Btrfs: fix race in run_clustered_refs
[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 #undef SCRAMBLE_DELAYED_REFS
38
39 /*
40  * control flags for do_chunk_alloc's force field
41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
42  * if we really need one.
43  *
44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
45  * if we have very few chunks already allocated.  This is
46  * used as part of the clustering code to help make sure
47  * we have a good pool of storage to cluster in, without
48  * filling the FS with empty chunks
49  *
50  * CHUNK_ALLOC_FORCE means it must try to allocate one
51  *
52  */
53 enum {
54         CHUNK_ALLOC_NO_FORCE = 0,
55         CHUNK_ALLOC_LIMITED = 1,
56         CHUNK_ALLOC_FORCE = 2,
57 };
58
59 /*
60  * Control how reservations are dealt with.
61  *
62  * RESERVE_FREE - freeing a reservation.
63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
64  *   ENOSPC accounting
65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
67  */
68 enum {
69         RESERVE_FREE = 0,
70         RESERVE_ALLOC = 1,
71         RESERVE_ALLOC_NO_ACCOUNT = 2,
72 };
73
74 static int update_block_group(struct btrfs_trans_handle *trans,
75                               struct btrfs_root *root,
76                               u64 bytenr, u64 num_bytes, int alloc);
77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
78                                 struct btrfs_root *root,
79                                 u64 bytenr, u64 num_bytes, u64 parent,
80                                 u64 root_objectid, u64 owner_objectid,
81                                 u64 owner_offset, int refs_to_drop,
82                                 struct btrfs_delayed_extent_op *extra_op);
83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
84                                     struct extent_buffer *leaf,
85                                     struct btrfs_extent_item *ei);
86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
87                                       struct btrfs_root *root,
88                                       u64 parent, u64 root_objectid,
89                                       u64 flags, u64 owner, u64 offset,
90                                       struct btrfs_key *ins, int ref_mod);
91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
92                                      struct btrfs_root *root,
93                                      u64 parent, u64 root_objectid,
94                                      u64 flags, struct btrfs_disk_key *key,
95                                      int level, struct btrfs_key *ins);
96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
97                           struct btrfs_root *extent_root, u64 alloc_bytes,
98                           u64 flags, int force);
99 static int find_next_key(struct btrfs_path *path, int level,
100                          struct btrfs_key *key);
101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
102                             int dump_block_groups);
103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
104                                        u64 num_bytes, int reserve);
105
106 static noinline int
107 block_group_cache_done(struct btrfs_block_group_cache *cache)
108 {
109         smp_mb();
110         return cache->cached == BTRFS_CACHE_FINISHED;
111 }
112
113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
114 {
115         return (cache->flags & bits) == bits;
116 }
117
118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
119 {
120         atomic_inc(&cache->count);
121 }
122
123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
124 {
125         if (atomic_dec_and_test(&cache->count)) {
126                 WARN_ON(cache->pinned > 0);
127                 WARN_ON(cache->reserved > 0);
128                 kfree(cache->free_space_ctl);
129                 kfree(cache);
130         }
131 }
132
133 /*
134  * this adds the block group to the fs_info rb tree for the block group
135  * cache
136  */
137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
138                                 struct btrfs_block_group_cache *block_group)
139 {
140         struct rb_node **p;
141         struct rb_node *parent = NULL;
142         struct btrfs_block_group_cache *cache;
143
144         spin_lock(&info->block_group_cache_lock);
145         p = &info->block_group_cache_tree.rb_node;
146
147         while (*p) {
148                 parent = *p;
149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
150                                  cache_node);
151                 if (block_group->key.objectid < cache->key.objectid) {
152                         p = &(*p)->rb_left;
153                 } else if (block_group->key.objectid > cache->key.objectid) {
154                         p = &(*p)->rb_right;
155                 } else {
156                         spin_unlock(&info->block_group_cache_lock);
157                         return -EEXIST;
158                 }
159         }
160
161         rb_link_node(&block_group->cache_node, parent, p);
162         rb_insert_color(&block_group->cache_node,
163                         &info->block_group_cache_tree);
164         spin_unlock(&info->block_group_cache_lock);
165
166         return 0;
167 }
168
169 /*
170  * This will return the block group at or after bytenr if contains is 0, else
171  * it will return the block group that contains the bytenr
172  */
173 static struct btrfs_block_group_cache *
174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
175                               int contains)
176 {
177         struct btrfs_block_group_cache *cache, *ret = NULL;
178         struct rb_node *n;
179         u64 end, start;
180
181         spin_lock(&info->block_group_cache_lock);
182         n = info->block_group_cache_tree.rb_node;
183
184         while (n) {
185                 cache = rb_entry(n, struct btrfs_block_group_cache,
186                                  cache_node);
187                 end = cache->key.objectid + cache->key.offset - 1;
188                 start = cache->key.objectid;
189
190                 if (bytenr < start) {
191                         if (!contains && (!ret || start < ret->key.objectid))
192                                 ret = cache;
193                         n = n->rb_left;
194                 } else if (bytenr > start) {
195                         if (contains && bytenr <= end) {
196                                 ret = cache;
197                                 break;
198                         }
199                         n = n->rb_right;
200                 } else {
201                         ret = cache;
202                         break;
203                 }
204         }
205         if (ret)
206                 btrfs_get_block_group(ret);
207         spin_unlock(&info->block_group_cache_lock);
208
209         return ret;
210 }
211
212 static int add_excluded_extent(struct btrfs_root *root,
213                                u64 start, u64 num_bytes)
214 {
215         u64 end = start + num_bytes - 1;
216         set_extent_bits(&root->fs_info->freed_extents[0],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         set_extent_bits(&root->fs_info->freed_extents[1],
219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
220         return 0;
221 }
222
223 static void free_excluded_extents(struct btrfs_root *root,
224                                   struct btrfs_block_group_cache *cache)
225 {
226         u64 start, end;
227
228         start = cache->key.objectid;
229         end = start + cache->key.offset - 1;
230
231         clear_extent_bits(&root->fs_info->freed_extents[0],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233         clear_extent_bits(&root->fs_info->freed_extents[1],
234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
235 }
236
237 static int exclude_super_stripes(struct btrfs_root *root,
238                                  struct btrfs_block_group_cache *cache)
239 {
240         u64 bytenr;
241         u64 *logical;
242         int stripe_len;
243         int i, nr, ret;
244
245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
247                 cache->bytes_super += stripe_len;
248                 ret = add_excluded_extent(root, cache->key.objectid,
249                                           stripe_len);
250                 BUG_ON(ret); /* -ENOMEM */
251         }
252
253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
254                 bytenr = btrfs_sb_offset(i);
255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
256                                        cache->key.objectid, bytenr,
257                                        0, &logical, &nr, &stripe_len);
258                 BUG_ON(ret); /* -ENOMEM */
259
260                 while (nr--) {
261                         cache->bytes_super += stripe_len;
262                         ret = add_excluded_extent(root, logical[nr],
263                                                   stripe_len);
264                         BUG_ON(ret); /* -ENOMEM */
265                 }
266
267                 kfree(logical);
268         }
269         return 0;
270 }
271
272 static struct btrfs_caching_control *
273 get_caching_control(struct btrfs_block_group_cache *cache)
274 {
275         struct btrfs_caching_control *ctl;
276
277         spin_lock(&cache->lock);
278         if (cache->cached != BTRFS_CACHE_STARTED) {
279                 spin_unlock(&cache->lock);
280                 return NULL;
281         }
282
283         /* We're loading it the fast way, so we don't have a caching_ctl. */
284         if (!cache->caching_ctl) {
285                 spin_unlock(&cache->lock);
286                 return NULL;
287         }
288
289         ctl = cache->caching_ctl;
290         atomic_inc(&ctl->count);
291         spin_unlock(&cache->lock);
292         return ctl;
293 }
294
295 static void put_caching_control(struct btrfs_caching_control *ctl)
296 {
297         if (atomic_dec_and_test(&ctl->count))
298                 kfree(ctl);
299 }
300
301 /*
302  * this is only called by cache_block_group, since we could have freed extents
303  * we need to check the pinned_extents for any extents that can't be used yet
304  * since their free space will be released as soon as the transaction commits.
305  */
306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
307                               struct btrfs_fs_info *info, u64 start, u64 end)
308 {
309         u64 extent_start, extent_end, size, total_added = 0;
310         int ret;
311
312         while (start < end) {
313                 ret = find_first_extent_bit(info->pinned_extents, start,
314                                             &extent_start, &extent_end,
315                                             EXTENT_DIRTY | EXTENT_UPTODATE);
316                 if (ret)
317                         break;
318
319                 if (extent_start <= start) {
320                         start = extent_end + 1;
321                 } else if (extent_start > start && extent_start < end) {
322                         size = extent_start - start;
323                         total_added += size;
324                         ret = btrfs_add_free_space(block_group, start,
325                                                    size);
326                         BUG_ON(ret); /* -ENOMEM or logic error */
327                         start = extent_end + 1;
328                 } else {
329                         break;
330                 }
331         }
332
333         if (start < end) {
334                 size = end - start;
335                 total_added += size;
336                 ret = btrfs_add_free_space(block_group, start, size);
337                 BUG_ON(ret); /* -ENOMEM or logic error */
338         }
339
340         return total_added;
341 }
342
343 static noinline void caching_thread(struct btrfs_work *work)
344 {
345         struct btrfs_block_group_cache *block_group;
346         struct btrfs_fs_info *fs_info;
347         struct btrfs_caching_control *caching_ctl;
348         struct btrfs_root *extent_root;
349         struct btrfs_path *path;
350         struct extent_buffer *leaf;
351         struct btrfs_key key;
352         u64 total_found = 0;
353         u64 last = 0;
354         u32 nritems;
355         int ret = 0;
356
357         caching_ctl = container_of(work, struct btrfs_caching_control, work);
358         block_group = caching_ctl->block_group;
359         fs_info = block_group->fs_info;
360         extent_root = fs_info->extent_root;
361
362         path = btrfs_alloc_path();
363         if (!path)
364                 goto out;
365
366         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
367
368         /*
369          * We don't want to deadlock with somebody trying to allocate a new
370          * extent for the extent root while also trying to search the extent
371          * root to add free space.  So we skip locking and search the commit
372          * root, since its read-only
373          */
374         path->skip_locking = 1;
375         path->search_commit_root = 1;
376         path->reada = 1;
377
378         key.objectid = last;
379         key.offset = 0;
380         key.type = BTRFS_EXTENT_ITEM_KEY;
381 again:
382         mutex_lock(&caching_ctl->mutex);
383         /* need to make sure the commit_root doesn't disappear */
384         down_read(&fs_info->extent_commit_sem);
385
386         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
387         if (ret < 0)
388                 goto err;
389
390         leaf = path->nodes[0];
391         nritems = btrfs_header_nritems(leaf);
392
393         while (1) {
394                 if (btrfs_fs_closing(fs_info) > 1) {
395                         last = (u64)-1;
396                         break;
397                 }
398
399                 if (path->slots[0] < nritems) {
400                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
401                 } else {
402                         ret = find_next_key(path, 0, &key);
403                         if (ret)
404                                 break;
405
406                         if (need_resched() ||
407                             btrfs_next_leaf(extent_root, path)) {
408                                 caching_ctl->progress = last;
409                                 btrfs_release_path(path);
410                                 up_read(&fs_info->extent_commit_sem);
411                                 mutex_unlock(&caching_ctl->mutex);
412                                 cond_resched();
413                                 goto again;
414                         }
415                         leaf = path->nodes[0];
416                         nritems = btrfs_header_nritems(leaf);
417                         continue;
418                 }
419
420                 if (key.objectid < block_group->key.objectid) {
421                         path->slots[0]++;
422                         continue;
423                 }
424
425                 if (key.objectid >= block_group->key.objectid +
426                     block_group->key.offset)
427                         break;
428
429                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
430                         total_found += add_new_free_space(block_group,
431                                                           fs_info, last,
432                                                           key.objectid);
433                         last = key.objectid + key.offset;
434
435                         if (total_found > (1024 * 1024 * 2)) {
436                                 total_found = 0;
437                                 wake_up(&caching_ctl->wait);
438                         }
439                 }
440                 path->slots[0]++;
441         }
442         ret = 0;
443
444         total_found += add_new_free_space(block_group, fs_info, last,
445                                           block_group->key.objectid +
446                                           block_group->key.offset);
447         caching_ctl->progress = (u64)-1;
448
449         spin_lock(&block_group->lock);
450         block_group->caching_ctl = NULL;
451         block_group->cached = BTRFS_CACHE_FINISHED;
452         spin_unlock(&block_group->lock);
453
454 err:
455         btrfs_free_path(path);
456         up_read(&fs_info->extent_commit_sem);
457
458         free_excluded_extents(extent_root, block_group);
459
460         mutex_unlock(&caching_ctl->mutex);
461 out:
462         wake_up(&caching_ctl->wait);
463
464         put_caching_control(caching_ctl);
465         btrfs_put_block_group(block_group);
466 }
467
468 static int cache_block_group(struct btrfs_block_group_cache *cache,
469                              struct btrfs_trans_handle *trans,
470                              struct btrfs_root *root,
471                              int load_cache_only)
472 {
473         DEFINE_WAIT(wait);
474         struct btrfs_fs_info *fs_info = cache->fs_info;
475         struct btrfs_caching_control *caching_ctl;
476         int ret = 0;
477
478         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
479         if (!caching_ctl)
480                 return -ENOMEM;
481
482         INIT_LIST_HEAD(&caching_ctl->list);
483         mutex_init(&caching_ctl->mutex);
484         init_waitqueue_head(&caching_ctl->wait);
485         caching_ctl->block_group = cache;
486         caching_ctl->progress = cache->key.objectid;
487         atomic_set(&caching_ctl->count, 1);
488         caching_ctl->work.func = caching_thread;
489
490         spin_lock(&cache->lock);
491         /*
492          * This should be a rare occasion, but this could happen I think in the
493          * case where one thread starts to load the space cache info, and then
494          * some other thread starts a transaction commit which tries to do an
495          * allocation while the other thread is still loading the space cache
496          * info.  The previous loop should have kept us from choosing this block
497          * group, but if we've moved to the state where we will wait on caching
498          * block groups we need to first check if we're doing a fast load here,
499          * so we can wait for it to finish, otherwise we could end up allocating
500          * from a block group who's cache gets evicted for one reason or
501          * another.
502          */
503         while (cache->cached == BTRFS_CACHE_FAST) {
504                 struct btrfs_caching_control *ctl;
505
506                 ctl = cache->caching_ctl;
507                 atomic_inc(&ctl->count);
508                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
509                 spin_unlock(&cache->lock);
510
511                 schedule();
512
513                 finish_wait(&ctl->wait, &wait);
514                 put_caching_control(ctl);
515                 spin_lock(&cache->lock);
516         }
517
518         if (cache->cached != BTRFS_CACHE_NO) {
519                 spin_unlock(&cache->lock);
520                 kfree(caching_ctl);
521                 return 0;
522         }
523         WARN_ON(cache->caching_ctl);
524         cache->caching_ctl = caching_ctl;
525         cache->cached = BTRFS_CACHE_FAST;
526         spin_unlock(&cache->lock);
527
528         /*
529          * We can't do the read from on-disk cache during a commit since we need
530          * to have the normal tree locking.  Also if we are currently trying to
531          * allocate blocks for the tree root we can't do the fast caching since
532          * we likely hold important locks.
533          */
534         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
535                 ret = load_free_space_cache(fs_info, cache);
536
537                 spin_lock(&cache->lock);
538                 if (ret == 1) {
539                         cache->caching_ctl = NULL;
540                         cache->cached = BTRFS_CACHE_FINISHED;
541                         cache->last_byte_to_unpin = (u64)-1;
542                 } else {
543                         if (load_cache_only) {
544                                 cache->caching_ctl = NULL;
545                                 cache->cached = BTRFS_CACHE_NO;
546                         } else {
547                                 cache->cached = BTRFS_CACHE_STARTED;
548                         }
549                 }
550                 spin_unlock(&cache->lock);
551                 wake_up(&caching_ctl->wait);
552                 if (ret == 1) {
553                         put_caching_control(caching_ctl);
554                         free_excluded_extents(fs_info->extent_root, cache);
555                         return 0;
556                 }
557         } else {
558                 /*
559                  * We are not going to do the fast caching, set cached to the
560                  * appropriate value and wakeup any waiters.
561                  */
562                 spin_lock(&cache->lock);
563                 if (load_cache_only) {
564                         cache->caching_ctl = NULL;
565                         cache->cached = BTRFS_CACHE_NO;
566                 } else {
567                         cache->cached = BTRFS_CACHE_STARTED;
568                 }
569                 spin_unlock(&cache->lock);
570                 wake_up(&caching_ctl->wait);
571         }
572
573         if (load_cache_only) {
574                 put_caching_control(caching_ctl);
575                 return 0;
576         }
577
578         down_write(&fs_info->extent_commit_sem);
579         atomic_inc(&caching_ctl->count);
580         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
581         up_write(&fs_info->extent_commit_sem);
582
583         btrfs_get_block_group(cache);
584
585         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
586
587         return ret;
588 }
589
590 /*
591  * return the block group that starts at or after bytenr
592  */
593 static struct btrfs_block_group_cache *
594 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
595 {
596         struct btrfs_block_group_cache *cache;
597
598         cache = block_group_cache_tree_search(info, bytenr, 0);
599
600         return cache;
601 }
602
603 /*
604  * return the block group that contains the given bytenr
605  */
606 struct btrfs_block_group_cache *btrfs_lookup_block_group(
607                                                  struct btrfs_fs_info *info,
608                                                  u64 bytenr)
609 {
610         struct btrfs_block_group_cache *cache;
611
612         cache = block_group_cache_tree_search(info, bytenr, 1);
613
614         return cache;
615 }
616
617 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
618                                                   u64 flags)
619 {
620         struct list_head *head = &info->space_info;
621         struct btrfs_space_info *found;
622
623         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
624
625         rcu_read_lock();
626         list_for_each_entry_rcu(found, head, list) {
627                 if (found->flags & flags) {
628                         rcu_read_unlock();
629                         return found;
630                 }
631         }
632         rcu_read_unlock();
633         return NULL;
634 }
635
636 /*
637  * after adding space to the filesystem, we need to clear the full flags
638  * on all the space infos.
639  */
640 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
641 {
642         struct list_head *head = &info->space_info;
643         struct btrfs_space_info *found;
644
645         rcu_read_lock();
646         list_for_each_entry_rcu(found, head, list)
647                 found->full = 0;
648         rcu_read_unlock();
649 }
650
651 static u64 div_factor(u64 num, int factor)
652 {
653         if (factor == 10)
654                 return num;
655         num *= factor;
656         do_div(num, 10);
657         return num;
658 }
659
660 static u64 div_factor_fine(u64 num, int factor)
661 {
662         if (factor == 100)
663                 return num;
664         num *= factor;
665         do_div(num, 100);
666         return num;
667 }
668
669 u64 btrfs_find_block_group(struct btrfs_root *root,
670                            u64 search_start, u64 search_hint, int owner)
671 {
672         struct btrfs_block_group_cache *cache;
673         u64 used;
674         u64 last = max(search_hint, search_start);
675         u64 group_start = 0;
676         int full_search = 0;
677         int factor = 9;
678         int wrapped = 0;
679 again:
680         while (1) {
681                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
682                 if (!cache)
683                         break;
684
685                 spin_lock(&cache->lock);
686                 last = cache->key.objectid + cache->key.offset;
687                 used = btrfs_block_group_used(&cache->item);
688
689                 if ((full_search || !cache->ro) &&
690                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
691                         if (used + cache->pinned + cache->reserved <
692                             div_factor(cache->key.offset, factor)) {
693                                 group_start = cache->key.objectid;
694                                 spin_unlock(&cache->lock);
695                                 btrfs_put_block_group(cache);
696                                 goto found;
697                         }
698                 }
699                 spin_unlock(&cache->lock);
700                 btrfs_put_block_group(cache);
701                 cond_resched();
702         }
703         if (!wrapped) {
704                 last = search_start;
705                 wrapped = 1;
706                 goto again;
707         }
708         if (!full_search && factor < 10) {
709                 last = search_start;
710                 full_search = 1;
711                 factor = 10;
712                 goto again;
713         }
714 found:
715         return group_start;
716 }
717
718 /* simple helper to search for an existing extent at a given offset */
719 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
720 {
721         int ret;
722         struct btrfs_key key;
723         struct btrfs_path *path;
724
725         path = btrfs_alloc_path();
726         if (!path)
727                 return -ENOMEM;
728
729         key.objectid = start;
730         key.offset = len;
731         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
732         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
733                                 0, 0);
734         btrfs_free_path(path);
735         return ret;
736 }
737
738 /*
739  * helper function to lookup reference count and flags of extent.
