Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
blk-core.c
Go to the documentation of this file.
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 <[email protected]> SuSE
5  * Queue request tables / lock, selectable elevator, Jens Axboe <[email protected]>
6  * kernel-doc documentation started by NeilBrown <[email protected]>
7  * - July2000
8  * bio rewrite, highmem i/o, etc, Jens Axboe <[email protected]> - 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>
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 
40 EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_remap);
41 EXPORT_TRACEPOINT_SYMBOL_GPL(block_rq_remap);
42 EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_complete);
43 
44 DEFINE_IDA(blk_queue_ida);
45 
46 /*
47  * For the allocated request tables
48  */
49 static struct kmem_cache *request_cachep;
50 
51 /*
52  * For queue allocation
53  */
55 
56 /*
57  * Controlling structure to kblockd
58  */
59 static struct workqueue_struct *kblockd_workqueue;
60 
61 static 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 
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 
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 }
131 
132 void 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 }
152 
153 static 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 
181 void 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 }
203 
204 static 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 
224 void 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 }
230 
241 {
243 
244  queue_flag_clear(QUEUE_FLAG_STOPPED, q);
245  __blk_run_queue(q);
246 }
248 
264 {
265  cancel_delayed_work(&q->delay_work);
266  queue_flag_set(QUEUE_FLAG_STOPPED, q);
267 }
269 
289 {
290  del_timer_sync(&q->timeout);
291  cancel_delayed_work_sync(&q->delay_work);
292 }
294 
304 {
305  if (unlikely(blk_queue_stopped(q)))
306  return;
307 
308  q->request_fn(q);
309 }
311 
321 {
322  if (likely(!blk_queue_stopped(q)))
323  mod_delayed_work(kblockd_workqueue, &q->delay_work, 0);
324 }
326 
336 {
337  unsigned long flags;
338 
339  spin_lock_irqsave(q->queue_lock, flags);
340  __blk_run_queue(q);
341  spin_unlock_irqrestore(q->queue_lock, flags);
342 }
344 
346 {
347  kobject_put(&q->kobj);
348 }
350 
360 void blk_drain_queue(struct request_queue *q, bool drain_all)
361 {
362  int i;
363 
364  while (true) {
365  bool drain = false;
366 
367  spin_lock_irq(q->queue_lock);
368 
369  /*
370  * The caller might be trying to drain @q before its
371  * elevator is initialized.
372  */
373  if (q->elevator)
375 
377 
378  /*
379  * This function might be called on a queue which failed
380  * driver init after queue creation or is not yet fully
381  * active yet. Some drivers (e.g. fd and loop) get unhappy
382  * in such cases. Kick queue iff dispatch queue has
383  * something on it and @q has request_fn set.
384  */
385  if (!list_empty(&q->queue_head) && q->request_fn)
386  __blk_run_queue(q);
387 
388  drain |= q->nr_rqs_elvpriv;
389 
390  /*
391  * Unfortunately, requests are queued at and tracked from
392  * multiple places and there's no single counter which can
393  * be drained. Check all the queues and counters.
394  */
395  if (drain_all) {
396  drain |= !list_empty(&q->queue_head);
397  for (i = 0; i < 2; i++) {
398  drain |= q->nr_rqs[i];
399  drain |= q->in_flight[i];
400  drain |= !list_empty(&q->flush_queue[i]);
401  }
402  }
403 
404  spin_unlock_irq(q->queue_lock);
405 
406  if (!drain)
407  break;
408  msleep(10);
409  }
410 
411  /*
412  * With queue marked dead, any woken up waiter will fail the
413  * allocation path, so the wakeup chaining is lost and we're
414  * left with hung waiters. We need to wake up those waiters.
415  */
416  if (q->request_fn) {
417  struct request_list *rl;
418 
419  spin_lock_irq(q->queue_lock);
420 
421  blk_queue_for_each_rl(rl, q)
422  for (i = 0; i < ARRAY_SIZE(rl->wait); i++)
423  wake_up_all(&rl->wait[i]);
424 
425  spin_unlock_irq(q->queue_lock);
426  }
427 }
428 
440 {
441  bool drain;
442 
443  spin_lock_irq(q->queue_lock);
444  drain = !q->bypass_depth++;
445  queue_flag_set(QUEUE_FLAG_BYPASS, q);
446  spin_unlock_irq(q->queue_lock);
447 
448  if (drain) {
449  blk_drain_queue(q, false);
450  /* ensure blk_queue_bypass() is %true inside RCU read lock */
451  synchronize_rcu();
452  }
453 }
455 
463 {
464  spin_lock_irq(q->queue_lock);
465  if (!--q->bypass_depth)
466  queue_flag_clear(QUEUE_FLAG_BYPASS, q);
467  WARN_ON_ONCE(q->bypass_depth < 0);
468  spin_unlock_irq(q->queue_lock);
469 }
471 
480 {
481  spinlock_t *lock = q->queue_lock;
482 
483  /* mark @q DEAD, no new request or merges will be allowed afterwards */
484  mutex_lock(&q->sysfs_lock);
485  queue_flag_set_unlocked(QUEUE_FLAG_DEAD, q);
486  spin_lock_irq(lock);
487 
488  /*
489  * Dead queue is permanently in bypass mode till released. Note
490  * that, unlike blk_queue_bypass_start(), we aren't performing
491  * synchronize_rcu() after entering bypass mode to avoid the delay
492  * as some drivers create and destroy a lot of queues while
493  * probing. This is still safe because blk_release_queue() will be
494  * called only after the queue refcnt drops to zero and nothing,
495  * RCU or not, would be traversing the queue by then.
496  */
497  q->bypass_depth++;
498  queue_flag_set(QUEUE_FLAG_BYPASS, q);
499 
500  queue_flag_set(QUEUE_FLAG_NOMERGES, q);
501  queue_flag_set(QUEUE_FLAG_NOXMERGES, q);
502  queue_flag_set(QUEUE_FLAG_DEAD, q);
503  spin_unlock_irq(lock);
504  mutex_unlock(&q->sysfs_lock);
505 
506  /* drain all requests queued before DEAD marking */
507  blk_drain_queue(q, true);
508 
509  /* @q won't process any more request, flush async actions */
510  del_timer_sync(&q->backing_dev_info.laptop_mode_wb_timer);
511  blk_sync_queue(q);
512 
513  spin_lock_irq(lock);
514  if (q->queue_lock != &q->__queue_lock)
515  q->queue_lock = &q->__queue_lock;
516  spin_unlock_irq(lock);
517 
518  /* @q is and will stay empty, shutdown and put */
519  blk_put_queue(q);
520 }
522 
523 int blk_init_rl(struct request_list *rl, struct request_queue *q,
524  gfp_t gfp_mask)
525 {
526  if (unlikely(rl->rq_pool))
527  return 0;
528 
529  rl->q = q;
530  rl->count[BLK_RW_SYNC] = rl->count[BLK_RW_ASYNC] = 0;
531  rl->starved[BLK_RW_SYNC] = rl->starved[BLK_RW_ASYNC] = 0;
532  init_waitqueue_head(&rl->wait[BLK_RW_SYNC]);
533  init_waitqueue_head(&rl->wait[BLK_RW_ASYNC]);
534 
535  rl->rq_pool = mempool_create_node(BLKDEV_MIN_RQ, mempool_alloc_slab,
536  mempool_free_slab, request_cachep,
537  gfp_mask, q->node);
538  if (!rl->rq_pool)
539  return -ENOMEM;
540 
541  return 0;
542 }
543 
544 void blk_exit_rl(struct request_list *rl)
545 {
546  if (rl->rq_pool)
547  mempool_destroy(rl->rq_pool);
548 }
549 
551 {
552  return blk_alloc_queue_node(gfp_mask, -1);
553 }
555 
557 {
558  struct request_queue *q;
559  int err;
560 
561  q = kmem_cache_alloc_node(blk_requestq_cachep,
562  gfp_mask | __GFP_ZERO, node_id);
563  if (!q)
564  return NULL;
565 
566  q->id = ida_simple_get(&blk_queue_ida, 0, 0, gfp_mask);
567  if (q->id < 0)
568  goto fail_q;
569 
570  q->backing_dev_info.ra_pages =
571  (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
572  q->backing_dev_info.state = 0;
573  q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
574  q->backing_dev_info.name = "block";
575  q->node = node_id;
576 
577  err = bdi_init(&q->backing_dev_info);
578  if (err)
579  goto fail_id;
580 
581  setup_timer(&q->backing_dev_info.laptop_mode_wb_timer,
582  laptop_mode_timer_fn, (unsigned long) q);
583  setup_timer(&q->timeout, blk_rq_timed_out_timer, (unsigned long) q);
584  INIT_LIST_HEAD(&q->queue_head);
585  INIT_LIST_HEAD(&q->timeout_list);
586  INIT_LIST_HEAD(&q->icq_list);
587 #ifdef CONFIG_BLK_CGROUP
588  INIT_LIST_HEAD(&q->blkg_list);
589 #endif
590  INIT_LIST_HEAD(&q->flush_queue[0]);
591  INIT_LIST_HEAD(&q->flush_queue[1]);
592  INIT_LIST_HEAD(&q->flush_data_in_flight);
593  INIT_DELAYED_WORK(&q->delay_work, blk_delay_work);
594 
595  kobject_init(&q->kobj, &blk_queue_ktype);
596 
597  mutex_init(&q->sysfs_lock);
598  spin_lock_init(&q->__queue_lock);
599 
600  /*
601  * By default initialize queue_lock to internal lock and driver can
602  * override it later if need be.
