283af7a676a39b4c2f31d7d9b7caf81607eee4ea
[linux-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36
37 /*
38  * control flags for do_chunk_alloc's force field
39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
40  * if we really need one.
41  *
42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
43  * if we have very few chunks already allocated.  This is
44  * used as part of the clustering code to help make sure
45  * we have a good pool of storage to cluster in, without
46  * filling the FS with empty chunks
47  *
48  * CHUNK_ALLOC_FORCE means it must try to allocate one
49  *
50  */
51 enum {
52         CHUNK_ALLOC_NO_FORCE = 0,
53         CHUNK_ALLOC_LIMITED = 1,
54         CHUNK_ALLOC_FORCE = 2,
55 };
56
57 /*
58  * Control how reservations are dealt with.
59  *
60  * RESERVE_FREE - freeing a reservation.
61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
62  *   ENOSPC accounting
63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
65  */
66 enum {
67         RESERVE_FREE = 0,
68         RESERVE_ALLOC = 1,
69         RESERVE_ALLOC_NO_ACCOUNT = 2,
70 };
71
72 static int update_block_group(struct btrfs_trans_handle *trans,
73                               struct btrfs_root *root,
74                               u64 bytenr, u64 num_bytes, int alloc);
75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
76                                 struct btrfs_root *root,
77                                 u64 bytenr, u64 num_bytes, u64 parent,
78                                 u64 root_objectid, u64 owner_objectid,
79                                 u64 owner_offset, int refs_to_drop,
80                                 struct btrfs_delayed_extent_op *extra_op);
81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
82                                     struct extent_buffer *leaf,
83                                     struct btrfs_extent_item *ei);
84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
85                                       struct btrfs_root *root,
86                                       u64 parent, u64 root_objectid,
87                                       u64 flags, u64 owner, u64 offset,
88                                       struct btrfs_key *ins, int ref_mod);
89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
90                                      struct btrfs_root *root,
91                                      u64 parent, u64 root_objectid,
92                                      u64 flags, struct btrfs_disk_key *key,
93                                      int level, struct btrfs_key *ins);
94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
95                           struct btrfs_root *extent_root, u64 alloc_bytes,
96                           u64 flags, int force);
97 static int find_next_key(struct btrfs_path *path, int level,
98                          struct btrfs_key *key);
99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
100                             int dump_block_groups);
101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
102                                        u64 num_bytes, int reserve);
103
104 static noinline int
105 block_group_cache_done(struct btrfs_block_group_cache *cache)
106 {
107         smp_mb();
108         return cache->cached == BTRFS_CACHE_FINISHED;
109 }
110
111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
112 {
113         return (cache->flags & bits) == bits;
114 }
115
116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
117 {
118         atomic_inc(&cache->count);
119 }
120
121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
122 {
123         if (atomic_dec_and_test(&cache->count)) {
124                 WARN_ON(cache->pinned > 0);
125                 WARN_ON(cache->reserved > 0);
126                 kfree(cache->free_space_ctl);
127                 kfree(cache);
128         }
129 }
130
131 /*
132  * this adds the block group to the fs_info rb tree for the block group
133  * cache
134  */
135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
136                                 struct btrfs_block_group_cache *block_group)
137 {
138         struct rb_node **p;
139         struct rb_node *parent = NULL;
140         struct btrfs_block_group_cache *cache;
141
142         spin_lock(&info->block_group_cache_lock);
143         p = &info->block_group_cache_tree.rb_node;
144
145         while (*p) {
146                 parent = *p;
147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
148                                  cache_node);
149                 if (block_group->key.objectid < cache->key.objectid) {
150                         p = &(*p)->rb_left;
151                 } else if (block_group->key.objectid > cache->key.objectid) {
152                         p = &(*p)->rb_right;
153                 } else {
154                         spin_unlock(&info->block_group_cache_lock);
155                         return -EEXIST;
156                 }
157         }
158
159         rb_link_node(&block_group->cache_node, parent, p);
160         rb_insert_color(&block_group->cache_node,
161                         &info->block_group_cache_tree);
162         spin_unlock(&info->block_group_cache_lock);
163
164         return 0;
165 }
166
167 /*
168  * This will return the block group at or after bytenr if contains is 0, else
169  * it will return the block group that contains the bytenr
170  */
171 static struct btrfs_block_group_cache *
172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
173                               int contains)
174 {
175         struct btrfs_block_group_cache *cache, *ret = NULL;
176         struct rb_node *n;
177         u64 end, start;
178
179         spin_lock(&info->block_group_cache_lock);
180         n = info->block_group_cache_tree.rb_node;
181
182         while (n) {
183                 cache = rb_entry(n, struct btrfs_block_group_cache,
184                                  cache_node);
185                 end = cache->key.objectid + cache->key.offset - 1;
186                 start = cache->key.objectid;
187
188                 if (bytenr < start) {
189                         if (!contains && (!ret || start < ret->key.objectid))
190                                 ret = cache;
191                         n = n->rb_left;
192                 } else if (bytenr > start) {
193                         if (contains && bytenr <= end) {
194                                 ret = cache;
195                                 break;
196                         }
197                         n = n->rb_right;
198                 } else {
199                         ret = cache;
200                         break;
201                 }
202         }
203         if (ret)
204                 btrfs_get_block_group(ret);
205         spin_unlock(&info->block_group_cache_lock);
206
207         return ret;
208 }
209
210 static int add_excluded_extent(struct btrfs_root *root,
211                                u64 start, u64 num_bytes)
212 {
213         u64 end = start + num_bytes - 1;
214         set_extent_bits(&root->fs_info->freed_extents[0],
215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
216         set_extent_bits(&root->fs_info->freed_extents[1],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         return 0;
219 }
220
221 static void free_excluded_extents(struct btrfs_root *root,
222                                   struct btrfs_block_group_cache *cache)
223 {
224         u64 start, end;
225
226         start = cache->key.objectid;
227         end = start + cache->key.offset - 1;
228
229         clear_extent_bits(&root->fs_info->freed_extents[0],
230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
231         clear_extent_bits(&root->fs_info->freed_extents[1],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233 }
234
235 static int exclude_super_stripes(struct btrfs_root *root,
236                                  struct btrfs_block_group_cache *cache)
237 {
238         u64 bytenr;
239         u64 *logical;
240         int stripe_len;
241         int i, nr, ret;
242
243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
245                 cache->bytes_super += stripe_len;
246                 ret = add_excluded_extent(root, cache->key.objectid,
247                                           stripe_len);
248                 BUG_ON(ret);
249         }
250
251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252                 bytenr = btrfs_sb_offset(i);
253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
254                                        cache->key.objectid, bytenr,
255                                        0, &logical, &nr, &stripe_len);
256                 BUG_ON(ret);
257
258                 while (nr--) {
259                         cache->bytes_super += stripe_len;
260                         ret = add_excluded_extent(root, logical[nr],
261                                                   stripe_len);
262                         BUG_ON(ret);
263                 }
264
265                 kfree(logical);
266         }
267         return 0;
268 }
269
270 static struct btrfs_caching_control *
271 get_caching_control(struct btrfs_block_group_cache *cache)
272 {
273         struct btrfs_caching_control *ctl;
274
275         spin_lock(&cache->lock);
276         if (cache->cached != BTRFS_CACHE_STARTED) {
277                 spin_unlock(&cache->lock);
278                 return NULL;
279         }
280
281         /* We're loading it the fast way, so we don't have a caching_ctl. */
282         if (!cache->caching_ctl) {
283                 spin_unlock(&cache->lock);
284                 return NULL;
285         }
286
287         ctl = cache->caching_ctl;
288         atomic_inc(&ctl->count);
289         spin_unlock(&cache->lock);
290         return ctl;
291 }
292
293 static void put_caching_control(struct btrfs_caching_control *ctl)
294 {
295         if (atomic_dec_and_test(&ctl->count))
296                 kfree(ctl);
297 }
298
299 /*
300  * this is only called by cache_block_group, since we could have freed extents
301  * we need to check the pinned_extents for any extents that can't be used yet
302  * since their free space will be released as soon as the transaction commits.
303  */
304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
305                               struct btrfs_fs_info *info, u64 start, u64 end)
306 {
307         u64 extent_start, extent_end, size, total_added = 0;
308         int ret;
309
310         while (start < end) {
311                 ret = find_first_extent_bit(info->pinned_extents, start,
312                                             &extent_start, &extent_end,
313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
314                 if (ret)
315                         break;
316
317                 if (extent_start <= start) {
318                         start = extent_end + 1;
319                 } else if (extent_start > start && extent_start < end) {
320                         size = extent_start - start;
321                         total_added += size;
322                         ret = btrfs_add_free_space(block_group, start,
323                                                    size);
324                         BUG_ON(ret);
325                         start = extent_end + 1;
326                 } else {
327                         break;
328                 }
329         }
330
331         if (start < end) {
332                 size = end - start;
333                 total_added += size;
334                 ret = btrfs_add_free_space(block_group, start, size);
335                 BUG_ON(ret);
336         }
337
338         return total_added;
339 }
340
341 static noinline void caching_thread(struct btrfs_work *work)
342 {
343         struct btrfs_block_group_cache *block_group;
344         struct btrfs_fs_info *fs_info;
345         struct btrfs_caching_control *caching_ctl;
346         struct btrfs_root *extent_root;
347         struct btrfs_path *path;
348         struct extent_buffer *leaf;
349         struct btrfs_key key;
350         u64 total_found = 0;
351         u64 last = 0;
352         u32 nritems;
353         int ret = 0;
354
355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
356         block_group = caching_ctl->block_group;
357         fs_info = block_group->fs_info;
358         extent_root = fs_info->extent_root;
359
360         path = btrfs_alloc_path();
361         if (!path)
362                 goto out;
363
364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
365
366         /*
367          * We don't want to deadlock with somebody trying to allocate a new
368          * extent for the extent root while also trying to search the extent
369          * root to add free space.  So we skip locking and search the commit
370          * root, since its read-only
371          */
372         path->skip_locking = 1;
373         path->search_commit_root = 1;
374         path->reada = 1;
375
376         key.objectid = last;
377         key.offset = 0;
378         key.type = BTRFS_EXTENT_ITEM_KEY;
379 again:
380         mutex_lock(&caching_ctl->mutex);
381         /* need to make sure the commit_root doesn't disappear */
382         down_read(&fs_info->extent_commit_sem);
383
384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
385         if (ret < 0)
386                 goto err;
387
388         leaf = path->nodes[0];
389         nritems = btrfs_header_nritems(leaf);
390
391         while (1) {
392                 if (btrfs_fs_closing(fs_info) > 1) {
393                         last = (u64)-1;
394                         break;
395                 }
396
397                 if (path->slots[0] < nritems) {
398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
399                 } else {
400                         ret = find_next_key(path, 0, &key);
401                         if (ret)
402                                 break;
403
404                         if (need_resched() ||
405                             btrfs_next_leaf(extent_root, path)) {
406                                 caching_ctl->progress = last;
407                                 btrfs_release_path(path);
408                                 up_read(&fs_info->extent_commit_sem);
409                                 mutex_unlock(&caching_ctl->mutex);
410                                 cond_resched();
411                                 goto again;
412                         }
413                         leaf = path->nodes[0];
414                         nritems = btrfs_header_nritems(leaf);
415                         continue;
416                 }
417
418                 if (key.objectid < block_group->key.objectid) {
419                         path->slots[0]++;
420                         continue;
421                 }
422
423                 if (key.objectid >= block_group->key.objectid +
424                     block_group->key.offset)
425                         break;
426
427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
428                         total_found += add_new_free_space(block_group,
429                                                           fs_info, last,
430                                                           key.objectid);
431                         last = key.objectid + key.offset;
432
433                         if (total_found > (1024 * 1024 * 2)) {
434                                 total_found = 0;
435                                 wake_up(&caching_ctl->wait);
436                         }
437                 }
438                 path->slots[0]++;
439         }
440         ret = 0;
441
442         total_found += add_new_free_space(block_group, fs_info, last,
443                                           block_group->key.objectid +
444                                           block_group->key.offset);
445         caching_ctl->progress = (u64)-1;
446
447         spin_lock(&block_group->lock);
448         block_group->caching_ctl = NULL;
449         block_group->cached = BTRFS_CACHE_FINISHED;
450         spin_unlock(&block_group->lock);
451
452 err:
453         btrfs_free_path(path);
454         up_read(&fs_info->extent_commit_sem);
455
456         free_excluded_extents(extent_root, block_group);
457
458         mutex_unlock(&caching_ctl->mutex);
459 out:
460         wake_up(&caching_ctl->wait);
461
462         put_caching_control(caching_ctl);
463         btrfs_put_block_group(block_group);
464 }
465
466 static int cache_block_group(struct btrfs_block_group_cache *cache,
467                              struct btrfs_trans_handle *trans,
468                              struct btrfs_root *root,
469                              int load_cache_only)
470 {
471         DEFINE_WAIT(wait);
472         struct btrfs_fs_info *fs_info = cache->fs_info;
473         struct btrfs_caching_control *caching_ctl;
474         int ret = 0;
475
476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
477         BUG_ON(!caching_ctl);
478
479         INIT_LIST_HEAD(&caching_ctl->list);
480         mutex_init(&caching_ctl->mutex);
481         init_waitqueue_head(&caching_ctl->wait);
482         caching_ctl->block_group = cache;
483         caching_ctl->progress = cache->key.objectid;
484         atomic_set(&caching_ctl->count, 1);
485         caching_ctl->work.func = caching_thread;
486
487         spin_lock(&cache->lock);
488         /*
489          * This should be a rare occasion, but this could happen I think in the
490          * case where one thread starts to load the space cache info, and then
491          * some other thread starts a transaction commit which tries to do an
492          * allocation while the other thread is still loading the space cache
493          * info.  The previous loop should have kept us from choosing this block
494          * group, but if we've moved to the state where we will wait on caching
495          * block groups we need to first check if we're doing a fast load here,
496          * so we can wait for it to finish, otherwise we could end up allocating
497          * from a block group who's cache gets evicted for one reason or
498          * another.
499          */
500         while (cache->cached == BTRFS_CACHE_FAST) {
501                 struct btrfs_caching_control *ctl;
502
503                 ctl = cache->caching_ctl;
504                 atomic_inc(&ctl->count);
505                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
506                 spin_unlock(&cache->lock);
507
508                 schedule();
509
510                 finish_wait(&ctl->wait, &wait);
511                 put_caching_control(ctl);
512                 spin_lock(&cache->lock);
513         }
514
515         if (cache->cached != BTRFS_CACHE_NO) {
516                 spin_unlock(&cache->lock);
517                 kfree(caching_ctl);
518                 return 0;
519         }
520         WARN_ON(cache->caching_ctl);
521         cache->caching_ctl = caching_ctl;
522         cache->cached = BTRFS_CACHE_FAST;
523         spin_unlock(&cache->lock);
524
525         /*
526          * We can't do the read from on-disk cache during a commit since we need
527          * to have the normal tree locking.  Also if we are currently trying to
528          * allocate blocks for the tree root we can't do the fast caching since
529          * we likely hold important locks.
530          */
531         if (trans && (!trans->transaction->in_commit) &&
532             (root && root != root->fs_info->tree_root) &&
533             btrfs_test_opt(root, SPACE_CACHE)) {
534                 ret = load_free_space_cache(fs_info, cache);
535
536                 spin_lock(&cache->lock);
537                 if (ret == 1) {
538                         cache->caching_ctl = NULL;
539                         cache->cached = BTRFS_CACHE_FINISHED;
540                         cache->last_byte_to_unpin = (u64)-1;
541                 } else {
542                         if (load_cache_only) {
543                                 cache->caching_ctl = NULL;
544                                 cache->cached = BTRFS_CACHE_NO;
545                         } else {
546                                 cache->cached = BTRFS_CACHE_STARTED;
547                         }
548                 }
549                 spin_unlock(&cache->lock);
550                 wake_up(&caching_ctl->wait);
551                 if (ret == 1) {
552                         put_caching_control(caching_ctl);
553                         free_excluded_extents(fs_info->extent_root, cache);
554                         return 0;
555                 }
556         } else {
557                 /*
558                  * We are not going to do the fast caching, set cached to the
559                  * appropriate value and wakeup any waiters.
560                  */
561                 spin_lock(&cache->lock);
562                 if (load_cache_only) {
563                         cache->caching_ctl = NULL;
564                         cache->cached = BTRFS_CACHE_NO;
565                 } else {
566                         cache->cached = BTRFS_CACHE_STARTED;
567                 }
568                 spin_unlock(&cache->lock);
569                 wake_up(&caching_ctl->wait);
570         }
571
572         if (load_cache_only) {
573                 put_caching_control(caching_ctl);
574                 return 0;
575         }
576
577         down_write(&fs_info->extent_commit_sem);
578         atomic_inc(&caching_ctl->count);
579         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
580         up_write(&fs_info->extent_commit_sem);
581
582         btrfs_get_block_group(cache);
583
584         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
585
586         return ret;
587 }
588
589 /*
590  * return the block group that starts at or after bytenr
591  */
592 static struct btrfs_block_group_cache *
593 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
594 {
595         struct btrfs_block_group_cache *cache;
596
597         cache = block_group_cache_tree_search(info, bytenr, 0);
598
599         return cache;
600 }
601
602 /*
603  * return the block group that contains the given bytenr
604  */
605 struct btrfs_block_group_cache *btrfs_lookup_block_group(
606                                                  struct btrfs_fs_info *info,
607                                                  u64 bytenr)
608 {
609         struct btrfs_block_group_cache *cache;
610
611         cache = block_group_cache_tree_search(info, bytenr, 1);
612
613         return cache;
614 }
615
616 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
617                                                   u64 flags)
618 {
619         struct list_head *head = &info->space_info;
620         struct btrfs_space_info *found;
621
622         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
623
624         rcu_read_lock();
625         list_for_each_entry_rcu(found, head, list) {
626                 if (found->flags & flags) {
627                         rcu_read_unlock();
628                         return found;
629                 }
630         }
631         rcu_read_unlock();
632         return NULL;
633 }
634
635 /*
636  * after adding space to the filesystem, we need to clear the full flags
637  * on all the space infos.
638  */
639 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
640 {
641         struct list_head *head = &info->space_info;
642         struct btrfs_space_info *found;
643
644         rcu_read_lock();
645         list_for_each_entry_rcu(found, head, list)
646                 found->full = 0;
647         rcu_read_unlock();
648 }
649
650 static u64 div_factor(u64 num, int factor)
651 {
652         if (factor == 10)
653                 return num;
654         num *= factor;
655         do_div(num, 10);
656         return num;
657 }
658
659 static u64 div_factor_fine(u64 num, int factor)
660 {
661         if (factor == 100)
662                 return num;
663         num *= factor;
664         do_div(num, 100);
665         return num;
666 }
667
668 u64 btrfs_find_block_group(struct btrfs_root *root,
669                            u64 search_start, u64 search_hint, int owner)
670 {
671         struct btrfs_block_group_cache *cache;
672         u64 used;
673         u64 last = max(search_hint, search_start);
674         u64 group_start = 0;
675         int full_search = 0;
676         int factor = 9;
677         int wrapped = 0;
678 again:
679         while (1) {
680                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
681                 if (!cache)
682                         break;
683
684                 spin_lock(&cache->lock);
685                 last = cache->key.objectid + cache->key.offset;
686                 used = btrfs_block_group_used(&cache->item);
687
688                 if ((full_search || !cache->ro) &&
689                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
690                         if (used + cache->pinned + cache->reserved <
691                             div_factor(cache->key.offset, factor)) {
692                                 group_start = cache->key.objectid;
693                                 spin_unlock(&cache->lock);
694                                 btrfs_put_block_group(cache);
695                                 goto found;
696                         }
697                 }
698                 spin_unlock(&cache->lock);
699                 btrfs_put_block_group(cache);
700                 cond_resched();
701         }
702         if (!wrapped) {
703                 last = search_start;
704                 wrapped = 1;
705                 goto again;
706         }
707         if (!full_search && factor < 10) {
708                 last = search_start;
709                 full_search = 1;
710                 factor = 10;
711                 goto again;
712         }
713 found:
714         return group_start;
715 }
716
717 /* simple helper to search for an existing extent at a given offset */
718 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
719 {
720         int ret;
721         struct btrfs_key key;
722         struct btrfs_path *path;
723
724         path = btrfs_alloc_path();
725         if (!path)
726                 return -ENOMEM;
727
728         key.objectid = start;
729         key.offset = len;
730         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
731         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
732                                 0, 0);
733         btrfs_free_path(path);
734         return ret;
735 }
736
737 /*
738  * helper function to lookup reference count and flags of extent.
739  *
740  * the head node for delayed ref is used to store the sum of all the
741  * reference count modifications queued up in the rbtree. the head
742  * node may also store the extent flags to set. This way you can check
743  * to see what the reference count and extent flags would be if all of
744  * the delayed refs are not processed.