740  *
741  * the head node for delayed ref is used to store the sum of all the
742  * reference count modifications queued up in the rbtree. the head
743  * node may also store the extent flags to set. This way you can check
744  * to see what the reference count and extent flags would be if all of
745  * the delayed refs are not processed.
746  */
747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
748                              struct btrfs_root *root, u64 bytenr,
749                              u64 num_bytes, u64 *refs, u64 *flags)
750 {
751         struct btrfs_delayed_ref_head *head;
752         struct btrfs_delayed_ref_root *delayed_refs;
753         struct btrfs_path *path;
754         struct btrfs_extent_item *ei;
755         struct extent_buffer *leaf;
756         struct btrfs_key key;
757         u32 item_size;
758         u64 num_refs;
759         u64 extent_flags;
760         int ret;
761
762         path = btrfs_alloc_path();
763         if (!path)
764                 return -ENOMEM;
765
766         key.objectid = bytenr;
767         key.type = BTRFS_EXTENT_ITEM_KEY;
768         key.offset = num_bytes;
769         if (!trans) {
770                 path->skip_locking = 1;
771                 path->search_commit_root = 1;
772         }
773 again:
774         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
775                                 &key, path, 0, 0);
776         if (ret < 0)
777                 goto out_free;
778
779         if (ret == 0) {
780                 leaf = path->nodes[0];
781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
782                 if (item_size >= sizeof(*ei)) {
783                         ei = btrfs_item_ptr(leaf, path->slots[0],
784                                             struct btrfs_extent_item);
785                         num_refs = btrfs_extent_refs(leaf, ei);
786                         extent_flags = btrfs_extent_flags(leaf, ei);
787                 } else {
788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
789                         struct btrfs_extent_item_v0 *ei0;
790                         BUG_ON(item_size != sizeof(*ei0));
791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
792                                              struct btrfs_extent_item_v0);
793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
794                         /* FIXME: this isn't correct for data */
795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
796 #else
797                         BUG();
798 #endif
799                 }
800                 BUG_ON(num_refs == 0);
801         } else {
802                 num_refs = 0;
803                 extent_flags = 0;
804                 ret = 0;
805         }
806
807         if (!trans)
808                 goto out;
809
810         delayed_refs = &trans->transaction->delayed_refs;
811         spin_lock(&delayed_refs->lock);
812         head = btrfs_find_delayed_ref_head(trans, bytenr);
813         if (head) {
814                 if (!mutex_trylock(&head->mutex)) {
815                         atomic_inc(&head->node.refs);
816                         spin_unlock(&delayed_refs->lock);
817
818                         btrfs_release_path(path);
819
820                         /*
821                          * Mutex was contended, block until it's released and try
822                          * again
823                          */
824                         mutex_lock(&head->mutex);
825                         mutex_unlock(&head->mutex);
826                         btrfs_put_delayed_ref(&head->node);
827                         goto again;
828                 }
829                 if (head->extent_op && head->extent_op->update_flags)
830                         extent_flags |= head->extent_op->flags_to_set;
831                 else
832                         BUG_ON(num_refs == 0);
833
834                 num_refs += head->node.ref_mod;
835                 mutex_unlock(&head->mutex);
836         }
837         spin_unlock(&delayed_refs->lock);
838 out:
839         WARN_ON(num_refs == 0);
840         if (refs)
841                 *refs = num_refs;
842         if (flags)
843                 *flags = extent_flags;
844 out_free:
845         btrfs_free_path(path);
846         return ret;
847 }
848
849 /*
850  * Back reference rules.  Back refs have three main goals:
851  *
852  * 1) differentiate between all holders of references to an extent so that
853  *    when a reference is dropped we can make sure it was a valid reference
854  *    before freeing the extent.
855  *
856  * 2) Provide enough information to quickly find the holders of an extent
857  *    if we notice a given block is corrupted or bad.
858  *
859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
860  *    maintenance.  This is actually the same as #2, but with a slightly
861  *    different use case.
862  *
863  * There are two kinds of back refs. The implicit back refs is optimized
864  * for pointers in non-shared tree blocks. For a given pointer in a block,
865  * back refs of this kind provide information about the block's owner tree
866  * and the pointer's key. These information allow us to find the block by
867  * b-tree searching. The full back refs is for pointers in tree blocks not
868  * referenced by their owner trees. The location of tree block is recorded
869  * in the back refs. Actually the full back refs is generic, and can be
870  * used in all cases the implicit back refs is used. The major shortcoming
871  * of the full back refs is its overhead. Every time a tree block gets
872  * COWed, we have to update back refs entry for all pointers in it.
873  *
874  * For a newly allocated tree block, we use implicit back refs for
875  * pointers in it. This means most tree related operations only involve
876  * implicit back refs. For a tree block created in old transaction, the
877  * only way to drop a reference to it is COW it. So we can detect the
878  * event that tree block loses its owner tree's reference and do the
879  * back refs conversion.
880  *
881  * When a tree block is COW'd through a tree, there are four cases:
882  *
883  * The reference count of the block is one and the tree is the block's
884  * owner tree. Nothing to do in this case.
885  *
886  * The reference count of the block is one and the tree is not the
887  * block's owner tree. In this case, full back refs is used for pointers
888  * in the block. Remove these full back refs, add implicit back refs for
889  * every pointers in the new block.
890  *
891  * The reference count of the block is greater than one and the tree is
892  * the block's owner tree. In this case, implicit back refs is used for
893  * pointers in the block. Add full back refs for every pointers in the
894  * block, increase lower level extents' reference counts. The original
895  * implicit back refs are entailed to the new block.
896  *
897  * The reference count of the block is greater than one and the tree is
898  * not the block's owner tree. Add implicit back refs for every pointer in
899  * the new block, increase lower level extents' reference count.
900  *
901  * Back Reference Key composing:
902  *
903  * The key objectid corresponds to the first byte in the extent,
904  * The key type is used to differentiate between types of back refs.
905  * There are different meanings of the key offset for different types
906  * of back refs.
907  *
908  * File extents can be referenced by:
909  *
910  * - multiple snapshots, subvolumes, or different generations in one subvol
911  * - different files inside a single subvolume
912  * - different offsets inside a file (bookend extents in file.c)
913  *
914  * The extent ref structure for the implicit back refs has fields for:
915  *
916  * - Objectid of the subvolume root
917  * - objectid of the file holding the reference
918  * - original offset in the file
919  * - how many bookend extents
920  *
921  * The key offset for the implicit back refs is hash of the first
922  * three fields.
923  *
924  * The extent ref structure for the full back refs has field for:
925  *
926  * - number of pointers in the tree leaf
927  *
928  * The key offset for the implicit back refs is the first byte of
929  * the tree leaf
930  *
931  * When a file extent is allocated, The implicit back refs is used.
932  * the fields are filled in:
933  *
934  *     (root_key.objectid, inode objectid, offset in file, 1)
935  *
936  * When a file extent is removed file truncation, we find the
937  * corresponding implicit back refs and check the following fields:
938  *
939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
940  *
941  * Btree extents can be referenced by:
942  *
943  * - Different subvolumes
944  *
945  * Both the implicit back refs and the full back refs for tree blocks
946  * only consist of key. The key offset for the implicit back refs is
947  * objectid of block's owner tree. The key offset for the full back refs
948  * is the first byte of parent block.
949  *
950  * When implicit back refs is used, information about the lowest key and
951  * level of the tree block are required. These information are stored in
952  * tree block info structure.
953  */
954
955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
957                                   struct btrfs_root *root,
958                                   struct btrfs_path *path,
959                                   u64 owner, u32 extra_size)
960 {
961         struct btrfs_extent_item *item;
962         struct btrfs_extent_item_v0 *ei0;
963         struct btrfs_extent_ref_v0 *ref0;
964         struct btrfs_tree_block_info *bi;
965         struct extent_buffer *leaf;
966         struct btrfs_key key;
967         struct btrfs_key found_key;
968         u32 new_size = sizeof(*item);
969         u64 refs;
970         int ret;
971
972         leaf = path->nodes[0];
973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
974
975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
977                              struct btrfs_extent_item_v0);
978         refs = btrfs_extent_refs_v0(leaf, ei0);
979
980         if (owner == (u64)-1) {
981                 while (1) {
982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
983                                 ret = btrfs_next_leaf(root, path);
984                                 if (ret < 0)
985                                         return ret;
986                                 BUG_ON(ret > 0); /* Corruption */
987                                 leaf = path->nodes[0];
988                         }
989                         btrfs_item_key_to_cpu(leaf, &found_key,
990                                               path->slots[0]);
991                         BUG_ON(key.objectid != found_key.objectid);
992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
993                                 path->slots[0]++;
994                                 continue;
995                         }
996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
997                                               struct btrfs_extent_ref_v0);
998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(trans, root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466
1467         key.objectid = bytenr;
1468         key.type = BTRFS_EXTENT_ITEM_KEY;
1469         key.offset = num_bytes;
1470
1471         want = extent_ref_type(parent, owner);
1472         if (insert) {
1473                 extra_size = btrfs_extent_inline_ref_size(want);
1474                 path->keep_locks = 1;
1475         } else
1476                 extra_size = -1;
1477         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1478         if (ret < 0) {
1479                 err = ret;
1480                 goto out;
1481         }
1482         if (ret && !insert) {
1483                 err = -ENOENT;
1484                 goto out;
1485         }
1486         BUG_ON(ret); /* Corruption */
1487
1488         leaf = path->nodes[0];
1489         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1490 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1491         if (item_size < sizeof(*ei)) {
1492                 if (!insert) {
1493                         err = -ENOENT;
1494                         goto out;
1495                 }
1496                 ret = convert_extent_item_v0(trans, root, path, owner,
1497                                              extra_size);
1498                 if (ret < 0) {
1499                         err = ret;
1500                         goto out;
1501                 }
1502                 leaf = path->nodes[0];
1503                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1504         }
1505 #endif
1506         BUG_ON(item_size < sizeof(*ei));
1507
1508         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1509         flags = btrfs_extent_flags(leaf, ei);
1510
1511         ptr = (unsigned long)(ei + 1);
1512         end = (unsigned long)ei + item_size;
1513
1514         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1515                 ptr += sizeof(struct btrfs_tree_block_info);
1516                 BUG_ON(ptr > end);
1517         } else {
1518                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1519         }
1520
1521         err = -ENOENT;
1522         while (1) {
1523                 if (ptr >= end) {
1524                         WARN_ON(ptr > end);
1525                         break;
1526                 }
1527                 iref = (struct btrfs_extent_inline_ref *)ptr;
1528                 type = btrfs_extent_inline_ref_type(leaf, iref);
1529                 if (want < type)
1530                         break;
1531                 if (want > type) {
1532                         ptr += btrfs_extent_inline_ref_size(type);
1533                         continue;
1534                 }
1535
1536                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1537                         struct btrfs_extent_data_ref *dref;
1538                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1539                         if (match_extent_data_ref(leaf, dref, root_objectid,
1540                                                   owner, offset)) {
1541                                 err = 0;
1542                                 break;
1543                         }
1544                         if (hash_extent_data_ref_item(leaf, dref) <
1545                             hash_extent_data_ref(root_objectid, owner, offset))
1546                                 break;
1547                 } else {
1548                         u64 ref_offset;
1549                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1550                         if (parent > 0) {
1551                                 if (parent == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < parent)
1556                                         break;
1557                         } else {
1558                                 if (root_objectid == ref_offset) {
1559                                         err = 0;
1560                                         break;
1561                                 }
1562                                 if (ref_offset < root_objectid)
1563                                         break;
1564                         }
1565                 }
1566                 ptr += btrfs_extent_inline_ref_size(type);
1567         }
1568         if (err == -ENOENT && insert) {
1569                 if (item_size + extra_size >=
1570                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574                 /*
1575                  * To add new inline back ref, we have to make sure
1576                  * there is no corresponding back ref item.
1577                  * For simplicity, we just do not add new inline back
1578                  * ref if there is any kind of item for this block
1579                  */
1580                 if (find_next_key(path, 0, &key) == 0 &&
1581                     key.objectid == bytenr &&
1582                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1583                         err = -EAGAIN;
1584                         goto out;
1585                 }
1586         }
1587         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1588 out:
1589         if (insert) {
1590                 path->keep_locks = 0;
1591                 btrfs_unlock_up_safe(path, 1);
1592         }
1593         return err;
1594 }
1595
1596 /*
1597  * helper to add new inline back ref
1598  */
1599 static noinline_for_stack
1600 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_root *root,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref *iref,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int refs_to_add,
1606                                  struct btrfs_delayed_extent_op *extent_op)
1607 {
1608         struct extent_buffer *leaf;
1609         struct btrfs_extent_item *ei;
1610         unsigned long ptr;
1611         unsigned long end;
1612         unsigned long item_offset;
1613         u64 refs;
1614         int size;
1615         int type;
1616
1617         leaf = path->nodes[0];
1618         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1619         item_offset = (unsigned long)iref - (unsigned long)ei;
1620
1621         type = extent_ref_type(parent, owner);
1622         size = btrfs_extent_inline_ref_size(type);
1623
1624         btrfs_extend_item(trans, root, path, size);
1625
1626         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1627         refs = btrfs_extent_refs(leaf, ei);
1628         refs += refs_to_add;
1629         btrfs_set_extent_refs(leaf, ei, refs);
1630         if (extent_op)
1631                 __run_delayed_extent_op(extent_op, leaf, ei);
1632
1633         ptr = (unsigned long)ei + item_offset;
1634         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1635         if (ptr < end - size)
1636                 memmove_extent_buffer(leaf, ptr + size, ptr,
1637                                       end - size - ptr);
1638
1639         iref = (struct btrfs_extent_inline_ref *)ptr;
1640         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1641         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1642                 struct btrfs_extent_data_ref *dref;
1643                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1644                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1645                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1646                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1647                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1648         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1649                 struct btrfs_shared_data_ref *sref;
1650                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1651                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1655         } else {
1656                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1657         }
1658         btrfs_mark_buffer_dirty(leaf);
1659 }
1660
1661 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1662                                  struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref **ref_ret,
1665                                  u64 bytenr, u64 num_bytes, u64 parent,
1666                                  u64 root_objectid, u64 owner, u64 offset)
1667 {
1668         int ret;
1669
1670         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1671                                            bytenr, num_bytes, parent,
1672                                            root_objectid, owner, offset, 0);
1673         if (ret != -ENOENT)
1674                 return ret;
1675
1676         btrfs_release_path(path);
1677         *ref_ret = NULL;
1678
1679         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1680                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1681                                             root_objectid);
1682         } else {
1683                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1684                                              root_objectid, owner, offset);
1685         }
1686         return ret;
1687 }
1688
1689 /*
1690  * helper to update/remove inline back ref
1691  */
1692 static noinline_for_stack
1693 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1694                                   struct btrfs_root *root,
1695                                   struct btrfs_path *path,
1696                                   struct btrfs_extent_inline_ref *iref,
1697                                   int refs_to_mod,
1698                                   struct btrfs_delayed_extent_op *extent_op)
1699 {
1700         struct extent_buffer *leaf;
1701         struct btrfs_extent_item *ei;
1702         struct btrfs_extent_data_ref *dref = NULL;
1703         struct btrfs_shared_data_ref *sref = NULL;
1704         unsigned long ptr;
1705         unsigned long end;
1706         u32 item_size;
1707         int size;
1708         int type;
1709         u64 refs;
1710
1711         leaf = path->nodes[0];
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1715         refs += refs_to_mod;
1716         btrfs_set_extent_refs(leaf, ei, refs);
1717         if (extent_op)
1718                 __run_delayed_extent_op(extent_op, leaf, ei);
1719
1720         type = btrfs_extent_inline_ref_type(leaf, iref);
1721
1722         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1723                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1724                 refs = btrfs_extent_data_ref_count(leaf, dref);
1725         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1726                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1727                 refs = btrfs_shared_data_ref_count(leaf, sref);
1728         } else {
1729                 refs = 1;
1730                 BUG_ON(refs_to_mod != -1);
1731         }
1732
1733         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1734         refs += refs_to_mod;
1735
1736         if (refs > 0) {
1737                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1738                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1739                 else
1740                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1741         } else {
1742                 size =  btrfs_extent_inline_ref_size(type);
1743                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1744                 ptr = (unsigned long)iref;
1745                 end = (unsigned long)ei + item_size;
1746                 if (ptr + size < end)
1747                         memmove_extent_buffer(leaf, ptr, ptr + size,
1748                                               end - ptr - size);
1749                 item_size -= size;
1750                 btrfs_truncate_item(trans, root, path, item_size, 1);
1751         }
1752         btrfs_mark_buffer_dirty(leaf);
1753 }
1754
1755 static noinline_for_stack
1756 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1757                                  struct btrfs_root *root,
1758                                  struct btrfs_path *path,
1759                                  u64 bytenr, u64 num_bytes, u64 parent,
1760                                  u64 root_objectid, u64 owner,
1761                                  u64 offset, int refs_to_add,
1762                                  struct btrfs_delayed_extent_op *extent_op)
1763 {
1764         struct btrfs_extent_inline_ref *iref;
1765         int ret;
1766
1767         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1768                                            bytenr, num_bytes, parent,
1769                                            root_objectid, owner, offset, 1);
1770         if (ret == 0) {
1771                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1772                 update_inline_extent_backref(trans, root, path, iref,
1773                                              refs_to_add, extent_op);
1774         } else if (ret == -ENOENT) {
1775                 setup_inline_extent_backref(trans, root, path, iref, parent,
1776                                             root_objectid, owner, offset,
1777                                             refs_to_add, extent_op);
1778                 ret = 0;
1779         }
1780         return ret;
1781 }
1782
1783 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1784                                  struct btrfs_root *root,
1785                                  struct btrfs_path *path,
1786                                  u64 bytenr, u64 parent, u64 root_objectid,
1787                                  u64 owner, u64 offset, int refs_to_add)
1788 {
1789         int ret;
1790         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1791                 BUG_ON(refs_to_add != 1);
1792                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1793                                             parent, root_objectid);
1794         } else {
1795                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1796                                              parent, root_objectid,
1797                                              owner, offset, refs_to_add);
1798         }
1799         return ret;
1800 }
1801
1802 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  int refs_to_drop, int is_data)
1807 {
1808         int ret = 0;
1809
1810         BUG_ON(!is_data && refs_to_drop != 1);
1811         if (iref) {
1812                 update_inline_extent_backref(trans, root, path, iref,
1813                                              -refs_to_drop, NULL);
1814         } else if (is_data) {
1815                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1816         } else {
1817                 ret = btrfs_del_item(trans, root, path);
1818         }
1819         return ret;
1820 }
1821
1822 static int btrfs_issue_discard(struct block_device *bdev,
1823                                 u64 start, u64 len)
1824 {
1825         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1826 }
1827
1828 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1829                                 u64 num_bytes, u64 *actual_bytes)
1830 {
1831         int ret;
1832         u64 discarded_bytes = 0;
1833         struct btrfs_bio *bbio = NULL;
1834
1835
1836         /* Tell the block device(s) that the sectors can be discarded */
1837         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1838                               bytenr, &num_bytes, &bbio, 0);
1839         /* Error condition is -ENOMEM */
1840         if (!ret) {
1841                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1842                 int i;
1843
1844
1845                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1846                         if (!stripe->dev->can_discard)
1847                                 continue;
1848
1849                         ret = btrfs_issue_discard(stripe->dev->bdev,
1850                                                   stripe->physical,
1851                                                   stripe->length);
1852                         if (!ret)
1853                                 discarded_bytes += stripe->length;
1854                         else if (ret != -EOPNOTSUPP)
1855                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1856
1857                         /*
1858                          * Just in case we get back EOPNOTSUPP for some reason,
1859                          * just ignore the return value so we don't screw up
1860                          * people calling discard_extent.
