Btrfs: fix deadlock during allocating chunks
[linux-3.10.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36
37 /*
38  * control flags for do_chunk_alloc's force field
39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
40  * if we really need one.
41  *
42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
43  * if we have very few chunks already allocated.  This is
44  * used as part of the clustering code to help make sure
45  * we have a good pool of storage to cluster in, without
46  * filling the FS with empty chunks
47  *
48  * CHUNK_ALLOC_FORCE means it must try to allocate one
49  *
50  */
51 enum {
52         CHUNK_ALLOC_NO_FORCE = 0,
53         CHUNK_ALLOC_LIMITED = 1,
54         CHUNK_ALLOC_FORCE = 2,
55 };
56
57 /*
58  * Control how reservations are dealt with.
59  *
60  * RESERVE_FREE - freeing a reservation.
61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
62  *   ENOSPC accounting
63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
65  */
66 enum {
67         RESERVE_FREE = 0,
68         RESERVE_ALLOC = 1,
69         RESERVE_ALLOC_NO_ACCOUNT = 2,
70 };
71
72 static int update_block_group(struct btrfs_trans_handle *trans,
73                               struct btrfs_root *root,
74                               u64 bytenr, u64 num_bytes, int alloc);
75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
76                                 struct btrfs_root *root,
77                                 u64 bytenr, u64 num_bytes, u64 parent,
78                                 u64 root_objectid, u64 owner_objectid,
79                                 u64 owner_offset, int refs_to_drop,
80                                 struct btrfs_delayed_extent_op *extra_op);
81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
82                                     struct extent_buffer *leaf,
83                                     struct btrfs_extent_item *ei);
84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
85                                       struct btrfs_root *root,
86                                       u64 parent, u64 root_objectid,
87                                       u64 flags, u64 owner, u64 offset,
88                                       struct btrfs_key *ins, int ref_mod);
89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
90                                      struct btrfs_root *root,
91                                      u64 parent, u64 root_objectid,
92                                      u64 flags, struct btrfs_disk_key *key,
93                                      int level, struct btrfs_key *ins);
94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
95                           struct btrfs_root *extent_root, u64 alloc_bytes,
96                           u64 flags, int force);
97 static int find_next_key(struct btrfs_path *path, int level,
98                          struct btrfs_key *key);
99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
100                             int dump_block_groups);
101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
102                                        u64 num_bytes, int reserve);
103
104 static noinline int
105 block_group_cache_done(struct btrfs_block_group_cache *cache)
106 {
107         smp_mb();
108         return cache->cached == BTRFS_CACHE_FINISHED;
109 }
110
111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
112 {
113         return (cache->flags & bits) == bits;
114 }
115
116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
117 {
118         atomic_inc(&cache->count);
119 }
120
121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
122 {
123         if (atomic_dec_and_test(&cache->count)) {
124                 WARN_ON(cache->pinned > 0);
125                 WARN_ON(cache->reserved > 0);
126                 kfree(cache->free_space_ctl);
127                 kfree(cache);
128         }
129 }
130
131 /*
132  * this adds the block group to the fs_info rb tree for the block group
133  * cache
134  */
135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
136                                 struct btrfs_block_group_cache *block_group)
137 {
138         struct rb_node **p;
139         struct rb_node *parent = NULL;
140         struct btrfs_block_group_cache *cache;
141
142         spin_lock(&info->block_group_cache_lock);
143         p = &info->block_group_cache_tree.rb_node;
144
145         while (*p) {
146                 parent = *p;
147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
148                                  cache_node);
149                 if (block_group->key.objectid < cache->key.objectid) {
150                         p = &(*p)->rb_left;
151                 } else if (block_group->key.objectid > cache->key.objectid) {
152                         p = &(*p)->rb_right;
153                 } else {
154                         spin_unlock(&info->block_group_cache_lock);
155                         return -EEXIST;
156                 }
157         }
158
159         rb_link_node(&block_group->cache_node, parent, p);
160         rb_insert_color(&block_group->cache_node,
161                         &info->block_group_cache_tree);
162         spin_unlock(&info->block_group_cache_lock);
163
164         return 0;
165 }
166
167 /*
168  * This will return the block group at or after bytenr if contains is 0, else
169  * it will return the block group that contains the bytenr
170  */
171 static struct btrfs_block_group_cache *
172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
173                               int contains)
174 {
175         struct btrfs_block_group_cache *cache, *ret = NULL;
176         struct rb_node *n;
177         u64 end, start;
178
179         spin_lock(&info->block_group_cache_lock);
180         n = info->block_group_cache_tree.rb_node;
181
182         while (n) {
183                 cache = rb_entry(n, struct btrfs_block_group_cache,
184                                  cache_node);
185                 end = cache->key.objectid + cache->key.offset - 1;
186                 start = cache->key.objectid;
187
188                 if (bytenr < start) {
189                         if (!contains && (!ret || start < ret->key.objectid))
190                                 ret = cache;
191                         n = n->rb_left;
192                 } else if (bytenr > start) {
193                         if (contains && bytenr <= end) {
194                                 ret = cache;
195                                 break;
196                         }
197                         n = n->rb_right;
198                 } else {
199                         ret = cache;
200                         break;
201                 }
202         }
203         if (ret)
204                 btrfs_get_block_group(ret);
205         spin_unlock(&info->block_group_cache_lock);
206
207         return ret;
208 }
209
210 static int add_excluded_extent(struct btrfs_root *root,
211                                u64 start, u64 num_bytes)
212 {
213         u64 end = start + num_bytes - 1;
214         set_extent_bits(&root->fs_info->freed_extents[0],
215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
216         set_extent_bits(&root->fs_info->freed_extents[1],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         return 0;
219 }
220
221 static void free_excluded_extents(struct btrfs_root *root,
222                                   struct btrfs_block_group_cache *cache)
223 {
224         u64 start, end;
225
226         start = cache->key.objectid;
227         end = start + cache->key.offset - 1;
228
229         clear_extent_bits(&root->fs_info->freed_extents[0],
230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
231         clear_extent_bits(&root->fs_info->freed_extents[1],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233 }
234
235 static int exclude_super_stripes(struct btrfs_root *root,
236                                  struct btrfs_block_group_cache *cache)
237 {
238         u64 bytenr;
239         u64 *logical;
240         int stripe_len;
241         int i, nr, ret;
242
243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
245                 cache->bytes_super += stripe_len;
246                 ret = add_excluded_extent(root, cache->key.objectid,
247                                           stripe_len);
248                 BUG_ON(ret); /* -ENOMEM */
249         }
250
251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252                 bytenr = btrfs_sb_offset(i);
253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
254                                        cache->key.objectid, bytenr,
255                                        0, &logical, &nr, &stripe_len);
256                 BUG_ON(ret); /* -ENOMEM */
257
258                 while (nr--) {
259                         cache->bytes_super += stripe_len;
260                         ret = add_excluded_extent(root, logical[nr],
261                                                   stripe_len);
262                         BUG_ON(ret); /* -ENOMEM */
263                 }
264
265                 kfree(logical);
266         }
267         return 0;
268 }
269
270 static struct btrfs_caching_control *
271 get_caching_control(struct btrfs_block_group_cache *cache)
272 {
273         struct btrfs_caching_control *ctl;
274
275         spin_lock(&cache->lock);
276         if (cache->cached != BTRFS_CACHE_STARTED) {
277                 spin_unlock(&cache->lock);
278                 return NULL;
279         }
280
281         /* We're loading it the fast way, so we don't have a caching_ctl. */
282         if (!cache->caching_ctl) {
283                 spin_unlock(&cache->lock);
284                 return NULL;
285         }
286
287         ctl = cache->caching_ctl;
288         atomic_inc(&ctl->count);
289         spin_unlock(&cache->lock);
290         return ctl;
291 }
292
293 static void put_caching_control(struct btrfs_caching_control *ctl)
294 {
295         if (atomic_dec_and_test(&ctl->count))
296                 kfree(ctl);
297 }
298
299 /*
300  * this is only called by cache_block_group, since we could have freed extents
301  * we need to check the pinned_extents for any extents that can't be used yet
302  * since their free space will be released as soon as the transaction commits.
303  */
304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
305                               struct btrfs_fs_info *info, u64 start, u64 end)
306 {
307         u64 extent_start, extent_end, size, total_added = 0;
308         int ret;
309
310         while (start < end) {
311                 ret = find_first_extent_bit(info->pinned_extents, start,
312                                             &extent_start, &extent_end,
313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
314                 if (ret)
315                         break;
316
317                 if (extent_start <= start) {
318                         start = extent_end + 1;
319                 } else if (extent_start > start && extent_start < end) {
320                         size = extent_start - start;
321                         total_added += size;
322                         ret = btrfs_add_free_space(block_group, start,
323                                                    size);
324                         BUG_ON(ret); /* -ENOMEM or logic error */
325                         start = extent_end + 1;
326                 } else {
327                         break;
328                 }
329         }
330
331         if (start < end) {
332                 size = end - start;
333                 total_added += size;
334                 ret = btrfs_add_free_space(block_group, start, size);
335                 BUG_ON(ret); /* -ENOMEM or logic error */
336         }
337
338         return total_added;
339 }
340
341 static noinline void caching_thread(struct btrfs_work *work)
342 {
343         struct btrfs_block_group_cache *block_group;
344         struct btrfs_fs_info *fs_info;
345         struct btrfs_caching_control *caching_ctl;
346         struct btrfs_root *extent_root;
347         struct btrfs_path *path;
348         struct extent_buffer *leaf;
349         struct btrfs_key key;
350         u64 total_found = 0;
351         u64 last = 0;
352         u32 nritems;
353         int ret = 0;
354
355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
356         block_group = caching_ctl->block_group;
357         fs_info = block_group->fs_info;
358         extent_root = fs_info->extent_root;
359
360         path = btrfs_alloc_path();
361         if (!path)
362                 goto out;
363
364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
365
366         /*
367          * We don't want to deadlock with somebody trying to allocate a new
368          * extent for the extent root while also trying to search the extent
369          * root to add free space.  So we skip locking and search the commit
370          * root, since its read-only
371          */
372         path->skip_locking = 1;
373         path->search_commit_root = 1;
374         path->reada = 1;
375
376         key.objectid = last;
377         key.offset = 0;
378         key.type = BTRFS_EXTENT_ITEM_KEY;
379 again:
380         mutex_lock(&caching_ctl->mutex);
381         /* need to make sure the commit_root doesn't disappear */
382         down_read(&fs_info->extent_commit_sem);
383
384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
385         if (ret < 0)
386                 goto err;
387
388         leaf = path->nodes[0];
389         nritems = btrfs_header_nritems(leaf);
390
391         while (1) {
392                 if (btrfs_fs_closing(fs_info) > 1) {
393                         last = (u64)-1;
394                         break;
395                 }
396
397                 if (path->slots[0] < nritems) {
398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
399                 } else {
400                         ret = find_next_key(path, 0, &key);
401                         if (ret)
402                                 break;
403
404                         if (need_resched() ||
405                             btrfs_next_leaf(extent_root, path)) {
406                                 caching_ctl->progress = last;
407                                 btrfs_release_path(path);
408                                 up_read(&fs_info->extent_commit_sem);
409                                 mutex_unlock(&caching_ctl->mutex);
410                                 cond_resched();
411                                 goto again;
412                         }
413                         leaf = path->nodes[0];
414                         nritems = btrfs_header_nritems(leaf);
415                         continue;
416                 }
417
418                 if (key.objectid < block_group->key.objectid) {
419                         path->slots[0]++;
420                         continue;
421                 }
422
423                 if (key.objectid >= block_group->key.objectid +
424                     block_group->key.offset)
425                         break;
426
427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
428                         total_found += add_new_free_space(block_group,
429                                                           fs_info, last,
430                                                           key.objectid);
431                         last = key.objectid + key.offset;
432
433                         if (total_found > (1024 * 1024 * 2)) {
434                                 total_found = 0;
435                                 wake_up(&caching_ctl->wait);
436                         }
437                 }
438                 path->slots[0]++;
439         }
440         ret = 0;
441
442         total_found += add_new_free_space(block_group, fs_info, last,
443                                           block_group->key.objectid +
444                                           block_group->key.offset);
445         caching_ctl->progress = (u64)-1;
446
447         spin_lock(&block_group->lock);
448         block_group->caching_ctl = NULL;
449         block_group->cached = BTRFS_CACHE_FINISHED;
450         spin_unlock(&block_group->lock);
451
452 err:
453         btrfs_free_path(path);
454         up_read(&fs_info->extent_commit_sem);
455
456         free_excluded_extents(extent_root, block_group);
457
458         mutex_unlock(&caching_ctl->mutex);
459 out:
460         wake_up(&caching_ctl->wait);
461
462         put_caching_control(caching_ctl);
463         btrfs_put_block_group(block_group);
464 }
465
466 static int cache_block_group(struct btrfs_block_group_cache *cache,
467                              struct btrfs_trans_handle *trans,
468                              struct btrfs_root *root,
469                              int load_cache_only)
470 {
471         DEFINE_WAIT(wait);
472         struct btrfs_fs_info *fs_info = cache->fs_info;
473         struct btrfs_caching_control *caching_ctl;
474         int ret = 0;
475
476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
477         if (!caching_ctl)
478                 return -ENOMEM;
479
480         INIT_LIST_HEAD(&caching_ctl->list);
481         mutex_init(&caching_ctl->mutex);
482         init_waitqueue_head(&caching_ctl->wait);
483         caching_ctl->block_group = cache;
484         caching_ctl->progress = cache->key.objectid;
485         atomic_set(&caching_ctl->count, 1);
486         caching_ctl->work.func = caching_thread;
487
488         spin_lock(&cache->lock);
489         /*
490          * This should be a rare occasion, but this could happen I think in the
491          * case where one thread starts to load the space cache info, and then
492          * some other thread starts a transaction commit which tries to do an
493          * allocation while the other thread is still loading the space cache
494          * info.  The previous loop should have kept us from choosing this block
495          * group, but if we've moved to the state where we will wait on caching
496          * block groups we need to first check if we're doing a fast load here,
497          * so we can wait for it to finish, otherwise we could end up allocating
498          * from a block group who's cache gets evicted for one reason or
499          * another.
500          */
501         while (cache->cached == BTRFS_CACHE_FAST) {
502                 struct btrfs_caching_control *ctl;
503
504                 ctl = cache->caching_ctl;
505                 atomic_inc(&ctl->count);
506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
507                 spin_unlock(&cache->lock);
508
509                 schedule();
510
511                 finish_wait(&ctl->wait, &wait);
512                 put_caching_control(ctl);
513                 spin_lock(&cache->lock);
514         }
515
516         if (cache->cached != BTRFS_CACHE_NO) {
517                 spin_unlock(&cache->lock);
518                 kfree(caching_ctl);
519                 return 0;
520         }
521         WARN_ON(cache->caching_ctl);
522         cache->caching_ctl = caching_ctl;
523         cache->cached = BTRFS_CACHE_FAST;
524         spin_unlock(&cache->lock);
525
526         /*
527          * We can't do the read from on-disk cache during a commit since we need
528          * to have the normal tree locking.  Also if we are currently trying to
529          * allocate blocks for the tree root we can't do the fast caching since
530          * we likely hold important locks.