745  */
746 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
747                              struct btrfs_root *root, u64 bytenr,
748                              u64 num_bytes, u64 *refs, u64 *flags)
749 {
750         struct btrfs_delayed_ref_head *head;
751         struct btrfs_delayed_ref_root *delayed_refs;
752         struct btrfs_path *path;
753         struct btrfs_extent_item *ei;
754         struct extent_buffer *leaf;
755         struct btrfs_key key;
756         u32 item_size;
757         u64 num_refs;
758         u64 extent_flags;
759         int ret;
760
761         path = btrfs_alloc_path();
762         if (!path)
763                 return -ENOMEM;
764
765         key.objectid = bytenr;
766         key.type = BTRFS_EXTENT_ITEM_KEY;
767         key.offset = num_bytes;
768         if (!trans) {
769                 path->skip_locking = 1;
770                 path->search_commit_root = 1;
771         }
772 again:
773         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
774                                 &key, path, 0, 0);
775         if (ret < 0)
776                 goto out_free;
777
778         if (ret == 0) {
779                 leaf = path->nodes[0];
780                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
781                 if (item_size >= sizeof(*ei)) {
782                         ei = btrfs_item_ptr(leaf, path->slots[0],
783                                             struct btrfs_extent_item);
784                         num_refs = btrfs_extent_refs(leaf, ei);
785                         extent_flags = btrfs_extent_flags(leaf, ei);
786                 } else {
787 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
788                         struct btrfs_extent_item_v0 *ei0;
789                         BUG_ON(item_size != sizeof(*ei0));
790                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
791                                              struct btrfs_extent_item_v0);
792                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
793                         /* FIXME: this isn't correct for data */
794                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
795 #else
796                         BUG();
797 #endif
798                 }
799                 BUG_ON(num_refs == 0);
800         } else {
801                 num_refs = 0;
802                 extent_flags = 0;
803                 ret = 0;
804         }
805
806         if (!trans)
807                 goto out;
808
809         delayed_refs = &trans->transaction->delayed_refs;
810         spin_lock(&delayed_refs->lock);
811         head = btrfs_find_delayed_ref_head(trans, bytenr);
812         if (head) {
813                 if (!mutex_trylock(&head->mutex)) {
814                         atomic_inc(&head->node.refs);
815                         spin_unlock(&delayed_refs->lock);
816
817                         btrfs_release_path(path);
818
819                         /*
820                          * Mutex was contended, block until it's released and try
821                          * again
822                          */
823                         mutex_lock(&head->mutex);
824                         mutex_unlock(&head->mutex);
825                         btrfs_put_delayed_ref(&head->node);
826                         goto again;
827                 }
828                 if (head->extent_op && head->extent_op->update_flags)
829                         extent_flags |= head->extent_op->flags_to_set;
830                 else
831                         BUG_ON(num_refs == 0);
832
833                 num_refs += head->node.ref_mod;
834                 mutex_unlock(&head->mutex);
835         }
836         spin_unlock(&delayed_refs->lock);
837 out:
838         WARN_ON(num_refs == 0);
839         if (refs)
840                 *refs = num_refs;
841         if (flags)
842                 *flags = extent_flags;
843 out_free:
844         btrfs_free_path(path);
845         return ret;
846 }
847
848 /*
849  * Back reference rules.  Back refs have three main goals:
850  *
851  * 1) differentiate between all holders of references to an extent so that
852  *    when a reference is dropped we can make sure it was a valid reference
853  *    before freeing the extent.
854  *
855  * 2) Provide enough information to quickly find the holders of an extent
856  *    if we notice a given block is corrupted or bad.
857  *
858  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
859  *    maintenance.  This is actually the same as #2, but with a slightly
860  *    different use case.
861  *
862  * There are two kinds of back refs. The implicit back refs is optimized
863  * for pointers in non-shared tree blocks. For a given pointer in a block,
864  * back refs of this kind provide information about the block's owner tree
865  * and the pointer's key. These information allow us to find the block by
866  * b-tree searching. The full back refs is for pointers in tree blocks not
867  * referenced by their owner trees. The location of tree block is recorded
868  * in the back refs. Actually the full back refs is generic, and can be
869  * used in all cases the implicit back refs is used. The major shortcoming
870  * of the full back refs is its overhead. Every time a tree block gets
871  * COWed, we have to update back refs entry for all pointers in it.
872  *
873  * For a newly allocated tree block, we use implicit back refs for
874  * pointers in it. This means most tree related operations only involve
875  * implicit back refs. For a tree block created in old transaction, the
876  * only way to drop a reference to it is COW it. So we can detect the
877  * event that tree block loses its owner tree's reference and do the
878  * back refs conversion.
879  *
880  * When a tree block is COW'd through a tree, there are four cases:
881  *
882  * The reference count of the block is one and the tree is the block's
883  * owner tree. Nothing to do in this case.
884  *
885  * The reference count of the block is one and the tree is not the
886  * block's owner tree. In this case, full back refs is used for pointers
887  * in the block. Remove these full back refs, add implicit back refs for
888  * every pointers in the new block.
889  *
890  * The reference count of the block is greater than one and the tree is
891  * the block's owner tree. In this case, implicit back refs is used for
892  * pointers in the block. Add full back refs for every pointers in the
893  * block, increase lower level extents' reference counts. The original
894  * implicit back refs are entailed to the new block.
895  *
896  * The reference count of the block is greater than one and the tree is
897  * not the block's owner tree. Add implicit back refs for every pointer in
898  * the new block, increase lower level extents' reference count.
899  *
900  * Back Reference Key composing:
901  *
902  * The key objectid corresponds to the first byte in the extent,
903  * The key type is used to differentiate between types of back refs.
904  * There are different meanings of the key offset for different types
905  * of back refs.
906  *
907  * File extents can be referenced by:
908  *
909  * - multiple snapshots, subvolumes, or different generations in one subvol
910  * - different files inside a single subvolume
911  * - different offsets inside a file (bookend extents in file.c)
912  *
913  * The extent ref structure for the implicit back refs has fields for:
914  *
915  * - Objectid of the subvolume root
916  * - objectid of the file holding the reference
917  * - original offset in the file
918  * - how many bookend extents
919  *
920  * The key offset for the implicit back refs is hash of the first
921  * three fields.
922  *
923  * The extent ref structure for the full back refs has field for:
924  *
925  * - number of pointers in the tree leaf
926  *
927  * The key offset for the implicit back refs is the first byte of
928  * the tree leaf
929  *
930  * When a file extent is allocated, The implicit back refs is used.
931  * the fields are filled in:
932  *
933  *     (root_key.objectid, inode objectid, offset in file, 1)
934  *
935  * When a file extent is removed file truncation, we find the
936  * corresponding implicit back refs and check the following fields:
937  *
938  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
939  *
940  * Btree extents can be referenced by:
941  *
942  * - Different subvolumes
943  *
944  * Both the implicit back refs and the full back refs for tree blocks
945  * only consist of key. The key offset for the implicit back refs is
946  * objectid of block's owner tree. The key offset for the full back refs
947  * is the first byte of parent block.
948  *
949  * When implicit back refs is used, information about the lowest key and
950  * level of the tree block are required. These information are stored in
951  * tree block info structure.
952  */
953
954 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
955 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
956                                   struct btrfs_root *root,
957                                   struct btrfs_path *path,
958                                   u64 owner, u32 extra_size)
959 {
960         struct btrfs_extent_item *item;
961         struct btrfs_extent_item_v0 *ei0;
962         struct btrfs_extent_ref_v0 *ref0;
963         struct btrfs_tree_block_info *bi;
964         struct extent_buffer *leaf;
965         struct btrfs_key key;
966         struct btrfs_key found_key;
967         u32 new_size = sizeof(*item);
968         u64 refs;
969         int ret;
970
971         leaf = path->nodes[0];
972         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
973
974         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
975         ei0 = btrfs_item_ptr(leaf, path->slots[0],
976                              struct btrfs_extent_item_v0);
977         refs = btrfs_extent_refs_v0(leaf, ei0);
978
979         if (owner == (u64)-1) {
980                 while (1) {
981                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
982                                 ret = btrfs_next_leaf(root, path);
983                                 if (ret < 0)
984                                         return ret;
985                                 BUG_ON(ret > 0);
986                                 leaf = path->nodes[0];
987                         }
988                         btrfs_item_key_to_cpu(leaf, &found_key,
989                                               path->slots[0]);
990                         BUG_ON(key.objectid != found_key.objectid);
991                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
992                                 path->slots[0]++;
993                                 continue;
994                         }
995                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
996                                               struct btrfs_extent_ref_v0);
997                         owner = btrfs_ref_objectid_v0(leaf, ref0);
998                         break;
999                 }
1000         }
1001         btrfs_release_path(path);
1002
1003         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1004                 new_size += sizeof(*bi);
1005
1006         new_size -= sizeof(*ei0);
1007         ret = btrfs_search_slot(trans, root, &key, path,
1008                                 new_size + extra_size, 1);
1009         if (ret < 0)
1010                 return ret;
1011         BUG_ON(ret);
1012
1013         ret = btrfs_extend_item(trans, root, path, new_size);
1014
1015         leaf = path->nodes[0];
1016         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1017         btrfs_set_extent_refs(leaf, item, refs);
1018         /* FIXME: get real generation */
1019         btrfs_set_extent_generation(leaf, item, 0);
1020         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1021                 btrfs_set_extent_flags(leaf, item,
1022                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1023                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1024                 bi = (struct btrfs_tree_block_info *)(item + 1);
1025                 /* FIXME: get first key of the block */
1026                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1027                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1028         } else {
1029                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1030         }
1031         btrfs_mark_buffer_dirty(leaf);
1032         return 0;
1033 }
1034 #endif
1035
1036 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1037 {
1038         u32 high_crc = ~(u32)0;
1039         u32 low_crc = ~(u32)0;
1040         __le64 lenum;
1041
1042         lenum = cpu_to_le64(root_objectid);
1043         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1044         lenum = cpu_to_le64(owner);
1045         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1046         lenum = cpu_to_le64(offset);
1047         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1048
1049         return ((u64)high_crc << 31) ^ (u64)low_crc;
1050 }
1051
1052 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1053                                      struct btrfs_extent_data_ref *ref)
1054 {
1055         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1056                                     btrfs_extent_data_ref_objectid(leaf, ref),
1057                                     btrfs_extent_data_ref_offset(leaf, ref));
1058 }
1059
1060 static int match_extent_data_ref(struct extent_buffer *leaf,
1061                                  struct btrfs_extent_data_ref *ref,
1062                                  u64 root_objectid, u64 owner, u64 offset)
1063 {
1064         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1065             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1066             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1067                 return 0;
1068         return 1;
1069 }
1070
1071 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1072                                            struct btrfs_root *root,
1073                                            struct btrfs_path *path,
1074                                            u64 bytenr, u64 parent,
1075                                            u64 root_objectid,
1076                                            u64 owner, u64 offset)
1077 {
1078         struct btrfs_key key;
1079         struct btrfs_extent_data_ref *ref;
1080         struct extent_buffer *leaf;
1081         u32 nritems;
1082         int ret;
1083         int recow;
1084         int err = -ENOENT;
1085
1086         key.objectid = bytenr;
1087         if (parent) {
1088                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1089                 key.offset = parent;
1090         } else {
1091                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1092                 key.offset = hash_extent_data_ref(root_objectid,
1093                                                   owner, offset);
1094         }
1095 again:
1096         recow = 0;
1097         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1098         if (ret < 0) {
1099                 err = ret;
1100                 goto fail;
1101         }
1102
1103         if (parent) {
1104                 if (!ret)
1105                         return 0;
1106 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1107                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1108                 btrfs_release_path(path);
1109                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1110                 if (ret < 0) {
1111                         err = ret;
1112                         goto fail;
1113                 }
1114                 if (!ret)
1115                         return 0;
1116 #endif
1117                 goto fail;
1118         }
1119
1120         leaf = path->nodes[0];
1121         nritems = btrfs_header_nritems(leaf);
1122         while (1) {
1123                 if (path->slots[0] >= nritems) {
1124                         ret = btrfs_next_leaf(root, path);
1125                         if (ret < 0)
1126                                 err = ret;
1127                         if (ret)
1128                                 goto fail;
1129
1130                         leaf = path->nodes[0];
1131                         nritems = btrfs_header_nritems(leaf);
1132                         recow = 1;
1133                 }
1134
1135                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1136                 if (key.objectid != bytenr ||
1137                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1138                         goto fail;
1139
1140                 ref = btrfs_item_ptr(leaf, path->slots[0],
1141                                      struct btrfs_extent_data_ref);
1142
1143                 if (match_extent_data_ref(leaf, ref, root_objectid,
1144                                           owner, offset)) {
1145                         if (recow) {
1146                                 btrfs_release_path(path);
1147                                 goto again;
1148                         }
1149                         err = 0;
1150                         break;
1151                 }
1152                 path->slots[0]++;
1153         }
1154 fail:
1155         return err;
1156 }
1157
1158 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1159                                            struct btrfs_root *root,
1160                                            struct btrfs_path *path,
1161                                            u64 bytenr, u64 parent,
1162                                            u64 root_objectid, u64 owner,
1163                                            u64 offset, int refs_to_add)
1164 {
1165         struct btrfs_key key;
1166         struct extent_buffer *leaf;
1167         u32 size;
1168         u32 num_refs;
1169         int ret;
1170
1171         key.objectid = bytenr;
1172         if (parent) {
1173                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1174                 key.offset = parent;
1175                 size = sizeof(struct btrfs_shared_data_ref);
1176         } else {
1177                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1178                 key.offset = hash_extent_data_ref(root_objectid,
1179                                                   owner, offset);
1180                 size = sizeof(struct btrfs_extent_data_ref);
1181         }
1182
1183         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1184         if (ret && ret != -EEXIST)
1185                 goto fail;
1186
1187         leaf = path->nodes[0];
1188         if (parent) {
1189                 struct btrfs_shared_data_ref *ref;
1190                 ref = btrfs_item_ptr(leaf, path->slots[0],
1191                                      struct btrfs_shared_data_ref);
1192                 if (ret == 0) {
1193                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1194                 } else {
1195                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1196                         num_refs += refs_to_add;
1197                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1198                 }
1199         } else {
1200                 struct btrfs_extent_data_ref *ref;
1201                 while (ret == -EEXIST) {
1202                         ref = btrfs_item_ptr(leaf, path->slots[0],
1203                                              struct btrfs_extent_data_ref);
1204                         if (match_extent_data_ref(leaf, ref, root_objectid,
1205                                                   owner, offset))
1206                                 break;
1207                         btrfs_release_path(path);
1208                         key.offset++;
1209                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1210                                                       size);
1211                         if (ret && ret != -EEXIST)
1212                                 goto fail;
1213
1214                         leaf = path->nodes[0];
1215                 }
1216                 ref = btrfs_item_ptr(leaf, path->slots[0],
1217                                      struct btrfs_extent_data_ref);
1218                 if (ret == 0) {
1219                         btrfs_set_extent_data_ref_root(leaf, ref,
1220                                                        root_objectid);
1221                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1222                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1223                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1224                 } else {
1225                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1226                         num_refs += refs_to_add;
1227                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1228                 }
1229         }
1230         btrfs_mark_buffer_dirty(leaf);
1231         ret = 0;
1232 fail:
1233         btrfs_release_path(path);
1234         return ret;
1235 }
1236
1237 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1238                                            struct btrfs_root *root,
1239                                            struct btrfs_path *path,
1240                                            int refs_to_drop)
1241 {
1242         struct btrfs_key key;
1243         struct btrfs_extent_data_ref *ref1 = NULL;
1244         struct btrfs_shared_data_ref *ref2 = NULL;
1245         struct extent_buffer *leaf;
1246         u32 num_refs = 0;
1247         int ret = 0;
1248
1249         leaf = path->nodes[0];
1250         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1251
1252         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1253                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1254                                       struct btrfs_extent_data_ref);
1255                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1256         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1257                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_shared_data_ref);
1259                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1260 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1261         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1262                 struct btrfs_extent_ref_v0 *ref0;
1263                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1264                                       struct btrfs_extent_ref_v0);
1265                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1266 #endif
1267         } else {
1268                 BUG();
1269         }
1270
1271         BUG_ON(num_refs < refs_to_drop);
1272         num_refs -= refs_to_drop;
1273
1274         if (num_refs == 0) {
1275                 ret = btrfs_del_item(trans, root, path);
1276         } else {
1277                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1278                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1279                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1280                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1281 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1282                 else {
1283                         struct btrfs_extent_ref_v0 *ref0;
1284                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1285                                         struct btrfs_extent_ref_v0);
1286                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1287                 }
1288 #endif
1289                 btrfs_mark_buffer_dirty(leaf);
1290         }
1291         return ret;
1292 }
1293
1294 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1295                                           struct btrfs_path *path,
1296                                           struct btrfs_extent_inline_ref *iref)
1297 {
1298         struct btrfs_key key;
1299         struct extent_buffer *leaf;
1300         struct btrfs_extent_data_ref *ref1;
1301         struct btrfs_shared_data_ref *ref2;
1302         u32 num_refs = 0;
1303
1304         leaf = path->nodes[0];
1305         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1306         if (iref) {
1307                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1308                     BTRFS_EXTENT_DATA_REF_KEY) {
1309                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1310                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1311                 } else {
1312                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1313                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1314                 }
1315         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1316                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1317                                       struct btrfs_extent_data_ref);
1318                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1319         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1320                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1321                                       struct btrfs_shared_data_ref);
1322                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1323 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1324         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1325                 struct btrfs_extent_ref_v0 *ref0;
1326                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1327                                       struct btrfs_extent_ref_v0);
1328                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1329 #endif
1330         } else {
1331                 WARN_ON(1);
1332         }
1333         return num_refs;
1334 }
1335
1336 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1337                                           struct btrfs_root *root,
1338                                           struct btrfs_path *path,
1339                                           u64 bytenr, u64 parent,
1340                                           u64 root_objectid)
1341 {
1342         struct btrfs_key key;
1343         int ret;
1344
1345         key.objectid = bytenr;
1346         if (parent) {
1347                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1348                 key.offset = parent;
1349         } else {
1350                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1351                 key.offset = root_objectid;
1352         }
1353
1354         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1355         if (ret > 0)
1356                 ret = -ENOENT;
1357 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1358         if (ret == -ENOENT && parent) {
1359                 btrfs_release_path(path);
1360                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1361                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1362                 if (ret > 0)
1363                         ret = -ENOENT;
1364         }
1365 #endif
1366         return ret;
1367 }
1368
1369 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1370                                           struct btrfs_root *root,
1371                                           struct btrfs_path *path,
1372                                           u64 bytenr, u64 parent,
1373                                           u64 root_objectid)
1374 {
1375         struct btrfs_key key;
1376         int ret;
1377
1378         key.objectid = bytenr;
1379         if (parent) {
1380                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1381                 key.offset = parent;
1382         } else {
1383                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1384                 key.offset = root_objectid;
1385         }
1386
1387         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1388         btrfs_release_path(path);
1389         return ret;
1390 }
1391
1392 static inline int extent_ref_type(u64 parent, u64 owner)
1393 {
1394         int type;
1395         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1396                 if (parent > 0)
1397                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1398                 else
1399                         type = BTRFS_TREE_BLOCK_REF_KEY;
1400         } else {
1401                 if (parent > 0)
1402                         type = BTRFS_SHARED_DATA_REF_KEY;
1403                 else
1404                         type = BTRFS_EXTENT_DATA_REF_KEY;
1405         }
1406         return type;
1407 }
1408
1409 static int find_next_key(struct btrfs_path *path, int level,
1410                          struct btrfs_key *key)
1411
1412 {
1413         for (; level < BTRFS_MAX_LEVEL; level++) {
1414                 if (!path->nodes[level])
1415                         break;
1416                 if (path->slots[level] + 1 >=
1417                     btrfs_header_nritems(path->nodes[level]))
1418                         continue;
1419                 if (level == 0)
1420                         btrfs_item_key_to_cpu(path->nodes[level], key,
1421                                               path->slots[level] + 1);
1422                 else
1423                         btrfs_node_key_to_cpu(path->nodes[level], key,
1424                                               path->slots[level] + 1);
1425                 return 0;
1426         }
1427         return 1;
1428 }
1429
1430 /*
1431  * look for inline back ref. if back ref is found, *ref_ret is set
1432  * to the address of inline back ref, and 0 is returned.
1433  *
1434  * if back ref isn't found, *ref_ret is set to the address where it
1435  * should be inserted, and -ENOENT is returned.
1436  *
1437  * if insert is true and there are too many inline back refs, the path
1438  * points to the extent item, and -EAGAIN is returned.
1439  *
1440  * NOTE: inline back refs are ordered in the same way that back ref
1441  *       items in the tree are ordered.
