507 * them, which gets them added to hctx->dispatch. 2508 * 2509 * If a passthrough request is required to unblock the queues, 2510 * and it is added to the scheduler queue, there is no chance to 2511 * dispatch it given we prioritize requests in hctx->dispatch. 2512 */ 2513 blk_mq_request_bypass_insert(rq, flags); 2514 } else if (req_op(rq) == REQ_OP_FLUSH) { 2515 /* 2516 * Firstly normal IO request is inserted to scheduler queue or 2517 * sw queue, meantime we add flush request to dispatch queue( 2518 * hctx->dispatch) directly and there is at most one in-flight 2519 * flush request for each hw queue, so it doesn't matter to add 2520 * flush request to tail or front of the dispatch queue. 2521 * 2522 * Secondly in case of NCQ, flush request belongs to non-NCQ 2523 * command, and queueing it will fail when there is any 2524 * in-flight normal IO request(NCQ command). When adding flush 2525 * rq to the front of hctx->dispatch, it is easier to introduce 2526 * extra time to flush rq's latency because of S_SCHED_RESTART 2527 * compared with adding to the tail of dispatch queue, then 2528 * chance of flush merge is increased, and less flush requests 2529 * will be issued to controller. It is observed that ~10% time 2530 * is saved in blktests block/004 on disk attached to AHCI/NCQ 2531 * drive when adding flush rq to the front of hctx->dispatch. 2532 * 2533 * Simply queue flush rq to the front of hctx->dispatch so that 2534 * intensive flush workloads can benefit in case of NCQ HW. 2535 */ 2536 blk_mq_request_bypass_insert(rq, BLK_MQ_INSERT_AT_HEAD); 2537 } else if (q->elevator) { 2538 LIST_HEAD(list); 2539 2540 WARN_ON_ONCE(rq->tag != BLK_MQ_NO_TAG); 2541 2542 list_add(&rq->queuelist, &list); 2543 q->elevator->type->ops.insert_requests(hctx, &list, flags); 2544 } else { 2545 trace_block_rq_insert(rq); 2546 2547 spin_lock(&ctx->lock); 2548 if (flags & BLK_MQ_INSERT_AT_HEAD) 2549 list_add(&rq->queuelist, &ctx->rq_lists[hctx->type]); 2550 else 2551 list_add_tail(&rq->queuelist, 2552 &ctx->rq_lists[hctx->type]); 2553 blk_mq_hctx_mark_pending(hctx, ctx); 2554 spin_unlock(&ctx->lock); 2555 } 2556 } 2557 2558 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio, 2559 unsigned int nr_segs) 2560 { 2561 int err; 2562 2563 if (bio->bi_opf & REQ_RAHEAD) 2564 rq->cmd_flags |= REQ_FAILFAST_MASK; 2565 2566 rq->__sector = bio->bi_iter.bi_sector; 2567 blk_rq_bio_prep(rq, bio, nr_segs); 2568 2569 /* This can't fail, since GFP_NOIO includes __GFP_DIRECT_RECLAIM. */ 2570 err = blk_crypto_rq_bio_prep(rq, bio, GFP_NOIO); 2571 WARN_ON_ONCE(err); 2572 2573 blk_account_io_start(rq); 2574 } 2575 2576 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx, 2577 struct request *rq, bool last) 2578 { 2579 struct request_queue *q = rq->q; 2580 struct blk_mq_queue_data bd = { 2581 .rq = rq, 2582 .last = last, 2583 }; 2584 blk_status_t ret; 2585 2586 /* 2587 * For OK queue, we are done. For error, caller may kill it. 2588 * Any other error (busy), just add it to our list as we 2589 * previously would have done. 2590 */ 2591 ret = q->mq_ops->queue_rq(hctx, &bd); 2592 switch (ret) { 2593 case BLK_STS_OK: 2594 blk_mq_update_dispatch_busy(hctx, false); 2595 break; 2596 case BLK_STS_RESOURCE: 2597 case BLK_STS_DEV_RESOURCE: 2598 blk_mq_update_dispatch_busy(hctx, true); 2599 __blk_mq_requeue_request(rq); 2600 break; 2601 default: 2602 blk_mq_update_dispatch_busy(hctx, false); 2603 break; 2604 } 2605 2606 return ret; 2607 } 2608 2609 static bool blk_mq_get_budget_and_tag(struct request *rq) 2610 { 2611 int budget_token; 2612 2613 budget_token = blk_mq_get_dispatch_budget(rq->q); 2614 if (budget_token < 0) 2615 return false; 2616 blk_mq_set_rq_budget_token(rq, budget_token); 2617 if (!blk_mq_get_driver_tag(rq)) { 2618 blk_mq_put_dispatch_budget(rq->q, budget_token); 2619 return false; 2620 } 2621 return true; 2622 } 2623 2624 /** 2625 * blk_mq_try_issue_directly - Try to send a request directly to device driver. 2626 * @hctx: Pointer of the associated hardware queue. 2627 * @rq: Pointer to request to be sent. 2628 * 2629 * If the device has enough resources to accept a new request now, send the 2630 * request directly to device driver. Else, insert at hctx->dispatch queue, so 2631 * we can try send it another time in the future. Requests inserted at this 2632 * queue have higher priority. 2633 */ 2634 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, 2635 struct request *rq) 2636 { 2637 blk_status_t ret; 2638 2639 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { 2640 blk_mq_insert_request(rq, 0); 2641 return; 2642 } 2643 2644 if ((rq->rq_flags & RQF_USE_SCHED) || !blk_mq_get_budget_and_tag(rq)) { 2645 blk_mq_insert_request(rq, 0); 2646 blk_mq_run_hw_queue(hctx, rq->cmd_flags & REQ_NOWAIT); 2647 return; 2648 } 2649 2650 ret = __blk_mq_issue_directly(hctx, rq, true); 2651 switch (ret) { 2652 case BLK_STS_OK: 2653 break; 2654 case BLK_STS_RESOURCE: 2655 case BLK_STS_DEV_RESOURCE: 2656 blk_mq_request_bypass_insert(rq, 0); 2657 blk_mq_run_hw_queue(hctx, false); 2658 break; 2659 default: 2660 blk_mq_end_request(rq, ret); 2661 break; 2662 } 2663 } 2664 2665 static blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last) 2666 { 2667 struct blk_mq_hw_ctx *hctx = rq->mq_hctx; 2668 2669 if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(rq->q)) { 2670 blk_mq_insert_request(rq, 0); 2671 return BLK_STS_OK; 2672 } 2673 2674 if (!blk_mq_get_budget_and_tag(rq)) 2675 return BLK_STS_RESOURCE; 2676 return __blk_mq_issue_directly(hctx, rq, last); 2677 } 2678 2679 static void blk_mq_plug_issue_direct(struct blk_plug *plug) 2680 { 2681 struct blk_mq_hw_ctx *hctx = NULL; 2682 struct request *rq; 2683 int queued = 0; 2684 blk_status_t ret = BLK_STS_OK; 2685 2686 while ((rq = rq_list_pop(&plug->mq_list))) { 2687 bool last = rq_list_empty(plug->mq_list); 2688 2689 if (hctx != rq->mq_hctx) { 2690 if (hctx) { 2691 blk_mq_commit_rqs(hctx, queued, false); 2692 queued = 0; 2693 } 2694 hctx = rq->mq_hctx; 2695 } 2696 2697 ret = blk_mq_request_issue_directly(rq, last); 2698 switch (ret) { 2699 case BLK_STS_OK: 2700 queued++; 2701 break; 2702 case BLK_STS_RESOURCE: 2703 case BLK_STS_DEV_RESOURCE: 2704 blk_mq_request_bypass_insert(rq, 0); 2705 blk_mq_run_hw_queue(hctx, false); 2706 goto out; 2707 default: 2708 blk_mq_end_request(rq, ret); 2709 break; 2710 } 2711 } 2712 2713 out: 2714 if (ret != BLK_STS_OK) 2715 blk_mq_commit_rqs(hctx, queued, false); 2716 } 2717 2718 static void __blk_mq_flush_plug_list(struct request_queue *q, 2719 struct blk_plug *plug) 2720 { 2721 if (blk_queue_quiesced(q)) 2722 return; 2723 q->mq_ops->queue_rqs(&plug->mq_list); 2724 } 2725 2726 static void blk_mq_dispatch_plug_list(struct blk_plug *plug, bool from_sched) 2727 { 2728 struct blk_mq_hw_ctx *this_hctx = NULL; 2729 struct blk_mq_ctx *this_ctx = NULL; 2730 struct request *requeue_list = NULL; 2731 struct request **requeue_lastp = &requeue_list; 2732 unsigned int depth = 0; 2733 bool is_passthrough = false; 2734 LIST_HEAD(list); 2735 2736 do { 2737 struct request *rq = rq_list_pop(&plug->mq_list); 2738 2739 if (!this_hctx) { 2740 this_hctx = rq->mq_hctx; 2741 this_ctx = rq->mq_ctx; 2742 is_passthrough = blk_rq_is_passthrough(rq); 2743 } else if (this_hctx != rq->mq_hctx || this_ctx != rq->mq_ctx || 2744 is_passthrough != blk_rq_is_passthrough(rq)) { 2745 rq_list_add_tail(&requeue_lastp, rq); 2746 continue; 2747 } 2748 list_add(&rq->queuelist, &list); 2749 depth++; 2750 } while (!rq_list_empty(plug->mq_list)); 2751 2752 plug->mq_list = requeue_list; 2753 trace_block_unplug(this_hctx->queue, depth, !from_sched); 2754 2755 percpu_ref_get(&this_hctx->queue->q_usage_counter); 2756 /* passthrough requests should never be issued to the I/O scheduler */ 2757 if (is_passthrough) { 2758 spin_lock(&this_hctx->lock); 2759 list_splice_tail_init(&list, &this_hctx->dispatch); 2760 spin_unlock(&this_hctx->lock); 2761 blk_mq_run_hw_queue(this_hctx, from_sched); 2762 } else if (this_hctx->queue->elevator) { 2763 this_hctx->queue->elevator->type->ops.insert_requests(this_hctx, 2764 &list, 0); 2765 blk_mq_run_hw_queue(this_hctx, from_sched); 2766 } else { 2767 blk_mq_insert_requests(this_hctx, this_ctx, &list, from_sched); 2768 } 2769 percpu_ref_put(&this_hctx->queue->q_usage_counter); 2770 } 2771 2772 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule) 2773 { 2774 struct request *rq; 2775 2776 /* 2777 * We may have been called recursively midway through handling 2778 * plug->mq_list via a schedule() in the driver's queue_rq() callback. 