btrfs: cleanup for open-coded alignment
[linux-3.10.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 <linux/prefetch.h>
14 #include <linux/cleancache.h>
15 #include "extent_io.h"
16 #include "extent_map.h"
17 #include "compat.h"
18 #include "ctree.h"
19 #include "btrfs_inode.h"
20 #include "volumes.h"
21 #include "check-integrity.h"
22 #include "locking.h"
23 #include "rcu-string.h"
24
25 static struct kmem_cache *extent_state_cache;
26 static struct kmem_cache *extent_buffer_cache;
27
28 static LIST_HEAD(buffers);
29 static LIST_HEAD(states);
30
31 #define LEAK_DEBUG 0
32 #if LEAK_DEBUG
33 static DEFINE_SPINLOCK(leak_lock);
34 #endif
35
36 #define BUFFER_LRU_MAX 64
37
38 struct tree_entry {
39         u64 start;
40         u64 end;
41         struct rb_node rb_node;
42 };
43
44 struct extent_page_data {
45         struct bio *bio;
46         struct extent_io_tree *tree;
47         get_extent_t *get_extent;
48         unsigned long bio_flags;
49
50         /* tells writepage not to lock the state bits for this range
51          * it still does the unlocking
52          */
53         unsigned int extent_locked:1;
54
55         /* tells the submit_bio code to use a WRITE_SYNC */
56         unsigned int sync_io:1;
57 };
58
59 static noinline void flush_write_bio(void *data);
60 static inline struct btrfs_fs_info *
61 tree_fs_info(struct extent_io_tree *tree)
62 {
63         return btrfs_sb(tree->mapping->host->i_sb);
64 }
65
66 int __init extent_io_init(void)
67 {
68         extent_state_cache = kmem_cache_create("btrfs_extent_state",
69                         sizeof(struct extent_state), 0,
70                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
71         if (!extent_state_cache)
72                 return -ENOMEM;
73
74         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
75                         sizeof(struct extent_buffer), 0,
76                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
77         if (!extent_buffer_cache)
78                 goto free_state_cache;
79         return 0;
80
81 free_state_cache:
82         kmem_cache_destroy(extent_state_cache);
83         return -ENOMEM;
84 }
85
86 void extent_io_exit(void)
87 {
88         struct extent_state *state;
89         struct extent_buffer *eb;
90
91         while (!list_empty(&states)) {
92                 state = list_entry(states.next, struct extent_state, leak_list);
93                 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
94                        "state %lu in tree %p refs %d\n",
95                        (unsigned long long)state->start,
96                        (unsigned long long)state->end,
97                        state->state, state->tree, atomic_read(&state->refs));
98                 list_del(&state->leak_list);
99                 kmem_cache_free(extent_state_cache, state);
100
101         }
102
103         while (!list_empty(&buffers)) {
104                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
105                 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
106                        "refs %d\n", (unsigned long long)eb->start,
107                        eb->len, atomic_read(&eb->refs));
108                 list_del(&eb->leak_list);
109                 kmem_cache_free(extent_buffer_cache, eb);
110         }
111
112         /*
113          * Make sure all delayed rcu free are flushed before we
114          * destroy caches.
115          */
116         rcu_barrier();
117         if (extent_state_cache)
118                 kmem_cache_destroy(extent_state_cache);
119         if (extent_buffer_cache)
120                 kmem_cache_destroy(extent_buffer_cache);
121 }
122
123 void extent_io_tree_init(struct extent_io_tree *tree,
124                          struct address_space *mapping)
125 {
126         tree->state = RB_ROOT;
127         INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
128         tree->ops = NULL;
129         tree->dirty_bytes = 0;
130         spin_lock_init(&tree->lock);
131         spin_lock_init(&tree->buffer_lock);
132         tree->mapping = mapping;
133 }
134
135 static struct extent_state *alloc_extent_state(gfp_t mask)
136 {
137         struct extent_state *state;
138 #if LEAK_DEBUG
139         unsigned long flags;
140 #endif
141
142         state = kmem_cache_alloc(extent_state_cache, mask);
143         if (!state)
144                 return state;
145         state->state = 0;
146         state->private = 0;
147         state->tree = NULL;
148 #if LEAK_DEBUG
149         spin_lock_irqsave(&leak_lock, flags);
150         list_add(&state->leak_list, &states);
151         spin_unlock_irqrestore(&leak_lock, flags);
152 #endif
153         atomic_set(&state->refs, 1);
154         init_waitqueue_head(&state->wq);
155         trace_alloc_extent_state(state, mask, _RET_IP_);
156         return state;
157 }
158
159 void free_extent_state(struct extent_state *state)
160 {
161         if (!state)
162                 return;
163         if (atomic_dec_and_test(&state->refs)) {
164 #if LEAK_DEBUG
165                 unsigned long flags;
166 #endif
167                 WARN_ON(state->tree);
168 #if LEAK_DEBUG
169                 spin_lock_irqsave(&leak_lock, flags);
170                 list_del(&state->leak_list);
171                 spin_unlock_irqrestore(&leak_lock, flags);
172 #endif
173                 trace_free_extent_state(state, _RET_IP_);
174                 kmem_cache_free(extent_state_cache, state);
175         }
176 }
177
178 static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
179                                    struct rb_node *node)
180 {
181         struct rb_node **p = &root->rb_node;
182         struct rb_node *parent = NULL;
183         struct tree_entry *entry;
184
185         while (*p) {
186                 parent = *p;
187                 entry = rb_entry(parent, struct tree_entry, rb_node);
188
189                 if (offset < entry->start)
190                         p = &(*p)->rb_left;
191                 else if (offset > entry->end)
192                         p = &(*p)->rb_right;
193                 else
194                         return parent;
195         }
196
197         rb_link_node(node, parent, p);
198         rb_insert_color(node, root);
199         return NULL;
200 }
201
202 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
203                                      struct rb_node **prev_ret,
204                                      struct rb_node **next_ret)
205 {
206         struct rb_root *root = &tree->state;
207         struct rb_node *n = root->rb_node;
208         struct rb_node *prev = NULL;
209         struct rb_node *orig_prev = NULL;
210         struct tree_entry *entry;
211         struct tree_entry *prev_entry = NULL;
212
213         while (n) {
214                 entry = rb_entry(n, struct tree_entry, rb_node);
215                 prev = n;
216                 prev_entry = entry;
217
218                 if (offset < entry->start)
219                         n = n->rb_left;
220                 else if (offset > entry->end)
221                         n = n->rb_right;
222                 else
223                         return n;
224         }
225
226         if (prev_ret) {
227                 orig_prev = prev;
228                 while (prev && offset > prev_entry->end) {
229                         prev = rb_next(prev);
230                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
231                 }
232                 *prev_ret = prev;
233                 prev = orig_prev;
234         }
235
236         if (next_ret) {
237                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
238                 while (prev && offset < prev_entry->start) {
239                         prev = rb_prev(prev);
240                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
241                 }
242                 *next_ret = prev;
243         }
244         return NULL;
245 }
246
247 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
248                                           u64 offset)
249 {
250         struct rb_node *prev = NULL;
251         struct rb_node *ret;
252
253         ret = __etree_search(tree, offset, &prev, NULL);
254         if (!ret)
255                 return prev;
256         return ret;
257 }
258
259 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
260                      struct extent_state *other)
261 {
262         if (tree->ops && tree->ops->merge_extent_hook)
263                 tree->ops->merge_extent_hook(tree->mapping->host, new,
264                                              other);
265 }
266
267 /*
268  * utility function to look for merge candidates inside a given range.
269  * Any extents with matching state are merged together into a single
270  * extent in the tree.  Extents with EXTENT_IO in their state field
271  * are not merged because the end_io handlers need to be able to do
272  * operations on them without sleeping (or doing allocations/splits).
273  *
274  * This should be called with the tree lock held.
275  */
276 static void merge_state(struct extent_io_tree *tree,
277                         struct extent_state *state)
278 {
279         struct extent_state *other;
280         struct rb_node *other_node;
281
282         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
283                 return;
284
285         other_node = rb_prev(&state->rb_node);
286         if (other_node) {
287                 other = rb_entry(other_node, struct extent_state, rb_node);
288                 if (other->end == state->start - 1 &&
289                     other->state == state->state) {
290                         merge_cb(tree, state, other);
291                         state->start = other->start;
292                         other->tree = NULL;
293                         rb_erase(&other->rb_node, &tree->state);
294                         free_extent_state(other);
295                 }
296         }
297         other_node = rb_next(&state->rb_node);
298         if (other_node) {
299                 other = rb_entry(other_node, struct extent_state, rb_node);
300                 if (other->start == state->end + 1 &&
301                     other->state == state->state) {
302                         merge_cb(tree, state, other);
303                         state->end = other->end;
304                         other->tree = NULL;
305                         rb_erase(&other->rb_node, &tree->state);
306                         free_extent_state(other);
307                 }
308         }
309 }
310
311 static void set_state_cb(struct extent_io_tree *tree,
312                          struct extent_state *state, int *bits)
313 {
314         if (tree->ops && tree->ops->set_bit_hook)
315                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
316 }
317
318 static void clear_state_cb(struct extent_io_tree *tree,
319                            struct extent_state *state, int *bits)
320 {
321         if (tree->ops && tree->ops->clear_bit_hook)
322                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
323 }
324
325 static void set_state_bits(struct extent_io_tree *tree,
326                            struct extent_state *state, int *bits);
327
328 /*
329  * insert an extent_state struct into the tree.  'bits' are set on the
330  * struct before it is inserted.
331  *
332  * This may return -EEXIST if the extent is already there, in which case the
333  * state struct is freed.
334  *
335  * The tree lock is not taken internally.  This is a utility function and
336  * probably isn't what you want to call (see set/clear_extent_bit).
337  */
338 static int insert_state(struct extent_io_tree *tree,
339                         struct extent_state *state, u64 start, u64 end,
340                         int *bits)
341 {
342         struct rb_node *node;
343
344         if (end < start)
345                 WARN(1, KERN_ERR "btrfs end < start %llu %llu\n",
346                        (unsigned long long)end,
347                        (unsigned long long)start);
348         state->start = start;
349         state->end = end;
350
351         set_state_bits(tree, state, bits);
352
353         node = tree_insert(&tree->state, end, &state->rb_node);
354         if (node) {
355                 struct extent_state *found;
356                 found = rb_entry(node, struct extent_state, rb_node);
357                 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
358                        "%llu %llu\n", (unsigned long long)found->start,
359                        (unsigned long long)found->end,
360                        (unsigned long long)start, (unsigned long long)end);
361                 return -EEXIST;
362         }
363         state->tree = tree;
364         merge_state(tree, state);
365         return 0;
366 }
367
368 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
369                      u64 split)
370 {
371         if (tree->ops && tree->ops->split_extent_hook)
372                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
373 }
374
375 /*
376  * split a given extent state struct in two, inserting the preallocated
377  * struct 'prealloc' as the newly created second half.  'split' indicates an
378  * offset inside 'orig' where it should be split.
379  *
380  * Before calling,
381  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
382  * are two extent state structs in the tree:
383  * prealloc: [orig->start, split - 1]
384  * orig: [ split, orig->end ]
385  *
386  * The tree locks are not taken by this function. They need to be held
387  * by the caller.
388  */
389 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
390                        struct extent_state *prealloc, u64 split)
391 {
392         struct rb_node *node;
393
394         split_cb(tree, orig, split);
395
396         prealloc->start = orig->start;
397         prealloc->end = split - 1;
398         prealloc->state = orig->state;
399         orig->start = split;
400
401         node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
402         if (node) {
403                 free_extent_state(prealloc);
404                 return -EEXIST;
405         }
406         prealloc->tree = tree;
407         return 0;
408 }
409
410 static struct extent_state *next_state(struct extent_state *state)
411 {
412         struct rb_node *next = rb_next(&state->rb_node);
413         if (next)
414                 return rb_entry(next, struct extent_state, rb_node);
415         else
416                 return NULL;
417 }
418
419 /*
420  * utility function to clear some bits in an extent state struct.
421  * it will optionally wake up any one waiting on this state (wake == 1).
422  *
423  * If no bits are set on the state struct after clearing things, the
424  * struct is freed and removed from the tree
425  */
426 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
427                                             struct extent_state *state,
428                                             int *bits, int wake)
429 {
430         struct extent_state *next;
431         int bits_to_clear = *bits & ~EXTENT_CTLBITS;
432
433         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
434                 u64 range = state->end - state->start + 1;
435                 WARN_ON(range > tree->dirty_bytes);
436                 tree->dirty_bytes -= range;
437         }
438         clear_state_cb(tree, state, bits);
439         state->state &= ~bits_to_clear;
440         if (wake)
441                 wake_up(&state->wq);
442         if (state->state == 0) {
443                 next = next_state(state);
444                 if (state->tree) {
445                         rb_erase(&state->rb_node, &tree->state);
446                         state->tree = NULL;
447                         free_extent_state(state);
448                 } else {
449                         WARN_ON(1);
450                 }
451         } else {
452                 merge_state(tree, state);
453                 next = next_state(state);
454         }
455         return next;
456 }
457
458 static struct extent_state *
459 alloc_extent_state_atomic(struct extent_state *prealloc)
460 {
461         if (!prealloc)
462                 prealloc = alloc_extent_state(GFP_ATOMIC);
463
464         return prealloc;
465 }
466
467 void extent_io_tree_panic(struct extent_io_tree *tree, int err)
468 {
469         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
470                     "Extent tree was modified by another "
471                     "thread while locked.");
472 }
473
474 /*
475  * clear some bits on a range in the tree.  This may require splitting
476  * or inserting elements in the tree, so the gfp mask is used to
477  * indicate which allocations or sleeping are allowed.
478  *
479  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
480  * the given range from the tree regardless of state (ie for truncate).
481  *
482  * the range [start, end] is inclusive.
483  *
484  * This takes the tree lock, and returns 0 on success and < 0 on error.
485  */
486 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
487                      int bits, int wake, int delete,
488                      struct extent_state **cached_state,
489                      gfp_t mask)
490 {
491         struct extent_state *state;
492         struct extent_state *cached;
493         struct extent_state *prealloc = NULL;
494         struct rb_node *node;
495         u64 last_end;
496         int err;
497         int clear = 0;
498
499         if (delete)
500                 bits |= ~EXTENT_CTLBITS;
501         bits |= EXTENT_FIRST_DELALLOC;
502
503         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
504                 clear = 1;
505 again:
506         if (!prealloc && (mask & __GFP_WAIT)) {
507                 prealloc = alloc_extent_state(mask);
508                 if (!prealloc)
509                         return -ENOMEM;
510         }
511
512         spin_lock(&tree->lock);
513         if (cached_state) {
514                 cached = *cached_state;
515
516                 if (clear) {
517                         *cached_state = NULL;
518                         cached_state = NULL;
519                 }
520
521                 if (cached && cached->tree && cached->start <= start &&
522                     cached->end > start) {
523                         if (clear)
524                                 atomic_dec(&cached->refs);
525                         state = cached;
526                         goto hit_next;
527                 }
528                 if (clear)
529                         free_extent_state(cached);
530         }
531         /*
532          * this search will find the extents that end after
533          * our range starts
534          */
535         node = tree_search(tree, start);
536         if (!node)
537                 goto out;
538         state = rb_entry(node, struct extent_state, rb_node);
539 hit_next:
540         if (state->start > end)
541                 goto out;
542         WARN_ON(state->end < start);
543         last_end = state->end;
544
545         /* the state doesn't have the wanted bits, go ahead */
546         if (!(state->state & bits)) {
547                 state = next_state(state);
548                 goto next;
549         }
550
551         /*
552          *     | ---- desired range ---- |
553          *  | state | or
554          *  | ------------- state -------------- |
555          *
556          * We need to split the extent we found, and may flip
557          * bits on second half.
558          *
559          * If the extent we found extends past our range, we
560          * just split and search again.  It'll get split again
561          * the next time though.
562          *
563          * If the extent we found is inside our range, we clear
564          * the desired bit on it.
565          */
566
567         if (state->start < start) {
568                 prealloc = alloc_extent_state_atomic(prealloc);
569                 BUG_ON(!prealloc);
570                 err = split_state(tree, state, prealloc, start);
571                 if (err)
572                         extent_io_tree_panic(tree, err);
573
574                 prealloc = NULL;
575                 if (err)
576                         goto out;
577                 if (state->end <= end) {
578                         state = clear_state_bit(tree, state, &bits, wake);
579                         goto next;
580                 }
581                 goto search_again;
582         }
583         /*
584          * | ---- desired range ---- |
585          *                        | state |
586          * We need to split the extent, and clear the bit
587          * on the first half
588          */
589         if (state->start <= end && state->end > end) {
590                 prealloc = alloc_extent_state_atomic(prealloc);
591                 BUG_ON(!prealloc);
592                 err = split_state(tree, state, prealloc, end + 1);
593                 if (err)
594                         extent_io_tree_panic(tree, err);
595
596                 if (wake)
597                         wake_up(&state->wq);
598
599                 clear_state_bit(tree, prealloc, &bits, wake);
600
601                 prealloc = NULL;
602                 goto out;
603         }
604
605         state = clear_state_bit(tree, state, &bits, wake);
606 next:
607         if (last_end == (u64)-1)
608                 goto out;
609         start = last_end + 1;
610         if (start <= end && state && !need_resched())
611                 goto hit_next;
612         goto search_again;
613
614 out:
615         spin_unlock(&tree->lock);
616         if (prealloc)
617                 free_extent_state(prealloc);
618
619         return 0;
620
621 search_again:
622         if (start > end)
623                 goto out;
624         spin_unlock(&tree->lock);
625         if (mask & __GFP_WAIT)
626                 cond_resched();
627         goto again;
628 }
629
630 static void wait_on_state(struct extent_io_tree *tree,
631                           struct extent_state *state)
632                 __releases(tree->lock)
633                 __acquires(tree->lock)
634 {
635         DEFINE_WAIT(wait);
636         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
637         spin_unlock(&tree->lock);
638         schedule();
639         spin_lock(&tree->lock);
640         finish_wait(&state->wq, &wait);
641 }
642
643 /*
644  * waits for one or more bits to clear on a range in the state tree.
645  * The range [start, end] is inclusive.
