Loading...
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/highmem.h>
20#include <linux/mm.h>
21#include <linux/kernel_stat.h>
22#include <linux/string.h>
23#include <linux/init.h>
24#include <linux/completion.h>
25#include <linux/slab.h>
26#include <linux/swap.h>
27#include <linux/writeback.h>
28#include <linux/task_io_accounting_ops.h>
29#include <linux/fault-inject.h>
30#include <linux/list_sort.h>
31#include <linux/delay.h>
32#include <linux/ratelimit.h>
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/block.h>
36
37#include "blk.h"
38#include "blk-cgroup.h"
39
40EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_remap);
41EXPORT_TRACEPOINT_SYMBOL_GPL(block_rq_remap);
42EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_complete);
43
44DEFINE_IDA(blk_queue_ida);
45
46/*
47 * For the allocated request tables
48 */
49static struct kmem_cache *request_cachep;
50
51/*
52 * For queue allocation
53 */
54struct kmem_cache *blk_requestq_cachep;
55
56/*
57 * Controlling structure to kblockd
58 */
59static struct workqueue_struct *kblockd_workqueue;
60
61static void drive_stat_acct(struct request *rq, int new_io)
62{
63 struct hd_struct *part;
64 int rw = rq_data_dir(rq);
65 int cpu;
66
67 if (!blk_do_io_stat(rq))
68 return;
69
70 cpu = part_stat_lock();
71
72 if (!new_io) {
73 part = rq->part;
74 part_stat_inc(cpu, part, merges[rw]);
75 } else {
76 part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
77 if (!hd_struct_try_get(part)) {
78 /*
79 * The partition is already being removed,
80 * the request will be accounted on the disk only
81 *
82 * We take a reference on disk->part0 although that
83 * partition will never be deleted, so we can treat
84 * it as any other partition.
85 */
86 part = &rq->rq_disk->part0;
87 hd_struct_get(part);
88 }
89 part_round_stats(cpu, part);
90 part_inc_in_flight(part, rw);
91 rq->part = part;
92 }
93
94 part_stat_unlock();
95}
96
97void blk_queue_congestion_threshold(struct request_queue *q)
98{
99 int nr;
100
101 nr = q->nr_requests - (q->nr_requests / 8) + 1;
102 if (nr > q->nr_requests)
103 nr = q->nr_requests;
104 q->nr_congestion_on = nr;
105
106 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
107 if (nr < 1)
108 nr = 1;
109 q->nr_congestion_off = nr;
110}
111
112/**
113 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
114 * @bdev: device
115 *
116 * Locates the passed device's request queue and returns the address of its
117 * backing_dev_info
118 *
119 * Will return NULL if the request queue cannot be located.
120 */
121struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
122{
123 struct backing_dev_info *ret = NULL;
124 struct request_queue *q = bdev_get_queue(bdev);
125
126 if (q)
127 ret = &q->backing_dev_info;
128 return ret;
129}
130EXPORT_SYMBOL(blk_get_backing_dev_info);
131
132void blk_rq_init(struct request_queue *q, struct request *rq)
133{
134 memset(rq, 0, sizeof(*rq));
135
136 INIT_LIST_HEAD(&rq->queuelist);
137 INIT_LIST_HEAD(&rq->timeout_list);
138 rq->cpu = -1;
139 rq->q = q;
140 rq->__sector = (sector_t) -1;
141 INIT_HLIST_NODE(&rq->hash);
142 RB_CLEAR_NODE(&rq->rb_node);
143 rq->cmd = rq->__cmd;
144 rq->cmd_len = BLK_MAX_CDB;
145 rq->tag = -1;
146 rq->ref_count = 1;
147 rq->start_time = jiffies;
148 set_start_time_ns(rq);
149 rq->part = NULL;
150}
151EXPORT_SYMBOL(blk_rq_init);
152
153static void req_bio_endio(struct request *rq, struct bio *bio,
154 unsigned int nbytes, int error)
155{
156 if (error)
157 clear_bit(BIO_UPTODATE, &bio->bi_flags);
158 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
159 error = -EIO;
160
161 if (unlikely(nbytes > bio->bi_size)) {
162 printk(KERN_ERR "%s: want %u bytes done, %u left\n",
163 __func__, nbytes, bio->bi_size);
164 nbytes = bio->bi_size;
165 }
166
167 if (unlikely(rq->cmd_flags & REQ_QUIET))
168 set_bit(BIO_QUIET, &bio->bi_flags);
169
170 bio->bi_size -= nbytes;
171 bio->bi_sector += (nbytes >> 9);
172
173 if (bio_integrity(bio))
174 bio_integrity_advance(bio, nbytes);
175
176 /* don't actually finish bio if it's part of flush sequence */
177 if (bio->bi_size == 0 && !(rq->cmd_flags & REQ_FLUSH_SEQ))
178 bio_endio(bio, error);
179}
180
181void blk_dump_rq_flags(struct request *rq, char *msg)
182{
183 int bit;
184
185 printk(KERN_INFO "%s: dev %s: type=%x, flags=%x\n", msg,
186 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
187 rq->cmd_flags);
188
189 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
190 (unsigned long long)blk_rq_pos(rq),
191 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
192 printk(KERN_INFO " bio %p, biotail %p, buffer %p, len %u\n",
193 rq->bio, rq->biotail, rq->buffer, blk_rq_bytes(rq));
194
195 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
196 printk(KERN_INFO " cdb: ");
197 for (bit = 0; bit < BLK_MAX_CDB; bit++)
198 printk("%02x ", rq->cmd[bit]);
199 printk("\n");
200 }
201}
202EXPORT_SYMBOL(blk_dump_rq_flags);
203
204static void blk_delay_work(struct work_struct *work)
205{
206 struct request_queue *q;
207
208 q = container_of(work, struct request_queue, delay_work.work);
209 spin_lock_irq(q->queue_lock);
210 __blk_run_queue(q);
211 spin_unlock_irq(q->queue_lock);
212}
213
214/**
215 * blk_delay_queue - restart queueing after defined interval
216 * @q: The &struct request_queue in question
217 * @msecs: Delay in msecs
218 *
219 * Description:
220 * Sometimes queueing needs to be postponed for a little while, to allow
221 * resources to come back. This function will make sure that queueing is
222 * restarted around the specified time.
223 */
224void blk_delay_queue(struct request_queue *q, unsigned long msecs)
225{
226 queue_delayed_work(kblockd_workqueue, &q->delay_work,
227 msecs_to_jiffies(msecs));
228}
229EXPORT_SYMBOL(blk_delay_queue);
230
231/**
232 * blk_start_queue - restart a previously stopped queue
233 * @q: The &struct request_queue in question
234 *
235 * Description:
236 * blk_start_queue() will clear the stop flag on the queue, and call
237 * the request_fn for the queue if it was in a stopped state when
238 * entered. Also see blk_stop_queue(). Queue lock must be held.
239 **/
240void blk_start_queue(struct request_queue *q)
241{
242 WARN_ON(!irqs_disabled());
243
244 queue_flag_clear(QUEUE_FLAG_STOPPED, q);
245 __blk_run_queue(q);
246}
247EXPORT_SYMBOL(blk_start_queue);
248
249/**
250 * blk_stop_queue - stop a queue
251 * @q: The &struct request_queue in question
252 *
253 * Description:
254 * The Linux block layer assumes that a block driver will consume all
255 * entries on the request queue when the request_fn strategy is called.
256 * Often this will not happen, because of hardware limitations (queue
257 * depth settings). If a device driver gets a 'queue full' response,
258 * or if it simply chooses not to queue more I/O at one point, it can
259 * call this function to prevent the request_fn from being called until
260 * the driver has signalled it's ready to go again. This happens by calling
261 * blk_start_queue() to restart queue operations. Queue lock must be held.
262 **/
263void blk_stop_queue(struct request_queue *q)
264{
265 __cancel_delayed_work(&q->delay_work);
266 queue_flag_set(QUEUE_FLAG_STOPPED, q);
267}
268EXPORT_SYMBOL(blk_stop_queue);
269
270/**
271 * blk_sync_queue - cancel any pending callbacks on a queue
272 * @q: the queue
273 *
274 * Description:
275 * The block layer may perform asynchronous callback activity
276 * on a queue, such as calling the unplug function after a timeout.
277 * A block device may call blk_sync_queue to ensure that any
278 * such activity is cancelled, thus allowing it to release resources
279 * that the callbacks might use. The caller must already have made sure
280 * that its ->make_request_fn will not re-add plugging prior to calling
281 * this function.
282 *
283 * This function does not cancel any asynchronous activity arising
284 * out of elevator or throttling code. That would require elevaotor_exit()
285 * and blkcg_exit_queue() to be called with queue lock initialized.
286 *
287 */
288void blk_sync_queue(struct request_queue *q)
289{
290 del_timer_sync(&q->timeout);
291 cancel_delayed_work_sync(&q->delay_work);
292}
293EXPORT_SYMBOL(blk_sync_queue);
294
295/**
296 * __blk_run_queue - run a single device queue
297 * @q: The queue to run
298 *
299 * Description:
300 * See @blk_run_queue. This variant must be called with the queue lock
301 * held and interrupts disabled.
302 */
303void __blk_run_queue(struct request_queue *q)
304{
305 if (unlikely(blk_queue_stopped(q)))
306 return;
307
308 q->request_fn(q);
309}
310EXPORT_SYMBOL(__blk_run_queue);
311
312/**
313 * blk_run_queue_async - run a single device queue in workqueue context
314 * @q: The queue to run
315 *
316 * Description:
317 * Tells kblockd to perform the equivalent of @blk_run_queue on behalf
318 * of us.
319 */
320void blk_run_queue_async(struct request_queue *q)
321{
322 if (likely(!blk_queue_stopped(q))) {
323 __cancel_delayed_work(&q->delay_work);
324 queue_delayed_work(kblockd_workqueue, &q->delay_work, 0);
325 }
326}
327EXPORT_SYMBOL(blk_run_queue_async);
328
329/**
330 * blk_run_queue - run a single device queue
331 * @q: The queue to run
332 *
333 * Description:
334 * Invoke request handling on this queue, if it has pending work to do.
335 * May be used to restart queueing when a request has completed.
336 */
337void blk_run_queue(struct request_queue *q)
338{
339 unsigned long flags;
340
341 spin_lock_irqsave(q->queue_lock, flags);
342 __blk_run_queue(q);
343 spin_unlock_irqrestore(q->queue_lock, flags);
344}
345EXPORT_SYMBOL(blk_run_queue);
346
347void blk_put_queue(struct request_queue *q)
348{
349 kobject_put(&q->kobj);
350}
351EXPORT_SYMBOL(blk_put_queue);
352
353/**
354 * blk_drain_queue - drain requests from request_queue
355 * @q: queue to drain
356 * @drain_all: whether to drain all requests or only the ones w/ ELVPRIV
357 *
358 * Drain requests from @q. If @drain_all is set, all requests are drained.
359 * If not, only ELVPRIV requests are drained. The caller is responsible
360 * for ensuring that no new requests which need to be drained are queued.
361 */
362void blk_drain_queue(struct request_queue *q, bool drain_all)
363{
364 int i;
365
366 while (true) {
367 bool drain = false;
368
369 spin_lock_irq(q->queue_lock);
370
371 /*
372 * The caller might be trying to drain @q before its
373 * elevator is initialized.
374 */
375 if (q->elevator)
376 elv_drain_elevator(q);
377
378 blkcg_drain_queue(q);
379
380 /*
381 * This function might be called on a queue which failed
382 * driver init after queue creation or is not yet fully
383 * active yet. Some drivers (e.g. fd and loop) get unhappy
384 * in such cases. Kick queue iff dispatch queue has
385 * something on it and @q has request_fn set.
386 */
387 if (!list_empty(&q->queue_head) && q->request_fn)
388 __blk_run_queue(q);
389
390 drain |= q->rq.elvpriv;
391
392 /*
393 * Unfortunately, requests are queued at and tracked from
394 * multiple places and there's no single counter which can
395 * be drained. Check all the queues and counters.
396 */
397 if (drain_all) {
398 drain |= !list_empty(&q->queue_head);
399 for (i = 0; i < 2; i++) {
400 drain |= q->rq.count[i];
401 drain |= q->in_flight[i];
402 drain |= !list_empty(&q->flush_queue[i]);
403 }
404 }
405
406 spin_unlock_irq(q->queue_lock);
407
408 if (!drain)
409 break;
410 msleep(10);
411 }
412
413 /*
414 * With queue marked dead, any woken up waiter will fail the
415 * allocation path, so the wakeup chaining is lost and we're
416 * left with hung waiters. We need to wake up those waiters.
417 */
418 if (q->request_fn) {
419 spin_lock_irq(q->queue_lock);
420 for (i = 0; i < ARRAY_SIZE(q->rq.wait); i++)
421 wake_up_all(&q->rq.wait[i]);
422 spin_unlock_irq(q->queue_lock);
423 }
424}
425
426/**
427 * blk_queue_bypass_start - enter queue bypass mode
428 * @q: queue of interest
429 *
430 * In bypass mode, only the dispatch FIFO queue of @q is used. This
431 * function makes @q enter bypass mode and drains all requests which were
432 * throttled or issued before. On return, it's guaranteed that no request
433 * is being throttled or has ELVPRIV set and blk_queue_bypass() %true
434 * inside queue or RCU read lock.
435 */
436void blk_queue_bypass_start(struct request_queue *q)
437{
438 bool drain;
439
440 spin_lock_irq(q->queue_lock);
441 drain = !q->bypass_depth++;
442 queue_flag_set(QUEUE_FLAG_BYPASS, q);
443 spin_unlock_irq(q->queue_lock);
444
445 if (drain) {
446 blk_drain_queue(q, false);
447 /* ensure blk_queue_bypass() is %true inside RCU read lock */
448 synchronize_rcu();
449 }
450}
451EXPORT_SYMBOL_GPL(blk_queue_bypass_start);
452
453/**
454 * blk_queue_bypass_end - leave queue bypass mode
455 * @q: queue of interest
456 *
457 * Leave bypass mode and restore the normal queueing behavior.
458 */
459void blk_queue_bypass_end(struct request_queue *q)
460{
461 spin_lock_irq(q->queue_lock);
462 if (!--q->bypass_depth)
463 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
464 WARN_ON_ONCE(q->bypass_depth < 0);
465 spin_unlock_irq(q->queue_lock);
466}
467EXPORT_SYMBOL_GPL(blk_queue_bypass_end);
468
469/**
470 * blk_cleanup_queue - shutdown a request queue
471 * @q: request queue to shutdown
472 *
473 * Mark @q DEAD, drain all pending requests, destroy and put it. All
474 * future requests will be failed immediately with -ENODEV.
475 */
476void blk_cleanup_queue(struct request_queue *q)
477{
478 spinlock_t *lock = q->queue_lock;
479
480 /* mark @q DEAD, no new request or merges will be allowed afterwards */
481 mutex_lock(&q->sysfs_lock);
482 queue_flag_set_unlocked(QUEUE_FLAG_DEAD, q);
483 spin_lock_irq(lock);
484
485 /*
486 * Dead queue is permanently in bypass mode till released. Note
487 * that, unlike blk_queue_bypass_start(), we aren't performing
488 * synchronize_rcu() after entering bypass mode to avoid the delay
489 * as some drivers create and destroy a lot of queues while
490 * probing. This is still safe because blk_release_queue() will be
491 * called only after the queue refcnt drops to zero and nothing,
492 * RCU or not, would be traversing the queue by then.
493 */
494 q->bypass_depth++;
495 queue_flag_set(QUEUE_FLAG_BYPASS, q);
496
497 queue_flag_set(QUEUE_FLAG_NOMERGES, q);
498 queue_flag_set(QUEUE_FLAG_NOXMERGES, q);
499 queue_flag_set(QUEUE_FLAG_DEAD, q);
500 spin_unlock_irq(lock);
501 mutex_unlock(&q->sysfs_lock);
502
503 /* drain all requests queued before DEAD marking */
504 blk_drain_queue(q, true);
505
506 /* @q won't process any more request, flush async actions */
507 del_timer_sync(&q->backing_dev_info.laptop_mode_wb_timer);
508 blk_sync_queue(q);
509
510 spin_lock_irq(lock);
511 if (q->queue_lock != &q->__queue_lock)
512 q->queue_lock = &q->__queue_lock;
513 spin_unlock_irq(lock);
514
515 /* @q is and will stay empty, shutdown and put */
516 blk_put_queue(q);
517}
518EXPORT_SYMBOL(blk_cleanup_queue);
519
520static int blk_init_free_list(struct request_queue *q)
521{
522 struct request_list *rl = &q->rq;
523
524 if (unlikely(rl->rq_pool))
525 return 0;
526
527 rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
528 rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
529 rl->elvpriv = 0;
530 init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
531 init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
532
533 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
534 mempool_free_slab, request_cachep, q->node);
535
536 if (!rl->rq_pool)
537 return -ENOMEM;
538
539 return 0;
540}
541
542struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
543{
544 return blk_alloc_queue_node(gfp_mask, -1);
545}
546EXPORT_SYMBOL(blk_alloc_queue);
547
548struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
549{
550 struct request_queue *q;
551 int err;
552
553 q = kmem_cache_alloc_node(blk_requestq_cachep,
554 gfp_mask | __GFP_ZERO, node_id);
555 if (!q)
556 return NULL;
557
558 q->id = ida_simple_get(&blk_queue_ida, 0, 0, gfp_mask);
559 if (q->id < 0)
560 goto fail_q;
561
562 q->backing_dev_info.ra_pages =
563 (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
564 q->backing_dev_info.state = 0;
565 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
566 q->backing_dev_info.name = "block";
567 q->node = node_id;
568
569 err = bdi_init(&q->backing_dev_info);
570 if (err)
571 goto fail_id;
572
573 setup_timer(&q->backing_dev_info.laptop_mode_wb_timer,
574 laptop_mode_timer_fn, (unsigned long) q);
575 setup_timer(&q->timeout, blk_rq_timed_out_timer, (unsigned long) q);
576 INIT_LIST_HEAD(&q->queue_head);
577 INIT_LIST_HEAD(&q->timeout_list);
578 INIT_LIST_HEAD(&q->icq_list);
579#ifdef CONFIG_BLK_CGROUP
580 INIT_LIST_HEAD(&q->blkg_list);
581#endif
582 INIT_LIST_HEAD(&q->flush_queue[0]);
583 INIT_LIST_HEAD(&q->flush_queue[1]);
584 INIT_LIST_HEAD(&q->flush_data_in_flight);
585 INIT_DELAYED_WORK(&q->delay_work, blk_delay_work);
586
587 kobject_init(&q->kobj, &blk_queue_ktype);
588
589 mutex_init(&q->sysfs_lock);
590 spin_lock_init(&q->__queue_lock);
591
592 /*
593 * By default initialize queue_lock to internal lock and driver can
594 * override it later if need be.
595 */
596 q->queue_lock = &q->__queue_lock;
597
598 /*
599 * A queue starts its life with bypass turned on to avoid
600 * unnecessary bypass on/off overhead and nasty surprises during
601 * init. The initial bypass will be finished at the end of
602 * blk_init_allocated_queue().
603 */
604 q->bypass_depth = 1;
605 __set_bit(QUEUE_FLAG_BYPASS, &q->queue_flags);
606
607 if (blkcg_init_queue(q))
608 goto fail_id;
609
610 return q;
611
612fail_id:
613 ida_simple_remove(&blk_queue_ida, q->id);
614fail_q:
615 kmem_cache_free(blk_requestq_cachep, q);
616 return NULL;
617}
618EXPORT_SYMBOL(blk_alloc_queue_node);
619
620/**
621 * blk_init_queue - prepare a request queue for use with a block device
622 * @rfn: The function to be called to process requests that have been
623 * placed on the queue.
624 * @lock: Request queue spin lock
625 *
626 * Description:
627 * If a block device wishes to use the standard request handling procedures,
628 * which sorts requests and coalesces adjacent requests, then it must
629 * call blk_init_queue(). The function @rfn will be called when there
630 * are requests on the queue that need to be processed. If the device
631 * supports plugging, then @rfn may not be called immediately when requests
632 * are available on the queue, but may be called at some time later instead.
633 * Plugged queues are generally unplugged when a buffer belonging to one
634 * of the requests on the queue is needed, or due to memory pressure.
635 *
636 * @rfn is not required, or even expected, to remove all requests off the
637 * queue, but only as many as it can handle at a time. If it does leave
638 * requests on the queue, it is responsible for arranging that the requests
639 * get dealt with eventually.
640 *
641 * The queue spin lock must be held while manipulating the requests on the
642 * request queue; this lock will be taken also from interrupt context, so irq
643 * disabling is needed for it.
644 *
645 * Function returns a pointer to the initialized request queue, or %NULL if
646 * it didn't succeed.
647 *
648 * Note:
649 * blk_init_queue() must be paired with a blk_cleanup_queue() call
650 * when the block device is deactivated (such as at module unload).
651 **/
652
653struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
654{
655 return blk_init_queue_node(rfn, lock, -1);
656}
657EXPORT_SYMBOL(blk_init_queue);
658
659struct request_queue *
660blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
661{
662 struct request_queue *uninit_q, *q;
663
664 uninit_q = blk_alloc_queue_node(GFP_KERNEL, node_id);
665 if (!uninit_q)
666 return NULL;
667
668 q = blk_init_allocated_queue(uninit_q, rfn, lock);
669 if (!q)
670 blk_cleanup_queue(uninit_q);
671
672 return q;
673}
674EXPORT_SYMBOL(blk_init_queue_node);
675
676struct request_queue *
677blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
678 spinlock_t *lock)
679{
680 if (!q)
681 return NULL;
682
683 if (blk_init_free_list(q))
684 return NULL;
685
686 q->request_fn = rfn;
687 q->prep_rq_fn = NULL;
688 q->unprep_rq_fn = NULL;
689 q->queue_flags = QUEUE_FLAG_DEFAULT;
690
691 /* Override internal queue lock with supplied lock pointer */
692 if (lock)
693 q->queue_lock = lock;
694
695 /*
696 * This also sets hw/phys segments, boundary and size
697 */
698 blk_queue_make_request(q, blk_queue_bio);
699
700 q->sg_reserved_size = INT_MAX;
701
702 /* init elevator */
703 if (elevator_init(q, NULL))
704 return NULL;
705
706 blk_queue_congestion_threshold(q);
707
708 /* all done, end the initial bypass */
709 blk_queue_bypass_end(q);
710 return q;
711}
712EXPORT_SYMBOL(blk_init_allocated_queue);
713
714bool blk_get_queue(struct request_queue *q)
715{
716 if (likely(!blk_queue_dead(q))) {
717 __blk_get_queue(q);
718 return true;
719 }
720
721 return false;
722}
723EXPORT_SYMBOL(blk_get_queue);
724
725static inline void blk_free_request(struct request_queue *q, struct request *rq)
726{
727 if (rq->cmd_flags & REQ_ELVPRIV) {
728 elv_put_request(q, rq);
729 if (rq->elv.icq)
730 put_io_context(rq->elv.icq->ioc);
731 }
732
733 mempool_free(rq, q->rq.rq_pool);
734}
735
736/*
737 * ioc_batching returns true if the ioc is a valid batching request and
738 * should be given priority access to a request.
739 */
740static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
741{
742 if (!ioc)
743 return 0;
744
745 /*
746 * Make sure the process is able to allocate at least 1 request
747 * even if the batch times out, otherwise we could theoretically
748 * lose wakeups.
749 */
750 return ioc->nr_batch_requests == q->nr_batching ||
751 (ioc->nr_batch_requests > 0
752 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
753}
754
755/*
756 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
757 * will cause the process to be a "batcher" on all queues in the system. This
758 * is the behaviour we want though - once it gets a wakeup it should be given
759 * a nice run.