2779 * To avoid mq_list changing under our feet, clear rq_count early and 2780 * bail out specifically if rq_count is 0 rather than checking 2781 * whether the mq_list is empty. 2782 */ 2783 if (plug->rq_count == 0) 2784 return; 2785 plug->rq_count = 0; 2786 2787 if (!plug->multiple_queues && !plug->has_elevator && !from_schedule) { 2788 struct request_queue *q; 2789 2790 rq = rq_list_peek(&plug->mq_list); 2791 q = rq->q; 2792 2793 /* 2794 * Peek first request and see if we have a ->queue_rqs() hook. 2795 * If we do, we can dispatch the whole plug list in one go. We 2796 * already know at this point that all requests belong to the 2797 * same queue, caller must ensure that's the case. 2798 */ 2799 if (q->mq_ops->queue_rqs) { 2800 blk_mq_run_dispatch_ops(q, 2801 __blk_mq_flush_plug_list(q, plug)); 2802 if (rq_list_empty(plug->mq_list)) 2803 return; 2804 } 2805 2806 blk_mq_run_dispatch_ops(q, 2807 blk_mq_plug_issue_direct(plug)); 2808 if (rq_list_empty(plug->mq_list)) 2809 return; 2810 } 2811 2812 do { 2813 blk_mq_dispatch_plug_list(plug, from_schedule); 2814 } while (!rq_list_empty(plug->mq_list)); 2815 } 2816 2817 static void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx, 2818 struct list_head *list) 2819 { 2820 int queued = 0; 2821 blk_status_t ret = BLK_STS_OK; 2822 2823 while (!list_empty(list)) { 2824 struct request *rq = list_first_entry(list, struct request, 2825 queuelist); 2826 2827 list_del_init(&rq->queuelist); 2828 ret = blk_mq_request_issue_directly(rq, list_empty(list)); 2829 switch (ret) { 2830 case BLK_STS_OK: 2831 queued++; 2832 break; 2833 case BLK_STS_RESOURCE: 2834 case BLK_STS_DEV_RESOURCE: 2835 blk_mq_request_bypass_insert(rq, 0); 2836 if (list_empty(list)) 2837 blk_mq_run_hw_queue(hctx, false); 2838 goto out; 2839 default: 2840 blk_mq_end_request(rq, ret); 2841 break; 2842 } 2843 } 2844 2845 out: 2846 if (ret != BLK_STS_OK) 2847 blk_mq_commit_rqs(hctx, queued, false); 2848 } 2849 2850 static bool blk_mq_attempt_bio_merge(struct request_queue *q, 2851 struct bio *bio, unsigned int nr_segs) 2852 { 2853 if (!blk_queue_nomerges(q) && bio_mergeable(bio)) { 2854 if (blk_attempt_plug_merge(q, bio, nr_segs)) 2855 return true; 2856 if (blk_mq_sched_bio_merge(q, bio, nr_segs)) 2857 return true; 2858 } 2859 return false; 2860 } 2861 2862 static struct request *blk_mq_get_new_requests(struct request_queue *q, 2863 struct blk_plug *plug, 2864 struct bio *bio, 2865 unsigned int nsegs) 2866 { 2867 struct blk_mq_alloc_data data = { 2868 .q = q, 2869 .nr_tags = 1, 2870 .cmd_flags = bio->bi_opf, 2871 }; 2872 struct request *rq; 2873 2874 if (blk_mq_attempt_bio_merge(q, bio, nsegs)) 2875 return NULL; 2876 2877 rq_qos_throttle(q, bio); 2878 2879 if (plug) { 2880 data.nr_tags = plug->nr_ios; 2881 plug->nr_ios = 1; 2882 data.cached_rq = &plug->cached_rq; 2883 } 2884 2885 rq = __blk_mq_alloc_requests(&data); 2886 if (rq) 2887 return rq; 2888 rq_qos_cleanup(q, bio); 2889 if (bio->bi_opf & REQ_NOWAIT) 2890 bio_wouldblock_error(bio); 2891 return NULL; 2892 } 2893 2894 /* return true if this @rq can be used for @bio */ 2895 static bool blk_mq_can_use_cached_rq(struct request *rq, struct blk_plug *plug, 2896 struct bio *bio) 2897 { 2898 enum hctx_type type = blk_mq_get_hctx_type(bio->bi_opf); 2899 enum hctx_type hctx_type = rq->mq_hctx->type; 2900 2901 WARN_ON_ONCE(rq_list_peek(&plug->cached_rq) != rq); 2902 2903 if (type != hctx_type && 2904 !(type == HCTX_TYPE_READ && hctx_type == HCTX_TYPE_DEFAULT)) 2905 return false; 2906 if (op_is_flush(rq->cmd_flags) != op_is_flush(bio->bi_opf)) 2907 return false; 2908 2909 /* 2910 * If any qos ->throttle() end up blocking, we will have flushed the 2911 * plug and hence killed the cached_rq list as well. Pop this entry 2912 * before we throttle. 2913 */ 2914 plug->cached_rq = rq_list_next(rq); 2915 rq_qos_throttle(rq->q, bio); 2916 2917 blk_mq_rq_time_init(rq, 0); 2918 rq->cmd_flags = bio->bi_opf; 2919 INIT_LIST_HEAD(&rq->queuelist); 2920 return true; 2921 } 2922 2923 static void bio_set_ioprio(struct bio *bio) 2924 { 2925 /* Nobody set ioprio so far? Initialize it based on task's nice value */ 2926 if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE) 2927 bio->bi_ioprio = get_current_ioprio(); 2928 blkcg_set_ioprio(bio); 2929 } 2930 2931 /** 2932 * blk_mq_submit_bio - Create and send a request to block device. 2933 * @bio: Bio pointer. 2934 * 2935 * Builds up a request structure from @q and @bio and send to the device. The 2936 * request may not be queued directly to hardware if: 2937 * * This request can be merged with another one 2938 * * We want to place request at plug queue for possible future merging 2939 * * There is an IO scheduler active at this queue 2940 * 2941 * It will not queue the request if there is an error with the bio, or at the 2942 * request creation. 2943 */ 2944 void blk_mq_submit_bio(struct bio *bio) 2945 { 2946 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 2947 struct blk_plug *plug = blk_mq_plug(bio); 2948 const int is_sync = op_is_sync(bio->bi_opf); 2949 struct blk_mq_hw_ctx *hctx; 2950 struct request *rq = NULL; 2951 unsigned int nr_segs = 1; 2952 blk_status_t ret; 2953 2954 bio = blk_queue_bounce(bio, q); 2955 if (bio_may_exceed_limits(bio, &q->limits)) { 2956 bio = __bio_split_to_limits(bio, &q->limits, &nr_segs); 2957 if (!bio) 2958 return; 2959 } 2960 2961 bio_set_ioprio(bio); 2962 2963 if (plug) { 2964 rq = rq_list_peek(&plug->cached_rq); 2965 if (rq && rq->q != q) 2966 rq = NULL; 2967 } 2968 if (rq) { 2969 if (!bio_integrity_prep(bio)) 2970 return; 2971 if (blk_mq_attempt_bio_merge(q, bio, nr_segs)) 2972 return; 2973 if (blk_mq_can_use_cached_rq(rq, plug, bio)) 2974 goto done; 2975 percpu_ref_get(&q->q_usage_counter); 2976 } else { 2977 if (unlikely(bio_queue_enter(bio))) 2978 return; 2979 if (!bio_integrity_prep(bio)) 2980 goto fail; 2981 } 2982 2983 rq = blk_mq_get_new_requests(q, plug, bio, nr_segs); 2984 if (unlikely(!rq)) { 2985 fail: 2986 blk_queue_exit(q); 2987 return; 2988 } 2989 2990 done: 2991 trace_block_getrq(bio); 2992 2993 rq_qos_track(q, rq, bio); 2994 2995 blk_mq_bio_to_request(rq, bio, nr_segs); 2996 2997 ret = blk_crypto_rq_get_keyslot(rq); 2998 if (ret != BLK_STS_OK) { 2999 bio->bi_status = ret; 3000 bio_endio(bio); 3001 blk_mq_free_request(rq); 3002 return; 3003 } 3004 3005 if (op_is_flush(bio->bi_opf) && blk_insert_flush(rq)) 3006 return; 3007 3008 if (plug) { 3009 blk_add_rq_to_plug(plug, rq); 3010 return; 3011 } 3012 3013 hctx = rq->mq_hctx; 3014 if ((rq->rq_flags & RQF_USE_SCHED) || 3015 (hctx->dispatch_busy && (q->nr_hw_queues == 1 || !is_sync))) { 3016 blk_mq_insert_request(rq, 0); 3017 blk_mq_run_hw_queue(hctx, true); 3018 } else { 3019 blk_mq_run_dispatch_ops(q, blk_mq_try_issue_directly(hctx, rq)); 3020 } 3021 } 3022 3023 #ifdef CONFIG_BLK_MQ_STACKING 3024 /** 3025 * blk_insert_cloned_request - Helper for stacking drivers to submit a request 3026 * @rq: the request being queued 3027 */ 3028 blk_status_t blk_insert_cloned_request(struct request *rq) 3029 { 3030 struct request_queue *q = rq->q; 3031 unsigned int max_sectors = blk_queue_get_max_sectors(q, req_op(rq)); 3032 unsigned int max_segments = blk_rq_get_max_segments(rq); 3033 blk_status_t ret; 3034 3035 if (blk_rq_sectors(rq) > max_sectors) { 3036 /* 3037 * SCSI device does not have a good way to return if 3038 * Write Same/Zero is actually supported. If a device rejects 3039 * a non-read/write command (discard, write same,etc.) the 3040 * low-level device driver will set the relevant queue limit to 3041 * 0 to prevent blk-lib from issuing more of the offending 3042 * operations. Commands queued prior to the queue limit being 3043 * reset need to be completed with BLK_STS_NOTSUPP to avoid I/O 3044 * errors being propagated to upper layers. 3045 */ 3046 if (max_sectors == 0) 3047 return BLK_STS_NOTSUPP; 3048 3049 printk(KERN_ERR "%s: over max size limit. (%u > %u)\n", 3050 __func__, blk_rq_sectors(rq), max_sectors); 3051 return BLK_STS_IOERR; 3052 } 3053 3054 /* 3055 * The queue settings related to segment counting may differ from the 3056 * original queue. 