1442  */
1443 static noinline_for_stack
1444 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1445                                  struct btrfs_root *root,
1446                                  struct btrfs_path *path,
1447                                  struct btrfs_extent_inline_ref **ref_ret,
1448                                  u64 bytenr, u64 num_bytes,
1449                                  u64 parent, u64 root_objectid,
1450                                  u64 owner, u64 offset, int insert)
1451 {
1452         struct btrfs_key key;
1453         struct extent_buffer *leaf;
1454         struct btrfs_extent_item *ei;
1455         struct btrfs_extent_inline_ref *iref;
1456         u64 flags;
1457         u64 item_size;
1458         unsigned long ptr;
1459         unsigned long end;
1460         int extra_size;
1461         int type;
1462         int want;
1463         int ret;
1464         int err = 0;
1465
1466         key.objectid = bytenr;
1467         key.type = BTRFS_EXTENT_ITEM_KEY;
1468         key.offset = num_bytes;
1469
1470         want = extent_ref_type(parent, owner);
1471         if (insert) {
1472                 extra_size = btrfs_extent_inline_ref_size(want);
1473                 path->keep_locks = 1;
1474         } else
1475                 extra_size = -1;
1476         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1477         if (ret < 0) {
1478                 err = ret;
1479                 goto out;
1480         }
1481         BUG_ON(ret);
1482
1483         leaf = path->nodes[0];
1484         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1485 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1486         if (item_size < sizeof(*ei)) {
1487                 if (!insert) {
1488                         err = -ENOENT;
1489                         goto out;
1490                 }
1491                 ret = convert_extent_item_v0(trans, root, path, owner,
1492                                              extra_size);
1493                 if (ret < 0) {
1494                         err = ret;
1495                         goto out;
1496                 }
1497                 leaf = path->nodes[0];
1498                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1499         }
1500 #endif
1501         BUG_ON(item_size < sizeof(*ei));
1502
1503         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1504         flags = btrfs_extent_flags(leaf, ei);
1505
1506         ptr = (unsigned long)(ei + 1);
1507         end = (unsigned long)ei + item_size;
1508
1509         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1510                 ptr += sizeof(struct btrfs_tree_block_info);
1511                 BUG_ON(ptr > end);
1512         } else {
1513                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1514         }
1515
1516         err = -ENOENT;
1517         while (1) {
1518                 if (ptr >= end) {
1519                         WARN_ON(ptr > end);
1520                         break;
1521                 }
1522                 iref = (struct btrfs_extent_inline_ref *)ptr;
1523                 type = btrfs_extent_inline_ref_type(leaf, iref);
1524                 if (want < type)
1525                         break;
1526                 if (want > type) {
1527                         ptr += btrfs_extent_inline_ref_size(type);
1528                         continue;
1529                 }
1530
1531                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1532                         struct btrfs_extent_data_ref *dref;
1533                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1534                         if (match_extent_data_ref(leaf, dref, root_objectid,
1535                                                   owner, offset)) {
1536                                 err = 0;
1537                                 break;
1538                         }
1539                         if (hash_extent_data_ref_item(leaf, dref) <
1540                             hash_extent_data_ref(root_objectid, owner, offset))
1541                                 break;
1542                 } else {
1543                         u64 ref_offset;
1544                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1545                         if (parent > 0) {
1546                                 if (parent == ref_offset) {
1547                                         err = 0;
1548                                         break;
1549                                 }
1550                                 if (ref_offset < parent)
1551                                         break;
1552                         } else {
1553                                 if (root_objectid == ref_offset) {
1554                                         err = 0;
1555                                         break;
1556                                 }
1557                                 if (ref_offset < root_objectid)
1558                                         break;
1559                         }
1560                 }
1561                 ptr += btrfs_extent_inline_ref_size(type);
1562         }
1563         if (err == -ENOENT && insert) {
1564                 if (item_size + extra_size >=
1565                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1566                         err = -EAGAIN;
1567                         goto out;
1568                 }
1569                 /*
1570                  * To add new inline back ref, we have to make sure
1571                  * there is no corresponding back ref item.
1572                  * For simplicity, we just do not add new inline back
1573                  * ref if there is any kind of item for this block
1574                  */
1575                 if (find_next_key(path, 0, &key) == 0 &&
1576                     key.objectid == bytenr &&
1577                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1578                         err = -EAGAIN;
1579                         goto out;
1580                 }
1581         }
1582         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1583 out:
1584         if (insert) {
1585                 path->keep_locks = 0;
1586                 btrfs_unlock_up_safe(path, 1);
1587         }
1588         return err;
1589 }
1590
1591 /*
1592  * helper to add new inline back ref
1593  */
1594 static noinline_for_stack
1595 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1596                                 struct btrfs_root *root,
1597                                 struct btrfs_path *path,
1598                                 struct btrfs_extent_inline_ref *iref,
1599                                 u64 parent, u64 root_objectid,
1600                                 u64 owner, u64 offset, int refs_to_add,
1601                                 struct btrfs_delayed_extent_op *extent_op)
1602 {
1603         struct extent_buffer *leaf;
1604         struct btrfs_extent_item *ei;
1605         unsigned long ptr;
1606         unsigned long end;
1607         unsigned long item_offset;
1608         u64 refs;
1609         int size;
1610         int type;
1611         int ret;
1612
1613         leaf = path->nodes[0];
1614         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1615         item_offset = (unsigned long)iref - (unsigned long)ei;
1616
1617         type = extent_ref_type(parent, owner);
1618         size = btrfs_extent_inline_ref_size(type);
1619
1620         ret = btrfs_extend_item(trans, root, path, size);
1621
1622         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1623         refs = btrfs_extent_refs(leaf, ei);
1624         refs += refs_to_add;
1625         btrfs_set_extent_refs(leaf, ei, refs);
1626         if (extent_op)
1627                 __run_delayed_extent_op(extent_op, leaf, ei);
1628
1629         ptr = (unsigned long)ei + item_offset;
1630         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1631         if (ptr < end - size)
1632                 memmove_extent_buffer(leaf, ptr + size, ptr,
1633                                       end - size - ptr);
1634
1635         iref = (struct btrfs_extent_inline_ref *)ptr;
1636         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1637         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1638                 struct btrfs_extent_data_ref *dref;
1639                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1640                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1641                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1642                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1643                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1644         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1645                 struct btrfs_shared_data_ref *sref;
1646                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1647                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1648                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1649         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1653         }
1654         btrfs_mark_buffer_dirty(leaf);
1655         return 0;
1656 }
1657
1658 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1659                                  struct btrfs_root *root,
1660                                  struct btrfs_path *path,
1661                                  struct btrfs_extent_inline_ref **ref_ret,
1662                                  u64 bytenr, u64 num_bytes, u64 parent,
1663                                  u64 root_objectid, u64 owner, u64 offset)
1664 {
1665         int ret;
1666
1667         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1668                                            bytenr, num_bytes, parent,
1669                                            root_objectid, owner, offset, 0);
1670         if (ret != -ENOENT)
1671                 return ret;
1672
1673         btrfs_release_path(path);
1674         *ref_ret = NULL;
1675
1676         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1677                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1678                                             root_objectid);
1679         } else {
1680                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1681                                              root_objectid, owner, offset);
1682         }
1683         return ret;
1684 }
1685
1686 /*
1687  * helper to update/remove inline back ref
1688  */
1689 static noinline_for_stack
1690 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1691                                  struct btrfs_root *root,
1692                                  struct btrfs_path *path,
1693                                  struct btrfs_extent_inline_ref *iref,
1694                                  int refs_to_mod,
1695                                  struct btrfs_delayed_extent_op *extent_op)
1696 {
1697         struct extent_buffer *leaf;
1698         struct btrfs_extent_item *ei;
1699         struct btrfs_extent_data_ref *dref = NULL;
1700         struct btrfs_shared_data_ref *sref = NULL;
1701         unsigned long ptr;
1702         unsigned long end;
1703         u32 item_size;
1704         int size;
1705         int type;
1706         int ret;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751         return 0;
1752 }
1753
1754 static noinline_for_stack
1755 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1756                                  struct btrfs_root *root,
1757                                  struct btrfs_path *path,
1758                                  u64 bytenr, u64 num_bytes, u64 parent,
1759                                  u64 root_objectid, u64 owner,
1760                                  u64 offset, int refs_to_add,
1761                                  struct btrfs_delayed_extent_op *extent_op)
1762 {
1763         struct btrfs_extent_inline_ref *iref;
1764         int ret;
1765
1766         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1767                                            bytenr, num_bytes, parent,
1768                                            root_objectid, owner, offset, 1);
1769         if (ret == 0) {
1770                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1771                 ret = update_inline_extent_backref(trans, root, path, iref,
1772                                                    refs_to_add, extent_op);
1773         } else if (ret == -ENOENT) {
1774                 ret = setup_inline_extent_backref(trans, root, path, iref,
1775                                                   parent, root_objectid,
1776                                                   owner, offset, refs_to_add,
1777                                                   extent_op);
1778         }
1779         return ret;
1780 }
1781
1782 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1783                                  struct btrfs_root *root,
1784                                  struct btrfs_path *path,
1785                                  u64 bytenr, u64 parent, u64 root_objectid,
1786                                  u64 owner, u64 offset, int refs_to_add)
1787 {
1788         int ret;
1789         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1790                 BUG_ON(refs_to_add != 1);
1791                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1792                                             parent, root_objectid);
1793         } else {
1794                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1795                                              parent, root_objectid,
1796                                              owner, offset, refs_to_add);
1797         }
1798         return ret;
1799 }
1800
1801 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1802                                  struct btrfs_root *root,
1803                                  struct btrfs_path *path,
1804                                  struct btrfs_extent_inline_ref *iref,
1805                                  int refs_to_drop, int is_data)
1806 {
1807         int ret;
1808
1809         BUG_ON(!is_data && refs_to_drop != 1);
1810         if (iref) {
1811                 ret = update_inline_extent_backref(trans, root, path, iref,
1812                                                    -refs_to_drop, NULL);
1813         } else if (is_data) {
1814                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1815         } else {
1816                 ret = btrfs_del_item(trans, root, path);
1817         }
1818         return ret;
1819 }
1820
1821 static int btrfs_issue_discard(struct block_device *bdev,
1822                                 u64 start, u64 len)
1823 {
1824         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1825 }
1826
1827 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1828                                 u64 num_bytes, u64 *actual_bytes)
1829 {
1830         int ret;
1831         u64 discarded_bytes = 0;
1832         struct btrfs_bio *bbio = NULL;
1833
1834
1835         /* Tell the block device(s) that the sectors can be discarded */
1836         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1837                               bytenr, &num_bytes, &bbio, 0);
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break;
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1873                          struct btrfs_root *root,
1874                          u64 bytenr, u64 num_bytes, u64 parent,
1875                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1876 {
1877         int ret;
1878         struct btrfs_fs_info *fs_info = root->fs_info;
1879
1880         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1881                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1882
1883         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1884                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1885                                         num_bytes,
1886                                         parent, root_objectid, (int)owner,
1887                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1888         } else {
1889                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1890                                         num_bytes,
1891                                         parent, root_objectid, owner, offset,
1892                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1893         }
1894         return ret;
1895 }
1896
1897 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1898                                   struct btrfs_root *root,
1899                                   u64 bytenr, u64 num_bytes,
1900                                   u64 parent, u64 root_objectid,
1901                                   u64 owner, u64 offset, int refs_to_add,
1902                                   struct btrfs_delayed_extent_op *extent_op)
1903 {
1904         struct btrfs_path *path;
1905         struct extent_buffer *leaf;
1906         struct btrfs_extent_item *item;
1907         u64 refs;
1908         int ret;
1909         int err = 0;
1910
1911         path = btrfs_alloc_path();
1912         if (!path)
1913                 return -ENOMEM;
1914
1915         path->reada = 1;
1916         path->leave_spinning = 1;
1917         /* this will setup the path even if it fails to insert the back ref */
1918         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1919                                            path, bytenr, num_bytes, parent,
1920                                            root_objectid, owner, offset,
1921                                            refs_to_add, extent_op);
1922         if (ret == 0)
1923                 goto out;
1924
1925         if (ret != -EAGAIN) {
1926                 err = ret;
1927                 goto out;
1928         }
1929
1930         leaf = path->nodes[0];
1931         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1932         refs = btrfs_extent_refs(leaf, item);
1933         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1934         if (extent_op)
1935                 __run_delayed_extent_op(extent_op, leaf, item);
1936
1937         btrfs_mark_buffer_dirty(leaf);
1938         btrfs_release_path(path);
1939
1940         path->reada = 1;
1941         path->leave_spinning = 1;
1942
1943         /* now insert the actual backref */
1944         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1945                                     path, bytenr, parent, root_objectid,
1946                                     owner, offset, refs_to_add);
1947         BUG_ON(ret);
1948 out:
1949         btrfs_free_path(path);
1950         return err;
1951 }
1952
1953 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1954                                 struct btrfs_root *root,
1955                                 struct btrfs_delayed_ref_node *node,
1956                                 struct btrfs_delayed_extent_op *extent_op,
1957                                 int insert_reserved)
1958 {
1959         int ret = 0;
1960         struct btrfs_delayed_data_ref *ref;
1961         struct btrfs_key ins;
1962         u64 parent = 0;
1963         u64 ref_root = 0;
1964         u64 flags = 0;
1965
1966         ins.objectid = node->bytenr;
1967         ins.offset = node->num_bytes;
1968         ins.type = BTRFS_EXTENT_ITEM_KEY;
1969
1970         ref = btrfs_delayed_node_to_data_ref(node);
1971         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1972                 parent = ref->parent;
1973         else
1974                 ref_root = ref->root;
1975
1976         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1977                 if (extent_op) {
1978                         BUG_ON(extent_op->update_key);
1979                         flags |= extent_op->flags_to_set;
1980                 }
1981                 ret = alloc_reserved_file_extent(trans, root,
1982                                                  parent, ref_root, flags,
1983                                                  ref->objectid, ref->offset,
1984                                                  &ins, node->ref_mod);
1985         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1986                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1987                                              node->num_bytes, parent,
1988                                              ref_root, ref->objectid,
1989                                              ref->offset, node->ref_mod,
1990                                              extent_op);
1991         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1992                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1993                                           node->num_bytes, parent,
1994                                           ref_root, ref->objectid,
1995                                           ref->offset, node->ref_mod,
1996                                           extent_op);
1997         } else {
1998                 BUG();
1999         }
2000         return ret;
2001 }
2002
2003 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2004                                     struct extent_buffer *leaf,
2005                                     struct btrfs_extent_item *ei)
2006 {
2007         u64 flags = btrfs_extent_flags(leaf, ei);
2008         if (extent_op->update_flags) {
2009                 flags |= extent_op->flags_to_set;
2010                 btrfs_set_extent_flags(leaf, ei, flags);
2011         }
2012
2013         if (extent_op->update_key) {
2014                 struct btrfs_tree_block_info *bi;
2015                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2016                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2017                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2018         }
2019 }
2020
2021 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2022                                  struct btrfs_root *root,
2023                                  struct btrfs_delayed_ref_node *node,
2024                                  struct btrfs_delayed_extent_op *extent_op)
2025 {
2026         struct btrfs_key key;
2027         struct btrfs_path *path;
2028         struct btrfs_extent_item *ei;
2029         struct extent_buffer *leaf;
2030         u32 item_size;
2031         int ret;
2032         int err = 0;
2033
2034         path = btrfs_alloc_path();
2035         if (!path)
2036                 return -ENOMEM;
2037
2038         key.objectid = node->bytenr;
2039         key.type = BTRFS_EXTENT_ITEM_KEY;
2040         key.offset = node->num_bytes;
2041
2042         path->reada = 1;
2043         path->leave_spinning = 1;
2044         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2045                                 path, 0, 1);
2046         if (ret < 0) {
2047                 err = ret;
2048                 goto out;
2049         }
2050         if (ret > 0) {
2051                 err = -EIO;
2052                 goto out;
2053         }
2054
2055         leaf = path->nodes[0];
2056         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2057 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2058         if (item_size < sizeof(*ei)) {
2059                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2060                                              path, (u64)-1, 0);
2061                 if (ret < 0) {
2062                         err = ret;
2063                         goto out;
2064                 }
2065                 leaf = path->nodes[0];
2066                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2067         }
2068 #endif
2069         BUG_ON(item_size < sizeof(*ei));
2070         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2071         __run_delayed_extent_op(extent_op, leaf, ei);
2072
2073         btrfs_mark_buffer_dirty(leaf);
2074 out:
2075         btrfs_free_path(path);
2076         return err;
2077 }
2078
2079 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2080                                 struct btrfs_root *root,
2081                                 struct btrfs_delayed_ref_node *node,
2082                                 struct btrfs_delayed_extent_op *extent_op,
2083                                 int insert_reserved)
2084 {
2085         int ret = 0;
2086         struct btrfs_delayed_tree_ref *ref;
2087         struct btrfs_key ins;
2088         u64 parent = 0;
2089         u64 ref_root = 0;
2090
2091         ins.objectid = node->bytenr;
2092         ins.offset = node->num_bytes;
2093         ins.type = BTRFS_EXTENT_ITEM_KEY;
2094
2095         ref = btrfs_delayed_node_to_tree_ref(node);
2096         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2097                 parent = ref->parent;
2098         else
2099                 ref_root = ref->root;
2100
2101         BUG_ON(node->ref_mod != 1);
2102         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2103                 BUG_ON(!extent_op || !extent_op->update_flags ||
2104                        !extent_op->update_key);
2105                 ret = alloc_reserved_tree_block(trans, root,
2106                                                 parent, ref_root,
2107                                                 extent_op->flags_to_set,
2108                                                 &extent_op->key,
2109                                                 ref->level, &ins);
2110         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2111                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2112                                              node->num_bytes, parent, ref_root,
2113                                              ref->level, 0, 1, extent_op);
2114         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2115                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2116                                           node->num_bytes, parent, ref_root,
2117                                           ref->level, 0, 1, extent_op);
2118         } else {
2119                 BUG();
2120         }
2121         return ret;
2122 }
2123
2124 /* helper function to actually process a single delayed ref entry */
2125 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2126                                struct btrfs_root *root,
2127                                struct btrfs_delayed_ref_node *node,
2128                                struct btrfs_delayed_extent_op *extent_op,
2129                                int insert_reserved)
2130 {
2131         int ret;
2132         if (btrfs_delayed_ref_is_head(node)) {
2133                 struct btrfs_delayed_ref_head *head;
2134                 /*
2135                  * we've hit the end of the chain and we were supposed
2136                  * to insert this extent into the tree.  But, it got
2137                  * deleted before we ever needed to insert it, so all
2138                  * we have to do is clean up the accounting
2139                  */
2140                 BUG_ON(extent_op);
2141                 head = btrfs_delayed_node_to_head(node);
2142                 if (insert_reserved) {
2143                         btrfs_pin_extent(root, node->bytenr,
2144                                          node->num_bytes, 1);
2145                         if (head->is_data) {
2146                                 ret = btrfs_del_csums(trans, root,
2147                                                       node->bytenr,
2148                                                       node->num_bytes);
2149                                 BUG_ON(ret);
2150                         }
2151                 }
2152                 mutex_unlock(&head->mutex);
2153                 return 0;
2154         }
2155
2156         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2157             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2158                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2159                                            insert_reserved);
2160         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2161                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2162                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2163                                            insert_reserved);
2164         else
2165                 BUG();
2166         return ret;
2167 }
2168
2169 static noinline struct btrfs_delayed_ref_node *
2170 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2171 {
2172         struct rb_node *node;
2173         struct btrfs_delayed_ref_node *ref;
2174         int action = BTRFS_ADD_DELAYED_REF;
2175 again:
2176         /*
2177          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2178          * this prevents ref count from going down to zero when
2179          * there still are pending delayed ref.
2180          */
2181         node = rb_prev(&head->node.rb_node);
2182         while (1) {
2183                 if (!node)
2184                         break;
2185                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2186                                 rb_node);
2187                 if (ref->bytenr != head->node.bytenr)
2188                         break;
2189                 if (ref->action == action)
2190                         return ref;
2191                 node = rb_prev(node);
2192         }
2193         if (action == BTRFS_ADD_DELAYED_REF) {
2194                 action = BTRFS_DROP_DELAYED_REF;
2195                 goto again;
2196         }
2197         return NULL;
2198 }
2199
2200 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2201                                        struct btrfs_root *root,
2202                                        struct list_head *cluster)
2203 {
2204         struct btrfs_delayed_ref_root *delayed_refs;
2205         struct btrfs_delayed_ref_node *ref;
2206         struct btrfs_delayed_ref_head *locked_ref = NULL;
2207         struct btrfs_delayed_extent_op *extent_op;
2208         int ret;
2209         int count = 0;
2210         int must_insert_reserved = 0;
2211
2212         delayed_refs = &trans->transaction->delayed_refs;
2213         while (1) {
2214                 if (!locked_ref) {
2215                         /* pick a new head ref from the cluster list */
2216                         if (list_empty(cluster))
2217                                 break;
2218
2219                         locked_ref = list_entry(cluster->next,
2220                                      struct btrfs_delayed_ref_head, cluster);
2221
2222                         /* grab the lock that says we are going to process
2223                          * all the refs for this head */
2224                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2225
2226                         /*
2227                          * we may have dropped the spin lock to get the head
2228                          * mutex lock, and that might have given someone else
2229                          * time to free the head.  If that's true, it has been
2230                          * removed from our list and we can move on.