603  */
604  q->queue_lock = &q->__queue_lock;
605 
606  /*
607  * A queue starts its life with bypass turned on to avoid
608  * unnecessary bypass on/off overhead and nasty surprises during
609  * init. The initial bypass will be finished when the queue is
610  * registered by blk_register_queue().
611  */
612  q->bypass_depth = 1;
613  __set_bit(QUEUE_FLAG_BYPASS, &q->queue_flags);
614 
615  if (blkcg_init_queue(q))
616  goto fail_id;
617 
618  return q;
619 
620 fail_id:
622 fail_q:
623  kmem_cache_free(blk_requestq_cachep, q);
624  return NULL;
625 }
627 
661 struct request_queue *blk_init_queue(request_fn_proc *rfn, spinlock_t *lock)
662 {
663  return blk_init_queue_node(rfn, lock, -1);
664 }
666 
667 struct request_queue *
668 blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
669 {
670  struct request_queue *uninit_q, *q;
671 
672  uninit_q = blk_alloc_queue_node(GFP_KERNEL, node_id);
673  if (!uninit_q)
674  return NULL;
675 
676  q = blk_init_allocated_queue(uninit_q, rfn, lock);
677  if (!q)
678  blk_cleanup_queue(uninit_q);
679 
680  return q;
681 }
683 
684 struct request_queue *
685 blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
686  spinlock_t *lock)
687 {
688  if (!q)
689  return NULL;
690 
691  if (blk_init_rl(&q->root_rl, q, GFP_KERNEL))
692  return NULL;
693 
694  q->request_fn = rfn;
695  q->prep_rq_fn = NULL;
696  q->unprep_rq_fn = NULL;
697  q->queue_flags |= QUEUE_FLAG_DEFAULT;
698 
699  /* Override internal queue lock with supplied lock pointer */
700  if (lock)
701  q->queue_lock = lock;
702 
703  /*
704  * This also sets hw/phys segments, boundary and size
705  */
707 
708  q->sg_reserved_size = INT_MAX;
709 
710  /* init elevator */
711  if (elevator_init(q, NULL))
712  return NULL;
713  return q;
714 }
716 
718 {
719  if (likely(!blk_queue_dead(q))) {
720  __blk_get_queue(q);
721  return true;
722  }
723 
724  return false;
725 }
727 
728 static inline void blk_free_request(struct request_list *rl, struct request *rq)
729 {
730  if (rq->cmd_flags & REQ_ELVPRIV) {
731  elv_put_request(rl->q, rq);
732  if (rq->elv.icq)
733  put_io_context(rq->elv.icq->ioc);
734  }
735 
736  mempool_free(rq, rl->rq_pool);
737 }
738 
739 /*
740  * ioc_batching returns true if the ioc is a valid batching request and
741  * should be given priority access to a request.
742  */
743 static inline int ioc_batching(struct request_queue *q, struct io_context *ioc)
744 {
745  if (!ioc)
746  return 0;
747 
748  /*
749  * Make sure the process is able to allocate at least 1 request
750  * even if the batch times out, otherwise we could theoretically
751  * lose wakeups.
752  */
753  return ioc->nr_batch_requests == q->nr_batching ||
754  (ioc->nr_batch_requests > 0
755  && time_before(jiffies, ioc->last_waited + BLK_BATCH_TIME));
756 }
757 
758 /*
759  * ioc_set_batching sets ioc to be a new "batcher" if it is not one. This
760  * will cause the process to be a "batcher" on all queues in the system. This
761  * is the behaviour we want though - once it gets a wakeup it should be given
762  * a nice run.
763  */
764 static void ioc_set_batching(struct request_queue *q, struct io_context *ioc)
765 {
766  if (!ioc || ioc_batching(q, ioc))
767  return;
768 
769  ioc->nr_batch_requests = q->nr_batching;
770  ioc->last_waited = jiffies;
771 }
772 
773 static void __freed_request(struct request_list *rl, int sync)
774 {
775  struct request_queue *q = rl->q;
776 
777  /*
778  * bdi isn't aware of blkcg yet. As all async IOs end up root
779  * blkcg anyway, just use root blkcg state.
780  */
781  if (rl == &q->root_rl &&
782  rl->count[sync] < queue_congestion_off_threshold(q))
783  blk_clear_queue_congested(q, sync);
784 
785  if (rl->count[sync] + 1 <= q->nr_requests) {
786  if (waitqueue_active(&rl->wait[sync]))
787  wake_up(&rl->wait[sync]);
788 
789  blk_clear_rl_full(rl, sync);
790  }
791 }
792 
793 /*
794  * A request has just been released. Account for it, update the full and
795  * congestion status, wake up any waiters. Called under q->queue_lock.
796  */
797 static void freed_request(struct request_list *rl, unsigned int flags)
798 {
799  struct request_queue *q = rl->q;
800  int sync = rw_is_sync(flags);
801 
802  q->nr_rqs[sync]--;
803  rl->count[sync]--;
804  if (flags & REQ_ELVPRIV)
805  q->nr_rqs_elvpriv--;
806 
807  __freed_request(rl, sync);
808 
809  if (unlikely(rl->starved[sync ^ 1]))
810  __freed_request(rl, sync ^ 1);
811 }
812 
813 /*
814  * Determine if elevator data should be initialized when allocating the
815  * request associated with @bio.
816  */
817 static bool blk_rq_should_init_elevator(struct bio *bio)
818 {
819  if (!bio)
820  return true;
821 
822  /*
823  * Flush requests do not use the elevator so skip initialization.
824  * This allows a request to share the flush and elevator data.
825  */
826  if (bio->bi_rw & (REQ_FLUSH | REQ_FUA))
827  return false;
828 
829  return true;
830 }
831 
839 static struct io_context *rq_ioc(struct bio *bio)
840 {
841 #ifdef CONFIG_BLK_CGROUP
842  if (bio && bio->bi_ioc)
843  return bio->bi_ioc;
844 #endif
845  return current->io_context;
846 }
847 
862 static struct request *__get_request(struct request_list *rl, int rw_flags,
863  struct bio *bio, gfp_t gfp_mask)
864 {
865  struct request_queue *q = rl->q;
866  struct request *rq;
867  struct elevator_type *et = q->elevator->type;
868  struct io_context *ioc = rq_ioc(bio);
869  struct io_cq *icq = NULL;
870  const bool is_sync = rw_is_sync(rw_flags) != 0;
871  int may_queue;
872 
873  if (unlikely(blk_queue_dead(q)))
874  return NULL;
875 
876  may_queue = elv_may_queue(q, rw_flags);
877  if (may_queue == ELV_MQUEUE_NO)
878  goto rq_starved;
879 
880  if (rl->count[is_sync]+1 >= queue_congestion_on_threshold(q)) {
881  if (rl->count[is_sync]+1 >= q->nr_requests) {
882  /*
883  * The queue will fill after this allocation, so set
884  * it as full, and mark this process as "batching".
885  * This process will be allowed to complete a batch of
886  * requests, others will be blocked.
887  */
888  if (!blk_rl_full(rl, is_sync)) {
889  ioc_set_batching(q, ioc);
890  blk_set_rl_full(rl, is_sync);
891  } else {
892  if (may_queue != ELV_MQUEUE_MUST
893  && !ioc_batching(q, ioc)) {
894  /*
895  * The queue is full and the allocating
896  * process is not a "batcher", and not
897  * exempted by the IO scheduler
898  */
899  return NULL;
900  }
901  }
902  }
903  /*
904  * bdi isn't aware of blkcg yet. As all async IOs end up
905  * root blkcg anyway, just use root blkcg state.
906  */
907  if (rl == &q->root_rl)
908  blk_set_queue_congested(q, is_sync);
909  }
910 
911  /*
912  * Only allow batching queuers to allocate up to 50% over the defined
913  * limit of requests, otherwise we could have thousands of requests
914  * allocated with any setting of ->nr_requests
915  */
916  if (rl->count[is_sync] >= (3 * q->nr_requests / 2))
917  return NULL;
918 
919  q->nr_rqs[is_sync]++;
920  rl->count[is_sync]++;
921  rl->starved[is_sync] = 0;
922 
923  /*
924  * Decide whether the new request will be managed by elevator. If
925  * so, mark @rw_flags and increment elvpriv. Non-zero elvpriv will
926  * prevent the current elevator from being destroyed until the new
927  * request is freed. This guarantees icq's won't be destroyed and
928  * makes creating new ones safe.
929  *
930  * Also, lookup icq while holding queue_lock. If it doesn't exist,
931  * it will be created after releasing queue_lock.
