1 /* 2 * Copyright (C) 1991, 1992 Linus Torvalds 3 * Copyright (C) 1994, Karl Keyte: Added support for disk statistics 4 * Elevator latency, (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE 5 * Queue request tables / lock, selectable elevator, Jens Axboe <axboe@suse.de> 6 * kernel-doc documentation started by NeilBrown <neilb@cse.unsw.edu.au> 7 * - July2000 8 * bio rewrite, highmem i/o, etc, Jens Axboe <axboe@suse.de> - may 2001 9 */ 10 11 /* 12 * This handles all read/write requests to block devices 13 */ 14 #include <linux/kernel.h> 15 #include <linux/module.h> 16 #include <linux/backing-dev.h> 17 #include <linux/bio.h> 18 #include <linux/blkdev.h> 19 #include <linux/blk-mq.h> 20 #include <linux/highmem.h> 21 #include <linux/mm.h> 22 #include <linux/kernel_stat.h> 23 #include <linux/string.h> 24 #include <linux/init.h> 25 #include <linux/completion.h> 26 #include <linux/slab.h> 27 #include <linux/swap.h> 28 #include <linux/writeback.h> 29 #include <linux/task_io_accounting_ops.h> 30 #include <linux/fault-inject.h> 31 #include <linux/list_sort.h> 32 #include <linux/delay.h> 33 #include <linux/ratelimit.h> 34 #include <linux/pm_runtime.h> 35 #include <linux/blk-cgroup.h> 36 #include <linux/debugfs.h> 37 #include <linux/bpf.h> 38 39 #define CREATE_TRACE_POINTS 40 #include <trace/events/block.h> 41 42 #include "blk.h" 43 #include "blk-mq.h" 44 #include "blk-mq-sched.h" 45 #include "blk-pm.h" 46 #include "blk-rq-qos.h" 47 48 #ifdef CONFIG_DEBUG_FS 49 struct dentry *blk_debugfs_root; 50 #endif 51 52 EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_remap); 53 EXPORT_TRACEPOINT_SYMBOL_GPL(block_rq_remap); 54 EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_complete); 55 EXPORT_TRACEPOINT_SYMBOL_GPL(block_split); 56 EXPORT_TRACEPOINT_SYMBOL_GPL(block_unplug); 57 58 DEFINE_IDA(blk_queue_ida); 59 60 /* 61 * For the allocated request tables 62 */ 63 struct kmem_cache *request_cachep; 64 65 /* 66 * For queue allocation 67 */ 68 struct kmem_cache *blk_requestq_cachep; 69 70 /* 71 * Controlling structure to kblockd 72 */ 73 static struct workqueue_struct *kblockd_workqueue; 74 75 /** 76 * blk_queue_flag_set - atomically set a queue flag 77 * @flag: flag to be set 78 * @q: request queue 79 */ 80 void blk_queue_flag_set(unsigned int flag, struct request_queue *q) 81 { 82 unsigned long flags; 83 84 spin_lock_irqsave(q->queue_lock, flags); 85 queue_flag_set(flag, q); 86 spin_unlock_irqrestore(q->queue_lock, flags); 87 } 88 EXPORT_SYMBOL(blk_queue_flag_set); 89 90 /** 91 * blk_queue_flag_clear - atomically clear a queue flag 92 * @flag: flag to be cleared 93 * @q: request queue 94 */ 95 void blk_queue_flag_clear(unsigned int flag, struct request_queue *q) 96 { 97 unsigned long flags; 98 99 spin_lock_irqsave(q->queue_lock, flags); 100 queue_flag_clear(flag, q); 101 spin_unlock_irqrestore(q->queue_lock, flags); 102 } 103 EXPORT_SYMBOL(blk_queue_flag_clear); 104 105 /** 106 * blk_queue_flag_test_and_set - atomically test and set a queue flag 107 * @flag: flag to be set 108 * @q: request queue 109 * 110 * Returns the previous value of @flag - 0 if the flag was not set and 1 if 111 * the flag was already set. 112 */ 113 bool blk_queue_flag_test_and_set(unsigned int flag, struct request_queue *q) 114 { 115 unsigned long flags; 116 bool res; 117 118 spin_lock_irqsave(q->queue_lock, flags); 119 res = queue_flag_test_and_set(flag, q); 120 spin_unlock_irqrestore(q->queue_lock, flags); 121 122 return res; 123 } 124 EXPORT_SYMBOL_GPL(blk_queue_flag_test_and_set); 125 126 /** 127 * blk_queue_flag_test_and_clear - atomically test and clear a queue flag 128 * @flag: flag to be cleared 129 * @q: request queue 130 * 131 * Returns the previous value of @flag - 0 if the flag was not set and 1 if 132 * the flag was set. 133 */ 134 bool blk_queue_flag_test_and_clear(unsigned int flag, struct request_queue *q) 135 { 136 unsigned long flags; 137 bool res; 138 139 spin_lock_irqsave(q->queue_lock, flags); 140 res = queue_flag_test_and_clear(flag, q); 141 spin_unlock_irqrestore(q->queue_lock, flags); 142 143 return res; 144 } 145 EXPORT_SYMBOL_GPL(blk_queue_flag_test_and_clear); 146 147 static void blk_clear_congested(struct request_list *rl, int sync) 148 { 149 #ifdef CONFIG_CGROUP_WRITEBACK 150 clear_wb_congested(rl->blkg->wb_congested, sync); 151 #else 152 /* 153 * If !CGROUP_WRITEBACK, all blkg's map to bdi->wb and we shouldn't 154 * flip its congestion state for events on other blkcgs. 155 */ 156 if (rl == &rl->q->root_rl) 157 clear_wb_congested(rl->q->backing_dev_info->wb.congested, sync); 158 #endif 159 } 160 161 static void blk_set_congested(struct request_list *rl, int sync) 162 { 163 #ifdef CONFIG_CGROUP_WRITEBACK 164 set_wb_congested(rl->blkg->wb_congested, sync); 165 #else 166 /* see blk_clear_congested() */ 167 if (rl == &rl->q->root_rl) 168 set_wb_congested(rl->q->backing_dev_info->wb.congested, sync); 169 #endif 170 } 171 172 void blk_queue_congestion_threshold(struct request_queue *q) 173 { 174 int nr; 175 176 nr = q->nr_requests - (q->nr_requests / 8) + 1; 177 if (nr > q->nr_requests) 178 nr = q->nr_requests; 179 q->nr_congestion_on = nr; 180 181 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1; 182 if (nr < 1) 183 nr = 1; 184 q->nr_congestion_off = nr; 185 } 186 187 void blk_rq_init(struct request_queue *q, struct request *rq) 188 { 189 memset(rq, 0, sizeof(*rq)); 190 191 INIT_LIST_HEAD(&rq->queuelist); 192 INIT_LIST_HEAD(&rq->timeout_list); 193 rq->cpu = -1; 194 rq->q = q; 195 rq->__sector = (sector_t) -1; 196 INIT_HLIST_NODE(&rq->hash); 197 RB_CLEAR_NODE(&rq->rb_node); 198 rq->tag = -1; 199 rq->internal_tag = -1; 200 rq->start_time_ns = ktime_get_ns(); 201 rq->part = NULL; 202 } 203 EXPORT_SYMBOL(blk_rq_init); 204 205 static const struct { 206 int errno; 207 const char *name; 208 } blk_errors[] = { 209 [BLK_STS_OK] = { 0, "" }, 210 [BLK_STS_NOTSUPP] = { -EOPNOTSUPP, "operation not supported" }, 211 [BLK_STS_TIMEOUT] = { -ETIMEDOUT, "timeout" }, 212 [BLK_STS_NOSPC] = { -ENOSPC, "critical space allocation" }, 213 [BLK_STS_TRANSPORT] = { -ENOLINK, "recoverable transport" }, 214 [BLK_STS_TARGET] = { -EREMOTEIO, "critical target" }, 215 [BLK_STS_NEXUS] = { -EBADE, "critical nexus" }, 216 [BLK_STS_MEDIUM] = { -ENODATA, "critical medium" }, 217 [BLK_STS_PROTECTION] = { -EILSEQ, "protection" }, 218 [BLK_STS_RESOURCE] = { -ENOMEM, "kernel resource" }, 219 [BLK_STS_DEV_RESOURCE] = { -EBUSY, "device resource" }, 220 [BLK_STS_AGAIN] = { -EAGAIN, "nonblocking retry" }, 221 222 /* device mapper special case, should not leak out: */ 223 [BLK_STS_DM_REQUEUE] = { -EREMCHG, "dm internal retry" }, 224 225 /* everything else not covered above: */ 226 [BLK_STS_IOERR] = { -EIO, "I/O" }, 227 }; 228 229 blk_status_t errno_to_blk_status(int errno) 230 { 231 int i; 232 233 for (i = 0; i < ARRAY_SIZE(blk_errors); i++) { 234 if (blk_errors[i].errno == errno) 235 return (__force blk_status_t)i; 236 } 237 238 return BLK_STS_IOERR; 239 } 240 EXPORT_SYMBOL_GPL(errno_to_blk_status); 241 242 int blk_status_to_errno(blk_status_t status) 243 { 244 int idx = (__force int)status; 245 246 if (WARN_ON_ONCE(idx >= ARRAY_SIZE(blk_errors))) 247 return -EIO; 248 return blk_errors[idx].errno; 249 } 250 EXPORT_SYMBOL_GPL(blk_status_to_errno); 251 252 static void print_req_error(struct request *req, blk_status_t status) 253 { 254 int idx = (__force int)status; 255 256 if (WARN_ON_ONCE(idx >= ARRAY_SIZE(blk_errors))) 257 return; 258 259 printk_ratelimited(KERN_ERR "%s: %s error, dev %s, sector %llu\n", 260 __func__, blk_errors[idx].name, req->rq_disk ? 261 req->rq_disk->disk_name : "?", 262 (unsigned long long)blk_rq_pos(req)); 263 } 264 265 static void req_bio_endio(struct request *rq, struct bio *bio, 266 unsigned int nbytes, blk_status_t error) 267 { 268 if (error) 269 bio->bi_status = error; 270 271 if (unlikely(rq->rq_flags & RQF_QUIET)) 272 bio_set_flag(bio, BIO_QUIET); 273 274 bio_advance(bio, nbytes); 275 276 /* don't actually finish bio if it's part of flush sequence */ 277 if (bio->bi_iter.bi_size == 0 && !(rq->rq_flags & RQF_FLUSH_SEQ)) 278 bio_endio(bio); 279 } 280 281 void blk_dump_rq_flags(struct request *rq, char *msg) 282 { 283 printk(KERN_INFO "%s: dev %s: flags=%llx\n", msg, 284 rq->rq_disk ? rq->rq_disk->disk_name : "?", 285 (unsigned long long) rq->cmd_flags); 286 287 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n", 288 (unsigned long long)blk_rq_pos(rq), 289 blk_rq_sectors(rq), blk_rq_cur_sectors(rq)); 290 printk(KERN_INFO " bio %p, biotail %p, len %u\n", 291 rq->bio, rq->biotail, blk_rq_bytes(rq)); 292 } 293 EXPORT_SYMBOL(blk_dump_rq_flags); 294 295 static void blk_delay_work(struct work_struct *work) 296 { 297 struct request_queue *q; 298 299 q = container_of(work, struct request_queue, delay_work.work); 300 spin_lock_irq(q->queue_lock); 301 __blk_run_queue(q); 302 spin_unlock_irq(q->queue_lock); 303 } 304 305 /** 306 * blk_delay_queue - restart queueing after defined interval 307 * @q: The &struct request_queue in question 308 * @msecs: Delay in msecs 309 * 310 * Description: 311 * Sometimes queueing needs to be postponed for a little while, to allow 312 * resources to come back. This function will make sure that queueing is 313 * restarted around the specified time. 314 */ 315 void blk_delay_queue(struct request_queue *q, unsigned long msecs) 316 { 317 lockdep_assert_held(q->queue_lock); 318 WARN_ON_ONCE(q->mq_ops); 319 320 if (likely(!blk_queue_dead(q))) 321 queue_delayed_work(kblockd_workqueue, &q->delay_work, 322 msecs_to_jiffies(msecs)); 323 } 324 EXPORT_SYMBOL(blk_delay_queue); 325 326 /** 327 * blk_start_queue_async - asynchronously restart a previously stopped queue 328 * @q: The &struct request_queue in question 329 * 330 * Description: 331 * blk_start_queue_async() will clear the stop flag on the queue, and 332 * ensure that the request_fn for the queue is run from an async 333 * context. 334 **/ 335 void blk_start_queue_async(struct request_queue *q) 336 { 337 lockdep_assert_held(q->queue_lock); 338 WARN_ON_ONCE(q->mq_ops); 339 340 queue_flag_clear(QUEUE_FLAG_STOPPED, q); 341 blk_run_queue_async(q); 342 } 343 EXPORT_SYMBOL(blk_start_queue_async); 344 345 /** 346 * blk_start_queue - restart a previously stopped queue 347 * @q: The &struct request_queue in question 348 * 349 * Description: 350 * blk_start_queue() will clear the stop flag on the queue, and call 351 * the request_fn for the queue if it was in a stopped state when 352 * entered. Also see blk_stop_queue(). 353 **/ 354 void blk_start_queue(struct request_queue *q) 355 { 356 lockdep_assert_held(q->queue_lock); 357 WARN_ON_ONCE(q->mq_ops); 358 359 queue_flag_clear(QUEUE_FLAG_STOPPED, q); 360 __blk_run_queue(q); 361 } 362 EXPORT_SYMBOL(blk_start_queue); 363 364 /** 365 * blk_stop_queue - stop a queue 366 * @q: The &struct request_queue in question 367 * 368 * Description: 369 * The Linux block layer assumes that a block driver will consume all 370 * entries on the request queue when the request_fn strategy is called. 371 * Often this will not happen, because of hardware limitations (queue 372 * depth settings). If a device driver gets a 'queue full' response, 373 * or if it simply chooses not to queue more I/O at one point, it can 374 * call this function to prevent the request_fn from being called until 375 * the driver has signalled it's ready to go again. This happens by calling 376 * blk_start_queue() to restart queue operations. 377 **/ 378 void blk_stop_queue(struct request_queue *q) 379 { 380 lockdep_assert_held(q->queue_lock); 381 WARN_ON_ONCE(q->mq_ops); 382 383 cancel_delayed_work(&q->delay_work); 384 queue_flag_set(QUEUE_FLAG_STOPPED, q); 385 } 386 EXPORT_SYMBOL(blk_stop_queue); 387 388 /** 389 * blk_sync_queue - cancel any pending callbacks on a queue 390 * @q: the queue 391 * 392 * Description: 393 * The block layer may perform asynchronous callback activity 394 * on a queue, such as calling the unplug function after a timeout. 395 * A block device may call blk_sync_queue to ensure that any 396 * such activity is cancelled, thus allowing it to release resources 397 * that the callbacks might use. The caller must already have made sure 398 * that its ->make_request_fn will not re-add plugging prior to calling 399 * this function. 400 * 401 * This function does not cancel any asynchronous activity arising 402 * out of elevator or throttling code. That would require elevator_exit() 403 * and blkcg_exit_queue() to be called with queue lock initialized. 404 * 405 */ 406 void blk_sync_queue(struct request_queue *q) 407 { 408 del_timer_sync(&q->timeout); 409 cancel_work_sync(&q->timeout_work); 410 411 if (q->mq_ops) { 412 struct blk_mq_hw_ctx *hctx; 413 int i; 414 415 cancel_delayed_work_sync(&q->requeue_work); 416 queue_for_each_hw_ctx(q, hctx, i) 417 cancel_delayed_work_sync(&hctx->run_work); 418 } else { 419 cancel_delayed_work_sync(&q->delay_work); 420 } 421 } 422 EXPORT_SYMBOL(blk_sync_queue); 423 424 /** 425 * blk_set_pm_only - increment pm_only counter 426 * @q: request queue pointer 427 */ 428 void blk_set_pm_only(struct request_queue *q) 429 { 430 atomic_inc(&q->pm_only); 431 } 432 EXPORT_SYMBOL_GPL(blk_set_pm_only); 433 434 void blk_clear_pm_only(struct request_queue *q) 435 { 436 int pm_only; 437 438 pm_only = atomic_dec_return(&q->pm_only); 439 WARN_ON_ONCE(pm_only < 0); 440 if (pm_only == 0) 441 wake_up_all(&q->mq_freeze_wq); 442 } 443 EXPORT_SYMBOL_GPL(blk_clear_pm_only); 444 445 /** 446 * __blk_run_queue_uncond - run a queue whether or not it has been stopped 447 * @q: The queue to run 448 * 449 * Description: 450 * Invoke request handling on a queue if there are any pending requests. 451 * May be used to restart request handling after a request has completed. 452 * This variant runs the queue whether or not the queue has been 453 * stopped. Must be called with the queue lock held and interrupts 454 * disabled. See also @blk_run_queue. 