760 */
761static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
762{
763 if (!ioc || ioc_batching(q, ioc))
764 return;
765
766 ioc->nr_batch_requests = q->nr_batching;
767 ioc->last_waited = jiffies;
768}
769
770static void __freed_request(struct request_queue *q, int sync)
771{
772 struct request_list *rl = &q->rq;
773
774 if (rl->count[sync] < queue_congestion_off_threshold(q))
775 blk_clear_queue_congested(q, sync);
776
777 if (rl->count[sync] + 1 <= q->nr_requests) {
778 if (waitqueue_active(&rl->wait[sync]))
779 wake_up(&rl->wait[sync]);
780
781 blk_clear_queue_full(q, sync);
782 }
783}
784
785/*
786 * A request has just been released. Account for it, update the full and
787 * congestion status, wake up any waiters. Called under q->queue_lock.
788 */
789static void freed_request(struct request_queue *q, unsigned int flags)
790{
791 struct request_list *rl = &q->rq;
792 int sync = rw_is_sync(flags);
793
794 rl->count[sync]--;
795 if (flags & REQ_ELVPRIV)
796 rl->elvpriv--;
797
798 __freed_request(q, sync);
799
800 if (unlikely(rl->starved[sync ^ 1]))
801 __freed_request(q, sync ^ 1);
802}
803
804/*
805 * Determine if elevator data should be initialized when allocating the
806 * request associated with @bio.
807 */
808static bool blk_rq_should_init_elevator(struct bio *bio)
809{
810 if (!bio)
811 return true;
812
813 /*
814 * Flush requests do not use the elevator so skip initialization.
815 * This allows a request to share the flush and elevator data.
816 */
817 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA))
818 return false;
819
820 return true;
821}
822
823/**
824 * rq_ioc - determine io_context for request allocation
825 * @bio: request being allocated is for this bio (can be %NULL)
826 *
827 * Determine io_context to use for request allocation for @bio. May return
828 * %NULL if %current->io_context doesn't exist.
829 */
830static struct io_context *rq_ioc(struct bio *bio)
831{
832#ifdef CONFIG_BLK_CGROUP
833 if (bio && bio->bi_ioc)
834 return bio->bi_ioc;
835#endif
836 return current->io_context;
837}
838
839/**
840 * get_request - get a free request
841 * @q: request_queue to allocate request from
842 * @rw_flags: RW and SYNC flags
843 * @bio: bio to allocate request for (can be %NULL)
844 * @gfp_mask: allocation mask
845 *
846 * Get a free request from @q. This function may fail under memory
847 * pressure or if @q is dead.
848 *
849 * Must be callled with @q->queue_lock held and,
850 * Returns %NULL on failure, with @q->queue_lock held.
851 * Returns !%NULL on success, with @q->queue_lock *not held*.
852 */
853static struct request *get_request(struct request_queue *q, int rw_flags,
854 struct bio *bio, gfp_t gfp_mask)
855{
856 struct request *rq;
857 struct request_list *rl = &q->rq;
858 struct elevator_type *et;
859 struct io_context *ioc;
860 struct io_cq *icq = NULL;
861 const bool is_sync = rw_is_sync(rw_flags) != 0;
862 bool retried = false;
863 int may_queue;
864retry:
865 et = q->elevator->type;
866 ioc = rq_ioc(bio);
867
868 if (unlikely(blk_queue_dead(q)))
869 return NULL;
870
871 may_queue = elv_may_queue(q, rw_flags);
872 if (may_queue == ELV_MQUEUE_NO)
873 goto rq_starved;
874
875 if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) {
876 if (rl->count[is_sync]+1 >= q->nr_requests) {
877 /*
878 * We want ioc to record batching state. If it's
879 * not already there, creating a new one requires
880 * dropping queue_lock, which in turn requires
881 * retesting conditions to avoid queue hang.
882 */
883 if (!ioc && !retried) {
884 spin_unlock_irq(q->queue_lock);
885 create_io_context(gfp_mask, q->node);
886 spin_lock_irq(q->queue_lock);
887 retried = true;
888 goto retry;
889 }
890
891 /*
892 * The queue will fill after this allocation, so set
893 * it as full, and mark this process as "batching".
894 * This process will be allowed to complete a batch of
895 * requests, others will be blocked.
896 */
897 if (!blk_queue_full(q, is_sync)) {
898 ioc_set_batching(q, ioc);
899 blk_set_queue_full(q, is_sync);
900 } else {
901 if (may_queue != ELV_MQUEUE_MUST
902 && !ioc_batching(q, ioc)) {
903 /*
904 * The queue is full and the allocating
905 * process is not a "batcher", and not
906 * exempted by the IO scheduler
907 */
908 return NULL;
909 }
910 }
911 }
912 blk_set_queue_congested(q, is_sync);
913 }
914
915 /*
916 * Only allow batching queuers to allocate up to 50% over the defined
917 * limit of requests, otherwise we could have thousands of requests
918 * allocated with any setting of ->nr_requests
919 */
920 if (rl->count[is_sync] >= (3 * q->nr_requests / 2))
921 return NULL;
922
923 rl->count[is_sync]++;
924 rl->starved[is_sync] = 0;
925
926 /*
927 * Decide whether the new request will be managed by elevator. If
928 * so, mark @rw_flags and increment elvpriv. Non-zero elvpriv will
929 * prevent the current elevator from being destroyed until the new
930 * request is freed. This guarantees icq's won't be destroyed and
931 * makes creating new ones safe.
932 *
933 * Also, lookup icq while holding queue_lock. If it doesn't exist,
934 * it will be created after releasing queue_lock.
935 */
936 if (blk_rq_should_init_elevator(bio) && !blk_queue_bypass(q)) {
937 rw_flags |= REQ_ELVPRIV;
938 rl->elvpriv++;
939 if (et->icq_cache && ioc)
940 icq = ioc_lookup_icq(ioc, q);
941 }
942
943 if (blk_queue_io_stat(q))
944 rw_flags |= REQ_IO_STAT;
945 spin_unlock_irq(q->queue_lock);
946
947 /* allocate and init request */
948 rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
949 if (!rq)
950 goto fail_alloc;
951
952 blk_rq_init(q, rq);
953 rq->cmd_flags = rw_flags | REQ_ALLOCED;
954
955 /* init elvpriv */
956 if (rw_flags & REQ_ELVPRIV) {
957 if (unlikely(et->icq_cache && !icq)) {
958 create_io_context(gfp_mask, q->node);
959 ioc = rq_ioc(bio);
960 if (!ioc)
961 goto fail_elvpriv;
962
963 icq = ioc_create_icq(ioc, q, gfp_mask);
964 if (!icq)
965 goto fail_elvpriv;
966 }
967
968 rq->elv.icq = icq;
969 if (unlikely(elv_set_request(q, rq, bio, gfp_mask)))
970 goto fail_elvpriv;
971
972 /* @rq->elv.icq holds io_context until @rq is freed */
973 if (icq)
974 get_io_context(icq->ioc);
975 }
976out:
977 /*
978 * ioc may be NULL here, and ioc_batching will be false. That's
979 * OK, if the queue is under the request limit then requests need
980 * not count toward the nr_batch_requests limit. There will always
981 * be some limit enforced by BLK_BATCH_TIME.
982 */
983 if (ioc_batching(q, ioc))
984 ioc->nr_batch_requests--;
985
986 trace_block_getrq(q, bio, rw_flags & 1);
987 return rq;
988
989fail_elvpriv:
990 /*
991 * elvpriv init failed. ioc, icq and elvpriv aren't mempool backed
992 * and may fail indefinitely under memory pressure and thus
993 * shouldn't stall IO. Treat this request as !elvpriv. This will
994 * disturb iosched and blkcg but weird is bettern than dead.
995 */
996 printk_ratelimited(KERN_WARNING "%s: request aux data allocation failed, iosched may be disturbed\n",
997 dev_name(q->backing_dev_info.dev));
998
999 rq->cmd_flags &= ~REQ_ELVPRIV;
1000 rq->elv.icq = NULL;
1001
1002 spin_lock_irq(q->queue_lock);
1003 rl->elvpriv--;
1004 spin_unlock_irq(q->queue_lock);
1005 goto out;
1006
1007fail_alloc:
1008 /*
1009 * Allocation failed presumably due to memory. Undo anything we
1010 * might have messed up.
1011 *
1012 * Allocating task should really be put onto the front of the wait
1013 * queue, but this is pretty rare.
1014 */
1015 spin_lock_irq(q->queue_lock);
1016 freed_request(q, rw_flags);
1017
1018 /*
1019 * in the very unlikely event that allocation failed and no
1020 * requests for this direction was pending, mark us starved so that
1021 * freeing of a request in the other direction will notice
1022 * us. another possible fix would be to split the rq mempool into
1023 * READ and WRITE
1024 */
1025rq_starved:
1026 if (unlikely(rl->count[is_sync] == 0))
1027 rl->starved[is_sync] = 1;
1028 return NULL;
1029}
1030
1031/**
1032 * get_request_wait - get a free request with retry
1033 * @q: request_queue to allocate request from
1034 * @rw_flags: RW and SYNC flags
1035 * @bio: bio to allocate request for (can be %NULL)
1036 *
1037 * Get a free request from @q. This function keeps retrying under memory
1038 * pressure and fails iff @q is dead.
1039 *
1040 * Must be callled with @q->queue_lock held and,
1041 * Returns %NULL on failure, with @q->queue_lock held.
1042 * Returns !%NULL on success, with @q->queue_lock *not held*.
1043 */
1044static struct request *get_request_wait(struct request_queue *q, int rw_flags,
1045 struct bio *bio)
1046{
1047 const bool is_sync = rw_is_sync(rw_flags) != 0;
1048 struct request *rq;
1049
1050 rq = get_request(q, rw_flags, bio, GFP_NOIO);
1051 while (!rq) {
1052 DEFINE_WAIT(wait);
1053 struct request_list *rl = &q->rq;
1054
1055 if (unlikely(blk_queue_dead(q)))
1056 return NULL;
1057
1058 prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
1059 TASK_UNINTERRUPTIBLE);
1060
1061 trace_block_sleeprq(q, bio, rw_flags & 1);
1062
1063 spin_unlock_irq(q->queue_lock);
1064 io_schedule();
1065
1066 /*
1067 * After sleeping, we become a "batching" process and
1068 * will be able to allocate at least one request, and
1069 * up to a big batch of them for a small period time.
1070 * See ioc_batching, ioc_set_batching
1071 */
1072 create_io_context(GFP_NOIO, q->node);
1073 ioc_set_batching(q, current->io_context);
1074
1075 spin_lock_irq(q->queue_lock);
1076 finish_wait(&rl->wait[is_sync], &wait);
1077
1078 rq = get_request(q, rw_flags, bio, GFP_NOIO);
1079 };
1080
1081 return rq;
1082}
1083
1084struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
1085{
1086 struct request *rq;
1087
1088 BUG_ON(rw != READ && rw != WRITE);
1089
1090 spin_lock_irq(q->queue_lock);
1091 if (gfp_mask & __GFP_WAIT)
1092 rq = get_request_wait(q, rw, NULL);
1093 else
1094 rq = get_request(q, rw, NULL, gfp_mask);
1095 if (!rq)
1096 spin_unlock_irq(q->queue_lock);
1097 /* q->queue_lock is unlocked at this point */
1098
1099 return rq;
1100}
1101EXPORT_SYMBOL(blk_get_request);
1102
1103/**
1104 * blk_make_request - given a bio, allocate a corresponding struct request.
1105 * @q: target request queue
1106 * @bio: The bio describing the memory mappings that will be submitted for IO.
1107 * It may be a chained-bio properly constructed by block/bio layer.
1108 * @gfp_mask: gfp flags to be used for memory allocation
1109 *
1110 * blk_make_request is the parallel of generic_make_request for BLOCK_PC
1111 * type commands. Where the struct request needs to be farther initialized by
1112 * the caller. It is passed a &struct bio, which describes the memory info of
1113 * the I/O transfer.
1114 *
1115 * The caller of blk_make_request must make sure that bi_io_vec
1116 * are set to describe the memory buffers. That bio_data_dir() will return
1117 * the needed direction of the request. (And all bio's in the passed bio-chain
1118 * are properly set accordingly)
1119 *
1120 * If called under none-sleepable conditions, mapped bio buffers must not
1121 * need bouncing, by calling the appropriate masked or flagged allocator,
1122 * suitable for the target device. Otherwise the call to blk_queue_bounce will
1123 * BUG.
1124 *
1125 * WARNING: When allocating/cloning a bio-chain, careful consideration should be
1126 * given to how you allocate bios. In particular, you cannot use __GFP_WAIT for
1127 * anything but the first bio in the chain. Otherwise you risk waiting for IO
1128 * completion of a bio that hasn't been submitted yet, thus resulting in a
1129 * deadlock. Alternatively bios should be allocated using bio_kmalloc() instead
1130 * of bio_alloc(), as that avoids the mempool deadlock.
1131 * If possible a big IO should be split into smaller parts when allocation
1132 * fails. Partial allocation should not be an error, or you risk a live-lock.
1133 */
1134struct request *blk_make_request(struct request_queue *q, struct bio *bio,
1135 gfp_t gfp_mask)
1136{
1137 struct request *rq = blk_get_request(q, bio_data_dir(bio), gfp_mask);
1138
1139 if (unlikely(!rq))
1140 return ERR_PTR(-ENOMEM);
1141
1142 for_each_bio(bio) {
1143 struct bio *bounce_bio = bio;
1144 int ret;
1145
1146 blk_queue_bounce(q, &bounce_bio);
1147 ret = blk_rq_append_bio(q, rq, bounce_bio);
1148 if (unlikely(ret)) {
1149 blk_put_request(rq);
1150 return ERR_PTR(ret);
1151 }
1152 }
1153
1154 return rq;
1155}
1156EXPORT_SYMBOL(blk_make_request);
1157
1158/**
1159 * blk_requeue_request - put a request back on queue
1160 * @q: request queue where request should be inserted
1161 * @rq: request to be inserted
1162 *
1163 * Description:
1164 * Drivers often keep queueing requests until the hardware cannot accept
1165 * more, when that condition happens we need to put the request back
1166 * on the queue. Must be called with queue lock held.
1167 */
1168void blk_requeue_request(struct request_queue *q, struct request *rq)
1169{
1170 blk_delete_timer(rq);
1171 blk_clear_rq_complete(rq);
1172 trace_block_rq_requeue(q, rq);
1173
1174 if (blk_rq_tagged(rq))
1175 blk_queue_end_tag(q, rq);
1176
1177 BUG_ON(blk_queued_rq(rq));
1178
1179 elv_requeue_request(q, rq);
1180}
1181EXPORT_SYMBOL(blk_requeue_request);
1182
1183static void add_acct_request(struct request_queue *q, struct request *rq,
1184 int where)
1185{
1186 drive_stat_acct(rq, 1);
1187 __elv_add_request(q, rq, where);
1188}
1189
1190static void part_round_stats_single(int cpu, struct hd_struct *part,
1191 unsigned long now)
1192{
1193 if (now == part->stamp)
1194 return;
1195
1196 if (part_in_flight(part)) {
1197 __part_stat_add(cpu, part, time_in_queue,
1198 part_in_flight(part) * (now - part->stamp));
1199 __part_stat_add(cpu, part, io_ticks, (now - part->stamp));
1200 }
1201 part->stamp = now;
1202}
1203
1204/**
1205 * part_round_stats() - Round off the performance stats on a struct disk_stats.
1206 * @cpu: cpu number for stats access
1207 * @part: target partition
1208 *
1209 * The average IO queue length and utilisation statistics are maintained
1210 * by observing the current state of the queue length and the amount of
1211 * time it has been in this state for.
1212 *
1213 * Normally, that accounting is done on IO completion, but that can result
1214 * in more than a second's worth of IO being accounted for within any one
1215 * second, leading to >100% utilisation. To deal with that, we call this
1216 * function to do a round-off before returning the results when reading
1217 * /proc/diskstats. This accounts immediately for all queue usage up to
1218 * the current jiffies and restarts the counters again.
1219 */
1220void part_round_stats(int cpu, struct hd_struct *part)
1221{
1222 unsigned long now = jiffies;
1223
1224 if (part->partno)
1225 part_round_stats_single(cpu, &part_to_disk(part)->part0, now);
1226 part_round_stats_single(cpu, part, now);
1227}
1228EXPORT_SYMBOL_GPL(part_round_stats);
1229
1230/*
1231 * queue lock must be held
1232 */
1233void __blk_put_request(struct request_queue *q, struct request *req)
1234{
1235 if (unlikely(!q))
1236 return;
1237 if (unlikely(--req->ref_count))
1238 return;
1239
1240 elv_completed_request(q, req);
1241
1242 /* this is a bio leak */
1243 WARN_ON(req->bio != NULL);
1244
1245 /*
1246 * Request may not have originated from ll_rw_blk. if not,
1247 * it didn't come out of our reserved rq pools
1248 */
1249 if (req->cmd_flags & REQ_ALLOCED) {
1250 unsigned int flags = req->cmd_flags;
1251
1252 BUG_ON(!list_empty(&req->queuelist));
1253 BUG_ON(!hlist_unhashed(&req->hash));
1254
1255 blk_free_request(q, req);
1256 freed_request(q, flags);
1257 }
1258}
1259EXPORT_SYMBOL_GPL(__blk_put_request);
1260
1261void blk_put_request(struct request *req)
1262{
1263 unsigned long flags;
1264 struct request_queue *q = req->q;
1265
1266 spin_lock_irqsave(q->queue_lock, flags);
1267 __blk_put_request(q, req);
1268 spin_unlock_irqrestore(q->queue_lock, flags);
1269}
1270EXPORT_SYMBOL(blk_put_request);
1271
1272/**
1273 * blk_add_request_payload - add a payload to a request
1274 * @rq: request to update
1275 * @page: page backing the payload
1276 * @len: length of the payload.
1277 *
1278 * This allows to later add a payload to an already submitted request by
1279 * a block driver. The driver needs to take care of freeing the payload
1280 * itself.
1281 *
1282 * Note that this is a quite horrible hack and nothing but handling of
1283 * discard requests should ever use it.
1284 */
1285void blk_add_request_payload(struct request *rq, struct page *page,
1286 unsigned int len)
1287{
1288 struct bio *bio = rq->bio;
1289
1290 bio->bi_io_vec->bv_page = page;
1291 bio->bi_io_vec->bv_offset = 0;
1292 bio->bi_io_vec->bv_len = len;
1293
1294 bio->bi_size = len;
1295 bio->bi_vcnt = 1;
1296 bio->bi_phys_segments = 1;
1297
1298 rq->__data_len = rq->resid_len = len;
1299 rq->nr_phys_segments = 1;
1300 rq->buffer = bio_data(bio);
1301}
1302EXPORT_SYMBOL_GPL(blk_add_request_payload);
1303
1304static bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
1305 struct bio *bio)
1306{
1307 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1308
1309 if (!ll_back_merge_fn(q, req, bio))
1310 return false;
1311
1312 trace_block_bio_backmerge(q, bio);
1313
1314 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1315 blk_rq_set_mixed_merge(req);
1316
1317 req->biotail->bi_next = bio;
1318 req->biotail = bio;
1319 req->__data_len += bio->bi_size;
1320 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1321
1322 drive_stat_acct(req, 0);
1323 return true;
1324}
1325
1326static bool bio_attempt_front_merge(struct request_queue *q,
1327 struct request *req, struct bio *bio)
1328{
1329 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1330
1331 if (!ll_front_merge_fn(q, req, bio))
1332 return false;
1333
1334 trace_block_bio_frontmerge(q, bio);
1335
1336 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1337 blk_rq_set_mixed_merge(req);
1338
1339 bio->bi_next = req->bio;
1340 req->bio = bio;
1341
1342 /*
1343 * may not be valid. if the low level driver said
1344 * it didn't need a bounce buffer then it better
1345 * not touch req->buffer either...
1346 */
1347 req->buffer = bio_data(bio);
1348 req->__sector = bio->bi_sector;
1349 req->__data_len += bio->bi_size;
1350 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1351
1352 drive_stat_acct(req, 0);
1353 return true;
1354}
1355
1356/**
1357 * attempt_plug_merge - try to merge with %current's plugged list
1358 * @q: request_queue new bio is being queued at
1359 * @bio: new bio being queued
1360 * @request_count: out parameter for number of traversed plugged requests
1361 *
1362 * Determine whether @bio being queued on @q can be merged with a request
1363 * on %current's plugged list. Returns %true if merge was successful,
1364 * otherwise %false.
1365 *
1366 * Plugging coalesces IOs from the same issuer for the same purpose without
1367 * going through @q->queue_lock. As such it's more of an issuing mechanism
1368 * than scheduling, and the request, while may have elvpriv data, is not
1369 * added on the elevator at this point. In addition, we don't have
1370 * reliable access to the elevator outside queue lock. Only check basic
1371 * merging parameters without querying the elevator.
1372 */
1373static bool attempt_plug_merge(struct request_queue *q, struct bio *bio,
1374 unsigned int *request_count)
1375{
1376 struct blk_plug *plug;
1377 struct request *rq;
1378 bool ret = false;
1379
1380 plug = current->plug;
1381 if (!plug)
1382 goto out;
1383 *request_count = 0;
1384
1385 list_for_each_entry_reverse(rq, &plug->list, queuelist) {
1386 int el_ret;
1387
1388 if (rq->q == q)
1389 (*request_count)++;
1390
1391 if (rq->q != q || !blk_rq_merge_ok(rq, bio))
1392 continue;
1393
1394 el_ret = blk_try_merge(rq, bio);
1395 if (el_ret == ELEVATOR_BACK_MERGE) {
1396 ret = bio_attempt_back_merge(q, rq, bio);
1397 if (ret)
1398 break;
1399 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1400 ret = bio_attempt_front_merge(q, rq, bio);
1401 if (ret)
1402 break;
1403 }
1404 }
1405out:
1406 return ret;
1407}
1408
1409void init_request_from_bio(struct request *req, struct bio *bio)
1410{
1411 req->cmd_type = REQ_TYPE_FS;
1412
1413 req->cmd_flags |= bio->bi_rw & REQ_COMMON_MASK;
1414 if (bio->bi_rw & REQ_RAHEAD)
1415 req->cmd_flags |= REQ_FAILFAST_MASK;
1416
1417 req->errors = 0;
1418 req->__sector = bio->bi_sector;
1419 req->ioprio = bio_prio(bio);
1420 blk_rq_bio_prep(req->q, req, bio);
1421}
1422
1423void blk_queue_bio(struct request_queue *q, struct bio *bio)
1424{
1425 const bool sync = !!(bio->bi_rw & REQ_SYNC);
1426 struct blk_plug *plug;
1427 int el_ret, rw_flags, where = ELEVATOR_INSERT_SORT;
1428 struct request *req;
1429 unsigned int request_count = 0;
1430
1431 /*
1432 * low level driver can indicate that it wants pages above a
1433 * certain limit bounced to low memory (ie for highmem, or even
1434 * ISA dma in theory)
1435 */
1436 blk_queue_bounce(q, &bio);
1437
1438 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
1439 spin_lock_irq(q->queue_lock);
1440 where = ELEVATOR_INSERT_FLUSH;
1441 goto get_rq;
1442 }
1443
1444 /*
1445 * Check if we can merge with the plugged list before grabbing
1446 * any locks.
