6621ed72f3c3f2b9ed09dcd8132482b39fbcb695
[linux-3.10.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36
37 /*
38  * control flags for do_chunk_alloc's force field
39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
40  * if we really need one.
41  *
42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
43  * if we have very few chunks already allocated.  This is
44  * used as part of the clustering code to help make sure
45  * we have a good pool of storage to cluster in, without
46  * filling the FS with empty chunks
47  *
48  * CHUNK_ALLOC_FORCE means it must try to allocate one
49  *
50  */
51 enum {
52         CHUNK_ALLOC_NO_FORCE = 0,
53         CHUNK_ALLOC_LIMITED = 1,
54         CHUNK_ALLOC_FORCE = 2,
55 };
56
57 /*
58  * Control how reservations are dealt with.
59  *
60  * RESERVE_FREE - freeing a reservation.
61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
62  *   ENOSPC accounting
63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
65  */
66 enum {
67         RESERVE_FREE = 0,
68         RESERVE_ALLOC = 1,
69         RESERVE_ALLOC_NO_ACCOUNT = 2,
70 };
71
72 static int update_block_group(struct btrfs_trans_handle *trans,
73                               struct btrfs_root *root,
74                               u64 bytenr, u64 num_bytes, int alloc);
75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
76                                 struct btrfs_root *root,
77                                 u64 bytenr, u64 num_bytes, u64 parent,
78                                 u64 root_objectid, u64 owner_objectid,
79                                 u64 owner_offset, int refs_to_drop,
80                                 struct btrfs_delayed_extent_op *extra_op);
81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
82                                     struct extent_buffer *leaf,
83                                     struct btrfs_extent_item *ei);
84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
85                                       struct btrfs_root *root,
86                                       u64 parent, u64 root_objectid,
87                                       u64 flags, u64 owner, u64 offset,
88                                       struct btrfs_key *ins, int ref_mod);
89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
90                                      struct btrfs_root *root,
91                                      u64 parent, u64 root_objectid,
92                                      u64 flags, struct btrfs_disk_key *key,
93                                      int level, struct btrfs_key *ins);
94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
95                           struct btrfs_root *extent_root, u64 alloc_bytes,
96                           u64 flags, int force);
97 static int find_next_key(struct btrfs_path *path, int level,
98                          struct btrfs_key *key);
99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
100                             int dump_block_groups);
101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
102                                        u64 num_bytes, int reserve);
103
104 static noinline int
105 block_group_cache_done(struct btrfs_block_group_cache *cache)
106 {
107         smp_mb();
108         return cache->cached == BTRFS_CACHE_FINISHED;
109 }
110
111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
112 {
113         return (cache->flags & bits) == bits;
114 }
115
116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
117 {
118         atomic_inc(&cache->count);
119 }
120
121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
122 {
123         if (atomic_dec_and_test(&cache->count)) {
124                 WARN_ON(cache->pinned > 0);
125                 WARN_ON(cache->reserved > 0);
126                 kfree(cache->free_space_ctl);
127                 kfree(cache);
128         }
129 }
130
131 /*
132  * this adds the block group to the fs_info rb tree for the block group
133  * cache
134  */
135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
136                                 struct btrfs_block_group_cache *block_group)
137 {
138         struct rb_node **p;
139         struct rb_node *parent = NULL;
140         struct btrfs_block_group_cache *cache;
141
142         spin_lock(&info->block_group_cache_lock);
143         p = &info->block_group_cache_tree.rb_node;
144
145         while (*p) {
146                 parent = *p;
147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
148                                  cache_node);
149                 if (block_group->key.objectid < cache->key.objectid) {
150                         p = &(*p)->rb_left;
151                 } else if (block_group->key.objectid > cache->key.objectid) {
152                         p = &(*p)->rb_right;
153                 } else {
154                         spin_unlock(&info->block_group_cache_lock);
155                         return -EEXIST;
156                 }
157         }
158
159         rb_link_node(&block_group->cache_node, parent, p);
160         rb_insert_color(&block_group->cache_node,
161                         &info->block_group_cache_tree);
162         spin_unlock(&info->block_group_cache_lock);
163
164         return 0;
165 }
166
167 /*
168  * This will return the block group at or after bytenr if contains is 0, else
169  * it will return the block group that contains the bytenr
170  */
171 static struct btrfs_block_group_cache *
172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
173                               int contains)
174 {
175         struct btrfs_block_group_cache *cache, *ret = NULL;
176         struct rb_node *n;
177         u64 end, start;
178
179         spin_lock(&info->block_group_cache_lock);
180         n = info->block_group_cache_tree.rb_node;
181
182         while (n) {
183                 cache = rb_entry(n, struct btrfs_block_group_cache,
184                                  cache_node);
185                 end = cache->key.objectid + cache->key.offset - 1;
186                 start = cache->key.objectid;
187
188                 if (bytenr < start) {
189                         if (!contains && (!ret || start < ret->key.objectid))
190                                 ret = cache;
191                         n = n->rb_left;
192                 } else if (bytenr > start) {
193                         if (contains && bytenr <= end) {
194                                 ret = cache;
195                                 break;
196                         }
197                         n = n->rb_right;
198                 } else {
199                         ret = cache;
200                         break;
201                 }
202         }
203         if (ret)
204                 btrfs_get_block_group(ret);
205         spin_unlock(&info->block_group_cache_lock);
206
207         return ret;
208 }
209
210 static int add_excluded_extent(struct btrfs_root *root,
211                                u64 start, u64 num_bytes)
212 {
213         u64 end = start + num_bytes - 1;
214         set_extent_bits(&root->fs_info->freed_extents[0],
215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
216         set_extent_bits(&root->fs_info->freed_extents[1],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         return 0;
219 }
220
221 static void free_excluded_extents(struct btrfs_root *root,
222                                   struct btrfs_block_group_cache *cache)
223 {
224         u64 start, end;
225
226         start = cache->key.objectid;
227         end = start + cache->key.offset - 1;
228
229         clear_extent_bits(&root->fs_info->freed_extents[0],
230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
231         clear_extent_bits(&root->fs_info->freed_extents[1],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233 }
234
235 static int exclude_super_stripes(struct btrfs_root *root,
236                                  struct btrfs_block_group_cache *cache)
237 {
238         u64 bytenr;
239         u64 *logical;
240         int stripe_len;
241         int i, nr, ret;
242
243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
245                 cache->bytes_super += stripe_len;
246                 ret = add_excluded_extent(root, cache->key.objectid,
247                                           stripe_len);
248                 BUG_ON(ret); /* -ENOMEM */
249         }
250
251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252                 bytenr = btrfs_sb_offset(i);
253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
254                                        cache->key.objectid, bytenr,
255                                        0, &logical, &nr, &stripe_len);
256                 BUG_ON(ret); /* -ENOMEM */
257
258                 while (nr--) {
259                         cache->bytes_super += stripe_len;
260                         ret = add_excluded_extent(root, logical[nr],
261                                                   stripe_len);
262                         BUG_ON(ret); /* -ENOMEM */
263                 }
264
265                 kfree(logical);
266         }
267         return 0;
268 }
269
270 static struct btrfs_caching_control *
271 get_caching_control(struct btrfs_block_group_cache *cache)
272 {
273         struct btrfs_caching_control *ctl;
274
275         spin_lock(&cache->lock);
276         if (cache->cached != BTRFS_CACHE_STARTED) {
277                 spin_unlock(&cache->lock);
278                 return NULL;
279         }
280
281         /* We're loading it the fast way, so we don't have a caching_ctl. */
282         if (!cache->caching_ctl) {
283                 spin_unlock(&cache->lock);
284                 return NULL;
285         }
286
287         ctl = cache->caching_ctl;
288         atomic_inc(&ctl->count);
289         spin_unlock(&cache->lock);
290         return ctl;
291 }
292
293 static void put_caching_control(struct btrfs_caching_control *ctl)
294 {
295         if (atomic_dec_and_test(&ctl->count))
296                 kfree(ctl);
297 }
298
299 /*
300  * this is only called by cache_block_group, since we could have freed extents
301  * we need to check the pinned_extents for any extents that can't be used yet
302  * since their free space will be released as soon as the transaction commits.
303  */
304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
305                               struct btrfs_fs_info *info, u64 start, u64 end)
306 {
307         u64 extent_start, extent_end, size, total_added = 0;
308         int ret;
309
310         while (start < end) {
311                 ret = find_first_extent_bit(info->pinned_extents, start,
312                                             &extent_start, &extent_end,
313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
314                 if (ret)
315                         break;
316
317                 if (extent_start <= start) {
318                         start = extent_end + 1;
319                 } else if (extent_start > start && extent_start < end) {
320                         size = extent_start - start;
321                         total_added += size;
322                         ret = btrfs_add_free_space(block_group, start,
323                                                    size);
324                         BUG_ON(ret); /* -ENOMEM or logic error */
325                         start = extent_end + 1;
326                 } else {
327                         break;
328                 }
329         }
330
331         if (start < end) {
332                 size = end - start;
333                 total_added += size;
334                 ret = btrfs_add_free_space(block_group, start, size);
335                 BUG_ON(ret); /* -ENOMEM or logic error */
336         }
337
338         return total_added;
339 }
340
341 static noinline void caching_thread(struct btrfs_work *work)
342 {
343         struct btrfs_block_group_cache *block_group;
344         struct btrfs_fs_info *fs_info;
345         struct btrfs_caching_control *caching_ctl;
346         struct btrfs_root *extent_root;
347         struct btrfs_path *path;
348         struct extent_buffer *leaf;
349         struct btrfs_key key;
350         u64 total_found = 0;
351         u64 last = 0;
352         u32 nritems;
353         int ret = 0;
354
355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
356         block_group = caching_ctl->block_group;
357         fs_info = block_group->fs_info;
358         extent_root = fs_info->extent_root;
359
360         path = btrfs_alloc_path();
361         if (!path)
362                 goto out;
363
364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
365
366         /*
367          * We don't want to deadlock with somebody trying to allocate a new
368          * extent for the extent root while also trying to search the extent
369          * root to add free space.  So we skip locking and search the commit
370          * root, since its read-only
371          */
372         path->skip_locking = 1;
373         path->search_commit_root = 1;
374         path->reada = 1;
375
376         key.objectid = last;
377         key.offset = 0;
378         key.type = BTRFS_EXTENT_ITEM_KEY;
379 again:
380         mutex_lock(&caching_ctl->mutex);
381         /* need to make sure the commit_root doesn't disappear */
382         down_read(&fs_info->extent_commit_sem);
383
384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
385         if (ret < 0)
386                 goto err;
387
388         leaf = path->nodes[0];
389         nritems = btrfs_header_nritems(leaf);
390
391         while (1) {
392                 if (btrfs_fs_closing(fs_info) > 1) {
393                         last = (u64)-1;
394                         break;
395                 }
396
397                 if (path->slots[0] < nritems) {
398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
399                 } else {
400                         ret = find_next_key(path, 0, &key);
401                         if (ret)
402                                 break;
403
404                         if (need_resched() ||
405                             btrfs_next_leaf(extent_root, path)) {
406                                 caching_ctl->progress = last;
407                                 btrfs_release_path(path);
408                                 up_read(&fs_info->extent_commit_sem);
409                                 mutex_unlock(&caching_ctl->mutex);
410                                 cond_resched();
411                                 goto again;
412                         }
413                         leaf = path->nodes[0];
414                         nritems = btrfs_header_nritems(leaf);
415                         continue;
416                 }
417
418                 if (key.objectid < block_group->key.objectid) {
419                         path->slots[0]++;
420                         continue;
421                 }
422
423                 if (key.objectid >= block_group->key.objectid +
424                     block_group->key.offset)
425                         break;
426
427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
428                         total_found += add_new_free_space(block_group,
429                                                           fs_info, last,
430                                                           key.objectid);
431                         last = key.objectid + key.offset;
432
433                         if (total_found > (1024 * 1024 * 2)) {
434                                 total_found = 0;
435                                 wake_up(&caching_ctl->wait);
436                         }
437                 }
438                 path->slots[0]++;
439         }
440         ret = 0;
441
442         total_found += add_new_free_space(block_group, fs_info, last,
443                                           block_group->key.objectid +
444                                           block_group->key.offset);
445         caching_ctl->progress = (u64)-1;
446
447         spin_lock(&block_group->lock);
448         block_group->caching_ctl = NULL;
449         block_group->cached = BTRFS_CACHE_FINISHED;
450         spin_unlock(&block_group->lock);
451
452 err:
453         btrfs_free_path(path);
454         up_read(&fs_info->extent_commit_sem);
455
456         free_excluded_extents(extent_root, block_group);
457
458         mutex_unlock(&caching_ctl->mutex);
459 out:
460         wake_up(&caching_ctl->wait);
461
462         put_caching_control(caching_ctl);
463         btrfs_put_block_group(block_group);
464 }
465
466 static int cache_block_group(struct btrfs_block_group_cache *cache,
467                              struct btrfs_trans_handle *trans,
468                              struct btrfs_root *root,
469                              int load_cache_only)
470 {
471         DEFINE_WAIT(wait);
472         struct btrfs_fs_info *fs_info = cache->fs_info;
473         struct btrfs_caching_control *caching_ctl;
474         int ret = 0;
475
476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
477         if (!caching_ctl)
478                 return -ENOMEM;
479
480         INIT_LIST_HEAD(&caching_ctl->list);
481         mutex_init(&caching_ctl->mutex);
482         init_waitqueue_head(&caching_ctl->wait);
483         caching_ctl->block_group = cache;
484         caching_ctl->progress = cache->key.objectid;
485         atomic_set(&caching_ctl->count, 1);
486         caching_ctl->work.func = caching_thread;
487
488         spin_lock(&cache->lock);
489         /*
490          * This should be a rare occasion, but this could happen I think in the
491          * case where one thread starts to load the space cache info, and then
492          * some other thread starts a transaction commit which tries to do an
493          * allocation while the other thread is still loading the space cache
494          * info.  The previous loop should have kept us from choosing this block
495          * group, but if we've moved to the state where we will wait on caching
496          * block groups we need to first check if we're doing a fast load here,
497          * so we can wait for it to finish, otherwise we could end up allocating
498          * from a block group who's cache gets evicted for one reason or
499          * another.
500          */
501         while (cache->cached == BTRFS_CACHE_FAST) {
502                 struct btrfs_caching_control *ctl;
503
504                 ctl = cache->caching_ctl;
505                 atomic_inc(&ctl->count);
506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
507                 spin_unlock(&cache->lock);
508
509                 schedule();
510
511                 finish_wait(&ctl->wait, &wait);
512                 put_caching_control(ctl);
513                 spin_lock(&cache->lock);
514         }
515
516         if (cache->cached != BTRFS_CACHE_NO) {
517                 spin_unlock(&cache->lock);
518                 kfree(caching_ctl);
519                 return 0;
520         }
521         WARN_ON(cache->caching_ctl);
522         cache->caching_ctl = caching_ctl;
523         cache->cached = BTRFS_CACHE_FAST;
524         spin_unlock(&cache->lock);
525
526         /*
527          * We can't do the read from on-disk cache during a commit since we need
528          * to have the normal tree locking.  Also if we are currently trying to
529          * allocate blocks for the tree root we can't do the fast caching since
530          * we likely hold important locks.
531          */
532         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
533                 ret = load_free_space_cache(fs_info, cache);
534
535                 spin_lock(&cache->lock);
536                 if (ret == 1) {
537                         cache->caching_ctl = NULL;
538                         cache->cached = BTRFS_CACHE_FINISHED;
539                         cache->last_byte_to_unpin = (u64)-1;
540                 } else {
541                         if (load_cache_only) {
542                                 cache->caching_ctl = NULL;
543                                 cache->cached = BTRFS_CACHE_NO;
544                         } else {
545                                 cache->cached = BTRFS_CACHE_STARTED;
546                         }
547                 }
548                 spin_unlock(&cache->lock);
549                 wake_up(&caching_ctl->wait);
550                 if (ret == 1) {
551                         put_caching_control(caching_ctl);
552                         free_excluded_extents(fs_info->extent_root, cache);
553                         return 0;
554                 }
555         } else {
556                 /*
557                  * We are not going to do the fast caching, set cached to the
558                  * appropriate value and wakeup any waiters.
559                  */
560                 spin_lock(&cache->lock);
561                 if (load_cache_only) {
562                         cache->caching_ctl = NULL;
563                         cache->cached = BTRFS_CACHE_NO;
564                 } else {
565                         cache->cached = BTRFS_CACHE_STARTED;
566                 }
567                 spin_unlock(&cache->lock);
568                 wake_up(&caching_ctl->wait);
569         }
570
571         if (load_cache_only) {
572                 put_caching_control(caching_ctl);
573                 return 0;
574         }
575
576         down_write(&fs_info->extent_commit_sem);
577         atomic_inc(&caching_ctl->count);
578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
579         up_write(&fs_info->extent_commit_sem);
580
581         btrfs_get_block_group(cache);
582
583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
584
585         return ret;
586 }
587
588 /*
589  * return the block group that starts at or after bytenr
590  */
591 static struct btrfs_block_group_cache *
592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
593 {
594         struct btrfs_block_group_cache *cache;
595
596         cache = block_group_cache_tree_search(info, bytenr, 0);
597
598         return cache;
599 }
600
601 /*
602  * return the block group that contains the given bytenr
603  */
604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
605                                                  struct btrfs_fs_info *info,
606                                                  u64 bytenr)
607 {
608         struct btrfs_block_group_cache *cache;
609
610         cache = block_group_cache_tree_search(info, bytenr, 1);
611
612         return cache;
613 }
614
615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
616                                                   u64 flags)
617 {
618         struct list_head *head = &info->space_info;
619         struct btrfs_space_info *found;
620
621         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
622
623         rcu_read_lock();
624         list_for_each_entry_rcu(found, head, list) {
625                 if (found->flags & flags) {
626                         rcu_read_unlock();
627                         return found;
628                 }
629         }
630         rcu_read_unlock();
631         return NULL;
632 }
633
634 /*
635  * after adding space to the filesystem, we need to clear the full flags
636  * on all the space infos.
637  */
638 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
639 {
640         struct list_head *head = &info->space_info;
641         struct btrfs_space_info *found;
642
643         rcu_read_lock();
644         list_for_each_entry_rcu(found, head, list)
645                 found->full = 0;
646         rcu_read_unlock();
647 }
648
649 static u64 div_factor(u64 num, int factor)
650 {
651         if (factor == 10)
652                 return num;
653         num *= factor;
654         do_div(num, 10);
655         return num;
656 }
657
658 static u64 div_factor_fine(u64 num, int factor)
659 {
660         if (factor == 100)
661                 return num;
662         num *= factor;
663         do_div(num, 100);
664         return num;
665 }
666
667 u64 btrfs_find_block_group(struct btrfs_root *root,
668                            u64 search_start, u64 search_hint, int owner)
669 {
670         struct btrfs_block_group_cache *cache;
671         u64 used;
672         u64 last = max(search_hint, search_start);
673         u64 group_start = 0;
674         int full_search = 0;
675         int factor = 9;
676         int wrapped = 0;
677 again:
678         while (1) {
679                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
680                 if (!cache)
681                         break;
682
683                 spin_lock(&cache->lock);
684                 last = cache->key.objectid + cache->key.offset;
685                 used = btrfs_block_group_used(&cache->item);
686
687                 if ((full_search || !cache->ro) &&
688                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
689                         if (used + cache->pinned + cache->reserved <
690                             div_factor(cache->key.offset, factor)) {
691                                 group_start = cache->key.objectid;
692                                 spin_unlock(&cache->lock);
693                                 btrfs_put_block_group(cache);
694                                 goto found;
695                         }
696                 }
697                 spin_unlock(&cache->lock);
698                 btrfs_put_block_group(cache);
699                 cond_resched();
700         }
701         if (!wrapped) {
702                 last = search_start;
703                 wrapped = 1;
704                 goto again;
705         }
706         if (!full_search && factor < 10) {
707                 last = search_start;
708                 full_search = 1;
709                 factor = 10;
710                 goto again;
711         }
712 found:
713         return group_start;
714 }
715
716 /* simple helper to search for an existing extent at a given offset */
717 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
718 {
719         int ret;
720         struct btrfs_key key;
721         struct btrfs_path *path;
722
723         path = btrfs_alloc_path();
724         if (!path)
725                 return -ENOMEM;
726
727         key.objectid = start;
728         key.offset = len;
729         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
730         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
731                                 0, 0);
732         btrfs_free_path(path);
733         return ret;
734 }
735
736 /*
737  * helper function to lookup reference count and flags of extent.
738  *
739  * the head node for delayed ref is used to store the sum of all the
740  * reference count modifications queued up in the rbtree. the head
741  * node may also store the extent flags to set. This way you can check
742  * to see what the reference count and extent flags would be if all of
743  * the delayed refs are not processed.
