Merge git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable
[pandora-kernel.git] / fs / btrfs / extent_io.c
1 #include <linux/bitops.h>
2 #include <linux/slab.h>
3 #include <linux/bio.h>
4 #include <linux/mm.h>
5 #include <linux/pagemap.h>
6 #include <linux/page-flags.h>
7 #include <linux/module.h>
8 #include <linux/spinlock.h>
9 #include <linux/blkdev.h>
10 #include <linux/swap.h>
11 #include <linux/writeback.h>
12 #include <linux/pagevec.h>
13 #include "extent_io.h"
14 #include "extent_map.h"
15 #include "compat.h"
16 #include "ctree.h"
17 #include "btrfs_inode.h"
18
19 static struct kmem_cache *extent_state_cache;
20 static struct kmem_cache *extent_buffer_cache;
21
22 static LIST_HEAD(buffers);
23 static LIST_HEAD(states);
24
25 #define LEAK_DEBUG 0
26 #if LEAK_DEBUG
27 static DEFINE_SPINLOCK(leak_lock);
28 #endif
29
30 #define BUFFER_LRU_MAX 64
31
32 struct tree_entry {
33         u64 start;
34         u64 end;
35         struct rb_node rb_node;
36 };
37
38 struct extent_page_data {
39         struct bio *bio;
40         struct extent_io_tree *tree;
41         get_extent_t *get_extent;
42
43         /* tells writepage not to lock the state bits for this range
44          * it still does the unlocking
45          */
46         unsigned int extent_locked:1;
47
48         /* tells the submit_bio code to use a WRITE_SYNC */
49         unsigned int sync_io:1;
50 };
51
52 int __init extent_io_init(void)
53 {
54         extent_state_cache = kmem_cache_create("extent_state",
55                         sizeof(struct extent_state), 0,
56                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
57         if (!extent_state_cache)
58                 return -ENOMEM;
59
60         extent_buffer_cache = kmem_cache_create("extent_buffers",
61                         sizeof(struct extent_buffer), 0,
62                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
63         if (!extent_buffer_cache)
64                 goto free_state_cache;
65         return 0;
66
67 free_state_cache:
68         kmem_cache_destroy(extent_state_cache);
69         return -ENOMEM;
70 }
71
72 void extent_io_exit(void)
73 {
74         struct extent_state *state;
75         struct extent_buffer *eb;
76
77         while (!list_empty(&states)) {
78                 state = list_entry(states.next, struct extent_state, leak_list);
79                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
80                        "state %lu in tree %p refs %d\n",
81                        (unsigned long long)state->start,
82                        (unsigned long long)state->end,
83                        state->state, state->tree, atomic_read(&state->refs));
84                 list_del(&state->leak_list);
85                 kmem_cache_free(extent_state_cache, state);
86
87         }
88
89         while (!list_empty(&buffers)) {
90                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
91                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
92                        "refs %d\n", (unsigned long long)eb->start,
93                        eb->len, atomic_read(&eb->refs));
94                 list_del(&eb->leak_list);
95                 kmem_cache_free(extent_buffer_cache, eb);
96         }
97         if (extent_state_cache)
98                 kmem_cache_destroy(extent_state_cache);
99         if (extent_buffer_cache)
100                 kmem_cache_destroy(extent_buffer_cache);
101 }
102
103 void extent_io_tree_init(struct extent_io_tree *tree,
104                           struct address_space *mapping, gfp_t mask)
105 {
106         tree->state = RB_ROOT;
107         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
108         tree->ops = NULL;
109         tree->dirty_bytes = 0;
110         spin_lock_init(&tree->lock);
111         spin_lock_init(&tree->buffer_lock);
112         tree->mapping = mapping;
113 }
114
115 static struct extent_state *alloc_extent_state(gfp_t mask)
116 {
117         struct extent_state *state;
118 #if LEAK_DEBUG
119         unsigned long flags;
120 #endif
121
122         state = kmem_cache_alloc(extent_state_cache, mask);
123         if (!state)
124                 return state;
125         state->state = 0;
126         state->private = 0;
127         state->tree = NULL;
128 #if LEAK_DEBUG
129         spin_lock_irqsave(&leak_lock, flags);
130         list_add(&state->leak_list, &states);
131         spin_unlock_irqrestore(&leak_lock, flags);
132 #endif
133         atomic_set(&state->refs, 1);
134         init_waitqueue_head(&state->wq);
135         return state;
136 }
137
138 void free_extent_state(struct extent_state *state)
139 {
140         if (!state)
141                 return;
142         if (atomic_dec_and_test(&state->refs)) {
143 #if LEAK_DEBUG
144                 unsigned long flags;
145 #endif
146                 WARN_ON(state->tree);
147 #if LEAK_DEBUG
148                 spin_lock_irqsave(&leak_lock, flags);
149                 list_del(&state->leak_list);
150                 spin_unlock_irqrestore(&leak_lock, flags);
151 #endif
152                 kmem_cache_free(extent_state_cache, state);
153         }
154 }
155
156 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
157                                    struct rb_node *node)
158 {
159         struct rb_node **p = &root->rb_node;
160         struct rb_node *parent = NULL;
161         struct tree_entry *entry;
162
163         while (*p) {
164                 parent = *p;
165                 entry = rb_entry(parent, struct tree_entry, rb_node);
166
167                 if (offset < entry->start)
168                         p = &(*p)->rb_left;
169                 else if (offset > entry->end)
170                         p = &(*p)->rb_right;
171                 else
172                         return parent;
173         }
174
175         entry = rb_entry(node, struct tree_entry, rb_node);
176         rb_link_node(node, parent, p);
177         rb_insert_color(node, root);
178         return NULL;
179 }
180
181 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
182                                      struct rb_node **prev_ret,
183                                      struct rb_node **next_ret)
184 {
185         struct rb_root *root = &tree->state;
186         struct rb_node *n = root->rb_node;
187         struct rb_node *prev = NULL;
188         struct rb_node *orig_prev = NULL;
189         struct tree_entry *entry;
190         struct tree_entry *prev_entry = NULL;
191
192         while (n) {
193                 entry = rb_entry(n, struct tree_entry, rb_node);
194                 prev = n;
195                 prev_entry = entry;
196
197                 if (offset < entry->start)
198                         n = n->rb_left;
199                 else if (offset > entry->end)
200                         n = n->rb_right;
201                 else
202                         return n;
203         }
204
205         if (prev_ret) {
206                 orig_prev = prev;
207                 while (prev && offset > prev_entry->end) {
208                         prev = rb_next(prev);
209                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
210                 }
211                 *prev_ret = prev;
212                 prev = orig_prev;
213         }
214
215         if (next_ret) {
216                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
217                 while (prev && offset < prev_entry->start) {
218                         prev = rb_prev(prev);
219                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
220                 }
221                 *next_ret = prev;
222         }
223         return NULL;
224 }
225
226 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
227                                           u64 offset)
228 {
229         struct rb_node *prev = NULL;
230         struct rb_node *ret;
231
232         ret = __etree_search(tree, offset, &prev, NULL);
233         if (!ret)
234                 return prev;
235         return ret;
236 }
237
238 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
239                      struct extent_state *other)
240 {
241         if (tree->ops && tree->ops->merge_extent_hook)
242                 tree->ops->merge_extent_hook(tree->mapping->host, new,
243                                              other);
244 }
245
246 /*
247  * utility function to look for merge candidates inside a given range.
248  * Any extents with matching state are merged together into a single
249  * extent in the tree.  Extents with EXTENT_IO in their state field
250  * are not merged because the end_io handlers need to be able to do
251  * operations on them without sleeping (or doing allocations/splits).
252  *
253  * This should be called with the tree lock held.
254  */
255 static int merge_state(struct extent_io_tree *tree,
256                        struct extent_state *state)
257 {
258         struct extent_state *other;
259         struct rb_node *other_node;
260
261         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
262                 return 0;
263
264         other_node = rb_prev(&state->rb_node);
265         if (other_node) {
266                 other = rb_entry(other_node, struct extent_state, rb_node);
267                 if (other->end == state->start - 1 &&
268                     other->state == state->state) {
269                         merge_cb(tree, state, other);
270                         state->start = other->start;
271                         other->tree = NULL;
272                         rb_erase(&other->rb_node, &tree->state);
273                         free_extent_state(other);
274                 }
275         }
276         other_node = rb_next(&state->rb_node);
277         if (other_node) {
278                 other = rb_entry(other_node, struct extent_state, rb_node);
279                 if (other->start == state->end + 1 &&
280                     other->state == state->state) {
281                         merge_cb(tree, state, other);
282                         other->start = state->start;
283                         state->tree = NULL;
284                         rb_erase(&state->rb_node, &tree->state);
285                         free_extent_state(state);
286                         state = NULL;
287                 }
288         }
289
290         return 0;
291 }
292
293 static int set_state_cb(struct extent_io_tree *tree,
294                          struct extent_state *state, int *bits)
295 {
296         if (tree->ops && tree->ops->set_bit_hook) {
297                 return tree->ops->set_bit_hook(tree->mapping->host,
298                                                state, bits);
299         }
300
301         return 0;
302 }
303
304 static void clear_state_cb(struct extent_io_tree *tree,
305                            struct extent_state *state, int *bits)
306 {
307         if (tree->ops && tree->ops->clear_bit_hook)
308                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
309 }
310
311 /*
312  * insert an extent_state struct into the tree.  'bits' are set on the
313  * struct before it is inserted.
314  *
315  * This may return -EEXIST if the extent is already there, in which case the
316  * state struct is freed.
317  *
318  * The tree lock is not taken internally.  This is a utility function and
319  * probably isn't what you want to call (see set/clear_extent_bit).
320  */
321 static int insert_state(struct extent_io_tree *tree,
322                         struct extent_state *state, u64 start, u64 end,
323                         int *bits)
324 {
325         struct rb_node *node;
326         int bits_to_set = *bits & ~EXTENT_CTLBITS;
327         int ret;
328
329         if (end < start) {
330                 printk(KERN_ERR "btrfs end < start %llu %llu\n",
331                        (unsigned long long)end,
332                        (unsigned long long)start);
333                 WARN_ON(1);
334         }
335         state->start = start;
336         state->end = end;
337         ret = set_state_cb(tree, state, bits);
338         if (ret)
339                 return ret;
340
341         if (bits_to_set & EXTENT_DIRTY)
342                 tree->dirty_bytes += end - start + 1;
343         state->state |= bits_to_set;
344         node = tree_insert(&tree->state, end, &state->rb_node);
345         if (node) {
346                 struct extent_state *found;
347                 found = rb_entry(node, struct extent_state, rb_node);
348                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
349                        "%llu %llu\n", (unsigned long long)found->start,
350                        (unsigned long long)found->end,
351                        (unsigned long long)start, (unsigned long long)end);
352                 free_extent_state(state);
353                 return -EEXIST;
354         }
355         state->tree = tree;
356         merge_state(tree, state);
357         return 0;
358 }
359
360 static int split_cb(struct extent_io_tree *tree, struct extent_state *orig,
361                      u64 split)
362 {
363         if (tree->ops && tree->ops->split_extent_hook)
364                 return tree->ops->split_extent_hook(tree->mapping->host,
365                                                     orig, split);
366         return 0;
367 }
368
369 /*
370  * split a given extent state struct in two, inserting the preallocated
371  * struct 'prealloc' as the newly created second half.  'split' indicates an
372  * offset inside 'orig' where it should be split.
373  *
374  * Before calling,
375  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
376  * are two extent state structs in the tree:
377  * prealloc: [orig->start, split - 1]
378  * orig: [ split, orig->end ]
379  *
380  * The tree locks are not taken by this function. They need to be held
381  * by the caller.
