Merge branch 'qgroup' of git://git.jan-o-sch.net/btrfs-unstable into for-linus
[linux-3.10.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36
37 #undef SCRAMBLE_DELAYED_REFS
38
39 /*
40  * control flags for do_chunk_alloc's force field
41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
42  * if we really need one.
43  *
44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
45  * if we have very few chunks already allocated.  This is
46  * used as part of the clustering code to help make sure
47  * we have a good pool of storage to cluster in, without
48  * filling the FS with empty chunks
49  *
50  * CHUNK_ALLOC_FORCE means it must try to allocate one
51  *
52  */
53 enum {
54         CHUNK_ALLOC_NO_FORCE = 0,
55         CHUNK_ALLOC_LIMITED = 1,
56         CHUNK_ALLOC_FORCE = 2,
57 };
58
59 /*
60  * Control how reservations are dealt with.
61  *
62  * RESERVE_FREE - freeing a reservation.
63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
64  *   ENOSPC accounting
65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
67  */
68 enum {
69         RESERVE_FREE = 0,
70         RESERVE_ALLOC = 1,
71         RESERVE_ALLOC_NO_ACCOUNT = 2,
72 };
73
74 static int update_block_group(struct btrfs_trans_handle *trans,
75                               struct btrfs_root *root,
76                               u64 bytenr, u64 num_bytes, int alloc);
77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
78                                 struct btrfs_root *root,
79                                 u64 bytenr, u64 num_bytes, u64 parent,
80                                 u64 root_objectid, u64 owner_objectid,
81                                 u64 owner_offset, int refs_to_drop,
82                                 struct btrfs_delayed_extent_op *extra_op);
83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
84                                     struct extent_buffer *leaf,
85                                     struct btrfs_extent_item *ei);
86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
87                                       struct btrfs_root *root,
88                                       u64 parent, u64 root_objectid,
89                                       u64 flags, u64 owner, u64 offset,
90                                       struct btrfs_key *ins, int ref_mod);
91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
92                                      struct btrfs_root *root,
93                                      u64 parent, u64 root_objectid,
94                                      u64 flags, struct btrfs_disk_key *key,
95                                      int level, struct btrfs_key *ins);
96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
97                           struct btrfs_root *extent_root, u64 alloc_bytes,
98                           u64 flags, int force);
99 static int find_next_key(struct btrfs_path *path, int level,
100                          struct btrfs_key *key);
101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
102                             int dump_block_groups);
103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
104                                        u64 num_bytes, int reserve);
105
106 static noinline int
107 block_group_cache_done(struct btrfs_block_group_cache *cache)
108 {
109         smp_mb();
110         return cache->cached == BTRFS_CACHE_FINISHED;
111 }
112
113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
114 {
115         return (cache->flags & bits) == bits;
116 }
117
118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
119 {
120         atomic_inc(&cache->count);
121 }
122
123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
124 {
125         if (atomic_dec_and_test(&cache->count)) {
126                 WARN_ON(cache->pinned > 0);
127                 WARN_ON(cache->reserved > 0);
128                 kfree(cache->free_space_ctl);
129                 kfree(cache);
130         }
131 }
132
133 /*
134  * this adds the block group to the fs_info rb tree for the block group
135  * cache
136  */
137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
138                                 struct btrfs_block_group_cache *block_group)
139 {
140         struct rb_node **p;
141         struct rb_node *parent = NULL;
142         struct btrfs_block_group_cache *cache;
143
144         spin_lock(&info->block_group_cache_lock);
145         p = &info->block_group_cache_tree.rb_node;
146
147         while (*p) {
148                 parent = *p;
149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
150                                  cache_node);
151                 if (block_group->key.objectid < cache->key.objectid) {
152                         p = &(*p)->rb_left;
153                 } else if (block_group->key.objectid > cache->key.objectid) {
154                         p = &(*p)->rb_right;
155                 } else {
156                         spin_unlock(&info->block_group_cache_lock);
157                         return -EEXIST;
158                 }
159         }
160
161         rb_link_node(&block_group->cache_node, parent, p);
162         rb_insert_color(&block_group->cache_node,
163                         &info->block_group_cache_tree);
164         spin_unlock(&info->block_group_cache_lock);
165
166         return 0;
167 }
168
169 /*
170  * This will return the block group at or after bytenr if contains is 0, else
171  * it will return the block group that contains the bytenr
172  */
173 static struct btrfs_block_group_cache *
174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
175                               int contains)
176 {
177         struct btrfs_block_group_cache *cache, *ret = NULL;
178         struct rb_node *n;
179         u64 end, start;
180
181         spin_lock(&info->block_group_cache_lock);
182         n = info->block_group_cache_tree.rb_node;
183
184         while (n) {
185                 cache = rb_entry(n, struct btrfs_block_group_cache,
186                                  cache_node);
187                 end = cache->key.objectid + cache->key.offset - 1;
188                 start = cache->key.objectid;
189
190                 if (bytenr < start) {
191                         if (!contains && (!ret || start < ret->key.objectid))
192                                 ret = cache;
193                         n = n->rb_left;
194                 } else if (bytenr > start) {
195                         if (contains && bytenr <= end) {
196                                 ret = cache;
197                                 break;
198                         }
199                         n = n->rb_right;
200                 } else {
201                         ret = cache;
202                         break;
203                 }
204         }
205         if (ret)
206                 btrfs_get_block_group(ret);
207         spin_unlock(&info->block_group_cache_lock);
208
209         return ret;
210 }
211
212 static int add_excluded_extent(struct btrfs_root *root,
213                                u64 start, u64 num_bytes)
214 {
215         u64 end = start + num_bytes - 1;
216         set_extent_bits(&root->fs_info->freed_extents[0],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         set_extent_bits(&root->fs_info->freed_extents[1],
219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
220         return 0;
221 }
222
223 static void free_excluded_extents(struct btrfs_root *root,
224                                   struct btrfs_block_group_cache *cache)
225 {
226         u64 start, end;
227
228         start = cache->key.objectid;
229         end = start + cache->key.offset - 1;
230
231         clear_extent_bits(&root->fs_info->freed_extents[0],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233         clear_extent_bits(&root->fs_info->freed_extents[1],
234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
235 }
236
237 static int exclude_super_stripes(struct btrfs_root *root,
238                                  struct btrfs_block_group_cache *cache)
239 {
240         u64 bytenr;
241         u64 *logical;
242         int stripe_len;
243         int i, nr, ret;
244
245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
247                 cache->bytes_super += stripe_len;
248                 ret = add_excluded_extent(root, cache->key.objectid,
249                                           stripe_len);
250                 BUG_ON(ret); /* -ENOMEM */
251         }
252
253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
254                 bytenr = btrfs_sb_offset(i);
255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
256                                        cache->key.objectid, bytenr,
257                                        0, &logical, &nr, &stripe_len);
258                 BUG_ON(ret); /* -ENOMEM */
259
260                 while (nr--) {
261                         cache->bytes_super += stripe_len;
262                         ret = add_excluded_extent(root, logical[nr],
263                                                   stripe_len);
264                         BUG_ON(ret); /* -ENOMEM */
265                 }
266
267                 kfree(logical);
268         }
269         return 0;
270 }
271
272 static struct btrfs_caching_control *
273 get_caching_control(struct btrfs_block_group_cache *cache)
274 {
275         struct btrfs_caching_control *ctl;
276
277         spin_lock(&cache->lock);
278         if (cache->cached != BTRFS_CACHE_STARTED) {
279                 spin_unlock(&cache->lock);
280                 return NULL;
281         }
282
283         /* We're loading it the fast way, so we don't have a caching_ctl. */
284         if (!cache->caching_ctl) {
285                 spin_unlock(&cache->lock);
286                 return NULL;
287         }
288
289         ctl = cache->caching_ctl;
290         atomic_inc(&ctl->count);
291         spin_unlock(&cache->lock);
292         return ctl;
293 }
294
295 static void put_caching_control(struct btrfs_caching_control *ctl)
296 {
297         if (atomic_dec_and_test(&ctl->count))
298                 kfree(ctl);
299 }
300
301 /*
302  * this is only called by cache_block_group, since we could have freed extents
303  * we need to check the pinned_extents for any extents that can't be used yet
304  * since their free space will be released as soon as the transaction commits.