932  */
933  if (blk_rq_should_init_elevator(bio) && !blk_queue_bypass(q)) {
934  rw_flags |= REQ_ELVPRIV;
935  q->nr_rqs_elvpriv++;
936  if (et->icq_cache && ioc)
937  icq = ioc_lookup_icq(ioc, q);
938  }
939 
940  if (blk_queue_io_stat(q))
941  rw_flags |= REQ_IO_STAT;
942  spin_unlock_irq(q->queue_lock);
943 
944  /* allocate and init request */
945  rq = mempool_alloc(rl->rq_pool, gfp_mask);
946  if (!rq)
947  goto fail_alloc;
948 
949  blk_rq_init(q, rq);
950  blk_rq_set_rl(rq, rl);
951  rq->cmd_flags = rw_flags | REQ_ALLOCED;
952 
953  /* init elvpriv */
954  if (rw_flags & REQ_ELVPRIV) {
955  if (unlikely(et->icq_cache && !icq)) {
956  if (ioc)
957  icq = ioc_create_icq(ioc, q, gfp_mask);
958  if (!icq)
959  goto fail_elvpriv;
960  }
961 
962  rq->elv.icq = icq;
963  if (unlikely(elv_set_request(q, rq, bio, gfp_mask)))
964  goto fail_elvpriv;
965 
966  /* @rq->elv.icq holds io_context until @rq is freed */
967  if (icq)
968  get_io_context(icq->ioc);
969  }
970 out:
971  /*
972  * ioc may be NULL here, and ioc_batching will be false. That's
973  * OK, if the queue is under the request limit then requests need
974  * not count toward the nr_batch_requests limit. There will always
975  * be some limit enforced by BLK_BATCH_TIME.
976  */
977  if (ioc_batching(q, ioc))
978  ioc->nr_batch_requests--;
979 
980  trace_block_getrq(q, bio, rw_flags & 1);
981  return rq;
982 
983 fail_elvpriv:
984  /*
985  * elvpriv init failed. ioc, icq and elvpriv aren't mempool backed
986  * and may fail indefinitely under memory pressure and thus
987  * shouldn't stall IO. Treat this request as !elvpriv. This will
988  * disturb iosched and blkcg but weird is bettern than dead.
989  */
990  printk_ratelimited(KERN_WARNING "%s: request aux data allocation failed, iosched may be disturbed\n",
991  dev_name(q->backing_dev_info.dev));
992 
993  rq->cmd_flags &= ~REQ_ELVPRIV;
994  rq->elv.icq = NULL;
995 
996  spin_lock_irq(q->queue_lock);
997  q->nr_rqs_elvpriv--;
998  spin_unlock_irq(q->queue_lock);
999  goto out;
1000 
1001 fail_alloc:
1002  /*
1003  * Allocation failed presumably due to memory. Undo anything we
1004  * might have messed up.
1005  *
1006  * Allocating task should really be put onto the front of the wait
1007  * queue, but this is pretty rare.
1008  */
1009  spin_lock_irq(q->queue_lock);
1010  freed_request(rl, rw_flags);
1011 
1012  /*
1013  * in the very unlikely event that allocation failed and no
1014  * requests for this direction was pending, mark us starved so that
1015  * freeing of a request in the other direction will notice
1016  * us. another possible fix would be to split the rq mempool into
1017  * READ and WRITE
1018  */
1019 rq_starved:
1020  if (unlikely(rl->count[is_sync] == 0))
1021  rl->starved[is_sync] = 1;
1022  return NULL;
1023 }
1024 
1039 static struct request *get_request(struct request_queue *q, int rw_flags,
1040  struct bio *bio, gfp_t gfp_mask)
1041 {
1042  const bool is_sync = rw_is_sync(rw_flags) != 0;
1043  DEFINE_WAIT(wait);
1044  struct request_list *rl;
1045  struct request *rq;
1046 
1047  rl = blk_get_rl(q, bio); /* transferred to @rq on success */
1048 retry:
1049  rq = __get_request(rl, rw_flags, bio, gfp_mask);
1050  if (rq)
1051  return rq;
1052 
1053  if (!(gfp_mask & __GFP_WAIT) || unlikely(blk_queue_dead(q))) {
1054  blk_put_rl(rl);
1055  return NULL;
1056  }
1057 
1058  /* wait on @rl and retry */
1059  prepare_to_wait_exclusive(&rl->wait[is_sync], &wait,
1061 
1062  trace_block_sleeprq(q, bio, rw_flags & 1);
1063 
1064  spin_unlock_irq(q->queue_lock);
1065  io_schedule();
1066 
1067  /*
1068  * After sleeping, we become a "batching" process and will be able
1069  * to allocate at least one request, and up to a big batch of them
1070  * for a small period time. See ioc_batching, ioc_set_batching
1071  */
1072  ioc_set_batching(q, current->io_context);
1073 
1074  spin_lock_irq(q->queue_lock);
1075  finish_wait(&rl->wait[is_sync], &wait);
1076 
1077  goto retry;
1078 }
1079 
1080 struct request *blk_get_request(struct request_queue *q, int rw, gfp_t gfp_mask)
1081 {
1082  struct request *rq;
1083 
1084  BUG_ON(rw != READ && rw != WRITE);
1085 
1086  /* create ioc upfront */
1087  create_io_context(gfp_mask, q->node);
1088 
1089  spin_lock_irq(q->queue_lock);
1090  rq = get_request(q, rw, NULL, gfp_mask);
1091  if (!rq)
1092  spin_unlock_irq(q->queue_lock);
1093  /* q->queue_lock is unlocked at this point */
1094 
1095  return rq;
1096 }
1098 
1130 struct request *blk_make_request(struct request_queue *q, struct bio *bio,
1131  gfp_t gfp_mask)
1132 {
1133  struct request *rq = blk_get_request(q, bio_data_dir(bio), gfp_mask);
1134 
1135  if (unlikely(!rq))
1136  return ERR_PTR(-ENOMEM);
1137 
1138  for_each_bio(bio) {
1139  struct bio *bounce_bio = bio;
1140  int ret;
1141 
1142  blk_queue_bounce(q, &bounce_bio);
1143  ret = blk_rq_append_bio(q, rq, bounce_bio);
1144  if (unlikely(ret)) {
1145  blk_put_request(rq);
1146  return ERR_PTR(ret);
1147  }
1148  }
1149 
1150  return rq;
1151 }
1153 
1164 void blk_requeue_request(struct request_queue *q, struct request *rq)
1165 {
1166  blk_delete_timer(rq);
1167  blk_clear_rq_complete(rq);
1168  trace_block_rq_requeue(q, rq);
1169 
1170  if (blk_rq_tagged(rq))
1171  blk_queue_end_tag(q, rq);
1172 
1173  BUG_ON(blk_queued_rq(rq));
1174 
1175  elv_requeue_request(q, rq);
1176 }
1178 
1179 static void add_acct_request(struct request_queue *q, struct request *rq,
1180  int where)
1181 {
1182  drive_stat_acct(rq, 1);
1183  __elv_add_request(q, rq, where);
1184 }
1185 
1186 static void part_round_stats_single(int cpu, struct hd_struct *part,
1187  unsigned long now)
1188 {
1189  if (now == part->stamp)
1190  return;
1191 
1192  if (part_in_flight(part)) {
1193  __part_stat_add(cpu, part, time_in_queue,
1194  part_in_flight(part) * (now - part->stamp));
1195  __part_stat_add(cpu, part, io_ticks, (now - part->stamp));
1196  }
1197  part->stamp = now;
1198 }
1199 
1216 void part_round_stats(int cpu, struct hd_struct *part)
1217 {
1218  unsigned long now = jiffies;
1219 
1220  if (part->partno)
1221  part_round_stats_single(cpu, &part_to_disk(part)->part0, now);
1222  part_round_stats_single(cpu, part, now);
1223 }
1225 
1226 /*
1227  * queue lock must be held
1228  */
1229 void __blk_put_request(struct request_queue *q, struct request *req)
1230 {
1231  if (unlikely(!q))
1232  return;
1233  if (unlikely(--req->ref_count))
1234  return;
1235 
1236  elv_completed_request(q, req);
1237 
1238  /* this is a bio leak */
1239  WARN_ON(req->bio != NULL);
1240 
1241  /*
1242  * Request may not have originated from ll_rw_blk. if not,
1243  * it didn't come out of our reserved rq pools
1244  */
1245  if (req->cmd_flags & REQ_ALLOCED) {
1246  unsigned int flags = req->cmd_flags;
1247  struct request_list *rl = blk_rq_rl(req);
1248 
1249  BUG_ON(!list_empty(&req->queuelist));
1250  BUG_ON(!hlist_unhashed(&req->hash));
1251 
1252  blk_free_request(rl, req);
1253  freed_request(rl, flags);
1254  blk_put_rl(rl);
1255  }
1256 }
1258 
1260 {
1261  unsigned long flags;
1262  struct request_queue *q = req->q;
1263 
1264  spin_lock_irqsave(q->queue_lock, flags);
1265  __blk_put_request(q, req);
1266  spin_unlock_irqrestore(q->queue_lock, flags);
1267 }
1269 
1283 void blk_add_request_payload(struct request *rq, struct page *page,
1284  unsigned int len)
1285 {
1286  struct bio *bio = rq->bio;
1287 
1288  bio->bi_io_vec->bv_page = page;
1289  bio->bi_io_vec->bv_offset = 0;
1290  bio->bi_io_vec->bv_len = len;
1291 
1292  bio->bi_size = len;
1293  bio->bi_vcnt = 1;
1294  bio->bi_phys_segments = 1;
1295 
1296  rq->__data_len = rq->resid_len = len;
1297  rq->nr_phys_segments = 1;
1298  rq->buffer = bio_data(bio);
1299 }
1301 
1302 static bool bio_attempt_back_merge(struct request_queue *q, struct request *req,
1303  struct bio *bio)
1304 {
1305  const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1306 
1307  if (!ll_back_merge_fn(q, req, bio))
1308  return false;
1309 
1310  trace_block_bio_backmerge(q, bio);
1311 
1312  if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1314 
1315  req->biotail->bi_next = bio;
1316  req->biotail = bio;
1317  req->__data_len += bio->bi_size;
1318  req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1319 
1320  drive_stat_acct(req, 0);
1321  return true;
1322 }
1323 
1324 static bool bio_attempt_front_merge(struct request_queue *q,
1325  struct request *req, struct bio *bio)
1326 {
1327  const int ff = bio->bi_rw & REQ_FAILFAST_MASK;
1328 
1329  if (!ll_front_merge_fn(q, req, bio))
1330  return false;
1331 
1332  trace_block_bio_frontmerge(q, bio);
1333 
1334  if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
1336 
1337  bio->bi_next = req->bio;
1338  req->bio = bio;
1339 
1340  /*
1341  * may not be valid. if the low level driver said
1342  * it didn't need a bounce buffer then it better
1343  * not touch req->buffer either...
