Merge branch 'for-chris' of git://git.jan-o-sch.net/btrfs-unstable into for-linus
[linux-3.10.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); /* -ENOMEM */
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); /* -ENOMEM */
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); /* -ENOMEM */
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); /* -ENOMEM or logic error */
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); /* -ENOMEM or logic error */
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         if (!caching_ctl)
478                 return -ENOMEM;
479
480         INIT_LIST_HEAD(&caching_ctl->list);
481         mutex_init(&caching_ctl->mutex);
482         init_waitqueue_head(&caching_ctl->wait);
483         caching_ctl->block_group = cache;
484         caching_ctl->progress = cache->key.objectid;
485         atomic_set(&caching_ctl->count, 1);
486         caching_ctl->work.func = caching_thread;
487
488         spin_lock(&cache->lock);
489         /*
490          * This should be a rare occasion, but this could happen I think in the
491          * case where one thread starts to load the space cache info, and then
492          * some other thread starts a transaction commit which tries to do an
493          * allocation while the other thread is still loading the space cache
494          * info.  The previous loop should have kept us from choosing this block
495          * group, but if we've moved to the state where we will wait on caching
496          * block groups we need to first check if we're doing a fast load here,
497          * so we can wait for it to finish, otherwise we could end up allocating
498          * from a block group who's cache gets evicted for one reason or
499          * another.
500          */
501         while (cache->cached == BTRFS_CACHE_FAST) {
502                 struct btrfs_caching_control *ctl;
503
504                 ctl = cache->caching_ctl;
505                 atomic_inc(&ctl->count);
506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
507                 spin_unlock(&cache->lock);
508
509                 schedule();
510
511                 finish_wait(&ctl->wait, &wait);
512                 put_caching_control(ctl);
513                 spin_lock(&cache->lock);
514         }
515
516         if (cache->cached != BTRFS_CACHE_NO) {
517                 spin_unlock(&cache->lock);
518                 kfree(caching_ctl);
519                 return 0;
520         }
521         WARN_ON(cache->caching_ctl);
522         cache->caching_ctl = caching_ctl;
523         cache->cached = BTRFS_CACHE_FAST;
524         spin_unlock(&cache->lock);
525
526         /*
527          * We can't do the read from on-disk cache during a commit since we need
528          * to have the normal tree locking.  Also if we are currently trying to
529          * allocate blocks for the tree root we can't do the fast caching since
530          * we likely hold important locks.
531          */
532         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
533                 ret = load_free_space_cache(fs_info, cache);
534
535                 spin_lock(&cache->lock);
536                 if (ret == 1) {
537                         cache->caching_ctl = NULL;
538                         cache->cached = BTRFS_CACHE_FINISHED;
539                         cache->last_byte_to_unpin = (u64)-1;
540                 } else {
541                         if (load_cache_only) {
542                                 cache->caching_ctl = NULL;
543                                 cache->cached = BTRFS_CACHE_NO;
544                         } else {
545                                 cache->cached = BTRFS_CACHE_STARTED;
546                         }
547                 }
548                 spin_unlock(&cache->lock);
549                 wake_up(&caching_ctl->wait);
550                 if (ret == 1) {
551                         put_caching_control(caching_ctl);
552                         free_excluded_extents(fs_info->extent_root, cache);
553                         return 0;
554                 }
555         } else {
556                 /*
557                  * We are not going to do the fast caching, set cached to the
558                  * appropriate value and wakeup any waiters.
559                  */
560                 spin_lock(&cache->lock);
561                 if (load_cache_only) {
562                         cache->caching_ctl = NULL;
563                         cache->cached = BTRFS_CACHE_NO;
564                 } else {
565                         cache->cached = BTRFS_CACHE_STARTED;
566                 }
567                 spin_unlock(&cache->lock);
568                 wake_up(&caching_ctl->wait);
569         }
570
571         if (load_cache_only) {
572                 put_caching_control(caching_ctl);
573                 return 0;
574         }
575
576         down_write(&fs_info->extent_commit_sem);
577         atomic_inc(&caching_ctl->count);
578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
579         up_write(&fs_info->extent_commit_sem);
580
581         btrfs_get_block_group(cache);
582
583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
584
585         return ret;
586 }
587
588 /*
589  * return the block group that starts at or after bytenr
590  */
591 static struct btrfs_block_group_cache *
592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
593 {
594         struct btrfs_block_group_cache *cache;
595
596         cache = block_group_cache_tree_search(info, bytenr, 0);
597
598         return cache;
599 }
600
601 /*
602  * return the block group that contains the given bytenr
603  */
604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
605                                                  struct btrfs_fs_info *info,
606                                                  u64 bytenr)
607 {
608         struct btrfs_block_group_cache *cache;
609
610         cache = block_group_cache_tree_search(info, bytenr, 1);
611
612         return cache;
613 }
614
615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
616                                                   u64 flags)
617 {
618         struct list_head *head = &info->space_info;
619         struct btrfs_space_info *found;
620
621         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
622
623         rcu_read_lock();
624         list_for_each_entry_rcu(found, head, list) {
625                 if (found->flags & flags) {
626                         rcu_read_unlock();
627                         return found;
628                 }
629         }
630         rcu_read_unlock();
631         return NULL;
632 }
633
634 /*
635  * after adding space to the filesystem, we need to clear the full flags
636  * on all the space infos.
637  */
638 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
639 {
640         struct list_head *head = &info->space_info;
641         struct btrfs_space_info *found;
642
643         rcu_read_lock();
644         list_for_each_entry_rcu(found, head, list)
645                 found->full = 0;
646         rcu_read_unlock();
647 }
648
649 static u64 div_factor(u64 num, int factor)
650 {
651         if (factor == 10)
652                 return num;
653         num *= factor;
654         do_div(num, 10);
655         return num;
656 }
657
658 static u64 div_factor_fine(u64 num, int factor)
659 {
660         if (factor == 100)
661                 return num;
662         num *= factor;
663         do_div(num, 100);
664         return num;
665 }
666
667 u64 btrfs_find_block_group(struct btrfs_root *root,
668                            u64 search_start, u64 search_hint, int owner)
669 {
670         struct btrfs_block_group_cache *cache;
671         u64 used;
672         u64 last = max(search_hint, search_start);
673         u64 group_start = 0;
674         int full_search = 0;
675         int factor = 9;
676         int wrapped = 0;
677 again:
678         while (1) {
679                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
680                 if (!cache)
681                         break;
682
683                 spin_lock(&cache->lock);
684                 last = cache->key.objectid + cache->key.offset;
685                 used = btrfs_block_group_used(&cache->item);
686
687                 if ((full_search || !cache->ro) &&
688                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
689                         if (used + cache->pinned + cache->reserved <
690                             div_factor(cache->key.offset, factor)) {
691                                 group_start = cache->key.objectid;
692                                 spin_unlock(&cache->lock);
693                                 btrfs_put_block_group(cache);
694                                 goto found;
695                         }
696                 }
697                 spin_unlock(&cache->lock);
698                 btrfs_put_block_group(cache);
699                 cond_resched();
700         }
701         if (!wrapped) {
702                 last = search_start;
703                 wrapped = 1;
704                 goto again;
705         }
706         if (!full_search && factor < 10) {
707                 last = search_start;
708                 full_search = 1;
709                 factor = 10;
710                 goto again;
711         }
712 found:
713         return group_start;
714 }
715
716 /* simple helper to search for an existing extent at a given offset */
717 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
718 {
719         int ret;
720         struct btrfs_key key;
721         struct btrfs_path *path;
722
723         path = btrfs_alloc_path();
724         if (!path)
725                 return -ENOMEM;
726
727         key.objectid = start;
728         key.offset = len;
729         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
730         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
731                                 0, 0);
732         btrfs_free_path(path);
733         return ret;
734 }
735
736 /*
737  * helper function to lookup reference count and flags of extent.
738  *
739  * the head node for delayed ref is used to store the sum of all the
740  * reference count modifications queued up in the rbtree. the head
741  * node may also store the extent flags to set. This way you can check
742  * to see what the reference count and extent flags would be if all of
743  * the delayed refs are not processed.
744  */
745 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
746                              struct btrfs_root *root, u64 bytenr,
747                              u64 num_bytes, u64 *refs, u64 *flags)
748 {
749         struct btrfs_delayed_ref_head *head;
750         struct btrfs_delayed_ref_root *delayed_refs;
751         struct btrfs_path *path;
752         struct btrfs_extent_item *ei;
753         struct extent_buffer *leaf;
754         struct btrfs_key key;
755         u32 item_size;
756         u64 num_refs;
757         u64 extent_flags;
758         int ret;
759
760         path = btrfs_alloc_path();
761         if (!path)
762                 return -ENOMEM;
763
764         key.objectid = bytenr;
765         key.type = BTRFS_EXTENT_ITEM_KEY;
766         key.offset = num_bytes;
767         if (!trans) {
768                 path->skip_locking = 1;
769                 path->search_commit_root = 1;
770         }
771 again:
772         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
773                                 &key, path, 0, 0);
774         if (ret < 0)
775                 goto out_free;
776
777         if (ret == 0) {
778                 leaf = path->nodes[0];
779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
780                 if (item_size >= sizeof(*ei)) {
781                         ei = btrfs_item_ptr(leaf, path->slots[0],
782                                             struct btrfs_extent_item);
783                         num_refs = btrfs_extent_refs(leaf, ei);
784                         extent_flags = btrfs_extent_flags(leaf, ei);
785                 } else {
786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
787                         struct btrfs_extent_item_v0 *ei0;
788                         BUG_ON(item_size != sizeof(*ei0));
789                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
790                                              struct btrfs_extent_item_v0);
791                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
792                         /* FIXME: this isn't correct for data */
793                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
794 #else
795                         BUG();
796 #endif
797                 }
798                 BUG_ON(num_refs == 0);
799         } else {
800                 num_refs = 0;
801                 extent_flags = 0;
802                 ret = 0;
803         }
804
805         if (!trans)
806                 goto out;
807
808         delayed_refs = &trans->transaction->delayed_refs;
809         spin_lock(&delayed_refs->lock);
810         head = btrfs_find_delayed_ref_head(trans, bytenr);
811         if (head) {
812                 if (!mutex_trylock(&head->mutex)) {
813                         atomic_inc(&head->node.refs);
814                         spin_unlock(&delayed_refs->lock);
815
816                         btrfs_release_path(path);
817
818                         /*
819                          * Mutex was contended, block until it's released and try
820                          * again
821                          */
822                         mutex_lock(&head->mutex);
823                         mutex_unlock(&head->mutex);
824                         btrfs_put_delayed_ref(&head->node);
825                         goto again;
826                 }
827                 if (head->extent_op && head->extent_op->update_flags)
828                         extent_flags |= head->extent_op->flags_to_set;
829                 else
830                         BUG_ON(num_refs == 0);
831
832                 num_refs += head->node.ref_mod;
833                 mutex_unlock(&head->mutex);
834         }
835         spin_unlock(&delayed_refs->lock);
836 out:
837         WARN_ON(num_refs == 0);
838         if (refs)
839                 *refs = num_refs;
840         if (flags)
841                 *flags = extent_flags;
842 out_free:
843         btrfs_free_path(path);
844         return ret;
845 }
846
847 /*
848  * Back reference rules.  Back refs have three main goals:
849  *
850  * 1) differentiate between all holders of references to an extent so that
851  *    when a reference is dropped we can make sure it was a valid reference
852  *    before freeing the extent.
853  *
854  * 2) Provide enough information to quickly find the holders of an extent
855  *    if we notice a given block is corrupted or bad.
856  *
857  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
858  *    maintenance.  This is actually the same as #2, but with a slightly
859  *    different use case.
860  *
861  * There are two kinds of back refs. The implicit back refs is optimized
862  * for pointers in non-shared tree blocks. For a given pointer in a block,
863  * back refs of this kind provide information about the block's owner tree
864  * and the pointer's key. These information allow us to find the block by
865  * b-tree searching. The full back refs is for pointers in tree blocks not
866  * referenced by their owner trees. The location of tree block is recorded
867  * in the back refs. Actually the full back refs is generic, and can be
868  * used in all cases the implicit back refs is used. The major shortcoming
869  * of the full back refs is its overhead. Every time a tree block gets
870  * COWed, we have to update back refs entry for all pointers in it.
871  *
872  * For a newly allocated tree block, we use implicit back refs for
873  * pointers in it. This means most tree related operations only involve
874  * implicit back refs. For a tree block created in old transaction, the
875  * only way to drop a reference to it is COW it. So we can detect the
876  * event that tree block loses its owner tree's reference and do the
877  * back refs conversion.
878  *
879  * When a tree block is COW'd through a tree, there are four cases:
880  *
881  * The reference count of the block is one and the tree is the block's
882  * owner tree. Nothing to do in this case.
883  *
884  * The reference count of the block is one and the tree is not the
885  * block's owner tree. In this case, full back refs is used for pointers
886  * in the block. Remove these full back refs, add implicit back refs for
887  * every pointers in the new block.
888  *
889  * The reference count of the block is greater than one and the tree is
890  * the block's owner tree. In this case, implicit back refs is used for
891  * pointers in the block. Add full back refs for every pointers in the
892  * block, increase lower level extents' reference counts. The original
893  * implicit back refs are entailed to the new block.
894  *
895  * The reference count of the block is greater than one and the tree is
896  * not the block's owner tree. Add implicit back refs for every pointer in
897  * the new block, increase lower level extents' reference count.
898  *
899  * Back Reference Key composing:
900  *
901  * The key objectid corresponds to the first byte in the extent,
902  * The key type is used to differentiate between types of back refs.
903  * There are different meanings of the key offset for different types
904  * of back refs.
905  *
906  * File extents can be referenced by:
907  *
908  * - multiple snapshots, subvolumes, or different generations in one subvol
909  * - different files inside a single subvolume
910  * - different offsets inside a file (bookend extents in file.c)
911  *
912  * The extent ref structure for the implicit back refs has fields for:
913  *
914  * - Objectid of the subvolume root
915  * - objectid of the file holding the reference
916  * - original offset in the file
917  * - how many bookend extents
918  *
919  * The key offset for the implicit back refs is hash of the first
920  * three fields.
921  *
922  * The extent ref structure for the full back refs has field for:
923  *
924  * - number of pointers in the tree leaf
925  *
926  * The key offset for the implicit back refs is the first byte of
927  * the tree leaf
928  *
929  * When a file extent is allocated, The implicit back refs is used.
930  * the fields are filled in:
931  *
932  *     (root_key.objectid, inode objectid, offset in file, 1)
933  *
934  * When a file extent is removed file truncation, we find the
935  * corresponding implicit back refs and check the following fields:
936  *
937  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
938  *
939  * Btree extents can be referenced by:
940  *
941  * - Different subvolumes
942  *
943  * Both the implicit back refs and the full back refs for tree blocks
944  * only consist of key. The key offset for the implicit back refs is
945  * objectid of block's owner tree. The key offset for the full back refs
946  * is the first byte of parent block.
947  *
948  * When implicit back refs is used, information about the lowest key and
949  * level of the tree block are required. These information are stored in
950  * tree block info structure.