305  */
306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
307                               struct btrfs_fs_info *info, u64 start, u64 end)
308 {
309         u64 extent_start, extent_end, size, total_added = 0;
310         int ret;
311
312         while (start < end) {
313                 ret = find_first_extent_bit(info->pinned_extents, start,
314                                             &extent_start, &extent_end,
315                                             EXTENT_DIRTY | EXTENT_UPTODATE);
316                 if (ret)
317                         break;
318
319                 if (extent_start <= start) {
320                         start = extent_end + 1;
321                 } else if (extent_start > start && extent_start < end) {
322                         size = extent_start - start;
323                         total_added += size;
324                         ret = btrfs_add_free_space(block_group, start,
325                                                    size);
326                         BUG_ON(ret); /* -ENOMEM or logic error */
327                         start = extent_end + 1;
328                 } else {
329                         break;
330                 }
331         }
332
333         if (start < end) {
334                 size = end - start;
335                 total_added += size;
336                 ret = btrfs_add_free_space(block_group, start, size);
337                 BUG_ON(ret); /* -ENOMEM or logic error */
338         }
339
340         return total_added;
341 }
342
343 static noinline void caching_thread(struct btrfs_work *work)
344 {
345         struct btrfs_block_group_cache *block_group;
346         struct btrfs_fs_info *fs_info;
347         struct btrfs_caching_control *caching_ctl;
348         struct btrfs_root *extent_root;
349         struct btrfs_path *path;
350         struct extent_buffer *leaf;
351         struct btrfs_key key;
352         u64 total_found = 0;
353         u64 last = 0;
354         u32 nritems;
355         int ret = 0;
356
357         caching_ctl = container_of(work, struct btrfs_caching_control, work);
358         block_group = caching_ctl->block_group;
359         fs_info = block_group->fs_info;
360         extent_root = fs_info->extent_root;
361
362         path = btrfs_alloc_path();
363         if (!path)
364                 goto out;
365
366         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
367
368         /*
369          * We don't want to deadlock with somebody trying to allocate a new
370          * extent for the extent root while also trying to search the extent
371          * root to add free space.  So we skip locking and search the commit
372          * root, since its read-only
373          */
374         path->skip_locking = 1;
375         path->search_commit_root = 1;
376         path->reada = 1;
377
378         key.objectid = last;
379         key.offset = 0;
380         key.type = BTRFS_EXTENT_ITEM_KEY;
381 again:
382         mutex_lock(&caching_ctl->mutex);
383         /* need to make sure the commit_root doesn't disappear */
384         down_read(&fs_info->extent_commit_sem);
385
386         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
387         if (ret < 0)
388                 goto err;
389
390         leaf = path->nodes[0];
391         nritems = btrfs_header_nritems(leaf);
392
393         while (1) {
394                 if (btrfs_fs_closing(fs_info) > 1) {
395                         last = (u64)-1;
396                         break;
397                 }
398
399                 if (path->slots[0] < nritems) {
400                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
401                 } else {
402                         ret = find_next_key(path, 0, &key);
403                         if (ret)
404                                 break;
405
406                         if (need_resched() ||
407                             btrfs_next_leaf(extent_root, path)) {
408                                 caching_ctl->progress = last;
409                                 btrfs_release_path(path);
410                                 up_read(&fs_info->extent_commit_sem);
411                                 mutex_unlock(&caching_ctl->mutex);
412                                 cond_resched();
413                                 goto again;
414                         }
415                         leaf = path->nodes[0];
416                         nritems = btrfs_header_nritems(leaf);
417                         continue;
418                 }
419
420                 if (key.objectid < block_group->key.objectid) {
421                         path->slots[0]++;
422                         continue;
423                 }
424
425                 if (key.objectid >= block_group->key.objectid +
426                     block_group->key.offset)
427                         break;
428
429                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
430                         total_found += add_new_free_space(block_group,
431                                                           fs_info, last,
432                                                           key.objectid);
433                         last = key.objectid + key.offset;
434
435                         if (total_found > (1024 * 1024 * 2)) {
436                                 total_found = 0;
437                                 wake_up(&caching_ctl->wait);
438                         }
439                 }
440                 path->slots[0]++;
441         }
442         ret = 0;
443
444         total_found += add_new_free_space(block_group, fs_info, last,
445                                           block_group->key.objectid +
446                                           block_group->key.offset);
447         caching_ctl->progress = (u64)-1;
448
449         spin_lock(&block_group->lock);
450         block_group->caching_ctl = NULL;
451         block_group->cached = BTRFS_CACHE_FINISHED;
452         spin_unlock(&block_group->lock);
453
454 err:
455         btrfs_free_path(path);
456         up_read(&fs_info->extent_commit_sem);
457
458         free_excluded_extents(extent_root, block_group);
459
460         mutex_unlock(&caching_ctl->mutex);
461 out:
462         wake_up(&caching_ctl->wait);
463
464         put_caching_control(caching_ctl);
465         btrfs_put_block_group(block_group);
466 }
467
468 static int cache_block_group(struct btrfs_block_group_cache *cache,
469                              struct btrfs_trans_handle *trans,
470                              struct btrfs_root *root,
471                              int load_cache_only)
472 {
473         DEFINE_WAIT(wait);
474         struct btrfs_fs_info *fs_info = cache->fs_info;
475         struct btrfs_caching_control *caching_ctl;
476         int ret = 0;
477
478         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
479         if (!caching_ctl)
480                 return -ENOMEM;
481
482         INIT_LIST_HEAD(&caching_ctl->list);
483         mutex_init(&caching_ctl->mutex);
484         init_waitqueue_head(&caching_ctl->wait);
485         caching_ctl->block_group = cache;
486         caching_ctl->progress = cache->key.objectid;
487         atomic_set(&caching_ctl->count, 1);
488         caching_ctl->work.func = caching_thread;
489
490         spin_lock(&cache->lock);
491         /*
492          * This should be a rare occasion, but this could happen I think in the
493          * case where one thread starts to load the space cache info, and then
494          * some other thread starts a transaction commit which tries to do an
495          * allocation while the other thread is still loading the space cache
496          * info.  The previous loop should have kept us from choosing this block
497          * group, but if we've moved to the state where we will wait on caching
498          * block groups we need to first check if we're doing a fast load here,
499          * so we can wait for it to finish, otherwise we could end up allocating
500          * from a block group who's cache gets evicted for one reason or
501          * another.
502          */
503         while (cache->cached == BTRFS_CACHE_FAST) {
504                 struct btrfs_caching_control *ctl;
505
506                 ctl = cache->caching_ctl;
507                 atomic_inc(&ctl->count);
508                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
509                 spin_unlock(&cache->lock);
510
511                 schedule();
512
513                 finish_wait(&ctl->wait, &wait);
514                 put_caching_control(ctl);
515                 spin_lock(&cache->lock);
516         }
517
518         if (cache->cached != BTRFS_CACHE_NO) {
519                 spin_unlock(&cache->lock);
520                 kfree(caching_ctl);
521                 return 0;
522         }
523         WARN_ON(cache->caching_ctl);
524         cache->caching_ctl = caching_ctl;
525         cache->cached = BTRFS_CACHE_FAST;
526         spin_unlock(&cache->lock);
527
528         /*
529          * We can't do the read from on-disk cache during a commit since we need
530          * to have the normal tree locking.  Also if we are currently trying to
531          * allocate blocks for the tree root we can't do the fast caching since
532          * we likely hold important locks.
533          */
534         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
535                 ret = load_free_space_cache(fs_info, cache);
536
537                 spin_lock(&cache->lock);
538                 if (ret == 1) {
539                         cache->caching_ctl = NULL;
540                         cache->cached = BTRFS_CACHE_FINISHED;
541                         cache->last_byte_to_unpin = (u64)-1;
542                 } else {
543                         if (load_cache_only) {
544                                 cache->caching_ctl = NULL;
545                                 cache->cached = BTRFS_CACHE_NO;
546                         } else {
547                                 cache->cached = BTRFS_CACHE_STARTED;
548                         }
549                 }
550                 spin_unlock(&cache->lock);
551                 wake_up(&caching_ctl->wait);
552                 if (ret == 1) {
553                         put_caching_control(caching_ctl);
554                         free_excluded_extents(fs_info->extent_root, cache);
555                         return 0;
556                 }
557         } else {
558                 /*
559                  * We are not going to do the fast caching, set cached to the
560                  * appropriate value and wakeup any waiters.
561                  */
562                 spin_lock(&cache->lock);
563                 if (load_cache_only) {
564                         cache->caching_ctl = NULL;
565                         cache->cached = BTRFS_CACHE_NO;
566                 } else {
567                         cache->cached = BTRFS_CACHE_STARTED;
568                 }
569                 spin_unlock(&cache->lock);
570                 wake_up(&caching_ctl->wait);
571         }
572
573         if (load_cache_only) {
574                 put_caching_control(caching_ctl);
575                 return 0;
576         }
577
578         down_write(&fs_info->extent_commit_sem);
579         atomic_inc(&caching_ctl->count);
580         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
581         up_write(&fs_info->extent_commit_sem);
582
583         btrfs_get_block_group(cache);
584
585         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
586
587         return ret;
588 }
589
590 /*
591  * return the block group that starts at or after bytenr
592  */
593 static struct btrfs_block_group_cache *
594 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
595 {
596         struct btrfs_block_group_cache *cache;
597
598         cache = block_group_cache_tree_search(info, bytenr, 0);
599
600         return cache;
601 }
602
603 /*
604  * return the block group that contains the given bytenr
605  */
606 struct btrfs_block_group_cache *btrfs_lookup_block_group(
607                                                  struct btrfs_fs_info *info,
608                                                  u64 bytenr)
609 {
610         struct btrfs_block_group_cache *cache;
611
612         cache = block_group_cache_tree_search(info, bytenr, 1);
613
614         return cache;
615 }
616
617 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
618                                                   u64 flags)
619 {
620         struct list_head *head = &info->space_info;
621         struct btrfs_space_info *found;
622
623         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
624
625         rcu_read_lock();
626         list_for_each_entry_rcu(found, head, list) {
627                 if (found->flags & flags) {
628                         rcu_read_unlock();
629                         return found;
630                 }
631         }
632         rcu_read_unlock();
633         return NULL;
634 }
635
636 /*
637  * after adding space to the filesystem, we need to clear the full flags
638  * on all the space infos.