646  * The tree lock is taken by this function
647  */
648 void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
649 {
650         struct extent_state *state;
651         struct rb_node *node;
652
653         spin_lock(&tree->lock);
654 again:
655         while (1) {
656                 /*
657                  * this search will find all the extents that end after
658                  * our range starts
659                  */
660                 node = tree_search(tree, start);
661                 if (!node)
662                         break;
663
664                 state = rb_entry(node, struct extent_state, rb_node);
665
666                 if (state->start > end)
667                         goto out;
668
669                 if (state->state & bits) {
670                         start = state->start;
671                         atomic_inc(&state->refs);
672                         wait_on_state(tree, state);
673                         free_extent_state(state);
674                         goto again;
675                 }
676                 start = state->end + 1;
677
678                 if (start > end)
679                         break;
680
681                 cond_resched_lock(&tree->lock);
682         }
683 out:
684         spin_unlock(&tree->lock);
685 }
686
687 static void set_state_bits(struct extent_io_tree *tree,
688                            struct extent_state *state,
689                            int *bits)
690 {
691         int bits_to_set = *bits & ~EXTENT_CTLBITS;
692
693         set_state_cb(tree, state, bits);
694         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
695                 u64 range = state->end - state->start + 1;
696                 tree->dirty_bytes += range;
697         }
698         state->state |= bits_to_set;
699 }
700
701 static void cache_state(struct extent_state *state,
702                         struct extent_state **cached_ptr)
703 {
704         if (cached_ptr && !(*cached_ptr)) {
705                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
706                         *cached_ptr = state;
707                         atomic_inc(&state->refs);
708                 }
709         }
710 }
711
712 static void uncache_state(struct extent_state **cached_ptr)
713 {
714         if (cached_ptr && (*cached_ptr)) {
715                 struct extent_state *state = *cached_ptr;
716                 *cached_ptr = NULL;
717                 free_extent_state(state);
718         }
719 }
720
721 /*
722  * set some bits on a range in the tree.  This may require allocations or
723  * sleeping, so the gfp mask is used to indicate what is allowed.
724  *
725  * If any of the exclusive bits are set, this will fail with -EEXIST if some
726  * part of the range already has the desired bits set.  The start of the
727  * existing range is returned in failed_start in this case.
728  *
729  * [start, end] is inclusive This takes the tree lock.
730  */
731
732 static int __must_check
733 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
734                  int bits, int exclusive_bits, u64 *failed_start,
735                  struct extent_state **cached_state, gfp_t mask)
736 {
737         struct extent_state *state;
738         struct extent_state *prealloc = NULL;
739         struct rb_node *node;
740         int err = 0;
741         u64 last_start;
742         u64 last_end;
743
744         bits |= EXTENT_FIRST_DELALLOC;
745 again:
746         if (!prealloc && (mask & __GFP_WAIT)) {
747                 prealloc = alloc_extent_state(mask);
748                 BUG_ON(!prealloc);
749         }
750
751         spin_lock(&tree->lock);
752         if (cached_state && *cached_state) {
753                 state = *cached_state;
754                 if (state->start <= start && state->end > start &&
755                     state->tree) {
756                         node = &state->rb_node;
757                         goto hit_next;
758                 }
759         }
760         /*
761          * this search will find all the extents that end after
762          * our range starts.
763          */
764         node = tree_search(tree, start);
765         if (!node) {
766                 prealloc = alloc_extent_state_atomic(prealloc);
767                 BUG_ON(!prealloc);
768                 err = insert_state(tree, prealloc, start, end, &bits);
769                 if (err)
770                         extent_io_tree_panic(tree, err);
771
772                 prealloc = NULL;
773                 goto out;
774         }
775         state = rb_entry(node, struct extent_state, rb_node);
776 hit_next:
777         last_start = state->start;
778         last_end = state->end;
779
780         /*
781          * | ---- desired range ---- |
782          * | state |
783          *
784          * Just lock what we found and keep going
785          */
786         if (state->start == start && state->end <= end) {
787                 if (state->state & exclusive_bits) {
788                         *failed_start = state->start;
789                         err = -EEXIST;
790                         goto out;
791                 }
792
793                 set_state_bits(tree, state, &bits);
794                 cache_state(state, cached_state);
795                 merge_state(tree, state);
796                 if (last_end == (u64)-1)
797                         goto out;
798                 start = last_end + 1;
799                 state = next_state(state);
800                 if (start < end && state && state->start == start &&
801                     !need_resched())
802                         goto hit_next;
803                 goto search_again;
804         }
805
806         /*
807          *     | ---- desired range ---- |
808          * | state |
809          *   or
810          * | ------------- state -------------- |
811          *
812          * We need to split the extent we found, and may flip bits on
813          * second half.
814          *
815          * If the extent we found extends past our
816          * range, we just split and search again.  It'll get split
817          * again the next time though.
818          *
819          * If the extent we found is inside our range, we set the
820          * desired bit on it.
821          */
822         if (state->start < start) {
823                 if (state->state & exclusive_bits) {
824                         *failed_start = start;
825                         err = -EEXIST;
826                         goto out;
827                 }
828
829                 prealloc = alloc_extent_state_atomic(prealloc);
830                 BUG_ON(!prealloc);
831                 err = split_state(tree, state, prealloc, start);
832                 if (err)
833                         extent_io_tree_panic(tree, err);
834
835                 prealloc = NULL;
836                 if (err)
837                         goto out;
838                 if (state->end <= end) {
839                         set_state_bits(tree, state, &bits);
840                         cache_state(state, cached_state);
841                         merge_state(tree, state);
842                         if (last_end == (u64)-1)
843                                 goto out;
844                         start = last_end + 1;
845                         state = next_state(state);
846                         if (start < end && state && state->start == start &&
847                             !need_resched())
848                                 goto hit_next;
849                 }
850                 goto search_again;
851         }
852         /*
853          * | ---- desired range ---- |
854          *     | state | or               | state |
855          *
856          * There's a hole, we need to insert something in it and
857          * ignore the extent we found.
858          */
859         if (state->start > start) {
860                 u64 this_end;
861                 if (end < last_start)
862                         this_end = end;
863                 else
864                         this_end = last_start - 1;
865
866                 prealloc = alloc_extent_state_atomic(prealloc);
867                 BUG_ON(!prealloc);
868
869                 /*
870                  * Avoid to free 'prealloc' if it can be merged with
871                  * the later extent.
872                  */
873                 err = insert_state(tree, prealloc, start, this_end,
874                                    &bits);
875                 if (err)
876                         extent_io_tree_panic(tree, err);
877
878                 cache_state(prealloc, cached_state);
879                 prealloc = NULL;
880                 start = this_end + 1;
881                 goto search_again;
882         }
883         /*
884          * | ---- desired range ---- |
885          *                        | state |
886          * We need to split the extent, and set the bit
887          * on the first half
888          */
889         if (state->start <= end && state->end > end) {
890                 if (state->state & exclusive_bits) {
891                         *failed_start = start;
892                         err = -EEXIST;
893                         goto out;
894                 }
895
896                 prealloc = alloc_extent_state_atomic(prealloc);
897                 BUG_ON(!prealloc);
898                 err = split_state(tree, state, prealloc, end + 1);
899                 if (err)
900                         extent_io_tree_panic(tree, err);
901
902                 set_state_bits(tree, prealloc, &bits);
903                 cache_state(prealloc, cached_state);
904                 merge_state(tree, prealloc);
905                 prealloc = NULL;
906                 goto out;
907         }
908
909         goto search_again;
910
911 out:
912         spin_unlock(&tree->lock);
913         if (prealloc)
914                 free_extent_state(prealloc);
915
916         return err;
917
918 search_again:
919         if (start > end)
920                 goto out;
921         spin_unlock(&tree->lock);
922         if (mask & __GFP_WAIT)
923                 cond_resched();
924         goto again;
925 }
926
927 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits,
928                    u64 *failed_start, struct extent_state **cached_state,
929                    gfp_t mask)
930 {
931         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
932                                 cached_state, mask);
933 }
934
935
936 /**
937  * convert_extent_bit - convert all bits in a given range from one bit to
938  *                      another
939  * @tree:       the io tree to search
940  * @start:      the start offset in bytes
941  * @end:        the end offset in bytes (inclusive)
942  * @bits:       the bits to set in this range
943  * @clear_bits: the bits to clear in this range
944  * @cached_state:       state that we're going to cache
945  * @mask:       the allocation mask
946  *
947  * This will go through and set bits for the given range.  If any states exist
948  * already in this range they are set with the given bit and cleared of the
949  * clear_bits.  This is only meant to be used by things that are mergeable, ie
950  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
951  * boundary bits like LOCK.
952  */
953 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
954                        int bits, int clear_bits,
955                        struct extent_state **cached_state, gfp_t mask)
956 {
957         struct extent_state *state;
958         struct extent_state *prealloc = NULL;
959         struct rb_node *node;
960         int err = 0;
961         u64 last_start;
962         u64 last_end;
963
964 again:
965         if (!prealloc && (mask & __GFP_WAIT)) {
966                 prealloc = alloc_extent_state(mask);
967                 if (!prealloc)
968                         return -ENOMEM;
969         }
970
971         spin_lock(&tree->lock);
972         if (cached_state && *cached_state) {
973                 state = *cached_state;
974                 if (state->start <= start && state->end > start &&
975                     state->tree) {
976                         node = &state->rb_node;
977                         goto hit_next;
978                 }
979         }
980
981         /*
982          * this search will find all the extents that end after
983          * our range starts.
984          */
985         node = tree_search(tree, start);
986         if (!node) {
987                 prealloc = alloc_extent_state_atomic(prealloc);
988                 if (!prealloc) {
989                         err = -ENOMEM;
990                         goto out;
991                 }
992                 err = insert_state(tree, prealloc, start, end, &bits);
993                 prealloc = NULL;
994                 if (err)
995                         extent_io_tree_panic(tree, err);
996                 goto out;
997         }
998         state = rb_entry(node, struct extent_state, rb_node);
999 hit_next:
1000         last_start = state->start;
1001         last_end = state->end;
1002
1003         /*
1004          * | ---- desired range ---- |
1005          * | state |
1006          *
1007          * Just lock what we found and keep going
1008          */
1009         if (state->start == start && state->end <= end) {
1010                 set_state_bits(tree, state, &bits);
1011                 cache_state(state, cached_state);
1012                 state = clear_state_bit(tree, state, &clear_bits, 0);
1013                 if (last_end == (u64)-1)
1014                         goto out;
1015                 start = last_end + 1;
1016                 if (start < end && state && state->start == start &&
1017                     !need_resched())
1018                         goto hit_next;
1019                 goto search_again;
1020         }
1021
1022         /*
1023          *     | ---- desired range ---- |
1024          * | state |
1025          *   or
1026          * | ------------- state -------------- |
1027          *
1028          * We need to split the extent we found, and may flip bits on
1029          * second half.
1030          *
1031          * If the extent we found extends past our
1032          * range, we just split and search again.  It'll get split
1033          * again the next time though.
1034          *
1035          * If the extent we found is inside our range, we set the
1036          * desired bit on it.
1037          */
1038         if (state->start < start) {
1039                 prealloc = alloc_extent_state_atomic(prealloc);
1040                 if (!prealloc) {
1041                         err = -ENOMEM;
1042                         goto out;
1043                 }
1044                 err = split_state(tree, state, prealloc, start);
1045                 if (err)
1046                         extent_io_tree_panic(tree, err);
1047                 prealloc = NULL;
1048                 if (err)
1049                         goto out;
1050                 if (state->end <= end) {
1051                         set_state_bits(tree, state, &bits);
1052                         cache_state(state, cached_state);
1053                         state = clear_state_bit(tree, state, &clear_bits, 0);
1054                         if (last_end == (u64)-1)
1055                                 goto out;
1056                         start = last_end + 1;
1057                         if (start < end && state && state->start == start &&
1058                             !need_resched())
1059                                 goto hit_next;
1060                 }
1061                 goto search_again;
1062         }
1063         /*
1064          * | ---- desired range ---- |
1065          *     | state | or               | state |
1066          *
1067          * There's a hole, we need to insert something in it and
1068          * ignore the extent we found.
1069          */
1070         if (state->start > start) {
1071                 u64 this_end;
1072                 if (end < last_start)
1073                         this_end = end;
1074                 else
1075                         this_end = last_start - 1;
1076
1077                 prealloc = alloc_extent_state_atomic(prealloc);
1078                 if (!prealloc) {
1079                         err = -ENOMEM;
1080                         goto out;
1081                 }
1082
1083                 /*
1084                  * Avoid to free 'prealloc' if it can be merged with
1085                  * the later extent.
1086                  */
1087                 err = insert_state(tree, prealloc, start, this_end,
1088                                    &bits);
1089                 if (err)
1090                         extent_io_tree_panic(tree, err);
1091                 cache_state(prealloc, cached_state);
1092                 prealloc = NULL;
1093                 start = this_end + 1;
1094                 goto search_again;
1095         }
1096         /*
1097          * | ---- desired range ---- |
1098          *                        | state |
1099          * We need to split the extent, and set the bit
1100          * on the first half
1101          */
1102         if (state->start <= end && state->end > end) {
1103                 prealloc = alloc_extent_state_atomic(prealloc);
1104                 if (!prealloc) {
1105                         err = -ENOMEM;
1106                         goto out;
1107                 }
1108
1109                 err = split_state(tree, state, prealloc, end + 1);
1110                 if (err)
1111                         extent_io_tree_panic(tree, err);
1112
1113                 set_state_bits(tree, prealloc, &bits);
1114                 cache_state(prealloc, cached_state);
1115                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1116                 prealloc = NULL;
1117                 goto out;
1118         }
1119
1120         goto search_again;
1121
1122 out:
1123         spin_unlock(&tree->lock);
1124         if (prealloc)
1125                 free_extent_state(prealloc);
1126
1127         return err;
1128
1129 search_again:
1130         if (start > end)
1131                 goto out;
1132         spin_unlock(&tree->lock);
1133         if (mask & __GFP_WAIT)
1134                 cond_resched();
1135         goto again;
1136 }
1137
1138 /* wrappers around set/clear extent bit */
1139 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1140                      gfp_t mask)
1141 {
1142         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1143                               NULL, mask);
1144 }
1145
1146 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1147                     int bits, gfp_t mask)
1148 {
1149         return set_extent_bit(tree, start, end, bits, NULL,
1150                               NULL, mask);
1151 }
1152
1153 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1154                       int bits, gfp_t mask)
1155 {
1156         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1157 }
1158
1159 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1160                         struct extent_state **cached_state, gfp_t mask)
1161 {
1162         return set_extent_bit(tree, start, end,
1163                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1164                               NULL, cached_state, mask);
1165 }
1166
1167 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1168                       struct extent_state **cached_state, gfp_t mask)
1169 {
1170         return set_extent_bit(tree, start, end,
1171                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1172                               NULL, cached_state, mask);
1173 }
1174
1175 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1176                        gfp_t mask)
1177 {
1178         return clear_extent_bit(tree, start, end,
1179                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1180                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1181 }
1182
1183 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1184                      gfp_t mask)
1185 {
1186         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1187                               NULL, mask);
1188 }
1189
1190 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1191                         struct extent_state **cached_state, gfp_t mask)
1192 {
1193         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
1194                               cached_state, mask);
1195 }
1196
1197 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1198                           struct extent_state **cached_state, gfp_t mask)
1199 {
1200         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1201                                 cached_state, mask);
1202 }
1203
1204 /*
1205  * either insert or lock state struct between start and end use mask to tell
1206  * us if waiting is desired.
1207  */
1208 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1209                      int bits, struct extent_state **cached_state)
1210 {
1211         int err;
1212         u64 failed_start;
1213         while (1) {
1214                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1215                                        EXTENT_LOCKED, &failed_start,
1216                                        cached_state, GFP_NOFS);
1217                 if (err == -EEXIST) {
1218                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1219                         start = failed_start;
1220                 } else
1221                         break;
1222                 WARN_ON(start > end);
1223         }
1224         return err;
1225 }
1226
1227 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1228 {
1229         return lock_extent_bits(tree, start, end, 0, NULL);
1230 }
1231
1232 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1233 {
1234         int err;
1235         u64 failed_start;
1236
1237         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1238                                &failed_start, NULL, GFP_NOFS);
1239         if (err == -EEXIST) {
1240                 if (failed_start > start)
1241                         clear_extent_bit(tree, start, failed_start - 1,
1242                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1243                 return 0;
1244         }
1245         return 1;
1246 }
1247
1248 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1249                          struct extent_state **cached, gfp_t mask)
1250 {
1251         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1252                                 mask);
1253 }
1254
1255 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1256 {
1257         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1258                                 GFP_NOFS);
1259 }
1260
1261 /*
1262  * helper function to set both pages and extents in the tree writeback
1263  */
1264 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1265 {
1266         unsigned long index = start >> PAGE_CACHE_SHIFT;
1267         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1268         struct page *page;
1269
1270         while (index <= end_index) {
1271                 page = find_get_page(tree->mapping, index);
1272                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1273                 set_page_writeback(page);
1274                 page_cache_release(page);
1275                 index++;
1276         }
1277         return 0;
1278 }
1279
1280 /* find the first state struct with 'bits' set after 'start', and
1281  * return it.  tree->lock must be held.  NULL will returned if
1282  * nothing was found after 'start'
1283  */
1284 struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1285                                                  u64 start, int bits)
1286 {
1287         struct rb_node *node;
1288         struct extent_state *state;
1289
1290         /*
1291          * this search will find all the extents that end after
1292          * our range starts.
1293          */
1294         node = tree_search(tree, start);
1295         if (!node)
1296                 goto out;
1297
1298         while (1) {
1299                 state = rb_entry(node, struct extent_state, rb_node);
1300                 if (state->end >= start && (state->state & bits))
1301                         return state;
1302
1303                 node = rb_next(node);
1304                 if (!node)
1305                         break;
1306         }
1307 out:
1308         return NULL;
1309 }
1310
1311 /*
1312  * find the first offset in the io tree with 'bits' set. zero is
1313  * returned if we find something, and *start_ret and *end_ret are
1314  * set to reflect the state struct that was found.
1315  *
1316  * If nothing was found, 1 is returned. If found something, return 0.
1317  */
1318 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1319                           u64 *start_ret, u64 *end_ret, int bits,
1320                           struct extent_state **cached_state)
1321 {
1322         struct extent_state *state;
1323         struct rb_node *n;
1324         int ret = 1;
1325
1326         spin_lock(&tree->lock);
1327         if (cached_state && *cached_state) {
1328                 state = *cached_state;
1329                 if (state->end == start - 1 && state->tree) {
1330                         n = rb_next(&state->rb_node);
1331                         while (n) {
1332                                 state = rb_entry(n, struct extent_state,
1333                                                  rb_node);
1334                                 if (state->state & bits)
1335                                         goto got_it;
1336                                 n = rb_next(n);
1337                         }
1338                         free_extent_state(*cached_state);
1339                         *cached_state = NULL;
1340                         goto out;
1341                 }
1342                 free_extent_state(*cached_state);
1343                 *cached_state = NULL;
1344         }
1345
1346         state = find_first_extent_bit_state(tree, start, bits);
1347 got_it:
1348         if (state) {
1349                 cache_state(state, cached_state);
1350                 *start_ret = state->start;
1351                 *end_ret = state->end;
1352                 ret = 0;
1353         }
1354 out:
1355         spin_unlock(&tree->lock);
1356         return ret;
1357 }
1358
1359 /*
1360  * find a contiguous range of bytes in the file marked as delalloc, not
1361  * more than 'max_bytes'.  start and end are used to return the range,
1362  *
1363  * 1 is returned if we find something, 0 if nothing was in the tree
1364  */
1365 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1366                                         u64 *start, u64 *end, u64 max_bytes,
1367                                         struct extent_state **cached_state)
1368 {
1369         struct rb_node *node;
1370         struct extent_state *state;
1371         u64 cur_start = *start;
1372         u64 found = 0;
1373         u64 total_bytes = 0;
1374
1375         spin_lock(&tree->lock);
1376
1377         /*
1378          * this search will find all the extents that end after
1379          * our range starts.