951  */
952
953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
954 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
955                                   struct btrfs_root *root,
956                                   struct btrfs_path *path,
957                                   u64 owner, u32 extra_size)
958 {
959         struct btrfs_extent_item *item;
960         struct btrfs_extent_item_v0 *ei0;
961         struct btrfs_extent_ref_v0 *ref0;
962         struct btrfs_tree_block_info *bi;
963         struct extent_buffer *leaf;
964         struct btrfs_key key;
965         struct btrfs_key found_key;
966         u32 new_size = sizeof(*item);
967         u64 refs;
968         int ret;
969
970         leaf = path->nodes[0];
971         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
972
973         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
974         ei0 = btrfs_item_ptr(leaf, path->slots[0],
975                              struct btrfs_extent_item_v0);
976         refs = btrfs_extent_refs_v0(leaf, ei0);
977
978         if (owner == (u64)-1) {
979                 while (1) {
980                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
981                                 ret = btrfs_next_leaf(root, path);
982                                 if (ret < 0)
983                                         return ret;
984                                 BUG_ON(ret > 0); /* Corruption */
985                                 leaf = path->nodes[0];
986                         }
987                         btrfs_item_key_to_cpu(leaf, &found_key,
988                                               path->slots[0]);
989                         BUG_ON(key.objectid != found_key.objectid);
990                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
991                                 path->slots[0]++;
992                                 continue;
993                         }
994                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
995                                               struct btrfs_extent_ref_v0);
996                         owner = btrfs_ref_objectid_v0(leaf, ref0);
997                         break;
998                 }
999         }
1000         btrfs_release_path(path);
1001
1002         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1003                 new_size += sizeof(*bi);
1004
1005         new_size -= sizeof(*ei0);
1006         ret = btrfs_search_slot(trans, root, &key, path,
1007                                 new_size + extra_size, 1);
1008         if (ret < 0)
1009                 return ret;
1010         BUG_ON(ret); /* Corruption */
1011
1012         btrfs_extend_item(trans, root, path, new_size);
1013
1014         leaf = path->nodes[0];
1015         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1016         btrfs_set_extent_refs(leaf, item, refs);
1017         /* FIXME: get real generation */
1018         btrfs_set_extent_generation(leaf, item, 0);
1019         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1020                 btrfs_set_extent_flags(leaf, item,
1021                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1022                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1023                 bi = (struct btrfs_tree_block_info *)(item + 1);
1024                 /* FIXME: get first key of the block */
1025                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1026                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1027         } else {
1028                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1029         }
1030         btrfs_mark_buffer_dirty(leaf);
1031         return 0;
1032 }
1033 #endif
1034
1035 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1036 {
1037         u32 high_crc = ~(u32)0;
1038         u32 low_crc = ~(u32)0;
1039         __le64 lenum;
1040
1041         lenum = cpu_to_le64(root_objectid);
1042         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1043         lenum = cpu_to_le64(owner);
1044         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(offset);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047
1048         return ((u64)high_crc << 31) ^ (u64)low_crc;
1049 }
1050
1051 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1052                                      struct btrfs_extent_data_ref *ref)
1053 {
1054         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1055                                     btrfs_extent_data_ref_objectid(leaf, ref),
1056                                     btrfs_extent_data_ref_offset(leaf, ref));
1057 }
1058
1059 static int match_extent_data_ref(struct extent_buffer *leaf,
1060                                  struct btrfs_extent_data_ref *ref,
1061                                  u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1064             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1065             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1066                 return 0;
1067         return 1;
1068 }
1069
1070 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1071                                            struct btrfs_root *root,
1072                                            struct btrfs_path *path,
1073                                            u64 bytenr, u64 parent,
1074                                            u64 root_objectid,
1075                                            u64 owner, u64 offset)
1076 {
1077         struct btrfs_key key;
1078         struct btrfs_extent_data_ref *ref;
1079         struct extent_buffer *leaf;
1080         u32 nritems;
1081         int ret;
1082         int recow;
1083         int err = -ENOENT;
1084
1085         key.objectid = bytenr;
1086         if (parent) {
1087                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1088                 key.offset = parent;
1089         } else {
1090                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1091                 key.offset = hash_extent_data_ref(root_objectid,
1092                                                   owner, offset);
1093         }
1094 again:
1095         recow = 0;
1096         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1097         if (ret < 0) {
1098                 err = ret;
1099                 goto fail;
1100         }
1101
1102         if (parent) {
1103                 if (!ret)
1104                         return 0;
1105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1106                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1107                 btrfs_release_path(path);
1108                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109                 if (ret < 0) {
1110                         err = ret;
1111                         goto fail;
1112                 }
1113                 if (!ret)
1114                         return 0;
1115 #endif
1116                 goto fail;
1117         }
1118
1119         leaf = path->nodes[0];
1120         nritems = btrfs_header_nritems(leaf);
1121         while (1) {
1122                 if (path->slots[0] >= nritems) {
1123                         ret = btrfs_next_leaf(root, path);
1124                         if (ret < 0)
1125                                 err = ret;
1126                         if (ret)
1127                                 goto fail;
1128
1129                         leaf = path->nodes[0];
1130                         nritems = btrfs_header_nritems(leaf);
1131                         recow = 1;
1132                 }
1133
1134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135                 if (key.objectid != bytenr ||
1136                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1137                         goto fail;
1138
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141
1142                 if (match_extent_data_ref(leaf, ref, root_objectid,
1143                                           owner, offset)) {
1144                         if (recow) {
1145                                 btrfs_release_path(path);
1146                                 goto again;
1147                         }
1148                         err = 0;
1149                         break;
1150                 }
1151                 path->slots[0]++;
1152         }
1153 fail:
1154         return err;
1155 }
1156
1157 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1158                                            struct btrfs_root *root,
1159                                            struct btrfs_path *path,
1160                                            u64 bytenr, u64 parent,
1161                                            u64 root_objectid, u64 owner,
1162                                            u64 offset, int refs_to_add)
1163 {
1164         struct btrfs_key key;
1165         struct extent_buffer *leaf;
1166         u32 size;
1167         u32 num_refs;
1168         int ret;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174                 size = sizeof(struct btrfs_shared_data_ref);
1175         } else {
1176                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1177                 key.offset = hash_extent_data_ref(root_objectid,
1178                                                   owner, offset);
1179                 size = sizeof(struct btrfs_extent_data_ref);
1180         }
1181
1182         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1183         if (ret && ret != -EEXIST)
1184                 goto fail;
1185
1186         leaf = path->nodes[0];
1187         if (parent) {
1188                 struct btrfs_shared_data_ref *ref;
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_shared_data_ref);
1191                 if (ret == 0) {
1192                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1193                 } else {
1194                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1195                         num_refs += refs_to_add;
1196                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1197                 }
1198         } else {
1199                 struct btrfs_extent_data_ref *ref;
1200                 while (ret == -EEXIST) {
1201                         ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                              struct btrfs_extent_data_ref);
1203                         if (match_extent_data_ref(leaf, ref, root_objectid,
1204                                                   owner, offset))
1205                                 break;
1206                         btrfs_release_path(path);
1207                         key.offset++;
1208                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1209                                                       size);
1210                         if (ret && ret != -EEXIST)
1211                                 goto fail;
1212
1213                         leaf = path->nodes[0];
1214                 }
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_extent_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_extent_data_ref_root(leaf, ref,
1219                                                        root_objectid);
1220                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1221                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1222                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1223                 } else {
1224                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1225                         num_refs += refs_to_add;
1226                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1227                 }
1228         }
1229         btrfs_mark_buffer_dirty(leaf);
1230         ret = 0;
1231 fail:
1232         btrfs_release_path(path);
1233         return ret;
1234 }
1235
1236 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1237                                            struct btrfs_root *root,
1238                                            struct btrfs_path *path,
1239                                            int refs_to_drop)
1240 {
1241         struct btrfs_key key;
1242         struct btrfs_extent_data_ref *ref1 = NULL;
1243         struct btrfs_shared_data_ref *ref2 = NULL;
1244         struct extent_buffer *leaf;
1245         u32 num_refs = 0;
1246         int ret = 0;
1247
1248         leaf = path->nodes[0];
1249         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1250
1251         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1252                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1253                                       struct btrfs_extent_data_ref);
1254                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1255         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1256                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1257                                       struct btrfs_shared_data_ref);
1258                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1259 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1260         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1261                 struct btrfs_extent_ref_v0 *ref0;
1262                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1263                                       struct btrfs_extent_ref_v0);
1264                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1265 #endif
1266         } else {
1267                 BUG();
1268         }
1269
1270         BUG_ON(num_refs < refs_to_drop);
1271         num_refs -= refs_to_drop;
1272
1273         if (num_refs == 0) {
1274                 ret = btrfs_del_item(trans, root, path);
1275         } else {
1276                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1277                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1278                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1279                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281                 else {
1282                         struct btrfs_extent_ref_v0 *ref0;
1283                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                         struct btrfs_extent_ref_v0);
1285                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1286                 }
1287 #endif
1288                 btrfs_mark_buffer_dirty(leaf);
1289         }
1290         return ret;
1291 }
1292
1293 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           struct btrfs_extent_inline_ref *iref)
1296 {
1297         struct btrfs_key key;
1298         struct extent_buffer *leaf;
1299         struct btrfs_extent_data_ref *ref1;
1300         struct btrfs_shared_data_ref *ref2;
1301         u32 num_refs = 0;
1302
1303         leaf = path->nodes[0];
1304         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1305         if (iref) {
1306                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1307                     BTRFS_EXTENT_DATA_REF_KEY) {
1308                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310                 } else {
1311                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1313                 }
1314         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1315                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1316                                       struct btrfs_extent_data_ref);
1317                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1318         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1319                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_shared_data_ref);
1321                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1322 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1323         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1324                 struct btrfs_extent_ref_v0 *ref0;
1325                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1326                                       struct btrfs_extent_ref_v0);
1327                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1328 #endif
1329         } else {
1330                 WARN_ON(1);
1331         }
1332         return num_refs;
1333 }
1334
1335 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1336                                           struct btrfs_root *root,
1337                                           struct btrfs_path *path,
1338                                           u64 bytenr, u64 parent,
1339                                           u64 root_objectid)
1340 {
1341         struct btrfs_key key;
1342         int ret;
1343
1344         key.objectid = bytenr;
1345         if (parent) {
1346                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1347                 key.offset = parent;
1348         } else {
1349                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1350                 key.offset = root_objectid;
1351         }
1352
1353         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1354         if (ret > 0)
1355                 ret = -ENOENT;
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         if (ret == -ENOENT && parent) {
1358                 btrfs_release_path(path);
1359                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1360                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1361                 if (ret > 0)
1362                         ret = -ENOENT;
1363         }
1364 #endif
1365         return ret;
1366 }
1367
1368 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1369                                           struct btrfs_root *root,
1370                                           struct btrfs_path *path,
1371                                           u64 bytenr, u64 parent,
1372                                           u64 root_objectid)
1373 {
1374         struct btrfs_key key;
1375         int ret;
1376
1377         key.objectid = bytenr;
1378         if (parent) {
1379                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1380                 key.offset = parent;
1381         } else {
1382                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1383                 key.offset = root_objectid;
1384         }
1385
1386         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1387         btrfs_release_path(path);
1388         return ret;
1389 }
1390
1391 static inline int extent_ref_type(u64 parent, u64 owner)
1392 {
1393         int type;
1394         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1395                 if (parent > 0)
1396                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1397                 else
1398                         type = BTRFS_TREE_BLOCK_REF_KEY;
1399         } else {
1400                 if (parent > 0)
1401                         type = BTRFS_SHARED_DATA_REF_KEY;
1402                 else
1403                         type = BTRFS_EXTENT_DATA_REF_KEY;
1404         }
1405         return type;
1406 }
1407
1408 static int find_next_key(struct btrfs_path *path, int level,
1409                          struct btrfs_key *key)
1410
1411 {
1412         for (; level < BTRFS_MAX_LEVEL; level++) {
1413                 if (!path->nodes[level])
1414                         break;
1415                 if (path->slots[level] + 1 >=
1416                     btrfs_header_nritems(path->nodes[level]))
1417                         continue;
1418                 if (level == 0)
1419                         btrfs_item_key_to_cpu(path->nodes[level], key,
1420                                               path->slots[level] + 1);
1421                 else
1422                         btrfs_node_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 return 0;
1425         }
1426         return 1;
1427 }
1428
1429 /*
1430  * look for inline back ref. if back ref is found, *ref_ret is set
1431  * to the address of inline back ref, and 0 is returned.
1432  *
1433  * if back ref isn't found, *ref_ret is set to the address where it
1434  * should be inserted, and -ENOENT is returned.
1435  *
1436  * if insert is true and there are too many inline back refs, the path
1437  * points to the extent item, and -EAGAIN is returned.
1438  *
1439  * NOTE: inline back refs are ordered in the same way that back ref
1440  *       items in the tree are ordered.
1441  */
1442 static noinline_for_stack
1443 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1444                                  struct btrfs_root *root,
1445                                  struct btrfs_path *path,
1446                                  struct btrfs_extent_inline_ref **ref_ret,
1447                                  u64 bytenr, u64 num_bytes,
1448                                  u64 parent, u64 root_objectid,
1449                                  u64 owner, u64 offset, int insert)
1450 {
1451         struct btrfs_key key;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *ei;
1454         struct btrfs_extent_inline_ref *iref;
1455         u64 flags;
1456         u64 item_size;
1457         unsigned long ptr;
1458         unsigned long end;
1459         int extra_size;
1460         int type;
1461         int want;
1462         int ret;
1463         int err = 0;
1464
1465         key.objectid = bytenr;
1466         key.type = BTRFS_EXTENT_ITEM_KEY;
1467         key.offset = num_bytes;
1468
1469         want = extent_ref_type(parent, owner);
1470         if (insert) {
1471                 extra_size = btrfs_extent_inline_ref_size(want);
1472                 path->keep_locks = 1;
1473         } else
1474                 extra_size = -1;
1475         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1476         if (ret < 0) {
1477                 err = ret;
1478                 goto out;
1479         }
1480         if (ret && !insert) {
1481                 err = -ENOENT;
1482                 goto out;
1483         }
1484         BUG_ON(ret); /* Corruption */
1485
1486         leaf = path->nodes[0];
1487         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1489         if (item_size < sizeof(*ei)) {
1490                 if (!insert) {
1491                         err = -ENOENT;
1492                         goto out;
1493                 }
1494                 ret = convert_extent_item_v0(trans, root, path, owner,
1495                                              extra_size);
1496                 if (ret < 0) {
1497                         err = ret;
1498                         goto out;
1499                 }
1500                 leaf = path->nodes[0];
1501                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1502         }
1503 #endif
1504         BUG_ON(item_size < sizeof(*ei));
1505
1506         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1507         flags = btrfs_extent_flags(leaf, ei);
1508
1509         ptr = (unsigned long)(ei + 1);
1510         end = (unsigned long)ei + item_size;
1511
1512         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1513                 ptr += sizeof(struct btrfs_tree_block_info);
1514                 BUG_ON(ptr > end);
1515         } else {
1516                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1517         }
1518
1519         err = -ENOENT;
1520         while (1) {
1521                 if (ptr >= end) {
1522                         WARN_ON(ptr > end);
1523                         break;
1524                 }
1525                 iref = (struct btrfs_extent_inline_ref *)ptr;
1526                 type = btrfs_extent_inline_ref_type(leaf, iref);
1527                 if (want < type)
1528                         break;
1529                 if (want > type) {
1530                         ptr += btrfs_extent_inline_ref_size(type);
1531                         continue;
1532                 }
1533
1534                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1535                         struct btrfs_extent_data_ref *dref;
1536                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1537                         if (match_extent_data_ref(leaf, dref, root_objectid,
1538                                                   owner, offset)) {
1539                                 err = 0;
1540                                 break;
1541                         }
1542                         if (hash_extent_data_ref_item(leaf, dref) <
1543                             hash_extent_data_ref(root_objectid, owner, offset))
1544                                 break;
1545                 } else {
1546                         u64 ref_offset;
1547                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1548                         if (parent > 0) {
1549                                 if (parent == ref_offset) {
1550                                         err = 0;
1551                                         break;
1552                                 }
1553                                 if (ref_offset < parent)
1554                                         break;
1555                         } else {
1556                                 if (root_objectid == ref_offset) {
1557                                         err = 0;
1558                                         break;
1559                                 }
1560                                 if (ref_offset < root_objectid)
1561                                         break;
1562                         }
1563                 }
1564                 ptr += btrfs_extent_inline_ref_size(type);
1565         }
1566         if (err == -ENOENT && insert) {
1567                 if (item_size + extra_size >=
1568                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1569                         err = -EAGAIN;
1570                         goto out;
1571                 }
1572                 /*
1573                  * To add new inline back ref, we have to make sure
1574                  * there is no corresponding back ref item.
1575                  * For simplicity, we just do not add new inline back
1576                  * ref if there is any kind of item for this block
1577                  */
1578                 if (find_next_key(path, 0, &key) == 0 &&
1579                     key.objectid == bytenr &&
1580                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1581                         err = -EAGAIN;
1582                         goto out;
1583                 }
1584         }
1585         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1586 out:
1587         if (insert) {
1588                 path->keep_locks = 0;
1589                 btrfs_unlock_up_safe(path, 1);
1590         }
1591         return err;
1592 }
1593
1594 /*
1595  * helper to add new inline back ref
1596  */
1597 static noinline_for_stack
1598 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1599                                  struct btrfs_root *root,
1600                                  struct btrfs_path *path,
1601                                  struct btrfs_extent_inline_ref *iref,
1602                                  u64 parent, u64 root_objectid,
1603                                  u64 owner, u64 offset, int refs_to_add,
1604                                  struct btrfs_delayed_extent_op *extent_op)
1605 {
1606         struct extent_buffer *leaf;
1607         struct btrfs_extent_item *ei;
1608         unsigned long ptr;
1609         unsigned long end;
1610         unsigned long item_offset;
1611         u64 refs;
1612         int size;
1613         int type;
1614
1615         leaf = path->nodes[0];
1616         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1617         item_offset = (unsigned long)iref - (unsigned long)ei;
1618
1619         type = extent_ref_type(parent, owner);
1620         size = btrfs_extent_inline_ref_size(type);
1621
1622         btrfs_extend_item(trans, root, path, size);
1623
1624         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1625         refs = btrfs_extent_refs(leaf, ei);
1626         refs += refs_to_add;
1627         btrfs_set_extent_refs(leaf, ei, refs);
1628         if (extent_op)
1629                 __run_delayed_extent_op(extent_op, leaf, ei);
1630
1631         ptr = (unsigned long)ei + item_offset;
1632         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1633         if (ptr < end - size)
1634                 memmove_extent_buffer(leaf, ptr + size, ptr,
1635                                       end - size - ptr);
1636
1637         iref = (struct btrfs_extent_inline_ref *)ptr;
1638         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1639         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1640                 struct btrfs_extent_data_ref *dref;
1641                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1642                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1643                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1644                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1645                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1646         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1647                 struct btrfs_shared_data_ref *sref;
1648                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1649                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1655         }
1656         btrfs_mark_buffer_dirty(leaf);
1657 }
1658
1659 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1660                                  struct btrfs_root *root,
1661                                  struct btrfs_path *path,
1662                                  struct btrfs_extent_inline_ref **ref_ret,
1663                                  u64 bytenr, u64 num_bytes, u64 parent,
1664                                  u64 root_objectid, u64 owner, u64 offset)
1665 {
1666         int ret;
1667
1668         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1669                                            bytenr, num_bytes, parent,
1670                                            root_objectid, owner, offset, 0);