744  */
745 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
746                              struct btrfs_root *root, u64 bytenr,
747                              u64 num_bytes, u64 *refs, u64 *flags)
748 {
749         struct btrfs_delayed_ref_head *head;
750         struct btrfs_delayed_ref_root *delayed_refs;
751         struct btrfs_path *path;
752         struct btrfs_extent_item *ei;
753         struct extent_buffer *leaf;
754         struct btrfs_key key;
755         u32 item_size;
756         u64 num_refs;
757         u64 extent_flags;
758         int ret;
759
760         path = btrfs_alloc_path();
761         if (!path)
762                 return -ENOMEM;
763
764         key.objectid = bytenr;
765         key.type = BTRFS_EXTENT_ITEM_KEY;
766         key.offset = num_bytes;
767         if (!trans) {
768                 path->skip_locking = 1;
769                 path->search_commit_root = 1;
770         }
771 again:
772         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
773                                 &key, path, 0, 0);
774         if (ret < 0)
775                 goto out_free;
776
777         if (ret == 0) {
778                 leaf = path->nodes[0];
779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
780                 if (item_size >= sizeof(*ei)) {
781                         ei = btrfs_item_ptr(leaf, path->slots[0],
782                                             struct btrfs_extent_item);
783                         num_refs = btrfs_extent_refs(leaf, ei);
784                         extent_flags = btrfs_extent_flags(leaf, ei);
785                 } else {
786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
787                         struct btrfs_extent_item_v0 *ei0;
788                         BUG_ON(item_size != sizeof(*ei0));
789                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
790                                              struct btrfs_extent_item_v0);
791                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
792                         /* FIXME: this isn't correct for data */
793                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
794 #else
795                         BUG();
796 #endif
797                 }
798                 BUG_ON(num_refs == 0);
799         } else {
800                 num_refs = 0;
801                 extent_flags = 0;
802                 ret = 0;
803         }
804
805         if (!trans)
806                 goto out;
807
808         delayed_refs = &trans->transaction->delayed_refs;
809         spin_lock(&delayed_refs->lock);
810         head = btrfs_find_delayed_ref_head(trans, bytenr);
811         if (head) {
812                 if (!mutex_trylock(&head->mutex)) {
813                         atomic_inc(&head->node.refs);
814                         spin_unlock(&delayed_refs->lock);
815
816                         btrfs_release_path(path);
817
818                         /*
819                          * Mutex was contended, block until it's released and try
820                          * again
821                          */
822                         mutex_lock(&head->mutex);
823                         mutex_unlock(&head->mutex);
824                         btrfs_put_delayed_ref(&head->node);
825                         goto again;
826                 }
827                 if (head->extent_op && head->extent_op->update_flags)
828                         extent_flags |= head->extent_op->flags_to_set;
829                 else
830                         BUG_ON(num_refs == 0);
831
832                 num_refs += head->node.ref_mod;
833                 mutex_unlock(&head->mutex);
834         }
835         spin_unlock(&delayed_refs->lock);
836 out:
837         WARN_ON(num_refs == 0);
838         if (refs)
839                 *refs = num_refs;
840         if (flags)
841                 *flags = extent_flags;
842 out_free:
843         btrfs_free_path(path);
844         return ret;
845 }
846
847 /*
848  * Back reference rules.  Back refs have three main goals:
849  *
850  * 1) differentiate between all holders of references to an extent so that
851  *    when a reference is dropped we can make sure it was a valid reference
852  *    before freeing the extent.
853  *
854  * 2) Provide enough information to quickly find the holders of an extent
855  *    if we notice a given block is corrupted or bad.
856  *
857  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
858  *    maintenance.  This is actually the same as #2, but with a slightly
859  *    different use case.
860  *
861  * There are two kinds of back refs. The implicit back refs is optimized
862  * for pointers in non-shared tree blocks. For a given pointer in a block,
863  * back refs of this kind provide information about the block's owner tree
864  * and the pointer's key. These information allow us to find the block by
865  * b-tree searching. The full back refs is for pointers in tree blocks not
866  * referenced by their owner trees. The location of tree block is recorded
867  * in the back refs. Actually the full back refs is generic, and can be
868  * used in all cases the implicit back refs is used. The major shortcoming
869  * of the full back refs is its overhead. Every time a tree block gets
870  * COWed, we have to update back refs entry for all pointers in it.
871  *
872  * For a newly allocated tree block, we use implicit back refs for
873  * pointers in it. This means most tree related operations only involve
874  * implicit back refs. For a tree block created in old transaction, the
875  * only way to drop a reference to it is COW it. So we can detect the
876  * event that tree block loses its owner tree's reference and do the
877  * back refs conversion.
878  *
879  * When a tree block is COW'd through a tree, there are four cases:
880  *
881  * The reference count of the block is one and the tree is the block's
882  * owner tree. Nothing to do in this case.
883  *
884  * The reference count of the block is one and the tree is not the
885  * block's owner tree. In this case, full back refs is used for pointers
886  * in the block. Remove these full back refs, add implicit back refs for
887  * every pointers in the new block.
888  *
889  * The reference count of the block is greater than one and the tree is
890  * the block's owner tree. In this case, implicit back refs is used for
891  * pointers in the block. Add full back refs for every pointers in the
892  * block, increase lower level extents' reference counts. The original
893  * implicit back refs are entailed to the new block.
894  *
895  * The reference count of the block is greater than one and the tree is
896  * not the block's owner tree. Add implicit back refs for every pointer in
897  * the new block, increase lower level extents' reference count.
898  *
899  * Back Reference Key composing:
900  *
901  * The key objectid corresponds to the first byte in the extent,
902  * The key type is used to differentiate between types of back refs.
903  * There are different meanings of the key offset for different types
904  * of back refs.
905  *
906  * File extents can be referenced by:
907  *
908  * - multiple snapshots, subvolumes, or different generations in one subvol
909  * - different files inside a single subvolume
910  * - different offsets inside a file (bookend extents in file.c)
911  *
912  * The extent ref structure for the implicit back refs has fields for:
913  *
914  * - Objectid of the subvolume root
915  * - objectid of the file holding the reference
916  * - original offset in the file
917  * - how many bookend extents
918  *
919  * The key offset for the implicit back refs is hash of the first
920  * three fields.
921  *
922  * The extent ref structure for the full back refs has field for:
923  *
924  * - number of pointers in the tree leaf
925  *
926  * The key offset for the implicit back refs is the first byte of
927  * the tree leaf
928  *
929  * When a file extent is allocated, The implicit back refs is used.
930  * the fields are filled in:
931  *
932  *     (root_key.objectid, inode objectid, offset in file, 1)
933  *
934  * When a file extent is removed file truncation, we find the
935  * corresponding implicit back refs and check the following fields:
936  *
937  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
938  *
939  * Btree extents can be referenced by:
940  *
941  * - Different subvolumes
942  *
943  * Both the implicit back refs and the full back refs for tree blocks
944  * only consist of key. The key offset for the implicit back refs is
945  * objectid of block's owner tree. The key offset for the full back refs
946  * is the first byte of parent block.
947  *
948  * When implicit back refs is used, information about the lowest key and
949  * level of the tree block are required. These information are stored in
950  * tree block info structure.
951  */
952
953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
954 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
955                                   struct btrfs_root *root,
956                                   struct btrfs_path *path,
957                                   u64 owner, u32 extra_size)
958 {
959         struct btrfs_extent_item *item;
960         struct btrfs_extent_item_v0 *ei0;
961         struct btrfs_extent_ref_v0 *ref0;
962         struct btrfs_tree_block_info *bi;
963         struct extent_buffer *leaf;
964         struct btrfs_key key;
965         struct btrfs_key found_key;
966         u32 new_size = sizeof(*item);
967         u64 refs;
968         int ret;
969
970         leaf = path->nodes[0];
971         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
972
973         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
974         ei0 = btrfs_item_ptr(leaf, path->slots[0],
975                              struct btrfs_extent_item_v0);
976         refs = btrfs_extent_refs_v0(leaf, ei0);
977
978         if (owner == (u64)-1) {
979                 while (1) {
980                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
981                                 ret = btrfs_next_leaf(root, path);
982                                 if (ret < 0)
983                                         return ret;
984                                 BUG_ON(ret > 0); /* Corruption */
985                                 leaf = path->nodes[0];
986                         }
987                         btrfs_item_key_to_cpu(leaf, &found_key,
988                                               path->slots[0]);
989                         BUG_ON(key.objectid != found_key.objectid);
990                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
991                                 path->slots[0]++;
992                                 continue;
993                         }
994                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
995                                               struct btrfs_extent_ref_v0);
996                         owner = btrfs_ref_objectid_v0(leaf, ref0);
997                         break;
998                 }
999         }
1000         btrfs_release_path(path);
1001
1002         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1003                 new_size += sizeof(*bi);
1004
1005         new_size -= sizeof(*ei0);
1006         ret = btrfs_search_slot(trans, root, &key, path,
1007                                 new_size + extra_size, 1);
1008         if (ret < 0)
1009                 return ret;
1010         BUG_ON(ret); /* Corruption */
1011
1012         btrfs_extend_item(trans, root, path, new_size);
1013
1014         leaf = path->nodes[0];
1015         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1016         btrfs_set_extent_refs(leaf, item, refs);
1017         /* FIXME: get real generation */
1018         btrfs_set_extent_generation(leaf, item, 0);
1019         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1020                 btrfs_set_extent_flags(leaf, item,
1021                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1022                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1023                 bi = (struct btrfs_tree_block_info *)(item + 1);
1024                 /* FIXME: get first key of the block */
1025                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1026                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1027         } else {
1028                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1029         }
1030         btrfs_mark_buffer_dirty(leaf);
1031         return 0;
1032 }
1033 #endif
1034
1035 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1036 {
1037         u32 high_crc = ~(u32)0;
1038         u32 low_crc = ~(u32)0;
1039         __le64 lenum;
1040
1041         lenum = cpu_to_le64(root_objectid);
1042         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1043         lenum = cpu_to_le64(owner);
1044         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(offset);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047
1048         return ((u64)high_crc << 31) ^ (u64)low_crc;
1049 }
1050
1051 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1052                                      struct btrfs_extent_data_ref *ref)
1053 {
1054         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1055                                     btrfs_extent_data_ref_objectid(leaf, ref),
1056                                     btrfs_extent_data_ref_offset(leaf, ref));
1057 }
1058
1059 static int match_extent_data_ref(struct extent_buffer *leaf,
1060                                  struct btrfs_extent_data_ref *ref,
1061                                  u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1064             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1065             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1066                 return 0;
1067         return 1;
1068 }
1069
1070 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1071                                            struct btrfs_root *root,
1072                                            struct btrfs_path *path,
1073                                            u64 bytenr, u64 parent,
1074                                            u64 root_objectid,
1075                                            u64 owner, u64 offset)
1076 {
1077         struct btrfs_key key;
1078         struct btrfs_extent_data_ref *ref;
1079         struct extent_buffer *leaf;
1080         u32 nritems;
1081         int ret;
1082         int recow;
1083         int err = -ENOENT;
1084
1085         key.objectid = bytenr;
1086         if (parent) {
1087                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1088                 key.offset = parent;
1089         } else {
1090                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1091                 key.offset = hash_extent_data_ref(root_objectid,
1092                                                   owner, offset);
1093         }
1094 again:
1095         recow = 0;
1096         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1097         if (ret < 0) {
1098                 err = ret;
1099                 goto fail;
1100         }
1101
1102         if (parent) {
1103                 if (!ret)
1104                         return 0;
1105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1106                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1107                 btrfs_release_path(path);
1108                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109                 if (ret < 0) {
1110                         err = ret;
1111                         goto fail;
1112                 }
1113                 if (!ret)
1114                         return 0;
1115 #endif
1116                 goto fail;
1117         }
1118
1119         leaf = path->nodes[0];
1120         nritems = btrfs_header_nritems(leaf);
1121         while (1) {
1122                 if (path->slots[0] >= nritems) {
1123                         ret = btrfs_next_leaf(root, path);
1124                         if (ret < 0)
1125                                 err = ret;
1126                         if (ret)
1127                                 goto fail;
1128
1129                         leaf = path->nodes[0];
1130                         nritems = btrfs_header_nritems(leaf);
1131                         recow = 1;
1132                 }
1133
1134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135                 if (key.objectid != bytenr ||
1136                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1137                         goto fail;
1138
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141
1142                 if (match_extent_data_ref(leaf, ref, root_objectid,
1143                                           owner, offset)) {
1144                         if (recow) {
1145                                 btrfs_release_path(path);
1146                                 goto again;
1147                         }
1148                         err = 0;
1149                         break;
1150                 }
1151                 path->slots[0]++;
1152         }
1153 fail:
1154         return err;
1155 }
1156
1157 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1158                                            struct btrfs_root *root,
1159                                            struct btrfs_path *path,
1160                                            u64 bytenr, u64 parent,
1161                                            u64 root_objectid, u64 owner,
1162                                            u64 offset, int refs_to_add)
1163 {
1164         struct btrfs_key key;
1165         struct extent_buffer *leaf;
1166         u32 size;
1167         u32 num_refs;
1168         int ret;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174                 size = sizeof(struct btrfs_shared_data_ref);
1175         } else {
1176                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1177                 key.offset = hash_extent_data_ref(root_objectid,
1178                                                   owner, offset);
1179                 size = sizeof(struct btrfs_extent_data_ref);
1180         }
1181
1182         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1183         if (ret && ret != -EEXIST)
1184                 goto fail;
1185
1186         leaf = path->nodes[0];
1187         if (parent) {
1188                 struct btrfs_shared_data_ref *ref;
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_shared_data_ref);
1191                 if (ret == 0) {
1192                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1193                 } else {
1194                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1195                         num_refs += refs_to_add;
1196                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1197                 }
1198         } else {
1199                 struct btrfs_extent_data_ref *ref;
1200                 while (ret == -EEXIST) {
1201                         ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                              struct btrfs_extent_data_ref);
1203                         if (match_extent_data_ref(leaf, ref, root_objectid,
1204                                                   owner, offset))
1205                                 break;
1206                         btrfs_release_path(path);
1207                         key.offset++;
1208                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1209                                                       size);
1210                         if (ret && ret != -EEXIST)
1211                                 goto fail;
1212
1213                         leaf = path->nodes[0];
1214                 }
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_extent_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_extent_data_ref_root(leaf, ref,
1219                                                        root_objectid);
1220                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1221                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1222                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1223                 } else {
1224                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1225                         num_refs += refs_to_add;
1226                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1227                 }
1228         }
1229         btrfs_mark_buffer_dirty(leaf);
1230         ret = 0;
1231 fail:
1232         btrfs_release_path(path);
1233         return ret;
1234 }
1235
1236 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1237                                            struct btrfs_root *root,
1238                                            struct btrfs_path *path,
1239                                            int refs_to_drop)
1240 {
1241         struct btrfs_key key;
1242         struct btrfs_extent_data_ref *ref1 = NULL;
1243         struct btrfs_shared_data_ref *ref2 = NULL;
1244         struct extent_buffer *leaf;
1245         u32 num_refs = 0;
1246         int ret = 0;
1247
1248         leaf = path->nodes[0];
1249         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1250
1251         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1252                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1253                                       struct btrfs_extent_data_ref);
1254                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1255         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1256                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1257                                       struct btrfs_shared_data_ref);
1258                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1259 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1260         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1261                 struct btrfs_extent_ref_v0 *ref0;
1262                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1263                                       struct btrfs_extent_ref_v0);
1264                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1265 #endif
1266         } else {
1267                 BUG();
1268         }
1269
1270         BUG_ON(num_refs < refs_to_drop);
1271         num_refs -= refs_to_drop;
1272
1273         if (num_refs == 0) {
1274                 ret = btrfs_del_item(trans, root, path);
1275         } else {
1276                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1277                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1278                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1279                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281                 else {
1282                         struct btrfs_extent_ref_v0 *ref0;
1283                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                         struct btrfs_extent_ref_v0);
1285                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1286                 }
1287 #endif
1288                 btrfs_mark_buffer_dirty(leaf);
1289         }
1290         return ret;
1291 }
1292
1293 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           struct btrfs_extent_inline_ref *iref)
1296 {
1297         struct btrfs_key key;
1298         struct extent_buffer *leaf;
1299         struct btrfs_extent_data_ref *ref1;
1300         struct btrfs_shared_data_ref *ref2;
1301         u32 num_refs = 0;
1302
1303         leaf = path->nodes[0];
1304         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1305         if (iref) {
1306                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1307                     BTRFS_EXTENT_DATA_REF_KEY) {
1308                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310                 } else {
1311                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1313                 }
1314         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1315                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1316                                       struct btrfs_extent_data_ref);
1317                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1318         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1319                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_shared_data_ref);
1321                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1322 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1323         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1324                 struct btrfs_extent_ref_v0 *ref0;
1325                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1326                                       struct btrfs_extent_ref_v0);
1327                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1328 #endif
1329         } else {
1330                 WARN_ON(1);
1331         }
1332         return num_refs;
1333 }
1334
1335 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1336                                           struct btrfs_root *root,
1337                                           struct btrfs_path *path,
1338                                           u64 bytenr, u64 parent,
1339                                           u64 root_objectid)
1340 {
1341         struct btrfs_key key;
1342         int ret;
1343
1344         key.objectid = bytenr;
1345         if (parent) {
1346                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1347                 key.offset = parent;
1348         } else {
1349                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1350                 key.offset = root_objectid;
1351         }
1352
1353         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1354         if (ret > 0)
1355                 ret = -ENOENT;
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         if (ret == -ENOENT && parent) {
1358                 btrfs_release_path(path);
1359                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1360                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1361                 if (ret > 0)
1362                         ret = -ENOENT;
1363         }
1364 #endif
1365         return ret;
1366 }
1367
1368 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1369                                           struct btrfs_root *root,
1370                                           struct btrfs_path *path,
1371                                           u64 bytenr, u64 parent,
1372                                           u64 root_objectid)
1373 {
1374         struct btrfs_key key;
1375         int ret;
1376
1377         key.objectid = bytenr;
1378         if (parent) {
1379                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1380                 key.offset = parent;
1381         } else {
1382                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1383                 key.offset = root_objectid;
1384         }
1385
1386         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1387         btrfs_release_path(path);
1388         return ret;
1389 }
1390
1391 static inline int extent_ref_type(u64 parent, u64 owner)
1392 {
1393         int type;
1394         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1395                 if (parent > 0)
1396                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1397                 else
1398                         type = BTRFS_TREE_BLOCK_REF_KEY;
1399         } else {
1400                 if (parent > 0)
1401                         type = BTRFS_SHARED_DATA_REF_KEY;
1402                 else
1403                         type = BTRFS_EXTENT_DATA_REF_KEY;
1404         }
1405         return type;
1406 }
1407
1408 static int find_next_key(struct btrfs_path *path, int level,
1409                          struct btrfs_key *key)
1410
1411 {
1412         for (; level < BTRFS_MAX_LEVEL; level++) {
1413                 if (!path->nodes[level])
1414                         break;
1415                 if (path->slots[level] + 1 >=
1416                     btrfs_header_nritems(path->nodes[level]))
1417                         continue;
1418                 if (level == 0)
1419                         btrfs_item_key_to_cpu(path->nodes[level], key,
1420                                               path->slots[level] + 1);
1421                 else
1422                         btrfs_node_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 return 0;
1425         }
1426         return 1;
1427 }
1428
1429 /*
1430  * look for inline back ref. if back ref is found, *ref_ret is set
1431  * to the address of inline back ref, and 0 is returned.
1432  *
1433  * if back ref isn't found, *ref_ret is set to the address where it
1434  * should be inserted, and -ENOENT is returned.
1435  *
1436  * if insert is true and there are too many inline back refs, the path
1437  * points to the extent item, and -EAGAIN is returned.
1438  *
1439  * NOTE: inline back refs are ordered in the same way that back ref
1440  *       items in the tree are ordered.