1344  */
1345  req->buffer = bio_data(bio);
1346  req->__sector = bio->bi_sector;
1347  req->__data_len += bio->bi_size;
1348  req->ioprio = ioprio_best(req->ioprio, bio_prio(bio));
1349 
1350  drive_stat_acct(req, 0);
1351  return true;
1352 }
1353 
1371 static bool attempt_plug_merge(struct request_queue *q, struct bio *bio,
1372  unsigned int *request_count)
1373 {
1374  struct blk_plug *plug;
1375  struct request *rq;
1376  bool ret = false;
1377 
1378  plug = current->plug;
1379  if (!plug)
1380  goto out;
1381  *request_count = 0;
1382 
1383  list_for_each_entry_reverse(rq, &plug->list, queuelist) {
1384  int el_ret;
1385 
1386  if (rq->q == q)
1387  (*request_count)++;
1388 
1389  if (rq->q != q || !blk_rq_merge_ok(rq, bio))
1390  continue;
1391 
1392  el_ret = blk_try_merge(rq, bio);
1393  if (el_ret == ELEVATOR_BACK_MERGE) {
1394  ret = bio_attempt_back_merge(q, rq, bio);
1395  if (ret)
1396  break;
1397  } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1398  ret = bio_attempt_front_merge(q, rq, bio);
1399  if (ret)
1400  break;
1401  }
1402  }
1403 out:
1404  return ret;
1405 }
1406 
1407 void init_request_from_bio(struct request *req, struct bio *bio)
1408 {
1409  req->cmd_type = REQ_TYPE_FS;
1410 
1411  req->cmd_flags |= bio->bi_rw & REQ_COMMON_MASK;
1412  if (bio->bi_rw & REQ_RAHEAD)
1413  req->cmd_flags |= REQ_FAILFAST_MASK;
1414 
1415  req->errors = 0;
1416  req->__sector = bio->bi_sector;
1417  req->ioprio = bio_prio(bio);
1418  blk_rq_bio_prep(req->q, req, bio);
1419 }
1420 
1421 void blk_queue_bio(struct request_queue *q, struct bio *bio)
1422 {
1423  const bool sync = !!(bio->bi_rw & REQ_SYNC);
1424  struct blk_plug *plug;
1425  int el_ret, rw_flags, where = ELEVATOR_INSERT_SORT;
1426  struct request *req;
1427  unsigned int request_count = 0;
1428 
1429  /*
1430  * low level driver can indicate that it wants pages above a
1431  * certain limit bounced to low memory (ie for highmem, or even
1432  * ISA dma in theory)
1433  */
1434  blk_queue_bounce(q, &bio);
1435 
1436  if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
1437  spin_lock_irq(q->queue_lock);
1438  where = ELEVATOR_INSERT_FLUSH;
1439  goto get_rq;
1440  }
1441 
1442  /*
1443  * Check if we can merge with the plugged list before grabbing
1444  * any locks.
1445  */
1446  if (attempt_plug_merge(q, bio, &request_count))
1447  return;
1448 
1449  spin_lock_irq(q->queue_lock);
1450 
1451  el_ret = elv_merge(q, &req, bio);
1452  if (el_ret == ELEVATOR_BACK_MERGE) {
1453  if (bio_attempt_back_merge(q, req, bio)) {
1454  elv_bio_merged(q, req, bio);
1455  if (!attempt_back_merge(q, req))
1456  elv_merged_request(q, req, el_ret);
1457  goto out_unlock;
1458  }
1459  } else if (el_ret == ELEVATOR_FRONT_MERGE) {
1460  if (bio_attempt_front_merge(q, req, bio)) {
1461  elv_bio_merged(q, req, bio);
1462  if (!attempt_front_merge(q, req))
1463  elv_merged_request(q, req, el_ret);
1464  goto out_unlock;
1465  }
1466  }
1467 
1468 get_rq:
1469  /*
1470  * This sync check and mask will be re-done in init_request_from_bio(),
1471  * but we need to set it earlier to expose the sync flag to the
1472  * rq allocator and io schedulers.
1473  */
1474  rw_flags = bio_data_dir(bio);
1475  if (sync)
1476  rw_flags |= REQ_SYNC;
1477 
1478  /*
1479  * Grab a free request. This is might sleep but can not fail.
1480  * Returns with the queue unlocked.
1481  */
1482  req = get_request(q, rw_flags, bio, GFP_NOIO);
1483  if (unlikely(!req)) {
1484  bio_endio(bio, -ENODEV); /* @q is dead */
1485  goto out_unlock;
1486  }
1487 
1488  /*
1489  * After dropping the lock and possibly sleeping here, our request
1490  * may now be mergeable after it had proven unmergeable (above).
1491  * We don't worry about that case for efficiency. It won't happen
1492  * often, and the elevators are able to handle it.
1493  */
1494  init_request_from_bio(req, bio);
1495 
1496  if (test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags))
1497  req->cpu = raw_smp_processor_id();
1498 
1499  plug = current->plug;
1500  if (plug) {
1501  /*
1502  * If this is the first request added after a plug, fire
1503  * of a plug trace. If others have been added before, check
1504  * if we have multiple devices in this plug. If so, make a
1505  * note to sort the list before dispatch.
1506  */
1507  if (list_empty(&plug->list))
1508  trace_block_plug(q);
1509  else {
1510  if (!plug->should_sort) {
1511  struct request *__rq;
1512 
1513  __rq = list_entry_rq(plug->list.prev);
1514  if (__rq->q != q)
1515  plug->should_sort = 1;
1516  }
1517  if (request_count >= BLK_MAX_REQUEST_COUNT) {
1518  blk_flush_plug_list(plug, false);
1519  trace_block_plug(q);
1520  }
1521  }
1522  list_add_tail(&req->queuelist, &plug->list);
1523  drive_stat_acct(req, 1);
1524  } else {
1525  spin_lock_irq(q->queue_lock);
1526  add_acct_request(q, req, where);
1527  __blk_run_queue(q);
1528 out_unlock:
1529  spin_unlock_irq(q->queue_lock);
1530  }
1531 }
1532 EXPORT_SYMBOL_GPL(blk_queue_bio); /* for device mapper only */
1533 
1534 /*
1535  * If bio->bi_dev is a partition, remap the location
1536  */
1537 static inline void blk_partition_remap(struct bio *bio)
1538 {
1539  struct block_device *bdev = bio->bi_bdev;
1540 
1541  if (bio_sectors(bio) && bdev != bdev->bd_contains) {
1542  struct hd_struct *p = bdev->bd_part;
1543 
1544  bio->bi_sector += p->start_sect;
1545  bio->bi_bdev = bdev->bd_contains;
1546 
1547  trace_block_bio_remap(bdev_get_queue(bio->bi_bdev), bio,
1548  bdev->bd_dev,
1549  bio->bi_sector - p->start_sect);
1550  }
1551 }
1552 
1553 static void handle_bad_sector(struct bio *bio)
1554 {
1555  char b[BDEVNAME_SIZE];
1556 
1557  printk(KERN_INFO "attempt to access beyond end of device\n");
1558  printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
1559  bdevname(bio->bi_bdev, b),
1560  bio->bi_rw,
1561  (unsigned long long)bio->bi_sector + bio_sectors(bio),
1562  (long long)(i_size_read(bio->bi_bdev->bd_inode) >> 9));
1563 
1564  set_bit(BIO_EOF, &bio->bi_flags);
1565 }
1566 
1567 #ifdef CONFIG_FAIL_MAKE_REQUEST
1568 
1569 static DECLARE_FAULT_ATTR(fail_make_request);
1570 
1571 static int __init setup_fail_make_request(char *str)
1572 {
1573  return setup_fault_attr(&fail_make_request, str);
1574 }
1575 __setup("fail_make_request=", setup_fail_make_request);
1576 
1577 static bool should_fail_request(struct hd_struct *part, unsigned int bytes)
1578 {
1579  return part->make_it_fail && should_fail(&fail_make_request, bytes);
1580 }
1581 
1582 static int __init fail_make_request_debugfs(void)
1583 {
1584  struct dentry *dir = fault_create_debugfs_attr("fail_make_request",
1585  NULL, &fail_make_request);
1586 
1587  return IS_ERR(dir) ? PTR_ERR(dir) : 0;
1588 }
1589 
1590 late_initcall(fail_make_request_debugfs);
1591 
1592 #else /* CONFIG_FAIL_MAKE_REQUEST */
1593 
1594 static inline bool should_fail_request(struct hd_struct *part,
1595  unsigned int bytes)
1596 {
1597  return false;
1598 }
1599 
1600 #endif /* CONFIG_FAIL_MAKE_REQUEST */
1601 
1602 /*
1603  * Check whether this bio extends beyond the end of the device.