531          */
532         if (trans && (!trans->transaction->in_commit) &&
533             (root && root != root->fs_info->tree_root) &&
534             btrfs_test_opt(root, 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         int ret;
2223         int count = 0;
2224         int must_insert_reserved = 0;
2225
2226         delayed_refs = &trans->transaction->delayed_refs;
2227         while (1) {
2228                 if (!locked_ref) {
2229                         /* pick a new head ref from the cluster list */
2230                         if (list_empty(cluster))
2231                                 break;
2232
2233                         locked_ref = list_entry(cluster->next,
2234                                      struct btrfs_delayed_ref_head, cluster);
2235
2236                         /* grab the lock that says we are going to process
2237                          * all the refs for this head */
2238                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2239
2240                         /*
2241                          * we may have dropped the spin lock to get the head
2242                          * mutex lock, and that might have given someone else
2243                          * time to free the head.  If that's true, it has been
2244                          * removed from our list and we can move on.
2245                          */
2246                         if (ret == -EAGAIN) {
2247                                 locked_ref = NULL;
2248                                 count++;
2249                                 continue;
2250                         }
2251                 }
2252
2253                 /*
2254                  * locked_ref is the head node, so we have to go one
2255                  * node back for any delayed ref updates
2256                  */
2257                 ref = select_delayed_ref(locked_ref);
2258
2259                 if (ref && ref->seq &&
2260                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2261                         /*
2262                          * there are still refs with lower seq numbers in the
2263                          * process of being added. Don't run this ref yet.
2264                          */
2265                         list_del_init(&locked_ref->cluster);
2266                         mutex_unlock(&locked_ref->mutex);
2267                         locked_ref = NULL;
2268                         delayed_refs->num_heads_ready++;
2269                         spin_unlock(&delayed_refs->lock);
2270                         cond_resched();
2271                         spin_lock(&delayed_refs->lock);
2272                         continue;
2273                 }
2274
2275                 /*
2276                  * record the must insert reserved flag before we
2277                  * drop the spin lock.
2278                  */
2279                 must_insert_reserved = locked_ref->must_insert_reserved;
2280                 locked_ref->must_insert_reserved = 0;
2281
2282                 extent_op = locked_ref->extent_op;
2283                 locked_ref->extent_op = NULL;
2284
2285                 if (!ref) {
2286                         /* All delayed refs have been processed, Go ahead
2287                          * and send the head node to run_one_delayed_ref,
2288                          * so that any accounting fixes can happen
2289                          */
2290                         ref = &locked_ref->node;
2291
2292                         if (extent_op && must_insert_reserved) {
2293                                 kfree(extent_op);
2294                                 extent_op = NULL;
2295                         }
2296
2297                         if (extent_op) {
2298                                 spin_unlock(&delayed_refs->lock);
2299
2300                                 ret = run_delayed_extent_op(trans, root,
2301                                                             ref, extent_op);
2302                                 kfree(extent_op);
2303
2304                                 if (ret) {
2305                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2306                                         return ret;
2307                                 }
2308
2309                                 goto next;
2310                         }
2311
2312                         list_del_init(&locked_ref->cluster);
2313                         locked_ref = NULL;
2314                 }
2315
2316                 ref->in_tree = 0;
2317                 rb_erase(&ref->rb_node, &delayed_refs->root);
2318                 delayed_refs->num_entries--;
2319                 /*
2320                  * we modified num_entries, but as we're currently running
2321                  * delayed refs, skip
2322                  *     wake_up(&delayed_refs->seq_wait);
2323                  * here.
2324                  */
2325                 spin_unlock(&delayed_refs->lock);
2326
2327                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2328                                           must_insert_reserved);
2329
2330                 btrfs_put_delayed_ref(ref);
2331                 kfree(extent_op);
2332                 count++;
2333
2334                 if (ret) {
2335                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350
2351 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2352                         unsigned long num_refs)
2353 {
2354         struct list_head *first_seq = delayed_refs->seq_head.next;
2355
2356         spin_unlock(&delayed_refs->lock);
2357         pr_debug("waiting for more refs (num %ld, first %p)\n",
2358                  num_refs, first_seq);
2359         wait_event(delayed_refs->seq_wait,
2360                    num_refs != delayed_refs->num_entries ||
2361                    delayed_refs->seq_head.next != first_seq);
2362         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2363                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2364         spin_lock(&delayed_refs->lock);
2365 }
2366
2367 /*
2368  * this starts processing the delayed reference count updates and
2369  * extent insertions we have queued up so far.  count can be
2370  * 0, which means to process everything in the tree at the start
2371  * of the run (but not newly added entries), or it can be some target
2372  * number you'd like to process.
2373  *
2374  * Returns 0 on success or if called with an aborted transaction
2375  * Returns <0 on error and aborts the transaction
2376  */
2377 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2378                            struct btrfs_root *root, unsigned long count)
2379 {
2380         struct rb_node *node;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct btrfs_delayed_ref_node *ref;
2383         struct list_head cluster;
2384         int ret;
2385         u64 delayed_start;
2386         int run_all = count == (unsigned long)-1;
2387         int run_most = 0;
2388         unsigned long num_refs = 0;
2389         int consider_waiting;
2390
2391         /* We'll clean this up in btrfs_cleanup_transaction */
2392         if (trans->aborted)
2393                 return 0;
2394
2395         if (root == root->fs_info->extent_root)
2396                 root = root->fs_info->tree_root;
2397
2398         do_chunk_alloc(trans, root->fs_info->extent_root,
2399                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2400                        CHUNK_ALLOC_NO_FORCE);
2401
2402         delayed_refs = &trans->transaction->delayed_refs;
2403         INIT_LIST_HEAD(&cluster);
2404 again:
2405         consider_waiting = 0;
2406         spin_lock(&delayed_refs->lock);
2407         if (count == 0) {
2408                 count = delayed_refs->num_entries * 2;
2409                 run_most = 1;
2410         }
2411         while (1) {
2412                 if (!(run_all || run_most) &&
2413                     delayed_refs->num_heads_ready < 64)
2414                         break;
2415
2416                 /*
2417                  * go find something we can process in the rbtree.  We start at
2418                  * the beginning of the tree, and then build a cluster
2419                  * of refs to process starting at the first one we are able to
2420                  * lock
2421                  */
2422                 delayed_start = delayed_refs->run_delayed_start;
2423                 ret = btrfs_find_ref_cluster(trans, &cluster,
2424                                              delayed_refs->run_delayed_start);
2425                 if (ret)
2426                         break;
2427
2428                 if (delayed_start >= delayed_refs->run_delayed_start) {
2429                         if (consider_waiting == 0) {
2430                                 /*
2431                                  * btrfs_find_ref_cluster looped. let's do one
2432                                  * more cycle. if we don't run any delayed ref
2433                                  * during that cycle (because we can't because
2434                                  * all of them are blocked) and if the number of
2435                                  * refs doesn't change, we avoid busy waiting.
2436                                  */
2437                                 consider_waiting = 1;
2438                                 num_refs = delayed_refs->num_entries;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs, num_refs);
2441                                 /*
2442                                  * after waiting, things have changed. we
2443                                  * dropped the lock and someone else might have
2444                                  * run some refs, built new clusters and so on.
2445                                  * therefore, we restart staleness detection.
2446                                  */
2447                                 consider_waiting = 0;
2448                         }
2449                 }
2450
2451                 ret = run_clustered_refs(trans, root, &cluster);
2452                 if (ret < 0) {
2453                         spin_unlock(&delayed_refs->lock);
2454                         btrfs_abort_transaction(trans, root, ret);
2455                         return ret;
2456                 }
2457
2458                 count -= min_t(unsigned long, ret, count);
2459
2460                 if (count == 0)
2461                         break;
2462
2463                 if (ret || delayed_refs->run_delayed_start == 0) {
2464                         /* refs were run, let's reset staleness detection */
2465                         consider_waiting = 0;
2466                 }
2467         }
2468
2469         if (run_all) {
2470                 node = rb_first(&delayed_refs->root);
2471                 if (!node)
2472                         goto out;
2473                 count = (unsigned long)-1;
2474
2475                 while (node) {
2476                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2477                                        rb_node);
2478                         if (btrfs_delayed_ref_is_head(ref)) {
2479                                 struct btrfs_delayed_ref_head *head;
2480
2481                                 head = btrfs_delayed_node_to_head(ref);
2482                                 atomic_inc(&ref->refs);
2483
2484                                 spin_unlock(&delayed_refs->lock);
2485                                 /*
2486                                  * Mutex was contended, block until it's
2487                                  * released and try again
2488                                  */
2489                                 mutex_lock(&head->mutex);
2490                                 mutex_unlock(&head->mutex);
2491
2492                                 btrfs_put_delayed_ref(ref);
2493                                 cond_resched();
2494                                 goto again;
2495                         }
2496                         node = rb_next(node);
2497                 }
2498                 spin_unlock(&delayed_refs->lock);
2499                 schedule_timeout(1);
2500                 goto again;
2501         }
2502 out:
2503         spin_unlock(&delayed_refs->lock);
2504         return 0;
2505 }
2506
2507 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2508                                 struct btrfs_root *root,
2509                                 u64 bytenr, u64 num_bytes, u64 flags,
2510                                 int is_data)
2511 {
2512         struct btrfs_delayed_extent_op *extent_op;
2513         int ret;
2514
2515         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2516         if (!extent_op)
2517                 return -ENOMEM;
2518
2519         extent_op->flags_to_set = flags;
2520         extent_op->update_flags = 1;
2521         extent_op->update_key = 0;
2522         extent_op->is_data = is_data ? 1 : 0;
2523
2524         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2525                                           num_bytes, extent_op);
2526         if (ret)
2527                 kfree(extent_op);
2528         return ret;
2529 }
2530
2531 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2532                                       struct btrfs_root *root,
2533                                       struct btrfs_path *path,
2534                                       u64 objectid, u64 offset, u64 bytenr)
2535 {
2536         struct btrfs_delayed_ref_head *head;
2537         struct btrfs_delayed_ref_node *ref;
2538         struct btrfs_delayed_data_ref *data_ref;
2539         struct btrfs_delayed_ref_root *delayed_refs;
2540         struct rb_node *node;
2541         int ret = 0;
2542
2543         ret = -ENOENT;
2544         delayed_refs = &trans->transaction->delayed_refs;
2545         spin_lock(&delayed_refs->lock);
2546         head = btrfs_find_delayed_ref_head(trans, bytenr);
2547         if (!head)
2548                 goto out;
2549
2550         if (!mutex_trylock(&head->mutex)) {
2551                 atomic_inc(&head->node.refs);
2552                 spin_unlock(&delayed_refs->lock);
2553
2554                 btrfs_release_path(path);
2555
2556                 /*
2557                  * Mutex was contended, block until it's released and let
2558                  * caller try again
2559                  */
2560                 mutex_lock(&head->mutex);
2561                 mutex_unlock(&head->mutex);
2562                 btrfs_put_delayed_ref(&head->node);
2563                 return -EAGAIN;
2564         }
2565
2566         node = rb_prev(&head->node.rb_node);
2567         if (!node)
2568                 goto out_unlock;
2569
2570         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2571
2572         if (ref->bytenr != bytenr)
2573                 goto out_unlock;
2574
2575         ret = 1;
2576         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2577                 goto out_unlock;
2578
2579         data_ref = btrfs_delayed_node_to_data_ref(ref);
2580
2581         node = rb_prev(node);
2582         if (node) {
2583                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2584                 if (ref->bytenr == bytenr)
2585                         goto out_unlock;
2586         }
2587
2588         if (data_ref->root != root->root_key.objectid ||
2589             data_ref->objectid != objectid || data_ref->offset != offset)
2590                 goto out_unlock;
2591
2592         ret = 0;
2593 out_unlock:
2594         mutex_unlock(&head->mutex);
2595 out:
2596         spin_unlock(&delayed_refs->lock);
2597         return ret;
2598 }
2599
2600 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2601                                         struct btrfs_root *root,
2602                                         struct btrfs_path *path,
2603                                         u64 objectid, u64 offset, u64 bytenr)
2604 {
2605         struct btrfs_root *extent_root = root->fs_info->extent_root;
2606         struct extent_buffer *leaf;
2607         struct btrfs_extent_data_ref *ref;
2608         struct btrfs_extent_inline_ref *iref;
2609         struct btrfs_extent_item *ei;
2610         struct btrfs_key key;
2611         u32 item_size;
2612         int ret;
2613
2614         key.objectid = bytenr;
2615         key.offset = (u64)-1;
2616         key.type = BTRFS_EXTENT_ITEM_KEY;
2617
2618         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2619         if (ret < 0)
2620                 goto out;
2621         BUG_ON(ret == 0); /* Corruption */
2622
2623         ret = -ENOENT;
2624         if (path->slots[0] == 0)
2625                 goto out;
2626
2627         path->slots[0]--;
2628         leaf = path->nodes[0];
2629         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2630
2631         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2632                 goto out;
2633
2634         ret = 1;
2635         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2637         if (item_size < sizeof(*ei)) {
2638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2639                 goto out;
2640         }
2641 #endif
2642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2643
2644         if (item_size != sizeof(*ei) +
2645             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2646                 goto out;
2647
2648         if (btrfs_extent_generation(leaf, ei) <=
2649             btrfs_root_last_snapshot(&root->root_item))
2650                 goto out;
2651
2652         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2653         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2654             BTRFS_EXTENT_DATA_REF_KEY)
2655                 goto out;
2656
2657         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2658         if (btrfs_extent_refs(leaf, ei) !=
2659             btrfs_extent_data_ref_count(leaf, ref) ||
2660             btrfs_extent_data_ref_root(leaf, ref) !=
2661             root->root_key.objectid ||
2662             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2663             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2664                 goto out;
2665
2666         ret = 0;
2667 out:
2668         return ret;
2669 }
2670
2671 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2672                           struct btrfs_root *root,
2673                           u64 objectid, u64 offset, u64 bytenr)
2674 {
2675         struct btrfs_path *path;
2676         int ret;
2677         int ret2;
2678
2679         path = btrfs_alloc_path();
2680         if (!path)
2681                 return -ENOENT;
2682
2683         do {
2684                 ret = check_committed_ref(trans, root, path, objectid,
2685                                           offset, bytenr);
2686                 if (ret && ret != -ENOENT)
2687                         goto out;
2688
2689                 ret2 = check_delayed_ref(trans, root, path, objectid,
2690                                          offset, bytenr);
2691         } while (ret2 == -EAGAIN);
2692
2693         if (ret2 && ret2 != -ENOENT) {
2694                 ret = ret2;
2695                 goto out;
2696         }
2697
2698         if (ret != -ENOENT || ret2 != -ENOENT)
2699                 ret = 0;
2700 out:
2701         btrfs_free_path(path);
2702         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2703                 WARN_ON(ret > 0);
2704         return ret;
2705 }
2706
2707 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2708                            struct btrfs_root *root,
2709                            struct extent_buffer *buf,
2710                            int full_backref, int inc, int for_cow)
2711 {
2712         u64 bytenr;
2713         u64 num_bytes;
2714         u64 parent;
2715         u64 ref_root;
2716         u32 nritems;
2717         struct btrfs_key key;
2718         struct btrfs_file_extent_item *fi;
2719         int i;
2720         int level;
2721         int ret = 0;
2722         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2723                             u64, u64, u64, u64, u64, u64, int);
2724
2725         ref_root = btrfs_header_owner(buf);
2726         nritems = btrfs_header_nritems(buf);
2727         level = btrfs_header_level(buf);
2728
2729         if (!root->ref_cows && level == 0)
2730                 return 0;
2731
2732         if (inc)
2733                 process_func = btrfs_inc_extent_ref;
2734         else
2735                 process_func = btrfs_free_extent;
2736
2737         if (full_backref)
2738                 parent = buf->start;
2739         else
2740                 parent = 0;
2741
2742         for (i = 0; i < nritems; i++) {
2743                 if (level == 0) {
2744                         btrfs_item_key_to_cpu(buf, &key, i);
2745                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2746                                 continue;
2747                         fi = btrfs_item_ptr(buf, i,
2748                                             struct btrfs_file_extent_item);
2749                         if (btrfs_file_extent_type(buf, fi) ==
2750                             BTRFS_FILE_EXTENT_INLINE)
2751                                 continue;
2752                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2753                         if (bytenr == 0)
2754                                 continue;
2755
2756                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2757                         key.offset -= btrfs_file_extent_offset(buf, fi);
2758                         ret = process_func(trans, root, bytenr, num_bytes,
2759                                            parent, ref_root, key.objectid,
2760                                            key.offset, for_cow);
2761                         if (ret)
2762                                 goto fail;
2763                 } else {
2764                         bytenr = btrfs_node_blockptr(buf, i);
2765                         num_bytes = btrfs_level_size(root, level - 1);
2766                         ret = process_func(trans, root, bytenr, num_bytes,
2767                                            parent, ref_root, level - 1, 0,
2768                                            for_cow);
2769                         if (ret)
2770                                 goto fail;
2771                 }
2772         }
2773         return 0;
2774 fail:
2775         return ret;
2776 }
2777
2778 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2779                   struct extent_buffer *buf, int full_backref, int for_cow)
2780 {
2781         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2782 }
2783
2784 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2785                   struct extent_buffer *buf, int full_backref, int for_cow)
2786 {
2787         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2788 }
2789
2790 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2791                                  struct btrfs_root *root,
2792                                  struct btrfs_path *path,
2793                                  struct btrfs_block_group_cache *cache)
2794 {
2795         int ret;
2796         struct btrfs_root *extent_root = root->fs_info->extent_root;
2797         unsigned long bi;
2798         struct extent_buffer *leaf;
2799
2800         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2801         if (ret < 0)
2802                 goto fail;
2803         BUG_ON(ret); /* Corruption */
2804
2805         leaf = path->nodes[0];
2806         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2807         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2808         btrfs_mark_buffer_dirty(leaf);
2809         btrfs_release_path(path);
2810 fail:
2811         if (ret) {
2812                 btrfs_abort_transaction(trans, root, ret);
2813                 return ret;
2814         }
2815         return 0;
2816
2817 }
2818
2819 static struct btrfs_block_group_cache *
2820 next_block_group(struct btrfs_root *root,
2821                  struct btrfs_block_group_cache *cache)
2822 {
2823         struct rb_node *node;
2824         spin_lock(&root->fs_info->block_group_cache_lock);
2825         node = rb_next(&cache->cache_node);
2826         btrfs_put_block_group(cache);
2827         if (node) {
2828                 cache = rb_entry(node, struct btrfs_block_group_cache,
2829                                  cache_node);
2830                 btrfs_get_block_group(cache);
2831         } else
2832                 cache = NULL;
2833         spin_unlock(&root->fs_info->block_group_cache_lock);
2834         return cache;
2835 }
2836
2837 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2838                             struct btrfs_trans_handle *trans,
2839                             struct btrfs_path *path)
2840 {
2841         struct btrfs_root *root = block_group->fs_info->tree_root;
2842         struct inode *inode = NULL;
2843         u64 alloc_hint = 0;
2844         int dcs = BTRFS_DC_ERROR;
2845         int num_pages = 0;
2846         int retries = 0;
2847         int ret = 0;
2848
2849         /*
2850          * If this block group is smaller than 100 megs don't bother caching the
2851          * block group.