1380          */
1381         node = tree_search(tree, cur_start);
1382         if (!node) {
1383                 if (!found)
1384                         *end = (u64)-1;
1385                 goto out;
1386         }
1387
1388         while (1) {
1389                 state = rb_entry(node, struct extent_state, rb_node);
1390                 if (found && (state->start != cur_start ||
1391                               (state->state & EXTENT_BOUNDARY))) {
1392                         goto out;
1393                 }
1394                 if (!(state->state & EXTENT_DELALLOC)) {
1395                         if (!found)
1396                                 *end = state->end;
1397                         goto out;
1398                 }
1399                 if (!found) {
1400                         *start = state->start;
1401                         *cached_state = state;
1402                         atomic_inc(&state->refs);
1403                 }
1404                 found++;
1405                 *end = state->end;
1406                 cur_start = state->end + 1;
1407                 node = rb_next(node);
1408                 if (!node)
1409                         break;
1410                 total_bytes += state->end - state->start + 1;
1411                 if (total_bytes >= max_bytes)
1412                         break;
1413         }
1414 out:
1415         spin_unlock(&tree->lock);
1416         return found;
1417 }
1418
1419 static noinline void __unlock_for_delalloc(struct inode *inode,
1420                                            struct page *locked_page,
1421                                            u64 start, u64 end)
1422 {
1423         int ret;
1424         struct page *pages[16];
1425         unsigned long index = start >> PAGE_CACHE_SHIFT;
1426         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1427         unsigned long nr_pages = end_index - index + 1;
1428         int i;
1429
1430         if (index == locked_page->index && end_index == index)
1431                 return;
1432
1433         while (nr_pages > 0) {
1434                 ret = find_get_pages_contig(inode->i_mapping, index,
1435                                      min_t(unsigned long, nr_pages,
1436                                      ARRAY_SIZE(pages)), pages);
1437                 for (i = 0; i < ret; i++) {
1438                         if (pages[i] != locked_page)
1439                                 unlock_page(pages[i]);
1440                         page_cache_release(pages[i]);
1441                 }
1442                 nr_pages -= ret;
1443                 index += ret;
1444                 cond_resched();
1445         }
1446 }
1447
1448 static noinline int lock_delalloc_pages(struct inode *inode,
1449                                         struct page *locked_page,
1450                                         u64 delalloc_start,
1451                                         u64 delalloc_end)
1452 {
1453         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1454         unsigned long start_index = index;
1455         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1456         unsigned long pages_locked = 0;
1457         struct page *pages[16];
1458         unsigned long nrpages;
1459         int ret;
1460         int i;
1461
1462         /* the caller is responsible for locking the start index */
1463         if (index == locked_page->index && index == end_index)
1464                 return 0;
1465
1466         /* skip the page at the start index */
1467         nrpages = end_index - index + 1;
1468         while (nrpages > 0) {
1469                 ret = find_get_pages_contig(inode->i_mapping, index,
1470                                      min_t(unsigned long,
1471                                      nrpages, ARRAY_SIZE(pages)), pages);
1472                 if (ret == 0) {
1473                         ret = -EAGAIN;
1474                         goto done;
1475                 }
1476                 /* now we have an array of pages, lock them all */
1477                 for (i = 0; i < ret; i++) {
1478                         /*
1479                          * the caller is taking responsibility for
1480                          * locked_page
1481                          */
1482                         if (pages[i] != locked_page) {
1483                                 lock_page(pages[i]);
1484                                 if (!PageDirty(pages[i]) ||
1485                                     pages[i]->mapping != inode->i_mapping) {
1486                                         ret = -EAGAIN;
1487                                         unlock_page(pages[i]);
1488                                         page_cache_release(pages[i]);
1489                                         goto done;
1490                                 }
1491                         }
1492                         page_cache_release(pages[i]);
1493                         pages_locked++;
1494                 }
1495                 nrpages -= ret;
1496                 index += ret;
1497                 cond_resched();
1498         }
1499         ret = 0;
1500 done:
1501         if (ret && pages_locked) {
1502                 __unlock_for_delalloc(inode, locked_page,
1503                               delalloc_start,
1504                               ((u64)(start_index + pages_locked - 1)) <<
1505                               PAGE_CACHE_SHIFT);
1506         }
1507         return ret;
1508 }
1509
1510 /*
1511  * find a contiguous range of bytes in the file marked as delalloc, not
1512  * more than 'max_bytes'.  start and end are used to return the range,
1513  *
1514  * 1 is returned if we find something, 0 if nothing was in the tree
1515  */
1516 static noinline u64 find_lock_delalloc_range(struct inode *inode,
1517                                              struct extent_io_tree *tree,
1518                                              struct page *locked_page,
1519                                              u64 *start, u64 *end,
1520                                              u64 max_bytes)
1521 {
1522         u64 delalloc_start;
1523         u64 delalloc_end;
1524         u64 found;
1525         struct extent_state *cached_state = NULL;
1526         int ret;
1527         int loops = 0;
1528
1529 again:
1530         /* step one, find a bunch of delalloc bytes starting at start */
1531         delalloc_start = *start;
1532         delalloc_end = 0;
1533         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1534                                     max_bytes, &cached_state);
1535         if (!found || delalloc_end <= *start) {
1536                 *start = delalloc_start;
1537                 *end = delalloc_end;
1538                 free_extent_state(cached_state);
1539                 return found;
1540         }
1541
1542         /*
1543          * start comes from the offset of locked_page.  We have to lock
1544          * pages in order, so we can't process delalloc bytes before
1545          * locked_page
1546          */
1547         if (delalloc_start < *start)
1548                 delalloc_start = *start;
1549
1550         /*
1551          * make sure to limit the number of pages we try to lock down
1552          * if we're looping.
1553          */
1554         if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1555                 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1556
1557         /* step two, lock all the pages after the page that has start */
1558         ret = lock_delalloc_pages(inode, locked_page,
1559                                   delalloc_start, delalloc_end);
1560         if (ret == -EAGAIN) {
1561                 /* some of the pages are gone, lets avoid looping by
1562                  * shortening the size of the delalloc range we're searching
1563                  */
1564                 free_extent_state(cached_state);
1565                 if (!loops) {
1566                         unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1567                         max_bytes = PAGE_CACHE_SIZE - offset;
1568                         loops = 1;
1569                         goto again;
1570                 } else {
1571                         found = 0;
1572                         goto out_failed;
1573                 }
1574         }
1575         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1576
1577         /* step three, lock the state bits for the whole range */
1578         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1579
1580         /* then test to make sure it is all still delalloc */
1581         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1582                              EXTENT_DELALLOC, 1, cached_state);
1583         if (!ret) {
1584                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1585                                      &cached_state, GFP_NOFS);
1586                 __unlock_for_delalloc(inode, locked_page,
1587                               delalloc_start, delalloc_end);
1588                 cond_resched();
1589                 goto again;
1590         }
1591         free_extent_state(cached_state);
1592         *start = delalloc_start;
1593         *end = delalloc_end;
1594 out_failed:
1595         return found;
1596 }
1597
1598 int extent_clear_unlock_delalloc(struct inode *inode,
1599                                 struct extent_io_tree *tree,
1600                                 u64 start, u64 end, struct page *locked_page,
1601                                 unsigned long op)
1602 {
1603         int ret;
1604         struct page *pages[16];
1605         unsigned long index = start >> PAGE_CACHE_SHIFT;
1606         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1607         unsigned long nr_pages = end_index - index + 1;
1608         int i;
1609         int clear_bits = 0;
1610
1611         if (op & EXTENT_CLEAR_UNLOCK)
1612                 clear_bits |= EXTENT_LOCKED;
1613         if (op & EXTENT_CLEAR_DIRTY)
1614                 clear_bits |= EXTENT_DIRTY;
1615
1616         if (op & EXTENT_CLEAR_DELALLOC)
1617                 clear_bits |= EXTENT_DELALLOC;
1618
1619         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1620         if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1621                     EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1622                     EXTENT_SET_PRIVATE2)))
1623                 return 0;
1624
1625         while (nr_pages > 0) {
1626                 ret = find_get_pages_contig(inode->i_mapping, index,
1627                                      min_t(unsigned long,
1628                                      nr_pages, ARRAY_SIZE(pages)), pages);
1629                 for (i = 0; i < ret; i++) {
1630
1631                         if (op & EXTENT_SET_PRIVATE2)
1632                                 SetPagePrivate2(pages[i]);
1633
1634                         if (pages[i] == locked_page) {
1635                                 page_cache_release(pages[i]);
1636                                 continue;
1637                         }
1638                         if (op & EXTENT_CLEAR_DIRTY)
1639                                 clear_page_dirty_for_io(pages[i]);
1640                         if (op & EXTENT_SET_WRITEBACK)
1641                                 set_page_writeback(pages[i]);
1642                         if (op & EXTENT_END_WRITEBACK)
1643                                 end_page_writeback(pages[i]);
1644                         if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1645                                 unlock_page(pages[i]);
1646                         page_cache_release(pages[i]);
1647                 }
1648                 nr_pages -= ret;
1649                 index += ret;
1650                 cond_resched();
1651         }
1652         return 0;
1653 }
1654
1655 /*
1656  * count the number of bytes in the tree that have a given bit(s)
1657  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1658  * cached.  The total number found is returned.
1659  */
1660 u64 count_range_bits(struct extent_io_tree *tree,
1661                      u64 *start, u64 search_end, u64 max_bytes,
1662                      unsigned long bits, int contig)
1663 {
1664         struct rb_node *node;
1665         struct extent_state *state;
1666         u64 cur_start = *start;
1667         u64 total_bytes = 0;
1668         u64 last = 0;
1669         int found = 0;
1670
1671         if (search_end <= cur_start) {
1672                 WARN_ON(1);
1673                 return 0;
1674         }
1675
1676         spin_lock(&tree->lock);
1677         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1678                 total_bytes = tree->dirty_bytes;
1679                 goto out;
1680         }
1681         /*
1682          * this search will find all the extents that end after
1683          * our range starts.
1684          */
1685         node = tree_search(tree, cur_start);
1686         if (!node)
1687                 goto out;
1688
1689         while (1) {
1690                 state = rb_entry(node, struct extent_state, rb_node);
1691                 if (state->start > search_end)
1692                         break;
1693                 if (contig && found && state->start > last + 1)
1694                         break;
1695                 if (state->end >= cur_start && (state->state & bits) == bits) {
1696                         total_bytes += min(search_end, state->end) + 1 -
1697                                        max(cur_start, state->start);
1698                         if (total_bytes >= max_bytes)
1699                                 break;
1700                         if (!found) {
1701                                 *start = max(cur_start, state->start);
1702                                 found = 1;
1703                         }
1704                         last = state->end;
1705                 } else if (contig && found) {
1706                         break;
1707                 }
1708                 node = rb_next(node);
1709                 if (!node)
1710                         break;
1711         }
1712 out:
1713         spin_unlock(&tree->lock);
1714         return total_bytes;
1715 }
1716
1717 /*
1718  * set the private field for a given byte offset in the tree.  If there isn't
1719  * an extent_state there already, this does nothing.
1720  */
1721 int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1722 {
1723         struct rb_node *node;
1724         struct extent_state *state;
1725         int ret = 0;
1726
1727         spin_lock(&tree->lock);
1728         /*
1729          * this search will find all the extents that end after
1730          * our range starts.
1731          */
1732         node = tree_search(tree, start);
1733         if (!node) {
1734                 ret = -ENOENT;
1735                 goto out;
1736         }
1737         state = rb_entry(node, struct extent_state, rb_node);
1738         if (state->start != start) {
1739                 ret = -ENOENT;
1740                 goto out;
1741         }
1742         state->private = private;
1743 out:
1744         spin_unlock(&tree->lock);
1745         return ret;
1746 }
1747
1748 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1749 {
1750         struct rb_node *node;
1751         struct extent_state *state;
1752         int ret = 0;
1753
1754         spin_lock(&tree->lock);
1755         /*
1756          * this search will find all the extents that end after
1757          * our range starts.
1758          */
1759         node = tree_search(tree, start);
1760         if (!node) {
1761                 ret = -ENOENT;
1762                 goto out;
1763         }
1764         state = rb_entry(node, struct extent_state, rb_node);
1765         if (state->start != start) {
1766                 ret = -ENOENT;
1767                 goto out;
1768         }
1769         *private = state->private;
1770 out:
1771         spin_unlock(&tree->lock);
1772         return ret;
1773 }
1774
1775 /*
1776  * searches a range in the state tree for a given mask.
1777  * If 'filled' == 1, this returns 1 only if every extent in the tree
1778  * has the bits set.  Otherwise, 1 is returned if any bit in the
1779  * range is found set.
1780  */
1781 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1782                    int bits, int filled, struct extent_state *cached)
1783 {
1784         struct extent_state *state = NULL;
1785         struct rb_node *node;
1786         int bitset = 0;
1787
1788         spin_lock(&tree->lock);
1789         if (cached && cached->tree && cached->start <= start &&
1790             cached->end > start)
1791                 node = &cached->rb_node;
1792         else
1793                 node = tree_search(tree, start);
1794         while (node && start <= end) {
1795                 state = rb_entry(node, struct extent_state, rb_node);
1796
1797                 if (filled && state->start > start) {
1798                         bitset = 0;
1799                         break;
1800                 }
1801
1802                 if (state->start > end)
1803                         break;
1804
1805                 if (state->state & bits) {
1806                         bitset = 1;
1807                         if (!filled)
1808                                 break;
1809                 } else if (filled) {
1810                         bitset = 0;
1811                         break;
1812                 }
1813
1814                 if (state->end == (u64)-1)
1815                         break;
1816
1817                 start = state->end + 1;
1818                 if (start > end)
1819                         break;
1820                 node = rb_next(node);
1821                 if (!node) {
1822                         if (filled)
1823                                 bitset = 0;
1824                         break;
1825                 }
1826         }
1827         spin_unlock(&tree->lock);
1828         return bitset;
1829 }
1830
1831 /*
1832  * helper function to set a given page up to date if all the
1833  * extents in the tree for that page are up to date
1834  */
1835 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1836 {
1837         u64 start = page_offset(page);
1838         u64 end = start + PAGE_CACHE_SIZE - 1;
1839         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1840                 SetPageUptodate(page);
1841 }
1842
1843 /*
1844  * helper function to unlock a page if all the extents in the tree
1845  * for that page are unlocked
1846  */
1847 static void check_page_locked(struct extent_io_tree *tree, struct page *page)
1848 {
1849         u64 start = page_offset(page);
1850         u64 end = start + PAGE_CACHE_SIZE - 1;
1851         if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1852                 unlock_page(page);
1853 }
1854
1855 /*
1856  * helper function to end page writeback if all the extents
1857  * in the tree for that page are done with writeback
1858  */
1859 static void check_page_writeback(struct extent_io_tree *tree,
1860                                  struct page *page)
1861 {
1862         end_page_writeback(page);
1863 }
1864
1865 /*
1866  * When IO fails, either with EIO or csum verification fails, we
1867  * try other mirrors that might have a good copy of the data.  This
1868  * io_failure_record is used to record state as we go through all the
1869  * mirrors.  If another mirror has good data, the page is set up to date
1870  * and things continue.  If a good mirror can't be found, the original
1871  * bio end_io callback is called to indicate things have failed.
1872  */
1873 struct io_failure_record {
1874         struct page *page;
1875         u64 start;
1876         u64 len;
1877         u64 logical;
1878         unsigned long bio_flags;
1879         int this_mirror;
1880         int failed_mirror;
1881         int in_validation;
1882 };
1883
1884 static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1885                                 int did_repair)
1886 {
1887         int ret;
1888         int err = 0;
1889         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1890
1891         set_state_private(failure_tree, rec->start, 0);
1892         ret = clear_extent_bits(failure_tree, rec->start,
1893                                 rec->start + rec->len - 1,
1894                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1895         if (ret)
1896                 err = ret;
1897
1898         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1899                                 rec->start + rec->len - 1,
1900                                 EXTENT_DAMAGED, GFP_NOFS);
1901         if (ret && !err)
1902                 err = ret;
1903
1904         kfree(rec);
1905         return err;
1906 }
1907
1908 static void repair_io_failure_callback(struct bio *bio, int err)
1909 {
1910         complete(bio->bi_private);
1911 }
1912
1913 /*
1914  * this bypasses the standard btrfs submit functions deliberately, as
1915  * the standard behavior is to write all copies in a raid setup. here we only
1916  * want to write the one bad copy. so we do the mapping for ourselves and issue
1917  * submit_bio directly.
1918  * to avoid any synchronization issues, wait for the data after writing, which
1919  * actually prevents the read that triggered the error from finishing.
1920  * currently, there can be no more than two copies of every data bit. thus,
1921  * exactly one rewrite is required.