1671         if (ret != -ENOENT)
1672                 return ret;
1673
1674         btrfs_release_path(path);
1675         *ref_ret = NULL;
1676
1677         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1678                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1679                                             root_objectid);
1680         } else {
1681                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1682                                              root_objectid, owner, offset);
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * helper to update/remove inline back ref
1689  */
1690 static noinline_for_stack
1691 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1692                                   struct btrfs_root *root,
1693                                   struct btrfs_path *path,
1694                                   struct btrfs_extent_inline_ref *iref,
1695                                   int refs_to_mod,
1696                                   struct btrfs_delayed_extent_op *extent_op)
1697 {
1698         struct extent_buffer *leaf;
1699         struct btrfs_extent_item *ei;
1700         struct btrfs_extent_data_ref *dref = NULL;
1701         struct btrfs_shared_data_ref *sref = NULL;
1702         unsigned long ptr;
1703         unsigned long end;
1704         u32 item_size;
1705         int size;
1706         int type;
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                 btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751 }
1752
1753 static noinline_for_stack
1754 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1755                                  struct btrfs_root *root,
1756                                  struct btrfs_path *path,
1757                                  u64 bytenr, u64 num_bytes, u64 parent,
1758                                  u64 root_objectid, u64 owner,
1759                                  u64 offset, int refs_to_add,
1760                                  struct btrfs_delayed_extent_op *extent_op)
1761 {
1762         struct btrfs_extent_inline_ref *iref;
1763         int ret;
1764
1765         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1766                                            bytenr, num_bytes, parent,
1767                                            root_objectid, owner, offset, 1);
1768         if (ret == 0) {
1769                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1770                 update_inline_extent_backref(trans, root, path, iref,
1771                                              refs_to_add, extent_op);
1772         } else if (ret == -ENOENT) {
1773                 setup_inline_extent_backref(trans, root, path, iref, parent,
1774                                             root_objectid, owner, offset,
1775                                             refs_to_add, extent_op);
1776                 ret = 0;
1777         }
1778         return ret;
1779 }
1780
1781 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1782                                  struct btrfs_root *root,
1783                                  struct btrfs_path *path,
1784                                  u64 bytenr, u64 parent, u64 root_objectid,
1785                                  u64 owner, u64 offset, int refs_to_add)
1786 {
1787         int ret;
1788         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1789                 BUG_ON(refs_to_add != 1);
1790                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1791                                             parent, root_objectid);
1792         } else {
1793                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1794                                              parent, root_objectid,
1795                                              owner, offset, refs_to_add);
1796         }
1797         return ret;
1798 }
1799
1800 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref *iref,
1804                                  int refs_to_drop, int is_data)
1805 {
1806         int ret = 0;
1807
1808         BUG_ON(!is_data && refs_to_drop != 1);
1809         if (iref) {
1810                 update_inline_extent_backref(trans, root, path, iref,
1811                                              -refs_to_drop, NULL);
1812         } else if (is_data) {
1813                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1814         } else {
1815                 ret = btrfs_del_item(trans, root, path);
1816         }
1817         return ret;
1818 }
1819
1820 static int btrfs_issue_discard(struct block_device *bdev,
1821                                 u64 start, u64 len)
1822 {
1823         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1824 }
1825
1826 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1827                                 u64 num_bytes, u64 *actual_bytes)
1828 {
1829         int ret;
1830         u64 discarded_bytes = 0;
1831         struct btrfs_bio *bbio = NULL;
1832
1833
1834         /* Tell the block device(s) that the sectors can be discarded */
1835         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1836                               bytenr, &num_bytes, &bbio, 0);
1837         /* Error condition is -ENOMEM */
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; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
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 /* Can return -ENOMEM */
1873 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1874                          struct btrfs_root *root,
1875                          u64 bytenr, u64 num_bytes, u64 parent,
1876                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1877 {
1878         int ret;
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1882                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1883
1884         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1885                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1886                                         num_bytes,
1887                                         parent, root_objectid, (int)owner,
1888                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1889         } else {
1890                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1891                                         num_bytes,
1892                                         parent, root_objectid, owner, offset,
1893                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1894         }
1895         return ret;
1896 }
1897
1898 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1899                                   struct btrfs_root *root,
1900                                   u64 bytenr, u64 num_bytes,
1901                                   u64 parent, u64 root_objectid,
1902                                   u64 owner, u64 offset, int refs_to_add,
1903                                   struct btrfs_delayed_extent_op *extent_op)
1904 {
1905         struct btrfs_path *path;
1906         struct extent_buffer *leaf;
1907         struct btrfs_extent_item *item;
1908         u64 refs;
1909         int ret;
1910         int err = 0;
1911
1912         path = btrfs_alloc_path();
1913         if (!path)
1914                 return -ENOMEM;
1915
1916         path->reada = 1;
1917         path->leave_spinning = 1;
1918         /* this will setup the path even if it fails to insert the back ref */
1919         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1920                                            path, bytenr, num_bytes, parent,
1921                                            root_objectid, owner, offset,
1922                                            refs_to_add, extent_op);
1923         if (ret == 0)
1924                 goto out;
1925
1926         if (ret != -EAGAIN) {
1927                 err = ret;
1928                 goto out;
1929         }
1930
1931         leaf = path->nodes[0];
1932         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1933         refs = btrfs_extent_refs(leaf, item);
1934         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1935         if (extent_op)
1936                 __run_delayed_extent_op(extent_op, leaf, item);
1937
1938         btrfs_mark_buffer_dirty(leaf);
1939         btrfs_release_path(path);
1940
1941         path->reada = 1;
1942         path->leave_spinning = 1;
1943
1944         /* now insert the actual backref */
1945         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1946                                     path, bytenr, parent, root_objectid,
1947                                     owner, offset, refs_to_add);
1948         if (ret)
1949                 btrfs_abort_transaction(trans, root, ret);
1950 out:
1951         btrfs_free_path(path);
1952         return err;
1953 }
1954
1955 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1956                                 struct btrfs_root *root,
1957                                 struct btrfs_delayed_ref_node *node,
1958                                 struct btrfs_delayed_extent_op *extent_op,
1959                                 int insert_reserved)
1960 {
1961         int ret = 0;
1962         struct btrfs_delayed_data_ref *ref;
1963         struct btrfs_key ins;
1964         u64 parent = 0;
1965         u64 ref_root = 0;
1966         u64 flags = 0;
1967
1968         ins.objectid = node->bytenr;
1969         ins.offset = node->num_bytes;
1970         ins.type = BTRFS_EXTENT_ITEM_KEY;
1971
1972         ref = btrfs_delayed_node_to_data_ref(node);
1973         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1974                 parent = ref->parent;
1975         else
1976                 ref_root = ref->root;
1977
1978         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1979                 if (extent_op) {
1980                         BUG_ON(extent_op->update_key);
1981                         flags |= extent_op->flags_to_set;
1982                 }
1983                 ret = alloc_reserved_file_extent(trans, root,
1984                                                  parent, ref_root, flags,
1985                                                  ref->objectid, ref->offset,
1986                                                  &ins, node->ref_mod);
1987         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1988                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1989                                              node->num_bytes, parent,
1990                                              ref_root, ref->objectid,
1991                                              ref->offset, node->ref_mod,
1992                                              extent_op);
1993         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1994                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1995                                           node->num_bytes, parent,
1996                                           ref_root, ref->objectid,
1997                                           ref->offset, node->ref_mod,
1998                                           extent_op);
1999         } else {
2000                 BUG();
2001         }
2002         return ret;
2003 }
2004
2005 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2006                                     struct extent_buffer *leaf,
2007                                     struct btrfs_extent_item *ei)
2008 {
2009         u64 flags = btrfs_extent_flags(leaf, ei);
2010         if (extent_op->update_flags) {
2011                 flags |= extent_op->flags_to_set;
2012                 btrfs_set_extent_flags(leaf, ei, flags);
2013         }
2014
2015         if (extent_op->update_key) {
2016                 struct btrfs_tree_block_info *bi;
2017                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2018                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2019                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2020         }
2021 }
2022
2023 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2024                                  struct btrfs_root *root,
2025                                  struct btrfs_delayed_ref_node *node,
2026                                  struct btrfs_delayed_extent_op *extent_op)
2027 {
2028         struct btrfs_key key;
2029         struct btrfs_path *path;
2030         struct btrfs_extent_item *ei;
2031         struct extent_buffer *leaf;
2032         u32 item_size;
2033         int ret;
2034         int err = 0;
2035
2036         if (trans->aborted)
2037                 return 0;
2038
2039         path = btrfs_alloc_path();
2040         if (!path)
2041                 return -ENOMEM;
2042
2043         key.objectid = node->bytenr;
2044         key.type = BTRFS_EXTENT_ITEM_KEY;
2045         key.offset = node->num_bytes;
2046
2047         path->reada = 1;
2048         path->leave_spinning = 1;
2049         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2050                                 path, 0, 1);
2051         if (ret < 0) {
2052                 err = ret;
2053                 goto out;
2054         }
2055         if (ret > 0) {
2056                 err = -EIO;
2057                 goto out;
2058         }
2059
2060         leaf = path->nodes[0];
2061         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2062 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2063         if (item_size < sizeof(*ei)) {
2064                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2065                                              path, (u64)-1, 0);
2066                 if (ret < 0) {
2067                         err = ret;
2068                         goto out;
2069                 }
2070                 leaf = path->nodes[0];
2071                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2072         }
2073 #endif
2074         BUG_ON(item_size < sizeof(*ei));
2075         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2076         __run_delayed_extent_op(extent_op, leaf, ei);
2077
2078         btrfs_mark_buffer_dirty(leaf);
2079 out:
2080         btrfs_free_path(path);
2081         return err;
2082 }
2083
2084 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2085                                 struct btrfs_root *root,
2086                                 struct btrfs_delayed_ref_node *node,
2087                                 struct btrfs_delayed_extent_op *extent_op,
2088                                 int insert_reserved)
2089 {
2090         int ret = 0;
2091         struct btrfs_delayed_tree_ref *ref;
2092         struct btrfs_key ins;
2093         u64 parent = 0;
2094         u64 ref_root = 0;
2095
2096         ins.objectid = node->bytenr;
2097         ins.offset = node->num_bytes;
2098         ins.type = BTRFS_EXTENT_ITEM_KEY;
2099
2100         ref = btrfs_delayed_node_to_tree_ref(node);
2101         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2102                 parent = ref->parent;
2103         else
2104                 ref_root = ref->root;
2105
2106         BUG_ON(node->ref_mod != 1);
2107         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2108                 BUG_ON(!extent_op || !extent_op->update_flags ||
2109                        !extent_op->update_key);
2110                 ret = alloc_reserved_tree_block(trans, root,
2111                                                 parent, ref_root,
2112                                                 extent_op->flags_to_set,
2113                                                 &extent_op->key,
2114                                                 ref->level, &ins);
2115         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2116                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2117                                              node->num_bytes, parent, ref_root,
2118                                              ref->level, 0, 1, extent_op);
2119         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2120                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2121                                           node->num_bytes, parent, ref_root,
2122                                           ref->level, 0, 1, extent_op);
2123         } else {
2124                 BUG();
2125         }
2126         return ret;
2127 }
2128
2129 /* helper function to actually process a single delayed ref entry */
2130 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2131                                struct btrfs_root *root,
2132                                struct btrfs_delayed_ref_node *node,
2133                                struct btrfs_delayed_extent_op *extent_op,
2134                                int insert_reserved)
2135 {
2136         int ret = 0;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (btrfs_delayed_ref_is_head(node)) {
2142                 struct btrfs_delayed_ref_head *head;
2143                 /*
2144                  * we've hit the end of the chain and we were supposed
2145                  * to insert this extent into the tree.  But, it got
2146                  * deleted before we ever needed to insert it, so all
2147                  * we have to do is clean up the accounting
2148                  */
2149                 BUG_ON(extent_op);
2150                 head = btrfs_delayed_node_to_head(node);
2151                 if (insert_reserved) {
2152                         btrfs_pin_extent(root, node->bytenr,
2153                                          node->num_bytes, 1);
2154                         if (head->is_data) {
2155                                 ret = btrfs_del_csums(trans, root,
2156                                                       node->bytenr,
2157                                                       node->num_bytes);
2158                         }
2159                 }
2160                 mutex_unlock(&head->mutex);
2161                 return ret;
2162         }
2163
2164         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2165             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2166                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2167                                            insert_reserved);
2168         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2169                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2170                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2171                                            insert_reserved);
2172         else
2173                 BUG();
2174         return ret;
2175 }
2176
2177 static noinline struct btrfs_delayed_ref_node *
2178 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2179 {
2180         struct rb_node *node;
2181         struct btrfs_delayed_ref_node *ref;
2182         int action = BTRFS_ADD_DELAYED_REF;
2183 again:
2184         /*
2185          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2186          * this prevents ref count from going down to zero when
2187          * there still are pending delayed ref.
2188          */
2189         node = rb_prev(&head->node.rb_node);
2190         while (1) {
2191                 if (!node)
2192                         break;
2193                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2194                                 rb_node);
2195                 if (ref->bytenr != head->node.bytenr)
2196                         break;
2197                 if (ref->action == action)
2198                         return ref;
2199                 node = rb_prev(node);
2200         }
2201         if (action == BTRFS_ADD_DELAYED_REF) {
2202                 action = BTRFS_DROP_DELAYED_REF;
2203                 goto again;
2204         }
2205         return NULL;
2206 }
2207
2208 /*
2209  * Returns 0 on success or if called with an already aborted transaction.
2210  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2211  */
2212 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2213                                        struct btrfs_root *root,
2214                                        struct list_head *cluster)
2215 {
2216         struct btrfs_delayed_ref_root *delayed_refs;
2217         struct btrfs_delayed_ref_node *ref;
2218         struct btrfs_delayed_ref_head *locked_ref = NULL;
2219         struct btrfs_delayed_extent_op *extent_op;
2220         int ret;
2221         int count = 0;
2222         int must_insert_reserved = 0;
2223
2224         delayed_refs = &trans->transaction->delayed_refs;
2225         while (1) {
2226                 if (!locked_ref) {
2227                         /* pick a new head ref from the cluster list */
2228                         if (list_empty(cluster))
2229                                 break;
2230
2231                         locked_ref = list_entry(cluster->next,
2232                                      struct btrfs_delayed_ref_head, cluster);
2233
2234                         /* grab the lock that says we are going to process
2235                          * all the refs for this head */
2236                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2237
2238                         /*
2239                          * we may have dropped the spin lock to get the head
2240                          * mutex lock, and that might have given someone else
2241                          * time to free the head.  If that's true, it has been
2242                          * removed from our list and we can move on.
2243                          */
2244                         if (ret == -EAGAIN) {
2245                                 locked_ref = NULL;
2246                                 count++;
2247                                 continue;
2248                         }
2249                 }
2250
2251                 /*
2252                  * locked_ref is the head node, so we have to go one
2253                  * node back for any delayed ref updates
2254                  */
2255                 ref = select_delayed_ref(locked_ref);
2256
2257                 if (ref && ref->seq &&
2258                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2259                         /*
2260                          * there are still refs with lower seq numbers in the
2261                          * process of being added. Don't run this ref yet.
2262                          */
2263                         list_del_init(&locked_ref->cluster);
2264                         mutex_unlock(&locked_ref->mutex);
2265                         locked_ref = NULL;
2266                         delayed_refs->num_heads_ready++;
2267                         spin_unlock(&delayed_refs->lock);
2268                         cond_resched();
2269                         spin_lock(&delayed_refs->lock);
2270                         continue;
2271                 }
2272
2273                 /*
2274                  * record the must insert reserved flag before we
2275                  * drop the spin lock.
2276                  */
2277                 must_insert_reserved = locked_ref->must_insert_reserved;
2278                 locked_ref->must_insert_reserved = 0;
2279
2280                 extent_op = locked_ref->extent_op;
2281                 locked_ref->extent_op = NULL;
2282
2283                 if (!ref) {
2284                         /* All delayed refs have been processed, Go ahead
2285                          * and send the head node to run_one_delayed_ref,
2286                          * so that any accounting fixes can happen
2287                          */
2288                         ref = &locked_ref->node;
2289
2290                         if (extent_op && must_insert_reserved) {
2291                                 kfree(extent_op);
2292                                 extent_op = NULL;
2293                         }
2294
2295                         if (extent_op) {
2296                                 spin_unlock(&delayed_refs->lock);
2297
2298                                 ret = run_delayed_extent_op(trans, root,
2299                                                             ref, extent_op);
2300                                 kfree(extent_op);
2301
2302                                 if (ret) {
2303                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2304                                         spin_lock(&delayed_refs->lock);
2305                                         return ret;
2306                                 }
2307
2308                                 goto next;
2309                         }
2310
2311                         list_del_init(&locked_ref->cluster);
2312                         locked_ref = NULL;
2313                 }
2314
2315                 ref->in_tree = 0;
2316                 rb_erase(&ref->rb_node, &delayed_refs->root);
2317                 delayed_refs->num_entries--;
2318                 /*
2319                  * we modified num_entries, but as we're currently running
2320                  * delayed refs, skip
2321                  *     wake_up(&delayed_refs->seq_wait);
2322                  * here.
2323                  */
2324                 spin_unlock(&delayed_refs->lock);
2325
2326                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2327                                           must_insert_reserved);
2328
2329                 btrfs_put_delayed_ref(ref);
2330                 kfree(extent_op);
2331                 count++;
2332
2333                 if (ret) {
2334                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2335                         spin_lock(&delayed_refs->lock);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350
2351 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2352                         unsigned long num_refs)
2353 {
2354         struct list_head *first_seq = delayed_refs->seq_head.next;
2355
2356         spin_unlock(&delayed_refs->lock);
2357         pr_debug("waiting for more refs (num %ld, first %p)\n",
2358                  num_refs, first_seq);
2359         wait_event(delayed_refs->seq_wait,
2360                    num_refs != delayed_refs->num_entries ||
2361                    delayed_refs->seq_head.next != first_seq);
2362         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2363                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2364         spin_lock(&delayed_refs->lock);
2365 }
2366
2367 /*
2368  * this starts processing the delayed reference count updates and
2369  * extent insertions we have queued up so far.  count can be
2370  * 0, which means to process everything in the tree at the start
2371  * of the run (but not newly added entries), or it can be some target
2372  * number you'd like to process.
2373  *
2374  * Returns 0 on success or if called with an aborted transaction
2375  * Returns <0 on error and aborts the transaction
2376  */
2377 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2378                            struct btrfs_root *root, unsigned long count)
2379 {
2380         struct rb_node *node;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct btrfs_delayed_ref_node *ref;
2383         struct list_head cluster;
2384         int ret;
2385         u64 delayed_start;
2386         int run_all = count == (unsigned long)-1;
2387         int run_most = 0;
2388         unsigned long num_refs = 0;
2389         int consider_waiting;
2390
2391         /* We'll clean this up in btrfs_cleanup_transaction */
2392         if (trans->aborted)
2393                 return 0;
2394
2395         if (root == root->fs_info->extent_root)
2396                 root = root->fs_info->tree_root;
2397
2398         do_chunk_alloc(trans, root->fs_info->extent_root,
2399                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2400                        CHUNK_ALLOC_NO_FORCE);
2401
2402         delayed_refs = &trans->transaction->delayed_refs;
2403         INIT_LIST_HEAD(&cluster);
2404 again:
2405         consider_waiting = 0;
2406         spin_lock(&delayed_refs->lock);
2407         if (count == 0) {
2408                 count = delayed_refs->num_entries * 2;
2409                 run_most = 1;
2410         }
2411         while (1) {
2412                 if (!(run_all || run_most) &&
2413                     delayed_refs->num_heads_ready < 64)
2414                         break;
2415
2416                 /*
2417                  * go find something we can process in the rbtree.  We start at
2418                  * the beginning of the tree, and then build a cluster
2419                  * of refs to process starting at the first one we are able to
2420                  * lock
2421                  */
2422                 delayed_start = delayed_refs->run_delayed_start;
2423                 ret = btrfs_find_ref_cluster(trans, &cluster,
2424                                              delayed_refs->run_delayed_start);
2425                 if (ret)
2426                         break;
2427
2428                 if (delayed_start >= delayed_refs->run_delayed_start) {
2429                         if (consider_waiting == 0) {
2430                                 /*
2431                                  * btrfs_find_ref_cluster looped. let's do one
2432                                  * more cycle. if we don't run any delayed ref
2433                                  * during that cycle (because we can't because
2434                                  * all of them are blocked) and if the number of
2435                                  * refs doesn't change, we avoid busy waiting.
2436                                  */
2437                                 consider_waiting = 1;
2438                                 num_refs = delayed_refs->num_entries;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs, num_refs);
2441                                 /*
2442                                  * after waiting, things have changed. we
2443                                  * dropped the lock and someone else might have
2444                                  * run some refs, built new clusters and so on.
2445                                  * therefore, we restart staleness detection.