2231                          */
2232                         if (ret == -EAGAIN) {
2233                                 locked_ref = NULL;
2234                                 count++;
2235                                 continue;
2236                         }
2237                 }
2238
2239                 /*
2240                  * locked_ref is the head node, so we have to go one
2241                  * node back for any delayed ref updates
2242                  */
2243                 ref = select_delayed_ref(locked_ref);
2244
2245                 if (ref && ref->seq &&
2246                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2247                         /*
2248                          * there are still refs with lower seq numbers in the
2249                          * process of being added. Don't run this ref yet.
2250                          */
2251                         list_del_init(&locked_ref->cluster);
2252                         mutex_unlock(&locked_ref->mutex);
2253                         locked_ref = NULL;
2254                         delayed_refs->num_heads_ready++;
2255                         spin_unlock(&delayed_refs->lock);
2256                         cond_resched();
2257                         spin_lock(&delayed_refs->lock);
2258                         continue;
2259                 }
2260
2261                 /*
2262                  * record the must insert reserved flag before we
2263                  * drop the spin lock.
2264                  */
2265                 must_insert_reserved = locked_ref->must_insert_reserved;
2266                 locked_ref->must_insert_reserved = 0;
2267
2268                 extent_op = locked_ref->extent_op;
2269                 locked_ref->extent_op = NULL;
2270
2271                 if (!ref) {
2272                         /* All delayed refs have been processed, Go ahead
2273                          * and send the head node to run_one_delayed_ref,
2274                          * so that any accounting fixes can happen
2275                          */
2276                         ref = &locked_ref->node;
2277
2278                         if (extent_op && must_insert_reserved) {
2279                                 kfree(extent_op);
2280                                 extent_op = NULL;
2281                         }
2282
2283                         if (extent_op) {
2284                                 spin_unlock(&delayed_refs->lock);
2285
2286                                 ret = run_delayed_extent_op(trans, root,
2287                                                             ref, extent_op);
2288                                 BUG_ON(ret);
2289                                 kfree(extent_op);
2290
2291                                 goto next;
2292                         }
2293
2294                         list_del_init(&locked_ref->cluster);
2295                         locked_ref = NULL;
2296                 }
2297
2298                 ref->in_tree = 0;
2299                 rb_erase(&ref->rb_node, &delayed_refs->root);
2300                 delayed_refs->num_entries--;
2301                 /*
2302                  * we modified num_entries, but as we're currently running
2303                  * delayed refs, skip
2304                  *     wake_up(&delayed_refs->seq_wait);
2305                  * here.
2306                  */
2307                 spin_unlock(&delayed_refs->lock);
2308
2309                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2310                                           must_insert_reserved);
2311                 BUG_ON(ret);
2312
2313                 btrfs_put_delayed_ref(ref);
2314                 kfree(extent_op);
2315                 count++;
2316 next:
2317                 do_chunk_alloc(trans, root->fs_info->extent_root,
2318                                2 * 1024 * 1024,
2319                                btrfs_get_alloc_profile(root, 0),
2320                                CHUNK_ALLOC_NO_FORCE);
2321                 cond_resched();
2322                 spin_lock(&delayed_refs->lock);
2323         }
2324         return count;
2325 }
2326
2327
2328 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2329                         unsigned long num_refs)
2330 {
2331         struct list_head *first_seq = delayed_refs->seq_head.next;
2332
2333         spin_unlock(&delayed_refs->lock);
2334         pr_debug("waiting for more refs (num %ld, first %p)\n",
2335                  num_refs, first_seq);
2336         wait_event(delayed_refs->seq_wait,
2337                    num_refs != delayed_refs->num_entries ||
2338                    delayed_refs->seq_head.next != first_seq);
2339         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2340                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2341         spin_lock(&delayed_refs->lock);
2342 }
2343
2344 /*
2345  * this starts processing the delayed reference count updates and
2346  * extent insertions we have queued up so far.  count can be
2347  * 0, which means to process everything in the tree at the start
2348  * of the run (but not newly added entries), or it can be some target
2349  * number you'd like to process.
2350  */
2351 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2352                            struct btrfs_root *root, unsigned long count)
2353 {
2354         struct rb_node *node;
2355         struct btrfs_delayed_ref_root *delayed_refs;
2356         struct btrfs_delayed_ref_node *ref;
2357         struct list_head cluster;
2358         int ret;
2359         u64 delayed_start;
2360         int run_all = count == (unsigned long)-1;
2361         int run_most = 0;
2362         unsigned long num_refs = 0;
2363         int consider_waiting;
2364
2365         if (root == root->fs_info->extent_root)
2366                 root = root->fs_info->tree_root;
2367
2368         do_chunk_alloc(trans, root->fs_info->extent_root,
2369                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2370                        CHUNK_ALLOC_NO_FORCE);
2371
2372         delayed_refs = &trans->transaction->delayed_refs;
2373         INIT_LIST_HEAD(&cluster);
2374 again:
2375         consider_waiting = 0;
2376         spin_lock(&delayed_refs->lock);
2377         if (count == 0) {
2378                 count = delayed_refs->num_entries * 2;
2379                 run_most = 1;
2380         }
2381         while (1) {
2382                 if (!(run_all || run_most) &&
2383                     delayed_refs->num_heads_ready < 64)
2384                         break;
2385
2386                 /*
2387                  * go find something we can process in the rbtree.  We start at
2388                  * the beginning of the tree, and then build a cluster
2389                  * of refs to process starting at the first one we are able to
2390                  * lock
2391                  */
2392                 delayed_start = delayed_refs->run_delayed_start;
2393                 ret = btrfs_find_ref_cluster(trans, &cluster,
2394                                              delayed_refs->run_delayed_start);
2395                 if (ret)
2396                         break;
2397
2398                 if (delayed_start >= delayed_refs->run_delayed_start) {
2399                         if (consider_waiting == 0) {
2400                                 /*
2401                                  * btrfs_find_ref_cluster looped. let's do one
2402                                  * more cycle. if we don't run any delayed ref
2403                                  * during that cycle (because we can't because
2404                                  * all of them are blocked) and if the number of
2405                                  * refs doesn't change, we avoid busy waiting.
2406                                  */
2407                                 consider_waiting = 1;
2408                                 num_refs = delayed_refs->num_entries;
2409                         } else {
2410                                 wait_for_more_refs(delayed_refs, num_refs);
2411                                 /*
2412                                  * after waiting, things have changed. we
2413                                  * dropped the lock and someone else might have
2414                                  * run some refs, built new clusters and so on.
2415                                  * therefore, we restart staleness detection.
2416                                  */
2417                                 consider_waiting = 0;
2418                         }
2419                 }
2420
2421                 ret = run_clustered_refs(trans, root, &cluster);
2422                 BUG_ON(ret < 0);
2423
2424                 count -= min_t(unsigned long, ret, count);
2425
2426                 if (count == 0)
2427                         break;
2428
2429                 if (ret || delayed_refs->run_delayed_start == 0) {
2430                         /* refs were run, let's reset staleness detection */
2431                         consider_waiting = 0;
2432                 }
2433         }
2434
2435         if (run_all) {
2436                 node = rb_first(&delayed_refs->root);
2437                 if (!node)
2438                         goto out;
2439                 count = (unsigned long)-1;
2440
2441                 while (node) {
2442                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2443                                        rb_node);
2444                         if (btrfs_delayed_ref_is_head(ref)) {
2445                                 struct btrfs_delayed_ref_head *head;
2446
2447                                 head = btrfs_delayed_node_to_head(ref);
2448                                 atomic_inc(&ref->refs);
2449
2450                                 spin_unlock(&delayed_refs->lock);
2451                                 /*
2452                                  * Mutex was contended, block until it's
2453                                  * released and try again
2454                                  */
2455                                 mutex_lock(&head->mutex);
2456                                 mutex_unlock(&head->mutex);
2457
2458                                 btrfs_put_delayed_ref(ref);
2459                                 cond_resched();
2460                                 goto again;
2461                         }
2462                         node = rb_next(node);
2463                 }
2464                 spin_unlock(&delayed_refs->lock);
2465                 schedule_timeout(1);
2466                 goto again;
2467         }
2468 out:
2469         spin_unlock(&delayed_refs->lock);
2470         return 0;
2471 }
2472
2473 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2474                                 struct btrfs_root *root,
2475                                 u64 bytenr, u64 num_bytes, u64 flags,
2476                                 int is_data)
2477 {
2478         struct btrfs_delayed_extent_op *extent_op;
2479         int ret;
2480
2481         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2482         if (!extent_op)
2483                 return -ENOMEM;
2484
2485         extent_op->flags_to_set = flags;
2486         extent_op->update_flags = 1;
2487         extent_op->update_key = 0;
2488         extent_op->is_data = is_data ? 1 : 0;
2489
2490         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2491                                           num_bytes, extent_op);
2492         if (ret)
2493                 kfree(extent_op);
2494         return ret;
2495 }
2496
2497 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2498                                       struct btrfs_root *root,
2499                                       struct btrfs_path *path,
2500                                       u64 objectid, u64 offset, u64 bytenr)
2501 {
2502         struct btrfs_delayed_ref_head *head;
2503         struct btrfs_delayed_ref_node *ref;
2504         struct btrfs_delayed_data_ref *data_ref;
2505         struct btrfs_delayed_ref_root *delayed_refs;
2506         struct rb_node *node;
2507         int ret = 0;
2508
2509         ret = -ENOENT;
2510         delayed_refs = &trans->transaction->delayed_refs;
2511         spin_lock(&delayed_refs->lock);
2512         head = btrfs_find_delayed_ref_head(trans, bytenr);
2513         if (!head)
2514                 goto out;
2515
2516         if (!mutex_trylock(&head->mutex)) {
2517                 atomic_inc(&head->node.refs);
2518                 spin_unlock(&delayed_refs->lock);
2519
2520                 btrfs_release_path(path);
2521
2522                 /*
2523                  * Mutex was contended, block until it's released and let
2524                  * caller try again
2525                  */
2526                 mutex_lock(&head->mutex);
2527                 mutex_unlock(&head->mutex);
2528                 btrfs_put_delayed_ref(&head->node);
2529                 return -EAGAIN;
2530         }
2531
2532         node = rb_prev(&head->node.rb_node);
2533         if (!node)
2534                 goto out_unlock;
2535
2536         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2537
2538         if (ref->bytenr != bytenr)
2539                 goto out_unlock;
2540
2541         ret = 1;
2542         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2543                 goto out_unlock;
2544
2545         data_ref = btrfs_delayed_node_to_data_ref(ref);
2546
2547         node = rb_prev(node);
2548         if (node) {
2549                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2550                 if (ref->bytenr == bytenr)
2551                         goto out_unlock;
2552         }
2553
2554         if (data_ref->root != root->root_key.objectid ||
2555             data_ref->objectid != objectid || data_ref->offset != offset)
2556                 goto out_unlock;
2557
2558         ret = 0;
2559 out_unlock:
2560         mutex_unlock(&head->mutex);
2561 out:
2562         spin_unlock(&delayed_refs->lock);
2563         return ret;
2564 }
2565
2566 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2567                                         struct btrfs_root *root,
2568                                         struct btrfs_path *path,
2569                                         u64 objectid, u64 offset, u64 bytenr)
2570 {
2571         struct btrfs_root *extent_root = root->fs_info->extent_root;
2572         struct extent_buffer *leaf;
2573         struct btrfs_extent_data_ref *ref;
2574         struct btrfs_extent_inline_ref *iref;
2575         struct btrfs_extent_item *ei;
2576         struct btrfs_key key;
2577         u32 item_size;
2578         int ret;
2579
2580         key.objectid = bytenr;
2581         key.offset = (u64)-1;
2582         key.type = BTRFS_EXTENT_ITEM_KEY;
2583
2584         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2585         if (ret < 0)
2586                 goto out;
2587         BUG_ON(ret == 0);
2588
2589         ret = -ENOENT;
2590         if (path->slots[0] == 0)
2591                 goto out;
2592
2593         path->slots[0]--;
2594         leaf = path->nodes[0];
2595         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2596
2597         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2598                 goto out;
2599
2600         ret = 1;
2601         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2602 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2603         if (item_size < sizeof(*ei)) {
2604                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2605                 goto out;
2606         }
2607 #endif
2608         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2609
2610         if (item_size != sizeof(*ei) +
2611             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2612                 goto out;
2613
2614         if (btrfs_extent_generation(leaf, ei) <=
2615             btrfs_root_last_snapshot(&root->root_item))
2616                 goto out;
2617
2618         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2619         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2620             BTRFS_EXTENT_DATA_REF_KEY)
2621                 goto out;
2622
2623         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2624         if (btrfs_extent_refs(leaf, ei) !=
2625             btrfs_extent_data_ref_count(leaf, ref) ||
2626             btrfs_extent_data_ref_root(leaf, ref) !=
2627             root->root_key.objectid ||
2628             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2629             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2630                 goto out;
2631
2632         ret = 0;
2633 out:
2634         return ret;
2635 }
2636
2637 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2638                           struct btrfs_root *root,
2639                           u64 objectid, u64 offset, u64 bytenr)
2640 {
2641         struct btrfs_path *path;
2642         int ret;
2643         int ret2;
2644
2645         path = btrfs_alloc_path();
2646         if (!path)
2647                 return -ENOENT;
2648
2649         do {
2650                 ret = check_committed_ref(trans, root, path, objectid,
2651                                           offset, bytenr);
2652                 if (ret && ret != -ENOENT)
2653                         goto out;
2654
2655                 ret2 = check_delayed_ref(trans, root, path, objectid,
2656                                          offset, bytenr);
2657         } while (ret2 == -EAGAIN);
2658
2659         if (ret2 && ret2 != -ENOENT) {
2660                 ret = ret2;
2661                 goto out;
2662         }
2663
2664         if (ret != -ENOENT || ret2 != -ENOENT)
2665                 ret = 0;
2666 out:
2667         btrfs_free_path(path);
2668         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2669                 WARN_ON(ret > 0);
2670         return ret;
2671 }
2672
2673 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2674                            struct btrfs_root *root,
2675                            struct extent_buffer *buf,
2676                            int full_backref, int inc, int for_cow)
2677 {
2678         u64 bytenr;
2679         u64 num_bytes;
2680         u64 parent;
2681         u64 ref_root;
2682         u32 nritems;
2683         struct btrfs_key key;
2684         struct btrfs_file_extent_item *fi;
2685         int i;
2686         int level;
2687         int ret = 0;
2688         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2689                             u64, u64, u64, u64, u64, u64, int);
2690
2691         ref_root = btrfs_header_owner(buf);
2692         nritems = btrfs_header_nritems(buf);
2693         level = btrfs_header_level(buf);
2694
2695         if (!root->ref_cows && level == 0)
2696                 return 0;
2697
2698         if (inc)
2699                 process_func = btrfs_inc_extent_ref;
2700         else
2701                 process_func = btrfs_free_extent;
2702
2703         if (full_backref)
2704                 parent = buf->start;
2705         else
2706                 parent = 0;
2707
2708         for (i = 0; i < nritems; i++) {
2709                 if (level == 0) {
2710                         btrfs_item_key_to_cpu(buf, &key, i);
2711                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2712                                 continue;
2713                         fi = btrfs_item_ptr(buf, i,
2714                                             struct btrfs_file_extent_item);
2715                         if (btrfs_file_extent_type(buf, fi) ==
2716                             BTRFS_FILE_EXTENT_INLINE)
2717                                 continue;
2718                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2719                         if (bytenr == 0)
2720                                 continue;
2721
2722                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2723                         key.offset -= btrfs_file_extent_offset(buf, fi);
2724                         ret = process_func(trans, root, bytenr, num_bytes,
2725                                            parent, ref_root, key.objectid,
2726                                            key.offset, for_cow);
2727                         if (ret)
2728                                 goto fail;
2729                 } else {
2730                         bytenr = btrfs_node_blockptr(buf, i);
2731                         num_bytes = btrfs_level_size(root, level - 1);
2732                         ret = process_func(trans, root, bytenr, num_bytes,
2733                                            parent, ref_root, level - 1, 0,
2734                                            for_cow);
2735                         if (ret)
2736                                 goto fail;
2737                 }
2738         }
2739         return 0;
2740 fail:
2741         BUG();
2742         return ret;
2743 }
2744
2745 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2746                   struct extent_buffer *buf, int full_backref, int for_cow)
2747 {
2748         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2749 }
2750
2751 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2752                   struct extent_buffer *buf, int full_backref, int for_cow)
2753 {
2754         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2755 }
2756
2757 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2758                                  struct btrfs_root *root,
2759                                  struct btrfs_path *path,
2760                                  struct btrfs_block_group_cache *cache)
2761 {
2762         int ret;
2763         struct btrfs_root *extent_root = root->fs_info->extent_root;
2764         unsigned long bi;
2765         struct extent_buffer *leaf;
2766
2767         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2768         if (ret < 0)
2769                 goto fail;
2770         BUG_ON(ret);
2771
2772         leaf = path->nodes[0];
2773         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2774         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2775         btrfs_mark_buffer_dirty(leaf);
2776         btrfs_release_path(path);
2777 fail:
2778         if (ret)
2779                 return ret;
2780         return 0;
2781
2782 }
2783
2784 static struct btrfs_block_group_cache *
2785 next_block_group(struct btrfs_root *root,
2786                  struct btrfs_block_group_cache *cache)
2787 {
2788         struct rb_node *node;
2789         spin_lock(&root->fs_info->block_group_cache_lock);
2790         node = rb_next(&cache->cache_node);
2791         btrfs_put_block_group(cache);
2792         if (node) {
2793                 cache = rb_entry(node, struct btrfs_block_group_cache,
2794                                  cache_node);
2795                 btrfs_get_block_group(cache);
2796         } else
2797                 cache = NULL;
2798         spin_unlock(&root->fs_info->block_group_cache_lock);
2799         return cache;
2800 }
2801
2802 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2803                             struct btrfs_trans_handle *trans,
2804                             struct btrfs_path *path)
2805 {
2806         struct btrfs_root *root = block_group->fs_info->tree_root;
2807         struct inode *inode = NULL;
2808         u64 alloc_hint = 0;
2809         int dcs = BTRFS_DC_ERROR;
2810         int num_pages = 0;
2811         int retries = 0;
2812         int ret = 0;
2813
2814         /*
2815          * If this block group is smaller than 100 megs don't bother caching the
2816          * block group.
2817          */
2818         if (block_group->key.offset < (100 * 1024 * 1024)) {
2819                 spin_lock(&block_group->lock);
2820                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2821                 spin_unlock(&block_group->lock);
2822                 return 0;
2823         }
2824
2825 again:
2826         inode = lookup_free_space_inode(root, block_group, path);
2827         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2828                 ret = PTR_ERR(inode);
2829                 btrfs_release_path(path);
2830                 goto out;
2831         }
2832
2833         if (IS_ERR(inode)) {
2834                 BUG_ON(retries);
2835                 retries++;
2836
2837                 if (block_group->ro)
2838                         goto out_free;
2839
2840                 ret = create_free_space_inode(root, trans, block_group, path);
2841                 if (ret)
2842                         goto out_free;
2843                 goto again;
2844         }
2845
2846         /* We've already setup this transaction, go ahead and exit */
2847         if (block_group->cache_generation == trans->transid &&
2848             i_size_read(inode)) {
2849                 dcs = BTRFS_DC_SETUP;
2850                 goto out_put;
2851         }
2852
2853         /*
2854          * We want to set the generation to 0, that way if anything goes wrong
2855          * from here on out we know not to trust this cache when we load up next
2856          * time.
2857          */
2858         BTRFS_I(inode)->generation = 0;
2859         ret = btrfs_update_inode(trans, root, inode);
2860         WARN_ON(ret);
2861
2862         if (i_size_read(inode) > 0) {
2863                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2864                                                       inode);
2865                 if (ret)
2866                         goto out_put;
2867         }
2868
2869         spin_lock(&block_group->lock);
2870         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2871                 /* We're not cached, don't bother trying to write stuff out */
2872                 dcs = BTRFS_DC_WRITTEN;
2873                 spin_unlock(&block_group->lock);
2874                 goto out_put;
2875         }
2876         spin_unlock(&block_group->lock);
2877
2878         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2879         if (!num_pages)
2880                 num_pages = 1;
2881
2882         /*
2883          * Just to make absolutely sure we have enough space, we're going to
2884          * preallocate 12 pages worth of space for each block group.  In
2885          * practice we ought to use at most 8, but we need extra space so we can
2886          * add our header and have a terminator between the extents and the
2887          * bitmaps.