455 */ 456 inline void __blk_run_queue_uncond(struct request_queue *q) 457 { 458 lockdep_assert_held(q->queue_lock); 459 WARN_ON_ONCE(q->mq_ops); 460 461 if (unlikely(blk_queue_dead(q))) 462 return; 463 464 /* 465 * Some request_fn implementations, e.g. scsi_request_fn(), unlock 466 * the queue lock internally. As a result multiple threads may be 467 * running such a request function concurrently. Keep track of the 468 * number of active request_fn invocations such that blk_drain_queue() 469 * can wait until all these request_fn calls have finished. 470 */ 471 q->request_fn_active++; 472 q->request_fn(q); 473 q->request_fn_active--; 474 } 475 EXPORT_SYMBOL_GPL(__blk_run_queue_uncond); 476 477 /** 478 * __blk_run_queue - run a single device queue 479 * @q: The queue to run 480 * 481 * Description: 482 * See @blk_run_queue. 483 */ 484 void __blk_run_queue(struct request_queue *q) 485 { 486 lockdep_assert_held(q->queue_lock); 487 WARN_ON_ONCE(q->mq_ops); 488 489 if (unlikely(blk_queue_stopped(q))) 490 return; 491 492 __blk_run_queue_uncond(q); 493 } 494 EXPORT_SYMBOL(__blk_run_queue); 495 496 /** 497 * blk_run_queue_async - run a single device queue in workqueue context 498 * @q: The queue to run 499 * 500 * Description: 501 * Tells kblockd to perform the equivalent of @blk_run_queue on behalf 502 * of us. 503 * 504 * Note: 505 * Since it is not allowed to run q->delay_work after blk_cleanup_queue() 506 * has canceled q->delay_work, callers must hold the queue lock to avoid 507 * race conditions between blk_cleanup_queue() and blk_run_queue_async(). 508 */ 509 void blk_run_queue_async(struct request_queue *q) 510 { 511 lockdep_assert_held(q->queue_lock); 512 WARN_ON_ONCE(q->mq_ops); 513 514 if (likely(!blk_queue_stopped(q) && !blk_queue_dead(q))) 515 mod_delayed_work(kblockd_workqueue, &q->delay_work, 0); 516 } 517 EXPORT_SYMBOL(blk_run_queue_async); 518 519 /** 520 * blk_run_queue - run a single device queue 521 * @q: The queue to run 522 * 523 * Description: 524 * Invoke request handling on this queue, if it has pending work to do. 525 * May be used to restart queueing when a request has completed. 526 */ 527 void blk_run_queue(struct request_queue *q) 528 { 529 unsigned long flags; 530 531 WARN_ON_ONCE(q->mq_ops); 532 533 spin_lock_irqsave(q->queue_lock, flags); 534 __blk_run_queue(q); 535 spin_unlock_irqrestore(q->queue_lock, flags); 536 } 537 EXPORT_SYMBOL(blk_run_queue); 538 539 void blk_put_queue(struct request_queue *q) 540 { 541 kobject_put(&q->kobj); 542 } 543 EXPORT_SYMBOL(blk_put_queue); 544 545 /** 546 * __blk_drain_queue - drain requests from request_queue 547 * @q: queue to drain 548 * @drain_all: whether to drain all requests or only the ones w/ ELVPRIV 549 * 550 * Drain requests from @q. If @drain_all is set, all requests are drained. 551 * If not, only ELVPRIV requests are drained. The caller is responsible 552 * for ensuring that no new requests which need to be drained are queued. 553 */ 554 static void __blk_drain_queue(struct request_queue *q, bool drain_all) 555 __releases(q->queue_lock) 556 __acquires(q->queue_lock) 557 { 558 int i; 559 560 lockdep_assert_held(q->queue_lock); 561 WARN_ON_ONCE(q->mq_ops); 562 563 while (true) { 564 bool drain = false; 565 566 /* 567 * The caller might be trying to drain @q before its 568 * elevator is initialized. 569 */ 570 if (q->elevator) 571 elv_drain_elevator(q); 572 573 blkcg_drain_queue(q); 574 575 /* 576 * This function might be called on a queue which failed 577 * driver init after queue creation or is not yet fully 578 * active yet. Some drivers (e.g. fd and loop) get unhappy 579 * in such cases. Kick queue iff dispatch queue has 580 * something on it and @q has request_fn set. 581 */ 582 if (!list_empty(&q->queue_head) && q->request_fn) 583 __blk_run_queue(q); 584 585 drain |= q->nr_rqs_elvpriv; 586 drain |= q->request_fn_active; 587 588 /* 589 * Unfortunately, requests are queued at and tracked from 590 * multiple places and there's no single counter which can 591 * be drained. Check all the queues and counters. 592 */ 593 if (drain_all) { 594 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL); 595 drain |= !list_empty(&q->queue_head); 596 for (i = 0; i < 2; i++) { 597 drain |= q->nr_rqs[i]; 598 drain |= q->in_flight[i]; 599 if (fq) 600 drain |= !list_empty(&fq->flush_queue[i]); 601 } 602 } 603 604 if (!drain) 605 break; 606 607 spin_unlock_irq(q->queue_lock); 608 609 msleep(10); 610 611 spin_lock_irq(q->queue_lock); 612 } 613 614 /* 615 * With queue marked dead, any woken up waiter will fail the 616 * allocation path, so the wakeup chaining is lost and we're 617 * left with hung waiters. We need to wake up those waiters. 618 */ 619 if (q->request_fn) { 620 struct request_list *rl; 621 622 blk_queue_for_each_rl(rl, q) 623 for (i = 0; i < ARRAY_SIZE(rl->wait); i++) 624 wake_up_all(&rl->wait[i]); 625 } 626 } 627 628 void blk_drain_queue(struct request_queue *q) 629 { 630 spin_lock_irq(q->queue_lock); 631 __blk_drain_queue(q, true); 632 spin_unlock_irq(q->queue_lock); 633 } 634 635 /** 636 * blk_queue_bypass_start - enter queue bypass mode 637 * @q: queue of interest 638 * 639 * In bypass mode, only the dispatch FIFO queue of @q is used. This 640 * function makes @q enter bypass mode and drains all requests which were 641 * throttled or issued before. On return, it's guaranteed that no request 642 * is being throttled or has ELVPRIV set and blk_queue_bypass() %true 643 * inside queue or RCU read lock. 644 */ 645 void blk_queue_bypass_start(struct request_queue *q) 646 { 647 WARN_ON_ONCE(q->mq_ops); 648 649 spin_lock_irq(q->queue_lock); 650 q->bypass_depth++; 651 queue_flag_set(QUEUE_FLAG_BYPASS, q); 652 spin_unlock_irq(q->queue_lock); 653 654 /* 655 * Queues start drained. Skip actual draining till init is 656 * complete. This avoids lenghty delays during queue init which 657 * can happen many times during boot. 658 */ 659 if (blk_queue_init_done(q)) { 660 spin_lock_irq(q->queue_lock); 661 __blk_drain_queue(q, false); 662 spin_unlock_irq(q->queue_lock); 663 664 /* ensure blk_queue_bypass() is %true inside RCU read lock */ 665 synchronize_rcu(); 666 } 667 } 668 EXPORT_SYMBOL_GPL(blk_queue_bypass_start); 669 670 /** 671 * blk_queue_bypass_end - leave queue bypass mode 672 * @q: queue of interest 673 * 674 * Leave bypass mode and restore the normal queueing behavior. 675 * 676 * Note: although blk_queue_bypass_start() is only called for blk-sq queues, 677 * this function is called for both blk-sq and blk-mq queues. 678 */ 679 void blk_queue_bypass_end(struct request_queue *q) 680 { 681 spin_lock_irq(q->queue_lock); 682 if (!--q->bypass_depth) 683 queue_flag_clear(QUEUE_FLAG_BYPASS, q); 684 WARN_ON_ONCE(q->bypass_depth < 0); 685 spin_unlock_irq(q->queue_lock); 686 } 687 EXPORT_SYMBOL_GPL(blk_queue_bypass_end); 688 689 void blk_set_queue_dying(struct request_queue *q) 690 { 691 blk_queue_flag_set(QUEUE_FLAG_DYING, q); 692 693 /* 694 * When queue DYING flag is set, we need to block new req 695 * entering queue, so we call blk_freeze_queue_start() to 696 * prevent I/O from crossing blk_queue_enter(). 697 */ 698 blk_freeze_queue_start(q); 699 700 if (q->mq_ops) 701 blk_mq_wake_waiters(q); 702 else { 703 struct request_list *rl; 704 705 spin_lock_irq(q->queue_lock); 706 blk_queue_for_each_rl(rl, q) { 707 if (rl->rq_pool) { 708 wake_up_all(&rl->wait[BLK_RW_SYNC]); 709 wake_up_all(&rl->wait[BLK_RW_ASYNC]); 710 } 711 } 712 spin_unlock_irq(q->queue_lock); 713 } 714 715 /* Make blk_queue_enter() reexamine the DYING flag. */ 716 wake_up_all(&q->mq_freeze_wq); 717 } 718 EXPORT_SYMBOL_GPL(blk_set_queue_dying); 719 720 /* Unconfigure the I/O scheduler and dissociate from the cgroup controller. */ 721 void blk_exit_queue(struct request_queue *q) 722 { 723 /* 724 * Since the I/O scheduler exit code may access cgroup information, 725 * perform I/O scheduler exit before disassociating from the block 726 * cgroup controller. 727 */ 728 if (q->elevator) { 729 ioc_clear_queue(q); 730 elevator_exit(q, q->elevator); 731 q->elevator = NULL; 732 } 733 734 /* 735 * Remove all references to @q from the block cgroup controller before 736 * restoring @q->queue_lock to avoid that restoring this pointer causes 737 * e.g. blkcg_print_blkgs() to crash. 738 */ 739 blkcg_exit_queue(q); 740 741 /* 742 * Since the cgroup code may dereference the @q->backing_dev_info 743 * pointer, only decrease its reference count after having removed the 744 * association with the block cgroup controller. 745 */ 746 bdi_put(q->backing_dev_info); 747 } 748 749 /** 750 * blk_cleanup_queue - shutdown a request queue 751 * @q: request queue to shutdown 752 * 753 * Mark @q DYING, drain all pending requests, mark @q DEAD, destroy and 754 * put it. All future requests will be failed immediately with -ENODEV. 755 */ 756 void blk_cleanup_queue(struct request_queue *q) 757 { 758 spinlock_t *lock = q->queue_lock; 759 760 /* mark @q DYING, no new request or merges will be allowed afterwards */ 761 mutex_lock(&q->sysfs_lock); 762 blk_set_queue_dying(q); 763 spin_lock_irq(lock); 764 765 /* 766 * A dying queue is permanently in bypass mode till released. Note 767 * that, unlike blk_queue_bypass_start(), we aren't performing 768 * synchronize_rcu() after entering bypass mode to avoid the delay 769 * as some drivers create and destroy a lot of queues while 770 * probing. This is still safe because blk_release_queue() will be 771 * called only after the queue refcnt drops to zero and nothing, 772 * RCU or not, would be traversing the queue by then. 773 */ 774 q->bypass_depth++; 775 queue_flag_set(QUEUE_FLAG_BYPASS, q); 776 777 queue_flag_set(QUEUE_FLAG_NOMERGES, q); 778 queue_flag_set(QUEUE_FLAG_NOXMERGES, q); 779 queue_flag_set(QUEUE_FLAG_DYING, q); 780 spin_unlock_irq(lock); 781 mutex_unlock(&q->sysfs_lock); 782 783 /* 784 * Drain all requests queued before DYING marking. Set DEAD flag to 785 * prevent that q->request_fn() gets invoked after draining finished. 786 */ 787 blk_freeze_queue(q); 788 789 rq_qos_exit(q); 790 791 spin_lock_irq(lock); 792 queue_flag_set(QUEUE_FLAG_DEAD, q); 793 spin_unlock_irq(lock); 794 795 /* 796 * make sure all in-progress dispatch are completed because 797 * blk_freeze_queue() can only complete all requests, and 798 * dispatch may still be in-progress since we dispatch requests 799 * from more than one contexts. 800 * 801 * No need to quiesce queue if it isn't initialized yet since 802 * blk_freeze_queue() should be enough for cases of passthrough 803 * request. 804 */ 805 if (q->mq_ops && blk_queue_init_done(q)) 806 blk_mq_quiesce_queue(q); 807 808 /* for synchronous bio-based driver finish in-flight integrity i/o */ 809 blk_flush_integrity(); 810 811 /* @q won't process any more request, flush async actions */ 812 del_timer_sync(&q->backing_dev_info->laptop_mode_wb_timer); 813 blk_sync_queue(q); 814 815 /* 816 * I/O scheduler exit is only safe after the sysfs scheduler attribute 817 * has been removed. 818 */ 819 WARN_ON_ONCE(q->kobj.state_in_sysfs); 820 821 blk_exit_queue(q); 822 823 if (q->mq_ops) 824 blk_mq_free_queue(q); 825 percpu_ref_exit(&q->q_usage_counter); 826 827 spin_lock_irq(lock); 828 if (q->queue_lock != &q->__queue_lock) 829 q->queue_lock = &q->__queue_lock; 830 spin_unlock_irq(lock); 831 832 /* @q is and will stay empty, shutdown and put */ 833 blk_put_queue(q); 834 } 835 EXPORT_SYMBOL(blk_cleanup_queue); 836 837 /* Allocate memory local to the request queue */ 838 static void *alloc_request_simple(gfp_t gfp_mask, void *data) 839 { 840 struct request_queue *q = data; 841 842 return kmem_cache_alloc_node(request_cachep, gfp_mask, q->node); 843 } 844 845 static void free_request_simple(void *element, void *data) 846 { 847 kmem_cache_free(request_cachep, element); 848 } 849 850 static void *alloc_request_size(gfp_t gfp_mask, void *data) 851 { 852 struct request_queue *q = data; 853 struct request *rq; 854 855 rq = kmalloc_node(sizeof(struct request) + q->cmd_size, gfp_mask, 856 q->node); 857 if (rq && q->init_rq_fn && q->init_rq_fn(q, rq, gfp_mask) < 0) { 858 kfree(rq); 859 rq = NULL; 860 } 861 return rq; 862 } 863 864 static void free_request_size(void *element, void *data) 865 { 866 struct request_queue *q = data; 867 868 if (q->exit_rq_fn) 869 q->exit_rq_fn(q, element); 870 kfree(element); 871 } 872 873 int blk_init_rl(struct request_list *rl, struct request_queue *q, 874 gfp_t gfp_mask) 875 { 876 if (unlikely(rl->rq_pool) || q->mq_ops) 877 return 0; 878 879 rl->q = q; 880 rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0; 881 rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0; 882 init_waitqueue_head(&rl->wait[BLK_RW_SYNC]); 883 init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]); 884 885 if (q->cmd_size) { 886 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, 887 alloc_request_size, free_request_size, 888 q, gfp_mask, q->node); 889 } else { 890 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, 891 alloc_request_simple, free_request_simple, 892 q, gfp_mask, q->node); 893 } 894 if (!rl->rq_pool) 895 return -ENOMEM; 896 897 if (rl != &q->root_rl) 898 WARN_ON_ONCE(!blk_get_queue(q)); 899 900 return 0; 901 } 902 903 void blk_exit_rl(struct request_queue *q, struct request_list *rl) 904 { 905 if (rl->rq_pool) { 906 mempool_destroy(rl->rq_pool); 907 if (rl != &q->root_rl) 908 blk_put_queue(q); 909 } 910 } 911 912 struct request_queue *blk_alloc_queue(gfp_t gfp_mask) 913 { 914 return blk_alloc_queue_node(gfp_mask, NUMA_NO_NODE, NULL); 915 } 916 EXPORT_SYMBOL(blk_alloc_queue); 917 918 /** 919 * blk_queue_enter() - try to increase q->q_usage_counter 920 * @q: request queue pointer 921 * @flags: BLK_MQ_REQ_NOWAIT and/or BLK_MQ_REQ_PREEMPT 922 */ 923 int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags) 924 { 925 const bool pm = flags & BLK_MQ_REQ_PREEMPT; 926 927 while (true) { 928 bool success = false; 929 930 rcu_read_lock(); 931 if (percpu_ref_tryget_live(&q->q_usage_counter)) { 932 /* 933 * The code that increments the pm_only counter is 934 * responsible for ensuring that that counter is 935 * globally visible before the queue is unfrozen. 936 */ 937 if (pm || !blk_queue_pm_only(q)) { 938 success = true; 939 } else { 940 percpu_ref_put(&q->q_usage_counter); 941 } 942 } 943 rcu_read_unlock(); 944 945 if (success) 946 return 0; 947 948 if (flags & BLK_MQ_REQ_NOWAIT) 949 return -EBUSY; 950 951 /* 952 * read pair of barrier in blk_freeze_queue_start(), 953 * we need to order reading __PERCPU_REF_DEAD flag of 954 * .q_usage_counter and reading .mq_freeze_depth or 955 * queue dying flag, otherwise the following wait may 956 * never return if the two reads are reordered. 