1447 */
1448 if (attempt_plug_merge(q, bio, &request_count))
1449 return;
1450
1451 spin_lock_irq(q->queue_lock);
1452
1453 el_ret = elv_merge(q, &req, bio);
1454 if (el_ret == ELEVATOR_BACK_MERGE) {
1455 if (bio_attempt_back_merge(q, req, bio)) {
1456 elv_bio_merged(q, req, bio);
1457 if (!attempt_back_merge(q, req))
1458 elv_merged_request(q, req, el_ret);
1459 goto out_unlock;
1460 }
1461 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1462 if (bio_attempt_front_merge(q, req, bio)) {
1463 elv_bio_merged(q, req, bio);
1464 if (!attempt_front_merge(q, req))
1465 elv_merged_request(q, req, el_ret);
1466 goto out_unlock;
1467 }
1468 }
1469
1470get_rq:
1471 /*
1472 * This sync check and mask will be re-done in init_request_from_bio(),
1473 * but we need to set it earlier to expose the sync flag to the
1474 * rq allocator and io schedulers.
1475 */
1476 rw_flags = bio_data_dir(bio);
1477 if (sync)
1478 rw_flags |= REQ_SYNC;
1479
1480 /*
1481 * Grab a free request. This is might sleep but can not fail.
1482 * Returns with the queue unlocked.
1483 */
1484 req = get_request_wait(q, rw_flags, bio);
1485 if (unlikely(!req)) {
1486 bio_endio(bio, -ENODEV); /* @q is dead */
1487 goto out_unlock;
1488 }
1489
1490 /*
1491 * After dropping the lock and possibly sleeping here, our request
1492 * may now be mergeable after it had proven unmergeable (above).
1493 * We don't worry about that case for efficiency. It won't happen
1494 * often, and the elevators are able to handle it.
1495 */
1496 init_request_from_bio(req, bio);
1497
1498 if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags))
1499 req->cpu = raw_smp_processor_id();
1500
1501 plug = current->plug;
1502 if (plug) {
1503 /*
1504 * If this is the first request added after a plug, fire
1505 * of a plug trace. If others have been added before, check
1506 * if we have multiple devices in this plug. If so, make a
1507 * note to sort the list before dispatch.
1508 */
1509 if (list_empty(&plug->list))
1510 trace_block_plug(q);
1511 else {
1512 if (!plug->should_sort) {
1513 struct request *__rq;
1514
1515 __rq = list_entry_rq(plug->list.prev);
1516 if (__rq->q != q)
1517 plug->should_sort = 1;
1518 }
1519 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1520 blk_flush_plug_list(plug, false);
1521 trace_block_plug(q);
1522 }
1523 }
1524 list_add_tail(&req->queuelist, &plug->list);
1525 drive_stat_acct(req, 1);
1526 } else {
1527 spin_lock_irq(q->queue_lock);
1528 add_acct_request(q, req, where);
1529 __blk_run_queue(q);
1530out_unlock:
1531 spin_unlock_irq(q->queue_lock);
1532 }
1533}
1534EXPORT_SYMBOL_GPL(blk_queue_bio); /* for device mapper only */
1535
1536/*
1537 * If bio->bi_dev is a partition, remap the location
1538 */
1539static inline void blk_partition_remap(struct bio *bio)
1540{
1541 struct block_device *bdev = bio->bi_bdev;
1542
1543 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
1544 struct hd_struct *p = bdev->bd_part;
1545
1546 bio->bi_sector += p->start_sect;
1547 bio->bi_bdev = bdev->bd_contains;
1548
1549 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev), bio,
1550 bdev->bd_dev,
1551 bio->bi_sector - p->start_sect);
1552 }
1553}
1554
1555static void handle_bad_sector(struct bio *bio)
1556{
1557 char b[BDEVNAME_SIZE];
1558
1559 printk(KERN_INFO "attempt to access beyond end of device\n");
1560 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
1561 bdevname(bio->bi_bdev, b),
1562 bio->bi_rw,
1563 (unsigned long long)bio->bi_sector + bio_sectors(bio),
1564 (long long)(i_size_read(bio->bi_bdev->bd_inode) >> 9));
1565
1566 set_bit(BIO_EOF, &bio->bi_flags);
1567}
1568
1569#ifdef CONFIG_FAIL_MAKE_REQUEST
1570
1571static DECLARE_FAULT_ATTR(fail_make_request);
1572
1573static int __init setup_fail_make_request(char *str)
1574{
1575 return setup_fault_attr(&fail_make_request, str);
1576}
1577__setup("fail_make_request=", setup_fail_make_request);
1578
1579static bool should_fail_request(struct hd_struct *part, unsigned int bytes)
1580{
1581 return part->make_it_fail && should_fail(&fail_make_request, bytes);
1582}
1583
1584static int __init fail_make_request_debugfs(void)
1585{
1586 struct dentry *dir = fault_create_debugfs_attr("fail_make_request",
1587 NULL, &fail_make_request);
1588
1589 return IS_ERR(dir) ? PTR_ERR(dir) : 0;
1590}
1591
1592late_initcall(fail_make_request_debugfs);
1593
1594#else /* CONFIG_FAIL_MAKE_REQUEST */
1595
1596static inline bool should_fail_request(struct hd_struct *part,
1597 unsigned int bytes)
1598{
1599 return false;
1600}
1601
1602#endif /* CONFIG_FAIL_MAKE_REQUEST */
1603
1604/*
1605 * Check whether this bio extends beyond the end of the device.
1606 */
1607static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
1608{
1609 sector_t maxsector;
1610
1611 if (!nr_sectors)
1612 return 0;
1613
1614 /* Test device or partition size, when known. */
1615 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
1616 if (maxsector) {
1617 sector_t sector = bio->bi_sector;
1618
1619 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
1620 /*
1621 * This may well happen - the kernel calls bread()
1622 * without checking the size of the device, e.g., when
1623 * mounting a device.
1624 */
1625 handle_bad_sector(bio);
1626 return 1;
1627 }
1628 }
1629
1630 return 0;
1631}
1632
1633static noinline_for_stack bool
1634generic_make_request_checks(struct bio *bio)
1635{
1636 struct request_queue *q;
1637 int nr_sectors = bio_sectors(bio);
1638 int err = -EIO;
1639 char b[BDEVNAME_SIZE];
1640 struct hd_struct *part;
1641
1642 might_sleep();
1643
1644 if (bio_check_eod(bio, nr_sectors))
1645 goto end_io;
1646
1647 q = bdev_get_queue(bio->bi_bdev);
1648 if (unlikely(!q)) {
1649 printk(KERN_ERR
1650 "generic_make_request: Trying to access "
1651 "nonexistent block-device %s (%Lu)\n",
1652 bdevname(bio->bi_bdev, b),
1653 (long long) bio->bi_sector);
1654 goto end_io;
1655 }
1656
1657 if (unlikely(!(bio->bi_rw & REQ_DISCARD) &&
1658 nr_sectors > queue_max_hw_sectors(q))) {
1659 printk(KERN_ERR "bio too big device %s (%u > %u)\n",
1660 bdevname(bio->bi_bdev, b),
1661 bio_sectors(bio),
1662 queue_max_hw_sectors(q));
1663 goto end_io;
1664 }
1665
1666 part = bio->bi_bdev->bd_part;
1667 if (should_fail_request(part, bio->bi_size) ||
1668 should_fail_request(&part_to_disk(part)->part0,
1669 bio->bi_size))
1670 goto end_io;
1671
1672 /*
1673 * If this device has partitions, remap block n
1674 * of partition p to block n+start(p) of the disk.
1675 */
1676 blk_partition_remap(bio);
1677
1678 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio))
1679 goto end_io;
1680
1681 if (bio_check_eod(bio, nr_sectors))
1682 goto end_io;
1683
1684 /*
1685 * Filter flush bio's early so that make_request based
1686 * drivers without flush support don't have to worry
1687 * about them.
1688 */
1689 if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) && !q->flush_flags) {
1690 bio->bi_rw &= ~(REQ_FLUSH | REQ_FUA);
1691 if (!nr_sectors) {
1692 err = 0;
1693 goto end_io;
1694 }
1695 }
1696
1697 if ((bio->bi_rw & REQ_DISCARD) &&
1698 (!blk_queue_discard(q) ||
1699 ((bio->bi_rw & REQ_SECURE) &&
1700 !blk_queue_secdiscard(q)))) {
1701 err = -EOPNOTSUPP;
1702 goto end_io;
1703 }
1704
1705 if (blk_throtl_bio(q, bio))
1706 return false; /* throttled, will be resubmitted later */
1707
1708 trace_block_bio_queue(q, bio);
1709 return true;
1710
1711end_io:
1712 bio_endio(bio, err);
1713 return false;
1714}
1715
1716/**
1717 * generic_make_request - hand a buffer to its device driver for I/O
1718 * @bio: The bio describing the location in memory and on the device.
1719 *
1720 * generic_make_request() is used to make I/O requests of block
1721 * devices. It is passed a &struct bio, which describes the I/O that needs
1722 * to be done.
1723 *
1724 * generic_make_request() does not return any status. The
1725 * success/failure status of the request, along with notification of
1726 * completion, is delivered asynchronously through the bio->bi_end_io
1727 * function described (one day) else where.
1728 *
1729 * The caller of generic_make_request must make sure that bi_io_vec
1730 * are set to describe the memory buffer, and that bi_dev and bi_sector are
1731 * set to describe the device address, and the
1732 * bi_end_io and optionally bi_private are set to describe how
1733 * completion notification should be signaled.
1734 *
1735 * generic_make_request and the drivers it calls may use bi_next if this
1736 * bio happens to be merged with someone else, and may resubmit the bio to
1737 * a lower device by calling into generic_make_request recursively, which
1738 * means the bio should NOT be touched after the call to ->make_request_fn.
1739 */
1740void generic_make_request(struct bio *bio)
1741{
1742 struct bio_list bio_list_on_stack;
1743
1744 if (!generic_make_request_checks(bio))
1745 return;
1746
1747 /*
1748 * We only want one ->make_request_fn to be active at a time, else
1749 * stack usage with stacked devices could be a problem. So use
1750 * current->bio_list to keep a list of requests submited by a
1751 * make_request_fn function. current->bio_list is also used as a
1752 * flag to say if generic_make_request is currently active in this
1753 * task or not. If it is NULL, then no make_request is active. If
1754 * it is non-NULL, then a make_request is active, and new requests
1755 * should be added at the tail
1756 */
1757 if (current->bio_list) {
1758 bio_list_add(current->bio_list, bio);
1759 return;
1760 }
1761
1762 /* following loop may be a bit non-obvious, and so deserves some
1763 * explanation.
1764 * Before entering the loop, bio->bi_next is NULL (as all callers
1765 * ensure that) so we have a list with a single bio.
1766 * We pretend that we have just taken it off a longer list, so
1767 * we assign bio_list to a pointer to the bio_list_on_stack,
1768 * thus initialising the bio_list of new bios to be
1769 * added. ->make_request() may indeed add some more bios
1770 * through a recursive call to generic_make_request. If it
1771 * did, we find a non-NULL value in bio_list and re-enter the loop
1772 * from the top. In this case we really did just take the bio
1773 * of the top of the list (no pretending) and so remove it from
1774 * bio_list, and call into ->make_request() again.
1775 */
1776 BUG_ON(bio->bi_next);
1777 bio_list_init(&bio_list_on_stack);
1778 current->bio_list = &bio_list_on_stack;
1779 do {
1780 struct request_queue *q = bdev_get_queue(bio->bi_bdev);
1781
1782 q->make_request_fn(q, bio);
1783
1784 bio = bio_list_pop(current->bio_list);
1785 } while (bio);
1786 current->bio_list = NULL; /* deactivate */
1787}
1788EXPORT_SYMBOL(generic_make_request);
1789
1790/**
1791 * submit_bio - submit a bio to the block device layer for I/O
1792 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
1793 * @bio: The &struct bio which describes the I/O
1794 *
1795 * submit_bio() is very similar in purpose to generic_make_request(), and
1796 * uses that function to do most of the work. Both are fairly rough
1797 * interfaces; @bio must be presetup and ready for I/O.
1798 *
1799 */
1800void submit_bio(int rw, struct bio *bio)
1801{
1802 int count = bio_sectors(bio);
1803
1804 bio->bi_rw |= rw;
1805
1806 /*
1807 * If it's a regular read/write or a barrier with data attached,
1808 * go through the normal accounting stuff before submission.
1809 */
1810 if (bio_has_data(bio) && !(rw & REQ_DISCARD)) {
1811 if (rw & WRITE) {
1812 count_vm_events(PGPGOUT, count);
1813 } else {
1814 task_io_account_read(bio->bi_size);
1815 count_vm_events(PGPGIN, count);
1816 }
1817
1818 if (unlikely(block_dump)) {
1819 char b[BDEVNAME_SIZE];
1820 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s (%u sectors)\n",
1821 current->comm, task_pid_nr(current),
1822 (rw & WRITE) ? "WRITE" : "READ",
1823 (unsigned long long)bio->bi_sector,
1824 bdevname(bio->bi_bdev, b),
1825 count);
1826 }
1827 }
1828
1829 generic_make_request(bio);
1830}
1831EXPORT_SYMBOL(submit_bio);
1832
1833/**
1834 * blk_rq_check_limits - Helper function to check a request for the queue limit
1835 * @q: the queue
1836 * @rq: the request being checked
1837 *
1838 * Description:
1839 * @rq may have been made based on weaker limitations of upper-level queues
1840 * in request stacking drivers, and it may violate the limitation of @q.
1841 * Since the block layer and the underlying device driver trust @rq
1842 * after it is inserted to @q, it should be checked against @q before
1843 * the insertion using this generic function.
1844 *
1845 * This function should also be useful for request stacking drivers
1846 * in some cases below, so export this function.
1847 * Request stacking drivers like request-based dm may change the queue
1848 * limits while requests are in the queue (e.g. dm's table swapping).
1849 * Such request stacking drivers should check those requests agaist
1850 * the new queue limits again when they dispatch those requests,
1851 * although such checkings are also done against the old queue limits
1852 * when submitting requests.
1853 */
1854int blk_rq_check_limits(struct request_queue *q, struct request *rq)
1855{
1856 if (rq->cmd_flags & REQ_DISCARD)
1857 return 0;
1858
1859 if (blk_rq_sectors(rq) > queue_max_sectors(q) ||
1860 blk_rq_bytes(rq) > queue_max_hw_sectors(q) << 9) {
1861 printk(KERN_ERR "%s: over max size limit.\n", __func__);
1862 return -EIO;
1863 }
1864
1865 /*
1866 * queue's settings related to segment counting like q->bounce_pfn
1867 * may differ from that of other stacking queues.
1868 * Recalculate it to check the request correctly on this queue's
1869 * limitation.
1870 */
1871 blk_recalc_rq_segments(rq);
1872 if (rq->nr_phys_segments > queue_max_segments(q)) {
1873 printk(KERN_ERR "%s: over max segments limit.\n", __func__);
1874 return -EIO;
1875 }
1876
1877 return 0;
1878}
1879EXPORT_SYMBOL_GPL(blk_rq_check_limits);
1880
1881/**
1882 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
1883 * @q: the queue to submit the request
1884 * @rq: the request being queued
1885 */
1886int blk_insert_cloned_request(struct request_queue *q, struct request *rq)
1887{
1888 unsigned long flags;
1889 int where = ELEVATOR_INSERT_BACK;
1890
1891 if (blk_rq_check_limits(q, rq))
1892 return -EIO;
1893
1894 if (rq->rq_disk &&
1895 should_fail_request(&rq->rq_disk->part0, blk_rq_bytes(rq)))
1896 return -EIO;
1897
1898 spin_lock_irqsave(q->queue_lock, flags);
1899 if (unlikely(blk_queue_dead(q))) {
1900 spin_unlock_irqrestore(q->queue_lock, flags);
1901 return -ENODEV;
1902 }
1903
1904 /*
1905 * Submitting request must be dequeued before calling this function
1906 * because it will be linked to another request_queue
1907 */
1908 BUG_ON(blk_queued_rq(rq));
1909
1910 if (rq->cmd_flags & (REQ_FLUSH|REQ_FUA))
1911 where = ELEVATOR_INSERT_FLUSH;
1912
1913 add_acct_request(q, rq, where);
1914 if (where == ELEVATOR_INSERT_FLUSH)
1915 __blk_run_queue(q);
1916 spin_unlock_irqrestore(q->queue_lock, flags);
1917
1918 return 0;
1919}
1920EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
1921
1922/**
1923 * blk_rq_err_bytes - determine number of bytes till the next failure boundary
1924 * @rq: request to examine
1925 *
1926 * Description:
1927 * A request could be merge of IOs which require different failure
1928 * handling. This function determines the number of bytes which
1929 * can be failed from the beginning of the request without
1930 * crossing into area which need to be retried further.
1931 *
1932 * Return:
1933 * The number of bytes to fail.
1934 *
1935 * Context:
1936 * queue_lock must be held.
1937 */
1938unsigned int blk_rq_err_bytes(const struct request *rq)
1939{
1940 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
1941 unsigned int bytes = 0;
1942 struct bio *bio;
1943
1944 if (!(rq->cmd_flags & REQ_MIXED_MERGE))
1945 return blk_rq_bytes(rq);
1946
1947 /*
1948 * Currently the only 'mixing' which can happen is between
1949 * different fastfail types. We can safely fail portions
1950 * which have all the failfast bits that the first one has -
1951 * the ones which are at least as eager to fail as the first
1952 * one.
1953 */
1954 for (bio = rq->bio; bio; bio = bio->bi_next) {
1955 if ((bio->bi_rw & ff) != ff)
1956 break;
1957 bytes += bio->bi_size;
1958 }
1959
1960 /* this could lead to infinite loop */
1961 BUG_ON(blk_rq_bytes(rq) && !bytes);
1962 return bytes;
1963}
1964EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
1965
1966static void blk_account_io_completion(struct request *req, unsigned int bytes)
1967{
1968 if (blk_do_io_stat(req)) {
1969 const int rw = rq_data_dir(req);
1970 struct hd_struct *part;
1971 int cpu;
1972
1973 cpu = part_stat_lock();
1974 part = req->part;
1975 part_stat_add(cpu, part, sectors[rw], bytes >> 9);
1976 part_stat_unlock();
1977 }
1978}
1979
1980static void blk_account_io_done(struct request *req)
1981{
1982 /*
1983 * Account IO completion. flush_rq isn't accounted as a
1984 * normal IO on queueing nor completion. Accounting the
1985 * containing request is enough.
1986 */
1987 if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_FLUSH_SEQ)) {
1988 unsigned long duration = jiffies - req->start_time;
1989 const int rw = rq_data_dir(req);
1990 struct hd_struct *part;
1991 int cpu;
1992
1993 cpu = part_stat_lock();
1994 part = req->part;
1995
1996 part_stat_inc(cpu, part, ios[rw]);
1997 part_stat_add(cpu, part, ticks[rw], duration);
1998 part_round_stats(cpu, part);
1999 part_dec_in_flight(part, rw);
2000
2001 hd_struct_put(part);
2002 part_stat_unlock();
2003 }
2004}
2005
2006/**
2007 * blk_peek_request - peek at the top of a request queue
2008 * @q: request queue to peek at
2009 *
2010 * Description:
2011 * Return the request at the top of @q. The returned request
2012 * should be started using blk_start_request() before LLD starts
2013 * processing it.
2014 *
2015 * Return:
2016 * Pointer to the request at the top of @q if available. Null
2017 * otherwise.
2018 *
2019 * Context:
2020 * queue_lock must be held.
2021 */
2022struct request *blk_peek_request(struct request_queue *q)
2023{
2024 struct request *rq;
2025 int ret;
2026
2027 while ((rq = __elv_next_request(q)) != NULL) {
2028 if (!(rq->cmd_flags & REQ_STARTED)) {
2029 /*
2030 * This is the first time the device driver
2031 * sees this request (possibly after
2032 * requeueing). Notify IO scheduler.
2033 */
2034 if (rq->cmd_flags & REQ_SORTED)
2035 elv_activate_rq(q, rq);
2036
2037 /*
2038 * just mark as started even if we don't start
2039 * it, a request that has been delayed should
2040 * not be passed by new incoming requests
2041 */
2042 rq->cmd_flags |= REQ_STARTED;
2043 trace_block_rq_issue(q, rq);
2044 }
2045
2046 if (!q->boundary_rq || q->boundary_rq == rq) {
2047 q->end_sector = rq_end_sector(rq);
2048 q->boundary_rq = NULL;
2049 }
2050
2051 if (rq->cmd_flags & REQ_DONTPREP)
2052 break;
2053
2054 if (q->dma_drain_size && blk_rq_bytes(rq)) {
2055 /*
2056 * make sure space for the drain appears we
2057 * know we can do this because max_hw_segments
2058 * has been adjusted to be one fewer than the
2059 * device can handle
2060 */
2061 rq->nr_phys_segments++;
2062 }
2063
2064 if (!q->prep_rq_fn)
2065 break;
2066
2067 ret = q->prep_rq_fn(q, rq);
2068 if (ret == BLKPREP_OK) {
2069 break;
2070 } else if (ret == BLKPREP_DEFER) {
2071 /*
2072 * the request may have been (partially) prepped.
2073 * we need to keep this request in the front to
2074 * avoid resource deadlock. REQ_STARTED will
2075 * prevent other fs requests from passing this one.
2076 */
2077 if (q->dma_drain_size && blk_rq_bytes(rq) &&
2078 !(rq->cmd_flags & REQ_DONTPREP)) {
2079 /*
2080 * remove the space for the drain we added
2081 * so that we don't add it again
2082 */
2083 --rq->nr_phys_segments;
2084 }
2085
2086 rq = NULL;
2087 break;
2088 } else if (ret == BLKPREP_KILL) {
2089 rq->cmd_flags |= REQ_QUIET;
2090 /*
2091 * Mark this request as started so we don't trigger
2092 * any debug logic in the end I/O path.
2093 */
2094 blk_start_request(rq);
2095 __blk_end_request_all(rq, -EIO);
2096 } else {
2097 printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
2098 break;
2099 }
2100 }
2101
2102 return rq;
2103}
2104EXPORT_SYMBOL(blk_peek_request);
2105
2106void blk_dequeue_request(struct request *rq)
2107{
2108 struct request_queue *q = rq->q;
2109
2110 BUG_ON(list_empty(&rq->queuelist));
2111 BUG_ON(ELV_ON_HASH(rq));
2112
2113 list_del_init(&rq->queuelist);
2114
2115 /*
2116 * the time frame between a request being removed from the lists
2117 * and to it is freed is accounted as io that is in progress at
2118 * the driver side.
2119 */
2120 if (blk_account_rq(rq)) {
2121 q->in_flight[rq_is_sync(rq)]++;
2122 set_io_start_time_ns(rq);
2123 }
2124}
2125
2126/**
2127 * blk_start_request - start request processing on the driver
2128 * @req: request to dequeue
2129 *
2130 * Description:
2131 * Dequeue @req and start timeout timer on it. This hands off the
2132 * request to the driver.
2133 *
2134 * Block internal functions which don't want to start timer should
2135 * call blk_dequeue_request().
2136 *
2137 * Context:
2138 * queue_lock must be held.
2139 */
2140void blk_start_request(struct request *req)
2141{
2142 blk_dequeue_request(req);
2143
2144 /*
2145 * We are now handing the request to the hardware, initialize
2146 * resid_len to full count and add the timeout handler.
2147 */
2148 req->resid_len = blk_rq_bytes(req);
2149 if (unlikely(blk_bidi_rq(req)))
2150 req->next_rq->resid_len = blk_rq_bytes(req->next_rq);
2151
2152 blk_add_timer(req);
2153}
2154EXPORT_SYMBOL(blk_start_request);
2155
2156/**
2157 * blk_fetch_request - fetch a request from a request queue
2158 * @q: request queue to fetch a request from
2159 *
2160 * Description:
2161 * Return the request at the top of @q. The request is started on
2162 * return and LLD can start processing it immediately.
2163 *
2164 * Return:
2165 * Pointer to the request at the top of @q if available. Null
2166 * otherwise.
2167 *
2168 * Context:
2169 * queue_lock must be held.