2852          */
2853         if (block_group->key.offset < (100 * 1024 * 1024)) {
2854                 spin_lock(&block_group->lock);
2855                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2856                 spin_unlock(&block_group->lock);
2857                 return 0;
2858         }
2859
2860 again:
2861         inode = lookup_free_space_inode(root, block_group, path);
2862         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2863                 ret = PTR_ERR(inode);
2864                 btrfs_release_path(path);
2865                 goto out;
2866         }
2867
2868         if (IS_ERR(inode)) {
2869                 BUG_ON(retries);
2870                 retries++;
2871
2872                 if (block_group->ro)
2873                         goto out_free;
2874
2875                 ret = create_free_space_inode(root, trans, block_group, path);
2876                 if (ret)
2877                         goto out_free;
2878                 goto again;
2879         }
2880
2881         /* We've already setup this transaction, go ahead and exit */
2882         if (block_group->cache_generation == trans->transid &&
2883             i_size_read(inode)) {
2884                 dcs = BTRFS_DC_SETUP;
2885                 goto out_put;
2886         }
2887
2888         /*
2889          * We want to set the generation to 0, that way if anything goes wrong
2890          * from here on out we know not to trust this cache when we load up next
2891          * time.
2892          */
2893         BTRFS_I(inode)->generation = 0;
2894         ret = btrfs_update_inode(trans, root, inode);
2895         WARN_ON(ret);
2896
2897         if (i_size_read(inode) > 0) {
2898                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2899                                                       inode);
2900                 if (ret)
2901                         goto out_put;
2902         }
2903
2904         spin_lock(&block_group->lock);
2905         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2906                 /* We're not cached, don't bother trying to write stuff out */
2907                 dcs = BTRFS_DC_WRITTEN;
2908                 spin_unlock(&block_group->lock);
2909                 goto out_put;
2910         }
2911         spin_unlock(&block_group->lock);
2912
2913         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2914         if (!num_pages)
2915                 num_pages = 1;
2916
2917         /*
2918          * Just to make absolutely sure we have enough space, we're going to
2919          * preallocate 12 pages worth of space for each block group.  In
2920          * practice we ought to use at most 8, but we need extra space so we can
2921          * add our header and have a terminator between the extents and the
2922          * bitmaps.
2923          */
2924         num_pages *= 16;
2925         num_pages *= PAGE_CACHE_SIZE;
2926
2927         ret = btrfs_check_data_free_space(inode, num_pages);
2928         if (ret)
2929                 goto out_put;
2930
2931         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2932                                               num_pages, num_pages,
2933                                               &alloc_hint);
2934         if (!ret)
2935                 dcs = BTRFS_DC_SETUP;
2936         btrfs_free_reserved_data_space(inode, num_pages);
2937
2938 out_put:
2939         iput(inode);
2940 out_free:
2941         btrfs_release_path(path);
2942 out:
2943         spin_lock(&block_group->lock);
2944         if (!ret && dcs == BTRFS_DC_SETUP)
2945                 block_group->cache_generation = trans->transid;
2946         block_group->disk_cache_state = dcs;
2947         spin_unlock(&block_group->lock);
2948
2949         return ret;
2950 }
2951
2952 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2953                                    struct btrfs_root *root)
2954 {
2955         struct btrfs_block_group_cache *cache;
2956         int err = 0;
2957         struct btrfs_path *path;
2958         u64 last = 0;
2959
2960         path = btrfs_alloc_path();
2961         if (!path)
2962                 return -ENOMEM;
2963
2964 again:
2965         while (1) {
2966                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2967                 while (cache) {
2968                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2969                                 break;
2970                         cache = next_block_group(root, cache);
2971                 }
2972                 if (!cache) {
2973                         if (last == 0)
2974                                 break;
2975                         last = 0;
2976                         continue;
2977                 }
2978                 err = cache_save_setup(cache, trans, path);
2979                 last = cache->key.objectid + cache->key.offset;
2980                 btrfs_put_block_group(cache);
2981         }
2982
2983         while (1) {
2984                 if (last == 0) {
2985                         err = btrfs_run_delayed_refs(trans, root,
2986                                                      (unsigned long)-1);
2987                         if (err) /* File system offline */
2988                                 goto out;
2989                 }
2990
2991                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2992                 while (cache) {
2993                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2994                                 btrfs_put_block_group(cache);
2995                                 goto again;
2996                         }
2997
2998                         if (cache->dirty)
2999                                 break;
3000                         cache = next_block_group(root, cache);
3001                 }
3002                 if (!cache) {
3003                         if (last == 0)
3004                                 break;
3005                         last = 0;
3006                         continue;
3007                 }
3008
3009                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3010                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3011                 cache->dirty = 0;
3012                 last = cache->key.objectid + cache->key.offset;
3013
3014                 err = write_one_cache_group(trans, root, path, cache);
3015                 if (err) /* File system offline */
3016                         goto out;
3017
3018                 btrfs_put_block_group(cache);
3019         }
3020
3021         while (1) {
3022                 /*
3023                  * I don't think this is needed since we're just marking our
3024                  * preallocated extent as written, but just in case it can't
3025                  * hurt.
3026                  */
3027                 if (last == 0) {
3028                         err = btrfs_run_delayed_refs(trans, root,
3029                                                      (unsigned long)-1);
3030                         if (err) /* File system offline */
3031                                 goto out;
3032                 }
3033
3034                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3035                 while (cache) {
3036                         /*
3037                          * Really this shouldn't happen, but it could if we
3038                          * couldn't write the entire preallocated extent and
3039                          * splitting the extent resulted in a new block.
3040                          */
3041                         if (cache->dirty) {
3042                                 btrfs_put_block_group(cache);
3043                                 goto again;
3044                         }
3045                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3046                                 break;
3047                         cache = next_block_group(root, cache);
3048                 }
3049                 if (!cache) {
3050                         if (last == 0)
3051                                 break;
3052                         last = 0;
3053                         continue;
3054                 }
3055
3056                 err = btrfs_write_out_cache(root, trans, cache, path);
3057
3058                 /*
3059                  * If we didn't have an error then the cache state is still
3060                  * NEED_WRITE, so we can set it to WRITTEN.
3061                  */
3062                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3063                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3064                 last = cache->key.objectid + cache->key.offset;
3065                 btrfs_put_block_group(cache);
3066         }
3067 out:
3068
3069         btrfs_free_path(path);
3070         return err;
3071 }
3072
3073 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3074 {
3075         struct btrfs_block_group_cache *block_group;
3076         int readonly = 0;
3077
3078         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3079         if (!block_group || block_group->ro)
3080                 readonly = 1;
3081         if (block_group)
3082                 btrfs_put_block_group(block_group);
3083         return readonly;
3084 }
3085
3086 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3087                              u64 total_bytes, u64 bytes_used,
3088                              struct btrfs_space_info **space_info)
3089 {
3090         struct btrfs_space_info *found;
3091         int i;
3092         int factor;
3093
3094         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3095                      BTRFS_BLOCK_GROUP_RAID10))
3096                 factor = 2;
3097         else
3098                 factor = 1;
3099
3100         found = __find_space_info(info, flags);
3101         if (found) {
3102                 spin_lock(&found->lock);
3103                 found->total_bytes += total_bytes;
3104                 found->disk_total += total_bytes * factor;
3105                 found->bytes_used += bytes_used;
3106                 found->disk_used += bytes_used * factor;
3107                 found->full = 0;
3108                 spin_unlock(&found->lock);
3109                 *space_info = found;
3110                 return 0;
3111         }
3112         found = kzalloc(sizeof(*found), GFP_NOFS);
3113         if (!found)
3114                 return -ENOMEM;
3115
3116         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3117                 INIT_LIST_HEAD(&found->block_groups[i]);
3118         init_rwsem(&found->groups_sem);
3119         spin_lock_init(&found->lock);
3120         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3121         found->total_bytes = total_bytes;
3122         found->disk_total = total_bytes * factor;
3123         found->bytes_used = bytes_used;
3124         found->disk_used = bytes_used * factor;
3125         found->bytes_pinned = 0;
3126         found->bytes_reserved = 0;
3127         found->bytes_readonly = 0;
3128         found->bytes_may_use = 0;
3129         found->full = 0;
3130         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3131         found->chunk_alloc = 0;
3132         found->flush = 0;
3133         init_waitqueue_head(&found->wait);
3134         *space_info = found;
3135         list_add_rcu(&found->list, &info->space_info);
3136         return 0;
3137 }
3138
3139 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3140 {
3141         u64 extra_flags = chunk_to_extended(flags) &
3142                                 BTRFS_EXTENDED_PROFILE_MASK;
3143
3144         if (flags & BTRFS_BLOCK_GROUP_DATA)
3145                 fs_info->avail_data_alloc_bits |= extra_flags;
3146         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3147                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3148         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3149                 fs_info->avail_system_alloc_bits |= extra_flags;
3150 }
3151
3152 /*
3153  * returns target flags in extended format or 0 if restripe for this
3154  * chunk_type is not in progress
3155  */
3156 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3157 {
3158         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3159         u64 target = 0;
3160
3161         BUG_ON(!mutex_is_locked(&fs_info->volume_mutex) &&
3162                !spin_is_locked(&fs_info->balance_lock));
3163
3164         if (!bctl)
3165                 return 0;
3166
3167         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3168             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3169                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3170         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3171                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3172                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3173         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3174                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3175                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3176         }
3177
3178         return target;
3179 }
3180
3181 /*
3182  * @flags: available profiles in extended format (see ctree.h)
3183  *
3184  * Returns reduced profile in chunk format.  If profile changing is in
3185  * progress (either running or paused) picks the target profile (if it's
3186  * already available), otherwise falls back to plain reducing.
3187  */
3188 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3189 {
3190         /*
3191          * we add in the count of missing devices because we want
3192          * to make sure that any RAID levels on a degraded FS
3193          * continue to be honored.
3194          */
3195         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3196                 root->fs_info->fs_devices->missing_devices;
3197         u64 target;
3198
3199         /*
3200          * see if restripe for this chunk_type is in progress, if so
3201          * try to reduce to the target profile
3202          */
3203         spin_lock(&root->fs_info->balance_lock);
3204         target = get_restripe_target(root->fs_info, flags);
3205         if (target) {
3206                 /* pick target profile only if it's already available */
3207                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3208                         spin_unlock(&root->fs_info->balance_lock);
3209                         return extended_to_chunk(target);
3210                 }
3211         }
3212         spin_unlock(&root->fs_info->balance_lock);
3213
3214         if (num_devices == 1)
3215                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3216         if (num_devices < 4)
3217                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3218
3219         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3220             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3221                       BTRFS_BLOCK_GROUP_RAID10))) {
3222                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3223         }
3224
3225         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3226             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3227                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3228         }
3229
3230         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3231             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3232              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3233              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3234                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3235         }
3236
3237         return extended_to_chunk(flags);
3238 }
3239
3240 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3241 {
3242         if (flags & BTRFS_BLOCK_GROUP_DATA)
3243                 flags |= root->fs_info->avail_data_alloc_bits;
3244         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3245                 flags |= root->fs_info->avail_system_alloc_bits;
3246         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3247                 flags |= root->fs_info->avail_metadata_alloc_bits;
3248
3249         return btrfs_reduce_alloc_profile(root, flags);
3250 }
3251
3252 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3253 {
3254         u64 flags;
3255
3256         if (data)
3257                 flags = BTRFS_BLOCK_GROUP_DATA;
3258         else if (root == root->fs_info->chunk_root)
3259                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3260         else
3261                 flags = BTRFS_BLOCK_GROUP_METADATA;
3262
3263         return get_alloc_profile(root, flags);
3264 }
3265
3266 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3267 {
3268         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3269                                                        BTRFS_BLOCK_GROUP_DATA);
3270 }
3271
3272 /*
3273  * This will check the space that the inode allocates from to make sure we have
3274  * enough space for bytes.