1922  */
1923 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
1924                         u64 length, u64 logical, struct page *page,
1925                         int mirror_num)
1926 {
1927         struct bio *bio;
1928         struct btrfs_device *dev;
1929         DECLARE_COMPLETION_ONSTACK(compl);
1930         u64 map_length = 0;
1931         u64 sector;
1932         struct btrfs_bio *bbio = NULL;
1933         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
1934         int ret;
1935
1936         BUG_ON(!mirror_num);
1937
1938         /* we can't repair anything in raid56 yet */
1939         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
1940                 return 0;
1941
1942         bio = bio_alloc(GFP_NOFS, 1);
1943         if (!bio)
1944                 return -EIO;
1945         bio->bi_private = &compl;
1946         bio->bi_end_io = repair_io_failure_callback;
1947         bio->bi_size = 0;
1948         map_length = length;
1949
1950         ret = btrfs_map_block(fs_info, WRITE, logical,
1951                               &map_length, &bbio, mirror_num);
1952         if (ret) {
1953                 bio_put(bio);
1954                 return -EIO;
1955         }
1956         BUG_ON(mirror_num != bbio->mirror_num);
1957         sector = bbio->stripes[mirror_num-1].physical >> 9;
1958         bio->bi_sector = sector;
1959         dev = bbio->stripes[mirror_num-1].dev;
1960         kfree(bbio);
1961         if (!dev || !dev->bdev || !dev->writeable) {
1962                 bio_put(bio);
1963                 return -EIO;
1964         }
1965         bio->bi_bdev = dev->bdev;
1966         bio_add_page(bio, page, length, start - page_offset(page));
1967         btrfsic_submit_bio(WRITE_SYNC, bio);
1968         wait_for_completion(&compl);
1969
1970         if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
1971                 /* try to remap that extent elsewhere? */
1972                 bio_put(bio);
1973                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
1974                 return -EIO;
1975         }
1976
1977         printk_ratelimited_in_rcu(KERN_INFO "btrfs read error corrected: ino %lu off %llu "
1978                       "(dev %s sector %llu)\n", page->mapping->host->i_ino,
1979                       start, rcu_str_deref(dev->name), sector);
1980
1981         bio_put(bio);
1982         return 0;
1983 }
1984
1985 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
1986                          int mirror_num)
1987 {
1988         u64 start = eb->start;
1989         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
1990         int ret = 0;
1991
1992         for (i = 0; i < num_pages; i++) {
1993                 struct page *p = extent_buffer_page(eb, i);
1994                 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
1995                                         start, p, mirror_num);
1996                 if (ret)
1997                         break;
1998                 start += PAGE_CACHE_SIZE;
1999         }
2000
2001         return ret;
2002 }
2003
2004 /*
2005  * each time an IO finishes, we do a fast check in the IO failure tree
2006  * to see if we need to process or clean up an io_failure_record
2007  */
2008 static int clean_io_failure(u64 start, struct page *page)
2009 {
2010         u64 private;
2011         u64 private_failure;
2012         struct io_failure_record *failrec;
2013         struct btrfs_fs_info *fs_info;
2014         struct extent_state *state;
2015         int num_copies;
2016         int did_repair = 0;
2017         int ret;
2018         struct inode *inode = page->mapping->host;
2019
2020         private = 0;
2021         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2022                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2023         if (!ret)
2024                 return 0;
2025
2026         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2027                                 &private_failure);
2028         if (ret)
2029                 return 0;
2030
2031         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2032         BUG_ON(!failrec->this_mirror);
2033
2034         if (failrec->in_validation) {
2035                 /* there was no real error, just free the record */
2036                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2037                          failrec->start);
2038                 did_repair = 1;
2039                 goto out;
2040         }
2041
2042         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2043         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2044                                             failrec->start,
2045                                             EXTENT_LOCKED);
2046         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2047
2048         if (state && state->start == failrec->start) {
2049                 fs_info = BTRFS_I(inode)->root->fs_info;
2050                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2051                                               failrec->len);
2052                 if (num_copies > 1)  {
2053                         ret = repair_io_failure(fs_info, start, failrec->len,
2054                                                 failrec->logical, page,
2055                                                 failrec->failed_mirror);
2056                         did_repair = !ret;
2057                 }
2058                 ret = 0;
2059         }
2060
2061 out:
2062         if (!ret)
2063                 ret = free_io_failure(inode, failrec, did_repair);
2064
2065         return ret;
2066 }
2067
2068 /*
2069  * this is a generic handler for readpage errors (default
2070  * readpage_io_failed_hook). if other copies exist, read those and write back
2071  * good data to the failed position. does not investigate in remapping the
2072  * failed extent elsewhere, hoping the device will be smart enough to do this as
2073  * needed
2074  */
2075
2076 static int bio_readpage_error(struct bio *failed_bio, struct page *page,
2077                                 u64 start, u64 end, int failed_mirror,
2078                                 struct extent_state *state)
2079 {
2080         struct io_failure_record *failrec = NULL;
2081         u64 private;
2082         struct extent_map *em;
2083         struct inode *inode = page->mapping->host;
2084         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2085         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2086         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2087         struct bio *bio;
2088         int num_copies;
2089         int ret;
2090         int read_mode;
2091         u64 logical;
2092
2093         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2094
2095         ret = get_state_private(failure_tree, start, &private);
2096         if (ret) {
2097                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2098                 if (!failrec)
2099                         return -ENOMEM;
2100                 failrec->start = start;
2101                 failrec->len = end - start + 1;
2102                 failrec->this_mirror = 0;
2103                 failrec->bio_flags = 0;
2104                 failrec->in_validation = 0;
2105
2106                 read_lock(&em_tree->lock);
2107                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2108                 if (!em) {
2109                         read_unlock(&em_tree->lock);
2110                         kfree(failrec);
2111                         return -EIO;
2112                 }
2113
2114                 if (em->start > start || em->start + em->len < start) {
2115                         free_extent_map(em);
2116                         em = NULL;
2117                 }
2118                 read_unlock(&em_tree->lock);
2119
2120                 if (!em) {
2121                         kfree(failrec);
2122                         return -EIO;
2123                 }
2124                 logical = start - em->start;
2125                 logical = em->block_start + logical;
2126                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2127                         logical = em->block_start;
2128                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2129                         extent_set_compress_type(&failrec->bio_flags,
2130                                                  em->compress_type);
2131                 }
2132                 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2133                          "len=%llu\n", logical, start, failrec->len);
2134                 failrec->logical = logical;
2135                 free_extent_map(em);
2136
2137                 /* set the bits in the private failure tree */
2138                 ret = set_extent_bits(failure_tree, start, end,
2139                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2140                 if (ret >= 0)
2141                         ret = set_state_private(failure_tree, start,
2142                                                 (u64)(unsigned long)failrec);
2143                 /* set the bits in the inode's tree */
2144                 if (ret >= 0)
2145                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2146                                                 GFP_NOFS);
2147                 if (ret < 0) {
2148                         kfree(failrec);
2149                         return ret;
2150                 }
2151         } else {
2152                 failrec = (struct io_failure_record *)(unsigned long)private;
2153                 pr_debug("bio_readpage_error: (found) logical=%llu, "
2154                          "start=%llu, len=%llu, validation=%d\n",
2155                          failrec->logical, failrec->start, failrec->len,
2156                          failrec->in_validation);
2157                 /*
2158                  * when data can be on disk more than twice, add to failrec here
2159                  * (e.g. with a list for failed_mirror) to make
2160                  * clean_io_failure() clean all those errors at once.
2161                  */
2162         }
2163         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2164                                       failrec->logical, failrec->len);
2165         if (num_copies == 1) {
2166                 /*
2167                  * we only have a single copy of the data, so don't bother with
2168                  * all the retry and error correction code that follows. no
2169                  * matter what the error is, it is very likely to persist.
2170                  */
2171                 pr_debug("bio_readpage_error: cannot repair, num_copies == 1. "
2172                          "state=%p, num_copies=%d, next_mirror %d, "
2173                          "failed_mirror %d\n", state, num_copies,
2174                          failrec->this_mirror, failed_mirror);
2175                 free_io_failure(inode, failrec, 0);
2176                 return -EIO;
2177         }
2178
2179         if (!state) {
2180                 spin_lock(&tree->lock);
2181                 state = find_first_extent_bit_state(tree, failrec->start,
2182                                                     EXTENT_LOCKED);
2183                 if (state && state->start != failrec->start)
2184                         state = NULL;
2185                 spin_unlock(&tree->lock);
2186         }
2187
2188         /*
2189          * there are two premises:
2190          *      a) deliver good data to the caller
2191          *      b) correct the bad sectors on disk
2192          */
2193         if (failed_bio->bi_vcnt > 1) {
2194                 /*
2195                  * to fulfill b), we need to know the exact failing sectors, as
2196                  * we don't want to rewrite any more than the failed ones. thus,
2197                  * we need separate read requests for the failed bio
2198                  *
2199                  * if the following BUG_ON triggers, our validation request got
2200                  * merged. we need separate requests for our algorithm to work.
2201                  */
2202                 BUG_ON(failrec->in_validation);
2203                 failrec->in_validation = 1;
2204                 failrec->this_mirror = failed_mirror;
2205                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2206         } else {
2207                 /*
2208                  * we're ready to fulfill a) and b) alongside. get a good copy
2209                  * of the failed sector and if we succeed, we have setup
2210                  * everything for repair_io_failure to do the rest for us.
2211                  */
2212                 if (failrec->in_validation) {
2213                         BUG_ON(failrec->this_mirror != failed_mirror);
2214                         failrec->in_validation = 0;
2215                         failrec->this_mirror = 0;
2216                 }
2217                 failrec->failed_mirror = failed_mirror;
2218                 failrec->this_mirror++;
2219                 if (failrec->this_mirror == failed_mirror)
2220                         failrec->this_mirror++;
2221                 read_mode = READ_SYNC;
2222         }
2223
2224         if (!state || failrec->this_mirror > num_copies) {
2225                 pr_debug("bio_readpage_error: (fail) state=%p, num_copies=%d, "
2226                          "next_mirror %d, failed_mirror %d\n", state,
2227                          num_copies, failrec->this_mirror, failed_mirror);
2228                 free_io_failure(inode, failrec, 0);
2229                 return -EIO;
2230         }
2231
2232         bio = bio_alloc(GFP_NOFS, 1);
2233         if (!bio) {
2234                 free_io_failure(inode, failrec, 0);
2235                 return -EIO;
2236         }
2237         bio->bi_private = state;
2238         bio->bi_end_io = failed_bio->bi_end_io;
2239         bio->bi_sector = failrec->logical >> 9;
2240         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2241         bio->bi_size = 0;
2242
2243         bio_add_page(bio, page, failrec->len, start - page_offset(page));
2244
2245         pr_debug("bio_readpage_error: submitting new read[%#x] to "
2246                  "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2247                  failrec->this_mirror, num_copies, failrec->in_validation);
2248
2249         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2250                                          failrec->this_mirror,
2251                                          failrec->bio_flags, 0);
2252         return ret;
2253 }
2254
2255 /* lots and lots of room for performance fixes in the end_bio funcs */
2256
2257 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2258 {
2259         int uptodate = (err == 0);
2260         struct extent_io_tree *tree;
2261         int ret;
2262
2263         tree = &BTRFS_I(page->mapping->host)->io_tree;
2264
2265         if (tree->ops && tree->ops->writepage_end_io_hook) {
2266                 ret = tree->ops->writepage_end_io_hook(page, start,
2267                                                end, NULL, uptodate);
2268                 if (ret)
2269                         uptodate = 0;
2270         }
2271
2272         if (!uptodate) {
2273                 ClearPageUptodate(page);
2274                 SetPageError(page);
2275         }
2276         return 0;
2277 }
2278
2279 /*
2280  * after a writepage IO is done, we need to:
2281  * clear the uptodate bits on error
2282  * clear the writeback bits in the extent tree for this IO
2283  * end_page_writeback if the page has no more pending IO
2284  *
2285  * Scheduling is not allowed, so the extent state tree is expected
2286  * to have one and only one object corresponding to this IO.
2287  */
2288 static void end_bio_extent_writepage(struct bio *bio, int err)
2289 {
2290         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2291         struct extent_io_tree *tree;
2292         u64 start;
2293         u64 end;
2294         int whole_page;
2295
2296         do {
2297                 struct page *page = bvec->bv_page;
2298                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2299
2300                 start = page_offset(page) + bvec->bv_offset;
2301                 end = start + bvec->bv_len - 1;
2302
2303                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2304                         whole_page = 1;
2305                 else
2306                         whole_page = 0;
2307
2308                 if (--bvec >= bio->bi_io_vec)
2309                         prefetchw(&bvec->bv_page->flags);
2310
2311                 if (end_extent_writepage(page, err, start, end))
2312                         continue;
2313
2314                 if (whole_page)
2315                         end_page_writeback(page);
2316                 else
2317                         check_page_writeback(tree, page);
2318         } while (bvec >= bio->bi_io_vec);
2319
2320         bio_put(bio);
2321 }
2322
2323 /*
2324  * after a readpage IO is done, we need to:
2325  * clear the uptodate bits on error
2326  * set the uptodate bits if things worked
2327  * set the page up to date if all extents in the tree are uptodate
2328  * clear the lock bit in the extent tree
2329  * unlock the page if there are no other extents locked for it
2330  *
2331  * Scheduling is not allowed, so the extent state tree is expected
2332  * to have one and only one object corresponding to this IO.
2333  */
2334 static void end_bio_extent_readpage(struct bio *bio, int err)
2335 {
2336         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2337         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
2338         struct bio_vec *bvec = bio->bi_io_vec;
2339         struct extent_io_tree *tree;
2340         u64 start;
2341         u64 end;
2342         int whole_page;
2343         int mirror;
2344         int ret;
2345
2346         if (err)
2347                 uptodate = 0;
2348
2349         do {
2350                 struct page *page = bvec->bv_page;
2351                 struct extent_state *cached = NULL;
2352                 struct extent_state *state;
2353
2354                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2355                          "mirror=%ld\n", (u64)bio->bi_sector, err,
2356                          (long int)bio->bi_bdev);
2357                 tree = &BTRFS_I(page->mapping->host)->io_tree;
2358
2359                 start = page_offset(page) + bvec->bv_offset;
2360                 end = start + bvec->bv_len - 1;
2361
2362                 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
2363                         whole_page = 1;
2364                 else
2365                         whole_page = 0;
2366
2367                 if (++bvec <= bvec_end)
2368                         prefetchw(&bvec->bv_page->flags);
2369
2370                 spin_lock(&tree->lock);
2371                 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
2372                 if (state && state->start == start) {
2373                         /*
2374                          * take a reference on the state, unlock will drop
2375                          * the ref
2376                          */
2377                         cache_state(state, &cached);
2378                 }
2379                 spin_unlock(&tree->lock);
2380
2381                 mirror = (int)(unsigned long)bio->bi_bdev;
2382                 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
2383                         ret = tree->ops->readpage_end_io_hook(page, start, end,
2384                                                               state, mirror);
2385                         if (ret)
2386                                 uptodate = 0;
2387                         else
2388                                 clean_io_failure(start, page);
2389                 }
2390
2391                 if (!uptodate && tree->ops && tree->ops->readpage_io_failed_hook) {
2392                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2393                         if (!ret && !err &&
2394                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2395                                 uptodate = 1;
2396                 } else if (!uptodate) {
2397                         /*
2398                          * The generic bio_readpage_error handles errors the
2399                          * following way: If possible, new read requests are
2400                          * created and submitted and will end up in
2401                          * end_bio_extent_readpage as well (if we're lucky, not
2402                          * in the !uptodate case). In that case it returns 0 and
2403                          * we just go on with the next page in our bio. If it
2404                          * can't handle the error it will return -EIO and we
2405                          * remain responsible for that page.
2406                          */
2407                         ret = bio_readpage_error(bio, page, start, end, mirror, NULL);
2408                         if (ret == 0) {
2409                                 uptodate =
2410                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2411                                 if (err)
2412                                         uptodate = 0;
2413                                 uncache_state(&cached);
2414                                 continue;
2415                         }
2416                 }
2417
2418                 if (uptodate && tree->track_uptodate) {
2419                         set_extent_uptodate(tree, start, end, &cached,
2420                                             GFP_ATOMIC);
2421                 }
2422                 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2423
2424                 if (whole_page) {
2425                         if (uptodate) {
2426                                 SetPageUptodate(page);
2427                         } else {
2428                                 ClearPageUptodate(page);
2429                                 SetPageError(page);
2430                         }
2431                         unlock_page(page);
2432                 } else {
2433                         if (uptodate) {
2434                                 check_page_uptodate(tree, page);
2435                         } else {
2436                                 ClearPageUptodate(page);
2437                                 SetPageError(page);
2438                         }
2439                         check_page_locked(tree, page);
2440                 }
2441         } while (bvec <= bvec_end);
2442
2443         bio_put(bio);
2444 }
2445
2446 struct bio *
2447 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2448                 gfp_t gfp_flags)
2449 {
2450         struct bio *bio;
2451
2452         bio = bio_alloc(gfp_flags, nr_vecs);
2453
2454         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2455                 while (!bio && (nr_vecs /= 2))
2456                         bio = bio_alloc(gfp_flags, nr_vecs);
2457         }
2458
2459         if (bio) {
2460                 bio->bi_size = 0;
2461                 bio->bi_bdev = bdev;
2462                 bio->bi_sector = first_sector;
2463         }
2464         return bio;
2465 }
2466
2467 static int __must_check submit_one_bio(int rw, struct bio *bio,
2468                                        int mirror_num, unsigned long bio_flags)
2469 {
2470         int ret = 0;
2471         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2472         struct page *page = bvec->bv_page;
2473         struct extent_io_tree *tree = bio->bi_private;
2474         u64 start;
2475
2476         start = page_offset(page) + bvec->bv_offset;
2477
2478         bio->bi_private = NULL;
2479
2480         bio_get(bio);
2481
2482         if (tree->ops && tree->ops->submit_bio_hook)
2483                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2484                                            mirror_num, bio_flags, start);
2485         else
2486                 btrfsic_submit_bio(rw, bio);
2487
2488         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2489                 ret = -EOPNOTSUPP;
2490         bio_put(bio);
2491         return ret;
2492 }
2493
2494 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2495                      unsigned long offset, size_t size, struct bio *bio,
2496                      unsigned long bio_flags)
2497 {
2498         int ret = 0;
2499         if (tree->ops && tree->ops->merge_bio_hook)
2500                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2501                                                 bio_flags);
2502         BUG_ON(ret < 0);
2503         return ret;
2504
2505 }
2506
2507 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2508                               struct page *page, sector_t sector,
2509                               size_t size, unsigned long offset,
2510                               struct block_device *bdev,
2511                               struct bio **bio_ret,
2512                               unsigned long max_pages,
2513                               bio_end_io_t end_io_func,
2514                               int mirror_num,
2515                               unsigned long prev_bio_flags,
2516                               unsigned long bio_flags)
2517 {
2518         int ret = 0;
2519         struct bio *bio;
2520         int nr;
2521         int contig = 0;
2522         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2523         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2524         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2525