639  */
640 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
641 {
642         struct list_head *head = &info->space_info;
643         struct btrfs_space_info *found;
644
645         rcu_read_lock();
646         list_for_each_entry_rcu(found, head, list)
647                 found->full = 0;
648         rcu_read_unlock();
649 }
650
651 static u64 div_factor(u64 num, int factor)
652 {
653         if (factor == 10)
654                 return num;
655         num *= factor;
656         do_div(num, 10);
657         return num;
658 }
659
660 static u64 div_factor_fine(u64 num, int factor)
661 {
662         if (factor == 100)
663                 return num;
664         num *= factor;
665         do_div(num, 100);
666         return num;
667 }
668
669 u64 btrfs_find_block_group(struct btrfs_root *root,
670                            u64 search_start, u64 search_hint, int owner)
671 {
672         struct btrfs_block_group_cache *cache;
673         u64 used;
674         u64 last = max(search_hint, search_start);
675         u64 group_start = 0;
676         int full_search = 0;
677         int factor = 9;
678         int wrapped = 0;
679 again:
680         while (1) {
681                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
682                 if (!cache)
683                         break;
684
685                 spin_lock(&cache->lock);
686                 last = cache->key.objectid + cache->key.offset;
687                 used = btrfs_block_group_used(&cache->item);
688
689                 if ((full_search || !cache->ro) &&
690                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
691                         if (used + cache->pinned + cache->reserved <
692                             div_factor(cache->key.offset, factor)) {
693                                 group_start = cache->key.objectid;
694                                 spin_unlock(&cache->lock);
695                                 btrfs_put_block_group(cache);
696                                 goto found;
697                         }
698                 }
699                 spin_unlock(&cache->lock);
700                 btrfs_put_block_group(cache);
701                 cond_resched();
702         }
703         if (!wrapped) {
704                 last = search_start;
705                 wrapped = 1;
706                 goto again;
707         }
708         if (!full_search && factor < 10) {
709                 last = search_start;
710                 full_search = 1;
711                 factor = 10;
712                 goto again;
713         }
714 found:
715         return group_start;
716 }
717
718 /* simple helper to search for an existing extent at a given offset */
719 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
720 {
721         int ret;
722         struct btrfs_key key;
723         struct btrfs_path *path;
724
725         path = btrfs_alloc_path();
726         if (!path)
727                 return -ENOMEM;
728
729         key.objectid = start;
730         key.offset = len;
731         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
732         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
733                                 0, 0);
734         btrfs_free_path(path);
735         return ret;
736 }
737
738 /*
739  * helper function to lookup reference count and flags of extent.
740  *
741  * the head node for delayed ref is used to store the sum of all the
742  * reference count modifications queued up in the rbtree. the head
743  * node may also store the extent flags to set. This way you can check
744  * to see what the reference count and extent flags would be if all of
745  * the delayed refs are not processed.
746  */
747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
748                              struct btrfs_root *root, u64 bytenr,
749                              u64 num_bytes, u64 *refs, u64 *flags)
750 {
751         struct btrfs_delayed_ref_head *head;
752         struct btrfs_delayed_ref_root *delayed_refs;
753         struct btrfs_path *path;
754         struct btrfs_extent_item *ei;
755         struct extent_buffer *leaf;
756         struct btrfs_key key;
757         u32 item_size;
758         u64 num_refs;
759         u64 extent_flags;
760         int ret;
761
762         path = btrfs_alloc_path();
763         if (!path)
764                 return -ENOMEM;
765
766         key.objectid = bytenr;
767         key.type = BTRFS_EXTENT_ITEM_KEY;
768         key.offset = num_bytes;
769         if (!trans) {
770                 path->skip_locking = 1;
771                 path->search_commit_root = 1;
772         }
773 again:
774         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
775                                 &key, path, 0, 0);
776         if (ret < 0)
777                 goto out_free;
778
779         if (ret == 0) {
780                 leaf = path->nodes[0];
781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
782                 if (item_size >= sizeof(*ei)) {
783                         ei = btrfs_item_ptr(leaf, path->slots[0],
784                                             struct btrfs_extent_item);
785                         num_refs = btrfs_extent_refs(leaf, ei);
786                         extent_flags = btrfs_extent_flags(leaf, ei);
787                 } else {
788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
789                         struct btrfs_extent_item_v0 *ei0;
790                         BUG_ON(item_size != sizeof(*ei0));
791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
792                                              struct btrfs_extent_item_v0);
793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
794                         /* FIXME: this isn't correct for data */
795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
796 #else
797                         BUG();
798 #endif
799                 }
800                 BUG_ON(num_refs == 0);
801         } else {
802                 num_refs = 0;
803                 extent_flags = 0;
804                 ret = 0;
805         }
806
807         if (!trans)
808                 goto out;
809
810         delayed_refs = &trans->transaction->delayed_refs;
811         spin_lock(&delayed_refs->lock);
812         head = btrfs_find_delayed_ref_head(trans, bytenr);
813         if (head) {
814                 if (!mutex_trylock(&head->mutex)) {
815                         atomic_inc(&head->node.refs);
816                         spin_unlock(&delayed_refs->lock);
817
818                         btrfs_release_path(path);
819
820                         /*
821                          * Mutex was contended, block until it's released and try
822                          * again
823                          */
824                         mutex_lock(&head->mutex);
825                         mutex_unlock(&head->mutex);
826                         btrfs_put_delayed_ref(&head->node);
827                         goto again;
828                 }
829                 if (head->extent_op && head->extent_op->update_flags)
830                         extent_flags |= head->extent_op->flags_to_set;
831                 else
832                         BUG_ON(num_refs == 0);
833
834                 num_refs += head->node.ref_mod;
835                 mutex_unlock(&head->mutex);
836         }
837         spin_unlock(&delayed_refs->lock);
838 out:
839         WARN_ON(num_refs == 0);
840         if (refs)
841                 *refs = num_refs;
842         if (flags)
843                 *flags = extent_flags;
844 out_free:
845         btrfs_free_path(path);
846         return ret;
847 }
848
849 /*
850  * Back reference rules.  Back refs have three main goals:
851  *
852  * 1) differentiate between all holders of references to an extent so that
853  *    when a reference is dropped we can make sure it was a valid reference
854  *    before freeing the extent.
855  *
856  * 2) Provide enough information to quickly find the holders of an extent
857  *    if we notice a given block is corrupted or bad.
858  *
859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
860  *    maintenance.  This is actually the same as #2, but with a slightly
861  *    different use case.
862  *
863  * There are two kinds of back refs. The implicit back refs is optimized
864  * for pointers in non-shared tree blocks. For a given pointer in a block,
865  * back refs of this kind provide information about the block's owner tree
866  * and the pointer's key. These information allow us to find the block by
867  * b-tree searching. The full back refs is for pointers in tree blocks not
868  * referenced by their owner trees. The location of tree block is recorded
869  * in the back refs. Actually the full back refs is generic, and can be
870  * used in all cases the implicit back refs is used. The major shortcoming
871  * of the full back refs is its overhead. Every time a tree block gets
872  * COWed, we have to update back refs entry for all pointers in it.
873  *
874  * For a newly allocated tree block, we use implicit back refs for
875  * pointers in it. This means most tree related operations only involve
876  * implicit back refs. For a tree block created in old transaction, the
877  * only way to drop a reference to it is COW it. So we can detect the
878  * event that tree block loses its owner tree's reference and do the
879  * back refs conversion.
880  *
881  * When a tree block is COW'd through a tree, there are four cases:
882  *
883  * The reference count of the block is one and the tree is the block's
884  * owner tree. Nothing to do in this case.
885  *
886  * The reference count of the block is one and the tree is not the
887  * block's owner tree. In this case, full back refs is used for pointers
888  * in the block. Remove these full back refs, add implicit back refs for
889  * every pointers in the new block.
890  *
891  * The reference count of the block is greater than one and the tree is
892  * the block's owner tree. In this case, implicit back refs is used for
893  * pointers in the block. Add full back refs for every pointers in the
894  * block, increase lower level extents' reference counts. The original
895  * implicit back refs are entailed to the new block.
896  *
897  * The reference count of the block is greater than one and the tree is
898  * not the block's owner tree. Add implicit back refs for every pointer in
899  * the new block, increase lower level extents' reference count.
900  *
901  * Back Reference Key composing:
902  *
903  * The key objectid corresponds to the first byte in the extent,
904  * The key type is used to differentiate between types of back refs.
905  * There are different meanings of the key offset for different types
906  * of back refs.
907  *
908  * File extents can be referenced by:
909  *
910  * - multiple snapshots, subvolumes, or different generations in one subvol
911  * - different files inside a single subvolume
912  * - different offsets inside a file (bookend extents in file.c)
913  *
914  * The extent ref structure for the implicit back refs has fields for:
915  *
916  * - Objectid of the subvolume root
917  * - objectid of the file holding the reference
918  * - original offset in the file
919  * - how many bookend extents
920  *
921  * The key offset for the implicit back refs is hash of the first
922  * three fields.
923  *
924  * The extent ref structure for the full back refs has field for:
925  *
926  * - number of pointers in the tree leaf
927  *
928  * The key offset for the implicit back refs is the first byte of
929  * the tree leaf
930  *
931  * When a file extent is allocated, The implicit back refs is used.
932  * the fields are filled in:
933  *
934  *     (root_key.objectid, inode objectid, offset in file, 1)
935  *
936  * When a file extent is removed file truncation, we find the
937  * corresponding implicit back refs and check the following fields:
938  *
939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
940  *
941  * Btree extents can be referenced by:
942  *
943  * - Different subvolumes
944  *
945  * Both the implicit back refs and the full back refs for tree blocks
946  * only consist of key. The key offset for the implicit back refs is
947  * objectid of block's owner tree. The key offset for the full back refs
948  * is the first byte of parent block.
949  *
950  * When implicit back refs is used, information about the lowest key and
951  * level of the tree block are required. These information are stored in
952  * tree block info structure.