3057 */ 3058 rq->nr_phys_segments = blk_recalc_rq_segments(rq); 3059 if (rq->nr_phys_segments > max_segments) { 3060 printk(KERN_ERR "%s: over max segments limit. (%u > %u)\n", 3061 __func__, rq->nr_phys_segments, max_segments); 3062 return BLK_STS_IOERR; 3063 } 3064 3065 if (q->disk && should_fail_request(q->disk->part0, blk_rq_bytes(rq))) 3066 return BLK_STS_IOERR; 3067 3068 ret = blk_crypto_rq_get_keyslot(rq); 3069 if (ret != BLK_STS_OK) 3070 return ret; 3071 3072 blk_account_io_start(rq); 3073 3074 /* 3075 * Since we have a scheduler attached on the top device, 3076 * bypass a potential scheduler on the bottom device for 3077 * insert. 3078 */ 3079 blk_mq_run_dispatch_ops(q, 3080 ret = blk_mq_request_issue_directly(rq, true)); 3081 if (ret) 3082 blk_account_io_done(rq, ktime_get_ns()); 3083 return ret; 3084 } 3085 EXPORT_SYMBOL_GPL(blk_insert_cloned_request); 3086 3087 /** 3088 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request 3089 * @rq: the clone request to be cleaned up 3090 * 3091 * Description: 3092 * Free all bios in @rq for a cloned request. 3093 */ 3094 void blk_rq_unprep_clone(struct request *rq) 3095 { 3096 struct bio *bio; 3097 3098 while ((bio = rq->bio) != NULL) { 3099 rq->bio = bio->bi_next; 3100 3101 bio_put(bio); 3102 } 3103 } 3104 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); 3105 3106 /** 3107 * blk_rq_prep_clone - Helper function to setup clone request 3108 * @rq: the request to be setup 3109 * @rq_src: original request to be cloned 3110 * @bs: bio_set that bios for clone are allocated from 3111 * @gfp_mask: memory allocation mask for bio 3112 * @bio_ctr: setup function to be called for each clone bio. 3113 * Returns %0 for success, non %0 for failure. 3114 * @data: private data to be passed to @bio_ctr 3115 * 3116 * Description: 3117 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. 3118 * Also, pages which the original bios are pointing to are not copied 3119 * and the cloned bios just point same pages. 3120 * So cloned bios must be completed before original bios, which means 3121 * the caller must complete @rq before @rq_src. 3122 */ 3123 int blk_rq_prep_clone(struct request *rq, struct request *rq_src, 3124 struct bio_set *bs, gfp_t gfp_mask, 3125 int (*bio_ctr)(struct bio *, struct bio *, void *), 3126 void *data) 3127 { 3128 struct bio *bio, *bio_src; 3129 3130 if (!bs) 3131 bs = &fs_bio_set; 3132 3133 __rq_for_each_bio(bio_src, rq_src) { 3134 bio = bio_alloc_clone(rq->q->disk->part0, bio_src, gfp_mask, 3135 bs); 3136 if (!bio) 3137 goto free_and_out; 3138 3139 if (bio_ctr && bio_ctr(bio, bio_src, data)) 3140 goto free_and_out; 3141 3142 if (rq->bio) { 3143 rq->biotail->bi_next = bio; 3144 rq->biotail = bio; 3145 } else { 3146 rq->bio = rq->biotail = bio; 3147 } 3148 bio = NULL; 3149 } 3150 3151 /* Copy attributes of the original request to the clone request. */ 3152 rq->__sector = blk_rq_pos(rq_src); 3153 rq->__data_len = blk_rq_bytes(rq_src); 3154 if (rq_src->rq_flags & RQF_SPECIAL_PAYLOAD) { 3155 rq->rq_flags |= RQF_SPECIAL_PAYLOAD; 3156 rq->special_vec = rq_src->special_vec; 3157 } 3158 rq->nr_phys_segments = rq_src->nr_phys_segments; 3159 rq->ioprio = rq_src->ioprio; 3160 3161 if (rq->bio && blk_crypto_rq_bio_prep(rq, rq->bio, gfp_mask) < 0) 3162 goto free_and_out; 3163 3164 return 0; 3165 3166 free_and_out: 3167 if (bio) 3168 bio_put(bio); 3169 blk_rq_unprep_clone(rq); 3170 3171 return -ENOMEM; 3172 } 3173 EXPORT_SYMBOL_GPL(blk_rq_prep_clone); 3174 #endif /* CONFIG_BLK_MQ_STACKING */ 3175 3176 /* 3177 * Steal bios from a request and add them to a bio list. 3178 * The request must not have been partially completed before. 3179 */ 3180 void blk_steal_bios(struct bio_list *list, struct request *rq) 3181 { 3182 if (rq->bio) { 3183 if (list->tail) 3184 list->tail->bi_next = rq->bio; 3185 else 3186 list->head = rq->bio; 3187 list->tail = rq->biotail; 3188 3189 rq->bio = NULL; 3190 rq->biotail = NULL; 3191 } 3192 3193 rq->__data_len = 0; 3194 } 3195 EXPORT_SYMBOL_GPL(blk_steal_bios); 3196 3197 static size_t order_to_size(unsigned int order) 3198 { 3199 return (size_t)PAGE_SIZE << order; 3200 } 3201 3202 /* called before freeing request pool in @tags */ 3203 static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags, 3204 struct blk_mq_tags *tags) 3205 { 3206 struct page *page; 3207 unsigned long flags; 3208 3209 /* 3210 * There is no need to clear mapping if driver tags is not initialized 3211 * or the mapping belongs to the driver tags. 3212 */ 3213 if (!drv_tags || drv_tags == tags) 3214 return; 3215 3216 list_for_each_entry(page, &tags->page_list, lru) { 3217 unsigned long start = (unsigned long)page_address(page); 3218 unsigned long end = start + order_to_size(page->private); 3219 int i; 3220 3221 for (i = 0; i < drv_tags->nr_tags; i++) { 3222 struct request *rq = drv_tags->rqs[i]; 3223 unsigned long rq_addr = (unsigned long)rq; 3224 3225 if (rq_addr >= start && rq_addr < end) { 3226 WARN_ON_ONCE(req_ref_read(rq) != 0); 3227 cmpxchg(&drv_tags->rqs[i], rq, NULL); 3228 } 3229 } 3230 } 3231 3232 /* 3233 * Wait until all pending iteration is done. 3234 * 3235 * Request reference is cleared and it is guaranteed to be observed 3236 * after the ->lock is released. 3237 */ 3238 spin_lock_irqsave(&drv_tags->lock, flags); 3239 spin_unlock_irqrestore(&drv_tags->lock, flags); 3240 } 3241 3242 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, 3243 unsigned int hctx_idx) 3244 { 3245 struct blk_mq_tags *drv_tags; 3246 struct page *page; 3247 3248 if (list_empty(&tags->page_list)) 3249 return; 3250 3251 if (blk_mq_is_shared_tags(set->flags)) 3252 drv_tags = set->shared_tags; 3253 else 3254 drv_tags = set->tags[hctx_idx]; 3255 3256 if (tags->static_rqs && set->ops->exit_request) { 3257 int i; 3258 3259 for (i = 0; i < tags->nr_tags; i++) { 3260 struct request *rq = tags->static_rqs[i]; 3261 3262 if (!rq) 3263 continue; 3264 set->ops->exit_request(set, rq, hctx_idx); 3265 tags->static_rqs[i] = NULL; 3266 } 3267 } 3268 3269 blk_mq_clear_rq_mapping(drv_tags, tags); 3270 3271 while (!list_empty(&tags->page_list)) { 3272 page = list_first_entry(&tags->page_list, struct page, lru); 3273 list_del_init(&page->lru); 3274 /* 3275 * Remove kmemleak object previously allocated in 3276 * blk_mq_alloc_rqs(). 3277 */ 3278 kmemleak_free(page_address(page)); 3279 __free_pages(page, page->private); 3280 } 3281 } 3282 3283 void blk_mq_free_rq_map(struct blk_mq_tags *tags) 3284 { 3285 kfree(tags->rqs); 3286 tags->rqs = NULL; 3287 kfree(tags->static_rqs); 3288 tags->static_rqs = NULL; 3289 3290 blk_mq_free_tags(tags); 3291 } 3292 3293 static enum hctx_type hctx_idx_to_type(struct blk_mq_tag_set *set, 3294 unsigned int hctx_idx) 3295 { 3296 int i; 3297 3298 for (i = 0; i < set->nr_maps; i++) { 3299 unsigned int start = set->map[i].queue_offset; 3300 unsigned int end = start + set->map[i].nr_queues; 3301 3302 if (hctx_idx >= start && hctx_idx < end) 3303 break; 3304 } 3305 3306 if (i >= set->nr_maps) 3307 i = HCTX_TYPE_DEFAULT; 3308 3309 return i; 3310 } 3311 3312 static int blk_mq_get_hctx_node(struct blk_mq_tag_set *set, 3313 unsigned int hctx_idx) 3314 { 3315 enum hctx_type type = hctx_idx_to_type(set, hctx_idx); 3316 3317 return blk_mq_hw_queue_to_node(&set->map[type], hctx_idx); 3318 } 3319 3320 static struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, 3321 unsigned int hctx_idx, 3322 unsigned int nr_tags, 3323 unsigned int reserved_tags) 3324 { 3325 int node = blk_mq_get_hctx_node(set, hctx_idx); 3326 struct blk_mq_tags *tags; 3327 3328 if (node == NUMA_NO_NODE) 3329 node = set->numa_node; 3330 3331 tags = blk_mq_init_tags(nr_tags, reserved_tags, node, 3332 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags)); 3333 if (!tags) 3334 return NULL; 3335 3336 tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *), 3337 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, 3338 node); 3339 if (!tags->rqs) 3340 goto err_free_tags; 3341 3342 tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *), 3343 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, 3344 node); 3345 if (!tags->static_rqs) 3346 goto err_free_rqs; 3347 3348 return tags; 3349 3350 err_free_rqs: 3351 kfree(tags->rqs); 3352 err_free_tags: 3353 blk_mq_free_tags(tags); 3354 return NULL; 3355 } 3356 3357 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq, 3358 unsigned int hctx_idx, int node) 3359 { 3360 int ret; 3361 3362 if (set->ops->init_request) { 3363 ret = set->ops->init_request(set, rq, hctx_idx, node); 3364 if (ret) 3365 return ret; 3366 } 3367 3368 