2446                                  */
2447                                 consider_waiting = 0;
2448                         }
2449                 }
2450
2451                 ret = run_clustered_refs(trans, root, &cluster);
2452                 if (ret < 0) {
2453                         spin_unlock(&delayed_refs->lock);
2454                         btrfs_abort_transaction(trans, root, ret);
2455                         return ret;
2456                 }
2457
2458                 count -= min_t(unsigned long, ret, count);
2459
2460                 if (count == 0)
2461                         break;
2462
2463                 if (ret || delayed_refs->run_delayed_start == 0) {
2464                         /* refs were run, let's reset staleness detection */
2465                         consider_waiting = 0;
2466                 }
2467         }
2468
2469         if (run_all) {
2470                 node = rb_first(&delayed_refs->root);
2471                 if (!node)
2472                         goto out;
2473                 count = (unsigned long)-1;
2474
2475                 while (node) {
2476                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2477                                        rb_node);
2478                         if (btrfs_delayed_ref_is_head(ref)) {
2479                                 struct btrfs_delayed_ref_head *head;
2480
2481                                 head = btrfs_delayed_node_to_head(ref);
2482                                 atomic_inc(&ref->refs);
2483
2484                                 spin_unlock(&delayed_refs->lock);
2485                                 /*
2486                                  * Mutex was contended, block until it's
2487                                  * released and try again
2488                                  */
2489                                 mutex_lock(&head->mutex);
2490                                 mutex_unlock(&head->mutex);
2491
2492                                 btrfs_put_delayed_ref(ref);
2493                                 cond_resched();
2494                                 goto again;
2495                         }
2496                         node = rb_next(node);
2497                 }
2498                 spin_unlock(&delayed_refs->lock);
2499                 schedule_timeout(1);
2500                 goto again;
2501         }
2502 out:
2503         spin_unlock(&delayed_refs->lock);
2504         return 0;
2505 }
2506
2507 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2508                                 struct btrfs_root *root,
2509                                 u64 bytenr, u64 num_bytes, u64 flags,
2510                                 int is_data)
2511 {
2512         struct btrfs_delayed_extent_op *extent_op;
2513         int ret;
2514
2515         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2516         if (!extent_op)
2517                 return -ENOMEM;
2518
2519         extent_op->flags_to_set = flags;
2520         extent_op->update_flags = 1;
2521         extent_op->update_key = 0;
2522         extent_op->is_data = is_data ? 1 : 0;
2523
2524         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2525                                           num_bytes, extent_op);
2526         if (ret)
2527                 kfree(extent_op);
2528         return ret;
2529 }
2530
2531 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2532                                       struct btrfs_root *root,
2533                                       struct btrfs_path *path,
2534                                       u64 objectid, u64 offset, u64 bytenr)
2535 {
2536         struct btrfs_delayed_ref_head *head;
2537         struct btrfs_delayed_ref_node *ref;
2538         struct btrfs_delayed_data_ref *data_ref;
2539         struct btrfs_delayed_ref_root *delayed_refs;
2540         struct rb_node *node;
2541         int ret = 0;
2542
2543         ret = -ENOENT;
2544         delayed_refs = &trans->transaction->delayed_refs;
2545         spin_lock(&delayed_refs->lock);
2546         head = btrfs_find_delayed_ref_head(trans, bytenr);
2547         if (!head)
2548                 goto out;
2549
2550         if (!mutex_trylock(&head->mutex)) {
2551                 atomic_inc(&head->node.refs);
2552                 spin_unlock(&delayed_refs->lock);
2553
2554                 btrfs_release_path(path);
2555
2556                 /*
2557                  * Mutex was contended, block until it's released and let
2558                  * caller try again
2559                  */
2560                 mutex_lock(&head->mutex);
2561                 mutex_unlock(&head->mutex);
2562                 btrfs_put_delayed_ref(&head->node);
2563                 return -EAGAIN;
2564         }
2565
2566         node = rb_prev(&head->node.rb_node);
2567         if (!node)
2568                 goto out_unlock;
2569
2570         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2571
2572         if (ref->bytenr != bytenr)
2573                 goto out_unlock;
2574
2575         ret = 1;
2576         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2577                 goto out_unlock;
2578
2579         data_ref = btrfs_delayed_node_to_data_ref(ref);
2580
2581         node = rb_prev(node);
2582         if (node) {
2583                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2584                 if (ref->bytenr == bytenr)
2585                         goto out_unlock;
2586         }
2587
2588         if (data_ref->root != root->root_key.objectid ||
2589             data_ref->objectid != objectid || data_ref->offset != offset)
2590                 goto out_unlock;
2591
2592         ret = 0;
2593 out_unlock:
2594         mutex_unlock(&head->mutex);
2595 out:
2596         spin_unlock(&delayed_refs->lock);
2597         return ret;
2598 }
2599
2600 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2601                                         struct btrfs_root *root,
2602                                         struct btrfs_path *path,
2603                                         u64 objectid, u64 offset, u64 bytenr)
2604 {
2605         struct btrfs_root *extent_root = root->fs_info->extent_root;
2606         struct extent_buffer *leaf;
2607         struct btrfs_extent_data_ref *ref;
2608         struct btrfs_extent_inline_ref *iref;
2609         struct btrfs_extent_item *ei;
2610         struct btrfs_key key;
2611         u32 item_size;
2612         int ret;
2613
2614         key.objectid = bytenr;
2615         key.offset = (u64)-1;
2616         key.type = BTRFS_EXTENT_ITEM_KEY;
2617
2618         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2619         if (ret < 0)
2620                 goto out;
2621         BUG_ON(ret == 0); /* Corruption */
2622
2623         ret = -ENOENT;
2624         if (path->slots[0] == 0)
2625                 goto out;
2626
2627         path->slots[0]--;
2628         leaf = path->nodes[0];
2629         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2630
2631         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2632                 goto out;
2633
2634         ret = 1;
2635         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2637         if (item_size < sizeof(*ei)) {
2638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2639                 goto out;
2640         }
2641 #endif
2642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2643
2644         if (item_size != sizeof(*ei) +
2645             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2646                 goto out;
2647
2648         if (btrfs_extent_generation(leaf, ei) <=
2649             btrfs_root_last_snapshot(&root->root_item))
2650                 goto out;
2651
2652         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2653         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2654             BTRFS_EXTENT_DATA_REF_KEY)
2655                 goto out;
2656
2657         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2658         if (btrfs_extent_refs(leaf, ei) !=
2659             btrfs_extent_data_ref_count(leaf, ref) ||
2660             btrfs_extent_data_ref_root(leaf, ref) !=
2661             root->root_key.objectid ||
2662             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2663             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2664                 goto out;
2665
2666         ret = 0;
2667 out:
2668         return ret;
2669 }
2670
2671 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2672                           struct btrfs_root *root,
2673                           u64 objectid, u64 offset, u64 bytenr)
2674 {
2675         struct btrfs_path *path;
2676         int ret;
2677         int ret2;
2678
2679         path = btrfs_alloc_path();
2680         if (!path)
2681                 return -ENOENT;
2682
2683         do {
2684                 ret = check_committed_ref(trans, root, path, objectid,
2685                                           offset, bytenr);
2686                 if (ret && ret != -ENOENT)
2687                         goto out;
2688
2689                 ret2 = check_delayed_ref(trans, root, path, objectid,
2690                                          offset, bytenr);
2691         } while (ret2 == -EAGAIN);
2692
2693         if (ret2 && ret2 != -ENOENT) {
2694                 ret = ret2;
2695                 goto out;
2696         }
2697
2698         if (ret != -ENOENT || ret2 != -ENOENT)
2699                 ret = 0;
2700 out:
2701         btrfs_free_path(path);
2702         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2703                 WARN_ON(ret > 0);
2704         return ret;
2705 }
2706
2707 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2708                            struct btrfs_root *root,
2709                            struct extent_buffer *buf,
2710                            int full_backref, int inc, int for_cow)
2711 {
2712         u64 bytenr;
2713         u64 num_bytes;
2714         u64 parent;
2715         u64 ref_root;
2716         u32 nritems;
2717         struct btrfs_key key;
2718         struct btrfs_file_extent_item *fi;
2719         int i;
2720         int level;
2721         int ret = 0;
2722         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2723                             u64, u64, u64, u64, u64, u64, int);
2724
2725         ref_root = btrfs_header_owner(buf);
2726         nritems = btrfs_header_nritems(buf);
2727         level = btrfs_header_level(buf);
2728
2729         if (!root->ref_cows && level == 0)
2730                 return 0;
2731
2732         if (inc)
2733                 process_func = btrfs_inc_extent_ref;
2734         else
2735                 process_func = btrfs_free_extent;
2736
2737         if (full_backref)
2738                 parent = buf->start;
2739         else
2740                 parent = 0;
2741
2742         for (i = 0; i < nritems; i++) {
2743                 if (level == 0) {
2744                         btrfs_item_key_to_cpu(buf, &key, i);
2745                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2746                                 continue;
2747                         fi = btrfs_item_ptr(buf, i,
2748                                             struct btrfs_file_extent_item);
2749                         if (btrfs_file_extent_type(buf, fi) ==
2750                             BTRFS_FILE_EXTENT_INLINE)
2751                                 continue;
2752                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2753                         if (bytenr == 0)
2754                                 continue;
2755
2756                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2757                         key.offset -= btrfs_file_extent_offset(buf, fi);
2758                         ret = process_func(trans, root, bytenr, num_bytes,
2759                                            parent, ref_root, key.objectid,
2760                                            key.offset, for_cow);
2761                         if (ret)
2762                                 goto fail;
2763                 } else {
2764                         bytenr = btrfs_node_blockptr(buf, i);
2765                         num_bytes = btrfs_level_size(root, level - 1);
2766                         ret = process_func(trans, root, bytenr, num_bytes,
2767                                            parent, ref_root, level - 1, 0,
2768                                            for_cow);
2769                         if (ret)
2770                                 goto fail;
2771                 }
2772         }
2773         return 0;
2774 fail:
2775         return ret;
2776 }
2777
2778 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2779                   struct extent_buffer *buf, int full_backref, int for_cow)
2780 {
2781         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2782 }
2783
2784 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2785                   struct extent_buffer *buf, int full_backref, int for_cow)
2786 {
2787         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2788 }
2789
2790 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2791                                  struct btrfs_root *root,
2792                                  struct btrfs_path *path,
2793                                  struct btrfs_block_group_cache *cache)
2794 {
2795         int ret;
2796         struct btrfs_root *extent_root = root->fs_info->extent_root;
2797         unsigned long bi;
2798         struct extent_buffer *leaf;
2799
2800         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2801         if (ret < 0)
2802                 goto fail;
2803         BUG_ON(ret); /* Corruption */
2804
2805         leaf = path->nodes[0];
2806         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2807         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2808         btrfs_mark_buffer_dirty(leaf);
2809         btrfs_release_path(path);
2810 fail:
2811         if (ret) {
2812                 btrfs_abort_transaction(trans, root, ret);
2813                 return ret;
2814         }
2815         return 0;
2816
2817 }
2818
2819 static struct btrfs_block_group_cache *
2820 next_block_group(struct btrfs_root *root,
2821                  struct btrfs_block_group_cache *cache)
2822 {
2823         struct rb_node *node;
2824         spin_lock(&root->fs_info->block_group_cache_lock);
2825         node = rb_next(&cache->cache_node);
2826         btrfs_put_block_group(cache);
2827         if (node) {
2828                 cache = rb_entry(node, struct btrfs_block_group_cache,
2829                                  cache_node);
2830                 btrfs_get_block_group(cache);
2831         } else
2832                 cache = NULL;
2833         spin_unlock(&root->fs_info->block_group_cache_lock);
2834         return cache;
2835 }
2836
2837 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2838                             struct btrfs_trans_handle *trans,
2839                             struct btrfs_path *path)
2840 {
2841         struct btrfs_root *root = block_group->fs_info->tree_root;
2842         struct inode *inode = NULL;
2843         u64 alloc_hint = 0;
2844         int dcs = BTRFS_DC_ERROR;
2845         int num_pages = 0;
2846         int retries = 0;
2847         int ret = 0;
2848
2849         /*
2850          * If this block group is smaller than 100 megs don't bother caching the
2851          * block group.
2852          */
2853         if (block_group->key.offset < (100 * 1024 * 1024)) {
2854                 spin_lock(&block_group->lock);
2855                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2856                 spin_unlock(&block_group->lock);
2857                 return 0;
2858         }
2859
2860 again:
2861         inode = lookup_free_space_inode(root, block_group, path);
2862         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2863                 ret = PTR_ERR(inode);
2864                 btrfs_release_path(path);
2865                 goto out;
2866         }
2867
2868         if (IS_ERR(inode)) {
2869                 BUG_ON(retries);
2870                 retries++;
2871
2872                 if (block_group->ro)
2873                         goto out_free;
2874
2875                 ret = create_free_space_inode(root, trans, block_group, path);
2876                 if (ret)
2877                         goto out_free;
2878                 goto again;
2879         }
2880
2881         /* We've already setup this transaction, go ahead and exit */
2882         if (block_group->cache_generation == trans->transid &&
2883             i_size_read(inode)) {
2884                 dcs = BTRFS_DC_SETUP;
2885                 goto out_put;
2886         }
2887
2888         /*
2889          * We want to set the generation to 0, that way if anything goes wrong
2890          * from here on out we know not to trust this cache when we load up next
2891          * time.
2892          */
2893         BTRFS_I(inode)->generation = 0;
2894         ret = btrfs_update_inode(trans, root, inode);
2895         WARN_ON(ret);
2896
2897         if (i_size_read(inode) > 0) {
2898                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2899                                                       inode);
2900                 if (ret)
2901                         goto out_put;
2902         }
2903
2904         spin_lock(&block_group->lock);
2905         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2906                 /* We're not cached, don't bother trying to write stuff out */
2907                 dcs = BTRFS_DC_WRITTEN;
2908                 spin_unlock(&block_group->lock);
2909                 goto out_put;
2910         }
2911         spin_unlock(&block_group->lock);
2912
2913         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2914         if (!num_pages)
2915                 num_pages = 1;
2916
2917         /*
2918          * Just to make absolutely sure we have enough space, we're going to
2919          * preallocate 12 pages worth of space for each block group.  In
2920          * practice we ought to use at most 8, but we need extra space so we can
2921          * add our header and have a terminator between the extents and the
2922          * bitmaps.
2923          */
2924         num_pages *= 16;
2925         num_pages *= PAGE_CACHE_SIZE;
2926
2927         ret = btrfs_check_data_free_space(inode, num_pages);
2928         if (ret)
2929                 goto out_put;
2930
2931         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2932                                               num_pages, num_pages,
2933                                               &alloc_hint);
2934         if (!ret)
2935                 dcs = BTRFS_DC_SETUP;
2936         btrfs_free_reserved_data_space(inode, num_pages);
2937
2938 out_put:
2939         iput(inode);
2940 out_free:
2941         btrfs_release_path(path);
2942 out:
2943         spin_lock(&block_group->lock);
2944         if (!ret && dcs == BTRFS_DC_SETUP)
2945                 block_group->cache_generation = trans->transid;
2946         block_group->disk_cache_state = dcs;
2947         spin_unlock(&block_group->lock);
2948
2949         return ret;
2950 }
2951
2952 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2953                                    struct btrfs_root *root)
2954 {
2955         struct btrfs_block_group_cache *cache;
2956         int err = 0;
2957         struct btrfs_path *path;
2958         u64 last = 0;
2959
2960         path = btrfs_alloc_path();
2961         if (!path)
2962                 return -ENOMEM;
2963
2964 again:
2965         while (1) {
2966                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2967                 while (cache) {
2968                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2969                                 break;
2970                         cache = next_block_group(root, cache);
2971                 }
2972                 if (!cache) {
2973                         if (last == 0)
2974                                 break;
2975                         last = 0;
2976                         continue;
2977                 }
2978                 err = cache_save_setup(cache, trans, path);
2979                 last = cache->key.objectid + cache->key.offset;
2980                 btrfs_put_block_group(cache);
2981         }
2982
2983         while (1) {
2984                 if (last == 0) {
2985                         err = btrfs_run_delayed_refs(trans, root,
2986                                                      (unsigned long)-1);
2987                         if (err) /* File system offline */
2988                                 goto out;
2989                 }
2990
2991                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2992                 while (cache) {
2993                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2994                                 btrfs_put_block_group(cache);
2995                                 goto again;
2996                         }
2997
2998                         if (cache->dirty)
2999                                 break;
3000                         cache = next_block_group(root, cache);
3001                 }
3002                 if (!cache) {
3003                         if (last == 0)
3004                                 break;
3005                         last = 0;
3006                         continue;
3007                 }
3008
3009                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3010                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3011                 cache->dirty = 0;
3012                 last = cache->key.objectid + cache->key.offset;
3013
3014                 err = write_one_cache_group(trans, root, path, cache);
3015                 if (err) /* File system offline */
3016                         goto out;
3017
3018                 btrfs_put_block_group(cache);
3019         }
3020
3021         while (1) {
3022                 /*
3023                  * I don't think this is needed since we're just marking our
3024                  * preallocated extent as written, but just in case it can't
3025                  * hurt.
3026                  */
3027                 if (last == 0) {
3028                         err = btrfs_run_delayed_refs(trans, root,
3029                                                      (unsigned long)-1);
3030                         if (err) /* File system offline */
3031                                 goto out;
3032                 }
3033
3034                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3035                 while (cache) {
3036                         /*
3037                          * Really this shouldn't happen, but it could if we
3038                          * couldn't write the entire preallocated extent and
3039                          * splitting the extent resulted in a new block.
3040                          */
3041                         if (cache->dirty) {
3042                                 btrfs_put_block_group(cache);
3043                                 goto again;
3044                         }
3045                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3046                                 break;
3047                         cache = next_block_group(root, cache);
3048                 }
3049                 if (!cache) {
3050                         if (last == 0)
3051                                 break;
3052                         last = 0;
3053                         continue;
3054                 }
3055
3056                 err = btrfs_write_out_cache(root, trans, cache, path);
3057
3058                 /*
3059                  * If we didn't have an error then the cache state is still
3060                  * NEED_WRITE, so we can set it to WRITTEN.
3061                  */
3062                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3063                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3064                 last = cache->key.objectid + cache->key.offset;
3065                 btrfs_put_block_group(cache);
3066         }
3067 out:
3068
3069         btrfs_free_path(path);
3070         return err;
3071 }
3072
3073 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3074 {
3075         struct btrfs_block_group_cache *block_group;
3076         int readonly = 0;
3077
3078         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3079         if (!block_group || block_group->ro)
3080                 readonly = 1;
3081         if (block_group)
3082                 btrfs_put_block_group(block_group);
3083         return readonly;
3084 }
3085
3086 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3087                              u64 total_bytes, u64 bytes_used,
3088                              struct btrfs_space_info **space_info)
3089 {
3090         struct btrfs_space_info *found;
3091         int i;
3092         int factor;
3093
3094         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3095                      BTRFS_BLOCK_GROUP_RAID10))
3096                 factor = 2;
3097         else
3098                 factor = 1;
3099
3100         found = __find_space_info(info, flags);
3101         if (found) {
3102                 spin_lock(&found->lock);
3103                 found->total_bytes += total_bytes;
3104                 found->disk_total += total_bytes * factor;
3105                 found->bytes_used += bytes_used;
3106                 found->disk_used += bytes_used * factor;
3107                 found->full = 0;
3108                 spin_unlock(&found->lock);
3109                 *space_info = found;
3110                 return 0;
3111         }
3112         found = kzalloc(sizeof(*found), GFP_NOFS);
3113         if (!found)
3114                 return -ENOMEM;
3115
3116         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3117                 INIT_LIST_HEAD(&found->block_groups[i]);
3118         init_rwsem(&found->groups_sem);
3119         spin_lock_init(&found->lock);
3120         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3121         found->total_bytes = total_bytes;
3122         found->disk_total = total_bytes * factor;
3123         found->bytes_used = bytes_used;
3124         found->disk_used = bytes_used * factor;
3125         found->bytes_pinned = 0;
3126         found->bytes_reserved = 0;
3127         found->bytes_readonly = 0;
3128         found->bytes_may_use = 0;
3129         found->full = 0;
3130         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3131         found->chunk_alloc = 0;
3132         found->flush = 0;
3133         init_waitqueue_head(&found->wait);
3134         *space_info = found;
3135         list_add_rcu(&found->list, &info->space_info);
3136         return 0;
3137 }
3138
3139 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3140 {
3141         u64 extra_flags = chunk_to_extended(flags) &
3142                                 BTRFS_EXTENDED_PROFILE_MASK;
3143
3144         if (flags & BTRFS_BLOCK_GROUP_DATA)
3145                 fs_info->avail_data_alloc_bits |= extra_flags;
3146         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3147                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3148         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3149                 fs_info->avail_system_alloc_bits |= extra_flags;
3150 }
3151
3152 /*
3153  * returns target flags in extended format or 0 if restripe for this
3154  * chunk_type is not in progress
3155  *
3156  * should be called with either volume_mutex or balance_lock held
3157  */
3158 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3159 {
3160         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3161         u64 target = 0;
3162
3163         if (!bctl)
3164                 return 0;
3165
3166         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3167             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3168                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3169         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3170                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3171                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3172         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3173                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3174                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3175         }
3176
3177         return target;
3178 }
3179
3180 /*
3181  * @flags: available profiles in extended format (see ctree.h)
3182  *
3183  * Returns reduced profile in chunk format.  If profile changing is in
3184  * progress (either running or paused) picks the target profile (if it's
3185  * already available), otherwise falls back to plain reducing.
3186  */
3187 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3188 {
3189         /*
3190          * we add in the count of missing devices because we want
3191          * to make sure that any RAID levels on a degraded FS
3192          * continue to be honored.