2170 */
2171struct request *blk_fetch_request(struct request_queue *q)
2172{
2173 struct request *rq;
2174
2175 rq = blk_peek_request(q);
2176 if (rq)
2177 blk_start_request(rq);
2178 return rq;
2179}
2180EXPORT_SYMBOL(blk_fetch_request);
2181
2182/**
2183 * blk_update_request - Special helper function for request stacking drivers
2184 * @req: the request being processed
2185 * @error: %0 for success, < %0 for error
2186 * @nr_bytes: number of bytes to complete @req
2187 *
2188 * Description:
2189 * Ends I/O on a number of bytes attached to @req, but doesn't complete
2190 * the request structure even if @req doesn't have leftover.
2191 * If @req has leftover, sets it up for the next range of segments.
2192 *
2193 * This special helper function is only for request stacking drivers
2194 * (e.g. request-based dm) so that they can handle partial completion.
2195 * Actual device drivers should use blk_end_request instead.
2196 *
2197 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
2198 * %false return from this function.
2199 *
2200 * Return:
2201 * %false - this request doesn't have any more data
2202 * %true - this request has more data
2203 **/
2204bool blk_update_request(struct request *req, int error, unsigned int nr_bytes)
2205{
2206 int total_bytes, bio_nbytes, next_idx = 0;
2207 struct bio *bio;
2208
2209 if (!req->bio)
2210 return false;
2211
2212 trace_block_rq_complete(req->q, req);
2213
2214 /*
2215 * For fs requests, rq is just carrier of independent bio's
2216 * and each partial completion should be handled separately.
2217 * Reset per-request error on each partial completion.
2218 *
2219 * TODO: tj: This is too subtle. It would be better to let
2220 * low level drivers do what they see fit.
2221 */
2222 if (req->cmd_type == REQ_TYPE_FS)
2223 req->errors = 0;
2224
2225 if (error && req->cmd_type == REQ_TYPE_FS &&
2226 !(req->cmd_flags & REQ_QUIET)) {
2227 char *error_type;
2228
2229 switch (error) {
2230 case -ENOLINK:
2231 error_type = "recoverable transport";
2232 break;
2233 case -EREMOTEIO:
2234 error_type = "critical target";
2235 break;
2236 case -EBADE:
2237 error_type = "critical nexus";
2238 break;
2239 case -EIO:
2240 default:
2241 error_type = "I/O";
2242 break;
2243 }
2244 printk(KERN_ERR "end_request: %s error, dev %s, sector %llu\n",
2245 error_type, req->rq_disk ? req->rq_disk->disk_name : "?",
2246 (unsigned long long)blk_rq_pos(req));
2247 }
2248
2249 blk_account_io_completion(req, nr_bytes);
2250
2251 total_bytes = bio_nbytes = 0;
2252 while ((bio = req->bio) != NULL) {
2253 int nbytes;
2254
2255 if (nr_bytes >= bio->bi_size) {
2256 req->bio = bio->bi_next;
2257 nbytes = bio->bi_size;
2258 req_bio_endio(req, bio, nbytes, error);
2259 next_idx = 0;
2260 bio_nbytes = 0;
2261 } else {
2262 int idx = bio->bi_idx + next_idx;
2263
2264 if (unlikely(idx >= bio->bi_vcnt)) {
2265 blk_dump_rq_flags(req, "__end_that");
2266 printk(KERN_ERR "%s: bio idx %d >= vcnt %d\n",
2267 __func__, idx, bio->bi_vcnt);
2268 break;
2269 }
2270
2271 nbytes = bio_iovec_idx(bio, idx)->bv_len;
2272 BIO_BUG_ON(nbytes > bio->bi_size);
2273
2274 /*
2275 * not a complete bvec done
2276 */
2277 if (unlikely(nbytes > nr_bytes)) {
2278 bio_nbytes += nr_bytes;
2279 total_bytes += nr_bytes;
2280 break;
2281 }
2282
2283 /*
2284 * advance to the next vector
2285 */
2286 next_idx++;
2287 bio_nbytes += nbytes;
2288 }
2289
2290 total_bytes += nbytes;
2291 nr_bytes -= nbytes;
2292
2293 bio = req->bio;
2294 if (bio) {
2295 /*
2296 * end more in this run, or just return 'not-done'
2297 */
2298 if (unlikely(nr_bytes <= 0))
2299 break;
2300 }
2301 }
2302
2303 /*
2304 * completely done
2305 */
2306 if (!req->bio) {
2307 /*
2308 * Reset counters so that the request stacking driver
2309 * can find how many bytes remain in the request
2310 * later.
2311 */
2312 req->__data_len = 0;
2313 return false;
2314 }
2315
2316 /*
2317 * if the request wasn't completed, update state
2318 */
2319 if (bio_nbytes) {
2320 req_bio_endio(req, bio, bio_nbytes, error);
2321 bio->bi_idx += next_idx;
2322 bio_iovec(bio)->bv_offset += nr_bytes;
2323 bio_iovec(bio)->bv_len -= nr_bytes;
2324 }
2325
2326 req->__data_len -= total_bytes;
2327 req->buffer = bio_data(req->bio);
2328
2329 /* update sector only for requests with clear definition of sector */
2330 if (req->cmd_type == REQ_TYPE_FS || (req->cmd_flags & REQ_DISCARD))
2331 req->__sector += total_bytes >> 9;
2332
2333 /* mixed attributes always follow the first bio */
2334 if (req->cmd_flags & REQ_MIXED_MERGE) {
2335 req->cmd_flags &= ~REQ_FAILFAST_MASK;
2336 req->cmd_flags |= req->bio->bi_rw & REQ_FAILFAST_MASK;
2337 }
2338
2339 /*
2340 * If total number of sectors is less than the first segment
2341 * size, something has gone terribly wrong.
2342 */
2343 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
2344 blk_dump_rq_flags(req, "request botched");
2345 req->__data_len = blk_rq_cur_bytes(req);
2346 }
2347
2348 /* recalculate the number of segments */
2349 blk_recalc_rq_segments(req);
2350
2351 return true;
2352}
2353EXPORT_SYMBOL_GPL(blk_update_request);
2354
2355static bool blk_update_bidi_request(struct request *rq, int error,
2356 unsigned int nr_bytes,
2357 unsigned int bidi_bytes)
2358{
2359 if (blk_update_request(rq, error, nr_bytes))
2360 return true;
2361
2362 /* Bidi request must be completed as a whole */
2363 if (unlikely(blk_bidi_rq(rq)) &&
2364 blk_update_request(rq->next_rq, error, bidi_bytes))
2365 return true;
2366
2367 if (blk_queue_add_random(rq->q))
2368 add_disk_randomness(rq->rq_disk);
2369
2370 return false;
2371}
2372
2373/**
2374 * blk_unprep_request - unprepare a request
2375 * @req: the request
2376 *
2377 * This function makes a request ready for complete resubmission (or
2378 * completion). It happens only after all error handling is complete,
2379 * so represents the appropriate moment to deallocate any resources
2380 * that were allocated to the request in the prep_rq_fn. The queue
2381 * lock is held when calling this.
2382 */
2383void blk_unprep_request(struct request *req)
2384{
2385 struct request_queue *q = req->q;
2386
2387 req->cmd_flags &= ~REQ_DONTPREP;
2388 if (q->unprep_rq_fn)
2389 q->unprep_rq_fn(q, req);
2390}
2391EXPORT_SYMBOL_GPL(blk_unprep_request);
2392
2393/*
2394 * queue lock must be held
2395 */
2396static void blk_finish_request(struct request *req, int error)
2397{
2398 if (blk_rq_tagged(req))
2399 blk_queue_end_tag(req->q, req);
2400
2401 BUG_ON(blk_queued_rq(req));
2402
2403 if (unlikely(laptop_mode) && req->cmd_type == REQ_TYPE_FS)
2404 laptop_io_completion(&req->q->backing_dev_info);
2405
2406 blk_delete_timer(req);
2407
2408 if (req->cmd_flags & REQ_DONTPREP)
2409 blk_unprep_request(req);
2410
2411
2412 blk_account_io_done(req);
2413
2414 if (req->end_io)
2415 req->end_io(req, error);
2416 else {
2417 if (blk_bidi_rq(req))
2418 __blk_put_request(req->next_rq->q, req->next_rq);
2419
2420 __blk_put_request(req->q, req);
2421 }
2422}
2423
2424/**
2425 * blk_end_bidi_request - Complete a bidi request
2426 * @rq: the request to complete
2427 * @error: %0 for success, < %0 for error
2428 * @nr_bytes: number of bytes to complete @rq
2429 * @bidi_bytes: number of bytes to complete @rq->next_rq
2430 *
2431 * Description:
2432 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
2433 * Drivers that supports bidi can safely call this member for any
2434 * type of request, bidi or uni. In the later case @bidi_bytes is
2435 * just ignored.
2436 *
2437 * Return:
2438 * %false - we are done with this request
2439 * %true - still buffers pending for this request
2440 **/
2441static bool blk_end_bidi_request(struct request *rq, int error,
2442 unsigned int nr_bytes, unsigned int bidi_bytes)
2443{
2444 struct request_queue *q = rq->q;
2445 unsigned long flags;
2446
2447 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2448 return true;
2449
2450 spin_lock_irqsave(q->queue_lock, flags);
2451 blk_finish_request(rq, error);
2452 spin_unlock_irqrestore(q->queue_lock, flags);
2453
2454 return false;
2455}
2456
2457/**
2458 * __blk_end_bidi_request - Complete a bidi request with queue lock held
2459 * @rq: the request to complete
2460 * @error: %0 for success, < %0 for error
2461 * @nr_bytes: number of bytes to complete @rq
2462 * @bidi_bytes: number of bytes to complete @rq->next_rq
2463 *
2464 * Description:
2465 * Identical to blk_end_bidi_request() except that queue lock is
2466 * assumed to be locked on entry and remains so on return.
2467 *
2468 * Return:
2469 * %false - we are done with this request
2470 * %true - still buffers pending for this request
2471 **/
2472bool __blk_end_bidi_request(struct request *rq, int error,
2473 unsigned int nr_bytes, unsigned int bidi_bytes)
2474{
2475 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2476 return true;
2477
2478 blk_finish_request(rq, error);
2479
2480 return false;
2481}
2482
2483/**
2484 * blk_end_request - Helper function for drivers to complete the request.
2485 * @rq: the request being processed
2486 * @error: %0 for success, < %0 for error
2487 * @nr_bytes: number of bytes to complete
2488 *
2489 * Description:
2490 * Ends I/O on a number of bytes attached to @rq.
2491 * If @rq has leftover, sets it up for the next range of segments.
2492 *
2493 * Return:
2494 * %false - we are done with this request
2495 * %true - still buffers pending for this request
2496 **/
2497bool blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2498{
2499 return blk_end_bidi_request(rq, error, nr_bytes, 0);
2500}
2501EXPORT_SYMBOL(blk_end_request);
2502
2503/**
2504 * blk_end_request_all - Helper function for drives to finish the request.
2505 * @rq: the request to finish
2506 * @error: %0 for success, < %0 for error
2507 *
2508 * Description:
2509 * Completely finish @rq.
2510 */
2511void blk_end_request_all(struct request *rq, int error)
2512{
2513 bool pending;
2514 unsigned int bidi_bytes = 0;
2515
2516 if (unlikely(blk_bidi_rq(rq)))
2517 bidi_bytes = blk_rq_bytes(rq->next_rq);
2518
2519 pending = blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2520 BUG_ON(pending);
2521}
2522EXPORT_SYMBOL(blk_end_request_all);
2523
2524/**
2525 * blk_end_request_cur - Helper function to finish the current request chunk.
2526 * @rq: the request to finish the current chunk for
2527 * @error: %0 for success, < %0 for error
2528 *
2529 * Description:
2530 * Complete the current consecutively mapped chunk from @rq.
2531 *
2532 * Return:
2533 * %false - we are done with this request
2534 * %true - still buffers pending for this request
2535 */
2536bool blk_end_request_cur(struct request *rq, int error)
2537{
2538 return blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2539}
2540EXPORT_SYMBOL(blk_end_request_cur);
2541
2542/**
2543 * blk_end_request_err - Finish a request till the next failure boundary.
2544 * @rq: the request to finish till the next failure boundary for
2545 * @error: must be negative errno
2546 *
2547 * Description:
2548 * Complete @rq till the next failure boundary.
2549 *
2550 * Return:
2551 * %false - we are done with this request
2552 * %true - still buffers pending for this request
2553 */
2554bool blk_end_request_err(struct request *rq, int error)
2555{
2556 WARN_ON(error >= 0);
2557 return blk_end_request(rq, error, blk_rq_err_bytes(rq));
2558}
2559EXPORT_SYMBOL_GPL(blk_end_request_err);
2560
2561/**
2562 * __blk_end_request - Helper function for drivers to complete the request.
2563 * @rq: the request being processed
2564 * @error: %0 for success, < %0 for error
2565 * @nr_bytes: number of bytes to complete
2566 *
2567 * Description:
2568 * Must be called with queue lock held unlike blk_end_request().
2569 *
2570 * Return:
2571 * %false - we are done with this request
2572 * %true - still buffers pending for this request
2573 **/
2574bool __blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2575{
2576 return __blk_end_bidi_request(rq, error, nr_bytes, 0);
2577}
2578EXPORT_SYMBOL(__blk_end_request);
2579
2580/**
2581 * __blk_end_request_all - Helper function for drives to finish the request.
2582 * @rq: the request to finish
2583 * @error: %0 for success, < %0 for error
2584 *
2585 * Description:
2586 * Completely finish @rq. Must be called with queue lock held.
2587 */
2588void __blk_end_request_all(struct request *rq, int error)
2589{
2590 bool pending;
2591 unsigned int bidi_bytes = 0;
2592
2593 if (unlikely(blk_bidi_rq(rq)))
2594 bidi_bytes = blk_rq_bytes(rq->next_rq);
2595
2596 pending = __blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2597 BUG_ON(pending);
2598}
2599EXPORT_SYMBOL(__blk_end_request_all);
2600
2601/**
2602 * __blk_end_request_cur - Helper function to finish the current request chunk.
2603 * @rq: the request to finish the current chunk for
2604 * @error: %0 for success, < %0 for error
2605 *
2606 * Description:
2607 * Complete the current consecutively mapped chunk from @rq. Must
2608 * be called with queue lock held.
2609 *
2610 * Return:
2611 * %false - we are done with this request
2612 * %true - still buffers pending for this request
2613 */
2614bool __blk_end_request_cur(struct request *rq, int error)
2615{
2616 return __blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2617}
2618EXPORT_SYMBOL(__blk_end_request_cur);
2619
2620/**
2621 * __blk_end_request_err - Finish a request till the next failure boundary.
2622 * @rq: the request to finish till the next failure boundary for
2623 * @error: must be negative errno
2624 *
2625 * Description:
2626 * Complete @rq till the next failure boundary. Must be called
2627 * with queue lock held.
2628 *
2629 * Return:
2630 * %false - we are done with this request
2631 * %true - still buffers pending for this request
2632 */
2633bool __blk_end_request_err(struct request *rq, int error)
2634{
2635 WARN_ON(error >= 0);
2636 return __blk_end_request(rq, error, blk_rq_err_bytes(rq));
2637}
2638EXPORT_SYMBOL_GPL(__blk_end_request_err);
2639
2640void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
2641 struct bio *bio)
2642{
2643 /* Bit 0 (R/W) is identical in rq->cmd_flags and bio->bi_rw */
2644 rq->cmd_flags |= bio->bi_rw & REQ_WRITE;
2645
2646 if (bio_has_data(bio)) {
2647 rq->nr_phys_segments = bio_phys_segments(q, bio);
2648 rq->buffer = bio_data(bio);
2649 }
2650 rq->__data_len = bio->bi_size;
2651 rq->bio = rq->biotail = bio;
2652
2653 if (bio->bi_bdev)
2654 rq->rq_disk = bio->bi_bdev->bd_disk;
2655}
2656
2657#if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
2658/**
2659 * rq_flush_dcache_pages - Helper function to flush all pages in a request
2660 * @rq: the request to be flushed
2661 *
2662 * Description:
2663 * Flush all pages in @rq.
2664 */
2665void rq_flush_dcache_pages(struct request *rq)
2666{
2667 struct req_iterator iter;
2668 struct bio_vec *bvec;
2669
2670 rq_for_each_segment(bvec, rq, iter)
2671 flush_dcache_page(bvec->bv_page);
2672}
2673EXPORT_SYMBOL_GPL(rq_flush_dcache_pages);
2674#endif
2675
2676/**
2677 * blk_lld_busy - Check if underlying low-level drivers of a device are busy
2678 * @q : the queue of the device being checked
2679 *
2680 * Description:
2681 * Check if underlying low-level drivers of a device are busy.
2682 * If the drivers want to export their busy state, they must set own
2683 * exporting function using blk_queue_lld_busy() first.
2684 *
2685 * Basically, this function is used only by request stacking drivers
2686 * to stop dispatching requests to underlying devices when underlying
2687 * devices are busy. This behavior helps more I/O merging on the queue
2688 * of the request stacking driver and prevents I/O throughput regression
2689 * on burst I/O load.
2690 *
2691 * Return:
2692 * 0 - Not busy (The request stacking driver should dispatch request)
2693 * 1 - Busy (The request stacking driver should stop dispatching request)
2694 */
2695int blk_lld_busy(struct request_queue *q)
2696{
2697 if (q->lld_busy_fn)
2698 return q->lld_busy_fn(q);
2699
2700 return 0;
2701}
2702EXPORT_SYMBOL_GPL(blk_lld_busy);
2703
2704/**
2705 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2706 * @rq: the clone request to be cleaned up
2707 *
2708 * Description:
2709 * Free all bios in @rq for a cloned request.
2710 */
2711void blk_rq_unprep_clone(struct request *rq)
2712{
2713 struct bio *bio;
2714
2715 while ((bio = rq->bio) != NULL) {
2716 rq->bio = bio->bi_next;
2717
2718 bio_put(bio);
2719 }
2720}
2721EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2722
2723/*
2724 * Copy attributes of the original request to the clone request.
2725 * The actual data parts (e.g. ->cmd, ->buffer, ->sense) are not copied.
2726 */
2727static void __blk_rq_prep_clone(struct request *dst, struct request *src)
2728{
2729 dst->cpu = src->cpu;
2730 dst->cmd_flags = (src->cmd_flags & REQ_CLONE_MASK) | REQ_NOMERGE;
2731 dst->cmd_type = src->cmd_type;
2732 dst->__sector = blk_rq_pos(src);
2733 dst->__data_len = blk_rq_bytes(src);
2734 dst->nr_phys_segments = src->nr_phys_segments;
2735 dst->ioprio = src->ioprio;
2736 dst->extra_len = src->extra_len;
2737}
2738
2739/**
2740 * blk_rq_prep_clone - Helper function to setup clone request
2741 * @rq: the request to be setup
2742 * @rq_src: original request to be cloned
2743 * @bs: bio_set that bios for clone are allocated from
2744 * @gfp_mask: memory allocation mask for bio
2745 * @bio_ctr: setup function to be called for each clone bio.
2746 * Returns %0 for success, non %0 for failure.
2747 * @data: private data to be passed to @bio_ctr
2748 *
2749 * Description:
2750 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2751 * The actual data parts of @rq_src (e.g. ->cmd, ->buffer, ->sense)
2752 * are not copied, and copying such parts is the caller's responsibility.
2753 * Also, pages which the original bios are pointing to are not copied
2754 * and the cloned bios just point same pages.
2755 * So cloned bios must be completed before original bios, which means
2756 * the caller must complete @rq before @rq_src.
2757 */
2758int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2759 struct bio_set *bs, gfp_t gfp_mask,
2760 int (*bio_ctr)(struct bio *, struct bio *, void *),
2761 void *data)
2762{
2763 struct bio *bio, *bio_src;
2764
2765 if (!bs)
2766 bs = fs_bio_set;
2767
2768 blk_rq_init(NULL, rq);
2769
2770 __rq_for_each_bio(bio_src, rq_src) {
2771 bio = bio_alloc_bioset(gfp_mask, bio_src->bi_max_vecs, bs);
2772 if (!bio)
2773 goto free_and_out;
2774
2775 __bio_clone(bio, bio_src);
2776
2777 if (bio_integrity(bio_src) &&
2778 bio_integrity_clone(bio, bio_src, gfp_mask, bs))
2779 goto free_and_out;
2780
2781 if (bio_ctr && bio_ctr(bio, bio_src, data))
2782 goto free_and_out;
2783
2784 if (rq->bio) {
2785 rq->biotail->bi_next = bio;
2786 rq->biotail = bio;
2787 } else
2788 rq->bio = rq->biotail = bio;
2789 }
2790
2791 __blk_rq_prep_clone(rq, rq_src);
2792
2793 return 0;
2794
2795free_and_out:
2796 if (bio)
2797 bio_free(bio, bs);
2798 blk_rq_unprep_clone(rq);
2799
2800 return -ENOMEM;
2801}
2802EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
2803
2804int kblockd_schedule_work(struct request_queue *q, struct work_struct *work)
2805{
2806 return queue_work(kblockd_workqueue, work);
2807}
2808EXPORT_SYMBOL(kblockd_schedule_work);
2809
2810int kblockd_schedule_delayed_work(struct request_queue *q,
2811 struct delayed_work *dwork, unsigned long delay)
2812{
2813 return queue_delayed_work(kblockd_workqueue, dwork, delay);
2814}
2815EXPORT_SYMBOL(kblockd_schedule_delayed_work);
2816
2817#define PLUG_MAGIC 0x91827364
2818
2819/**
2820 * blk_start_plug - initialize blk_plug and track it inside the task_struct
2821 * @plug: The &struct blk_plug that needs to be initialized
2822 *
2823 * Description:
2824 * Tracking blk_plug inside the task_struct will help with auto-flushing the
2825 * pending I/O should the task end up blocking between blk_start_plug() and
2826 * blk_finish_plug(). This is important from a performance perspective, but
2827 * also ensures that we don't deadlock. For instance, if the task is blocking
2828 * for a memory allocation, memory reclaim could end up wanting to free a
2829 * page belonging to that request that is currently residing in our private
2830 * plug. By flushing the pending I/O when the process goes to sleep, we avoid
2831 * this kind of deadlock.
2832 */
2833void blk_start_plug(struct blk_plug *plug)
2834{
2835 struct task_struct *tsk = current;
2836
2837 plug->magic = PLUG_MAGIC;
2838 INIT_LIST_HEAD(&plug->list);
2839 INIT_LIST_HEAD(&plug->cb_list);
2840 plug->should_sort = 0;
2841
2842 /*
2843 * If this is a nested plug, don't actually assign it. It will be
2844 * flushed on its own.
2845 */
2846 if (!tsk->plug) {
2847 /*
2848 * Store ordering should not be needed here, since a potential
2849 * preempt will imply a full memory barrier
2850 */
2851 tsk->plug = plug;
2852 }
2853}
2854EXPORT_SYMBOL(blk_start_plug);
2855
2856static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
2857{
2858 struct request *rqa = container_of(a, struct request, queuelist);
2859 struct request *rqb = container_of(b, struct request, queuelist);
2860
2861 return !(rqa->q <= rqb->q);
2862}
2863
2864/*
2865 * If 'from_schedule' is true, then postpone the dispatch of requests
2866 * until a safe kblockd context. We due this to avoid accidental big
2867 * additional stack usage in driver dispatch, in places where the originally
2868 * plugger did not intend it.
2869 */
2870static void queue_unplugged(struct request_queue *q, unsigned int depth,
2871 bool from_schedule)
2872 __releases(q->queue_lock)
2873{
2874 trace_block_unplug(q, depth, !from_schedule);
2875
2876 /*
2877 * Don't mess with dead queue.