953  */
954
955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
957                                   struct btrfs_root *root,
958                                   struct btrfs_path *path,
959                                   u64 owner, u32 extra_size)
960 {
961         struct btrfs_extent_item *item;
962         struct btrfs_extent_item_v0 *ei0;
963         struct btrfs_extent_ref_v0 *ref0;
964         struct btrfs_tree_block_info *bi;
965         struct extent_buffer *leaf;
966         struct btrfs_key key;
967         struct btrfs_key found_key;
968         u32 new_size = sizeof(*item);
969         u64 refs;
970         int ret;
971
972         leaf = path->nodes[0];
973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
974
975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
977                              struct btrfs_extent_item_v0);
978         refs = btrfs_extent_refs_v0(leaf, ei0);
979
980         if (owner == (u64)-1) {
981                 while (1) {
982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
983                                 ret = btrfs_next_leaf(root, path);
984                                 if (ret < 0)
985                                         return ret;
986                                 BUG_ON(ret > 0); /* Corruption */
987                                 leaf = path->nodes[0];
988                         }
989                         btrfs_item_key_to_cpu(leaf, &found_key,
990                                               path->slots[0]);
991                         BUG_ON(key.objectid != found_key.objectid);
992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
993                                 path->slots[0]++;
994                                 continue;
995                         }
996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
997                                               struct btrfs_extent_ref_v0);
998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(trans, root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466
1467         key.objectid = bytenr;
1468         key.type = BTRFS_EXTENT_ITEM_KEY;
1469         key.offset = num_bytes;
1470
1471         want = extent_ref_type(parent, owner);
1472         if (insert) {
1473                 extra_size = btrfs_extent_inline_ref_size(want);
1474                 path->keep_locks = 1;
1475         } else
1476                 extra_size = -1;
1477         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1478         if (ret < 0) {
1479                 err = ret;
1480                 goto out;
1481         }
1482         if (ret && !insert) {
1483                 err = -ENOENT;
1484                 goto out;
1485         }
1486         BUG_ON(ret); /* Corruption */
1487
1488         leaf = path->nodes[0];
1489         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1490 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1491         if (item_size < sizeof(*ei)) {
1492                 if (!insert) {
1493                         err = -ENOENT;
1494                         goto out;
1495                 }
1496                 ret = convert_extent_item_v0(trans, root, path, owner,
1497                                              extra_size);
1498                 if (ret < 0) {
1499                         err = ret;
1500                         goto out;
1501                 }
1502                 leaf = path->nodes[0];
1503                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1504         }
1505 #endif
1506         BUG_ON(item_size < sizeof(*ei));
1507
1508         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1509         flags = btrfs_extent_flags(leaf, ei);
1510
1511         ptr = (unsigned long)(ei + 1);
1512         end = (unsigned long)ei + item_size;
1513
1514         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1515                 ptr += sizeof(struct btrfs_tree_block_info);
1516                 BUG_ON(ptr > end);
1517         } else {
1518                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1519         }
1520
1521         err = -ENOENT;
1522         while (1) {
1523                 if (ptr >= end) {
1524                         WARN_ON(ptr > end);
1525                         break;
1526                 }
1527                 iref = (struct btrfs_extent_inline_ref *)ptr;
1528                 type = btrfs_extent_inline_ref_type(leaf, iref);
1529                 if (want < type)
1530                         break;
1531                 if (want > type) {
1532                         ptr += btrfs_extent_inline_ref_size(type);
1533                         continue;
1534                 }
1535
1536                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1537                         struct btrfs_extent_data_ref *dref;
1538                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1539                         if (match_extent_data_ref(leaf, dref, root_objectid,
1540                                                   owner, offset)) {
1541                                 err = 0;
1542                                 break;
1543                         }
1544                         if (hash_extent_data_ref_item(leaf, dref) <
1545                             hash_extent_data_ref(root_objectid, owner, offset))
1546                                 break;
1547                 } else {
1548                         u64 ref_offset;
1549                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1550                         if (parent > 0) {
1551                                 if (parent == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < parent)
1556                                         break;
1557                         } else {
1558                                 if (root_objectid == ref_offset) {
1559                                         err = 0;
1560                                         break;
1561                                 }
1562                                 if (ref_offset < root_objectid)
1563                                         break;
1564                         }
1565                 }
1566                 ptr += btrfs_extent_inline_ref_size(type);
1567         }
1568         if (err == -ENOENT && insert) {
1569                 if (item_size + extra_size >=
1570                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574                 /*
1575                  * To add new inline back ref, we have to make sure
1576                  * there is no corresponding back ref item.
1577                  * For simplicity, we just do not add new inline back
1578                  * ref if there is any kind of item for this block
1579                  */
1580                 if (find_next_key(path, 0, &key) == 0 &&
1581                     key.objectid == bytenr &&
1582                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1583                         err = -EAGAIN;
1584                         goto out;
1585                 }
1586         }
1587         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1588 out:
1589         if (insert) {
1590                 path->keep_locks = 0;
1591                 btrfs_unlock_up_safe(path, 1);
1592         }
1593         return err;
1594 }
1595
1596 /*
1597  * helper to add new inline back ref
1598  */
1599 static noinline_for_stack
1600 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_root *root,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref *iref,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int refs_to_add,
1606                                  struct btrfs_delayed_extent_op *extent_op)
1607 {
1608         struct extent_buffer *leaf;
1609         struct btrfs_extent_item *ei;
1610         unsigned long ptr;
1611         unsigned long end;
1612         unsigned long item_offset;
1613         u64 refs;
1614         int size;
1615         int type;
1616
1617         leaf = path->nodes[0];
1618         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1619         item_offset = (unsigned long)iref - (unsigned long)ei;
1620
1621         type = extent_ref_type(parent, owner);
1622         size = btrfs_extent_inline_ref_size(type);
1623
1624         btrfs_extend_item(trans, root, path, size);
1625
1626         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1627         refs = btrfs_extent_refs(leaf, ei);
1628         refs += refs_to_add;
1629         btrfs_set_extent_refs(leaf, ei, refs);
1630         if (extent_op)
1631                 __run_delayed_extent_op(extent_op, leaf, ei);
1632
1633         ptr = (unsigned long)ei + item_offset;
1634         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1635         if (ptr < end - size)
1636                 memmove_extent_buffer(leaf, ptr + size, ptr,
1637                                       end - size - ptr);
1638
1639         iref = (struct btrfs_extent_inline_ref *)ptr;
1640         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1641         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1642                 struct btrfs_extent_data_ref *dref;
1643                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1644                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1645                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1646                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1647                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1648         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1649                 struct btrfs_shared_data_ref *sref;
1650                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1651                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1655         } else {
1656                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1657         }
1658         btrfs_mark_buffer_dirty(leaf);
1659 }
1660
1661 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1662                                  struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref **ref_ret,
1665                                  u64 bytenr, u64 num_bytes, u64 parent,
1666                                  u64 root_objectid, u64 owner, u64 offset)
1667 {
1668         int ret;
1669
1670         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1671                                            bytenr, num_bytes, parent,
1672                                            root_objectid, owner, offset, 0);
1673         if (ret != -ENOENT)
1674                 return ret;
1675
1676         btrfs_release_path(path);
1677         *ref_ret = NULL;
1678
1679         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1680                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1681                                             root_objectid);
1682         } else {
1683                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1684                                              root_objectid, owner, offset);
1685         }
1686         return ret;
1687 }
1688
1689 /*
1690  * helper to update/remove inline back ref
1691  */
1692 static noinline_for_stack
1693 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1694                                   struct btrfs_root *root,
1695                                   struct btrfs_path *path,
1696                                   struct btrfs_extent_inline_ref *iref,
1697                                   int refs_to_mod,
1698                                   struct btrfs_delayed_extent_op *extent_op)
1699 {
1700         struct extent_buffer *leaf;
1701         struct btrfs_extent_item *ei;
1702         struct btrfs_extent_data_ref *dref = NULL;
1703         struct btrfs_shared_data_ref *sref = NULL;
1704         unsigned long ptr;
1705         unsigned long end;
1706         u32 item_size;
1707         int size;
1708         int type;
1709         u64 refs;
1710
1711         leaf = path->nodes[0];
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1715         refs += refs_to_mod;
1716         btrfs_set_extent_refs(leaf, ei, refs);
1717         if (extent_op)
1718                 __run_delayed_extent_op(extent_op, leaf, ei);
1719
1720         type = btrfs_extent_inline_ref_type(leaf, iref);
1721
1722         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1723                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1724                 refs = btrfs_extent_data_ref_count(leaf, dref);
1725         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1726                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1727                 refs = btrfs_shared_data_ref_count(leaf, sref);
1728         } else {
1729                 refs = 1;
1730                 BUG_ON(refs_to_mod != -1);
1731         }
1732
1733         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1734         refs += refs_to_mod;
1735
1736         if (refs > 0) {
1737                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1738                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1739                 else
1740                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1741         } else {
1742                 size =  btrfs_extent_inline_ref_size(type);
1743                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1744                 ptr = (unsigned long)iref;
1745                 end = (unsigned long)ei + item_size;
1746                 if (ptr + size < end)
1747                         memmove_extent_buffer(leaf, ptr, ptr + size,
1748                                               end - ptr - size);
1749                 item_size -= size;
1750                 btrfs_truncate_item(trans, root, path, item_size, 1);
1751         }
1752         btrfs_mark_buffer_dirty(leaf);
1753 }
1754
1755 static noinline_for_stack
1756 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1757                                  struct btrfs_root *root,
1758                                  struct btrfs_path *path,
1759                                  u64 bytenr, u64 num_bytes, u64 parent,
1760                                  u64 root_objectid, u64 owner,
1761                                  u64 offset, int refs_to_add,
1762                                  struct btrfs_delayed_extent_op *extent_op)
1763 {
1764         struct btrfs_extent_inline_ref *iref;
1765         int ret;
1766
1767         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1768                                            bytenr, num_bytes, parent,
1769                                            root_objectid, owner, offset, 1);
1770         if (ret == 0) {
1771                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1772                 update_inline_extent_backref(trans, root, path, iref,
1773                                              refs_to_add, extent_op);
1774         } else if (ret == -ENOENT) {
1775                 setup_inline_extent_backref(trans, root, path, iref, parent,
1776                                             root_objectid, owner, offset,
1777                                             refs_to_add, extent_op);
1778                 ret = 0;
1779         }
1780         return ret;
1781 }
1782
1783 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1784                                  struct btrfs_root *root,
1785                                  struct btrfs_path *path,
1786                                  u64 bytenr, u64 parent, u64 root_objectid,
1787                                  u64 owner, u64 offset, int refs_to_add)
1788 {
1789         int ret;
1790         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1791                 BUG_ON(refs_to_add != 1);
1792                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1793                                             parent, root_objectid);
1794         } else {
1795                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1796                                              parent, root_objectid,
1797                                              owner, offset, refs_to_add);
1798         }
1799         return ret;
1800 }
1801
1802 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  int refs_to_drop, int is_data)
1807 {
1808         int ret = 0;
1809
1810         BUG_ON(!is_data && refs_to_drop != 1);
1811         if (iref) {
1812                 update_inline_extent_backref(trans, root, path, iref,
1813                                              -refs_to_drop, NULL);
1814         } else if (is_data) {
1815                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1816         } else {
1817                 ret = btrfs_del_item(trans, root, path);
1818         }
1819         return ret;
1820 }
1821
1822 static int btrfs_issue_discard(struct block_device *bdev,
1823                                 u64 start, u64 len)
1824 {
1825         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1826 }
1827
1828 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1829                                 u64 num_bytes, u64 *actual_bytes)
1830 {
1831         int ret;
1832         u64 discarded_bytes = 0;
1833         struct btrfs_bio *bbio = NULL;
1834
1835
1836         /* Tell the block device(s) that the sectors can be discarded */
1837         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1838                               bytenr, &num_bytes, &bbio, 0);
1839         /* Error condition is -ENOMEM */
1840         if (!ret) {
1841                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1842                 int i;
1843
1844
1845                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1846                         if (!stripe->dev->can_discard)
1847                                 continue;
1848
1849                         ret = btrfs_issue_discard(stripe->dev->bdev,
1850                                                   stripe->physical,
1851                                                   stripe->length);
1852                         if (!ret)
1853                                 discarded_bytes += stripe->length;
1854                         else if (ret != -EOPNOTSUPP)
1855                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1856
1857                         /*
1858                          * Just in case we get back EOPNOTSUPP for some reason,
1859                          * just ignore the return value so we don't screw up
1860                          * people calling discard_extent.