382  */
383 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
384                        struct extent_state *prealloc, u64 split)
385 {
386         struct rb_node *node;
387
388         split_cb(tree, orig, split);
389
390         prealloc->start = orig->start;
391         prealloc->end = split - 1;
392         prealloc->state = orig->state;
393         orig->start = split;
394
395         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
396         if (node) {
397                 free_extent_state(prealloc);
398                 return -EEXIST;
399         }
400         prealloc->tree = tree;
401         return 0;
402 }
403
404 /*
405  * utility function to clear some bits in an extent state struct.
406  * it will optionally wake up any one waiting on this state (wake == 1), or
407  * forcibly remove the state from the tree (delete == 1).
408  *
409  * If no bits are set on the state struct after clearing things, the
410  * struct is freed and removed from the tree
411  */
412 static int clear_state_bit(struct extent_io_tree *tree,
413                             struct extent_state *state,
414                             int *bits, int wake)
415 {
416         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
417         int ret = state->state & bits_to_clear;
418
419         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
420                 u64 range = state->end - state->start + 1;
421                 WARN_ON(range > tree->dirty_bytes);
422                 tree->dirty_bytes -= range;
423         }
424         clear_state_cb(tree, state, bits);
425         state->state &= ~bits_to_clear;
426         if (wake)
427                 wake_up(&state->wq);
428         if (state->state == 0) {
429                 if (state->tree) {
430                         rb_erase(&state->rb_node, &tree->state);
431                         state->tree = NULL;
432                         free_extent_state(state);
433                 } else {
434                         WARN_ON(1);
435                 }
436         } else {
437                 merge_state(tree, state);
438         }
439         return ret;
440 }
441
442 /*
443  * clear some bits on a range in the tree.  This may require splitting
444  * or inserting elements in the tree, so the gfp mask is used to
445  * indicate which allocations or sleeping are allowed.
446  *
447  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
448  * the given range from the tree regardless of state (ie for truncate).
449  *
450  * the range [start, end] is inclusive.
451  *
452  * This takes the tree lock, and returns < 0 on error, > 0 if any of the
453  * bits were already set, or zero if none of the bits were already set.
454  */
455 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
456                      int bits, int wake, int delete,
457                      struct extent_state **cached_state,
458                      gfp_t mask)
459 {
460         struct extent_state *state;
461         struct extent_state *cached;
462         struct extent_state *prealloc = NULL;
463         struct rb_node *next_node;
464         struct rb_node *node;
465         u64 last_end;
466         int err;
467         int set = 0;
468         int clear = 0;
469
470         if (delete)
471                 bits |= ~EXTENT_CTLBITS;
472         bits |= EXTENT_FIRST_DELALLOC;
473
474         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
475                 clear = 1;
476 again:
477         if (!prealloc && (mask & __GFP_WAIT)) {
478                 prealloc = alloc_extent_state(mask);
479                 if (!prealloc)
480                         return -ENOMEM;
481         }
482
483         spin_lock(&tree->lock);
484         if (cached_state) {
485                 cached = *cached_state;
486
487                 if (clear) {
488                         *cached_state = NULL;
489                         cached_state = NULL;
490                 }
491
492                 if (cached && cached->tree && cached->start == start) {
493                         if (clear)
494                                 atomic_dec(&cached->refs);
495                         state = cached;
496                         goto hit_next;
497                 }
498                 if (clear)
499                         free_extent_state(cached);
500         }
501         /*
502          * this search will find the extents that end after
503          * our range starts
504          */
505         node = tree_search(tree, start);
506         if (!node)
507                 goto out;
508         state = rb_entry(node, struct extent_state, rb_node);
509 hit_next:
510         if (state->start > end)
511                 goto out;
512         WARN_ON(state->end < start);
513         last_end = state->end;
514
515         /*
516          *     | ---- desired range ---- |
517          *  | state | or
518          *  | ------------- state -------------- |
519          *
520          * We need to split the extent we found, and may flip
521          * bits on second half.
522          *
523          * If the extent we found extends past our range, we
524          * just split and search again.  It'll get split again
525          * the next time though.
526          *
527          * If the extent we found is inside our range, we clear
528          * the desired bit on it.
529          */
530
531         if (state->start < start) {
532                 if (!prealloc)
533                         prealloc = alloc_extent_state(GFP_ATOMIC);
534                 err = split_state(tree, state, prealloc, start);
535                 BUG_ON(err == -EEXIST);
536                 prealloc = NULL;
537                 if (err)
538                         goto out;
539                 if (state->end <= end) {
540                         set |= clear_state_bit(tree, state, &bits, wake);
541                         if (last_end == (u64)-1)
542                                 goto out;
543                         start = last_end + 1;
544                 }
545                 goto search_again;
546         }
547         /*
548          * | ---- desired range ---- |
549          *                        | state |
550          * We need to split the extent, and clear the bit
551          * on the first half
552          */
553         if (state->start <= end && state->end > end) {
554                 if (!prealloc)
555                         prealloc = alloc_extent_state(GFP_ATOMIC);
556                 err = split_state(tree, state, prealloc, end + 1);
557                 BUG_ON(err == -EEXIST);
558                 if (wake)
559                         wake_up(&state->wq);
560
561                 set |= clear_state_bit(tree, prealloc, &bits, wake);
562
563                 prealloc = NULL;
564                 goto out;
565         }
566
567         if (state->end < end && prealloc && !need_resched())
568                 next_node = rb_next(&state->rb_node);
569         else
570                 next_node = NULL;
571
572         set |= clear_state_bit(tree, state, &bits, wake);
573         if (last_end == (u64)-1)
574                 goto out;
575         start = last_end + 1;
576         if (start <= end && next_node) {
577                 state = rb_entry(next_node, struct extent_state,
578                                  rb_node);
579                 if (state->start == start)
580                         goto hit_next;
581         }
582         goto search_again;
583
584 out:
585         spin_unlock(&tree->lock);
586         if (prealloc)
587                 free_extent_state(prealloc);
588
589         return set;
590
591 search_again:
592         if (start > end)
593                 goto out;
594         spin_unlock(&tree->lock);
595         if (mask & __GFP_WAIT)
596                 cond_resched();
597         goto again;
598 }
599
600 static int wait_on_state(struct extent_io_tree *tree,
601                          struct extent_state *state)
602                 __releases(tree->lock)
603                 __acquires(tree->lock)
604 {
605         DEFINE_WAIT(wait);
606         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
607         spin_unlock(&tree->lock);
608         schedule();
609         spin_lock(&tree->lock);
610         finish_wait(&state->wq, &wait);
611         return 0;
612 }
613
614 /*
615  * waits for one or more bits to clear on a range in the state tree.
616  * The range [start, end] is inclusive.
617  * The tree lock is taken by this function
618  */
619 int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
620 {
621         struct extent_state *state;
622         struct rb_node *node;
623
624         spin_lock(&tree->lock);
625 again:
626         while (1) {
627                 /*
628                  * this search will find all the extents that end after
629                  * our range starts
630                  */
631                 node = tree_search(tree, start);
632                 if (!node)
633                         break;
634
635                 state = rb_entry(node, struct extent_state, rb_node);
636
637                 if (state->start > end)
638                         goto out;
639
640                 if (state->state & bits) {
641                         start = state->start;
642                         atomic_inc(&state->refs);
643                         wait_on_state(tree, state);
644                         free_extent_state(state);
645                         goto again;
646                 }
647                 start = state->end + 1;
648
649                 if (start > end)
650                         break;
651
652                 if (need_resched()) {
653                         spin_unlock(&tree->lock);
654                         cond_resched();
655                         spin_lock(&tree->lock);
656                 }
657         }
658 out:
659         spin_unlock(&tree->lock);
660         return 0;
661 }
662
663 static int set_state_bits(struct extent_io_tree *tree,
664                            struct extent_state *state,
665                            int *bits)
666 {
667         int ret;
668         int bits_to_set = *bits & ~EXTENT_CTLBITS;
669
670         ret = set_state_cb(tree, state, bits);
671         if (ret)
672                 return ret;
673         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
674                 u64 range = state->end - state->start + 1;
675                 tree->dirty_bytes += range;
676         }
677         state->state |= bits_to_set;
678
679         return 0;
680 }
681
682 static void cache_state(struct extent_state *state,
683                         struct extent_state **cached_ptr)
684 {
685         if (cached_ptr && !(*cached_ptr)) {
686                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
687                         *cached_ptr = state;
688                         atomic_inc(&state->refs);
689                 }
690         }
691 }
692
693 /*
694  * set some bits on a range in the tree.  This may require allocations or
695  * sleeping, so the gfp mask is used to indicate what is allowed.
696  *
697  * If any of the exclusive bits are set, this will fail with -EEXIST if some
698  * part of the range already has the desired bits set.  The start of the
699  * existing range is returned in failed_start in this case.
700  *
701  * [start, end] is inclusive This takes the tree lock.
702  */
703
704 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
705                    int bits, int exclusive_bits, u64 *failed_start,
706                    struct extent_state **cached_state, gfp_t mask)
707 {
708         struct extent_state *state;
709         struct extent_state *prealloc = NULL;
710         struct rb_node *node;
711         int err = 0;
712         u64 last_start;
713         u64 last_end;
714
715         bits |= EXTENT_FIRST_DELALLOC;
716 again:
717         if (!prealloc && (mask & __GFP_WAIT)) {
718                 prealloc = alloc_extent_state(mask);
719                 if (!prealloc)
720                         return -ENOMEM;
721         }
722
723         spin_lock(&tree->lock);
724         if (cached_state && *cached_state) {
725                 state = *cached_state;
726                 if (state->start == start && state->tree) {
727                         node = &state->rb_node;
728                         goto hit_next;
729                 }
730         }
731         /*
732          * this search will find all the extents that end after
733          * our range starts.
734          */
735         node = tree_search(tree, start);
736         if (!node) {
737                 err = insert_state(tree, prealloc, start, end, &bits);
738                 prealloc = NULL;
739                 BUG_ON(err == -EEXIST);
740                 goto out;
741         }
742         state = rb_entry(node, struct extent_state, rb_node);
743 hit_next:
744         last_start = state->start;
745         last_end = state->end;
746
747         /*
748          * | ---- desired range ---- |
749          * | state |
750          *
751          * Just lock what we found and keep going
752          */
753         if (state->start == start && state->end <= end) {
754                 struct rb_node *next_node;
755                 if (state->state & exclusive_bits) {
756                         *failed_start = state->start;
757                         err = -EEXIST;
758                         goto out;
759                 }
760
761                 err = set_state_bits(tree, state, &bits);
762                 if (err)
763                         goto out;
764
765                 cache_state(state, cached_state);
766                 merge_state(tree, state);
767                 if (last_end == (u64)-1)
768                         goto out;
769
770                 start = last_end + 1;
771                 if (start < end && prealloc && !need_resched()) {
772                         next_node = rb_next(node);
773                         if (next_node) {
774                                 state = rb_entry(next_node, struct extent_state,
775                                                  rb_node);
776                                 if (state->start == start)
777                                         goto hit_next;
778                         }
779                 }
780                 goto search_again;
781         }
782
783         /*
784          *     | ---- desired range ---- |
785          * | state |
786          *   or
787          * | ------------- state -------------- |
788          *
789          * We need to split the extent we found, and may flip bits on
790          * second half.
791          *
792          * If the extent we found extends past our
793          * range, we just split and search again.  It'll get split
794          * again the next time though.
795          *
796          * If the extent we found is inside our range, we set the
797          * desired bit on it.