1441  */
1442 static noinline_for_stack
1443 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1444                                  struct btrfs_root *root,
1445                                  struct btrfs_path *path,
1446                                  struct btrfs_extent_inline_ref **ref_ret,
1447                                  u64 bytenr, u64 num_bytes,
1448                                  u64 parent, u64 root_objectid,
1449                                  u64 owner, u64 offset, int insert)
1450 {
1451         struct btrfs_key key;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *ei;
1454         struct btrfs_extent_inline_ref *iref;
1455         u64 flags;
1456         u64 item_size;
1457         unsigned long ptr;
1458         unsigned long end;
1459         int extra_size;
1460         int type;
1461         int want;
1462         int ret;
1463         int err = 0;
1464
1465         key.objectid = bytenr;
1466         key.type = BTRFS_EXTENT_ITEM_KEY;
1467         key.offset = num_bytes;
1468
1469         want = extent_ref_type(parent, owner);
1470         if (insert) {
1471                 extra_size = btrfs_extent_inline_ref_size(want);
1472                 path->keep_locks = 1;
1473         } else
1474                 extra_size = -1;
1475         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1476         if (ret < 0) {
1477                 err = ret;
1478                 goto out;
1479         }
1480         if (ret && !insert) {
1481                 err = -ENOENT;
1482                 goto out;
1483         }
1484         BUG_ON(ret); /* Corruption */
1485
1486         leaf = path->nodes[0];
1487         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1489         if (item_size < sizeof(*ei)) {
1490                 if (!insert) {
1491                         err = -ENOENT;
1492                         goto out;
1493                 }
1494                 ret = convert_extent_item_v0(trans, root, path, owner,
1495                                              extra_size);
1496                 if (ret < 0) {
1497                         err = ret;
1498                         goto out;
1499                 }
1500                 leaf = path->nodes[0];
1501                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1502         }
1503 #endif
1504         BUG_ON(item_size < sizeof(*ei));
1505
1506         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1507         flags = btrfs_extent_flags(leaf, ei);
1508
1509         ptr = (unsigned long)(ei + 1);
1510         end = (unsigned long)ei + item_size;
1511
1512         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1513                 ptr += sizeof(struct btrfs_tree_block_info);
1514                 BUG_ON(ptr > end);
1515         } else {
1516                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1517         }
1518
1519         err = -ENOENT;
1520         while (1) {
1521                 if (ptr >= end) {
1522                         WARN_ON(ptr > end);
1523                         break;
1524                 }
1525                 iref = (struct btrfs_extent_inline_ref *)ptr;
1526                 type = btrfs_extent_inline_ref_type(leaf, iref);
1527                 if (want < type)
1528                         break;
1529                 if (want > type) {
1530                         ptr += btrfs_extent_inline_ref_size(type);
1531                         continue;
1532                 }
1533
1534                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1535                         struct btrfs_extent_data_ref *dref;
1536                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1537                         if (match_extent_data_ref(leaf, dref, root_objectid,
1538                                                   owner, offset)) {
1539                                 err = 0;
1540                                 break;
1541                         }
1542                         if (hash_extent_data_ref_item(leaf, dref) <
1543                             hash_extent_data_ref(root_objectid, owner, offset))
1544                                 break;
1545                 } else {
1546                         u64 ref_offset;
1547                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1548                         if (parent > 0) {
1549                                 if (parent == ref_offset) {
1550                                         err = 0;
1551                                         break;
1552                                 }
1553                                 if (ref_offset < parent)
1554                                         break;
1555                         } else {
1556                                 if (root_objectid == ref_offset) {
1557                                         err = 0;
1558                                         break;
1559                                 }
1560                                 if (ref_offset < root_objectid)
1561                                         break;
1562                         }
1563                 }
1564                 ptr += btrfs_extent_inline_ref_size(type);
1565         }
1566         if (err == -ENOENT && insert) {
1567                 if (item_size + extra_size >=
1568                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1569                         err = -EAGAIN;
1570                         goto out;
1571                 }
1572                 /*
1573                  * To add new inline back ref, we have to make sure
1574                  * there is no corresponding back ref item.
1575                  * For simplicity, we just do not add new inline back
1576                  * ref if there is any kind of item for this block
1577                  */
1578                 if (find_next_key(path, 0, &key) == 0 &&
1579                     key.objectid == bytenr &&
1580                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1581                         err = -EAGAIN;
1582                         goto out;
1583                 }
1584         }
1585         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1586 out:
1587         if (insert) {
1588                 path->keep_locks = 0;
1589                 btrfs_unlock_up_safe(path, 1);
1590         }
1591         return err;
1592 }
1593
1594 /*
1595  * helper to add new inline back ref
1596  */
1597 static noinline_for_stack
1598 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1599                                  struct btrfs_root *root,
1600                                  struct btrfs_path *path,
1601                                  struct btrfs_extent_inline_ref *iref,
1602                                  u64 parent, u64 root_objectid,
1603                                  u64 owner, u64 offset, int refs_to_add,
1604                                  struct btrfs_delayed_extent_op *extent_op)
1605 {
1606         struct extent_buffer *leaf;
1607         struct btrfs_extent_item *ei;
1608         unsigned long ptr;
1609         unsigned long end;
1610         unsigned long item_offset;
1611         u64 refs;
1612         int size;
1613         int type;
1614
1615         leaf = path->nodes[0];
1616         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1617         item_offset = (unsigned long)iref - (unsigned long)ei;
1618
1619         type = extent_ref_type(parent, owner);
1620         size = btrfs_extent_inline_ref_size(type);
1621
1622         btrfs_extend_item(trans, root, path, size);
1623
1624         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1625         refs = btrfs_extent_refs(leaf, ei);
1626         refs += refs_to_add;
1627         btrfs_set_extent_refs(leaf, ei, refs);
1628         if (extent_op)
1629                 __run_delayed_extent_op(extent_op, leaf, ei);
1630
1631         ptr = (unsigned long)ei + item_offset;
1632         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1633         if (ptr < end - size)
1634                 memmove_extent_buffer(leaf, ptr + size, ptr,
1635                                       end - size - ptr);
1636
1637         iref = (struct btrfs_extent_inline_ref *)ptr;
1638         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1639         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1640                 struct btrfs_extent_data_ref *dref;
1641                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1642                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1643                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1644                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1645                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1646         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1647                 struct btrfs_shared_data_ref *sref;
1648                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1649                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1655         }
1656         btrfs_mark_buffer_dirty(leaf);
1657 }
1658
1659 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1660                                  struct btrfs_root *root,
1661                                  struct btrfs_path *path,
1662                                  struct btrfs_extent_inline_ref **ref_ret,
1663                                  u64 bytenr, u64 num_bytes, u64 parent,
1664                                  u64 root_objectid, u64 owner, u64 offset)
1665 {
1666         int ret;
1667
1668         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1669                                            bytenr, num_bytes, parent,
1670                                            root_objectid, owner, offset, 0);
1671         if (ret != -ENOENT)
1672                 return ret;
1673
1674         btrfs_release_path(path);
1675         *ref_ret = NULL;
1676
1677         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1678                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1679                                             root_objectid);
1680         } else {
1681                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1682                                              root_objectid, owner, offset);
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * helper to update/remove inline back ref
1689  */
1690 static noinline_for_stack
1691 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1692                                   struct btrfs_root *root,
1693                                   struct btrfs_path *path,
1694                                   struct btrfs_extent_inline_ref *iref,
1695                                   int refs_to_mod,
1696                                   struct btrfs_delayed_extent_op *extent_op)
1697 {
1698         struct extent_buffer *leaf;
1699         struct btrfs_extent_item *ei;
1700         struct btrfs_extent_data_ref *dref = NULL;
1701         struct btrfs_shared_data_ref *sref = NULL;
1702         unsigned long ptr;
1703         unsigned long end;
1704         u32 item_size;
1705         int size;
1706         int type;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751 }
1752
1753 static noinline_for_stack
1754 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1755                                  struct btrfs_root *root,
1756                                  struct btrfs_path *path,
1757                                  u64 bytenr, u64 num_bytes, u64 parent,
1758                                  u64 root_objectid, u64 owner,
1759                                  u64 offset, int refs_to_add,
1760                                  struct btrfs_delayed_extent_op *extent_op)
1761 {
1762         struct btrfs_extent_inline_ref *iref;
1763         int ret;
1764
1765         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1766                                            bytenr, num_bytes, parent,
1767                                            root_objectid, owner, offset, 1);
1768         if (ret == 0) {
1769                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1770                 update_inline_extent_backref(trans, root, path, iref,
1771                                              refs_to_add, extent_op);
1772         } else if (ret == -ENOENT) {
1773                 setup_inline_extent_backref(trans, root, path, iref, parent,
1774                                             root_objectid, owner, offset,
1775                                             refs_to_add, extent_op);
1776                 ret = 0;
1777         }
1778         return ret;
1779 }
1780
1781 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1782                                  struct btrfs_root *root,
1783                                  struct btrfs_path *path,
1784                                  u64 bytenr, u64 parent, u64 root_objectid,
1785                                  u64 owner, u64 offset, int refs_to_add)
1786 {
1787         int ret;
1788         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1789                 BUG_ON(refs_to_add != 1);
1790                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1791                                             parent, root_objectid);
1792         } else {
1793                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1794                                              parent, root_objectid,
1795                                              owner, offset, refs_to_add);
1796         }
1797         return ret;
1798 }
1799
1800 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref *iref,
1804                                  int refs_to_drop, int is_data)
1805 {
1806         int ret = 0;
1807
1808         BUG_ON(!is_data && refs_to_drop != 1);
1809         if (iref) {
1810                 update_inline_extent_backref(trans, root, path, iref,
1811                                              -refs_to_drop, NULL);
1812         } else if (is_data) {
1813                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1814         } else {
1815                 ret = btrfs_del_item(trans, root, path);
1816         }
1817         return ret;
1818 }
1819
1820 static int btrfs_issue_discard(struct block_device *bdev,
1821                                 u64 start, u64 len)
1822 {
1823         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1824 }
1825
1826 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1827                                 u64 num_bytes, u64 *actual_bytes)
1828 {
1829         int ret;
1830         u64 discarded_bytes = 0;
1831         struct btrfs_bio *bbio = NULL;
1832
1833
1834         /* Tell the block device(s) that the sectors can be discarded */
1835         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1836                               bytenr, &num_bytes, &bbio, 0);
1837         /* Error condition is -ENOMEM */
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 /* Can return -ENOMEM */
1873 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1874                          struct btrfs_root *root,
1875                          u64 bytenr, u64 num_bytes, u64 parent,
1876                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1877 {
1878         int ret;
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1882                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1883
1884         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1885                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1886                                         num_bytes,
1887                                         parent, root_objectid, (int)owner,
1888                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1889         } else {
1890                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1891                                         num_bytes,
1892                                         parent, root_objectid, owner, offset,
1893                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1894         }
1895         return ret;
1896 }
1897
1898 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1899                                   struct btrfs_root *root,
1900                                   u64 bytenr, u64 num_bytes,
1901                                   u64 parent, u64 root_objectid,
1902                                   u64 owner, u64 offset, int refs_to_add,
1903                                   struct btrfs_delayed_extent_op *extent_op)
1904 {
1905         struct btrfs_path *path;
1906         struct extent_buffer *leaf;
1907         struct btrfs_extent_item *item;
1908         u64 refs;
1909         int ret;
1910         int err = 0;
1911
1912         path = btrfs_alloc_path();
1913         if (!path)
1914                 return -ENOMEM;
1915
1916         path->reada = 1;
1917         path->leave_spinning = 1;
1918         /* this will setup the path even if it fails to insert the back ref */
1919         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1920                                            path, bytenr, num_bytes, parent,
1921                                            root_objectid, owner, offset,
1922                                            refs_to_add, extent_op);
1923         if (ret == 0)
1924                 goto out;
1925
1926         if (ret != -EAGAIN) {
1927                 err = ret;
1928                 goto out;
1929         }
1930
1931         leaf = path->nodes[0];
1932         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1933         refs = btrfs_extent_refs(leaf, item);
1934         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1935         if (extent_op)
1936                 __run_delayed_extent_op(extent_op, leaf, item);
1937
1938         btrfs_mark_buffer_dirty(leaf);
1939         btrfs_release_path(path);
1940
1941         path->reada = 1;
1942         path->leave_spinning = 1;
1943
1944         /* now insert the actual backref */
1945         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1946                                     path, bytenr, parent, root_objectid,
1947                                     owner, offset, refs_to_add);
1948         if (ret)
1949                 btrfs_abort_transaction(trans, root, ret);
1950 out:
1951         btrfs_free_path(path);
1952         return err;
1953 }
1954
1955 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1956                                 struct btrfs_root *root,
1957                                 struct btrfs_delayed_ref_node *node,
1958                                 struct btrfs_delayed_extent_op *extent_op,
1959                                 int insert_reserved)
1960 {
1961         int ret = 0;
1962         struct btrfs_delayed_data_ref *ref;
1963         struct btrfs_key ins;
1964         u64 parent = 0;
1965         u64 ref_root = 0;
1966         u64 flags = 0;
1967
1968         ins.objectid = node->bytenr;
1969         ins.offset = node->num_bytes;
1970         ins.type = BTRFS_EXTENT_ITEM_KEY;
1971
1972         ref = btrfs_delayed_node_to_data_ref(node);
1973         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1974                 parent = ref->parent;
1975         else
1976                 ref_root = ref->root;
1977
1978         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1979                 if (extent_op) {
1980                         BUG_ON(extent_op->update_key);
1981                         flags |= extent_op->flags_to_set;
1982                 }
1983                 ret = alloc_reserved_file_extent(trans, root,
1984                                                  parent, ref_root, flags,
1985                                                  ref->objectid, ref->offset,
1986                                                  &ins, node->ref_mod);
1987         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1988                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1989                                              node->num_bytes, parent,
1990                                              ref_root, ref->objectid,
1991                                              ref->offset, node->ref_mod,
1992                                              extent_op);
1993         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1994                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1995                                           node->num_bytes, parent,
1996                                           ref_root, ref->objectid,
1997                                           ref->offset, node->ref_mod,
1998                                           extent_op);
1999         } else {
2000                 BUG();
2001         }
2002         return ret;
2003 }
2004
2005 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2006                                     struct extent_buffer *leaf,
2007                                     struct btrfs_extent_item *ei)
2008 {
2009         u64 flags = btrfs_extent_flags(leaf, ei);
2010         if (extent_op->update_flags) {
2011                 flags |= extent_op->flags_to_set;
2012                 btrfs_set_extent_flags(leaf, ei, flags);
2013         }
2014
2015         if (extent_op->update_key) {
2016                 struct btrfs_tree_block_info *bi;
2017                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2018                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2019                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2020         }
2021 }
2022
2023 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2024                                  struct btrfs_root *root,
2025                                  struct btrfs_delayed_ref_node *node,
2026                                  struct btrfs_delayed_extent_op *extent_op)
2027 {
2028         struct btrfs_key key;
2029         struct btrfs_path *path;
2030         struct btrfs_extent_item *ei;
2031         struct extent_buffer *leaf;
2032         u32 item_size;
2033         int ret;
2034         int err = 0;
2035
2036         if (trans->aborted)
2037                 return 0;
2038
2039         path = btrfs_alloc_path();
2040         if (!path)
2041                 return -ENOMEM;
2042
2043         key.objectid = node->bytenr;
2044         key.type = BTRFS_EXTENT_ITEM_KEY;
2045         key.offset = node->num_bytes;
2046
2047         path->reada = 1;
2048         path->leave_spinning = 1;
2049         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2050                                 path, 0, 1);
2051         if (ret < 0) {
2052                 err = ret;
2053                 goto out;
2054         }
2055         if (ret > 0) {
2056                 err = -EIO;
2057                 goto out;
2058         }
2059
2060         leaf = path->nodes[0];
2061         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2062 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2063         if (item_size < sizeof(*ei)) {
2064                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2065                                              path, (u64)-1, 0);
2066                 if (ret < 0) {
2067                         err = ret;
2068                         goto out;
2069                 }
2070                 leaf = path->nodes[0];
2071                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2072         }
2073 #endif
2074         BUG_ON(item_size < sizeof(*ei));
2075         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2076         __run_delayed_extent_op(extent_op, leaf, ei);
2077
2078         btrfs_mark_buffer_dirty(leaf);
2079 out:
2080         btrfs_free_path(path);
2081         return err;
2082 }
2083
2084 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2085                                 struct btrfs_root *root,
2086                                 struct btrfs_delayed_ref_node *node,
2087                                 struct btrfs_delayed_extent_op *extent_op,
2088                                 int insert_reserved)
2089 {
2090         int ret = 0;
2091         struct btrfs_delayed_tree_ref *ref;
2092         struct btrfs_key ins;
2093         u64 parent = 0;
2094         u64 ref_root = 0;
2095
2096         ins.objectid = node->bytenr;
2097         ins.offset = node->num_bytes;
2098         ins.type = BTRFS_EXTENT_ITEM_KEY;
2099
2100         ref = btrfs_delayed_node_to_tree_ref(node);
2101         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2102                 parent = ref->parent;
2103         else
2104                 ref_root = ref->root;
2105
2106         BUG_ON(node->ref_mod != 1);
2107         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2108                 BUG_ON(!extent_op || !extent_op->update_flags ||
2109                        !extent_op->update_key);
2110                 ret = alloc_reserved_tree_block(trans, root,
2111                                                 parent, ref_root,
2112                                                 extent_op->flags_to_set,
2113                                                 &extent_op->key,
2114                                                 ref->level, &ins);
2115         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2116                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2117                                              node->num_bytes, parent, ref_root,
2118                                              ref->level, 0, 1, extent_op);
2119         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2120                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2121                                           node->num_bytes, parent, ref_root,
2122                                           ref->level, 0, 1, extent_op);
2123         } else {
2124                 BUG();
2125         }
2126         return ret;
2127 }
2128
2129 /* helper function to actually process a single delayed ref entry */
2130 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2131                                struct btrfs_root *root,
2132                                struct btrfs_delayed_ref_node *node,
2133                                struct btrfs_delayed_extent_op *extent_op,
2134                                int insert_reserved)
2135 {
2136         int ret = 0;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (btrfs_delayed_ref_is_head(node)) {
2142                 struct btrfs_delayed_ref_head *head;
2143                 /*
2144                  * we've hit the end of the chain and we were supposed
2145                  * to insert this extent into the tree.  But, it got
2146                  * deleted before we ever needed to insert it, so all
2147                  * we have to do is clean up the accounting
2148                  */
2149                 BUG_ON(extent_op);
2150                 head = btrfs_delayed_node_to_head(node);
2151                 if (insert_reserved) {
2152                         btrfs_pin_extent(root, node->bytenr,
2153                                          node->num_bytes, 1);
2154                         if (head->is_data) {
2155                                 ret = btrfs_del_csums(trans, root,
2156                                                       node->bytenr,
2157                                                       node->num_bytes);
2158                         }
2159                 }
2160                 mutex_unlock(&head->mutex);
2161                 return ret;
2162         }
2163
2164         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2165             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2166                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2167                                            insert_reserved);
2168         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2169                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2170                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2171                                            insert_reserved);
2172         else
2173                 BUG();
2174         return ret;
2175 }
2176
2177 static noinline struct btrfs_delayed_ref_node *
2178 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2179 {
2180         struct rb_node *node;
2181         struct btrfs_delayed_ref_node *ref;
2182         int action = BTRFS_ADD_DELAYED_REF;
2183 again:
2184         /*
2185          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2186          * this prevents ref count from going down to zero when
2187          * there still are pending delayed ref.
2188          */
2189         node = rb_prev(&head->node.rb_node);
2190         while (1) {
2191                 if (!node)
2192                         break;
2193                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2194                                 rb_node);
2195                 if (ref->bytenr != head->node.bytenr)
2196                         break;
2197                 if (ref->action == action)
2198                         return ref;
2199                 node = rb_prev(node);
2200         }
2201         if (action == BTRFS_ADD_DELAYED_REF) {
2202                 action = BTRFS_DROP_DELAYED_REF;
2203                 goto again;
2204         }
2205         return NULL;
2206 }
2207
2208 /*
2209  * Returns 0 on success or if called with an already aborted transaction.
2210  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2211  */
2212 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2213                                        struct btrfs_root *root,
2214                                        struct list_head *cluster)
2215 {
2216         struct btrfs_delayed_ref_root *delayed_refs;
2217         struct btrfs_delayed_ref_node *ref;
2218         struct btrfs_delayed_ref_head *locked_ref = NULL;
2219         struct btrfs_delayed_extent_op *extent_op;
2220         int ret;
2221         int count = 0;
2222         int must_insert_reserved = 0;
2223
2224         delayed_refs = &trans->transaction->delayed_refs;
2225         while (1) {
2226                 if (!locked_ref) {
2227                         /* pick a new head ref from the cluster list */
2228                         if (list_empty(cluster))
2229                                 break;
2230
2231                         locked_ref = list_entry(cluster->next,
2232                                      struct btrfs_delayed_ref_head, cluster);
2233
2234                         /* grab the lock that says we are going to process
2235                          * all the refs for this head */
2236                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2237
2238                         /*
2239                          * we may have dropped the spin lock to get the head
2240                          * mutex lock, and that might have given someone else
2241                          * time to free the head.  If that's true, it has been
2242                          * removed from our list and we can move on.
2243                          */
2244                         if (ret == -EAGAIN) {
2245                                 locked_ref = NULL;
2246                                 count++;
2247                                 continue;
2248                         }
2249                 }
2250
2251                 /*
2252                  * locked_ref is the head node, so we have to go one
2253                  * node back for any delayed ref updates
2254                  */
2255                 ref = select_delayed_ref(locked_ref);
2256
2257                 if (ref && ref->seq &&
2258                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2259                         /*
2260                          * there are still refs with lower seq numbers in the
2261                          * process of being added. Don't run this ref yet.
2262                          */
2263                         list_del_init(&locked_ref->cluster);
2264                         mutex_unlock(&locked_ref->mutex);
2265                         locked_ref = NULL;
2266                         delayed_refs->num_heads_ready++;
2267                         spin_unlock(&delayed_refs->lock);
2268                         cond_resched();
2269                         spin_lock(&delayed_refs->lock);
2270                         continue;
2271                 }
2272
2273                 /*
2274                  * record the must insert reserved flag before we
2275                  * drop the spin lock.