1861                          */
1862                         ret = 0;
1863                 }
1864                 kfree(bbio);
1865         }
1866
1867         if (actual_bytes)
1868                 *actual_bytes = discarded_bytes;
1869
1870
1871         return ret;
1872 }
1873
1874 /* Can return -ENOMEM */
1875 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          u64 bytenr, u64 num_bytes, u64 parent,
1878                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1879 {
1880         int ret;
1881         struct btrfs_fs_info *fs_info = root->fs_info;
1882
1883         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1884                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1885
1886         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, (int)owner,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         } else {
1892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1893                                         num_bytes,
1894                                         parent, root_objectid, owner, offset,
1895                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1896         }
1897         return ret;
1898 }
1899
1900 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1901                                   struct btrfs_root *root,
1902                                   u64 bytenr, u64 num_bytes,
1903                                   u64 parent, u64 root_objectid,
1904                                   u64 owner, u64 offset, int refs_to_add,
1905                                   struct btrfs_delayed_extent_op *extent_op)
1906 {
1907         struct btrfs_path *path;
1908         struct extent_buffer *leaf;
1909         struct btrfs_extent_item *item;
1910         u64 refs;
1911         int ret;
1912         int err = 0;
1913
1914         path = btrfs_alloc_path();
1915         if (!path)
1916                 return -ENOMEM;
1917
1918         path->reada = 1;
1919         path->leave_spinning = 1;
1920         /* this will setup the path even if it fails to insert the back ref */
1921         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1922                                            path, bytenr, num_bytes, parent,
1923                                            root_objectid, owner, offset,
1924                                            refs_to_add, extent_op);
1925         if (ret == 0)
1926                 goto out;
1927
1928         if (ret != -EAGAIN) {
1929                 err = ret;
1930                 goto out;
1931         }
1932
1933         leaf = path->nodes[0];
1934         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1935         refs = btrfs_extent_refs(leaf, item);
1936         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1937         if (extent_op)
1938                 __run_delayed_extent_op(extent_op, leaf, item);
1939
1940         btrfs_mark_buffer_dirty(leaf);
1941         btrfs_release_path(path);
1942
1943         path->reada = 1;
1944         path->leave_spinning = 1;
1945
1946         /* now insert the actual backref */
1947         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1948                                     path, bytenr, parent, root_objectid,
1949                                     owner, offset, refs_to_add);
1950         if (ret)
1951                 btrfs_abort_transaction(trans, root, ret);
1952 out:
1953         btrfs_free_path(path);
1954         return err;
1955 }
1956
1957 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1958                                 struct btrfs_root *root,
1959                                 struct btrfs_delayed_ref_node *node,
1960                                 struct btrfs_delayed_extent_op *extent_op,
1961                                 int insert_reserved)
1962 {
1963         int ret = 0;
1964         struct btrfs_delayed_data_ref *ref;
1965         struct btrfs_key ins;
1966         u64 parent = 0;
1967         u64 ref_root = 0;
1968         u64 flags = 0;
1969
1970         ins.objectid = node->bytenr;
1971         ins.offset = node->num_bytes;
1972         ins.type = BTRFS_EXTENT_ITEM_KEY;
1973
1974         ref = btrfs_delayed_node_to_data_ref(node);
1975         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1976                 parent = ref->parent;
1977         else
1978                 ref_root = ref->root;
1979
1980         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1981                 if (extent_op) {
1982                         BUG_ON(extent_op->update_key);
1983                         flags |= extent_op->flags_to_set;
1984                 }
1985                 ret = alloc_reserved_file_extent(trans, root,
1986                                                  parent, ref_root, flags,
1987                                                  ref->objectid, ref->offset,
1988                                                  &ins, node->ref_mod);
1989         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1990                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1991                                              node->num_bytes, parent,
1992                                              ref_root, ref->objectid,
1993                                              ref->offset, node->ref_mod,
1994                                              extent_op);
1995         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1996                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1997                                           node->num_bytes, parent,
1998                                           ref_root, ref->objectid,
1999                                           ref->offset, node->ref_mod,
2000                                           extent_op);
2001         } else {
2002                 BUG();
2003         }
2004         return ret;
2005 }
2006
2007 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2008                                     struct extent_buffer *leaf,
2009                                     struct btrfs_extent_item *ei)
2010 {
2011         u64 flags = btrfs_extent_flags(leaf, ei);
2012         if (extent_op->update_flags) {
2013                 flags |= extent_op->flags_to_set;
2014                 btrfs_set_extent_flags(leaf, ei, flags);
2015         }
2016
2017         if (extent_op->update_key) {
2018                 struct btrfs_tree_block_info *bi;
2019                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2020                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2021                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2022         }
2023 }
2024
2025 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2026                                  struct btrfs_root *root,
2027                                  struct btrfs_delayed_ref_node *node,
2028                                  struct btrfs_delayed_extent_op *extent_op)
2029 {
2030         struct btrfs_key key;
2031         struct btrfs_path *path;
2032         struct btrfs_extent_item *ei;
2033         struct extent_buffer *leaf;
2034         u32 item_size;
2035         int ret;
2036         int err = 0;
2037
2038         if (trans->aborted)
2039                 return 0;
2040
2041         path = btrfs_alloc_path();
2042         if (!path)
2043                 return -ENOMEM;
2044
2045         key.objectid = node->bytenr;
2046         key.type = BTRFS_EXTENT_ITEM_KEY;
2047         key.offset = node->num_bytes;
2048
2049         path->reada = 1;
2050         path->leave_spinning = 1;
2051         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2052                                 path, 0, 1);
2053         if (ret < 0) {
2054                 err = ret;
2055                 goto out;
2056         }
2057         if (ret > 0) {
2058                 err = -EIO;
2059                 goto out;
2060         }
2061
2062         leaf = path->nodes[0];
2063         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2065         if (item_size < sizeof(*ei)) {
2066                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2067                                              path, (u64)-1, 0);
2068                 if (ret < 0) {
2069                         err = ret;
2070                         goto out;
2071                 }
2072                 leaf = path->nodes[0];
2073                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2074         }
2075 #endif
2076         BUG_ON(item_size < sizeof(*ei));
2077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         __run_delayed_extent_op(extent_op, leaf, ei);
2079
2080         btrfs_mark_buffer_dirty(leaf);
2081 out:
2082         btrfs_free_path(path);
2083         return err;
2084 }
2085
2086 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_root *root,
2088                                 struct btrfs_delayed_ref_node *node,
2089                                 struct btrfs_delayed_extent_op *extent_op,
2090                                 int insert_reserved)
2091 {
2092         int ret = 0;
2093         struct btrfs_delayed_tree_ref *ref;
2094         struct btrfs_key ins;
2095         u64 parent = 0;
2096         u64 ref_root = 0;
2097
2098         ins.objectid = node->bytenr;
2099         ins.offset = node->num_bytes;
2100         ins.type = BTRFS_EXTENT_ITEM_KEY;
2101
2102         ref = btrfs_delayed_node_to_tree_ref(node);
2103         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2104                 parent = ref->parent;
2105         else
2106                 ref_root = ref->root;
2107
2108         BUG_ON(node->ref_mod != 1);
2109         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2110                 BUG_ON(!extent_op || !extent_op->update_flags ||
2111                        !extent_op->update_key);
2112                 ret = alloc_reserved_tree_block(trans, root,
2113                                                 parent, ref_root,
2114                                                 extent_op->flags_to_set,
2115                                                 &extent_op->key,
2116                                                 ref->level, &ins);
2117         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2118                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2119                                              node->num_bytes, parent, ref_root,
2120                                              ref->level, 0, 1, extent_op);
2121         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2122                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2123                                           node->num_bytes, parent, ref_root,
2124                                           ref->level, 0, 1, extent_op);
2125         } else {
2126                 BUG();
2127         }
2128         return ret;
2129 }
2130
2131 /* helper function to actually process a single delayed ref entry */
2132 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2133                                struct btrfs_root *root,
2134                                struct btrfs_delayed_ref_node *node,
2135                                struct btrfs_delayed_extent_op *extent_op,
2136                                int insert_reserved)
2137 {
2138         int ret = 0;
2139
2140         if (trans->aborted)
2141                 return 0;
2142
2143         if (btrfs_delayed_ref_is_head(node)) {
2144                 struct btrfs_delayed_ref_head *head;
2145                 /*
2146                  * we've hit the end of the chain and we were supposed
2147                  * to insert this extent into the tree.  But, it got
2148                  * deleted before we ever needed to insert it, so all
2149                  * we have to do is clean up the accounting
2150                  */
2151                 BUG_ON(extent_op);
2152                 head = btrfs_delayed_node_to_head(node);
2153                 if (insert_reserved) {
2154                         btrfs_pin_extent(root, node->bytenr,
2155                                          node->num_bytes, 1);
2156                         if (head->is_data) {
2157                                 ret = btrfs_del_csums(trans, root,
2158                                                       node->bytenr,
2159                                                       node->num_bytes);
2160                         }
2161                 }
2162                 mutex_unlock(&head->mutex);
2163                 return ret;
2164         }
2165
2166         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2167             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2168                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2169                                            insert_reserved);
2170         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2171                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2172                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2173                                            insert_reserved);
2174         else
2175                 BUG();
2176         return ret;
2177 }
2178
2179 static noinline struct btrfs_delayed_ref_node *
2180 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2181 {
2182         struct rb_node *node;
2183         struct btrfs_delayed_ref_node *ref;
2184         int action = BTRFS_ADD_DELAYED_REF;
2185 again:
2186         /*
2187          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2188          * this prevents ref count from going down to zero when
2189          * there still are pending delayed ref.