798          */
799         if (state->start < start) {
800                 if (state->state & exclusive_bits) {
801                         *failed_start = start;
802                         err = -EEXIST;
803                         goto out;
804                 }
805                 err = split_state(tree, state, prealloc, start);
806                 BUG_ON(err == -EEXIST);
807                 prealloc = NULL;
808                 if (err)
809                         goto out;
810                 if (state->end <= end) {
811                         err = set_state_bits(tree, state, &bits);
812                         if (err)
813                                 goto out;
814                         cache_state(state, cached_state);
815                         merge_state(tree, state);
816                         if (last_end == (u64)-1)
817                                 goto out;
818                         start = last_end + 1;
819                 }
820                 goto search_again;
821         }
822         /*
823          * | ---- desired range ---- |
824          *     | state | or               | state |
825          *
826          * There's a hole, we need to insert something in it and
827          * ignore the extent we found.
828          */
829         if (state->start > start) {
830                 u64 this_end;
831                 if (end < last_start)
832                         this_end = end;
833                 else
834                         this_end = last_start - 1;
835                 err = insert_state(tree, prealloc, start, this_end,
836                                    &bits);
837                 BUG_ON(err == -EEXIST);
838                 if (err) {
839                         prealloc = NULL;
840                         goto out;
841                 }
842                 cache_state(prealloc, cached_state);
843                 prealloc = NULL;
844                 start = this_end + 1;
845                 goto search_again;
846         }
847         /*
848          * | ---- desired range ---- |
849          *                        | state |
850          * We need to split the extent, and set the bit
851          * on the first half
852          */
853         if (state->start <= end && state->end > end) {
854                 if (state->state & exclusive_bits) {
855                         *failed_start = start;
856                         err = -EEXIST;
857                         goto out;
858                 }
859                 err = split_state(tree, state, prealloc, end + 1);
860                 BUG_ON(err == -EEXIST);
861
862                 err = set_state_bits(tree, prealloc, &bits);
863                 if (err) {
864                         prealloc = NULL;
865                         goto out;
866                 }
867                 cache_state(prealloc, cached_state);
868                 merge_state(tree, prealloc);
869                 prealloc = NULL;
870                 goto out;
871         }
872
873         goto search_again;
874
875 out:
876         spin_unlock(&tree->lock);
877         if (prealloc)
878                 free_extent_state(prealloc);
879
880         return err;
881
882 search_again:
883         if (start > end)
884                 goto out;
885         spin_unlock(&tree->lock);
886         if (mask & __GFP_WAIT)
887                 cond_resched();
888         goto again;
889 }
890
891 /* wrappers around set/clear extent bit */
892 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
893                      gfp_t mask)
894 {
895         return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
896                               NULL, mask);
897 }
898
899 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
900                     int bits, gfp_t mask)
901 {
902         return set_extent_bit(tree, start, end, bits, 0, NULL,
903                               NULL, mask);
904 }
905
906 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
907                       int bits, gfp_t mask)
908 {
909         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
910 }
911
912 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
913                         struct extent_state **cached_state, gfp_t mask)
914 {
915         return set_extent_bit(tree, start, end,
916                               EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
917                               0, NULL, cached_state, mask);
918 }
919
920 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
921                        gfp_t mask)
922 {
923         return clear_extent_bit(tree, start, end,
924                                 EXTENT_DIRTY | EXTENT_DELALLOC |
925                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
926 }
927
928 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
929                      gfp_t mask)
930 {
931         return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
932                               NULL, mask);
933 }
934
935 static int clear_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
936                        gfp_t mask)
937 {
938         return clear_extent_bit(tree, start, end, EXTENT_NEW, 0, 0,
939                                 NULL, mask);
940 }
941
942 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
943                         gfp_t mask)
944 {
945         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, NULL,
946                               NULL, mask);
947 }
948
949 static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
950                                  u64 end, struct extent_state **cached_state,
951                                  gfp_t mask)
952 {
953         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
954                                 cached_state, mask);
955 }
956
957 int wait_on_extent_writeback(struct extent_io_tree *tree, u64 start, u64 end)
958 {
959         return wait_extent_bit(tree, start, end, EXTENT_WRITEBACK);
960 }
961
962 /*
963  * either insert or lock state struct between start and end use mask to tell
964  * us if waiting is desired.
965  */
966 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
967                      int bits, struct extent_state **cached_state, gfp_t mask)
968 {
969         int err;
970         u64 failed_start;
971         while (1) {
972                 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
973                                      EXTENT_LOCKED, &failed_start,
974                                      cached_state, mask);
975                 if (err == -EEXIST && (mask & __GFP_WAIT)) {
976                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
977                         start = failed_start;
978                 } else {
979                         break;
980                 }
981                 WARN_ON(start > end);
982         }
983         return err;
984 }
985
986 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
987 {
988         return lock_extent_bits(tree, start, end, 0, NULL, mask);
989 }
990
991 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
992                     gfp_t mask)
993 {
994         int err;
995         u64 failed_start;
996
997         err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
998                              &failed_start, NULL, mask);
999         if (err == -EEXIST) {
1000                 if (failed_start > start)
1001                         clear_extent_bit(tree, start, failed_start - 1,
1002                                          EXTENT_LOCKED, 1, 0, NULL, mask);
1003                 return 0;
1004         }
1005         return 1;
1006 }
1007
1008 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1009                          struct extent_state **cached, gfp_t mask)
1010 {
1011         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1012                                 mask);
1013 }
1014
1015 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end,
1016                   gfp_t mask)
1017 {
1018         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1019                                 mask);
1020 }
1021
1022 /*
1023  * helper function to set pages and extents in the tree dirty
1024  */
1025 int set_range_dirty(struct extent_io_tree *tree, u64 start, u64 end)
1026 {
1027         unsigned long index = start >> PAGE_CACHE_SHIFT;
1028         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1029         struct page *page;
1030
1031         while (index <= end_index) {
1032                 page = find_get_page(tree->mapping, index);
1033                 BUG_ON(!page);
1034                 __set_page_dirty_nobuffers(page);
1035                 page_cache_release(page);
1036                 index++;
1037         }
1038         return 0;
1039 }
1040
1041 /*
1042  * helper function to set both pages and extents in the tree writeback
1043  */
1044 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1045 {
1046         unsigned long index = start >> PAGE_CACHE_SHIFT;
1047         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1048         struct page *page;
1049
1050         while (index <= end_index) {
1051                 page = find_get_page(tree->mapping, index);
1052                 BUG_ON(!page);
1053                 set_page_writeback(page);
1054                 page_cache_release(page);
1055                 index++;
1056         }
1057         return 0;
1058 }
1059
1060 /*
1061  * find the first offset in the io tree with 'bits' set. zero is
1062  * returned if we find something, and *start_ret and *end_ret are
1063  * set to reflect the state struct that was found.
1064  *
1065  * If nothing was found, 1 is returned, < 0 on error
1066  */
1067 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1068                           u64 *start_ret, u64 *end_ret, int bits)
1069 {
1070         struct rb_node *node;
1071         struct extent_state *state;
1072         int ret = 1;
1073
1074         spin_lock(&tree->lock);
1075         /*
1076          * this search will find all the extents that end after
1077          * our range starts.
1078          */
1079         node = tree_search(tree, start);
1080         if (!node)
1081                 goto out;
1082
1083         while (1) {
1084                 state = rb_entry(node, struct extent_state, rb_node);
1085                 if (state->end >= start && (state->state & bits)) {
1086                         *start_ret = state->start;
1087                         *end_ret = state->end;
1088                         ret = 0;
1089                         break;
1090                 }
1091                 node = rb_next(node);
1092                 if (!node)
1093                         break;
1094         }
1095 out:
1096         spin_unlock(&tree->lock);
1097         return ret;
1098 }
1099
1100 /* find the first state struct with 'bits' set after 'start', and
1101  * return it.  tree->lock must be held.  NULL will returned if
1102  * nothing was found after 'start'
1103  */
1104 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1105                                                  u64 start, int bits)
1106 {
1107         struct rb_node *node;
1108         struct extent_state *state;
1109
1110         /*
1111          * this search will find all the extents that end after
1112          * our range starts.
1113          */
1114         node = tree_search(tree, start);
1115         if (!node)
1116                 goto out;
1117
1118         while (1) {
1119                 state = rb_entry(node, struct extent_state, rb_node);
1120                 if (state->end >= start && (state->state & bits))
1121                         return state;
1122
1123                 node = rb_next(node);
1124                 if (!node)
1125                         break;
1126         }
1127 out:
1128         return NULL;
1129 }
1130
1131 /*
1132  * find a contiguous range of bytes in the file marked as delalloc, not
1133  * more than 'max_bytes'.  start and end are used to return the range,
1134  *
1135  * 1 is returned if we find something, 0 if nothing was in the tree
1136  */
1137 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1138                                         u64 *start, u64 *end, u64 max_bytes,
1139                                         struct extent_state **cached_state)
1140 {
1141         struct rb_node *node;
1142         struct extent_state *state;
1143         u64 cur_start = *start;
1144         u64 found = 0;
1145         u64 total_bytes = 0;
1146
1147         spin_lock(&tree->lock);
1148
1149         /*
1150          * this search will find all the extents that end after
1151          * our range starts.