2878 */
2879 if (unlikely(blk_queue_dead(q))) {
2880 spin_unlock(q->queue_lock);
2881 return;
2882 }
2883
2884 /*
2885 * If we are punting this to kblockd, then we can safely drop
2886 * the queue_lock before waking kblockd (which needs to take
2887 * this lock).
2888 */
2889 if (from_schedule) {
2890 spin_unlock(q->queue_lock);
2891 blk_run_queue_async(q);
2892 } else {
2893 __blk_run_queue(q);
2894 spin_unlock(q->queue_lock);
2895 }
2896
2897}
2898
2899static void flush_plug_callbacks(struct blk_plug *plug)
2900{
2901 LIST_HEAD(callbacks);
2902
2903 if (list_empty(&plug->cb_list))
2904 return;
2905
2906 list_splice_init(&plug->cb_list, &callbacks);
2907
2908 while (!list_empty(&callbacks)) {
2909 struct blk_plug_cb *cb = list_first_entry(&callbacks,
2910 struct blk_plug_cb,
2911 list);
2912 list_del(&cb->list);
2913 cb->callback(cb);
2914 }
2915}
2916
2917void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2918{
2919 struct request_queue *q;
2920 unsigned long flags;
2921 struct request *rq;
2922 LIST_HEAD(list);
2923 unsigned int depth;
2924
2925 BUG_ON(plug->magic != PLUG_MAGIC);
2926
2927 flush_plug_callbacks(plug);
2928 if (list_empty(&plug->list))
2929 return;
2930
2931 list_splice_init(&plug->list, &list);
2932
2933 if (plug->should_sort) {
2934 list_sort(NULL, &list, plug_rq_cmp);
2935 plug->should_sort = 0;
2936 }
2937
2938 q = NULL;
2939 depth = 0;
2940
2941 /*
2942 * Save and disable interrupts here, to avoid doing it for every
2943 * queue lock we have to take.
2944 */
2945 local_irq_save(flags);
2946 while (!list_empty(&list)) {
2947 rq = list_entry_rq(list.next);
2948 list_del_init(&rq->queuelist);
2949 BUG_ON(!rq->q);
2950 if (rq->q != q) {
2951 /*
2952 * This drops the queue lock
2953 */
2954 if (q)
2955 queue_unplugged(q, depth, from_schedule);
2956 q = rq->q;
2957 depth = 0;
2958 spin_lock(q->queue_lock);
2959 }
2960
2961 /*
2962 * Short-circuit if @q is dead
2963 */
2964 if (unlikely(blk_queue_dead(q))) {
2965 __blk_end_request_all(rq, -ENODEV);
2966 continue;
2967 }
2968
2969 /*
2970 * rq is already accounted, so use raw insert
2971 */
2972 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA))
2973 __elv_add_request(q, rq, ELEVATOR_INSERT_FLUSH);
2974 else
2975 __elv_add_request(q, rq, ELEVATOR_INSERT_SORT_MERGE);
2976
2977 depth++;
2978 }
2979
2980 /*
2981 * This drops the queue lock
2982 */
2983 if (q)
2984 queue_unplugged(q, depth, from_schedule);
2985
2986 local_irq_restore(flags);
2987}
2988
2989void blk_finish_plug(struct blk_plug *plug)
2990{
2991 blk_flush_plug_list(plug, false);
2992
2993 if (plug == current->plug)
2994 current->plug = NULL;
2995}
2996EXPORT_SYMBOL(blk_finish_plug);
2997
2998int __init blk_dev_init(void)
2999{
3000 BUILD_BUG_ON(__REQ_NR_BITS > 8 *
3001 sizeof(((struct request *)0)->cmd_flags));
3002
3003 /* used for unplugging and affects IO latency/throughput - HIGHPRI */
3004 kblockd_workqueue = alloc_workqueue("kblockd",
3005 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
3006 if (!kblockd_workqueue)
3007 panic("Failed to create kblockd\n");
3008
3009 request_cachep = kmem_cache_create("blkdev_requests",
3010 sizeof(struct request), 0, SLAB_PANIC, NULL);
3011
3012 blk_requestq_cachep = kmem_cache_create("blkdev_queue",
3013 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3014
3015 return 0;
3016}
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/highmem.h>
20#include <linux/mm.h>
21#include <linux/kernel_stat.h>
22#include <linux/string.h>
23#include <linux/init.h>
24#include <linux/completion.h>
25#include <linux/slab.h>
26#include <linux/swap.h>
27#include <linux/writeback.h>
28#include <linux/task_io_accounting_ops.h>
29#include <linux/fault-inject.h>
30#include <linux/list_sort.h>
31
32#define CREATE_TRACE_POINTS
33#include <trace/events/block.h>
34
35#include "blk.h"
36
37EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_remap);
38EXPORT_TRACEPOINT_SYMBOL_GPL(block_rq_remap);
39EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_complete);
40
41static int __make_request(struct request_queue *q, struct bio *bio);
42
43/*
44 * For the allocated request tables
45 */
46static struct kmem_cache *request_cachep;
47
48/*
49 * For queue allocation
50 */
51struct kmem_cache *blk_requestq_cachep;
52
53/*
54 * Controlling structure to kblockd
55 */
56static struct workqueue_struct *kblockd_workqueue;
57
58static void drive_stat_acct(struct request *rq, int new_io)
59{
60 struct hd_struct *part;
61 int rw = rq_data_dir(rq);
62 int cpu;
63
64 if (!blk_do_io_stat(rq))
65 return;
66
67 cpu = part_stat_lock();
68
69 if (!new_io) {
70 part = rq->part;
71 part_stat_inc(cpu, part, merges[rw]);
72 } else {
73 part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
74 if (!hd_struct_try_get(part)) {
75 /*
76 * The partition is already being removed,
77 * the request will be accounted on the disk only
78 *
79 * We take a reference on disk->part0 although that
80 * partition will never be deleted, so we can treat
81 * it as any other partition.
82 */
83 part = &rq->rq_disk->part0;
84 hd_struct_get(part);
85 }
86 part_round_stats(cpu, part);
87 part_inc_in_flight(part, rw);
88 rq->part = part;
89 }
90
91 part_stat_unlock();
92}
93
94void blk_queue_congestion_threshold(struct request_queue *q)
95{
96 int nr;
97
98 nr = q->nr_requests - (q->nr_requests / 8) + 1;
99 if (nr > q->nr_requests)
100 nr = q->nr_requests;
101 q->nr_congestion_on = nr;
102
103 nr = q->nr_requests - (q->nr_requests / 8) - (q->nr_requests / 16) - 1;
104 if (nr < 1)
105 nr = 1;
106 q->nr_congestion_off = nr;
107}
108
109/**
110 * blk_get_backing_dev_info - get the address of a queue's backing_dev_info
111 * @bdev: device
112 *
113 * Locates the passed device's request queue and returns the address of its
114 * backing_dev_info
115 *
116 * Will return NULL if the request queue cannot be located.
117 */
118struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev)
119{
120 struct backing_dev_info *ret = NULL;
121 struct request_queue *q = bdev_get_queue(bdev);
122
123 if (q)
124 ret = &q->backing_dev_info;
125 return ret;
126}
127EXPORT_SYMBOL(blk_get_backing_dev_info);
128
129void blk_rq_init(struct request_queue *q, struct request *rq)
130{
131 memset(rq, 0, sizeof(*rq));
132
133 INIT_LIST_HEAD(&rq->queuelist);
134 INIT_LIST_HEAD(&rq->timeout_list);
135 rq->cpu = -1;
136 rq->q = q;
137 rq->__sector = (sector_t) -1;
138 INIT_HLIST_NODE(&rq->hash);
139 RB_CLEAR_NODE(&rq->rb_node);
140 rq->cmd = rq->__cmd;
141 rq->cmd_len = BLK_MAX_CDB;
142 rq->tag = -1;
143 rq->ref_count = 1;
144 rq->start_time = jiffies;
145 set_start_time_ns(rq);
146 rq->part = NULL;
147}
148EXPORT_SYMBOL(blk_rq_init);
149
150static void req_bio_endio(struct request *rq, struct bio *bio,
151 unsigned int nbytes, int error)
152{
153 if (error)
154 clear_bit(BIO_UPTODATE, &bio->bi_flags);
155 else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
156 error = -EIO;
157
158 if (unlikely(nbytes > bio->bi_size)) {
159 printk(KERN_ERR "%s: want %u bytes done, %u left\n",
160 __func__, nbytes, bio->bi_size);
161 nbytes = bio->bi_size;
162 }
163
164 if (unlikely(rq->cmd_flags & REQ_QUIET))
165 set_bit(BIO_QUIET, &bio->bi_flags);
166
167 bio->bi_size -= nbytes;
168 bio->bi_sector += (nbytes >> 9);
169
170 if (bio_integrity(bio))
171 bio_integrity_advance(bio, nbytes);
172
173 /* don't actually finish bio if it's part of flush sequence */
174 if (bio->bi_size == 0 && !(rq->cmd_flags & REQ_FLUSH_SEQ))
175 bio_endio(bio, error);
176}
177
178void blk_dump_rq_flags(struct request *rq, char *msg)
179{
180 int bit;
181
182 printk(KERN_INFO "%s: dev %s: type=%x, flags=%x\n", msg,
183 rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->cmd_type,
184 rq->cmd_flags);
185
186 printk(KERN_INFO " sector %llu, nr/cnr %u/%u\n",
187 (unsigned long long)blk_rq_pos(rq),
188 blk_rq_sectors(rq), blk_rq_cur_sectors(rq));
189 printk(KERN_INFO " bio %p, biotail %p, buffer %p, len %u\n",
190 rq->bio, rq->biotail, rq->buffer, blk_rq_bytes(rq));
191
192 if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
193 printk(KERN_INFO " cdb: ");
194 for (bit = 0; bit < BLK_MAX_CDB; bit++)
195 printk("%02x ", rq->cmd[bit]);
196 printk("\n");
197 }
198}
199EXPORT_SYMBOL(blk_dump_rq_flags);
200
201static void blk_delay_work(struct work_struct *work)
202{
203 struct request_queue *q;
204
205 q = container_of(work, struct request_queue, delay_work.work);
206 spin_lock_irq(q->queue_lock);
207 __blk_run_queue(q);
208 spin_unlock_irq(q->queue_lock);
209}
210
211/**
212 * blk_delay_queue - restart queueing after defined interval
213 * @q: The &struct request_queue in question
214 * @msecs: Delay in msecs
215 *
216 * Description:
217 * Sometimes queueing needs to be postponed for a little while, to allow
218 * resources to come back. This function will make sure that queueing is
219 * restarted around the specified time.
220 */
221void blk_delay_queue(struct request_queue *q, unsigned long msecs)
222{
223 queue_delayed_work(kblockd_workqueue, &q->delay_work,
224 msecs_to_jiffies(msecs));
225}
226EXPORT_SYMBOL(blk_delay_queue);
227
228/**
229 * blk_start_queue - restart a previously stopped queue
230 * @q: The &struct request_queue in question
231 *
232 * Description:
233 * blk_start_queue() will clear the stop flag on the queue, and call
234 * the request_fn for the queue if it was in a stopped state when
235 * entered. Also see blk_stop_queue(). Queue lock must be held.
236 **/
237void blk_start_queue(struct request_queue *q)
238{
239 WARN_ON(!irqs_disabled());
240
241 queue_flag_clear(QUEUE_FLAG_STOPPED, q);
242 __blk_run_queue(q);
243}
244EXPORT_SYMBOL(blk_start_queue);
245
246/**
247 * blk_stop_queue - stop a queue
248 * @q: The &struct request_queue in question
249 *
250 * Description:
251 * The Linux block layer assumes that a block driver will consume all
252 * entries on the request queue when the request_fn strategy is called.
253 * Often this will not happen, because of hardware limitations (queue
254 * depth settings). If a device driver gets a 'queue full' response,
255 * or if it simply chooses not to queue more I/O at one point, it can
256 * call this function to prevent the request_fn from being called until
257 * the driver has signalled it's ready to go again. This happens by calling
258 * blk_start_queue() to restart queue operations. Queue lock must be held.
259 **/
260void blk_stop_queue(struct request_queue *q)
261{
262 __cancel_delayed_work(&q->delay_work);
263 queue_flag_set(QUEUE_FLAG_STOPPED, q);
264}
265EXPORT_SYMBOL(blk_stop_queue);
266
267/**
268 * blk_sync_queue - cancel any pending callbacks on a queue
269 * @q: the queue
270 *
271 * Description:
272 * The block layer may perform asynchronous callback activity
273 * on a queue, such as calling the unplug function after a timeout.
274 * A block device may call blk_sync_queue to ensure that any
275 * such activity is cancelled, thus allowing it to release resources
276 * that the callbacks might use. The caller must already have made sure
277 * that its ->make_request_fn will not re-add plugging prior to calling
278 * this function.
279 *
280 * This function does not cancel any asynchronous activity arising
281 * out of elevator or throttling code. That would require elevaotor_exit()
282 * and blk_throtl_exit() to be called with queue lock initialized.
283 *
284 */
285void blk_sync_queue(struct request_queue *q)
286{
287 del_timer_sync(&q->timeout);
288 cancel_delayed_work_sync(&q->delay_work);
289}
290EXPORT_SYMBOL(blk_sync_queue);
291
292/**
293 * __blk_run_queue - run a single device queue
294 * @q: The queue to run
295 *
296 * Description:
297 * See @blk_run_queue. This variant must be called with the queue lock
298 * held and interrupts disabled.
299 */
300void __blk_run_queue(struct request_queue *q)
301{
302 if (unlikely(blk_queue_stopped(q)))
303 return;
304
305 q->request_fn(q);
306}
307EXPORT_SYMBOL(__blk_run_queue);
308
309/**
310 * blk_run_queue_async - run a single device queue in workqueue context
311 * @q: The queue to run
312 *
313 * Description:
314 * Tells kblockd to perform the equivalent of @blk_run_queue on behalf
315 * of us.
316 */
317void blk_run_queue_async(struct request_queue *q)
318{
319 if (likely(!blk_queue_stopped(q))) {
320 __cancel_delayed_work(&q->delay_work);
321 queue_delayed_work(kblockd_workqueue, &q->delay_work, 0);
322 }
323}
324EXPORT_SYMBOL(blk_run_queue_async);
325
326/**
327 * blk_run_queue - run a single device queue
328 * @q: The queue to run
329 *
330 * Description:
331 * Invoke request handling on this queue, if it has pending work to do.
332 * May be used to restart queueing when a request has completed.
333 */
334void blk_run_queue(struct request_queue *q)
335{
336 unsigned long flags;
337
338 spin_lock_irqsave(q->queue_lock, flags);
339 __blk_run_queue(q);
340 spin_unlock_irqrestore(q->queue_lock, flags);
341}
342EXPORT_SYMBOL(blk_run_queue);
343
344void blk_put_queue(struct request_queue *q)
345{
346 kobject_put(&q->kobj);
347}
348EXPORT_SYMBOL(blk_put_queue);
349
350/*
351 * Note: If a driver supplied the queue lock, it is disconnected
352 * by this function. The actual state of the lock doesn't matter
353 * here as the request_queue isn't accessible after this point
354 * (QUEUE_FLAG_DEAD is set) and no other requests will be queued.
355 */
356void blk_cleanup_queue(struct request_queue *q)
357{
358 /*
359 * We know we have process context here, so we can be a little
360 * cautious and ensure that pending block actions on this device
361 * are done before moving on. Going into this function, we should
362 * not have processes doing IO to this device.
363 */
364 blk_sync_queue(q);
365
366 del_timer_sync(&q->backing_dev_info.laptop_mode_wb_timer);
367 mutex_lock(&q->sysfs_lock);
368 queue_flag_set_unlocked(QUEUE_FLAG_DEAD, q);
369 mutex_unlock(&q->sysfs_lock);
370
371 if (q->queue_lock != &q->__queue_lock)
372 q->queue_lock = &q->__queue_lock;
373
374 blk_put_queue(q);
375}
376EXPORT_SYMBOL(blk_cleanup_queue);
377
378static int blk_init_free_list(struct request_queue *q)
379{
380 struct request_list *rl = &q->rq;
381
382 if (unlikely(rl->rq_pool))
383 return 0;
384
385 rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
386 rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
387 rl->elvpriv = 0;
388 init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
389 init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
390
391 rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
392 mempool_free_slab, request_cachep, q->node);
393
394 if (!rl->rq_pool)
395 return -ENOMEM;
396
397 return 0;
398}
399
400struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
401{
402 return blk_alloc_queue_node(gfp_mask, -1);
403}
404EXPORT_SYMBOL(blk_alloc_queue);
405
406struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
407{
408 struct request_queue *q;
409 int err;
410
411 q = kmem_cache_alloc_node(blk_requestq_cachep,
412 gfp_mask | __GFP_ZERO, node_id);
413 if (!q)
414 return NULL;
415
416 q->backing_dev_info.ra_pages =
417 (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
418 q->backing_dev_info.state = 0;
419 q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
420 q->backing_dev_info.name = "block";
421
422 err = bdi_init(&q->backing_dev_info);
423 if (err) {
424 kmem_cache_free(blk_requestq_cachep, q);
425 return NULL;
426 }
427
428 if (blk_throtl_init(q)) {
429 kmem_cache_free(blk_requestq_cachep, q);
430 return NULL;
431 }
432
433 setup_timer(&q->backing_dev_info.laptop_mode_wb_timer,
434 laptop_mode_timer_fn, (unsigned long) q);
435 setup_timer(&q->timeout, blk_rq_timed_out_timer, (unsigned long) q);
436 INIT_LIST_HEAD(&q->timeout_list);
437 INIT_LIST_HEAD(&q->flush_queue[0]);
438 INIT_LIST_HEAD(&q->flush_queue[1]);
439 INIT_LIST_HEAD(&q->flush_data_in_flight);
440 INIT_DELAYED_WORK(&q->delay_work, blk_delay_work);
441
442 kobject_init(&q->kobj, &blk_queue_ktype);
443
444 mutex_init(&q->sysfs_lock);
445 spin_lock_init(&q->__queue_lock);
446
447 /*
448 * By default initialize queue_lock to internal lock and driver can
449 * override it later if need be.
450 */
451 q->queue_lock = &q->__queue_lock;
452
453 return q;
454}
455EXPORT_SYMBOL(blk_alloc_queue_node);
456
457/**
458 * blk_init_queue - prepare a request queue for use with a block device
459 * @rfn: The function to be called to process requests that have been
460 * placed on the queue.
461 * @lock: Request queue spin lock
462 *
463 * Description:
464 * If a block device wishes to use the standard request handling procedures,
465 * which sorts requests and coalesces adjacent requests, then it must
466 * call blk_init_queue(). The function @rfn will be called when there
467 * are requests on the queue that need to be processed. If the device
468 * supports plugging, then @rfn may not be called immediately when requests
469 * are available on the queue, but may be called at some time later instead.
470 * Plugged queues are generally unplugged when a buffer belonging to one
471 * of the requests on the queue is needed, or due to memory pressure.
472 *
473 * @rfn is not required, or even expected, to remove all requests off the
474 * queue, but only as many as it can handle at a time. If it does leave
475 * requests on the queue, it is responsible for arranging that the requests
476 * get dealt with eventually.
477 *
478 * The queue spin lock must be held while manipulating the requests on the
479 * request queue; this lock will be taken also from interrupt context, so irq
480 * disabling is needed for it.
481 *
482 * Function returns a pointer to the initialized request queue, or %NULL if
483 * it didn't succeed.
484 *
485 * Note:
486 * blk_init_queue() must be paired with a blk_cleanup_queue() call
487 * when the block device is deactivated (such as at module unload).
488 **/
489
490struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
491{
492 return blk_init_queue_node(rfn, lock, -1);
493}
494EXPORT_SYMBOL(blk_init_queue);
495
496struct request_queue *
497blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
498{
499 struct request_queue *uninit_q, *q;
500
501 uninit_q = blk_alloc_queue_node(GFP_KERNEL, node_id);
502 if (!uninit_q)
503 return NULL;
504
505 q = blk_init_allocated_queue_node(uninit_q, rfn, lock, node_id);
506 if (!q)
507 blk_cleanup_queue(uninit_q);
508
509 return q;
510}
511EXPORT_SYMBOL(blk_init_queue_node);
512
513struct request_queue *
514blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
515 spinlock_t *lock)
516{
517 return blk_init_allocated_queue_node(q, rfn, lock, -1);
518}
519EXPORT_SYMBOL(blk_init_allocated_queue);
520
521struct request_queue *
522blk_init_allocated_queue_node(struct request_queue *q, request_fn_proc *rfn,
523 spinlock_t *lock, int node_id)
524{
525 if (!q)
526 return NULL;
527
528 q->node = node_id;
529 if (blk_init_free_list(q))
530 return NULL;
531
532 q->request_fn = rfn;
533 q->prep_rq_fn = NULL;
534 q->unprep_rq_fn = NULL;
535 q->queue_flags = QUEUE_FLAG_DEFAULT;
536
537 /* Override internal queue lock with supplied lock pointer */
538 if (lock)
539 q->queue_lock = lock;
540
541 /*
542 * This also sets hw/phys segments, boundary and size
543 */
544 blk_queue_make_request(q, __make_request);
545
546 q->sg_reserved_size = INT_MAX;
547
548 /*
549 * all done
550 */
551 if (!elevator_init(q, NULL)) {
552 blk_queue_congestion_threshold(q);
553 return q;
554 }
555
556 return NULL;
557}
558EXPORT_SYMBOL(blk_init_allocated_queue_node);
559
560int blk_get_queue(struct request_queue *q)
561{
562 if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
563 kobject_get(&q->kobj);
564 return 0;
565 }
566
567 return 1;
568}
569EXPORT_SYMBOL(blk_get_queue);
570
571static inline void blk_free_request(struct request_queue *q, struct request *rq)
572{
573 if (rq->cmd_flags & REQ_ELVPRIV)
574 elv_put_request(q, rq);
575 mempool_free(rq, q->rq.rq_pool);
576}
577
578static struct request *
579blk_alloc_request(struct request_queue *q, int flags, int priv, gfp_t gfp_mask)
580{
581 struct request *rq = mempool_alloc(q->rq.rq_pool, gfp_mask);
582
583 if (!rq)
584 return NULL;
585
586 blk_rq_init(q, rq);
587
588 rq->cmd_flags = flags | REQ_ALLOCED;
589
590 if (priv) {
591 if (unlikely(elv_set_request(q, rq, gfp_mask))) {
592 mempool_free(rq, q->rq.rq_pool);
593 return NULL;
594 }
595 rq->cmd_flags |= REQ_ELVPRIV;
596 }
597
598 return rq;
599}
600
601/*
602 * ioc_batching returns true if the ioc is a valid batching request and
603 * should be given priority access to a request.
604 */
605static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
606{
607 if (!ioc)
608 return 0;
609
610 /*
611 * Make sure the process is able to allocate at least 1 request
612 * even if the batch times out, otherwise we could theoretically
613 * lose wakeups.
614 */
615 return ioc->nr_batch_requests == q->nr_batching ||
616 (ioc->nr_batch_requests > 0
617 && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
618}
619
620/*
621 * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
622 * will cause the process to be a "batcher" on all queues in the system. This
623 * is the behaviour we want though - once it gets a wakeup it should be given
624 * a nice run.
625 */
626static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
627{
628 if (!ioc || ioc_batching(q, ioc))
629 return;
630
631 ioc->nr_batch_requests = q->nr_batching;
632 ioc->last_waited = jiffies;
633}
634
635static void __freed_request(struct request_queue *q, int sync)
636{
637 struct request_list *rl = &q->rq;
638
639 if (rl->count[sync] < queue_congestion_off_threshold(q))
640 blk_clear_queue_congested(q, sync);
641
642 if (rl->count[sync] + 1 <= q->nr_requests) {
643 if (waitqueue_active(&rl->wait[sync]))
644 wake_up(&rl->wait[sync]);
645
646 blk_clear_queue_full(q, sync);
647 }
648}
649
650/*
651 * A request has just been released. Account for it, update the full and
652 * congestion status, wake up any waiters. Called under q->queue_lock.