1861                          */
1862                         ret = 0;
1863                 }
1864                 kfree(bbio);
1865         }
1866
1867         if (actual_bytes)
1868                 *actual_bytes = discarded_bytes;
1869
1870
1871         return ret;
1872 }
1873
1874 /* Can return -ENOMEM */
1875 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          u64 bytenr, u64 num_bytes, u64 parent,
1878                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1879 {
1880         int ret;
1881         struct btrfs_fs_info *fs_info = root->fs_info;
1882
1883         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1884                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1885
1886         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, (int)owner,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         } else {
1892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1893                                         num_bytes,
1894                                         parent, root_objectid, owner, offset,
1895                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1896         }
1897         return ret;
1898 }
1899
1900 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1901                                   struct btrfs_root *root,
1902                                   u64 bytenr, u64 num_bytes,
1903                                   u64 parent, u64 root_objectid,
1904                                   u64 owner, u64 offset, int refs_to_add,
1905                                   struct btrfs_delayed_extent_op *extent_op)
1906 {
1907         struct btrfs_path *path;
1908         struct extent_buffer *leaf;
1909         struct btrfs_extent_item *item;
1910         u64 refs;
1911         int ret;
1912         int err = 0;
1913
1914         path = btrfs_alloc_path();
1915         if (!path)
1916                 return -ENOMEM;
1917
1918         path->reada = 1;
1919         path->leave_spinning = 1;
1920         /* this will setup the path even if it fails to insert the back ref */
1921         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1922                                            path, bytenr, num_bytes, parent,
1923                                            root_objectid, owner, offset,
1924                                            refs_to_add, extent_op);
1925         if (ret == 0)
1926                 goto out;
1927
1928         if (ret != -EAGAIN) {
1929                 err = ret;
1930                 goto out;
1931         }
1932
1933         leaf = path->nodes[0];
1934         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1935         refs = btrfs_extent_refs(leaf, item);
1936         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1937         if (extent_op)
1938                 __run_delayed_extent_op(extent_op, leaf, item);
1939
1940         btrfs_mark_buffer_dirty(leaf);
1941         btrfs_release_path(path);
1942
1943         path->reada = 1;
1944         path->leave_spinning = 1;
1945
1946         /* now insert the actual backref */
1947         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1948                                     path, bytenr, parent, root_objectid,
1949                                     owner, offset, refs_to_add);
1950         if (ret)
1951                 btrfs_abort_transaction(trans, root, ret);
1952 out:
1953         btrfs_free_path(path);
1954         return err;
1955 }
1956
1957 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1958                                 struct btrfs_root *root,
1959                                 struct btrfs_delayed_ref_node *node,
1960                                 struct btrfs_delayed_extent_op *extent_op,
1961                                 int insert_reserved)
1962 {
1963         int ret = 0;
1964         struct btrfs_delayed_data_ref *ref;
1965         struct btrfs_key ins;
1966         u64 parent = 0;
1967         u64 ref_root = 0;
1968         u64 flags = 0;
1969
1970         ins.objectid = node->bytenr;
1971         ins.offset = node->num_bytes;
1972         ins.type = BTRFS_EXTENT_ITEM_KEY;
1973
1974         ref = btrfs_delayed_node_to_data_ref(node);
1975         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1976                 parent = ref->parent;
1977         else
1978                 ref_root = ref->root;
1979
1980         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1981                 if (extent_op) {
1982                         BUG_ON(extent_op->update_key);
1983                         flags |= extent_op->flags_to_set;
1984                 }
1985                 ret = alloc_reserved_file_extent(trans, root,
1986                                                  parent, ref_root, flags,
1987                                                  ref->objectid, ref->offset,
1988                                                  &ins, node->ref_mod);
1989         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1990                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1991                                              node->num_bytes, parent,
1992                                              ref_root, ref->objectid,
1993                                              ref->offset, node->ref_mod,
1994                                              extent_op);
1995         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1996                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1997                                           node->num_bytes, parent,
1998                                           ref_root, ref->objectid,
1999                                           ref->offset, node->ref_mod,
2000                                           extent_op);
2001         } else {
2002                 BUG();
2003         }
2004         return ret;
2005 }
2006
2007 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2008                                     struct extent_buffer *leaf,
2009                                     struct btrfs_extent_item *ei)
2010 {
2011         u64 flags = btrfs_extent_flags(leaf, ei);
2012         if (extent_op->update_flags) {
2013                 flags |= extent_op->flags_to_set;
2014                 btrfs_set_extent_flags(leaf, ei, flags);
2015         }
2016
2017         if (extent_op->update_key) {
2018                 struct btrfs_tree_block_info *bi;
2019                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2020                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2021                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2022         }
2023 }
2024
2025 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2026                                  struct btrfs_root *root,
2027                                  struct btrfs_delayed_ref_node *node,
2028                                  struct btrfs_delayed_extent_op *extent_op)
2029 {
2030         struct btrfs_key key;
2031         struct btrfs_path *path;
2032         struct btrfs_extent_item *ei;
2033         struct extent_buffer *leaf;
2034         u32 item_size;
2035         int ret;
2036         int err = 0;
2037
2038         if (trans->aborted)
2039                 return 0;
2040
2041         path = btrfs_alloc_path();
2042         if (!path)
2043                 return -ENOMEM;
2044
2045         key.objectid = node->bytenr;
2046         key.type = BTRFS_EXTENT_ITEM_KEY;
2047         key.offset = node->num_bytes;
2048
2049         path->reada = 1;
2050         path->leave_spinning = 1;
2051         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2052                                 path, 0, 1);
2053         if (ret < 0) {
2054                 err = ret;
2055                 goto out;
2056         }
2057         if (ret > 0) {
2058                 err = -EIO;
2059                 goto out;
2060         }
2061
2062         leaf = path->nodes[0];
2063         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2065         if (item_size < sizeof(*ei)) {
2066                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2067                                              path, (u64)-1, 0);
2068                 if (ret < 0) {
2069                         err = ret;
2070                         goto out;
2071                 }
2072                 leaf = path->nodes[0];
2073                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2074         }
2075 #endif
2076         BUG_ON(item_size < sizeof(*ei));
2077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         __run_delayed_extent_op(extent_op, leaf, ei);
2079
2080         btrfs_mark_buffer_dirty(leaf);
2081 out:
2082         btrfs_free_path(path);
2083         return err;
2084 }
2085
2086 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_root *root,
2088                                 struct btrfs_delayed_ref_node *node,
2089                                 struct btrfs_delayed_extent_op *extent_op,
2090                                 int insert_reserved)
2091 {
2092         int ret = 0;
2093         struct btrfs_delayed_tree_ref *ref;
2094         struct btrfs_key ins;
2095         u64 parent = 0;
2096         u64 ref_root = 0;
2097
2098         ins.objectid = node->bytenr;
2099         ins.offset = node->num_bytes;
2100         ins.type = BTRFS_EXTENT_ITEM_KEY;
2101
2102         ref = btrfs_delayed_node_to_tree_ref(node);
2103         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2104                 parent = ref->parent;
2105         else
2106                 ref_root = ref->root;
2107
2108         BUG_ON(node->ref_mod != 1);
2109         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2110                 BUG_ON(!extent_op || !extent_op->update_flags ||
2111                        !extent_op->update_key);
2112                 ret = alloc_reserved_tree_block(trans, root,
2113                                                 parent, ref_root,
2114                                                 extent_op->flags_to_set,
2115                                                 &extent_op->key,
2116                                                 ref->level, &ins);
2117         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2118                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2119                                              node->num_bytes, parent, ref_root,
2120                                              ref->level, 0, 1, extent_op);
2121         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2122                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2123                                           node->num_bytes, parent, ref_root,
2124                                           ref->level, 0, 1, extent_op);
2125         } else {
2126                 BUG();
2127         }
2128         return ret;
2129 }
2130
2131 /* helper function to actually process a single delayed ref entry */
2132 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2133                                struct btrfs_root *root,
2134                                struct btrfs_delayed_ref_node *node,
2135                                struct btrfs_delayed_extent_op *extent_op,
2136                                int insert_reserved)
2137 {
2138         int ret = 0;
2139
2140         if (trans->aborted)
2141                 return 0;
2142
2143         if (btrfs_delayed_ref_is_head(node)) {
2144                 struct btrfs_delayed_ref_head *head;
2145                 /*
2146                  * we've hit the end of the chain and we were supposed
2147                  * to insert this extent into the tree.  But, it got
2148                  * deleted before we ever needed to insert it, so all
2149                  * we have to do is clean up the accounting
2150                  */
2151                 BUG_ON(extent_op);
2152                 head = btrfs_delayed_node_to_head(node);
2153                 if (insert_reserved) {
2154                         btrfs_pin_extent(root, node->bytenr,
2155                                          node->num_bytes, 1);
2156                         if (head->is_data) {
2157                                 ret = btrfs_del_csums(trans, root,
2158                                                       node->bytenr,
2159                                                       node->num_bytes);
2160                         }
2161                 }
2162                 mutex_unlock(&head->mutex);
2163                 return ret;
2164         }
2165
2166         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2167             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2168                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2169                                            insert_reserved);
2170         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2171                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2172                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2173                                            insert_reserved);
2174         else
2175                 BUG();
2176         return ret;
2177 }
2178
2179 static noinline struct btrfs_delayed_ref_node *
2180 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2181 {
2182         struct rb_node *node;
2183         struct btrfs_delayed_ref_node *ref;
2184         int action = BTRFS_ADD_DELAYED_REF;
2185 again:
2186         /*
2187          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2188          * this prevents ref count from going down to zero when
2189          * there still are pending delayed ref.
2190          */
2191         node = rb_prev(&head->node.rb_node);
2192         while (1) {
2193                 if (!node)
2194                         break;
2195                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2196                                 rb_node);
2197                 if (ref->bytenr != head->node.bytenr)
2198                         break;
2199                 if (ref->action == action)
2200                         return ref;
2201                 node = rb_prev(node);
2202         }
2203         if (action == BTRFS_ADD_DELAYED_REF) {
2204                 action = BTRFS_DROP_DELAYED_REF;
2205                 goto again;
2206         }
2207         return NULL;
2208 }
2209
2210 /*
2211  * Returns 0 on success or if called with an already aborted transaction.
2212  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2213  */
2214 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2215                                        struct btrfs_root *root,
2216                                        struct list_head *cluster)
2217 {
2218         struct btrfs_delayed_ref_root *delayed_refs;
2219         struct btrfs_delayed_ref_node *ref;
2220         struct btrfs_delayed_ref_head *locked_ref = NULL;
2221         struct btrfs_delayed_extent_op *extent_op;
2222         struct btrfs_fs_info *fs_info = root->fs_info;
2223         int ret;
2224         int count = 0;
2225         int must_insert_reserved = 0;
2226
2227         delayed_refs = &trans->transaction->delayed_refs;
2228         while (1) {
2229                 if (!locked_ref) {
2230                         /* pick a new head ref from the cluster list */
2231                         if (list_empty(cluster))
2232                                 break;
2233
2234                         locked_ref = list_entry(cluster->next,
2235                                      struct btrfs_delayed_ref_head, cluster);
2236
2237                         /* grab the lock that says we are going to process
2238                          * all the refs for this head */
2239                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2240
2241                         /*
2242                          * we may have dropped the spin lock to get the head
2243                          * mutex lock, and that might have given someone else
2244                          * time to free the head.  If that's true, it has been
2245                          * removed from our list and we can move on.
2246                          */
2247                         if (ret == -EAGAIN) {
2248                                 locked_ref = NULL;
2249                                 count++;
2250                                 continue;
2251                         }
2252                 }
2253
2254                 /*
2255                  * locked_ref is the head node, so we have to go one
2256                  * node back for any delayed ref updates
2257                  */
2258                 ref = select_delayed_ref(locked_ref);
2259
2260                 if (ref && ref->seq &&
2261                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2262                         /*
2263                          * there are still refs with lower seq numbers in the
2264                          * process of being added. Don't run this ref yet.
2265                          */
2266                         list_del_init(&locked_ref->cluster);
2267                         mutex_unlock(&locked_ref->mutex);
2268                         locked_ref = NULL;
2269                         delayed_refs->num_heads_ready++;
2270                         spin_unlock(&delayed_refs->lock);
2271                         cond_resched();
2272                         spin_lock(&delayed_refs->lock);
2273                         continue;
2274                 }
2275
2276                 /*
2277                  * record the must insert reserved flag before we
2278                  * drop the spin lock.
2279                  */
2280                 must_insert_reserved = locked_ref->must_insert_reserved;
2281                 locked_ref->must_insert_reserved = 0;
2282
2283                 extent_op = locked_ref->extent_op;
2284                 locked_ref->extent_op = NULL;
2285
2286                 if (!ref) {
2287                         /* All delayed refs have been processed, Go ahead
2288                          * and send the head node to run_one_delayed_ref,
2289                          * so that any accounting fixes can happen
2290                          */
2291                         ref = &locked_ref->node;
2292
2293                         if (extent_op && must_insert_reserved) {
2294                                 kfree(extent_op);
2295                                 extent_op = NULL;
2296                         }
2297
2298                         if (extent_op) {
2299                                 spin_unlock(&delayed_refs->lock);
2300
2301                                 ret = run_delayed_extent_op(trans, root,
2302                                                             ref, extent_op);
2303                                 kfree(extent_op);
2304
2305                                 if (ret) {
2306                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2307                                         spin_lock(&delayed_refs->lock);
2308                                         return ret;
2309                                 }
2310
2311                                 goto next;
2312                         }
2313
2314                         list_del_init(&locked_ref->cluster);
2315                         locked_ref = NULL;
2316                 }
2317
2318                 ref->in_tree = 0;
2319                 rb_erase(&ref->rb_node, &delayed_refs->root);
2320                 delayed_refs->num_entries--;
2321                 if (locked_ref) {
2322                         /*
2323                          * when we play the delayed ref, also correct the
2324                          * ref_mod on head
2325                          */
2326                         switch (ref->action) {
2327                         case BTRFS_ADD_DELAYED_REF:
2328                         case BTRFS_ADD_DELAYED_EXTENT:
2329                                 locked_ref->node.ref_mod -= ref->ref_mod;
2330                                 break;
2331                         case BTRFS_DROP_DELAYED_REF:
2332                                 locked_ref->node.ref_mod += ref->ref_mod;
2333                                 break;
2334                         default:
2335                                 WARN_ON(1);
2336                         }
2337                 }
2338                 spin_unlock(&delayed_refs->lock);
2339
2340                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2341                                           must_insert_reserved);
2342
2343                 btrfs_put_delayed_ref(ref);
2344                 kfree(extent_op);
2345                 count++;
2346
2347                 if (ret) {
2348                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2349                         spin_lock(&delayed_refs->lock);
2350                         return ret;
2351                 }
2352
2353 next:
2354                 do_chunk_alloc(trans, fs_info->extent_root,
2355                                2 * 1024 * 1024,
2356                                btrfs_get_alloc_profile(root, 0),
2357                                CHUNK_ALLOC_NO_FORCE);
2358                 cond_resched();
2359                 spin_lock(&delayed_refs->lock);
2360         }
2361         return count;
2362 }
2363
2364 #ifdef SCRAMBLE_DELAYED_REFS
2365 /*
2366  * Normally delayed refs get processed in ascending bytenr order. This
2367  * correlates in most cases to the order added. To expose dependencies on this
2368  * order, we start to process the tree in the middle instead of the beginning
2369  */
2370 static u64 find_middle(struct rb_root *root)
2371 {
2372         struct rb_node *n = root->rb_node;
2373         struct btrfs_delayed_ref_node *entry;
2374         int alt = 1;
2375         u64 middle;
2376         u64 first = 0, last = 0;
2377
2378         n = rb_first(root);
2379         if (n) {
2380                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2381                 first = entry->bytenr;
2382         }
2383         n = rb_last(root);
2384         if (n) {
2385                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2386                 last = entry->bytenr;
2387         }
2388         n = root->rb_node;
2389
2390         while (n) {
2391                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2392                 WARN_ON(!entry->in_tree);
2393
2394                 middle = entry->bytenr;
2395
2396                 if (alt)
2397                         n = n->rb_left;
2398                 else
2399                         n = n->rb_right;
2400
2401                 alt = 1 - alt;
2402         }
2403         return middle;
2404 }
2405 #endif
2406
2407 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2408                                          struct btrfs_fs_info *fs_info)
2409 {
2410         struct qgroup_update *qgroup_update;
2411         int ret = 0;
2412
2413         if (list_empty(&trans->qgroup_ref_list) !=
2414             !trans->delayed_ref_elem.seq) {
2415                 /* list without seq or seq without list */
2416                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2417                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2418                         trans->delayed_ref_elem.seq);
2419                 BUG();
2420         }
2421
2422         if (!trans->delayed_ref_elem.seq)
2423                 return 0;
2424
2425         while (!list_empty(&trans->qgroup_ref_list)) {
2426                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2427                                                  struct qgroup_update, list);
2428                 list_del(&qgroup_update->list);
2429                 if (!ret)
2430                         ret = btrfs_qgroup_account_ref(
2431                                         trans, fs_info, qgroup_update->node,
2432                                         qgroup_update->extent_op);
2433                 kfree(qgroup_update);
2434         }
2435
2436         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2437
2438         return ret;
2439 }
2440
2441 /*
2442  * this starts processing the delayed reference count updates and
2443  * extent insertions we have queued up so far.  count can be
2444  * 0, which means to process everything in the tree at the start
2445  * of the run (but not newly added entries), or it can be some target
2446  * number you'd like to process.
2447  *
2448  * Returns 0 on success or if called with an aborted transaction
2449  * Returns <0 on error and aborts the transaction
2450  */
2451 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2452                            struct btrfs_root *root, unsigned long count)
2453 {
2454         struct rb_node *node;
2455         struct btrfs_delayed_ref_root *delayed_refs;
2456         struct btrfs_delayed_ref_node *ref;
2457         struct list_head cluster;
2458         int ret;
2459         u64 delayed_start;
2460         int run_all = count == (unsigned long)-1;
2461         int run_most = 0;
2462         int loops;
2463
2464         /* We'll clean this up in btrfs_cleanup_transaction */
2465         if (trans->aborted)
2466                 return 0;
2467
2468         if (root == root->fs_info->extent_root)
2469                 root = root->fs_info->tree_root;
2470
2471         do_chunk_alloc(trans, root->fs_info->extent_root,
2472                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2473                        CHUNK_ALLOC_NO_FORCE);
2474
2475         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2476
2477         delayed_refs = &trans->transaction->delayed_refs;
2478         INIT_LIST_HEAD(&cluster);
2479 again:
2480         loops = 0;
2481         spin_lock(&delayed_refs->lock);
2482
2483 #ifdef SCRAMBLE_DELAYED_REFS
2484         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2485 #endif
2486
2487         if (count == 0) {
2488                 count = delayed_refs->num_entries * 2;
2489                 run_most = 1;
2490         }
2491         while (1) {
2492                 if (!(run_all || run_most) &&
2493                     delayed_refs->num_heads_ready < 64)
2494                         break;
2495
2496                 /*
2497                  * go find something we can process in the rbtree.  We start at
2498                  * the beginning of the tree, and then build a cluster
2499                  * of refs to process starting at the first one we are able to
2500                  * lock
2501                  */
2502                 delayed_start = delayed_refs->run_delayed_start;
2503                 ret = btrfs_find_ref_cluster(trans, &cluster,
2504                                              delayed_refs->run_delayed_start);
2505                 if (ret)
2506                         break;
2507
2508                 ret = run_clustered_refs(trans, root, &cluster);
2509                 if (ret < 0) {
2510                         spin_unlock(&delayed_refs->lock);
2511                         btrfs_abort_transaction(trans, root, ret);
2512                         return ret;
2513                 }
2514
2515                 count -= min_t(unsigned long, ret, count);
2516
2517                 if (count == 0)
2518                         break;
2519
2520                 if (delayed_start >= delayed_refs->run_delayed_start) {
2521                         if (loops == 0) {
2522                                 /*
2523                                  * btrfs_find_ref_cluster looped. let's do one
2524                                  * more cycle. if we don't run any delayed ref
2525                                  * during that cycle (because we can't because
2526                                  * all of them are blocked), bail out.