1152          */
1153         node = tree_search(tree, cur_start);
1154         if (!node) {
1155                 if (!found)
1156                         *end = (u64)-1;
1157                 goto out;
1158         }
1159
1160         while (1) {
1161                 state = rb_entry(node, struct extent_state, rb_node);
1162                 if (found && (state->start != cur_start ||
1163                               (state->state & EXTENT_BOUNDARY))) {
1164                         goto out;
1165                 }
1166                 if (!(state->state & EXTENT_DELALLOC)) {
1167                         if (!found)
1168                                 *end = state->end;
1169                         goto out;
1170                 }
1171                 if (!found) {
1172                         *start = state->start;
1173                         *cached_state = state;
1174                         atomic_inc(&state->refs);
1175                 }
1176                 found++;
1177                 *end = state->end;
1178                 cur_start = state->end + 1;
1179                 node = rb_next(node);
1180                 if (!node)
1181                         break;
1182                 total_bytes += state->end - state->start + 1;
1183                 if (total_bytes >= max_bytes)
1184                         break;
1185         }
1186 out:
1187         spin_unlock(&tree->lock);
1188         return found;
1189 }
1190
1191 static noinline int __unlock_for_delalloc(struct inode *inode,
1192                                           struct page *locked_page,
1193                                           u64 start, u64 end)
1194 {
1195         int ret;
1196         struct page *pages[16];
1197         unsigned long index = start >> PAGE_CACHE_SHIFT;
1198         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1199         unsigned long nr_pages = end_index - index + 1;
1200         int i;
1201
1202         if (index == locked_page->index && end_index == index)
1203                 return 0;
1204
1205         while (nr_pages > 0) {
1206                 ret = find_get_pages_contig(inode->i_mapping, index,
1207                                      min_t(unsigned long, nr_pages,
1208                                      ARRAY_SIZE(pages)), pages);
1209                 for (i = 0; i < ret; i++) {
1210                         if (pages[i] != locked_page)
1211                                 unlock_page(pages[i]);
1212                         page_cache_release(pages[i]);
1213                 }
1214                 nr_pages -= ret;
1215                 index += ret;
1216                 cond_resched();
1217         }
1218         return 0;
1219 }
1220
1221 static noinline int lock_delalloc_pages(struct inode *inode,
1222                                         struct page *locked_page,
1223                                         u64 delalloc_start,
1224                                         u64 delalloc_end)
1225 {
1226         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1227         unsigned long start_index = index;
1228         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1229         unsigned long pages_locked = 0;
1230         struct page *pages[16];
1231         unsigned long nrpages;
1232         int ret;
1233         int i;
1234
1235         /* the caller is responsible for locking the start index */
1236         if (index == locked_page->index && index == end_index)
1237                 return 0;
1238
1239         /* skip the page at the start index */
1240         nrpages = end_index - index + 1;
1241         while (nrpages > 0) {
1242                 ret = find_get_pages_contig(inode->i_mapping, index,
1243                                      min_t(unsigned long,
1244                                      nrpages, ARRAY_SIZE(pages)), pages);
1245                 if (ret == 0) {
1246                         ret = -EAGAIN;
1247                         goto done;
1248                 }
1249                 /* now we have an array of pages, lock them all */
1250                 for (i = 0; i < ret; i++) {
1251                         /*
1252                          * the caller is taking responsibility for
1253                          * locked_page
1254                          */
1255                         if (pages[i] != locked_page) {
1256                                 lock_page(pages[i]);
1257                                 if (!PageDirty(pages[i]) ||
1258                                     pages[i]->mapping != inode->i_mapping) {
1259                                         ret = -EAGAIN;
1260                                         unlock_page(pages[i]);
1261                                         page_cache_release(pages[i]);
1262                                         goto done;
1263                                 }
1264                         }
1265                         page_cache_release(pages[i]);
1266                         pages_locked++;
1267                 }
1268                 nrpages -= ret;
1269                 index += ret;
1270                 cond_resched();
1271         }
1272         ret = 0;
1273 done:
1274         if (ret && pages_locked) {
1275                 __unlock_for_delalloc(inode, locked_page,
1276                               delalloc_start,
1277                               ((u64)(start_index + pages_locked - 1)) <<
1278                               PAGE_CACHE_SHIFT);
1279         }
1280         return ret;
1281 }
1282
1283 /*
1284  * find a contiguous range of bytes in the file marked as delalloc, not
1285  * more than 'max_bytes'.  start and end are used to return the range,
1286  *
1287  * 1 is returned if we find something, 0 if nothing was in the tree
1288  */
1289 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1290                                              struct extent_io_tree *tree,
1291                                              struct page *locked_page,
1292                                              u64 *start, u64 *end,
1293                                              u64 max_bytes)
1294 {
1295         u64 delalloc_start;
1296         u64 delalloc_end;
1297         u64 found;
1298         struct extent_state *cached_state = NULL;
1299         int ret;
1300         int loops = 0;
1301
1302 again:
1303         /* step one, find a bunch of delalloc bytes starting at start */
1304         delalloc_start = *start;
1305         delalloc_end = 0;
1306         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1307                                     max_bytes, &cached_state);
1308         if (!found || delalloc_end <= *start) {
1309                 *start = delalloc_start;
1310                 *end = delalloc_end;
1311                 free_extent_state(cached_state);
1312                 return found;
1313         }
1314
1315         /*
1316          * start comes from the offset of locked_page.  We have to lock
1317          * pages in order, so we can't process delalloc bytes before
1318          * locked_page
1319          */
1320         if (delalloc_start < *start)
1321                 delalloc_start = *start;
1322
1323         /*
1324          * make sure to limit the number of pages we try to lock down
1325          * if we're looping.
1326          */
1327         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1328                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1329
1330         /* step two, lock all the pages after the page that has start */
1331         ret = lock_delalloc_pages(inode, locked_page,
1332                                   delalloc_start, delalloc_end);
1333         if (ret == -EAGAIN) {
1334                 /* some of the pages are gone, lets avoid looping by
1335                  * shortening the size of the delalloc range we're searching
1336                  */
1337                 free_extent_state(cached_state);
1338                 if (!loops) {
1339                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1340                         max_bytes = PAGE_CACHE_SIZE - offset;
1341                         loops = 1;
1342                         goto again;
1343                 } else {
1344                         found = 0;
1345                         goto out_failed;
1346                 }
1347         }
1348         BUG_ON(ret);
1349
1350         /* step three, lock the state bits for the whole range */
1351         lock_extent_bits(tree, delalloc_start, delalloc_end,
1352                          0, &cached_state, GFP_NOFS);
1353
1354         /* then test to make sure it is all still delalloc */
1355         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1356                              EXTENT_DELALLOC, 1, cached_state);
1357         if (!ret) {
1358                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1359                                      &cached_state, GFP_NOFS);
1360                 __unlock_for_delalloc(inode, locked_page,
1361                               delalloc_start, delalloc_end);
1362                 cond_resched();
1363                 goto again;
1364         }
1365         free_extent_state(cached_state);
1366         *start = delalloc_start;
1367         *end = delalloc_end;
1368 out_failed:
1369         return found;
1370 }
1371
1372 int extent_clear_unlock_delalloc(struct inode *inode,
1373                                 struct extent_io_tree *tree,
1374                                 u64 start, u64 end, struct page *locked_page,
1375                                 unsigned long op)
1376 {
1377         int ret;
1378         struct page *pages[16];
1379         unsigned long index = start >> PAGE_CACHE_SHIFT;
1380         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1381         unsigned long nr_pages = end_index - index + 1;
1382         int i;
1383         int clear_bits = 0;
1384
1385         if (op & EXTENT_CLEAR_UNLOCK)
1386                 clear_bits |= EXTENT_LOCKED;
1387         if (op & EXTENT_CLEAR_DIRTY)
1388                 clear_bits |= EXTENT_DIRTY;
1389
1390         if (op & EXTENT_CLEAR_DELALLOC)
1391                 clear_bits |= EXTENT_DELALLOC;
1392
1393         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1394         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1395                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1396                     EXTENT_SET_PRIVATE2)))
1397                 return 0;
1398
1399         while (nr_pages > 0) {
1400                 ret = find_get_pages_contig(inode->i_mapping, index,
1401                                      min_t(unsigned long,
1402                                      nr_pages, ARRAY_SIZE(pages)), pages);
1403                 for (i = 0; i < ret; i++) {
1404
1405                         if (op & EXTENT_SET_PRIVATE2)
1406                                 SetPagePrivate2(pages[i]);
1407
1408                         if (pages[i] == locked_page) {
1409                                 page_cache_release(pages[i]);
1410                                 continue;
1411                         }
1412                         if (op & EXTENT_CLEAR_DIRTY)
1413                                 clear_page_dirty_for_io(pages[i]);
1414                         if (op & EXTENT_SET_WRITEBACK)
1415                                 set_page_writeback(pages[i]);
1416                         if (op & EXTENT_END_WRITEBACK)
1417                                 end_page_writeback(pages[i]);
1418                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1419                                 unlock_page(pages[i]);
1420                         page_cache_release(pages[i]);
1421                 }
1422                 nr_pages -= ret;
1423                 index += ret;
1424                 cond_resched();
1425         }
1426         return 0;
1427 }
1428
1429 /*
1430  * count the number of bytes in the tree that have a given bit(s)
1431  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1432  * cached.  The total number found is returned.
1433  */
1434 u64 count_range_bits(struct extent_io_tree *tree,
1435                      u64 *start, u64 search_end, u64 max_bytes,
1436                      unsigned long bits)
1437 {
1438         struct rb_node *node;
1439         struct extent_state *state;
1440         u64 cur_start = *start;
1441         u64 total_bytes = 0;
1442         int found = 0;
1443
1444         if (search_end <= cur_start) {
1445                 WARN_ON(1);
1446                 return 0;
1447         }
1448
1449         spin_lock(&tree->lock);
1450         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1451                 total_bytes = tree->dirty_bytes;
1452                 goto out;
1453         }
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, cur_start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->start > search_end)
1465                         break;
1466                 if (state->end >= cur_start && (state->state & bits)) {
1467                         total_bytes += min(search_end, state->end) + 1 -
1468                                        max(cur_start, state->start);
1469                         if (total_bytes >= max_bytes)
1470                                 break;
1471                         if (!found) {
1472                                 *start = state->start;
1473                                 found = 1;
1474                         }
1475                 }
1476                 node = rb_next(node);
1477                 if (!node)
1478                         break;
1479         }
1480 out:
1481         spin_unlock(&tree->lock);
1482         return total_bytes;
1483 }
1484
1485 /*
1486  * set the private field for a given byte offset in the tree.  If there isn't
1487  * an extent_state there already, this does nothing.
1488  */
1489 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1490 {
1491         struct rb_node *node;
1492         struct extent_state *state;
1493         int ret = 0;
1494
1495         spin_lock(&tree->lock);
1496         /*
1497          * this search will find all the extents that end after
1498          * our range starts.
1499          */
1500         node = tree_search(tree, start);
1501         if (!node) {
1502                 ret = -ENOENT;
1503                 goto out;
1504         }
1505         state = rb_entry(node, struct extent_state, rb_node);
1506         if (state->start != start) {
1507                 ret = -ENOENT;
1508                 goto out;
1509         }
1510         state->private = private;
1511 out:
1512         spin_unlock(&tree->lock);
1513         return ret;
1514 }
1515
1516 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1517 {
1518         struct rb_node *node;
1519         struct extent_state *state;
1520         int ret = 0;
1521
1522         spin_lock(&tree->lock);
1523         /*
1524          * this search will find all the extents that end after
1525          * our range starts.
1526          */
1527         node = tree_search(tree, start);
1528         if (!node) {
1529                 ret = -ENOENT;
1530                 goto out;
1531         }
1532         state = rb_entry(node, struct extent_state, rb_node);
1533         if (state->start != start) {
1534                 ret = -ENOENT;
1535                 goto out;
1536         }
1537         *private = state->private;
1538 out:
1539         spin_unlock(&tree->lock);
1540         return ret;
1541 }
1542
1543 /*
1544  * searches a range in the state tree for a given mask.
1545  * If 'filled' == 1, this returns 1 only if every extent in the tree
1546  * has the bits set.  Otherwise, 1 is returned if any bit in the
1547  * range is found set.
1548  */
1549 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1550                    int bits, int filled, struct extent_state *cached)
1551 {
1552         struct extent_state *state = NULL;
1553         struct rb_node *node;
1554         int bitset = 0;
1555
1556         spin_lock(&tree->lock);
1557         if (cached && cached->tree && cached->start == start)
1558                 node = &cached->rb_node;
1559         else
1560                 node = tree_search(tree, start);
1561         while (node && start <= end) {
1562                 state = rb_entry(node, struct extent_state, rb_node);
1563
1564                 if (filled && state->start > start) {
1565                         bitset = 0;
1566                         break;
1567                 }
1568
1569                 if (state->start > end)
1570                         break;
1571
1572                 if (state->state & bits) {
1573                         bitset = 1;
1574                         if (!filled)
1575                                 break;
1576                 } else if (filled) {
1577                         bitset = 0;
1578                         break;
1579                 }
1580
1581                 if (state->end == (u64)-1)
1582                         break;
1583
1584                 start = state->end + 1;
1585                 if (start > end)
1586                         break;
1587                 node = rb_next(node);
1588                 if (!node) {
1589                         if (filled)
1590                                 bitset = 0;
1591                         break;
1592                 }
1593         }
1594         spin_unlock(&tree->lock);
1595         return bitset;
1596 }
1597
1598 /*
1599  * helper function to set a given page up to date if all the
1600  * extents in the tree for that page are up to date
1601  */
1602 static int check_page_uptodate(struct extent_io_tree *tree,
1603                                struct page *page)
1604 {
1605         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1606         u64 end = start + PAGE_CACHE_SIZE - 1;
1607         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1608                 SetPageUptodate(page);
1609         return 0;
1610 }
1611
1612 /*
1613  * helper function to unlock a page if all the extents in the tree
1614  * for that page are unlocked
1615  */
1616 static int check_page_locked(struct extent_io_tree *tree,
1617                              struct page *page)
1618 {
1619         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1620         u64 end = start + PAGE_CACHE_SIZE - 1;
1621         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1622                 unlock_page(page);
1623         return 0;
1624 }
1625
1626 /*
1627  * helper function to end page writeback if all the extents
1628  * in the tree for that page are done with writeback
1629  */
1630 static int check_page_writeback(struct extent_io_tree *tree,
1631                              struct page *page)
1632 {
1633         end_page_writeback(page);
1634         return 0;
1635 }
1636
1637 /* lots and lots of room for performance fixes in the end_bio funcs */
1638
1639 /*
1640  * after a writepage IO is done, we need to:
1641  * clear the uptodate bits on error
1642  * clear the writeback bits in the extent tree for this IO
1643  * end_page_writeback if the page has no more pending IO
1644  *
1645  * Scheduling is not allowed, so the extent state tree is expected
1646  * to have one and only one object corresponding to this IO.