3275  */
3276 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3277 {
3278         struct btrfs_space_info *data_sinfo;
3279         struct btrfs_root *root = BTRFS_I(inode)->root;
3280         u64 used;
3281         int ret = 0, committed = 0, alloc_chunk = 1;
3282
3283         /* make sure bytes are sectorsize aligned */
3284         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3285
3286         if (root == root->fs_info->tree_root ||
3287             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3288                 alloc_chunk = 0;
3289                 committed = 1;
3290         }
3291
3292         data_sinfo = BTRFS_I(inode)->space_info;
3293         if (!data_sinfo)
3294                 goto alloc;
3295
3296 again:
3297         /* make sure we have enough space to handle the data first */
3298         spin_lock(&data_sinfo->lock);
3299         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3300                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3301                 data_sinfo->bytes_may_use;
3302
3303         if (used + bytes > data_sinfo->total_bytes) {
3304                 struct btrfs_trans_handle *trans;
3305
3306                 /*
3307                  * if we don't have enough free bytes in this space then we need
3308                  * to alloc a new chunk.
3309                  */
3310                 if (!data_sinfo->full && alloc_chunk) {
3311                         u64 alloc_target;
3312
3313                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3314                         spin_unlock(&data_sinfo->lock);
3315 alloc:
3316                         alloc_target = btrfs_get_alloc_profile(root, 1);
3317                         trans = btrfs_join_transaction(root);
3318                         if (IS_ERR(trans))
3319                                 return PTR_ERR(trans);
3320
3321                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3322                                              bytes + 2 * 1024 * 1024,
3323                                              alloc_target,
3324                                              CHUNK_ALLOC_NO_FORCE);
3325                         btrfs_end_transaction(trans, root);
3326                         if (ret < 0) {
3327                                 if (ret != -ENOSPC)
3328                                         return ret;
3329                                 else
3330                                         goto commit_trans;
3331                         }
3332
3333                         if (!data_sinfo) {
3334                                 btrfs_set_inode_space_info(root, inode);
3335                                 data_sinfo = BTRFS_I(inode)->space_info;
3336                         }
3337                         goto again;
3338                 }
3339
3340                 /*
3341                  * If we have less pinned bytes than we want to allocate then
3342                  * don't bother committing the transaction, it won't help us.
3343                  */
3344                 if (data_sinfo->bytes_pinned < bytes)
3345                         committed = 1;
3346                 spin_unlock(&data_sinfo->lock);
3347
3348                 /* commit the current transaction and try again */
3349 commit_trans:
3350                 if (!committed &&
3351                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3352                         committed = 1;
3353                         trans = btrfs_join_transaction(root);
3354                         if (IS_ERR(trans))
3355                                 return PTR_ERR(trans);
3356                         ret = btrfs_commit_transaction(trans, root);
3357                         if (ret)
3358                                 return ret;
3359                         goto again;
3360                 }
3361
3362                 return -ENOSPC;
3363         }
3364         data_sinfo->bytes_may_use += bytes;
3365         trace_btrfs_space_reservation(root->fs_info, "space_info",
3366                                       data_sinfo->flags, bytes, 1);
3367         spin_unlock(&data_sinfo->lock);
3368
3369         return 0;
3370 }
3371
3372 /*
3373  * Called if we need to clear a data reservation for this inode.
3374  */
3375 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3376 {
3377         struct btrfs_root *root = BTRFS_I(inode)->root;
3378         struct btrfs_space_info *data_sinfo;
3379
3380         /* make sure bytes are sectorsize aligned */
3381         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3382
3383         data_sinfo = BTRFS_I(inode)->space_info;
3384         spin_lock(&data_sinfo->lock);
3385         data_sinfo->bytes_may_use -= bytes;
3386         trace_btrfs_space_reservation(root->fs_info, "space_info",
3387                                       data_sinfo->flags, bytes, 0);
3388         spin_unlock(&data_sinfo->lock);
3389 }
3390
3391 static void force_metadata_allocation(struct btrfs_fs_info *info)
3392 {
3393         struct list_head *head = &info->space_info;
3394         struct btrfs_space_info *found;
3395
3396         rcu_read_lock();
3397         list_for_each_entry_rcu(found, head, list) {
3398                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3399                         found->force_alloc = CHUNK_ALLOC_FORCE;
3400         }
3401         rcu_read_unlock();
3402 }
3403
3404 static int should_alloc_chunk(struct btrfs_root *root,
3405                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3406                               int force)
3407 {
3408         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3409         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3410         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3411         u64 thresh;
3412
3413         if (force == CHUNK_ALLOC_FORCE)
3414                 return 1;
3415
3416         /*
3417          * We need to take into account the global rsv because for all intents
3418          * and purposes it's used space.  Don't worry about locking the
3419          * global_rsv, it doesn't change except when the transaction commits.
3420          */
3421         num_allocated += global_rsv->size;
3422
3423         /*
3424          * in limited mode, we want to have some free space up to
3425          * about 1% of the FS size.
3426          */
3427         if (force == CHUNK_ALLOC_LIMITED) {
3428                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3429                 thresh = max_t(u64, 64 * 1024 * 1024,
3430                                div_factor_fine(thresh, 1));
3431
3432                 if (num_bytes - num_allocated < thresh)
3433                         return 1;
3434         }
3435         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3436
3437         /* 256MB or 2% of the FS */
3438         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3439         /* system chunks need a much small threshold */
3440         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3441                 thresh = 32 * 1024 * 1024;
3442
3443         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3444                 return 0;
3445         return 1;
3446 }
3447
3448 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3449 {
3450         u64 num_dev;
3451
3452         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3453             type & BTRFS_BLOCK_GROUP_RAID0)
3454                 num_dev = root->fs_info->fs_devices->rw_devices;
3455         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3456                 num_dev = 2;
3457         else
3458                 num_dev = 1;    /* DUP or single */
3459
3460         /* metadata for updaing devices and chunk tree */
3461         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3462 }
3463
3464 static void check_system_chunk(struct btrfs_trans_handle *trans,
3465                                struct btrfs_root *root, u64 type)
3466 {
3467         struct btrfs_space_info *info;
3468         u64 left;
3469         u64 thresh;
3470
3471         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3472         spin_lock(&info->lock);
3473         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3474                 info->bytes_reserved - info->bytes_readonly;
3475         spin_unlock(&info->lock);
3476
3477         thresh = get_system_chunk_thresh(root, type);
3478         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3479                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3480                        left, thresh, type);
3481                 dump_space_info(info, 0, 0);
3482         }
3483
3484         if (left < thresh) {
3485                 u64 flags;
3486
3487                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3488                 btrfs_alloc_chunk(trans, root, flags);
3489         }
3490 }
3491
3492 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3493                           struct btrfs_root *extent_root, u64 alloc_bytes,
3494                           u64 flags, int force)
3495 {
3496         struct btrfs_space_info *space_info;
3497         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3498         int wait_for_alloc = 0;
3499         int ret = 0;
3500
3501         space_info = __find_space_info(extent_root->fs_info, flags);
3502         if (!space_info) {
3503                 ret = update_space_info(extent_root->fs_info, flags,
3504                                         0, 0, &space_info);
3505                 BUG_ON(ret); /* -ENOMEM */
3506         }
3507         BUG_ON(!space_info); /* Logic error */
3508
3509 again:
3510         spin_lock(&space_info->lock);
3511         if (force < space_info->force_alloc)
3512                 force = space_info->force_alloc;
3513         if (space_info->full) {
3514                 spin_unlock(&space_info->lock);
3515                 return 0;
3516         }
3517
3518         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3519                 spin_unlock(&space_info->lock);
3520                 return 0;
3521         } else if (space_info->chunk_alloc) {
3522                 wait_for_alloc = 1;
3523         } else {
3524                 space_info->chunk_alloc = 1;
3525         }
3526
3527         spin_unlock(&space_info->lock);
3528
3529         mutex_lock(&fs_info->chunk_mutex);
3530
3531         /*
3532          * The chunk_mutex is held throughout the entirety of a chunk
3533          * allocation, so once we've acquired the chunk_mutex we know that the
3534          * other guy is done and we need to recheck and see if we should
3535          * allocate.
3536          */
3537         if (wait_for_alloc) {
3538                 mutex_unlock(&fs_info->chunk_mutex);
3539                 wait_for_alloc = 0;
3540                 goto again;
3541         }
3542
3543         /*
3544          * If we have mixed data/metadata chunks we want to make sure we keep
3545          * allocating mixed chunks instead of individual chunks.
3546          */
3547         if (btrfs_mixed_space_info(space_info))
3548                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3549
3550         /*
3551          * if we're doing a data chunk, go ahead and make sure that
3552          * we keep a reasonable number of metadata chunks allocated in the
3553          * FS as well.
3554          */
3555         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3556                 fs_info->data_chunk_allocations++;
3557                 if (!(fs_info->data_chunk_allocations %
3558                       fs_info->metadata_ratio))
3559                         force_metadata_allocation(fs_info);
3560         }
3561
3562         /*
3563          * Check if we have enough space in SYSTEM chunk because we may need
3564          * to update devices.
3565          */
3566         check_system_chunk(trans, extent_root, flags);
3567
3568         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3569         if (ret < 0 && ret != -ENOSPC)
3570                 goto out;
3571
3572         spin_lock(&space_info->lock);
3573         if (ret)
3574                 space_info->full = 1;
3575         else
3576                 ret = 1;
3577
3578         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3579         space_info->chunk_alloc = 0;
3580         spin_unlock(&space_info->lock);
3581 out:
3582         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3583         return ret;
3584 }
3585
3586 /*
3587  * shrink metadata reservation for delalloc
3588  */
3589 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3590                            bool wait_ordered)
3591 {
3592         struct btrfs_block_rsv *block_rsv;
3593         struct btrfs_space_info *space_info;
3594         struct btrfs_trans_handle *trans;
3595         u64 reserved;
3596         u64 max_reclaim;
3597         u64 reclaimed = 0;
3598         long time_left;
3599         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3600         int loops = 0;
3601         unsigned long progress;
3602
3603         trans = (struct btrfs_trans_handle *)current->journal_info;
3604         block_rsv = &root->fs_info->delalloc_block_rsv;
3605         space_info = block_rsv->space_info;
3606
3607         smp_mb();
3608         reserved = space_info->bytes_may_use;
3609         progress = space_info->reservation_progress;
3610
3611         if (reserved == 0)
3612                 return 0;
3613
3614         smp_mb();
3615         if (root->fs_info->delalloc_bytes == 0) {
3616                 if (trans)
3617                         return 0;
3618                 btrfs_wait_ordered_extents(root, 0, 0);
3619                 return 0;
3620         }
3621
3622         max_reclaim = min(reserved, to_reclaim);
3623         nr_pages = max_t(unsigned long, nr_pages,
3624                          max_reclaim >> PAGE_CACHE_SHIFT);
3625         while (loops < 1024) {
3626                 /* have the flusher threads jump in and do some IO */
3627                 smp_mb();
3628                 nr_pages = min_t(unsigned long, nr_pages,
3629                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3630                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3631                                                 WB_REASON_FS_FREE_SPACE);
3632
3633                 spin_lock(&space_info->lock);
3634                 if (reserved > space_info->bytes_may_use)
3635                         reclaimed += reserved - space_info->bytes_may_use;
3636                 reserved = space_info->bytes_may_use;
3637                 spin_unlock(&space_info->lock);
3638
3639                 loops++;
3640
3641                 if (reserved == 0 || reclaimed >= max_reclaim)
3642                         break;
3643
3644                 if (trans && trans->transaction->blocked)
3645                         return -EAGAIN;
3646
3647                 if (wait_ordered && !trans) {
3648                         btrfs_wait_ordered_extents(root, 0, 0);
3649                 } else {
3650                         time_left = schedule_timeout_interruptible(1);
3651
3652                         /* We were interrupted, exit */
3653                         if (time_left)
3654                                 break;
3655                 }
3656
3657                 /* we've kicked the IO a few times, if anything has been freed,
3658                  * exit.  There is no sense in looping here for a long time
3659                  * when we really need to commit the transaction, or there are
3660                  * just too many writers without enough free space
3661                  */
3662
3663                 if (loops > 3) {
3664                         smp_mb();
3665                         if (progress != space_info->reservation_progress)
3666                                 break;
3667                 }
3668
3669         }
3670
3671         return reclaimed >= to_reclaim;
3672 }
3673
3674 /**
3675  * maybe_commit_transaction - possibly commit the transaction if its ok to
3676  * @root - the root we're allocating for
3677  * @bytes - the number of bytes we want to reserve
3678  * @force - force the commit
3679  *
3680  * This will check to make sure that committing the transaction will actually
3681  * get us somewhere and then commit the transaction if it does.  Otherwise it
3682  * will return -ENOSPC.
3683  */
3684 static int may_commit_transaction(struct btrfs_root *root,
3685                                   struct btrfs_space_info *space_info,
3686                                   u64 bytes, int force)
3687 {
3688         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3689         struct btrfs_trans_handle *trans;
3690
3691         trans = (struct btrfs_trans_handle *)current->journal_info;
3692         if (trans)
3693                 return -EAGAIN;
3694
3695         if (force)
3696                 goto commit;
3697
3698         /* See if there is enough pinned space to make this reservation */
3699         spin_lock(&space_info->lock);
3700         if (space_info->bytes_pinned >= bytes) {
3701                 spin_unlock(&space_info->lock);
3702                 goto commit;
3703         }
3704         spin_unlock(&space_info->lock);
3705
3706         /*
3707          * See if there is some space in the delayed insertion reservation for
3708          * this reservation.
3709          */
3710         if (space_info != delayed_rsv->space_info)
3711                 return -ENOSPC;
3712
3713         spin_lock(&space_info->lock);
3714         spin_lock(&delayed_rsv->lock);
3715         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3716                 spin_unlock(&delayed_rsv->lock);
3717                 spin_unlock(&space_info->lock);
3718                 return -ENOSPC;
3719         }
3720         spin_unlock(&delayed_rsv->lock);
3721         spin_unlock(&space_info->lock);
3722
3723 commit:
3724         trans = btrfs_join_transaction(root);
3725         if (IS_ERR(trans))
3726                 return -ENOSPC;
3727
3728         return btrfs_commit_transaction(trans, root);
3729 }
3730
3731 /**
3732  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3733  * @root - the root we're allocating for
3734  * @block_rsv - the block_rsv we're allocating for
3735  * @orig_bytes - the number of bytes we want
3736  * @flush - wether or not we can flush to make our reservation
3737  *
3738  * This will reserve orgi_bytes number of bytes from the space info associated
3739  * with the block_rsv.  If there is not enough space it will make an attempt to
3740  * flush out space to make room.  It will do this by flushing delalloc if
3741  * possible or committing the transaction.  If flush is 0 then no attempts to
3742  * regain reservations will be made and this will fail if there is not enough
3743  * space already.