653 */
654static void freed_request(struct request_queue *q, int sync, int priv)
655{
656 struct request_list *rl = &q->rq;
657
658 rl->count[sync]--;
659 if (priv)
660 rl->elvpriv--;
661
662 __freed_request(q, sync);
663
664 if (unlikely(rl->starved[sync ^ 1]))
665 __freed_request(q, sync ^ 1);
666}
667
668/*
669 * Determine if elevator data should be initialized when allocating the
670 * request associated with @bio.
671 */
672static bool blk_rq_should_init_elevator(struct bio *bio)
673{
674 if (!bio)
675 return true;
676
677 /*
678 * Flush requests do not use the elevator so skip initialization.
679 * This allows a request to share the flush and elevator data.
680 */
681 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA))
682 return false;
683
684 return true;
685}
686
687/*
688 * Get a free request, queue_lock must be held.
689 * Returns NULL on failure, with queue_lock held.
690 * Returns !NULL on success, with queue_lock *not held*.
691 */
692static struct request *get_request(struct request_queue *q, int rw_flags,
693 struct bio *bio, gfp_t gfp_mask)
694{
695 struct request *rq = NULL;
696 struct request_list *rl = &q->rq;
697 struct io_context *ioc = NULL;
698 const bool is_sync = rw_is_sync(rw_flags) != 0;
699 int may_queue, priv = 0;
700
701 may_queue = elv_may_queue(q, rw_flags);
702 if (may_queue == ELV_MQUEUE_NO)
703 goto rq_starved;
704
705 if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) {
706 if (rl->count[is_sync]+1 >= q->nr_requests) {
707 ioc = current_io_context(GFP_ATOMIC, q->node);
708 /*
709 * The queue will fill after this allocation, so set
710 * it as full, and mark this process as "batching".
711 * This process will be allowed to complete a batch of
712 * requests, others will be blocked.
713 */
714 if (!blk_queue_full(q, is_sync)) {
715 ioc_set_batching(q, ioc);
716 blk_set_queue_full(q, is_sync);
717 } else {
718 if (may_queue != ELV_MQUEUE_MUST
719 && !ioc_batching(q, ioc)) {
720 /*
721 * The queue is full and the allocating
722 * process is not a "batcher", and not
723 * exempted by the IO scheduler
724 */
725 goto out;
726 }
727 }
728 }
729 blk_set_queue_congested(q, is_sync);
730 }
731
732 /*
733 * Only allow batching queuers to allocate up to 50% over the defined
734 * limit of requests, otherwise we could have thousands of requests
735 * allocated with any setting of ->nr_requests
736 */
737 if (rl->count[is_sync] >= (3 * q->nr_requests / 2))
738 goto out;
739
740 rl->count[is_sync]++;
741 rl->starved[is_sync] = 0;
742
743 if (blk_rq_should_init_elevator(bio)) {
744 priv = !test_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
745 if (priv)
746 rl->elvpriv++;
747 }
748
749 if (blk_queue_io_stat(q))
750 rw_flags |= REQ_IO_STAT;
751 spin_unlock_irq(q->queue_lock);
752
753 rq = blk_alloc_request(q, rw_flags, priv, gfp_mask);
754 if (unlikely(!rq)) {
755 /*
756 * Allocation failed presumably due to memory. Undo anything
757 * we might have messed up.
758 *
759 * Allocating task should really be put onto the front of the
760 * wait queue, but this is pretty rare.
761 */
762 spin_lock_irq(q->queue_lock);
763 freed_request(q, is_sync, priv);
764
765 /*
766 * in the very unlikely event that allocation failed and no
767 * requests for this direction was pending, mark us starved
768 * so that freeing of a request in the other direction will
769 * notice us. another possible fix would be to split the
770 * rq mempool into READ and WRITE
771 */
772rq_starved:
773 if (unlikely(rl->count[is_sync] == 0))
774 rl->starved[is_sync] = 1;
775
776 goto out;
777 }
778
779 /*
780 * ioc may be NULL here, and ioc_batching will be false. That's
781 * OK, if the queue is under the request limit then requests need
782 * not count toward the nr_batch_requests limit. There will always
783 * be some limit enforced by BLK_BATCH_TIME.
784 */
785 if (ioc_batching(q, ioc))
786 ioc->nr_batch_requests--;
787
788 trace_block_getrq(q, bio, rw_flags & 1);
789out:
790 return rq;
791}
792
793/*
794 * No available requests for this queue, wait for some requests to become
795 * available.
796 *
797 * Called with q->queue_lock held, and returns with it unlocked.
798 */
799static struct request *get_request_wait(struct request_queue *q, int rw_flags,
800 struct bio *bio)
801{
802 const bool is_sync = rw_is_sync(rw_flags) != 0;
803 struct request *rq;
804
805 rq = get_request(q, rw_flags, bio, GFP_NOIO);
806 while (!rq) {
807 DEFINE_WAIT(wait);
808 struct io_context *ioc;
809 struct request_list *rl = &q->rq;
810
811 prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
812 TASK_UNINTERRUPTIBLE);
813
814 trace_block_sleeprq(q, bio, rw_flags & 1);
815
816 spin_unlock_irq(q->queue_lock);
817 io_schedule();
818
819 /*
820 * After sleeping, we become a "batching" process and
821 * will be able to allocate at least one request, and
822 * up to a big batch of them for a small period time.
823 * See ioc_batching, ioc_set_batching
824 */
825 ioc = current_io_context(GFP_NOIO, q->node);
826 ioc_set_batching(q, ioc);
827
828 spin_lock_irq(q->queue_lock);
829 finish_wait(&rl->wait[is_sync], &wait);
830
831 rq = get_request(q, rw_flags, bio, GFP_NOIO);
832 };
833
834 return rq;
835}
836
837struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
838{
839 struct request *rq;
840
841 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
842 return NULL;
843
844 BUG_ON(rw != READ && rw != WRITE);
845
846 spin_lock_irq(q->queue_lock);
847 if (gfp_mask & __GFP_WAIT) {
848 rq = get_request_wait(q, rw, NULL);
849 } else {
850 rq = get_request(q, rw, NULL, gfp_mask);
851 if (!rq)
852 spin_unlock_irq(q->queue_lock);
853 }
854 /* q->queue_lock is unlocked at this point */
855
856 return rq;
857}
858EXPORT_SYMBOL(blk_get_request);
859
860/**
861 * blk_make_request - given a bio, allocate a corresponding struct request.
862 * @q: target request queue
863 * @bio: The bio describing the memory mappings that will be submitted for IO.
864 * It may be a chained-bio properly constructed by block/bio layer.
865 * @gfp_mask: gfp flags to be used for memory allocation
866 *
867 * blk_make_request is the parallel of generic_make_request for BLOCK_PC
868 * type commands. Where the struct request needs to be farther initialized by
869 * the caller. It is passed a &struct bio, which describes the memory info of
870 * the I/O transfer.
871 *
872 * The caller of blk_make_request must make sure that bi_io_vec
873 * are set to describe the memory buffers. That bio_data_dir() will return
874 * the needed direction of the request. (And all bio's in the passed bio-chain
875 * are properly set accordingly)
876 *
877 * If called under none-sleepable conditions, mapped bio buffers must not
878 * need bouncing, by calling the appropriate masked or flagged allocator,
879 * suitable for the target device. Otherwise the call to blk_queue_bounce will
880 * BUG.
881 *
882 * WARNING: When allocating/cloning a bio-chain, careful consideration should be
883 * given to how you allocate bios. In particular, you cannot use __GFP_WAIT for
884 * anything but the first bio in the chain. Otherwise you risk waiting for IO
885 * completion of a bio that hasn't been submitted yet, thus resulting in a
886 * deadlock. Alternatively bios should be allocated using bio_kmalloc() instead
887 * of bio_alloc(), as that avoids the mempool deadlock.
888 * If possible a big IO should be split into smaller parts when allocation
889 * fails. Partial allocation should not be an error, or you risk a live-lock.
890 */
891struct request *blk_make_request(struct request_queue *q, struct bio *bio,
892 gfp_t gfp_mask)
893{
894 struct request *rq = blk_get_request(q, bio_data_dir(bio), gfp_mask);
895
896 if (unlikely(!rq))
897 return ERR_PTR(-ENOMEM);
898
899 for_each_bio(bio) {
900 struct bio *bounce_bio = bio;
901 int ret;
902
903 blk_queue_bounce(q, &bounce_bio);
904 ret = blk_rq_append_bio(q, rq, bounce_bio);
905 if (unlikely(ret)) {
906 blk_put_request(rq);
907 return ERR_PTR(ret);
908 }
909 }
910
911 return rq;
912}
913EXPORT_SYMBOL(blk_make_request);
914
915/**
916 * blk_requeue_request - put a request back on queue
917 * @q: request queue where request should be inserted
918 * @rq: request to be inserted
919 *
920 * Description:
921 * Drivers often keep queueing requests until the hardware cannot accept
922 * more, when that condition happens we need to put the request back
923 * on the queue. Must be called with queue lock held.
924 */
925void blk_requeue_request(struct request_queue *q, struct request *rq)
926{
927 blk_delete_timer(rq);
928 blk_clear_rq_complete(rq);
929 trace_block_rq_requeue(q, rq);
930
931 if (blk_rq_tagged(rq))
932 blk_queue_end_tag(q, rq);
933
934 BUG_ON(blk_queued_rq(rq));
935
936 elv_requeue_request(q, rq);
937}
938EXPORT_SYMBOL(blk_requeue_request);
939
940static void add_acct_request(struct request_queue *q, struct request *rq,
941 int where)
942{
943 drive_stat_acct(rq, 1);
944 __elv_add_request(q, rq, where);
945}
946
947/**
948 * blk_insert_request - insert a special request into a request queue
949 * @q: request queue where request should be inserted
950 * @rq: request to be inserted
951 * @at_head: insert request at head or tail of queue
952 * @data: private data
953 *
954 * Description:
955 * Many block devices need to execute commands asynchronously, so they don't
956 * block the whole kernel from preemption during request execution. This is
957 * accomplished normally by inserting aritficial requests tagged as
958 * REQ_TYPE_SPECIAL in to the corresponding request queue, and letting them
959 * be scheduled for actual execution by the request queue.
960 *
961 * We have the option of inserting the head or the tail of the queue.
962 * Typically we use the tail for new ioctls and so forth. We use the head
963 * of the queue for things like a QUEUE_FULL message from a device, or a
964 * host that is unable to accept a particular command.
965 */
966void blk_insert_request(struct request_queue *q, struct request *rq,
967 int at_head, void *data)
968{
969 int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
970 unsigned long flags;
971
972 /*
973 * tell I/O scheduler that this isn't a regular read/write (ie it
974 * must not attempt merges on this) and that it acts as a soft
975 * barrier
976 */
977 rq->cmd_type = REQ_TYPE_SPECIAL;
978
979 rq->special = data;
980
981 spin_lock_irqsave(q->queue_lock, flags);
982
983 /*
984 * If command is tagged, release the tag
985 */
986 if (blk_rq_tagged(rq))
987 blk_queue_end_tag(q, rq);
988
989 add_acct_request(q, rq, where);
990 __blk_run_queue(q);
991 spin_unlock_irqrestore(q->queue_lock, flags);
992}
993EXPORT_SYMBOL(blk_insert_request);
994
995static void part_round_stats_single(int cpu, struct hd_struct *part,
996 unsigned long now)
997{
998 if (now == part->stamp)
999 return;
1000
1001 if (part_in_flight(part)) {
1002 __part_stat_add(cpu, part, time_in_queue,
1003 part_in_flight(part) * (now - part->stamp));
1004 __part_stat_add(cpu, part, io_ticks, (now - part->stamp));
1005 }
1006 part->stamp = now;
1007}
1008
1009/**
1010 * part_round_stats() - Round off the performance stats on a struct disk_stats.
1011 * @cpu: cpu number for stats access
1012 * @part: target partition
1013 *
1014 * The average IO queue length and utilisation statistics are maintained
1015 * by observing the current state of the queue length and the amount of
1016 * time it has been in this state for.
1017 *
1018 * Normally, that accounting is done on IO completion, but that can result
1019 * in more than a second's worth of IO being accounted for within any one
1020 * second, leading to >100% utilisation. To deal with that, we call this
1021 * function to do a round-off before returning the results when reading
1022 * /proc/diskstats. This accounts immediately for all queue usage up to
1023 * the current jiffies and restarts the counters again.
1024 */
1025void part_round_stats(int cpu, struct hd_struct *part)
1026{
1027 unsigned long now = jiffies;
1028
1029 if (part->partno)
1030 part_round_stats_single(cpu, &part_to_disk(part)->part0, now);
1031 part_round_stats_single(cpu, part, now);
1032}
1033EXPORT_SYMBOL_GPL(part_round_stats);
1034
1035/*
1036 * queue lock must be held
1037 */
1038void __blk_put_request(struct request_queue *q, struct request *req)
1039{
1040 if (unlikely(!q))
1041 return;
1042 if (unlikely(--req->ref_count))
1043 return;
1044
1045 elv_completed_request(q, req);
1046
1047 /* this is a bio leak */
1048 WARN_ON(req->bio != NULL);
1049
1050 /*
1051 * Request may not have originated from ll_rw_blk. if not,
1052 * it didn't come out of our reserved rq pools
1053 */
1054 if (req->cmd_flags & REQ_ALLOCED) {
1055 int is_sync = rq_is_sync(req) != 0;
1056 int priv = req->cmd_flags & REQ_ELVPRIV;
1057
1058 BUG_ON(!list_empty(&req->queuelist));
1059 BUG_ON(!hlist_unhashed(&req->hash));
1060
1061 blk_free_request(q, req);
1062 freed_request(q, is_sync, priv);
1063 }
1064}
1065EXPORT_SYMBOL_GPL(__blk_put_request);
1066
1067void blk_put_request(struct request *req)
1068{
1069 unsigned long flags;
1070 struct request_queue *q = req->q;
1071
1072 spin_lock_irqsave(q->queue_lock, flags);
1073 __blk_put_request(q, req);
1074 spin_unlock_irqrestore(q->queue_lock, flags);
1075}
1076EXPORT_SYMBOL(blk_put_request);
1077
1078/**
1079 * blk_add_request_payload - add a payload to a request
1080 * @rq: request to update
1081 * @page: page backing the payload
1082 * @len: length of the payload.
1083 *
1084 * This allows to later add a payload to an already submitted request by
1085 * a block driver. The driver needs to take care of freeing the payload
1086 * itself.
1087 *
1088 * Note that this is a quite horrible hack and nothing but handling of
1089 * discard requests should ever use it.
1090 */
1091void blk_add_request_payload(struct request *rq, struct page *page,
1092 unsigned int len)
1093{
1094 struct bio *bio = rq->bio;
1095
1096 bio->bi_io_vec->bv_page = page;
1097 bio->bi_io_vec->bv_offset = 0;
1098 bio->bi_io_vec->bv_len = len;
1099
1100 bio->bi_size = len;
1101 bio->bi_vcnt = 1;
1102 bio->bi_phys_segments = 1;
1103
1104 rq->__data_len = rq->resid_len = len;
1105 rq->nr_phys_segments = 1;
1106 rq->buffer = bio_data(bio);
1107}
1108EXPORT_SYMBOL_GPL(blk_add_request_payload);
1109
1110static bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
1111 struct bio *bio)
1112{
1113 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1114
1115 if (!ll_back_merge_fn(q, req, bio))
1116 return false;
1117
1118 trace_block_bio_backmerge(q, bio);
1119
1120 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1121 blk_rq_set_mixed_merge(req);
1122
1123 req->biotail->bi_next = bio;
1124 req->biotail = bio;
1125 req->__data_len += bio->bi_size;
1126 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1127
1128 drive_stat_acct(req, 0);
1129 elv_bio_merged(q, req, bio);
1130 return true;
1131}
1132
1133static bool bio_attempt_front_merge(struct request_queue *q,
1134 struct request *req, struct bio *bio)
1135{
1136 const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1137
1138 if (!ll_front_merge_fn(q, req, bio))
1139 return false;
1140
1141 trace_block_bio_frontmerge(q, bio);
1142
1143 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1144 blk_rq_set_mixed_merge(req);
1145
1146 bio->bi_next = req->bio;
1147 req->bio = bio;
1148
1149 /*
1150 * may not be valid. if the low level driver said
1151 * it didn't need a bounce buffer then it better
1152 * not touch req->buffer either...
1153 */
1154 req->buffer = bio_data(bio);
1155 req->__sector = bio->bi_sector;
1156 req->__data_len += bio->bi_size;
1157 req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1158
1159 drive_stat_acct(req, 0);
1160 elv_bio_merged(q, req, bio);
1161 return true;
1162}
1163
1164/*
1165 * Attempts to merge with the plugged list in the current process. Returns
1166 * true if merge was successful, otherwise false.
1167 */
1168static bool attempt_plug_merge(struct task_struct *tsk, struct request_queue *q,
1169 struct bio *bio, unsigned int *request_count)
1170{
1171 struct blk_plug *plug;
1172 struct request *rq;
1173 bool ret = false;
1174
1175 plug = tsk->plug;
1176 if (!plug)
1177 goto out;
1178 *request_count = 0;
1179
1180 list_for_each_entry_reverse(rq, &plug->list, queuelist) {
1181 int el_ret;
1182
1183 (*request_count)++;
1184
1185 if (rq->q != q)
1186 continue;
1187
1188 el_ret = elv_try_merge(rq, bio);
1189 if (el_ret == ELEVATOR_BACK_MERGE) {
1190 ret = bio_attempt_back_merge(q, rq, bio);
1191 if (ret)
1192 break;
1193 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1194 ret = bio_attempt_front_merge(q, rq, bio);
1195 if (ret)
1196 break;
1197 }
1198 }
1199out:
1200 return ret;
1201}
1202
1203void init_request_from_bio(struct request *req, struct bio *bio)
1204{
1205 req->cpu = bio->bi_comp_cpu;
1206 req->cmd_type = REQ_TYPE_FS;
1207
1208 req->cmd_flags |= bio->bi_rw & REQ_COMMON_MASK;
1209 if (bio->bi_rw & REQ_RAHEAD)
1210 req->cmd_flags |= REQ_FAILFAST_MASK;
1211
1212 req->errors = 0;
1213 req->__sector = bio->bi_sector;
1214 req->ioprio = bio_prio(bio);
1215 blk_rq_bio_prep(req->q, req, bio);
1216}
1217
1218static int __make_request(struct request_queue *q, struct bio *bio)
1219{
1220 const bool sync = !!(bio->bi_rw & REQ_SYNC);
1221 struct blk_plug *plug;
1222 int el_ret, rw_flags, where = ELEVATOR_INSERT_SORT;
1223 struct request *req;
1224 unsigned int request_count = 0;
1225
1226 /*
1227 * low level driver can indicate that it wants pages above a
1228 * certain limit bounced to low memory (ie for highmem, or even
1229 * ISA dma in theory)
1230 */
1231 blk_queue_bounce(q, &bio);
1232
1233 if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
1234 spin_lock_irq(q->queue_lock);
1235 where = ELEVATOR_INSERT_FLUSH;
1236 goto get_rq;
1237 }
1238
1239 /*
1240 * Check if we can merge with the plugged list before grabbing
1241 * any locks.
1242 */
1243 if (attempt_plug_merge(current, q, bio, &request_count))
1244 goto out;
1245
1246 spin_lock_irq(q->queue_lock);
1247
1248 el_ret = elv_merge(q, &req, bio);
1249 if (el_ret == ELEVATOR_BACK_MERGE) {
1250 if (bio_attempt_back_merge(q, req, bio)) {
1251 if (!attempt_back_merge(q, req))
1252 elv_merged_request(q, req, el_ret);
1253 goto out_unlock;
1254 }
1255 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1256 if (bio_attempt_front_merge(q, req, bio)) {
1257 if (!attempt_front_merge(q, req))
1258 elv_merged_request(q, req, el_ret);
1259 goto out_unlock;
1260 }
1261 }
1262
1263get_rq:
1264 /*
1265 * This sync check and mask will be re-done in init_request_from_bio(),
1266 * but we need to set it earlier to expose the sync flag to the
1267 * rq allocator and io schedulers.
1268 */
1269 rw_flags = bio_data_dir(bio);
1270 if (sync)
1271 rw_flags |= REQ_SYNC;
1272
1273 /*
1274 * Grab a free request. This is might sleep but can not fail.
1275 * Returns with the queue unlocked.
1276 */
1277 req = get_request_wait(q, rw_flags, bio);
1278
1279 /*
1280 * After dropping the lock and possibly sleeping here, our request
1281 * may now be mergeable after it had proven unmergeable (above).
1282 * We don't worry about that case for efficiency. It won't happen
1283 * often, and the elevators are able to handle it.
1284 */
1285 init_request_from_bio(req, bio);
1286
1287 if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags) ||
1288 bio_flagged(bio, BIO_CPU_AFFINE))
1289 req->cpu = raw_smp_processor_id();
1290
1291 plug = current->plug;
1292 if (plug) {
1293 /*
1294 * If this is the first request added after a plug, fire
1295 * of a plug trace. If others have been added before, check
1296 * if we have multiple devices in this plug. If so, make a
1297 * note to sort the list before dispatch.
1298 */
1299 if (list_empty(&plug->list))
1300 trace_block_plug(q);
1301 else if (!plug->should_sort) {
1302 struct request *__rq;
1303
1304 __rq = list_entry_rq(plug->list.prev);
1305 if (__rq->q != q)
1306 plug->should_sort = 1;
1307 }
1308 if (request_count >= BLK_MAX_REQUEST_COUNT)
1309 blk_flush_plug_list(plug, false);
1310 list_add_tail(&req->queuelist, &plug->list);
1311 drive_stat_acct(req, 1);
1312 } else {
1313 spin_lock_irq(q->queue_lock);
1314 add_acct_request(q, req, where);
1315 __blk_run_queue(q);
1316out_unlock:
1317 spin_unlock_irq(q->queue_lock);
1318 }
1319out:
1320 return 0;
1321}
1322
1323/*
1324 * If bio->bi_dev is a partition, remap the location
1325 */
1326static inline void blk_partition_remap(struct bio *bio)
1327{
1328 struct block_device *bdev = bio->bi_bdev;
1329
1330 if (bio_sectors(bio) && bdev != bdev->bd_contains) {
1331 struct hd_struct *p = bdev->bd_part;
1332
1333 bio->bi_sector += p->start_sect;
1334 bio->bi_bdev = bdev->bd_contains;
1335
1336 trace_block_bio_remap(bdev_get_queue(bio->bi_bdev), bio,
1337 bdev->bd_dev,
1338 bio->bi_sector - p->start_sect);
1339 }
1340}
1341
1342static void handle_bad_sector(struct bio *bio)
1343{
1344 char b[BDEVNAME_SIZE];
1345
1346 printk(KERN_INFO "attempt to access beyond end of device\n");
1347 printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
1348 bdevname(bio->bi_bdev, b),
1349 bio->bi_rw,
1350 (unsigned long long)bio->bi_sector + bio_sectors(bio),
1351 (long long)(i_size_read(bio->bi_bdev->bd_inode) >> 9));
1352
1353 set_bit(BIO_EOF, &bio->bi_flags);
1354}
1355
1356#ifdef CONFIG_FAIL_MAKE_REQUEST
1357
1358static DECLARE_FAULT_ATTR(fail_make_request);
1359
1360static int __init setup_fail_make_request(char *str)
1361{
1362 return setup_fault_attr(&fail_make_request, str);
1363}
1364__setup("fail_make_request=", setup_fail_make_request);
1365
1366static bool should_fail_request(struct hd_struct *part, unsigned int bytes)
1367{
1368 return part->make_it_fail && should_fail(&fail_make_request, bytes);
1369}
1370
1371static int __init fail_make_request_debugfs(void)
1372{
1373 struct dentry *dir = fault_create_debugfs_attr("fail_make_request",
1374 NULL, &fail_make_request);
1375
1376 return IS_ERR(dir) ? PTR_ERR(dir) : 0;
1377}
1378
1379late_initcall(fail_make_request_debugfs);
1380
1381#else /* CONFIG_FAIL_MAKE_REQUEST */
1382
1383static inline bool should_fail_request(struct hd_struct *part,
1384 unsigned int bytes)
1385{
1386 return false;
1387}
1388
1389#endif /* CONFIG_FAIL_MAKE_REQUEST */
1390
1391/*
1392 * Check whether this bio extends beyond the end of the device.