1647  */
1648 static void end_bio_extent_writepage(struct bio *bio, int err)
1649 {
1650         int uptodate = err == 0;
1651         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1652         struct extent_io_tree *tree;
1653         u64 start;
1654         u64 end;
1655         int whole_page;
1656         int ret;
1657
1658         do {
1659                 struct page *page = bvec->bv_page;
1660                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1661
1662                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1663                          bvec->bv_offset;
1664                 end = start + bvec->bv_len - 1;
1665
1666                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1667                         whole_page = 1;
1668                 else
1669                         whole_page = 0;
1670
1671                 if (--bvec >= bio->bi_io_vec)
1672                         prefetchw(&bvec->bv_page->flags);
1673                 if (tree->ops && tree->ops->writepage_end_io_hook) {
1674                         ret = tree->ops->writepage_end_io_hook(page, start,
1675                                                        end, NULL, uptodate);
1676                         if (ret)
1677                                 uptodate = 0;
1678                 }
1679
1680                 if (!uptodate && tree->ops &&
1681                     tree->ops->writepage_io_failed_hook) {
1682                         ret = tree->ops->writepage_io_failed_hook(bio, page,
1683                                                          start, end, NULL);
1684                         if (ret == 0) {
1685                                 uptodate = (err == 0);
1686                                 continue;
1687                         }
1688                 }
1689
1690                 if (!uptodate) {
1691                         clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1692                         ClearPageUptodate(page);
1693                         SetPageError(page);
1694                 }
1695
1696                 if (whole_page)
1697                         end_page_writeback(page);
1698                 else
1699                         check_page_writeback(tree, page);
1700         } while (bvec >= bio->bi_io_vec);
1701
1702         bio_put(bio);
1703 }
1704
1705 /*
1706  * after a readpage IO is done, we need to:
1707  * clear the uptodate bits on error
1708  * set the uptodate bits if things worked
1709  * set the page up to date if all extents in the tree are uptodate
1710  * clear the lock bit in the extent tree
1711  * unlock the page if there are no other extents locked for it
1712  *
1713  * Scheduling is not allowed, so the extent state tree is expected
1714  * to have one and only one object corresponding to this IO.
1715  */
1716 static void end_bio_extent_readpage(struct bio *bio, int err)
1717 {
1718         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1719         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1720         struct bio_vec *bvec = bio->bi_io_vec;
1721         struct extent_io_tree *tree;
1722         u64 start;
1723         u64 end;
1724         int whole_page;
1725         int ret;
1726
1727         if (err)
1728                 uptodate = 0;
1729
1730         do {
1731                 struct page *page = bvec->bv_page;
1732                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1733
1734                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1735                         bvec->bv_offset;
1736                 end = start + bvec->bv_len - 1;
1737
1738                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1739                         whole_page = 1;
1740                 else
1741                         whole_page = 0;
1742
1743                 if (++bvec <= bvec_end)
1744                         prefetchw(&bvec->bv_page->flags);
1745
1746                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1747                         ret = tree->ops->readpage_end_io_hook(page, start, end,
1748                                                               NULL);
1749                         if (ret)
1750                                 uptodate = 0;
1751                 }
1752                 if (!uptodate && tree->ops &&
1753                     tree->ops->readpage_io_failed_hook) {
1754                         ret = tree->ops->readpage_io_failed_hook(bio, page,
1755                                                          start, end, NULL);
1756                         if (ret == 0) {
1757                                 uptodate =
1758                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
1759                                 if (err)
1760                                         uptodate = 0;
1761                                 continue;
1762                         }
1763                 }
1764
1765                 if (uptodate) {
1766                         set_extent_uptodate(tree, start, end,
1767                                             GFP_ATOMIC);
1768                 }
1769                 unlock_extent(tree, start, end, GFP_ATOMIC);
1770
1771                 if (whole_page) {
1772                         if (uptodate) {
1773                                 SetPageUptodate(page);
1774                         } else {
1775                                 ClearPageUptodate(page);
1776                                 SetPageError(page);
1777                         }
1778                         unlock_page(page);
1779                 } else {
1780                         if (uptodate) {
1781                                 check_page_uptodate(tree, page);
1782                         } else {
1783                                 ClearPageUptodate(page);
1784                                 SetPageError(page);
1785                         }
1786                         check_page_locked(tree, page);
1787                 }
1788         } while (bvec <= bvec_end);
1789
1790         bio_put(bio);
1791 }
1792
1793 /*
1794  * IO done from prepare_write is pretty simple, we just unlock
1795  * the structs in the extent tree when done, and set the uptodate bits
1796  * as appropriate.
1797  */
1798 static void end_bio_extent_preparewrite(struct bio *bio, int err)
1799 {
1800         const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1801         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1802         struct extent_io_tree *tree;
1803         u64 start;
1804         u64 end;
1805
1806         do {
1807                 struct page *page = bvec->bv_page;
1808                 tree = &BTRFS_I(page->mapping->host)->io_tree;
1809
1810                 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1811                         bvec->bv_offset;
1812                 end = start + bvec->bv_len - 1;
1813
1814                 if (--bvec >= bio->bi_io_vec)
1815                         prefetchw(&bvec->bv_page->flags);
1816
1817                 if (uptodate) {
1818                         set_extent_uptodate(tree, start, end, GFP_ATOMIC);
1819                 } else {
1820                         ClearPageUptodate(page);
1821                         SetPageError(page);
1822                 }
1823
1824                 unlock_extent(tree, start, end, GFP_ATOMIC);
1825
1826         } while (bvec >= bio->bi_io_vec);
1827
1828         bio_put(bio);
1829 }
1830
1831 struct bio *
1832 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1833                 gfp_t gfp_flags)
1834 {
1835         struct bio *bio;
1836
1837         bio = bio_alloc(gfp_flags, nr_vecs);
1838
1839         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1840                 while (!bio && (nr_vecs /= 2))
1841                         bio = bio_alloc(gfp_flags, nr_vecs);
1842         }
1843
1844         if (bio) {
1845                 bio->bi_size = 0;
1846                 bio->bi_bdev = bdev;
1847                 bio->bi_sector = first_sector;
1848         }
1849         return bio;
1850 }
1851
1852 static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1853                           unsigned long bio_flags)
1854 {
1855         int ret = 0;
1856         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1857         struct page *page = bvec->bv_page;
1858         struct extent_io_tree *tree = bio->bi_private;
1859         u64 start;
1860
1861         start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1862
1863         bio->bi_private = NULL;
1864
1865         bio_get(bio);
1866
1867         if (tree->ops && tree->ops->submit_bio_hook)
1868                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1869                                            mirror_num, bio_flags, start);
1870         else
1871                 submit_bio(rw, bio);
1872         if (bio_flagged(bio, BIO_EOPNOTSUPP))
1873                 ret = -EOPNOTSUPP;
1874         bio_put(bio);
1875         return ret;
1876 }
1877
1878 static int submit_extent_page(int rw, struct extent_io_tree *tree,
1879                               struct page *page, sector_t sector,
1880                               size_t size, unsigned long offset,
1881                               struct block_device *bdev,
1882                               struct bio **bio_ret,
1883                               unsigned long max_pages,
1884                               bio_end_io_t end_io_func,
1885                               int mirror_num,
1886                               unsigned long prev_bio_flags,
1887                               unsigned long bio_flags)
1888 {
1889         int ret = 0;
1890         struct bio *bio;
1891         int nr;
1892         int contig = 0;
1893         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1894         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1895         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1896
1897         if (bio_ret && *bio_ret) {
1898                 bio = *bio_ret;
1899                 if (old_compressed)
1900                         contig = bio->bi_sector == sector;
1901                 else
1902                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
1903                                 sector;
1904
1905                 if (prev_bio_flags != bio_flags || !contig ||
1906                     (tree->ops && tree->ops->merge_bio_hook &&
1907                      tree->ops->merge_bio_hook(page, offset, page_size, bio,
1908                                                bio_flags)) ||
1909                     bio_add_page(bio, page, page_size, offset) < page_size) {
1910                         ret = submit_one_bio(rw, bio, mirror_num,
1911                                              prev_bio_flags);
1912                         bio = NULL;
1913                 } else {
1914                         return 0;
1915                 }
1916         }
1917         if (this_compressed)
1918                 nr = BIO_MAX_PAGES;
1919         else
1920                 nr = bio_get_nr_vecs(bdev);
1921
1922         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1923         if (!bio)
1924                 return -ENOMEM;
1925
1926         bio_add_page(bio, page, page_size, offset);
1927         bio->bi_end_io = end_io_func;
1928         bio->bi_private = tree;
1929
1930         if (bio_ret)
1931                 *bio_ret = bio;
1932         else
1933                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1934
1935         return ret;
1936 }
1937
1938 void set_page_extent_mapped(struct page *page)
1939 {
1940         if (!PagePrivate(page)) {
1941                 SetPagePrivate(page);
1942                 page_cache_get(page);
1943                 set_page_private(page, EXTENT_PAGE_PRIVATE);
1944         }
1945 }
1946
1947 static void set_page_extent_head(struct page *page, unsigned long len)
1948 {
1949         WARN_ON(!PagePrivate(page));
1950         set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1951 }
1952
1953 /*
1954  * basic readpage implementation.  Locked extent state structs are inserted
1955  * into the tree that are removed when the IO is done (by the end_io
1956  * handlers)
1957  */
1958 static int __extent_read_full_page(struct extent_io_tree *tree,
1959                                    struct page *page,
1960                                    get_extent_t *get_extent,
1961                                    struct bio **bio, int mirror_num,
1962                                    unsigned long *bio_flags)
1963 {
1964         struct inode *inode = page->mapping->host;
1965         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1966         u64 page_end = start + PAGE_CACHE_SIZE - 1;
1967         u64 end;
1968         u64 cur = start;
1969         u64 extent_offset;
1970         u64 last_byte = i_size_read(inode);
1971         u64 block_start;
1972         u64 cur_end;
1973         sector_t sector;
1974         struct extent_map *em;
1975         struct block_device *bdev;
1976         struct btrfs_ordered_extent *ordered;
1977         int ret;
1978         int nr = 0;
1979         size_t page_offset = 0;
1980         size_t iosize;
1981         size_t disk_io_size;
1982         size_t blocksize = inode->i_sb->s_blocksize;
1983         unsigned long this_bio_flag = 0;
1984
1985         set_page_extent_mapped(page);
1986
1987         end = page_end;
1988         while (1) {
1989                 lock_extent(tree, start, end, GFP_NOFS);
1990                 ordered = btrfs_lookup_ordered_extent(inode, start);
1991                 if (!ordered)
1992                         break;
1993                 unlock_extent(tree, start, end, GFP_NOFS);
1994                 btrfs_start_ordered_extent(inode, ordered, 1);
1995                 btrfs_put_ordered_extent(ordered);
1996         }
1997
1998         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1999                 char *userpage;
2000                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2001
2002                 if (zero_offset) {
2003                         iosize = PAGE_CACHE_SIZE - zero_offset;
2004                         userpage = kmap_atomic(page, KM_USER0);
2005                         memset(userpage + zero_offset, 0, iosize);
2006                         flush_dcache_page(page);
2007                         kunmap_atomic(userpage, KM_USER0);
2008                 }
2009         }
2010         while (cur <= end) {
2011                 if (cur >= last_byte) {
2012                         char *userpage;
2013                         iosize = PAGE_CACHE_SIZE - page_offset;
2014                         userpage = kmap_atomic(page, KM_USER0);
2015                         memset(userpage + page_offset, 0, iosize);
2016                         flush_dcache_page(page);
2017                         kunmap_atomic(userpage, KM_USER0);
2018                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2019                                             GFP_NOFS);
2020                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2021                         break;
2022                 }
2023                 em = get_extent(inode, page, page_offset, cur,
2024                                 end - cur + 1, 0);
2025                 if (IS_ERR(em) || !em) {
2026                         SetPageError(page);
2027                         unlock_extent(tree, cur, end, GFP_NOFS);
2028                         break;
2029                 }
2030                 extent_offset = cur - em->start;
2031                 BUG_ON(extent_map_end(em) <= cur);
2032                 BUG_ON(end < cur);
2033
2034                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2035                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2036                         extent_set_compress_type(&this_bio_flag,
2037                                                  em->compress_type);
2038                 }
2039
2040                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2041                 cur_end = min(extent_map_end(em) - 1, end);
2042                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2043                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2044                         disk_io_size = em->block_len;
2045                         sector = em->block_start >> 9;
2046                 } else {
2047                         sector = (em->block_start + extent_offset) >> 9;
2048                         disk_io_size = iosize;
2049                 }
2050                 bdev = em->bdev;
2051                 block_start = em->block_start;
2052                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2053                         block_start = EXTENT_MAP_HOLE;