WRITE_ONCE(rq->state, MQ_RQ_IDLE); 3369 return 0; 3370 } 3371 3372 static int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, 3373 struct blk_mq_tags *tags, 3374 unsigned int hctx_idx, unsigned int depth) 3375 { 3376 unsigned int i, j, entries_per_page, max_order = 4; 3377 int node = blk_mq_get_hctx_node(set, hctx_idx); 3378 size_t rq_size, left; 3379 3380 if (node == NUMA_NO_NODE) 3381 node = set->numa_node; 3382 3383 INIT_LIST_HEAD(&tags->page_list); 3384 3385 /* 3386 * rq_size is the size of the request plus driver payload, rounded 3387 * to the cacheline size 3388 */ 3389 rq_size = round_up(sizeof(struct request) + set->cmd_size, 3390 cache_line_size()); 3391 left = rq_size * depth; 3392 3393 for (i = 0; i < depth; ) { 3394 int this_order = max_order; 3395 struct page *page; 3396 int to_do; 3397 void *p; 3398 3399 while (this_order && left < order_to_size(this_order - 1)) 3400 this_order--; 3401 3402 do { 3403 page = alloc_pages_node(node, 3404 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO, 3405 this_order); 3406 if (page) 3407 break; 3408 if (!this_order--) 3409 break; 3410 if (order_to_size(this_order) < rq_size) 3411 break; 3412 } while (1); 3413 3414 if (!page) 3415 goto fail; 3416 3417 page->private = this_order; 3418 list_add_tail(&page->lru, &tags->page_list); 3419 3420 p = page_address(page); 3421 /* 3422 * Allow kmemleak to scan these pages as they contain pointers 3423 * to additional allocations like via ops->init_request(). 3424 */ 3425 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO); 3426 entries_per_page = order_to_size(this_order) / rq_size; 3427 to_do = min(entries_per_page, depth - i); 3428 left -= to_do * rq_size; 3429 for (j = 0; j < to_do; j++) { 3430 struct request *rq = p; 3431 3432 tags->static_rqs[i] = rq; 3433 if (blk_mq_init_request(set, rq, hctx_idx, node)) { 3434 tags->static_rqs[i] = NULL; 3435 goto fail; 3436 } 3437 3438 p += rq_size; 3439 i++; 3440 } 3441 } 3442 return 0; 3443 3444 fail: 3445 blk_mq_free_rqs(set, tags, hctx_idx); 3446 return -ENOMEM; 3447 } 3448 3449 struct rq_iter_data { 3450 struct blk_mq_hw_ctx *hctx; 3451 bool has_rq; 3452 }; 3453 3454 static bool blk_mq_has_request(struct request *rq, void *data) 3455 { 3456 struct rq_iter_data *iter_data = data; 3457 3458 if (rq->mq_hctx != iter_data->hctx) 3459 return true; 3460 iter_data->has_rq = true; 3461 return false; 3462 } 3463 3464 static bool blk_mq_hctx_has_requests(struct blk_mq_hw_ctx *hctx) 3465 { 3466 struct blk_mq_tags *tags = hctx->sched_tags ? 3467 hctx->sched_tags : hctx->tags; 3468 struct rq_iter_data data = { 3469 .hctx = hctx, 3470 }; 3471 3472 blk_mq_all_tag_iter(tags, blk_mq_has_request, &data); 3473 return data.has_rq; 3474 } 3475 3476 static inline bool blk_mq_last_cpu_in_hctx(unsigned int cpu, 3477 struct blk_mq_hw_ctx *hctx) 3478 { 3479 if (cpumask_first_and(hctx->cpumask, cpu_online_mask) != cpu) 3480 return false; 3481 if (cpumask_next_and(cpu, hctx->cpumask, cpu_online_mask) < nr_cpu_ids) 3482 return false; 3483 return true; 3484 } 3485 3486 static int blk_mq_hctx_notify_offline(unsigned int cpu, struct hlist_node *node) 3487 { 3488 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, 3489 struct blk_mq_hw_ctx, cpuhp_online); 3490 3491 if (!cpumask_test_cpu(cpu, hctx->cpumask) || 3492 !blk_mq_last_cpu_in_hctx(cpu, hctx)) 3493 return 0; 3494 3495 /* 3496 * Prevent new request from being allocated on the current hctx. 3497 * 3498 * The smp_mb__after_atomic() Pairs with the implied barrier in 3499 * test_and_set_bit_lock in sbitmap_get(). Ensures the inactive flag is 3500 * seen once we return from the tag allocator. 3501 */ 3502 set_bit(BLK_MQ_S_INACTIVE, &hctx->state); 3503 smp_mb__after_atomic(); 3504 3505 /* 3506 * Try to grab a reference to the queue and wait for any outstanding 3507 * requests. If we could not grab a reference the queue has been 3508 * frozen and there are no requests. 3509 */ 3510 if (percpu_ref_tryget(&hctx->queue->q_usage_counter)) { 3511 while (blk_mq_hctx_has_requests(hctx)) 3512 msleep(5); 3513 percpu_ref_put(&hctx->queue->q_usage_counter); 3514 } 3515 3516 return 0; 3517 } 3518 3519 static int blk_mq_hctx_notify_online(unsigned int cpu, struct hlist_node *node) 3520 { 3521 struct blk_mq_hw_ctx *hctx = hlist_entry_safe(node, 3522 struct blk_mq_hw_ctx, cpuhp_online); 3523 3524 if (cpumask_test_cpu(cpu, hctx->cpumask)) 3525 clear_bit(BLK_MQ_S_INACTIVE, &hctx->state); 3526 return 0; 3527 } 3528 3529 /* 3530 * 'cpu' is going away. splice any existing rq_list entries from this 3531 * software queue to the hw queue dispatch list, and ensure that it 3532 * gets run. 3533 */ 3534 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node) 3535 { 3536 struct blk_mq_hw_ctx *hctx; 3537 struct blk_mq_ctx *ctx; 3538 LIST_HEAD(tmp); 3539 enum hctx_type type; 3540 3541 hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead); 3542 if (!cpumask_test_cpu(cpu, hctx->cpumask)) 3543 return 0; 3544 3545 ctx = __blk_mq_get_ctx(hctx->queue, cpu); 3546 type = hctx->type; 3547 3548 spin_lock(&ctx->lock); 3549 if (!list_empty(&ctx->rq_lists[type])) { 3550 list_splice_init(&ctx->rq_lists[type], &tmp); 3551 blk_mq_hctx_clear_pending(hctx, ctx); 3552 } 3553 spin_unlock(&ctx->lock); 3554 3555 if (list_empty(&tmp)) 3556 return 0; 3557 3558 spin_lock(&hctx->lock); 3559 list_splice_tail_init(&tmp, &hctx->dispatch); 3560 spin_unlock(&hctx->lock); 3561 3562 blk_mq_run_hw_queue(hctx, true); 3563 return 0; 3564 } 3565 3566 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx) 3567 { 3568 if (!(hctx->flags & BLK_MQ_F_STACKING)) 3569 cpuhp_state_remove_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, 3570 &hctx->cpuhp_online); 3571 cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD, 3572 &hctx->cpuhp_dead); 3573 } 3574 3575 /* 3576 * Before freeing hw queue, clearing the flush request reference in 3577 * tags->rqs[] for avoiding potential UAF. 3578 */ 3579 static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags, 3580 unsigned int queue_depth, struct request *flush_rq) 3581 { 3582 int i; 3583 unsigned long flags; 3584 3585 /* The hw queue may not be mapped yet */ 3586 if (!tags) 3587 return; 3588 3589 WARN_ON_ONCE(req_ref_read(flush_rq) != 0); 3590 3591 for (i = 0; i < queue_depth; i++) 3592 cmpxchg(&tags->rqs[i], flush_rq, NULL); 3593 3594 /* 3595 * Wait until all pending iteration is done. 3596 * 3597 * Request reference is cleared and it is guaranteed to be observed 3598 * after the ->lock is released. 3599 */ 3600 spin_lock_irqsave(&tags->lock, flags); 3601 spin_unlock_irqrestore(&tags->lock, flags); 3602 } 3603 3604 /* hctx->ctxs will be freed in queue's release handler */ 3605 static void blk_mq_exit_hctx(struct request_queue *q, 3606 struct blk_mq_tag_set *set, 3607 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) 3608 { 3609 struct request *flush_rq = hctx->fq->flush_rq; 3610 3611 if (blk_mq_hw_queue_mapped(hctx)) 3612 blk_mq_tag_idle(hctx); 3613 3614 if (blk_queue_init_done(q)) 3615 blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx], 3616 set->queue_depth, flush_rq); 3617 if (set->ops->exit_request) 3618 set->ops->exit_request(set, flush_rq, hctx_idx); 3619 3620 if (set->ops->exit_hctx) 3621 set->ops->exit_hctx(hctx, hctx_idx); 3622 3623 blk_mq_remove_cpuhp(hctx); 3624 3625 xa_erase(&q->hctx_table, hctx_idx); 3626 3627 spin_lock(&q->unused_hctx_lock); 3628 list_add(&hctx->hctx_list, &q->unused_hctx_list); 3629 spin_unlock(&q->unused_hctx_lock); 3630 } 3631 3632 static void blk_mq_exit_hw_queues(struct request_queue *q, 3633 struct blk_mq_tag_set *set, int nr_queue) 3634 { 3635 struct blk_mq_hw_ctx *hctx; 3636 unsigned long i; 3637 3638 queue_for_each_hw_ctx(q, hctx, i) { 3639 if (i == nr_queue) 3640 break; 3641 blk_mq_exit_hctx(q, set, hctx, i); 3642 } 3643 } 3644 3645 static int blk_mq_init_hctx(struct request_queue *q, 3646 struct blk_mq_tag_set *set, 3647 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx) 3648 { 3649 hctx->queue_num = hctx_idx; 3650 3651 if (!