1604  */
1605 static inline int bio_check_eod(struct bio *bio, unsigned int nr_sectors)
1606 {
1607  sector_t maxsector;
1608 
1609  if (!nr_sectors)
1610  return 0;
1611 
1612  /* Test device or partition size, when known. */
1613  maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
1614  if (maxsector) {
1615  sector_t sector = bio->bi_sector;
1616 
1617  if (maxsector < nr_sectors || maxsector - nr_sectors < sector) {
1618  /*
1619  * This may well happen - the kernel calls bread()
1620  * without checking the size of the device, e.g., when
1621  * mounting a device.
1622  */
1623  handle_bad_sector(bio);
1624  return 1;
1625  }
1626  }
1627 
1628  return 0;
1629 }
1630 
1631 static noinline_for_stack bool
1632 generic_make_request_checks(struct bio *bio)
1633 {
1634  struct request_queue *q;
1635  int nr_sectors = bio_sectors(bio);
1636  int err = -EIO;
1637  char b[BDEVNAME_SIZE];
1638  struct hd_struct *part;
1639 
1640  might_sleep();
1641 
1642  if (bio_check_eod(bio, nr_sectors))
1643  goto end_io;
1644 
1645  q = bdev_get_queue(bio->bi_bdev);
1646  if (unlikely(!q)) {
1648  "generic_make_request: Trying to access "
1649  "nonexistent block-device %s (%Lu)\n",
1650  bdevname(bio->bi_bdev, b),
1651  (long long) bio->bi_sector);
1652  goto end_io;
1653  }
1654 
1655  if (likely(bio_is_rw(bio) &&
1656  nr_sectors > queue_max_hw_sectors(q))) {
1657  printk(KERN_ERR "bio too big device %s (%u > %u)\n",
1658  bdevname(bio->bi_bdev, b),
1659  bio_sectors(bio),
1660  queue_max_hw_sectors(q));
1661  goto end_io;
1662  }
1663 
1664  part = bio->bi_bdev->bd_part;
1665  if (should_fail_request(part, bio->bi_size) ||
1666  should_fail_request(&part_to_disk(part)->part0,
1667  bio->bi_size))
1668  goto end_io;
1669 
1670  /*
1671  * If this device has partitions, remap block n
1672  * of partition p to block n+start(p) of the disk.
1673  */
1674  blk_partition_remap(bio);
1675 
1676  if (bio_integrity_enabled(bio) && bio_integrity_prep(bio))
1677  goto end_io;
1678 
1679  if (bio_check_eod(bio, nr_sectors))
1680  goto end_io;
1681 
1682  /*
1683  * Filter flush bio's early so that make_request based
1684  * drivers without flush support don't have to worry
1685  * about them.
1686  */
1687  if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) && !q->flush_flags) {
1688  bio->bi_rw &= ~(REQ_FLUSH | REQ_FUA);
1689  if (!nr_sectors) {
1690  err = 0;
1691  goto end_io;
1692  }
1693  }
1694 
1695  if ((bio->bi_rw & REQ_DISCARD) &&
1696  (!blk_queue_discard(q) ||
1697  ((bio->bi_rw & REQ_SECURE) && !blk_queue_secdiscard(q)))) {
1698  err = -EOPNOTSUPP;
1699  goto end_io;
1700  }
1701 
1702  if (bio->bi_rw & REQ_WRITE_SAME && !bdev_write_same(bio->bi_bdev)) {
1703  err = -EOPNOTSUPP;
1704  goto end_io;
1705  }
1706 
1707  /*
1708  * Various block parts want %current->io_context and lazy ioc
1709  * allocation ends up trading a lot of pain for a small amount of
1710  * memory. Just allocate it upfront. This may fail and block
1711  * layer knows how to live with it.
1712  */
1713  create_io_context(GFP_ATOMIC, q->node);
1714 
1715  if (blk_throtl_bio(q, bio))
1716  return false; /* throttled, will be resubmitted later */
1717 
1718  trace_block_bio_queue(q, bio);
1719  return true;
1720 
1721 end_io:
1722  bio_endio(bio, err);
1723  return false;
1724 }
1725 
1750 void generic_make_request(struct bio *bio)
1751 {
1752  struct bio_list bio_list_on_stack;
1753 
1754  if (!generic_make_request_checks(bio))
1755  return;
1756 
1757  /*
1758  * We only want one ->make_request_fn to be active at a time, else
1759  * stack usage with stacked devices could be a problem. So use
1760  * current->bio_list to keep a list of requests submited by a
1761  * make_request_fn function. current->bio_list is also used as a
1762  * flag to say if generic_make_request is currently active in this
1763  * task or not. If it is NULL, then no make_request is active. If
1764  * it is non-NULL, then a make_request is active, and new requests
1765  * should be added at the tail
1766  */
1767  if (current->bio_list) {
1768  bio_list_add(current->bio_list, bio);
1769  return;
1770  }
1771 
1772  /* following loop may be a bit non-obvious, and so deserves some
1773  * explanation.
1774  * Before entering the loop, bio->bi_next is NULL (as all callers
1775  * ensure that) so we have a list with a single bio.
1776  * We pretend that we have just taken it off a longer list, so
1777  * we assign bio_list to a pointer to the bio_list_on_stack,
1778  * thus initialising the bio_list of new bios to be
1779  * added. ->make_request() may indeed add some more bios
1780  * through a recursive call to generic_make_request. If it
1781  * did, we find a non-NULL value in bio_list and re-enter the loop
1782  * from the top. In this case we really did just take the bio
1783  * of the top of the list (no pretending) and so remove it from
1784  * bio_list, and call into ->make_request() again.
1785  */
1786  BUG_ON(bio->bi_next);
1787  bio_list_init(&bio_list_on_stack);
1788  current->bio_list = &bio_list_on_stack;
1789  do {
1790  struct request_queue *q = bdev_get_queue(bio->bi_bdev);
1791 
1792  q->make_request_fn(q, bio);
1793 
1794  bio = bio_list_pop(current->bio_list);
1795  } while (bio);
1796  current->bio_list = NULL; /* deactivate */
1797 }
1799 
1810 void submit_bio(int rw, struct bio *bio)
1811 {
1812  bio->bi_rw |= rw;
1813 
1814  /*
1815  * If it's a regular read/write or a barrier with data attached,
1816  * go through the normal accounting stuff before submission.
1817  */
1818  if (bio_has_data(bio)) {
1819  unsigned int count;
1820 
1821  if (unlikely(rw & REQ_WRITE_SAME))
1822  count = bdev_logical_block_size(bio->bi_bdev) >> 9;
1823  else
1824  count = bio_sectors(bio);
1825 
1826  if (rw & WRITE) {
1827  count_vm_events(PGPGOUT, count);
1828  } else {
1829  task_io_account_read(bio->bi_size);
1830  count_vm_events(PGPGIN, count);
1831  }
1832 
1833  if (unlikely(block_dump)) {
1834  char b[BDEVNAME_SIZE];
1835  printk(KERN_DEBUG "%s(%d): %s block %Lu on %s (%u sectors)\n",
1836  current->comm, task_pid_nr(current),
1837  (rw & WRITE) ? "WRITE" : "READ",
1838  (unsigned long long)bio->bi_sector,
1839  bdevname(bio->bi_bdev, b),
1840  count);
1841  }
1842  }
1843 
1844  generic_make_request(bio);
1845 }
1847 
1869 int blk_rq_check_limits(struct request_queue *q, struct request *rq)
1870 {
1871  if (!rq_mergeable(rq))
1872  return 0;
1873 
1874  if (blk_rq_sectors(rq) > blk_queue_get_max_sectors(q, rq->cmd_flags)) {
1875  printk(KERN_ERR "%s: over max size limit.\n", __func__);
1876  return -EIO;
1877  }
1878 
1879  /*
1880  * queue's settings related to segment counting like q->bounce_pfn
1881  * may differ from that of other stacking queues.
1882  * Recalculate it to check the request correctly on this queue's
1883  * limitation.