2054                 free_extent_map(em);
2055                 em = NULL;
2056
2057                 /* we've found a hole, just zero and go on */
2058                 if (block_start == EXTENT_MAP_HOLE) {
2059                         char *userpage;
2060                         userpage = kmap_atomic(page, KM_USER0);
2061                         memset(userpage + page_offset, 0, iosize);
2062                         flush_dcache_page(page);
2063                         kunmap_atomic(userpage, KM_USER0);
2064
2065                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2066                                             GFP_NOFS);
2067                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2068                         cur = cur + iosize;
2069                         page_offset += iosize;
2070                         continue;
2071                 }
2072                 /* the get_extent function already copied into the page */
2073                 if (test_range_bit(tree, cur, cur_end,
2074                                    EXTENT_UPTODATE, 1, NULL)) {
2075                         check_page_uptodate(tree, page);
2076                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2077                         cur = cur + iosize;
2078                         page_offset += iosize;
2079                         continue;
2080                 }
2081                 /* we have an inline extent but it didn't get marked up
2082                  * to date.  Error out
2083                  */
2084                 if (block_start == EXTENT_MAP_INLINE) {
2085                         SetPageError(page);
2086                         unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2087                         cur = cur + iosize;
2088                         page_offset += iosize;
2089                         continue;
2090                 }
2091
2092                 ret = 0;
2093                 if (tree->ops && tree->ops->readpage_io_hook) {
2094                         ret = tree->ops->readpage_io_hook(page, cur,
2095                                                           cur + iosize - 1);
2096                 }
2097                 if (!ret) {
2098                         unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2099                         pnr -= page->index;
2100                         ret = submit_extent_page(READ, tree, page,
2101                                          sector, disk_io_size, page_offset,
2102                                          bdev, bio, pnr,
2103                                          end_bio_extent_readpage, mirror_num,
2104                                          *bio_flags,
2105                                          this_bio_flag);
2106                         nr++;
2107                         *bio_flags = this_bio_flag;
2108                 }
2109                 if (ret)
2110                         SetPageError(page);
2111                 cur = cur + iosize;
2112                 page_offset += iosize;
2113         }
2114         if (!nr) {
2115                 if (!PageError(page))
2116                         SetPageUptodate(page);
2117                 unlock_page(page);
2118         }
2119         return 0;
2120 }
2121
2122 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2123                             get_extent_t *get_extent)
2124 {
2125         struct bio *bio = NULL;
2126         unsigned long bio_flags = 0;
2127         int ret;
2128
2129         ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2130                                       &bio_flags);
2131         if (bio)
2132                 ret = submit_one_bio(READ, bio, 0, bio_flags);
2133         return ret;
2134 }
2135
2136 static noinline void update_nr_written(struct page *page,
2137                                       struct writeback_control *wbc,
2138                                       unsigned long nr_written)
2139 {
2140         wbc->nr_to_write -= nr_written;
2141         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2142             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2143                 page->mapping->writeback_index = page->index + nr_written;
2144 }
2145
2146 /*
2147  * the writepage semantics are similar to regular writepage.  extent
2148  * records are inserted to lock ranges in the tree, and as dirty areas
2149  * are found, they are marked writeback.  Then the lock bits are removed
2150  * and the end_io handler clears the writeback ranges
2151  */
2152 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2153                               void *data)
2154 {
2155         struct inode *inode = page->mapping->host;
2156         struct extent_page_data *epd = data;
2157         struct extent_io_tree *tree = epd->tree;
2158         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2159         u64 delalloc_start;
2160         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2161         u64 end;
2162         u64 cur = start;
2163         u64 extent_offset;
2164         u64 last_byte = i_size_read(inode);
2165         u64 block_start;
2166         u64 iosize;
2167         sector_t sector;
2168         struct extent_state *cached_state = NULL;
2169         struct extent_map *em;
2170         struct block_device *bdev;
2171         int ret;
2172         int nr = 0;
2173         size_t pg_offset = 0;
2174         size_t blocksize;
2175         loff_t i_size = i_size_read(inode);
2176         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2177         u64 nr_delalloc;
2178         u64 delalloc_end;
2179         int page_started;
2180         int compressed;
2181         int write_flags;
2182         unsigned long nr_written = 0;
2183
2184         if (wbc->sync_mode == WB_SYNC_ALL)
2185                 write_flags = WRITE_SYNC_PLUG;
2186         else
2187                 write_flags = WRITE;
2188
2189         WARN_ON(!PageLocked(page));
2190         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2191         if (page->index > end_index ||
2192            (page->index == end_index && !pg_offset)) {
2193                 page->mapping->a_ops->invalidatepage(page, 0);
2194                 unlock_page(page);
2195                 return 0;
2196         }
2197
2198         if (page->index == end_index) {
2199                 char *userpage;
2200
2201                 userpage = kmap_atomic(page, KM_USER0);
2202                 memset(userpage + pg_offset, 0,
2203                        PAGE_CACHE_SIZE - pg_offset);
2204                 kunmap_atomic(userpage, KM_USER0);
2205                 flush_dcache_page(page);
2206         }
2207         pg_offset = 0;
2208
2209         set_page_extent_mapped(page);
2210
2211         delalloc_start = start;
2212         delalloc_end = 0;
2213         page_started = 0;
2214         if (!epd->extent_locked) {
2215                 u64 delalloc_to_write = 0;
2216                 /*
2217                  * make sure the wbc mapping index is at least updated
2218                  * to this page.
2219                  */
2220                 update_nr_written(page, wbc, 0);
2221
2222                 while (delalloc_end < page_end) {
2223                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2224                                                        page,
2225                                                        &delalloc_start,
2226                                                        &delalloc_end,
2227                                                        128 * 1024 * 1024);
2228                         if (nr_delalloc == 0) {
2229                                 delalloc_start = delalloc_end + 1;
2230                                 continue;
2231                         }
2232                         tree->ops->fill_delalloc(inode, page, delalloc_start,
2233                                                  delalloc_end, &page_started,
2234                                                  &nr_written);
2235                         /*
2236                          * delalloc_end is already one less than the total
2237                          * length, so we don't subtract one from
2238                          * PAGE_CACHE_SIZE
2239                          */
2240                         delalloc_to_write += (delalloc_end - delalloc_start +
2241                                               PAGE_CACHE_SIZE) >>
2242                                               PAGE_CACHE_SHIFT;
2243                         delalloc_start = delalloc_end + 1;
2244                 }
2245                 if (wbc->nr_to_write < delalloc_to_write) {
2246                         int thresh = 8192;
2247
2248                         if (delalloc_to_write < thresh * 2)
2249                                 thresh = delalloc_to_write;
2250                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2251                                                  thresh);
2252                 }
2253
2254                 /* did the fill delalloc function already unlock and start
2255                  * the IO?
2256                  */
2257                 if (page_started) {
2258                         ret = 0;
2259                         /*
2260                          * we've unlocked the page, so we can't update
2261                          * the mapping's writeback index, just update
2262                          * nr_to_write.
2263                          */
2264                         wbc->nr_to_write -= nr_written;
2265                         goto done_unlocked;
2266                 }
2267         }
2268         if (tree->ops && tree->ops->writepage_start_hook) {
2269                 ret = tree->ops->writepage_start_hook(page, start,
2270                                                       page_end);
2271                 if (ret == -EAGAIN) {
2272                         redirty_page_for_writepage(wbc, page);
2273                         update_nr_written(page, wbc, nr_written);
2274                         unlock_page(page);
2275                         ret = 0;
2276                         goto done_unlocked;
2277                 }
2278         }
2279
2280         /*
2281          * we don't want to touch the inode after unlocking the page,
2282          * so we update the mapping writeback index now
2283          */
2284         update_nr_written(page, wbc, nr_written + 1);
2285
2286         end = page_end;
2287         if (last_byte <= start) {
2288                 if (tree->ops && tree->ops->writepage_end_io_hook)
2289                         tree->ops->writepage_end_io_hook(page, start,
2290                                                          page_end, NULL, 1);
2291                 goto done;
2292         }
2293
2294         blocksize = inode->i_sb->s_blocksize;
2295
2296         while (cur <= end) {
2297                 if (cur >= last_byte) {
2298                         if (tree->ops && tree->ops->writepage_end_io_hook)
2299                                 tree->ops->writepage_end_io_hook(page, cur,
2300                                                          page_end, NULL, 1);
2301                         break;
2302                 }
2303                 em = epd->get_extent(inode, page, pg_offset, cur,
2304                                      end - cur + 1, 1);
2305                 if (IS_ERR(em) || !em) {
2306                         SetPageError(page);
2307                         break;
2308                 }
2309
2310                 extent_offset = cur - em->start;
2311                 BUG_ON(extent_map_end(em) <= cur);
2312                 BUG_ON(end < cur);
2313                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2314                 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2315                 sector = (em->block_start + extent_offset) >> 9;
2316                 bdev = em->bdev;
2317                 block_start = em->block_start;
2318                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2319                 free_extent_map(em);
2320                 em = NULL;
2321
2322                 /*
2323                  * compressed and inline extents are written through other
2324                  * paths in the FS
2325                  */
2326                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2327                     block_start == EXTENT_MAP_INLINE) {
2328                         /*
2329                          * end_io notification does not happen here for
2330                          * compressed extents
2331                          */
2332                         if (!compressed && tree->ops &&
2333                             tree->ops->writepage_end_io_hook)
2334                                 tree->ops->writepage_end_io_hook(page, cur,
2335                                                          cur + iosize - 1,
2336                                                          NULL, 1);
2337                         else if (compressed) {
2338                                 /* we don't want to end_page_writeback on
2339                                  * a compressed extent.  this happens
2340                                  * elsewhere
2341                                  */
2342                                 nr++;
2343                         }
2344
2345                         cur += iosize;
2346                         pg_offset += iosize;
2347                         continue;
2348                 }
2349                 /* leave this out until we have a page_mkwrite call */
2350                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2351                                    EXTENT_DIRTY, 0, NULL)) {
2352                         cur = cur + iosize;
2353                         pg_offset += iosize;
2354                         continue;
2355                 }
2356
2357                 if (tree->ops && tree->ops->writepage_io_hook) {
2358                         ret = tree->ops->writepage_io_hook(page, cur,
2359                                                 cur + iosize - 1);
2360                 } else {
2361                         ret = 0;
2362                 }
2363                 if (ret) {
2364                         SetPageError(page);
2365                 } else {
2366                         unsigned long max_nr = end_index + 1;
2367
2368                         set_range_writeback(tree, cur, cur + iosize - 1);
2369                         if (!PageWriteback(page)) {
2370                                 printk(KERN_ERR "btrfs warning page %lu not "
2371                                        "writeback, cur %llu end %llu\n",
2372                                        page->index, (unsigned long long)cur,
2373                                        (unsigned long long)end);
2374                         }
2375
2376                         ret = submit_extent_page(write_flags, tree, page,
2377                                                  sector, iosize, pg_offset,
2378                                                  bdev, &epd->bio, max_nr,
2379                                                  end_bio_extent_writepage,
2380                                                  0, 0, 0);
2381                         if (ret)
2382                                 SetPageError(page);
2383                 }
2384                 cur = cur + iosize;
2385                 pg_offset += iosize;
2386                 nr++;
2387         }
2388 done:
2389         if (nr == 0) {
2390                 /* make sure the mapping tag for page dirty gets cleared */
2391                 set_page_writeback(page);
2392                 end_page_writeback(page);
2393         }
2394         unlock_page(page);
2395
2396 done_unlocked:
2397
2398         /* drop our reference on any cached states */
2399         free_extent_state(cached_state);
2400         return 0;
2401 }
2402
2403 /**
2404  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2405  * @mapping: address space structure to write
2406  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2407  * @writepage: function called for each page
2408  * @data: data passed to writepage function
2409  *
2410  * If a page is already under I/O, write_cache_pages() skips it, even
2411  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2412  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2413  * and msync() need to guarantee that all the data which was dirty at the time
2414  * the call was made get new I/O started against them.  If wbc->sync_mode is
2415  * WB_SYNC_ALL then we were called for data integrity and we must wait for
2416  * existing IO to complete.