3744  */
3745 static int reserve_metadata_bytes(struct btrfs_root *root,
3746                                   struct btrfs_block_rsv *block_rsv,
3747                                   u64 orig_bytes, int flush)
3748 {
3749         struct btrfs_space_info *space_info = block_rsv->space_info;
3750         u64 used;
3751         u64 num_bytes = orig_bytes;
3752         int retries = 0;
3753         int ret = 0;
3754         bool committed = false;
3755         bool flushing = false;
3756         bool wait_ordered = false;
3757
3758 again:
3759         ret = 0;
3760         spin_lock(&space_info->lock);
3761         /*
3762          * We only want to wait if somebody other than us is flushing and we are
3763          * actually alloed to flush.
3764          */
3765         while (flush && !flushing && space_info->flush) {
3766                 spin_unlock(&space_info->lock);
3767                 /*
3768                  * If we have a trans handle we can't wait because the flusher
3769                  * may have to commit the transaction, which would mean we would
3770                  * deadlock since we are waiting for the flusher to finish, but
3771                  * hold the current transaction open.
3772                  */
3773                 if (current->journal_info)
3774                         return -EAGAIN;
3775                 ret = wait_event_interruptible(space_info->wait,
3776                                                !space_info->flush);
3777                 /* Must have been interrupted, return */
3778                 if (ret) {
3779                         printk(KERN_DEBUG "btrfs: %s returning -EINTR\n", __func__);
3780                         return -EINTR;
3781                 }
3782
3783                 spin_lock(&space_info->lock);
3784         }
3785
3786         ret = -ENOSPC;
3787         used = space_info->bytes_used + space_info->bytes_reserved +
3788                 space_info->bytes_pinned + space_info->bytes_readonly +
3789                 space_info->bytes_may_use;
3790
3791         /*
3792          * The idea here is that we've not already over-reserved the block group
3793          * then we can go ahead and save our reservation first and then start
3794          * flushing if we need to.  Otherwise if we've already overcommitted
3795          * lets start flushing stuff first and then come back and try to make
3796          * our reservation.
3797          */
3798         if (used <= space_info->total_bytes) {
3799                 if (used + orig_bytes <= space_info->total_bytes) {
3800                         space_info->bytes_may_use += orig_bytes;
3801                         trace_btrfs_space_reservation(root->fs_info,
3802                                 "space_info", space_info->flags, orig_bytes, 1);
3803                         ret = 0;
3804                 } else {
3805                         /*
3806                          * Ok set num_bytes to orig_bytes since we aren't
3807                          * overocmmitted, this way we only try and reclaim what
3808                          * we need.
3809                          */
3810                         num_bytes = orig_bytes;
3811                 }
3812         } else {
3813                 /*
3814                  * Ok we're over committed, set num_bytes to the overcommitted
3815                  * amount plus the amount of bytes that we need for this
3816                  * reservation.
3817                  */
3818                 wait_ordered = true;
3819                 num_bytes = used - space_info->total_bytes +
3820                         (orig_bytes * (retries + 1));
3821         }
3822
3823         if (ret) {
3824                 u64 profile = btrfs_get_alloc_profile(root, 0);
3825                 u64 avail;
3826
3827                 /*
3828                  * If we have a lot of space that's pinned, don't bother doing
3829                  * the overcommit dance yet and just commit the transaction.
3830                  */
3831                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3832                 do_div(avail, 10);
3833                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3834                         space_info->flush = 1;
3835                         flushing = true;
3836                         spin_unlock(&space_info->lock);
3837                         ret = may_commit_transaction(root, space_info,
3838                                                      orig_bytes, 1);
3839                         if (ret)
3840                                 goto out;
3841                         committed = true;
3842                         goto again;
3843                 }
3844
3845                 spin_lock(&root->fs_info->free_chunk_lock);
3846                 avail = root->fs_info->free_chunk_space;
3847
3848                 /*
3849                  * If we have dup, raid1 or raid10 then only half of the free
3850                  * space is actually useable.
3851                  */
3852                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3853                                BTRFS_BLOCK_GROUP_RAID1 |
3854                                BTRFS_BLOCK_GROUP_RAID10))
3855                         avail >>= 1;
3856
3857                 /*
3858                  * If we aren't flushing don't let us overcommit too much, say
3859                  * 1/8th of the space.  If we can flush, let it overcommit up to
3860                  * 1/2 of the space.
3861                  */
3862                 if (flush)
3863                         avail >>= 3;
3864                 else
3865                         avail >>= 1;
3866                  spin_unlock(&root->fs_info->free_chunk_lock);
3867
3868                 if (used + num_bytes < space_info->total_bytes + avail) {
3869                         space_info->bytes_may_use += orig_bytes;
3870                         trace_btrfs_space_reservation(root->fs_info,
3871                                 "space_info", space_info->flags, orig_bytes, 1);
3872                         ret = 0;
3873                 } else {
3874                         wait_ordered = true;
3875                 }
3876         }
3877
3878         /*
3879          * Couldn't make our reservation, save our place so while we're trying
3880          * to reclaim space we can actually use it instead of somebody else
3881          * stealing it from us.
3882          */
3883         if (ret && flush) {
3884                 flushing = true;
3885                 space_info->flush = 1;
3886         }
3887
3888         spin_unlock(&space_info->lock);
3889
3890         if (!ret || !flush)
3891                 goto out;
3892
3893         /*
3894          * We do synchronous shrinking since we don't actually unreserve
3895          * metadata until after the IO is completed.
3896          */
3897         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3898         if (ret < 0)
3899                 goto out;
3900
3901         ret = 0;
3902
3903         /*
3904          * So if we were overcommitted it's possible that somebody else flushed
3905          * out enough space and we simply didn't have enough space to reclaim,
3906          * so go back around and try again.
3907          */
3908         if (retries < 2) {
3909                 wait_ordered = true;
3910                 retries++;
3911                 goto again;
3912         }
3913
3914         ret = -ENOSPC;
3915         if (committed)
3916                 goto out;
3917
3918         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3919         if (!ret) {
3920                 committed = true;
3921                 goto again;
3922         }
3923
3924 out:
3925         if (flushing) {
3926                 spin_lock(&space_info->lock);
3927                 space_info->flush = 0;
3928                 wake_up_all(&space_info->wait);
3929                 spin_unlock(&space_info->lock);
3930         }
3931         return ret;
3932 }
3933
3934 static struct btrfs_block_rsv *get_block_rsv(
3935                                         const struct btrfs_trans_handle *trans,
3936                                         const struct btrfs_root *root)
3937 {
3938         struct btrfs_block_rsv *block_rsv = NULL;
3939
3940         if (root->ref_cows || root == root->fs_info->csum_root)
3941                 block_rsv = trans->block_rsv;
3942
3943         if (!block_rsv)
3944                 block_rsv = root->block_rsv;
3945
3946         if (!block_rsv)
3947                 block_rsv = &root->fs_info->empty_block_rsv;
3948
3949         return block_rsv;
3950 }
3951
3952 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3953                                u64 num_bytes)
3954 {
3955         int ret = -ENOSPC;
3956         spin_lock(&block_rsv->lock);
3957         if (block_rsv->reserved >= num_bytes) {
3958                 block_rsv->reserved -= num_bytes;
3959                 if (block_rsv->reserved < block_rsv->size)
3960                         block_rsv->full = 0;
3961                 ret = 0;
3962         }
3963         spin_unlock(&block_rsv->lock);
3964         return ret;
3965 }
3966
3967 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3968                                 u64 num_bytes, int update_size)
3969 {
3970         spin_lock(&block_rsv->lock);
3971         block_rsv->reserved += num_bytes;
3972         if (update_size)
3973                 block_rsv->size += num_bytes;
3974         else if (block_rsv->reserved >= block_rsv->size)
3975                 block_rsv->full = 1;
3976         spin_unlock(&block_rsv->lock);
3977 }
3978
3979 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3980                                     struct btrfs_block_rsv *block_rsv,
3981                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3982 {
3983         struct btrfs_space_info *space_info = block_rsv->space_info;
3984
3985         spin_lock(&block_rsv->lock);
3986         if (num_bytes == (u64)-1)
3987                 num_bytes = block_rsv->size;
3988         block_rsv->size -= num_bytes;
3989         if (block_rsv->reserved >= block_rsv->size) {
3990                 num_bytes = block_rsv->reserved - block_rsv->size;
3991                 block_rsv->reserved = block_rsv->size;
3992                 block_rsv->full = 1;
3993         } else {
3994                 num_bytes = 0;
3995         }
3996         spin_unlock(&block_rsv->lock);
3997
3998         if (num_bytes > 0) {
3999                 if (dest) {
4000                         spin_lock(&dest->lock);
4001                         if (!dest->full) {
4002                                 u64 bytes_to_add;
4003
4004                                 bytes_to_add = dest->size - dest->reserved;
4005                                 bytes_to_add = min(num_bytes, bytes_to_add);
4006                                 dest->reserved += bytes_to_add;
4007                                 if (dest->reserved >= dest->size)
4008                                         dest->full = 1;
4009                                 num_bytes -= bytes_to_add;
4010                         }
4011                         spin_unlock(&dest->lock);
4012                 }
4013                 if (num_bytes) {
4014                         spin_lock(&space_info->lock);
4015                         space_info->bytes_may_use -= num_bytes;
4016                         trace_btrfs_space_reservation(fs_info, "space_info",
4017                                         space_info->flags, num_bytes, 0);
4018                         space_info->reservation_progress++;
4019                         spin_unlock(&space_info->lock);
4020                 }
4021         }
4022 }
4023
4024 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4025                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4026 {
4027         int ret;
4028
4029         ret = block_rsv_use_bytes(src, num_bytes);
4030         if (ret)
4031                 return ret;
4032
4033         block_rsv_add_bytes(dst, num_bytes, 1);
4034         return 0;
4035 }
4036
4037 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4038 {
4039         memset(rsv, 0, sizeof(*rsv));
4040         spin_lock_init(&rsv->lock);
4041 }
4042
4043 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4044 {
4045         struct btrfs_block_rsv *block_rsv;
4046         struct btrfs_fs_info *fs_info = root->fs_info;
4047
4048         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4049         if (!block_rsv)
4050                 return NULL;
4051
4052         btrfs_init_block_rsv(block_rsv);
4053         block_rsv->space_info = __find_space_info(fs_info,
4054                                                   BTRFS_BLOCK_GROUP_METADATA);
4055         return block_rsv;
4056 }
4057
4058 void btrfs_free_block_rsv(struct btrfs_root *root,
4059                           struct btrfs_block_rsv *rsv)
4060 {
4061         btrfs_block_rsv_release(root, rsv, (u64)-1);
4062         kfree(rsv);
4063 }
4064
4065 static inline int __block_rsv_add(struct btrfs_root *root,
4066                                   struct btrfs_block_rsv *block_rsv,
4067                                   u64 num_bytes, int flush)
4068 {
4069         int ret;
4070
4071         if (num_bytes == 0)
4072                 return 0;
4073
4074         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4075         if (!ret) {
4076                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4077                 return 0;
4078         }
4079
4080         return ret;
4081 }
4082
4083 int btrfs_block_rsv_add(struct btrfs_root *root,
4084                         struct btrfs_block_rsv *block_rsv,
4085                         u64 num_bytes)
4086 {
4087         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4088 }
4089
4090 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4091                                 struct btrfs_block_rsv *block_rsv,
4092                                 u64 num_bytes)
4093 {
4094         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4095 }
4096
4097 int btrfs_block_rsv_check(struct btrfs_root *root,
4098                           struct btrfs_block_rsv *block_rsv, int min_factor)
4099 {
4100         u64 num_bytes = 0;
4101         int ret = -ENOSPC;
4102
4103         if (!block_rsv)
4104                 return 0;
4105
4106         spin_lock(&block_rsv->lock);
4107         num_bytes = div_factor(block_rsv->size, min_factor);
4108         if (block_rsv->reserved >= num_bytes)
4109                 ret = 0;
4110         spin_unlock(&block_rsv->lock);
4111
4112         return ret;
4113 }
4114
4115 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4116                                            struct btrfs_block_rsv *block_rsv,
4117                                            u64 min_reserved, int flush)
4118 {
4119         u64 num_bytes = 0;
4120         int ret = -ENOSPC;
4121
4122         if (!block_rsv)
4123                 return 0;
4124
4125         spin_lock(&block_rsv->lock);
4126         num_bytes = min_reserved;
4127         if (block_rsv->reserved >= num_bytes)
4128                 ret = 0;
4129         else
4130                 num_bytes -= block_rsv->reserved;
4131         spin_unlock(&block_rsv->lock);
4132
4133         if (!ret)
4134                 return 0;
4135
4136         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4137         if (!ret) {
4138                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4139                 return 0;
4140         }
4141
4142         return ret;
4143 }
4144
4145 int btrfs_block_rsv_refill(struct btrfs_root *root,
4146                            struct btrfs_block_rsv *block_rsv,
4147                            u64 min_reserved)
4148 {
4149         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4150 }
4151
4152 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4153                                    struct btrfs_block_rsv *block_rsv,
4154                                    u64 min_reserved)
4155 {
4156         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4157 }
4158
4159 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4160                             struct btrfs_block_rsv *dst_rsv,
4161                             u64 num_bytes)
4162 {
4163         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4164 }
4165
4166 void btrfs_block_rsv_release(struct btrfs_root *root,
4167                              struct btrfs_block_rsv *block_rsv,
4168                              u64 num_bytes)
4169 {
4170         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4171         if (global_rsv->full || global_rsv == block_rsv ||
4172             block_rsv->space_info != global_rsv->space_info)
4173                 global_rsv = NULL;
4174         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4175                                 num_bytes);
4176 }
4177
4178 /*
4179  * helper to calculate size of global block reservation.