1884  */
1886  if (rq->nr_phys_segments > queue_max_segments(q)) {
1887  printk(KERN_ERR "%s: over max segments limit.\n", __func__);
1888  return -EIO;
1889  }
1890 
1891  return 0;
1892 }
1894 
1901 {
1902  unsigned long flags;
1903  int where = ELEVATOR_INSERT_BACK;
1904 
1905  if (blk_rq_check_limits(q, rq))
1906  return -EIO;
1907 
1908  if (rq->rq_disk &&
1909  should_fail_request(&rq->rq_disk->part0, blk_rq_bytes(rq)))
1910  return -EIO;
1911 
1912  spin_lock_irqsave(q->queue_lock, flags);
1913  if (unlikely(blk_queue_dead(q))) {
1914  spin_unlock_irqrestore(q->queue_lock, flags);
1915  return -ENODEV;
1916  }
1917 
1918  /*
1919  * Submitting request must be dequeued before calling this function
1920  * because it will be linked to another request_queue
1921  */
1922  BUG_ON(blk_queued_rq(rq));
1923 
1924  if (rq->cmd_flags & (REQ_FLUSH|REQ_FUA))
1925  where = ELEVATOR_INSERT_FLUSH;
1926 
1927  add_acct_request(q, rq, where);
1928  if (where == ELEVATOR_INSERT_FLUSH)
1929  __blk_run_queue(q);
1930  spin_unlock_irqrestore(q->queue_lock, flags);
1931 
1932  return 0;
1933 }
1935 
1952 unsigned int blk_rq_err_bytes(const struct request *rq)
1953 {
1954  unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
1955  unsigned int bytes = 0;
1956  struct bio *bio;
1957 
1958  if (!(rq->cmd_flags & REQ_MIXED_MERGE))
1959  return blk_rq_bytes(rq);
1960 
1961  /*
1962  * Currently the only 'mixing' which can happen is between
1963  * different fastfail types. We can safely fail portions
1964  * which have all the failfast bits that the first one has -
1965  * the ones which are at least as eager to fail as the first
1966  * one.
1967  */
1968  for (bio = rq->bio; bio; bio = bio->bi_next) {
1969  if ((bio->bi_rw & ff) != ff)
1970  break;
1971  bytes += bio->bi_size;
1972  }
1973 
1974  /* this could lead to infinite loop */
1975  BUG_ON(blk_rq_bytes(rq) && !bytes);
1976  return bytes;
1977 }
1979 
1980 static void blk_account_io_completion(struct request *req, unsigned int bytes)
1981 {
1982  if (blk_do_io_stat(req)) {
1983  const int rw = rq_data_dir(req);
1984  struct hd_struct *part;
1985  int cpu;
1986 
1987  cpu = part_stat_lock();
1988  part = req->part;
1989  part_stat_add(cpu, part, sectors[rw], bytes >> 9);
1990  part_stat_unlock();
1991  }
1992 }
1993 
1994 static void blk_account_io_done(struct request *req)
1995 {
1996  /*
1997  * Account IO completion. flush_rq isn't accounted as a
1998  * normal IO on queueing nor completion. Accounting the
1999  * containing request is enough.
2000  */
2001  if (blk_do_io_stat(req) && !(req->cmd_flags & REQ_FLUSH_SEQ)) {
2002  unsigned long duration = jiffies - req->start_time;
2003  const int rw = rq_data_dir(req);
2004  struct hd_struct *part;
2005  int cpu;
2006 
2007  cpu = part_stat_lock();
2008  part = req->part;
2009 
2010  part_stat_inc(cpu, part, ios[rw]);
2011  part_stat_add(cpu, part, ticks[rw], duration);
2012  part_round_stats(cpu, part);
2013  part_dec_in_flight(part, rw);
2014 
2015  hd_struct_put(part);
2016  part_stat_unlock();
2017  }
2018 }
2019 
2037 {
2038  struct request *rq;
2039  int ret;
2040 
2041  while ((rq = __elv_next_request(q)) != NULL) {
2042  if (!(rq->cmd_flags & REQ_STARTED)) {
2043  /*
2044  * This is the first time the device driver
2045  * sees this request (possibly after
2046  * requeueing). Notify IO scheduler.
2047  */
2048  if (rq->cmd_flags & REQ_SORTED)
2049  elv_activate_rq(q, rq);
2050 
2051  /*
2052  * just mark as started even if we don't start
2053  * it, a request that has been delayed should
2054  * not be passed by new incoming requests
2055  */
2056  rq->cmd_flags |= REQ_STARTED;
2057  trace_block_rq_issue(q, rq);
2058  }
2059 
2060  if (!q->boundary_rq || q->boundary_rq == rq) {
2061  q->end_sector = rq_end_sector(rq);
2062  q->boundary_rq = NULL;
2063  }
2064 
2065  if (rq->cmd_flags & REQ_DONTPREP)
2066  break;
2067 
2068  if (q->dma_drain_size && blk_rq_bytes(rq)) {
2069  /*
2070  * make sure space for the drain appears we
2071  * know we can do this because max_hw_segments
2072  * has been adjusted to be one fewer than the
2073  * device can handle
2074  */
2075  rq->nr_phys_segments++;
2076  }
2077 
2078  if (!q->prep_rq_fn)
2079  break;
2080 
2081  ret = q->prep_rq_fn(q, rq);
2082  if (ret == BLKPREP_OK) {
2083  break;
2084  } else if (ret == BLKPREP_DEFER) {
2085  /*
2086  * the request may have been (partially) prepped.
2087  * we need to keep this request in the front to
2088  * avoid resource deadlock. REQ_STARTED will
2089  * prevent other fs requests from passing this one.
2090  */
2091  if (q->dma_drain_size && blk_rq_bytes(rq) &&
2092  !(rq->cmd_flags & REQ_DONTPREP)) {
2093  /*
2094  * remove the space for the drain we added
2095  * so that we don't add it again
2096  */
2097  --rq->nr_phys_segments;
2098  }
2099 
2100  rq = NULL;
2101  break;
2102  } else if (ret == BLKPREP_KILL) {
2103  rq->cmd_flags |= REQ_QUIET;
2104  /*
2105  * Mark this request as started so we don't trigger
2106  * any debug logic in the end I/O path.
2107  */
2108  blk_start_request(rq);
2109  __blk_end_request_all(rq, -EIO);
2110  } else {
2111  printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
2112  break;
2113  }
2114  }
2115 
2116  return rq;
2117 }
2119 
2121 {
2122  struct request_queue *q = rq->q;
2123 
2124  BUG_ON(list_empty(&rq->queuelist));
2125  BUG_ON(ELV_ON_HASH(rq));
2126 
2127  list_del_init(&rq->queuelist);
2128 
2129  /*
2130  * the time frame between a request being removed from the lists
2131  * and to it is freed is accounted as io that is in progress at
2132  * the driver side.
2133  */
2134  if (blk_account_rq(rq)) {
2135  q->in_flight[rq_is_sync(rq)]++;
2136  set_io_start_time_ns(rq);
2137  }
2138 }
2139 
2154 void blk_start_request(struct request *req)
2155 {
2156  blk_dequeue_request(req);
2157 
2158  /*
2159  * We are now handing the request to the hardware, initialize
2160  * resid_len to full count and add the timeout handler.
2161  */
2162  req->resid_len = blk_rq_bytes(req);
2163  if (unlikely(blk_bidi_rq(req)))
2164  req->next_rq->resid_len = blk_rq_bytes(req->next_rq);
2165 
2166  blk_add_timer(req);
2167 }
2169 
2186 {
2187  struct request *rq;
2188 
2189  rq = blk_peek_request(q);
2190  if (rq)
2191  blk_start_request(rq);
2192  return rq;
2193 }
2195 
2218 bool blk_update_request(struct request *req, int error, unsigned int nr_bytes)
2219 {
2220  int total_bytes, bio_nbytes, next_idx = 0;
2221  struct bio *bio;
2222 
2223  if (!req->bio)
2224  return false;
2225 
2226  trace_block_rq_complete(req->q, req);
2227 
2228  /*
2229  * For fs requests, rq is just carrier of independent bio's
2230  * and each partial completion should be handled separately.
2231  * Reset per-request error on each partial completion.
2232  *
2233  * TODO: tj: This is too subtle. It would be better to let
2234  * low level drivers do what they see fit.
2235  */
2236  if (req->cmd_type == REQ_TYPE_FS)
2237  req->errors = 0;
2238 
2239  if (error && req->cmd_type == REQ_TYPE_FS &&
2240  !(req->cmd_flags & REQ_QUIET)) {
2241  char *error_type;
2242 
2243  switch (error) {
2244  case -ENOLINK:
2245  error_type = "recoverable transport";
2246  break;
2247  case -EREMOTEIO:
2248  error_type = "critical target";
2249  break;
2250  case -EBADE:
2251  error_type = "critical nexus";
2252  break;
2253  case -EIO:
2254  default:
2255  error_type = "I/O";
2256  break;
2257  }
2258  printk_ratelimited(KERN_ERR "end_request: %s error, dev %s, sector %llu\n",
2259  error_type, req->rq_disk ?