2526         if (bio_ret && *bio_ret) {
2527                 bio = *bio_ret;
2528                 if (old_compressed)
2529                         contig = bio->bi_sector == sector;
2530                 else
2531                         contig = bio->bi_sector + (bio->bi_size >> 9) ==
2532                                 sector;
2533
2534                 if (prev_bio_flags != bio_flags || !contig ||
2535                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2536                     bio_add_page(bio, page, page_size, offset) < page_size) {
2537                         ret = submit_one_bio(rw, bio, mirror_num,
2538                                              prev_bio_flags);
2539                         if (ret < 0)
2540                                 return ret;
2541                         bio = NULL;
2542                 } else {
2543                         return 0;
2544                 }
2545         }
2546         if (this_compressed)
2547                 nr = BIO_MAX_PAGES;
2548         else
2549                 nr = bio_get_nr_vecs(bdev);
2550
2551         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2552         if (!bio)
2553                 return -ENOMEM;
2554
2555         bio_add_page(bio, page, page_size, offset);
2556         bio->bi_end_io = end_io_func;
2557         bio->bi_private = tree;
2558
2559         if (bio_ret)
2560                 *bio_ret = bio;
2561         else
2562                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2563
2564         return ret;
2565 }
2566
2567 void attach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
2568 {
2569         if (!PagePrivate(page)) {
2570                 SetPagePrivate(page);
2571                 page_cache_get(page);
2572                 set_page_private(page, (unsigned long)eb);
2573         } else {
2574                 WARN_ON(page->private != (unsigned long)eb);
2575         }
2576 }
2577
2578 void set_page_extent_mapped(struct page *page)
2579 {
2580         if (!PagePrivate(page)) {
2581                 SetPagePrivate(page);
2582                 page_cache_get(page);
2583                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2584         }
2585 }
2586
2587 /*
2588  * basic readpage implementation.  Locked extent state structs are inserted
2589  * into the tree that are removed when the IO is done (by the end_io
2590  * handlers)
2591  * XXX JDM: This needs looking at to ensure proper page locking
2592  */
2593 static int __extent_read_full_page(struct extent_io_tree *tree,
2594                                    struct page *page,
2595                                    get_extent_t *get_extent,
2596                                    struct bio **bio, int mirror_num,
2597                                    unsigned long *bio_flags)
2598 {
2599         struct inode *inode = page->mapping->host;
2600         u64 start = page_offset(page);
2601         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2602         u64 end;
2603         u64 cur = start;
2604         u64 extent_offset;
2605         u64 last_byte = i_size_read(inode);
2606         u64 block_start;
2607         u64 cur_end;
2608         sector_t sector;
2609         struct extent_map *em;
2610         struct block_device *bdev;
2611         struct btrfs_ordered_extent *ordered;
2612         int ret;
2613         int nr = 0;
2614         size_t pg_offset = 0;
2615         size_t iosize;
2616         size_t disk_io_size;
2617         size_t blocksize = inode->i_sb->s_blocksize;
2618         unsigned long this_bio_flag = 0;
2619
2620         set_page_extent_mapped(page);
2621
2622         if (!PageUptodate(page)) {
2623                 if (cleancache_get_page(page) == 0) {
2624                         BUG_ON(blocksize != PAGE_SIZE);
2625                         goto out;
2626                 }
2627         }
2628
2629         end = page_end;
2630         while (1) {
2631                 lock_extent(tree, start, end);
2632                 ordered = btrfs_lookup_ordered_extent(inode, start);
2633                 if (!ordered)
2634                         break;
2635                 unlock_extent(tree, start, end);
2636                 btrfs_start_ordered_extent(inode, ordered, 1);
2637                 btrfs_put_ordered_extent(ordered);
2638         }
2639
2640         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2641                 char *userpage;
2642                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2643
2644                 if (zero_offset) {
2645                         iosize = PAGE_CACHE_SIZE - zero_offset;
2646                         userpage = kmap_atomic(page);
2647                         memset(userpage + zero_offset, 0, iosize);
2648                         flush_dcache_page(page);
2649                         kunmap_atomic(userpage);
2650                 }
2651         }
2652         while (cur <= end) {
2653                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2654
2655                 if (cur >= last_byte) {
2656                         char *userpage;
2657                         struct extent_state *cached = NULL;
2658
2659                         iosize = PAGE_CACHE_SIZE - pg_offset;
2660                         userpage = kmap_atomic(page);
2661                         memset(userpage + pg_offset, 0, iosize);
2662                         flush_dcache_page(page);
2663                         kunmap_atomic(userpage);
2664                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2665                                             &cached, GFP_NOFS);
2666                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2667                                              &cached, GFP_NOFS);
2668                         break;
2669                 }
2670                 em = get_extent(inode, page, pg_offset, cur,
2671                                 end - cur + 1, 0);
2672                 if (IS_ERR_OR_NULL(em)) {
2673                         SetPageError(page);
2674                         unlock_extent(tree, cur, end);
2675                         break;
2676                 }
2677                 extent_offset = cur - em->start;
2678                 BUG_ON(extent_map_end(em) <= cur);
2679                 BUG_ON(end < cur);
2680
2681                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2682                         this_bio_flag = EXTENT_BIO_COMPRESSED;
2683                         extent_set_compress_type(&this_bio_flag,
2684                                                  em->compress_type);
2685                 }
2686
2687                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2688                 cur_end = min(extent_map_end(em) - 1, end);
2689                 iosize = ALIGN(iosize, blocksize);
2690                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2691                         disk_io_size = em->block_len;
2692                         sector = em->block_start >> 9;
2693                 } else {
2694                         sector = (em->block_start + extent_offset) >> 9;
2695                         disk_io_size = iosize;
2696                 }
2697                 bdev = em->bdev;
2698                 block_start = em->block_start;
2699                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2700                         block_start = EXTENT_MAP_HOLE;
2701                 free_extent_map(em);
2702                 em = NULL;
2703
2704                 /* we've found a hole, just zero and go on */
2705                 if (block_start == EXTENT_MAP_HOLE) {
2706                         char *userpage;
2707                         struct extent_state *cached = NULL;
2708
2709                         userpage = kmap_atomic(page);
2710                         memset(userpage + pg_offset, 0, iosize);
2711                         flush_dcache_page(page);
2712                         kunmap_atomic(userpage);
2713
2714                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2715                                             &cached, GFP_NOFS);
2716                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2717                                              &cached, GFP_NOFS);
2718                         cur = cur + iosize;
2719                         pg_offset += iosize;
2720                         continue;
2721                 }
2722                 /* the get_extent function already copied into the page */
2723                 if (test_range_bit(tree, cur, cur_end,
2724                                    EXTENT_UPTODATE, 1, NULL)) {
2725                         check_page_uptodate(tree, page);
2726                         unlock_extent(tree, cur, cur + iosize - 1);
2727                         cur = cur + iosize;
2728                         pg_offset += iosize;
2729                         continue;
2730                 }
2731                 /* we have an inline extent but it didn't get marked up
2732                  * to date.  Error out
2733                  */
2734                 if (block_start == EXTENT_MAP_INLINE) {
2735                         SetPageError(page);
2736                         unlock_extent(tree, cur, cur + iosize - 1);
2737                         cur = cur + iosize;
2738                         pg_offset += iosize;
2739                         continue;
2740                 }
2741
2742                 pnr -= page->index;
2743                 ret = submit_extent_page(READ, tree, page,
2744                                          sector, disk_io_size, pg_offset,
2745                                          bdev, bio, pnr,
2746                                          end_bio_extent_readpage, mirror_num,
2747                                          *bio_flags,
2748                                          this_bio_flag);
2749                 if (!ret) {
2750                         nr++;
2751                         *bio_flags = this_bio_flag;
2752                 } else {
2753                         SetPageError(page);
2754                         unlock_extent(tree, cur, cur + iosize - 1);
2755                 }
2756                 cur = cur + iosize;
2757                 pg_offset += iosize;
2758         }
2759 out:
2760         if (!nr) {
2761                 if (!PageError(page))
2762                         SetPageUptodate(page);
2763                 unlock_page(page);
2764         }
2765         return 0;
2766 }
2767
2768 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2769                             get_extent_t *get_extent, int mirror_num)
2770 {
2771         struct bio *bio = NULL;
2772         unsigned long bio_flags = 0;
2773         int ret;
2774
2775         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
2776                                       &bio_flags);
2777         if (bio)
2778                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
2779         return ret;
2780 }
2781
2782 static noinline void update_nr_written(struct page *page,
2783                                       struct writeback_control *wbc,
2784                                       unsigned long nr_written)
2785 {
2786         wbc->nr_to_write -= nr_written;
2787         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2788             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2789                 page->mapping->writeback_index = page->index + nr_written;
2790 }
2791
2792 /*
2793  * the writepage semantics are similar to regular writepage.  extent
2794  * records are inserted to lock ranges in the tree, and as dirty areas
2795  * are found, they are marked writeback.  Then the lock bits are removed
2796  * and the end_io handler clears the writeback ranges
2797  */
2798 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2799                               void *data)
2800 {
2801         struct inode *inode = page->mapping->host;
2802         struct extent_page_data *epd = data;
2803         struct extent_io_tree *tree = epd->tree;
2804         u64 start = page_offset(page);
2805         u64 delalloc_start;
2806         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2807         u64 end;
2808         u64 cur = start;
2809         u64 extent_offset;
2810         u64 last_byte = i_size_read(inode);
2811         u64 block_start;
2812         u64 iosize;
2813         sector_t sector;
2814         struct extent_state *cached_state = NULL;
2815         struct extent_map *em;
2816         struct block_device *bdev;
2817         int ret;
2818         int nr = 0;
2819         size_t pg_offset = 0;
2820         size_t blocksize;
2821         loff_t i_size = i_size_read(inode);
2822         unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2823         u64 nr_delalloc;
2824         u64 delalloc_end;
2825         int page_started;
2826         int compressed;
2827         int write_flags;
2828         unsigned long nr_written = 0;
2829         bool fill_delalloc = true;
2830
2831         if (wbc->sync_mode == WB_SYNC_ALL)
2832                 write_flags = WRITE_SYNC;
2833         else
2834                 write_flags = WRITE;
2835
2836         trace___extent_writepage(page, inode, wbc);
2837
2838         WARN_ON(!PageLocked(page));
2839
2840         ClearPageError(page);
2841
2842         pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2843         if (page->index > end_index ||
2844            (page->index == end_index && !pg_offset)) {
2845                 page->mapping->a_ops->invalidatepage(page, 0);
2846                 unlock_page(page);
2847                 return 0;
2848         }
2849
2850         if (page->index == end_index) {
2851                 char *userpage;
2852
2853                 userpage = kmap_atomic(page);
2854                 memset(userpage + pg_offset, 0,
2855                        PAGE_CACHE_SIZE - pg_offset);
2856                 kunmap_atomic(userpage);
2857                 flush_dcache_page(page);
2858         }
2859         pg_offset = 0;
2860
2861         set_page_extent_mapped(page);
2862
2863         if (!tree->ops || !tree->ops->fill_delalloc)
2864                 fill_delalloc = false;
2865
2866         delalloc_start = start;
2867         delalloc_end = 0;
2868         page_started = 0;
2869         if (!epd->extent_locked && fill_delalloc) {
2870                 u64 delalloc_to_write = 0;
2871                 /*
2872                  * make sure the wbc mapping index is at least updated
2873                  * to this page.
2874                  */
2875                 update_nr_written(page, wbc, 0);
2876
2877                 while (delalloc_end < page_end) {
2878                         nr_delalloc = find_lock_delalloc_range(inode, tree,
2879                                                        page,
2880                                                        &delalloc_start,
2881                                                        &delalloc_end,
2882                                                        128 * 1024 * 1024);
2883                         if (nr_delalloc == 0) {
2884                                 delalloc_start = delalloc_end + 1;
2885                                 continue;
2886                         }
2887                         ret = tree->ops->fill_delalloc(inode, page,
2888                                                        delalloc_start,
2889                                                        delalloc_end,
2890                                                        &page_started,
2891                                                        &nr_written);
2892                         /* File system has been set read-only */
2893                         if (ret) {
2894                                 SetPageError(page);
2895                                 goto done;
2896                         }
2897                         /*
2898                          * delalloc_end is already one less than the total
2899                          * length, so we don't subtract one from
2900                          * PAGE_CACHE_SIZE
2901                          */
2902                         delalloc_to_write += (delalloc_end - delalloc_start +
2903                                               PAGE_CACHE_SIZE) >>
2904                                               PAGE_CACHE_SHIFT;
2905                         delalloc_start = delalloc_end + 1;
2906                 }
2907                 if (wbc->nr_to_write < delalloc_to_write) {
2908                         int thresh = 8192;
2909
2910                         if (delalloc_to_write < thresh * 2)
2911                                 thresh = delalloc_to_write;
2912                         wbc->nr_to_write = min_t(u64, delalloc_to_write,
2913                                                  thresh);
2914                 }
2915
2916                 /* did the fill delalloc function already unlock and start
2917                  * the IO?
2918                  */
2919                 if (page_started) {
2920                         ret = 0;
2921                         /*
2922                          * we've unlocked the page, so we can't update
2923                          * the mapping's writeback index, just update
2924                          * nr_to_write.
2925                          */
2926                         wbc->nr_to_write -= nr_written;
2927                         goto done_unlocked;
2928                 }
2929         }
2930         if (tree->ops && tree->ops->writepage_start_hook) {
2931                 ret = tree->ops->writepage_start_hook(page, start,
2932                                                       page_end);
2933                 if (ret) {
2934                         /* Fixup worker will requeue */
2935                         if (ret == -EBUSY)
2936                                 wbc->pages_skipped++;
2937                         else
2938                                 redirty_page_for_writepage(wbc, page);
2939                         update_nr_written(page, wbc, nr_written);
2940                         unlock_page(page);
2941                         ret = 0;
2942                         goto done_unlocked;
2943                 }
2944         }
2945
2946         /*
2947          * we don't want to touch the inode after unlocking the page,
2948          * so we update the mapping writeback index now
2949          */
2950         update_nr_written(page, wbc, nr_written + 1);
2951
2952         end = page_end;
2953         if (last_byte <= start) {
2954                 if (tree->ops && tree->ops->writepage_end_io_hook)
2955                         tree->ops->writepage_end_io_hook(page, start,
2956                                                          page_end, NULL, 1);
2957                 goto done;
2958         }
2959
2960         blocksize = inode->i_sb->s_blocksize;
2961
2962         while (cur <= end) {
2963                 if (cur >= last_byte) {
2964                         if (tree->ops && tree->ops->writepage_end_io_hook)
2965                                 tree->ops->writepage_end_io_hook(page, cur,
2966                                                          page_end, NULL, 1);
2967                         break;
2968                 }
2969                 em = epd->get_extent(inode, page, pg_offset, cur,
2970                                      end - cur + 1, 1);
2971                 if (IS_ERR_OR_NULL(em)) {
2972                         SetPageError(page);
2973                         break;
2974                 }
2975
2976                 extent_offset = cur - em->start;
2977                 BUG_ON(extent_map_end(em) <= cur);
2978                 BUG_ON(end < cur);
2979                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2980                 iosize = ALIGN(iosize, blocksize);
2981                 sector = (em->block_start + extent_offset) >> 9;
2982                 bdev = em->bdev;
2983                 block_start = em->block_start;
2984                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2985                 free_extent_map(em);
2986                 em = NULL;
2987
2988                 /*
2989                  * compressed and inline extents are written through other
2990                  * paths in the FS
2991                  */
2992                 if (compressed || block_start == EXTENT_MAP_HOLE ||
2993                     block_start == EXTENT_MAP_INLINE) {
2994                         /*
2995                          * end_io notification does not happen here for
2996                          * compressed extents
2997                          */
2998                         if (!compressed && tree->ops &&
2999                             tree->ops->writepage_end_io_hook)
3000                                 tree->ops->writepage_end_io_hook(page, cur,
3001                                                          cur + iosize - 1,
3002                                                          NULL, 1);
3003                         else if (compressed) {
3004                                 /* we don't want to end_page_writeback on
3005                                  * a compressed extent.  this happens
3006                                  * elsewhere
3007                                  */
3008                                 nr++;
3009                         }
3010
3011                         cur += iosize;
3012                         pg_offset += iosize;
3013                         continue;
3014                 }
3015                 /* leave this out until we have a page_mkwrite call */
3016                 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3017                                    EXTENT_DIRTY, 0, NULL)) {
3018                         cur = cur + iosize;
3019                         pg_offset += iosize;
3020                         continue;
3021                 }
3022
3023                 if (tree->ops && tree->ops->writepage_io_hook) {
3024                         ret = tree->ops->writepage_io_hook(page, cur,
3025                                                 cur + iosize - 1);
3026                 } else {
3027                         ret = 0;
3028                 }
3029                 if (ret) {
3030                         SetPageError(page);
3031                 } else {
3032                         unsigned long max_nr = end_index + 1;
3033
3034                         set_range_writeback(tree, cur, cur + iosize - 1);
3035                         if (!PageWriteback(page)) {
3036                                 printk(KERN_ERR "btrfs warning page %lu not "
3037                                        "writeback, cur %llu end %llu\n",
3038                                        page->index, (unsigned long long)cur,
3039                                        (unsigned long long)end);
3040                         }
3041
3042                         ret = submit_extent_page(write_flags, tree, page,
3043                                                  sector, iosize, pg_offset,
3044                                                  bdev, &epd->bio, max_nr,
3045                                                  end_bio_extent_writepage,
3046                                                  0, 0, 0);
3047                         if (ret)
3048                                 SetPageError(page);
3049                 }
3050                 cur = cur + iosize;
3051                 pg_offset += iosize;
3052                 nr++;
3053         }
3054 done:
3055         if (nr == 0) {
3056                 /* make sure the mapping tag for page dirty gets cleared */
3057                 set_page_writeback(page);
3058                 end_page_writeback(page);
3059         }
3060         unlock_page(page);
3061
3062 done_unlocked:
3063
3064         /* drop our reference on any cached states */
3065         free_extent_state(cached_state);
3066         return 0;
3067 }
3068
3069 static int eb_wait(void *word)
3070 {
3071         io_schedule();
3072         return 0;
3073 }
3074
3075 static void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3076 {
3077         wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3078                     TASK_UNINTERRUPTIBLE);
3079 }
3080
3081 static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3082                                      struct btrfs_fs_info *fs_info,
3083                                      struct extent_page_data *epd)
3084 {
3085         unsigned long i, num_pages;
3086         int flush = 0;
3087         int ret = 0;
3088
3089         if (!btrfs_try_tree_write_lock(eb)) {
3090                 flush = 1;
3091                 flush_write_bio(epd);
3092                 btrfs_tree_lock(eb);
3093         }
3094
3095         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3096                 btrfs_tree_unlock(eb);
3097                 if (!epd->sync_io)
3098                         return 0;
3099                 if (!flush) {
3100                         flush_write_bio(epd);
3101                         flush = 1;
3102                 }
3103                 while (1) {
3104                         wait_on_extent_buffer_writeback(eb);
3105                         btrfs_tree_lock(eb);
3106                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3107                                 break;
3108                         btrfs_tree_unlock(eb);
3109                 }
3110         }
3111
3112         /*
3113          * We need to do this to prevent races in people who check if the eb is
3114          * under IO since we can end up having no IO bits set for a short period
3115          * of time.