2417  */
2418 static int extent_write_cache_pages(struct extent_io_tree *tree,
2419                              struct address_space *mapping,
2420                              struct writeback_control *wbc,
2421                              writepage_t writepage, void *data,
2422                              void (*flush_fn)(void *))
2423 {
2424         int ret = 0;
2425         int done = 0;
2426         int nr_to_write_done = 0;
2427         struct pagevec pvec;
2428         int nr_pages;
2429         pgoff_t index;
2430         pgoff_t end;            /* Inclusive */
2431         int scanned = 0;
2432
2433         pagevec_init(&pvec, 0);
2434         if (wbc->range_cyclic) {
2435                 index = mapping->writeback_index; /* Start from prev offset */
2436                 end = -1;
2437         } else {
2438                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2439                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2440                 scanned = 1;
2441         }
2442 retry:
2443         while (!done && !nr_to_write_done && (index <= end) &&
2444                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
2445                               PAGECACHE_TAG_DIRTY, min(end - index,
2446                                   (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2447                 unsigned i;
2448
2449                 scanned = 1;
2450                 for (i = 0; i < nr_pages; i++) {
2451                         struct page *page = pvec.pages[i];
2452
2453                         /*
2454                          * At this point we hold neither mapping->tree_lock nor
2455                          * lock on the page itself: the page may be truncated or
2456                          * invalidated (changing page->mapping to NULL), or even
2457                          * swizzled back from swapper_space to tmpfs file
2458                          * mapping
2459                          */
2460                         if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2461                                 tree->ops->write_cache_pages_lock_hook(page);
2462                         else
2463                                 lock_page(page);
2464
2465                         if (unlikely(page->mapping != mapping)) {
2466                                 unlock_page(page);
2467                                 continue;
2468                         }
2469
2470                         if (!wbc->range_cyclic && page->index > end) {
2471                                 done = 1;
2472                                 unlock_page(page);
2473                                 continue;
2474                         }
2475
2476                         if (wbc->sync_mode != WB_SYNC_NONE) {
2477                                 if (PageWriteback(page))
2478                                         flush_fn(data);
2479                                 wait_on_page_writeback(page);
2480                         }
2481
2482                         if (PageWriteback(page) ||
2483                             !clear_page_dirty_for_io(page)) {
2484                                 unlock_page(page);
2485                                 continue;
2486                         }
2487
2488                         ret = (*writepage)(page, wbc, data);
2489
2490                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2491                                 unlock_page(page);
2492                                 ret = 0;
2493                         }
2494                         if (ret)
2495                                 done = 1;
2496
2497                         /*
2498                          * the filesystem may choose to bump up nr_to_write.
2499                          * We have to make sure to honor the new nr_to_write
2500                          * at any time
2501                          */
2502                         nr_to_write_done = wbc->nr_to_write <= 0;
2503                 }
2504                 pagevec_release(&pvec);
2505                 cond_resched();
2506         }
2507         if (!scanned && !done) {
2508                 /*
2509                  * We hit the last page and there is more work to be done: wrap
2510                  * back to the start of the file
2511                  */
2512                 scanned = 1;
2513                 index = 0;
2514                 goto retry;
2515         }
2516         return ret;
2517 }
2518
2519 static void flush_epd_write_bio(struct extent_page_data *epd)
2520 {
2521         if (epd->bio) {
2522                 if (epd->sync_io)
2523                         submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2524                 else
2525                         submit_one_bio(WRITE, epd->bio, 0, 0);
2526                 epd->bio = NULL;
2527         }
2528 }
2529
2530 static noinline void flush_write_bio(void *data)
2531 {
2532         struct extent_page_data *epd = data;
2533         flush_epd_write_bio(epd);
2534 }
2535
2536 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2537                           get_extent_t *get_extent,
2538                           struct writeback_control *wbc)
2539 {
2540         int ret;
2541         struct address_space *mapping = page->mapping;
2542         struct extent_page_data epd = {
2543                 .bio = NULL,
2544                 .tree = tree,
2545                 .get_extent = get_extent,
2546                 .extent_locked = 0,
2547                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2548         };
2549         struct writeback_control wbc_writepages = {
2550                 .sync_mode      = wbc->sync_mode,
2551                 .older_than_this = NULL,
2552                 .nr_to_write    = 64,
2553                 .range_start    = page_offset(page) + PAGE_CACHE_SIZE,
2554                 .range_end      = (loff_t)-1,
2555         };
2556
2557         ret = __extent_writepage(page, wbc, &epd);
2558
2559         extent_write_cache_pages(tree, mapping, &wbc_writepages,
2560                                  __extent_writepage, &epd, flush_write_bio);
2561         flush_epd_write_bio(&epd);
2562         return ret;
2563 }
2564
2565 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2566                               u64 start, u64 end, get_extent_t *get_extent,
2567                               int mode)
2568 {
2569         int ret = 0;
2570         struct address_space *mapping = inode->i_mapping;
2571         struct page *page;
2572         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2573                 PAGE_CACHE_SHIFT;
2574
2575         struct extent_page_data epd = {
2576                 .bio = NULL,
2577                 .tree = tree,
2578                 .get_extent = get_extent,
2579                 .extent_locked = 1,
2580                 .sync_io = mode == WB_SYNC_ALL,
2581         };
2582         struct writeback_control wbc_writepages = {
2583                 .sync_mode      = mode,
2584                 .older_than_this = NULL,
2585                 .nr_to_write    = nr_pages * 2,
2586                 .range_start    = start,
2587                 .range_end      = end + 1,
2588         };
2589
2590         while (start <= end) {
2591                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2592                 if (clear_page_dirty_for_io(page))
2593                         ret = __extent_writepage(page, &wbc_writepages, &epd);
2594                 else {
2595                         if (tree->ops && tree->ops->writepage_end_io_hook)
2596                                 tree->ops->writepage_end_io_hook(page, start,
2597                                                  start + PAGE_CACHE_SIZE - 1,
2598                                                  NULL, 1);
2599                         unlock_page(page);
2600                 }
2601                 page_cache_release(page);
2602                 start += PAGE_CACHE_SIZE;
2603         }
2604
2605         flush_epd_write_bio(&epd);
2606         return ret;
2607 }
2608
2609 int extent_writepages(struct extent_io_tree *tree,
2610                       struct address_space *mapping,
2611                       get_extent_t *get_extent,
2612                       struct writeback_control *wbc)
2613 {
2614         int ret = 0;
2615         struct extent_page_data epd = {
2616                 .bio = NULL,
2617                 .tree = tree,
2618                 .get_extent = get_extent,
2619                 .extent_locked = 0,
2620                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2621         };
2622
2623         ret = extent_write_cache_pages(tree, mapping, wbc,
2624                                        __extent_writepage, &epd,
2625                                        flush_write_bio);
2626         flush_epd_write_bio(&epd);
2627         return ret;
2628 }
2629
2630 int extent_readpages(struct extent_io_tree *tree,
2631                      struct address_space *mapping,
2632                      struct list_head *pages, unsigned nr_pages,
2633                      get_extent_t get_extent)
2634 {
2635         struct bio *bio = NULL;
2636         unsigned page_idx;
2637         unsigned long bio_flags = 0;
2638
2639         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2640                 struct page *page = list_entry(pages->prev, struct page, lru);
2641
2642                 prefetchw(&page->flags);
2643                 list_del(&page->lru);
2644                 if (!add_to_page_cache_lru(page, mapping,
2645                                         page->index, GFP_KERNEL)) {
2646                         __extent_read_full_page(tree, page, get_extent,
2647                                                 &bio, 0, &bio_flags);
2648                 }
2649                 page_cache_release(page);
2650         }
2651         BUG_ON(!list_empty(pages));
2652         if (bio)
2653                 submit_one_bio(READ, bio, 0, bio_flags);
2654         return 0;
2655 }
2656
2657 /*
2658  * basic invalidatepage code, this waits on any locked or writeback
2659  * ranges corresponding to the page, and then deletes any extent state
2660  * records from the tree
2661  */
2662 int extent_invalidatepage(struct extent_io_tree *tree,
2663                           struct page *page, unsigned long offset)
2664 {
2665         struct extent_state *cached_state = NULL;
2666         u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2667         u64 end = start + PAGE_CACHE_SIZE - 1;
2668         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2669
2670         start += (offset + blocksize - 1) & ~(blocksize - 1);
2671         if (start > end)
2672                 return 0;
2673
2674         lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2675         wait_on_page_writeback(page);
2676         clear_extent_bit(tree, start, end,
2677                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2678                          EXTENT_DO_ACCOUNTING,
2679                          1, 1, &cached_state, GFP_NOFS);
2680         return 0;
2681 }
2682
2683 /*
2684  * simple commit_write call, set_range_dirty is used to mark both
2685  * the pages and the extent records as dirty
2686  */
2687 int extent_commit_write(struct extent_io_tree *tree,
2688                         struct inode *inode, struct page *page,
2689                         unsigned from, unsigned to)
2690 {
2691         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2692
2693         set_page_extent_mapped(page);
2694         set_page_dirty(page);
2695
2696         if (pos > inode->i_size) {
2697                 i_size_write(inode, pos);
2698                 mark_inode_dirty(inode);
2699         }
2700         return 0;
2701 }
2702
2703 int extent_prepare_write(struct extent_io_tree *tree,
2704                          struct inode *inode, struct page *page,
2705                          unsigned from, unsigned to, get_extent_t *get_extent)
2706 {
2707         u64 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
2708         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
2709         u64 block_start;
2710         u64 orig_block_start;
2711         u64 block_end;
2712         u64 cur_end;
2713         struct extent_map *em;
2714         unsigned blocksize = 1 << inode->i_blkbits;
2715         size_t page_offset = 0;
2716         size_t block_off_start;
2717         size_t block_off_end;
2718         int err = 0;
2719         int iocount = 0;
2720         int ret = 0;
2721         int isnew;
2722
2723         set_page_extent_mapped(page);
2724
2725         block_start = (page_start + from) & ~((u64)blocksize - 1);
2726         block_end = (page_start + to - 1) | (blocksize - 1);
2727         orig_block_start = block_start;
2728
2729         lock_extent(tree, page_start, page_end, GFP_NOFS);
2730         while (block_start <= block_end) {
2731                 em = get_extent(inode, page, page_offset, block_start,
2732                                 block_end - block_start + 1, 1);
2733                 if (IS_ERR(em) || !em)
2734                         goto err;
2735
2736                 cur_end = min(block_end, extent_map_end(em) - 1);
2737                 block_off_start = block_start & (PAGE_CACHE_SIZE - 1);
2738                 block_off_end = block_off_start + blocksize;
2739                 isnew = clear_extent_new(tree, block_start, cur_end, GFP_NOFS);
2740
2741                 if (!PageUptodate(page) && isnew &&
2742                     (block_off_end > to || block_off_start < from)) {
2743                         void *kaddr;
2744
2745                         kaddr = kmap_atomic(page, KM_USER0);
2746                         if (block_off_end > to)
2747                                 memset(kaddr + to, 0, block_off_end - to);
2748                         if (block_off_start < from)
2749                                 memset(kaddr + block_off_start, 0,
2750                                        from - block_off_start);
2751                         flush_dcache_page(page);
2752                         kunmap_atomic(kaddr, KM_USER0);
2753                 }
2754                 if ((em->block_start != EXTENT_MAP_HOLE &&
2755                      em->block_start != EXTENT_MAP_INLINE) &&
2756                     !isnew && !PageUptodate(page) &&
2757                     (block_off_end > to || block_off_start < from) &&
2758                     !test_range_bit(tree, block_start, cur_end,
2759                                     EXTENT_UPTODATE, 1, NULL)) {
2760                         u64 sector;
2761                         u64 extent_offset = block_start - em->start;
2762                         size_t iosize;
2763                         sector = (em->block_start + extent_offset) >> 9;
2764                         iosize = (cur_end - block_start + blocksize) &
2765                                 ~((u64)blocksize - 1);
2766                         /*
2767                          * we've already got the extent locked, but we
2768                          * need to split the state such that our end_bio
2769                          * handler can clear the lock.