2888          */
2889         num_pages *= 16;
2890         num_pages *= PAGE_CACHE_SIZE;
2891
2892         ret = btrfs_check_data_free_space(inode, num_pages);
2893         if (ret)
2894                 goto out_put;
2895
2896         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2897                                               num_pages, num_pages,
2898                                               &alloc_hint);
2899         if (!ret)
2900                 dcs = BTRFS_DC_SETUP;
2901         btrfs_free_reserved_data_space(inode, num_pages);
2902
2903 out_put:
2904         iput(inode);
2905 out_free:
2906         btrfs_release_path(path);
2907 out:
2908         spin_lock(&block_group->lock);
2909         if (!ret && dcs == BTRFS_DC_SETUP)
2910                 block_group->cache_generation = trans->transid;
2911         block_group->disk_cache_state = dcs;
2912         spin_unlock(&block_group->lock);
2913
2914         return ret;
2915 }
2916
2917 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2918                                    struct btrfs_root *root)
2919 {
2920         struct btrfs_block_group_cache *cache;
2921         int err = 0;
2922         struct btrfs_path *path;
2923         u64 last = 0;
2924
2925         path = btrfs_alloc_path();
2926         if (!path)
2927                 return -ENOMEM;
2928
2929 again:
2930         while (1) {
2931                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2932                 while (cache) {
2933                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2934                                 break;
2935                         cache = next_block_group(root, cache);
2936                 }
2937                 if (!cache) {
2938                         if (last == 0)
2939                                 break;
2940                         last = 0;
2941                         continue;
2942                 }
2943                 err = cache_save_setup(cache, trans, path);
2944                 last = cache->key.objectid + cache->key.offset;
2945                 btrfs_put_block_group(cache);
2946         }
2947
2948         while (1) {
2949                 if (last == 0) {
2950                         err = btrfs_run_delayed_refs(trans, root,
2951                                                      (unsigned long)-1);
2952                         BUG_ON(err);
2953                 }
2954
2955                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2956                 while (cache) {
2957                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2958                                 btrfs_put_block_group(cache);
2959                                 goto again;
2960                         }
2961
2962                         if (cache->dirty)
2963                                 break;
2964                         cache = next_block_group(root, cache);
2965                 }
2966                 if (!cache) {
2967                         if (last == 0)
2968                                 break;
2969                         last = 0;
2970                         continue;
2971                 }
2972
2973                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2974                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2975                 cache->dirty = 0;
2976                 last = cache->key.objectid + cache->key.offset;
2977
2978                 err = write_one_cache_group(trans, root, path, cache);
2979                 BUG_ON(err);
2980                 btrfs_put_block_group(cache);
2981         }
2982
2983         while (1) {
2984                 /*
2985                  * I don't think this is needed since we're just marking our
2986                  * preallocated extent as written, but just in case it can't
2987                  * hurt.
2988                  */
2989                 if (last == 0) {
2990                         err = btrfs_run_delayed_refs(trans, root,
2991                                                      (unsigned long)-1);
2992                         BUG_ON(err);
2993                 }
2994
2995                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2996                 while (cache) {
2997                         /*
2998                          * Really this shouldn't happen, but it could if we
2999                          * couldn't write the entire preallocated extent and
3000                          * splitting the extent resulted in a new block.
3001                          */
3002                         if (cache->dirty) {
3003                                 btrfs_put_block_group(cache);
3004                                 goto again;
3005                         }
3006                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3007                                 break;
3008                         cache = next_block_group(root, cache);
3009                 }
3010                 if (!cache) {
3011                         if (last == 0)
3012                                 break;
3013                         last = 0;
3014                         continue;
3015                 }
3016
3017                 btrfs_write_out_cache(root, trans, cache, path);
3018
3019                 /*
3020                  * If we didn't have an error then the cache state is still
3021                  * NEED_WRITE, so we can set it to WRITTEN.
3022                  */
3023                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3024                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3025                 last = cache->key.objectid + cache->key.offset;
3026                 btrfs_put_block_group(cache);
3027         }
3028
3029         btrfs_free_path(path);
3030         return 0;
3031 }
3032
3033 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3034 {
3035         struct btrfs_block_group_cache *block_group;
3036         int readonly = 0;
3037
3038         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3039         if (!block_group || block_group->ro)
3040                 readonly = 1;
3041         if (block_group)
3042                 btrfs_put_block_group(block_group);
3043         return readonly;
3044 }
3045
3046 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3047                              u64 total_bytes, u64 bytes_used,
3048                              struct btrfs_space_info **space_info)
3049 {
3050         struct btrfs_space_info *found;
3051         int i;
3052         int factor;
3053
3054         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3055                      BTRFS_BLOCK_GROUP_RAID10))
3056                 factor = 2;
3057         else
3058                 factor = 1;
3059
3060         found = __find_space_info(info, flags);
3061         if (found) {
3062                 spin_lock(&found->lock);
3063                 found->total_bytes += total_bytes;
3064                 found->disk_total += total_bytes * factor;
3065                 found->bytes_used += bytes_used;
3066                 found->disk_used += bytes_used * factor;
3067                 found->full = 0;
3068                 spin_unlock(&found->lock);
3069                 *space_info = found;
3070                 return 0;
3071         }
3072         found = kzalloc(sizeof(*found), GFP_NOFS);
3073         if (!found)
3074                 return -ENOMEM;
3075
3076         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3077                 INIT_LIST_HEAD(&found->block_groups[i]);
3078         init_rwsem(&found->groups_sem);
3079         spin_lock_init(&found->lock);
3080         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3081         found->total_bytes = total_bytes;
3082         found->disk_total = total_bytes * factor;
3083         found->bytes_used = bytes_used;
3084         found->disk_used = bytes_used * factor;
3085         found->bytes_pinned = 0;
3086         found->bytes_reserved = 0;
3087         found->bytes_readonly = 0;
3088         found->bytes_may_use = 0;
3089         found->full = 0;
3090         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3091         found->chunk_alloc = 0;
3092         found->flush = 0;
3093         init_waitqueue_head(&found->wait);
3094         *space_info = found;
3095         list_add_rcu(&found->list, &info->space_info);
3096         return 0;
3097 }
3098
3099 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3100 {
3101         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3102
3103         /* chunk -> extended profile */
3104         if (extra_flags == 0)
3105                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3106
3107         if (flags & BTRFS_BLOCK_GROUP_DATA)
3108                 fs_info->avail_data_alloc_bits |= extra_flags;
3109         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3110                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3111         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3112                 fs_info->avail_system_alloc_bits |= extra_flags;
3113 }
3114
3115 /*
3116  * @flags: available profiles in extended format (see ctree.h)
3117  *
3118  * Returns reduced profile in chunk format.  If profile changing is in
3119  * progress (either running or paused) picks the target profile (if it's
3120  * already available), otherwise falls back to plain reducing.
3121  */
3122 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3123 {
3124         /*
3125          * we add in the count of missing devices because we want
3126          * to make sure that any RAID levels on a degraded FS
3127          * continue to be honored.
3128          */
3129         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3130                 root->fs_info->fs_devices->missing_devices;
3131
3132         /* pick restriper's target profile if it's available */
3133         spin_lock(&root->fs_info->balance_lock);
3134         if (root->fs_info->balance_ctl) {
3135                 struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
3136                 u64 tgt = 0;
3137
3138                 if ((flags & BTRFS_BLOCK_GROUP_DATA) &&
3139                     (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3140                     (flags & bctl->data.target)) {
3141                         tgt = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3142                 } else if ((flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
3143                            (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3144                            (flags & bctl->sys.target)) {
3145                         tgt = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3146                 } else if ((flags & BTRFS_BLOCK_GROUP_METADATA) &&
3147                            (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3148                            (flags & bctl->meta.target)) {
3149                         tgt = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3150                 }
3151
3152                 if (tgt) {
3153                         spin_unlock(&root->fs_info->balance_lock);
3154                         flags = tgt;
3155                         goto out;
3156                 }
3157         }
3158         spin_unlock(&root->fs_info->balance_lock);
3159
3160         if (num_devices == 1)
3161                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3162         if (num_devices < 4)
3163                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3164
3165         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3166             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3167                       BTRFS_BLOCK_GROUP_RAID10))) {
3168                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3169         }
3170
3171         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3172             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3173                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3174         }
3175
3176         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3177             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3178              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3179              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3180                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3181         }
3182
3183 out:
3184         /* extended -> chunk profile */
3185         flags &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3186         return flags;
3187 }
3188
3189 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3190 {
3191         if (flags & BTRFS_BLOCK_GROUP_DATA)
3192                 flags |= root->fs_info->avail_data_alloc_bits;
3193         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3194                 flags |= root->fs_info->avail_system_alloc_bits;
3195         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3196                 flags |= root->fs_info->avail_metadata_alloc_bits;
3197
3198         return btrfs_reduce_alloc_profile(root, flags);
3199 }
3200
3201 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3202 {
3203         u64 flags;
3204
3205         if (data)
3206                 flags = BTRFS_BLOCK_GROUP_DATA;
3207         else if (root == root->fs_info->chunk_root)
3208                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3209         else
3210                 flags = BTRFS_BLOCK_GROUP_METADATA;
3211
3212         return get_alloc_profile(root, flags);
3213 }
3214
3215 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3216 {
3217         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3218                                                        BTRFS_BLOCK_GROUP_DATA);
3219 }
3220
3221 /*
3222  * This will check the space that the inode allocates from to make sure we have
3223  * enough space for bytes.
3224  */
3225 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3226 {
3227         struct btrfs_space_info *data_sinfo;
3228         struct btrfs_root *root = BTRFS_I(inode)->root;
3229         u64 used;
3230         int ret = 0, committed = 0, alloc_chunk = 1;
3231
3232         /* make sure bytes are sectorsize aligned */
3233         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3234
3235         if (root == root->fs_info->tree_root ||
3236             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3237                 alloc_chunk = 0;
3238                 committed = 1;
3239         }
3240
3241         data_sinfo = BTRFS_I(inode)->space_info;
3242         if (!data_sinfo)
3243                 goto alloc;
3244
3245 again:
3246         /* make sure we have enough space to handle the data first */
3247         spin_lock(&data_sinfo->lock);
3248         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3249                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3250                 data_sinfo->bytes_may_use;
3251
3252         if (used + bytes > data_sinfo->total_bytes) {
3253                 struct btrfs_trans_handle *trans;
3254
3255                 /*
3256                  * if we don't have enough free bytes in this space then we need
3257                  * to alloc a new chunk.
3258                  */
3259                 if (!data_sinfo->full && alloc_chunk) {
3260                         u64 alloc_target;
3261
3262                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3263                         spin_unlock(&data_sinfo->lock);
3264 alloc:
3265                         alloc_target = btrfs_get_alloc_profile(root, 1);
3266                         trans = btrfs_join_transaction(root);
3267                         if (IS_ERR(trans))
3268                                 return PTR_ERR(trans);
3269
3270                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3271                                              bytes + 2 * 1024 * 1024,
3272                                              alloc_target,
3273                                              CHUNK_ALLOC_NO_FORCE);
3274                         btrfs_end_transaction(trans, root);
3275                         if (ret < 0) {
3276                                 if (ret != -ENOSPC)
3277                                         return ret;
3278                                 else
3279                                         goto commit_trans;
3280                         }
3281
3282                         if (!data_sinfo) {
3283                                 btrfs_set_inode_space_info(root, inode);
3284                                 data_sinfo = BTRFS_I(inode)->space_info;
3285                         }
3286                         goto again;
3287                 }
3288
3289                 /*
3290                  * If we have less pinned bytes than we want to allocate then
3291                  * don't bother committing the transaction, it won't help us.
3292                  */
3293                 if (data_sinfo->bytes_pinned < bytes)
3294                         committed = 1;
3295                 spin_unlock(&data_sinfo->lock);
3296
3297                 /* commit the current transaction and try again */
3298 commit_trans:
3299                 if (!committed &&
3300                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3301                         committed = 1;
3302                         trans = btrfs_join_transaction(root);
3303                         if (IS_ERR(trans))
3304                                 return PTR_ERR(trans);
3305                         ret = btrfs_commit_transaction(trans, root);
3306                         if (ret)
3307                                 return ret;
3308                         goto again;
3309                 }
3310
3311                 return -ENOSPC;
3312         }
3313         data_sinfo->bytes_may_use += bytes;
3314         trace_btrfs_space_reservation(root->fs_info, "space_info",
3315                                       (u64)data_sinfo, bytes, 1);
3316         spin_unlock(&data_sinfo->lock);
3317
3318         return 0;
3319 }
3320
3321 /*
3322  * Called if we need to clear a data reservation for this inode.
3323  */
3324 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3325 {
3326         struct btrfs_root *root = BTRFS_I(inode)->root;
3327         struct btrfs_space_info *data_sinfo;
3328
3329         /* make sure bytes are sectorsize aligned */
3330         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3331
3332         data_sinfo = BTRFS_I(inode)->space_info;
3333         spin_lock(&data_sinfo->lock);
3334         data_sinfo->bytes_may_use -= bytes;
3335         trace_btrfs_space_reservation(root->fs_info, "space_info",
3336                                       (u64)data_sinfo, bytes, 0);
3337         spin_unlock(&data_sinfo->lock);
3338 }
3339
3340 static void force_metadata_allocation(struct btrfs_fs_info *info)
3341 {
3342         struct list_head *head = &info->space_info;
3343         struct btrfs_space_info *found;
3344
3345         rcu_read_lock();
3346         list_for_each_entry_rcu(found, head, list) {
3347                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3348                         found->force_alloc = CHUNK_ALLOC_FORCE;
3349         }
3350         rcu_read_unlock();
3351 }
3352
3353 static int should_alloc_chunk(struct btrfs_root *root,
3354                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3355                               int force)
3356 {
3357         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3358         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3359         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3360         u64 thresh;
3361
3362         if (force == CHUNK_ALLOC_FORCE)
3363                 return 1;
3364
3365         /*
3366          * We need to take into account the global rsv because for all intents
3367          * and purposes it's used space.  Don't worry about locking the
3368          * global_rsv, it doesn't change except when the transaction commits.
3369          */
3370         num_allocated += global_rsv->size;
3371
3372         /*
3373          * in limited mode, we want to have some free space up to
3374          * about 1% of the FS size.
3375          */
3376         if (force == CHUNK_ALLOC_LIMITED) {
3377                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3378                 thresh = max_t(u64, 64 * 1024 * 1024,
3379                                div_factor_fine(thresh, 1));
3380
3381                 if (num_bytes - num_allocated < thresh)
3382                         return 1;
3383         }
3384         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3385
3386         /* 256MB or 2% of the FS */
3387         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3388         /* system chunks need a much small threshold */
3389         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3390                 thresh = 32 * 1024 * 1024;
3391
3392         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3393                 return 0;
3394         return 1;
3395 }
3396
3397 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3398                           struct btrfs_root *extent_root, u64 alloc_bytes,
3399                           u64 flags, int force)
3400 {
3401         struct btrfs_space_info *space_info;
3402         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3403         int wait_for_alloc = 0;
3404         int ret = 0;
3405
3406         BUG_ON(!profile_is_valid(flags, 0));
3407
3408         space_info = __find_space_info(extent_root->fs_info, flags);
3409         if (!space_info) {
3410                 ret = update_space_info(extent_root->fs_info, flags,
3411                                         0, 0, &space_info);
3412                 BUG_ON(ret);
3413         }
3414         BUG_ON(!space_info);
3415
3416 again:
3417         spin_lock(&space_info->lock);
3418         if (force < space_info->force_alloc)
3419                 force = space_info->force_alloc;
3420         if (space_info->full) {
3421                 spin_unlock(&space_info->lock);
3422                 return 0;
3423         }
3424
3425         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3426                 spin_unlock(&space_info->lock);
3427                 return 0;
3428         } else if (space_info->chunk_alloc) {
3429                 wait_for_alloc = 1;
3430         } else {
3431                 space_info->chunk_alloc = 1;
3432         }
3433
3434         spin_unlock(&space_info->lock);
3435
3436         mutex_lock(&fs_info->chunk_mutex);
3437
3438         /*
3439          * The chunk_mutex is held throughout the entirety of a chunk
3440          * allocation, so once we've acquired the chunk_mutex we know that the
3441          * other guy is done and we need to recheck and see if we should
3442          * allocate.
3443          */
3444         if (wait_for_alloc) {
3445                 mutex_unlock(&fs_info->chunk_mutex);
3446                 wait_for_alloc = 0;
3447                 goto again;
3448         }
3449
3450         /*
3451          * If we have mixed data/metadata chunks we want to make sure we keep
3452          * allocating mixed chunks instead of individual chunks.
3453          */
3454         if (btrfs_mixed_space_info(space_info))
3455                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3456
3457         /*
3458          * if we're doing a data chunk, go ahead and make sure that
3459          * we keep a reasonable number of metadata chunks allocated in the
3460          * FS as well.
3461          */
3462         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3463                 fs_info->data_chunk_allocations++;
3464                 if (!(fs_info->data_chunk_allocations %
3465                       fs_info->metadata_ratio))
3466                         force_metadata_allocation(fs_info);
3467         }
3468
3469         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3470         if (ret < 0 && ret != -ENOSPC)
3471                 goto out;
3472
3473         spin_lock(&space_info->lock);
3474         if (ret)
3475                 space_info->full = 1;
3476         else
3477                 ret = 1;
3478
3479         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3480         space_info->chunk_alloc = 0;
3481         spin_unlock(&space_info->lock);
3482 out:
3483         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3484         return ret;
3485 }
3486
3487 /*
3488  * shrink metadata reservation for delalloc
3489  */
3490 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3491                            bool wait_ordered)
3492 {
3493         struct btrfs_block_rsv *block_rsv;
3494         struct btrfs_space_info *space_info;
3495         struct btrfs_trans_handle *trans;
3496         u64 reserved;
3497         u64 max_reclaim;
3498         u64 reclaimed = 0;
3499         long time_left;
3500         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3501         int loops = 0;
3502         unsigned long progress;
3503
3504         trans = (struct btrfs_trans_handle *)current->journal_info;
3505         block_rsv = &root->fs_info->delalloc_block_rsv;
3506         space_info = block_rsv->space_info;
3507
3508         smp_mb();
3509         reserved = space_info->bytes_may_use;
3510         progress = space_info->reservation_progress;
3511
3512         if (reserved == 0)
3513                 return 0;
3514
3515         smp_mb();
3516         if (root->fs_info->delalloc_bytes == 0) {
3517                 if (trans)
3518                         return 0;
3519                 btrfs_wait_ordered_extents(root, 0, 0);
3520                 return 0;
3521         }
3522
3523         max_reclaim = min(reserved, to_reclaim);
3524         nr_pages = max_t(unsigned long, nr_pages,
3525                          max_reclaim >> PAGE_CACHE_SHIFT);
3526         while (loops < 1024) {
3527                 /* have the flusher threads jump in and do some IO */
3528                 smp_mb();
3529                 nr_pages = min_t(unsigned long, nr_pages,
3530                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3531                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3532                                                 WB_REASON_FS_FREE_SPACE);
3533
3534                 spin_lock(&space_info->lock);
3535                 if (reserved > space_info->bytes_may_use)
3536                         reclaimed += reserved - space_info->bytes_may_use;
3537                 reserved = space_info->bytes_may_use;
3538                 spin_unlock(&space_info->lock);
3539
3540                 loops++;
3541
3542                 if (reserved == 0 || reclaimed >= max_reclaim)
3543                         break;
3544
3545                 if (trans && trans->transaction->blocked)
3546                         return -EAGAIN;
3547
3548                 if (wait_ordered && !trans) {
3549                         btrfs_wait_ordered_extents(root, 0, 0);
3550                 } else {
3551                         time_left = schedule_timeout_interruptible(1);
3552
3553                         /* We were interrupted, exit */
3554                         if (time_left)
3555                                 break;
3556                 }
3557
3558                 /* we've kicked the IO a few times, if anything has been freed,
3559                  * exit.  There is no sense in looping here for a long time
3560                  * when we really need to commit the transaction, or there are
3561                  * just too many writers without enough free space
3562                  */
3563
3564                 if (loops > 3) {
3565                         smp_mb();
3566                         if (progress != space_info->reservation_progress)
3567                                 break;
3568                 }
3569
3570         }
3571
3572         return reclaimed >= to_reclaim;
3573 }
3574
3575 /**
3576  * maybe_commit_transaction - possibly commit the transaction if its ok to
3577  * @root - the root we're allocating for
3578  * @bytes - the number of bytes we want to reserve
3579  * @force - force the commit
3580  *
3581  * This will check to make sure that committing the transaction will actually
3582  * get us somewhere and then commit the transaction if it does.  Otherwise it
3583  * will return -ENOSPC.
3584  */
3585 static int may_commit_transaction(struct btrfs_root *root,
3586                                   struct btrfs_space_info *space_info,
3587                                   u64 bytes, int force)
3588 {
3589         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3590         struct btrfs_trans_handle *trans;
3591
3592         trans = (struct btrfs_trans_handle *)current->journal_info;
3593         if (trans)
3594                 return -EAGAIN;
3595
3596         if (force)
3597                 goto commit;
3598
3599         /* See if there is enough pinned space to make this reservation */
3600         spin_lock(&space_info->lock);
3601         if (space_info->bytes_pinned >= bytes) {
3602                 spin_unlock(&space_info->lock);
3603                 goto commit;
3604         }
3605         spin_unlock(&space_info->lock);
3606
3607         /*
3608          * See if there is some space in the delayed insertion reservation for
3609          * this reservation.