957 */ 958 smp_rmb(); 959 960 wait_event(q->mq_freeze_wq, 961 (atomic_read(&q->mq_freeze_depth) == 0 && 962 (pm || (blk_pm_request_resume(q), 963 !blk_queue_pm_only(q)))) || 964 blk_queue_dying(q)); 965 if (blk_queue_dying(q)) 966 return -ENODEV; 967 } 968 } 969 970 void blk_queue_exit(struct request_queue *q) 971 { 972 percpu_ref_put(&q->q_usage_counter); 973 } 974 975 static void blk_queue_usage_counter_release(struct percpu_ref *ref) 976 { 977 struct request_queue *q = 978 container_of(ref, struct request_queue, q_usage_counter); 979 980 wake_up_all(&q->mq_freeze_wq); 981 } 982 983 static void blk_rq_timed_out_timer(struct timer_list *t) 984 { 985 struct request_queue *q = from_timer(q, t, timeout); 986 987 kblockd_schedule_work(&q->timeout_work); 988 } 989 990 /** 991 * blk_alloc_queue_node - allocate a request queue 992 * @gfp_mask: memory allocation flags 993 * @node_id: NUMA node to allocate memory from 994 * @lock: For legacy queues, pointer to a spinlock that will be used to e.g. 995 * serialize calls to the legacy .request_fn() callback. Ignored for 996 * blk-mq request queues. 997 * 998 * Note: pass the queue lock as the third argument to this function instead of 999 * setting the queue lock pointer explicitly to avoid triggering a sporadic 1000 * crash in the blkcg code. This function namely calls blkcg_init_queue() and 1001 * the queue lock pointer must be set before blkcg_init_queue() is called. 1002 */ 1003 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id, 1004 spinlock_t *lock) 1005 { 1006 struct request_queue *q; 1007 int ret; 1008 1009 q = kmem_cache_alloc_node(blk_requestq_cachep, 1010 gfp_mask | __GFP_ZERO, node_id); 1011 if (!q) 1012 return NULL; 1013 1014 INIT_LIST_HEAD(&q->queue_head); 1015 q->last_merge = NULL; 1016 q->end_sector = 0; 1017 q->boundary_rq = NULL; 1018 1019 q->id = ida_simple_get(&blk_queue_ida, 0, 0, gfp_mask); 1020 if (q->id < 0) 1021 goto fail_q; 1022 1023 ret = bioset_init(&q->bio_split, BIO_POOL_SIZE, 0, BIOSET_NEED_BVECS); 1024 if (ret) 1025 goto fail_id; 1026 1027 q->backing_dev_info = bdi_alloc_node(gfp_mask, node_id); 1028 if (!q->backing_dev_info) 1029 goto fail_split; 1030 1031 q->stats = blk_alloc_queue_stats(); 1032 if (!q->stats) 1033 goto fail_stats; 1034 1035 q->backing_dev_info->ra_pages = 1036 (VM_MAX_READAHEAD * 1024) / PAGE_SIZE; 1037 q->backing_dev_info->capabilities = BDI_CAP_CGROUP_WRITEBACK; 1038 q->backing_dev_info->name = "block"; 1039 q->node = node_id; 1040 1041 timer_setup(&q->backing_dev_info->laptop_mode_wb_timer, 1042 laptop_mode_timer_fn, 0); 1043 timer_setup(&q->timeout, blk_rq_timed_out_timer, 0); 1044 INIT_WORK(&q->timeout_work, NULL); 1045 INIT_LIST_HEAD(&q->timeout_list); 1046 INIT_LIST_HEAD(&q->icq_list); 1047 #ifdef CONFIG_BLK_CGROUP 1048 INIT_LIST_HEAD(&q->blkg_list); 1049 #endif 1050 INIT_DELAYED_WORK(&q->delay_work, blk_delay_work); 1051 1052 kobject_init(&q->kobj, &blk_queue_ktype); 1053 1054 #ifdef CONFIG_BLK_DEV_IO_TRACE 1055 mutex_init(&q->blk_trace_mutex); 1056 #endif 1057 mutex_init(&q->sysfs_lock); 1058 spin_lock_init(&q->__queue_lock); 1059 1060 q->queue_lock = lock ? : &q->__queue_lock; 1061 1062 /* 1063 * A queue starts its life with bypass turned on to avoid 1064 * unnecessary bypass on/off overhead and nasty surprises during 1065 * init. The initial bypass will be finished when the queue is 1066 * registered by blk_register_queue(). 1067 */ 1068 q->bypass_depth = 1; 1069 queue_flag_set_unlocked(QUEUE_FLAG_BYPASS, q); 1070 1071 init_waitqueue_head(&q->mq_freeze_wq); 1072 1073 /* 1074 * Init percpu_ref in atomic mode so that it's faster to shutdown. 1075 * See blk_register_queue() for details. 1076 */ 1077 if (percpu_ref_init(&q->q_usage_counter, 1078 blk_queue_usage_counter_release, 1079 PERCPU_REF_INIT_ATOMIC, GFP_KERNEL)) 1080 goto fail_bdi; 1081 1082 if (blkcg_init_queue(q)) 1083 goto fail_ref; 1084 1085 return q; 1086 1087 fail_ref: 1088 percpu_ref_exit(&q->q_usage_counter); 1089 fail_bdi: 1090 blk_free_queue_stats(q->stats); 1091 fail_stats: 1092 bdi_put(q->backing_dev_info); 1093 fail_split: 1094 bioset_exit(&q->bio_split); 1095 fail_id: 1096 ida_simple_remove(&blk_queue_ida, q->id); 1097 fail_q: 1098 kmem_cache_free(blk_requestq_cachep, q); 1099 return NULL; 1100 } 1101 EXPORT_SYMBOL(blk_alloc_queue_node); 1102 1103 /** 1104 * blk_init_queue - prepare a request queue for use with a block device 1105 * @rfn: The function to be called to process requests that have been 1106 * placed on the queue. 1107 * @lock: Request queue spin lock 1108 * 1109 * Description: 1110 * If a block device wishes to use the standard request handling procedures, 1111 * which sorts requests and coalesces adjacent requests, then it must 1112 * call blk_init_queue(). The function @rfn will be called when there 1113 * are requests on the queue that need to be processed. If the device 1114 * supports plugging, then @rfn may not be called immediately when requests 1115 * are available on the queue, but may be called at some time later instead. 1116 * Plugged queues are generally unplugged when a buffer belonging to one 1117 * of the requests on the queue is needed, or due to memory pressure. 1118 * 1119 * @rfn is not required, or even expected, to remove all requests off the 1120 * queue, but only as many as it can handle at a time. If it does leave 1121 * requests on the queue, it is responsible for arranging that the requests 1122 * get dealt with eventually. 1123 * 1124 * The queue spin lock must be held while manipulating the requests on the 1125 * request queue; this lock will be taken also from interrupt context, so irq 1126 * disabling is needed for it. 1127 * 1128 * Function returns a pointer to the initialized request queue, or %NULL if 1129 * it didn't succeed. 1130 * 1131 * Note: 1132 * blk_init_queue() must be paired with a blk_cleanup_queue() call 1133 * when the block device is deactivated (such as at module unload). 1134 **/ 1135 1136 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock) 1137 { 1138 return blk_init_queue_node(rfn, lock, NUMA_NO_NODE); 1139 } 1140 EXPORT_SYMBOL(blk_init_queue); 1141 1142 struct request_queue * 1143 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id) 1144 { 1145 struct request_queue *q; 1146 1147 q = blk_alloc_queue_node(GFP_KERNEL, node_id, lock); 1148 if (!q) 1149 return NULL; 1150 1151 q->request_fn = rfn; 1152 if (blk_init_allocated_queue(q) < 0) { 1153 blk_cleanup_queue(q); 1154 return NULL; 1155 } 1156 1157 return q; 1158 } 1159 EXPORT_SYMBOL(blk_init_queue_node); 1160 1161 static blk_qc_t blk_queue_bio(struct request_queue *q, struct bio *bio); 1162 1163 1164 int blk_init_allocated_queue(struct request_queue *q) 1165 { 1166 WARN_ON_ONCE(q->mq_ops); 1167 1168 q->fq = blk_alloc_flush_queue(q, NUMA_NO_NODE, q->cmd_size, GFP_KERNEL); 1169 if (!q->fq) 1170 return -ENOMEM; 1171 1172 if (q->init_rq_fn && q->init_rq_fn(q, q->fq->flush_rq, GFP_KERNEL)) 1173 goto out_free_flush_queue; 1174 1175 if (blk_init_rl(&q->root_rl, q, GFP_KERNEL)) 1176 goto out_exit_flush_rq; 1177 1178 INIT_WORK(&q->timeout_work, blk_timeout_work); 1179 q->queue_flags |= QUEUE_FLAG_DEFAULT; 1180 1181 /* 1182 * This also sets hw/phys segments, boundary and size 1183 */ 1184 blk_queue_make_request(q, blk_queue_bio); 1185 1186 q->sg_reserved_size = INT_MAX; 1187 1188 if (elevator_init(q)) 1189 goto out_exit_flush_rq; 1190 return 0; 1191 1192 out_exit_flush_rq: 1193 if (q->exit_rq_fn) 1194 q->exit_rq_fn(q, q->fq->flush_rq); 1195 out_free_flush_queue: 1196 blk_free_flush_queue(q->fq); 1197 q->fq = NULL; 1198 return -ENOMEM; 1199 } 1200 EXPORT_SYMBOL(blk_init_allocated_queue); 1201 1202 bool blk_get_queue(struct request_queue *q) 1203 { 1204 if (likely(!blk_queue_dying(q))) { 1205 __blk_get_queue(q); 1206 return true; 1207 } 1208 1209 return false; 1210 } 1211 EXPORT_SYMBOL(blk_get_queue); 1212 1213 static inline void blk_free_request(struct request_list *rl, struct request *rq) 1214 { 1215 if (rq->rq_flags & RQF_ELVPRIV) { 1216 elv_put_request(rl->q, rq); 1217 if (rq->elv.icq) 1218 put_io_context(rq->elv.icq->ioc); 1219 } 1220 1221 mempool_free(rq, rl->rq_pool); 1222 } 1223 1224 /* 1225 * ioc_batching returns true if the ioc is a valid batching request and 1226 * should be given priority access to a request. 1227 */ 1228 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc) 1229 { 1230 if (!ioc) 1231 return 0; 1232 1233 /* 1234 * Make sure the process is able to allocate at least 1 request 1235 * even if the batch times out, otherwise we could theoretically 1236 * lose wakeups. 1237 */ 1238 return ioc->nr_batch_requests == q->nr_batching || 1239 (ioc->nr_batch_requests > 0 1240 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME)); 1241 } 1242 1243 /* 1244 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This 1245 * will cause the process to be a "batcher" on all queues in the system. This 1246 * is the behaviour we want though - once it gets a wakeup it should be given 1247 * a nice run. 1248 */ 1249 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc) 1250 { 1251 if (!ioc || ioc_batching(q, ioc)) 1252 return; 1253 1254 ioc->nr_batch_requests = q->nr_batching; 1255 ioc->last_waited = jiffies; 1256 } 1257 1258 static void __freed_request(struct request_list *rl, int sync) 1259 { 1260 struct request_queue *q = rl->q; 1261 1262 if (rl->count[sync] < queue_congestion_off_threshold(q)) 1263 blk_clear_congested(rl, sync); 1264 1265 if (rl->count[sync] + 1 <= q->nr_requests) { 1266 if (waitqueue_active(&rl->wait[sync])) 1267 wake_up(&rl->wait[sync]); 1268 1269 blk_clear_rl_full(rl, sync); 1270 } 1271 } 1272 1273 /* 1274 * A request has just been released. Account for it, update the full and 1275 * congestion status, wake up any waiters. Called under q->queue_lock. 1276 */ 1277 static void freed_request(struct request_list *rl, bool sync, 1278 req_flags_t rq_flags) 1279 { 1280 struct request_queue *q = rl->q; 1281 1282 q->nr_rqs[sync]--; 1283 rl->count[sync]--; 1284 if (rq_flags & RQF_ELVPRIV) 1285 q->nr_rqs_elvpriv--; 1286 1287 __freed_request(rl, sync); 1288 1289 if (unlikely(rl->starved[sync ^ 1])) 1290 __freed_request(rl, sync ^ 1); 1291 } 1292 1293 int blk_update_nr_requests(struct request_queue *q, unsigned int nr) 1294 { 1295 struct request_list *rl; 1296 int on_thresh, off_thresh; 1297 1298 WARN_ON_ONCE(q->mq_ops); 1299 1300 spin_lock_irq(q->queue_lock); 1301 q->nr_requests = nr; 1302 blk_queue_congestion_threshold(q); 1303 on_thresh = queue_congestion_on_threshold(q); 1304 off_thresh = queue_congestion_off_threshold(q); 1305 1306 blk_queue_for_each_rl(rl, q) { 1307 if (rl->count[BLK_RW_SYNC] >= on_thresh) 1308 blk_set_congested(rl, BLK_RW_SYNC); 1309 else if (rl->count[BLK_RW_SYNC] < off_thresh) 1310 blk_clear_congested(rl, BLK_RW_SYNC); 1311 1312 if (rl->count[BLK_RW_ASYNC] >= on_thresh) 1313 blk_set_congested(rl, BLK_RW_ASYNC); 1314 else if (rl->count[BLK_RW_ASYNC] < off_thresh) 1315 blk_clear_congested(rl, BLK_RW_ASYNC); 1316 1317 if (rl->count[BLK_RW_SYNC] >= q->nr_requests) { 1318 blk_set_rl_full(rl, BLK_RW_SYNC); 1319 } else { 1320 blk_clear_rl_full(rl, BLK_RW_SYNC); 1321 wake_up(&rl->wait[BLK_RW_SYNC]); 1322 } 1323 1324 if (rl->count[BLK_RW_ASYNC] >= q->nr_requests) { 1325 blk_set_rl_full(rl, BLK_RW_ASYNC); 1326 } else { 1327 blk_clear_rl_full(rl, BLK_RW_ASYNC); 1328 wake_up(&rl->wait[BLK_RW_ASYNC]); 1329 } 1330 } 1331 1332 spin_unlock_irq(q->queue_lock); 1333 return 0; 1334 } 1335 1336 /** 1337 * __get_request - get a free request 1338 * @rl: request list to allocate from 1339 * @op: operation and flags 1340 * @bio: bio to allocate request for (can be %NULL) 1341 * @flags: BLQ_MQ_REQ_* flags 1342 * @gfp_mask: allocator flags 1343 * 1344 * Get a free request from @q. This function may fail under memory 1345 * pressure or if @q is dead. 1346 * 1347 * Must be called with @q->queue_lock held and, 1348 * Returns ERR_PTR on failure, with @q->queue_lock held. 1349 * Returns request pointer on success, with @q->queue_lock *not held*. 1350 */ 1351 static struct request *__get_request(struct request_list *rl, unsigned int op, 1352 struct bio *bio, blk_mq_req_flags_t flags, gfp_t gfp_mask) 1353 { 1354 struct request_queue *q = rl->q; 1355 struct request *rq; 1356 struct elevator_type *et = q->elevator->type; 1357 struct io_context *ioc = rq_ioc(bio); 1358 struct io_cq *icq = NULL; 1359 const bool is_sync = op_is_sync(op); 1360 int may_queue; 1361 req_flags_t rq_flags = RQF_ALLOCED; 1362 1363 lockdep_assert_held(q->queue_lock); 1364 1365 if (unlikely(blk_queue_dying(q))) 1366 return ERR_PTR(-ENODEV); 1367 1368 may_queue = elv_may_queue(q, op); 1369 if (may_queue == ELV_MQUEUE_NO) 1370 goto rq_starved; 1371 1372 if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) { 1373 if (rl->count[is_sync]+1 >= q->nr_requests) { 1374 /* 1375 * The queue will fill after this allocation, so set 1376 * it as full, and mark this process as "batching". 1377 * This process will be allowed to complete a batch of 1378 * requests, others will be blocked. 1379 */ 1380 if (!blk_rl_full(rl, is_sync)) { 1381 ioc_set_batching(q, ioc); 1382 blk_set_rl_full(rl, is_sync); 1383 } else { 1384 if (may_queue != ELV_MQUEUE_MUST 1385 && !ioc_batching(q, ioc)) { 1386 /* 1387 * The queue is full and the allocating 1388 * process is not a "batcher", and not 1389 * exempted by the IO scheduler 1390 */ 1391 return ERR_PTR(-ENOMEM); 1392 } 1393 } 1394 } 1395 blk_set_congested(rl, is_sync); 1396 } 1397 1398 /* 1399 * Only allow batching queuers to allocate up to 50% over the defined 1400 * limit of requests, otherwise we could have thousands of requests 1401 * allocated with any setting of ->nr_requests 1402 */ 1403 if (rl->count[is_sync] >= (3 * q->nr_requests / 2)) 1404 return ERR_PTR(-ENOMEM); 1405 1406 q->nr_rqs[is_sync]++; 1407 rl->count[is_sync]++; 1408 rl->starved[is_sync] = 0; 1409 1410 /* 1411 * Decide whether the new request will be managed by elevator. If 1412 * so, mark @rq_flags and increment elvpriv. Non-zero elvpriv will 1413 * prevent the current elevator from being destroyed until the new 1414 * request is freed. This guarantees icq's won't be destroyed and 1415 * makes creating new ones safe. 1416 * 1417 * Flush requests do not use the elevator so skip initialization. 