2527                                  */
2528                                 loops = 1;
2529                         } else {
2530                                 /*
2531                                  * no runnable refs left, stop trying
2532                                  */
2533                                 BUG_ON(run_all);
2534                                 break;
2535                         }
2536                 }
2537                 if (ret) {
2538                         /* refs were run, let's reset staleness detection */
2539                         loops = 0;
2540                 }
2541         }
2542
2543         if (run_all) {
2544                 node = rb_first(&delayed_refs->root);
2545                 if (!node)
2546                         goto out;
2547                 count = (unsigned long)-1;
2548
2549                 while (node) {
2550                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2551                                        rb_node);
2552                         if (btrfs_delayed_ref_is_head(ref)) {
2553                                 struct btrfs_delayed_ref_head *head;
2554
2555                                 head = btrfs_delayed_node_to_head(ref);
2556                                 atomic_inc(&ref->refs);
2557
2558                                 spin_unlock(&delayed_refs->lock);
2559                                 /*
2560                                  * Mutex was contended, block until it's
2561                                  * released and try again
2562                                  */
2563                                 mutex_lock(&head->mutex);
2564                                 mutex_unlock(&head->mutex);
2565
2566                                 btrfs_put_delayed_ref(ref);
2567                                 cond_resched();
2568                                 goto again;
2569                         }
2570                         node = rb_next(node);
2571                 }
2572                 spin_unlock(&delayed_refs->lock);
2573                 schedule_timeout(1);
2574                 goto again;
2575         }
2576 out:
2577         spin_unlock(&delayed_refs->lock);
2578         assert_qgroups_uptodate(trans);
2579         return 0;
2580 }
2581
2582 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2583                                 struct btrfs_root *root,
2584                                 u64 bytenr, u64 num_bytes, u64 flags,
2585                                 int is_data)
2586 {
2587         struct btrfs_delayed_extent_op *extent_op;
2588         int ret;
2589
2590         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2591         if (!extent_op)
2592                 return -ENOMEM;
2593
2594         extent_op->flags_to_set = flags;
2595         extent_op->update_flags = 1;
2596         extent_op->update_key = 0;
2597         extent_op->is_data = is_data ? 1 : 0;
2598
2599         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2600                                           num_bytes, extent_op);
2601         if (ret)
2602                 kfree(extent_op);
2603         return ret;
2604 }
2605
2606 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2607                                       struct btrfs_root *root,
2608                                       struct btrfs_path *path,
2609                                       u64 objectid, u64 offset, u64 bytenr)
2610 {
2611         struct btrfs_delayed_ref_head *head;
2612         struct btrfs_delayed_ref_node *ref;
2613         struct btrfs_delayed_data_ref *data_ref;
2614         struct btrfs_delayed_ref_root *delayed_refs;
2615         struct rb_node *node;
2616         int ret = 0;
2617
2618         ret = -ENOENT;
2619         delayed_refs = &trans->transaction->delayed_refs;
2620         spin_lock(&delayed_refs->lock);
2621         head = btrfs_find_delayed_ref_head(trans, bytenr);
2622         if (!head)
2623                 goto out;
2624
2625         if (!mutex_trylock(&head->mutex)) {
2626                 atomic_inc(&head->node.refs);
2627                 spin_unlock(&delayed_refs->lock);
2628
2629                 btrfs_release_path(path);
2630
2631                 /*
2632                  * Mutex was contended, block until it's released and let
2633                  * caller try again
2634                  */
2635                 mutex_lock(&head->mutex);
2636                 mutex_unlock(&head->mutex);
2637                 btrfs_put_delayed_ref(&head->node);
2638                 return -EAGAIN;
2639         }
2640
2641         node = rb_prev(&head->node.rb_node);
2642         if (!node)
2643                 goto out_unlock;
2644
2645         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2646
2647         if (ref->bytenr != bytenr)
2648                 goto out_unlock;
2649
2650         ret = 1;
2651         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2652                 goto out_unlock;
2653
2654         data_ref = btrfs_delayed_node_to_data_ref(ref);
2655
2656         node = rb_prev(node);
2657         if (node) {
2658                 int seq = ref->seq;
2659
2660                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2661                 if (ref->bytenr == bytenr && ref->seq == seq)
2662                         goto out_unlock;
2663         }
2664
2665         if (data_ref->root != root->root_key.objectid ||
2666             data_ref->objectid != objectid || data_ref->offset != offset)
2667                 goto out_unlock;
2668
2669         ret = 0;
2670 out_unlock:
2671         mutex_unlock(&head->mutex);
2672 out:
2673         spin_unlock(&delayed_refs->lock);
2674         return ret;
2675 }
2676
2677 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2678                                         struct btrfs_root *root,
2679                                         struct btrfs_path *path,
2680                                         u64 objectid, u64 offset, u64 bytenr)
2681 {
2682         struct btrfs_root *extent_root = root->fs_info->extent_root;
2683         struct extent_buffer *leaf;
2684         struct btrfs_extent_data_ref *ref;
2685         struct btrfs_extent_inline_ref *iref;
2686         struct btrfs_extent_item *ei;
2687         struct btrfs_key key;
2688         u32 item_size;
2689         int ret;
2690
2691         key.objectid = bytenr;
2692         key.offset = (u64)-1;
2693         key.type = BTRFS_EXTENT_ITEM_KEY;
2694
2695         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2696         if (ret < 0)
2697                 goto out;
2698         BUG_ON(ret == 0); /* Corruption */
2699
2700         ret = -ENOENT;
2701         if (path->slots[0] == 0)
2702                 goto out;
2703
2704         path->slots[0]--;
2705         leaf = path->nodes[0];
2706         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2707
2708         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2709                 goto out;
2710
2711         ret = 1;
2712         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2713 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2714         if (item_size < sizeof(*ei)) {
2715                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2716                 goto out;
2717         }
2718 #endif
2719         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2720
2721         if (item_size != sizeof(*ei) +
2722             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2723                 goto out;
2724
2725         if (btrfs_extent_generation(leaf, ei) <=
2726             btrfs_root_last_snapshot(&root->root_item))
2727                 goto out;
2728
2729         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2730         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2731             BTRFS_EXTENT_DATA_REF_KEY)
2732                 goto out;
2733
2734         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2735         if (btrfs_extent_refs(leaf, ei) !=
2736             btrfs_extent_data_ref_count(leaf, ref) ||
2737             btrfs_extent_data_ref_root(leaf, ref) !=
2738             root->root_key.objectid ||
2739             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2740             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2741                 goto out;
2742
2743         ret = 0;
2744 out:
2745         return ret;
2746 }
2747
2748 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2749                           struct btrfs_root *root,
2750                           u64 objectid, u64 offset, u64 bytenr)
2751 {
2752         struct btrfs_path *path;
2753         int ret;
2754         int ret2;
2755
2756         path = btrfs_alloc_path();
2757         if (!path)
2758                 return -ENOENT;
2759
2760         do {
2761                 ret = check_committed_ref(trans, root, path, objectid,
2762                                           offset, bytenr);
2763                 if (ret && ret != -ENOENT)
2764                         goto out;
2765
2766                 ret2 = check_delayed_ref(trans, root, path, objectid,
2767                                          offset, bytenr);
2768         } while (ret2 == -EAGAIN);
2769
2770         if (ret2 && ret2 != -ENOENT) {
2771                 ret = ret2;
2772                 goto out;
2773         }
2774
2775         if (ret != -ENOENT || ret2 != -ENOENT)
2776                 ret = 0;
2777 out:
2778         btrfs_free_path(path);
2779         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2780                 WARN_ON(ret > 0);
2781         return ret;
2782 }
2783
2784 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2785                            struct btrfs_root *root,
2786                            struct extent_buffer *buf,
2787                            int full_backref, int inc, int for_cow)
2788 {
2789         u64 bytenr;
2790         u64 num_bytes;
2791         u64 parent;
2792         u64 ref_root;
2793         u32 nritems;
2794         struct btrfs_key key;
2795         struct btrfs_file_extent_item *fi;
2796         int i;
2797         int level;
2798         int ret = 0;
2799         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2800                             u64, u64, u64, u64, u64, u64, int);
2801
2802         ref_root = btrfs_header_owner(buf);
2803         nritems = btrfs_header_nritems(buf);
2804         level = btrfs_header_level(buf);
2805
2806         if (!root->ref_cows && level == 0)
2807                 return 0;
2808
2809         if (inc)
2810                 process_func = btrfs_inc_extent_ref;
2811         else
2812                 process_func = btrfs_free_extent;
2813
2814         if (full_backref)
2815                 parent = buf->start;
2816         else
2817                 parent = 0;
2818
2819         for (i = 0; i < nritems; i++) {
2820                 if (level == 0) {
2821                         btrfs_item_key_to_cpu(buf, &key, i);
2822                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2823                                 continue;
2824                         fi = btrfs_item_ptr(buf, i,
2825                                             struct btrfs_file_extent_item);
2826                         if (btrfs_file_extent_type(buf, fi) ==
2827                             BTRFS_FILE_EXTENT_INLINE)
2828                                 continue;
2829                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2830                         if (bytenr == 0)
2831                                 continue;
2832
2833                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2834                         key.offset -= btrfs_file_extent_offset(buf, fi);
2835                         ret = process_func(trans, root, bytenr, num_bytes,
2836                                            parent, ref_root, key.objectid,
2837                                            key.offset, for_cow);
2838                         if (ret)
2839                                 goto fail;
2840                 } else {
2841                         bytenr = btrfs_node_blockptr(buf, i);
2842                         num_bytes = btrfs_level_size(root, level - 1);
2843                         ret = process_func(trans, root, bytenr, num_bytes,
2844                                            parent, ref_root, level - 1, 0,
2845                                            for_cow);
2846                         if (ret)
2847                                 goto fail;
2848                 }
2849         }
2850         return 0;
2851 fail:
2852         return ret;
2853 }
2854
2855 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2856                   struct extent_buffer *buf, int full_backref, int for_cow)
2857 {
2858         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2859 }
2860
2861 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2862                   struct extent_buffer *buf, int full_backref, int for_cow)
2863 {
2864         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2865 }
2866
2867 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2868                                  struct btrfs_root *root,
2869                                  struct btrfs_path *path,
2870                                  struct btrfs_block_group_cache *cache)
2871 {
2872         int ret;
2873         struct btrfs_root *extent_root = root->fs_info->extent_root;
2874         unsigned long bi;
2875         struct extent_buffer *leaf;
2876
2877         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2878         if (ret < 0)
2879                 goto fail;
2880         BUG_ON(ret); /* Corruption */
2881
2882         leaf = path->nodes[0];
2883         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2884         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2885         btrfs_mark_buffer_dirty(leaf);
2886         btrfs_release_path(path);
2887 fail:
2888         if (ret) {
2889                 btrfs_abort_transaction(trans, root, ret);
2890                 return ret;
2891         }
2892         return 0;
2893
2894 }
2895
2896 static struct btrfs_block_group_cache *
2897 next_block_group(struct btrfs_root *root,
2898                  struct btrfs_block_group_cache *cache)
2899 {
2900         struct rb_node *node;
2901         spin_lock(&root->fs_info->block_group_cache_lock);
2902         node = rb_next(&cache->cache_node);
2903         btrfs_put_block_group(cache);
2904         if (node) {
2905                 cache = rb_entry(node, struct btrfs_block_group_cache,
2906                                  cache_node);
2907                 btrfs_get_block_group(cache);
2908         } else
2909                 cache = NULL;
2910         spin_unlock(&root->fs_info->block_group_cache_lock);
2911         return cache;
2912 }
2913
2914 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2915                             struct btrfs_trans_handle *trans,
2916                             struct btrfs_path *path)
2917 {
2918         struct btrfs_root *root = block_group->fs_info->tree_root;
2919         struct inode *inode = NULL;
2920         u64 alloc_hint = 0;
2921         int dcs = BTRFS_DC_ERROR;
2922         int num_pages = 0;
2923         int retries = 0;
2924         int ret = 0;
2925
2926         /*
2927          * If this block group is smaller than 100 megs don't bother caching the
2928          * block group.
2929          */
2930         if (block_group->key.offset < (100 * 1024 * 1024)) {
2931                 spin_lock(&block_group->lock);
2932                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2933                 spin_unlock(&block_group->lock);
2934                 return 0;
2935         }
2936
2937 again:
2938         inode = lookup_free_space_inode(root, block_group, path);
2939         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2940                 ret = PTR_ERR(inode);
2941                 btrfs_release_path(path);
2942                 goto out;
2943         }
2944
2945         if (IS_ERR(inode)) {
2946                 BUG_ON(retries);
2947                 retries++;
2948
2949                 if (block_group->ro)
2950                         goto out_free;
2951
2952                 ret = create_free_space_inode(root, trans, block_group, path);
2953                 if (ret)
2954                         goto out_free;
2955                 goto again;
2956         }
2957
2958         /* We've already setup this transaction, go ahead and exit */
2959         if (block_group->cache_generation == trans->transid &&
2960             i_size_read(inode)) {
2961                 dcs = BTRFS_DC_SETUP;
2962                 goto out_put;
2963         }
2964
2965         /*
2966          * We want to set the generation to 0, that way if anything goes wrong
2967          * from here on out we know not to trust this cache when we load up next
2968          * time.
2969          */
2970         BTRFS_I(inode)->generation = 0;
2971         ret = btrfs_update_inode(trans, root, inode);
2972         WARN_ON(ret);
2973
2974         if (i_size_read(inode) > 0) {
2975                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2976                                                       inode);
2977                 if (ret)
2978                         goto out_put;
2979         }
2980
2981         spin_lock(&block_group->lock);
2982         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2983             !btrfs_test_opt(root, SPACE_CACHE)) {
2984                 /*
2985                  * don't bother trying to write stuff out _if_
2986                  * a) we're not cached,
2987                  * b) we're with nospace_cache mount option.
2988                  */
2989                 dcs = BTRFS_DC_WRITTEN;
2990                 spin_unlock(&block_group->lock);
2991                 goto out_put;
2992         }
2993         spin_unlock(&block_group->lock);
2994
2995         /*
2996          * Try to preallocate enough space based on how big the block group is.
2997          * Keep in mind this has to include any pinned space which could end up
2998          * taking up quite a bit since it's not folded into the other space
2999          * cache.
3000          */
3001         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3002         if (!num_pages)
3003                 num_pages = 1;
3004
3005         num_pages *= 16;
3006         num_pages *= PAGE_CACHE_SIZE;
3007
3008         ret = btrfs_check_data_free_space(inode, num_pages);
3009         if (ret)
3010                 goto out_put;
3011
3012         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3013                                               num_pages, num_pages,
3014                                               &alloc_hint);
3015         if (!ret)
3016                 dcs = BTRFS_DC_SETUP;
3017         btrfs_free_reserved_data_space(inode, num_pages);
3018
3019 out_put:
3020         iput(inode);
3021 out_free:
3022         btrfs_release_path(path);
3023 out:
3024         spin_lock(&block_group->lock);
3025         if (!ret && dcs == BTRFS_DC_SETUP)
3026                 block_group->cache_generation = trans->transid;
3027         block_group->disk_cache_state = dcs;
3028         spin_unlock(&block_group->lock);
3029
3030         return ret;
3031 }
3032
3033 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3034                                    struct btrfs_root *root)
3035 {
3036         struct btrfs_block_group_cache *cache;
3037         int err = 0;
3038         struct btrfs_path *path;
3039         u64 last = 0;
3040
3041         path = btrfs_alloc_path();
3042         if (!path)
3043                 return -ENOMEM;
3044
3045 again:
3046         while (1) {
3047                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3048                 while (cache) {
3049                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3050                                 break;
3051                         cache = next_block_group(root, cache);
3052                 }
3053                 if (!cache) {
3054                         if (last == 0)
3055                                 break;
3056                         last = 0;
3057                         continue;
3058                 }
3059                 err = cache_save_setup(cache, trans, path);
3060                 last = cache->key.objectid + cache->key.offset;
3061                 btrfs_put_block_group(cache);
3062         }
3063
3064         while (1) {
3065                 if (last == 0) {
3066                         err = btrfs_run_delayed_refs(trans, root,
3067                                                      (unsigned long)-1);
3068                         if (err) /* File system offline */
3069                                 goto out;
3070                 }
3071
3072                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3073                 while (cache) {
3074                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3075                                 btrfs_put_block_group(cache);
3076                                 goto again;
3077                         }
3078
3079                         if (cache->dirty)
3080                                 break;
3081                         cache = next_block_group(root, cache);
3082                 }
3083                 if (!cache) {
3084                         if (last == 0)
3085                                 break;
3086                         last = 0;
3087                         continue;
3088                 }
3089
3090                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3091                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3092                 cache->dirty = 0;
3093                 last = cache->key.objectid + cache->key.offset;
3094
3095                 err = write_one_cache_group(trans, root, path, cache);
3096                 if (err) /* File system offline */
3097                         goto out;
3098
3099                 btrfs_put_block_group(cache);
3100         }
3101
3102         while (1) {
3103                 /*
3104                  * I don't think this is needed since we're just marking our
3105                  * preallocated extent as written, but just in case it can't
3106                  * hurt.
3107                  */
3108                 if (last == 0) {
3109                         err = btrfs_run_delayed_refs(trans, root,
3110                                                      (unsigned long)-1);
3111                         if (err) /* File system offline */
3112                                 goto out;
3113                 }
3114
3115                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3116                 while (cache) {
3117                         /*
3118                          * Really this shouldn't happen, but it could if we
3119                          * couldn't write the entire preallocated extent and
3120                          * splitting the extent resulted in a new block.
3121                          */
3122                         if (cache->dirty) {
3123                                 btrfs_put_block_group(cache);
3124                                 goto again;
3125                         }
3126                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3127                                 break;
3128                         cache = next_block_group(root, cache);
3129                 }
3130                 if (!cache) {
3131                         if (last == 0)
3132                                 break;
3133                         last = 0;
3134                         continue;
3135                 }
3136
3137                 err = btrfs_write_out_cache(root, trans, cache, path);
3138
3139                 /*
3140                  * If we didn't have an error then the cache state is still
3141                  * NEED_WRITE, so we can set it to WRITTEN.