3610          */
3611         if (space_info != delayed_rsv->space_info)
3612                 return -ENOSPC;
3613
3614         spin_lock(&delayed_rsv->lock);
3615         if (delayed_rsv->size < bytes) {
3616                 spin_unlock(&delayed_rsv->lock);
3617                 return -ENOSPC;
3618         }
3619         spin_unlock(&delayed_rsv->lock);
3620
3621 commit:
3622         trans = btrfs_join_transaction(root);
3623         if (IS_ERR(trans))
3624                 return -ENOSPC;
3625
3626         return btrfs_commit_transaction(trans, root);
3627 }
3628
3629 /**
3630  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3631  * @root - the root we're allocating for
3632  * @block_rsv - the block_rsv we're allocating for
3633  * @orig_bytes - the number of bytes we want
3634  * @flush - wether or not we can flush to make our reservation
3635  *
3636  * This will reserve orgi_bytes number of bytes from the space info associated
3637  * with the block_rsv.  If there is not enough space it will make an attempt to
3638  * flush out space to make room.  It will do this by flushing delalloc if
3639  * possible or committing the transaction.  If flush is 0 then no attempts to
3640  * regain reservations will be made and this will fail if there is not enough
3641  * space already.
3642  */
3643 static int reserve_metadata_bytes(struct btrfs_root *root,
3644                                   struct btrfs_block_rsv *block_rsv,
3645                                   u64 orig_bytes, int flush)
3646 {
3647         struct btrfs_space_info *space_info = block_rsv->space_info;
3648         u64 used;
3649         u64 num_bytes = orig_bytes;
3650         int retries = 0;
3651         int ret = 0;
3652         bool committed = false;
3653         bool flushing = false;
3654         bool wait_ordered = false;
3655
3656 again:
3657         ret = 0;
3658         spin_lock(&space_info->lock);
3659         /*
3660          * We only want to wait if somebody other than us is flushing and we are
3661          * actually alloed to flush.
3662          */
3663         while (flush && !flushing && space_info->flush) {
3664                 spin_unlock(&space_info->lock);
3665                 /*
3666                  * If we have a trans handle we can't wait because the flusher
3667                  * may have to commit the transaction, which would mean we would
3668                  * deadlock since we are waiting for the flusher to finish, but
3669                  * hold the current transaction open.
3670                  */
3671                 if (current->journal_info)
3672                         return -EAGAIN;
3673                 ret = wait_event_interruptible(space_info->wait,
3674                                                !space_info->flush);
3675                 /* Must have been interrupted, return */
3676                 if (ret)
3677                         return -EINTR;
3678
3679                 spin_lock(&space_info->lock);
3680         }
3681
3682         ret = -ENOSPC;
3683         used = space_info->bytes_used + space_info->bytes_reserved +
3684                 space_info->bytes_pinned + space_info->bytes_readonly +
3685                 space_info->bytes_may_use;
3686
3687         /*
3688          * The idea here is that we've not already over-reserved the block group
3689          * then we can go ahead and save our reservation first and then start
3690          * flushing if we need to.  Otherwise if we've already overcommitted
3691          * lets start flushing stuff first and then come back and try to make
3692          * our reservation.
3693          */
3694         if (used <= space_info->total_bytes) {
3695                 if (used + orig_bytes <= space_info->total_bytes) {
3696                         space_info->bytes_may_use += orig_bytes;
3697                         trace_btrfs_space_reservation(root->fs_info,
3698                                                       "space_info",
3699                                                       (u64)space_info,
3700                                                       orig_bytes, 1);
3701                         ret = 0;
3702                 } else {
3703                         /*
3704                          * Ok set num_bytes to orig_bytes since we aren't
3705                          * overocmmitted, this way we only try and reclaim what
3706                          * we need.
3707                          */
3708                         num_bytes = orig_bytes;
3709                 }
3710         } else {
3711                 /*
3712                  * Ok we're over committed, set num_bytes to the overcommitted
3713                  * amount plus the amount of bytes that we need for this
3714                  * reservation.
3715                  */
3716                 wait_ordered = true;
3717                 num_bytes = used - space_info->total_bytes +
3718                         (orig_bytes * (retries + 1));
3719         }
3720
3721         if (ret) {
3722                 u64 profile = btrfs_get_alloc_profile(root, 0);
3723                 u64 avail;
3724
3725                 /*
3726                  * If we have a lot of space that's pinned, don't bother doing
3727                  * the overcommit dance yet and just commit the transaction.
3728                  */
3729                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3730                 do_div(avail, 10);
3731                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3732                         space_info->flush = 1;
3733                         flushing = true;
3734                         spin_unlock(&space_info->lock);
3735                         ret = may_commit_transaction(root, space_info,
3736                                                      orig_bytes, 1);
3737                         if (ret)
3738                                 goto out;
3739                         committed = true;
3740                         goto again;
3741                 }
3742
3743                 spin_lock(&root->fs_info->free_chunk_lock);
3744                 avail = root->fs_info->free_chunk_space;
3745
3746                 /*
3747                  * If we have dup, raid1 or raid10 then only half of the free
3748                  * space is actually useable.
3749                  */
3750                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3751                                BTRFS_BLOCK_GROUP_RAID1 |
3752                                BTRFS_BLOCK_GROUP_RAID10))
3753                         avail >>= 1;
3754
3755                 /*
3756                  * If we aren't flushing don't let us overcommit too much, say
3757                  * 1/8th of the space.  If we can flush, let it overcommit up to
3758                  * 1/2 of the space.
3759                  */
3760                 if (flush)
3761                         avail >>= 3;
3762                 else
3763                         avail >>= 1;
3764                  spin_unlock(&root->fs_info->free_chunk_lock);
3765
3766                 if (used + num_bytes < space_info->total_bytes + avail) {
3767                         space_info->bytes_may_use += orig_bytes;
3768                         trace_btrfs_space_reservation(root->fs_info,
3769                                                       "space_info",
3770                                                       (u64)space_info,
3771                                                       orig_bytes, 1);
3772                         ret = 0;
3773                 } else {
3774                         wait_ordered = true;
3775                 }
3776         }
3777
3778         /*
3779          * Couldn't make our reservation, save our place so while we're trying
3780          * to reclaim space we can actually use it instead of somebody else
3781          * stealing it from us.
3782          */
3783         if (ret && flush) {
3784                 flushing = true;
3785                 space_info->flush = 1;
3786         }
3787
3788         spin_unlock(&space_info->lock);
3789
3790         if (!ret || !flush)
3791                 goto out;
3792
3793         /*
3794          * We do synchronous shrinking since we don't actually unreserve
3795          * metadata until after the IO is completed.
3796          */
3797         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3798         if (ret < 0)
3799                 goto out;
3800
3801         ret = 0;
3802
3803         /*
3804          * So if we were overcommitted it's possible that somebody else flushed
3805          * out enough space and we simply didn't have enough space to reclaim,
3806          * so go back around and try again.
3807          */
3808         if (retries < 2) {
3809                 wait_ordered = true;
3810                 retries++;
3811                 goto again;
3812         }
3813
3814         ret = -ENOSPC;
3815         if (committed)
3816                 goto out;
3817
3818         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3819         if (!ret) {
3820                 committed = true;
3821                 goto again;
3822         }
3823
3824 out:
3825         if (flushing) {
3826                 spin_lock(&space_info->lock);
3827                 space_info->flush = 0;
3828                 wake_up_all(&space_info->wait);
3829                 spin_unlock(&space_info->lock);
3830         }
3831         return ret;
3832 }
3833
3834 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3835                                              struct btrfs_root *root)
3836 {
3837         struct btrfs_block_rsv *block_rsv = NULL;
3838
3839         if (root->ref_cows || root == root->fs_info->csum_root)
3840                 block_rsv = trans->block_rsv;
3841
3842         if (!block_rsv)
3843                 block_rsv = root->block_rsv;
3844
3845         if (!block_rsv)
3846                 block_rsv = &root->fs_info->empty_block_rsv;
3847
3848         return block_rsv;
3849 }
3850
3851 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3852                                u64 num_bytes)
3853 {
3854         int ret = -ENOSPC;
3855         spin_lock(&block_rsv->lock);
3856         if (block_rsv->reserved >= num_bytes) {
3857                 block_rsv->reserved -= num_bytes;
3858                 if (block_rsv->reserved < block_rsv->size)
3859                         block_rsv->full = 0;
3860                 ret = 0;
3861         }
3862         spin_unlock(&block_rsv->lock);
3863         return ret;
3864 }
3865
3866 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3867                                 u64 num_bytes, int update_size)
3868 {
3869         spin_lock(&block_rsv->lock);
3870         block_rsv->reserved += num_bytes;
3871         if (update_size)
3872                 block_rsv->size += num_bytes;
3873         else if (block_rsv->reserved >= block_rsv->size)
3874                 block_rsv->full = 1;
3875         spin_unlock(&block_rsv->lock);
3876 }
3877
3878 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3879                                     struct btrfs_block_rsv *block_rsv,
3880                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3881 {
3882         struct btrfs_space_info *space_info = block_rsv->space_info;
3883
3884         spin_lock(&block_rsv->lock);
3885         if (num_bytes == (u64)-1)
3886                 num_bytes = block_rsv->size;
3887         block_rsv->size -= num_bytes;
3888         if (block_rsv->reserved >= block_rsv->size) {
3889                 num_bytes = block_rsv->reserved - block_rsv->size;
3890                 block_rsv->reserved = block_rsv->size;
3891                 block_rsv->full = 1;
3892         } else {
3893                 num_bytes = 0;
3894         }
3895         spin_unlock(&block_rsv->lock);
3896
3897         if (num_bytes > 0) {
3898                 if (dest) {
3899                         spin_lock(&dest->lock);
3900                         if (!dest->full) {
3901                                 u64 bytes_to_add;
3902
3903                                 bytes_to_add = dest->size - dest->reserved;
3904                                 bytes_to_add = min(num_bytes, bytes_to_add);
3905                                 dest->reserved += bytes_to_add;
3906                                 if (dest->reserved >= dest->size)
3907                                         dest->full = 1;
3908                                 num_bytes -= bytes_to_add;
3909                         }
3910                         spin_unlock(&dest->lock);
3911                 }
3912                 if (num_bytes) {
3913                         spin_lock(&space_info->lock);
3914                         space_info->bytes_may_use -= num_bytes;
3915                         trace_btrfs_space_reservation(fs_info, "space_info",
3916                                                       (u64)space_info,
3917                                                       num_bytes, 0);
3918                         space_info->reservation_progress++;
3919                         spin_unlock(&space_info->lock);
3920                 }
3921         }
3922 }
3923
3924 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3925                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3926 {
3927         int ret;
3928
3929         ret = block_rsv_use_bytes(src, num_bytes);
3930         if (ret)
3931                 return ret;
3932
3933         block_rsv_add_bytes(dst, num_bytes, 1);
3934         return 0;
3935 }
3936
3937 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3938 {
3939         memset(rsv, 0, sizeof(*rsv));
3940         spin_lock_init(&rsv->lock);
3941 }
3942
3943 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3944 {
3945         struct btrfs_block_rsv *block_rsv;
3946         struct btrfs_fs_info *fs_info = root->fs_info;
3947
3948         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3949         if (!block_rsv)
3950                 return NULL;
3951
3952         btrfs_init_block_rsv(block_rsv);
3953         block_rsv->space_info = __find_space_info(fs_info,
3954                                                   BTRFS_BLOCK_GROUP_METADATA);
3955         return block_rsv;
3956 }
3957
3958 void btrfs_free_block_rsv(struct btrfs_root *root,
3959                           struct btrfs_block_rsv *rsv)
3960 {
3961         btrfs_block_rsv_release(root, rsv, (u64)-1);
3962         kfree(rsv);
3963 }
3964
3965 static inline int __block_rsv_add(struct btrfs_root *root,
3966                                   struct btrfs_block_rsv *block_rsv,
3967                                   u64 num_bytes, int flush)
3968 {
3969         int ret;
3970
3971         if (num_bytes == 0)
3972                 return 0;
3973
3974         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3975         if (!ret) {
3976                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3977                 return 0;
3978         }
3979
3980         return ret;
3981 }
3982
3983 int btrfs_block_rsv_add(struct btrfs_root *root,
3984                         struct btrfs_block_rsv *block_rsv,
3985                         u64 num_bytes)
3986 {
3987         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3988 }
3989
3990 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3991                                 struct btrfs_block_rsv *block_rsv,
3992                                 u64 num_bytes)
3993 {
3994         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3995 }
3996
3997 int btrfs_block_rsv_check(struct btrfs_root *root,
3998                           struct btrfs_block_rsv *block_rsv, int min_factor)
3999 {
4000         u64 num_bytes = 0;
4001         int ret = -ENOSPC;
4002
4003         if (!block_rsv)
4004                 return 0;
4005
4006         spin_lock(&block_rsv->lock);
4007         num_bytes = div_factor(block_rsv->size, min_factor);
4008         if (block_rsv->reserved >= num_bytes)
4009                 ret = 0;
4010         spin_unlock(&block_rsv->lock);
4011
4012         return ret;
4013 }
4014
4015 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4016                                            struct btrfs_block_rsv *block_rsv,
4017                                            u64 min_reserved, int flush)
4018 {
4019         u64 num_bytes = 0;
4020         int ret = -ENOSPC;
4021
4022         if (!block_rsv)
4023                 return 0;
4024
4025         spin_lock(&block_rsv->lock);
4026         num_bytes = min_reserved;
4027         if (block_rsv->reserved >= num_bytes)
4028                 ret = 0;
4029         else
4030                 num_bytes -= block_rsv->reserved;
4031         spin_unlock(&block_rsv->lock);
4032
4033         if (!ret)
4034                 return 0;
4035
4036         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4037         if (!ret) {
4038                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4039                 return 0;
4040         }
4041
4042         return ret;
4043 }
4044
4045 int btrfs_block_rsv_refill(struct btrfs_root *root,
4046                            struct btrfs_block_rsv *block_rsv,
4047                            u64 min_reserved)
4048 {
4049         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4050 }
4051
4052 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4053                                    struct btrfs_block_rsv *block_rsv,
4054                                    u64 min_reserved)
4055 {
4056         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4057 }
4058
4059 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4060                             struct btrfs_block_rsv *dst_rsv,
4061                             u64 num_bytes)
4062 {
4063         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4064 }
4065
4066 void btrfs_block_rsv_release(struct btrfs_root *root,
4067                              struct btrfs_block_rsv *block_rsv,
4068                              u64 num_bytes)
4069 {
4070         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4071         if (global_rsv->full || global_rsv == block_rsv ||
4072             block_rsv->space_info != global_rsv->space_info)
4073                 global_rsv = NULL;
4074         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4075                                 num_bytes);
4076 }
4077
4078 /*
4079  * helper to calculate size of global block reservation.
4080  * the desired value is sum of space used by extent tree,
4081  * checksum tree and root tree
4082  */
4083 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4084 {
4085         struct btrfs_space_info *sinfo;
4086         u64 num_bytes;
4087         u64 meta_used;
4088         u64 data_used;
4089         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4090
4091         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4092         spin_lock(&sinfo->lock);
4093         data_used = sinfo->bytes_used;
4094         spin_unlock(&sinfo->lock);
4095
4096         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4097         spin_lock(&sinfo->lock);
4098         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4099                 data_used = 0;
4100         meta_used = sinfo->bytes_used;
4101         spin_unlock(&sinfo->lock);
4102
4103         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4104                     csum_size * 2;
4105         num_bytes += div64_u64(data_used + meta_used, 50);
4106
4107         if (num_bytes * 3 > meta_used)
4108                 num_bytes = div64_u64(meta_used, 3);
4109
4110         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4111 }
4112
4113 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4114 {
4115         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4116         struct btrfs_space_info *sinfo = block_rsv->space_info;
4117         u64 num_bytes;
4118
4119         num_bytes = calc_global_metadata_size(fs_info);
4120
4121         spin_lock(&block_rsv->lock);
4122         spin_lock(&sinfo->lock);
4123
4124         block_rsv->size = num_bytes;
4125
4126         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4127                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4128                     sinfo->bytes_may_use;
4129
4130         if (sinfo->total_bytes > num_bytes) {
4131                 num_bytes = sinfo->total_bytes - num_bytes;
4132                 block_rsv->reserved += num_bytes;
4133                 sinfo->bytes_may_use += num_bytes;
4134                 trace_btrfs_space_reservation(fs_info, "space_info",
4135                                               (u64)sinfo, num_bytes, 1);
4136         }
4137
4138         if (block_rsv->reserved >= block_rsv->size) {
4139                 num_bytes = block_rsv->reserved - block_rsv->size;
4140                 sinfo->bytes_may_use -= num_bytes;
4141                 trace_btrfs_space_reservation(fs_info, "space_info",
4142                                               (u64)sinfo, num_bytes, 0);
4143                 sinfo->reservation_progress++;
4144                 block_rsv->reserved = block_rsv->size;
4145                 block_rsv->full = 1;
4146         }
4147
4148         spin_unlock(&sinfo->lock);
4149         spin_unlock(&block_rsv->lock);
4150 }
4151
4152 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4153 {
4154         struct btrfs_space_info *space_info;
4155
4156         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4157         fs_info->chunk_block_rsv.space_info = space_info;
4158
4159         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4160         fs_info->global_block_rsv.space_info = space_info;
4161         fs_info->delalloc_block_rsv.space_info = space_info;
4162         fs_info->trans_block_rsv.space_info = space_info;
4163         fs_info->empty_block_rsv.space_info = space_info;
4164         fs_info->delayed_block_rsv.space_info = space_info;
4165
4166         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4167         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4168         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4169         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4170         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4171
4172         update_global_block_rsv(fs_info);
4173 }
4174
4175 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4176 {
4177         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4178                                 (u64)-1);
4179         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4180         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4181         WARN_ON(fs_info->trans_block_rsv.size > 0);
4182         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4183         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4184         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4185         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4186         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4187 }
4188
4189 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4190                                   struct btrfs_root *root)
4191 {
4192         if (!trans->bytes_reserved)
4193                 return;
4194
4195         trace_btrfs_space_reservation(root->fs_info, "transaction", (u64)trans,
4196                                       trans->bytes_reserved, 0);
4197         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4198         trans->bytes_reserved = 0;
4199 }
4200
4201 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4202                                   struct inode *inode)
4203 {
4204         struct btrfs_root *root = BTRFS_I(inode)->root;
4205         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4206         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4207
4208         /*
4209          * We need to hold space in order to delete our orphan item once we've
4210          * added it, so this takes the reservation so we can release it later
4211          * when we are truly done with the orphan item.
4212          */
4213         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4214         trace_btrfs_space_reservation(root->fs_info, "orphan",
4215                                       btrfs_ino(inode), num_bytes, 1);
4216         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4217 }
4218
4219 void btrfs_orphan_release_metadata(struct inode *inode)
4220 {
4221         struct btrfs_root *root = BTRFS_I(inode)->root;
4222         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4223         trace_btrfs_space_reservation(root->fs_info, "orphan",
4224                                       btrfs_ino(inode), num_bytes, 0);
4225         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4226 }
4227
4228 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4229                                 struct btrfs_pending_snapshot *pending)
4230 {
4231         struct btrfs_root *root = pending->root;
4232         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4233         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4234         /*
4235          * two for root back/forward refs, two for directory entries
4236          * and one for root of the snapshot.
4237          */
4238         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4239         dst_rsv->space_info = src_rsv->space_info;
4240         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4241 }
4242
4243 /**
4244  * drop_outstanding_extent - drop an outstanding extent
4245  * @inode: the inode we're dropping the extent for
4246  *
4247  * This is called when we are freeing up an outstanding extent, either called
4248  * after an error or after an extent is written.  This will return the number of
4249  * reserved extents that need to be freed.  This must be called with
4250  * BTRFS_I(inode)->lock held.
4251  */
4252 static unsigned drop_outstanding_extent(struct inode *inode)
4253 {
4254         unsigned drop_inode_space = 0;
4255         unsigned dropped_extents = 0;
4256
4257         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4258         BTRFS_I(inode)->outstanding_extents--;
4259
4260         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4261             BTRFS_I(inode)->delalloc_meta_reserved) {
4262                 drop_inode_space = 1;
4263                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4264         }
4265
4266         /*
4267          * If we have more or the same amount of outsanding extents than we have
4268          * reserved then we need to leave the reserved extents count alone.
4269          */
4270         if (BTRFS_I(inode)->outstanding_extents >=
4271             BTRFS_I(inode)->reserved_extents)
4272                 return drop_inode_space;
4273
4274         dropped_extents = BTRFS_I(inode)->reserved_extents -
4275                 BTRFS_I(inode)->outstanding_extents;
4276         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4277         return dropped_extents + drop_inode_space;
4278 }
4279
4280 /**
4281  * calc_csum_metadata_size - return the amount of metada space that must be
4282  *      reserved/free'd for the given bytes.
4283  * @inode: the inode we're manipulating
4284  * @num_bytes: the number of bytes in question
4285  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4286  *
4287  * This adjusts the number of csum_bytes in the inode and then returns the
4288  * correct amount of metadata that must either be reserved or freed.  We
4289  * calculate how many checksums we can fit into one leaf and then divide the
4290  * number of bytes that will need to be checksumed by this value to figure out
4291  * how many checksums will be required.  If we are adding bytes then the number
4292  * may go up and we will return the number of additional bytes that must be
4293  * reserved.  If it is going down we will return the number of bytes that must
4294  * be freed.
4295  *
4296  * This must be called with BTRFS_I(inode)->lock held.