1418 * This allows a request to share the flush and elevator data. 1419 * 1420 * Also, lookup icq while holding queue_lock. If it doesn't exist, 1421 * it will be created after releasing queue_lock. 1422 */ 1423 if (!op_is_flush(op) && !blk_queue_bypass(q)) { 1424 rq_flags |= RQF_ELVPRIV; 1425 q->nr_rqs_elvpriv++; 1426 if (et->icq_cache && ioc) 1427 icq = ioc_lookup_icq(ioc, q); 1428 } 1429 1430 if (blk_queue_io_stat(q)) 1431 rq_flags |= RQF_IO_STAT; 1432 spin_unlock_irq(q->queue_lock); 1433 1434 /* allocate and init request */ 1435 rq = mempool_alloc(rl->rq_pool, gfp_mask); 1436 if (!rq) 1437 goto fail_alloc; 1438 1439 blk_rq_init(q, rq); 1440 blk_rq_set_rl(rq, rl); 1441 rq->cmd_flags = op; 1442 rq->rq_flags = rq_flags; 1443 if (flags & BLK_MQ_REQ_PREEMPT) 1444 rq->rq_flags |= RQF_PREEMPT; 1445 1446 /* init elvpriv */ 1447 if (rq_flags & RQF_ELVPRIV) { 1448 if (unlikely(et->icq_cache && !icq)) { 1449 if (ioc) 1450 icq = ioc_create_icq(ioc, q, gfp_mask); 1451 if (!icq) 1452 goto fail_elvpriv; 1453 } 1454 1455 rq->elv.icq = icq; 1456 if (unlikely(elv_set_request(q, rq, bio, gfp_mask))) 1457 goto fail_elvpriv; 1458 1459 /* @rq->elv.icq holds io_context until @rq is freed */ 1460 if (icq) 1461 get_io_context(icq->ioc); 1462 } 1463 out: 1464 /* 1465 * ioc may be NULL here, and ioc_batching will be false. That's 1466 * OK, if the queue is under the request limit then requests need 1467 * not count toward the nr_batch_requests limit. There will always 1468 * be some limit enforced by BLK_BATCH_TIME. 1469 */ 1470 if (ioc_batching(q, ioc)) 1471 ioc->nr_batch_requests--; 1472 1473 trace_block_getrq(q, bio, op); 1474 return rq; 1475 1476 fail_elvpriv: 1477 /* 1478 * elvpriv init failed. ioc, icq and elvpriv aren't mempool backed 1479 * and may fail indefinitely under memory pressure and thus 1480 * shouldn't stall IO. Treat this request as !elvpriv. This will 1481 * disturb iosched and blkcg but weird is bettern than dead. 1482 */ 1483 printk_ratelimited(KERN_WARNING "%s: dev %s: request aux data allocation failed, iosched may be disturbed\n", 1484 __func__, dev_name(q->backing_dev_info->dev)); 1485 1486 rq->rq_flags &= ~RQF_ELVPRIV; 1487 rq->elv.icq = NULL; 1488 1489 spin_lock_irq(q->queue_lock); 1490 q->nr_rqs_elvpriv--; 1491 spin_unlock_irq(q->queue_lock); 1492 goto out; 1493 1494 fail_alloc: 1495 /* 1496 * Allocation failed presumably due to memory. Undo anything we 1497 * might have messed up. 1498 * 1499 * Allocating task should really be put onto the front of the wait 1500 * queue, but this is pretty rare. 1501 */ 1502 spin_lock_irq(q->queue_lock); 1503 freed_request(rl, is_sync, rq_flags); 1504 1505 /* 1506 * in the very unlikely event that allocation failed and no 1507 * requests for this direction was pending, mark us starved so that 1508 * freeing of a request in the other direction will notice 1509 * us. another possible fix would be to split the rq mempool into 1510 * READ and WRITE 1511 */ 1512 rq_starved: 1513 if (unlikely(rl->count[is_sync] == 0)) 1514 rl->starved[is_sync] = 1; 1515 return ERR_PTR(-ENOMEM); 1516 } 1517 1518 /** 1519 * get_request - get a free request 1520 * @q: request_queue to allocate request from 1521 * @op: operation and flags 1522 * @bio: bio to allocate request for (can be %NULL) 1523 * @flags: BLK_MQ_REQ_* flags. 1524 * @gfp: allocator flags 1525 * 1526 * Get a free request from @q. If %BLK_MQ_REQ_NOWAIT is set in @flags, 1527 * this function keeps retrying under memory pressure and fails iff @q is dead. 1528 * 1529 * Must be called with @q->queue_lock held and, 1530 * Returns ERR_PTR on failure, with @q->queue_lock held. 1531 * Returns request pointer on success, with @q->queue_lock *not held*. 1532 */ 1533 static struct request *get_request(struct request_queue *q, unsigned int op, 1534 struct bio *bio, blk_mq_req_flags_t flags, gfp_t gfp) 1535 { 1536 const bool is_sync = op_is_sync(op); 1537 DEFINE_WAIT(wait); 1538 struct request_list *rl; 1539 struct request *rq; 1540 1541 lockdep_assert_held(q->queue_lock); 1542 WARN_ON_ONCE(q->mq_ops); 1543 1544 rl = blk_get_rl(q, bio); /* transferred to @rq on success */ 1545 retry: 1546 rq = __get_request(rl, op, bio, flags, gfp); 1547 if (!IS_ERR(rq)) 1548 return rq; 1549 1550 if (op & REQ_NOWAIT) { 1551 blk_put_rl(rl); 1552 return ERR_PTR(-EAGAIN); 1553 } 1554 1555 if ((flags & BLK_MQ_REQ_NOWAIT) || unlikely(blk_queue_dying(q))) { 1556 blk_put_rl(rl); 1557 return rq; 1558 } 1559 1560 /* wait on @rl and retry */ 1561 prepare_to_wait_exclusive(&rl->wait[is_sync], &wait, 1562 TASK_UNINTERRUPTIBLE); 1563 1564 trace_block_sleeprq(q, bio, op); 1565 1566 spin_unlock_irq(q->queue_lock); 1567 io_schedule(); 1568 1569 /* 1570 * After sleeping, we become a "batching" process and will be able 1571 * to allocate at least one request, and up to a big batch of them 1572 * for a small period time. See ioc_batching, ioc_set_batching 1573 */ 1574 ioc_set_batching(q, current->io_context); 1575 1576 spin_lock_irq(q->queue_lock); 1577 finish_wait(&rl->wait[is_sync], &wait); 1578 1579 goto retry; 1580 } 1581 1582 /* flags: BLK_MQ_REQ_PREEMPT and/or BLK_MQ_REQ_NOWAIT. */ 1583 static struct request *blk_old_get_request(struct request_queue *q, 1584 unsigned int op, blk_mq_req_flags_t flags) 1585 { 1586 struct request *rq; 1587 gfp_t gfp_mask = flags & BLK_MQ_REQ_NOWAIT ? GFP_ATOMIC : GFP_NOIO; 1588 int ret = 0; 1589 1590 WARN_ON_ONCE(q->mq_ops); 1591 1592 /* create ioc upfront */ 1593 create_io_context(gfp_mask, q->node); 1594 1595 ret = blk_queue_enter(q, flags); 1596 if (ret) 1597 return ERR_PTR(ret); 1598 spin_lock_irq(q->queue_lock); 1599 rq = get_request(q, op, NULL, flags, gfp_mask); 1600 if (IS_ERR(rq)) { 1601 spin_unlock_irq(q->queue_lock); 1602 blk_queue_exit(q); 1603 return rq; 1604 } 1605 1606 /* q->queue_lock is unlocked at this point */ 1607 rq->__data_len = 0; 1608 rq->__sector = (sector_t) -1; 1609 rq->bio = rq->biotail = NULL; 1610 return rq; 1611 } 1612 1613 /** 1614 * blk_get_request - allocate a request 1615 * @q: request queue to allocate a request for 1616 * @op: operation (REQ_OP_*) and REQ_* flags, e.g. REQ_SYNC. 1617 * @flags: BLK_MQ_REQ_* flags, e.g. BLK_MQ_REQ_NOWAIT. 1618 */ 1619 struct request *blk_get_request(struct request_queue *q, unsigned int op, 1620 blk_mq_req_flags_t flags) 1621 { 1622 struct request *req; 1623 1624 WARN_ON_ONCE(op & REQ_NOWAIT); 1625 WARN_ON_ONCE(flags & ~(BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_PREEMPT)); 1626 1627 if (q->mq_ops) { 1628 req = blk_mq_alloc_request(q, op, flags); 1629 if (!IS_ERR(req) && q->mq_ops->initialize_rq_fn) 1630 q->mq_ops->initialize_rq_fn(req); 1631 } else { 1632 req = blk_old_get_request(q, op, flags); 1633 if (!IS_ERR(req) && q->initialize_rq_fn) 1634 q->initialize_rq_fn(req); 1635 } 1636 1637 return req; 1638 } 1639 EXPORT_SYMBOL(blk_get_request); 1640 1641 /** 1642 * blk_requeue_request - put a request back on queue 1643 * @q: request queue where request should be inserted 1644 * @rq: request to be inserted 1645 * 1646 * Description: 1647 * Drivers often keep queueing requests until the hardware cannot accept 1648 * more, when that condition happens we need to put the request back 1649 * on the queue. Must be called with queue lock held. 1650 */ 1651 void blk_requeue_request(struct request_queue *q, struct request *rq) 1652 { 1653 lockdep_assert_held(q->queue_lock); 1654 WARN_ON_ONCE(q->mq_ops); 1655 1656 blk_delete_timer(rq); 1657 blk_clear_rq_complete(rq); 1658 trace_block_rq_requeue(q, rq); 1659 rq_qos_requeue(q, rq); 1660 1661 if (rq->rq_flags & RQF_QUEUED) 1662 blk_queue_end_tag(q, rq); 1663 1664 BUG_ON(blk_queued_rq(rq)); 1665 1666 elv_requeue_request(q, rq); 1667 } 1668 EXPORT_SYMBOL(blk_requeue_request); 1669 1670 static void add_acct_request(struct request_queue *q, struct request *rq, 1671 int where) 1672 { 1673 blk_account_io_start(rq, true); 1674 __elv_add_request(q, rq, where); 1675 } 1676 1677 static void part_round_stats_single(struct request_queue *q, int cpu, 1678 struct hd_struct *part, unsigned long now, 1679 unsigned int inflight) 1680 { 1681 if (inflight) { 1682 __part_stat_add(cpu, part, time_in_queue, 1683 inflight * (now - part->stamp)); 1684 __part_stat_add(cpu, part, io_ticks, (now - part->stamp)); 1685 } 1686 part->stamp = now; 1687 } 1688 1689 /** 1690 * part_round_stats() - Round off the performance stats on a struct disk_stats. 1691 * @q: target block queue 1692 * @cpu: cpu number for stats access 1693 * @part: target partition 1694 * 1695 * The average IO queue length and utilisation statistics are maintained 1696 * by observing the current state of the queue length and the amount of 1697 * time it has been in this state for. 1698 * 1699 * Normally, that accounting is done on IO completion, but that can result 1700 * in more than a second's worth of IO being accounted for within any one 1701 * second, leading to >100% utilisation. To deal with that, we call this 1702 * function to do a round-off before returning the results when reading 1703 * /proc/diskstats. This accounts immediately for all queue usage up to 1704 * the current jiffies and restarts the counters again. 1705 */ 1706 void part_round_stats(struct request_queue *q, int cpu, struct hd_struct *part) 1707 { 1708 struct hd_struct *part2 = NULL; 1709 unsigned long now = jiffies; 1710 unsigned int inflight[2]; 1711 int stats = 0; 1712 1713 if (part->stamp != now) 1714 stats |= 1; 1715 1716 if (part->partno) { 1717 part2 = &part_to_disk(part)->part0; 1718 if (part2->stamp != now) 1719 stats |= 2; 1720 } 1721 1722 if (!stats) 1723 return; 1724 1725 part_in_flight(q, part, inflight); 1726 1727 if (stats & 2) 1728 part_round_stats_single(q, cpu, part2, now, inflight[1]); 1729 if (stats & 1) 1730 part_round_stats_single(q, cpu, part, now, inflight[0]); 1731 } 1732 EXPORT_SYMBOL_GPL(part_round_stats); 1733 1734 void __blk_put_request(struct request_queue *q, struct request *req) 1735 { 1736 req_flags_t rq_flags = req->rq_flags; 1737 1738 if (unlikely(!q)) 1739 return; 1740 1741 if (q->mq_ops) { 1742 blk_mq_free_request(req); 1743 return; 1744 } 1745 1746 lockdep_assert_held(q->queue_lock); 1747 1748 blk_req_zone_write_unlock(req); 1749 blk_pm_put_request(req); 1750 blk_pm_mark_last_busy(req); 1751 1752 elv_completed_request(q, req); 1753 1754 /* this is a bio leak */ 1755 WARN_ON(req->bio != NULL); 1756 1757 rq_qos_done(q, req); 1758 1759 /* 1760 * Request may not have originated from ll_rw_blk. if not, 1761 * it didn't come out of our reserved rq pools 1762 */ 1763 if (rq_flags & RQF_ALLOCED) { 1764 struct request_list *rl = blk_rq_rl(req); 1765 bool sync = op_is_sync(req->cmd_flags); 1766 1767 BUG_ON(!list_empty(&req->queuelist)); 1768 BUG_ON(ELV_ON_HASH(req)); 1769 1770 blk_free_request(rl, req); 1771 freed_request(rl, sync, rq_flags); 1772 blk_put_rl(rl); 1773 blk_queue_exit(q); 1774 } 1775 } 1776 EXPORT_SYMBOL_GPL(__blk_put_request); 1777 1778 void blk_put_request(struct request *req) 1779 { 1780 struct request_queue *q = req->q; 1781 1782 if (q->mq_ops) 1783 blk_mq_free_request(req); 1784 else { 1785 unsigned long flags; 1786 1787 spin_lock_irqsave(q->queue_lock, flags); 1788 __blk_put_request(q, req); 1789 spin_unlock_irqrestore(q->queue_lock, flags); 1790 } 1791 } 1792 EXPORT_SYMBOL(blk_put_request); 1793 1794 bool bio_attempt_back_merge(struct request_queue *q, struct request *req, 1795 struct bio *bio) 1796 { 1797 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 1798 1799 if (!ll_back_merge_fn(q, req, bio)) 1800 return false; 1801 1802 trace_block_bio_backmerge(q, req, bio); 1803 1804 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 1805 blk_rq_set_mixed_merge(req); 1806 1807 req->biotail->bi_next = bio; 1808 req->biotail = bio; 1809 req->__data_len += bio->bi_iter.bi_size; 1810 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio)); 1811 1812 blk_account_io_start(req, false); 1813 return true; 1814 } 1815 1816 bool bio_attempt_front_merge(struct request_queue *q, struct request *req, 1817 struct bio *bio) 1818 { 1819 const int ff = bio->bi_opf & REQ_FAILFAST_MASK; 1820 1821 if (!ll_front_merge_fn(q, req, bio)) 1822 return false; 1823 1824 trace_block_bio_frontmerge(q, req, bio); 1825 1826 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff) 1827 blk_rq_set_mixed_merge(req); 1828 1829 bio->bi_next = req->bio; 1830 req->bio = bio; 1831 1832 req->__sector = bio->bi_iter.bi_sector; 1833 req->__data_len += bio->bi_iter.bi_size; 1834 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio)); 1835 1836 blk_account_io_start(req, false); 1837 return true; 1838 } 1839 1840 bool bio_attempt_discard_merge(struct request_queue *q, struct request *req, 1841 struct bio *bio) 1842 { 1843 unsigned short segments = blk_rq_nr_discard_segments(req); 1844 1845 if (segments >= queue_max_discard_segments(q)) 1846 goto no_merge; 1847 if (blk_rq_sectors(req) + bio_sectors(bio) > 1848 blk_rq_get_max_sectors(req, blk_rq_pos(req))) 1849 goto no_merge; 1850 1851 req->biotail->bi_next = bio; 1852 req->biotail = bio; 1853 req->__data_len += bio->bi_iter.bi_size; 1854 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio)); 1855 req->nr_phys_segments = segments + 1; 1856 1857 blk_account_io_start(req, false); 1858 return true; 1859 no_merge: 1860 req_set_nomerge(q, req); 1861 return false; 1862 } 1863 1864 /** 1865 * blk_attempt_plug_merge - try to merge with %current's plugged list 1866 * @q: request_queue new bio is being queued at 1867 * @bio: new bio being queued 1868 * @request_count: out parameter for number of traversed plugged requests 1869 * @same_queue_rq: pointer to &struct request that gets filled in when 1870 * another request associated with @q is found on the plug list 1871 * (optional, may be %NULL) 1872 * 1873 * Determine whether @bio being queued on @q can be merged with a request 1874 * on %current's plugged list. Returns %true if merge was successful, 1875 * otherwise %false. 