3116          */
3117         spin_lock(&eb->refs_lock);
3118         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3119                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3120                 spin_unlock(&eb->refs_lock);
3121                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3122                 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3123                                      -eb->len,
3124                                      fs_info->dirty_metadata_batch);
3125                 ret = 1;
3126         } else {
3127                 spin_unlock(&eb->refs_lock);
3128         }
3129
3130         btrfs_tree_unlock(eb);
3131
3132         if (!ret)
3133                 return ret;
3134
3135         num_pages = num_extent_pages(eb->start, eb->len);
3136         for (i = 0; i < num_pages; i++) {
3137                 struct page *p = extent_buffer_page(eb, i);
3138
3139                 if (!trylock_page(p)) {
3140                         if (!flush) {
3141                                 flush_write_bio(epd);
3142                                 flush = 1;
3143                         }
3144                         lock_page(p);
3145                 }
3146         }
3147
3148         return ret;
3149 }
3150
3151 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3152 {
3153         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3154         smp_mb__after_clear_bit();
3155         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3156 }
3157
3158 static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3159 {
3160         int uptodate = err == 0;
3161         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
3162         struct extent_buffer *eb;
3163         int done;
3164
3165         do {
3166                 struct page *page = bvec->bv_page;
3167
3168                 bvec--;
3169                 eb = (struct extent_buffer *)page->private;
3170                 BUG_ON(!eb);
3171                 done = atomic_dec_and_test(&eb->io_pages);
3172
3173                 if (!uptodate || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3174                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3175                         ClearPageUptodate(page);
3176                         SetPageError(page);
3177                 }
3178
3179                 end_page_writeback(page);
3180
3181                 if (!done)
3182                         continue;
3183
3184                 end_extent_buffer_writeback(eb);
3185         } while (bvec >= bio->bi_io_vec);
3186
3187         bio_put(bio);
3188
3189 }
3190
3191 static int write_one_eb(struct extent_buffer *eb,
3192                         struct btrfs_fs_info *fs_info,
3193                         struct writeback_control *wbc,
3194                         struct extent_page_data *epd)
3195 {
3196         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3197         u64 offset = eb->start;
3198         unsigned long i, num_pages;
3199         unsigned long bio_flags = 0;
3200         int rw = (epd->sync_io ? WRITE_SYNC : WRITE);
3201         int ret = 0;
3202
3203         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3204         num_pages = num_extent_pages(eb->start, eb->len);
3205         atomic_set(&eb->io_pages, num_pages);
3206         if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3207                 bio_flags = EXTENT_BIO_TREE_LOG;
3208
3209         for (i = 0; i < num_pages; i++) {
3210                 struct page *p = extent_buffer_page(eb, i);
3211
3212                 clear_page_dirty_for_io(p);
3213                 set_page_writeback(p);
3214                 ret = submit_extent_page(rw, eb->tree, p, offset >> 9,
3215                                          PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3216                                          -1, end_bio_extent_buffer_writepage,
3217                                          0, epd->bio_flags, bio_flags);
3218                 epd->bio_flags = bio_flags;
3219                 if (ret) {
3220                         set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3221                         SetPageError(p);
3222                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3223                                 end_extent_buffer_writeback(eb);
3224                         ret = -EIO;
3225                         break;
3226                 }
3227                 offset += PAGE_CACHE_SIZE;
3228                 update_nr_written(p, wbc, 1);
3229                 unlock_page(p);
3230         }
3231
3232         if (unlikely(ret)) {
3233                 for (; i < num_pages; i++) {
3234                         struct page *p = extent_buffer_page(eb, i);
3235                         unlock_page(p);
3236                 }
3237         }
3238
3239         return ret;
3240 }
3241
3242 int btree_write_cache_pages(struct address_space *mapping,
3243                                    struct writeback_control *wbc)
3244 {
3245         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3246         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3247         struct extent_buffer *eb, *prev_eb = NULL;
3248         struct extent_page_data epd = {
3249                 .bio = NULL,
3250                 .tree = tree,
3251                 .extent_locked = 0,
3252                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3253                 .bio_flags = 0,
3254         };
3255         int ret = 0;
3256         int done = 0;
3257         int nr_to_write_done = 0;
3258         struct pagevec pvec;
3259         int nr_pages;
3260         pgoff_t index;
3261         pgoff_t end;            /* Inclusive */
3262         int scanned = 0;
3263         int tag;
3264
3265         pagevec_init(&pvec, 0);
3266         if (wbc->range_cyclic) {
3267                 index = mapping->writeback_index; /* Start from prev offset */
3268                 end = -1;
3269         } else {
3270                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3271                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3272                 scanned = 1;
3273         }
3274         if (wbc->sync_mode == WB_SYNC_ALL)
3275                 tag = PAGECACHE_TAG_TOWRITE;
3276         else
3277                 tag = PAGECACHE_TAG_DIRTY;
3278 retry:
3279         if (wbc->sync_mode == WB_SYNC_ALL)
3280                 tag_pages_for_writeback(mapping, index, end);
3281         while (!done && !nr_to_write_done && (index <= end) &&
3282                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3283                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3284                 unsigned i;
3285
3286                 scanned = 1;
3287                 for (i = 0; i < nr_pages; i++) {
3288                         struct page *page = pvec.pages[i];
3289
3290                         if (!PagePrivate(page))
3291                                 continue;
3292
3293                         if (!wbc->range_cyclic && page->index > end) {
3294                                 done = 1;
3295                                 break;
3296                         }
3297
3298                         spin_lock(&mapping->private_lock);
3299                         if (!PagePrivate(page)) {
3300                                 spin_unlock(&mapping->private_lock);
3301                                 continue;
3302                         }
3303
3304                         eb = (struct extent_buffer *)page->private;
3305
3306                         /*
3307                          * Shouldn't happen and normally this would be a BUG_ON
3308                          * but no sense in crashing the users box for something
3309                          * we can survive anyway.
3310                          */
3311                         if (!eb) {
3312                                 spin_unlock(&mapping->private_lock);
3313                                 WARN_ON(1);
3314                                 continue;
3315                         }
3316
3317                         if (eb == prev_eb) {
3318                                 spin_unlock(&mapping->private_lock);
3319                                 continue;
3320                         }
3321
3322                         ret = atomic_inc_not_zero(&eb->refs);
3323                         spin_unlock(&mapping->private_lock);
3324                         if (!ret)
3325                                 continue;
3326
3327                         prev_eb = eb;
3328                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3329                         if (!ret) {
3330                                 free_extent_buffer(eb);
3331                                 continue;
3332                         }
3333
3334                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3335                         if (ret) {
3336                                 done = 1;
3337                                 free_extent_buffer(eb);
3338                                 break;
3339                         }
3340                         free_extent_buffer(eb);
3341
3342                         /*
3343                          * the filesystem may choose to bump up nr_to_write.
3344                          * We have to make sure to honor the new nr_to_write
3345                          * at any time
3346                          */
3347                         nr_to_write_done = wbc->nr_to_write <= 0;
3348                 }
3349                 pagevec_release(&pvec);
3350                 cond_resched();
3351         }
3352         if (!scanned && !done) {
3353                 /*
3354                  * We hit the last page and there is more work to be done: wrap
3355                  * back to the start of the file
3356                  */
3357                 scanned = 1;
3358                 index = 0;
3359                 goto retry;
3360         }
3361         flush_write_bio(&epd);
3362         return ret;
3363 }
3364
3365 /**
3366  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3367  * @mapping: address space structure to write
3368  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3369  * @writepage: function called for each page
3370  * @data: data passed to writepage function
3371  *
3372  * If a page is already under I/O, write_cache_pages() skips it, even
3373  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3374  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3375  * and msync() need to guarantee that all the data which was dirty at the time
3376  * the call was made get new I/O started against them.  If wbc->sync_mode is
3377  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3378  * existing IO to complete.
3379  */
3380 static int extent_write_cache_pages(struct extent_io_tree *tree,
3381                              struct address_space *mapping,
3382                              struct writeback_control *wbc,
3383                              writepage_t writepage, void *data,
3384                              void (*flush_fn)(void *))
3385 {
3386         struct inode *inode = mapping->host;
3387         int ret = 0;
3388         int done = 0;
3389         int nr_to_write_done = 0;
3390         struct pagevec pvec;
3391         int nr_pages;
3392         pgoff_t index;
3393         pgoff_t end;            /* Inclusive */
3394         int scanned = 0;
3395         int tag;
3396
3397         /*
3398          * We have to hold onto the inode so that ordered extents can do their
3399          * work when the IO finishes.  The alternative to this is failing to add
3400          * an ordered extent if the igrab() fails there and that is a huge pain
3401          * to deal with, so instead just hold onto the inode throughout the
3402          * writepages operation.  If it fails here we are freeing up the inode
3403          * anyway and we'd rather not waste our time writing out stuff that is
3404          * going to be truncated anyway.
3405          */
3406         if (!igrab(inode))
3407                 return 0;
3408
3409         pagevec_init(&pvec, 0);
3410         if (wbc->range_cyclic) {
3411                 index = mapping->writeback_index; /* Start from prev offset */
3412                 end = -1;
3413         } else {
3414                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3415                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3416                 scanned = 1;
3417         }
3418         if (wbc->sync_mode == WB_SYNC_ALL)
3419                 tag = PAGECACHE_TAG_TOWRITE;
3420         else
3421                 tag = PAGECACHE_TAG_DIRTY;
3422 retry:
3423         if (wbc->sync_mode == WB_SYNC_ALL)
3424                 tag_pages_for_writeback(mapping, index, end);
3425         while (!done && !nr_to_write_done && (index <= end) &&
3426                (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3427                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3428                 unsigned i;
3429
3430                 scanned = 1;
3431                 for (i = 0; i < nr_pages; i++) {
3432                         struct page *page = pvec.pages[i];
3433
3434                         /*
3435                          * At this point we hold neither mapping->tree_lock nor
3436                          * lock on the page itself: the page may be truncated or
3437                          * invalidated (changing page->mapping to NULL), or even
3438                          * swizzled back from swapper_space to tmpfs file
3439                          * mapping
3440                          */
3441                         if (!trylock_page(page)) {
3442                                 flush_fn(data);
3443                                 lock_page(page);
3444                         }
3445
3446                         if (unlikely(page->mapping != mapping)) {
3447                                 unlock_page(page);
3448                                 continue;
3449                         }
3450
3451                         if (!wbc->range_cyclic && page->index > end) {
3452                                 done = 1;
3453                                 unlock_page(page);
3454                                 continue;
3455                         }
3456
3457                         if (wbc->sync_mode != WB_SYNC_NONE) {
3458                                 if (PageWriteback(page))
3459                                         flush_fn(data);
3460                                 wait_on_page_writeback(page);
3461                         }
3462
3463                         if (PageWriteback(page) ||
3464                             !clear_page_dirty_for_io(page)) {
3465                                 unlock_page(page);
3466                                 continue;
3467                         }
3468
3469                         ret = (*writepage)(page, wbc, data);
3470
3471                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3472                                 unlock_page(page);
3473                                 ret = 0;
3474                         }
3475                         if (ret)
3476                                 done = 1;
3477
3478                         /*
3479                          * the filesystem may choose to bump up nr_to_write.
3480                          * We have to make sure to honor the new nr_to_write
3481                          * at any time
3482                          */
3483                         nr_to_write_done = wbc->nr_to_write <= 0;
3484                 }
3485                 pagevec_release(&pvec);
3486                 cond_resched();
3487         }
3488         if (!scanned && !done) {
3489                 /*
3490                  * We hit the last page and there is more work to be done: wrap
3491                  * back to the start of the file
3492                  */
3493                 scanned = 1;
3494                 index = 0;
3495                 goto retry;
3496         }
3497         btrfs_add_delayed_iput(inode);
3498         return ret;
3499 }
3500
3501 static void flush_epd_write_bio(struct extent_page_data *epd)
3502 {
3503         if (epd->bio) {
3504                 int rw = WRITE;
3505                 int ret;
3506
3507                 if (epd->sync_io)
3508                         rw = WRITE_SYNC;
3509
3510                 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3511                 BUG_ON(ret < 0); /* -ENOMEM */
3512                 epd->bio = NULL;
3513         }
3514 }
3515
3516 static noinline void flush_write_bio(void *data)
3517 {
3518         struct extent_page_data *epd = data;
3519         flush_epd_write_bio(epd);
3520 }
3521
3522 int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3523                           get_extent_t *get_extent,
3524                           struct writeback_control *wbc)
3525 {
3526         int ret;
3527         struct extent_page_data epd = {
3528                 .bio = NULL,
3529                 .tree = tree,
3530                 .get_extent = get_extent,
3531                 .extent_locked = 0,
3532                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3533                 .bio_flags = 0,
3534         };
3535
3536         ret = __extent_writepage(page, wbc, &epd);
3537
3538         flush_epd_write_bio(&epd);
3539         return ret;
3540 }
3541
3542 int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3543                               u64 start, u64 end, get_extent_t *get_extent,
3544                               int mode)
3545 {
3546         int ret = 0;
3547         struct address_space *mapping = inode->i_mapping;
3548         struct page *page;
3549         unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3550                 PAGE_CACHE_SHIFT;
3551
3552         struct extent_page_data epd = {
3553                 .bio = NULL,
3554                 .tree = tree,
3555                 .get_extent = get_extent,
3556                 .extent_locked = 1,
3557                 .sync_io = mode == WB_SYNC_ALL,
3558                 .bio_flags = 0,
3559         };
3560         struct writeback_control wbc_writepages = {
3561                 .sync_mode      = mode,
3562                 .nr_to_write    = nr_pages * 2,
3563                 .range_start    = start,
3564                 .range_end      = end + 1,
3565         };
3566
3567         while (start <= end) {
3568                 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3569                 if (clear_page_dirty_for_io(page))
3570                         ret = __extent_writepage(page, &wbc_writepages, &epd);
3571                 else {
3572                         if (tree->ops && tree->ops->writepage_end_io_hook)
3573                                 tree->ops->writepage_end_io_hook(page, start,
3574                                                  start + PAGE_CACHE_SIZE - 1,
3575                                                  NULL, 1);
3576                         unlock_page(page);
3577                 }
3578                 page_cache_release(page);
3579                 start += PAGE_CACHE_SIZE;
3580         }
3581
3582         flush_epd_write_bio(&epd);
3583         return ret;
3584 }
3585
3586 int extent_writepages(struct extent_io_tree *tree,
3587                       struct address_space *mapping,
3588                       get_extent_t *get_extent,
3589                       struct writeback_control *wbc)
3590 {
3591         int ret = 0;
3592         struct extent_page_data epd = {
3593                 .bio = NULL,
3594                 .tree = tree,
3595                 .get_extent = get_extent,
3596                 .extent_locked = 0,
3597                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3598                 .bio_flags = 0,
3599         };
3600
3601         ret = extent_write_cache_pages(tree, mapping, wbc,
3602                                        __extent_writepage, &epd,
3603                                        flush_write_bio);
3604         flush_epd_write_bio(&epd);
3605         return ret;
3606 }
3607
3608 int extent_readpages(struct extent_io_tree *tree,
3609                      struct address_space *mapping,
3610                      struct list_head *pages, unsigned nr_pages,
3611                      get_extent_t get_extent)
3612 {
3613         struct bio *bio = NULL;
3614         unsigned page_idx;
3615         unsigned long bio_flags = 0;
3616         struct page *pagepool[16];
3617         struct page *page;
3618         int i = 0;
3619         int nr = 0;
3620
3621         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3622                 page = list_entry(pages->prev, struct page, lru);
3623
3624                 prefetchw(&page->flags);
3625                 list_del(&page->lru);
3626                 if (add_to_page_cache_lru(page, mapping,
3627                                         page->index, GFP_NOFS)) {
3628                         page_cache_release(page);
3629                         continue;
3630                 }
3631
3632                 pagepool[nr++] = page;
3633                 if (nr < ARRAY_SIZE(pagepool))
3634                         continue;
3635                 for (i = 0; i < nr; i++) {
3636                         __extent_read_full_page(tree, pagepool[i], get_extent,
3637                                         &bio, 0, &bio_flags);
3638                         page_cache_release(pagepool[i]);
3639                 }
3640                 nr = 0;
3641         }
3642         for (i = 0; i < nr; i++) {
3643                 __extent_read_full_page(tree, pagepool[i], get_extent,
3644                                         &bio, 0, &bio_flags);
3645                 page_cache_release(pagepool[i]);
3646         }
3647
3648         BUG_ON(!list_empty(pages));
3649         if (bio)
3650                 return submit_one_bio(READ, bio, 0, bio_flags);
3651         return 0;
3652 }
3653
3654 /*
3655  * basic invalidatepage code, this waits on any locked or writeback
3656  * ranges corresponding to the page, and then deletes any extent state
3657  * records from the tree
3658  */
3659 int extent_invalidatepage(struct extent_io_tree *tree,
3660                           struct page *page, unsigned long offset)
3661 {
3662         struct extent_state *cached_state = NULL;
3663         u64 start = page_offset(page);
3664         u64 end = start + PAGE_CACHE_SIZE - 1;
3665         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3666
3667         start += ALIGN(offset, blocksize);
3668         if (start > end)
3669                 return 0;
3670
3671         lock_extent_bits(tree, start, end, 0, &cached_state);
3672         wait_on_page_writeback(page);
3673         clear_extent_bit(tree, start, end,
3674                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3675                          EXTENT_DO_ACCOUNTING,
3676                          1, 1, &cached_state, GFP_NOFS);
3677         return 0;
3678 }
3679
3680 /*
3681  * a helper for releasepage, this tests for areas of the page that
3682  * are locked or under IO and drops the related state bits if it is safe
3683  * to drop the page.
3684  */
3685 int try_release_extent_state(struct extent_map_tree *map,
3686                              struct extent_io_tree *tree, struct page *page,
3687                              gfp_t mask)
3688 {
3689         u64 start = page_offset(page);
3690         u64 end = start + PAGE_CACHE_SIZE - 1;
3691         int ret = 1;
3692
3693         if (test_range_bit(tree, start, end,
3694                            EXTENT_IOBITS, 0, NULL))
3695                 ret = 0;
3696         else {
3697                 if ((mask & GFP_NOFS) == GFP_NOFS)
3698                         mask = GFP_NOFS;
3699                 /*
3700                  * at this point we can safely clear everything except the
3701                  * locked bit and the nodatasum bit
3702                  */
3703                 ret = clear_extent_bit(tree, start, end,
3704                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
3705                                  0, 0, NULL, mask);
3706
3707                 /* if clear_extent_bit failed for enomem reasons,
3708                  * we can't allow the release to continue.
3709                  */
3710                 if (ret < 0)
3711                         ret = 0;
3712                 else
3713                         ret = 1;
3714         }
3715         return ret;
3716 }
3717
3718 /*
3719  * a helper for releasepage.  As long as there are no locked extents
3720  * in the range corresponding to the page, both state records and extent
3721  * map records are removed
3722  */
3723 int try_release_extent_mapping(struct extent_map_tree *map,
3724                                struct extent_io_tree *tree, struct page *page,
3725                                gfp_t mask)
3726 {
3727         struct extent_map *em;
3728         u64 start = page_offset(page);
3729         u64 end = start + PAGE_CACHE_SIZE - 1;
3730
3731         if ((mask & __GFP_WAIT) &&
3732             page->mapping->host->i_size > 16 * 1024 * 1024) {
3733                 u64 len;
3734                 while (start <= end) {
3735                         len = end - start + 1;
3736                         write_lock(&map->lock);
3737                         em = lookup_extent_mapping(map, start, len);
3738                         if (!em) {
3739                                 write_unlock(&map->lock);
3740                                 break;
3741                         }
3742                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3743                             em->start != start) {
3744                                 write_unlock(&map->lock);
3745                                 free_extent_map(em);
3746                                 break;
3747                         }
3748                         if (!test_range_bit(tree, em->start,
3749                                             extent_map_end(em) - 1,
3750                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
3751                                             0, NULL)) {
3752                                 remove_extent_mapping(map, em);
3753                                 /* once for the rb tree */
3754                                 free_extent_map(em);
3755                         }
3756                         start = extent_map_end(em);
3757                         write_unlock(&map->lock);
3758
3759                         /* once for us */
3760                         free_extent_map(em);
3761                 }
3762         }
3763         return try_release_extent_state(map, tree, page, mask);
3764 }
3765
3766 /*
3767  * helper function for fiemap, which doesn't want to see any holes.