(hctx->flags & BLK_MQ_F_STACKING)) 3652 cpuhp_state_add_instance_nocalls(CPUHP_AP_BLK_MQ_ONLINE, 3653 &hctx->cpuhp_online); 3654 cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead); 3655 3656 hctx->tags = set->tags[hctx_idx]; 3657 3658 if (set->ops->init_hctx && 3659 set->ops->init_hctx(hctx, set->driver_data, hctx_idx)) 3660 goto unregister_cpu_notifier; 3661 3662 if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, 3663 hctx->numa_node)) 3664 goto exit_hctx; 3665 3666 if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL)) 3667 goto exit_flush_rq; 3668 3669 return 0; 3670 3671 exit_flush_rq: 3672 if (set->ops->exit_request) 3673 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx); 3674 exit_hctx: 3675 if (set->ops->exit_hctx) 3676 set->ops->exit_hctx(hctx, hctx_idx); 3677 unregister_cpu_notifier: 3678 blk_mq_remove_cpuhp(hctx); 3679 return -1; 3680 } 3681 3682 static struct blk_mq_hw_ctx * 3683 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set, 3684 int node) 3685 { 3686 struct blk_mq_hw_ctx *hctx; 3687 gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY; 3688 3689 hctx = kzalloc_node(sizeof(struct blk_mq_hw_ctx), gfp, node); 3690 if (!hctx) 3691 goto fail_alloc_hctx; 3692 3693 if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node)) 3694 goto free_hctx; 3695 3696 atomic_set(&hctx->nr_active, 0); 3697 if (node == NUMA_NO_NODE) 3698 node = set->numa_node; 3699 hctx->numa_node = node; 3700 3701 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn); 3702 spin_lock_init(&hctx->lock); 3703 INIT_LIST_HEAD(&hctx->dispatch); 3704 hctx->queue = q; 3705 hctx->flags = set->flags & ~BLK_MQ_F_TAG_QUEUE_SHARED; 3706 3707 INIT_LIST_HEAD(&hctx->hctx_list); 3708 3709 /* 3710 * Allocate space for all possible cpus to avoid allocation at 3711 * runtime 3712 */ 3713 hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *), 3714 gfp, node); 3715 if (!hctx->ctxs) 3716 goto free_cpumask; 3717 3718 if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), 3719 gfp, node, false, false)) 3720 goto free_ctxs; 3721 hctx->nr_ctx = 0; 3722 3723 spin_lock_init(&hctx->dispatch_wait_lock); 3724 init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake); 3725 INIT_LIST_HEAD(&hctx->dispatch_wait.entry); 3726 3727 hctx->fq = blk_alloc_flush_queue(hctx->numa_node, set->cmd_size, gfp); 3728 if (!hctx->fq) 3729 goto free_bitmap; 3730 3731 blk_mq_hctx_kobj_init(hctx); 3732 3733 return hctx; 3734 3735 free_bitmap: 3736 sbitmap_free(&hctx->ctx_map); 3737 free_ctxs: 3738 kfree(hctx->ctxs); 3739 free_cpumask: 3740 free_cpumask_var(hctx->cpumask); 3741 free_hctx: 3742 kfree(hctx); 3743 fail_alloc_hctx: 3744 return NULL; 3745 } 3746 3747 static void blk_mq_init_cpu_queues(struct request_queue *q, 3748 unsigned int nr_hw_queues) 3749 { 3750 struct blk_mq_tag_set *set = q->tag_set; 3751 unsigned int i, j; 3752 3753 for_each_possible_cpu(i) { 3754 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i); 3755 struct blk_mq_hw_ctx *hctx; 3756 int k; 3757 3758 __ctx->cpu = i; 3759 spin_lock_init(&__ctx->lock); 3760 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) 3761 INIT_LIST_HEAD(&__ctx->rq_lists[k]); 3762 3763 __ctx->queue = q; 3764 3765 /* 3766 * Set local node, IFF we have more than one hw queue. If 3767 * not, we remain on the home node of the device 3768 */ 3769 for (j = 0; j < set->nr_maps; j++) { 3770 hctx = blk_mq_map_queue_type(q, j, i); 3771 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE) 3772 hctx->numa_node = cpu_to_node(i); 3773 } 3774 } 3775 } 3776 3777 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, 3778 unsigned int hctx_idx, 3779 unsigned int depth) 3780 { 3781 struct blk_mq_tags *tags; 3782 int ret; 3783 3784 tags = blk_mq_alloc_rq_map(set, hctx_idx, depth, set->reserved_tags); 3785 if (!tags) 3786 return NULL; 3787 3788 ret = blk_mq_alloc_rqs(set, tags, hctx_idx, depth); 3789 if (ret) { 3790 blk_mq_free_rq_map(tags); 3791 return NULL; 3792 } 3793 3794 return tags; 3795 } 3796 3797 static bool __blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, 3798 int hctx_idx) 3799 { 3800 if (blk_mq_is_shared_tags(set->flags)) { 3801 set->tags[hctx_idx] = set->shared_tags; 3802 3803 return true; 3804 } 3805 3806 set->tags[hctx_idx] = blk_mq_alloc_map_and_rqs(set, hctx_idx, 3807 set->queue_depth); 3808 3809 return set->tags[hctx_idx]; 3810 } 3811 3812 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, 3813 struct blk_mq_tags *tags, 3814 unsigned int hctx_idx) 3815 { 3816 if (tags) { 3817 blk_mq_free_rqs(set, tags, hctx_idx); 3818 blk_mq_free_rq_map(tags); 3819 } 3820 } 3821 3822 static void __blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, 3823 unsigned int hctx_idx) 3824 { 3825 if (!blk_mq_is_shared_tags(set->flags)) 3826 blk_mq_free_map_and_rqs(set, set->tags[hctx_idx], hctx_idx); 3827 3828 set->tags[hctx_idx] = NULL; 3829 } 3830 3831 static void blk_mq_map_swqueue(struct request_queue *q) 3832 { 3833 unsigned int j, hctx_idx; 3834 unsigned long i; 3835 struct blk_mq_hw_ctx *hctx; 3836 struct blk_mq_ctx *ctx; 3837 struct blk_mq_tag_set *set = q->tag_set; 3838 3839 queue_for_each_hw_ctx(q, hctx, i) { 3840 cpumask_clear(hctx->cpumask); 3841 hctx->nr_ctx = 0; 3842 hctx->dispatch_from = NULL; 3843 } 3844 3845 /* 3846 * Map software to hardware queues. 3847 * 3848 * If the cpu isn't present, the cpu is mapped to first hctx. 3849 */ 3850 for_each_possible_cpu(i) { 3851 3852 ctx = per_cpu_ptr(q->queue_ctx, i); 3853 for (j = 0; j < set->nr_maps; j++) { 3854 if (!set->map[j].nr_queues) { 3855 ctx->hctxs[j] = blk_mq_map_queue_type(q, 3856 HCTX_TYPE_DEFAULT, i); 3857 continue; 3858 } 3859 hctx_idx = set->map[j].mq_map[i]; 3860 /* unmapped hw queue can be remapped after CPU topo changed */ 3861 if (!set->tags[hctx_idx] && 3862 !__blk_mq_alloc_map_and_rqs(set, hctx_idx)) { 3863 /* 3864 * If tags initialization fail for some hctx, 3865 * that hctx won't be brought online. In this 3866 * case, remap the current ctx to hctx[0] which 3867 * is guaranteed to always have tags allocated 3868 */ 3869 set->map[j].mq_map[i] = 0; 3870 } 3871 3872 hctx = blk_mq_map_queue_type(q, j, i); 3873 ctx->hctxs[j] = hctx; 3874 /* 3875 * If the CPU is already set in the mask, then we've 3876 * mapped this one already. This can happen if 3877 * devices share queues across queue maps. 3878 */ 3879 if (cpumask_test_cpu(i, hctx->cpumask)) 3880 continue; 3881 3882 cpumask_set_cpu(i, hctx->cpumask); 3883 hctx->type = j; 3884 ctx->index_hw[hctx->type] = hctx->nr_ctx; 3885 hctx->ctxs[hctx->nr_ctx++] = ctx; 3886 3887 /* 3888 * If the nr_ctx type overflows, we have exceeded the 3889 * amount of sw queues we can support. 3890 */ 3891 BUG_ON(!hctx->nr_ctx); 3892 } 3893 3894 for (; j < HCTX_MAX_TYPES; j++) 3895 ctx->hctxs[j] = blk_mq_map_queue_type(q, 3896 HCTX_TYPE_DEFAULT, i); 3897 } 3898 3899 queue_for_each_hw_ctx(q, hctx, i) { 3900 /* 3901 * If no software queues are mapped to this hardware queue, 3902 * disable it and free the request entries. 3903 */ 3904 if (!hctx->nr_ctx) { 3905 /* Never unmap queue 0. We need it as a 3906 * fallback in case of a new remap fails 3907 * allocation 3908 */ 3909 if (i) 3910 __blk_mq_free_map_and_rqs(set, i); 3911 3912 hctx->tags = NULL; 3913 continue; 3914 } 3915 3916 hctx->tags = set->tags[i]; 3917 WARN_ON(!hctx->tags); 3918 3919 /* 3920 * Set the map size to the number of mapped software queues. 3921 * This is more accurate and more efficient than looping 3922 * over all possibly mapped software queues. 3923 */ 3924 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx); 3925 3926 /* 3927 * Initialize batch roundrobin counts 3928 */ 3929 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx); 3930 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH; 3931 } 3932 } 3933 3934 /* 3935 * Caller needs to ensure that we're either frozen/quiesced, or that 3936 * the queue isn't live yet. 3937 */ 3938 static void queue_set_hctx_shared(struct request_queue *q, bool shared) 3939 { 3940 struct blk_mq_hw_ctx *hctx; 3941 unsigned long i; 3942 3943 queue_for_each_hw_ctx(q, hctx, i) { 3944 if (shared) { 3945 hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; 3946 } else { 3947 blk_mq_tag_idle(hctx); 3948 hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; 3949 } 3950 } 3951 } 3952 3953 static void blk_mq_update_tag_set_shared(struct blk_mq_tag_set *set, 3954 bool shared) 3955 { 3956 struct request_queue *q; 3957 3958 lockdep_assert_held(&set->tag_list_lock); 3959 3960 list_for_each_entry(q, &set->tag_list, tag_set_list) { 3961 blk_mq_freeze_queue(q); 3962 queue_set_hctx_shared(q, shared); 3963 blk_mq_unfreeze_queue(q); 3964 } 3965 } 3966 3967 static void blk_mq_del_queue_tag_set(struct request_queue *q) 3968 { 3969 struct blk_mq_tag_set *set = q->tag_set; 3970 3971 mutex_lock(&set->tag_list_lock); 3972 list_del(&q->tag_set_list); 3973 if (list_is_singular(&set->tag_list)) { 3974 /* just transitioned to unshared */ 3975 set->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED; 3976 /* update existing queue */ 3977 blk_mq_update_tag_set_shared(set, false); 3978 } 3979 mutex_unlock(&set->tag_list_lock); 3980 INIT_LIST_HEAD(&q->tag_set_list); 3981 } 3982 3983 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set, 3984 struct request_queue *q) 3985 { 3986 mutex_lock(&set->tag_list_lock); 3987 3988 /* 3989 * Check to see if we're transitioning to shared (from 1 to 2 queues). 