1876 * 1877 * Plugging coalesces IOs from the same issuer for the same purpose without 1878 * going through @q->queue_lock. As such it's more of an issuing mechanism 1879 * than scheduling, and the request, while may have elvpriv data, is not 1880 * added on the elevator at this point. In addition, we don't have 1881 * reliable access to the elevator outside queue lock. Only check basic 1882 * merging parameters without querying the elevator. 1883 * 1884 * Caller must ensure !blk_queue_nomerges(q) beforehand. 1885 */ 1886 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio, 1887 unsigned int *request_count, 1888 struct request **same_queue_rq) 1889 { 1890 struct blk_plug *plug; 1891 struct request *rq; 1892 struct list_head *plug_list; 1893 1894 plug = current->plug; 1895 if (!plug) 1896 return false; 1897 *request_count = 0; 1898 1899 if (q->mq_ops) 1900 plug_list = &plug->mq_list; 1901 else 1902 plug_list = &plug->list; 1903 1904 list_for_each_entry_reverse(rq, plug_list, queuelist) { 1905 bool merged = false; 1906 1907 if (rq->q == q) { 1908 (*request_count)++; 1909 /* 1910 * Only blk-mq multiple hardware queues case checks the 1911 * rq in the same queue, there should be only one such 1912 * rq in a queue 1913 **/ 1914 if (same_queue_rq) 1915 *same_queue_rq = rq; 1916 } 1917 1918 if (rq->q != q || !blk_rq_merge_ok(rq, bio)) 1919 continue; 1920 1921 switch (blk_try_merge(rq, bio)) { 1922 case ELEVATOR_BACK_MERGE: 1923 merged = bio_attempt_back_merge(q, rq, bio); 1924 break; 1925 case ELEVATOR_FRONT_MERGE: 1926 merged = bio_attempt_front_merge(q, rq, bio); 1927 break; 1928 case ELEVATOR_DISCARD_MERGE: 1929 merged = bio_attempt_discard_merge(q, rq, bio); 1930 break; 1931 default: 1932 break; 1933 } 1934 1935 if (merged) 1936 return true; 1937 } 1938 1939 return false; 1940 } 1941 1942 unsigned int blk_plug_queued_count(struct request_queue *q) 1943 { 1944 struct blk_plug *plug; 1945 struct request *rq; 1946 struct list_head *plug_list; 1947 unsigned int ret = 0; 1948 1949 plug = current->plug; 1950 if (!plug) 1951 goto out; 1952 1953 if (q->mq_ops) 1954 plug_list = &plug->mq_list; 1955 else 1956 plug_list = &plug->list; 1957 1958 list_for_each_entry(rq, plug_list, queuelist) { 1959 if (rq->q == q) 1960 ret++; 1961 } 1962 out: 1963 return ret; 1964 } 1965 1966 void blk_init_request_from_bio(struct request *req, struct bio *bio) 1967 { 1968 struct io_context *ioc = rq_ioc(bio); 1969 1970 if (bio->bi_opf & REQ_RAHEAD) 1971 req->cmd_flags |= REQ_FAILFAST_MASK; 1972 1973 req->__sector = bio->bi_iter.bi_sector; 1974 if (ioprio_valid(bio_prio(bio))) 1975 req->ioprio = bio_prio(bio); 1976 else if (ioc) 1977 req->ioprio = ioc->ioprio; 1978 else 1979 req->ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0); 1980 req->write_hint = bio->bi_write_hint; 1981 blk_rq_bio_prep(req->q, req, bio); 1982 } 1983 EXPORT_SYMBOL_GPL(blk_init_request_from_bio); 1984 1985 static blk_qc_t blk_queue_bio(struct request_queue *q, struct bio *bio) 1986 { 1987 struct blk_plug *plug; 1988 int where = ELEVATOR_INSERT_SORT; 1989 struct request *req, *free; 1990 unsigned int request_count = 0; 1991 1992 /* 1993 * low level driver can indicate that it wants pages above a 1994 * certain limit bounced to low memory (ie for highmem, or even 1995 * ISA dma in theory) 1996 */ 1997 blk_queue_bounce(q, &bio); 1998 1999 blk_queue_split(q, &bio); 2000 2001 if (!bio_integrity_prep(bio)) 2002 return BLK_QC_T_NONE; 2003 2004 if (op_is_flush(bio->bi_opf)) { 2005 spin_lock_irq(q->queue_lock); 2006 where = ELEVATOR_INSERT_FLUSH; 2007 goto get_rq; 2008 } 2009 2010 /* 2011 * Check if we can merge with the plugged list before grabbing 2012 * any locks. 2013 */ 2014 if (!blk_queue_nomerges(q)) { 2015 if (blk_attempt_plug_merge(q, bio, &request_count, NULL)) 2016 return BLK_QC_T_NONE; 2017 } else 2018 request_count = blk_plug_queued_count(q); 2019 2020 spin_lock_irq(q->queue_lock); 2021 2022 switch (elv_merge(q, &req, bio)) { 2023 case ELEVATOR_BACK_MERGE: 2024 if (!bio_attempt_back_merge(q, req, bio)) 2025 break; 2026 elv_bio_merged(q, req, bio); 2027 free = attempt_back_merge(q, req); 2028 if (free) 2029 __blk_put_request(q, free); 2030 else 2031 elv_merged_request(q, req, ELEVATOR_BACK_MERGE); 2032 goto out_unlock; 2033 case ELEVATOR_FRONT_MERGE: 2034 if (!bio_attempt_front_merge(q, req, bio)) 2035 break; 2036 elv_bio_merged(q, req, bio); 2037 free = attempt_front_merge(q, req); 2038 if (free) 2039 __blk_put_request(q, free); 2040 else 2041 elv_merged_request(q, req, ELEVATOR_FRONT_MERGE); 2042 goto out_unlock; 2043 default: 2044 break; 2045 } 2046 2047 get_rq: 2048 rq_qos_throttle(q, bio, q->queue_lock); 2049 2050 /* 2051 * Grab a free request. This is might sleep but can not fail. 2052 * Returns with the queue unlocked. 2053 */ 2054 blk_queue_enter_live(q); 2055 req = get_request(q, bio->bi_opf, bio, 0, GFP_NOIO); 2056 if (IS_ERR(req)) { 2057 blk_queue_exit(q); 2058 rq_qos_cleanup(q, bio); 2059 if (PTR_ERR(req) == -ENOMEM) 2060 bio->bi_status = BLK_STS_RESOURCE; 2061 else 2062 bio->bi_status = BLK_STS_IOERR; 2063 bio_endio(bio); 2064 goto out_unlock; 2065 } 2066 2067 rq_qos_track(q, req, bio); 2068 2069 /* 2070 * After dropping the lock and possibly sleeping here, our request 2071 * may now be mergeable after it had proven unmergeable (above). 2072 * We don't worry about that case for efficiency. It won't happen 2073 * often, and the elevators are able to handle it. 2074 */ 2075 blk_init_request_from_bio(req, bio); 2076 2077 if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) 2078 req->cpu = raw_smp_processor_id(); 2079 2080 plug = current->plug; 2081 if (plug) { 2082 /* 2083 * If this is the first request added after a plug, fire 2084 * of a plug trace. 2085 * 2086 * @request_count may become stale because of schedule 2087 * out, so check plug list again. 2088 */ 2089 if (!request_count || list_empty(&plug->list)) 2090 trace_block_plug(q); 2091 else { 2092 struct request *last = list_entry_rq(plug->list.prev); 2093 if (request_count >= BLK_MAX_REQUEST_COUNT || 2094 blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE) { 2095 blk_flush_plug_list(plug, false); 2096 trace_block_plug(q); 2097 } 2098 } 2099 list_add_tail(&req->queuelist, &plug->list); 2100 blk_account_io_start(req, true); 2101 } else { 2102 spin_lock_irq(q->queue_lock); 2103 add_acct_request(q, req, where); 2104 __blk_run_queue(q); 2105 out_unlock: 2106 spin_unlock_irq(q->queue_lock); 2107 } 2108 2109 return BLK_QC_T_NONE; 2110 } 2111 2112 static void handle_bad_sector(struct bio *bio, sector_t maxsector) 2113 { 2114 char b[BDEVNAME_SIZE]; 2115 2116 printk(KERN_INFO "attempt to access beyond end of device\n"); 2117 printk(KERN_INFO "%s: rw=%d, want=%Lu, limit=%Lu\n", 2118 bio_devname(bio, b), bio->bi_opf, 2119 (unsigned long long)bio_end_sector(bio), 2120 (long long)maxsector); 2121 } 2122 2123 #ifdef CONFIG_FAIL_MAKE_REQUEST 2124 2125 static DECLARE_FAULT_ATTR(fail_make_request); 2126 2127 static int __init setup_fail_make_request(char *str) 2128 { 2129 return setup_fault_attr(&fail_make_request, str); 2130 } 2131 __setup("fail_make_request=", setup_fail_make_request); 2132 2133 static bool should_fail_request(struct hd_struct *part, unsigned int bytes) 2134 { 2135 return part->make_it_fail && should_fail(&fail_make_request, bytes); 2136 } 2137 2138 static int __init fail_make_request_debugfs(void) 2139 { 2140 struct dentry *dir = fault_create_debugfs_attr("fail_make_request", 2141 NULL, &fail_make_request); 2142 2143 return PTR_ERR_OR_ZERO(dir); 2144 } 2145 2146 late_initcall(fail_make_request_debugfs); 2147 2148 #else /* CONFIG_FAIL_MAKE_REQUEST */ 2149 2150 static inline bool should_fail_request(struct hd_struct *part, 2151 unsigned int bytes) 2152 { 2153 return false; 2154 } 2155 2156 #endif /* CONFIG_FAIL_MAKE_REQUEST */ 2157 2158 static inline bool bio_check_ro(struct bio *bio, struct hd_struct *part) 2159 { 2160 const int op = bio_op(bio); 2161 2162 if (part->policy && op_is_write(op)) { 2163 char b[BDEVNAME_SIZE]; 2164 2165 if (op_is_flush(bio->bi_opf) && !bio_sectors(bio)) 2166 return false; 2167 2168 WARN_ONCE(1, 2169 "generic_make_request: Trying to write " 2170 "to read-only block-device %s (partno %d)\n", 2171 bio_devname(bio, b), part->partno); 2172 /* Older lvm-tools actually trigger this */ 2173 return false; 2174 } 2175 2176 return false; 2177 } 2178 2179 static noinline int should_fail_bio(struct bio *bio) 2180 { 2181 if (should_fail_request(&bio->bi_disk->part0, bio->bi_iter.bi_size)) 2182 return -EIO; 2183 return 0; 2184 } 2185 ALLOW_ERROR_INJECTION(should_fail_bio, ERRNO); 2186 2187 /* 2188 * Check whether this bio extends beyond the end of the device or partition. 2189 * This may well happen - the kernel calls bread() without checking the size of 2190 * the device, e.g., when mounting a file system. 2191 */ 2192 static inline int bio_check_eod(struct bio *bio, sector_t maxsector) 2193 { 2194 unsigned int nr_sectors = bio_sectors(bio); 2195 2196 if (nr_sectors && maxsector && 2197 (nr_sectors > maxsector || 2198 bio->bi_iter.bi_sector > maxsector - nr_sectors)) { 2199 handle_bad_sector(bio, maxsector); 2200 return -EIO; 2201 } 2202 return 0; 2203 } 2204 2205 /* 2206 * Remap block n of partition p to block n+start(p) of the disk. 2207 */ 2208 static inline int blk_partition_remap(struct bio *bio) 2209 { 2210 struct hd_struct *p; 2211 int ret = -EIO; 2212 2213 rcu_read_lock(); 2214 p = __disk_get_part(bio->bi_disk, bio->bi_partno); 2215 if (unlikely(!p)) 2216 goto out; 2217 if (unlikely(should_fail_request(p, bio->bi_iter.bi_size))) 2218 goto out; 2219 if (unlikely(bio_check_ro(bio, p))) 2220 goto out; 2221 2222 /* 2223 * Zone reset does not include bi_size so bio_sectors() is always 0. 2224 * Include a test for the reset op code and perform the remap if needed. 2225 */ 2226 if (bio_sectors(bio) || bio_op(bio) == REQ_OP_ZONE_RESET) { 2227 if (bio_check_eod(bio, part_nr_sects_read(p))) 2228 goto out; 2229 bio->bi_iter.bi_sector += p->start_sect; 2230 trace_block_bio_remap(bio->bi_disk->queue, bio, part_devt(p), 2231 bio->bi_iter.bi_sector - p->start_sect); 2232 } 2233 bio->bi_partno = 0; 2234 ret = 0; 2235 out: 2236 rcu_read_unlock(); 2237 return ret; 2238 } 2239 2240 static noinline_for_stack bool 2241 generic_make_request_checks(struct bio *bio) 2242 { 2243 struct request_queue *q; 2244 int nr_sectors = bio_sectors(bio); 2245 blk_status_t status = BLK_STS_IOERR; 2246 char b[BDEVNAME_SIZE]; 2247 2248 might_sleep(); 2249 2250 q = bio->bi_disk->queue; 2251 if (unlikely(!q)) { 2252 printk(KERN_ERR 2253 "generic_make_request: Trying to access " 2254 "nonexistent block-device %s (%Lu)\n", 2255 bio_devname(bio, b), (long long)bio->bi_iter.bi_sector); 2256 goto end_io; 2257 } 2258 2259 /* 2260 * For a REQ_NOWAIT based request, return -EOPNOTSUPP 2261 * if queue is not a request based queue. 2262 */ 2263 if ((bio->bi_opf & REQ_NOWAIT) && !queue_is_rq_based(q)) 2264 goto not_supported; 2265 2266 if (should_fail_bio(bio)) 2267 goto end_io; 2268 2269 if (bio->bi_partno) { 2270 if (unlikely(blk_partition_remap(bio))) 2271 goto end_io; 2272 } else { 2273 if (unlikely(bio_check_ro(bio, &bio->bi_disk->part0))) 2274 goto end_io; 2275 if (unlikely(bio_check_eod(bio, get_capacity(bio->bi_disk)))) 2276 goto end_io; 2277 } 2278 2279 /* 2280 * Filter flush bio's early so that make_request based 2281 * drivers without flush support don't have to worry 2282 * about them. 2283 */ 2284 if (op_is_flush(bio->bi_opf) && 2285 !test_bit(QUEUE_FLAG_WC, &q->queue_flags)) { 2286 bio->bi_opf &= ~(REQ_PREFLUSH | REQ_FUA); 2287 if (!nr_sectors) { 2288 status = BLK_STS_OK; 2289 goto end_io; 2290 } 2291 } 2292 2293 switch (bio_op(bio)) { 2294 case REQ_OP_DISCARD: 2295 if (!blk_queue_discard(q)) 2296 goto not_supported; 2297 break; 2298 case REQ_OP_SECURE_ERASE: 2299 if (!blk_queue_secure_erase(q)) 2300 goto not_supported; 2301 break; 2302 case REQ_OP_WRITE_SAME: 2303 if (!q->limits.max_write_same_sectors) 2304 goto not_supported; 2305 break; 2306 case REQ_OP_ZONE_RESET: 2307 if (!blk_queue_is_zoned(q)) 2308 goto not_supported; 2309 break; 2310 case REQ_OP_WRITE_ZEROES: 2311 if (!q->limits.max_write_zeroes_sectors) 2312 goto not_supported; 2313 break; 2314 default: 2315 break; 2316 } 2317 2318 /* 2319 * Various block parts want %current->io_context and lazy ioc 2320 * allocation ends up trading a lot of pain for a small amount of 2321 * memory. Just allocate it upfront. This may fail and block 2322 * layer knows how to live with it. 2323 */ 2324 create_io_context(GFP_ATOMIC, q->node); 2325 2326 if (!blkcg_bio_issue_check(q, bio)) 2327 return false; 2328 2329 if (!bio_flagged(bio, BIO_TRACE_COMPLETION)) { 2330 trace_block_bio_queue(q, bio); 2331 /* Now that enqueuing has been traced, we need to trace 2332 * completion as well. 2333 */ 2334 bio_set_flag(bio, BIO_TRACE_COMPLETION); 2335 } 2336 return true; 2337 2338 not_supported: 2339 status = BLK_STS_NOTSUPP; 2340 end_io: 2341 bio->bi_status = status; 2342 bio_endio(bio); 2343 return false; 2344 } 2345 2346 /** 2347 * generic_make_request - hand a buffer to its device driver for I/O 2348 * @bio: The bio describing the location in memory and on the device. 2349 * 2350 * generic_make_request() is used to make I/O requests of block 2351 * devices. It is passed a &struct bio, which describes the I/O that needs 2352 * to be done. 2353 * 2354 * generic_make_request() does not return any status. The 2355 * success/failure status of the request, along with notification of 2356 * completion, is delivered asynchronously through the bio->bi_end_io 2357 * function described (one day) else where. 2358 * 2359 * The caller of generic_make_request must make sure that bi_io_vec 2360 * are set to describe the memory buffer, and that bi_dev and bi_sector are 2361 * set to describe the device address, and the 2362 * bi_end_io and optionally bi_private are set to describe how 2363 * completion notification should be signaled. 