1393 */
1394static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
1395{
1396 sector_t maxsector;
1397
1398 if (!nr_sectors)
1399 return 0;
1400
1401 /* Test device or partition size, when known. */
1402 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
1403 if (maxsector) {
1404 sector_t sector = bio->bi_sector;
1405
1406 if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
1407 /*
1408 * This may well happen - the kernel calls bread()
1409 * without checking the size of the device, e.g., when
1410 * mounting a device.
1411 */
1412 handle_bad_sector(bio);
1413 return 1;
1414 }
1415 }
1416
1417 return 0;
1418}
1419
1420/**
1421 * generic_make_request - hand a buffer to its device driver for I/O
1422 * @bio: The bio describing the location in memory and on the device.
1423 *
1424 * generic_make_request() is used to make I/O requests of block
1425 * devices. It is passed a &struct bio, which describes the I/O that needs
1426 * to be done.
1427 *
1428 * generic_make_request() does not return any status. The
1429 * success/failure status of the request, along with notification of
1430 * completion, is delivered asynchronously through the bio->bi_end_io
1431 * function described (one day) else where.
1432 *
1433 * The caller of generic_make_request must make sure that bi_io_vec
1434 * are set to describe the memory buffer, and that bi_dev and bi_sector are
1435 * set to describe the device address, and the
1436 * bi_end_io and optionally bi_private are set to describe how
1437 * completion notification should be signaled.
1438 *
1439 * generic_make_request and the drivers it calls may use bi_next if this
1440 * bio happens to be merged with someone else, and may change bi_dev and
1441 * bi_sector for remaps as it sees fit. So the values of these fields
1442 * should NOT be depended on after the call to generic_make_request.
1443 */
1444static inline void __generic_make_request(struct bio *bio)
1445{
1446 struct request_queue *q;
1447 sector_t old_sector;
1448 int ret, nr_sectors = bio_sectors(bio);
1449 dev_t old_dev;
1450 int err = -EIO;
1451
1452 might_sleep();
1453
1454 if (bio_check_eod(bio, nr_sectors))
1455 goto end_io;
1456
1457 /*
1458 * Resolve the mapping until finished. (drivers are
1459 * still free to implement/resolve their own stacking
1460 * by explicitly returning 0)
1461 *
1462 * NOTE: we don't repeat the blk_size check for each new device.
1463 * Stacking drivers are expected to know what they are doing.
1464 */
1465 old_sector = -1;
1466 old_dev = 0;
1467 do {
1468 char b[BDEVNAME_SIZE];
1469 struct hd_struct *part;
1470
1471 q = bdev_get_queue(bio->bi_bdev);
1472 if (unlikely(!q)) {
1473 printk(KERN_ERR
1474 "generic_make_request: Trying to access "
1475 "nonexistent block-device %s (%Lu)\n",
1476 bdevname(bio->bi_bdev, b),
1477 (long long) bio->bi_sector);
1478 goto end_io;
1479 }
1480
1481 if (unlikely(!(bio->bi_rw & REQ_DISCARD) &&
1482 nr_sectors > queue_max_hw_sectors(q))) {
1483 printk(KERN_ERR "bio too big device %s (%u > %u)\n",
1484 bdevname(bio->bi_bdev, b),
1485 bio_sectors(bio),
1486 queue_max_hw_sectors(q));
1487 goto end_io;
1488 }
1489
1490 if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags)))
1491 goto end_io;
1492
1493 part = bio->bi_bdev->bd_part;
1494 if (should_fail_request(part, bio->bi_size) ||
1495 should_fail_request(&part_to_disk(part)->part0,
1496 bio->bi_size))
1497 goto end_io;
1498
1499 /*
1500 * If this device has partitions, remap block n
1501 * of partition p to block n+start(p) of the disk.
1502 */
1503 blk_partition_remap(bio);
1504
1505 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio))
1506 goto end_io;
1507
1508 if (old_sector != -1)
1509 trace_block_bio_remap(q, bio, old_dev, old_sector);
1510
1511 old_sector = bio->bi_sector;
1512 old_dev = bio->bi_bdev->bd_dev;
1513
1514 if (bio_check_eod(bio, nr_sectors))
1515 goto end_io;
1516
1517 /*
1518 * Filter flush bio's early so that make_request based
1519 * drivers without flush support don't have to worry
1520 * about them.
1521 */
1522 if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) && !q->flush_flags) {
1523 bio->bi_rw &= ~(REQ_FLUSH | REQ_FUA);
1524 if (!nr_sectors) {
1525 err = 0;
1526 goto end_io;
1527 }
1528 }
1529
1530 if ((bio->bi_rw & REQ_DISCARD) &&
1531 (!blk_queue_discard(q) ||
1532 ((bio->bi_rw & REQ_SECURE) &&
1533 !blk_queue_secdiscard(q)))) {
1534 err = -EOPNOTSUPP;
1535 goto end_io;
1536 }
1537
1538 if (blk_throtl_bio(q, &bio))
1539 goto end_io;
1540
1541 /*
1542 * If bio = NULL, bio has been throttled and will be submitted
1543 * later.
1544 */
1545 if (!bio)
1546 break;
1547
1548 trace_block_bio_queue(q, bio);
1549
1550 ret = q->make_request_fn(q, bio);
1551 } while (ret);
1552
1553 return;
1554
1555end_io:
1556 bio_endio(bio, err);
1557}
1558
1559/*
1560 * We only want one ->make_request_fn to be active at a time,
1561 * else stack usage with stacked devices could be a problem.
1562 * So use current->bio_list to keep a list of requests
1563 * submited by a make_request_fn function.
1564 * current->bio_list is also used as a flag to say if
1565 * generic_make_request is currently active in this task or not.
1566 * If it is NULL, then no make_request is active. If it is non-NULL,
1567 * then a make_request is active, and new requests should be added
1568 * at the tail
1569 */
1570void generic_make_request(struct bio *bio)
1571{
1572 struct bio_list bio_list_on_stack;
1573
1574 if (current->bio_list) {
1575 /* make_request is active */
1576 bio_list_add(current->bio_list, bio);
1577 return;
1578 }
1579 /* following loop may be a bit non-obvious, and so deserves some
1580 * explanation.
1581 * Before entering the loop, bio->bi_next is NULL (as all callers
1582 * ensure that) so we have a list with a single bio.
1583 * We pretend that we have just taken it off a longer list, so
1584 * we assign bio_list to a pointer to the bio_list_on_stack,
1585 * thus initialising the bio_list of new bios to be
1586 * added. __generic_make_request may indeed add some more bios
1587 * through a recursive call to generic_make_request. If it
1588 * did, we find a non-NULL value in bio_list and re-enter the loop
1589 * from the top. In this case we really did just take the bio
1590 * of the top of the list (no pretending) and so remove it from
1591 * bio_list, and call into __generic_make_request again.
1592 *
1593 * The loop was structured like this to make only one call to
1594 * __generic_make_request (which is important as it is large and
1595 * inlined) and to keep the structure simple.
1596 */
1597 BUG_ON(bio->bi_next);
1598 bio_list_init(&bio_list_on_stack);
1599 current->bio_list = &bio_list_on_stack;
1600 do {
1601 __generic_make_request(bio);
1602 bio = bio_list_pop(current->bio_list);
1603 } while (bio);
1604 current->bio_list = NULL; /* deactivate */
1605}
1606EXPORT_SYMBOL(generic_make_request);
1607
1608/**
1609 * submit_bio - submit a bio to the block device layer for I/O
1610 * @rw: whether to %READ or %WRITE, or maybe to %READA (read ahead)
1611 * @bio: The &struct bio which describes the I/O
1612 *
1613 * submit_bio() is very similar in purpose to generic_make_request(), and
1614 * uses that function to do most of the work. Both are fairly rough
1615 * interfaces; @bio must be presetup and ready for I/O.
1616 *
1617 */
1618void submit_bio(int rw, struct bio *bio)
1619{
1620 int count = bio_sectors(bio);
1621
1622 bio->bi_rw |= rw;
1623
1624 /*
1625 * If it's a regular read/write or a barrier with data attached,
1626 * go through the normal accounting stuff before submission.
1627 */
1628 if (bio_has_data(bio) && !(rw & REQ_DISCARD)) {
1629 if (rw & WRITE) {
1630 count_vm_events(PGPGOUT, count);
1631 } else {
1632 task_io_account_read(bio->bi_size);
1633 count_vm_events(PGPGIN, count);
1634 }
1635
1636 if (unlikely(block_dump)) {
1637 char b[BDEVNAME_SIZE];
1638 printk(KERN_DEBUG "%s(%d): %s block %Lu on %s (%u sectors)\n",
1639 current->comm, task_pid_nr(current),
1640 (rw & WRITE) ? "WRITE" : "READ",
1641 (unsigned long long)bio->bi_sector,
1642 bdevname(bio->bi_bdev, b),
1643 count);
1644 }
1645 }
1646
1647 generic_make_request(bio);
1648}
1649EXPORT_SYMBOL(submit_bio);
1650
1651/**
1652 * blk_rq_check_limits - Helper function to check a request for the queue limit
1653 * @q: the queue
1654 * @rq: the request being checked
1655 *
1656 * Description:
1657 * @rq may have been made based on weaker limitations of upper-level queues
1658 * in request stacking drivers, and it may violate the limitation of @q.
1659 * Since the block layer and the underlying device driver trust @rq
1660 * after it is inserted to @q, it should be checked against @q before
1661 * the insertion using this generic function.
1662 *
1663 * This function should also be useful for request stacking drivers
1664 * in some cases below, so export this function.
1665 * Request stacking drivers like request-based dm may change the queue
1666 * limits while requests are in the queue (e.g. dm's table swapping).
1667 * Such request stacking drivers should check those requests agaist
1668 * the new queue limits again when they dispatch those requests,
1669 * although such checkings are also done against the old queue limits
1670 * when submitting requests.
1671 */
1672int blk_rq_check_limits(struct request_queue *q, struct request *rq)
1673{
1674 if (rq->cmd_flags & REQ_DISCARD)
1675 return 0;
1676
1677 if (blk_rq_sectors(rq) > queue_max_sectors(q) ||
1678 blk_rq_bytes(rq) > queue_max_hw_sectors(q) << 9) {
1679 printk(KERN_ERR "%s: over max size limit.\n", __func__);
1680 return -EIO;
1681 }
1682
1683 /*
1684 * queue's settings related to segment counting like q->bounce_pfn
1685 * may differ from that of other stacking queues.
1686 * Recalculate it to check the request correctly on this queue's
1687 * limitation.
1688 */
1689 blk_recalc_rq_segments(rq);
1690 if (rq->nr_phys_segments > queue_max_segments(q)) {
1691 printk(KERN_ERR "%s: over max segments limit.\n", __func__);
1692 return -EIO;
1693 }
1694
1695 return 0;
1696}
1697EXPORT_SYMBOL_GPL(blk_rq_check_limits);
1698
1699/**
1700 * blk_insert_cloned_request - Helper for stacking drivers to submit a request
1701 * @q: the queue to submit the request
1702 * @rq: the request being queued
1703 */
1704int blk_insert_cloned_request(struct request_queue *q, struct request *rq)
1705{
1706 unsigned long flags;
1707 int where = ELEVATOR_INSERT_BACK;
1708
1709 if (blk_rq_check_limits(q, rq))
1710 return -EIO;
1711
1712 if (rq->rq_disk &&
1713 should_fail_request(&rq->rq_disk->part0, blk_rq_bytes(rq)))
1714 return -EIO;
1715
1716 spin_lock_irqsave(q->queue_lock, flags);
1717
1718 /*
1719 * Submitting request must be dequeued before calling this function
1720 * because it will be linked to another request_queue
1721 */
1722 BUG_ON(blk_queued_rq(rq));
1723
1724 if (rq->cmd_flags & (REQ_FLUSH|REQ_FUA))
1725 where = ELEVATOR_INSERT_FLUSH;
1726
1727 add_acct_request(q, rq, where);
1728 spin_unlock_irqrestore(q->queue_lock, flags);
1729
1730 return 0;
1731}
1732EXPORT_SYMBOL_GPL(blk_insert_cloned_request);
1733
1734/**
1735 * blk_rq_err_bytes - determine number of bytes till the next failure boundary
1736 * @rq: request to examine
1737 *
1738 * Description:
1739 * A request could be merge of IOs which require different failure
1740 * handling. This function determines the number of bytes which
1741 * can be failed from the beginning of the request without
1742 * crossing into area which need to be retried further.
1743 *
1744 * Return:
1745 * The number of bytes to fail.
1746 *
1747 * Context:
1748 * queue_lock must be held.
1749 */
1750unsigned int blk_rq_err_bytes(const struct request *rq)
1751{
1752 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
1753 unsigned int bytes = 0;
1754 struct bio *bio;
1755
1756 if (!(rq->cmd_flags & REQ_MIXED_MERGE))
1757 return blk_rq_bytes(rq);
1758
1759 /*
1760 * Currently the only 'mixing' which can happen is between
1761 * different fastfail types. We can safely fail portions
1762 * which have all the failfast bits that the first one has -
1763 * the ones which are at least as eager to fail as the first
1764 * one.
1765 */
1766 for (bio = rq->bio; bio; bio = bio->bi_next) {
1767 if ((bio->bi_rw & ff) != ff)
1768 break;
1769 bytes += bio->bi_size;
1770 }
1771
1772 /* this could lead to infinite loop */
1773 BUG_ON(blk_rq_bytes(rq) && !bytes);
1774 return bytes;
1775}
1776EXPORT_SYMBOL_GPL(blk_rq_err_bytes);
1777
1778static void blk_account_io_completion(struct request *req, unsigned int bytes)
1779{
1780 if (blk_do_io_stat(req)) {
1781 const int rw = rq_data_dir(req);
1782 struct hd_struct *part;
1783 int cpu;
1784
1785 cpu = part_stat_lock();
1786 part = req->part;
1787 part_stat_add(cpu, part, sectors[rw], bytes >> 9);
1788 part_stat_unlock();
1789 }
1790}
1791
1792static void blk_account_io_done(struct request *req)
1793{
1794 /*
1795 * Account IO completion. flush_rq isn't accounted as a
1796 * normal IO on queueing nor completion. Accounting the
1797 * containing request is enough.
1798 */
1799 if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_FLUSH_SEQ)) {
1800 unsigned long duration = jiffies - req->start_time;
1801 const int rw = rq_data_dir(req);
1802 struct hd_struct *part;
1803 int cpu;
1804
1805 cpu = part_stat_lock();
1806 part = req->part;
1807
1808 part_stat_inc(cpu, part, ios[rw]);
1809 part_stat_add(cpu, part, ticks[rw], duration);
1810 part_round_stats(cpu, part);
1811 part_dec_in_flight(part, rw);
1812
1813 hd_struct_put(part);
1814 part_stat_unlock();
1815 }
1816}
1817
1818/**
1819 * blk_peek_request - peek at the top of a request queue
1820 * @q: request queue to peek at
1821 *
1822 * Description:
1823 * Return the request at the top of @q. The returned request
1824 * should be started using blk_start_request() before LLD starts
1825 * processing it.
1826 *
1827 * Return:
1828 * Pointer to the request at the top of @q if available. Null
1829 * otherwise.
1830 *
1831 * Context:
1832 * queue_lock must be held.
1833 */
1834struct request *blk_peek_request(struct request_queue *q)
1835{
1836 struct request *rq;
1837 int ret;
1838
1839 while ((rq = __elv_next_request(q)) != NULL) {
1840 if (!(rq->cmd_flags & REQ_STARTED)) {
1841 /*
1842 * This is the first time the device driver
1843 * sees this request (possibly after
1844 * requeueing). Notify IO scheduler.
1845 */
1846 if (rq->cmd_flags & REQ_SORTED)
1847 elv_activate_rq(q, rq);
1848
1849 /*
1850 * just mark as started even if we don't start
1851 * it, a request that has been delayed should
1852 * not be passed by new incoming requests
1853 */
1854 rq->cmd_flags |= REQ_STARTED;
1855 trace_block_rq_issue(q, rq);
1856 }
1857
1858 if (!q->boundary_rq || q->boundary_rq == rq) {
1859 q->end_sector = rq_end_sector(rq);
1860 q->boundary_rq = NULL;
1861 }
1862
1863 if (rq->cmd_flags & REQ_DONTPREP)
1864 break;
1865
1866 if (q->dma_drain_size && blk_rq_bytes(rq)) {
1867 /*
1868 * make sure space for the drain appears we
1869 * know we can do this because max_hw_segments
1870 * has been adjusted to be one fewer than the
1871 * device can handle
1872 */
1873 rq->nr_phys_segments++;
1874 }
1875
1876 if (!q->prep_rq_fn)
1877 break;
1878
1879 ret = q->prep_rq_fn(q, rq);
1880 if (ret == BLKPREP_OK) {
1881 break;
1882 } else if (ret == BLKPREP_DEFER) {
1883 /*
1884 * the request may have been (partially) prepped.
1885 * we need to keep this request in the front to
1886 * avoid resource deadlock. REQ_STARTED will
1887 * prevent other fs requests from passing this one.
1888 */
1889 if (q->dma_drain_size && blk_rq_bytes(rq) &&
1890 !(rq->cmd_flags & REQ_DONTPREP)) {
1891 /*
1892 * remove the space for the drain we added
1893 * so that we don't add it again
1894 */
1895 --rq->nr_phys_segments;
1896 }
1897
1898 rq = NULL;
1899 break;
1900 } else if (ret == BLKPREP_KILL) {
1901 rq->cmd_flags |= REQ_QUIET;
1902 /*
1903 * Mark this request as started so we don't trigger
1904 * any debug logic in the end I/O path.
1905 */
1906 blk_start_request(rq);
1907 __blk_end_request_all(rq, -EIO);
1908 } else {
1909 printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
1910 break;
1911 }
1912 }
1913
1914 return rq;
1915}
1916EXPORT_SYMBOL(blk_peek_request);
1917
1918void blk_dequeue_request(struct request *rq)
1919{
1920 struct request_queue *q = rq->q;
1921
1922 BUG_ON(list_empty(&rq->queuelist));
1923 BUG_ON(ELV_ON_HASH(rq));
1924
1925 list_del_init(&rq->queuelist);
1926
1927 /*
1928 * the time frame between a request being removed from the lists
1929 * and to it is freed is accounted as io that is in progress at
1930 * the driver side.
1931 */
1932 if (blk_account_rq(rq)) {
1933 q->in_flight[rq_is_sync(rq)]++;
1934 set_io_start_time_ns(rq);
1935 }
1936}
1937
1938/**
1939 * blk_start_request - start request processing on the driver
1940 * @req: request to dequeue
1941 *
1942 * Description:
1943 * Dequeue @req and start timeout timer on it. This hands off the
1944 * request to the driver.
1945 *
1946 * Block internal functions which don't want to start timer should
1947 * call blk_dequeue_request().
1948 *
1949 * Context:
1950 * queue_lock must be held.
1951 */
1952void blk_start_request(struct request *req)
1953{
1954 blk_dequeue_request(req);
1955
1956 /*
1957 * We are now handing the request to the hardware, initialize
1958 * resid_len to full count and add the timeout handler.
1959 */
1960 req->resid_len = blk_rq_bytes(req);
1961 if (unlikely(blk_bidi_rq(req)))
1962 req->next_rq->resid_len = blk_rq_bytes(req->next_rq);
1963
1964 blk_add_timer(req);
1965}
1966EXPORT_SYMBOL(blk_start_request);
1967
1968/**
1969 * blk_fetch_request - fetch a request from a request queue
1970 * @q: request queue to fetch a request from
1971 *
1972 * Description:
1973 * Return the request at the top of @q. The request is started on
1974 * return and LLD can start processing it immediately.
1975 *
1976 * Return:
1977 * Pointer to the request at the top of @q if available. Null
1978 * otherwise.
1979 *
1980 * Context:
1981 * queue_lock must be held.
1982 */
1983struct request *blk_fetch_request(struct request_queue *q)
1984{
1985 struct request *rq;
1986
1987 rq = blk_peek_request(q);
1988 if (rq)
1989 blk_start_request(rq);
1990 return rq;
1991}
1992EXPORT_SYMBOL(blk_fetch_request);
1993
1994/**
1995 * blk_update_request - Special helper function for request stacking drivers
1996 * @req: the request being processed
1997 * @error: %0 for success, < %0 for error
1998 * @nr_bytes: number of bytes to complete @req
1999 *
2000 * Description:
2001 * Ends I/O on a number of bytes attached to @req, but doesn't complete
2002 * the request structure even if @req doesn't have leftover.
2003 * If @req has leftover, sets it up for the next range of segments.
2004 *
2005 * This special helper function is only for request stacking drivers
2006 * (e.g. request-based dm) so that they can handle partial completion.
2007 * Actual device drivers should use blk_end_request instead.
2008 *
2009 * Passing the result of blk_rq_bytes() as @nr_bytes guarantees
2010 * %false return from this function.
2011 *
2012 * Return:
2013 * %false - this request doesn't have any more data
2014 * %true - this request has more data
2015 **/
2016bool blk_update_request(struct request *req, int error, unsigned int nr_bytes)
2017{
2018 int total_bytes, bio_nbytes, next_idx = 0;
2019 struct bio *bio;
2020
2021 if (!req->bio)
2022 return false;
2023
2024 trace_block_rq_complete(req->q, req);
2025
2026 /*
2027 * For fs requests, rq is just carrier of independent bio's
2028 * and each partial completion should be handled separately.
2029 * Reset per-request error on each partial completion.
2030 *
2031 * TODO: tj: This is too subtle. It would be better to let
2032 * low level drivers do what they see fit.