3193          */
3194         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3195                 root->fs_info->fs_devices->missing_devices;
3196         u64 target;
3197
3198         /*
3199          * see if restripe for this chunk_type is in progress, if so
3200          * try to reduce to the target profile
3201          */
3202         spin_lock(&root->fs_info->balance_lock);
3203         target = get_restripe_target(root->fs_info, flags);
3204         if (target) {
3205                 /* pick target profile only if it's already available */
3206                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3207                         spin_unlock(&root->fs_info->balance_lock);
3208                         return extended_to_chunk(target);
3209                 }
3210         }
3211         spin_unlock(&root->fs_info->balance_lock);
3212
3213         if (num_devices == 1)
3214                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3215         if (num_devices < 4)
3216                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3217
3218         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3219             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3220                       BTRFS_BLOCK_GROUP_RAID10))) {
3221                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3222         }
3223
3224         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3225             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3226                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3227         }
3228
3229         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3230             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3231              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3232              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3233                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3234         }
3235
3236         return extended_to_chunk(flags);
3237 }
3238
3239 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3240 {
3241         if (flags & BTRFS_BLOCK_GROUP_DATA)
3242                 flags |= root->fs_info->avail_data_alloc_bits;
3243         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3244                 flags |= root->fs_info->avail_system_alloc_bits;
3245         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3246                 flags |= root->fs_info->avail_metadata_alloc_bits;
3247
3248         return btrfs_reduce_alloc_profile(root, flags);
3249 }
3250
3251 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3252 {
3253         u64 flags;
3254
3255         if (data)
3256                 flags = BTRFS_BLOCK_GROUP_DATA;
3257         else if (root == root->fs_info->chunk_root)
3258                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3259         else
3260                 flags = BTRFS_BLOCK_GROUP_METADATA;
3261
3262         return get_alloc_profile(root, flags);
3263 }
3264
3265 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3266 {
3267         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3268                                                        BTRFS_BLOCK_GROUP_DATA);
3269 }
3270
3271 /*
3272  * This will check the space that the inode allocates from to make sure we have
3273  * enough space for bytes.
3274  */
3275 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3276 {
3277         struct btrfs_space_info *data_sinfo;
3278         struct btrfs_root *root = BTRFS_I(inode)->root;
3279         u64 used;
3280         int ret = 0, committed = 0, alloc_chunk = 1;
3281
3282         /* make sure bytes are sectorsize aligned */
3283         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3284
3285         if (root == root->fs_info->tree_root ||
3286             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3287                 alloc_chunk = 0;
3288                 committed = 1;
3289         }
3290
3291         data_sinfo = BTRFS_I(inode)->space_info;
3292         if (!data_sinfo)
3293                 goto alloc;
3294
3295 again:
3296         /* make sure we have enough space to handle the data first */
3297         spin_lock(&data_sinfo->lock);
3298         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3299                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3300                 data_sinfo->bytes_may_use;
3301
3302         if (used + bytes > data_sinfo->total_bytes) {
3303                 struct btrfs_trans_handle *trans;
3304
3305                 /*
3306                  * if we don't have enough free bytes in this space then we need
3307                  * to alloc a new chunk.
3308                  */
3309                 if (!data_sinfo->full && alloc_chunk) {
3310                         u64 alloc_target;
3311
3312                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3313                         spin_unlock(&data_sinfo->lock);
3314 alloc:
3315                         alloc_target = btrfs_get_alloc_profile(root, 1);
3316                         trans = btrfs_join_transaction(root);
3317                         if (IS_ERR(trans))
3318                                 return PTR_ERR(trans);
3319
3320                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3321                                              bytes + 2 * 1024 * 1024,
3322                                              alloc_target,
3323                                              CHUNK_ALLOC_NO_FORCE);
3324                         btrfs_end_transaction(trans, root);
3325                         if (ret < 0) {
3326                                 if (ret != -ENOSPC)
3327                                         return ret;
3328                                 else
3329                                         goto commit_trans;
3330                         }
3331
3332                         if (!data_sinfo) {
3333                                 btrfs_set_inode_space_info(root, inode);
3334                                 data_sinfo = BTRFS_I(inode)->space_info;
3335                         }
3336                         goto again;
3337                 }
3338
3339                 /*
3340                  * If we have less pinned bytes than we want to allocate then
3341                  * don't bother committing the transaction, it won't help us.
3342                  */
3343                 if (data_sinfo->bytes_pinned < bytes)
3344                         committed = 1;
3345                 spin_unlock(&data_sinfo->lock);
3346
3347                 /* commit the current transaction and try again */
3348 commit_trans:
3349                 if (!committed &&
3350                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3351                         committed = 1;
3352                         trans = btrfs_join_transaction(root);
3353                         if (IS_ERR(trans))
3354                                 return PTR_ERR(trans);
3355                         ret = btrfs_commit_transaction(trans, root);
3356                         if (ret)
3357                                 return ret;
3358                         goto again;
3359                 }
3360
3361                 return -ENOSPC;
3362         }
3363         data_sinfo->bytes_may_use += bytes;
3364         trace_btrfs_space_reservation(root->fs_info, "space_info",
3365                                       data_sinfo->flags, bytes, 1);
3366         spin_unlock(&data_sinfo->lock);
3367
3368         return 0;
3369 }
3370
3371 /*
3372  * Called if we need to clear a data reservation for this inode.
3373  */
3374 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3375 {
3376         struct btrfs_root *root = BTRFS_I(inode)->root;
3377         struct btrfs_space_info *data_sinfo;
3378
3379         /* make sure bytes are sectorsize aligned */
3380         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3381
3382         data_sinfo = BTRFS_I(inode)->space_info;
3383         spin_lock(&data_sinfo->lock);
3384         data_sinfo->bytes_may_use -= bytes;
3385         trace_btrfs_space_reservation(root->fs_info, "space_info",
3386                                       data_sinfo->flags, bytes, 0);
3387         spin_unlock(&data_sinfo->lock);
3388 }
3389
3390 static void force_metadata_allocation(struct btrfs_fs_info *info)
3391 {
3392         struct list_head *head = &info->space_info;
3393         struct btrfs_space_info *found;
3394
3395         rcu_read_lock();
3396         list_for_each_entry_rcu(found, head, list) {
3397                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3398                         found->force_alloc = CHUNK_ALLOC_FORCE;
3399         }
3400         rcu_read_unlock();
3401 }
3402
3403 static int should_alloc_chunk(struct btrfs_root *root,
3404                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3405                               int force)
3406 {
3407         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3408         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3409         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3410         u64 thresh;
3411
3412         if (force == CHUNK_ALLOC_FORCE)
3413                 return 1;
3414
3415         /*
3416          * We need to take into account the global rsv because for all intents
3417          * and purposes it's used space.  Don't worry about locking the
3418          * global_rsv, it doesn't change except when the transaction commits.
3419          */
3420         num_allocated += global_rsv->size;
3421
3422         /*
3423          * in limited mode, we want to have some free space up to
3424          * about 1% of the FS size.
3425          */
3426         if (force == CHUNK_ALLOC_LIMITED) {
3427                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3428                 thresh = max_t(u64, 64 * 1024 * 1024,
3429                                div_factor_fine(thresh, 1));
3430
3431                 if (num_bytes - num_allocated < thresh)
3432                         return 1;
3433         }
3434         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3435
3436         /* 256MB or 2% of the FS */
3437         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3438         /* system chunks need a much small threshold */
3439         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3440                 thresh = 32 * 1024 * 1024;
3441
3442         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3443                 return 0;
3444         return 1;
3445 }
3446
3447 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3448 {
3449         u64 num_dev;
3450
3451         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3452             type & BTRFS_BLOCK_GROUP_RAID0)
3453                 num_dev = root->fs_info->fs_devices->rw_devices;
3454         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3455                 num_dev = 2;
3456         else
3457                 num_dev = 1;    /* DUP or single */
3458
3459         /* metadata for updaing devices and chunk tree */
3460         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3461 }
3462
3463 static void check_system_chunk(struct btrfs_trans_handle *trans,
3464                                struct btrfs_root *root, u64 type)
3465 {
3466         struct btrfs_space_info *info;
3467         u64 left;
3468         u64 thresh;
3469
3470         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3471         spin_lock(&info->lock);
3472         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3473                 info->bytes_reserved - info->bytes_readonly;
3474         spin_unlock(&info->lock);
3475
3476         thresh = get_system_chunk_thresh(root, type);
3477         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3478                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3479                        left, thresh, type);
3480                 dump_space_info(info, 0, 0);
3481         }
3482
3483         if (left < thresh) {
3484                 u64 flags;
3485
3486                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3487                 btrfs_alloc_chunk(trans, root, flags);
3488         }
3489 }
3490
3491 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3492                           struct btrfs_root *extent_root, u64 alloc_bytes,
3493                           u64 flags, int force)
3494 {
3495         struct btrfs_space_info *space_info;
3496         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3497         int wait_for_alloc = 0;
3498         int ret = 0;
3499
3500         space_info = __find_space_info(extent_root->fs_info, flags);
3501         if (!space_info) {
3502                 ret = update_space_info(extent_root->fs_info, flags,
3503                                         0, 0, &space_info);
3504                 BUG_ON(ret); /* -ENOMEM */
3505         }
3506         BUG_ON(!space_info); /* Logic error */
3507
3508 again:
3509         spin_lock(&space_info->lock);
3510         if (force < space_info->force_alloc)
3511                 force = space_info->force_alloc;
3512         if (space_info->full) {
3513                 spin_unlock(&space_info->lock);
3514                 return 0;
3515         }
3516
3517         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3518                 spin_unlock(&space_info->lock);
3519                 return 0;
3520         } else if (space_info->chunk_alloc) {
3521                 wait_for_alloc = 1;
3522         } else {
3523                 space_info->chunk_alloc = 1;
3524         }
3525
3526         spin_unlock(&space_info->lock);
3527
3528         mutex_lock(&fs_info->chunk_mutex);
3529
3530         /*
3531          * The chunk_mutex is held throughout the entirety of a chunk
3532          * allocation, so once we've acquired the chunk_mutex we know that the
3533          * other guy is done and we need to recheck and see if we should
3534          * allocate.
3535          */
3536         if (wait_for_alloc) {
3537                 mutex_unlock(&fs_info->chunk_mutex);
3538                 wait_for_alloc = 0;
3539                 goto again;
3540         }
3541
3542         /*
3543          * If we have mixed data/metadata chunks we want to make sure we keep
3544          * allocating mixed chunks instead of individual chunks.
3545          */
3546         if (btrfs_mixed_space_info(space_info))
3547                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3548
3549         /*
3550          * if we're doing a data chunk, go ahead and make sure that
3551          * we keep a reasonable number of metadata chunks allocated in the
3552          * FS as well.
3553          */
3554         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3555                 fs_info->data_chunk_allocations++;
3556                 if (!(fs_info->data_chunk_allocations %
3557                       fs_info->metadata_ratio))
3558                         force_metadata_allocation(fs_info);
3559         }
3560
3561         /*
3562          * Check if we have enough space in SYSTEM chunk because we may need
3563          * to update devices.
3564          */
3565         check_system_chunk(trans, extent_root, flags);
3566
3567         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3568         if (ret < 0 && ret != -ENOSPC)
3569                 goto out;
3570
3571         spin_lock(&space_info->lock);
3572         if (ret)
3573                 space_info->full = 1;
3574         else
3575                 ret = 1;
3576
3577         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3578         space_info->chunk_alloc = 0;
3579         spin_unlock(&space_info->lock);
3580 out:
3581         mutex_unlock(&fs_info->chunk_mutex);
3582         return ret;
3583 }
3584
3585 /*
3586  * shrink metadata reservation for delalloc
3587  */
3588 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3589                            bool wait_ordered)
3590 {
3591         struct btrfs_block_rsv *block_rsv;
3592         struct btrfs_space_info *space_info;
3593         struct btrfs_trans_handle *trans;
3594         u64 reserved;
3595         u64 max_reclaim;
3596         u64 reclaimed = 0;
3597         long time_left;
3598         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3599         int loops = 0;
3600         unsigned long progress;
3601
3602         trans = (struct btrfs_trans_handle *)current->journal_info;
3603         block_rsv = &root->fs_info->delalloc_block_rsv;
3604         space_info = block_rsv->space_info;
3605
3606         smp_mb();
3607         reserved = space_info->bytes_may_use;
3608         progress = space_info->reservation_progress;
3609
3610         if (reserved == 0)
3611                 return 0;
3612
3613         smp_mb();
3614         if (root->fs_info->delalloc_bytes == 0) {
3615                 if (trans)
3616                         return 0;
3617                 btrfs_wait_ordered_extents(root, 0, 0);
3618                 return 0;
3619         }
3620
3621         max_reclaim = min(reserved, to_reclaim);
3622         nr_pages = max_t(unsigned long, nr_pages,
3623                          max_reclaim >> PAGE_CACHE_SHIFT);
3624         while (loops < 1024) {
3625                 /* have the flusher threads jump in and do some IO */
3626                 smp_mb();
3627                 nr_pages = min_t(unsigned long, nr_pages,
3628                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3629                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3630                                                 WB_REASON_FS_FREE_SPACE);
3631
3632                 spin_lock(&space_info->lock);
3633                 if (reserved > space_info->bytes_may_use)
3634                         reclaimed += reserved - space_info->bytes_may_use;
3635                 reserved = space_info->bytes_may_use;
3636                 spin_unlock(&space_info->lock);
3637
3638                 loops++;
3639
3640                 if (reserved == 0 || reclaimed >= max_reclaim)
3641                         break;
3642
3643                 if (trans && trans->transaction->blocked)
3644                         return -EAGAIN;
3645
3646                 if (wait_ordered && !trans) {
3647                         btrfs_wait_ordered_extents(root, 0, 0);
3648                 } else {
3649                         time_left = schedule_timeout_interruptible(1);
3650
3651                         /* We were interrupted, exit */
3652                         if (time_left)
3653                                 break;
3654                 }
3655
3656                 /* we've kicked the IO a few times, if anything has been freed,
3657                  * exit.  There is no sense in looping here for a long time
3658                  * when we really need to commit the transaction, or there are
3659                  * just too many writers without enough free space
3660                  */
3661
3662                 if (loops > 3) {
3663                         smp_mb();
3664                         if (progress != space_info->reservation_progress)
3665                                 break;
3666                 }
3667
3668         }
3669
3670         return reclaimed >= to_reclaim;
3671 }
3672
3673 /**
3674  * maybe_commit_transaction - possibly commit the transaction if its ok to
3675  * @root - the root we're allocating for
3676  * @bytes - the number of bytes we want to reserve
3677  * @force - force the commit
3678  *
3679  * This will check to make sure that committing the transaction will actually
3680  * get us somewhere and then commit the transaction if it does.  Otherwise it
3681  * will return -ENOSPC.
3682  */
3683 static int may_commit_transaction(struct btrfs_root *root,
3684                                   struct btrfs_space_info *space_info,
3685                                   u64 bytes, int force)
3686 {
3687         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3688         struct btrfs_trans_handle *trans;
3689
3690         trans = (struct btrfs_trans_handle *)current->journal_info;
3691         if (trans)
3692                 return -EAGAIN;
3693
3694         if (force)
3695                 goto commit;
3696
3697         /* See if there is enough pinned space to make this reservation */
3698         spin_lock(&space_info->lock);
3699         if (space_info->bytes_pinned >= bytes) {
3700                 spin_unlock(&space_info->lock);
3701                 goto commit;
3702         }
3703         spin_unlock(&space_info->lock);
3704
3705         /*
3706          * See if there is some space in the delayed insertion reservation for
3707          * this reservation.
3708          */
3709         if (space_info != delayed_rsv->space_info)
3710                 return -ENOSPC;
3711
3712         spin_lock(&space_info->lock);
3713         spin_lock(&delayed_rsv->lock);
3714         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3715                 spin_unlock(&delayed_rsv->lock);
3716                 spin_unlock(&space_info->lock);
3717                 return -ENOSPC;
3718         }
3719         spin_unlock(&delayed_rsv->lock);
3720         spin_unlock(&space_info->lock);
3721
3722 commit:
3723         trans = btrfs_join_transaction(root);
3724         if (IS_ERR(trans))
3725                 return -ENOSPC;
3726
3727         return btrfs_commit_transaction(trans, root);
3728 }
3729
3730 /**
3731  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3732  * @root - the root we're allocating for
3733  * @block_rsv - the block_rsv we're allocating for
3734  * @orig_bytes - the number of bytes we want
3735  * @flush - wether or not we can flush to make our reservation
3736  *
3737  * This will reserve orgi_bytes number of bytes from the space info associated
3738  * with the block_rsv.  If there is not enough space it will make an attempt to
3739  * flush out space to make room.  It will do this by flushing delalloc if
3740  * possible or committing the transaction.  If flush is 0 then no attempts to
3741  * regain reservations will be made and this will fail if there is not enough
3742  * space already.
3743  */
3744 static int reserve_metadata_bytes(struct btrfs_root *root,
3745                                   struct btrfs_block_rsv *block_rsv,
3746                                   u64 orig_bytes, int flush)
3747 {
3748         struct btrfs_space_info *space_info = block_rsv->space_info;
3749         u64 used;
3750         u64 num_bytes = orig_bytes;
3751         int retries = 0;
3752         int ret = 0;
3753         bool committed = false;
3754         bool flushing = false;
3755         bool wait_ordered = false;
3756
3757 again:
3758         ret = 0;
3759         spin_lock(&space_info->lock);
3760         /*
3761          * We only want to wait if somebody other than us is flushing and we are
3762          * actually alloed to flush.
3763          */
3764         while (flush && !flushing && space_info->flush) {
3765                 spin_unlock(&space_info->lock);
3766                 /*
3767                  * If we have a trans handle we can't wait because the flusher
3768                  * may have to commit the transaction, which would mean we would
3769                  * deadlock since we are waiting for the flusher to finish, but
3770                  * hold the current transaction open.
3771                  */
3772                 if (current->journal_info)
3773                         return -EAGAIN;
3774                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3775                 /* Must have been killed, return */
3776                 if (ret)
3777                         return -EINTR;
3778
3779                 spin_lock(&space_info->lock);
3780         }
3781
3782         ret = -ENOSPC;
3783         used = space_info->bytes_used + space_info->bytes_reserved +
3784                 space_info->bytes_pinned + space_info->bytes_readonly +
3785                 space_info->bytes_may_use;
3786
3787         /*
3788          * The idea here is that we've not already over-reserved the block group
3789          * then we can go ahead and save our reservation first and then start
3790          * flushing if we need to.  Otherwise if we've already overcommitted
3791          * lets start flushing stuff first and then come back and try to make
3792          * our reservation.
3793          */
3794         if (used <= space_info->total_bytes) {
3795                 if (used + orig_bytes <= space_info->total_bytes) {
3796                         space_info->bytes_may_use += orig_bytes;
3797                         trace_btrfs_space_reservation(root->fs_info,
3798                                 "space_info", space_info->flags, orig_bytes, 1);
3799                         ret = 0;
3800                 } else {
3801                         /*
3802                          * Ok set num_bytes to orig_bytes since we aren't
3803                          * overocmmitted, this way we only try and reclaim what
3804                          * we need.
3805                          */
3806                         num_bytes = orig_bytes;
3807                 }
3808         } else {
3809                 /*
3810                  * Ok we're over committed, set num_bytes to the overcommitted
3811                  * amount plus the amount of bytes that we need for this
3812                  * reservation.