4297  */
4298 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4299                                    int reserve)
4300 {
4301         struct btrfs_root *root = BTRFS_I(inode)->root;
4302         u64 csum_size;
4303         int num_csums_per_leaf;
4304         int num_csums;
4305         int old_csums;
4306
4307         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4308             BTRFS_I(inode)->csum_bytes == 0)
4309                 return 0;
4310
4311         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4312         if (reserve)
4313                 BTRFS_I(inode)->csum_bytes += num_bytes;
4314         else
4315                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4316         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4317         num_csums_per_leaf = (int)div64_u64(csum_size,
4318                                             sizeof(struct btrfs_csum_item) +
4319                                             sizeof(struct btrfs_disk_key));
4320         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4321         num_csums = num_csums + num_csums_per_leaf - 1;
4322         num_csums = num_csums / num_csums_per_leaf;
4323
4324         old_csums = old_csums + num_csums_per_leaf - 1;
4325         old_csums = old_csums / num_csums_per_leaf;
4326
4327         /* No change, no need to reserve more */
4328         if (old_csums == num_csums)
4329                 return 0;
4330
4331         if (reserve)
4332                 return btrfs_calc_trans_metadata_size(root,
4333                                                       num_csums - old_csums);
4334
4335         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4336 }
4337
4338 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4339 {
4340         struct btrfs_root *root = BTRFS_I(inode)->root;
4341         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4342         u64 to_reserve = 0;
4343         u64 csum_bytes;
4344         unsigned nr_extents = 0;
4345         int extra_reserve = 0;
4346         int flush = 1;
4347         int ret;
4348
4349         /* Need to be holding the i_mutex here if we aren't free space cache */
4350         if (btrfs_is_free_space_inode(root, inode))
4351                 flush = 0;
4352
4353         if (flush && btrfs_transaction_in_commit(root->fs_info))
4354                 schedule_timeout(1);
4355
4356         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4357         num_bytes = ALIGN(num_bytes, root->sectorsize);
4358
4359         spin_lock(&BTRFS_I(inode)->lock);
4360         BTRFS_I(inode)->outstanding_extents++;
4361
4362         if (BTRFS_I(inode)->outstanding_extents >
4363             BTRFS_I(inode)->reserved_extents)
4364                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4365                         BTRFS_I(inode)->reserved_extents;
4366
4367         /*
4368          * Add an item to reserve for updating the inode when we complete the
4369          * delalloc io.
4370          */
4371         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4372                 nr_extents++;
4373                 extra_reserve = 1;
4374         }
4375
4376         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4377         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4378         csum_bytes = BTRFS_I(inode)->csum_bytes;
4379         spin_unlock(&BTRFS_I(inode)->lock);
4380
4381         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4382         if (ret) {
4383                 u64 to_free = 0;
4384                 unsigned dropped;
4385
4386                 spin_lock(&BTRFS_I(inode)->lock);
4387                 dropped = drop_outstanding_extent(inode);
4388                 /*
4389                  * If the inodes csum_bytes is the same as the original
4390                  * csum_bytes then we know we haven't raced with any free()ers
4391                  * so we can just reduce our inodes csum bytes and carry on.
4392                  * Otherwise we have to do the normal free thing to account for
4393                  * the case that the free side didn't free up its reserve
4394                  * because of this outstanding reservation.
4395                  */
4396                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4397                         calc_csum_metadata_size(inode, num_bytes, 0);
4398                 else
4399                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4400                 spin_unlock(&BTRFS_I(inode)->lock);
4401                 if (dropped)
4402                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4403
4404                 if (to_free) {
4405                         btrfs_block_rsv_release(root, block_rsv, to_free);
4406                         trace_btrfs_space_reservation(root->fs_info,
4407                                                       "delalloc",
4408                                                       btrfs_ino(inode),
4409                                                       to_free, 0);
4410                 }
4411                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4412                 return ret;
4413         }
4414
4415         spin_lock(&BTRFS_I(inode)->lock);
4416         if (extra_reserve) {
4417                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4418                 nr_extents--;
4419         }
4420         BTRFS_I(inode)->reserved_extents += nr_extents;
4421         spin_unlock(&BTRFS_I(inode)->lock);
4422         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4423
4424         if (to_reserve)
4425                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4426                                               btrfs_ino(inode), to_reserve, 1);
4427         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4428
4429         return 0;
4430 }
4431
4432 /**
4433  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4434  * @inode: the inode to release the reservation for
4435  * @num_bytes: the number of bytes we're releasing
4436  *
4437  * This will release the metadata reservation for an inode.  This can be called
4438  * once we complete IO for a given set of bytes to release their metadata
4439  * reservations.
4440  */
4441 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4442 {
4443         struct btrfs_root *root = BTRFS_I(inode)->root;
4444         u64 to_free = 0;
4445         unsigned dropped;
4446
4447         num_bytes = ALIGN(num_bytes, root->sectorsize);
4448         spin_lock(&BTRFS_I(inode)->lock);
4449         dropped = drop_outstanding_extent(inode);
4450
4451         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4452         spin_unlock(&BTRFS_I(inode)->lock);
4453         if (dropped > 0)
4454                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4455
4456         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4457                                       btrfs_ino(inode), to_free, 0);
4458         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4459                                 to_free);
4460 }
4461
4462 /**
4463  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4464  * @inode: inode we're writing to
4465  * @num_bytes: the number of bytes we want to allocate
4466  *
4467  * This will do the following things
4468  *
4469  * o reserve space in the data space info for num_bytes
4470  * o reserve space in the metadata space info based on number of outstanding
4471  *   extents and how much csums will be needed
4472  * o add to the inodes ->delalloc_bytes
4473  * o add it to the fs_info's delalloc inodes list.
4474  *
4475  * This will return 0 for success and -ENOSPC if there is no space left.
4476  */
4477 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4478 {
4479         int ret;
4480
4481         ret = btrfs_check_data_free_space(inode, num_bytes);
4482         if (ret)
4483                 return ret;
4484
4485         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4486         if (ret) {
4487                 btrfs_free_reserved_data_space(inode, num_bytes);
4488                 return ret;
4489         }
4490
4491         return 0;
4492 }
4493
4494 /**
4495  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4496  * @inode: inode we're releasing space for
4497  * @num_bytes: the number of bytes we want to free up
4498  *
4499  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4500  * called in the case that we don't need the metadata AND data reservations
4501  * anymore.  So if there is an error or we insert an inline extent.
4502  *
4503  * This function will release the metadata space that was not used and will
4504  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4505  * list if there are no delalloc bytes left.
4506  */
4507 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4508 {
4509         btrfs_delalloc_release_metadata(inode, num_bytes);
4510         btrfs_free_reserved_data_space(inode, num_bytes);
4511 }
4512
4513 static int update_block_group(struct btrfs_trans_handle *trans,
4514                               struct btrfs_root *root,
4515                               u64 bytenr, u64 num_bytes, int alloc)
4516 {
4517         struct btrfs_block_group_cache *cache = NULL;
4518         struct btrfs_fs_info *info = root->fs_info;
4519         u64 total = num_bytes;
4520         u64 old_val;
4521         u64 byte_in_group;
4522         int factor;
4523
4524         /* block accounting for super block */
4525         spin_lock(&info->delalloc_lock);
4526         old_val = btrfs_super_bytes_used(info->super_copy);
4527         if (alloc)
4528                 old_val += num_bytes;
4529         else
4530                 old_val -= num_bytes;
4531         btrfs_set_super_bytes_used(info->super_copy, old_val);
4532         spin_unlock(&info->delalloc_lock);
4533
4534         while (total) {
4535                 cache = btrfs_lookup_block_group(info, bytenr);
4536                 if (!cache)
4537                         return -1;
4538                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4539                                     BTRFS_BLOCK_GROUP_RAID1 |
4540                                     BTRFS_BLOCK_GROUP_RAID10))
4541                         factor = 2;
4542                 else
4543                         factor = 1;
4544                 /*
4545                  * If this block group has free space cache written out, we
4546                  * need to make sure to load it if we are removing space.  This
4547                  * is because we need the unpinning stage to actually add the
4548                  * space back to the block group, otherwise we will leak space.
4549                  */
4550                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4551                         cache_block_group(cache, trans, NULL, 1);
4552
4553                 byte_in_group = bytenr - cache->key.objectid;
4554                 WARN_ON(byte_in_group > cache->key.offset);
4555
4556                 spin_lock(&cache->space_info->lock);
4557                 spin_lock(&cache->lock);
4558
4559                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4560                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4561                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4562
4563                 cache->dirty = 1;
4564                 old_val = btrfs_block_group_used(&cache->item);
4565                 num_bytes = min(total, cache->key.offset - byte_in_group);
4566                 if (alloc) {
4567                         old_val += num_bytes;
4568                         btrfs_set_block_group_used(&cache->item, old_val);
4569                         cache->reserved -= num_bytes;
4570                         cache->space_info->bytes_reserved -= num_bytes;
4571                         cache->space_info->bytes_used += num_bytes;
4572                         cache->space_info->disk_used += num_bytes * factor;
4573                         spin_unlock(&cache->lock);
4574                         spin_unlock(&cache->space_info->lock);
4575                 } else {
4576                         old_val -= num_bytes;
4577                         btrfs_set_block_group_used(&cache->item, old_val);
4578                         cache->pinned += num_bytes;
4579                         cache->space_info->bytes_pinned += num_bytes;
4580                         cache->space_info->bytes_used -= num_bytes;
4581                         cache->space_info->disk_used -= num_bytes * factor;
4582                         spin_unlock(&cache->lock);
4583                         spin_unlock(&cache->space_info->lock);
4584
4585                         set_extent_dirty(info->pinned_extents,
4586                                          bytenr, bytenr + num_bytes - 1,
4587                                          GFP_NOFS | __GFP_NOFAIL);
4588                 }
4589                 btrfs_put_block_group(cache);
4590                 total -= num_bytes;
4591                 bytenr += num_bytes;
4592         }
4593         return 0;
4594 }
4595
4596 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4597 {
4598         struct btrfs_block_group_cache *cache;
4599         u64 bytenr;
4600
4601         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4602         if (!cache)
4603                 return 0;
4604
4605         bytenr = cache->key.objectid;
4606         btrfs_put_block_group(cache);
4607
4608         return bytenr;
4609 }
4610
4611 static int pin_down_extent(struct btrfs_root *root,
4612                            struct btrfs_block_group_cache *cache,
4613                            u64 bytenr, u64 num_bytes, int reserved)
4614 {
4615         spin_lock(&cache->space_info->lock);
4616         spin_lock(&cache->lock);
4617         cache->pinned += num_bytes;
4618         cache->space_info->bytes_pinned += num_bytes;
4619         if (reserved) {
4620                 cache->reserved -= num_bytes;
4621                 cache->space_info->bytes_reserved -= num_bytes;
4622         }
4623         spin_unlock(&cache->lock);
4624         spin_unlock(&cache->space_info->lock);
4625
4626         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4627                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4628         return 0;
4629 }
4630
4631 /*
4632  * this function must be called within transaction
4633  */
4634 int btrfs_pin_extent(struct btrfs_root *root,
4635                      u64 bytenr, u64 num_bytes, int reserved)
4636 {
4637         struct btrfs_block_group_cache *cache;
4638
4639         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4640         BUG_ON(!cache);
4641
4642         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4643
4644         btrfs_put_block_group(cache);
4645         return 0;
4646 }
4647
4648 /*
4649  * this function must be called within transaction
4650  */
4651 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4652                                     struct btrfs_root *root,
4653                                     u64 bytenr, u64 num_bytes)
4654 {
4655         struct btrfs_block_group_cache *cache;
4656
4657         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4658         BUG_ON(!cache);
4659
4660         /*
4661          * pull in the free space cache (if any) so that our pin
4662          * removes the free space from the cache.  We have load_only set
4663          * to one because the slow code to read in the free extents does check
4664          * the pinned extents.
4665          */
4666         cache_block_group(cache, trans, root, 1);
4667
4668         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4669
4670         /* remove us from the free space cache (if we're there at all) */
4671         btrfs_remove_free_space(cache, bytenr, num_bytes);
4672         btrfs_put_block_group(cache);
4673         return 0;
4674 }
4675
4676 /**
4677  * btrfs_update_reserved_bytes - update the block_group and space info counters
4678  * @cache:      The cache we are manipulating
4679  * @num_bytes:  The number of bytes in question
4680  * @reserve:    One of the reservation enums
4681  *
4682  * This is called by the allocator when it reserves space, or by somebody who is
4683  * freeing space that was never actually used on disk.  For example if you
4684  * reserve some space for a new leaf in transaction A and before transaction A
4685  * commits you free that leaf, you call this with reserve set to 0 in order to
4686  * clear the reservation.
4687  *
4688  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4689  * ENOSPC accounting.  For data we handle the reservation through clearing the
4690  * delalloc bits in the io_tree.  We have to do this since we could end up
4691  * allocating less disk space for the amount of data we have reserved in the
4692  * case of compression.
4693  *
4694  * If this is a reservation and the block group has become read only we cannot
4695  * make the reservation and return -EAGAIN, otherwise this function always
4696  * succeeds.
4697  */
4698 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4699                                        u64 num_bytes, int reserve)
4700 {
4701         struct btrfs_space_info *space_info = cache->space_info;
4702         int ret = 0;
4703         spin_lock(&space_info->lock);
4704         spin_lock(&cache->lock);
4705         if (reserve != RESERVE_FREE) {
4706                 if (cache->ro) {
4707                         ret = -EAGAIN;
4708                 } else {
4709                         cache->reserved += num_bytes;
4710                         space_info->bytes_reserved += num_bytes;
4711                         if (reserve == RESERVE_ALLOC) {
4712                                 trace_btrfs_space_reservation(cache->fs_info,
4713                                                               "space_info",
4714                                                               (u64)space_info,
4715                                                               num_bytes, 0);
4716                                 space_info->bytes_may_use -= num_bytes;
4717                         }
4718                 }
4719         } else {
4720                 if (cache->ro)
4721                         space_info->bytes_readonly += num_bytes;
4722                 cache->reserved -= num_bytes;
4723                 space_info->bytes_reserved -= num_bytes;
4724                 space_info->reservation_progress++;
4725         }
4726         spin_unlock(&cache->lock);
4727         spin_unlock(&space_info->lock);
4728         return ret;
4729 }
4730
4731 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4732                                 struct btrfs_root *root)
4733 {
4734         struct btrfs_fs_info *fs_info = root->fs_info;
4735         struct btrfs_caching_control *next;
4736         struct btrfs_caching_control *caching_ctl;
4737         struct btrfs_block_group_cache *cache;
4738
4739         down_write(&fs_info->extent_commit_sem);
4740
4741         list_for_each_entry_safe(caching_ctl, next,
4742                                  &fs_info->caching_block_groups, list) {
4743                 cache = caching_ctl->block_group;
4744                 if (block_group_cache_done(cache)) {
4745                         cache->last_byte_to_unpin = (u64)-1;
4746                         list_del_init(&caching_ctl->list);
4747                         put_caching_control(caching_ctl);
4748                 } else {
4749                         cache->last_byte_to_unpin = caching_ctl->progress;
4750                 }
4751         }
4752
4753         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4754                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4755         else
4756                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4757
4758         up_write(&fs_info->extent_commit_sem);
4759
4760         update_global_block_rsv(fs_info);
4761         return 0;
4762 }
4763
4764 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4765 {
4766         struct btrfs_fs_info *fs_info = root->fs_info;
4767         struct btrfs_block_group_cache *cache = NULL;
4768         u64 len;
4769
4770         while (start <= end) {
4771                 if (!cache ||
4772                     start >= cache->key.objectid + cache->key.offset) {
4773                         if (cache)
4774                                 btrfs_put_block_group(cache);
4775                         cache = btrfs_lookup_block_group(fs_info, start);
4776                         BUG_ON(!cache);
4777                 }
4778
4779                 len = cache->key.objectid + cache->key.offset - start;
4780                 len = min(len, end + 1 - start);
4781
4782                 if (start < cache->last_byte_to_unpin) {
4783                         len = min(len, cache->last_byte_to_unpin - start);
4784                         btrfs_add_free_space(cache, start, len);
4785                 }
4786
4787                 start += len;
4788
4789                 spin_lock(&cache->space_info->lock);
4790                 spin_lock(&cache->lock);
4791                 cache->pinned -= len;
4792                 cache->space_info->bytes_pinned -= len;
4793                 if (cache->ro)
4794                         cache->space_info->bytes_readonly += len;
4795                 spin_unlock(&cache->lock);
4796                 spin_unlock(&cache->space_info->lock);
4797         }
4798
4799         if (cache)
4800                 btrfs_put_block_group(cache);
4801         return 0;
4802 }
4803
4804 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4805                                struct btrfs_root *root)
4806 {
4807         struct btrfs_fs_info *fs_info = root->fs_info;
4808         struct extent_io_tree *unpin;
4809         u64 start;
4810         u64 end;
4811         int ret;
4812
4813         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4814                 unpin = &fs_info->freed_extents[1];
4815         else
4816                 unpin = &fs_info->freed_extents[0];
4817
4818         while (1) {
4819                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4820                                             EXTENT_DIRTY);
4821                 if (ret)
4822                         break;
4823
4824                 if (btrfs_test_opt(root, DISCARD))
4825                         ret = btrfs_discard_extent(root, start,
4826                                                    end + 1 - start, NULL);
4827
4828                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4829                 unpin_extent_range(root, start, end);
4830                 cond_resched();
4831         }
4832
4833         return 0;
4834 }
4835
4836 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4837                                 struct btrfs_root *root,
4838                                 u64 bytenr, u64 num_bytes, u64 parent,
4839                                 u64 root_objectid, u64 owner_objectid,
4840                                 u64 owner_offset, int refs_to_drop,
4841                                 struct btrfs_delayed_extent_op *extent_op)
4842 {
4843         struct btrfs_key key;
4844         struct btrfs_path *path;
4845         struct btrfs_fs_info *info = root->fs_info;
4846         struct btrfs_root *extent_root = info->extent_root;
4847         struct extent_buffer *leaf;
4848         struct btrfs_extent_item *ei;
4849         struct btrfs_extent_inline_ref *iref;
4850         int ret;
4851         int is_data;
4852         int extent_slot = 0;
4853         int found_extent = 0;
4854         int num_to_del = 1;
4855         u32 item_size;
4856         u64 refs;
4857
4858         path = btrfs_alloc_path();
4859         if (!path)
4860                 return -ENOMEM;
4861
4862         path->reada = 1;
4863         path->leave_spinning = 1;
4864
4865         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4866         BUG_ON(!is_data && refs_to_drop != 1);
4867
4868         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4869                                     bytenr, num_bytes, parent,
4870                                     root_objectid, owner_objectid,
4871                                     owner_offset);
4872         if (ret == 0) {
4873                 extent_slot = path->slots[0];
4874                 while (extent_slot >= 0) {
4875                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4876                                               extent_slot);
4877                         if (key.objectid != bytenr)
4878                                 break;
4879                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4880                             key.offset == num_bytes) {
4881                                 found_extent = 1;
4882                                 break;
4883                         }
4884                         if (path->slots[0] - extent_slot > 5)
4885                                 break;
4886                         extent_slot--;
4887                 }
4888 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4889                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4890                 if (found_extent && item_size < sizeof(*ei))
4891                         found_extent = 0;
4892 #endif
4893                 if (!found_extent) {
4894                         BUG_ON(iref);
4895                         ret = remove_extent_backref(trans, extent_root, path,
4896                                                     NULL, refs_to_drop,
4897                                                     is_data);
4898                         BUG_ON(ret);
4899                         btrfs_release_path(path);
4900                         path->leave_spinning = 1;
4901
4902                         key.objectid = bytenr;
4903                         key.type = BTRFS_EXTENT_ITEM_KEY;
4904                         key.offset = num_bytes;
4905
4906                         ret = btrfs_search_slot(trans, extent_root,
4907                                                 &key, path, -1, 1);
4908                         if (ret) {
4909                                 printk(KERN_ERR "umm, got %d back from search"
4910                                        ", was looking for %llu\n", ret,
4911                                        (unsigned long long)bytenr);
4912                                 if (ret > 0)
4913                                         btrfs_print_leaf(extent_root,
4914                                                          path->nodes[0]);
4915                         }
4916                         BUG_ON(ret);
4917                         extent_slot = path->slots[0];
4918                 }
4919         } else {
4920                 btrfs_print_leaf(extent_root, path->nodes[0]);
4921                 WARN_ON(1);
4922                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4923                        "parent %llu root %llu  owner %llu offset %llu\n",
4924                        (unsigned long long)bytenr,
4925                        (unsigned long long)parent,
4926                        (unsigned long long)root_objectid,
4927                        (unsigned long long)owner_objectid,
4928                        (unsigned long long)owner_offset);
4929         }
4930
4931         leaf = path->nodes[0];
4932         item_size = btrfs_item_size_nr(leaf, extent_slot);
4933 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4934         if (item_size < sizeof(*ei)) {
4935                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4936                 ret = convert_extent_item_v0(trans, extent_root, path,
4937                                              owner_objectid, 0);
4938                 BUG_ON(ret < 0);
4939
4940                 btrfs_release_path(path);
4941                 path->leave_spinning = 1;
4942
4943                 key.objectid = bytenr;
4944                 key.type = BTRFS_EXTENT_ITEM_KEY;
4945                 key.offset = num_bytes;
4946
4947                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4948                                         -1, 1);
4949                 if (ret) {
4950                         printk(KERN_ERR "umm, got %d back from search"
4951                                ", was looking for %llu\n", ret,
4952                                (unsigned long long)bytenr);
4953                         btrfs_print_leaf(extent_root, path->nodes[0]);
4954                 }
4955                 BUG_ON(ret);
4956                 extent_slot = path->slots[0];
4957                 leaf = path->nodes[0];
4958                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4959         }
4960 #endif
4961         BUG_ON(item_size < sizeof(*ei));
4962         ei = btrfs_item_ptr(leaf, extent_slot,
4963                             struct btrfs_extent_item);
4964         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4965                 struct btrfs_tree_block_info *bi;
4966                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4967                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4968                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4969         }
4970
4971         refs = btrfs_extent_refs(leaf, ei);
4972         BUG_ON(refs < refs_to_drop);
4973         refs -= refs_to_drop;
4974
4975         if (refs > 0) {
4976                 if (extent_op)
4977                         __run_delayed_extent_op(extent_op, leaf, ei);
4978                 /*
4979                  * In the case of inline back ref, reference count will
4980                  * be updated by remove_extent_backref
4981                  */
4982                 if (iref) {
4983                         BUG_ON(!found_extent);
4984                 } else {
4985                         btrfs_set_extent_refs(leaf, ei, refs);
4986                         btrfs_mark_buffer_dirty(leaf);
4987                 }
4988                 if (found_extent) {
4989                         ret = remove_extent_backref(trans, extent_root, path,
4990                                                     iref, refs_to_drop,
4991                                                     is_data);
4992                         BUG_ON(ret);
4993                 }
4994         } else {
4995                 if (found_extent) {
4996                         BUG_ON(is_data && refs_to_drop !=
4997                                extent_data_ref_count(root, path, iref));
4998                         if (iref) {
4999                                 BUG_ON(path->slots[0] != extent_slot);
5000                         } else {
5001                                 BUG_ON(path->slots[0] != extent_slot + 1);
5002                                 path->slots[0] = extent_slot;
5003                                 num_to_del = 2;
5004                         }
5005                 }
5006
5007                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5008                                       num_to_del);
5009                 BUG_ON(ret);
5010                 btrfs_release_path(path);
5011
5012                 if (is_data) {
5013                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5014                         BUG_ON(ret);
5015                 } else {
5016                         invalidate_mapping_pages(info->btree_inode->i_mapping,
5017                              bytenr >> PAGE_CACHE_SHIFT,
5018                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
5019                 }
5020
5021                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5022                 BUG_ON(ret);
5023         }
5024         btrfs_free_path(path);
5025         return ret;
5026 }
5027
5028 /*
5029  * when we free an block, it is possible (and likely) that we free the last
5030  * delayed ref for that extent as well.  This searches the delayed ref tree for
5031  * a given extent, and if there are no other delayed refs to be processed, it
5032  * removes it from the tree.