2260  req->rq_disk->disk_name : "?",
2261  (unsigned long long)blk_rq_pos(req));
2262 
2263  }
2264 
2265  blk_account_io_completion(req, nr_bytes);
2266 
2267  total_bytes = bio_nbytes = 0;
2268  while ((bio = req->bio) != NULL) {
2269  int nbytes;
2270 
2271  if (nr_bytes >= bio->bi_size) {
2272  req->bio = bio->bi_next;
2273  nbytes = bio->bi_size;
2274  req_bio_endio(req, bio, nbytes, error);
2275  next_idx = 0;
2276  bio_nbytes = 0;
2277  } else {
2278  int idx = bio->bi_idx + next_idx;
2279 
2280  if (unlikely(idx >= bio->bi_vcnt)) {
2281  blk_dump_rq_flags(req, "__end_that");
2282  printk(KERN_ERR "%s: bio idx %d >= vcnt %d\n",
2283  __func__, idx, bio->bi_vcnt);
2284  break;
2285  }
2286 
2287  nbytes = bio_iovec_idx(bio, idx)->bv_len;
2288  BIO_BUG_ON(nbytes > bio->bi_size);
2289 
2290  /*
2291  * not a complete bvec done
2292  */
2293  if (unlikely(nbytes > nr_bytes)) {
2294  bio_nbytes += nr_bytes;
2295  total_bytes += nr_bytes;
2296  break;
2297  }
2298 
2299  /*
2300  * advance to the next vector
2301  */
2302  next_idx++;
2303  bio_nbytes += nbytes;
2304  }
2305 
2306  total_bytes += nbytes;
2307  nr_bytes -= nbytes;
2308 
2309  bio = req->bio;
2310  if (bio) {
2311  /*
2312  * end more in this run, or just return 'not-done'
2313  */
2314  if (unlikely(nr_bytes <= 0))
2315  break;
2316  }
2317  }
2318 
2319  /*
2320  * completely done
2321  */
2322  if (!req->bio) {
2323  /*
2324  * Reset counters so that the request stacking driver
2325  * can find how many bytes remain in the request
2326  * later.
2327  */
2328  req->__data_len = 0;
2329  return false;
2330  }
2331 
2332  /*
2333  * if the request wasn't completed, update state
2334  */
2335  if (bio_nbytes) {
2336  req_bio_endio(req, bio, bio_nbytes, error);
2337  bio->bi_idx += next_idx;
2338  bio_iovec(bio)->bv_offset += nr_bytes;
2339  bio_iovec(bio)->bv_len -= nr_bytes;
2340  }
2341 
2342  req->__data_len -= total_bytes;
2343  req->buffer = bio_data(req->bio);
2344 
2345  /* update sector only for requests with clear definition of sector */
2346  if (req->cmd_type == REQ_TYPE_FS)
2347  req->__sector += total_bytes >> 9;
2348 
2349  /* mixed attributes always follow the first bio */
2350  if (req->cmd_flags & REQ_MIXED_MERGE) {
2351  req->cmd_flags &= ~REQ_FAILFAST_MASK;
2352  req->cmd_flags |= req->bio->bi_rw & REQ_FAILFAST_MASK;
2353  }
2354 
2355  /*
2356  * If total number of sectors is less than the first segment
2357  * size, something has gone terribly wrong.
2358  */
2359  if (blk_rq_bytes(req) < blk_rq_cur_bytes(req)) {
2360  blk_dump_rq_flags(req, "request botched");
2361  req->__data_len = blk_rq_cur_bytes(req);
2362  }
2363 
2364  /* recalculate the number of segments */
2366 
2367  return true;
2368 }
2370 
2371 static bool blk_update_bidi_request(struct request *rq, int error,
2372  unsigned int nr_bytes,
2373  unsigned int bidi_bytes)
2374 {
2375  if (blk_update_request(rq, error, nr_bytes))
2376  return true;
2377 
2378  /* Bidi request must be completed as a whole */
2379  if (unlikely(blk_bidi_rq(rq)) &&
2380  blk_update_request(rq->next_rq, error, bidi_bytes))
2381  return true;
2382 
2383  if (blk_queue_add_random(rq->q))
2384  add_disk_randomness(rq->rq_disk);
2385 
2386  return false;
2387 }
2388 
2399 void blk_unprep_request(struct request *req)
2400 {
2401  struct request_queue *q = req->q;
2402 
2403  req->cmd_flags &= ~REQ_DONTPREP;
2404  if (q->unprep_rq_fn)
2405  q->unprep_rq_fn(q, req);
2406 }
2408 
2409 /*
2410  * queue lock must be held
2411  */
2412 static void blk_finish_request(struct request *req, int error)
2413 {
2414  if (blk_rq_tagged(req))
2415  blk_queue_end_tag(req->q, req);
2416 
2417  BUG_ON(blk_queued_rq(req));
2418 
2419  if (unlikely(laptop_mode) && req->cmd_type == REQ_TYPE_FS)
2420  laptop_io_completion(&req->q->backing_dev_info);
2421 
2422  blk_delete_timer(req);
2423 
2424  if (req->cmd_flags & REQ_DONTPREP)
2425  blk_unprep_request(req);
2426 
2427 
2428  blk_account_io_done(req);
2429 
2430  if (req->end_io)
2431  req->end_io(req, error);
2432  else {
2433  if (blk_bidi_rq(req))
2434  __blk_put_request(req->next_rq->q, req->next_rq);
2435 
2436  __blk_put_request(req->q, req);
2437  }
2438 }
2439 
2457 static bool blk_end_bidi_request(struct request *rq, int error,
2458  unsigned int nr_bytes, unsigned int bidi_bytes)
2459 {
2460  struct request_queue *q = rq->q;
2461  unsigned long flags;
2462 
2463  if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2464  return true;
2465 
2466  spin_lock_irqsave(q->queue_lock, flags);
2467  blk_finish_request(rq, error);
2468  spin_unlock_irqrestore(q->queue_lock, flags);
2469 
2470  return false;
2471 }
2472 
2488 bool __blk_end_bidi_request(struct request *rq, int error,
2489  unsigned int nr_bytes, unsigned int bidi_bytes)
2490 {
2491  if (blk_update_bidi_request(rq, error, nr_bytes, bidi_bytes))
2492  return true;
2493 
2494  blk_finish_request(rq, error);
2495 
2496  return false;
2497 }
2498 
2513 bool blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2514 {
2515  return blk_end_bidi_request(rq, error, nr_bytes, 0);
2516 }
2518 
2527 void blk_end_request_all(struct request *rq, int error)
2528 {
2529  bool pending;
2530  unsigned int bidi_bytes = 0;
2531 
2532  if (unlikely(blk_bidi_rq(rq)))
2533  bidi_bytes = blk_rq_bytes(rq->next_rq);
2534 
2535  pending = blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2536  BUG_ON(pending);
2537 }
2539 
2552 bool blk_end_request_cur(struct request *rq, int error)
2553 {
2554  return blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2555 }
2557 
2570 bool blk_end_request_err(struct request *rq, int error)
2571 {
2572  WARN_ON(error >= 0);
2573  return blk_end_request(rq, error, blk_rq_err_bytes(rq));
2574 }
2576 
2590 bool __blk_end_request(struct request *rq, int error, unsigned int nr_bytes)
2591 {
2592  return __blk_end_bidi_request(rq, error, nr_bytes, 0);
2593 }
2595 
2604 void __blk_end_request_all(struct request *rq, int error)
2605 {
2606  bool pending;
2607  unsigned int bidi_bytes = 0;
2608 
2609  if (unlikely(blk_bidi_rq(rq)))
2610  bidi_bytes = blk_rq_bytes(rq->next_rq);
2611 
2612  pending = __blk_end_bidi_request(rq, error, blk_rq_bytes(rq), bidi_bytes);
2613  BUG_ON(pending);
2614 }
2616 
2630 bool __blk_end_request_cur(struct request *rq, int error)
2631 {
2632  return __blk_end_request(rq, error, blk_rq_cur_bytes(rq));
2633 }
2635 
2649 bool __blk_end_request_err(struct request *rq, int error)
2650 {
2651  WARN_ON(error >= 0);
2652  return __blk_end_request(rq, error, blk_rq_err_bytes(rq));
2653 }
2655 
2656 void blk_rq_bio_prep(struct request_queue *q, struct request *rq,
2657  struct bio *bio)
2658 {
2659  /* Bit 0 (R/W) is identical in rq->cmd_flags and bio->bi_rw */
2660  rq->cmd_flags |= bio->bi_rw & REQ_WRITE;
2661 
2662  if (bio_has_data(bio)) {
2663  rq->nr_phys_segments = bio_phys_segments(q, bio);
2664  rq->buffer = bio_data(bio);
2665  }
2666  rq->__data_len = bio->bi_size;
2667  rq->bio = rq->biotail = bio;
2668 
2669  if (bio->bi_bdev)
2670  rq->rq_disk = bio->bi_bdev->bd_disk;
2671 }
2672 
2673 #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
2674 
2681 void rq_flush_dcache_pages(struct request *rq)
2682 {
2683  struct req_iterator iter;
2684  struct bio_vec *bvec;
2685 
2686  rq_for_each_segment(bvec, rq, iter)
2687  flush_dcache_page(bvec->bv_page);
2688 }
2689 EXPORT_SYMBOL_GPL(rq_flush_dcache_pages);
2690 #endif
2691 
2712 {
2713  if (q->lld_busy_fn)
2714  return q->lld_busy_fn(q);
2715 
2716  return 0;
2717 }
2719 
2728 {
2729  struct bio *bio;
2730 
2731  while ((bio = rq->bio) != NULL) {
2732  rq->bio = bio->bi_next;
2733 
2734  bio_put(bio);
2735  }
2736 }
2738 
2739 /*
2740  * Copy attributes of the original request to the clone request.