3813                  */
3814                 wait_ordered = true;
3815                 num_bytes = used - space_info->total_bytes +
3816                         (orig_bytes * (retries + 1));
3817         }
3818
3819         if (ret) {
3820                 u64 profile = btrfs_get_alloc_profile(root, 0);
3821                 u64 avail;
3822
3823                 /*
3824                  * If we have a lot of space that's pinned, don't bother doing
3825                  * the overcommit dance yet and just commit the transaction.
3826                  */
3827                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3828                 do_div(avail, 10);
3829                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3830                         space_info->flush = 1;
3831                         flushing = true;
3832                         spin_unlock(&space_info->lock);
3833                         ret = may_commit_transaction(root, space_info,
3834                                                      orig_bytes, 1);
3835                         if (ret)
3836                                 goto out;
3837                         committed = true;
3838                         goto again;
3839                 }
3840
3841                 spin_lock(&root->fs_info->free_chunk_lock);
3842                 avail = root->fs_info->free_chunk_space;
3843
3844                 /*
3845                  * If we have dup, raid1 or raid10 then only half of the free
3846                  * space is actually useable.
3847                  */
3848                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3849                                BTRFS_BLOCK_GROUP_RAID1 |
3850                                BTRFS_BLOCK_GROUP_RAID10))
3851                         avail >>= 1;
3852
3853                 /*
3854                  * If we aren't flushing don't let us overcommit too much, say
3855                  * 1/8th of the space.  If we can flush, let it overcommit up to
3856                  * 1/2 of the space.
3857                  */
3858                 if (flush)
3859                         avail >>= 3;
3860                 else
3861                         avail >>= 1;
3862                  spin_unlock(&root->fs_info->free_chunk_lock);
3863
3864                 if (used + num_bytes < space_info->total_bytes + avail) {
3865                         space_info->bytes_may_use += orig_bytes;
3866                         trace_btrfs_space_reservation(root->fs_info,
3867                                 "space_info", space_info->flags, orig_bytes, 1);
3868                         ret = 0;
3869                 } else {
3870                         wait_ordered = true;
3871                 }
3872         }
3873
3874         /*
3875          * Couldn't make our reservation, save our place so while we're trying
3876          * to reclaim space we can actually use it instead of somebody else
3877          * stealing it from us.
3878          */
3879         if (ret && flush) {
3880                 flushing = true;
3881                 space_info->flush = 1;
3882         }
3883
3884         spin_unlock(&space_info->lock);
3885
3886         if (!ret || !flush)
3887                 goto out;
3888
3889         /*
3890          * We do synchronous shrinking since we don't actually unreserve
3891          * metadata until after the IO is completed.
3892          */
3893         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3894         if (ret < 0)
3895                 goto out;
3896
3897         ret = 0;
3898
3899         /*
3900          * So if we were overcommitted it's possible that somebody else flushed
3901          * out enough space and we simply didn't have enough space to reclaim,
3902          * so go back around and try again.
3903          */
3904         if (retries < 2) {
3905                 wait_ordered = true;
3906                 retries++;
3907                 goto again;
3908         }
3909
3910         ret = -ENOSPC;
3911         if (committed)
3912                 goto out;
3913
3914         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3915         if (!ret) {
3916                 committed = true;
3917                 goto again;
3918         }
3919
3920 out:
3921         if (flushing) {
3922                 spin_lock(&space_info->lock);
3923                 space_info->flush = 0;
3924                 wake_up_all(&space_info->wait);
3925                 spin_unlock(&space_info->lock);
3926         }
3927         return ret;
3928 }
3929
3930 static struct btrfs_block_rsv *get_block_rsv(
3931                                         const struct btrfs_trans_handle *trans,
3932                                         const struct btrfs_root *root)
3933 {
3934         struct btrfs_block_rsv *block_rsv = NULL;
3935
3936         if (root->ref_cows || root == root->fs_info->csum_root)
3937                 block_rsv = trans->block_rsv;
3938
3939         if (!block_rsv)
3940                 block_rsv = root->block_rsv;
3941
3942         if (!block_rsv)
3943                 block_rsv = &root->fs_info->empty_block_rsv;
3944
3945         return block_rsv;
3946 }
3947
3948 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3949                                u64 num_bytes)
3950 {
3951         int ret = -ENOSPC;
3952         spin_lock(&block_rsv->lock);
3953         if (block_rsv->reserved >= num_bytes) {
3954                 block_rsv->reserved -= num_bytes;
3955                 if (block_rsv->reserved < block_rsv->size)
3956                         block_rsv->full = 0;
3957                 ret = 0;
3958         }
3959         spin_unlock(&block_rsv->lock);
3960         return ret;
3961 }
3962
3963 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3964                                 u64 num_bytes, int update_size)
3965 {
3966         spin_lock(&block_rsv->lock);
3967         block_rsv->reserved += num_bytes;
3968         if (update_size)
3969                 block_rsv->size += num_bytes;
3970         else if (block_rsv->reserved >= block_rsv->size)
3971                 block_rsv->full = 1;
3972         spin_unlock(&block_rsv->lock);
3973 }
3974
3975 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3976                                     struct btrfs_block_rsv *block_rsv,
3977                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3978 {
3979         struct btrfs_space_info *space_info = block_rsv->space_info;
3980
3981         spin_lock(&block_rsv->lock);
3982         if (num_bytes == (u64)-1)
3983                 num_bytes = block_rsv->size;
3984         block_rsv->size -= num_bytes;
3985         if (block_rsv->reserved >= block_rsv->size) {
3986                 num_bytes = block_rsv->reserved - block_rsv->size;
3987                 block_rsv->reserved = block_rsv->size;
3988                 block_rsv->full = 1;
3989         } else {
3990                 num_bytes = 0;
3991         }
3992         spin_unlock(&block_rsv->lock);
3993
3994         if (num_bytes > 0) {
3995                 if (dest) {
3996                         spin_lock(&dest->lock);
3997                         if (!dest->full) {
3998                                 u64 bytes_to_add;
3999
4000                                 bytes_to_add = dest->size - dest->reserved;
4001                                 bytes_to_add = min(num_bytes, bytes_to_add);
4002                                 dest->reserved += bytes_to_add;
4003                                 if (dest->reserved >= dest->size)
4004                                         dest->full = 1;
4005                                 num_bytes -= bytes_to_add;
4006                         }
4007                         spin_unlock(&dest->lock);
4008                 }
4009                 if (num_bytes) {
4010                         spin_lock(&space_info->lock);
4011                         space_info->bytes_may_use -= num_bytes;
4012                         trace_btrfs_space_reservation(fs_info, "space_info",
4013                                         space_info->flags, num_bytes, 0);
4014                         space_info->reservation_progress++;
4015                         spin_unlock(&space_info->lock);
4016                 }
4017         }
4018 }
4019
4020 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4021                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4022 {
4023         int ret;
4024
4025         ret = block_rsv_use_bytes(src, num_bytes);
4026         if (ret)
4027                 return ret;
4028
4029         block_rsv_add_bytes(dst, num_bytes, 1);
4030         return 0;
4031 }
4032
4033 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4034 {
4035         memset(rsv, 0, sizeof(*rsv));
4036         spin_lock_init(&rsv->lock);
4037 }
4038
4039 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4040 {
4041         struct btrfs_block_rsv *block_rsv;
4042         struct btrfs_fs_info *fs_info = root->fs_info;
4043
4044         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4045         if (!block_rsv)
4046                 return NULL;
4047
4048         btrfs_init_block_rsv(block_rsv);
4049         block_rsv->space_info = __find_space_info(fs_info,
4050                                                   BTRFS_BLOCK_GROUP_METADATA);
4051         return block_rsv;
4052 }
4053
4054 void btrfs_free_block_rsv(struct btrfs_root *root,
4055                           struct btrfs_block_rsv *rsv)
4056 {
4057         btrfs_block_rsv_release(root, rsv, (u64)-1);
4058         kfree(rsv);
4059 }
4060
4061 static inline int __block_rsv_add(struct btrfs_root *root,
4062                                   struct btrfs_block_rsv *block_rsv,
4063                                   u64 num_bytes, int flush)
4064 {
4065         int ret;
4066
4067         if (num_bytes == 0)
4068                 return 0;
4069
4070         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4071         if (!ret) {
4072                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4073                 return 0;
4074         }
4075
4076         return ret;
4077 }
4078
4079 int btrfs_block_rsv_add(struct btrfs_root *root,
4080                         struct btrfs_block_rsv *block_rsv,
4081                         u64 num_bytes)
4082 {
4083         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4084 }
4085
4086 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4087                                 struct btrfs_block_rsv *block_rsv,
4088                                 u64 num_bytes)
4089 {
4090         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4091 }
4092
4093 int btrfs_block_rsv_check(struct btrfs_root *root,
4094                           struct btrfs_block_rsv *block_rsv, int min_factor)
4095 {
4096         u64 num_bytes = 0;
4097         int ret = -ENOSPC;
4098
4099         if (!block_rsv)
4100                 return 0;
4101
4102         spin_lock(&block_rsv->lock);
4103         num_bytes = div_factor(block_rsv->size, min_factor);
4104         if (block_rsv->reserved >= num_bytes)
4105                 ret = 0;
4106         spin_unlock(&block_rsv->lock);
4107
4108         return ret;
4109 }
4110
4111 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4112                                            struct btrfs_block_rsv *block_rsv,
4113                                            u64 min_reserved, int flush)
4114 {
4115         u64 num_bytes = 0;
4116         int ret = -ENOSPC;
4117
4118         if (!block_rsv)
4119                 return 0;
4120
4121         spin_lock(&block_rsv->lock);
4122         num_bytes = min_reserved;
4123         if (block_rsv->reserved >= num_bytes)
4124                 ret = 0;
4125         else
4126                 num_bytes -= block_rsv->reserved;
4127         spin_unlock(&block_rsv->lock);
4128
4129         if (!ret)
4130                 return 0;
4131
4132         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4133         if (!ret) {
4134                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4135                 return 0;
4136         }
4137
4138         return ret;
4139 }
4140
4141 int btrfs_block_rsv_refill(struct btrfs_root *root,
4142                            struct btrfs_block_rsv *block_rsv,
4143                            u64 min_reserved)
4144 {
4145         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4146 }
4147
4148 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4149                                    struct btrfs_block_rsv *block_rsv,
4150                                    u64 min_reserved)
4151 {
4152         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4153 }
4154
4155 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4156                             struct btrfs_block_rsv *dst_rsv,
4157                             u64 num_bytes)
4158 {
4159         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4160 }
4161
4162 void btrfs_block_rsv_release(struct btrfs_root *root,
4163                              struct btrfs_block_rsv *block_rsv,
4164                              u64 num_bytes)
4165 {
4166         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4167         if (global_rsv->full || global_rsv == block_rsv ||
4168             block_rsv->space_info != global_rsv->space_info)
4169                 global_rsv = NULL;
4170         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4171                                 num_bytes);
4172 }
4173
4174 /*
4175  * helper to calculate size of global block reservation.
4176  * the desired value is sum of space used by extent tree,
4177  * checksum tree and root tree
4178  */
4179 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4180 {
4181         struct btrfs_space_info *sinfo;
4182         u64 num_bytes;
4183         u64 meta_used;
4184         u64 data_used;
4185         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4186
4187         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4188         spin_lock(&sinfo->lock);
4189         data_used = sinfo->bytes_used;
4190         spin_unlock(&sinfo->lock);
4191
4192         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4193         spin_lock(&sinfo->lock);
4194         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4195                 data_used = 0;
4196         meta_used = sinfo->bytes_used;
4197         spin_unlock(&sinfo->lock);
4198
4199         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4200                     csum_size * 2;
4201         num_bytes += div64_u64(data_used + meta_used, 50);
4202
4203         if (num_bytes * 3 > meta_used)
4204                 num_bytes = div64_u64(meta_used, 3);
4205
4206         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4207 }
4208
4209 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4210 {
4211         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4212         struct btrfs_space_info *sinfo = block_rsv->space_info;
4213         u64 num_bytes;
4214
4215         num_bytes = calc_global_metadata_size(fs_info);
4216
4217         spin_lock(&sinfo->lock);
4218         spin_lock(&block_rsv->lock);
4219
4220         block_rsv->size = num_bytes;
4221
4222         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4223                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4224                     sinfo->bytes_may_use;
4225
4226         if (sinfo->total_bytes > num_bytes) {
4227                 num_bytes = sinfo->total_bytes - num_bytes;
4228                 block_rsv->reserved += num_bytes;
4229                 sinfo->bytes_may_use += num_bytes;
4230                 trace_btrfs_space_reservation(fs_info, "space_info",
4231                                       sinfo->flags, num_bytes, 1);
4232         }
4233
4234         if (block_rsv->reserved >= block_rsv->size) {
4235                 num_bytes = block_rsv->reserved - block_rsv->size;
4236                 sinfo->bytes_may_use -= num_bytes;
4237                 trace_btrfs_space_reservation(fs_info, "space_info",
4238                                       sinfo->flags, num_bytes, 0);
4239                 sinfo->reservation_progress++;
4240                 block_rsv->reserved = block_rsv->size;
4241                 block_rsv->full = 1;
4242         }
4243
4244         spin_unlock(&block_rsv->lock);
4245         spin_unlock(&sinfo->lock);
4246 }
4247
4248 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4249 {
4250         struct btrfs_space_info *space_info;
4251
4252         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4253         fs_info->chunk_block_rsv.space_info = space_info;
4254
4255         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4256         fs_info->global_block_rsv.space_info = space_info;
4257         fs_info->delalloc_block_rsv.space_info = space_info;
4258         fs_info->trans_block_rsv.space_info = space_info;
4259         fs_info->empty_block_rsv.space_info = space_info;
4260         fs_info->delayed_block_rsv.space_info = space_info;
4261
4262         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4263         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4264         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4265         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4266         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4267
4268         update_global_block_rsv(fs_info);
4269 }
4270
4271 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4272 {
4273         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4274                                 (u64)-1);
4275         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4276         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4277         WARN_ON(fs_info->trans_block_rsv.size > 0);
4278         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4279         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4280         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4281         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4282         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4283 }
4284
4285 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4286                                   struct btrfs_root *root)
4287 {
4288         if (!trans->bytes_reserved)
4289                 return;
4290
4291         trace_btrfs_space_reservation(root->fs_info, "transaction",
4292                                       trans->transid, trans->bytes_reserved, 0);
4293         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4294         trans->bytes_reserved = 0;
4295 }
4296
4297 /* Can only return 0 or -ENOSPC */
4298 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4299                                   struct inode *inode)
4300 {
4301         struct btrfs_root *root = BTRFS_I(inode)->root;
4302         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4303         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4304
4305         /*
4306          * We need to hold space in order to delete our orphan item once we've
4307          * added it, so this takes the reservation so we can release it later
4308          * when we are truly done with the orphan item.
4309          */
4310         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4311         trace_btrfs_space_reservation(root->fs_info, "orphan",
4312                                       btrfs_ino(inode), num_bytes, 1);
4313         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4314 }
4315
4316 void btrfs_orphan_release_metadata(struct inode *inode)
4317 {
4318         struct btrfs_root *root = BTRFS_I(inode)->root;
4319         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4320         trace_btrfs_space_reservation(root->fs_info, "orphan",
4321                                       btrfs_ino(inode), num_bytes, 0);
4322         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4323 }
4324
4325 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4326                                 struct btrfs_pending_snapshot *pending)
4327 {
4328         struct btrfs_root *root = pending->root;
4329         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4330         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4331         /*
4332          * two for root back/forward refs, two for directory entries
4333          * and one for root of the snapshot.
4334          */
4335         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4336         dst_rsv->space_info = src_rsv->space_info;
4337         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4338 }
4339
4340 /**
4341  * drop_outstanding_extent - drop an outstanding extent
4342  * @inode: the inode we're dropping the extent for
4343  *
4344  * This is called when we are freeing up an outstanding extent, either called
4345  * after an error or after an extent is written.  This will return the number of
4346  * reserved extents that need to be freed.  This must be called with
4347  * BTRFS_I(inode)->lock held.
4348  */
4349 static unsigned drop_outstanding_extent(struct inode *inode)
4350 {
4351         unsigned drop_inode_space = 0;
4352         unsigned dropped_extents = 0;
4353
4354         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4355         BTRFS_I(inode)->outstanding_extents--;
4356
4357         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4358             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4359                                &BTRFS_I(inode)->runtime_flags))
4360                 drop_inode_space = 1;
4361
4362         /*
4363          * If we have more or the same amount of outsanding extents than we have
4364          * reserved then we need to leave the reserved extents count alone.
4365          */
4366         if (BTRFS_I(inode)->outstanding_extents >=
4367             BTRFS_I(inode)->reserved_extents)
4368                 return drop_inode_space;
4369
4370         dropped_extents = BTRFS_I(inode)->reserved_extents -
4371                 BTRFS_I(inode)->outstanding_extents;
4372         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4373         return dropped_extents + drop_inode_space;
4374 }
4375
4376 /**
4377  * calc_csum_metadata_size - return the amount of metada space that must be
4378  *      reserved/free'd for the given bytes.
4379  * @inode: the inode we're manipulating
4380  * @num_bytes: the number of bytes in question
4381  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4382  *
4383  * This adjusts the number of csum_bytes in the inode and then returns the
4384  * correct amount of metadata that must either be reserved or freed.  We
4385  * calculate how many checksums we can fit into one leaf and then divide the
4386  * number of bytes that will need to be checksumed by this value to figure out
4387  * how many checksums will be required.  If we are adding bytes then the number
4388  * may go up and we will return the number of additional bytes that must be
4389  * reserved.  If it is going down we will return the number of bytes that must
4390  * be freed.
4391  *
4392  * This must be called with BTRFS_I(inode)->lock held.
4393  */
4394 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4395                                    int reserve)
4396 {
4397         struct btrfs_root *root = BTRFS_I(inode)->root;
4398         u64 csum_size;
4399         int num_csums_per_leaf;
4400         int num_csums;
4401         int old_csums;
4402
4403         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4404             BTRFS_I(inode)->csum_bytes == 0)
4405                 return 0;
4406
4407         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4408         if (reserve)
4409                 BTRFS_I(inode)->csum_bytes += num_bytes;
4410         else
4411                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4412         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4413         num_csums_per_leaf = (int)div64_u64(csum_size,
4414                                             sizeof(struct btrfs_csum_item) +
4415                                             sizeof(struct btrfs_disk_key));
4416         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4417         num_csums = num_csums + num_csums_per_leaf - 1;
4418         num_csums = num_csums / num_csums_per_leaf;
4419
4420         old_csums = old_csums + num_csums_per_leaf - 1;
4421         old_csums = old_csums / num_csums_per_leaf;
4422
4423         /* No change, no need to reserve more */
4424         if (old_csums == num_csums)
4425                 return 0;
4426
4427         if (reserve)
4428                 return btrfs_calc_trans_metadata_size(root,
4429                                                       num_csums - old_csums);
4430
4431         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4432 }
4433
4434 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4435 {
4436         struct btrfs_root *root = BTRFS_I(inode)->root;
4437         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4438         u64 to_reserve = 0;
4439         u64 csum_bytes;
4440         unsigned nr_extents = 0;
4441         int extra_reserve = 0;
4442         int flush = 1;
4443         int ret;
4444
4445         /* Need to be holding the i_mutex here if we aren't free space cache */
4446         if (btrfs_is_free_space_inode(root, inode))
4447                 flush = 0;
4448
4449         if (flush && btrfs_transaction_in_commit(root->fs_info))
4450                 schedule_timeout(1);
4451
4452         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4453         num_bytes = ALIGN(num_bytes, root->sectorsize);
4454
4455         spin_lock(&BTRFS_I(inode)->lock);
4456         BTRFS_I(inode)->outstanding_extents++;
4457
4458         if (BTRFS_I(inode)->outstanding_extents >
4459             BTRFS_I(inode)->reserved_extents)
4460                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4461                         BTRFS_I(inode)->reserved_extents;
4462
4463         /*
4464          * Add an item to reserve for updating the inode when we complete the
4465          * delalloc io.