2364 * 2365 * generic_make_request and the drivers it calls may use bi_next if this 2366 * bio happens to be merged with someone else, and may resubmit the bio to 2367 * a lower device by calling into generic_make_request recursively, which 2368 * means the bio should NOT be touched after the call to ->make_request_fn. 2369 */ 2370 blk_qc_t generic_make_request(struct bio *bio) 2371 { 2372 /* 2373 * bio_list_on_stack[0] contains bios submitted by the current 2374 * make_request_fn. 2375 * bio_list_on_stack[1] contains bios that were submitted before 2376 * the current make_request_fn, but that haven't been processed 2377 * yet. 2378 */ 2379 struct bio_list bio_list_on_stack[2]; 2380 blk_mq_req_flags_t flags = 0; 2381 struct request_queue *q = bio->bi_disk->queue; 2382 blk_qc_t ret = BLK_QC_T_NONE; 2383 2384 if (bio->bi_opf & REQ_NOWAIT) 2385 flags = BLK_MQ_REQ_NOWAIT; 2386 if (bio_flagged(bio, BIO_QUEUE_ENTERED)) 2387 blk_queue_enter_live(q); 2388 else if (blk_queue_enter(q, flags) < 0) { 2389 if (!blk_queue_dying(q) && (bio->bi_opf & REQ_NOWAIT)) 2390 bio_wouldblock_error(bio); 2391 else 2392 bio_io_error(bio); 2393 return ret; 2394 } 2395 2396 if (!generic_make_request_checks(bio)) 2397 goto out; 2398 2399 /* 2400 * We only want one ->make_request_fn to be active at a time, else 2401 * stack usage with stacked devices could be a problem. So use 2402 * current->bio_list to keep a list of requests submited by a 2403 * make_request_fn function. current->bio_list is also used as a 2404 * flag to say if generic_make_request is currently active in this 2405 * task or not. If it is NULL, then no make_request is active. If 2406 * it is non-NULL, then a make_request is active, and new requests 2407 * should be added at the tail 2408 */ 2409 if (current->bio_list) { 2410 bio_list_add(¤t->bio_list[0], bio); 2411 goto out; 2412 } 2413 2414 /* following loop may be a bit non-obvious, and so deserves some 2415 * explanation. 2416 * Before entering the loop, bio->bi_next is NULL (as all callers 2417 * ensure that) so we have a list with a single bio. 2418 * We pretend that we have just taken it off a longer list, so 2419 * we assign bio_list to a pointer to the bio_list_on_stack, 2420 * thus initialising the bio_list of new bios to be 2421 * added. ->make_request() may indeed add some more bios 2422 * through a recursive call to generic_make_request. If it 2423 * did, we find a non-NULL value in bio_list and re-enter the loop 2424 * from the top. In this case we really did just take the bio 2425 * of the top of the list (no pretending) and so remove it from 2426 * bio_list, and call into ->make_request() again. 2427 */ 2428 BUG_ON(bio->bi_next); 2429 bio_list_init(&bio_list_on_stack[0]); 2430 current->bio_list = bio_list_on_stack; 2431 do { 2432 bool enter_succeeded = true; 2433 2434 if (unlikely(q != bio->bi_disk->queue)) { 2435 if (q) 2436 blk_queue_exit(q); 2437 q = bio->bi_disk->queue; 2438 flags = 0; 2439 if (bio->bi_opf & REQ_NOWAIT) 2440 flags = BLK_MQ_REQ_NOWAIT; 2441 if (blk_queue_enter(q, flags) < 0) { 2442 enter_succeeded = false; 2443 q = NULL; 2444 } 2445 } 2446 2447 if (enter_succeeded) { 2448 struct bio_list lower, same; 2449 2450 /* Create a fresh bio_list for all subordinate requests */ 2451 bio_list_on_stack[1] = bio_list_on_stack[0]; 2452 bio_list_init(&bio_list_on_stack[0]); 2453 ret = q->make_request_fn(q, bio); 2454 2455 /* sort new bios into those for a lower level 2456 * and those for the same level 2457 */ 2458 bio_list_init(&lower); 2459 bio_list_init(&same); 2460 while ((bio = bio_list_pop(&bio_list_on_stack[0])) != NULL) 2461 if (q == bio->bi_disk->queue) 2462 bio_list_add(&same, bio); 2463 else 2464 bio_list_add(&lower, bio); 2465 /* now assemble so we handle the lowest level first */ 2466 bio_list_merge(&bio_list_on_stack[0], &lower); 2467 bio_list_merge(&bio_list_on_stack[0], &same); 2468 bio_list_merge(&bio_list_on_stack[0], &bio_list_on_stack[1]); 2469 } else { 2470 if (unlikely(!blk_queue_dying(q) && 2471 (bio->bi_opf & REQ_NOWAIT))) 2472 bio_wouldblock_error(bio); 2473 else 2474 bio_io_error(bio); 2475 } 2476 bio = bio_list_pop(&bio_list_on_stack[0]); 2477 } while (bio); 2478 current->bio_list = NULL; /* deactivate */ 2479 2480 out: 2481 if (q) 2482 blk_queue_exit(q); 2483 return ret; 2484 } 2485 EXPORT_SYMBOL(generic_make_request); 2486 2487 /** 2488 * direct_make_request - hand a buffer directly to its device driver for I/O 2489 * @bio: The bio describing the location in memory and on the device. 2490 * 2491 * This function behaves like generic_make_request(), but does not protect 2492 * against recursion. Must only be used if the called driver is known 2493 * to not call generic_make_request (or direct_make_request) again from 2494 * its make_request function. (Calling direct_make_request again from 2495 * a workqueue is perfectly fine as that doesn't recurse). 2496 */ 2497 blk_qc_t direct_make_request(struct bio *bio) 2498 { 2499 struct request_queue *q = bio->bi_disk->queue; 2500 bool nowait = bio->bi_opf & REQ_NOWAIT; 2501 blk_qc_t ret; 2502 2503 if (!generic_make_request_checks(bio)) 2504 return BLK_QC_T_NONE; 2505 2506 if (unlikely(blk_queue_enter(q, nowait ? BLK_MQ_REQ_NOWAIT : 0))) { 2507 if (nowait && !blk_queue_dying(q)) 2508 bio->bi_status = BLK_STS_AGAIN; 2509 else 2510 bio->bi_status = BLK_STS_IOERR; 2511 bio_endio(bio); 2512 return BLK_QC_T_NONE; 2513 } 2514 2515 ret = q->make_request_fn(q, bio); 2516 blk_queue_exit(q); 2517 return ret; 2518 } 2519 EXPORT_SYMBOL_GPL(direct_make_request); 2520 2521 /** 2522 * submit_bio - submit a bio to the block device layer for I/O 2523 * @bio: The &struct bio which describes the I/O 2524 * 2525 * submit_bio() is very similar in purpose to generic_make_request(), and 2526 * uses that function to do most of the work. Both are fairly rough 2527 * interfaces; @bio must be presetup and ready for I/O. 2528 * 2529 */ 2530 blk_qc_t submit_bio(struct bio *bio) 2531 { 2532 /* 2533 * If it's a regular read/write or a barrier with data attached, 2534 * go through the normal accounting stuff before submission. 2535 */ 2536 if (bio_has_data(bio)) { 2537 unsigned int count; 2538 2539 if (unlikely(bio_op(bio) == REQ_OP_WRITE_SAME)) 2540 count = queue_logical_block_size(bio->bi_disk->queue) >> 9; 2541 else 2542 count = bio_sectors(bio); 2543 2544 if (op_is_write(bio_op(bio))) { 2545 count_vm_events(PGPGOUT, count); 2546 } else { 2547 task_io_account_read(bio->bi_iter.bi_size); 2548 count_vm_events(PGPGIN, count); 2549 } 2550 2551 if (unlikely(block_dump)) { 2552 char b[BDEVNAME_SIZE]; 2553 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s (%u sectors)\n", 2554 current->comm, task_pid_nr(current), 2555 op_is_write(bio_op(bio)) ? "WRITE" : "READ", 2556 (unsigned long long)bio->bi_iter.bi_sector, 2557 bio_devname(bio, b), count); 2558 } 2559 } 2560 2561 return generic_make_request(bio); 2562 } 2563 EXPORT_SYMBOL(submit_bio); 2564 2565 bool blk_poll(struct request_queue *q, blk_qc_t cookie) 2566 { 2567 if (!q->poll_fn || !blk_qc_t_valid(cookie)) 2568 return false; 2569 2570 if (current->plug) 2571 blk_flush_plug_list(current->plug, false); 2572 return q->poll_fn(q, cookie); 2573 } 2574 EXPORT_SYMBOL_GPL(blk_poll); 2575 2576 /** 2577 * blk_cloned_rq_check_limits - Helper function to check a cloned request 2578 * for new the queue limits 2579 * @q: the queue 2580 * @rq: the request being checked 2581 * 2582 * Description: 2583 * @rq may have been made based on weaker limitations of upper-level queues 2584 * in request stacking drivers, and it may violate the limitation of @q. 2585 * Since the block layer and the underlying device driver trust @rq 2586 * after it is inserted to @q, it should be checked against @q before 2587 * the insertion using this generic function. 2588 * 2589 * Request stacking drivers like request-based dm may change the queue 2590 * limits when retrying requests on other queues. Those requests need 2591 * to be checked against the new queue limits again during dispatch. 2592 */ 2593 static int blk_cloned_rq_check_limits(struct request_queue *q, 2594 struct request *rq) 2595 { 2596 if (blk_rq_sectors(rq) > blk_queue_get_max_sectors(q, req_op(rq))) { 2597 printk(KERN_ERR "%s: over max size limit.\n", __func__); 2598 return -EIO; 2599 } 2600 2601 /* 2602 * queue's settings related to segment counting like q->bounce_pfn 2603 * may differ from that of other stacking queues. 2604 * Recalculate it to check the request correctly on this queue's 2605 * limitation. 2606 */ 2607 blk_recalc_rq_segments(rq); 2608 if (rq->nr_phys_segments > queue_max_segments(q)) { 2609 printk(KERN_ERR "%s: over max segments limit.\n", __func__); 2610 return -EIO; 2611 } 2612 2613 return 0; 2614 } 2615 2616 /** 2617 * blk_insert_cloned_request - Helper for stacking drivers to submit a request 2618 * @q: the queue to submit the request 2619 * @rq: the request being queued 2620 */ 2621 blk_status_t blk_insert_cloned_request(struct request_queue *q, struct request *rq) 2622 { 2623 unsigned long flags; 2624 int where = ELEVATOR_INSERT_BACK; 2625 2626 if (blk_cloned_rq_check_limits(q, rq)) 2627 return BLK_STS_IOERR; 2628 2629 if (rq->rq_disk && 2630 should_fail_request(&rq->rq_disk->part0, blk_rq_bytes(rq))) 2631 return BLK_STS_IOERR; 2632 2633 if (q->mq_ops) { 2634 if (blk_queue_io_stat(q)) 2635 blk_account_io_start(rq, true); 2636 /* 2637 * Since we have a scheduler attached on the top device, 2638 * bypass a potential scheduler on the bottom device for 2639 * insert. 2640 */ 2641 return blk_mq_request_issue_directly(rq); 2642 } 2643 2644 spin_lock_irqsave(q->queue_lock, flags); 2645 if (unlikely(blk_queue_dying(q))) { 2646 spin_unlock_irqrestore(q->queue_lock, flags); 2647 return BLK_STS_IOERR; 2648 } 2649 2650 /* 2651 * Submitting request must be dequeued before calling this function 2652 * because it will be linked to another request_queue 2653 */ 2654 BUG_ON(blk_queued_rq(rq)); 2655 2656 if (op_is_flush(rq->cmd_flags)) 2657 where = ELEVATOR_INSERT_FLUSH; 2658 2659 add_acct_request(q, rq, where); 2660 if (where == ELEVATOR_INSERT_FLUSH) 2661 __blk_run_queue(q); 2662 spin_unlock_irqrestore(q->queue_lock, flags); 2663 2664 return BLK_STS_OK; 2665 } 2666 EXPORT_SYMBOL_GPL(blk_insert_cloned_request); 2667 2668 /** 2669 * blk_rq_err_bytes - determine number of bytes till the next failure boundary 2670 * @rq: request to examine 2671 * 2672 * Description: 2673 * A request could be merge of IOs which require different failure 2674 * handling. This function determines the number of bytes which 2675 * can be failed from the beginning of the request without 2676 * crossing into area which need to be retried further. 2677 * 2678 * Return: 2679 * The number of bytes to fail. 2680 */ 2681 unsigned int blk_rq_err_bytes(const struct request *rq) 2682 { 2683 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK; 2684 unsigned int bytes = 0; 2685 struct bio *bio; 2686 2687 if (!(rq->rq_flags & RQF_MIXED_MERGE)) 2688 return blk_rq_bytes(rq); 2689 2690 /* 2691 * Currently the only 'mixing' which can happen is between 2692 * different fastfail types. We can safely fail portions 2693 * which have all the failfast bits that the first one has - 2694 * the ones which are at least as eager to fail as the first 2695 * one. 2696 */ 2697 for (bio = rq->bio; bio; bio = bio->bi_next) { 2698 if ((bio->bi_opf & ff) != ff) 2699 break; 2700 bytes += bio->bi_iter.bi_size; 2701 } 2702 2703 /* this could lead to infinite loop */ 2704 BUG_ON(blk_rq_bytes(rq) && !bytes); 2705 return bytes; 2706 } 2707 EXPORT_SYMBOL_GPL(blk_rq_err_bytes); 2708 2709 void blk_account_io_completion(struct request *req, unsigned int bytes) 2710 { 2711 if (blk_do_io_stat(req)) { 2712 const int sgrp = op_stat_group(req_op(req)); 2713 struct hd_struct *part; 2714 int cpu; 2715 2716 cpu = part_stat_lock(); 2717 part = req->part; 2718 part_stat_add(cpu, part, sectors[sgrp], bytes >> 9); 2719 part_stat_unlock(); 2720 } 2721 } 2722 2723 void blk_account_io_done(struct request *req, u64 now) 2724 { 2725 /* 2726 * Account IO completion. flush_rq isn't accounted as a 2727 * normal IO on queueing nor completion. Accounting the 2728 * containing request is enough. 2729 */ 2730 if (blk_do_io_stat(req) && !(req->rq_flags & RQF_FLUSH_SEQ)) { 2731 const int sgrp = op_stat_group(req_op(req)); 2732 struct hd_struct *part; 2733 int cpu; 2734 2735 cpu = part_stat_lock(); 2736 part = req->part; 2737 2738 part_stat_inc(cpu, part, ios[sgrp]); 2739 part_stat_add(cpu, part, nsecs[sgrp], now - req->start_time_ns); 2740 part_round_stats(req->q, cpu, part); 2741 part_dec_in_flight(req->q, part, rq_data_dir(req)); 2742 2743 hd_struct_put(part); 2744 part_stat_unlock(); 2745 } 2746 } 2747 2748 void blk_account_io_start(struct request *rq, bool new_io) 2749 { 2750 struct hd_struct *part; 2751 int rw = rq_data_dir(rq); 2752 int cpu; 2753 2754 if (!blk_do_io_stat(rq)) 2755 return; 2756 2757 cpu = part_stat_lock(); 2758 2759 if (!new_io) { 2760 part = rq->part; 2761 part_stat_inc(cpu, part, merges[rw]); 2762 } else { 2763 part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq)); 2764 if (!hd_struct_try_get(part)) { 2765 /* 2766 * The partition is already being removed, 2767 * the request will be accounted on the disk only 2768 * 2769 * We take a reference on disk->part0 although that 2770 * partition will never be deleted, so we can treat 2771 * it as any other partition. 2772 */ 2773 part = &rq->rq_disk->part0; 2774 hd_struct_get(part); 2775 } 2776 part_round_stats(rq->q, cpu, part); 2777 part_inc_in_flight(rq->q, part, rw); 2778 rq->part = part; 2779 } 2780 2781 part_stat_unlock(); 2782 } 2783 2784 static struct request *elv_next_request(struct request_queue *q) 2785 { 2786 struct request *rq; 2787 struct blk_flush_queue *fq = blk_get_flush_queue(q, NULL); 2788 2789 WARN_ON_ONCE(q->mq_ops); 2790 2791 while (1) { 2792 list_for_each_entry(rq, &q->queue_head, queuelist) { 2793 #ifdef CONFIG_PM 2794 /* 2795 * If a request gets queued in state RPM_SUSPENDED 2796 * then that's a kernel bug. 2797 */ 2798 WARN_ON_ONCE(q->rpm_status == RPM_SUSPENDED); 2799 #endif 2800 return rq; 2801 } 2802 2803 /* 2804 * Flush request is running and flush request isn't queueable 2805 * in the drive, we can hold the queue till flush request is 2806 * finished. Even we don't do this, driver can't dispatch next 2807 * requests and will requeue them. And this can improve 2808 * throughput too. For example, we have request flush1, write1, 2809 * flush 2. flush1 is dispatched, then queue is hold, write1 2810 * isn't inserted to queue. After flush1 is finished, flush2 2811 * will be dispatched. Since disk cache is already clean, 2812 * flush2 will be finished very soon, so looks like flush2 is 2813 * folded to flush1. 