2276                  */
2277                 must_insert_reserved = locked_ref->must_insert_reserved;
2278                 locked_ref->must_insert_reserved = 0;
2279
2280                 extent_op = locked_ref->extent_op;
2281                 locked_ref->extent_op = NULL;
2282
2283                 if (!ref) {
2284                         /* All delayed refs have been processed, Go ahead
2285                          * and send the head node to run_one_delayed_ref,
2286                          * so that any accounting fixes can happen
2287                          */
2288                         ref = &locked_ref->node;
2289
2290                         if (extent_op && must_insert_reserved) {
2291                                 kfree(extent_op);
2292                                 extent_op = NULL;
2293                         }
2294
2295                         if (extent_op) {
2296                                 spin_unlock(&delayed_refs->lock);
2297
2298                                 ret = run_delayed_extent_op(trans, root,
2299                                                             ref, extent_op);
2300                                 kfree(extent_op);
2301
2302                                 if (ret) {
2303                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2304                                         spin_lock(&delayed_refs->lock);
2305                                         return ret;
2306                                 }
2307
2308                                 goto next;
2309                         }
2310
2311                         list_del_init(&locked_ref->cluster);
2312                         locked_ref = NULL;
2313                 }
2314
2315                 ref->in_tree = 0;
2316                 rb_erase(&ref->rb_node, &delayed_refs->root);
2317                 delayed_refs->num_entries--;
2318                 /*
2319                  * we modified num_entries, but as we're currently running
2320                  * delayed refs, skip
2321                  *     wake_up(&delayed_refs->seq_wait);
2322                  * here.
2323                  */
2324                 spin_unlock(&delayed_refs->lock);
2325
2326                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2327                                           must_insert_reserved);
2328
2329                 btrfs_put_delayed_ref(ref);
2330                 kfree(extent_op);
2331                 count++;
2332
2333                 if (ret) {
2334                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2335                         spin_lock(&delayed_refs->lock);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2351                                unsigned long num_refs,
2352                                struct list_head *first_seq)
2353 {
2354         spin_unlock(&delayed_refs->lock);
2355         pr_debug("waiting for more refs (num %ld, first %p)\n",
2356                  num_refs, first_seq);
2357         wait_event(delayed_refs->seq_wait,
2358                    num_refs != delayed_refs->num_entries ||
2359                    delayed_refs->seq_head.next != first_seq);
2360         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2361                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2362         spin_lock(&delayed_refs->lock);
2363 }
2364
2365 /*
2366  * this starts processing the delayed reference count updates and
2367  * extent insertions we have queued up so far.  count can be
2368  * 0, which means to process everything in the tree at the start
2369  * of the run (but not newly added entries), or it can be some target
2370  * number you'd like to process.
2371  *
2372  * Returns 0 on success or if called with an aborted transaction
2373  * Returns <0 on error and aborts the transaction
2374  */
2375 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2376                            struct btrfs_root *root, unsigned long count)
2377 {
2378         struct rb_node *node;
2379         struct btrfs_delayed_ref_root *delayed_refs;
2380         struct btrfs_delayed_ref_node *ref;
2381         struct list_head cluster;
2382         struct list_head *first_seq = NULL;
2383         int ret;
2384         u64 delayed_start;
2385         int run_all = count == (unsigned long)-1;
2386         int run_most = 0;
2387         unsigned long num_refs = 0;
2388         int consider_waiting;
2389
2390         /* We'll clean this up in btrfs_cleanup_transaction */
2391         if (trans->aborted)
2392                 return 0;
2393
2394         if (root == root->fs_info->extent_root)
2395                 root = root->fs_info->tree_root;
2396
2397         do_chunk_alloc(trans, root->fs_info->extent_root,
2398                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2399                        CHUNK_ALLOC_NO_FORCE);
2400
2401         delayed_refs = &trans->transaction->delayed_refs;
2402         INIT_LIST_HEAD(&cluster);
2403 again:
2404         consider_waiting = 0;
2405         spin_lock(&delayed_refs->lock);
2406         if (count == 0) {
2407                 count = delayed_refs->num_entries * 2;
2408                 run_most = 1;
2409         }
2410         while (1) {
2411                 if (!(run_all || run_most) &&
2412                     delayed_refs->num_heads_ready < 64)
2413                         break;
2414
2415                 /*
2416                  * go find something we can process in the rbtree.  We start at
2417                  * the beginning of the tree, and then build a cluster
2418                  * of refs to process starting at the first one we are able to
2419                  * lock
2420                  */
2421                 delayed_start = delayed_refs->run_delayed_start;
2422                 ret = btrfs_find_ref_cluster(trans, &cluster,
2423                                              delayed_refs->run_delayed_start);
2424                 if (ret)
2425                         break;
2426
2427                 if (delayed_start >= delayed_refs->run_delayed_start) {
2428                         if (consider_waiting == 0) {
2429                                 /*
2430                                  * btrfs_find_ref_cluster looped. let's do one
2431                                  * more cycle. if we don't run any delayed ref
2432                                  * during that cycle (because we can't because
2433                                  * all of them are blocked) and if the number of
2434                                  * refs doesn't change, we avoid busy waiting.
2435                                  */
2436                                 consider_waiting = 1;
2437                                 num_refs = delayed_refs->num_entries;
2438                                 first_seq = root->fs_info->tree_mod_seq_list.next;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs,
2441                                                    num_refs, first_seq);
2442                                 /*
2443                                  * after waiting, things have changed. we
2444                                  * dropped the lock and someone else might have
2445                                  * run some refs, built new clusters and so on.
2446                                  * therefore, we restart staleness detection.
2447                                  */
2448                                 consider_waiting = 0;
2449                         }
2450                 }
2451
2452                 ret = run_clustered_refs(trans, root, &cluster);
2453                 if (ret < 0) {
2454                         spin_unlock(&delayed_refs->lock);
2455                         btrfs_abort_transaction(trans, root, ret);
2456                         return ret;
2457                 }
2458
2459                 count -= min_t(unsigned long, ret, count);
2460
2461                 if (count == 0)
2462                         break;
2463
2464                 if (ret || delayed_refs->run_delayed_start == 0) {
2465                         /* refs were run, let's reset staleness detection */
2466                         consider_waiting = 0;
2467                 }
2468         }
2469
2470         if (run_all) {
2471                 node = rb_first(&delayed_refs->root);
2472                 if (!node)
2473                         goto out;
2474                 count = (unsigned long)-1;
2475
2476                 while (node) {
2477                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2478                                        rb_node);
2479                         if (btrfs_delayed_ref_is_head(ref)) {
2480                                 struct btrfs_delayed_ref_head *head;
2481
2482                                 head = btrfs_delayed_node_to_head(ref);
2483                                 atomic_inc(&ref->refs);
2484
2485                                 spin_unlock(&delayed_refs->lock);
2486                                 /*
2487                                  * Mutex was contended, block until it's
2488                                  * released and try again
2489                                  */
2490                                 mutex_lock(&head->mutex);
2491                                 mutex_unlock(&head->mutex);
2492
2493                                 btrfs_put_delayed_ref(ref);
2494                                 cond_resched();
2495                                 goto again;
2496                         }
2497                         node = rb_next(node);
2498                 }
2499                 spin_unlock(&delayed_refs->lock);
2500                 schedule_timeout(1);
2501                 goto again;
2502         }
2503 out:
2504         spin_unlock(&delayed_refs->lock);
2505         return 0;
2506 }
2507
2508 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2509                                 struct btrfs_root *root,
2510                                 u64 bytenr, u64 num_bytes, u64 flags,
2511                                 int is_data)
2512 {
2513         struct btrfs_delayed_extent_op *extent_op;
2514         int ret;
2515
2516         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2517         if (!extent_op)
2518                 return -ENOMEM;
2519
2520         extent_op->flags_to_set = flags;
2521         extent_op->update_flags = 1;
2522         extent_op->update_key = 0;
2523         extent_op->is_data = is_data ? 1 : 0;
2524
2525         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2526                                           num_bytes, extent_op);
2527         if (ret)
2528                 kfree(extent_op);
2529         return ret;
2530 }
2531
2532 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2533                                       struct btrfs_root *root,
2534                                       struct btrfs_path *path,
2535                                       u64 objectid, u64 offset, u64 bytenr)
2536 {
2537         struct btrfs_delayed_ref_head *head;
2538         struct btrfs_delayed_ref_node *ref;
2539         struct btrfs_delayed_data_ref *data_ref;
2540         struct btrfs_delayed_ref_root *delayed_refs;
2541         struct rb_node *node;
2542         int ret = 0;
2543
2544         ret = -ENOENT;
2545         delayed_refs = &trans->transaction->delayed_refs;
2546         spin_lock(&delayed_refs->lock);
2547         head = btrfs_find_delayed_ref_head(trans, bytenr);
2548         if (!head)
2549                 goto out;
2550
2551         if (!mutex_trylock(&head->mutex)) {
2552                 atomic_inc(&head->node.refs);
2553                 spin_unlock(&delayed_refs->lock);
2554
2555                 btrfs_release_path(path);
2556
2557                 /*
2558                  * Mutex was contended, block until it's released and let
2559                  * caller try again
2560                  */
2561                 mutex_lock(&head->mutex);
2562                 mutex_unlock(&head->mutex);
2563                 btrfs_put_delayed_ref(&head->node);
2564                 return -EAGAIN;
2565         }
2566
2567         node = rb_prev(&head->node.rb_node);
2568         if (!node)
2569                 goto out_unlock;
2570
2571         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2572
2573         if (ref->bytenr != bytenr)
2574                 goto out_unlock;
2575
2576         ret = 1;
2577         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2578                 goto out_unlock;
2579
2580         data_ref = btrfs_delayed_node_to_data_ref(ref);
2581
2582         node = rb_prev(node);
2583         if (node) {
2584                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2585                 if (ref->bytenr == bytenr)
2586                         goto out_unlock;
2587         }
2588
2589         if (data_ref->root != root->root_key.objectid ||
2590             data_ref->objectid != objectid || data_ref->offset != offset)
2591                 goto out_unlock;
2592
2593         ret = 0;
2594 out_unlock:
2595         mutex_unlock(&head->mutex);
2596 out:
2597         spin_unlock(&delayed_refs->lock);
2598         return ret;
2599 }
2600
2601 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2602                                         struct btrfs_root *root,
2603                                         struct btrfs_path *path,
2604                                         u64 objectid, u64 offset, u64 bytenr)
2605 {
2606         struct btrfs_root *extent_root = root->fs_info->extent_root;
2607         struct extent_buffer *leaf;
2608         struct btrfs_extent_data_ref *ref;
2609         struct btrfs_extent_inline_ref *iref;
2610         struct btrfs_extent_item *ei;
2611         struct btrfs_key key;
2612         u32 item_size;
2613         int ret;
2614
2615         key.objectid = bytenr;
2616         key.offset = (u64)-1;
2617         key.type = BTRFS_EXTENT_ITEM_KEY;
2618
2619         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2620         if (ret < 0)
2621                 goto out;
2622         BUG_ON(ret == 0); /* Corruption */
2623
2624         ret = -ENOENT;
2625         if (path->slots[0] == 0)
2626                 goto out;
2627
2628         path->slots[0]--;
2629         leaf = path->nodes[0];
2630         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2631
2632         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2633                 goto out;
2634
2635         ret = 1;
2636         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2637 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2638         if (item_size < sizeof(*ei)) {
2639                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2640                 goto out;
2641         }
2642 #endif
2643         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2644
2645         if (item_size != sizeof(*ei) +
2646             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2647                 goto out;
2648
2649         if (btrfs_extent_generation(leaf, ei) <=
2650             btrfs_root_last_snapshot(&root->root_item))
2651                 goto out;
2652
2653         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2654         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2655             BTRFS_EXTENT_DATA_REF_KEY)
2656                 goto out;
2657
2658         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2659         if (btrfs_extent_refs(leaf, ei) !=
2660             btrfs_extent_data_ref_count(leaf, ref) ||
2661             btrfs_extent_data_ref_root(leaf, ref) !=
2662             root->root_key.objectid ||
2663             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2664             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2665                 goto out;
2666
2667         ret = 0;
2668 out:
2669         return ret;
2670 }
2671
2672 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2673                           struct btrfs_root *root,
2674                           u64 objectid, u64 offset, u64 bytenr)
2675 {
2676         struct btrfs_path *path;
2677         int ret;
2678         int ret2;
2679
2680         path = btrfs_alloc_path();
2681         if (!path)
2682                 return -ENOENT;
2683
2684         do {
2685                 ret = check_committed_ref(trans, root, path, objectid,
2686                                           offset, bytenr);
2687                 if (ret && ret != -ENOENT)
2688                         goto out;
2689
2690                 ret2 = check_delayed_ref(trans, root, path, objectid,
2691                                          offset, bytenr);
2692         } while (ret2 == -EAGAIN);
2693
2694         if (ret2 && ret2 != -ENOENT) {
2695                 ret = ret2;
2696                 goto out;
2697         }
2698
2699         if (ret != -ENOENT || ret2 != -ENOENT)
2700                 ret = 0;
2701 out:
2702         btrfs_free_path(path);
2703         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2704                 WARN_ON(ret > 0);
2705         return ret;
2706 }
2707
2708 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2709                            struct btrfs_root *root,
2710                            struct extent_buffer *buf,
2711                            int full_backref, int inc, int for_cow)
2712 {
2713         u64 bytenr;
2714         u64 num_bytes;
2715         u64 parent;
2716         u64 ref_root;
2717         u32 nritems;
2718         struct btrfs_key key;
2719         struct btrfs_file_extent_item *fi;
2720         int i;
2721         int level;
2722         int ret = 0;
2723         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2724                             u64, u64, u64, u64, u64, u64, int);
2725
2726         ref_root = btrfs_header_owner(buf);
2727         nritems = btrfs_header_nritems(buf);
2728         level = btrfs_header_level(buf);
2729
2730         if (!root->ref_cows && level == 0)
2731                 return 0;
2732
2733         if (inc)
2734                 process_func = btrfs_inc_extent_ref;
2735         else
2736                 process_func = btrfs_free_extent;
2737
2738         if (full_backref)
2739                 parent = buf->start;
2740         else
2741                 parent = 0;
2742
2743         for (i = 0; i < nritems; i++) {
2744                 if (level == 0) {
2745                         btrfs_item_key_to_cpu(buf, &key, i);
2746                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2747                                 continue;
2748                         fi = btrfs_item_ptr(buf, i,
2749                                             struct btrfs_file_extent_item);
2750                         if (btrfs_file_extent_type(buf, fi) ==
2751                             BTRFS_FILE_EXTENT_INLINE)
2752                                 continue;
2753                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2754                         if (bytenr == 0)
2755                                 continue;
2756
2757                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2758                         key.offset -= btrfs_file_extent_offset(buf, fi);
2759                         ret = process_func(trans, root, bytenr, num_bytes,
2760                                            parent, ref_root, key.objectid,
2761                                            key.offset, for_cow);
2762                         if (ret)
2763                                 goto fail;
2764                 } else {
2765                         bytenr = btrfs_node_blockptr(buf, i);
2766                         num_bytes = btrfs_level_size(root, level - 1);
2767                         ret = process_func(trans, root, bytenr, num_bytes,
2768                                            parent, ref_root, level - 1, 0,
2769                                            for_cow);
2770                         if (ret)
2771                                 goto fail;
2772                 }
2773         }
2774         return 0;
2775 fail:
2776         return ret;
2777 }
2778
2779 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2780                   struct extent_buffer *buf, int full_backref, int for_cow)
2781 {
2782         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2783 }
2784
2785 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2786                   struct extent_buffer *buf, int full_backref, int for_cow)
2787 {
2788         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2789 }
2790
2791 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2792                                  struct btrfs_root *root,
2793                                  struct btrfs_path *path,
2794                                  struct btrfs_block_group_cache *cache)
2795 {
2796         int ret;
2797         struct btrfs_root *extent_root = root->fs_info->extent_root;
2798         unsigned long bi;
2799         struct extent_buffer *leaf;
2800
2801         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2802         if (ret < 0)
2803                 goto fail;
2804         BUG_ON(ret); /* Corruption */
2805
2806         leaf = path->nodes[0];
2807         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2808         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2809         btrfs_mark_buffer_dirty(leaf);
2810         btrfs_release_path(path);
2811 fail:
2812         if (ret) {
2813                 btrfs_abort_transaction(trans, root, ret);
2814                 return ret;
2815         }
2816         return 0;
2817
2818 }
2819
2820 static struct btrfs_block_group_cache *
2821 next_block_group(struct btrfs_root *root,
2822                  struct btrfs_block_group_cache *cache)
2823 {
2824         struct rb_node *node;
2825         spin_lock(&root->fs_info->block_group_cache_lock);
2826         node = rb_next(&cache->cache_node);
2827         btrfs_put_block_group(cache);
2828         if (node) {
2829                 cache = rb_entry(node, struct btrfs_block_group_cache,
2830                                  cache_node);
2831                 btrfs_get_block_group(cache);
2832         } else
2833                 cache = NULL;
2834         spin_unlock(&root->fs_info->block_group_cache_lock);
2835         return cache;
2836 }
2837
2838 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2839                             struct btrfs_trans_handle *trans,
2840                             struct btrfs_path *path)
2841 {
2842         struct btrfs_root *root = block_group->fs_info->tree_root;
2843         struct inode *inode = NULL;
2844         u64 alloc_hint = 0;
2845         int dcs = BTRFS_DC_ERROR;
2846         int num_pages = 0;
2847         int retries = 0;
2848         int ret = 0;
2849
2850         /*
2851          * If this block group is smaller than 100 megs don't bother caching the
2852          * block group.
2853          */
2854         if (block_group->key.offset < (100 * 1024 * 1024)) {
2855                 spin_lock(&block_group->lock);
2856                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2857                 spin_unlock(&block_group->lock);
2858                 return 0;
2859         }
2860
2861 again:
2862         inode = lookup_free_space_inode(root, block_group, path);
2863         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2864                 ret = PTR_ERR(inode);
2865                 btrfs_release_path(path);
2866                 goto out;
2867         }
2868
2869         if (IS_ERR(inode)) {
2870                 BUG_ON(retries);
2871                 retries++;
2872
2873                 if (block_group->ro)
2874                         goto out_free;
2875
2876                 ret = create_free_space_inode(root, trans, block_group, path);
2877                 if (ret)
2878                         goto out_free;
2879                 goto again;
2880         }
2881
2882         /* We've already setup this transaction, go ahead and exit */
2883         if (block_group->cache_generation == trans->transid &&
2884             i_size_read(inode)) {
2885                 dcs = BTRFS_DC_SETUP;
2886                 goto out_put;
2887         }
2888
2889         /*
2890          * We want to set the generation to 0, that way if anything goes wrong
2891          * from here on out we know not to trust this cache when we load up next
2892          * time.
2893          */
2894         BTRFS_I(inode)->generation = 0;
2895         ret = btrfs_update_inode(trans, root, inode);
2896         WARN_ON(ret);
2897
2898         if (i_size_read(inode) > 0) {
2899                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2900                                                       inode);
2901                 if (ret)
2902                         goto out_put;
2903         }
2904
2905         spin_lock(&block_group->lock);
2906         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2907             !btrfs_test_opt(root, SPACE_CACHE)) {
2908                 /*
2909                  * don't bother trying to write stuff out _if_
2910                  * a) we're not cached,
2911                  * b) we're with nospace_cache mount option.
2912                  */
2913                 dcs = BTRFS_DC_WRITTEN;
2914                 spin_unlock(&block_group->lock);
2915                 goto out_put;
2916         }
2917         spin_unlock(&block_group->lock);
2918
2919         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2920         if (!num_pages)
2921                 num_pages = 1;
2922
2923         /*
2924          * Just to make absolutely sure we have enough space, we're going to
2925          * preallocate 12 pages worth of space for each block group.  In
2926          * practice we ought to use at most 8, but we need extra space so we can
2927          * add our header and have a terminator between the extents and the
2928          * bitmaps.