2190          */
2191         node = rb_prev(&head->node.rb_node);
2192         while (1) {
2193                 if (!node)
2194                         break;
2195                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2196                                 rb_node);
2197                 if (ref->bytenr != head->node.bytenr)
2198                         break;
2199                 if (ref->action == action)
2200                         return ref;
2201                 node = rb_prev(node);
2202         }
2203         if (action == BTRFS_ADD_DELAYED_REF) {
2204                 action = BTRFS_DROP_DELAYED_REF;
2205                 goto again;
2206         }
2207         return NULL;
2208 }
2209
2210 /*
2211  * Returns 0 on success or if called with an already aborted transaction.
2212  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2213  */
2214 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2215                                        struct btrfs_root *root,
2216                                        struct list_head *cluster)
2217 {
2218         struct btrfs_delayed_ref_root *delayed_refs;
2219         struct btrfs_delayed_ref_node *ref;
2220         struct btrfs_delayed_ref_head *locked_ref = NULL;
2221         struct btrfs_delayed_extent_op *extent_op;
2222         struct btrfs_fs_info *fs_info = root->fs_info;
2223         int ret;
2224         int count = 0;
2225         int must_insert_reserved = 0;
2226
2227         delayed_refs = &trans->transaction->delayed_refs;
2228         while (1) {
2229                 if (!locked_ref) {
2230                         /* pick a new head ref from the cluster list */
2231                         if (list_empty(cluster))
2232                                 break;
2233
2234                         locked_ref = list_entry(cluster->next,
2235                                      struct btrfs_delayed_ref_head, cluster);
2236
2237                         /* grab the lock that says we are going to process
2238                          * all the refs for this head */
2239                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2240
2241                         /*
2242                          * we may have dropped the spin lock to get the head
2243                          * mutex lock, and that might have given someone else
2244                          * time to free the head.  If that's true, it has been
2245                          * removed from our list and we can move on.
2246                          */
2247                         if (ret == -EAGAIN) {
2248                                 locked_ref = NULL;
2249                                 count++;
2250                                 continue;
2251                         }
2252                 }
2253
2254                 /*
2255                  * locked_ref is the head node, so we have to go one
2256                  * node back for any delayed ref updates
2257                  */
2258                 ref = select_delayed_ref(locked_ref);
2259
2260                 if (ref && ref->seq &&
2261                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2262                         /*
2263                          * there are still refs with lower seq numbers in the
2264                          * process of being added. Don't run this ref yet.
2265                          */
2266                         list_del_init(&locked_ref->cluster);
2267                         mutex_unlock(&locked_ref->mutex);
2268                         locked_ref = NULL;
2269                         delayed_refs->num_heads_ready++;
2270                         spin_unlock(&delayed_refs->lock);
2271                         cond_resched();
2272                         spin_lock(&delayed_refs->lock);
2273                         continue;
2274                 }
2275
2276                 /*
2277                  * record the must insert reserved flag before we
2278                  * drop the spin lock.
2279                  */
2280                 must_insert_reserved = locked_ref->must_insert_reserved;
2281                 locked_ref->must_insert_reserved = 0;
2282
2283                 extent_op = locked_ref->extent_op;
2284                 locked_ref->extent_op = NULL;
2285
2286                 if (!ref) {
2287                         /* All delayed refs have been processed, Go ahead
2288                          * and send the head node to run_one_delayed_ref,
2289                          * so that any accounting fixes can happen
2290                          */
2291                         ref = &locked_ref->node;
2292
2293                         if (extent_op && must_insert_reserved) {
2294                                 kfree(extent_op);
2295                                 extent_op = NULL;
2296                         }
2297
2298                         if (extent_op) {
2299                                 spin_unlock(&delayed_refs->lock);
2300
2301                                 ret = run_delayed_extent_op(trans, root,
2302                                                             ref, extent_op);
2303                                 kfree(extent_op);
2304
2305                                 if (ret) {
2306                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2307                                         spin_lock(&delayed_refs->lock);
2308                                         return ret;
2309                                 }
2310
2311                                 goto next;
2312                         }
2313
2314                         list_del_init(&locked_ref->cluster);
2315                         locked_ref = NULL;
2316                 }
2317
2318                 ref->in_tree = 0;
2319                 rb_erase(&ref->rb_node, &delayed_refs->root);
2320                 delayed_refs->num_entries--;
2321                 /*
2322                  * we modified num_entries, but as we're currently running
2323                  * delayed refs, skip
2324                  *     wake_up(&delayed_refs->seq_wait);
2325                  * here.
2326                  */
2327                 spin_unlock(&delayed_refs->lock);
2328
2329                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2330                                           must_insert_reserved);
2331
2332                 btrfs_put_delayed_ref(ref);
2333                 kfree(extent_op);
2334                 count++;
2335
2336                 if (ret) {
2337                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2338                         spin_lock(&delayed_refs->lock);
2339                         return ret;
2340                 }
2341
2342 next:
2343                 do_chunk_alloc(trans, fs_info->extent_root,
2344                                2 * 1024 * 1024,
2345                                btrfs_get_alloc_profile(root, 0),
2346                                CHUNK_ALLOC_NO_FORCE);
2347                 cond_resched();
2348                 spin_lock(&delayed_refs->lock);
2349         }
2350         return count;
2351 }
2352
2353 static void wait_for_more_refs(struct btrfs_fs_info *fs_info,
2354                                struct btrfs_delayed_ref_root *delayed_refs,
2355                                unsigned long num_refs,
2356                                struct list_head *first_seq)
2357 {
2358         spin_unlock(&delayed_refs->lock);
2359         pr_debug("waiting for more refs (num %ld, first %p)\n",
2360                  num_refs, first_seq);
2361         wait_event(fs_info->tree_mod_seq_wait,
2362                    num_refs != delayed_refs->num_entries ||
2363                    fs_info->tree_mod_seq_list.next != first_seq);
2364         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2365                  delayed_refs->num_entries, fs_info->tree_mod_seq_list.next);
2366         spin_lock(&delayed_refs->lock);
2367 }
2368
2369 #ifdef SCRAMBLE_DELAYED_REFS
2370 /*
2371  * Normally delayed refs get processed in ascending bytenr order. This
2372  * correlates in most cases to the order added. To expose dependencies on this
2373  * order, we start to process the tree in the middle instead of the beginning
2374  */
2375 static u64 find_middle(struct rb_root *root)
2376 {
2377         struct rb_node *n = root->rb_node;
2378         struct btrfs_delayed_ref_node *entry;
2379         int alt = 1;
2380         u64 middle;
2381         u64 first = 0, last = 0;
2382
2383         n = rb_first(root);
2384         if (n) {
2385                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2386                 first = entry->bytenr;
2387         }
2388         n = rb_last(root);
2389         if (n) {
2390                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2391                 last = entry->bytenr;
2392         }
2393         n = root->rb_node;
2394
2395         while (n) {
2396                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2397                 WARN_ON(!entry->in_tree);
2398
2399                 middle = entry->bytenr;
2400
2401                 if (alt)
2402                         n = n->rb_left;
2403                 else
2404                         n = n->rb_right;
2405
2406                 alt = 1 - alt;
2407         }
2408         return middle;
2409 }
2410 #endif
2411
2412 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2413                                          struct btrfs_fs_info *fs_info)
2414 {
2415         struct qgroup_update *qgroup_update;
2416         int ret = 0;
2417
2418         if (list_empty(&trans->qgroup_ref_list) !=
2419             !trans->delayed_ref_elem.seq) {
2420                 /* list without seq or seq without list */
2421                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2422                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2423                         trans->delayed_ref_elem.seq);
2424                 BUG();
2425         }
2426
2427         if (!trans->delayed_ref_elem.seq)
2428                 return 0;
2429
2430         while (!list_empty(&trans->qgroup_ref_list)) {
2431                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2432                                                  struct qgroup_update, list);
2433                 list_del(&qgroup_update->list);
2434                 if (!ret)
2435                         ret = btrfs_qgroup_account_ref(
2436                                         trans, fs_info, qgroup_update->node,
2437                                         qgroup_update->extent_op);
2438                 kfree(qgroup_update);
2439         }
2440
2441         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2442
2443         return ret;
2444 }
2445
2446 /*
2447  * this starts processing the delayed reference count updates and
2448  * extent insertions we have queued up so far.  count can be
2449  * 0, which means to process everything in the tree at the start
2450  * of the run (but not newly added entries), or it can be some target
2451  * number you'd like to process.