3990 */ 3991 if (!list_empty(&set->tag_list) && 3992 !(set->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) { 3993 set->flags |= BLK_MQ_F_TAG_QUEUE_SHARED; 3994 /* update existing queue */ 3995 blk_mq_update_tag_set_shared(set, true); 3996 } 3997 if (set->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 3998 queue_set_hctx_shared(q, true); 3999 list_add_tail(&q->tag_set_list, &set->tag_list); 4000 4001 mutex_unlock(&set->tag_list_lock); 4002 } 4003 4004 /* All allocations will be freed in release handler of q->mq_kobj */ 4005 static int blk_mq_alloc_ctxs(struct request_queue *q) 4006 { 4007 struct blk_mq_ctxs *ctxs; 4008 int cpu; 4009 4010 ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL); 4011 if (!ctxs) 4012 return -ENOMEM; 4013 4014 ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx); 4015 if (!ctxs->queue_ctx) 4016 goto fail; 4017 4018 for_each_possible_cpu(cpu) { 4019 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu); 4020 ctx->ctxs = ctxs; 4021 } 4022 4023 q->mq_kobj = &ctxs->kobj; 4024 q->queue_ctx = ctxs->queue_ctx; 4025 4026 return 0; 4027 fail: 4028 kfree(ctxs); 4029 return -ENOMEM; 4030 } 4031 4032 /* 4033 * It is the actual release handler for mq, but we do it from 4034 * request queue's release handler for avoiding use-after-free 4035 * and headache because q->mq_kobj shouldn't have been introduced, 4036 * but we can't group ctx/kctx kobj without it. 4037 */ 4038 void blk_mq_release(struct request_queue *q) 4039 { 4040 struct blk_mq_hw_ctx *hctx, *next; 4041 unsigned long i; 4042 4043 queue_for_each_hw_ctx(q, hctx, i) 4044 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list)); 4045 4046 /* all hctx are in .unused_hctx_list now */ 4047 list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) { 4048 list_del_init(&hctx->hctx_list); 4049 kobject_put(&hctx->kobj); 4050 } 4051 4052 xa_destroy(&q->hctx_table); 4053 4054 /* 4055 * release .mq_kobj and sw queue's kobject now because 4056 * both share lifetime with request queue. 4057 */ 4058 blk_mq_sysfs_deinit(q); 4059 } 4060 4061 static struct request_queue *blk_mq_init_queue_data(struct blk_mq_tag_set *set, 4062 void *queuedata) 4063 { 4064 struct request_queue *q; 4065 int ret; 4066 4067 q = blk_alloc_queue(set->numa_node); 4068 if (!q) 4069 return ERR_PTR(-ENOMEM); 4070 q->queuedata = queuedata; 4071 ret = blk_mq_init_allocated_queue(set, q); 4072 if (ret) { 4073 blk_put_queue(q); 4074 return ERR_PTR(ret); 4075 } 4076 return q; 4077 } 4078 4079 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set) 4080 { 4081 return blk_mq_init_queue_data(set, NULL); 4082 } 4083 EXPORT_SYMBOL(blk_mq_init_queue); 4084 4085 /** 4086 * blk_mq_destroy_queue - shutdown a request queue 4087 * @q: request queue to shutdown 4088 * 4089 * This shuts down a request queue allocated by blk_mq_init_queue(). All future 4090 * requests will be failed with -ENODEV. The caller is responsible for dropping 4091 * the reference from blk_mq_init_queue() by calling blk_put_queue(). 4092 * 4093 * Context: can sleep 4094 */ 4095 void blk_mq_destroy_queue(struct request_queue *q) 4096 { 4097 WARN_ON_ONCE(!queue_is_mq(q)); 4098 WARN_ON_ONCE(blk_queue_registered(q)); 4099 4100 might_sleep(); 4101 4102 blk_queue_flag_set(QUEUE_FLAG_DYING, q); 4103 blk_queue_start_drain(q); 4104 blk_mq_freeze_queue_wait(q); 4105 4106 blk_sync_queue(q); 4107 blk_mq_cancel_work_sync(q); 4108 blk_mq_exit_queue(q); 4109 } 4110 EXPORT_SYMBOL(blk_mq_destroy_queue); 4111 4112 struct gendisk *__blk_mq_alloc_disk(struct blk_mq_tag_set *set, void *queuedata, 4113 struct lock_class_key *lkclass) 4114 { 4115 struct request_queue *q; 4116 struct gendisk *disk; 4117 4118 q = blk_mq_init_queue_data(set, queuedata); 4119 if (IS_ERR(q)) 4120 return ERR_CAST(q); 4121 4122 disk = __alloc_disk_node(q, set->numa_node, lkclass); 4123 if (!disk) { 4124 blk_mq_destroy_queue(q); 4125 blk_put_queue(q); 4126 return ERR_PTR(-ENOMEM); 4127 } 4128 set_bit(GD_OWNS_QUEUE, &disk->state); 4129 return disk; 4130 } 4131 EXPORT_SYMBOL(__blk_mq_alloc_disk); 4132 4133 struct gendisk *blk_mq_alloc_disk_for_queue(struct request_queue *q, 4134 struct lock_class_key *lkclass) 4135 { 4136 struct gendisk *disk; 4137 4138 if (!blk_get_queue(q)) 4139 return NULL; 4140 disk = __alloc_disk_node(q, NUMA_NO_NODE, lkclass); 4141 if (!disk) 4142 blk_put_queue(q); 4143 return disk; 4144 } 4145 EXPORT_SYMBOL(blk_mq_alloc_disk_for_queue); 4146 4147 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx( 4148 struct blk_mq_tag_set *set, struct request_queue *q, 4149 int hctx_idx, int node) 4150 { 4151 struct blk_mq_hw_ctx *hctx = NULL, *tmp; 4152 4153 /* reuse dead hctx first */ 4154 spin_lock(&q->unused_hctx_lock); 4155 list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) { 4156 if (tmp->numa_node == node) { 4157 hctx = tmp; 4158 break; 4159 } 4160 } 4161 if (hctx) 4162 list_del_init(&hctx->hctx_list); 4163 spin_unlock(&q->unused_hctx_lock); 4164 4165 if (!hctx) 4166 hctx = blk_mq_alloc_hctx(q, set, node); 4167 if (!hctx) 4168 goto fail; 4169 4170 if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) 4171 goto free_hctx; 4172 4173 return hctx; 4174 4175 free_hctx: 4176 kobject_put(&hctx->kobj); 4177 fail: 4178 return NULL; 4179 } 4180 4181 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set, 4182 struct request_queue *q) 4183 { 4184 struct blk_mq_hw_ctx *hctx; 4185 unsigned long i, j; 4186 4187 /* protect against switching io scheduler */ 4188 mutex_lock(&q->sysfs_lock); 4189 for (i = 0; i < set->nr_hw_queues; i++) { 4190 int old_node; 4191 int node = blk_mq_get_hctx_node(set, i); 4192 struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i); 4193 4194 if (old_hctx) { 4195 old_node = old_hctx->numa_node; 4196 blk_mq_exit_hctx(q, set, old_hctx, i); 4197 } 4198 4199 if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) { 4200 if (!old_hctx) 4201 break; 4202 pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n", 4203 node, old_node); 4204 hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node); 4205 WARN_ON_ONCE(!hctx); 4206 } 4207 } 4208 /* 4209 * Increasing nr_hw_queues fails. Free the newly allocated 4210 * hctxs and keep the previous q->nr_hw_queues. 4211 */ 4212 if (i != set->nr_hw_queues) { 4213 j = q->nr_hw_queues; 4214 } else { 4215 j = i; 4216 q->nr_hw_queues = set->nr_hw_queues; 4217 } 4218 4219 xa_for_each_start(&q->hctx_table, j, hctx, j) 4220 blk_mq_exit_hctx(q, set, hctx, j); 4221 mutex_unlock(&q->sysfs_lock); 4222 } 4223 4224 static void blk_mq_update_poll_flag(struct request_queue *q) 4225 { 4226 struct blk_mq_tag_set *set = q->tag_set; 4227 4228 if (set->nr_maps > HCTX_TYPE_POLL && 4229 set->map[HCTX_TYPE_POLL].nr_queues) 4230 blk_queue_flag_set(QUEUE_FLAG_POLL, q); 4231 else 4232 blk_queue_flag_clear(QUEUE_FLAG_POLL, q); 4233 } 4234 4235 int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set, 4236 struct request_queue *q) 4237 { 4238 /* mark the queue as mq asap */ 4239 q->mq_ops = set->ops; 4240 4241 if (blk_mq_alloc_ctxs(q)) 4242 goto err_exit; 4243 4244 /* init q->mq_kobj and sw queues' kobjects */ 4245 blk_mq_sysfs_init(q); 4246 4247 INIT_LIST_HEAD(&q->unused_hctx_list); 4248 spin_lock_init(&q->unused_hctx_lock); 4249 4250 xa_init(&q->hctx_table); 4251 4252 blk_mq_realloc_hw_ctxs(set, q); 4253 if (!q->nr_hw_queues) 4254 goto err_hctxs; 4255 4256 INIT_WORK(&q->timeout_work, blk_mq_timeout_work); 4257 blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ); 4258 4259 q->tag_set = set; 4260 4261 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT; 4262 blk_mq_update_poll_flag(q); 4263 4264 INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work); 4265 INIT_LIST_HEAD(&q->flush_list); 4266 INIT_LIST_HEAD(&q->requeue_list); 4267 spin_lock_init(&q->requeue_lock); 4268 4269 q->nr_requests = set->queue_depth; 4270 4271 blk_mq_init_cpu_queues(q, set->nr_hw_queues); 4272 blk_mq_add_queue_tag_set(set, q); 4273 blk_mq_map_swqueue(q); 4274 return 0; 4275 4276 err_hctxs: 4277 blk_mq_release(q); 4278 err_exit: 4279 q->mq_ops = NULL; 4280 return -ENOMEM; 4281 } 4282 EXPORT_SYMBOL(blk_mq_init_allocated_queue); 4283 4284 /* tags can _not_ be used after returning from blk_mq_exit_queue */ 4285 void blk_mq_exit_queue(struct request_queue *q) 4286 { 4287 struct blk_mq_tag_set *set = q->tag_set; 4288 4289 /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */ 4290 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues); 4291 /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */ 4292 blk_mq_del_queue_tag_set(q); 4293 } 4294 4295 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set) 4296 { 4297 int i; 4298 4299 if (blk_mq_is_shared_tags(set->flags)) { 4300 set->shared_tags = blk_mq_alloc_map_and_rqs(set, 4301 BLK_MQ_NO_HCTX_IDX, 4302 set->queue_depth); 4303 if (!set->shared_tags) 4304 return -ENOMEM; 4305 } 4306 4307 for (i = 0; i < set->nr_hw_queues; i++) { 4308 if (!