2814 * Since the queue is hold, a flag is set to indicate the queue 2815 * should be restarted later. Please see flush_end_io() for 2816 * details. 2817 */ 2818 if (fq->flush_pending_idx != fq->flush_running_idx && 2819 !queue_flush_queueable(q)) { 2820 fq->flush_queue_delayed = 1; 2821 return NULL; 2822 } 2823 if (unlikely(blk_queue_bypass(q)) || 2824 !q->elevator->type->ops.sq.elevator_dispatch_fn(q, 0)) 2825 return NULL; 2826 } 2827 } 2828 2829 /** 2830 * blk_peek_request - peek at the top of a request queue 2831 * @q: request queue to peek at 2832 * 2833 * Description: 2834 * Return the request at the top of @q. The returned request 2835 * should be started using blk_start_request() before LLD starts 2836 * processing it. 2837 * 2838 * Return: 2839 * Pointer to the request at the top of @q if available. Null 2840 * otherwise. 2841 */ 2842 struct request *blk_peek_request(struct request_queue *q) 2843 { 2844 struct request *rq; 2845 int ret; 2846 2847 lockdep_assert_held(q->queue_lock); 2848 WARN_ON_ONCE(q->mq_ops); 2849 2850 while ((rq = elv_next_request(q)) != NULL) { 2851 if (!(rq->rq_flags & RQF_STARTED)) { 2852 /* 2853 * This is the first time the device driver 2854 * sees this request (possibly after 2855 * requeueing). Notify IO scheduler. 2856 */ 2857 if (rq->rq_flags & RQF_SORTED) 2858 elv_activate_rq(q, rq); 2859 2860 /* 2861 * just mark as started even if we don't start 2862 * it, a request that has been delayed should 2863 * not be passed by new incoming requests 2864 */ 2865 rq->rq_flags |= RQF_STARTED; 2866 trace_block_rq_issue(q, rq); 2867 } 2868 2869 if (!q->boundary_rq || q->boundary_rq == rq) { 2870 q->end_sector = rq_end_sector(rq); 2871 q->boundary_rq = NULL; 2872 } 2873 2874 if (rq->rq_flags & RQF_DONTPREP) 2875 break; 2876 2877 if (q->dma_drain_size && blk_rq_bytes(rq)) { 2878 /* 2879 * make sure space for the drain appears we 2880 * know we can do this because max_hw_segments 2881 * has been adjusted to be one fewer than the 2882 * device can handle 2883 */ 2884 rq->nr_phys_segments++; 2885 } 2886 2887 if (!q->prep_rq_fn) 2888 break; 2889 2890 ret = q->prep_rq_fn(q, rq); 2891 if (ret == BLKPREP_OK) { 2892 break; 2893 } else if (ret == BLKPREP_DEFER) { 2894 /* 2895 * the request may have been (partially) prepped. 2896 * we need to keep this request in the front to 2897 * avoid resource deadlock. RQF_STARTED will 2898 * prevent other fs requests from passing this one. 2899 */ 2900 if (q->dma_drain_size && blk_rq_bytes(rq) && 2901 !(rq->rq_flags & RQF_DONTPREP)) { 2902 /* 2903 * remove the space for the drain we added 2904 * so that we don't add it again 2905 */ 2906 --rq->nr_phys_segments; 2907 } 2908 2909 rq = NULL; 2910 break; 2911 } else if (ret == BLKPREP_KILL || ret == BLKPREP_INVALID) { 2912 rq->rq_flags |= RQF_QUIET; 2913 /* 2914 * Mark this request as started so we don't trigger 2915 * any debug logic in the end I/O path. 2916 */ 2917 blk_start_request(rq); 2918 __blk_end_request_all(rq, ret == BLKPREP_INVALID ? 2919 BLK_STS_TARGET : BLK_STS_IOERR); 2920 } else { 2921 printk(KERN_ERR "%s: bad return=%d\n", __func__, ret); 2922 break; 2923 } 2924 } 2925 2926 return rq; 2927 } 2928 EXPORT_SYMBOL(blk_peek_request); 2929 2930 static void blk_dequeue_request(struct request *rq) 2931 { 2932 struct request_queue *q = rq->q; 2933 2934 BUG_ON(list_empty(&rq->queuelist)); 2935 BUG_ON(ELV_ON_HASH(rq)); 2936 2937 list_del_init(&rq->queuelist); 2938 2939 /* 2940 * the time frame between a request being removed from the lists 2941 * and to it is freed is accounted as io that is in progress at 2942 * the driver side. 2943 */ 2944 if (blk_account_rq(rq)) 2945 q->in_flight[rq_is_sync(rq)]++; 2946 } 2947 2948 /** 2949 * blk_start_request - start request processing on the driver 2950 * @req: request to dequeue 2951 * 2952 * Description: 2953 * Dequeue @req and start timeout timer on it. This hands off the 2954 * request to the driver. 2955 */ 2956 void blk_start_request(struct request *req) 2957 { 2958 lockdep_assert_held(req->q->queue_lock); 2959 WARN_ON_ONCE(req->q->mq_ops); 2960 2961 blk_dequeue_request(req); 2962 2963 if (test_bit(QUEUE_FLAG_STATS, &req->q->queue_flags)) { 2964 req->io_start_time_ns = ktime_get_ns(); 2965 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW 2966 req->throtl_size = blk_rq_sectors(req); 2967 #endif 2968 req->rq_flags |= RQF_STATS; 2969 rq_qos_issue(req->q, req); 2970 } 2971 2972 BUG_ON(blk_rq_is_complete(req)); 2973 blk_add_timer(req); 2974 } 2975 EXPORT_SYMBOL(blk_start_request); 2976 2977 /** 2978 * blk_fetch_request - fetch a request from a request queue 2979 * @q: request queue to fetch a request from 2980 * 2981 * Description: 2982 * Return the request at the top of @q. The request is started on 2983 * return and LLD can start processing it immediately. 2984 * 2985 * Return: 2986 * Pointer to the request at the top of @q if available. Null 2987 * otherwise. 2988 */ 2989 struct request *blk_fetch_request(struct request_queue *q) 2990 { 2991 struct request *rq; 2992 2993 lockdep_assert_held(q->queue_lock); 2994 WARN_ON_ONCE(q->mq_ops); 2995 2996 rq = blk_peek_request(q); 2997 if (rq) 2998 blk_start_request(rq); 2999 return rq; 3000 } 3001 EXPORT_SYMBOL(blk_fetch_request); 3002 3003 /* 3004 * Steal bios from a request and add them to a bio list. 3005 * The request must not have been partially completed before. 3006 */ 3007 void blk_steal_bios(struct bio_list *list, struct request *rq) 3008 { 3009 if (rq->bio) { 3010 if (list->tail) 3011 list->tail->bi_next = rq->bio; 3012 else 3013 list->head = rq->bio; 3014 list->tail = rq->biotail; 3015 3016 rq->bio = NULL; 3017 rq->biotail = NULL; 3018 } 3019 3020 rq->__data_len = 0; 3021 } 3022 EXPORT_SYMBOL_GPL(blk_steal_bios); 3023 3024 /** 3025 * blk_update_request - Special helper function for request stacking drivers 3026 * @req: the request being processed 3027 * @error: block status code 3028 * @nr_bytes: number of bytes to complete @req 3029 * 3030 * Description: 3031 * Ends I/O on a number of bytes attached to @req, but doesn't complete 3032 * the request structure even if @req doesn't have leftover. 3033 * If @req has leftover, sets it up for the next range of segments. 3034 * 3035 * This special helper function is only for request stacking drivers 3036 * (e.g. request-based dm) so that they can handle partial completion. 3037 * Actual device drivers should use blk_end_request instead. 3038 * 3039 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees 3040 * %false return from this function. 3041 * 3042 * Note: 3043 * The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in both 3044 * blk_rq_bytes() and in blk_update_request(). 3045 * 3046 * Return: 3047 * %false - this request doesn't have any more data 3048 * %true - this request has more data 3049 **/ 3050 bool blk_update_request(struct request *req, blk_status_t error, 3051 unsigned int nr_bytes) 3052 { 3053 int total_bytes; 3054 3055 trace_block_rq_complete(req, blk_status_to_errno(error), nr_bytes); 3056 3057 if (!req->bio) 3058 return false; 3059 3060 if (unlikely(error && !blk_rq_is_passthrough(req) && 3061 !(req->rq_flags & RQF_QUIET))) 3062 print_req_error(req, error); 3063 3064 blk_account_io_completion(req, nr_bytes); 3065 3066 total_bytes = 0; 3067 while (req->bio) { 3068 struct bio *bio = req->bio; 3069 unsigned bio_bytes = min(bio->bi_iter.bi_size, nr_bytes); 3070 3071 if (bio_bytes == bio->bi_iter.bi_size) 3072 req->bio = bio->bi_next; 3073 3074 /* Completion has already been traced */ 3075 bio_clear_flag(bio, BIO_TRACE_COMPLETION); 3076 req_bio_endio(req, bio, bio_bytes, error); 3077 3078 total_bytes += bio_bytes; 3079 nr_bytes -= bio_bytes; 3080 3081 if (!nr_bytes) 3082 break; 3083 } 3084 3085 /* 3086 * completely done 3087 */ 3088 if (!req->bio) { 3089 /* 3090 * Reset counters so that the request stacking driver 3091 * can find how many bytes remain in the request 3092 * later. 3093 */ 3094 req->__data_len = 0; 3095 return false; 3096 } 3097 3098 req->__data_len -= total_bytes; 3099 3100 /* update sector only for requests with clear definition of sector */ 3101 if (!blk_rq_is_passthrough(req)) 3102 req->__sector += total_bytes >> 9; 3103 3104 /* mixed attributes always follow the first bio */ 3105 if (req->rq_flags & RQF_MIXED_MERGE) { 3106 req->cmd_flags &= ~REQ_FAILFAST_MASK; 3107 req->cmd_flags |= req->bio->bi_opf & REQ_FAILFAST_MASK; 3108 } 3109 3110 if (!(req->rq_flags & RQF_SPECIAL_PAYLOAD)) { 3111 /* 3112 * If total number of sectors is less than the first segment 3113 * size, something has gone terribly wrong. 3114 */ 3115 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) { 3116 blk_dump_rq_flags(req, "request botched"); 3117 req->__data_len = blk_rq_cur_bytes(req); 3118 } 3119 3120 /* recalculate the number of segments */ 3121 blk_recalc_rq_segments(req); 3122 } 3123 3124 return true; 3125 } 3126 EXPORT_SYMBOL_GPL(blk_update_request); 3127 3128 static bool blk_update_bidi_request(struct request *rq, blk_status_t error, 3129 unsigned int nr_bytes, 3130 unsigned int bidi_bytes) 3131 { 3132 if (blk_update_request(rq, error, nr_bytes)) 3133 return true; 3134 3135 /* Bidi request must be completed as a whole */ 3136 if (unlikely(blk_bidi_rq(rq)) && 3137 blk_update_request(rq->next_rq, error, bidi_bytes)) 3138 return true; 3139 3140 if (blk_queue_add_random(rq->q)) 3141 add_disk_randomness(rq->rq_disk); 3142 3143 return false; 3144 } 3145 3146 /** 3147 * blk_unprep_request - unprepare a request 3148 * @req: the request 3149 * 3150 * This function makes a request ready for complete resubmission (or 3151 * completion). It happens only after all error handling is complete, 3152 * so represents the appropriate moment to deallocate any resources 3153 * that were allocated to the request in the prep_rq_fn. The queue 3154 * lock is held when calling this. 3155 */ 3156 void blk_unprep_request(struct request *req) 3157 { 3158 struct request_queue *q = req->q; 3159 3160 req->rq_flags &= ~RQF_DONTPREP; 3161 if (q->unprep_rq_fn) 3162 q->unprep_rq_fn(q, req); 3163 } 3164 EXPORT_SYMBOL_GPL(blk_unprep_request); 3165 3166 void blk_finish_request(struct request *req, blk_status_t error) 3167 { 3168 struct request_queue *q = req->q; 3169 u64 now = ktime_get_ns(); 3170 3171 lockdep_assert_held(req->q->queue_lock); 3172 WARN_ON_ONCE(q->mq_ops); 3173 3174 if (req->rq_flags & RQF_STATS) 3175 blk_stat_add(req, now); 3176 3177 if (req->rq_flags & RQF_QUEUED) 3178 blk_queue_end_tag(q, req); 3179 3180 BUG_ON(blk_queued_rq(req)); 3181 3182 if (unlikely(laptop_mode) && !blk_rq_is_passthrough(req)) 3183 laptop_io_completion(req->q->backing_dev_info); 3184 3185 blk_delete_timer(req); 3186 3187 if (req->rq_flags & RQF_DONTPREP) 3188 blk_unprep_request(req); 3189 3190 blk_account_io_done(req, now); 3191 3192 if (req->end_io) { 3193 rq_qos_done(q, req); 3194 req->end_io(req, error); 3195 } else { 3196 if (blk_bidi_rq(req)) 3197 __blk_put_request(req->next_rq->q, req->next_rq); 3198 3199 __blk_put_request(q, req); 3200 } 3201 } 3202 EXPORT_SYMBOL(blk_finish_request); 3203 3204 /** 3205 * blk_end_bidi_request - Complete a bidi request 3206 * @rq: the request to complete 3207 * @error: block status code 3208 * @nr_bytes: number of bytes to complete @rq 3209 * @bidi_bytes: number of bytes to complete @rq->next_rq 3210 * 3211 * Description: 3212 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq. 3213 * Drivers that supports bidi can safely call this member for any 3214 * type of request, bidi or uni. In the later case @bidi_bytes is 3215 * just ignored. 3216 * 3217 * Return: 3218 * %false - we are done with this request 3219 * %true - still buffers pending for this request 3220 **/ 3221 static bool blk_end_bidi_request(struct request *rq, blk_status_t error, 3222 unsigned int nr_bytes, unsigned int bidi_bytes) 3223 { 3224 struct request_queue *q = rq->q; 3225 unsigned long flags; 3226 3227 WARN_ON_ONCE(q->mq_ops); 3228 3229 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes)) 3230 return true; 3231 3232 spin_lock_irqsave(q->queue_lock, flags); 3233 blk_finish_request(rq, error); 3234 spin_unlock_irqrestore(q->queue_lock, flags); 3235 3236 return false; 3237 } 3238 3239 /** 3240 * __blk_end_bidi_request - Complete a bidi request with queue lock held 3241 * @rq: the request to complete 3242 * @error: block status code 3243 * @nr_bytes: number of bytes to complete @rq 3244 * @bidi_bytes: number of bytes to complete @rq->next_rq 3245 * 3246 * Description: 3247 * Identical to blk_end_bidi_request() except that queue lock is 3248 * assumed to be locked on entry and remains so on return. 3249 * 3250 * Return: 3251 * %false - we are done with this request 3252 * %true - still buffers pending for this request 3253 **/ 3254 static bool __blk_end_bidi_request(struct request *rq, blk_status_t error, 3255 unsigned int nr_bytes, unsigned int bidi_bytes) 3256 { 3257 lockdep_assert_held(rq->q->queue_lock); 3258 WARN_ON_ONCE(rq->q->mq_ops); 3259 3260 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes)) 3261 return true; 3262 3263 blk_finish_request(rq, error); 3264 3265 return false; 3266 } 3267 3268 /** 3269 * blk_end_request - Helper function for drivers to complete the request. 3270 * @rq: the request being processed 3271 * @error: block status code 3272 * @nr_bytes: number of bytes to complete 3273 * 3274 * Description: 3275 * Ends I/O on a number of bytes attached to @rq. 3276 * If @rq has leftover, sets it up for the next range of segments. 3277 * 3278 * Return: 3279 * %false - we are done with this request 3280 * %true - still buffers pending for this request 3281 **/ 3282 bool blk_end_request(struct request *rq, blk_status_t error, 3283 unsigned int nr_bytes) 3284 { 3285 WARN_ON_ONCE(rq->q->mq_ops); 3286 return blk_end_bidi_request(rq, error, nr_bytes, 0); 3287 } 3288 EXPORT_SYMBOL(blk_end_request); 3289 3290 /** 3291 * blk_end_request_all - Helper function for drives to finish the request. 