3142                  */
3143                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3144                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3145                 last = cache->key.objectid + cache->key.offset;
3146                 btrfs_put_block_group(cache);
3147         }
3148 out:
3149
3150         btrfs_free_path(path);
3151         return err;
3152 }
3153
3154 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3155 {
3156         struct btrfs_block_group_cache *block_group;
3157         int readonly = 0;
3158
3159         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3160         if (!block_group || block_group->ro)
3161                 readonly = 1;
3162         if (block_group)
3163                 btrfs_put_block_group(block_group);
3164         return readonly;
3165 }
3166
3167 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3168                              u64 total_bytes, u64 bytes_used,
3169                              struct btrfs_space_info **space_info)
3170 {
3171         struct btrfs_space_info *found;
3172         int i;
3173         int factor;
3174
3175         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3176                      BTRFS_BLOCK_GROUP_RAID10))
3177                 factor = 2;
3178         else
3179                 factor = 1;
3180
3181         found = __find_space_info(info, flags);
3182         if (found) {
3183                 spin_lock(&found->lock);
3184                 found->total_bytes += total_bytes;
3185                 found->disk_total += total_bytes * factor;
3186                 found->bytes_used += bytes_used;
3187                 found->disk_used += bytes_used * factor;
3188                 found->full = 0;
3189                 spin_unlock(&found->lock);
3190                 *space_info = found;
3191                 return 0;
3192         }
3193         found = kzalloc(sizeof(*found), GFP_NOFS);
3194         if (!found)
3195                 return -ENOMEM;
3196
3197         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3198                 INIT_LIST_HEAD(&found->block_groups[i]);
3199         init_rwsem(&found->groups_sem);
3200         spin_lock_init(&found->lock);
3201         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3202         found->total_bytes = total_bytes;
3203         found->disk_total = total_bytes * factor;
3204         found->bytes_used = bytes_used;
3205         found->disk_used = bytes_used * factor;
3206         found->bytes_pinned = 0;
3207         found->bytes_reserved = 0;
3208         found->bytes_readonly = 0;
3209         found->bytes_may_use = 0;
3210         found->full = 0;
3211         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3212         found->chunk_alloc = 0;
3213         found->flush = 0;
3214         init_waitqueue_head(&found->wait);
3215         *space_info = found;
3216         list_add_rcu(&found->list, &info->space_info);
3217         if (flags & BTRFS_BLOCK_GROUP_DATA)
3218                 info->data_sinfo = found;
3219         return 0;
3220 }
3221
3222 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3223 {
3224         u64 extra_flags = chunk_to_extended(flags) &
3225                                 BTRFS_EXTENDED_PROFILE_MASK;
3226
3227         if (flags & BTRFS_BLOCK_GROUP_DATA)
3228                 fs_info->avail_data_alloc_bits |= extra_flags;
3229         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3230                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3231         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3232                 fs_info->avail_system_alloc_bits |= extra_flags;
3233 }
3234
3235 /*
3236  * returns target flags in extended format or 0 if restripe for this
3237  * chunk_type is not in progress
3238  *
3239  * should be called with either volume_mutex or balance_lock held
3240  */
3241 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3242 {
3243         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3244         u64 target = 0;
3245
3246         if (!bctl)
3247                 return 0;
3248
3249         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3250             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3251                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3252         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3253                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3254                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3255         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3256                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3257                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3258         }
3259
3260         return target;
3261 }
3262
3263 /*
3264  * @flags: available profiles in extended format (see ctree.h)
3265  *
3266  * Returns reduced profile in chunk format.  If profile changing is in
3267  * progress (either running or paused) picks the target profile (if it's
3268  * already available), otherwise falls back to plain reducing.
3269  */
3270 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3271 {
3272         /*
3273          * we add in the count of missing devices because we want
3274          * to make sure that any RAID levels on a degraded FS
3275          * continue to be honored.
3276          */
3277         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3278                 root->fs_info->fs_devices->missing_devices;
3279         u64 target;
3280
3281         /*
3282          * see if restripe for this chunk_type is in progress, if so
3283          * try to reduce to the target profile
3284          */
3285         spin_lock(&root->fs_info->balance_lock);
3286         target = get_restripe_target(root->fs_info, flags);
3287         if (target) {
3288                 /* pick target profile only if it's already available */
3289                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3290                         spin_unlock(&root->fs_info->balance_lock);
3291                         return extended_to_chunk(target);
3292                 }
3293         }
3294         spin_unlock(&root->fs_info->balance_lock);
3295
3296         if (num_devices == 1)
3297                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3298         if (num_devices < 4)
3299                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3300
3301         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3302             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3303                       BTRFS_BLOCK_GROUP_RAID10))) {
3304                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3305         }
3306
3307         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3308             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3309                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3310         }
3311
3312         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3313             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3314              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3315              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3316                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3317         }
3318
3319         return extended_to_chunk(flags);
3320 }
3321
3322 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3323 {
3324         if (flags & BTRFS_BLOCK_GROUP_DATA)
3325                 flags |= root->fs_info->avail_data_alloc_bits;
3326         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3327                 flags |= root->fs_info->avail_system_alloc_bits;
3328         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3329                 flags |= root->fs_info->avail_metadata_alloc_bits;
3330
3331         return btrfs_reduce_alloc_profile(root, flags);
3332 }
3333
3334 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3335 {
3336         u64 flags;
3337
3338         if (data)
3339                 flags = BTRFS_BLOCK_GROUP_DATA;
3340         else if (root == root->fs_info->chunk_root)
3341                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3342         else
3343                 flags = BTRFS_BLOCK_GROUP_METADATA;
3344
3345         return get_alloc_profile(root, flags);
3346 }
3347
3348 /*
3349  * This will check the space that the inode allocates from to make sure we have
3350  * enough space for bytes.
3351  */
3352 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3353 {
3354         struct btrfs_space_info *data_sinfo;
3355         struct btrfs_root *root = BTRFS_I(inode)->root;
3356         struct btrfs_fs_info *fs_info = root->fs_info;
3357         u64 used;
3358         int ret = 0, committed = 0, alloc_chunk = 1;
3359
3360         /* make sure bytes are sectorsize aligned */
3361         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3362
3363         if (root == root->fs_info->tree_root ||
3364             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3365                 alloc_chunk = 0;
3366                 committed = 1;
3367         }
3368
3369         data_sinfo = fs_info->data_sinfo;
3370         if (!data_sinfo)
3371                 goto alloc;
3372
3373 again:
3374         /* make sure we have enough space to handle the data first */
3375         spin_lock(&data_sinfo->lock);
3376         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3377                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3378                 data_sinfo->bytes_may_use;
3379
3380         if (used + bytes > data_sinfo->total_bytes) {
3381                 struct btrfs_trans_handle *trans;
3382
3383                 /*
3384                  * if we don't have enough free bytes in this space then we need
3385                  * to alloc a new chunk.
3386                  */
3387                 if (!data_sinfo->full && alloc_chunk) {
3388                         u64 alloc_target;
3389
3390                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3391                         spin_unlock(&data_sinfo->lock);
3392 alloc:
3393                         alloc_target = btrfs_get_alloc_profile(root, 1);
3394                         trans = btrfs_join_transaction(root);
3395                         if (IS_ERR(trans))
3396                                 return PTR_ERR(trans);
3397
3398                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3399                                              bytes + 2 * 1024 * 1024,
3400                                              alloc_target,
3401                                              CHUNK_ALLOC_NO_FORCE);
3402                         btrfs_end_transaction(trans, root);
3403                         if (ret < 0) {
3404                                 if (ret != -ENOSPC)
3405                                         return ret;
3406                                 else
3407                                         goto commit_trans;
3408                         }
3409
3410                         if (!data_sinfo)
3411                                 data_sinfo = fs_info->data_sinfo;
3412
3413                         goto again;
3414                 }
3415
3416                 /*
3417                  * If we have less pinned bytes than we want to allocate then
3418                  * don't bother committing the transaction, it won't help us.
3419                  */
3420                 if (data_sinfo->bytes_pinned < bytes)
3421                         committed = 1;
3422                 spin_unlock(&data_sinfo->lock);
3423
3424                 /* commit the current transaction and try again */
3425 commit_trans:
3426                 if (!committed &&
3427                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3428                         committed = 1;
3429                         trans = btrfs_join_transaction(root);
3430                         if (IS_ERR(trans))
3431                                 return PTR_ERR(trans);
3432                         ret = btrfs_commit_transaction(trans, root);
3433                         if (ret)
3434                                 return ret;
3435                         goto again;
3436                 }
3437
3438                 return -ENOSPC;
3439         }
3440         data_sinfo->bytes_may_use += bytes;
3441         trace_btrfs_space_reservation(root->fs_info, "space_info",
3442                                       data_sinfo->flags, bytes, 1);
3443         spin_unlock(&data_sinfo->lock);
3444
3445         return 0;
3446 }
3447
3448 /*
3449  * Called if we need to clear a data reservation for this inode.
3450  */
3451 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3452 {
3453         struct btrfs_root *root = BTRFS_I(inode)->root;
3454         struct btrfs_space_info *data_sinfo;
3455
3456         /* make sure bytes are sectorsize aligned */
3457         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3458
3459         data_sinfo = root->fs_info->data_sinfo;
3460         spin_lock(&data_sinfo->lock);
3461         data_sinfo->bytes_may_use -= bytes;
3462         trace_btrfs_space_reservation(root->fs_info, "space_info",
3463                                       data_sinfo->flags, bytes, 0);
3464         spin_unlock(&data_sinfo->lock);
3465 }
3466
3467 static void force_metadata_allocation(struct btrfs_fs_info *info)
3468 {
3469         struct list_head *head = &info->space_info;
3470         struct btrfs_space_info *found;
3471
3472         rcu_read_lock();
3473         list_for_each_entry_rcu(found, head, list) {
3474                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3475                         found->force_alloc = CHUNK_ALLOC_FORCE;
3476         }
3477         rcu_read_unlock();
3478 }
3479
3480 static int should_alloc_chunk(struct btrfs_root *root,
3481                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3482                               int force)
3483 {
3484         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3485         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3486         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3487         u64 thresh;
3488
3489         if (force == CHUNK_ALLOC_FORCE)
3490                 return 1;
3491
3492         /*
3493          * We need to take into account the global rsv because for all intents
3494          * and purposes it's used space.  Don't worry about locking the
3495          * global_rsv, it doesn't change except when the transaction commits.
3496          */
3497         num_allocated += global_rsv->size;
3498
3499         /*
3500          * in limited mode, we want to have some free space up to
3501          * about 1% of the FS size.
3502          */
3503         if (force == CHUNK_ALLOC_LIMITED) {
3504                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3505                 thresh = max_t(u64, 64 * 1024 * 1024,
3506                                div_factor_fine(thresh, 1));
3507
3508                 if (num_bytes - num_allocated < thresh)
3509                         return 1;
3510         }
3511         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3512
3513         /* 256MB or 2% of the FS */
3514         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3515         /* system chunks need a much small threshold */
3516         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3517                 thresh = 32 * 1024 * 1024;
3518
3519         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3520                 return 0;
3521         return 1;
3522 }
3523
3524 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3525 {
3526         u64 num_dev;
3527
3528         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3529             type & BTRFS_BLOCK_GROUP_RAID0)
3530                 num_dev = root->fs_info->fs_devices->rw_devices;
3531         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3532                 num_dev = 2;
3533         else
3534                 num_dev = 1;    /* DUP or single */
3535
3536         /* metadata for updaing devices and chunk tree */
3537         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3538 }
3539
3540 static void check_system_chunk(struct btrfs_trans_handle *trans,
3541                                struct btrfs_root *root, u64 type)
3542 {
3543         struct btrfs_space_info *info;
3544         u64 left;
3545         u64 thresh;
3546
3547         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3548         spin_lock(&info->lock);
3549         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3550                 info->bytes_reserved - info->bytes_readonly;
3551         spin_unlock(&info->lock);
3552
3553         thresh = get_system_chunk_thresh(root, type);
3554         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3555                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3556                        left, thresh, type);
3557                 dump_space_info(info, 0, 0);
3558         }
3559
3560         if (left < thresh) {
3561                 u64 flags;
3562
3563                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3564                 btrfs_alloc_chunk(trans, root, flags);
3565         }
3566 }
3567
3568 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3569                           struct btrfs_root *extent_root, u64 alloc_bytes,
3570                           u64 flags, int force)
3571 {
3572         struct btrfs_space_info *space_info;
3573         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3574         int wait_for_alloc = 0;
3575         int ret = 0;
3576
3577         space_info = __find_space_info(extent_root->fs_info, flags);
3578         if (!space_info) {
3579                 ret = update_space_info(extent_root->fs_info, flags,
3580                                         0, 0, &space_info);
3581                 BUG_ON(ret); /* -ENOMEM */
3582         }
3583         BUG_ON(!space_info); /* Logic error */
3584
3585 again:
3586         spin_lock(&space_info->lock);
3587         if (force < space_info->force_alloc)
3588                 force = space_info->force_alloc;
3589         if (space_info->full) {
3590                 spin_unlock(&space_info->lock);
3591                 return 0;
3592         }
3593
3594         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3595                 spin_unlock(&space_info->lock);
3596                 return 0;
3597         } else if (space_info->chunk_alloc) {
3598                 wait_for_alloc = 1;
3599         } else {
3600                 space_info->chunk_alloc = 1;
3601         }
3602
3603         spin_unlock(&space_info->lock);
3604
3605         mutex_lock(&fs_info->chunk_mutex);
3606
3607         /*
3608          * The chunk_mutex is held throughout the entirety of a chunk
3609          * allocation, so once we've acquired the chunk_mutex we know that the
3610          * other guy is done and we need to recheck and see if we should
3611          * allocate.
3612          */
3613         if (wait_for_alloc) {
3614                 mutex_unlock(&fs_info->chunk_mutex);
3615                 wait_for_alloc = 0;
3616                 goto again;
3617         }
3618
3619         /*
3620          * If we have mixed data/metadata chunks we want to make sure we keep
3621          * allocating mixed chunks instead of individual chunks.
3622          */
3623         if (btrfs_mixed_space_info(space_info))
3624                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3625
3626         /*
3627          * if we're doing a data chunk, go ahead and make sure that
3628          * we keep a reasonable number of metadata chunks allocated in the
3629          * FS as well.
3630          */
3631         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3632                 fs_info->data_chunk_allocations++;
3633                 if (!(fs_info->data_chunk_allocations %
3634                       fs_info->metadata_ratio))
3635                         force_metadata_allocation(fs_info);
3636         }
3637
3638         /*
3639          * Check if we have enough space in SYSTEM chunk because we may need
3640          * to update devices.
3641          */
3642         check_system_chunk(trans, extent_root, flags);
3643
3644         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3645         if (ret < 0 && ret != -ENOSPC)
3646                 goto out;
3647
3648         spin_lock(&space_info->lock);
3649         if (ret)
3650                 space_info->full = 1;
3651         else
3652                 ret = 1;
3653
3654         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3655         space_info->chunk_alloc = 0;
3656         spin_unlock(&space_info->lock);
3657 out:
3658         mutex_unlock(&fs_info->chunk_mutex);
3659         return ret;
3660 }
3661
3662 /*
3663  * shrink metadata reservation for delalloc
3664  */
3665 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3666                             bool wait_ordered)
3667 {
3668         struct btrfs_block_rsv *block_rsv;
3669         struct btrfs_space_info *space_info;
3670         struct btrfs_trans_handle *trans;
3671         u64 delalloc_bytes;
3672         u64 max_reclaim;
3673         long time_left;
3674         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3675         int loops = 0;
3676
3677         trans = (struct btrfs_trans_handle *)current->journal_info;
3678         block_rsv = &root->fs_info->delalloc_block_rsv;
3679         space_info = block_rsv->space_info;
3680
3681         smp_mb();
3682         delalloc_bytes = root->fs_info->delalloc_bytes;
3683         if (delalloc_bytes == 0) {
3684                 if (trans)
3685                         return;
3686                 btrfs_wait_ordered_extents(root, 0, 0);
3687                 return;
3688         }
3689
3690         while (delalloc_bytes && loops < 3) {
3691                 max_reclaim = min(delalloc_bytes, to_reclaim);
3692                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3693                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3694                                                WB_REASON_FS_FREE_SPACE);
3695
3696                 spin_lock(&space_info->lock);
3697                 if (space_info->bytes_used + space_info->bytes_reserved +
3698                     space_info->bytes_pinned + space_info->bytes_readonly +
3699                     space_info->bytes_may_use + orig <=
3700                     space_info->total_bytes) {
3701                         spin_unlock(&space_info->lock);
3702                         break;
3703                 }
3704                 spin_unlock(&space_info->lock);
3705
3706                 loops++;
3707                 if (wait_ordered && !trans) {
3708                         btrfs_wait_ordered_extents(root, 0, 0);
3709                 } else {
3710                         time_left = schedule_timeout_killable(1);
3711                         if (time_left)
3712                                 break;
3713                 }
3714                 smp_mb();
3715                 delalloc_bytes = root->fs_info->delalloc_bytes;
3716         }
3717 }
3718
3719 /**
3720  * maybe_commit_transaction - possibly commit the transaction if its ok to
3721  * @root - the root we're allocating for
3722  * @bytes - the number of bytes we want to reserve
3723  * @force - force the commit
3724  *
3725  * This will check to make sure that committing the transaction will actually
3726  * get us somewhere and then commit the transaction if it does.  Otherwise it
3727  * will return -ENOSPC.
3728  */
3729 static int may_commit_transaction(struct btrfs_root *root,
3730                                   struct btrfs_space_info *space_info,
3731                                   u64 bytes, int force)
3732 {
3733         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3734         struct btrfs_trans_handle *trans;
3735
3736         trans = (struct btrfs_trans_handle *)current->journal_info;
3737         if (trans)
3738                 return -EAGAIN;
3739
3740         if (force)
3741                 goto commit;
3742
3743         /* See if there is enough pinned space to make this reservation */
3744         spin_lock(&space_info->lock);
3745         if (space_info->bytes_pinned >= bytes) {
3746                 spin_unlock(&space_info->lock);
3747                 goto commit;
3748         }
3749         spin_unlock(&space_info->lock);
3750
3751         /*
3752          * See if there is some space in the delayed insertion reservation for
3753          * this reservation.
3754          */
3755         if (space_info != delayed_rsv->space_info)
3756                 return -ENOSPC;
3757
3758         spin_lock(&space_info->lock);
3759         spin_lock(&delayed_rsv->lock);
3760         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3761                 spin_unlock(&delayed_rsv->lock);
3762                 spin_unlock(&space_info->lock);
3763                 return -ENOSPC;
3764         }
3765         spin_unlock(&delayed_rsv->lock);
3766         spin_unlock(&space_info->lock);
3767
3768 commit:
3769         trans = btrfs_join_transaction(root);
3770         if (IS_ERR(trans))
3771                 return -ENOSPC;
3772
3773         return btrfs_commit_transaction(trans, root);
3774 }
3775
3776 enum flush_state {
3777         FLUSH_DELALLOC          =       1,
3778         FLUSH_DELALLOC_WAIT     =       2,
3779         FLUSH_DELAYED_ITEMS_NR  =       3,
3780         FLUSH_DELAYED_ITEMS     =       4,
3781         COMMIT_TRANS            =       5,
3782 };
3783
3784 static int flush_space(struct btrfs_root *root,
3785                        struct btrfs_space_info *space_info, u64 num_bytes,
3786                        u64 orig_bytes, int state)
3787 {
3788         struct btrfs_trans_handle *trans;
3789         int nr;
3790         int ret = 0;
3791
3792         switch (state) {
3793         case FLUSH_DELALLOC:
3794         case FLUSH_DELALLOC_WAIT:
3795                 shrink_delalloc(root, num_bytes, orig_bytes,
3796                                 state == FLUSH_DELALLOC_WAIT);
3797                 break;
3798         case FLUSH_DELAYED_ITEMS_NR:
3799         case FLUSH_DELAYED_ITEMS:
3800                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3801                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3802
3803                         nr = (int)div64_u64(num_bytes, bytes);
3804                         if (!nr)
3805                                 nr = 1;
3806                         nr *= 2;
3807                 } else {
3808                         nr = -1;
3809                 }
3810                 trans = btrfs_join_transaction(root);
3811                 if (IS_ERR(trans)) {
3812                         ret = PTR_ERR(trans);
3813                         break;
3814                 }
3815                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3816                 btrfs_end_transaction(trans, root);
3817                 break;
3818         case COMMIT_TRANS:
3819                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3820                 break;
3821         default:
3822                 ret = -ENOSPC;
3823                 break;
3824         }
3825
3826         return ret;
3827 }
3828 /**
3829  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3830  * @root - the root we're allocating for
3831  * @block_rsv - the block_rsv we're allocating for
3832  * @orig_bytes - the number of bytes we want
3833  * @flush - wether or not we can flush to make our reservation
3834  *
3835  * This will reserve orgi_bytes number of bytes from the space info associated
3836  * with the block_rsv.  If there is not enough space it will make an attempt to
3837  * flush out space to make room.  It will do this by flushing delalloc if
3838  * possible or committing the transaction.  If flush is 0 then no attempts to
3839  * regain reservations will be made and this will fail if there is not enough
3840  * space already.