4180  * the desired value is sum of space used by extent tree,
4181  * checksum tree and root tree
4182  */
4183 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4184 {
4185         struct btrfs_space_info *sinfo;
4186         u64 num_bytes;
4187         u64 meta_used;
4188         u64 data_used;
4189         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4190
4191         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4192         spin_lock(&sinfo->lock);
4193         data_used = sinfo->bytes_used;
4194         spin_unlock(&sinfo->lock);
4195
4196         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4197         spin_lock(&sinfo->lock);
4198         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4199                 data_used = 0;
4200         meta_used = sinfo->bytes_used;
4201         spin_unlock(&sinfo->lock);
4202
4203         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4204                     csum_size * 2;
4205         num_bytes += div64_u64(data_used + meta_used, 50);
4206
4207         if (num_bytes * 3 > meta_used)
4208                 num_bytes = div64_u64(meta_used, 3) * 2;
4209
4210         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4211 }
4212
4213 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4214 {
4215         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4216         struct btrfs_space_info *sinfo = block_rsv->space_info;
4217         u64 num_bytes;
4218
4219         num_bytes = calc_global_metadata_size(fs_info);
4220
4221         spin_lock(&block_rsv->lock);
4222         spin_lock(&sinfo->lock);
4223
4224         block_rsv->size = num_bytes;
4225
4226         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4227                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4228                     sinfo->bytes_may_use;
4229
4230         if (sinfo->total_bytes > num_bytes) {
4231                 num_bytes = sinfo->total_bytes - num_bytes;
4232                 block_rsv->reserved += num_bytes;
4233                 sinfo->bytes_may_use += num_bytes;
4234                 trace_btrfs_space_reservation(fs_info, "space_info",
4235                                       sinfo->flags, num_bytes, 1);
4236         }
4237
4238         if (block_rsv->reserved >= block_rsv->size) {
4239                 num_bytes = block_rsv->reserved - block_rsv->size;
4240                 sinfo->bytes_may_use -= num_bytes;
4241                 trace_btrfs_space_reservation(fs_info, "space_info",
4242                                       sinfo->flags, num_bytes, 0);
4243                 sinfo->reservation_progress++;
4244                 block_rsv->reserved = block_rsv->size;
4245                 block_rsv->full = 1;
4246         }
4247
4248         spin_unlock(&sinfo->lock);
4249         spin_unlock(&block_rsv->lock);
4250 }
4251
4252 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4253 {
4254         struct btrfs_space_info *space_info;
4255
4256         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4257         fs_info->chunk_block_rsv.space_info = space_info;
4258
4259         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4260         fs_info->global_block_rsv.space_info = space_info;
4261         fs_info->delalloc_block_rsv.space_info = space_info;
4262         fs_info->trans_block_rsv.space_info = space_info;
4263         fs_info->empty_block_rsv.space_info = space_info;
4264         fs_info->delayed_block_rsv.space_info = space_info;
4265
4266         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4267         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4268         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4269         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4270         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4271
4272         update_global_block_rsv(fs_info);
4273 }
4274
4275 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4276 {
4277         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4278                                 (u64)-1);
4279         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4280         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4281         WARN_ON(fs_info->trans_block_rsv.size > 0);
4282         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4283         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4284         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4285         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4286         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4287 }
4288
4289 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4290                                   struct btrfs_root *root)
4291 {
4292         if (!trans->bytes_reserved)
4293                 return;
4294
4295         trace_btrfs_space_reservation(root->fs_info, "transaction",
4296                                       trans->transid, trans->bytes_reserved, 0);
4297         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4298         trans->bytes_reserved = 0;
4299 }
4300
4301 /* Can only return 0 or -ENOSPC */
4302 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4303                                   struct inode *inode)
4304 {
4305         struct btrfs_root *root = BTRFS_I(inode)->root;
4306         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4307         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4308
4309         /*
4310          * We need to hold space in order to delete our orphan item once we've
4311          * added it, so this takes the reservation so we can release it later
4312          * when we are truly done with the orphan item.
4313          */
4314         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4315         trace_btrfs_space_reservation(root->fs_info, "orphan",
4316                                       btrfs_ino(inode), num_bytes, 1);
4317         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4318 }
4319
4320 void btrfs_orphan_release_metadata(struct inode *inode)
4321 {
4322         struct btrfs_root *root = BTRFS_I(inode)->root;
4323         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4324         trace_btrfs_space_reservation(root->fs_info, "orphan",
4325                                       btrfs_ino(inode), num_bytes, 0);
4326         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4327 }
4328
4329 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4330                                 struct btrfs_pending_snapshot *pending)
4331 {
4332         struct btrfs_root *root = pending->root;
4333         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4334         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4335         /*
4336          * two for root back/forward refs, two for directory entries
4337          * and one for root of the snapshot.
4338          */
4339         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4340         dst_rsv->space_info = src_rsv->space_info;
4341         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4342 }
4343
4344 /**
4345  * drop_outstanding_extent - drop an outstanding extent
4346  * @inode: the inode we're dropping the extent for
4347  *
4348  * This is called when we are freeing up an outstanding extent, either called
4349  * after an error or after an extent is written.  This will return the number of
4350  * reserved extents that need to be freed.  This must be called with
4351  * BTRFS_I(inode)->lock held.
4352  */
4353 static unsigned drop_outstanding_extent(struct inode *inode)
4354 {
4355         unsigned drop_inode_space = 0;
4356         unsigned dropped_extents = 0;
4357
4358         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4359         BTRFS_I(inode)->outstanding_extents--;
4360
4361         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4362             BTRFS_I(inode)->delalloc_meta_reserved) {
4363                 drop_inode_space = 1;
4364                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4365         }
4366
4367         /*
4368          * If we have more or the same amount of outsanding extents than we have
4369          * reserved then we need to leave the reserved extents count alone.
4370          */
4371         if (BTRFS_I(inode)->outstanding_extents >=
4372             BTRFS_I(inode)->reserved_extents)
4373                 return drop_inode_space;
4374
4375         dropped_extents = BTRFS_I(inode)->reserved_extents -
4376                 BTRFS_I(inode)->outstanding_extents;
4377         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4378         return dropped_extents + drop_inode_space;
4379 }
4380
4381 /**
4382  * calc_csum_metadata_size - return the amount of metada space that must be
4383  *      reserved/free'd for the given bytes.
4384  * @inode: the inode we're manipulating
4385  * @num_bytes: the number of bytes in question
4386  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4387  *
4388  * This adjusts the number of csum_bytes in the inode and then returns the
4389  * correct amount of metadata that must either be reserved or freed.  We
4390  * calculate how many checksums we can fit into one leaf and then divide the
4391  * number of bytes that will need to be checksumed by this value to figure out
4392  * how many checksums will be required.  If we are adding bytes then the number
4393  * may go up and we will return the number of additional bytes that must be
4394  * reserved.  If it is going down we will return the number of bytes that must
4395  * be freed.
4396  *
4397  * This must be called with BTRFS_I(inode)->lock held.
4398  */
4399 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4400                                    int reserve)
4401 {
4402         struct btrfs_root *root = BTRFS_I(inode)->root;
4403         u64 csum_size;
4404         int num_csums_per_leaf;
4405         int num_csums;
4406         int old_csums;
4407
4408         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4409             BTRFS_I(inode)->csum_bytes == 0)
4410                 return 0;
4411
4412         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4413         if (reserve)
4414                 BTRFS_I(inode)->csum_bytes += num_bytes;
4415         else
4416                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4417         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4418         num_csums_per_leaf = (int)div64_u64(csum_size,
4419                                             sizeof(struct btrfs_csum_item) +
4420                                             sizeof(struct btrfs_disk_key));
4421         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4422         num_csums = num_csums + num_csums_per_leaf - 1;
4423         num_csums = num_csums / num_csums_per_leaf;
4424
4425         old_csums = old_csums + num_csums_per_leaf - 1;
4426         old_csums = old_csums / num_csums_per_leaf;
4427
4428         /* No change, no need to reserve more */
4429         if (old_csums == num_csums)
4430                 return 0;
4431
4432         if (reserve)
4433                 return btrfs_calc_trans_metadata_size(root,
4434                                                       num_csums - old_csums);
4435
4436         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4437 }
4438
4439 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4440 {
4441         struct btrfs_root *root = BTRFS_I(inode)->root;
4442         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4443         u64 to_reserve = 0;
4444         u64 csum_bytes;
4445         unsigned nr_extents = 0;
4446         int extra_reserve = 0;
4447         int flush = 1;
4448         int ret;
4449
4450         /* Need to be holding the i_mutex here if we aren't free space cache */
4451         if (btrfs_is_free_space_inode(root, inode))
4452                 flush = 0;
4453
4454         if (flush && btrfs_transaction_in_commit(root->fs_info))
4455                 schedule_timeout(1);
4456
4457         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4458         num_bytes = ALIGN(num_bytes, root->sectorsize);
4459
4460         spin_lock(&BTRFS_I(inode)->lock);
4461         BTRFS_I(inode)->outstanding_extents++;
4462
4463         if (BTRFS_I(inode)->outstanding_extents >
4464             BTRFS_I(inode)->reserved_extents)
4465                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4466                         BTRFS_I(inode)->reserved_extents;
4467
4468         /*
4469          * Add an item to reserve for updating the inode when we complete the
4470          * delalloc io.
4471          */
4472         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4473                 nr_extents++;
4474                 extra_reserve = 1;
4475         }
4476
4477         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4478         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4479         csum_bytes = BTRFS_I(inode)->csum_bytes;
4480         spin_unlock(&BTRFS_I(inode)->lock);
4481
4482         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4483         if (ret) {
4484                 u64 to_free = 0;
4485                 unsigned dropped;
4486
4487                 spin_lock(&BTRFS_I(inode)->lock);
4488                 dropped = drop_outstanding_extent(inode);
4489                 /*
4490                  * If the inodes csum_bytes is the same as the original
4491                  * csum_bytes then we know we haven't raced with any free()ers
4492                  * so we can just reduce our inodes csum bytes and carry on.
4493                  * Otherwise we have to do the normal free thing to account for
4494                  * the case that the free side didn't free up its reserve
4495                  * because of this outstanding reservation.
4496                  */
4497                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4498                         calc_csum_metadata_size(inode, num_bytes, 0);
4499                 else
4500                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4501                 spin_unlock(&BTRFS_I(inode)->lock);
4502                 if (dropped)
4503                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4504
4505                 if (to_free) {
4506                         btrfs_block_rsv_release(root, block_rsv, to_free);
4507                         trace_btrfs_space_reservation(root->fs_info,
4508                                                       "delalloc",
4509                                                       btrfs_ino(inode),
4510                                                       to_free, 0);
4511                 }
4512                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4513                 return ret;
4514         }
4515
4516         spin_lock(&BTRFS_I(inode)->lock);
4517         if (extra_reserve) {
4518                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4519                 nr_extents--;
4520         }
4521         BTRFS_I(inode)->reserved_extents += nr_extents;
4522         spin_unlock(&BTRFS_I(inode)->lock);
4523         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4524
4525         if (to_reserve)
4526                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4527                                               btrfs_ino(inode), to_reserve, 1);
4528         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4529
4530         return 0;
4531 }
4532
4533 /**
4534  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4535  * @inode: the inode to release the reservation for
4536  * @num_bytes: the number of bytes we're releasing
4537  *
4538  * This will release the metadata reservation for an inode.  This can be called
4539  * once we complete IO for a given set of bytes to release their metadata
4540  * reservations.
4541  */
4542 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4543 {
4544         struct btrfs_root *root = BTRFS_I(inode)->root;
4545         u64 to_free = 0;
4546         unsigned dropped;
4547
4548         num_bytes = ALIGN(num_bytes, root->sectorsize);
4549         spin_lock(&BTRFS_I(inode)->lock);
4550         dropped = drop_outstanding_extent(inode);
4551
4552         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4553         spin_unlock(&BTRFS_I(inode)->lock);
4554         if (dropped > 0)
4555                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4556
4557         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4558                                       btrfs_ino(inode), to_free, 0);
4559         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4560                                 to_free);
4561 }
4562
4563 /**
4564  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4565  * @inode: inode we're writing to
4566  * @num_bytes: the number of bytes we want to allocate
4567  *
4568  * This will do the following things
4569  *
4570  * o reserve space in the data space info for num_bytes
4571  * o reserve space in the metadata space info based on number of outstanding
4572  *   extents and how much csums will be needed
4573  * o add to the inodes ->delalloc_bytes
4574  * o add it to the fs_info's delalloc inodes list.
4575  *
4576  * This will return 0 for success and -ENOSPC if there is no space left.
4577  */
4578 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4579 {
4580         int ret;
4581
4582         ret = btrfs_check_data_free_space(inode, num_bytes);
4583         if (ret)
4584                 return ret;
4585
4586         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4587         if (ret) {
4588                 btrfs_free_reserved_data_space(inode, num_bytes);
4589                 return ret;
4590         }
4591
4592         return 0;
4593 }
4594
4595 /**
4596  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4597  * @inode: inode we're releasing space for
4598  * @num_bytes: the number of bytes we want to free up
4599  *
4600  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4601  * called in the case that we don't need the metadata AND data reservations
4602  * anymore.  So if there is an error or we insert an inline extent.
4603  *
4604  * This function will release the metadata space that was not used and will
4605  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4606  * list if there are no delalloc bytes left.
4607  */
4608 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4609 {
4610         btrfs_delalloc_release_metadata(inode, num_bytes);
4611         btrfs_free_reserved_data_space(inode, num_bytes);
4612 }
4613
4614 static int update_block_group(struct btrfs_trans_handle *trans,
4615                               struct btrfs_root *root,
4616                               u64 bytenr, u64 num_bytes, int alloc)
4617 {
4618         struct btrfs_block_group_cache *cache = NULL;
4619         struct btrfs_fs_info *info = root->fs_info;
4620         u64 total = num_bytes;
4621         u64 old_val;
4622         u64 byte_in_group;
4623         int factor;
4624
4625         /* block accounting for super block */
4626         spin_lock(&info->delalloc_lock);
4627         old_val = btrfs_super_bytes_used(info->super_copy);
4628         if (alloc)
4629                 old_val += num_bytes;
4630         else
4631                 old_val -= num_bytes;
4632         btrfs_set_super_bytes_used(info->super_copy, old_val);
4633         spin_unlock(&info->delalloc_lock);
4634
4635         while (total) {
4636                 cache = btrfs_lookup_block_group(info, bytenr);
4637                 if (!cache)
4638                         return -ENOENT;
4639                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4640                                     BTRFS_BLOCK_GROUP_RAID1 |
4641                                     BTRFS_BLOCK_GROUP_RAID10))
4642                         factor = 2;
4643                 else
4644                         factor = 1;
4645                 /*
4646                  * If this block group has free space cache written out, we
4647                  * need to make sure to load it if we are removing space.  This
4648                  * is because we need the unpinning stage to actually add the
4649                  * space back to the block group, otherwise we will leak space.