2741  * The actual data parts (e.g. ->cmd, ->buffer, ->sense) are not copied.
2742  */
2743 static void __blk_rq_prep_clone(struct request *dst, struct request *src)
2744 {
2745  dst->cpu = src->cpu;
2746  dst->cmd_flags = (src->cmd_flags & REQ_CLONE_MASK) | REQ_NOMERGE;
2747  dst->cmd_type = src->cmd_type;
2748  dst->__sector = blk_rq_pos(src);
2749  dst->__data_len = blk_rq_bytes(src);
2750  dst->nr_phys_segments = src->nr_phys_segments;
2751  dst->ioprio = src->ioprio;
2752  dst->extra_len = src->extra_len;
2753 }
2754 
2774 int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
2775  struct bio_set *bs, gfp_t gfp_mask,
2776  int (*bio_ctr)(struct bio *, struct bio *, void *),
2777  void *data)
2778 {
2779  struct bio *bio, *bio_src;
2780 
2781  if (!bs)
2782  bs = fs_bio_set;
2783 
2784  blk_rq_init(NULL, rq);
2785 
2786  __rq_for_each_bio(bio_src, rq_src) {
2787  bio = bio_clone_bioset(bio_src, gfp_mask, bs);
2788  if (!bio)
2789  goto free_and_out;
2790 
2791  if (bio_ctr && bio_ctr(bio, bio_src, data))
2792  goto free_and_out;
2793 
2794  if (rq->bio) {
2795  rq->biotail->bi_next = bio;
2796  rq->biotail = bio;
2797  } else
2798  rq->bio = rq->biotail = bio;
2799  }
2800 
2801  __blk_rq_prep_clone(rq, rq_src);
2802 
2803  return 0;
2804 
2805 free_and_out:
2806  if (bio)
2807  bio_put(bio);
2808  blk_rq_unprep_clone(rq);
2809 
2810  return -ENOMEM;
2811 }
2813 
2815 {
2816  return queue_work(kblockd_workqueue, work);
2817 }
2819 
2821  struct delayed_work *dwork, unsigned long delay)
2822 {
2823  return queue_delayed_work(kblockd_workqueue, dwork, delay);
2824 }
2826 
2827 #define PLUG_MAGIC 0x91827364
2828 
2843 void blk_start_plug(struct blk_plug *plug)
2844 {
2845  struct task_struct *tsk = current;
2846 
2847  plug->magic = PLUG_MAGIC;
2848  INIT_LIST_HEAD(&plug->list);
2849  INIT_LIST_HEAD(&plug->cb_list);
2850  plug->should_sort = 0;
2851 
2852  /*
2853  * If this is a nested plug, don't actually assign it. It will be
2854  * flushed on its own.
2855  */
2856  if (!tsk->plug) {
2857  /*
2858  * Store ordering should not be needed here, since a potential
2859  * preempt will imply a full memory barrier
2860  */
2861  tsk->plug = plug;
2862  }
2863 }
2865 
2866 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
2867 {
2868  struct request *rqa = container_of(a, struct request, queuelist);
2869  struct request *rqb = container_of(b, struct request, queuelist);
2870 
2871  return !(rqa->q < rqb->q ||
2872  (rqa->q == rqb->q && blk_rq_pos(rqa) < blk_rq_pos(rqb)));
2873 }
2874 
2875 /*
2876  * If 'from_schedule' is true, then postpone the dispatch of requests
2877  * until a safe kblockd context. We due this to avoid accidental big
2878  * additional stack usage in driver dispatch, in places where the originally
2879  * plugger did not intend it.
2880  */
2881 static void queue_unplugged(struct request_queue *q, unsigned int depth,
2882  bool from_schedule)
2883  __releases(q->queue_lock)
2884 {
2885  trace_block_unplug(q, depth, !from_schedule);
2886 
2887  /*
2888  * Don't mess with dead queue.
2889  */
2890  if (unlikely(blk_queue_dead(q))) {
2891  spin_unlock(q->queue_lock);
2892  return;
2893  }
2894 
2895  /*
2896  * If we are punting this to kblockd, then we can safely drop
2897  * the queue_lock before waking kblockd (which needs to take
2898  * this lock).
2899  */
2900  if (from_schedule) {
2901  spin_unlock(q->queue_lock);
2903  } else {
2904  __blk_run_queue(q);
2905  spin_unlock(q->queue_lock);
2906  }
2907 
2908 }
2909 
2910 static void flush_plug_callbacks(struct blk_plug *plug, bool from_schedule)
2911 {
2912  LIST_HEAD(callbacks);
2913 
2914  while (!list_empty(&plug->cb_list)) {
2915  list_splice_init(&plug->cb_list, &callbacks);
2916 
2917  while (!list_empty(&callbacks)) {
2918  struct blk_plug_cb *cb = list_first_entry(&callbacks,
2919  struct blk_plug_cb,
2920  list);
2921  list_del(&cb->list);
2922  cb->callback(cb, from_schedule);
2923  }
2924  }
2925 }
2926 
2927 struct blk_plug_cb *blk_check_plugged(blk_plug_cb_fn unplug, void *data,
2928  int size)
2929 {
2930  struct blk_plug *plug = current->plug;
2931  struct blk_plug_cb *cb;
2932 
2933  if (!plug)
2934  return NULL;
2935 
2936  list_for_each_entry(cb, &plug->cb_list, list)
2937  if (cb->callback == unplug && cb->data == data)
2938  return cb;
2939 
2940  /* Not currently on the callback list */
2941  BUG_ON(size < sizeof(*cb));
2942  cb = kzalloc(size, GFP_ATOMIC);
2943  if (cb) {
2944  cb->data = data;
2945  cb->callback = unplug;
2946  list_add(&cb->list, &plug->cb_list);
2947  }
2948  return cb;
2949 }
2951 
2952 void blk_flush_plug_list(struct blk_plug *plug, bool from_schedule)
2953 {
2954  struct request_queue *q;
2955  unsigned long flags;
2956  struct request *rq;
2957  LIST_HEAD(list);
2958  unsigned int depth;
2959 
2960  BUG_ON(plug->magic != PLUG_MAGIC);
2961 
2962  flush_plug_callbacks(plug, from_schedule);
2963  if (list_empty(&plug->list))
2964  return;
2965 
2966  list_splice_init(&plug->list, &list);
2967 
2968  if (plug->should_sort) {
2969  list_sort(NULL, &list, plug_rq_cmp);
2970  plug->should_sort = 0;
2971  }
2972 
2973  q = NULL;
2974  depth = 0;
2975 
2976  /*
2977  * Save and disable interrupts here, to avoid doing it for every
2978  * queue lock we have to take.
2979  */
2980  local_irq_save(flags);
2981  while (!list_empty(&list)) {
2982  rq = list_entry_rq(list.next);
2983  list_del_init(&rq->queuelist);
2984  BUG_ON(!rq->q);
2985  if (rq->q != q) {
2986  /*
2987  * This drops the queue lock
2988  */
2989  if (q)
2990  queue_unplugged(q, depth, from_schedule);
2991  q = rq->q;
2992  depth = 0;
2993  spin_lock(q->queue_lock);
2994  }
2995 
2996  /*
2997  * Short-circuit if @q is dead
2998  */
2999  if (unlikely(blk_queue_dead(q))) {
3001  continue;
3002  }
3003 
3004  /*
3005  * rq is already accounted, so use raw insert
3006  */
3007  if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA))
3008  __elv_add_request(q, rq, ELEVATOR_INSERT_FLUSH);
3009  else
3010  __elv_add_request(q, rq, ELEVATOR_INSERT_SORT_MERGE);
3011 
3012  depth++;
3013  }
3014 
3015  /*
3016  * This drops the queue lock
3017  */
3018  if (q)
3019  queue_unplugged(q, depth, from_schedule);
3020 
3021  local_irq_restore(flags);
3022 }
3023 
3024 void blk_finish_plug(struct blk_plug *plug)
3025 {
3026  blk_flush_plug_list(plug, false);
3027 
3028  if (plug == current->plug)
3029  current->plug = NULL;
3030 }
3032 
3034 {
3036  sizeof(((struct request *)0)->cmd_flags));
3037 
3038  /* used for unplugging and affects IO latency/throughput - HIGHPRI */
3039  kblockd_workqueue = alloc_workqueue("kblockd",
3040  WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
3041  if (!kblockd_workqueue)
3042  panic("Failed to create kblockd\n");
3043 
3044  request_cachep = kmem_cache_create("blkdev_requests",
3045  sizeof(struct request), 0, SLAB_PANIC, NULL);
3046 
3047  blk_requestq_cachep = kmem_cache_create("blkdev_queue",
3048  sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
3049 
3050  return 0;
3051 }