3841  */
3842 static int reserve_metadata_bytes(struct btrfs_root *root,
3843                                   struct btrfs_block_rsv *block_rsv,
3844                                   u64 orig_bytes, int flush)
3845 {
3846         struct btrfs_space_info *space_info = block_rsv->space_info;
3847         u64 used;
3848         u64 num_bytes = orig_bytes;
3849         int flush_state = FLUSH_DELALLOC;
3850         int ret = 0;
3851         bool flushing = false;
3852         bool committed = false;
3853
3854 again:
3855         ret = 0;
3856         spin_lock(&space_info->lock);
3857         /*
3858          * We only want to wait if somebody other than us is flushing and we are
3859          * actually alloed to flush.
3860          */
3861         while (flush && !flushing && space_info->flush) {
3862                 spin_unlock(&space_info->lock);
3863                 /*
3864                  * If we have a trans handle we can't wait because the flusher
3865                  * may have to commit the transaction, which would mean we would
3866                  * deadlock since we are waiting for the flusher to finish, but
3867                  * hold the current transaction open.
3868                  */
3869                 if (current->journal_info)
3870                         return -EAGAIN;
3871                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3872                 /* Must have been killed, return */
3873                 if (ret)
3874                         return -EINTR;
3875
3876                 spin_lock(&space_info->lock);
3877         }
3878
3879         ret = -ENOSPC;
3880         used = space_info->bytes_used + space_info->bytes_reserved +
3881                 space_info->bytes_pinned + space_info->bytes_readonly +
3882                 space_info->bytes_may_use;
3883
3884         /*
3885          * The idea here is that we've not already over-reserved the block group
3886          * then we can go ahead and save our reservation first and then start
3887          * flushing if we need to.  Otherwise if we've already overcommitted
3888          * lets start flushing stuff first and then come back and try to make
3889          * our reservation.
3890          */
3891         if (used <= space_info->total_bytes) {
3892                 if (used + orig_bytes <= space_info->total_bytes) {
3893                         space_info->bytes_may_use += orig_bytes;
3894                         trace_btrfs_space_reservation(root->fs_info,
3895                                 "space_info", space_info->flags, orig_bytes, 1);
3896                         ret = 0;
3897                 } else {
3898                         /*
3899                          * Ok set num_bytes to orig_bytes since we aren't
3900                          * overocmmitted, this way we only try and reclaim what
3901                          * we need.
3902                          */
3903                         num_bytes = orig_bytes;
3904                 }
3905         } else {
3906                 /*
3907                  * Ok we're over committed, set num_bytes to the overcommitted
3908                  * amount plus the amount of bytes that we need for this
3909                  * reservation.
3910                  */
3911                 num_bytes = used - space_info->total_bytes +
3912                         (orig_bytes * 2);
3913         }
3914
3915         if (ret) {
3916                 u64 profile = btrfs_get_alloc_profile(root, 0);
3917                 u64 avail;
3918
3919                 /*
3920                  * If we have a lot of space that's pinned, don't bother doing
3921                  * the overcommit dance yet and just commit the transaction.
3922                  */
3923                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3924                 do_div(avail, 10);
3925                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3926                         space_info->flush = 1;
3927                         flushing = true;
3928                         spin_unlock(&space_info->lock);
3929                         ret = may_commit_transaction(root, space_info,
3930                                                      orig_bytes, 1);
3931                         if (ret)
3932                                 goto out;
3933                         committed = true;
3934                         goto again;
3935                 }
3936
3937                 spin_lock(&root->fs_info->free_chunk_lock);
3938                 avail = root->fs_info->free_chunk_space;
3939
3940                 /*
3941                  * If we have dup, raid1 or raid10 then only half of the free
3942                  * space is actually useable.
3943                  */
3944                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3945                                BTRFS_BLOCK_GROUP_RAID1 |
3946                                BTRFS_BLOCK_GROUP_RAID10))
3947                         avail >>= 1;
3948
3949                 /*
3950                  * If we aren't flushing don't let us overcommit too much, say
3951                  * 1/8th of the space.  If we can flush, let it overcommit up to
3952                  * 1/2 of the space.
3953                  */
3954                 if (flush)
3955                         avail >>= 3;
3956                 else
3957                         avail >>= 1;
3958                  spin_unlock(&root->fs_info->free_chunk_lock);
3959
3960                 if (used + num_bytes < space_info->total_bytes + avail) {
3961                         space_info->bytes_may_use += orig_bytes;
3962                         trace_btrfs_space_reservation(root->fs_info,
3963                                 "space_info", space_info->flags, orig_bytes, 1);
3964                         ret = 0;
3965                 }
3966         }
3967
3968         /*
3969          * Couldn't make our reservation, save our place so while we're trying
3970          * to reclaim space we can actually use it instead of somebody else
3971          * stealing it from us.
3972          */
3973         if (ret && flush) {
3974                 flushing = true;
3975                 space_info->flush = 1;
3976         }
3977
3978         spin_unlock(&space_info->lock);
3979
3980         if (!ret || !flush)
3981                 goto out;
3982
3983         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3984                           flush_state);
3985         flush_state++;
3986         if (!ret)
3987                 goto again;
3988         else if (flush_state <= COMMIT_TRANS)
3989                 goto again;
3990
3991 out:
3992         if (flushing) {
3993                 spin_lock(&space_info->lock);
3994                 space_info->flush = 0;
3995                 wake_up_all(&space_info->wait);
3996                 spin_unlock(&space_info->lock);
3997         }
3998         return ret;
3999 }
4000
4001 static struct btrfs_block_rsv *get_block_rsv(
4002                                         const struct btrfs_trans_handle *trans,
4003                                         const struct btrfs_root *root)
4004 {
4005         struct btrfs_block_rsv *block_rsv = NULL;
4006
4007         if (root->ref_cows)
4008                 block_rsv = trans->block_rsv;
4009
4010         if (root == root->fs_info->csum_root && trans->adding_csums)
4011                 block_rsv = trans->block_rsv;
4012
4013         if (!block_rsv)
4014                 block_rsv = root->block_rsv;
4015
4016         if (!block_rsv)
4017                 block_rsv = &root->fs_info->empty_block_rsv;
4018
4019         return block_rsv;
4020 }
4021
4022 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4023                                u64 num_bytes)
4024 {
4025         int ret = -ENOSPC;
4026         spin_lock(&block_rsv->lock);
4027         if (block_rsv->reserved >= num_bytes) {
4028                 block_rsv->reserved -= num_bytes;
4029                 if (block_rsv->reserved < block_rsv->size)
4030                         block_rsv->full = 0;
4031                 ret = 0;
4032         }
4033         spin_unlock(&block_rsv->lock);
4034         return ret;
4035 }
4036
4037 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4038                                 u64 num_bytes, int update_size)
4039 {
4040         spin_lock(&block_rsv->lock);
4041         block_rsv->reserved += num_bytes;
4042         if (update_size)
4043                 block_rsv->size += num_bytes;
4044         else if (block_rsv->reserved >= block_rsv->size)
4045                 block_rsv->full = 1;
4046         spin_unlock(&block_rsv->lock);
4047 }
4048
4049 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4050                                     struct btrfs_block_rsv *block_rsv,
4051                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4052 {
4053         struct btrfs_space_info *space_info = block_rsv->space_info;
4054
4055         spin_lock(&block_rsv->lock);
4056         if (num_bytes == (u64)-1)
4057                 num_bytes = block_rsv->size;
4058         block_rsv->size -= num_bytes;
4059         if (block_rsv->reserved >= block_rsv->size) {
4060                 num_bytes = block_rsv->reserved - block_rsv->size;
4061                 block_rsv->reserved = block_rsv->size;
4062                 block_rsv->full = 1;
4063         } else {
4064                 num_bytes = 0;
4065         }
4066         spin_unlock(&block_rsv->lock);
4067
4068         if (num_bytes > 0) {
4069                 if (dest) {
4070                         spin_lock(&dest->lock);
4071                         if (!dest->full) {
4072                                 u64 bytes_to_add;
4073
4074                                 bytes_to_add = dest->size - dest->reserved;
4075                                 bytes_to_add = min(num_bytes, bytes_to_add);
4076                                 dest->reserved += bytes_to_add;
4077                                 if (dest->reserved >= dest->size)
4078                                         dest->full = 1;
4079                                 num_bytes -= bytes_to_add;
4080                         }
4081                         spin_unlock(&dest->lock);
4082                 }
4083                 if (num_bytes) {
4084                         spin_lock(&space_info->lock);
4085                         space_info->bytes_may_use -= num_bytes;
4086                         trace_btrfs_space_reservation(fs_info, "space_info",
4087                                         space_info->flags, num_bytes, 0);
4088                         space_info->reservation_progress++;
4089                         spin_unlock(&space_info->lock);
4090                 }
4091         }
4092 }
4093
4094 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4095                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4096 {
4097         int ret;
4098
4099         ret = block_rsv_use_bytes(src, num_bytes);
4100         if (ret)
4101                 return ret;
4102
4103         block_rsv_add_bytes(dst, num_bytes, 1);
4104         return 0;
4105 }
4106
4107 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4108 {
4109         memset(rsv, 0, sizeof(*rsv));
4110         spin_lock_init(&rsv->lock);
4111 }
4112
4113 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4114 {
4115         struct btrfs_block_rsv *block_rsv;
4116         struct btrfs_fs_info *fs_info = root->fs_info;
4117
4118         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4119         if (!block_rsv)
4120                 return NULL;
4121
4122         btrfs_init_block_rsv(block_rsv);
4123         block_rsv->space_info = __find_space_info(fs_info,
4124                                                   BTRFS_BLOCK_GROUP_METADATA);
4125         return block_rsv;
4126 }
4127
4128 void btrfs_free_block_rsv(struct btrfs_root *root,
4129                           struct btrfs_block_rsv *rsv)
4130 {
4131         btrfs_block_rsv_release(root, rsv, (u64)-1);
4132         kfree(rsv);
4133 }
4134
4135 static inline int __block_rsv_add(struct btrfs_root *root,
4136                                   struct btrfs_block_rsv *block_rsv,
4137                                   u64 num_bytes, int flush)
4138 {
4139         int ret;
4140
4141         if (num_bytes == 0)
4142                 return 0;
4143
4144         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4145         if (!ret) {
4146                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4147                 return 0;
4148         }
4149
4150         return ret;
4151 }
4152
4153 int btrfs_block_rsv_add(struct btrfs_root *root,
4154                         struct btrfs_block_rsv *block_rsv,
4155                         u64 num_bytes)
4156 {
4157         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4158 }
4159
4160 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4161                                 struct btrfs_block_rsv *block_rsv,
4162                                 u64 num_bytes)
4163 {
4164         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4165 }
4166
4167 int btrfs_block_rsv_check(struct btrfs_root *root,
4168                           struct btrfs_block_rsv *block_rsv, int min_factor)
4169 {
4170         u64 num_bytes = 0;
4171         int ret = -ENOSPC;
4172
4173         if (!block_rsv)
4174                 return 0;
4175
4176         spin_lock(&block_rsv->lock);
4177         num_bytes = div_factor(block_rsv->size, min_factor);
4178         if (block_rsv->reserved >= num_bytes)
4179                 ret = 0;
4180         spin_unlock(&block_rsv->lock);
4181
4182         return ret;
4183 }
4184
4185 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4186                                            struct btrfs_block_rsv *block_rsv,
4187                                            u64 min_reserved, int flush)
4188 {
4189         u64 num_bytes = 0;
4190         int ret = -ENOSPC;
4191
4192         if (!block_rsv)
4193                 return 0;
4194
4195         spin_lock(&block_rsv->lock);
4196         num_bytes = min_reserved;
4197         if (block_rsv->reserved >= num_bytes)
4198                 ret = 0;
4199         else
4200                 num_bytes -= block_rsv->reserved;
4201         spin_unlock(&block_rsv->lock);
4202
4203         if (!ret)
4204                 return 0;
4205
4206         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4207         if (!ret) {
4208                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4209                 return 0;
4210         }
4211
4212         return ret;
4213 }
4214
4215 int btrfs_block_rsv_refill(struct btrfs_root *root,
4216                            struct btrfs_block_rsv *block_rsv,
4217                            u64 min_reserved)
4218 {
4219         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4220 }
4221
4222 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4223                                    struct btrfs_block_rsv *block_rsv,
4224                                    u64 min_reserved)
4225 {
4226         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4227 }
4228
4229 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4230                             struct btrfs_block_rsv *dst_rsv,
4231                             u64 num_bytes)
4232 {
4233         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4234 }
4235
4236 void btrfs_block_rsv_release(struct btrfs_root *root,
4237                              struct btrfs_block_rsv *block_rsv,
4238                              u64 num_bytes)
4239 {
4240         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4241         if (global_rsv->full || global_rsv == block_rsv ||
4242             block_rsv->space_info != global_rsv->space_info)
4243                 global_rsv = NULL;
4244         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4245                                 num_bytes);
4246 }
4247
4248 /*
4249  * helper to calculate size of global block reservation.
4250  * the desired value is sum of space used by extent tree,
4251  * checksum tree and root tree
4252  */
4253 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4254 {
4255         struct btrfs_space_info *sinfo;
4256         u64 num_bytes;
4257         u64 meta_used;
4258         u64 data_used;
4259         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4260
4261         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4262         spin_lock(&sinfo->lock);
4263         data_used = sinfo->bytes_used;
4264         spin_unlock(&sinfo->lock);
4265
4266         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4267         spin_lock(&sinfo->lock);
4268         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4269                 data_used = 0;
4270         meta_used = sinfo->bytes_used;
4271         spin_unlock(&sinfo->lock);
4272
4273         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4274                     csum_size * 2;
4275         num_bytes += div64_u64(data_used + meta_used, 50);
4276
4277         if (num_bytes * 3 > meta_used)
4278                 num_bytes = div64_u64(meta_used, 3);
4279
4280         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4281 }
4282
4283 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4284 {
4285         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4286         struct btrfs_space_info *sinfo = block_rsv->space_info;
4287         u64 num_bytes;
4288
4289         num_bytes = calc_global_metadata_size(fs_info);
4290
4291         spin_lock(&sinfo->lock);
4292         spin_lock(&block_rsv->lock);
4293
4294         block_rsv->size = num_bytes;
4295
4296         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4297                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4298                     sinfo->bytes_may_use;
4299
4300         if (sinfo->total_bytes > num_bytes) {
4301                 num_bytes = sinfo->total_bytes - num_bytes;
4302                 block_rsv->reserved += num_bytes;
4303                 sinfo->bytes_may_use += num_bytes;
4304                 trace_btrfs_space_reservation(fs_info, "space_info",
4305                                       sinfo->flags, num_bytes, 1);
4306         }
4307
4308         if (block_rsv->reserved >= block_rsv->size) {
4309                 num_bytes = block_rsv->reserved - block_rsv->size;
4310                 sinfo->bytes_may_use -= num_bytes;
4311                 trace_btrfs_space_reservation(fs_info, "space_info",
4312                                       sinfo->flags, num_bytes, 0);
4313                 sinfo->reservation_progress++;
4314                 block_rsv->reserved = block_rsv->size;
4315                 block_rsv->full = 1;
4316         }
4317
4318         spin_unlock(&block_rsv->lock);
4319         spin_unlock(&sinfo->lock);
4320 }
4321
4322 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4323 {
4324         struct btrfs_space_info *space_info;
4325
4326         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4327         fs_info->chunk_block_rsv.space_info = space_info;
4328
4329         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4330         fs_info->global_block_rsv.space_info = space_info;
4331         fs_info->delalloc_block_rsv.space_info = space_info;
4332         fs_info->trans_block_rsv.space_info = space_info;
4333         fs_info->empty_block_rsv.space_info = space_info;
4334         fs_info->delayed_block_rsv.space_info = space_info;
4335
4336         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4337         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4338         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4339         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4340         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4341
4342         update_global_block_rsv(fs_info);
4343 }
4344
4345 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4346 {
4347         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4348                                 (u64)-1);
4349         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4350         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4351         WARN_ON(fs_info->trans_block_rsv.size > 0);
4352         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4353         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4354         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4355         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4356         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4357 }
4358
4359 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4360                                   struct btrfs_root *root)
4361 {
4362         if (!trans->block_rsv)
4363                 return;
4364
4365         if (!trans->bytes_reserved)
4366                 return;
4367
4368         trace_btrfs_space_reservation(root->fs_info, "transaction",
4369                                       trans->transid, trans->bytes_reserved, 0);
4370         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4371         trans->bytes_reserved = 0;
4372 }
4373
4374 /* Can only return 0 or -ENOSPC */
4375 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4376                                   struct inode *inode)
4377 {
4378         struct btrfs_root *root = BTRFS_I(inode)->root;
4379         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4380         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4381
4382         /*
4383          * We need to hold space in order to delete our orphan item once we've
4384          * added it, so this takes the reservation so we can release it later
4385          * when we are truly done with the orphan item.
4386          */
4387         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4388         trace_btrfs_space_reservation(root->fs_info, "orphan",
4389                                       btrfs_ino(inode), num_bytes, 1);
4390         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4391 }
4392
4393 void btrfs_orphan_release_metadata(struct inode *inode)
4394 {
4395         struct btrfs_root *root = BTRFS_I(inode)->root;
4396         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4397         trace_btrfs_space_reservation(root->fs_info, "orphan",
4398                                       btrfs_ino(inode), num_bytes, 0);
4399         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4400 }
4401
4402 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4403                                 struct btrfs_pending_snapshot *pending)
4404 {
4405         struct btrfs_root *root = pending->root;
4406         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4407         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4408         /*
4409          * two for root back/forward refs, two for directory entries
4410          * and one for root of the snapshot.
4411          */
4412         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4413         dst_rsv->space_info = src_rsv->space_info;
4414         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4415 }
4416
4417 /**
4418  * drop_outstanding_extent - drop an outstanding extent
4419  * @inode: the inode we're dropping the extent for
4420  *
4421  * This is called when we are freeing up an outstanding extent, either called
4422  * after an error or after an extent is written.  This will return the number of
4423  * reserved extents that need to be freed.  This must be called with
4424  * BTRFS_I(inode)->lock held.
4425  */
4426 static unsigned drop_outstanding_extent(struct inode *inode)
4427 {
4428         unsigned drop_inode_space = 0;
4429         unsigned dropped_extents = 0;
4430
4431         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4432         BTRFS_I(inode)->outstanding_extents--;
4433
4434         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4435             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4436                                &BTRFS_I(inode)->runtime_flags))
4437                 drop_inode_space = 1;
4438
4439         /*
4440          * If we have more or the same amount of outsanding extents than we have
4441          * reserved then we need to leave the reserved extents count alone.