__blk_mq_alloc_map_and_rqs(set, i)) 4309 goto out_unwind; 4310 cond_resched(); 4311 } 4312 4313 return 0; 4314 4315 out_unwind: 4316 while (--i >= 0) 4317 __blk_mq_free_map_and_rqs(set, i); 4318 4319 if (blk_mq_is_shared_tags(set->flags)) { 4320 blk_mq_free_map_and_rqs(set, set->shared_tags, 4321 BLK_MQ_NO_HCTX_IDX); 4322 } 4323 4324 return -ENOMEM; 4325 } 4326 4327 /* 4328 * Allocate the request maps associated with this tag_set. Note that this 4329 * may reduce the depth asked for, if memory is tight. set->queue_depth 4330 * will be updated to reflect the allocated depth. 4331 */ 4332 static int blk_mq_alloc_set_map_and_rqs(struct blk_mq_tag_set *set) 4333 { 4334 unsigned int depth; 4335 int err; 4336 4337 depth = set->queue_depth; 4338 do { 4339 err = __blk_mq_alloc_rq_maps(set); 4340 if (!err) 4341 break; 4342 4343 set->queue_depth >>= 1; 4344 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) { 4345 err = -ENOMEM; 4346 break; 4347 } 4348 } while (set->queue_depth); 4349 4350 if (!set->queue_depth || err) { 4351 pr_err("blk-mq: failed to allocate request map\n"); 4352 return -ENOMEM; 4353 } 4354 4355 if (depth != set->queue_depth) 4356 pr_info("blk-mq: reduced tag depth (%u -> %u)\n", 4357 depth, set->queue_depth); 4358 4359 return 0; 4360 } 4361 4362 static void blk_mq_update_queue_map(struct blk_mq_tag_set *set) 4363 { 4364 /* 4365 * blk_mq_map_queues() and multiple .map_queues() implementations 4366 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the 4367 * number of hardware queues. 4368 */ 4369 if (set->nr_maps == 1) 4370 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues; 4371 4372 if (set->ops->map_queues && !is_kdump_kernel()) { 4373 int i; 4374 4375 /* 4376 * transport .map_queues is usually done in the following 4377 * way: 4378 * 4379 * for (queue = 0; queue < set->nr_hw_queues; queue++) { 4380 * mask = get_cpu_mask(queue) 4381 * for_each_cpu(cpu, mask) 4382 * set->map[x].mq_map[cpu] = queue; 4383 * } 4384 * 4385 * When we need to remap, the table has to be cleared for 4386 * killing stale mapping since one CPU may not be mapped 4387 * to any hw queue. 4388 */ 4389 for (i = 0; i < set->nr_maps; i++) 4390 blk_mq_clear_mq_map(&set->map[i]); 4391 4392 set->ops->map_queues(set); 4393 } else { 4394 BUG_ON(set->nr_maps > 1); 4395 blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]); 4396 } 4397 } 4398 4399 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set, 4400 int new_nr_hw_queues) 4401 { 4402 struct blk_mq_tags **new_tags; 4403 int i; 4404 4405 if (set->nr_hw_queues >= new_nr_hw_queues) 4406 goto done; 4407 4408 new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *), 4409 GFP_KERNEL, set->numa_node); 4410 if (!new_tags) 4411 return -ENOMEM; 4412 4413 if (set->tags) 4414 memcpy(new_tags, set->tags, set->nr_hw_queues * 4415 sizeof(*set->tags)); 4416 kfree(set->tags); 4417 set->tags = new_tags; 4418 4419 for (i = set->nr_hw_queues; i < new_nr_hw_queues; i++) { 4420 if (!__blk_mq_alloc_map_and_rqs(set, i)) { 4421 while (--i >= set->nr_hw_queues) 4422 __blk_mq_free_map_and_rqs(set, i); 4423 return -ENOMEM; 4424 } 4425 cond_resched(); 4426 } 4427 4428 done: 4429 set->nr_hw_queues = new_nr_hw_queues; 4430 return 0; 4431 } 4432 4433 /* 4434 * Alloc a tag set to be associated with one or more request queues. 4435 * May fail with EINVAL for various error conditions. May adjust the 4436 * requested depth down, if it's too large. In that case, the set 4437 * value will be stored in set->queue_depth. 4438 */ 4439 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) 4440 { 4441 int i, ret; 4442 4443 BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS); 4444 4445 if (!set->nr_hw_queues) 4446 return -EINVAL; 4447 if (!set->queue_depth) 4448 return -EINVAL; 4449 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) 4450 return -EINVAL; 4451 4452 if (!set->ops->queue_rq) 4453 return -EINVAL; 4454 4455 if (!set->ops->get_budget ^ !set->ops->put_budget) 4456 return -EINVAL; 4457 4458 if (set->queue_depth > BLK_MQ_MAX_DEPTH) { 4459 pr_info("blk-mq: reduced tag depth to %u\n", 4460 BLK_MQ_MAX_DEPTH); 4461 set->queue_depth = BLK_MQ_MAX_DEPTH; 4462 } 4463 4464 if (!set->nr_maps) 4465 set->nr_maps = 1; 4466 else if (set->nr_maps > HCTX_MAX_TYPES) 4467 return -EINVAL; 4468 4469 /* 4470 * If a crashdump is active, then we are potentially in a very 4471 * memory constrained environment. Limit us to 1 queue and 4472 * 64 tags to prevent using too much memory. 4473 */ 4474 if (is_kdump_kernel()) { 4475 set->nr_hw_queues = 1; 4476 set->nr_maps = 1; 4477 set->queue_depth = min(64U, set->queue_depth); 4478 } 4479 /* 4480 * There is no use for more h/w queues than cpus if we just have 4481 * a single map 4482 */ 4483 if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids) 4484 set->nr_hw_queues = nr_cpu_ids; 4485 4486 if (set->flags & BLK_MQ_F_BLOCKING) { 4487 set->srcu = kmalloc(sizeof(*set->srcu), GFP_KERNEL); 4488 if (!set->srcu) 4489 return -ENOMEM; 4490 ret = init_srcu_struct(set->srcu); 4491 if (ret) 4492 goto out_free_srcu; 4493 } 4494 4495 ret = -ENOMEM; 4496 set->tags = kcalloc_node(set->nr_hw_queues, 4497 sizeof(struct blk_mq_tags *), GFP_KERNEL, 4498 set->numa_node); 4499 if (!set->tags) 4500 goto out_cleanup_srcu; 4501 4502 for (i = 0; i < set->nr_maps; i++) { 4503 set->map[i].mq_map = kcalloc_node(nr_cpu_ids, 4504 sizeof(set->map[i].mq_map[0]), 4505 GFP_KERNEL, set->numa_node); 4506 if (!set->map[i].mq_map) 4507 goto out_free_mq_map; 4508 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues; 4509 } 4510 4511 blk_mq_update_queue_map(set); 4512 4513 ret = blk_mq_alloc_set_map_and_rqs(set); 4514 if (ret) 4515 goto out_free_mq_map; 4516 4517 mutex_init(&set->tag_list_lock); 4518 INIT_LIST_HEAD(&set->tag_list); 4519 4520 return 0; 4521 4522 out_free_mq_map: 4523 for (i = 0; i < set->nr_maps; i++) { 4524 kfree(set->map[i].mq_map); 4525 set->map[i].mq_map = NULL; 4526 } 4527 kfree(set->tags); 4528 set->tags = NULL; 4529 out_cleanup_srcu: 4530 if (set->flags & BLK_MQ_F_BLOCKING) 4531 cleanup_srcu_struct(set->srcu); 4532 out_free_srcu: 4533 if (set->flags & BLK_MQ_F_BLOCKING) 4534 kfree(set->srcu); 4535 return ret; 4536 } 4537 EXPORT_SYMBOL(blk_mq_alloc_tag_set); 4538 4539 /* allocate and initialize a tagset for a simple single-queue device */ 4540 int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, 4541 const struct blk_mq_ops *ops, unsigned int queue_depth, 4542 unsigned int set_flags) 4543 { 4544 memset(set, 0, sizeof(*set)); 4545 set->ops = ops; 4546 set->nr_hw_queues = 1; 4547 set->nr_maps = 1; 4548 set->queue_depth = queue_depth; 4549 set->numa_node = NUMA_NO_NODE; 4550 set->flags = set_flags; 4551 return blk_mq_alloc_tag_set(set); 4552 } 4553 EXPORT_SYMBOL_GPL(blk_mq_alloc_sq_tag_set); 4554 4555 void blk_mq_free_tag_set(struct blk_mq_tag_set *set) 4556 { 4557 int i, j; 4558 4559 for (i = 0; i < set->nr_hw_queues; i++) 4560 __blk_mq_free_map_and_rqs(set, i); 4561 4562 if (blk_mq_is_shared_tags(set->flags)) { 4563 blk_mq_free_map_and_rqs(set, set->shared_tags, 4564 BLK_MQ_NO_HCTX_IDX); 4565 } 4566 4567 for (j = 0; j < set->nr_maps; j++) { 4568 kfree(set->map[j].mq_map); 4569 set->map[j].mq_map = NULL; 4570 } 4571 4572 kfree(set->tags); 4573 set->tags = NULL; 4574 if (set->flags & BLK_MQ_F_BLOCKING) { 