4650                  */
4651                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4652                         cache_block_group(cache, trans, NULL, 1);
4653
4654                 byte_in_group = bytenr - cache->key.objectid;
4655                 WARN_ON(byte_in_group > cache->key.offset);
4656
4657                 spin_lock(&cache->space_info->lock);
4658                 spin_lock(&cache->lock);
4659
4660                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4661                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4662                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4663
4664                 cache->dirty = 1;
4665                 old_val = btrfs_block_group_used(&cache->item);
4666                 num_bytes = min(total, cache->key.offset - byte_in_group);
4667                 if (alloc) {
4668                         old_val += num_bytes;
4669                         btrfs_set_block_group_used(&cache->item, old_val);
4670                         cache->reserved -= num_bytes;
4671                         cache->space_info->bytes_reserved -= num_bytes;
4672                         cache->space_info->bytes_used += num_bytes;
4673                         cache->space_info->disk_used += num_bytes * factor;
4674                         spin_unlock(&cache->lock);
4675                         spin_unlock(&cache->space_info->lock);
4676                 } else {
4677                         old_val -= num_bytes;
4678                         btrfs_set_block_group_used(&cache->item, old_val);
4679                         cache->pinned += num_bytes;
4680                         cache->space_info->bytes_pinned += num_bytes;
4681                         cache->space_info->bytes_used -= num_bytes;
4682                         cache->space_info->disk_used -= num_bytes * factor;
4683                         spin_unlock(&cache->lock);
4684                         spin_unlock(&cache->space_info->lock);
4685
4686                         set_extent_dirty(info->pinned_extents,
4687                                          bytenr, bytenr + num_bytes - 1,
4688                                          GFP_NOFS | __GFP_NOFAIL);
4689                 }
4690                 btrfs_put_block_group(cache);
4691                 total -= num_bytes;
4692                 bytenr += num_bytes;
4693         }
4694         return 0;
4695 }
4696
4697 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4698 {
4699         struct btrfs_block_group_cache *cache;
4700         u64 bytenr;
4701
4702         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4703         if (!cache)
4704                 return 0;
4705
4706         bytenr = cache->key.objectid;
4707         btrfs_put_block_group(cache);
4708
4709         return bytenr;
4710 }
4711
4712 static int pin_down_extent(struct btrfs_root *root,
4713                            struct btrfs_block_group_cache *cache,
4714                            u64 bytenr, u64 num_bytes, int reserved)
4715 {
4716         spin_lock(&cache->space_info->lock);
4717         spin_lock(&cache->lock);
4718         cache->pinned += num_bytes;
4719         cache->space_info->bytes_pinned += num_bytes;
4720         if (reserved) {
4721                 cache->reserved -= num_bytes;
4722                 cache->space_info->bytes_reserved -= num_bytes;
4723         }
4724         spin_unlock(&cache->lock);
4725         spin_unlock(&cache->space_info->lock);
4726
4727         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4728                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4729         return 0;
4730 }
4731
4732 /*
4733  * this function must be called within transaction
4734  */
4735 int btrfs_pin_extent(struct btrfs_root *root,
4736                      u64 bytenr, u64 num_bytes, int reserved)
4737 {
4738         struct btrfs_block_group_cache *cache;
4739
4740         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4741         BUG_ON(!cache); /* Logic error */
4742
4743         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4744
4745         btrfs_put_block_group(cache);
4746         return 0;
4747 }
4748
4749 /*
4750  * this function must be called within transaction
4751  */
4752 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4753                                     struct btrfs_root *root,
4754                                     u64 bytenr, u64 num_bytes)
4755 {
4756         struct btrfs_block_group_cache *cache;
4757
4758         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4759         BUG_ON(!cache); /* Logic error */
4760
4761         /*
4762          * pull in the free space cache (if any) so that our pin
4763          * removes the free space from the cache.  We have load_only set
4764          * to one because the slow code to read in the free extents does check
4765          * the pinned extents.
4766          */
4767         cache_block_group(cache, trans, root, 1);
4768
4769         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4770
4771         /* remove us from the free space cache (if we're there at all) */
4772         btrfs_remove_free_space(cache, bytenr, num_bytes);
4773         btrfs_put_block_group(cache);
4774         return 0;
4775 }
4776
4777 /**
4778  * btrfs_update_reserved_bytes - update the block_group and space info counters
4779  * @cache:      The cache we are manipulating
4780  * @num_bytes:  The number of bytes in question
4781  * @reserve:    One of the reservation enums
4782  *
4783  * This is called by the allocator when it reserves space, or by somebody who is
4784  * freeing space that was never actually used on disk.  For example if you
4785  * reserve some space for a new leaf in transaction A and before transaction A
4786  * commits you free that leaf, you call this with reserve set to 0 in order to
4787  * clear the reservation.
4788  *
4789  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4790  * ENOSPC accounting.  For data we handle the reservation through clearing the
4791  * delalloc bits in the io_tree.  We have to do this since we could end up
4792  * allocating less disk space for the amount of data we have reserved in the
4793  * case of compression.
4794  *
4795  * If this is a reservation and the block group has become read only we cannot
4796  * make the reservation and return -EAGAIN, otherwise this function always
4797  * succeeds.
4798  */
4799 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4800                                        u64 num_bytes, int reserve)
4801 {
4802         struct btrfs_space_info *space_info = cache->space_info;
4803         int ret = 0;
4804
4805         spin_lock(&space_info->lock);
4806         spin_lock(&cache->lock);
4807         if (reserve != RESERVE_FREE) {
4808                 if (cache->ro) {
4809                         ret = -EAGAIN;
4810                 } else {
4811                         cache->reserved += num_bytes;
4812                         space_info->bytes_reserved += num_bytes;
4813                         if (reserve == RESERVE_ALLOC) {
4814                                 trace_btrfs_space_reservation(cache->fs_info,
4815                                                 "space_info", space_info->flags,
4816                                                 num_bytes, 0);
4817                                 space_info->bytes_may_use -= num_bytes;
4818                         }
4819                 }
4820         } else {
4821                 if (cache->ro)
4822                         space_info->bytes_readonly += num_bytes;
4823                 cache->reserved -= num_bytes;
4824                 space_info->bytes_reserved -= num_bytes;
4825                 space_info->reservation_progress++;
4826         }
4827         spin_unlock(&cache->lock);
4828         spin_unlock(&space_info->lock);
4829         return ret;
4830 }
4831
4832 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4833                                 struct btrfs_root *root)
4834 {
4835         struct btrfs_fs_info *fs_info = root->fs_info;
4836         struct btrfs_caching_control *next;
4837         struct btrfs_caching_control *caching_ctl;
4838         struct btrfs_block_group_cache *cache;
4839
4840         down_write(&fs_info->extent_commit_sem);
4841
4842         list_for_each_entry_safe(caching_ctl, next,
4843                                  &fs_info->caching_block_groups, list) {
4844                 cache = caching_ctl->block_group;
4845                 if (block_group_cache_done(cache)) {
4846                         cache->last_byte_to_unpin = (u64)-1;
4847                         list_del_init(&caching_ctl->list);
4848                         put_caching_control(caching_ctl);
4849                 } else {
4850                         cache->last_byte_to_unpin = caching_ctl->progress;
4851                 }
4852         }
4853
4854         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4855                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4856         else
4857                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4858
4859         up_write(&fs_info->extent_commit_sem);
4860
4861         update_global_block_rsv(fs_info);
4862 }
4863
4864 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4865 {
4866         struct btrfs_fs_info *fs_info = root->fs_info;
4867         struct btrfs_block_group_cache *cache = NULL;
4868         u64 len;
4869
4870         while (start <= end) {
4871                 if (!cache ||
4872                     start >= cache->key.objectid + cache->key.offset) {
4873                         if (cache)
4874                                 btrfs_put_block_group(cache);
4875                         cache = btrfs_lookup_block_group(fs_info, start);
4876                         BUG_ON(!cache); /* Logic error */
4877                 }
4878
4879                 len = cache->key.objectid + cache->key.offset - start;
4880                 len = min(len, end + 1 - start);
4881
4882                 if (start < cache->last_byte_to_unpin) {
4883                         len = min(len, cache->last_byte_to_unpin - start);
4884                         btrfs_add_free_space(cache, start, len);
4885                 }
4886
4887                 start += len;
4888
4889                 spin_lock(&cache->space_info->lock);
4890                 spin_lock(&cache->lock);
4891                 cache->pinned -= len;
4892                 cache->space_info->bytes_pinned -= len;
4893                 if (cache->ro)
4894                         cache->space_info->bytes_readonly += len;
4895                 spin_unlock(&cache->lock);
4896                 spin_unlock(&cache->space_info->lock);
4897         }
4898
4899         if (cache)
4900                 btrfs_put_block_group(cache);
4901         return 0;
4902 }
4903
4904 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4905                                struct btrfs_root *root)
4906 {
4907         struct btrfs_fs_info *fs_info = root->fs_info;
4908         struct extent_io_tree *unpin;
4909         u64 start;
4910         u64 end;
4911         int ret;
4912
4913         if (trans->aborted)
4914                 return 0;
4915
4916         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4917                 unpin = &fs_info->freed_extents[1];
4918         else
4919                 unpin = &fs_info->freed_extents[0];
4920
4921         while (1) {
4922                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4923                                             EXTENT_DIRTY);
4924                 if (ret)
4925                         break;
4926
4927                 if (btrfs_test_opt(root, DISCARD))
4928                         ret = btrfs_discard_extent(root, start,
4929                                                    end + 1 - start, NULL);
4930
4931                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4932                 unpin_extent_range(root, start, end);
4933                 cond_resched();
4934         }
4935
4936         return 0;
4937 }
4938
4939 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4940                                 struct btrfs_root *root,
4941                                 u64 bytenr, u64 num_bytes, u64 parent,
4942                                 u64 root_objectid, u64 owner_objectid,
4943                                 u64 owner_offset, int refs_to_drop,
4944                                 struct btrfs_delayed_extent_op *extent_op)
4945 {
4946         struct btrfs_key key;
4947         struct btrfs_path *path;
4948         struct btrfs_fs_info *info = root->fs_info;
4949         struct btrfs_root *extent_root = info->extent_root;
4950         struct extent_buffer *leaf;
4951         struct btrfs_extent_item *ei;
4952         struct btrfs_extent_inline_ref *iref;
4953         int ret;
4954         int is_data;
4955         int extent_slot = 0;
4956         int found_extent = 0;
4957         int num_to_del = 1;
4958         u32 item_size;
4959         u64 refs;
4960
4961         path = btrfs_alloc_path();
4962         if (!path)
4963                 return -ENOMEM;
4964
4965         path->reada = 1;
4966         path->leave_spinning = 1;
4967
4968         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4969         BUG_ON(!is_data && refs_to_drop != 1);
4970
4971         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4972                                     bytenr, num_bytes, parent,
4973                                     root_objectid, owner_objectid,
4974                                     owner_offset);
4975         if (ret == 0) {
4976                 extent_slot = path->slots[0];
4977                 while (extent_slot >= 0) {
4978                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4979                                               extent_slot);
4980                         if (key.objectid != bytenr)
4981                                 break;
4982                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4983                             key.offset == num_bytes) {
4984                                 found_extent = 1;
4985                                 break;
4986                         }
4987                         if (path->slots[0] - extent_slot > 5)
4988                                 break;
4989                         extent_slot--;
4990                 }
4991 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4992                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4993                 if (found_extent && item_size < sizeof(*ei))
4994                         found_extent = 0;
4995 #endif
4996                 if (!found_extent) {
4997                         BUG_ON(iref);
4998                         ret = remove_extent_backref(trans, extent_root, path,
4999                                                     NULL, refs_to_drop,
5000                                                     is_data);
5001                         if (ret)
5002                                 goto abort;
5003                         btrfs_release_path(path);
5004                         path->leave_spinning = 1;
5005
5006                         key.objectid = bytenr;
5007                         key.type = BTRFS_EXTENT_ITEM_KEY;
5008                         key.offset = num_bytes;
5009
5010                         ret = btrfs_search_slot(trans, extent_root,
5011                                                 &key, path, -1, 1);
5012                         if (ret) {
5013                                 printk(KERN_ERR "umm, got %d back from search"
5014                                        ", was looking for %llu\n", ret,
5015                                        (unsigned long long)bytenr);
5016                                 if (ret > 0)
5017                                         btrfs_print_leaf(extent_root,
5018                                                          path->nodes[0]);
5019                         }
5020                         if (ret < 0)
5021                                 goto abort;
5022                         extent_slot = path->slots[0];
5023                 }
5024         } else if (ret == -ENOENT) {
5025                 btrfs_print_leaf(extent_root, path->nodes[0]);
5026                 WARN_ON(1);
5027                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5028                        "parent %llu root %llu  owner %llu offset %llu\n",
5029                        (unsigned long long)bytenr,
5030                        (unsigned long long)parent,
5031                        (unsigned long long)root_objectid,
5032                        (unsigned long long)owner_objectid,
5033                        (unsigned long long)owner_offset);
5034         } else {
5035                 goto abort;
5036         }
5037
5038         leaf = path->nodes[0];
5039         item_size = btrfs_item_size_nr(leaf, extent_slot);
5040 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5041         if (item_size < sizeof(*ei)) {
5042                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5043                 ret = convert_extent_item_v0(trans, extent_root, path,
5044                                              owner_objectid, 0);
5045                 if (ret < 0)
5046                         goto abort;
5047
5048                 btrfs_release_path(path);
5049                 path->leave_spinning = 1;
5050
5051                 key.objectid = bytenr;
5052                 key.type = BTRFS_EXTENT_ITEM_KEY;
5053                 key.offset = num_bytes;
5054
5055                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5056                                         -1, 1);
5057                 if (ret) {
5058                         printk(KERN_ERR "umm, got %d back from search"
5059                                ", was looking for %llu\n", ret,
5060                                (unsigned long long)bytenr);
5061                         btrfs_print_leaf(extent_root, path->nodes[0]);
5062                 }
5063                 if (ret < 0)
5064                         goto abort;
5065                 extent_slot = path->slots[0];
5066                 leaf = path->nodes[0];
5067                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5068         }
5069 #endif
5070         BUG_ON(item_size < sizeof(*ei));
5071         ei = btrfs_item_ptr(leaf, extent_slot,
5072                             struct btrfs_extent_item);
5073         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5074                 struct btrfs_tree_block_info *bi;
5075                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5076                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5077                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5078         }
5079
5080         refs = btrfs_extent_refs(leaf, ei);
5081         BUG_ON(refs < refs_to_drop);
5082         refs -= refs_to_drop;
5083
5084         if (refs > 0) {
5085                 if (extent_op)
5086                         __run_delayed_extent_op(extent_op, leaf, ei);
5087                 /*
5088                  * In the case of inline back ref, reference count will
5089                  * be updated by remove_extent_backref
5090                  */
5091                 if (iref) {
5092                         BUG_ON(!found_extent);
5093                 } else {
5094                         btrfs_set_extent_refs(leaf, ei, refs);
5095                         btrfs_mark_buffer_dirty(leaf);
5096                 }
5097                 if (found_extent) {
5098                         ret = remove_extent_backref(trans, extent_root, path,
5099                                                     iref, refs_to_drop,
5100                                                     is_data);
5101                         if (ret)
5102                                 goto abort;
5103                 }
5104         } else {
5105                 if (found_extent) {
5106                         BUG_ON(is_data && refs_to_drop !=
5107                                extent_data_ref_count(root, path, iref));
5108                         if (iref) {
5109                                 BUG_ON(path->slots[0] != extent_slot);
5110                         } else {
5111                                 BUG_ON(path->slots[0] != extent_slot + 1);
5112                                 path->slots[0] = extent_slot;
5113                                 num_to_del = 2;
5114                         }
5115                 }
5116
5117                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5118                                       num_to_del);
5119                 if (ret)
5120                         goto abort;
5121                 btrfs_release_path(path);
5122
5123                 if (is_data) {
5124                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5125                         if (ret)
5126                                 goto abort;
5127                 }
5128
5129                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5130                 if (ret)
5131                         goto abort;
5132         }
5133 out:
5134         btrfs_free_path(path);
5135         return ret;
5136
5137 abort:
5138         btrfs_abort_transaction(trans, extent_root, ret);
5139         goto out;
5140 }
5141
5142 /*
5143  * when we free an block, it is possible (and likely) that we free the last
5144  * delayed ref for that extent as well.  This searches the delayed ref tree for
5145  * a given extent, and if there are no other delayed refs to be processed, it
5146  * removes it from the tree.