3768  * This maps until we find something past 'last'
3769  */
3770 static struct extent_map *get_extent_skip_holes(struct inode *inode,
3771                                                 u64 offset,
3772                                                 u64 last,
3773                                                 get_extent_t *get_extent)
3774 {
3775         u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
3776         struct extent_map *em;
3777         u64 len;
3778
3779         if (offset >= last)
3780                 return NULL;
3781
3782         while(1) {
3783                 len = last - offset;
3784                 if (len == 0)
3785                         break;
3786                 len = ALIGN(len, sectorsize);
3787                 em = get_extent(inode, NULL, 0, offset, len, 0);
3788                 if (IS_ERR_OR_NULL(em))
3789                         return em;
3790
3791                 /* if this isn't a hole return it */
3792                 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
3793                     em->block_start != EXTENT_MAP_HOLE) {
3794                         return em;
3795                 }
3796
3797                 /* this is a hole, advance to the next extent */
3798                 offset = extent_map_end(em);
3799                 free_extent_map(em);
3800                 if (offset >= last)
3801                         break;
3802         }
3803         return NULL;
3804 }
3805
3806 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
3807                 __u64 start, __u64 len, get_extent_t *get_extent)
3808 {
3809         int ret = 0;
3810         u64 off = start;
3811         u64 max = start + len;
3812         u32 flags = 0;
3813         u32 found_type;
3814         u64 last;
3815         u64 last_for_get_extent = 0;
3816         u64 disko = 0;
3817         u64 isize = i_size_read(inode);
3818         struct btrfs_key found_key;
3819         struct extent_map *em = NULL;
3820         struct extent_state *cached_state = NULL;
3821         struct btrfs_path *path;
3822         struct btrfs_file_extent_item *item;
3823         int end = 0;
3824         u64 em_start = 0;
3825         u64 em_len = 0;
3826         u64 em_end = 0;
3827         unsigned long emflags;
3828
3829         if (len == 0)
3830                 return -EINVAL;
3831
3832         path = btrfs_alloc_path();
3833         if (!path)
3834                 return -ENOMEM;
3835         path->leave_spinning = 1;
3836
3837         start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
3838         len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
3839
3840         /*
3841          * lookup the last file extent.  We're not using i_size here
3842          * because there might be preallocation past i_size
3843          */
3844         ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
3845                                        path, btrfs_ino(inode), -1, 0);
3846         if (ret < 0) {
3847                 btrfs_free_path(path);
3848                 return ret;
3849         }
3850         WARN_ON(!ret);
3851         path->slots[0]--;
3852         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3853                               struct btrfs_file_extent_item);
3854         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
3855         found_type = btrfs_key_type(&found_key);
3856
3857         /* No extents, but there might be delalloc bits */
3858         if (found_key.objectid != btrfs_ino(inode) ||
3859             found_type != BTRFS_EXTENT_DATA_KEY) {
3860                 /* have to trust i_size as the end */
3861                 last = (u64)-1;
3862                 last_for_get_extent = isize;
3863         } else {
3864                 /*
3865                  * remember the start of the last extent.  There are a
3866                  * bunch of different factors that go into the length of the
3867                  * extent, so its much less complex to remember where it started
3868                  */
3869                 last = found_key.offset;
3870                 last_for_get_extent = last + 1;
3871         }
3872         btrfs_free_path(path);
3873
3874         /*
3875          * we might have some extents allocated but more delalloc past those
3876          * extents.  so, we trust isize unless the start of the last extent is
3877          * beyond isize
3878          */
3879         if (last < isize) {
3880                 last = (u64)-1;
3881                 last_for_get_extent = isize;
3882         }
3883
3884         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
3885                          &cached_state);
3886
3887         em = get_extent_skip_holes(inode, start, last_for_get_extent,
3888                                    get_extent);
3889         if (!em)
3890                 goto out;
3891         if (IS_ERR(em)) {
3892                 ret = PTR_ERR(em);
3893                 goto out;
3894         }
3895
3896         while (!end) {
3897                 u64 offset_in_extent;
3898
3899                 /* break if the extent we found is outside the range */
3900                 if (em->start >= max || extent_map_end(em) < off)
3901                         break;
3902
3903                 /*
3904                  * get_extent may return an extent that starts before our
3905                  * requested range.  We have to make sure the ranges
3906                  * we return to fiemap always move forward and don't
3907                  * overlap, so adjust the offsets here
3908                  */
3909                 em_start = max(em->start, off);
3910
3911                 /*
3912                  * record the offset from the start of the extent
3913                  * for adjusting the disk offset below
3914                  */
3915                 offset_in_extent = em_start - em->start;
3916                 em_end = extent_map_end(em);
3917                 em_len = em_end - em_start;
3918                 emflags = em->flags;
3919                 disko = 0;
3920                 flags = 0;
3921
3922                 /*
3923                  * bump off for our next call to get_extent
3924                  */
3925                 off = extent_map_end(em);
3926                 if (off >= max)
3927                         end = 1;
3928
3929                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
3930                         end = 1;
3931                         flags |= FIEMAP_EXTENT_LAST;
3932                 } else if (em->block_start == EXTENT_MAP_INLINE) {
3933                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
3934                                   FIEMAP_EXTENT_NOT_ALIGNED);
3935                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
3936                         flags |= (FIEMAP_EXTENT_DELALLOC |
3937                                   FIEMAP_EXTENT_UNKNOWN);
3938                 } else {
3939                         disko = em->block_start + offset_in_extent;
3940                 }
3941                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
3942                         flags |= FIEMAP_EXTENT_ENCODED;
3943
3944                 free_extent_map(em);
3945                 em = NULL;
3946                 if ((em_start >= last) || em_len == (u64)-1 ||
3947                    (last == (u64)-1 && isize <= em_end)) {
3948                         flags |= FIEMAP_EXTENT_LAST;
3949                         end = 1;
3950                 }
3951
3952                 /* now scan forward to see if this is really the last extent. */
3953                 em = get_extent_skip_holes(inode, off, last_for_get_extent,
3954                                            get_extent);
3955                 if (IS_ERR(em)) {
3956                         ret = PTR_ERR(em);
3957                         goto out;
3958                 }
3959                 if (!em) {
3960                         flags |= FIEMAP_EXTENT_LAST;
3961                         end = 1;
3962                 }
3963                 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
3964                                               em_len, flags);
3965                 if (ret)
3966                         goto out_free;
3967         }
3968 out_free:
3969         free_extent_map(em);
3970 out:
3971         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
3972                              &cached_state, GFP_NOFS);
3973         return ret;
3974 }
3975
3976 static void __free_extent_buffer(struct extent_buffer *eb)
3977 {
3978 #if LEAK_DEBUG
3979         unsigned long flags;
3980         spin_lock_irqsave(&leak_lock, flags);
3981         list_del(&eb->leak_list);
3982         spin_unlock_irqrestore(&leak_lock, flags);
3983 #endif
3984         if (eb->pages && eb->pages != eb->inline_pages)
3985                 kfree(eb->pages);
3986         kmem_cache_free(extent_buffer_cache, eb);
3987 }
3988
3989 static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
3990                                                    u64 start,
3991                                                    unsigned long len,
3992                                                    gfp_t mask)
3993 {
3994         struct extent_buffer *eb = NULL;
3995 #if LEAK_DEBUG
3996         unsigned long flags;
3997 #endif
3998
3999         eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4000         if (eb == NULL)
4001                 return NULL;
4002         eb->start = start;
4003         eb->len = len;
4004         eb->tree = tree;
4005         eb->bflags = 0;
4006         rwlock_init(&eb->lock);
4007         atomic_set(&eb->write_locks, 0);
4008         atomic_set(&eb->read_locks, 0);
4009         atomic_set(&eb->blocking_readers, 0);
4010         atomic_set(&eb->blocking_writers, 0);
4011         atomic_set(&eb->spinning_readers, 0);
4012         atomic_set(&eb->spinning_writers, 0);
4013         eb->lock_nested = 0;
4014         init_waitqueue_head(&eb->write_lock_wq);
4015         init_waitqueue_head(&eb->read_lock_wq);
4016
4017 #if LEAK_DEBUG
4018         spin_lock_irqsave(&leak_lock, flags);
4019         list_add(&eb->leak_list, &buffers);
4020         spin_unlock_irqrestore(&leak_lock, flags);
4021 #endif
4022         spin_lock_init(&eb->refs_lock);
4023         atomic_set(&eb->refs, 1);
4024         atomic_set(&eb->io_pages, 0);
4025
4026         if (len > MAX_INLINE_EXTENT_BUFFER_SIZE) {
4027                 struct page **pages;
4028                 int num_pages = (len + PAGE_CACHE_SIZE - 1) >>
4029                         PAGE_CACHE_SHIFT;
4030                 pages = kzalloc(num_pages, mask);
4031                 if (!pages) {
4032                         __free_extent_buffer(eb);
4033                         return NULL;
4034                 }
4035                 eb->pages = pages;
4036         } else {
4037                 eb->pages = eb->inline_pages;
4038         }
4039
4040         return eb;
4041 }
4042
4043 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4044 {
4045         unsigned long i;
4046         struct page *p;
4047         struct extent_buffer *new;
4048         unsigned long num_pages = num_extent_pages(src->start, src->len);
4049
4050         new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_ATOMIC);
4051         if (new == NULL)
4052                 return NULL;
4053
4054         for (i = 0; i < num_pages; i++) {
4055                 p = alloc_page(GFP_ATOMIC);
4056                 BUG_ON(!p);
4057                 attach_extent_buffer_page(new, p);
4058                 WARN_ON(PageDirty(p));
4059                 SetPageUptodate(p);
4060                 new->pages[i] = p;
4061         }
4062
4063         copy_extent_buffer(new, src, 0, 0, src->len);
4064         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4065         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4066
4067         return new;
4068 }
4069
4070 struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4071 {
4072         struct extent_buffer *eb;
4073         unsigned long num_pages = num_extent_pages(0, len);
4074         unsigned long i;
4075
4076         eb = __alloc_extent_buffer(NULL, start, len, GFP_ATOMIC);
4077         if (!eb)
4078                 return NULL;
4079
4080         for (i = 0; i < num_pages; i++) {
4081                 eb->pages[i] = alloc_page(GFP_ATOMIC);
4082                 if (!eb->pages[i])
4083                         goto err;
4084         }
4085         set_extent_buffer_uptodate(eb);
4086         btrfs_set_header_nritems(eb, 0);
4087         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4088
4089         return eb;
4090 err:
4091         for (; i > 0; i--)
4092                 __free_page(eb->pages[i - 1]);
4093         __free_extent_buffer(eb);
4094         return NULL;
4095 }
4096
4097 static int extent_buffer_under_io(struct extent_buffer *eb)
4098 {
4099         return (atomic_read(&eb->io_pages) ||
4100                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4101                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4102 }
4103
4104 /*
4105  * Helper for releasing extent buffer page.
4106  */
4107 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4108                                                 unsigned long start_idx)
4109 {
4110         unsigned long index;
4111         unsigned long num_pages;
4112         struct page *page;
4113         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4114
4115         BUG_ON(extent_buffer_under_io(eb));
4116
4117         num_pages = num_extent_pages(eb->start, eb->len);
4118         index = start_idx + num_pages;
4119         if (start_idx >= index)
4120                 return;
4121
4122         do {
4123                 index--;
4124                 page = extent_buffer_page(eb, index);
4125                 if (page && mapped) {
4126                         spin_lock(&page->mapping->private_lock);
4127                         /*
4128                          * We do this since we'll remove the pages after we've
4129                          * removed the eb from the radix tree, so we could race
4130                          * and have this page now attached to the new eb.  So
4131                          * only clear page_private if it's still connected to
4132                          * this eb.
4133                          */
4134                         if (PagePrivate(page) &&
4135                             page->private == (unsigned long)eb) {
4136                                 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4137                                 BUG_ON(PageDirty(page));
4138                                 BUG_ON(PageWriteback(page));
4139                                 /*
4140                                  * We need to make sure we haven't be attached
4141                                  * to a new eb.
4142                                  */
4143                                 ClearPagePrivate(page);
4144                                 set_page_private(page, 0);
4145                                 /* One for the page private */
4146                                 page_cache_release(page);
4147                         }
4148                         spin_unlock(&page->mapping->private_lock);
4149
4150                 }
4151                 if (page) {
4152                         /* One for when we alloced the page */
4153                         page_cache_release(page);
4154                 }
4155         } while (index != start_idx);
4156 }
4157
4158 /*
4159  * Helper for releasing the extent buffer.
4160  */
4161 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4162 {
4163         btrfs_release_extent_buffer_page(eb, 0);
4164         __free_extent_buffer(eb);
4165 }
4166
4167 static void check_buffer_tree_ref(struct extent_buffer *eb)
4168 {
4169         int refs;
4170         /* the ref bit is tricky.  We have to make sure it is set
4171          * if we have the buffer dirty.   Otherwise the
4172          * code to free a buffer can end up dropping a dirty
4173          * page
4174          *
4175          * Once the ref bit is set, it won't go away while the
4176          * buffer is dirty or in writeback, and it also won't
4177          * go away while we have the reference count on the
4178          * eb bumped.
4179          *
4180          * We can't just set the ref bit without bumping the
4181          * ref on the eb because free_extent_buffer might
4182          * see the ref bit and try to clear it.  If this happens
4183          * free_extent_buffer might end up dropping our original
4184          * ref by mistake and freeing the page before we are able
4185          * to add one more ref.
4186          *
4187          * So bump the ref count first, then set the bit.  If someone
4188          * beat us to it, drop the ref we added.
4189          */
4190         refs = atomic_read(&eb->refs);
4191         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4192                 return;
4193
4194         spin_lock(&eb->refs_lock);
4195         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4196                 atomic_inc(&eb->refs);
4197         spin_unlock(&eb->refs_lock);
4198 }
4199
4200 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4201 {
4202         unsigned long num_pages, i;
4203
4204         check_buffer_tree_ref(eb);
4205
4206         num_pages = num_extent_pages(eb->start, eb->len);
4207         for (i = 0; i < num_pages; i++) {
4208                 struct page *p = extent_buffer_page(eb, i);
4209                 mark_page_accessed(p);
4210         }
4211 }
4212
4213 struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
4214                                           u64 start, unsigned long len)
4215 {
4216         unsigned long num_pages = num_extent_pages(start, len);
4217         unsigned long i;
4218         unsigned long index = start >> PAGE_CACHE_SHIFT;
4219         struct extent_buffer *eb;
4220         struct extent_buffer *exists = NULL;
4221         struct page *p;
4222         struct address_space *mapping = tree->mapping;
4223         int uptodate = 1;
4224         int ret;
4225
4226         rcu_read_lock();
4227         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4228         if (eb && atomic_inc_not_zero(&eb->refs)) {
4229                 rcu_read_unlock();
4230                 mark_extent_buffer_accessed(eb);
4231                 return eb;
4232         }
4233         rcu_read_unlock();
4234
4235         eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
4236         if (!eb)
4237                 return NULL;
4238
4239         for (i = 0; i < num_pages; i++, index++) {
4240                 p = find_or_create_page(mapping, index, GFP_NOFS);
4241                 if (!p)
4242                         goto free_eb;
4243
4244                 spin_lock(&mapping->private_lock);
4245                 if (PagePrivate(p)) {
4246                         /*
4247                          * We could have already allocated an eb for this page
4248                          * and attached one so lets see if we can get a ref on
4249                          * the existing eb, and if we can we know it's good and
4250                          * we can just return that one, else we know we can just
4251                          * overwrite page->private.
4252                          */
4253                         exists = (struct extent_buffer *)p->private;
4254                         if (atomic_inc_not_zero(&exists->refs)) {
4255                                 spin_unlock(&mapping->private_lock);
4256                                 unlock_page(p);
4257                                 page_cache_release(p);
4258                                 mark_extent_buffer_accessed(exists);
4259                                 goto free_eb;
4260                         }
4261
4262                         /*
4263                          * Do this so attach doesn't complain and we need to
4264                          * drop the ref the old guy had.
4265                          */
4266                         ClearPagePrivate(p);
4267                         WARN_ON(PageDirty(p));
4268                         page_cache_release(p);
4269                 }
4270                 attach_extent_buffer_page(eb, p);
4271                 spin_unlock(&mapping->private_lock);
4272                 WARN_ON(PageDirty(p));
4273                 mark_page_accessed(p);
4274                 eb->pages[i] = p;
4275                 if (!PageUptodate(p))
4276                         uptodate = 0;
4277
4278                 /*
4279                  * see below about how we avoid a nasty race with release page
4280                  * and why we unlock later
4281                  */
4282         }
4283         if (uptodate)
4284                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4285 again:
4286         ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4287         if (ret)
4288                 goto free_eb;
4289
4290         spin_lock(&tree->buffer_lock);
4291         ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
4292         if (ret == -EEXIST) {
4293                 exists = radix_tree_lookup(&tree->buffer,
4294                                                 start >> PAGE_CACHE_SHIFT);
4295                 if (!atomic_inc_not_zero(&exists->refs)) {
4296                         spin_unlock(&tree->buffer_lock);
4297                         radix_tree_preload_end();
4298                         exists = NULL;
4299                         goto again;
4300                 }
4301                 spin_unlock(&tree->buffer_lock);
4302                 radix_tree_preload_end();
4303                 mark_extent_buffer_accessed(exists);
4304                 goto free_eb;
4305         }
4306         /* add one reference for the tree */
4307         check_buffer_tree_ref(eb);
4308         spin_unlock(&tree->buffer_lock);
4309         radix_tree_preload_end();
4310
4311         /*
4312          * there is a race where release page may have
4313          * tried to find this extent buffer in the radix
4314          * but failed.  It will tell the VM it is safe to
4315          * reclaim the, and it will clear the page private bit.
4316          * We must make sure to set the page private bit properly
4317          * after the extent buffer is in the radix tree so
4318          * it doesn't get lost
4319          */
4320         SetPageChecked(eb->pages[0]);
4321         for (i = 1; i < num_pages; i++) {
4322                 p = extent_buffer_page(eb, i);
4323                 ClearPageChecked(p);
4324                 unlock_page(p);
4325         }
4326         unlock_page(eb->pages[0]);
4327         return eb;
4328
4329 free_eb:
4330         for (i = 0; i < num_pages; i++) {
4331                 if (eb->pages[i])
4332                         unlock_page(eb->pages[i]);
4333         }
4334
4335         WARN_ON(!atomic_dec_and_test(&eb->refs));
4336         btrfs_release_extent_buffer(eb);
4337         return exists;
4338 }
4339
4340 struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
4341                                          u64 start, unsigned long len)
4342 {
4343         struct extent_buffer *eb;
4344
4345         rcu_read_lock();
4346         eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
4347         if (eb && atomic_inc_not_zero(&eb->refs)) {
4348                 rcu_read_unlock();
4349                 mark_extent_buffer_accessed(eb);
4350                 return eb;
4351         }
4352         rcu_read_unlock();
4353
4354         return NULL;
4355 }
4356
4357 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4358 {
4359         struct extent_buffer *eb =
4360                         container_of(head, struct extent_buffer, rcu_head);
4361
4362         __free_extent_buffer(eb);
4363 }
4364
4365 /* Expects to have eb->eb_lock already held */
4366 static int release_extent_buffer(struct extent_buffer *eb, gfp_t mask)
4367 {
4368         WARN_ON(atomic_read(&eb->refs) == 0);
4369         if (atomic_dec_and_test(&eb->refs)) {
4370                 if (test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags)) {
4371                         spin_unlock(&eb->refs_lock);
4372                 } else {
4373                         struct extent_io_tree *tree = eb->tree;
4374
4375                         spin_unlock(&eb->refs_lock);
4376
4377                         spin_lock(&tree->buffer_lock);
4378                         radix_tree_delete(&tree->buffer,
4379                                           eb->start >> PAGE_CACHE_SHIFT);
4380                         spin_unlock(&tree->buffer_lock);
4381                 }
4382
4383                 /* Should be safe to release our pages at this point */
4384                 btrfs_release_extent_buffer_page(eb, 0);
4385                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4386                 return 1;
4387         }
4388         spin_unlock(&eb->refs_lock);
4389
4390         return 0;
4391 }
4392
4393 void free_extent_buffer(struct extent_buffer *eb)
4394 {
4395         int refs;
4396         int old;
4397         if (!eb)
4398                 return;
4399
4400         while (1) {
4401                 refs = atomic_read(&eb->refs);
4402                 if (refs <= 3)
4403                         break;
4404                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4405                 if (old == refs)
4406                         return;
4407         }
4408
4409         spin_lock(&eb->refs_lock);
4410         if (atomic_read(&eb->refs) == 2 &&
4411             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4412                 atomic_dec(&eb->refs);
4413
4414         if (atomic_read(&eb->refs) == 2 &&
4415             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4416             !extent_buffer_under_io(eb) &&
4417             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4418                 atomic_dec(&eb->refs);
4419
4420         /*
4421          * I know this is terrible, but it's temporary until we stop tracking
4422          * the uptodate bits and such for the extent buffers.