4442          */
4443         if (BTRFS_I(inode)->outstanding_extents >=
4444             BTRFS_I(inode)->reserved_extents)
4445                 return drop_inode_space;
4446
4447         dropped_extents = BTRFS_I(inode)->reserved_extents -
4448                 BTRFS_I(inode)->outstanding_extents;
4449         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4450         return dropped_extents + drop_inode_space;
4451 }
4452
4453 /**
4454  * calc_csum_metadata_size - return the amount of metada space that must be
4455  *      reserved/free'd for the given bytes.
4456  * @inode: the inode we're manipulating
4457  * @num_bytes: the number of bytes in question
4458  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4459  *
4460  * This adjusts the number of csum_bytes in the inode and then returns the
4461  * correct amount of metadata that must either be reserved or freed.  We
4462  * calculate how many checksums we can fit into one leaf and then divide the
4463  * number of bytes that will need to be checksumed by this value to figure out
4464  * how many checksums will be required.  If we are adding bytes then the number
4465  * may go up and we will return the number of additional bytes that must be
4466  * reserved.  If it is going down we will return the number of bytes that must
4467  * be freed.
4468  *
4469  * This must be called with BTRFS_I(inode)->lock held.
4470  */
4471 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4472                                    int reserve)
4473 {
4474         struct btrfs_root *root = BTRFS_I(inode)->root;
4475         u64 csum_size;
4476         int num_csums_per_leaf;
4477         int num_csums;
4478         int old_csums;
4479
4480         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4481             BTRFS_I(inode)->csum_bytes == 0)
4482                 return 0;
4483
4484         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4485         if (reserve)
4486                 BTRFS_I(inode)->csum_bytes += num_bytes;
4487         else
4488                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4489         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4490         num_csums_per_leaf = (int)div64_u64(csum_size,
4491                                             sizeof(struct btrfs_csum_item) +
4492                                             sizeof(struct btrfs_disk_key));
4493         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4494         num_csums = num_csums + num_csums_per_leaf - 1;
4495         num_csums = num_csums / num_csums_per_leaf;
4496
4497         old_csums = old_csums + num_csums_per_leaf - 1;
4498         old_csums = old_csums / num_csums_per_leaf;
4499
4500         /* No change, no need to reserve more */
4501         if (old_csums == num_csums)
4502                 return 0;
4503
4504         if (reserve)
4505                 return btrfs_calc_trans_metadata_size(root,
4506                                                       num_csums - old_csums);
4507
4508         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4509 }
4510
4511 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4512 {
4513         struct btrfs_root *root = BTRFS_I(inode)->root;
4514         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4515         u64 to_reserve = 0;
4516         u64 csum_bytes;
4517         unsigned nr_extents = 0;
4518         int extra_reserve = 0;
4519         int flush = 1;
4520         int ret;
4521
4522         /* Need to be holding the i_mutex here if we aren't free space cache */
4523         if (btrfs_is_free_space_inode(inode))
4524                 flush = 0;
4525
4526         if (flush && btrfs_transaction_in_commit(root->fs_info))
4527                 schedule_timeout(1);
4528
4529         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4530         num_bytes = ALIGN(num_bytes, root->sectorsize);
4531
4532         spin_lock(&BTRFS_I(inode)->lock);
4533         BTRFS_I(inode)->outstanding_extents++;
4534
4535         if (BTRFS_I(inode)->outstanding_extents >
4536             BTRFS_I(inode)->reserved_extents)
4537                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4538                         BTRFS_I(inode)->reserved_extents;
4539
4540         /*
4541          * Add an item to reserve for updating the inode when we complete the
4542          * delalloc io.
4543          */
4544         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4545                       &BTRFS_I(inode)->runtime_flags)) {
4546                 nr_extents++;
4547                 extra_reserve = 1;
4548         }
4549
4550         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4551         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4552         csum_bytes = BTRFS_I(inode)->csum_bytes;
4553         spin_unlock(&BTRFS_I(inode)->lock);
4554
4555         if (root->fs_info->quota_enabled) {
4556                 ret = btrfs_qgroup_reserve(root, num_bytes +
4557                                            nr_extents * root->leafsize);
4558                 if (ret) {
4559                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4560                         return ret;
4561                 }
4562         }
4563
4564         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4565         if (ret) {
4566                 u64 to_free = 0;
4567                 unsigned dropped;
4568
4569                 spin_lock(&BTRFS_I(inode)->lock);
4570                 dropped = drop_outstanding_extent(inode);
4571                 /*
4572                  * If the inodes csum_bytes is the same as the original
4573                  * csum_bytes then we know we haven't raced with any free()ers
4574                  * so we can just reduce our inodes csum bytes and carry on.
4575                  * Otherwise we have to do the normal free thing to account for
4576                  * the case that the free side didn't free up its reserve
4577                  * because of this outstanding reservation.
4578                  */
4579                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4580                         calc_csum_metadata_size(inode, num_bytes, 0);
4581                 else
4582                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4583                 spin_unlock(&BTRFS_I(inode)->lock);
4584                 if (dropped)
4585                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4586
4587                 if (to_free) {
4588                         btrfs_block_rsv_release(root, block_rsv, to_free);
4589                         trace_btrfs_space_reservation(root->fs_info,
4590                                                       "delalloc",
4591                                                       btrfs_ino(inode),
4592                                                       to_free, 0);
4593                 }
4594                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4595                 return ret;
4596         }
4597
4598         spin_lock(&BTRFS_I(inode)->lock);
4599         if (extra_reserve) {
4600                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4601                         &BTRFS_I(inode)->runtime_flags);
4602                 nr_extents--;
4603         }
4604         BTRFS_I(inode)->reserved_extents += nr_extents;
4605         spin_unlock(&BTRFS_I(inode)->lock);
4606         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4607
4608         if (to_reserve)
4609                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4610                                               btrfs_ino(inode), to_reserve, 1);
4611         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4612
4613         return 0;
4614 }
4615
4616 /**
4617  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4618  * @inode: the inode to release the reservation for
4619  * @num_bytes: the number of bytes we're releasing
4620  *
4621  * This will release the metadata reservation for an inode.  This can be called
4622  * once we complete IO for a given set of bytes to release their metadata
4623  * reservations.
4624  */
4625 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4626 {
4627         struct btrfs_root *root = BTRFS_I(inode)->root;
4628         u64 to_free = 0;
4629         unsigned dropped;
4630
4631         num_bytes = ALIGN(num_bytes, root->sectorsize);
4632         spin_lock(&BTRFS_I(inode)->lock);
4633         dropped = drop_outstanding_extent(inode);
4634
4635         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4636         spin_unlock(&BTRFS_I(inode)->lock);
4637         if (dropped > 0)
4638                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4639
4640         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4641                                       btrfs_ino(inode), to_free, 0);
4642         if (root->fs_info->quota_enabled) {
4643                 btrfs_qgroup_free(root, num_bytes +
4644                                         dropped * root->leafsize);
4645         }
4646
4647         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4648                                 to_free);
4649 }
4650
4651 /**
4652  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4653  * @inode: inode we're writing to
4654  * @num_bytes: the number of bytes we want to allocate
4655  *
4656  * This will do the following things
4657  *
4658  * o reserve space in the data space info for num_bytes
4659  * o reserve space in the metadata space info based on number of outstanding
4660  *   extents and how much csums will be needed
4661  * o add to the inodes ->delalloc_bytes
4662  * o add it to the fs_info's delalloc inodes list.
4663  *
4664  * This will return 0 for success and -ENOSPC if there is no space left.
4665  */
4666 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4667 {
4668         int ret;
4669
4670         ret = btrfs_check_data_free_space(inode, num_bytes);
4671         if (ret)
4672                 return ret;
4673
4674         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4675         if (ret) {
4676                 btrfs_free_reserved_data_space(inode, num_bytes);
4677                 return ret;
4678         }
4679
4680         return 0;
4681 }
4682
4683 /**
4684  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4685  * @inode: inode we're releasing space for
4686  * @num_bytes: the number of bytes we want to free up
4687  *
4688  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4689  * called in the case that we don't need the metadata AND data reservations
4690  * anymore.  So if there is an error or we insert an inline extent.
4691  *
4692  * This function will release the metadata space that was not used and will
4693  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4694  * list if there are no delalloc bytes left.
4695  */
4696 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4697 {
4698         btrfs_delalloc_release_metadata(inode, num_bytes);
4699         btrfs_free_reserved_data_space(inode, num_bytes);
4700 }
4701
4702 static int update_block_group(struct btrfs_trans_handle *trans,
4703                               struct btrfs_root *root,
4704                               u64 bytenr, u64 num_bytes, int alloc)
4705 {
4706         struct btrfs_block_group_cache *cache = NULL;
4707         struct btrfs_fs_info *info = root->fs_info;
4708         u64 total = num_bytes;
4709         u64 old_val;
4710         u64 byte_in_group;
4711         int factor;
4712
4713         /* block accounting for super block */
4714         spin_lock(&info->delalloc_lock);
4715         old_val = btrfs_super_bytes_used(info->super_copy);
4716         if (alloc)
4717                 old_val += num_bytes;
4718         else
4719                 old_val -= num_bytes;
4720         btrfs_set_super_bytes_used(info->super_copy, old_val);
4721         spin_unlock(&info->delalloc_lock);
4722
4723         while (total) {
4724                 cache = btrfs_lookup_block_group(info, bytenr);
4725                 if (!cache)
4726                         return -ENOENT;
4727                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4728                                     BTRFS_BLOCK_GROUP_RAID1 |
4729                                     BTRFS_BLOCK_GROUP_RAID10))
4730                         factor = 2;
4731                 else
4732                         factor = 1;
4733                 /*
4734                  * If this block group has free space cache written out, we
4735                  * need to make sure to load it if we are removing space.  This
4736                  * is because we need the unpinning stage to actually add the
4737                  * space back to the block group, otherwise we will leak space.
4738                  */
4739                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4740                         cache_block_group(cache, trans, NULL, 1);
4741
4742                 byte_in_group = bytenr - cache->key.objectid;
4743                 WARN_ON(byte_in_group > cache->key.offset);
4744
4745                 spin_lock(&cache->space_info->lock);
4746                 spin_lock(&cache->lock);
4747
4748                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4749                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4750                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4751
4752                 cache->dirty = 1;
4753                 old_val = btrfs_block_group_used(&cache->item);
4754                 num_bytes = min(total, cache->key.offset - byte_in_group);
4755                 if (alloc) {
4756                         old_val += num_bytes;
4757                         btrfs_set_block_group_used(&cache->item, old_val);
4758                         cache->reserved -= num_bytes;
4759                         cache->space_info->bytes_reserved -= num_bytes;
4760                         cache->space_info->bytes_used += num_bytes;
4761                         cache->space_info->disk_used += num_bytes * factor;
4762                         spin_unlock(&cache->lock);
4763                         spin_unlock(&cache->space_info->lock);
4764                 } else {
4765                         old_val -= num_bytes;
4766                         btrfs_set_block_group_used(&cache->item, old_val);
4767                         cache->pinned += num_bytes;
4768                         cache->space_info->bytes_pinned += num_bytes;
4769                         cache->space_info->bytes_used -= num_bytes;
4770                         cache->space_info->disk_used -= num_bytes * factor;
4771                         spin_unlock(&cache->lock);
4772                         spin_unlock(&cache->space_info->lock);
4773
4774                         set_extent_dirty(info->pinned_extents,
4775                                          bytenr, bytenr + num_bytes - 1,
4776                                          GFP_NOFS | __GFP_NOFAIL);
4777                 }
4778                 btrfs_put_block_group(cache);
4779                 total -= num_bytes;
4780                 bytenr += num_bytes;
4781         }
4782         return 0;
4783 }
4784
4785 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4786 {
4787         struct btrfs_block_group_cache *cache;
4788         u64 bytenr;
4789
4790         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4791         if (!cache)
4792                 return 0;
4793
4794         bytenr = cache->key.objectid;
4795         btrfs_put_block_group(cache);
4796
4797         return bytenr;
4798 }
4799
4800 static int pin_down_extent(struct btrfs_root *root,
4801                            struct btrfs_block_group_cache *cache,
4802                            u64 bytenr, u64 num_bytes, int reserved)
4803 {
4804         spin_lock(&cache->space_info->lock);
4805         spin_lock(&cache->lock);
4806         cache->pinned += num_bytes;
4807         cache->space_info->bytes_pinned += num_bytes;
4808         if (reserved) {
4809                 cache->reserved -= num_bytes;
4810                 cache->space_info->bytes_reserved -= num_bytes;
4811         }
4812         spin_unlock(&cache->lock);
4813         spin_unlock(&cache->space_info->lock);
4814
4815         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4816                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4817         return 0;
4818 }
4819
4820 /*
4821  * this function must be called within transaction
4822  */
4823 int btrfs_pin_extent(struct btrfs_root *root,
4824                      u64 bytenr, u64 num_bytes, int reserved)
4825 {
4826         struct btrfs_block_group_cache *cache;
4827
4828         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4829         BUG_ON(!cache); /* Logic error */
4830
4831         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4832
4833         btrfs_put_block_group(cache);
4834         return 0;
4835 }
4836
4837 /*
4838  * this function must be called within transaction
4839  */
4840 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4841                                     struct btrfs_root *root,
4842                                     u64 bytenr, u64 num_bytes)
4843 {
4844         struct btrfs_block_group_cache *cache;
4845
4846         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4847         BUG_ON(!cache); /* Logic error */
4848
4849         /*
4850          * pull in the free space cache (if any) so that our pin
4851          * removes the free space from the cache.  We have load_only set
4852          * to one because the slow code to read in the free extents does check
4853          * the pinned extents.
4854          */
4855         cache_block_group(cache, trans, root, 1);
4856
4857         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4858
4859         /* remove us from the free space cache (if we're there at all) */
4860         btrfs_remove_free_space(cache, bytenr, num_bytes);
4861         btrfs_put_block_group(cache);
4862         return 0;
4863 }
4864
4865 /**
4866  * btrfs_update_reserved_bytes - update the block_group and space info counters
4867  * @cache:      The cache we are manipulating
4868  * @num_bytes:  The number of bytes in question
4869  * @reserve:    One of the reservation enums
4870  *
4871  * This is called by the allocator when it reserves space, or by somebody who is
4872  * freeing space that was never actually used on disk.  For example if you
4873  * reserve some space for a new leaf in transaction A and before transaction A
4874  * commits you free that leaf, you call this with reserve set to 0 in order to
4875  * clear the reservation.
4876  *
4877  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4878  * ENOSPC accounting.  For data we handle the reservation through clearing the
4879  * delalloc bits in the io_tree.  We have to do this since we could end up
4880  * allocating less disk space for the amount of data we have reserved in the
4881  * case of compression.
4882  *
4883  * If this is a reservation and the block group has become read only we cannot
4884  * make the reservation and return -EAGAIN, otherwise this function always
4885  * succeeds.