5033  */
5034 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5035                                       struct btrfs_root *root, u64 bytenr)
5036 {
5037         struct btrfs_delayed_ref_head *head;
5038         struct btrfs_delayed_ref_root *delayed_refs;
5039         struct btrfs_delayed_ref_node *ref;
5040         struct rb_node *node;
5041         int ret = 0;
5042
5043         delayed_refs = &trans->transaction->delayed_refs;
5044         spin_lock(&delayed_refs->lock);
5045         head = btrfs_find_delayed_ref_head(trans, bytenr);
5046         if (!head)
5047                 goto out;
5048
5049         node = rb_prev(&head->node.rb_node);
5050         if (!node)
5051                 goto out;
5052
5053         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5054
5055         /* there are still entries for this ref, we can't drop it */
5056         if (ref->bytenr == bytenr)
5057                 goto out;
5058
5059         if (head->extent_op) {
5060                 if (!head->must_insert_reserved)
5061                         goto out;
5062                 kfree(head->extent_op);
5063                 head->extent_op = NULL;
5064         }
5065
5066         /*
5067          * waiting for the lock here would deadlock.  If someone else has it
5068          * locked they are already in the process of dropping it anyway
5069          */
5070         if (!mutex_trylock(&head->mutex))
5071                 goto out;
5072
5073         /*
5074          * at this point we have a head with no other entries.  Go
5075          * ahead and process it.
5076          */
5077         head->node.in_tree = 0;
5078         rb_erase(&head->node.rb_node, &delayed_refs->root);
5079
5080         delayed_refs->num_entries--;
5081         if (waitqueue_active(&delayed_refs->seq_wait))
5082                 wake_up(&delayed_refs->seq_wait);
5083
5084         /*
5085          * we don't take a ref on the node because we're removing it from the
5086          * tree, so we just steal the ref the tree was holding.
5087          */
5088         delayed_refs->num_heads--;
5089         if (list_empty(&head->cluster))
5090                 delayed_refs->num_heads_ready--;
5091
5092         list_del_init(&head->cluster);
5093         spin_unlock(&delayed_refs->lock);
5094
5095         BUG_ON(head->extent_op);
5096         if (head->must_insert_reserved)
5097                 ret = 1;
5098
5099         mutex_unlock(&head->mutex);
5100         btrfs_put_delayed_ref(&head->node);
5101         return ret;
5102 out:
5103         spin_unlock(&delayed_refs->lock);
5104         return 0;
5105 }
5106
5107 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5108                            struct btrfs_root *root,
5109                            struct extent_buffer *buf,
5110                            u64 parent, int last_ref, int for_cow)
5111 {
5112         struct btrfs_block_group_cache *cache = NULL;
5113         int ret;
5114
5115         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5116                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5117                                         buf->start, buf->len,
5118                                         parent, root->root_key.objectid,
5119                                         btrfs_header_level(buf),
5120                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5121                 BUG_ON(ret);
5122         }
5123
5124         if (!last_ref)
5125                 return;
5126
5127         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5128
5129         if (btrfs_header_generation(buf) == trans->transid) {
5130                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5131                         ret = check_ref_cleanup(trans, root, buf->start);
5132                         if (!ret)
5133                                 goto out;
5134                 }
5135
5136                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5137                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5138                         goto out;
5139                 }
5140
5141                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5142
5143                 btrfs_add_free_space(cache, buf->start, buf->len);
5144                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5145         }
5146 out:
5147         /*
5148          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5149          * anymore.
5150          */
5151         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5152         btrfs_put_block_group(cache);
5153 }
5154
5155 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5156                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5157                       u64 owner, u64 offset, int for_cow)
5158 {
5159         int ret;
5160         struct btrfs_fs_info *fs_info = root->fs_info;
5161
5162         /*
5163          * tree log blocks never actually go into the extent allocation
5164          * tree, just update pinning info and exit early.
5165          */
5166         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5167                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5168                 /* unlocks the pinned mutex */
5169                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5170                 ret = 0;
5171         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5172                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5173                                         num_bytes,
5174                                         parent, root_objectid, (int)owner,
5175                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5176                 BUG_ON(ret);
5177         } else {
5178                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5179                                                 num_bytes,
5180                                                 parent, root_objectid, owner,
5181                                                 offset, BTRFS_DROP_DELAYED_REF,
5182                                                 NULL, for_cow);
5183                 BUG_ON(ret);
5184         }
5185         return ret;
5186 }
5187
5188 static u64 stripe_align(struct btrfs_root *root, u64 val)
5189 {
5190         u64 mask = ((u64)root->stripesize - 1);
5191         u64 ret = (val + mask) & ~mask;
5192         return ret;
5193 }
5194
5195 /*
5196  * when we wait for progress in the block group caching, its because
5197  * our allocation attempt failed at least once.  So, we must sleep
5198  * and let some progress happen before we try again.
5199  *
5200  * This function will sleep at least once waiting for new free space to
5201  * show up, and then it will check the block group free space numbers
5202  * for our min num_bytes.  Another option is to have it go ahead
5203  * and look in the rbtree for a free extent of a given size, but this
5204  * is a good start.
5205  */
5206 static noinline int
5207 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5208                                 u64 num_bytes)
5209 {
5210         struct btrfs_caching_control *caching_ctl;
5211         DEFINE_WAIT(wait);
5212
5213         caching_ctl = get_caching_control(cache);
5214         if (!caching_ctl)
5215                 return 0;
5216
5217         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5218                    (cache->free_space_ctl->free_space >= num_bytes));
5219
5220         put_caching_control(caching_ctl);
5221         return 0;
5222 }
5223
5224 static noinline int
5225 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5226 {
5227         struct btrfs_caching_control *caching_ctl;
5228         DEFINE_WAIT(wait);
5229
5230         caching_ctl = get_caching_control(cache);
5231         if (!caching_ctl)
5232                 return 0;
5233
5234         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5235
5236         put_caching_control(caching_ctl);
5237         return 0;
5238 }
5239
5240 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5241 {
5242         int index;
5243         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5244                 index = 0;
5245         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5246                 index = 1;
5247         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5248                 index = 2;
5249         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5250                 index = 3;
5251         else
5252                 index = 4;
5253         return index;
5254 }
5255
5256 enum btrfs_loop_type {
5257         LOOP_FIND_IDEAL = 0,
5258         LOOP_CACHING_NOWAIT = 1,
5259         LOOP_CACHING_WAIT = 2,
5260         LOOP_ALLOC_CHUNK = 3,
5261         LOOP_NO_EMPTY_SIZE = 4,
5262 };
5263
5264 /*
5265  * walks the btree of allocated extents and find a hole of a given size.
5266  * The key ins is changed to record the hole:
5267  * ins->objectid == block start
5268  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5269  * ins->offset == number of blocks
5270  * Any available blocks before search_start are skipped.
5271  */
5272 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5273                                      struct btrfs_root *orig_root,
5274                                      u64 num_bytes, u64 empty_size,
5275                                      u64 search_start, u64 search_end,
5276                                      u64 hint_byte, struct btrfs_key *ins,
5277                                      u64 data)
5278 {
5279         int ret = 0;
5280         struct btrfs_root *root = orig_root->fs_info->extent_root;
5281         struct btrfs_free_cluster *last_ptr = NULL;
5282         struct btrfs_block_group_cache *block_group = NULL;
5283         struct btrfs_block_group_cache *used_block_group;
5284         int empty_cluster = 2 * 1024 * 1024;
5285         int allowed_chunk_alloc = 0;
5286         int done_chunk_alloc = 0;
5287         struct btrfs_space_info *space_info;
5288         int loop = 0;
5289         int index = 0;
5290         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5291                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5292         bool found_uncached_bg = false;
5293         bool failed_cluster_refill = false;
5294         bool failed_alloc = false;
5295         bool use_cluster = true;
5296         bool have_caching_bg = false;
5297         u64 ideal_cache_percent = 0;
5298         u64 ideal_cache_offset = 0;
5299
5300         WARN_ON(num_bytes < root->sectorsize);
5301         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5302         ins->objectid = 0;
5303         ins->offset = 0;
5304
5305         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5306
5307         space_info = __find_space_info(root->fs_info, data);
5308         if (!space_info) {
5309                 printk(KERN_ERR "No space info for %llu\n", data);
5310                 return -ENOSPC;
5311         }
5312
5313         /*
5314          * If the space info is for both data and metadata it means we have a
5315          * small filesystem and we can't use the clustering stuff.
5316          */
5317         if (btrfs_mixed_space_info(space_info))
5318                 use_cluster = false;
5319
5320         if (orig_root->ref_cows || empty_size)
5321                 allowed_chunk_alloc = 1;
5322
5323         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5324                 last_ptr = &root->fs_info->meta_alloc_cluster;
5325                 if (!btrfs_test_opt(root, SSD))
5326                         empty_cluster = 64 * 1024;
5327         }
5328
5329         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5330             btrfs_test_opt(root, SSD)) {
5331                 last_ptr = &root->fs_info->data_alloc_cluster;
5332         }
5333
5334         if (last_ptr) {
5335                 spin_lock(&last_ptr->lock);
5336                 if (last_ptr->block_group)
5337                         hint_byte = last_ptr->window_start;
5338                 spin_unlock(&last_ptr->lock);
5339         }
5340
5341         search_start = max(search_start, first_logical_byte(root, 0));
5342         search_start = max(search_start, hint_byte);
5343
5344         if (!last_ptr)
5345                 empty_cluster = 0;
5346
5347         if (search_start == hint_byte) {
5348 ideal_cache:
5349                 block_group = btrfs_lookup_block_group(root->fs_info,
5350                                                        search_start);
5351                 used_block_group = block_group;
5352                 /*
5353                  * we don't want to use the block group if it doesn't match our
5354                  * allocation bits, or if its not cached.
5355                  *
5356                  * However if we are re-searching with an ideal block group
5357                  * picked out then we don't care that the block group is cached.
5358                  */
5359                 if (block_group && block_group_bits(block_group, data) &&
5360                     (block_group->cached != BTRFS_CACHE_NO ||
5361                      search_start == ideal_cache_offset)) {
5362                         down_read(&space_info->groups_sem);
5363                         if (list_empty(&block_group->list) ||
5364                             block_group->ro) {
5365                                 /*
5366                                  * someone is removing this block group,
5367                                  * we can't jump into the have_block_group
5368                                  * target because our list pointers are not
5369                                  * valid
5370                                  */
5371                                 btrfs_put_block_group(block_group);
5372                                 up_read(&space_info->groups_sem);
5373                         } else {
5374                                 index = get_block_group_index(block_group);
5375                                 goto have_block_group;
5376                         }
5377                 } else if (block_group) {
5378                         btrfs_put_block_group(block_group);
5379                 }
5380         }
5381 search:
5382         have_caching_bg = false;
5383         down_read(&space_info->groups_sem);
5384         list_for_each_entry(block_group, &space_info->block_groups[index],
5385                             list) {
5386                 u64 offset;
5387                 int cached;
5388
5389                 used_block_group = block_group;
5390                 btrfs_get_block_group(block_group);
5391                 search_start = block_group->key.objectid;
5392
5393                 /*
5394                  * this can happen if we end up cycling through all the
5395                  * raid types, but we want to make sure we only allocate
5396                  * for the proper type.
5397                  */
5398                 if (!block_group_bits(block_group, data)) {
5399                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5400                                 BTRFS_BLOCK_GROUP_RAID1 |
5401                                 BTRFS_BLOCK_GROUP_RAID10;
5402
5403                         /*
5404                          * if they asked for extra copies and this block group
5405                          * doesn't provide them, bail.  This does allow us to
5406                          * fill raid0 from raid1.
5407                          */
5408                         if ((data & extra) && !(block_group->flags & extra))
5409                                 goto loop;
5410                 }
5411
5412 have_block_group:
5413                 cached = block_group_cache_done(block_group);
5414                 if (unlikely(!cached)) {
5415                         u64 free_percent;
5416
5417                         found_uncached_bg = true;
5418                         ret = cache_block_group(block_group, trans,
5419                                                 orig_root, 1);
5420                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5421                                 goto alloc;
5422
5423                         free_percent = btrfs_block_group_used(&block_group->item);
5424                         free_percent *= 100;
5425                         free_percent = div64_u64(free_percent,
5426                                                  block_group->key.offset);
5427                         free_percent = 100 - free_percent;
5428                         if (free_percent > ideal_cache_percent &&
5429                             likely(!block_group->ro)) {
5430                                 ideal_cache_offset = block_group->key.objectid;
5431                                 ideal_cache_percent = free_percent;
5432                         }
5433
5434                         /*
5435                          * The caching workers are limited to 2 threads, so we
5436                          * can queue as much work as we care to.
5437                          */
5438                         if (loop > LOOP_FIND_IDEAL) {
5439                                 ret = cache_block_group(block_group, trans,
5440                                                         orig_root, 0);
5441                                 BUG_ON(ret);
5442                         }
5443
5444                         /*
5445                          * If loop is set for cached only, try the next block
5446                          * group.
5447                          */
5448                         if (loop == LOOP_FIND_IDEAL)
5449                                 goto loop;
5450                 }
5451
5452 alloc:
5453                 if (unlikely(block_group->ro))
5454                         goto loop;
5455
5456                 /*
5457                  * Ok we want to try and use the cluster allocator, so
5458                  * lets look there
5459                  */
5460                 if (last_ptr) {
5461                         /*
5462                          * the refill lock keeps out other
5463                          * people trying to start a new cluster
5464                          */
5465                         spin_lock(&last_ptr->refill_lock);
5466                         used_block_group = last_ptr->block_group;
5467                         if (used_block_group != block_group &&
5468                             (!used_block_group ||
5469                              used_block_group->ro ||
5470                              !block_group_bits(used_block_group, data))) {
5471                                 used_block_group = block_group;
5472                                 goto refill_cluster;
5473                         }
5474
5475                         if (used_block_group != block_group)
5476                                 btrfs_get_block_group(used_block_group);
5477
5478                         offset = btrfs_alloc_from_cluster(used_block_group,
5479                           last_ptr, num_bytes, used_block_group->key.objectid);
5480                         if (offset) {
5481                                 /* we have a block, we're done */
5482                                 spin_unlock(&last_ptr->refill_lock);
5483                                 trace_btrfs_reserve_extent_cluster(root,
5484                                         block_group, search_start, num_bytes);
5485                                 goto checks;
5486                         }
5487
5488                         WARN_ON(last_ptr->block_group != used_block_group);
5489                         if (used_block_group != block_group) {
5490                                 btrfs_put_block_group(used_block_group);
5491                                 used_block_group = block_group;
5492                         }
5493 refill_cluster:
5494                         BUG_ON(used_block_group != block_group);
5495                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5496                          * set up a new clusters, so lets just skip it
5497                          * and let the allocator find whatever block
5498                          * it can find.  If we reach this point, we
5499                          * will have tried the cluster allocator
5500                          * plenty of times and not have found
5501                          * anything, so we are likely way too
5502                          * fragmented for the clustering stuff to find
5503                          * anything.
5504                          *
5505                          * However, if the cluster is taken from the
5506                          * current block group, release the cluster
5507                          * first, so that we stand a better chance of
5508                          * succeeding in the unclustered
5509                          * allocation.  */
5510                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5511                             last_ptr->block_group != block_group) {
5512                                 spin_unlock(&last_ptr->refill_lock);
5513                                 goto unclustered_alloc;
5514                         }
5515
5516                         /*
5517                          * this cluster didn't work out, free it and
5518                          * start over
5519                          */
5520                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5521
5522                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5523                                 spin_unlock(&last_ptr->refill_lock);
5524                                 goto unclustered_alloc;
5525                         }
5526
5527                         /* allocate a cluster in this block group */
5528                         ret = btrfs_find_space_cluster(trans, root,
5529                                                block_group, last_ptr,
5530                                                search_start, num_bytes,
5531                                                empty_cluster + empty_size);
5532                         if (ret == 0) {
5533                                 /*
5534                                  * now pull our allocation out of this
5535                                  * cluster
5536                                  */
5537                                 offset = btrfs_alloc_from_cluster(block_group,
5538                                                   last_ptr, num_bytes,
5539                                                   search_start);
5540                                 if (offset) {
5541                                         /* we found one, proceed */
5542                                         spin_unlock(&last_ptr->refill_lock);
5543                                         trace_btrfs_reserve_extent_cluster(root,
5544                                                 block_group, search_start,
5545                                                 num_bytes);
5546                                         goto checks;
5547                                 }
5548                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5549                                    && !failed_cluster_refill) {
5550                                 spin_unlock(&last_ptr->refill_lock);
5551
5552                                 failed_cluster_refill = true;
5553                                 wait_block_group_cache_progress(block_group,
5554                                        num_bytes + empty_cluster + empty_size);
5555                                 goto have_block_group;
5556                         }
5557
5558                         /*
5559                          * at this point we either didn't find a cluster
5560                          * or we weren't able to allocate a block from our
5561                          * cluster.  Free the cluster we've been trying
5562                          * to use, and go to the next block group
5563                          */
5564                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5565                         spin_unlock(&last_ptr->refill_lock);
5566                         goto loop;
5567                 }
5568
5569 unclustered_alloc:
5570                 spin_lock(&block_group->free_space_ctl->tree_lock);
5571                 if (cached &&
5572                     block_group->free_space_ctl->free_space <
5573                     num_bytes + empty_cluster + empty_size) {
5574                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5575                         goto loop;
5576                 }
5577                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5578
5579                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5580                                                     num_bytes, empty_size);
5581                 /*
5582                  * If we didn't find a chunk, and we haven't failed on this
5583                  * block group before, and this block group is in the middle of
5584                  * caching and we are ok with waiting, then go ahead and wait
5585                  * for progress to be made, and set failed_alloc to true.
5586                  *
5587                  * If failed_alloc is true then we've already waited on this
5588                  * block group once and should move on to the next block group.
5589                  */
5590                 if (!offset && !failed_alloc && !cached &&
5591                     loop > LOOP_CACHING_NOWAIT) {
5592                         wait_block_group_cache_progress(block_group,
5593                                                 num_bytes + empty_size);
5594                         failed_alloc = true