4466          */
4467         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4468                       &BTRFS_I(inode)->runtime_flags)) {
4469                 nr_extents++;
4470                 extra_reserve = 1;
4471         }
4472
4473         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4474         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4475         csum_bytes = BTRFS_I(inode)->csum_bytes;
4476         spin_unlock(&BTRFS_I(inode)->lock);
4477
4478         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4479         if (ret) {
4480                 u64 to_free = 0;
4481                 unsigned dropped;
4482
4483                 spin_lock(&BTRFS_I(inode)->lock);
4484                 dropped = drop_outstanding_extent(inode);
4485                 /*
4486                  * If the inodes csum_bytes is the same as the original
4487                  * csum_bytes then we know we haven't raced with any free()ers
4488                  * so we can just reduce our inodes csum bytes and carry on.
4489                  * Otherwise we have to do the normal free thing to account for
4490                  * the case that the free side didn't free up its reserve
4491                  * because of this outstanding reservation.
4492                  */
4493                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4494                         calc_csum_metadata_size(inode, num_bytes, 0);
4495                 else
4496                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4497                 spin_unlock(&BTRFS_I(inode)->lock);
4498                 if (dropped)
4499                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4500
4501                 if (to_free) {
4502                         btrfs_block_rsv_release(root, block_rsv, to_free);
4503                         trace_btrfs_space_reservation(root->fs_info,
4504                                                       "delalloc",
4505                                                       btrfs_ino(inode),
4506                                                       to_free, 0);
4507                 }
4508                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4509                 return ret;
4510         }
4511
4512         spin_lock(&BTRFS_I(inode)->lock);
4513         if (extra_reserve) {
4514                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4515                         &BTRFS_I(inode)->runtime_flags);
4516                 nr_extents--;
4517         }
4518         BTRFS_I(inode)->reserved_extents += nr_extents;
4519         spin_unlock(&BTRFS_I(inode)->lock);
4520         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4521
4522         if (to_reserve)
4523                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4524                                               btrfs_ino(inode), to_reserve, 1);
4525         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4526
4527         return 0;
4528 }
4529
4530 /**
4531  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4532  * @inode: the inode to release the reservation for
4533  * @num_bytes: the number of bytes we're releasing
4534  *
4535  * This will release the metadata reservation for an inode.  This can be called
4536  * once we complete IO for a given set of bytes to release their metadata
4537  * reservations.
4538  */
4539 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4540 {
4541         struct btrfs_root *root = BTRFS_I(inode)->root;
4542         u64 to_free = 0;
4543         unsigned dropped;
4544
4545         num_bytes = ALIGN(num_bytes, root->sectorsize);
4546         spin_lock(&BTRFS_I(inode)->lock);
4547         dropped = drop_outstanding_extent(inode);
4548
4549         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4550         spin_unlock(&BTRFS_I(inode)->lock);
4551         if (dropped > 0)
4552                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4553
4554         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4555                                       btrfs_ino(inode), to_free, 0);
4556         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4557                                 to_free);
4558 }
4559
4560 /**
4561  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4562  * @inode: inode we're writing to
4563  * @num_bytes: the number of bytes we want to allocate
4564  *
4565  * This will do the following things
4566  *
4567  * o reserve space in the data space info for num_bytes
4568  * o reserve space in the metadata space info based on number of outstanding
4569  *   extents and how much csums will be needed
4570  * o add to the inodes ->delalloc_bytes
4571  * o add it to the fs_info's delalloc inodes list.
4572  *
4573  * This will return 0 for success and -ENOSPC if there is no space left.
4574  */
4575 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4576 {
4577         int ret;
4578
4579         ret = btrfs_check_data_free_space(inode, num_bytes);
4580         if (ret)
4581                 return ret;
4582
4583         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4584         if (ret) {
4585                 btrfs_free_reserved_data_space(inode, num_bytes);
4586                 return ret;
4587         }
4588
4589         return 0;
4590 }
4591
4592 /**
4593  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4594  * @inode: inode we're releasing space for
4595  * @num_bytes: the number of bytes we want to free up
4596  *
4597  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4598  * called in the case that we don't need the metadata AND data reservations
4599  * anymore.  So if there is an error or we insert an inline extent.
4600  *
4601  * This function will release the metadata space that was not used and will
4602  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4603  * list if there are no delalloc bytes left.
4604  */
4605 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4606 {
4607         btrfs_delalloc_release_metadata(inode, num_bytes);
4608         btrfs_free_reserved_data_space(inode, num_bytes);
4609 }
4610
4611 static int update_block_group(struct btrfs_trans_handle *trans,
4612                               struct btrfs_root *root,
4613                               u64 bytenr, u64 num_bytes, int alloc)
4614 {
4615         struct btrfs_block_group_cache *cache = NULL;
4616         struct btrfs_fs_info *info = root->fs_info;
4617         u64 total = num_bytes;
4618         u64 old_val;
4619         u64 byte_in_group;
4620         int factor;
4621
4622         /* block accounting for super block */
4623         spin_lock(&info->delalloc_lock);
4624         old_val = btrfs_super_bytes_used(info->super_copy);
4625         if (alloc)
4626                 old_val += num_bytes;
4627         else
4628                 old_val -= num_bytes;
4629         btrfs_set_super_bytes_used(info->super_copy, old_val);
4630         spin_unlock(&info->delalloc_lock);
4631
4632         while (total) {
4633                 cache = btrfs_lookup_block_group(info, bytenr);
4634                 if (!cache)
4635                         return -ENOENT;
4636                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4637                                     BTRFS_BLOCK_GROUP_RAID1 |
4638                                     BTRFS_BLOCK_GROUP_RAID10))
4639                         factor = 2;
4640                 else
4641                         factor = 1;
4642                 /*
4643                  * If this block group has free space cache written out, we
4644                  * need to make sure to load it if we are removing space.  This
4645                  * is because we need the unpinning stage to actually add the
4646                  * space back to the block group, otherwise we will leak space.
4647                  */
4648                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4649                         cache_block_group(cache, trans, NULL, 1);
4650
4651                 byte_in_group = bytenr - cache->key.objectid;
4652                 WARN_ON(byte_in_group > cache->key.offset);
4653
4654                 spin_lock(&cache->space_info->lock);
4655                 spin_lock(&cache->lock);
4656
4657                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4658                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4659                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4660
4661                 cache->dirty = 1;
4662                 old_val = btrfs_block_group_used(&cache->item);
4663                 num_bytes = min(total, cache->key.offset - byte_in_group);
4664                 if (alloc) {
4665                         old_val += num_bytes;
4666                         btrfs_set_block_group_used(&cache->item, old_val);
4667                         cache->reserved -= num_bytes;
4668                         cache->space_info->bytes_reserved -= num_bytes;
4669                         cache->space_info->bytes_used += num_bytes;
4670                         cache->space_info->disk_used += num_bytes * factor;
4671                         spin_unlock(&cache->lock);
4672                         spin_unlock(&cache->space_info->lock);
4673                 } else {
4674                         old_val -= num_bytes;
4675                         btrfs_set_block_group_used(&cache->item, old_val);
4676                         cache->pinned += num_bytes;
4677                         cache->space_info->bytes_pinned += num_bytes;
4678                         cache->space_info->bytes_used -= num_bytes;
4679                         cache->space_info->disk_used -= num_bytes * factor;
4680                         spin_unlock(&cache->lock);
4681                         spin_unlock(&cache->space_info->lock);
4682
4683                         set_extent_dirty(info->pinned_extents,
4684                                          bytenr, bytenr + num_bytes - 1,
4685                                          GFP_NOFS | __GFP_NOFAIL);
4686                 }
4687                 btrfs_put_block_group(cache);
4688                 total -= num_bytes;
4689                 bytenr += num_bytes;
4690         }
4691         return 0;
4692 }
4693
4694 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4695 {
4696         struct btrfs_block_group_cache *cache;
4697         u64 bytenr;
4698
4699         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4700         if (!cache)
4701                 return 0;
4702
4703         bytenr = cache->key.objectid;
4704         btrfs_put_block_group(cache);
4705
4706         return bytenr;
4707 }
4708
4709 static int pin_down_extent(struct btrfs_root *root,
4710                            struct btrfs_block_group_cache *cache,
4711                            u64 bytenr, u64 num_bytes, int reserved)
4712 {
4713         spin_lock(&cache->space_info->lock);
4714         spin_lock(&cache->lock);
4715         cache->pinned += num_bytes;
4716         cache->space_info->bytes_pinned += num_bytes;
4717         if (reserved) {
4718                 cache->reserved -= num_bytes;
4719                 cache->space_info->bytes_reserved -= num_bytes;
4720         }
4721         spin_unlock(&cache->lock);
4722         spin_unlock(&cache->space_info->lock);
4723
4724         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4725                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4726         return 0;
4727 }
4728
4729 /*
4730  * this function must be called within transaction
4731  */
4732 int btrfs_pin_extent(struct btrfs_root *root,
4733                      u64 bytenr, u64 num_bytes, int reserved)
4734 {
4735         struct btrfs_block_group_cache *cache;
4736
4737         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4738         BUG_ON(!cache); /* Logic error */
4739
4740         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4741
4742         btrfs_put_block_group(cache);
4743         return 0;
4744 }
4745
4746 /*
4747  * this function must be called within transaction
4748  */
4749 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4750                                     struct btrfs_root *root,
4751                                     u64 bytenr, u64 num_bytes)
4752 {
4753         struct btrfs_block_group_cache *cache;
4754
4755         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4756         BUG_ON(!cache); /* Logic error */
4757
4758         /*
4759          * pull in the free space cache (if any) so that our pin
4760          * removes the free space from the cache.  We have load_only set
4761          * to one because the slow code to read in the free extents does check
4762          * the pinned extents.
4763          */
4764         cache_block_group(cache, trans, root, 1);
4765
4766         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4767
4768         /* remove us from the free space cache (if we're there at all) */
4769         btrfs_remove_free_space(cache, bytenr, num_bytes);
4770         btrfs_put_block_group(cache);
4771         return 0;
4772 }
4773
4774 /**
4775  * btrfs_update_reserved_bytes - update the block_group and space info counters
4776  * @cache:      The cache we are manipulating
4777  * @num_bytes:  The number of bytes in question
4778  * @reserve:    One of the reservation enums
4779  *
4780  * This is called by the allocator when it reserves space, or by somebody who is
4781  * freeing space that was never actually used on disk.  For example if you
4782  * reserve some space for a new leaf in transaction A and before transaction A
4783  * commits you free that leaf, you call this with reserve set to 0 in order to
4784  * clear the reservation.
4785  *
4786  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4787  * ENOSPC accounting.  For data we handle the reservation through clearing the
4788  * delalloc bits in the io_tree.  We have to do this since we could end up
4789  * allocating less disk space for the amount of data we have reserved in the
4790  * case of compression.
4791  *
4792  * If this is a reservation and the block group has become read only we cannot
4793  * make the reservation and return -EAGAIN, otherwise this function always
4794  * succeeds.
4795  */
4796 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4797                                        u64 num_bytes, int reserve)
4798 {
4799         struct btrfs_space_info *space_info = cache->space_info;
4800         int ret = 0;
4801
4802         spin_lock(&space_info->lock);
4803         spin_lock(&cache->lock);
4804         if (reserve != RESERVE_FREE) {
4805                 if (cache->ro) {
4806                         ret = -EAGAIN;
4807                 } else {
4808                         cache->reserved += num_bytes;
4809                         space_info->bytes_reserved += num_bytes;
4810                         if (reserve == RESERVE_ALLOC) {
4811                                 trace_btrfs_space_reservation(cache->fs_info,
4812                                                 "space_info", space_info->flags,
4813                                                 num_bytes, 0);
4814                                 space_info->bytes_may_use -= num_bytes;
4815                         }
4816                 }
4817         } else {
4818                 if (cache->ro)
4819                         space_info->bytes_readonly += num_bytes;
4820                 cache->reserved -= num_bytes;
4821                 space_info->bytes_reserved -= num_bytes;
4822                 space_info->reservation_progress++;
4823         }
4824         spin_unlock(&cache->lock);
4825         spin_unlock(&space_info->lock);
4826         return ret;
4827 }
4828
4829 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4830                                 struct btrfs_root *root)
4831 {
4832         struct btrfs_fs_info *fs_info = root->fs_info;
4833         struct btrfs_caching_control *next;
4834         struct btrfs_caching_control *caching_ctl;
4835         struct btrfs_block_group_cache *cache;
4836
4837         down_write(&fs_info->extent_commit_sem);
4838
4839         list_for_each_entry_safe(caching_ctl, next,
4840                                  &fs_info->caching_block_groups, list) {
4841                 cache = caching_ctl->block_group;
4842                 if (block_group_cache_done(cache)) {
4843                         cache->last_byte_to_unpin = (u64)-1;
4844                         list_del_init(&caching_ctl->list);
4845                         put_caching_control(caching_ctl);
4846                 } else {
4847                         cache->last_byte_to_unpin = caching_ctl->progress;
4848                 }
4849         }
4850
4851         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4852                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4853         else
4854                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4855
4856         up_write(&fs_info->extent_commit_sem);
4857
4858         update_global_block_rsv(fs_info);
4859 }
4860
4861 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4862 {
4863         struct btrfs_fs_info *fs_info = root->fs_info;
4864         struct btrfs_block_group_cache *cache = NULL;
4865         u64 len;
4866
4867         while (start <= end) {
4868                 if (!cache ||
4869                     start >= cache->key.objectid + cache->key.offset) {
4870                         if (cache)
4871                                 btrfs_put_block_group(cache);
4872                         cache = btrfs_lookup_block_group(fs_info, start);
4873                         BUG_ON(!cache); /* Logic error */
4874                 }
4875
4876                 len = cache->key.objectid + cache->key.offset - start;
4877                 len = min(len, end + 1 - start);
4878
4879                 if (start < cache->last_byte_to_unpin) {
4880                         len = min(len, cache->last_byte_to_unpin - start);
4881                         btrfs_add_free_space(cache, start, len);
4882                 }
4883
4884                 start += len;
4885
4886                 spin_lock(&cache->space_info->lock);
4887                 spin_lock(&cache->lock);
4888                 cache->pinned -= len;
4889                 cache->space_info->bytes_pinned -= len;
4890                 if (cache->ro)
4891                         cache->space_info->bytes_readonly += len;
4892                 spin_unlock(&cache->lock);
4893                 spin_unlock(&cache->space_info->lock);
4894         }
4895
4896         if (cache)
4897                 btrfs_put_block_group(cache);
4898         return 0;
4899 }
4900
4901 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4902                                struct btrfs_root *root)
4903 {
4904         struct btrfs_fs_info *fs_info = root->fs_info;
4905         struct extent_io_tree *unpin;
4906         u64 start;
4907         u64 end;
4908         int ret;
4909
4910         if (trans->aborted)
4911                 return 0;
4912
4913         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4914                 unpin = &fs_info->freed_extents[1];
4915         else
4916                 unpin = &fs_info->freed_extents[0];
4917
4918         while (1) {
4919                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4920                                             EXTENT_DIRTY);
4921                 if (ret)
4922                         break;
4923
4924                 if (btrfs_test_opt(root, DISCARD))
4925                         ret = btrfs_discard_extent(root, start,
4926                                                    end + 1 - start, NULL);
4927
4928                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4929                 unpin_extent_range(root, start, end);
4930                 cond_resched();
4931         }
4932
4933         return 0;
4934 }
4935
4936 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4937                                 struct btrfs_root *root,
4938                                 u64 bytenr, u64 num_bytes, u64 parent,
4939                                 u64 root_objectid, u64 owner_objectid,
4940                                 u64 owner_offset, int refs_to_drop,
4941                                 struct btrfs_delayed_extent_op *extent_op)
4942 {
4943         struct btrfs_key key;
4944         struct btrfs_path *path;
4945         struct btrfs_fs_info *info = root->fs_info;
4946         struct btrfs_root *extent_root = info->extent_root;
4947         struct extent_buffer *leaf;
4948         struct btrfs_extent_item *ei;
4949         struct btrfs_extent_inline_ref *iref;
4950         int ret;
4951         int is_data;
4952         int extent_slot = 0;
4953         int found_extent = 0;
4954         int num_to_del = 1;
4955         u32 item_size;
4956         u64 refs;
4957
4958         path = btrfs_alloc_path();
4959         if (!path)
4960                 return -ENOMEM;
4961
4962         path->reada = 1;
4963         path->leave_spinning = 1;
4964
4965         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4966         BUG_ON(!is_data && refs_to_drop != 1);
4967
4968         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4969                                     bytenr, num_bytes, parent,
4970                                     root_objectid, owner_objectid,
4971                                     owner_offset);
4972         if (ret == 0) {
4973                 extent_slot = path->slots[0];
4974                 while (extent_slot >= 0) {
4975                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4976                                               extent_slot);
4977                         if (key.objectid != bytenr)
4978                                 break;
4979                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4980                             key.offset == num_bytes) {
4981                                 found_extent = 1;
4982                                 break;
4983                         }
4984                         if (path->slots[0] - extent_slot > 5)
4985                                 break;
4986                         extent_slot--;
4987                 }
4988 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4989                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4990                 if (found_extent && item_size < sizeof(*ei))
4991                         found_extent = 0;
4992 #endif
4993                 if (!found_extent) {
4994                         BUG_ON(iref);
4995                         ret = remove_extent_backref(trans, extent_root, path,
4996                                                     NULL, refs_to_drop,
4997                                                     is_data);
4998                         if (ret)
4999                                 goto abort;
5000                         btrfs_release_path(path);
5001                         path->leave_spinning = 1;
5002
5003                         key.objectid = bytenr;
5004                         key.type = BTRFS_EXTENT_ITEM_KEY;
5005                         key.offset = num_bytes;
5006
5007                         ret = btrfs_search_slot(trans, extent_root,
5008                                                 &key, path, -1, 1);
5009                         if (ret) {
5010                                 printk(KERN_ERR "umm, got %d back from search"
5011                                        ", was looking for %llu\n", ret,
5012                                        (unsigned long long)bytenr);
5013                                 if (ret > 0)
5014                                         btrfs_print_leaf(extent_root,
5015                                                          path->nodes[0]);
5016                         }
5017                         if (ret < 0)
5018                                 goto abort;
5019                         extent_slot = path->slots[0];
5020                 }
5021         } else if (ret == -ENOENT) {
5022                 btrfs_print_leaf(extent_root, path->nodes[0]);
5023                 WARN_ON(1);
5024                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5025                        "parent %llu root %llu  owner %llu offset %llu\n",
5026                        (unsigned long long)bytenr,
5027                        (unsigned long long)parent,
5028                        (unsigned long long)root_objectid,
5029                        (unsigned long long)owner_objectid,
5030                        (unsigned long long)owner_offset);
5031         } else {
5032                 goto abort;
5033         }
5034
5035         leaf = path->nodes[0];
5036         item_size = btrfs_item_size_nr(leaf, extent_slot);
5037 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5038         if (item_size < sizeof(*ei)) {
5039                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5040                 ret = convert_extent_item_v0(trans, extent_root, path,
5041                                              owner_objectid, 0);
5042                 if (ret < 0)
5043                         goto abort;
5044
5045                 btrfs_release_path(path);
5046                 path->leave_spinning = 1;
5047
5048                 key.objectid = bytenr;
5049                 key.type = BTRFS_EXTENT_ITEM_KEY;
5050                 key.offset = num_bytes;
5051
5052                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5053                                         -1, 1);
5054                 if (ret) {
5055                         printk(KERN_ERR "umm, got %d back from search"
5056                                ", was looking for %llu\n", ret,
5057                                (unsigned long long)bytenr);
5058                         btrfs_print_leaf(extent_root, path->nodes[0]);
5059                 }
5060                 if (ret < 0)
5061                         goto abort;
5062                 extent_slot = path->slots[0];
5063                 leaf = path->nodes[0];
5064                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5065         }
5066 #endif
5067         BUG_ON(item_size < sizeof(*ei));
5068         ei = btrfs_item_ptr(leaf, extent_slot,
5069                             struct btrfs_extent_item);
5070         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5071                 struct btrfs_tree_block_info *bi;
5072                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5073                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5074                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5075         }
5076
5077         refs = btrfs_extent_refs(leaf, ei);
5078         BUG_ON(refs < refs_to_drop);
5079         refs -= refs_to_drop;
5080
5081         if (refs > 0) {
5082                 if (extent_op)
5083                         __run_delayed_extent_op(extent_op, leaf, ei);
5084                 /*
5085                  * In the case of inline back ref, reference count will
5086                  * be updated by remove_extent_backref
5087                  */
5088                 if (iref) {
5089                         BUG_ON(!found_extent);
5090                 } else {
5091                         btrfs_set_extent_refs(leaf, ei, refs);
5092                         btrfs_mark_buffer_dirty(leaf);
5093                 }
5094                 if (found_extent) {
5095                         ret = remove_extent_backref(trans, extent_root, path,
5096                                                     iref, refs_to_drop,
5097                                                     is_data);
5098                         if (ret)
5099                                 goto abort;
5100                 }
5101         } else {
5102                 if (found_extent) {
5103                         BUG_ON(is_data && refs_to_drop !=
5104                                extent_data_ref_count(root, path, iref));
5105                         if (iref) {
5106                                 BUG_ON(path->slots[0] != extent_slot);
5107                         } else {
5108                                 BUG_ON(path->slots[0] != extent_slot + 1);
5109                                 path->slots[0] = extent_slot;
5110                                 num_to_del = 2;
5111                         }
5112                 }
5113
5114                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5115                                       num_to_del);
5116                 if (ret)
5117                         goto abort;
5118                 btrfs_release_path(path);
5119
5120                 if (is_data) {
5121                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5122                         if (ret)
5123                                 goto abort;
5124                 }
5125
5126                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5127                 if (ret)
5128                         goto abort;
5129         }
5130 out:
5131         btrfs_free_path(path);
5132         return ret;
5133
5134 abort:
5135         btrfs_abort_transaction(trans, extent_root, ret);
5136         goto out;
5137 }
5138
5139 /*
5140  * when we free an block, it is possible (and likely) that we free the last
5141  * delayed ref for that extent as well.  This searches the delayed ref tree for
5142  * a given extent, and if there are no other delayed refs to be processed, it
5143  * removes it from the tree.