4575 cleanup_srcu_struct(set->srcu); 4576 kfree(set->srcu); 4577 } 4578 } 4579 EXPORT_SYMBOL(blk_mq_free_tag_set); 4580 4581 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr) 4582 { 4583 struct blk_mq_tag_set *set = q->tag_set; 4584 struct blk_mq_hw_ctx *hctx; 4585 int ret; 4586 unsigned long i; 4587 4588 if (!set) 4589 return -EINVAL; 4590 4591 if (q->nr_requests == nr) 4592 return 0; 4593 4594 blk_mq_freeze_queue(q); 4595 blk_mq_quiesce_queue(q); 4596 4597 ret = 0; 4598 queue_for_each_hw_ctx(q, hctx, i) { 4599 if (!hctx->tags) 4600 continue; 4601 /* 4602 * If we're using an MQ scheduler, just update the scheduler 4603 * queue depth. This is similar to what the old code would do. 4604 */ 4605 if (hctx->sched_tags) { 4606 ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags, 4607 nr, true); 4608 } else { 4609 ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr, 4610 false); 4611 } 4612 if (ret) 4613 break; 4614 if (q->elevator && q->elevator->type->ops.depth_updated) 4615 q->elevator->type->ops.depth_updated(hctx); 4616 } 4617 if (!ret) { 4618 q->nr_requests = nr; 4619 if (blk_mq_is_shared_tags(set->flags)) { 4620 if (q->elevator) 4621 blk_mq_tag_update_sched_shared_tags(q); 4622 else 4623 blk_mq_tag_resize_shared_tags(set, nr); 4624 } 4625 } 4626 4627 blk_mq_unquiesce_queue(q); 4628 blk_mq_unfreeze_queue(q); 4629 4630 return ret; 4631 } 4632 4633 /* 4634 * request_queue and elevator_type pair. 4635 * It is just used by __blk_mq_update_nr_hw_queues to cache 4636 * the elevator_type associated with a request_queue. 4637 */ 4638 struct blk_mq_qe_pair { 4639 struct list_head node; 4640 struct request_queue *q; 4641 struct elevator_type *type; 4642 }; 4643 4644 /* 4645 * Cache the elevator_type in qe pair list and switch the 4646 * io scheduler to 'none' 4647 */ 4648 static bool blk_mq_elv_switch_none(struct list_head *head, 4649 struct request_queue *q) 4650 { 4651 struct blk_mq_qe_pair *qe; 4652 4653 qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY); 4654 if (!qe) 4655 return false; 4656 4657 /* q->elevator needs protection from ->sysfs_lock */ 4658 mutex_lock(&q->sysfs_lock); 4659 4660 /* the check has to be done with holding sysfs_lock */ 4661 if (!q->elevator) { 4662 kfree(qe); 4663 goto unlock; 4664 } 4665 4666 INIT_LIST_HEAD(&qe->node); 4667 qe->q = q; 4668 qe->type = q->elevator->type; 4669 /* keep a reference to the elevator module as we'll switch back */ 4670 __elevator_get(qe->type); 4671 list_add(&qe->node, head); 4672 elevator_disable(q); 4673 unlock: 4674 mutex_unlock(&q->sysfs_lock); 4675 4676 return true; 4677 } 4678 4679 static struct blk_mq_qe_pair *blk_lookup_qe_pair(struct list_head *head, 4680 struct request_queue *q) 4681 { 4682 struct blk_mq_qe_pair *qe; 4683 4684 list_for_each_entry(qe, head, node) 4685 if (qe->q == q) 4686 return qe; 4687 4688 return NULL; 4689 } 4690 4691 static void blk_mq_elv_switch_back(struct list_head *head, 4692 struct request_queue *q) 4693 { 4694 struct blk_mq_qe_pair *qe; 4695 struct elevator_type *t; 4696 4697 qe = blk_lookup_qe_pair(head, q); 4698 if (!qe) 4699 return; 4700 t = qe->type; 4701 list_del(&qe->node); 4702 kfree(qe); 4703 4704 mutex_lock(&q->sysfs_lock); 4705 elevator_switch(q, t); 4706 /* drop the reference acquired in blk_mq_elv_switch_none */ 4707 elevator_put(t); 4708 mutex_unlock(&q->sysfs_lock); 4709 } 4710 4711 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, 4712 int nr_hw_queues) 4713 { 4714 struct request_queue *q; 4715 LIST_HEAD(head); 4716 int prev_nr_hw_queues = set->nr_hw_queues; 4717 int i; 4718 4719 lockdep_assert_held(&set->tag_list_lock); 4720 4721 if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids) 4722 nr_hw_queues = nr_cpu_ids; 4723 if (nr_hw_queues < 1) 4724 return; 4725 if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues) 4726 return; 4727 4728 list_for_each_entry(q, &set->tag_list, tag_set_list) 4729 blk_mq_freeze_queue(q); 4730 /* 4731 * Switch IO scheduler to 'none', cleaning up the data associated 4732 * with the previous scheduler. We will switch back once we are done 4733 * updating the new sw to hw queue mappings. 4734 */ 4735 list_for_each_entry(q, &set->tag_list, tag_set_list) 4736 if (!blk_mq_elv_switch_none(&head, q)) 4737 goto switch_back; 4738 4739 list_for_each_entry(q, &set->tag_list, tag_set_list) { 4740 blk_mq_debugfs_unregister_hctxs(q); 4741 blk_mq_sysfs_unregister_hctxs(q); 4742 } 4743 4744 if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) 4745 goto reregister; 4746 4747 fallback: 4748 blk_mq_update_queue_map(set); 4749 list_for_each_entry(q, &set->tag_list, tag_set_list) { 4750 blk_mq_realloc_hw_ctxs(set, q); 4751 blk_mq_update_poll_flag(q); 4752 if (q->nr_hw_queues != set->nr_hw_queues) { 4753 int i = prev_nr_hw_queues; 4754 4755 pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n", 4756 nr_hw_queues, prev_nr_hw_queues); 4757 for (; i < set->nr_hw_queues; i++) 4758 __blk_mq_free_map_and_rqs(set, i); 4759 4760 set->nr_hw_queues = prev_nr_hw_queues; 4761 goto fallback; 4762 } 4763 blk_mq_map_swqueue(q); 4764 } 4765 4766 reregister: 4767 list_for_each_entry(q, &set->tag_list, tag_set_list) { 4768 blk_mq_sysfs_register_hctxs(q); 4769 blk_mq_debugfs_register_hctxs(q); 4770 } 4771 4772 switch_back: 4773 list_for_each_entry(q, &set->tag_list, tag_set_list) 4774 blk_mq_elv_switch_back(&head, q); 4775 4776 list_for_each_entry(q, &set->tag_list, tag_set_list) 4777 blk_mq_unfreeze_queue(q); 4778 4779 /* Free the excess tags when nr_hw_queues shrink. */ 4780 for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++) 4781 __blk_mq_free_map_and_rqs(set, i); 4782 } 4783 4784 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues) 4785 { 4786 mutex_lock(&set->tag_list_lock); 4787 __blk_mq_update_nr_hw_queues(set, nr_hw_queues); 4788 mutex_unlock(&set->tag_list_lock); 4789 } 4790 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues); 4791 4792 static int blk_hctx_poll(struct request_queue *q, struct blk_mq_hw_ctx *hctx, 4793 struct io_comp_batch *iob, unsigned int flags) 4794 { 4795 long state = get_current_state(); 4796 int ret; 4797 4798 do { 4799 ret = q->mq_ops->poll(hctx, iob); 4800 if (ret > 0) { 4801 __set_current_state(TASK_RUNNING); 4802 return ret; 4803 } 4804 4805 if (signal_pending_state(state, current)) 4806 __set_current_state(TASK_RUNNING); 4807 if (task_is_running(current)) 4808 return 1; 4809 4810 if (ret < 0 || (flags & BLK_POLL_ONESHOT)) 4811 break; 4812 cpu_relax(); 4813 } while (!need_resched()); 4814 4815 __set_current_state(TASK_RUNNING); 4816 return 0; 4817 } 4818 4819 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, 4820 struct io_comp_batch *iob, unsigned int flags) 4821 { 4822 struct blk_mq_hw_ctx *hctx = xa_load(&q->hctx_table, cookie); 4823 4824 return blk_hctx_poll(q, hctx, iob, flags); 4825 } 4826 4827 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob, 4828 unsigned int poll_flags) 4829 { 4830 struct request_queue *q = rq->q; 4831 int ret; 4832 4833 if (!blk_rq_is_poll(rq)) 4834 return 0; 4835 if (!percpu_ref_tryget(&q->q_usage_counter)) 4836 return 0; 4837 4838 ret = blk_hctx_poll(q, rq->mq_hctx, iob, poll_flags); 4839 blk_queue_exit(q); 4840 4841 return ret; 4842 } 4843 EXPORT_SYMBOL_GPL(blk_rq_poll); 4844 4845 unsigned int blk_mq_rq_cpu(struct request *rq) 4846 { 4847 return rq->mq_ctx->cpu; 4848 } 4849 EXPORT_SYMBOL(blk_mq_rq_cpu); 4850 4851 void blk_mq_cancel_work_sync(struct request_queue *q) 4852 { 4853 struct blk_mq_hw_ctx *hctx; 4854 unsigned long i; 4855 4856 cancel_delayed_work_sync(&q->requeue_work); 4857 4858 queue_for_each_hw_ctx(q, hctx, i) 4859 cancel_delayed_work_sync(&hctx->run_work); 4860 } 4861 4862 static int __init blk_mq_init(void) 4863 { 4864 int i; 4865 4866 for_each_possible_cpu(i) 4867 init_llist_head(&per_cpu(blk_cpu_done, i)); 4868 for_each_possible_cpu(i) 4869 INIT_CSD(&per_cpu(blk_cpu_csd, i), 4870 __blk_mq_complete_request_remote, NULL); 4871 open_softirq(BLOCK_SOFTIRQ, blk_done_softirq); 4872 4873 cpuhp_setup_state_nocalls(CPUHP_BLOCK_SOFTIRQ_DEAD, 4874 "block/softirq:dead", NULL, 4875 blk_softirq_cpu_dead); 4876 cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL, 4877 blk_mq_hctx_notify_dead); 4878 cpuhp_setup_state_multi(CPUHP_AP_BLK_MQ_ONLINE, "block/mq:online", 4879 blk_mq_hctx_notify_online, 4880 blk_mq_hctx_notify_offline); 4881 return 0; 4882 } 4883 subsys_initcall(blk_mq_init); 4884