2929          */
2930         num_pages *= 16;
2931         num_pages *= PAGE_CACHE_SIZE;
2932
2933         ret = btrfs_check_data_free_space(inode, num_pages);
2934         if (ret)
2935                 goto out_put;
2936
2937         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2938                                               num_pages, num_pages,
2939                                               &alloc_hint);
2940         if (!ret)
2941                 dcs = BTRFS_DC_SETUP;
2942         btrfs_free_reserved_data_space(inode, num_pages);
2943
2944 out_put:
2945         iput(inode);
2946 out_free:
2947         btrfs_release_path(path);
2948 out:
2949         spin_lock(&block_group->lock);
2950         if (!ret && dcs == BTRFS_DC_SETUP)
2951                 block_group->cache_generation = trans->transid;
2952         block_group->disk_cache_state = dcs;
2953         spin_unlock(&block_group->lock);
2954
2955         return ret;
2956 }
2957
2958 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2959                                    struct btrfs_root *root)
2960 {
2961         struct btrfs_block_group_cache *cache;
2962         int err = 0;
2963         struct btrfs_path *path;
2964         u64 last = 0;
2965
2966         path = btrfs_alloc_path();
2967         if (!path)
2968                 return -ENOMEM;
2969
2970 again:
2971         while (1) {
2972                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2973                 while (cache) {
2974                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2975                                 break;
2976                         cache = next_block_group(root, cache);
2977                 }
2978                 if (!cache) {
2979                         if (last == 0)
2980                                 break;
2981                         last = 0;
2982                         continue;
2983                 }
2984                 err = cache_save_setup(cache, trans, path);
2985                 last = cache->key.objectid + cache->key.offset;
2986                 btrfs_put_block_group(cache);
2987         }
2988
2989         while (1) {
2990                 if (last == 0) {
2991                         err = btrfs_run_delayed_refs(trans, root,
2992                                                      (unsigned long)-1);
2993                         if (err) /* File system offline */
2994                                 goto out;
2995                 }
2996
2997                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2998                 while (cache) {
2999                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3000                                 btrfs_put_block_group(cache);
3001                                 goto again;
3002                         }
3003
3004                         if (cache->dirty)
3005                                 break;
3006                         cache = next_block_group(root, cache);
3007                 }
3008                 if (!cache) {
3009                         if (last == 0)
3010                                 break;
3011                         last = 0;
3012                         continue;
3013                 }
3014
3015                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3016                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3017                 cache->dirty = 0;
3018                 last = cache->key.objectid + cache->key.offset;
3019
3020                 err = write_one_cache_group(trans, root, path, cache);
3021                 if (err) /* File system offline */
3022                         goto out;
3023
3024                 btrfs_put_block_group(cache);
3025         }
3026
3027         while (1) {
3028                 /*
3029                  * I don't think this is needed since we're just marking our
3030                  * preallocated extent as written, but just in case it can't
3031                  * hurt.
3032                  */
3033                 if (last == 0) {
3034                         err = btrfs_run_delayed_refs(trans, root,
3035                                                      (unsigned long)-1);
3036                         if (err) /* File system offline */
3037                                 goto out;
3038                 }
3039
3040                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3041                 while (cache) {
3042                         /*
3043                          * Really this shouldn't happen, but it could if we
3044                          * couldn't write the entire preallocated extent and
3045                          * splitting the extent resulted in a new block.
3046                          */
3047                         if (cache->dirty) {
3048                                 btrfs_put_block_group(cache);
3049                                 goto again;
3050                         }
3051                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3052                                 break;
3053                         cache = next_block_group(root, cache);
3054                 }
3055                 if (!cache) {
3056                         if (last == 0)
3057                                 break;
3058                         last = 0;
3059                         continue;
3060                 }
3061
3062                 err = btrfs_write_out_cache(root, trans, cache, path);
3063
3064                 /*
3065                  * If we didn't have an error then the cache state is still
3066                  * NEED_WRITE, so we can set it to WRITTEN.
3067                  */
3068                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3069                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3070                 last = cache->key.objectid + cache->key.offset;
3071                 btrfs_put_block_group(cache);
3072         }
3073 out:
3074
3075         btrfs_free_path(path);
3076         return err;
3077 }
3078
3079 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3080 {
3081         struct btrfs_block_group_cache *block_group;
3082         int readonly = 0;
3083
3084         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3085         if (!block_group || block_group->ro)
3086                 readonly = 1;
3087         if (block_group)
3088                 btrfs_put_block_group(block_group);
3089         return readonly;
3090 }
3091
3092 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3093                              u64 total_bytes, u64 bytes_used,
3094                              struct btrfs_space_info **space_info)
3095 {
3096         struct btrfs_space_info *found;
3097         int i;
3098         int factor;
3099
3100         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3101                      BTRFS_BLOCK_GROUP_RAID10))
3102                 factor = 2;
3103         else
3104                 factor = 1;
3105
3106         found = __find_space_info(info, flags);
3107         if (found) {
3108                 spin_lock(&found->lock);
3109                 found->total_bytes += total_bytes;
3110                 found->disk_total += total_bytes * factor;
3111                 found->bytes_used += bytes_used;
3112                 found->disk_used += bytes_used * factor;
3113                 found->full = 0;
3114                 spin_unlock(&found->lock);
3115                 *space_info = found;
3116                 return 0;
3117         }
3118         found = kzalloc(sizeof(*found), GFP_NOFS);
3119         if (!found)
3120                 return -ENOMEM;
3121
3122         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3123                 INIT_LIST_HEAD(&found->block_groups[i]);
3124         init_rwsem(&found->groups_sem);
3125         spin_lock_init(&found->lock);
3126         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3127         found->total_bytes = total_bytes;
3128         found->disk_total = total_bytes * factor;
3129         found->bytes_used = bytes_used;
3130         found->disk_used = bytes_used * factor;
3131         found->bytes_pinned = 0;
3132         found->bytes_reserved = 0;
3133         found->bytes_readonly = 0;
3134         found->bytes_may_use = 0;
3135         found->full = 0;
3136         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3137         found->chunk_alloc = 0;
3138         found->flush = 0;
3139         init_waitqueue_head(&found->wait);
3140         *space_info = found;
3141         list_add_rcu(&found->list, &info->space_info);
3142         if (flags & BTRFS_BLOCK_GROUP_DATA)
3143                 info->data_sinfo = found;
3144         return 0;
3145 }
3146
3147 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3148 {
3149         u64 extra_flags = chunk_to_extended(flags) &
3150                                 BTRFS_EXTENDED_PROFILE_MASK;
3151
3152         if (flags & BTRFS_BLOCK_GROUP_DATA)
3153                 fs_info->avail_data_alloc_bits |= extra_flags;
3154         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3155                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3156         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3157                 fs_info->avail_system_alloc_bits |= extra_flags;
3158 }
3159
3160 /*
3161  * returns target flags in extended format or 0 if restripe for this
3162  * chunk_type is not in progress
3163  *
3164  * should be called with either volume_mutex or balance_lock held
3165  */
3166 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3167 {
3168         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3169         u64 target = 0;
3170
3171         if (!bctl)
3172                 return 0;
3173
3174         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3175             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3176                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3177         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3178                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3179                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3180         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3181                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3182                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3183         }
3184
3185         return target;
3186 }
3187
3188 /*
3189  * @flags: available profiles in extended format (see ctree.h)
3190  *
3191  * Returns reduced profile in chunk format.  If profile changing is in
3192  * progress (either running or paused) picks the target profile (if it's
3193  * already available), otherwise falls back to plain reducing.
3194  */
3195 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3196 {
3197         /*
3198          * we add in the count of missing devices because we want
3199          * to make sure that any RAID levels on a degraded FS
3200          * continue to be honored.
3201          */
3202         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3203                 root->fs_info->fs_devices->missing_devices;
3204         u64 target;
3205
3206         /*
3207          * see if restripe for this chunk_type is in progress, if so
3208          * try to reduce to the target profile
3209          */
3210         spin_lock(&root->fs_info->balance_lock);
3211         target = get_restripe_target(root->fs_info, flags);
3212         if (target) {
3213                 /* pick target profile only if it's already available */
3214                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3215                         spin_unlock(&root->fs_info->balance_lock);
3216                         return extended_to_chunk(target);
3217                 }
3218         }
3219         spin_unlock(&root->fs_info->balance_lock);
3220
3221         if (num_devices == 1)
3222                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3223         if (num_devices < 4)
3224                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3225
3226         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3227             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3228                       BTRFS_BLOCK_GROUP_RAID10))) {
3229                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3230         }
3231
3232         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3233             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3234                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3235         }
3236
3237         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3238             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3239              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3240              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3241                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3242         }
3243
3244         return extended_to_chunk(flags);
3245 }
3246
3247 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3248 {
3249         if (flags & BTRFS_BLOCK_GROUP_DATA)
3250                 flags |= root->fs_info->avail_data_alloc_bits;
3251         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3252                 flags |= root->fs_info->avail_system_alloc_bits;
3253         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3254                 flags |= root->fs_info->avail_metadata_alloc_bits;
3255
3256         return btrfs_reduce_alloc_profile(root, flags);
3257 }
3258
3259 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3260 {
3261         u64 flags;
3262
3263         if (data)
3264                 flags = BTRFS_BLOCK_GROUP_DATA;
3265         else if (root == root->fs_info->chunk_root)
3266                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3267         else
3268                 flags = BTRFS_BLOCK_GROUP_METADATA;
3269
3270         return get_alloc_profile(root, flags);
3271 }
3272
3273 /*
3274  * This will check the space that the inode allocates from to make sure we have
3275  * enough space for bytes.
3276  */
3277 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3278 {
3279         struct btrfs_space_info *data_sinfo;
3280         struct btrfs_root *root = BTRFS_I(inode)->root;
3281         struct btrfs_fs_info *fs_info = root->fs_info;
3282         u64 used;
3283         int ret = 0, committed = 0, alloc_chunk = 1;
3284
3285         /* make sure bytes are sectorsize aligned */
3286         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3287
3288         if (root == root->fs_info->tree_root ||
3289             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3290                 alloc_chunk = 0;
3291                 committed = 1;
3292         }
3293
3294         data_sinfo = fs_info->data_sinfo;
3295         if (!data_sinfo)
3296                 goto alloc;
3297
3298 again:
3299         /* make sure we have enough space to handle the data first */
3300         spin_lock(&data_sinfo->lock);
3301         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3302                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3303                 data_sinfo->bytes_may_use;
3304
3305         if (used + bytes > data_sinfo->total_bytes) {
3306                 struct btrfs_trans_handle *trans;
3307
3308                 /*
3309                  * if we don't have enough free bytes in this space then we need
3310                  * to alloc a new chunk.
3311                  */
3312                 if (!data_sinfo->full && alloc_chunk) {
3313                         u64 alloc_target;
3314
3315                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3316                         spin_unlock(&data_sinfo->lock);
3317 alloc:
3318                         alloc_target = btrfs_get_alloc_profile(root, 1);
3319                         trans = btrfs_join_transaction(root);
3320                         if (IS_ERR(trans))
3321                                 return PTR_ERR(trans);
3322
3323                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3324                                              bytes + 2 * 1024 * 1024,
3325                                              alloc_target,
3326                                              CHUNK_ALLOC_NO_FORCE);
3327                         btrfs_end_transaction(trans, root);
3328                         if (ret < 0) {
3329                                 if (ret != -ENOSPC)
3330                                         return ret;
3331                                 else
3332                                         goto commit_trans;
3333                         }
3334
3335                         if (!data_sinfo)
3336                                 data_sinfo = fs_info->data_sinfo;
3337
3338                         goto again;
3339                 }
3340
3341                 /*
3342                  * If we have less pinned bytes than we want to allocate then
3343                  * don't bother committing the transaction, it won't help us.
3344                  */
3345                 if (data_sinfo->bytes_pinned < bytes)
3346                         committed = 1;
3347                 spin_unlock(&data_sinfo->lock);
3348
3349                 /* commit the current transaction and try again */
3350 commit_trans:
3351                 if (!committed &&
3352                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3353                         committed = 1;
3354                         trans = btrfs_join_transaction(root);
3355                         if (IS_ERR(trans))
3356                                 return PTR_ERR(trans);
3357                         ret = btrfs_commit_transaction(trans, root);
3358                         if (ret)
3359                                 return ret;
3360                         goto again;
3361                 }
3362
3363                 return -ENOSPC;
3364         }
3365         data_sinfo->bytes_may_use += bytes;
3366         trace_btrfs_space_reservation(root->fs_info, "space_info",
3367                                       data_sinfo->flags, bytes, 1);
3368         spin_unlock(&data_sinfo->lock);
3369
3370         return 0;
3371 }
3372
3373 /*
3374  * Called if we need to clear a data reservation for this inode.
3375  */
3376 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3377 {
3378         struct btrfs_root *root = BTRFS_I(inode)->root;
3379         struct btrfs_space_info *data_sinfo;
3380
3381         /* make sure bytes are sectorsize aligned */
3382         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3383
3384         data_sinfo = root->fs_info->data_sinfo;
3385         spin_lock(&data_sinfo->lock);
3386         data_sinfo->bytes_may_use -= bytes;
3387         trace_btrfs_space_reservation(root->fs_info, "space_info",
3388                                       data_sinfo->flags, bytes, 0);
3389         spin_unlock(&data_sinfo->lock);
3390 }
3391
3392 static void force_metadata_allocation(struct btrfs_fs_info *info)
3393 {
3394         struct list_head *head = &info->space_info;
3395         struct btrfs_space_info *found;
3396
3397         rcu_read_lock();
3398         list_for_each_entry_rcu(found, head, list) {
3399                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3400                         found->force_alloc = CHUNK_ALLOC_FORCE;
3401         }
3402         rcu_read_unlock();
3403 }
3404
3405 static int should_alloc_chunk(struct btrfs_root *root,
3406                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3407                               int force)
3408 {
3409         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3410         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3411         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3412         u64 thresh;
3413
3414         if (force == CHUNK_ALLOC_FORCE)
3415                 return 1;
3416
3417         /*
3418          * We need to take into account the global rsv because for all intents
3419          * and purposes it's used space.  Don't worry about locking the
3420          * global_rsv, it doesn't change except when the transaction commits.
3421          */
3422         num_allocated += global_rsv->size;
3423
3424         /*
3425          * in limited mode, we want to have some free space up to
3426          * about 1% of the FS size.
3427          */
3428         if (force == CHUNK_ALLOC_LIMITED) {
3429                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3430                 thresh = max_t(u64, 64 * 1024 * 1024,
3431                                div_factor_fine(thresh, 1));
3432
3433                 if (num_bytes - num_allocated < thresh)
3434                         return 1;
3435         }
3436         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3437
3438         /* 256MB or 2% of the FS */
3439         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3440         /* system chunks need a much small threshold */
3441         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3442                 thresh = 32 * 1024 * 1024;
3443
3444         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3445                 return 0;
3446         return 1;
3447 }
3448
3449 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3450 {
3451         u64 num_dev;
3452
3453         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3454             type & BTRFS_BLOCK_GROUP_RAID0)
3455                 num_dev = root->fs_info->fs_devices->rw_devices;
3456         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3457                 num_dev = 2;
3458         else
3459                 num_dev = 1;    /* DUP or single */
3460
3461         /* metadata for updaing devices and chunk tree */
3462         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3463 }
3464
3465 static void check_system_chunk(struct btrfs_trans_handle *trans,
3466                                struct btrfs_root *root, u64 type)
3467 {
3468         struct btrfs_space_info *info;
3469         u64 left;
3470         u64 thresh;
3471
3472         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3473         spin_lock(&info->lock);
3474         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3475                 info->bytes_reserved - info->bytes_readonly;
3476         spin_unlock(&info->lock);
3477
3478         thresh = get_system_chunk_thresh(root, type);
3479         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3480                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3481                        left, thresh, type);
3482                 dump_space_info(info, 0, 0);
3483         }
3484
3485         if (left < thresh) {
3486                 u64 flags;
3487
3488                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3489                 btrfs_alloc_chunk(trans, root, flags);
3490         }
3491 }
3492
3493 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3494                           struct btrfs_root *extent_root, u64 alloc_bytes,
3495                           u64 flags, int force)
3496 {
3497         struct btrfs_space_info *space_info;
3498         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3499         int wait_for_alloc = 0;
3500         int ret = 0;
3501
3502         space_info = __find_space_info(extent_root->fs_info, flags);
3503         if (!space_info) {
3504                 ret = update_space_info(extent_root->fs_info, flags,
3505                                         0, 0, &space_info);
3506                 BUG_ON(ret); /* -ENOMEM */
3507         }
3508         BUG_ON(!space_info); /* Logic error */
3509
3510 again:
3511         spin_lock(&space_info->lock);
3512         if (force < space_info->force_alloc)
3513                 force = space_info->force_alloc;
3514         if (space_info->full) {
3515                 spin_unlock(&space_info->lock);
3516                 return 0;
3517         }
3518
3519         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3520                 spin_unlock(&space_info->lock);
3521                 return 0;
3522         } else if (space_info->chunk_alloc) {
3523                 wait_for_alloc = 1;
3524         } else {
3525                 space_info->chunk_alloc = 1;
3526         }
3527
3528         spin_unlock(&space_info->lock);
3529
3530         mutex_lock(&fs_info->chunk_mutex);
3531
3532         /*
3533          * The chunk_mutex is held throughout the entirety of a chunk
3534          * allocation, so once we've acquired the chunk_mutex we know that the
3535          * other guy is done and we need to recheck and see if we should
3536          * allocate.
3537          */
3538         if (wait_for_alloc) {
3539                 mutex_unlock(&fs_info->chunk_mutex);
3540                 wait_for_alloc = 0;
3541                 goto again;
3542         }
3543
3544         /*
3545          * If we have mixed data/metadata chunks we want to make sure we keep
3546          * allocating mixed chunks instead of individual chunks.
3547          */
3548         if (btrfs_mixed_space_info(space_info))
3549                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3550
3551         /*
3552          * if we're doing a data chunk, go ahead and make sure that
3553          * we keep a reasonable number of metadata chunks allocated in the
3554          * FS as well.
3555          */
3556         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3557                 fs_info->data_chunk_allocations++;
3558                 if (!(fs_info->data_chunk_allocations %
3559                       fs_info->metadata_ratio))
3560                         force_metadata_allocation(fs_info);
3561         }
3562
3563         /*
3564          * Check if we have enough space in SYSTEM chunk because we may need
3565          * to update devices.
3566          */
3567         check_system_chunk(trans, extent_root, flags);
3568
3569         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3570         if (ret < 0 && ret != -ENOSPC)
3571                 goto out;
3572
3573         spin_lock(&space_info->lock);
3574         if (ret)
3575                 space_info->full = 1;
3576         else
3577                 ret = 1;
3578
3579         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3580         space_info->chunk_alloc = 0;
3581         spin_unlock(&space_info->lock);
3582 out:
3583         mutex_unlock(&fs_info->chunk_mutex);
3584         return ret;
3585 }
3586
3587 /*
3588  * shrink metadata reservation for delalloc
3589  */
3590 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3591                             bool wait_ordered)
3592 {
3593         struct btrfs_block_rsv *block_rsv;
3594         struct btrfs_space_info *space_info;
3595         struct btrfs_trans_handle *trans;
3596         u64 delalloc_bytes;
3597         u64 max_reclaim;
3598         long time_left;
3599         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3600         int loops = 0;
3601
3602         trans = (struct btrfs_trans_handle *)current->journal_info;
3603         block_rsv = &root->fs_info->delalloc_block_rsv;
3604         space_info = block_rsv->space_info;
3605
3606         smp_mb();
3607         delalloc_bytes = root->fs_info->delalloc_bytes;
3608         if (delalloc_bytes == 0) {
3609                 if (trans)
3610                         return;
3611                 btrfs_wait_ordered_extents(root, 0, 0);
3612                 return;
3613         }
3614
3615         while (delalloc_bytes && loops < 3) {
3616                 max_reclaim = min(delalloc_bytes, to_reclaim);
3617                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3618                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3619                                                WB_REASON_FS_FREE_SPACE);
3620
3621                 spin_lock(&space_info->lock);
3622                 if (space_info->bytes_used + space_info->bytes_reserved +
3623                     space_info->bytes_pinned + space_info->bytes_readonly +
3624                     space_info->bytes_may_use + orig <=
3625                     space_info->total_bytes) {
3626                         spin_unlock(&space_info->lock);
3627                         break;
3628                 }
3629                 spin_unlock(&space_info->lock);
3630
3631                 loops++;
3632                 if (wait_ordered && !trans) {
3633                         btrfs_wait_ordered_extents(root, 0, 0);
3634                 } else {
3635                         time_left = schedule_timeout_killable(1);
3636                         if (time_left)
3637                                 break;
3638                 }
3639                 smp_mb();
3640                 delalloc_bytes = root->fs_info->delalloc_bytes;
3641         }
3642 }
3643
3644 /**
3645  * maybe_commit_transaction - possibly commit the transaction if its ok to
3646  * @root - the root we're allocating for
3647  * @bytes - the number of bytes we want to reserve
3648  * @force - force the commit
3649  *
3650  * This will check to make sure that committing the transaction will actually
3651  * get us somewhere and then commit the transaction if it does.  Otherwise it
3652  * will return -ENOSPC.
3653  */
3654 static int may_commit_transaction(struct btrfs_root *root,
3655                                   struct btrfs_space_info *space_info,
3656                                   u64 bytes, int force)
3657 {
3658         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3659         struct btrfs_trans_handle *trans;
3660
3661         trans = (struct btrfs_trans_handle *)current->journal_info;
3662         if (trans)
3663                 return -EAGAIN;
3664
3665         if (force)
3666                 goto commit;
3667
3668         /* See if there is enough pinned space to make this reservation */
3669         spin_lock(&space_info->lock);
3670         if (space_info->bytes_pinned >= bytes) {
3671                 spin_unlock(&space_info->lock);
3672                 goto commit;
3673         }
3674         spin_unlock(&space_info->lock);
3675
3676         /*
3677          * See if there is some space in the delayed insertion reservation for
3678          * this reservation.