4423          */
4424         release_extent_buffer(eb, GFP_ATOMIC);
4425 }
4426
4427 void free_extent_buffer_stale(struct extent_buffer *eb)
4428 {
4429         if (!eb)
4430                 return;
4431
4432         spin_lock(&eb->refs_lock);
4433         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4434
4435         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4436             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4437                 atomic_dec(&eb->refs);
4438         release_extent_buffer(eb, GFP_NOFS);
4439 }
4440
4441 void clear_extent_buffer_dirty(struct extent_buffer *eb)
4442 {
4443         unsigned long i;
4444         unsigned long num_pages;
4445         struct page *page;
4446
4447         num_pages = num_extent_pages(eb->start, eb->len);
4448
4449         for (i = 0; i < num_pages; i++) {
4450                 page = extent_buffer_page(eb, i);
4451                 if (!PageDirty(page))
4452                         continue;
4453
4454                 lock_page(page);
4455                 WARN_ON(!PagePrivate(page));
4456
4457                 clear_page_dirty_for_io(page);
4458                 spin_lock_irq(&page->mapping->tree_lock);
4459                 if (!PageDirty(page)) {
4460                         radix_tree_tag_clear(&page->mapping->page_tree,
4461                                                 page_index(page),
4462                                                 PAGECACHE_TAG_DIRTY);
4463                 }
4464                 spin_unlock_irq(&page->mapping->tree_lock);
4465                 ClearPageError(page);
4466                 unlock_page(page);
4467         }
4468         WARN_ON(atomic_read(&eb->refs) == 0);
4469 }
4470
4471 int set_extent_buffer_dirty(struct extent_buffer *eb)
4472 {
4473         unsigned long i;
4474         unsigned long num_pages;
4475         int was_dirty = 0;
4476
4477         check_buffer_tree_ref(eb);
4478
4479         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4480
4481         num_pages = num_extent_pages(eb->start, eb->len);
4482         WARN_ON(atomic_read(&eb->refs) == 0);
4483         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4484
4485         for (i = 0; i < num_pages; i++)
4486                 set_page_dirty(extent_buffer_page(eb, i));
4487         return was_dirty;
4488 }
4489
4490 static int range_straddles_pages(u64 start, u64 len)
4491 {
4492         if (len < PAGE_CACHE_SIZE)
4493                 return 1;
4494         if (start & (PAGE_CACHE_SIZE - 1))
4495                 return 1;
4496         if ((start + len) & (PAGE_CACHE_SIZE - 1))
4497                 return 1;
4498         return 0;
4499 }
4500
4501 int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4502 {
4503         unsigned long i;
4504         struct page *page;
4505         unsigned long num_pages;
4506
4507         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4508         num_pages = num_extent_pages(eb->start, eb->len);
4509         for (i = 0; i < num_pages; i++) {
4510                 page = extent_buffer_page(eb, i);
4511                 if (page)
4512                         ClearPageUptodate(page);
4513         }
4514         return 0;
4515 }
4516
4517 int set_extent_buffer_uptodate(struct extent_buffer *eb)
4518 {
4519         unsigned long i;
4520         struct page *page;
4521         unsigned long num_pages;
4522
4523         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4524         num_pages = num_extent_pages(eb->start, eb->len);
4525         for (i = 0; i < num_pages; i++) {
4526                 page = extent_buffer_page(eb, i);
4527                 SetPageUptodate(page);
4528         }
4529         return 0;
4530 }
4531
4532 int extent_range_uptodate(struct extent_io_tree *tree,
4533                           u64 start, u64 end)
4534 {
4535         struct page *page;
4536         int ret;
4537         int pg_uptodate = 1;
4538         int uptodate;
4539         unsigned long index;
4540
4541         if (range_straddles_pages(start, end - start + 1)) {
4542                 ret = test_range_bit(tree, start, end,
4543                                      EXTENT_UPTODATE, 1, NULL);
4544                 if (ret)
4545                         return 1;
4546         }
4547         while (start <= end) {
4548                 index = start >> PAGE_CACHE_SHIFT;
4549                 page = find_get_page(tree->mapping, index);
4550                 if (!page)
4551                         return 1;
4552                 uptodate = PageUptodate(page);
4553                 page_cache_release(page);
4554                 if (!uptodate) {
4555                         pg_uptodate = 0;
4556                         break;
4557                 }
4558                 start += PAGE_CACHE_SIZE;
4559         }
4560         return pg_uptodate;
4561 }
4562
4563 int extent_buffer_uptodate(struct extent_buffer *eb)
4564 {
4565         return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4566 }
4567
4568 int read_extent_buffer_pages(struct extent_io_tree *tree,
4569                              struct extent_buffer *eb, u64 start, int wait,
4570                              get_extent_t *get_extent, int mirror_num)
4571 {
4572         unsigned long i;
4573         unsigned long start_i;
4574         struct page *page;
4575         int err;
4576         int ret = 0;
4577         int locked_pages = 0;
4578         int all_uptodate = 1;
4579         unsigned long num_pages;
4580         unsigned long num_reads = 0;
4581         struct bio *bio = NULL;
4582         unsigned long bio_flags = 0;
4583
4584         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4585                 return 0;
4586
4587         if (start) {
4588                 WARN_ON(start < eb->start);
4589                 start_i = (start >> PAGE_CACHE_SHIFT) -
4590                         (eb->start >> PAGE_CACHE_SHIFT);
4591         } else {
4592                 start_i = 0;
4593         }
4594
4595         num_pages = num_extent_pages(eb->start, eb->len);
4596         for (i = start_i; i < num_pages; i++) {
4597                 page = extent_buffer_page(eb, i);
4598                 if (wait == WAIT_NONE) {
4599                         if (!trylock_page(page))
4600                                 goto unlock_exit;
4601                 } else {
4602                         lock_page(page);
4603                 }
4604                 locked_pages++;
4605                 if (!PageUptodate(page)) {
4606                         num_reads++;
4607                         all_uptodate = 0;
4608                 }
4609         }
4610         if (all_uptodate) {
4611                 if (start_i == 0)
4612                         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4613                 goto unlock_exit;
4614         }
4615
4616         clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4617         eb->read_mirror = 0;
4618         atomic_set(&eb->io_pages, num_reads);
4619         for (i = start_i; i < num_pages; i++) {
4620                 page = extent_buffer_page(eb, i);
4621                 if (!PageUptodate(page)) {
4622                         ClearPageError(page);
4623                         err = __extent_read_full_page(tree, page,
4624                                                       get_extent, &bio,
4625                                                       mirror_num, &bio_flags);
4626                         if (err)
4627                                 ret = err;
4628                 } else {
4629                         unlock_page(page);
4630                 }
4631         }
4632
4633         if (bio) {
4634                 err = submit_one_bio(READ, bio, mirror_num, bio_flags);
4635                 if (err)
4636                         return err;
4637         }
4638
4639         if (ret || wait != WAIT_COMPLETE)
4640                 return ret;
4641
4642         for (i = start_i; i < num_pages; i++) {
4643                 page = extent_buffer_page(eb, i);
4644                 wait_on_page_locked(page);
4645                 if (!PageUptodate(page))
4646                         ret = -EIO;
4647         }
4648
4649         return ret;
4650
4651 unlock_exit:
4652         i = start_i;
4653         while (locked_pages > 0) {
4654                 page = extent_buffer_page(eb, i);
4655                 i++;
4656                 unlock_page(page);
4657                 locked_pages--;
4658         }
4659         return ret;
4660 }
4661
4662 void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4663                         unsigned long start,
4664                         unsigned long len)
4665 {
4666         size_t cur;
4667         size_t offset;
4668         struct page *page;
4669         char *kaddr;
4670         char *dst = (char *)dstv;
4671         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4672         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4673
4674         WARN_ON(start > eb->len);
4675         WARN_ON(start + len > eb->start + eb->len);
4676
4677         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4678
4679         while (len > 0) {
4680                 page = extent_buffer_page(eb, i);
4681
4682                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4683                 kaddr = page_address(page);
4684                 memcpy(dst, kaddr + offset, cur);
4685
4686                 dst += cur;
4687                 len -= cur;
4688                 offset = 0;
4689                 i++;
4690         }
4691 }
4692
4693 int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4694                                unsigned long min_len, char **map,
4695                                unsigned long *map_start,
4696                                unsigned long *map_len)
4697 {
4698         size_t offset = start & (PAGE_CACHE_SIZE - 1);
4699         char *kaddr;
4700         struct page *p;
4701         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4702         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4703         unsigned long end_i = (start_offset + start + min_len - 1) >>
4704                 PAGE_CACHE_SHIFT;
4705
4706         if (i != end_i)
4707                 return -EINVAL;
4708
4709         if (i == 0) {
4710                 offset = start_offset;
4711                 *map_start = 0;
4712         } else {
4713                 offset = 0;
4714                 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4715         }
4716
4717         if (start + min_len > eb->len) {
4718                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4719                        "wanted %lu %lu\n", (unsigned long long)eb->start,
4720                        eb->len, start, min_len);
4721                 return -EINVAL;
4722         }
4723
4724         p = extent_buffer_page(eb, i);
4725         kaddr = page_address(p);
4726         *map = kaddr + offset;
4727         *map_len = PAGE_CACHE_SIZE - offset;
4728         return 0;
4729 }
4730
4731 int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4732                           unsigned long start,
4733                           unsigned long len)
4734 {
4735         size_t cur;
4736         size_t offset;
4737         struct page *page;
4738         char *kaddr;
4739         char *ptr = (char *)ptrv;
4740         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4741         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4742         int ret = 0;
4743
4744         WARN_ON(start > eb->len);
4745         WARN_ON(start + len > eb->start + eb->len);
4746
4747         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4748
4749         while (len > 0) {
4750                 page = extent_buffer_page(eb, i);
4751
4752                 cur = min(len, (PAGE_CACHE_SIZE - offset));
4753
4754                 kaddr = page_address(page);
4755                 ret = memcmp(ptr, kaddr + offset, cur);
4756                 if (ret)
4757                         break;
4758
4759                 ptr += cur;
4760                 len -= cur;
4761                 offset = 0;
4762                 i++;
4763         }
4764         return ret;
4765 }
4766
4767 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
4768                          unsigned long start, unsigned long len)
4769 {
4770         size_t cur;
4771         size_t offset;
4772         struct page *page;
4773         char *kaddr;
4774         char *src = (char *)srcv;
4775         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4776         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4777
4778         WARN_ON(start > eb->len);
4779         WARN_ON(start + len > eb->start + eb->len);
4780
4781         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4782
4783         while (len > 0) {
4784                 page = extent_buffer_page(eb, i);
4785                 WARN_ON(!PageUptodate(page));
4786
4787                 cur = min(len, PAGE_CACHE_SIZE - offset);
4788                 kaddr = page_address(page);
4789                 memcpy(kaddr + offset, src, cur);
4790
4791                 src += cur;
4792                 len -= cur;
4793                 offset = 0;
4794                 i++;
4795         }
4796 }
4797
4798 void memset_extent_buffer(struct extent_buffer *eb, char c,
4799                           unsigned long start, unsigned long len)
4800 {
4801         size_t cur;
4802         size_t offset;
4803         struct page *page;
4804         char *kaddr;
4805         size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4806         unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4807
4808         WARN_ON(start > eb->len);
4809         WARN_ON(start + len > eb->start + eb->len);
4810
4811         offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
4812
4813         while (len > 0) {
4814                 page = extent_buffer_page(eb, i);
4815                 WARN_ON(!PageUptodate(page));
4816
4817                 cur = min(len, PAGE_CACHE_SIZE - offset);
4818                 kaddr = page_address(page);
4819                 memset(kaddr + offset, c, cur);
4820
4821                 len -= cur;
4822                 offset = 0;
4823                 i++;
4824         }
4825 }
4826
4827 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
4828                         unsigned long dst_offset, unsigned long src_offset,
4829                         unsigned long len)
4830 {
4831         u64 dst_len = dst->len;
4832         size_t cur;
4833         size_t offset;
4834         struct page *page;
4835         char *kaddr;
4836         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4837         unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4838
4839         WARN_ON(src->len != dst_len);
4840
4841         offset = (start_offset + dst_offset) &
4842                 ((unsigned long)PAGE_CACHE_SIZE - 1);
4843
4844         while (len > 0) {
4845                 page = extent_buffer_page(dst, i);
4846                 WARN_ON(!PageUptodate(page));
4847
4848                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
4849
4850                 kaddr = page_address(page);
4851                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4852
4853                 src_offset += cur;
4854                 len -= cur;
4855                 offset = 0;
4856                 i++;
4857         }
4858 }
4859
4860 static void move_pages(struct page *dst_page, struct page *src_page,
4861                        unsigned long dst_off, unsigned long src_off,
4862                        unsigned long len)
4863 {
4864         char *dst_kaddr = page_address(dst_page);
4865         if (dst_page == src_page) {
4866                 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
4867         } else {
4868                 char *src_kaddr = page_address(src_page);
4869                 char *p = dst_kaddr + dst_off + len;
4870                 char *s = src_kaddr + src_off + len;
4871
4872                 while (len--)
4873                         *--p = *--s;
4874         }
4875 }
4876
4877 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4878 {
4879         unsigned long distance = (src > dst) ? src - dst : dst - src;
4880         return distance < len;
4881 }
4882
4883 static void copy_pages(struct page *dst_page, struct page *src_page,
4884                        unsigned long dst_off, unsigned long src_off,
4885                        unsigned long len)
4886 {
4887         char *dst_kaddr = page_address(dst_page);
4888         char *src_kaddr;
4889         int must_memmove = 0;
4890
4891         if (dst_page != src_page) {
4892                 src_kaddr = page_address(src_page);
4893         } else {
4894                 src_kaddr = dst_kaddr;
4895                 if (areas_overlap(src_off, dst_off, len))
4896                         must_memmove = 1;
4897         }
4898
4899         if (must_memmove)
4900                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
4901         else
4902                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
4903 }
4904
4905 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4906                            unsigned long src_offset, unsigned long len)
4907 {
4908         size_t cur;
4909         size_t dst_off_in_page;
4910         size_t src_off_in_page;
4911         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4912         unsigned long dst_i;
4913         unsigned long src_i;
4914
4915         if (src_offset + len > dst->len) {
4916                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4917                        "len %lu dst len %lu\n", src_offset, len, dst->len);
4918                 BUG_ON(1);
4919         }
4920         if (dst_offset + len > dst->len) {
4921                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4922                        "len %lu dst len %lu\n", dst_offset, len, dst->len);
4923                 BUG_ON(1);
4924         }
4925
4926         while (len > 0) {
4927                 dst_off_in_page = (start_offset + dst_offset) &
4928                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4929                 src_off_in_page = (start_offset + src_offset) &
4930                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4931
4932                 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
4933                 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
4934
4935                 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
4936                                                src_off_in_page));
4937                 cur = min_t(unsigned long, cur,
4938                         (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
4939
4940                 copy_pages(extent_buffer_page(dst, dst_i),
4941                            extent_buffer_page(dst, src_i),
4942                            dst_off_in_page, src_off_in_page, cur);
4943
4944                 src_offset += cur;
4945                 dst_offset += cur;
4946                 len -= cur;
4947         }
4948 }
4949
4950 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
4951                            unsigned long src_offset, unsigned long len)
4952 {
4953         size_t cur;
4954         size_t dst_off_in_page;
4955         size_t src_off_in_page;
4956         unsigned long dst_end = dst_offset + len - 1;
4957         unsigned long src_end = src_offset + len - 1;
4958         size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
4959         unsigned long dst_i;
4960         unsigned long src_i;
4961
4962         if (src_offset + len > dst->len) {
4963                 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
4964                        "len %lu len %lu\n", src_offset, len, dst->len);
4965                 BUG_ON(1);
4966         }
4967         if (dst_offset + len > dst->len) {
4968                 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
4969                        "len %lu len %lu\n", dst_offset, len, dst->len);
4970                 BUG_ON(1);
4971         }
4972         if (dst_offset < src_offset) {
4973                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4974                 return;
4975         }
4976         while (len > 0) {
4977                 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
4978                 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
4979
4980                 dst_off_in_page = (start_offset + dst_end) &
4981                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4982                 src_off_in_page = (start_offset + src_end) &
4983                         ((unsigned long)PAGE_CACHE_SIZE - 1);
4984
4985                 cur = min_t(unsigned long, len, src_off_in_page + 1);
4986                 cur = min(cur, dst_off_in_page + 1);
4987                 move_pages(extent_buffer_page(dst, dst_i),
4988                            extent_buffer_page(dst, src_i),
4989                            dst_off_in_page - cur + 1,
4990                            src_off_in_page - cur + 1, cur);
4991
4992                 dst_end -= cur;
4993                 src_end -= cur;
4994                 len -= cur;
4995         }
4996 }
4997
4998 int try_release_extent_buffer(struct page *page, gfp_t mask)
4999 {
5000         struct extent_buffer *eb;
5001
5002         /*
5003          * We need to make sure noboody is attaching this page to an eb right
5004          * now.
5005          */
5006         spin_lock(&page->mapping->private_lock);
5007         if (!PagePrivate(page)) {
5008                 spin_unlock(&page->mapping->private_lock);
5009                 return 1;
5010         }
5011
5012         eb = (struct extent_buffer *)page->private;
5013         BUG_ON(!eb);
5014
5015         /*
5016          * This is a little awful but should be ok, we need to make sure that
5017          * the eb doesn't disappear out from under us while we're looking at
5018          * this page.
5019          */
5020         spin_lock(&eb->refs_lock);
5021         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5022                 spin_unlock(&eb->refs_lock);
5023                 spin_unlock(&page->mapping->private_lock);
5024                 return 0;
5025         }
5026         spin_unlock(&page->mapping->private_lock);
5027
5028         if ((mask & GFP_NOFS) == GFP_NOFS)
5029                 mask = GFP_NOFS;
5030
5031         /*
5032          * If tree ref isn't set then we know the ref on this eb is a real ref,
5033          * so just return, this page will likely be freed soon anyway.
5034          */
5035         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5036                 spin_unlock(&eb->refs_lock);
5037                 return 0;
5038         }
5039
5040         return release_extent_buffer(eb, mask);
5041 }