5144  */
5145 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5146                                       struct btrfs_root *root, u64 bytenr)
5147 {
5148         struct btrfs_delayed_ref_head *head;
5149         struct btrfs_delayed_ref_root *delayed_refs;
5150         struct btrfs_delayed_ref_node *ref;
5151         struct rb_node *node;
5152         int ret = 0;
5153
5154         delayed_refs = &trans->transaction->delayed_refs;
5155         spin_lock(&delayed_refs->lock);
5156         head = btrfs_find_delayed_ref_head(trans, bytenr);
5157         if (!head)
5158                 goto out;
5159
5160         node = rb_prev(&head->node.rb_node);
5161         if (!node)
5162                 goto out;
5163
5164         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5165
5166         /* there are still entries for this ref, we can't drop it */
5167         if (ref->bytenr == bytenr)
5168                 goto out;
5169
5170         if (head->extent_op) {
5171                 if (!head->must_insert_reserved)
5172                         goto out;
5173                 kfree(head->extent_op);
5174                 head->extent_op = NULL;
5175         }
5176
5177         /*
5178          * waiting for the lock here would deadlock.  If someone else has it
5179          * locked they are already in the process of dropping it anyway
5180          */
5181         if (!mutex_trylock(&head->mutex))
5182                 goto out;
5183
5184         /*
5185          * at this point we have a head with no other entries.  Go
5186          * ahead and process it.
5187          */
5188         head->node.in_tree = 0;
5189         rb_erase(&head->node.rb_node, &delayed_refs->root);
5190
5191         delayed_refs->num_entries--;
5192         if (waitqueue_active(&delayed_refs->seq_wait))
5193                 wake_up(&delayed_refs->seq_wait);
5194
5195         /*
5196          * we don't take a ref on the node because we're removing it from the
5197          * tree, so we just steal the ref the tree was holding.
5198          */
5199         delayed_refs->num_heads--;
5200         if (list_empty(&head->cluster))
5201                 delayed_refs->num_heads_ready--;
5202
5203         list_del_init(&head->cluster);
5204         spin_unlock(&delayed_refs->lock);
5205
5206         BUG_ON(head->extent_op);
5207         if (head->must_insert_reserved)
5208                 ret = 1;
5209
5210         mutex_unlock(&head->mutex);
5211         btrfs_put_delayed_ref(&head->node);
5212         return ret;
5213 out:
5214         spin_unlock(&delayed_refs->lock);
5215         return 0;
5216 }
5217
5218 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5219                            struct btrfs_root *root,
5220                            struct extent_buffer *buf,
5221                            u64 parent, int last_ref)
5222 {
5223         struct btrfs_block_group_cache *cache = NULL;
5224         int ret;
5225
5226         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5227                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5228                                         buf->start, buf->len,
5229                                         parent, root->root_key.objectid,
5230                                         btrfs_header_level(buf),
5231                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5232                 BUG_ON(ret); /* -ENOMEM */
5233         }
5234
5235         if (!last_ref)
5236                 return;
5237
5238         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5239
5240         if (btrfs_header_generation(buf) == trans->transid) {
5241                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5242                         ret = check_ref_cleanup(trans, root, buf->start);
5243                         if (!ret)
5244                                 goto out;
5245                 }
5246
5247                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5248                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5249                         goto out;
5250                 }
5251
5252                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5253
5254                 btrfs_add_free_space(cache, buf->start, buf->len);
5255                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5256         }
5257 out:
5258         /*
5259          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5260          * anymore.
5261          */
5262         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5263         btrfs_put_block_group(cache);
5264 }
5265
5266 /* Can return -ENOMEM */
5267 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5268                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5269                       u64 owner, u64 offset, int for_cow)
5270 {
5271         int ret;
5272         struct btrfs_fs_info *fs_info = root->fs_info;
5273
5274         /*
5275          * tree log blocks never actually go into the extent allocation
5276          * tree, just update pinning info and exit early.
5277          */
5278         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5279                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5280                 /* unlocks the pinned mutex */
5281                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5282                 ret = 0;
5283         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5284                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5285                                         num_bytes,
5286                                         parent, root_objectid, (int)owner,
5287                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5288         } else {
5289                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5290                                                 num_bytes,
5291                                                 parent, root_objectid, owner,
5292                                                 offset, BTRFS_DROP_DELAYED_REF,
5293                                                 NULL, for_cow);
5294         }
5295         return ret;
5296 }
5297
5298 static u64 stripe_align(struct btrfs_root *root, u64 val)
5299 {
5300         u64 mask = ((u64)root->stripesize - 1);
5301         u64 ret = (val + mask) & ~mask;
5302         return ret;
5303 }
5304
5305 /*
5306  * when we wait for progress in the block group caching, its because
5307  * our allocation attempt failed at least once.  So, we must sleep
5308  * and let some progress happen before we try again.
5309  *
5310  * This function will sleep at least once waiting for new free space to
5311  * show up, and then it will check the block group free space numbers
5312  * for our min num_bytes.  Another option is to have it go ahead
5313  * and look in the rbtree for a free extent of a given size, but this
5314  * is a good start.
5315  */
5316 static noinline int
5317 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5318                                 u64 num_bytes)
5319 {
5320         struct btrfs_caching_control *caching_ctl;
5321         DEFINE_WAIT(wait);
5322
5323         caching_ctl = get_caching_control(cache);
5324         if (!caching_ctl)
5325                 return 0;
5326
5327         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5328                    (cache->free_space_ctl->free_space >= num_bytes));
5329
5330         put_caching_control(caching_ctl);
5331         return 0;
5332 }
5333
5334 static noinline int
5335 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5336 {
5337         struct btrfs_caching_control *caching_ctl;
5338         DEFINE_WAIT(wait);
5339
5340         caching_ctl = get_caching_control(cache);
5341         if (!caching_ctl)
5342                 return 0;
5343
5344         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5345
5346         put_caching_control(caching_ctl);
5347         return 0;
5348 }
5349
5350 static int __get_block_group_index(u64 flags)
5351 {
5352         int index;
5353
5354         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5355                 index = 0;
5356         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5357                 index = 1;
5358         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5359                 index = 2;
5360         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5361                 index = 3;
5362         else
5363                 index = 4;
5364
5365         return index;
5366 }
5367
5368 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5369 {
5370         return __get_block_group_index(cache->flags);
5371 }
5372
5373 enum btrfs_loop_type {
5374         LOOP_CACHING_NOWAIT = 0,
5375         LOOP_CACHING_WAIT = 1,
5376         LOOP_ALLOC_CHUNK = 2,
5377         LOOP_NO_EMPTY_SIZE = 3,
5378 };
5379
5380 /*
5381  * walks the btree of allocated extents and find a hole of a given size.
5382  * The key ins is changed to record the hole:
5383  * ins->objectid == block start
5384  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5385  * ins->offset == number of blocks
5386  * Any available blocks before search_start are skipped.
5387  */
5388 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5389                                      struct btrfs_root *orig_root,
5390                                      u64 num_bytes, u64 empty_size,
5391                                      u64 hint_byte, struct btrfs_key *ins,
5392                                      u64 data)
5393 {
5394         int ret = 0;
5395         struct btrfs_root *root = orig_root->fs_info->extent_root;
5396         struct btrfs_free_cluster *last_ptr = NULL;
5397         struct btrfs_block_group_cache *block_group = NULL;
5398         struct btrfs_block_group_cache *used_block_group;
5399         u64 search_start = 0;
5400         int empty_cluster = 2 * 1024 * 1024;
5401         int allowed_chunk_alloc = 0;
5402         int done_chunk_alloc = 0;
5403         struct btrfs_space_info *space_info;
5404         int loop = 0;
5405         int index = 0;
5406         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5407                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5408         bool found_uncached_bg = false;
5409         bool failed_cluster_refill = false;
5410         bool failed_alloc = false;
5411         bool use_cluster = true;
5412         bool have_caching_bg = false;
5413
5414         WARN_ON(num_bytes < root->sectorsize);
5415         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5416         ins->objectid = 0;
5417         ins->offset = 0;
5418
5419         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5420
5421         space_info = __find_space_info(root->fs_info, data);
5422         if (!space_info) {
5423                 printk(KERN_ERR "No space info for %llu\n", data);
5424                 return -ENOSPC;
5425         }
5426
5427         /*
5428          * If the space info is for both data and metadata it means we have a
5429          * small filesystem and we can't use the clustering stuff.
5430          */
5431         if (btrfs_mixed_space_info(space_info))
5432                 use_cluster = false;
5433
5434         if (orig_root->ref_cows || empty_size)
5435                 allowed_chunk_alloc = 1;
5436
5437         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5438                 last_ptr = &root->fs_info->meta_alloc_cluster;
5439                 if (!btrfs_test_opt(root, SSD))
5440                         empty_cluster = 64 * 1024;
5441         }
5442
5443         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5444             btrfs_test_opt(root, SSD)) {
5445                 last_ptr = &root->fs_info->data_alloc_cluster;
5446         }
5447
5448         if (last_ptr) {
5449                 spin_lock(&last_ptr->lock);
5450                 if (last_ptr->block_group)
5451                         hint_byte = last_ptr->window_start;
5452                 spin_unlock(&last_ptr->lock);
5453         }
5454
5455         search_start = max(search_start, first_logical_byte(root, 0));
5456         search_start = max(search_start, hint_byte);
5457
5458         if (!last_ptr)
5459                 empty_cluster = 0;
5460
5461         if (search_start == hint_byte) {
5462                 block_group = btrfs_lookup_block_group(root->fs_info,
5463                                                        search_start);
5464                 used_block_group = block_group;
5465                 /*
5466                  * we don't want to use the block group if it doesn't match our
5467                  * allocation bits, or if its not cached.
5468                  *
5469                  * However if we are re-searching with an ideal block group
5470                  * picked out then we don't care that the block group is cached.
5471                  */
5472                 if (block_group && block_group_bits(block_group, data) &&
5473                     block_group->cached != BTRFS_CACHE_NO) {
5474                         down_read(&space_info->groups_sem);
5475                         if (list_empty(&block_group->list) ||
5476                             block_group->ro) {
5477                                 /*
5478                                  * someone is removing this block group,
5479                                  * we can't jump into the have_block_group
5480                                  * target because our list pointers are not
5481                                  * valid
5482                                  */
5483                                 btrfs_put_block_group(block_group);
5484                                 up_read(&space_info->groups_sem);
5485                         } else {
5486                                 index = get_block_group_index(block_group);
5487                                 goto have_block_group;
5488                         }
5489                 } else if (block_group) {
5490                         btrfs_put_block_group(block_group);
5491                 }
5492         }
5493 search:
5494         have_caching_bg = false;
5495         down_read(&space_info->groups_sem);
5496         list_for_each_entry(block_group, &space_info->block_groups[index],
5497                             list) {
5498                 u64 offset;
5499                 int cached;
5500
5501                 used_block_group = block_group;
5502                 btrfs_get_block_group(block_group);
5503                 search_start = block_group->key.objectid;
5504
5505                 /*
5506                  * this can happen if we end up cycling through all the
5507                  * raid types, but we want to make sure we only allocate
5508                  * for the proper type.
5509                  */
5510                 if (!block_group_bits(block_group, data)) {
5511                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5512                                 BTRFS_BLOCK_GROUP_RAID1 |
5513                                 BTRFS_BLOCK_GROUP_RAID10;
5514
5515                         /*
5516                          * if they asked for extra copies and this block group
5517                          * doesn't provide them, bail.  This does allow us to
5518                          * fill raid0 from raid1.
5519                          */
5520                         if ((data & extra) && !(block_group->flags & extra))
5521                                 goto loop;
5522                 }
5523
5524 have_block_group:
5525                 cached = block_group_cache_done(block_group);
5526                 if (unlikely(!cached)) {
5527                         found_uncached_bg = true;
5528                         ret = cache_block_group(block_group, trans,
5529                                                 orig_root, 0);
5530                         BUG_ON(ret < 0);
5531                         ret = 0;
5532                 }
5533
5534                 if (unlikely(block_group->ro))
5535                         goto loop;
5536
5537                 /*
5538                  * Ok we want to try and use the cluster allocator, so
5539                  * lets look there
5540                  */
5541                 if (last_ptr) {
5542                         /*
5543                          * the refill lock keeps out other
5544                          * people trying to start a new cluster
5545                          */
5546                         spin_lock(&last_ptr->refill_lock);
5547                         used_block_group = last_ptr->block_group;
5548                         if (used_block_group != block_group &&
5549                             (!used_block_group ||
5550                              used_block_group->ro ||
5551                              !block_group_bits(used_block_group, data))) {
5552                                 used_block_group = block_group;
5553                                 goto refill_cluster;
5554                         }
5555
5556                         if (used_block_group != block_group)
5557                                 btrfs_get_block_group(used_block_group);
5558
5559                         offset = btrfs_alloc_from_cluster(used_block_group,
5560                           last_ptr, num_bytes, used_block_group->key.objectid);
5561                         if (offset) {
5562                                 /* we have a block, we're done */
5563                                 spin_unlock(&last_ptr->refill_lock);
5564                                 trace_btrfs_reserve_extent_cluster(root,
5565                                         block_group, search_start, num_bytes);
5566                                 goto checks;
5567                         }
5568
5569                         WARN_ON(last_ptr->block_group != used_block_group);
5570                         if (used_block_group != block_group) {
5571                                 btrfs_put_block_group(used_block_group);
5572                                 used_block_group = block_group;
5573                         }
5574 refill_cluster:
5575                         BUG_ON(used_block_group != block_group);
5576                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5577                          * set up a new clusters, so lets just skip it
5578                          * and let the allocator find whatever block
5579                          * it can find.  If we reach this point, we
5580                          * will have tried the cluster allocator
5581                          * plenty of times and not have found
5582                          * anything, so we are likely way too
5583                          * fragmented for the clustering stuff to find
5584                          * anything.
5585                          *
5586                          * However, if the cluster is taken from the
5587                          * current block group, release the cluster
5588                          * first, so that we stand a better chance of
5589                          * succeeding in the unclustered
5590                          * allocation.  */
5591                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5592                             last_ptr->block_group != block_group) {
5593                                 spin_unlock(&last_ptr->refill_lock);
5594                                 goto unclustered_alloc;
5595                         }
5596
5597                         /*
5598                          * this cluster didn't work out, free it and
5599                          * start over
5600                          */
5601                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5602
5603                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5604                                 spin_unlock(&last_ptr->refill_lock);
5605                                 goto unclustered_alloc;
5606                         }
5607
5608                         /* allocate a cluster in this block group */
5609                         ret = btrfs_find_space_cluster(trans, root,
5610                                                block_group, last_ptr,
5611                                                search_start, num_bytes,
5612                                                empty_cluster + empty_size);
5613                         if (ret == 0) {
5614                                 /*
5615                                  * now pull our allocation out of this
5616                                  * cluster
5617                                  */
5618                                 offset = btrfs_alloc_from_cluster(block_group,
5619                                                   last_ptr, num_bytes,
5620                                                   search_start);
5621                                 if (offset) {
5622                                         /* we found one, proceed */
5623                                         spin_unlock(&last_ptr->refill_lock);
5624                                         trace_btrfs_reserve_extent_cluster(root,
5625                                                 block_group, search_start,
5626                                                 num_bytes);
5627                                         goto checks;
5628                                 }
5629                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5630                                    && !failed_cluster_refill) {
5631                                 spin_unlock(&last_ptr->refill_lock);
5632
5633                                 failed_cluster_refill = true;
5634                                 wait_block_group_cache_progress(block_group,
5635                                        num_bytes + empty_cluster + empty_size);
5636                   &n