3679          */
3680         if (space_info != delayed_rsv->space_info)
3681                 return -ENOSPC;
3682
3683         spin_lock(&space_info->lock);
3684         spin_lock(&delayed_rsv->lock);
3685         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3686                 spin_unlock(&delayed_rsv->lock);
3687                 spin_unlock(&space_info->lock);
3688                 return -ENOSPC;
3689         }
3690         spin_unlock(&delayed_rsv->lock);
3691         spin_unlock(&space_info->lock);
3692
3693 commit:
3694         trans = btrfs_join_transaction(root);
3695         if (IS_ERR(trans))
3696                 return -ENOSPC;
3697
3698         return btrfs_commit_transaction(trans, root);
3699 }
3700
3701 enum flush_state {
3702         FLUSH_DELALLOC          =       1,
3703         FLUSH_DELALLOC_WAIT     =       2,
3704         FLUSH_DELAYED_ITEMS_NR  =       3,
3705         FLUSH_DELAYED_ITEMS     =       4,
3706         COMMIT_TRANS            =       5,
3707 };
3708
3709 static int flush_space(struct btrfs_root *root,
3710                        struct btrfs_space_info *space_info, u64 num_bytes,
3711                        u64 orig_bytes, int state)
3712 {
3713         struct btrfs_trans_handle *trans;
3714         int nr;
3715         int ret = 0;
3716
3717         switch (state) {
3718         case FLUSH_DELALLOC:
3719         case FLUSH_DELALLOC_WAIT:
3720                 shrink_delalloc(root, num_bytes, orig_bytes,
3721                                 state == FLUSH_DELALLOC_WAIT);
3722                 break;
3723         case FLUSH_DELAYED_ITEMS_NR:
3724         case FLUSH_DELAYED_ITEMS:
3725                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3726                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3727
3728                         nr = (int)div64_u64(num_bytes, bytes);
3729                         if (!nr)
3730                                 nr = 1;
3731                         nr *= 2;
3732                 } else {
3733                         nr = -1;
3734                 }
3735                 trans = btrfs_join_transaction(root);
3736                 if (IS_ERR(trans)) {
3737                         ret = PTR_ERR(trans);
3738                         break;
3739                 }
3740                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3741                 btrfs_end_transaction(trans, root);
3742                 break;
3743         case COMMIT_TRANS:
3744                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3745                 break;
3746         default:
3747                 ret = -ENOSPC;
3748                 break;
3749         }
3750
3751         return ret;
3752 }
3753 /**
3754  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3755  * @root - the root we're allocating for
3756  * @block_rsv - the block_rsv we're allocating for
3757  * @orig_bytes - the number of bytes we want
3758  * @flush - wether or not we can flush to make our reservation
3759  *
3760  * This will reserve orgi_bytes number of bytes from the space info associated
3761  * with the block_rsv.  If there is not enough space it will make an attempt to
3762  * flush out space to make room.  It will do this by flushing delalloc if
3763  * possible or committing the transaction.  If flush is 0 then no attempts to
3764  * regain reservations will be made and this will fail if there is not enough
3765  * space already.
3766  */
3767 static int reserve_metadata_bytes(struct btrfs_root *root,
3768                                   struct btrfs_block_rsv *block_rsv,
3769                                   u64 orig_bytes, int flush)
3770 {
3771         struct btrfs_space_info *space_info = block_rsv->space_info;
3772         u64 used;
3773         u64 num_bytes = orig_bytes;
3774         int flush_state = FLUSH_DELALLOC;
3775         int ret = 0;
3776         bool flushing = false;
3777         bool committed = false;
3778
3779 again:
3780         ret = 0;
3781         spin_lock(&space_info->lock);
3782         /*
3783          * We only want to wait if somebody other than us is flushing and we are
3784          * actually alloed to flush.
3785          */
3786         while (flush && !flushing && space_info->flush) {
3787                 spin_unlock(&space_info->lock);
3788                 /*
3789                  * If we have a trans handle we can't wait because the flusher
3790                  * may have to commit the transaction, which would mean we would
3791                  * deadlock since we are waiting for the flusher to finish, but
3792                  * hold the current transaction open.
3793                  */
3794                 if (current->journal_info)
3795                         return -EAGAIN;
3796                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3797                 /* Must have been killed, return */
3798                 if (ret)
3799                         return -EINTR;
3800
3801                 spin_lock(&space_info->lock);
3802         }
3803
3804         ret = -ENOSPC;
3805         used = space_info->bytes_used + space_info->bytes_reserved +
3806                 space_info->bytes_pinned + space_info->bytes_readonly +
3807                 space_info->bytes_may_use;
3808
3809         /*
3810          * The idea here is that we've not already over-reserved the block group
3811          * then we can go ahead and save our reservation first and then start
3812          * flushing if we need to.  Otherwise if we've already overcommitted
3813          * lets start flushing stuff first and then come back and try to make
3814          * our reservation.
3815          */
3816         if (used <= space_info->total_bytes) {
3817                 if (used + orig_bytes <= space_info->total_bytes) {
3818                         space_info->bytes_may_use += orig_bytes;
3819                         trace_btrfs_space_reservation(root->fs_info,
3820                                 "space_info", space_info->flags, orig_bytes, 1);
3821                         ret = 0;
3822                 } else {
3823                         /*
3824                          * Ok set num_bytes to orig_bytes since we aren't
3825                          * overocmmitted, this way we only try and reclaim what
3826                          * we need.
3827                          */
3828                         num_bytes = orig_bytes;
3829                 }
3830         } else {
3831                 /*
3832                  * Ok we're over committed, set num_bytes to the overcommitted
3833                  * amount plus the amount of bytes that we need for this
3834                  * reservation.
3835                  */
3836                 num_bytes = used - space_info->total_bytes +
3837                         (orig_bytes * 2);
3838         }
3839
3840         if (ret) {
3841                 u64 profile = btrfs_get_alloc_profile(root, 0);
3842                 u64 avail;
3843
3844                 /*
3845                  * If we have a lot of space that's pinned, don't bother doing
3846                  * the overcommit dance yet and just commit the transaction.
3847                  */
3848                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3849                 do_div(avail, 10);
3850                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3851                         space_info->flush = 1;
3852                         flushing = true;
3853                         spin_unlock(&space_info->lock);
3854                         ret = may_commit_transaction(root, space_info,
3855                                                      orig_bytes, 1);
3856                         if (ret)
3857                                 goto out;
3858                         committed = true;
3859                         goto again;
3860                 }
3861
3862                 spin_lock(&root->fs_info->free_chunk_lock);
3863                 avail = root->fs_info->free_chunk_space;
3864
3865                 /*
3866                  * If we have dup, raid1 or raid10 then only half of the free
3867                  * space is actually useable.
3868                  */
3869                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3870                                BTRFS_BLOCK_GROUP_RAID1 |
3871                                BTRFS_BLOCK_GROUP_RAID10))
3872                         avail >>= 1;
3873
3874                 /*
3875                  * If we aren't flushing don't let us overcommit too much, say
3876                  * 1/8th of the space.  If we can flush, let it overcommit up to
3877                  * 1/2 of the space.
3878                  */
3879                 if (flush)
3880                         avail >>= 3;
3881                 else
3882                         avail >>= 1;
3883                  spin_unlock(&root->fs_info->free_chunk_lock);
3884
3885                 if (used + num_bytes < space_info->total_bytes + avail) {
3886                         space_info->bytes_may_use += orig_bytes;
3887                         trace_btrfs_space_reservation(root->fs_info,
3888                                 "space_info", space_info->flags, orig_bytes, 1);
3889                         ret = 0;
3890                 }
3891         }
3892
3893         /*
3894          * Couldn't make our reservation, save our place so while we're trying
3895          * to reclaim space we can actually use it instead of somebody else
3896          * stealing it from us.
3897          */
3898         if (ret && flush) {
3899                 flushing = true;
3900                 space_info->flush = 1;
3901         }
3902
3903         spin_unlock(&space_info->lock);
3904
3905         if (!ret || !flush)
3906                 goto out;
3907
3908         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3909                           flush_state);
3910         flush_state++;
3911         if (!ret)
3912                 goto again;
3913         else if (flush_state <= COMMIT_TRANS)
3914                 goto again;
3915
3916 out:
3917         if (flushing) {
3918                 spin_lock(&space_info->lock);
3919                 space_info->flush = 0;
3920                 wake_up_all(&space_info->wait);
3921                 spin_unlock(&space_info->lock);
3922         }
3923         return ret;
3924 }
3925
3926 static struct btrfs_block_rsv *get_block_rsv(
3927                                         const struct btrfs_trans_handle *trans,
3928                                         const struct btrfs_root *root)
3929 {
3930         struct btrfs_block_rsv *block_rsv = NULL;
3931
3932         if (root->ref_cows)
3933                 block_rsv = trans->block_rsv;
3934
3935         if (root == root->fs_info->csum_root && trans->adding_csums)
3936                 block_rsv = trans->block_rsv;
3937
3938         if (!block_rsv)
3939                 block_rsv = root->block_rsv;
3940
3941         if (!block_rsv)
3942                 block_rsv = &root->fs_info->empty_block_rsv;
3943
3944         return block_rsv;
3945 }
3946
3947 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3948                                u64 num_bytes)
3949 {
3950         int ret = -ENOSPC;
3951         spin_lock(&block_rsv->lock);
3952         if (block_rsv->reserved >= num_bytes) {
3953                 block_rsv->reserved -= num_bytes;
3954                 if (block_rsv->reserved < block_rsv->size)
3955                         block_rsv->full = 0;
3956                 ret = 0;
3957         }
3958         spin_unlock(&block_rsv->lock);
3959         return ret;
3960 }
3961
3962 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3963                                 u64 num_bytes, int update_size)
3964 {
3965         spin_lock(&block_rsv->lock);
3966         block_rsv->reserved += num_bytes;
3967         if (update_size)
3968                 block_rsv->size += num_bytes;
3969         else if (block_rsv->reserved >= block_rsv->size)
3970                 block_rsv->full = 1;
3971         spin_unlock(&block_rsv->lock);
3972 }
3973
3974 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3975                                     struct btrfs_block_rsv *block_rsv,
3976                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3977 {
3978         struct btrfs_space_info *space_info = block_rsv->space_info;
3979
3980         spin_lock(&block_rsv->lock);
3981         if (num_bytes == (u64)-1)
3982                 num_bytes = block_rsv->size;
3983         block_rsv->size -= num_bytes;
3984         if (block_rsv->reserved >= block_rsv->size) {
3985                 num_bytes = block_rsv->reserved - block_rsv->size;
3986                 block_rsv->reserved = block_rsv->size;
3987                 block_rsv->full = 1;
3988         } else {
3989                 num_bytes = 0;
3990         }
3991         spin_unlock(&block_rsv->lock);
3992
3993         if (num_bytes > 0) {
3994                 if (dest) {
3995                         spin_lock(&dest->lock);
3996                         if (!dest->full) {
3997                                 u64 bytes_to_add;
3998
3999                                 bytes_to_add = dest->size - dest->reserved;
4000                                 bytes_to_add = min(num_bytes, bytes_to_add);
4001                                 dest->reserved += bytes_to_add;
4002                                 if (dest->reserved >= dest->size)
4003                                         dest->full = 1;
4004                                 num_bytes -= bytes_to_add;
4005                         }
4006                         spin_unlock(&dest->lock);
4007                 }
4008                 if (num_bytes) {
4009                         spin_lock(&space_info->lock);
4010                         space_info->bytes_may_use -= num_bytes;
4011                         trace_btrfs_space_reservation(fs_info, "space_info",
4012                                         space_info->flags, num_bytes, 0);
4013                         space_info->reservation_progress++;
4014                         spin_unlock(&space_info->lock);
4015                 }
4016         }
4017 }
4018
4019 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4020                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4021 {
4022         int ret;
4023
4024         ret = block_rsv_use_bytes(src, num_bytes);
4025         if (ret)
4026                 return ret;
4027
4028         block_rsv_add_bytes(dst, num_bytes, 1);
4029         return 0;
4030 }
4031
4032 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4033 {
4034         memset(rsv, 0, sizeof(*rsv));
4035         spin_lock_init(&rsv->lock);
4036 }
4037
4038 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4039 {
4040         struct btrfs_block_rsv *block_rsv;
4041         struct btrfs_fs_info *fs_info = root->fs_info;
4042
4043         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4044         if (!block_rsv)
4045                 return NULL;
4046
4047         btrfs_init_block_rsv(block_rsv);
4048         block_rsv->space_info = __find_space_info(fs_info,
4049                                                   BTRFS_BLOCK_GROUP_METADATA);
4050         return block_rsv;
4051 }
4052
4053 void btrfs_free_block_rsv(struct btrfs_root *root,
4054                           struct btrfs_block_rsv *rsv)
4055 {
4056         btrfs_block_rsv_release(root, rsv, (u64)-1);
4057         kfree(rsv);
4058 }
4059
4060 static inline int __block_rsv_add(struct btrfs_root *root,
4061                                   struct btrfs_block_rsv *block_rsv,
4062                                   u64 num_bytes, int flush)
4063 {
4064         int ret;
4065
4066         if (num_bytes == 0)
4067                 return 0;
4068
4069         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4070         if (!ret) {
4071                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4072                 return 0;
4073         }
4074
4075         return ret;
4076 }
4077
4078 int btrfs_block_rsv_add(struct btrfs_root *root,
4079                         struct btrfs_block_rsv *block_rsv,
4080                         u64 num_bytes)
4081 {
4082         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4083 }
4084
4085 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4086                                 struct btrfs_block_rsv *block_rsv,
4087                                 u64 num_bytes)
4088 {
4089         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4090 }
4091
4092 int btrfs_block_rsv_check(struct btrfs_root *root,
4093                           struct btrfs_block_rsv *block_rsv, int min_factor)
4094 {
4095         u64 num_bytes = 0;
4096         int ret = -ENOSPC;
4097
4098         if (!block_rsv)
4099                 return 0;
4100
4101         spin_lock(&block_rsv->lock);
4102         num_bytes = div_factor(block_rsv->size, min_factor);
4103         if (block_rsv->reserved >= num_bytes)
4104                 ret = 0;
4105         spin_unlock(&block_rsv->lock);
4106
4107         return ret;
4108 }
4109
4110 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4111                                            struct btrfs_block_rsv *block_rsv,
4112                                            u64 min_reserved, int flush)
4113 {
4114         u64 num_bytes = 0;
4115         int ret = -ENOSPC;
4116
4117         if (!block_rsv)
4118                 return 0;
4119
4120         spin_lock(&block_rsv->lock);
4121         num_bytes = min_reserved;
4122         if (block_rsv->reserved >= num_bytes)
4123                 ret = 0;
4124         else
4125                 num_bytes -= block_rsv->reserved;
4126         spin_unlock(&block_rsv->lock);
4127
4128         if (!ret)
4129                 return 0;
4130
4131         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4132         if (!ret) {
4133                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4134                 return 0;
4135         }
4136
4137         return ret;
4138 }
4139
4140 int btrfs_block_rsv_refill(struct btrfs_root *root,
4141                            struct btrfs_block_rsv *block_rsv,
4142                            u64 min_reserved)
4143 {
4144         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4145 }
4146
4147 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4148                                    struct btrfs_block_rsv *block_rsv,
4149                                    u64 min_reserved)
4150 {
4151         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4152 }
4153
4154 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4155                             struct btrfs_block_rsv *dst_rsv,
4156                             u64 num_bytes)
4157 {
4158         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4159 }
4160
4161 void btrfs_block_rsv_release(struct btrfs_root *root,
4162                              struct btrfs_block_rsv *block_rsv,
4163                              u64 num_bytes)
4164 {
4165         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4166         if (global_rsv->full || global_rsv == block_rsv ||
4167             block_rsv->space_info != global_rsv->space_info)
4168                 global_rsv = NULL;
4169         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4170                                 num_bytes);
4171 }
4172
4173 /*
4174  * helper to calculate size of global block reservation.
4175  * the desired value is sum of space used by extent tree,
4176  * checksum tree and root tree
4177  */
4178 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4179 {
4180         struct btrfs_space_info *sinfo;
4181         u64 num_bytes;
4182         u64 meta_used;
4183         u64 data_used;
4184         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4185
4186         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4187         spin_lock(&sinfo->lock);
4188         data_used = sinfo->bytes_used;
4189         spin_unlock(&sinfo->lock);
4190
4191         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4192         spin_lock(&sinfo->lock);
4193         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4194                 data_used = 0;
4195         meta_used = sinfo->bytes_used;
4196         spin_unlock(&sinfo->lock);
4197
4198         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4199                     csum_size * 2;
4200         num_bytes += div64_u64(data_used + meta_used, 50);
4201
4202         if (num_bytes * 3 > meta_used)
4203                 num_bytes = div64_u64(meta_used, 3);
4204
4205         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4206 }
4207
4208 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4209 {
4210         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4211         struct btrfs_space_info *sinfo = block_rsv->space_info;
4212         u64 num_bytes;
4213
4214         num_bytes = calc_global_metadata_size(fs_info);
4215
4216         spin_lock(&sinfo->lock);
4217         spin_lock(&block_rsv->lock);
4218
4219         block_rsv->size = num_bytes;
4220
4221         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4222                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4223                     sinfo->bytes_may_use;
4224
4225         if (sinfo->total_bytes > num_bytes) {
4226                 num_bytes = sinfo->total_bytes - num_bytes;
4227                 block_rsv->reserved += num_bytes;
4228                 sinfo->bytes_may_use += num_bytes;
4229                 trace_btrfs_space_reservation(fs_info, "space_info",
4230                                       sinfo->flags, num_bytes, 1);
4231         }
4232
4233         if (block_rsv->reserved >= block_rsv->size) {
4234                 num_bytes = block_rsv->reserved - block_rsv->size;
4235                 sinfo->bytes_may_use -= num_bytes;
4236                 trace_btrfs_space_reservation(fs_info, "space_info",
4237                                       sinfo->flags, num_bytes, 0);
4238                 sinfo->reservation_progress++;
4239                 block_rsv->reserved = block_rsv->size;
4240                 block_rsv->full = 1;
4241         }
4242
4243         spin_unlock(&block_rsv->lock);
4244         spin_unlock(&sinfo->lock);
4245 }
4246
4247 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4248 {
4249         struct btrfs_space_info *space_info;
4250
4251         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4252         fs_info->chunk_block_rsv.space_info = space_info;
4253
4254         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4255         fs_info->global_block_rsv.space_info = space_info;
4256         fs_info->delalloc_block_rsv.space_info = space_info;
4257         fs_info->trans_block_rsv.space_info = space_info;
4258         fs_info->empty_block_rsv.space_info = space_info;
4259         fs_info->delayed_block_rsv.space_info = space_info;
4260
4261         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4262         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4263         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4264         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4265         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4266
4267         update_global_block_rsv(fs_info);
4268 }
4269
4270 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4271 {
4272         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4273                                 (u64)-1);
4274         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4275         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4276         WARN_ON(fs_info->trans_block_rsv.size > 0);
4277         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4278         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4279         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4280         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4281         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4282 }
4283
4284 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4285                                   struct btrfs_root *root)
4286 {
4287         if (!trans->block_rsv)
4288                 return;
4289
4290         if (!trans->bytes_reserved)
4291                 return;
4292
4293         trace_btrfs_space_reservation(root->fs_info, "transaction",
4294                                       trans->transid, trans->bytes_reserved, 0);
4295         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4296         trans->bytes_reserved = 0;
4297 }
4298
4299 /* Can only return 0 or -ENOSPC */
4300 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4301                                   struct inode *inode)
4302 {
4303         struct btrfs_root *root = BTRFS_I(inode)->root;
4304         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4305         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4306
4307         /*
4308          * We need to hold space in order to delete our orphan item once we've
4309          * added it, so this takes the reservation so we can release it later
4310          * when we are truly done with the orphan item.
4311          */
4312         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4313         trace_btrfs_space_reservation(root->fs_info, "orphan",
4314                                       btrfs_ino(inode), num_bytes, 1);
4315         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4316 }
4317
4318 void btrfs_orphan_release_metadata(struct inode *inode)
4319 {
4320         struct btrfs_root *root = BTRFS_I(inode)->root;
4321         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4322         trace_btrfs_space_reservation(root->fs_info, "orphan",
4323                                       btrfs_ino(inode), num_bytes, 0);
4324         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4325 }
4326
4327 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4328                                 struct btrfs_pending_snapshot *pending)
4329 {
4330         struct btrfs_root *root = pending->root;
4331         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4332         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4333         /*
4334          * two for root back/forward refs, two for directory entries
4335          * and one for root of the snapshot.
4336          */
4337         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4338         dst_rsv->space_info = src_rsv->space_info;
4339         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4340 }
4341
4342 /**
4343  * drop_outstanding_extent - drop an outstanding extent
4344  * @inode: the inode we're dropping the extent for
4345  *
4346  * This is called when we are freeing up an outstanding extent, either called
4347  * after an error or after an extent is written.  This will return the number of
4348  * reserved extents that need to be freed.  This must be called with
4349  * BTRFS_I(inode)->lock held.