3292 * @rq: the request to finish 3293 * @error: block status code 3294 * 3295 * Description: 3296 * Completely finish @rq. 3297 */ 3298 void blk_end_request_all(struct request *rq, blk_status_t error) 3299 { 3300 bool pending; 3301 unsigned int bidi_bytes = 0; 3302 3303 if (unlikely(blk_bidi_rq(rq))) 3304 bidi_bytes = blk_rq_bytes(rq->next_rq); 3305 3306 pending = blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes); 3307 BUG_ON(pending); 3308 } 3309 EXPORT_SYMBOL(blk_end_request_all); 3310 3311 /** 3312 * __blk_end_request - Helper function for drivers to complete the request. 3313 * @rq: the request being processed 3314 * @error: block status code 3315 * @nr_bytes: number of bytes to complete 3316 * 3317 * Description: 3318 * Must be called with queue lock held unlike blk_end_request(). 3319 * 3320 * Return: 3321 * %false - we are done with this request 3322 * %true - still buffers pending for this request 3323 **/ 3324 bool __blk_end_request(struct request *rq, blk_status_t error, 3325 unsigned int nr_bytes) 3326 { 3327 lockdep_assert_held(rq->q->queue_lock); 3328 WARN_ON_ONCE(rq->q->mq_ops); 3329 3330 return __blk_end_bidi_request(rq, error, nr_bytes, 0); 3331 } 3332 EXPORT_SYMBOL(__blk_end_request); 3333 3334 /** 3335 * __blk_end_request_all - Helper function for drives to finish the request. 3336 * @rq: the request to finish 3337 * @error: block status code 3338 * 3339 * Description: 3340 * Completely finish @rq. Must be called with queue lock held. 3341 */ 3342 void __blk_end_request_all(struct request *rq, blk_status_t error) 3343 { 3344 bool pending; 3345 unsigned int bidi_bytes = 0; 3346 3347 lockdep_assert_held(rq->q->queue_lock); 3348 WARN_ON_ONCE(rq->q->mq_ops); 3349 3350 if (unlikely(blk_bidi_rq(rq))) 3351 bidi_bytes = blk_rq_bytes(rq->next_rq); 3352 3353 pending = __blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes); 3354 BUG_ON(pending); 3355 } 3356 EXPORT_SYMBOL(__blk_end_request_all); 3357 3358 /** 3359 * __blk_end_request_cur - Helper function to finish the current request chunk. 3360 * @rq: the request to finish the current chunk for 3361 * @error: block status code 3362 * 3363 * Description: 3364 * Complete the current consecutively mapped chunk from @rq. Must 3365 * be called with queue lock held. 3366 * 3367 * Return: 3368 * %false - we are done with this request 3369 * %true - still buffers pending for this request 3370 */ 3371 bool __blk_end_request_cur(struct request *rq, blk_status_t error) 3372 { 3373 return __blk_end_request(rq, error, blk_rq_cur_bytes(rq)); 3374 } 3375 EXPORT_SYMBOL(__blk_end_request_cur); 3376 3377 void blk_rq_bio_prep(struct request_queue *q, struct request *rq, 3378 struct bio *bio) 3379 { 3380 if (bio_has_data(bio)) 3381 rq->nr_phys_segments = bio_phys_segments(q, bio); 3382 else if (bio_op(bio) == REQ_OP_DISCARD) 3383 rq->nr_phys_segments = 1; 3384 3385 rq->__data_len = bio->bi_iter.bi_size; 3386 rq->bio = rq->biotail = bio; 3387 3388 if (bio->bi_disk) 3389 rq->rq_disk = bio->bi_disk; 3390 } 3391 3392 #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE 3393 /** 3394 * rq_flush_dcache_pages - Helper function to flush all pages in a request 3395 * @rq: the request to be flushed 3396 * 3397 * Description: 3398 * Flush all pages in @rq. 3399 */ 3400 void rq_flush_dcache_pages(struct request *rq) 3401 { 3402 struct req_iterator iter; 3403 struct bio_vec bvec; 3404 3405 rq_for_each_segment(bvec, rq, iter) 3406 flush_dcache_page(bvec.bv_page); 3407 } 3408 EXPORT_SYMBOL_GPL(rq_flush_dcache_pages); 3409 #endif 3410 3411 /** 3412 * blk_lld_busy - Check if underlying low-level drivers of a device are busy 3413 * @q : the queue of the device being checked 3414 * 3415 * Description: 3416 * Check if underlying low-level drivers of a device are busy. 3417 * If the drivers want to export their busy state, they must set own 3418 * exporting function using blk_queue_lld_busy() first. 3419 * 3420 * Basically, this function is used only by request stacking drivers 3421 * to stop dispatching requests to underlying devices when underlying 3422 * devices are busy. This behavior helps more I/O merging on the queue 3423 * of the request stacking driver and prevents I/O throughput regression 3424 * on burst I/O load. 3425 * 3426 * Return: 3427 * 0 - Not busy (The request stacking driver should dispatch request) 3428 * 1 - Busy (The request stacking driver should stop dispatching request) 3429 */ 3430 int blk_lld_busy(struct request_queue *q) 3431 { 3432 if (q->lld_busy_fn) 3433 return q->lld_busy_fn(q); 3434 3435 return 0; 3436 } 3437 EXPORT_SYMBOL_GPL(blk_lld_busy); 3438 3439 /** 3440 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request 3441 * @rq: the clone request to be cleaned up 3442 * 3443 * Description: 3444 * Free all bios in @rq for a cloned request. 3445 */ 3446 void blk_rq_unprep_clone(struct request *rq) 3447 { 3448 struct bio *bio; 3449 3450 while ((bio = rq->bio) != NULL) { 3451 rq->bio = bio->bi_next; 3452 3453 bio_put(bio); 3454 } 3455 } 3456 EXPORT_SYMBOL_GPL(blk_rq_unprep_clone); 3457 3458 /* 3459 * Copy attributes of the original request to the clone request. 3460 * The actual data parts (e.g. ->cmd, ->sense) are not copied. 3461 */ 3462 static void __blk_rq_prep_clone(struct request *dst, struct request *src) 3463 { 3464 dst->cpu = src->cpu; 3465 dst->__sector = blk_rq_pos(src); 3466 dst->__data_len = blk_rq_bytes(src); 3467 if (src->rq_flags & RQF_SPECIAL_PAYLOAD) { 3468 dst->rq_flags |= RQF_SPECIAL_PAYLOAD; 3469 dst->special_vec = src->special_vec; 3470 } 3471 dst->nr_phys_segments = src->nr_phys_segments; 3472 dst->ioprio = src->ioprio; 3473 dst->extra_len = src->extra_len; 3474 } 3475 3476 /** 3477 * blk_rq_prep_clone - Helper function to setup clone request 3478 * @rq: the request to be setup 3479 * @rq_src: original request to be cloned 3480 * @bs: bio_set that bios for clone are allocated from 3481 * @gfp_mask: memory allocation mask for bio 3482 * @bio_ctr: setup function to be called for each clone bio. 3483 * Returns %0 for success, non %0 for failure. 3484 * @data: private data to be passed to @bio_ctr 3485 * 3486 * Description: 3487 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq. 3488 * The actual data parts of @rq_src (e.g. ->cmd, ->sense) 3489 * are not copied, and copying such parts is the caller's responsibility. 3490 * Also, pages which the original bios are pointing to are not copied 3491 * and the cloned bios just point same pages. 3492 * So cloned bios must be completed before original bios, which means 3493 * the caller must complete @rq before @rq_src. 3494 */ 3495 int blk_rq_prep_clone(struct request *rq, struct request *rq_src, 3496 struct bio_set *bs, gfp_t gfp_mask, 3497 int (*bio_ctr)(struct bio *, struct bio *, void *), 3498 void *data) 3499 { 3500 struct bio *bio, *bio_src; 3501 3502 if (!bs) 3503 bs = &fs_bio_set; 3504 3505 __rq_for_each_bio(bio_src, rq_src) { 3506 bio = bio_clone_fast(bio_src, gfp_mask, bs); 3507 if (!bio) 3508 goto free_and_out; 3509 3510 if (bio_ctr && bio_ctr(bio, bio_src, data)) 3511 goto free_and_out; 3512 3513 if (rq->bio) { 3514 rq->biotail->bi_next = bio; 3515 rq->biotail = bio; 3516 } else 3517 rq->bio = rq->biotail = bio; 3518 } 3519 3520 __blk_rq_prep_clone(rq, rq_src); 3521 3522 return 0; 3523 3524 free_and_out: 3525 if (bio) 3526 bio_put(bio); 3527 blk_rq_unprep_clone(rq); 3528 3529 return -ENOMEM; 3530 } 3531 EXPORT_SYMBOL_GPL(blk_rq_prep_clone); 3532 3533 int kblockd_schedule_work(struct work_struct *work) 3534 { 3535 return queue_work(kblockd_workqueue, work); 3536 } 3537 EXPORT_SYMBOL(kblockd_schedule_work); 3538 3539 int kblockd_schedule_work_on(int cpu, struct work_struct *work) 3540 { 3541 return queue_work_on(cpu, kblockd_workqueue, work); 3542 } 3543 EXPORT_SYMBOL(kblockd_schedule_work_on); 3544 3545 int kblockd_mod_delayed_work_on(int cpu, struct delayed_work *dwork, 3546 unsigned long delay) 3547 { 3548 return mod_delayed_work_on(cpu, kblockd_workqueue, dwork, delay); 3549 } 3550 EXPORT_SYMBOL(kblockd_mod_delayed_work_on); 3551 3552 /** 3553 * blk_start_plug - initialize blk_plug and track it inside the task_struct 3554 * @plug: The &struct blk_plug that needs to be initialized 3555 * 3556 * Description: 3557 * Tracking blk_plug inside the task_struct will help with auto-flushing the 3558 * pending I/O should the task end up blocking between blk_start_plug() and 3559 * blk_finish_plug(). This is important from a performance perspective, but 3560 * also ensures that we don't deadlock. For instance, if the task is blocking 3561 * for a memory allocation, memory reclaim could end up wanting to free a 3562 * page belonging to that request that is currently residing in our private 3563 * plug. By flushing the pending I/O when the process goes to sleep, we avoid 3564 * this kind of deadlock. 3565 */ 3566 void blk_start_plug(struct blk_plug *plug) 3567 { 3568 struct task_struct *tsk = current; 3569 3570 /* 3571 * If this is a nested plug, don't actually assign it. 3572 */ 3573 if (tsk->plug) 3574 return; 3575 3576 INIT_LIST_HEAD(&plug->list); 3577 INIT_LIST_HEAD(&plug->mq_list); 3578 INIT_LIST_HEAD(&plug->cb_list); 3579 /* 3580 * Store ordering should not be needed here, since a potential 3581 * preempt will imply a full memory barrier 3582 */ 3583 tsk->plug = plug; 3584 } 3585 EXPORT_SYMBOL(blk_start_plug); 3586 3587 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b) 3588 { 3589 struct request *rqa = container_of(a, struct request, queuelist); 3590 struct request *rqb = container_of(b, struct request, queuelist); 3591 3592 return !(rqa->q < rqb->q || 3593 (rqa->q == rqb->q && blk_rq_pos(rqa) < blk_rq_pos(rqb))); 3594 } 3595 3596 /* 3597 * If 'from_schedule' is true, then postpone the dispatch of requests 3598 * until a safe kblockd context. We due this to avoid accidental big 3599 * additional stack usage in driver dispatch, in places where the originally 3600 * plugger did not intend it. 3601 */ 3602 static void queue_unplugged(struct request_queue *q, unsigned int depth, 3603 bool from_schedule) 3604 __releases(q->queue_lock) 3605 { 3606 lockdep_assert_held(q->queue_lock); 3607 3608 trace_block_unplug(q, depth, !from_schedule); 3609 3610 if (from_schedule) 3611 blk_run_queue_async(q); 3612 else 3613 __blk_run_queue(q); 3614 spin_unlock_irq(q->queue_lock); 3615 } 3616 3617 static void flush_plug_callbacks(struct blk_plug *plug, bool from_schedule) 3618 { 3619 LIST_HEAD(callbacks); 3620 3621 while (!list_empty(&plug->cb_list)) { 3622 list_splice_init(&plug->cb_list, &callbacks); 3623 3624 while (!list_empty(&callbacks)) { 3625 struct blk_plug_cb *cb = list_first_entry(&callbacks, 3626 struct blk_plug_cb, 3627 list); 3628 list_del(&cb->list); 3629 cb->callback(cb, from_schedule); 3630 } 3631 } 3632 } 3633 3634 struct blk_plug_cb *blk_check_plugged(blk_plug_cb_fn unplug, void *data, 3635 int size) 3636 { 3637 struct blk_plug *plug = current->plug; 3638 struct blk_plug_cb *cb; 3639 3640 if (!plug) 3641 return NULL; 3642 3643 list_for_each_entry(cb, &plug->cb_list, list) 3644 if (cb->callback == unplug && cb->data == data) 3645 return cb; 3646 3647 /* Not currently on the callback list */ 3648 BUG_ON(size < sizeof(*cb)); 3649 cb = kzalloc(size, GFP_ATOMIC); 3650 if (cb) { 3651 cb->data = data; 3652 cb->callback = unplug; 3653 list_add(&cb->list, &plug->cb_list); 3654 } 3655 return cb; 3656 } 3657 EXPORT_SYMBOL(blk_check_plugged); 3658 3659 void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule) 3660 { 3661 struct request_queue *q; 3662 struct request *rq; 3663 LIST_HEAD(list); 3664 unsigned int depth; 3665 3666 flush_plug_callbacks(plug, from_schedule); 3667 3668 if (!list_empty(&plug->mq_list)) 3669 blk_mq_flush_plug_list(plug, from_schedule); 3670 3671 if (list_empty(&plug->list)) 3672 return; 3673 3674 list_splice_init(&plug->list, &list); 3675 3676 list_sort(NULL, &list, plug_rq_cmp); 3677 3678 q = NULL; 3679 depth = 0; 3680 3681 while (!list_empty(&list)) { 3682 rq = list_entry_rq(list.next); 3683 list_del_init(&rq->queuelist); 3684 BUG_ON(!rq->q); 3685 if (rq->q != q) { 3686 /* 3687 * This drops the queue lock 3688 */ 3689 if (q) 3690 queue_unplugged(q, depth, from_schedule); 3691 q = rq->q; 3692 depth = 0; 3693 spin_lock_irq(q->queue_lock); 3694 } 3695 3696 /* 3697 * Short-circuit if @q is dead 3698 */ 3699 if (unlikely(blk_queue_dying(q))) { 3700 __blk_end_request_all(rq, BLK_STS_IOERR); 3701 continue; 3702 } 3703 3704 /* 3705 * rq is already accounted, so use raw insert 3706 */ 3707 if (op_is_flush(rq->cmd_flags)) 3708 __elv_add_request(q, rq, ELEVATOR_INSERT_FLUSH); 3709 else 3710 __elv_add_request(q, rq, ELEVATOR_INSERT_SORT_MERGE); 3711 3712 depth++; 3713 } 3714 3715 /* 3716 * This drops the queue lock 3717 */ 3718 if (q) 3719 queue_unplugged(q, depth, from_schedule); 3720 } 3721 3722 void blk_finish_plug(struct blk_plug *plug) 3723 { 3724 if (plug != current->plug) 3725 return; 3726 blk_flush_plug_list(plug, false); 3727 3728 current->plug = NULL; 3729 } 3730 EXPORT_SYMBOL(blk_finish_plug); 3731 3732 int __init blk_dev_init(void) 3733 { 3734 BUILD_BUG_ON(REQ_OP_LAST >= (1 << REQ_OP_BITS)); 3735 BUILD_BUG_ON(REQ_OP_BITS + REQ_FLAG_BITS > 8 * 3736 FIELD_SIZEOF(struct request, cmd_flags)); 3737 BUILD_BUG_ON(REQ_OP_BITS + REQ_FLAG_BITS > 8 * 3738 FIELD_SIZEOF(struct bio, bi_opf)); 3739 3740 /* used for unplugging and affects IO latency/throughput - HIGHPRI */ 3741 kblockd_workqueue = alloc_workqueue("kblockd", 3742 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0); 3743 if (!kblockd_workqueue) 3744 panic("Failed to create kblockd\n"); 3745 3746 request_cachep = kmem_cache_create("blkdev_requests", 3747 sizeof(struct request), 0, SLAB_PANIC, NULL); 3748 3749 blk_requestq_cachep = kmem_cache_create("request_queue", 3750 sizeof(struct request_queue), 0, SLAB_PANIC, NULL); 3751 3752 #ifdef CONFIG_DEBUG_FS 3753 blk_debugfs_root = debugfs_create_dir("block", NULL); 3754 #endif 3755 3756 return 0; 3757 } 3758