2033 */
2034 if (req->cmd_type == REQ_TYPE_FS)
2035 req->errors = 0;
2036
2037 if (error && req->cmd_type == REQ_TYPE_FS &&
2038 !(req->cmd_flags & REQ_QUIET)) {
2039 char *error_type;
2040
2041 switch (error) {
2042 case -ENOLINK:
2043 error_type = "recoverable transport";
2044 break;
2045 case -EREMOTEIO:
2046 error_type = "critical target";
2047 break;
2048 case -EBADE:
2049 error_type = "critical nexus";
2050 break;
2051 case -EIO:
2052 default:
2053 error_type = "I/O";
2054 break;
2055 }
2056 printk(KERN_ERR "end_request: %s error, dev %s, sector %llu\n",
2057 error_type, req->rq_disk ? req->rq_disk->disk_name : "?",
2058 (unsigned long long)blk_rq_pos(req));
2059 }
2060
2061 blk_account_io_completion(req, nr_bytes);
2062
2063 total_bytes = bio_nbytes = 0;
2064 while ((bio = req->bio) != NULL) {
2065 int nbytes;
2066
2067 if (nr_bytes >= bio->bi_size) {
2068 req->bio = bio->bi_next;
2069 nbytes = bio->bi_size;
2070 req_bio_endio(req, bio, nbytes, error);
2071 next_idx = 0;
2072 bio_nbytes = 0;
2073 } else {
2074 int idx = bio->bi_idx + next_idx;
2075
2076 if (unlikely(idx >= bio->bi_vcnt)) {
2077 blk_dump_rq_flags(req, "__end_that");
2078 printk(KERN_ERR "%s: bio idx %d >= vcnt %d\n",
2079 __func__, idx, bio->bi_vcnt);
2080 break;
2081 }
2082
2083 nbytes = bio_iovec_idx(bio, idx)->bv_len;
2084 BIO_BUG_ON(nbytes > bio->bi_size);
2085
2086 /*
2087 * not a complete bvec done
2088 */
2089 if (unlikely(nbytes > nr_bytes)) {
2090 bio_nbytes += nr_bytes;
2091 total_bytes += nr_bytes;
2092 break;
2093 }
2094
2095 /*
2096 * advance to the next vector
2097 */
2098 next_idx++;
2099 bio_nbytes += nbytes;
2100 }
2101
2102 total_bytes += nbytes;
2103 nr_bytes -= nbytes;
2104
2105 bio = req->bio;
2106 if (bio) {
2107 /*
2108 * end more in this run, or just return 'not-done'
2109 */
2110 if (unlikely(nr_bytes <= 0))
2111 break;
2112 }
2113 }
2114
2115 /*
2116 * completely done
2117 */
2118 if (!req->bio) {
2119 /*
2120 * Reset counters so that the request stacking driver
2121 * can find how many bytes remain in the request
2122 * later.
2123 */
2124 req->__data_len = 0;
2125 return false;
2126 }
2127
2128 /*
2129 * if the request wasn't completed, update state
2130 */
2131 if (bio_nbytes) {
2132 req_bio_endio(req, bio, bio_nbytes, error);
2133 bio->bi_idx += next_idx;
2134 bio_iovec(bio)->bv_offset += nr_bytes;
2135 bio_iovec(bio)->bv_len -= nr_bytes;
2136 }
2137
2138 req->__data_len -= total_bytes;
2139 req->buffer = bio_data(req->bio);
2140
2141 /* update sector only for requests with clear definition of sector */
2142 if (req->cmd_type == REQ_TYPE_FS || (req->cmd_flags & REQ_DISCARD))
2143 req->__sector += total_bytes >> 9;
2144
2145 /* mixed attributes always follow the first bio */
2146 if (req->cmd_flags & REQ_MIXED_MERGE) {
2147 req->cmd_flags &= ~REQ_FAILFAST_MASK;
2148 req->cmd_flags |= req->bio->bi_rw & REQ_FAILFAST_MASK;
2149 }
2150
2151 /*
2152 * If total number of sectors is less than the first segment
2153 * size, something has gone terribly wrong.
2154 */
2155 if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
2156 blk_dump_rq_flags(req, "request botched");
2157 req->__data_len = blk_rq_cur_bytes(req);
2158 }
2159
2160 /* recalculate the number of segments */
2161 blk_recalc_rq_segments(req);
2162
2163 return true;
2164}
2165EXPORT_SYMBOL_GPL(blk_update_request);
2166
2167static bool blk_update_bidi_request(struct request *rq, int error,
2168 unsigned int nr_bytes,
2169 unsigned int bidi_bytes)
2170{
2171 if (blk_update_request(rq, error, nr_bytes))
2172 return true;
2173
2174 /* Bidi request must be completed as a whole */
2175 if (unlikely(blk_bidi_rq(rq)) &&
2176 blk_update_request(rq->next_rq, error, bidi_bytes))
2177 return true;
2178
2179 if (blk_queue_add_random(rq->q))
2180 add_disk_randomness(rq->rq_disk);
2181
2182 return false;
2183}
2184
2185/**
2186 * blk_unprep_request - unprepare a request
2187 * @req: the request
2188 *
2189 * This function makes a request ready for complete resubmission (or
2190 * completion). It happens only after all error handling is complete,
2191 * so represents the appropriate moment to deallocate any resources
2192 * that were allocated to the request in the prep_rq_fn. The queue
2193 * lock is held when calling this.
2194 */
2195void blk_unprep_request(struct request *req)
2196{
2197 struct request_queue *q = req->q;
2198
2199 req->cmd_flags &= ~REQ_DONTPREP;
2200 if (q->unprep_rq_fn)
2201 q->unprep_rq_fn(q, req);
2202}
2203EXPORT_SYMBOL_GPL(blk_unprep_request);
2204
2205/*
2206 * queue lock must be held
2207 */
2208static void blk_finish_request(struct request *req, int error)
2209{
2210 if (blk_rq_tagged(req))
2211 blk_queue_end_tag(req->q, req);
2212
2213 BUG_ON(blk_queued_rq(req));
2214
2215 if (unlikely(laptop_mode) && req->cmd_type == REQ_TYPE_FS)
2216 laptop_io_completion(&req->q->backing_dev_info);
2217
2218 blk_delete_timer(req);
2219
2220 if (req->cmd_flags & REQ_DONTPREP)
2221 blk_unprep_request(req);
2222
2223
2224 blk_account_io_done(req);
2225
2226 if (req->end_io)
2227 req->end_io(req, error);
2228 else {
2229 if (blk_bidi_rq(req))
2230 __blk_put_request(req->next_rq->q, req->next_rq);
2231
2232 __blk_put_request(req->q, req);
2233 }
2234}
2235
2236/**
2237 * blk_end_bidi_request - Complete a bidi request
2238 * @rq: the request to complete
2239 * @error: %0 for success, < %0 for error
2240 * @nr_bytes: number of bytes to complete @rq
2241 * @bidi_bytes: number of bytes to complete @rq->next_rq
2242 *
2243 * Description:
2244 * Ends I/O on a number of bytes attached to @rq and @rq->next_rq.
2245 * Drivers that supports bidi can safely call this member for any
2246 * type of request, bidi or uni. In the later case @bidi_bytes is
2247 * just ignored.
2248 *
2249 * Return:
2250 * %false - we are done with this request
2251 * %true - still buffers pending for this request
2252 **/
2253static bool blk_end_bidi_request(struct request *rq, int error,
2254 unsigned int nr_bytes, unsigned int bidi_bytes)
2255{
2256 struct request_queue *q = rq->q;
2257 unsigned long flags;
2258
2259 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2260 return true;
2261
2262 spin_lock_irqsave(q->queue_lock, flags);
2263 blk_finish_request(rq, error);
2264 spin_unlock_irqrestore(q->queue_lock, flags);
2265
2266 return false;
2267}
2268
2269/**
2270 * __blk_end_bidi_request - Complete a bidi request with queue lock held
2271 * @rq: the request to complete
2272 * @error: %0 for success, < %0 for error
2273 * @nr_bytes: number of bytes to complete @rq
2274 * @bidi_bytes: number of bytes to complete @rq->next_rq
2275 *
2276 * Description:
2277 * Identical to blk_end_bidi_request() except that queue lock is
2278 * assumed to be locked on entry and remains so on return.
2279 *
2280 * Return:
2281 * %false - we are done with this request
2282 * %true - still buffers pending for this request
2283 **/
2284bool __blk_end_bidi_request(struct request *rq, int error,
2285 unsigned int nr_bytes, unsigned int bidi_bytes)
2286{
2287 if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2288 return true;
2289
2290 blk_finish_request(rq, error);
2291
2292 return false;
2293}
2294
2295/**
2296 * blk_end_request - Helper function for drivers to complete the request.
2297 * @rq: the request being processed
2298 * @error: %0 for success, < %0 for error
2299 * @nr_bytes: number of bytes to complete
2300 *
2301 * Description:
2302 * Ends I/O on a number of bytes attached to @rq.
2303 * If @rq has leftover, sets it up for the next range of segments.
2304 *
2305 * Return:
2306 * %false - we are done with this request
2307 * %true - still buffers pending for this request
2308 **/
2309bool blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2310{
2311 return blk_end_bidi_request(rq, error, nr_bytes, 0);
2312}
2313EXPORT_SYMBOL(blk_end_request);
2314
2315/**
2316 * blk_end_request_all - Helper function for drives to finish the request.
2317 * @rq: the request to finish
2318 * @error: %0 for success, < %0 for error
2319 *
2320 * Description:
2321 * Completely finish @rq.
2322 */
2323void blk_end_request_all(struct request *rq, int error)
2324{
2325 bool pending;
2326 unsigned int bidi_bytes = 0;
2327
2328 if (unlikely(blk_bidi_rq(rq)))
2329 bidi_bytes = blk_rq_bytes(rq->next_rq);
2330
2331 pending = blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2332 BUG_ON(pending);
2333}
2334EXPORT_SYMBOL(blk_end_request_all);
2335
2336/**
2337 * blk_end_request_cur - Helper function to finish the current request chunk.
2338 * @rq: the request to finish the current chunk for
2339 * @error: %0 for success, < %0 for error
2340 *
2341 * Description:
2342 * Complete the current consecutively mapped chunk from @rq.
2343 *
2344 * Return:
2345 * %false - we are done with this request
2346 * %true - still buffers pending for this request
2347 */
2348bool blk_end_request_cur(struct request *rq, int error)
2349{
2350 return blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2351}
2352EXPORT_SYMBOL(blk_end_request_cur);
2353
2354/**
2355 * blk_end_request_err - Finish a request till the next failure boundary.
2356 * @rq: the request to finish till the next failure boundary for
2357 * @error: must be negative errno
2358 *
2359 * Description:
2360 * Complete @rq till the next failure boundary.
2361 *
2362 * Return:
2363 * %false - we are done with this request
2364 * %true - still buffers pending for this request
2365 */
2366bool blk_end_request_err(struct request *rq, int error)
2367{
2368 WARN_ON(error >= 0);
2369 return blk_end_request(rq, error, blk_rq_err_bytes(rq));
2370}
2371EXPORT_SYMBOL_GPL(blk_end_request_err);
2372
2373/**
2374 * __blk_end_request - Helper function for drivers to complete the request.
2375 * @rq: the request being processed
2376 * @error: %0 for success, < %0 for error
2377 * @nr_bytes: number of bytes to complete
2378 *
2379 * Description:
2380 * Must be called with queue lock held unlike blk_end_request().
2381 *
2382 * Return:
2383 * %false - we are done with this request
2384 * %true - still buffers pending for this request
2385 **/
2386bool __blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2387{
2388 return __blk_end_bidi_request(rq, error, nr_bytes, 0);
2389}
2390EXPORT_SYMBOL(__blk_end_request);
2391
2392/**
2393 * __blk_end_request_all - Helper function for drives to finish the request.
2394 * @rq: the request to finish
2395 * @error: %0 for success, < %0 for error
2396 *
2397 * Description:
2398 * Completely finish @rq. Must be called with queue lock held.
2399 */
2400void __blk_end_request_all(struct request *rq, int error)
2401{
2402 bool pending;
2403 unsigned int bidi_bytes = 0;
2404
2405 if (unlikely(blk_bidi_rq(rq)))
2406 bidi_bytes = blk_rq_bytes(rq->next_rq);
2407
2408 pending = __blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2409 BUG_ON(pending);
2410}
2411EXPORT_SYMBOL(__blk_end_request_all);
2412
2413/**
2414 * __blk_end_request_cur - Helper function to finish the current request chunk.
2415 * @rq: the request to finish the current chunk for
2416 * @error: %0 for success, < %0 for error
2417 *
2418 * Description:
2419 * Complete the current consecutively mapped chunk from @rq. Must
2420 * be called with queue lock held.
2421 *
2422 * Return:
2423 * %false - we are done with this request
2424 * %true - still buffers pending for this request
2425 */
2426bool __blk_end_request_cur(struct request *rq, int error)
2427{
2428 return __blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2429}
2430EXPORT_SYMBOL(__blk_end_request_cur);
2431
2432/**
2433 * __blk_end_request_err - Finish a request till the next failure boundary.
2434 * @rq: the request to finish till the next failure boundary for
2435 * @error: must be negative errno
2436 *
2437 * Description:
2438 * Complete @rq till the next failure boundary. Must be called
2439 * with queue lock held.
2440 *
2441 * Return:
2442 * %false - we are done with this request
2443 * %true - still buffers pending for this request
2444 */
2445bool __blk_end_request_err(struct request *rq, int error)
2446{
2447 WARN_ON(error >= 0);
2448 return __blk_end_request(rq, error, blk_rq_err_bytes(rq));
2449}
2450EXPORT_SYMBOL_GPL(__blk_end_request_err);
2451
2452void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
2453 struct bio *bio)
2454{
2455 /* Bit 0 (R/W) is identical in rq->cmd_flags and bio->bi_rw */
2456 rq->cmd_flags |= bio->bi_rw & REQ_WRITE;
2457
2458 if (bio_has_data(bio)) {
2459 rq->nr_phys_segments = bio_phys_segments(q, bio);
2460 rq->buffer = bio_data(bio);
2461 }
2462 rq->__data_len = bio->bi_size;
2463 rq->bio = rq->biotail = bio;
2464
2465 if (bio->bi_bdev)
2466 rq->rq_disk = bio->bi_bdev->bd_disk;
2467}
2468
2469#if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
2470/**
2471 * rq_flush_dcache_pages - Helper function to flush all pages in a request
2472 * @rq: the request to be flushed
2473 *
2474 * Description:
2475 * Flush all pages in @rq.
2476 */
2477void rq_flush_dcache_pages(struct request *rq)
2478{
2479 struct req_iterator iter;
2480 struct bio_vec *bvec;
2481
2482 rq_for_each_segment(bvec, rq, iter)
2483 flush_dcache_page(bvec->bv_page);
2484}
2485EXPORT_SYMBOL_GPL(rq_flush_dcache_pages);
2486#endif
2487
2488/**
2489 * blk_lld_busy - Check if underlying low-level drivers of a device are busy
2490 * @q : the queue of the device being checked
2491 *
2492 * Description:
2493 * Check if underlying low-level drivers of a device are busy.
2494 * If the drivers want to export their busy state, they must set own
2495 * exporting function using blk_queue_lld_busy() first.
2496 *
2497 * Basically, this function is used only by request stacking drivers
2498 * to stop dispatching requests to underlying devices when underlying
2499 * devices are busy. This behavior helps more I/O merging on the queue
2500 * of the request stacking driver and prevents I/O throughput regression
2501 * on burst I/O load.
2502 *
2503 * Return:
2504 * 0 - Not busy (The request stacking driver should dispatch request)
2505 * 1 - Busy (The request stacking driver should stop dispatching request)
2506 */
2507int blk_lld_busy(struct request_queue *q)
2508{
2509 if (q->lld_busy_fn)
2510 return q->lld_busy_fn(q);
2511
2512 return 0;
2513}
2514EXPORT_SYMBOL_GPL(blk_lld_busy);
2515
2516/**
2517 * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
2518 * @rq: the clone request to be cleaned up
2519 *
2520 * Description:
2521 * Free all bios in @rq for a cloned request.
2522 */
2523void blk_rq_unprep_clone(struct request *rq)
2524{
2525 struct bio *bio;
2526
2527 while ((bio = rq->bio) != NULL) {
2528 rq->bio = bio->bi_next;
2529
2530 bio_put(bio);
2531 }
2532}
2533EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
2534
2535/*
2536 * Copy attributes of the original request to the clone request.
2537 * The actual data parts (e.g. ->cmd, ->buffer, ->sense) are not copied.
2538 */
2539static void __blk_rq_prep_clone(struct request *dst, struct request *src)
2540{
2541 dst->cpu = src->cpu;
2542 dst->cmd_flags = (src->cmd_flags & REQ_CLONE_MASK) | REQ_NOMERGE;
2543 dst->cmd_type = src->cmd_type;
2544 dst->__sector = blk_rq_pos(src);
2545 dst->__data_len = blk_rq_bytes(src);
2546 dst->nr_phys_segments = src->nr_phys_segments;
2547 dst->ioprio = src->ioprio;
2548 dst->extra_len = src->extra_len;
2549}
2550
2551/**
2552 * blk_rq_prep_clone - Helper function to setup clone request
2553 * @rq: the request to be setup
2554 * @rq_src: original request to be cloned
2555 * @bs: bio_set that bios for clone are allocated from
2556 * @gfp_mask: memory allocation mask for bio
2557 * @bio_ctr: setup function to be called for each clone bio.
2558 * Returns %0 for success, non %0 for failure.
2559 * @data: private data to be passed to @bio_ctr
2560 *
2561 * Description:
2562 * Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
2563 * The actual data parts of @rq_src (e.g. ->cmd, ->buffer, ->sense)
2564 * are not copied, and copying such parts is the caller's responsibility.
2565 * Also, pages which the original bios are pointing to are not copied
2566 * and the cloned bios just point same pages.
2567 * So cloned bios must be completed before original bios, which means
2568 * the caller must complete @rq before @rq_src.
2569 */
2570int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2571 struct bio_set *bs, gfp_t gfp_mask,
2572 int (*bio_ctr)(struct bio *, struct bio *, void *),
2573 void *data)
2574{
2575 struct bio *bio, *bio_src;
2576
2577 if (!bs)
2578 bs = fs_bio_set;
2579
2580 blk_rq_init(NULL, rq);
2581
2582 __rq_for_each_bio(bio_src, rq_src) {
2583 bio = bio_alloc_bioset(gfp_mask, bio_src->bi_max_vecs, bs);
2584 if (!bio)
2585 goto free_and_out;
2586
2587 __bio_clone(bio, bio_src);
2588
2589 if (bio_integrity(bio_src) &&
2590 bio_integrity_clone(bio, bio_src, gfp_mask, bs))
2591 goto free_and_out;
2592
2593 if (bio_ctr && bio_ctr(bio, bio_src, data))
2594 goto free_and_out;
2595
2596 if (rq->bio) {
2597 rq->biotail->bi_next = bio;
2598 rq->biotail = bio;
2599 } else
2600 rq->bio = rq->biotail = bio;
2601 }
2602
2603 __blk_rq_prep_clone(rq, rq_src);
2604
2605 return 0;
2606
2607free_and_out:
2608 if (bio)
2609 bio_free(bio, bs);
2610 blk_rq_unprep_clone(rq);
2611
2612 return -ENOMEM;
2613}
2614EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
2615
2616int kblockd_schedule_work(struct request_queue *q, struct work_struct *work)
2617{
2618 return queue_work(kblockd_workqueue, work);
2619}
2620EXPORT_SYMBOL(kblockd_schedule_work);
2621
2622int kblockd_schedule_delayed_work(struct request_queue *q,
2623 struct delayed_work *dwork, unsigned long delay)
2624{
2625 return queue_delayed_work(kblockd_workqueue, dwork, delay);
2626}
2627EXPORT_SYMBOL(kblockd_schedule_delayed_work);
2628
2629#define PLUG_MAGIC 0x91827364
2630
2631void blk_start_plug(struct blk_plug *plug)
2632{
2633 struct task_struct *tsk = current;
2634
2635 plug->magic = PLUG_MAGIC;
2636 INIT_LIST_HEAD(&plug->list);
2637 INIT_LIST_HEAD(&plug->cb_list);
2638 plug->should_sort = 0;
2639
2640 /*
2641 * If this is a nested plug, don't actually assign it. It will be
2642 * flushed on its own.
2643 */
2644 if (!tsk->plug) {
2645 /*
2646 * Store ordering should not be needed here, since a potential
2647 * preempt will imply a full memory barrier
2648 */
2649 tsk->plug = plug;
2650 }
2651}
2652EXPORT_SYMBOL(blk_start_plug);
2653
2654static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
2655{
2656 struct request *rqa = container_of(a, struct request, queuelist);
2657 struct request *rqb = container_of(b, struct request, queuelist);
2658
2659 return !(rqa->q <= rqb->q);
2660}
2661
2662/*
2663 * If 'from_schedule' is true, then postpone the dispatch of requests
2664 * until a safe kblockd context. We due this to avoid accidental big
2665 * additional stack usage in driver dispatch, in places where the originally
2666 * plugger did not intend it.
2667 */
2668static void queue_unplugged(struct request_queue *q, unsigned int depth,
2669 bool from_schedule)
2670 __releases(q->queue_lock)
2671{
2672 trace_block_unplug(q, depth, !from_schedule);
2673
2674 /*
2675 * If we are punting this to kblockd, then we can safely drop
2676 * the queue_lock before waking kblockd (which needs to take
2677 * this lock).
2678 */
2679 if (from_schedule) {
2680 spin_unlock(q->queue_lock);
2681 blk_run_queue_async(q);
2682 } else {
2683 __blk_run_queue(q);
2684 spin_unlock(q->queue_lock);
2685 }
2686
2687}
2688
2689static void flush_plug_callbacks(struct blk_plug *plug)
2690{
2691 LIST_HEAD(callbacks);
2692
2693 if (list_empty(&plug->cb_list))
2694 return;
2695
2696 list_splice_init(&plug->cb_list, &callbacks);
2697
2698 while (!list_empty(&callbacks)) {
2699 struct blk_plug_cb *cb = list_first_entry(&callbacks,
2700 struct blk_plug_cb,
2701 list);
2702 list_del(&cb->list);
2703 cb->callback(cb);
2704 }
2705}
2706
2707void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2708{
2709 struct request_queue *q;
2710 unsigned long flags;
2711 struct request *rq;
2712 LIST_HEAD(list);
2713 unsigned int depth;
2714
2715 BUG_ON(plug->magic != PLUG_MAGIC);
2716
2717 flush_plug_callbacks(plug);
2718 if (list_empty(&plug->list))
2719 return;
2720
2721 list_splice_init(&plug->list, &list);
2722
2723 if (plug->should_sort) {
2724 list_sort(NULL, &list, plug_rq_cmp);
2725 plug->should_sort = 0;
2726 }
2727
2728 q = NULL;
2729 depth = 0;
2730
2731 /*
2732 * Save and disable interrupts here, to avoid doing it for every
2733 * queue lock we have to take.
2734 */
2735 local_irq_save(flags);
2736 while (!list_empty(&list)) {
2737 rq = list_entry_rq(list.next);
2738 list_del_init(&rq->queuelist);
2739 BUG_ON(!rq->q);
2740 if (rq->q != q) {
2741 /*
2742 * This drops the queue lock
2743 */
2744 if (q)
2745 queue_unplugged(q, depth, from_schedule);
2746 q = rq->q;
2747 depth = 0;
2748 spin_lock(q->queue_lock);
2749 }
2750 /*
2751 * rq is already accounted, so use raw insert
2752 */
2753 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA))
2754 __elv_add_request(q, rq, ELEVATOR_INSERT_FLUSH);
2755 else
2756 __elv_add_request(q, rq, ELEVATOR_INSERT_SORT_MERGE);
2757
2758 depth++;
2759 }
2760
2761 /*
2762 * This drops the queue lock
2763 */
2764 if (q)
2765 queue_unplugged(q, depth, from_schedule);
2766
2767 local_irq_restore(flags);
2768}
2769
2770void blk_finish_plug(struct blk_plug *plug)
2771{
2772 blk_flush_plug_list(plug, false);
2773
2774 if (plug == current->plug)
2775 current->plug = NULL;
2776}
2777EXPORT_SYMBOL(blk_finish_plug);
2778
2779int __init blk_dev_init(void)
2780{
2781 BUILD_BUG_ON(__REQ_NR_BITS > 8 *
2782 sizeof(((struct request *)0)->cmd_flags));
2783
2784 /* used for unplugging and affects IO latency/throughput - HIGHPRI */
2785 kblockd_workqueue = alloc_workqueue("kblockd",
2786 WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
2787 if (!kblockd_workqueue)
2788 panic("Failed to create kblockd\n");
2789
2790 request_cachep = kmem_cache_create("blkdev_requests",
2791 sizeof(struct request), 0, SLAB_PANIC, NULL);
2792
2793 blk_requestq_cachep = kmem_cache_create("blkdev_queue",
2794 sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
2795
2796 return 0;
2797}