4350  */
4351 static unsigned drop_outstanding_extent(struct inode *inode)
4352 {
4353         unsigned drop_inode_space = 0;
4354         unsigned dropped_extents = 0;
4355
4356         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4357         BTRFS_I(inode)->outstanding_extents--;
4358
4359         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4360             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4361                                &BTRFS_I(inode)->runtime_flags))
4362                 drop_inode_space = 1;
4363
4364         /*
4365          * If we have more or the same amount of outsanding extents than we have
4366          * reserved then we need to leave the reserved extents count alone.
4367          */
4368         if (BTRFS_I(inode)->outstanding_extents >=
4369             BTRFS_I(inode)->reserved_extents)
4370                 return drop_inode_space;
4371
4372         dropped_extents = BTRFS_I(inode)->reserved_extents -
4373                 BTRFS_I(inode)->outstanding_extents;
4374         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4375         return dropped_extents + drop_inode_space;
4376 }
4377
4378 /**
4379  * calc_csum_metadata_size - return the amount of metada space that must be
4380  *      reserved/free'd for the given bytes.
4381  * @inode: the inode we're manipulating
4382  * @num_bytes: the number of bytes in question
4383  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4384  *
4385  * This adjusts the number of csum_bytes in the inode and then returns the
4386  * correct amount of metadata that must either be reserved or freed.  We
4387  * calculate how many checksums we can fit into one leaf and then divide the
4388  * number of bytes that will need to be checksumed by this value to figure out
4389  * how many checksums will be required.  If we are adding bytes then the number
4390  * may go up and we will return the number of additional bytes that must be
4391  * reserved.  If it is going down we will return the number of bytes that must
4392  * be freed.
4393  *
4394  * This must be called with BTRFS_I(inode)->lock held.
4395  */
4396 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4397                                    int reserve)
4398 {
4399         struct btrfs_root *root = BTRFS_I(inode)->root;
4400         u64 csum_size;
4401         int num_csums_per_leaf;
4402         int num_csums;
4403         int old_csums;
4404
4405         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4406             BTRFS_I(inode)->csum_bytes == 0)
4407                 return 0;
4408
4409         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4410         if (reserve)
4411                 BTRFS_I(inode)->csum_bytes += num_bytes;
4412         else
4413                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4414         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4415         num_csums_per_leaf = (int)div64_u64(csum_size,
4416                                             sizeof(struct btrfs_csum_item) +
4417                                             sizeof(struct btrfs_disk_key));
4418         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4419         num_csums = num_csums + num_csums_per_leaf - 1;
4420         num_csums = num_csums / num_csums_per_leaf;
4421
4422         old_csums = old_csums + num_csums_per_leaf - 1;
4423         old_csums = old_csums / num_csums_per_leaf;
4424
4425         /* No change, no need to reserve more */
4426         if (old_csums == num_csums)
4427                 return 0;
4428
4429         if (reserve)
4430                 return btrfs_calc_trans_metadata_size(root,
4431                                                       num_csums - old_csums);
4432
4433         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4434 }
4435
4436 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4437 {
4438         struct btrfs_root *root = BTRFS_I(inode)->root;
4439         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4440         u64 to_reserve = 0;
4441         u64 csum_bytes;
4442         unsigned nr_extents = 0;
4443         int extra_reserve = 0;
4444         int flush = 1;
4445         int ret;
4446
4447         /* Need to be holding the i_mutex here if we aren't free space cache */
4448         if (btrfs_is_free_space_inode(inode))
4449                 flush = 0;
4450
4451         if (flush && btrfs_transaction_in_commit(root->fs_info))
4452                 schedule_timeout(1);
4453
4454         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4455         num_bytes = ALIGN(num_bytes, root->sectorsize);
4456
4457         spin_lock(&BTRFS_I(inode)->lock);
4458         BTRFS_I(inode)->outstanding_extents++;
4459
4460         if (BTRFS_I(inode)->outstanding_extents >
4461             BTRFS_I(inode)->reserved_extents)
4462                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4463                         BTRFS_I(inode)->reserved_extents;
4464
4465         /*
4466          * Add an item to reserve for updating the inode when we complete the
4467          * delalloc io.
4468          */
4469         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4470                       &BTRFS_I(inode)->runtime_flags)) {
4471                 nr_extents++;
4472                 extra_reserve = 1;
4473         }
4474
4475         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4476         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4477         csum_bytes = BTRFS_I(inode)->csum_bytes;
4478         spin_unlock(&BTRFS_I(inode)->lock);
4479
4480         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4481         if (ret) {
4482                 u64 to_free = 0;
4483                 unsigned dropped;
4484
4485                 spin_lock(&BTRFS_I(inode)->lock);
4486                 dropped = drop_outstanding_extent(inode);
4487                 /*
4488                  * If the inodes csum_bytes is the same as the original
4489                  * csum_bytes then we know we haven't raced with any free()ers
4490                  * so we can just reduce our inodes csum bytes and carry on.
4491                  * Otherwise we have to do the normal free thing to account for
4492                  * the case that the free side didn't free up its reserve
4493                  * because of this outstanding reservation.
4494                  */
4495                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4496                         calc_csum_metadata_size(inode, num_bytes, 0);
4497                 else
4498                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4499                 spin_unlock(&BTRFS_I(inode)->lock);
4500                 if (dropped)
4501                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4502
4503                 if (to_free) {
4504                         btrfs_block_rsv_release(root, block_rsv, to_free);
4505                         trace_btrfs_space_reservation(root->fs_info,
4506                                                       "delalloc",
4507                                                       btrfs_ino(inode),
4508                                                       to_free, 0);
4509                 }
4510                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4511                 return ret;
4512         }
4513
4514         spin_lock(&BTRFS_I(inode)->lock);
4515         if (extra_reserve) {
4516                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4517                         &BTRFS_I(inode)->runtime_flags);
4518                 nr_extents--;
4519         }
4520         BTRFS_I(inode)->reserved_extents += nr_extents;
4521         spin_unlock(&BTRFS_I(inode)->lock);
4522         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4523
4524         if (to_reserve)
4525                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4526                                               btrfs_ino(inode), to_reserve, 1);
4527         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4528
4529         return 0;
4530 }
4531
4532 /**
4533  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4534  * @inode: the inode to release the reservation for
4535  * @num_bytes: the number of bytes we're releasing
4536  *
4537  * This will release the metadata reservation for an inode.  This can be called
4538  * once we complete IO for a given set of bytes to release their metadata
4539  * reservations.
4540  */
4541 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4542 {
4543         struct btrfs_root *root = BTRFS_I(inode)->root;
4544         u64 to_free = 0;
4545         unsigned dropped;
4546
4547         num_bytes = ALIGN(num_bytes, root->sectorsize);
4548         spin_lock(&BTRFS_I(inode)->lock);
4549         dropped = drop_outstanding_extent(inode);
4550
4551         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4552         spin_unlock(&BTRFS_I(inode)->lock);
4553         if (dropped > 0)
4554                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4555
4556         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4557                                       btrfs_ino(inode), to_free, 0);
4558         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4559                                 to_free);
4560 }
4561
4562 /**
4563  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4564  * @inode: inode we're writing to
4565  * @num_bytes: the number of bytes we want to allocate
4566  *
4567  * This will do the following things
4568  *
4569  * o reserve space in the data space info for num_bytes
4570  * o reserve space in the metadata space info based on number of outstanding
4571  *   extents and how much csums will be needed
4572  * o add to the inodes ->delalloc_bytes
4573  * o add it to the fs_info's delalloc inodes list.
4574  *
4575  * This will return 0 for success and -ENOSPC if there is no space left.
4576  */
4577 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4578 {
4579         int ret;
4580
4581         ret = btrfs_check_data_free_space(inode, num_bytes);
4582         if (ret)
4583                 return ret;
4584
4585         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4586         if (ret) {
4587                 btrfs_free_reserved_data_space(inode, num_bytes);
4588                 return ret;
4589         }
4590
4591         return 0;
4592 }
4593
4594 /**
4595  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4596  * @inode: inode we're releasing space for
4597  * @num_bytes: the number of bytes we want to free up
4598  *
4599  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4600  * called in the case that we don't need the metadata AND data reservations
4601  * anymore.  So if there is an error or we insert an inline extent.
4602  *
4603  * This function will release the metadata space that was not used and will
4604  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4605  * list if there are no delalloc bytes left.
4606  */
4607 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4608 {
4609         btrfs_delalloc_release_metadata(inode, num_bytes);
4610         btrfs_free_reserved_data_space(inode, num_bytes);
4611 }
4612
4613 static int update_block_group(struct btrfs_trans_handle *trans,
4614                               struct btrfs_root *root,
4615                               u64 bytenr, u64 num_bytes, int alloc)
4616 {
4617         struct btrfs_block_group_cache *cache = NULL;
4618         struct btrfs_fs_info *info = root->fs_info;
4619         u64 total = num_bytes;
4620         u64 old_val;
4621         u64 byte_in_group;
4622         int factor;
4623
4624         /* block accounting for super block */
4625         spin_lock(&info->delalloc_lock);
4626         old_val = btrfs_super_bytes_used(info->super_copy);
4627         if (alloc)
4628                 old_val += num_bytes;
4629         else
4630                 old_val -= num_bytes;
4631         btrfs_set_super_bytes_used(info->super_copy, old_val);
4632         spin_unlock(&info->delalloc_lock);
4633
4634         while (total) {
4635                 cache = btrfs_lookup_block_group(info, bytenr);
4636                 if (!cache)
4637                         return -ENOENT;
4638                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4639                                     BTRFS_BLOCK_GROUP_RAID1 |
4640                                     BTRFS_BLOCK_GROUP_RAID10))
4641                         factor = 2;
4642                 else
4643                         factor = 1;
4644                 /*
4645                  * If this block group has free space cache written out, we
4646                  * need to make sure to load it if we are removing space.  This
4647                  * is because we need the unpinning stage to actually add the
4648                  * space back to the block group, otherwise we will leak space.
4649                  */
4650                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4651                         cache_block_group(cache, trans, NULL, 1);
4652
4653                 byte_in_group = bytenr - cache->key.objectid;
4654                 WARN_ON(byte_in_group > cache->key.offset);
4655
4656                 spin_lock(&cache->space_info->lock);
4657                 spin_lock(&cache->lock);
4658
4659                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4660                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4661                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4662
4663                 cache->dirty = 1;
4664                 old_val = btrfs_block_group_used(&cache->item);
4665                 num_bytes = min(total, cache->key.offset - byte_in_group);
4666                 if (alloc) {
4667                         old_val += num_bytes;
4668                         btrfs_set_block_group_used(&cache->item, old_val);
4669                         cache->reserved -= num_bytes;
4670                         cache->space_info->bytes_reserved -= num_bytes;
4671                         cache->space_info->bytes_used += num_bytes;
4672                         cache->space_info->disk_used += num_bytes * factor;
4673                         spin_unlock(&cache->lock);
4674                         spin_unlock(&cache->space_info->lock);
4675                 } else {
4676                         old_val -= num_bytes;
4677                         btrfs_set_block_group_used(&cache->item, old_val);
4678                         cache->pinned += num_bytes;
4679                         cache->space_info->bytes_pinned += num_bytes;
4680                         cache->space_info->bytes_used -= num_bytes;
4681                         cache->space_info->disk_used -= num_bytes * factor;
4682                         spin_unlock(&cache->lock);
4683                         spin_unlock(&cache->space_info->lock);
4684
4685                         set_extent_dirty(info->pinned_extents,
4686                                          bytenr, bytenr + num_bytes - 1,
4687                                          GFP_NOFS | __GFP_NOFAIL);
4688                 }
4689                 btrfs_put_block_group(cache);
4690                 total -= num_bytes;
4691                 bytenr += num_bytes;
4692         }
4693         return 0;
4694 }
4695
4696 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4697 {
4698         struct btrfs_block_group_cache *cache;
4699         u64 bytenr;
4700
4701         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4702         if (!cache)
4703                 return 0;
4704
4705         bytenr = cache->key.objectid;
4706         btrfs_put_block_group(cache);
4707
4708         return bytenr;
4709 }
4710
4711 static int pin_down_extent(struct btrfs_root *root,
4712                            struct btrfs_block_group_cache *cache,
4713                            u64 bytenr, u64 num_bytes, int reserved)
4714 {
4715         spin_lock(&cache->space_info->lock);
4716         spin_lock(&cache->lock);
4717         cache->pinned += num_bytes;
4718         cache->space_info->bytes_pinned += num_bytes;
4719         if (reserved) {
4720                 cache->reserved -= num_bytes;
4721                 cache->space_info->bytes_reserved -= num_bytes;
4722         }
4723         spin_unlock(&cache->lock);
4724         spin_unlock(&cache->space_info->lock);
4725
4726         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4727                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4728         return 0;
4729 }
4730
4731 /*
4732  * this function must be called within transaction
4733  */
4734 int btrfs_pin_extent(struct btrfs_root *root,
4735                      u64 bytenr, u64 num_bytes, int reserved)
4736 {
4737         struct btrfs_block_group_cache *cache;
4738
4739         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4740         BUG_ON(!cache); /* Logic error */
4741
4742         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4743
4744         btrfs_put_block_group(cache);
4745         return 0;
4746 }
4747
4748 /*
4749  * this function must be called within transaction
4750  */
4751 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4752                                     struct btrfs_root *root,
4753                                     u64 bytenr, u64 num_bytes)
4754 {
4755         struct btrfs_block_group_cache *cache;
4756
4757         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4758         BUG_ON(!cache); /* Logic error */
4759
4760         /*
4761          * pull in the free space cache (if any) so that our pin
4762          * removes the free space from the cache.  We have load_only set
4763          * to one because the slow code to read in the free extents does check
4764          * the pinned extents.
4765          */
4766         cache_block_group(cache, trans, root, 1);
4767
4768         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4769
4770         /* remove us from the free space cache (if we're there at all) */
4771         btrfs_remove_free_space(cache, bytenr, num_bytes);
4772         btrfs_put_block_group(cache);
4773         return 0;
4774 }
4775
4776 /**
4777  * btrfs_update_reserved_bytes - update the block_group and space info counters
4778  * @cache:      The cache we are manipulating
4779  * @num_bytes:  The number of bytes in question
4780  * @reserve:    One of the reservation enums
4781  *
4782  * This is called by the allocator when it reserves space, or by somebody who is
4783  * freeing space that was never actually used on disk.  For example if you
4784  * reserve some space for a new leaf in transaction A and before transaction A
4785  * commits you free that leaf, you call this with reserve set to 0 in order to
4786  * clear the reservation.
4787  *
4788  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4789  * ENOSPC accounting.  For data we handle the reservation through clearing the
4790  * delalloc bits in the io_tree.  We have to do this since we could end up
4791  * allocating less disk space for the amount of data we have reserved in the
4792  * case of compression.
4793  *
4794  * If this is a reservation and the block group has become read only we cannot
4795  * make the reservation and return -EAGAIN, otherwise this function always
4796  * succeeds.
4797  */
4798 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4799                                        u64 num_bytes, int reserve)
4800 {
4801         struct btrfs_space_info *space_info = cache->space_info;
4802         int ret = 0;
4803
4804         spin_lock(&space_info->lock);
4805         spin_lock(&cache->lock);
4806         if (reserve != RESERVE_FREE) {
4807                 if (cache->ro) {
4808                         ret = -EAGAIN;
4809                 } else {
4810                         cache->reserved += num_bytes;
4811                         space_info->bytes_reserved += num_bytes;
4812                         if (reserve == RESERVE_ALLOC) {
4813                                 trace_btrfs_space_reservation(cache->fs_info,
4814                                                 "space_info", space_info->flags,
4815                                                 num_bytes, 0);
4816                                 space_info->bytes_may_use -= num_bytes;
4817                         }
4818                 }
4819         } else {
4820                 if (cache->ro)
4821                         space_info->bytes_readonly += num_bytes;
4822                 cache->reserved -= num_bytes;
4823                 space_info->bytes_reserved -= num_bytes;
4824                 space_info->reservation_progress++;
4825         }
4826         spin_unlock(&cache->lock);
4827         spin_unlock(&space_info->lock);
4828         return ret;
4829 }
4830
4831 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4832                                 struct btrfs_root *root)
4833 {
4834         struct btrfs_fs_info *fs_info = root->fs_info;
4835         struct btrfs_caching_control *next;
4836         struct btrfs_caching_control *caching_ctl;
4837         struct btrfs_block_group_cache *cache;
4838
4839         down_write(&fs_info->extent_commit_sem);
4840
4841         list_for_each_entry_safe(caching_ctl, next,
4842                                  &fs_info->caching_block_groups, list) {
4843                 cache = caching_ctl->block_group;
4844                 if (block_group_cache_done(cache)) {
4845                         cache->last_byte_to_unpin = (u64)-1;
4846                         list_del_init(&caching_ctl->list);
4847                         put_caching_control(caching_ctl);
4848                 } else {
4849                         cache->last_byte_to_unpin = caching_ctl->progress;
4850                 }
4851         }
4852
4853         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4854                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4855         else
4856                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4857
4858         up_write(&fs_info->extent_commit_sem);
4859
4860         update_global_block_rsv(fs_info);
4861 }
4862
4863 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4864 {
4865         struct btrfs_fs_info *fs_info = root->fs_info;
4866         struct btrfs_block_group_cache *cache = NULL;
4867         u64 len;
4868
4869         while (start <= end) {
4870                 if (!cache ||
4871                     start >= cache->key.objectid + cache->key.offset) {
4872                         if (cache)
4873                                 btrfs_put_block_group(cache);
4874                         cache = btrfs_lookup_block_group(fs_info, start);
4875                         BUG_ON(!cache); /* Logic error */
4876                 }
4877
4878                 len = cache->key.objectid + cache->key.offset - start;
4879                 len = min(len, end + 1 - start);
4880
4881                 if (start < cache->last_byte_to_unpin) {
4882                         len = min(len, cache->last_byte_to_unpin - start);
4883                         btrfs_add_free_space(cache, start, len);
4884                 }
4885
4886                 start += len;
4887
4888                 spin_lock(&cache->space_info->lock);
4889                 spin_lock(&cache->lock);
4890                 cache->pinned -= len;
4891                 cache->space_info->bytes_pinned -= len;
4892                 if (cache->ro)
4893                         cache->space_info->bytes_readonly += len;
4894                 spin_unlock(&cache->lock);
4895                 spin_unlock(&cache->space_info->lock);
4896         }
4897
4898         if (cache)
4899                 btrfs_put_block_group(cache);
4900         return 0;
4901 }
4902
4903 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4904                                struct btrfs_root *root)
4905 {
4906         struct btrfs_fs_info *fs_info = root->fs_info;
4907         struct extent_io_tree *unpin;
4908         u64 start;
4909         u64 end;
4910         int ret;
4911
4912         if (trans->aborted)
4913                 return 0;
4914
4915         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4916                 unpin = &fs_info->freed_extents[1];
4917         else
4918                 unpin = &fs_info->freed_extents[0];
4919
4920         while (1) {
4921                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4922                                             EXTENT_DIRTY);
4923                 if (ret)
4924                         break;
4925
4926                 if (btrfs_test_opt(root, DISCARD))
4927                         ret = btrfs_discard_extent(root, start,
4928                                                    end + 1 - start, NULL);
4929
4930                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4931                 unpin_extent_range(root, start, end);
4932                 cond_resched();
4933         }
4934
4935         return 0;
4936 }
4937
4938 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4939                                 struct btrfs_root *root,
4940                                 u64 bytenr, u64 num_bytes, u64 parent,
4941                                 u64 root_objectid, u64 owner_objectid,
4942                                 u64 owner_offset, int refs_to_drop,
4943                                 struct btrfs_delayed_extent_op *extent_op)
4944 {
4945         struct btrfs_key key;
4946         struct btrfs_path *path;
4947         struct btrfs_fs_info *info = root->fs_info;
4948         struct btrfs_root *extent_root = info->extent_root;
4949         struct extent_buffer *leaf;
4950         struct btrfs_extent_item *ei;
4951         struct btrfs_extent_inline_ref *iref;
4952         int ret;
4953         int is_data;
4954         int extent_slot = 0;
4955         int found_extent = 0;
4956         int num_to_del = 1;
4957         u32 item_size;
4958         u64 refs;
4959
4960         path = btrfs_alloc_path();
4961         if (!path)
4962                 return -ENOMEM;
4963
4964         path->reada = 1;
4965         path->leave_spinning = 1;
4966
4967         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4968         BUG_ON(!is_data && refs_to_drop != 1);
4969
4970         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4971                                     bytenr, num_bytes, parent,
4972                                     root_objectid, owner_objectid,
4973                                     owner_offset);
4974         if (ret == 0) {
4975                 extent_slot = path->slots[0];
4976                 while (extent_slot >= 0) {
4977                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4978                                               extent_slot);
4979                         if (key.objectid != bytenr)
4980                                 break;
4981                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4982                             key.offset == num_bytes) {
4983                                 found_extent = 1;
4984                                 break;
4985                         }
4986                         if (path->slots[0] - extent_slot > 5)
4987                                 break;
4988                         extent_slot--;
4989                 }
4990 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4991                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4992                 if (found_extent && item_size < sizeof(*ei))
4993                         found_extent = 0;
4994 #endif
4995                 if (!found_extent) {
4996                         BUG_ON(iref);
4997                         ret = remove_extent_backref(trans, extent_root, path,
4998                                                     NULL, refs_to_drop,
4999                                                     is_data);
5000                         if (ret)
5001                                 goto abort;
5002                         btrfs_release_path(path);
5003                         path->leave_spinning = 1;
5004
5005                         key.objectid = bytenr;
5006                         key.type = BTRFS_EXTENT_ITEM_KEY;
5007                         key.offset = num_bytes;
5008
5009                         ret = btrfs_search_slot(trans, extent_root,
5010                                                 &key, path, -1, 1);
5011                         if (ret) {
5012                                 printk(KERN_ERR "umm, got %d back from search"
5013                                        ", was looking for %llu\n", ret,
5014                                        (unsigned long long)bytenr);
5015                                 if (ret > 0)
5016                                         btrfs_print_leaf(extent_root,
5017                                                          path->nodes[0]);
5018                         }
5019                         if (ret < 0)
5020                                 goto abort;
5021                         extent_slot = path->slots[0];
5022                 }
5023         } else if (ret == -ENOENT) {
5024                 btrfs_print_leaf(extent_root, path->nodes[0]);
5025                 WARN_ON(1);
5026                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5027                        "parent %llu root %llu  owner %llu offset %llu\n",
5028                        (unsigned long long)bytenr,
5029                        (unsigned long long)parent,
5030                        (unsigned long long)root_objectid,
5031                        (unsigned long long)owner_objectid,
5032                        (unsigned long long)owner_offset);
5033         } else {
5034                 goto abort;
5035         }
5036
5037         leaf = path->nodes[0];
5038         item_size = btrfs_item_size_nr(leaf, extent_slot);
5039 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5040         if (item_size < sizeof(*ei)) {
5041                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5042                 ret = convert_extent_item_v0(trans, extent_root, path,
5043                                              owner_objectid, 0);
5044                 if (ret < 0)
5045                         goto abort;
5046
5047                 btrfs_release_path(path);
5048                 path->leave_spinning = 1;
5049
5050                 key.objectid = bytenr;
5051                 key.type = BTRFS_EXTENT_ITEM_KEY;
5052                 key.offset = num_bytes;
5053
5054                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5055                                         -1, 1);
5056                 if (ret) {
5057                         printk(KERN_ERR "umm, got %d back from search"
5058                                ", was looking for %llu\n", ret,
5059                                (unsigned long long)bytenr);
5060                         btrfs_print_leaf(extent_root, path->nodes[0]);
5061                 }
5062                 if (ret < 0)
5063                         goto abort;
5064                 extent_slot = path->slots[0];
5065                 leaf = path->nodes[0];
5066                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5067         }
5068 #endif
5069         BUG_ON(item_size < sizeof(*ei));
5070         ei = btrfs_item_ptr(leaf, extent_slot,
5071                             struct btrfs_extent_item);
5072         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5073                 struct btrfs_tree_block_info *bi;
5074                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5075                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5076                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5077         }
5078
5079         refs = btrfs_extent_refs(leaf, ei);
5080         BUG_ON(refs < refs_to_drop);
5081         refs -= refs_to_drop;
5082
5083         if (refs > 0) {
5084                 if (extent_op)
5085                         __run_delayed_extent_op(extent_op, leaf, ei);
5086                 /*
5087                  * In the case of inline back ref, reference count will
5088                  * be updated by remove_extent_backref
5089                  */
5090                 if (iref) {
5091                         BUG_ON(!found_extent);
5092                 } else {
5093                         btrfs_set_extent_refs(leaf, ei, refs);
5094                         btrfs_mark_buffer_dirty(leaf);
5095                 }
5096                 if (found_extent) {
5097                         ret = remove_extent_backref(trans, extent_root, path,
5098                                                     iref, refs_to_drop,
5099                                                     is_data);
5100                         if (ret)
5101                                 goto abort;
5102                 }
5103         } else {
5104                 if (found_extent) {
5105                         BUG_ON(is_data && refs_to_drop !=
5106                                extent_data_ref_count(root, path, iref));
5107                         if (iref) {
5108                                 BUG_ON(path->slots[0] != extent_slot);
5109                         } else {
5110                                 BUG_ON(path->slots[0] != extent_slot + 1);
5111                                 path->slots[0] = extent_slot;
5112                                 num_to_del = 2;
5113                         }
5114                 }
5115
5116                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5117                                       num_to_del);
5118                 if (ret)
5119                         goto abort;
5120                 btrfs_release_path(path);
5121
5122                 if (is_data) {
5123                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5124                         if (ret)
5125                                 goto abort;
5126                 }
5127
5128                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5129                 if (ret)
5130                         goto abort;
5131         }
5132 out:
5133         btrfs_free_path(path);
5134         return ret;
5135
5136 abort:
5137         btrfs_abort_transaction(trans, extent_root, ret);
5138         goto out;
5139 }
5140
5141 /*
5142  * when we free an block, it is possible (and likely) that we free the last
5143  * delayed ref for that extent as well.  This searches the delayed ref tree for
5144  * a given extent, and if there are no other delayed refs to be processed, it
5145  * removes it from the tree.