4886  */
4887 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4888                                        u64 num_bytes, int reserve)
4889 {
4890         struct btrfs_space_info *space_info = cache->space_info;
4891         int ret = 0;
4892
4893         spin_lock(&space_info->lock);
4894         spin_lock(&cache->lock);
4895         if (reserve != RESERVE_FREE) {
4896                 if (cache->ro) {
4897                         ret = -EAGAIN;
4898                 } else {
4899                         cache->reserved += num_bytes;
4900                         space_info->bytes_reserved += num_bytes;
4901                         if (reserve == RESERVE_ALLOC) {
4902                                 trace_btrfs_space_reservation(cache->fs_info,
4903                                                 "space_info", space_info->flags,
4904                                                 num_bytes, 0);
4905                                 space_info->bytes_may_use -= num_bytes;
4906                         }
4907                 }
4908         } else {
4909                 if (cache->ro)
4910                         space_info->bytes_readonly += num_bytes;
4911                 cache->reserved -= num_bytes;
4912                 space_info->bytes_reserved -= num_bytes;
4913                 space_info->reservation_progress++;
4914         }
4915         spin_unlock(&cache->lock);
4916         spin_unlock(&space_info->lock);
4917         return ret;
4918 }
4919
4920 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4921                                 struct btrfs_root *root)
4922 {
4923         struct btrfs_fs_info *fs_info = root->fs_info;
4924         struct btrfs_caching_control *next;
4925         struct btrfs_caching_control *caching_ctl;
4926         struct btrfs_block_group_cache *cache;
4927
4928         down_write(&fs_info->extent_commit_sem);
4929
4930         list_for_each_entry_safe(caching_ctl, next,
4931                                  &fs_info->caching_block_groups, list) {
4932                 cache = caching_ctl->block_group;
4933                 if (block_group_cache_done(cache)) {
4934                         cache->last_byte_to_unpin = (u64)-1;
4935                         list_del_init(&caching_ctl->list);
4936                         put_caching_control(caching_ctl);
4937                 } else {
4938                         cache->last_byte_to_unpin = caching_ctl->progress;
4939                 }
4940         }
4941
4942         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4943                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4944         else
4945                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4946
4947         up_write(&fs_info->extent_commit_sem);
4948
4949         update_global_block_rsv(fs_info);
4950 }
4951
4952 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4953 {
4954         struct btrfs_fs_info *fs_info = root->fs_info;
4955         struct btrfs_block_group_cache *cache = NULL;
4956         u64 len;
4957
4958         while (start <= end) {
4959                 if (!cache ||
4960                     start >= cache->key.objectid + cache->key.offset) {
4961                         if (cache)
4962                                 btrfs_put_block_group(cache);
4963                         cache = btrfs_lookup_block_group(fs_info, start);
4964                         BUG_ON(!cache); /* Logic error */
4965                 }
4966
4967                 len = cache->key.objectid + cache->key.offset - start;
4968                 len = min(len, end + 1 - start);
4969
4970                 if (start < cache->last_byte_to_unpin) {
4971                         len = min(len, cache->last_byte_to_unpin - start);
4972                         btrfs_add_free_space(cache, start, len);
4973                 }
4974
4975                 start += len;
4976
4977                 spin_lock(&cache->space_info->lock);
4978                 spin_lock(&cache->lock);
4979                 cache->pinned -= len;
4980                 cache->space_info->bytes_pinned -= len;
4981                 if (cache->ro)
4982                         cache->space_info->bytes_readonly += len;
4983                 spin_unlock(&cache->lock);
4984                 spin_unlock(&cache->space_info->lock);
4985         }
4986
4987         if (cache)
4988                 btrfs_put_block_group(cache);
4989         return 0;
4990 }
4991
4992 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4993                                struct btrfs_root *root)
4994 {
4995         struct btrfs_fs_info *fs_info = root->fs_info;
4996         struct extent_io_tree *unpin;
4997         u64 start;
4998         u64 end;
4999         int ret;
5000
5001         if (trans->aborted)
5002                 return 0;
5003
5004         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5005                 unpin = &fs_info->freed_extents[1];
5006         else
5007                 unpin = &fs_info->freed_extents[0];
5008
5009         while (1) {
5010                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5011                                             EXTENT_DIRTY);
5012                 if (ret)
5013                         break;
5014
5015                 if (btrfs_test_opt(root, DISCARD))
5016                         ret = btrfs_discard_extent(root, start,
5017                                                    end + 1 - start, NULL);
5018
5019                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5020                 unpin_extent_range(root, start, end);
5021                 cond_resched();
5022         }
5023
5024         return 0;
5025 }
5026
5027 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5028                                 struct btrfs_root *root,
5029                                 u64 bytenr, u64 num_bytes, u64 parent,
5030                                 u64 root_objectid, u64 owner_objectid,
5031                                 u64 owner_offset, int refs_to_drop,
5032                                 struct btrfs_delayed_extent_op *extent_op)
5033 {
5034         struct btrfs_key key;
5035         struct btrfs_path *path;
5036         struct btrfs_fs_info *info = root->fs_info;
5037         struct btrfs_root *extent_root = info->extent_root;
5038         struct extent_buffer *leaf;
5039         struct btrfs_extent_item *ei;
5040         struct btrfs_extent_inline_ref *iref;
5041         int ret;
5042         int is_data;
5043         int extent_slot = 0;
5044         int found_extent = 0;
5045         int num_to_del = 1;
5046         u32 item_size;
5047         u64 refs;
5048
5049         path = btrfs_alloc_path();
5050         if (!path)
5051                 return -ENOMEM;
5052
5053         path->reada = 1;
5054         path->leave_spinning = 1;
5055
5056         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5057         BUG_ON(!is_data && refs_to_drop != 1);
5058
5059         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5060                                     bytenr, num_bytes, parent,
5061                                     root_objectid, owner_objectid,
5062                                     owner_offset);
5063         if (ret == 0) {
5064                 extent_slot = path->slots[0];
5065                 while (extent_slot >= 0) {
5066                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5067                                               extent_slot);
5068                         if (key.objectid != bytenr)
5069                                 break;
5070                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5071                             key.offset == num_bytes) {
5072                                 found_extent = 1;
5073                                 break;
5074                         }
5075                         if (path->slots[0] - extent_slot > 5)
5076                                 break;
5077                         extent_slot--;
5078                 }
5079 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5080                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5081                 if (found_extent && item_size < sizeof(*ei))
5082                         found_extent = 0;
5083 #endif
5084                 if (!found_extent) {
5085                         BUG_ON(iref);
5086                         ret = remove_extent_backref(trans, extent_root, path,
5087                                                     NULL, refs_to_drop,
5088                                                     is_data);
5089                         if (ret)
5090                                 goto abort;
5091                         btrfs_release_path(path);
5092                         path->leave_spinning = 1;
5093
5094                         key.objectid = bytenr;
5095                         key.type = BTRFS_EXTENT_ITEM_KEY;
5096                         key.offset = num_bytes;
5097
5098                         ret = btrfs_search_slot(trans, extent_root,
5099                                                 &key, path, -1, 1);
5100                         if (ret) {
5101                                 printk(KERN_ERR "umm, got %d back from search"
5102                                        ", was looking for %llu\n", ret,
5103                                        (unsigned long long)bytenr);
5104                                 if (ret > 0)
5105                                         btrfs_print_leaf(extent_root,
5106                                                          path->nodes[0]);
5107                         }
5108                         if (ret < 0)
5109                                 goto abort;
5110                         extent_slot = path->slots[0];
5111                 }
5112         } else if (ret == -ENOENT) {
5113                 btrfs_print_leaf(extent_root, path->nodes[0]);
5114                 WARN_ON(1);
5115                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5116                        "parent %llu root %llu  owner %llu offset %llu\n",
5117                        (unsigned long long)bytenr,
5118                        (unsigned long long)parent,
5119                        (unsigned long long)root_objectid,
5120                        (unsigned long long)owner_objectid,
5121                        (unsigned long long)owner_offset);
5122         } else {
5123                 goto abort;
5124         }
5125
5126         leaf = path->nodes[0];
5127         item_size = btrfs_item_size_nr(leaf, extent_slot);
5128 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5129         if (item_size < sizeof(*ei)) {
5130                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5131                 ret = convert_extent_item_v0(trans, extent_root, path,
5132                                              owner_objectid, 0);
5133                 if (ret < 0)
5134                         goto abort;
5135
5136                 btrfs_release_path(path);
5137                 path->leave_spinning = 1;
5138
5139                 key.objectid = bytenr;
5140                 key.type = BTRFS_EXTENT_ITEM_KEY;
5141                 key.offset = num_bytes;
5142
5143                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5144                                         -1, 1);
5145                 if (ret) {
5146                         printk(KERN_ERR "umm, got %d back from search"
5147                                ", was looking for %llu\n", ret,
5148                                (unsigned long long)bytenr);
5149                         btrfs_print_leaf(extent_root, path->nodes[0]);
5150                 }
5151                 if (ret < 0)
5152                         goto abort;
5153                 extent_slot = path->slots[0];
5154                 leaf = path->nodes[0];
5155                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5156         }
5157 #endif
5158         BUG_ON(item_size < sizeof(*ei));
5159         ei = btrfs_item_ptr(leaf, extent_slot,
5160                             struct btrfs_extent_item);
5161         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5162                 struct btrfs_tree_block_info *bi;
5163                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5164                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5165                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5166         }
5167
5168         refs = btrfs_extent_refs(leaf, ei);
5169         BUG_ON(refs < refs_to_drop);
5170         refs -= refs_to_drop;
5171
5172         if (refs > 0) {
5173                 if (extent_op)
5174                         __run_delayed_extent_op(extent_op, leaf, ei);
5175                 /*
5176                  * In the case of inline back ref, reference count will
5177                  * be updated by remove_extent_backref
5178                  */
5179                 if (iref) {
5180                         BUG_ON(!found_extent);
5181                 } else {
5182                         btrfs_set_extent_refs(leaf, ei, refs);
5183                         btrfs_mark_buffer_dirty(leaf);
5184                 }
5185                 if (found_extent) {
5186                         ret = remove_extent_backref(trans, extent_root, path,
5187                                                     iref, refs_to_drop,
5188                                                     is_data);
5189                         if (ret)
5190                                 goto abort;
5191                 }
5192         } else {
5193                 if (found_extent) {
5194                         BUG_ON(is_data && refs_to_drop !=
5195                                extent_data_ref_count(root, path, iref));
5196                         if (iref) {
5197                                 BUG_ON(path->slots[0] != extent_slot);
5198                         } else {
5199                                 BUG_ON(path->slots[0] != extent_slot + 1);
5200                                 path->slots[0] = extent_slot;
5201                                 num_to_del = 2;
5202                         }
5203                 }
5204
5205                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5206                                       num_to_del);
5207                 if (ret)
5208                         goto abort;
5209                 btrfs_release_path(path);
5210
5211                 if (is_data) {
5212                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5213                         if (ret)
5214                                 goto abort;
5215                 }
5216
5217                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5218                 if (ret)
5219                         goto abort;
5220         }
5221 out:
5222         btrfs_free_path(path);
5223         return ret;
5224
5225 abort:
5226         btrfs_abort_transaction(trans, extent_root, ret);
5227         goto out;
5228 }
5229
5230 /*
5231  * when we free an block, it is possible (and likely) that we free the last
5232  * delayed ref for that extent as well.  This searches the delayed ref tree for
5233  * a given extent, and if there are no other delayed refs to be processed, it
5234  * removes it from the tree.
5235  */
5236 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5237                                       struct btrfs_root *root, u64 bytenr)
5238 {
5239         struct btrfs_delayed_ref_head *head;
5240         struct btrfs_delayed_ref_root *delayed_refs;
5241         struct btrfs_delayed_ref_node *ref;
5242         struct rb_node *node;
5243         int ret = 0;
5244
5245         delayed_refs = &trans->transaction->delayed_refs;
5246         spin_lock(&delayed_refs->lock);
5247         head = btrfs_find_delayed_ref_head(trans, bytenr);
5248         if (!head)
5249                 goto out;
5250
5251         node = rb_prev(&head->node.rb_node);
5252         if (!node)
5253                 goto out;
5254
5255         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5256
5257         /* there are still entries for this ref, we can't drop it */
5258         if (ref->bytenr == bytenr)
5259                 goto out;
5260
5261         if (head->extent_op) {
5262                 if (!head->must_insert_reserved)
5263                         goto out;
5264                 kfree(head->extent_op);
5265                 head->extent_op = NULL;
5266         }
5267
5268         /*
5269          * waiting for the lock here would deadlock.  If someone else has it
5270          * locked they are already in the process of dropping it anyway
5271          */
5272         if (!mutex_trylock(&head->mutex))
5273                 goto out;
5274
5275         /*
5276          * at this point we have a head with no other entries.  Go
5277          * ahead and process it.
5278          */
5279         head->node.in_tree = 0;
5280         rb_erase(&head->node.rb_node, &delayed_refs->root);
5281
5282         delayed_refs->num_entries--;
5283
5284         /*
5285          * we don't take a ref on the node because we're removing it from the
5286          * tree, so we just steal the ref the tree was holding.
5287          */
5288         delayed_refs->num_heads--;
5289         if (list_empty(&head->cluster))
5290                 delayed_refs->num_heads_ready--;
5291
5292         list_del_init(&head->cluster);
5293         spin_unlock(&delayed_refs->lock);
5294
5295         BUG_ON(head->extent_op);
5296         if (head->must_insert_reserved)
5297                 ret = 1;
5298
5299         mutex_unlock(&head->mutex);
5300         btrfs_put_delayed_ref(&head->node);
5301         return ret;
5302 out:
5303         spin_unlock(&delayed_refs->lock);
5304         return 0;
5305 }
5306
5307 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5308                            struct btrfs_root *root,
5309                            struct extent_buffer *buf,
5310                            u64 parent, int last_ref)
5311 {
5312         struct btrfs_block_group_cache *cache = NULL;
5313         int ret;
5314
5315         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5316                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5317                                         buf->start, buf->len,
5318                                         parent, root->root_key.objectid,
5319                                         btrfs_header_level(buf),
5320                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5321                 BUG_ON(ret); /* -ENOMEM */
5322         }
5323
5324         if (!last_ref)
5325                 return;
5326
5327         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5328
5329         if (btrfs_header_generation(buf) == trans->transid) {
5330                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5331                         ret = check_ref_cleanup(trans, root, buf->start);
5332                         if (!ret)
5333                                 goto out;
5334                 }
5335
5336                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5337                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5338                         goto out;
5339                 }
5340
5341                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5342
5343                 btrfs_add_free_space(cache, buf->start, buf->len);
5344                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5345         }
5346 out:
5347         /*
5348          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5349          * anymore.
5350          */
5351         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5352         btrfs_put_block_group(cache);
5353 }
5354
5355 /* Can return -ENOMEM */
5356 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5357                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5358                       u64 owner, u64 offset, int for_cow)
5359 {
5360         int ret;
5361         struct btrfs_fs_info *fs_info = root->fs_info;
5362
5363         /*
5364          * tree log blocks never actually go into the extent allocation
5365          * tree, just update pinning info and exit early.
5366          */
5367         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5368                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5369                 /* unlocks the pinned mutex */
5370                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5371                 ret = 0;
5372         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5373                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5374                                         num_bytes,
5375                                         parent, root_objectid, (int)owner,
5376                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5377         } else {
5378                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5379                                                 num_bytes,
5380                                                 parent, root_objectid, owner,
5381                                                 offset, BTRFS_DROP_DELAYED_REF,
5382                                                 NULL, for_cow);
5383         }
5384         return ret;
5385 }
5386
5387 static u64 stripe_align(struct btrfs_root *root, u64 val)
5388 {
5389         u64 mask = ((u64)root->stripesize - 1);
5390         u64 ret = (val + mask) & ~mask;
5391         return ret;
5392 }
5393
5394 /*
5395  * when we wait for progress in the block group caching, its because
5396  * our allocation attempt failed at least once.  So, we must sleep
5397  * and let some progress happen before we try again.
5398  *
5399  * This function will sleep at least once waiting for new free space to
5400  * show up, and then it will check the block group free space numbers
5401  * for our min num_bytes.  Another option is to have it go ahead
5402  * and look in the rbtree for a free extent of a given size, but this
5403  * is a good start.
5404  */
5405 static noinline int
5406 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5407                                 u64 num_bytes)
5408 {
5409         struct btrfs_caching_control *caching_ctl;
5410         DEFINE_WAIT(wait);
5411
5412         caching_ctl = get_caching_control(cache);
5413         if (!caching_ctl)
5414                 return 0;
5415
5416         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5417                    (cache->free_space_ctl->free_space >= num_bytes));
5418
5419         put_caching_control(caching_ctl);
5420         return 0;
5421 }
5422
5423 static noinline int
5424 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5425 {
5426         struct btrfs_caching_control *caching_ctl;
5427         DEFINE_WAIT(wait);
5428
5429         caching_ctl = get_caching_control(cache);
5430         if (!caching_ctl)
5431                 return 0;
5432
5433         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5434
5435         put_caching_control(caching_ctl);
5436         return 0;
5437 }
5438
5439 static int __get_block_group_index(u64 flags)
5440 {
5441         int index;
5442
5443         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5444                 index = 0;
5445         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5446                 index = 1;
5447         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5448                 index = 2;
5449         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5450                 index = 3;
5451         else
5452                 index = 4;
5453
5454         return index;
5455 }
5456
5457 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5458 {
5459         return __get_block_group_index(cache->flags);
5460 }
5461
5462 enum btrfs_loop_type {
5463         LOOP_CACHING_NOWAIT = 0,
5464         LOOP_CACHING_WAIT = 1,
5465         LOOP_ALLOC_CHUNK = 2,
5466         LOOP_NO_EMPTY_SIZE = 3,
5467 };
5468
5469 /*
5470  * walks the btree of allocated extents and find a hole of a given size.
5471  * The key ins is changed to record the hole:
5472  * ins->objectid == block start
5473  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5474  * ins->offset == number of blocks
5475  * Any available blocks before search_start are skipped.
5476  */
5477 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5478                                      struct btrfs_root *orig_root,
5479                                      u64 num_bytes, u64 empty_size,
5480                                      u64 hint_byte, struct btrfs_key *ins,
5481                                      u64 data)
5482 {
5483         int ret = 0;
5484         struct btrfs_root *root = orig_root->fs_info->extent_root;
5485         struct btrfs_free_cluster *last_ptr = NULL;
5486         struct btrfs_block_group_cache *block_group = NULL;
5487         struct btrfs_block_group_cache *used_block_group;
5488         u64 search_start = 0;
5489         int empty_cluster = 2 * 1024 * 1024;
5490         int allowed_chunk_alloc = 0;
5491         int done_chunk_alloc = 0;
5492         struct btrfs_space_info *space_info;
5493         int loop = 0;
5494         int index = 0;
5495         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5496                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5497         bool found_uncached_bg = false;
5498         bool failed_cluster_refill = false;
5499         bool failed_alloc = false;
5500         bool use_cluster = true;
5501         bool have_caching_bg = false;
5502
5503         WARN_ON(num_bytes < root->sectorsize);
5504         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5505         ins->objectid = 0;
5506         ins->offset = 0;
5507
5508         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5509
5510         space_info = __find_space_info(root->fs_info, data);
5511         if (!space_info) {
5512                 printk(KERN_ERR "No space info for %llu\n", data);
5513                 return -ENOSPC;
5514         }
5515
5516         /*
5517          * If the space info is for both data and metadata it means we have a
5518          * small filesystem and we can't use the clustering stuff.
5519          */
5520         if (btrfs_mixed_space_info(space_info))
5521                 use_cluster = false;
5522
5523         if (orig_root->ref_cows || empty_size)
5524                 allowed_chunk_alloc = 1;
5525
5526         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5527                 last_ptr = &root->fs_info->meta_alloc_cluster;
5528                 if (!btrfs_test_opt(root, SSD))
5529                         empty_cluster = 64 * 1024;
5530         }
5531
5532         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5533             btrfs_test_opt(root, SSD)) {
5534                 last_ptr = &root->fs_info->data_alloc_cluster;
5535         }
5536
5537         if (last_ptr) {
5538                 spin_lock(&last_ptr->lock);
5539                 if (last_ptr->block_group)
5540                         hint_byte = last_ptr->window_start;
5541                 spin_unlock(&last_ptr->lock);
5542         }
5543
5544         search_start = max(search_start, first_logical_byte(root, 0));
5545         search_start = max(search_start, hint_byte);
5546
5547         if (!last_ptr)
5548                 empty_cluster = 0;
5549
5550         if (search_start == hint_byte) {
5551                 block_group = btrfs_lookup_block_group(root->fs_info,
5552                                                        search_start);
5553                 used_block_group = block_group;
5554                 /*
5555                  * we don't want to use the block group if it doesn't match our
5556                  * allocation bits, or if its not cached.
5557                  *
5558                  * However if we are re-searching with an ideal block group
5559                  * picked out then we don't care that the block group is cached.
5560                  */
5561                 if (block_group && block_group_bits(block_group, data) &&
5562                     block_group->cached != BTRFS_CACHE_NO) {
5563                         down_read(&space_info->groups_sem);
5564                         if (list_empty(&block_group->list) ||
5565                             block_group->ro) {
5566                                 /*
5567                                  * someone is removing this block group,
5568                                  * we can't jump into the have_block_group
5569                                  * target because our list pointers are not
5570                                  * valid
5571                                  */
5572                                 btrfs_put_block_group(block_group);
5573                                 up_read(&space_info->groups_sem);
5574                         } else {
5575                                 index = get_block_group_index(block_group);
5576                                 goto have_block_group;
5577                         }
5578                 } else if (block_group) {
5579                         btrfs_put_block_group(block_group);
5580                 }
5581         }
5582 search:
5583         have_caching_bg = false;
5584         down_read(&space_info->groups_sem);
5585         list_for_each_entry(block_group, &space_info->block_groups[index],
5586                             list) {
5587                 u64 offset;
5588                 int cached;
5589
5590                 used_block_group = block_group;
5591                 btrfs_get_block_group(block_group);
5592                 search_start = block_group->key.objectid;
5593
5594                 /*
5595                  * this can happen if we end up cycling through all the
5596                  * raid types, but we want to make sure we only allocate
5597                  * for the proper type.
5598                  */
5599                 if (!block_group_bits(block_group, data)) {
5600                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5601                                 BTRFS_BLOCK_GROUP_RAID1 |
5602                                 BTRFS_BLOCK_GROUP_RAID10;
5603
5604                         /*
5605                          * if they asked for extra copies and this block group
5606                          * doesn't provide them, bail.  This does allow us to
5607                          * fill raid0 from raid1.
5608                          */
5609                         if ((data & extra) && !(block_group->flags & extra))
5610                                 goto loop;
5611                 }
5612
5613 have_block_group:
5614                 cached = block_group_cache_done(block_group);
5615                 if (unlikely(!cached)) {
5616                         found_uncached_bg = true;
5617                         ret = cache_block_group(block_group, trans,
5618                                                 orig_root, 0);
5619                         BUG_ON(ret < 0);
5620                         ret = 0;
5621                 }
5622
5623                 if (unlikely(block_group->ro))
5624                         goto loop;
5625
5626                 /*
5627                  * Ok we want to try and use the cluster allocator, so
5628                  * lets look there
5629                  */
5630                 if (last_ptr) {
5631                         /*
5632                          * the refill lock keeps out other
5633                          * people trying to start a new cluster
5634                          */
5635                         spin_lock(&last_ptr->refill_lock);
5636                         used_block_group = last_ptr->block_group;
5637                         if (used_block_group != block_group &&
5638                             (!used_block_group ||
5639                              used_block_group->ro ||
5640                              !block_group_bits(used_block_group, data))) {
5641                                 used_block_group = block_group;
5642                                 goto refill_cluster;
5643                         }
5644
5645                         if (used_block_group != block_group)
5646                                 btrfs_get_block_group(used_block_group);
5647
5648                         offset = btrfs_alloc_from_cluster(used_block_group,
5649                           last_ptr, num_bytes, used_block_group->key.objectid);
5650                         if (offset) {
5651                                 /* we have a block, we're done */
5652                                 spin_unlock(&last_ptr->refill_lock);
5653                                 trace_btrfs_reserve_extent_cluster(root,
5654                                         block_group, search_start, num_bytes);
5655                                 goto checks;
5656                         }
5657
5658                         WARN_ON(last_ptr->block_group != used_block_group);
5659                         if (used_block_group != block_group) {
5660                                 btrfs_put_block_group(used_block_group);
5661                                 used_block_group = block_group;
5662                         }
5663 refill_cluster:
5664                         BUG_ON(used_block_group != block_group);
5665                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5666                          * set up a new clusters, so lets just skip it
5667                          * and let the allocator find whatever block
5668                          * it can find.  If we reach this point, we
5669                          * will have tried the cluster allocator
5670                          * plenty of times and not have found
5671