2452  *
2453  * Returns 0 on success or if called with an aborted transaction
2454  * Returns <0 on error and aborts the transaction
2455  */
2456 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2457                            struct btrfs_root *root, unsigned long count)
2458 {
2459         struct rb_node *node;
2460         struct btrfs_delayed_ref_root *delayed_refs;
2461         struct btrfs_delayed_ref_node *ref;
2462         struct list_head cluster;
2463         struct list_head *first_seq = NULL;
2464         int ret;
2465         u64 delayed_start;
2466         int run_all = count == (unsigned long)-1;
2467         int run_most = 0;
2468         unsigned long num_refs = 0;
2469         int consider_waiting;
2470
2471         /* We'll clean this up in btrfs_cleanup_transaction */
2472         if (trans->aborted)
2473                 return 0;
2474
2475         if (root == root->fs_info->extent_root)
2476                 root = root->fs_info->tree_root;
2477
2478         do_chunk_alloc(trans, root->fs_info->extent_root,
2479                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2480                        CHUNK_ALLOC_NO_FORCE);
2481
2482         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2483
2484         delayed_refs = &trans->transaction->delayed_refs;
2485         INIT_LIST_HEAD(&cluster);
2486 again:
2487         consider_waiting = 0;
2488         spin_lock(&delayed_refs->lock);
2489
2490 #ifdef SCRAMBLE_DELAYED_REFS
2491         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2492 #endif
2493
2494         if (count == 0) {
2495                 count = delayed_refs->num_entries * 2;
2496                 run_most = 1;
2497         }
2498         while (1) {
2499                 if (!(run_all || run_most) &&
2500                     delayed_refs->num_heads_ready < 64)
2501                         break;
2502
2503                 /*
2504                  * go find something we can process in the rbtree.  We start at
2505                  * the beginning of the tree, and then build a cluster
2506                  * of refs to process starting at the first one we are able to
2507                  * lock
2508                  */
2509                 delayed_start = delayed_refs->run_delayed_start;
2510                 ret = btrfs_find_ref_cluster(trans, &cluster,
2511                                              delayed_refs->run_delayed_start);
2512                 if (ret)
2513                         break;
2514
2515                 if (delayed_start >= delayed_refs->run_delayed_start) {
2516                         if (consider_waiting == 0) {
2517                                 /*
2518                                  * btrfs_find_ref_cluster looped. let's do one
2519                                  * more cycle. if we don't run any delayed ref
2520                                  * during that cycle (because we can't because
2521                                  * all of them are blocked) and if the number of
2522                                  * refs doesn't change, we avoid busy waiting.
2523                                  */
2524                                 consider_waiting = 1;
2525                                 num_refs = delayed_refs->num_entries;
2526                                 first_seq = root->fs_info->tree_mod_seq_list.next;
2527                         } else {
2528                                 wait_for_more_refs(root->fs_info, delayed_refs,
2529                                                    num_refs, first_seq);
2530                                 /*
2531                                  * after waiting, things have changed. we
2532                                  * dropped the lock and someone else might have
2533                                  * run some refs, built new clusters and so on.
2534                                  * therefore, we restart staleness detection.
2535                                  */
2536                                 consider_waiting = 0;
2537                         }
2538                 }
2539
2540                 ret = run_clustered_refs(trans, root, &cluster);
2541                 if (ret < 0) {
2542                         spin_unlock(&delayed_refs->lock);
2543                         btrfs_abort_transaction(trans, root, ret);
2544                         return ret;
2545                 }
2546
2547                 count -= min_t(unsigned long, ret, count);
2548
2549                 if (count == 0)
2550                         break;
2551
2552                 if (ret || delayed_refs->run_delayed_start == 0) {
2553                         /* refs were run, let's reset staleness detection */
2554                         consider_waiting = 0;
2555                 }
2556         }
2557
2558         if (run_all) {
2559                 node = rb_first(&delayed_refs->root);
2560                 if (!node)
2561                         goto out;
2562                 count = (unsigned long)-1;
2563
2564                 while (node) {
2565                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2566                                        rb_node);
2567                         if (btrfs_delayed_ref_is_head(ref)) {
2568                                 struct btrfs_delayed_ref_head *head;
2569
2570                                 head = btrfs_delayed_node_to_head(ref);
2571                                 atomic_inc(&ref->refs);
2572
2573                                 spin_unlock(&delayed_refs->lock);
2574                                 /*
2575                                  * Mutex was contended, block until it's
2576                                  * released and try again
2577                                  */
2578                                 mutex_lock(&head->mutex);
2579                                 mutex_unlock(&head->mutex);
2580
2581                                 btrfs_put_delayed_ref(ref);
2582                                 cond_resched();
2583                                 goto again;
2584                         }
2585                         node = rb_next(node);
2586                 }
2587                 spin_unlock(&delayed_refs->lock);
2588                 schedule_timeout(1);
2589                 goto again;
2590         }
2591 out:
2592         spin_unlock(&delayed_refs->lock);
2593         assert_qgroups_uptodate(trans);
2594         return 0;
2595 }
2596
2597 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2598                                 struct btrfs_root *root,
2599                                 u64 bytenr, u64 num_bytes, u64 flags,
2600                                 int is_data)
2601 {
2602         struct btrfs_delayed_extent_op *extent_op;
2603         int ret;
2604
2605         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2606         if (!extent_op)
2607                 return -ENOMEM;
2608
2609         extent_op->flags_to_set = flags;
2610         extent_op->update_flags = 1;
2611         extent_op->update_key = 0;
2612         extent_op->is_data = is_data ? 1 : 0;
2613
2614         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2615                                           num_bytes, extent_op);
2616         if (ret)
2617                 kfree(extent_op);
2618         return ret;
2619 }
2620
2621 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2622                                       struct btrfs_root *root,
2623                                       struct btrfs_path *path,
2624                                       u64 objectid, u64 offset, u64 bytenr)
2625 {
2626         struct btrfs_delayed_ref_head *head;
2627         struct btrfs_delayed_ref_node *ref;
2628         struct btrfs_delayed_data_ref *data_ref;
2629         struct btrfs_delayed_ref_root *delayed_refs;
2630         struct rb_node *node;
2631         int ret = 0;
2632
2633         ret = -ENOENT;
2634         delayed_refs = &trans->transaction->delayed_refs;
2635         spin_lock(&delayed_refs->lock);
2636         head = btrfs_find_delayed_ref_head(trans, bytenr);
2637         if (!head)
2638                 goto out;
2639
2640         if (!mutex_trylock(&head->mutex)) {
2641                 atomic_inc(&head->node.refs);
2642                 spin_unlock(&delayed_refs->lock);
2643
2644                 btrfs_release_path(path);
2645
2646                 /*
2647                  * Mutex was contended, block until it's released and let
2648                  * caller try again
2649                  */
2650                 mutex_lock(&head->mutex);
2651                 mutex_unlock(&head->mutex);
2652                 btrfs_put_delayed_ref(&head->node);
2653                 return -EAGAIN;
2654         }
2655
2656         node = rb_prev(&head->node.rb_node);
2657         if (!node)
2658                 goto out_unlock;
2659
2660         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2661
2662         if (ref->bytenr != bytenr)
2663                 goto out_unlock;
2664
2665         ret = 1;
2666         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2667                 goto out_unlock;
2668
2669         data_ref = btrfs_delayed_node_to_data_ref(ref);
2670
2671         node = rb_prev(node);
2672         if (node) {
2673                 int seq = ref->seq;
2674
2675                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2676                 if (ref->bytenr == bytenr && ref->seq == seq)
2677                         goto out_unlock;
2678         }
2679
2680         if (data_ref->root != root->root_key.objectid ||
2681             data_ref->objectid != objectid || data_ref->offset != offset)
2682                 goto out_unlock;
2683
2684         ret = 0;
2685 out_unlock:
2686         mutex_unlock(&head->mutex);
2687 out:
2688         spin_unlock(&delayed_refs->lock);
2689         return ret;
2690 }
2691
2692 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2693                                         struct btrfs_root *root,
2694                                         struct btrfs_path *path,
2695                                         u64 objectid, u64 offset, u64 bytenr)
2696 {
2697         struct btrfs_root *extent_root = root->fs_info->extent_root;
2698         struct extent_buffer *leaf;
2699         struct btrfs_extent_data_ref *ref;
2700         struct btrfs_extent_inline_ref *iref;
2701         struct btrfs_extent_item *ei;
2702         struct btrfs_key key;
2703         u32 item_size;
2704         int ret;
2705
2706         key.objectid = bytenr;
2707         key.offset = (u64)-1;
2708         key.type = BTRFS_EXTENT_ITEM_KEY;
2709
2710         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2711         if (ret < 0)
2712                 goto out;
2713         BUG_ON(ret == 0); /* Corruption */
2714
2715         ret = -ENOENT;
2716         if (path->slots[0] == 0)
2717                 goto out;
2718
2719         path->slots[0]--;
2720         leaf = path->nodes[0];
2721         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2722
2723         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2724                 goto out;
2725
2726         ret = 1;
2727         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2728 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2729         if (item_size < sizeof(*ei)) {
2730                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2731                 goto out;
2732         }
2733 #endif
2734         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2735
2736         if (item_size != sizeof(*ei) +
2737             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2738                 goto out;
2739
2740         if (btrfs_extent_generation(leaf, ei) <=
2741             btrfs_root_last_snapshot(&root->root_item))
2742                 goto out;
2743
2744         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2745         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2746             BTRFS_EXTENT_DATA_REF_KEY)
2747                 goto out;
2748
2749         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2750         if (btrfs_extent_refs(leaf, ei) !=
2751             btrfs_extent_data_ref_count(leaf, ref) ||
2752             btrfs_extent_data_ref_root(leaf, ref) !=
2753             root->root_key.objectid ||
2754             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2755             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2756                 goto out;