2770                          */
2771                         set_extent_bit(tree, block_start,
2772                                        block_start + iosize - 1,
2773                                        EXTENT_LOCKED, 0, NULL, NULL, GFP_NOFS);
2774                         ret = submit_extent_page(READ, tree, page,
2775                                          sector, iosize, page_offset, em->bdev,
2776                                          NULL, 1,
2777                                          end_bio_extent_preparewrite, 0,
2778                                          0, 0);
2779                         if (ret && !err)
2780                                 err = ret;
2781                         iocount++;
2782                         block_start = block_start + iosize;
2783                 } else {
2784                         set_extent_uptodate(tree, block_start, cur_end,
2785                                             GFP_NOFS);
2786                         unlock_extent(tree, block_start, cur_end, GFP_NOFS);
2787                         block_start = cur_end + 1;
2788                 }
2789                 page_offset = block_start & (PAGE_CACHE_SIZE - 1);
2790                 free_extent_map(em);
2791         }
2792         if (iocount) {
2793                 wait_extent_bit(tree, orig_block_start,
2794                                 block_end, EXTENT_LOCKED);
2795         }
2796         check_page_uptodate(tree, page);
2797 err:
2798         /* FIXME, zero out newly allocated blocks on error */
2799         return err;
2800 }
2801
2802 /*
2803  * a helper for releasepage, this tests for areas of the page that
2804  * are locked or under IO and drops the related state bits if it is safe
2805  * to drop the page.
2806  */
2807 int try_release_extent_state(struct extent_map_tree *map,
2808                              struct extent_io_tree *tree, struct page *page,
2809                              gfp_t mask)
2810 {
2811         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2812         u64 end = start + PAGE_CACHE_SIZE - 1;
2813         int ret = 1;
2814
2815         if (test_range_bit(tree, start, end,
2816                            EXTENT_IOBITS, 0, NULL))
2817                 ret = 0;
2818         else {
2819                 if ((mask & GFP_NOFS) == GFP_NOFS)
2820                         mask = GFP_NOFS;
2821                 /*
2822                  * at this point we can safely clear everything except the
2823                  * locked bit and the nodatasum bit
2824                  */
2825                 ret = clear_extent_bit(tree, start, end,
2826                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2827                                  0, 0, NULL, mask);
2828
2829                 /* if clear_extent_bit failed for enomem reasons,
2830                  * we can't allow the release to continue.
2831                  */
2832                 if (ret < 0)
2833                         ret = 0;
2834                 else
2835                         ret = 1;
2836         }
2837         return ret;
2838 }
2839
2840 /*
2841  * a helper for releasepage.  As long as there are no locked extents
2842  * in the range corresponding to the page, both state records and extent
2843  * map records are removed
2844  */
2845 int try_release_extent_mapping(struct extent_map_tree *map,
2846                                struct extent_io_tree *tree, struct page *page,
2847                                gfp_t mask)
2848 {
2849         struct extent_map *em;
2850         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2851         u64 end = start + PAGE_CACHE_SIZE - 1;
2852
2853         if ((mask & __GFP_WAIT) &&
2854             page->mapping->host->i_size > 16 * 1024 * 1024) {
2855                 u64 len;
2856                 while (start <= end) {
2857                         len = end - start + 1;
2858                         write_lock(&map->lock);
2859                         em = lookup_extent_mapping(map, start, len);
2860                         if (!em || IS_ERR(em)) {
2861                                 write_unlock(&map->lock);
2862                                 break;
2863                         }
2864                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2865                             em->start != start) {
2866                                 write_unlock(&map->lock);
2867                                 free_extent_map(em);
2868                                 break;
2869                         }
2870                         if (!test_range_bit(tree, em->start,
2871                                             extent_map_end(em) - 1,
2872                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
2873                                             0, NULL)) {
2874                                 remove_extent_mapping(map, em);
2875                                 /* once for the rb tree */
2876                                 free_extent_map(em);
2877                         }
2878                         start = extent_map_end(em);
2879                         write_unlock(&map->lock);
2880
2881                         /* once for us */
2882                         free_extent_map(em);
2883                 }
2884         }
2885         return try_release_extent_state(map, tree, page, mask);
2886 }
2887
2888 sector_t extent_bmap(struct address_space *mapping, sector_t iblock,
2889                 get_extent_t *get_extent)
2890 {
2891         struct inode *inode = mapping->host;
2892         struct extent_state *cached_state = NULL;
2893         u64 start = iblock << inode->i_blkbits;
2894         sector_t sector = 0;
2895         size_t blksize = (1 << inode->i_blkbits);
2896         struct extent_map *em;
2897
2898         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + blksize - 1,
2899                          0, &cached_state, GFP_NOFS);
2900         em = get_extent(inode, NULL, 0, start, blksize, 0);
2901         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start,
2902                              start + blksize - 1, &cached_state, GFP_NOFS);
2903         if (!em || IS_ERR(em))
2904                 return 0;
2905
2906         if (em->block_start > EXTENT_MAP_LAST_BYTE)
2907                 goto out;
2908
2909         sector = (em->block_start + start - em->start) >> inode->i_blkbits;
2910 out:
2911         free_extent_map(em);
2912         return sector;
2913 }
2914
2915 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2916                 __u64 start, __u64 len, get_extent_t *get_extent)
2917 {
2918         int ret = 0;
2919         u64 off = start;
2920         u64 max = start + len;
2921         u32 flags = 0;
2922         u32 found_type;
2923         u64 last;
2924         u64 disko = 0;
2925         struct btrfs_key found_key;
2926         struct extent_map *em = NULL;
2927         struct extent_state *cached_state = NULL;
2928         struct btrfs_path *path;
2929         struct btrfs_file_extent_item *item;
2930         int end = 0;
2931         u64 em_start = 0, em_len = 0;
2932         unsigned long emflags;
2933         int hole = 0;
2934
2935         if (len == 0)
2936                 return -EINVAL;
2937
2938         path = btrfs_alloc_path();
2939         if (!path)
2940                 return -ENOMEM;
2941         path->leave_spinning = 1;
2942
2943         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2944                                        path, inode->i_ino, -1, 0);
2945         if (ret < 0) {
2946                 btrfs_free_path(path);
2947                 return ret;
2948         }
2949         WARN_ON(!ret);
2950         path->slots[0]--;
2951         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2952                               struct btrfs_file_extent_item);
2953         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2954         found_type = btrfs_key_type(&found_key);
2955
2956         /* No extents, just return */
2957         if (found_key.objectid != inode->i_ino ||
2958             found_type != BTRFS_EXTENT_DATA_KEY) {
2959                 btrfs_free_path(path);
2960                 return 0;
2961         }
2962         last = found_key.offset;
2963         btrfs_free_path(path);
2964
2965         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2966                          &cached_state, GFP_NOFS);
2967         em = get_extent(inode, NULL, 0, off, max - off, 0);
2968         if (!em)
2969                 goto out;
2970         if (IS_ERR(em)) {
2971                 ret = PTR_ERR(em);
2972                 goto out;
2973         }
2974
2975         while (!end) {
2976                 hole = 0;
2977                 off = em->start + em->len;
2978                 if (off >= max)
2979                         end = 1;
2980
2981                 if (em->block_start == EXTENT_MAP_HOLE) {
2982                         hole = 1;
2983                         goto next;
2984                 }
2985
2986                 em_start = em->start;
2987                 em_len = em->len;
2988
2989                 disko = 0;
2990                 flags = 0;
2991
2992                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2993                         end = 1;
2994                         flags |= FIEMAP_EXTENT_LAST;
2995                 } else if (em->block_start == EXTENT_MAP_INLINE) {
2996                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
2997                                   FIEMAP_EXTENT_NOT_ALIGNED);
2998                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2999                         flags |= (FIEMAP_EXTENT_DELALLOC |
3000                                   FIEMAP_EXTENT_UNKNOWN);
3001                 } else {
3002                         disko = em->block_start;
3003                 }
3004                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3005                         flags |= FIEMAP_EXTENT_ENCODED;
3006
3007 next:
3008                 emflags = em->flags;
3009                 free_extent_map(em);
3010                 em = NULL;
3011                 if (!end) {
3012                         em = get_extent(inode, NULL, 0, off, max - off, 0);
3013                         if (!em)
3014                                 goto out;
3015                         if (IS_ERR(em)) {
3016                                 ret = PTR_ERR(em);
3017                                 goto out;
3018                         }
3019                         emflags = em->flags;
3020                 }
3021
3022                 if (test_bit(EXTENT_FLAG_VACANCY, &emflags)) {
3023                         flags |= FIEMAP_EXTENT_LAST;
3024                         end = 1;
3025                 }
3026
3027                 if (em_start == last) {
3028                         flags |= FIEMAP_EXTENT_LAST;
3029                         end = 1;
3030                 }
3031
3032                 if (!hole) {
3033                         ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3034                                                 em_len, flags);
3035                         if (ret)
3036                                 goto out_free;
3037                 }
3038         }
3039 out_free:
3040         free_extent_map(em);
3041 out:
3042         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3043                              &cached_state, GFP_NOFS);
3044         return ret;
3045 }
3046
3047 static inline struct page *extent_buffer_page(struct extent_buffer *eb,