5146  */
5147 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5148                                       struct btrfs_root *root, u64 bytenr)
5149 {
5150         struct btrfs_delayed_ref_head *head;
5151         struct btrfs_delayed_ref_root *delayed_refs;
5152         struct btrfs_delayed_ref_node *ref;
5153         struct rb_node *node;
5154         int ret = 0;
5155
5156         delayed_refs = &trans->transaction->delayed_refs;
5157         spin_lock(&delayed_refs->lock);
5158         head = btrfs_find_delayed_ref_head(trans, bytenr);
5159         if (!head)
5160                 goto out;
5161
5162         node = rb_prev(&head->node.rb_node);
5163         if (!node)
5164                 goto out;
5165
5166         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5167
5168         /* there are still entries for this ref, we can't drop it */
5169         if (ref->bytenr == bytenr)
5170                 goto out;
5171
5172         if (head->extent_op) {
5173                 if (!head->must_insert_reserved)
5174                         goto out;
5175                 kfree(head->extent_op);
5176                 head->extent_op = NULL;
5177         }
5178
5179         /*
5180          * waiting for the lock here would deadlock.  If someone else has it
5181          * locked they are already in the process of dropping it anyway
5182          */
5183         if (!mutex_trylock(&head->mutex))
5184                 goto out;
5185
5186         /*
5187          * at this point we have a head with no other entries.  Go
5188          * ahead and process it.
5189          */
5190         head->node.in_tree = 0;
5191         rb_erase(&head->node.rb_node, &delayed_refs->root);
5192
5193         delayed_refs->num_entries--;
5194         if (waitqueue_active(&delayed_refs->seq_wait))
5195                 wake_up(&delayed_refs->seq_wait);
5196
5197         /*
5198          * we don't take a ref on the node because we're removing it from the
5199          * tree, so we just steal the ref the tree was holding.
5200          */
5201         delayed_refs->num_heads--;
5202         if (list_empty(&head->cluster))
5203                 delayed_refs->num_heads_ready--;
5204
5205         list_del_init(&head->cluster);
5206         spin_unlock(&delayed_refs->lock);
5207
5208         BUG_ON(head->extent_op);
5209         if (head->must_insert_reserved)
5210                 ret = 1;
5211
5212         mutex_unlock(&head->mutex);
5213         btrfs_put_delayed_ref(&head->node);
5214         return ret;
5215 out:
5216         spin_unlock(&delayed_refs->lock);
5217         return 0;
5218 }
5219
5220 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5221                            struct btrfs_root *root,
5222                            struct extent_buffer *buf,
5223                            u64 parent, int last_ref)
5224 {
5225         struct btrfs_block_group_cache *cache = NULL;
5226         int ret;
5227
5228         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5229                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5230                                         buf->start, buf->len,
5231                                         parent, root->root_key.objectid,
5232                                         btrfs_header_level(buf),
5233                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5234                 BUG_ON(ret); /* -ENOMEM */
5235         }
5236
5237         if (!last_ref)
5238                 return;
5239
5240         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5241
5242         if (btrfs_header_generation(buf) == trans->transid) {
5243                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5244                         ret = check_ref_cleanup(trans, root, buf->start);
5245                         if (!ret)
5246                                 goto out;
5247                 }
5248
5249                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5250                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5251                         goto out;
5252                 }
5253
5254                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5255
5256                 btrfs_add_free_space(cache, buf->start, buf->len);
5257                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5258         }
5259 out:
5260         /*
5261          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5262          * anymore.
5263          */
5264         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5265         btrfs_put_block_group(cache);
5266 }
5267
5268 /* Can return -ENOMEM */
5269 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5270                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5271                       u64 owner, u64 offset, int for_cow)
5272 {
5273         int ret;
5274         struct btrfs_fs_info *fs_info = root->fs_info;
5275
5276         /*
5277          * tree log blocks never actually go into the extent allocation
5278          * tree, just update pinning info and exit early.
5279          */
5280         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5281                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5282                 /* unlocks the pinned mutex */
5283                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5284                 ret = 0;
5285         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5286                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5287                                         num_bytes,
5288                                         parent, root_objectid, (int)owner,
5289                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5290         } else {
5291                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5292                                                 num_bytes,
5293                                                 parent, root_objectid, owner,
5294                                                 offset, BTRFS_DROP_DELAYED_REF,
5295                                                 NULL, for_cow);
5296         }
5297         return ret;
5298 }
5299
5300 static u64 stripe_align(struct btrfs_root *root, u64 val)
5301 {
5302         u64 mask = ((u64)root->stripesize - 1);
5303         u64 ret = (val + mask) & ~mask;
5304         return ret;
5305 }
5306
5307 /*
5308  * when we wait for progress in the block group caching, its because
5309  * our allocation attempt failed at least once.  So, we must sleep
5310  * and let some progress happen before we try again.
5311  *
5312  * This function will sleep at least once waiting for new free space to
5313  * show up, and then it will check the block group free space numbers
5314  * for our min num_bytes.  Another option is to have it go ahead
5315  * and look in the rbtree for a free extent of a given size, but this
5316  * is a good start.
5317  */
5318 static noinline int
5319 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5320                                 u64 num_bytes)
5321 {
5322         struct btrfs_caching_control *caching_ctl;
5323         DEFINE_WAIT(wait);
5324
5325         caching_ctl = get_caching_control(cache);
5326         if (!caching_ctl)
5327                 return 0;
5328
5329         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5330                    (cache->free_space_ctl->free_space >= num_bytes));
5331
5332         put_caching_control(caching_ctl);
5333         return 0;
5334 }
5335
5336 static noinline int
5337 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5338 {
5339         struct btrfs_caching_control *caching_ctl;
5340         DEFINE_WAIT(wait);
5341
5342         caching_ctl = get_caching_control(cache);
5343         if (!caching_ctl)
5344                 return 0;
5345
5346         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5347
5348         put_caching_control(caching_ctl);
5349         return 0;
5350 }
5351
5352 static int __get_block_group_index(u64 flags)
5353 {
5354         int index;
5355
5356         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5357                 index = 0;
5358         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5359                 index = 1;
5360         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5361                 index = 2;
5362         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5363                 index = 3;
5364         else
5365                 index = 4;
5366
5367         return index;
5368 }
5369
5370 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5371 {
5372         return __get_block_group_index(cache->flags);
5373 }
5374
5375 enum btrfs_loop_type {
5376         LOOP_CACHING_NOWAIT = 0,
5377         LOOP_CACHING_WAIT = 1,
5378         LOOP_ALLOC_CHUNK = 2,
5379         LOOP_NO_EMPTY_SIZE = 3,
5380 };
5381
5382 /*
5383  * walks the btree of allocated extents and find a hole of a given size.
5384  * The key ins is changed to record the hole:
5385  * ins->objectid == block start
5386  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5387  * ins->offset == number of blocks
5388  * Any available blocks before search_start are skipped.
5389  */
5390 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5391                                      struct btrfs_root *orig_root,
5392                                      u64 num_bytes, u64 empty_size,
5393                                      u64 hint_byte, struct btrfs_key *ins,
5394                                      u64 data)
5395 {
5396         int ret = 0;
5397         struct btrfs_root *root = orig_root->fs_info->extent_root;
5398         struct btrfs_free_cluster *last_ptr = NULL;
5399         struct btrfs_block_group_cache *block_group = NULL;
5400         struct btrfs_block_group_cache *used_block_group;
5401         u64 search_start = 0;
5402         int empty_cluster = 2 * 1024 * 1024;
5403         int allowed_chunk_alloc = 0;
5404         int done_chunk_alloc = 0;
5405         struct btrfs_space_info *space_info;
5406         int loop = 0;
5407         int index = 0;
5408         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5409                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5410         bool found_uncached_bg = false;
5411         bool failed_cluster_refill = false;
5412         bool failed_alloc = false;
5413         bool use_cluster = true;
5414         bool have_caching_bg = false;
5415
5416         WARN_ON(num_bytes < root->sectorsize);
5417         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5418         ins->objectid = 0;
5419         ins->offset = 0;
5420
5421         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5422
5423         space_info = __find_space_info(root->fs_info, data);
5424         if (!space_info) {
5425                 printk(KERN_ERR "No space info for %llu\n", data);
5426                 return -ENOSPC;
5427         }
5428
5429         /*
5430          * If the space info is for both data and metadata it means we have a
5431          * small filesystem and we can't use the clustering stuff.
5432          */
5433         if (btrfs_mixed_space_info(space_info))
5434                 use_cluster = false;
5435
5436         if (orig_root->ref_cows || empty_size)
5437                 allowed_chunk_alloc = 1;
5438
5439         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5440                 last_ptr = &root->fs_info->meta_alloc_cluster;
5441                 if (!btrfs_test_opt(root, SSD))
5442                         empty_cluster = 64 * 1024;
5443         }
5444
5445         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5446             btrfs_test_opt(root, SSD)) {
5447                 last_ptr = &root->fs_info->data_alloc_cluster;
5448         }
5449
5450         if (last_ptr) {
5451                 spin_lock(&last_ptr->lock);
5452                 if (last_ptr->block_group)
5453                         hint_byte = last_ptr->window_start;
5454                 spin_unlock(&last_ptr->lock);
5455         }
5456
5457         search_start = max(search_start, first_logical_byte(root, 0));
5458         search_start = max(search_start, hint_byte);
5459
5460         if (!last_ptr)
5461                 empty_cluster = 0;
5462
5463         if (search_start == hint_byte) {
5464                 block_group = btrfs_lookup_block_group(root->fs_info,
5465                                                        search_start);
5466                 used_block_group = block_group;
5467                 /*
5468                  * we don't want to use the block group if it doesn't match our
5469                  * allocation bits, or if its not cached.
5470                  *
5471                  * However if we are re-searching with an ideal block group
5472                  * picked out then we don't care that the block group is cached.
5473                  */
5474                 if (block_group && block_group_bits(block_group, data) &&
5475                     block_group->cached != BTRFS_CACHE_NO) {
5476                         down_read(&space_info->groups_sem);
5477                         if (list_empty(&block_group->list) ||
5478                             block_group->ro) {
5479                                 /*
5480                                  * someone is removing this block group,
5481                                  * we can't jump into the have_block_group
5482                                  * target because our list pointers are not
5483                                  * valid
5484                                  */
5485                                 btrfs_put_block_group(block_group);
5486                                 up_read(&space_info->groups_sem);
5487                         } else {
5488                                 index = get_block_group_index(block_group);
5489                                 goto have_block_group;
5490                         }
5491                 } else if (block_group) {
5492                         btrfs_put_block_group(block_group);
5493                 }
5494         }
5495 search:
5496         have_caching_bg = false;
5497         down_read(&space_info->groups_sem);
5498         list_for_each_entry(block_group, &space_info->block_groups[index],
5499                             list) {
5500                 u64 offset;
5501                 int cached;
5502
5503                 used_block_group = block_group;
5504                 btrfs_get_block_group(block_group);
5505                 search_start = block_group->key.objectid;
5506
5507                 /*
5508                  * this can happen if we end up cycling through all the
5509                  * raid types, but we want to make sure we only allocate
5510                  * for the proper type.
5511                  */
5512                 if (!block_group_bits(block_group, data)) {
5513                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5514                                 BTRFS_BLOCK_GROUP_RAID1 |
5515                                 BTRFS_BLOCK_GROUP_RAID10;
5516
5517                         /*
5518                          * if they asked for extra copies and this block group
5519                          * doesn't provide them, bail.  This does allow us to
5520                          * fill raid0 from raid1.
5521                          */
5522                         if ((data & extra) && !(block_group->flags & extra))
5523                                 goto loop;
5524                 }
5525
5526 have_block_group:
5527                 cached = block_group_cache_done(block_group);
5528                 if (unlikely(!cached)) {
5529                         found_uncached_bg = true;
5530                         ret = cache_block_group(block_group, trans,
5531                                                 orig_root, 0);
5532                         BUG_ON(ret < 0);
5533                         ret = 0;
5534                 }
5535
5536                 if (unlikely(block_group->ro))
5537                         goto loop;
5538
5539                 /*
5540                  * Ok we want to try and use the cluster allocator, so
5541                  * lets look there
5542                  */
5543                 if (last_ptr) {
5544                         /*
5545                          * the refill lock keeps out other
5546                          * people trying to start a new cluster
5547                          */
5548                         spin_lock(&last_ptr->refill_lock);
5549                         used_block_group = last_ptr->block_group;
5550                         if (used_block_group != block_group &&
5551                             (!used_block_group ||
5552                              used_block_group->ro ||
5553                              !block_group_bits(used_block_group, data))) {
5554                                 used_block_group = block_group;
5555                                 goto refill_cluster;
5556                         }
5557
5558                         if (used_block_group != block_group)
5559                                 btrfs_get_block_group(used_block_group);
5560
5561                         offset = btrfs_alloc_from_cluster(used_block_group,
5562                           last_ptr, num_bytes, used_block_group->key.objectid);
5563                         if (offset) {
5564                                 /* we have a block, we're done */
5565                                 spin_unlock(&last_ptr->refill_lock);
5566                                 trace_btrfs_reserve_extent_cluster(root,
5567                                         block_group, search_start, num_bytes);
5568                                 goto checks;
5569                         }
5570
5571                         WARN_ON(last_ptr->block_group != used_block_group);
5572                         if (used_block_group != block_group) {
5573                                 btrfs_put_block_group(used_block_group);
5574                                 used_block_group = block_group;
5575                         }
5576 refill_cluster:
5577                         BUG_ON(used_block_group != block_group);
5578                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5579                          * set up a new clusters, so lets just skip it
5580                          * and let the allocator find whatever block
5581                          * it can find.  If we reach this point, we
5582                          * will have tried the cluster allocator
5583                          * plenty of times and not have found
5584                          * anything, so we are likely way too
5585                          * fragmented for the clustering stuff to find
5586                          * anything.
5587                          *
5588                          * However, if the cluster is taken from the
5589                          * current block group, release the cluster
5590                          * first, so that we stand a better chance of
5591                          * succeeding in the unclustered
5592                          * allocation.  */
5593                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5594                             last_ptr->block_group != block_group) {
5595                                 spin_unlock(&last_ptr->refill_lock);
5596                                 goto unclustered_alloc;
5597                         }
5598
5599                         /*
5600                          * this cluster didn't work out, free it and
5601                          * start over
5602                          */
5603                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5604
5605                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5606                                 spin_unlock(&last_ptr->refill_lock);
5607                                 goto unclustered_alloc;
5608                         }
5609
5610                         /* allocate a cluster in this block group */
5611                         ret = btrfs_find_space_cluster(trans, root,
5612                                                block_group, last_ptr,
5613                                                search_start, num_bytes,
5614                                                empty_cluster + empty_size);
5615                         if (ret == 0) {
5616                                 /*
5617                                  * now pull our allocation out of this
5618                                  * cluster
5619                                  */
5620                                 offset = btrfs_alloc_from_cluster(block_group,
5621                                                   last_ptr, num_bytes,
5622                                                   search_start);
5623                                 if (offset) {
5624                                         /* we found one, proceed */
5625                                         spin_unlock(&last_ptr->refill_lock);
5626                                         trace_btrfs_reserve_extent_cluster(root,
5627                                                 block_group, search_start,
5628                                                 num_bytes);
5629                                         goto checks;
5630                                 }
5631                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5632                                    && !failed_cluster_refill) {
5633                                 spin_unlock(&last_ptr->refill_lock);
5634
5635                                 failed_cluster_refill = true;
5636                                 wait_block_group_cache_progress(block_group,
5637