5147  */
5148 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5149                                       struct btrfs_root *root, u64 bytenr)
5150 {
5151         struct btrfs_delayed_ref_head *head;
5152         struct btrfs_delayed_ref_root *delayed_refs;
5153         struct btrfs_delayed_ref_node *ref;
5154         struct rb_node *node;
5155         int ret = 0;
5156
5157         delayed_refs = &trans->transaction->delayed_refs;
5158         spin_lock(&delayed_refs->lock);
5159         head = btrfs_find_delayed_ref_head(trans, bytenr);
5160         if (!head)
5161                 goto out;
5162
5163         node = rb_prev(&head->node.rb_node);
5164         if (!node)
5165                 goto out;
5166
5167         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5168
5169         /* there are still entries for this ref, we can't drop it */
5170         if (ref->bytenr == bytenr)
5171                 goto out;
5172
5173         if (head->extent_op) {
5174                 if (!head->must_insert_reserved)
5175                         goto out;
5176                 kfree(head->extent_op);
5177                 head->extent_op = NULL;
5178         }
5179
5180         /*
5181          * waiting for the lock here would deadlock.  If someone else has it
5182          * locked they are already in the process of dropping it anyway
5183          */
5184         if (!mutex_trylock(&head->mutex))
5185                 goto out;
5186
5187         /*
5188          * at this point we have a head with no other entries.  Go
5189          * ahead and process it.
5190          */
5191         head->node.in_tree = 0;
5192         rb_erase(&head->node.rb_node, &delayed_refs->root);
5193
5194         delayed_refs->num_entries--;
5195         if (waitqueue_active(&delayed_refs->seq_wait))
5196                 wake_up(&delayed_refs->seq_wait);
5197
5198         /*
5199          * we don't take a ref on the node because we're removing it from the
5200          * tree, so we just steal the ref the tree was holding.
5201          */
5202         delayed_refs->num_heads--;
5203         if (list_empty(&head->cluster))
5204                 delayed_refs->num_heads_ready--;
5205
5206         list_del_init(&head->cluster);
5207         spin_unlock(&delayed_refs->lock);
5208
5209         BUG_ON(head->extent_op);
5210         if (head->must_insert_reserved)
5211                 ret = 1;
5212
5213         mutex_unlock(&head->mutex);
5214         btrfs_put_delayed_ref(&head->node);
5215         return ret;
5216 out:
5217         spin_unlock(&delayed_refs->lock);
5218         return 0;
5219 }
5220
5221 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5222                            struct btrfs_root *root,
5223                            struct extent_buffer *buf,
5224                            u64 parent, int last_ref, int for_cow)
5225 {
5226         struct btrfs_block_group_cache *cache = NULL;
5227         int ret;
5228
5229         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5230                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5231                                         buf->start, buf->len,
5232                                         parent, root->root_key.objectid,
5233                                         btrfs_header_level(buf),
5234                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5235                 BUG_ON(ret); /* -ENOMEM */
5236         }
5237
5238         if (!last_ref)
5239                 return;
5240
5241         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5242
5243         if (btrfs_header_generation(buf) == trans->transid) {
5244                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5245                         ret = check_ref_cleanup(trans, root, buf->start);
5246                         if (!ret)
5247                                 goto out;
5248                 }
5249
5250                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5251                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5252                         goto out;
5253                 }
5254
5255                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5256
5257                 btrfs_add_free_space(cache, buf->start, buf->len);
5258                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5259         }
5260 out:
5261         /*
5262          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5263          * anymore.
5264          */
5265         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5266         btrfs_put_block_group(cache);
5267 }
5268
5269 /* Can return -ENOMEM */
5270 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5271                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5272                       u64 owner, u64 offset, int for_cow)
5273 {
5274         int ret;
5275         struct btrfs_fs_info *fs_info = root->fs_info;
5276
5277         /*
5278          * tree log blocks never actually go into the extent allocation
5279          * tree, just update pinning info and exit early.
5280          */
5281         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5282                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5283                 /* unlocks the pinned mutex */
5284                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5285                 ret = 0;
5286         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5287                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5288                                         num_bytes,
5289                                         parent, root_objectid, (int)owner,
5290                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5291         } else {
5292                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5293                                                 num_bytes,
5294                                                 parent, root_objectid, owner,
5295                                                 offset, BTRFS_DROP_DELAYED_REF,
5296                                                 NULL, for_cow);
5297         }
5298         return ret;
5299 }
5300
5301 static u64 stripe_align(struct btrfs_root *root, u64 val)
5302 {
5303         u64 mask = ((u64)root->stripesize - 1);
5304         u64 ret = (val + mask) & ~mask;
5305         return ret;
5306 }
5307
5308 /*
5309  * when we wait for progress in the block group caching, its because
5310  * our allocation attempt failed at least once.  So, we must sleep
5311  * and let some progress happen before we try again.
5312  *
5313  * This function will sleep at least once waiting for new free space to
5314  * show up, and then it will check the block group free space numbers
5315  * for our min num_bytes.  Another option is to have it go ahead
5316  * and look in the rbtree for a free extent of a given size, but this
5317  * is a good start.
5318  */
5319 static noinline int
5320 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5321                                 u64 num_bytes)
5322 {
5323         struct btrfs_caching_control *caching_ctl;
5324         DEFINE_WAIT(wait);
5325
5326         caching_ctl = get_caching_control(cache);
5327         if (!caching_ctl)
5328                 return 0;
5329
5330         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5331                    (cache->free_space_ctl->free_space >= num_bytes));
5332
5333         put_caching_control(caching_ctl);
5334         return 0;
5335 }
5336
5337 static noinline int
5338 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5339 {
5340         struct btrfs_caching_control *caching_ctl;
5341         DEFINE_WAIT(wait);
5342
5343         caching_ctl = get_caching_control(cache);
5344         if (!caching_ctl)
5345                 return 0;
5346
5347         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5348
5349         put_caching_control(caching_ctl);
5350         return 0;
5351 }
5352
5353 static int __get_block_group_index(u64 flags)
5354 {
5355         int index;
5356
5357         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5358                 index = 0;
5359         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5360                 index = 1;
5361         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5362                 index = 2;
5363         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5364                 index = 3;
5365         else
5366                 index = 4;
5367
5368         return index;
5369 }
5370
5371 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5372 {
5373         return __get_block_group_index(cache->flags);
5374 }
5375
5376 enum btrfs_loop_type {
5377         LOOP_CACHING_NOWAIT = 0,
5378         LOOP_CACHING_WAIT = 1,
5379         LOOP_ALLOC_CHUNK = 2,
5380         LOOP_NO_EMPTY_SIZE = 3,
5381 };
5382
5383 /*
5384  * walks the btree of allocated extents and find a hole of a given size.
5385  * The key ins is changed to record the hole:
5386  * ins->objectid == block start
5387  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5388  * ins->offset == number of blocks
5389  * Any available blocks before search_start are skipped.
5390  */
5391 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5392                                      struct btrfs_root *orig_root,
5393                                      u64 num_bytes, u64 empty_size,
5394                                      u64 hint_byte, struct btrfs_key *ins,
5395                                      u64 data)
5396 {
5397         int ret = 0;
5398         struct btrfs_root *root = orig_root->fs_info->extent_root;
5399         struct btrfs_free_cluster *last_ptr = NULL;
5400         struct btrfs_block_group_cache *block_group = NULL;
5401         struct btrfs_block_group_cache *used_block_group;
5402         u64 search_start = 0;
5403         int empty_cluster = 2 * 1024 * 1024;
5404         int allowed_chunk_alloc = 0;
5405         int done_chunk_alloc = 0;
5406         struct btrfs_space_info *space_info;
5407         int loop = 0;
5408         int index = 0;
5409         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5410                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5411         bool found_uncached_bg = false;
5412         bool failed_cluster_refill = false;
5413         bool failed_alloc = false;
5414         bool use_cluster = true;
5415         bool have_caching_bg = false;
5416
5417         WARN_ON(num_bytes < root->sectorsize);
5418         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5419         ins->objectid = 0;
5420         ins->offset = 0;
5421
5422         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5423
5424         space_info = __find_space_info(root->fs_info, data);
5425         if (!space_info) {
5426                 printk(KERN_ERR "No space info for %llu\n", data);
5427                 return -ENOSPC;
5428         }
5429
5430         /*
5431          * If the space info is for both data and metadata it means we have a
5432          * small filesystem and we can't use the clustering stuff.
5433          */
5434         if (btrfs_mixed_space_info(space_info))
5435                 use_cluster = false;
5436
5437         if (orig_root->ref_cows || empty_size)
5438                 allowed_chunk_alloc = 1;
5439
5440         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5441                 last_ptr = &root->fs_info->meta_alloc_cluster;
5442                 if (!btrfs_test_opt(root, SSD))
5443                         empty_cluster = 64 * 1024;
5444         }
5445
5446         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5447             btrfs_test_opt(root, SSD)) {
5448                 last_ptr = &root->fs_info->data_alloc_cluster;
5449         }
5450
5451         if (last_ptr) {
5452                 spin_lock(&last_ptr->lock);
5453                 if (last_ptr->block_group)
5454                         hint_byte = last_ptr->window_start;
5455                 spin_unlock(&last_ptr->lock);
5456         }
5457
5458         search_start = max(search_start, first_logical_byte(root, 0));
5459         search_start = max(search_start, hint_byte);
5460
5461         if (!last_ptr)
5462                 empty_cluster = 0;
5463
5464         if (search_start == hint_byte) {
5465                 block_group = btrfs_lookup_block_group(root->fs_info,
5466                                                        search_start);
5467                 used_block_group = block_group;
5468                 /*
5469                  * we don't want to use the block group if it doesn't match our
5470                  * allocation bits, or if its not cached.
5471                  *
5472                  * However if we are re-searching with an ideal block group
5473                  * picked out then we don't care that the block group is cached.
5474                  */
5475                 if (block_group && block_group_bits(block_group, data) &&
5476                     block_group->cached != BTRFS_CACHE_NO) {
5477                         down_read(&space_info->groups_sem);
5478                         if (list_empty(&block_group->list) ||
5479                             block_group->ro) {
5480                                 /*
5481                                  * someone is removing this block group,
5482                                  * we can't jump into the have_block_group
5483                                  * target because our list pointers are not
5484                                  * valid
5485                                  */
5486                                 btrfs_put_block_group(block_group);
5487                                 up_read(&space_info->groups_sem);
5488                         } else {
5489                                 index = get_block_group_index(block_group);
5490                                 goto have_block_group;
5491                         }
5492                 } else if (block_group) {
5493                         btrfs_put_block_group(block_group);
5494                 }
5495         }
5496 search:
5497         have_caching_bg = false;
5498         down_read(&space_info->groups_sem);
5499         list_for_each_entry(block_group, &space_info->block_groups[index],
5500                             list) {
5501                 u64 offset;
5502                 int cached;
5503
5504                 used_block_group = block_group;
5505                 btrfs_get_block_group(block_group);
5506                 search_start = block_group->key.objectid;
5507
5508                 /*
5509                  * this can happen if we end up cycling through all the
5510                  * raid types, but we want to make sure we only allocate
5511                  * for the proper type.
5512                  */
5513                 if (!block_group_bits(block_group, data)) {
5514                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5515                                 BTRFS_BLOCK_GROUP_RAID1 |
5516                                 BTRFS_BLOCK_GROUP_RAID10;
5517
5518                         /*
5519                          * if they asked for extra copies and this block group
5520                          * doesn't provide them, bail.  This does allow us to
5521                          * fill raid0 from raid1.
5522                          */
5523                         if ((data & extra) && !(block_group->flags & extra))
5524                                 goto loop;
5525                 }
5526
5527 have_block_group:
5528                 cached = block_group_cache_done(block_group);
5529                 if (unlikely(!cached)) {
5530                         found_uncached_bg = true;
5531                         ret = cache_block_group(block_group, trans,
5532                                                 orig_root, 0);
5533                         BUG_ON(ret < 0);
5534                         ret = 0;
5535                 }
5536
5537                 if (unlikely(block_group->ro))
5538                         goto loop;
5539
5540                 /*
5541                  * Ok we want to try and use the cluster allocator, so
5542                  * lets look there
5543                  */
5544                 if (last_ptr) {
5545                         /*
5546                          * the refill lock keeps out other
5547                          * people trying to start a new cluster
5548                          */
5549                         spin_lock(&last_ptr->refill_lock);
5550                         used_block_group = last_ptr->block_group;
5551                         if (used_block_group != block_group &&
5552                             (!used_block_group ||
5553                              used_block_group->ro ||
5554                              !block_group_bits(used_block_group, data))) {
5555                                 used_block_group = block_group;
5556                                 goto refill_cluster;
5557                         }
5558
5559                         if (used_block_group != block_group)
5560                                 btrfs_get_block_group(used_block_group);
5561
5562                         offset = btrfs_alloc_from_cluster(used_block_group,
5563                           last_ptr, num_bytes, used_block_group->key.objectid);
5564                         if (offset) {
5565                                 /* we have a block, we're done */
5566                                 spin_unlock(&last_ptr->refill_lock);
5567                                 trace_btrfs_reserve_extent_cluster(root,
5568                                         block_group, search_start, num_bytes);
5569                                 goto checks;
5570                         }
5571
5572                         WARN_ON(last_ptr->block_group != used_block_group);
5573                         if (used_block_group != block_group) {
5574                                 btrfs_put_block_group(used_block_group);
5575                                 used_block_group = block_group;
5576                         }
5577 refill_cluster:
5578                         BUG_ON(used_block_group != block_group);
5579                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5580                          * set up a new clusters, so lets just skip it
5581                          * and let the allocator find whatever block
5582                          * it can find.  If we reach this point, we
5583                          * will have tried the cluster allocator
5584                          * plenty of times and not have found
5585                          * anything, so we are likely way too
5586                          * fragmented for the clustering stuff to find
5587                          * anything.
5588                          *
5589                          * However, if the cluster is taken from the
5590                          * current block group, release the cluster
5591                          * first, so that we stand a better chance of
5592                          * succeeding in the unclustered
5593                          * allocation.  */
5594                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5595                             last_ptr->block_group != block_group) {
5596                                 spin_unlock(&last_ptr->refill_lock);
5597                                 goto unclustered_alloc;
5598                         }
5599
5600                         /*
5601                          * this cluster didn't work out, free it and
5602                          * start over
5603                          */
5604                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5605
5606                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5607                                 spin_unlock(&last_ptr->refill_lock);
5608                                 goto unclustered_alloc;
5609                         }
5610
5611                         /* allocate a cluster in this block group */
5612                         ret = btrfs_find_space_cluster(trans, root,
5613                                                block_group, last_ptr,
5614                                                search_start, num_bytes,
5615                                                empty_cluster + empty_size);
5616                         if (ret == 0) {
5617                                 /*
5618                                  * now pull our allocation out of this
5619                                  * cluster
5620                                  */
5621                                 offset = btrfs_alloc_from_cluster(block_group,
5622                                                   last_ptr, num_bytes,
5623                                                   search_start);
5624                                 if (offset) {
5625                                         /* we found one, proceed */
5626                                         spin_unlock(&last_ptr->refill_lock);
5627                                         trace_btrfs_reserve_extent_cluster(root,
5628                                                 block_group, search_start,
5629                                                 num_bytes);
5630                                         goto checks;
5631                                 }
5632                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5633                                    && !failed_cluster_refill) {
5634                                 spin_unlock(&last_ptr->refill_lock);
5635
5636                                 failed_cluster_refill = true;
5637                                 wait_block_group_cache_progress(block_group,
5638