2757
2758         ret = 0;
2759 out:
2760         return ret;
2761 }
2762
2763 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2764                           struct btrfs_root *root,
2765                           u64 objectid, u64 offset, u64 bytenr)
2766 {
2767         struct btrfs_path *path;
2768         int ret;
2769         int ret2;
2770
2771         path = btrfs_alloc_path();
2772         if (!path)
2773                 return -ENOENT;
2774
2775         do {
2776                 ret = check_committed_ref(trans, root, path, objectid,
2777                                           offset, bytenr);
2778                 if (ret && ret != -ENOENT)
2779                         goto out;
2780
2781                 ret2 = check_delayed_ref(trans, root, path, objectid,
2782                                          offset, bytenr);
2783         } while (ret2 == -EAGAIN);
2784
2785         if (ret2 && ret2 != -ENOENT) {
2786                 ret = ret2;
2787                 goto out;
2788         }
2789
2790         if (ret != -ENOENT || ret2 != -ENOENT)
2791                 ret = 0;
2792 out:
2793         btrfs_free_path(path);
2794         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2795                 WARN_ON(ret > 0);
2796         return ret;
2797 }
2798
2799 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2800                            struct btrfs_root *root,
2801                            struct extent_buffer *buf,
2802                            int full_backref, int inc, int for_cow)
2803 {
2804         u64 bytenr;
2805         u64 num_bytes;
2806         u64 parent;
2807         u64 ref_root;
2808         u32 nritems;
2809         struct btrfs_key key;
2810         struct btrfs_file_extent_item *fi;
2811         int i;
2812         int level;
2813         int ret = 0;
2814         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2815                             u64, u64, u64, u64, u64, u64, int);
2816
2817         ref_root = btrfs_header_owner(buf);
2818         nritems = btrfs_header_nritems(buf);
2819         level = btrfs_header_level(buf);
2820
2821         if (!root->ref_cows && level == 0)
2822                 return 0;
2823
2824         if (inc)
2825                 process_func = btrfs_inc_extent_ref;
2826         else
2827                 process_func = btrfs_free_extent;
2828
2829         if (full_backref)
2830                 parent = buf->start;
2831         else
2832                 parent = 0;
2833
2834         for (i = 0; i < nritems; i++) {
2835                 if (level == 0) {
2836                         btrfs_item_key_to_cpu(buf, &key, i);
2837                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2838                                 continue;
2839                         fi = btrfs_item_ptr(buf, i,
2840                                             struct btrfs_file_extent_item);
2841                         if (btrfs_file_extent_type(buf, fi) ==
2842                             BTRFS_FILE_EXTENT_INLINE)
2843                                 continue;
2844                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2845                         if (bytenr == 0)
2846                                 continue;
2847
2848                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2849                         key.offset -= btrfs_file_extent_offset(buf, fi);
2850                         ret = process_func(trans, root, bytenr, num_bytes,
2851                                            parent, ref_root, key.objectid,
2852                                            key.offset, for_cow);
2853                         if (ret)
2854                                 goto fail;
2855                 } else {
2856                         bytenr = btrfs_node_blockptr(buf, i);
2857                         num_bytes = btrfs_level_size(root, level - 1);
2858                         ret = process_func(trans, root, bytenr, num_bytes,
2859                                            parent, ref_root, level - 1, 0,
2860                                            for_cow);
2861                         if (ret)
2862                                 goto fail;
2863                 }
2864         }
2865         return 0;
2866 fail:
2867         return ret;
2868 }
2869
2870 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2871                   struct extent_buffer *buf, int full_backref, int for_cow)
2872 {
2873         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2874 }
2875
2876 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2877                   struct extent_buffer *buf, int full_backref, int for_cow)
2878 {
2879         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2880 }
2881
2882 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2883                                  struct btrfs_root *root,
2884                                  struct btrfs_path *path,
2885                                  struct btrfs_block_group_cache *cache)
2886 {
2887         int ret;
2888         struct btrfs_root *extent_root = root->fs_info->extent_root;
2889         unsigned long bi;
2890         struct extent_buffer *leaf;
2891
2892         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2893         if (ret < 0)
2894                 goto fail;
2895         BUG_ON(ret); /* Corruption */
2896
2897         leaf = path->nodes[0];
2898         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2899         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2900         btrfs_mark_buffer_dirty(leaf);
2901         btrfs_release_path(path);
2902 fail:
2903         if (ret) {
2904                 btrfs_abort_transaction(trans, root, ret);
2905                 return ret;
2906         }
2907         return 0;
2908
2909 }
2910
2911 static struct btrfs_block_group_cache *
2912 next_block_group(struct btrfs_root *root,
2913                  struct btrfs_block_group_cache *cache)
2914 {
2915         struct rb_node *node;
2916         spin_lock(&root->fs_info->block_group_cache_lock);
2917         node = rb_next(&cache->cache_node);
2918         btrfs_put_block_group(cache);
2919         if (node) {
2920                 cache = rb_entry(node, struct btrfs_block_group_cache,
2921                                  cache_node);
2922                 btrfs_get_block_group(cache);
2923         } else
2924                 cache = NULL;
2925         spin_unlock(&root->fs_info->block_group_cache_lock);
2926         return cache;
2927 }
2928
2929 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2930                             struct btrfs_trans_handle *trans,
2931                             struct btrfs_path *path)
2932 {
2933         struct btrfs_root *root = block_group->fs_info->tree_root;
2934         struct inode *inode = NULL;
2935         u64 alloc_hint = 0;
2936         int dcs = BTRFS_DC_ERROR;
2937         int num_pages = 0;
2938         int retries = 0;
2939         int ret = 0;
2940
2941         /*
2942          * If this block group is smaller than 100 megs don't bother caching the
2943          * block group.
2944          */
2945         if (block_group->key.offset < (100 * 1024 * 1024)) {
2946                 spin_lock(&block_group->lock);
2947                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2948                 spin_unlock(&block_group->lock);
2949                 return 0;
2950         }
2951
2952 again:
2953         inode = lookup_free_space_inode(root, block_group, path);
2954         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2955                 ret = PTR_ERR(inode);
2956                 btrfs_release_path(path);
2957                 goto out;
2958         }
2959
2960         if (IS_ERR(inode)) {
2961                 BUG_ON(retries);
2962                 retries++;
2963
2964                 if (block_group->ro)
2965                         goto out_free;
2966
2967                 ret = create_free_space_inode(root, trans, block_group, path);
2968                 if (ret)
2969                         goto out_free;
2970                 goto again;
2971         }
2972
2973         /* We've already setup this transaction, go ahead and exit */
2974         if (block_group->cache_generation == trans->transid &&
2975             i_size_read(inode)) {
2976                 dcs = BTRFS_DC_SETUP;
2977                 goto out_put;
2978         }
2979
2980         /*
2981          * We want to set the generation to 0, that way if anything goes wrong
2982          * from here on out we know not to trust this cache when we load up next
2983          * time.
2984          */
2985         BTRFS_I(inode)->generation = 0;
2986         ret = btrfs_update_inode(trans, root, inode);
2987         WARN_ON(ret);
2988
2989         if (i_size_read(inode) > 0) {
2990                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2991                                                       inode);
2992                 if (ret)
2993                         goto out_put;
2994         }
2995
2996         spin_lock(&block_group->lock);
2997         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2998             !btrfs_test_opt(root, SPACE_CACHE)) {
2999                 /*
3000                  * don't bother trying to write stuff out _if_
3001                  * a) we're not cached,
3002                  * b) we're with nospace_cache mount option.
3003                  */
3004                 dcs = BTRFS_DC_WRITTEN;
3005                 spin_unlock(&block_group->lock);
3006                 goto out_put;
3007         }
3008         spin_unlock(&block_group->lock);
3009
3010         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
3011         if (!num_pages)
3012                 num_pages = 1;
3013
3014         /*
3015          * Just to make absolutely sure we have enough space, we're going to
3016          * preallocate 12 pages worth of space for each block group.  In
3017          * practice we ought to use at most 8, but we need extra space so we can
3018          * add our header and have a terminator between the extents and the
3019          * bitmaps.
3020          */
3021         num_pages *= 16;
3022         num_pages *= PAGE_CACHE_SIZE;
3023
3024         ret = btrfs_check_data_free_space(inode, num_pages);
3025         if (ret)
3026                 goto out_put;
3027
3028         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3029                                               num_pages, num_pages,
3030                                               &alloc_hint);
3031         if (!ret)
3032                 dcs = BTRFS_DC_SETUP;
3033         btrfs_free_reserved_data_space(inode, num_pages);
3034
3035 out_put:
3036         iput(inode);
3037 out_free:
3038         btrfs_release_path(path);
3039 out:
3040         spin_lock(&block_group->lock);
3041         if (!ret && dcs == BTRFS_DC_SETUP)
3042                 block_group->cache_generation = trans->transid;
3043         block_group->disk_cache_state = dcs;
3044         spin_unlock(&block_group->lock);
3045
3046         return ret;
3047 }
3048
3049 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3050                                    struct btrfs_root *root)
3051 {
3052         struct btrfs_block_group_cache *cache;
3053         int err = 0;
3054         struct btrfs_path *path;
3055         u64 last = 0;
3056
3057         path = btrfs_alloc_path();
3058         if (!path)
3059                 return -ENOMEM;
3060
3061 again:
3062         while (1) {
3063                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3064                 while (cache) {
3065                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3066                                 break;
3067                         cache = next_block_group(root, cache);
3068                 }
3069                 if (!cache) {
3070                         if (last == 0)
3071                                 break;
3072                         last = 0;
3073                         continue;
3074                 }
3075                 err = cache_save_setup(cache, trans, path);
3076                 last = cache->key.objectid + cache->key.offset;
3077                 btrfs_put_block_group(cache);
3078         }
3079
3080         while (1) {
3081                 if (last == 0) {
3082                         err = btrfs_run_delayed_refs(trans, root,
3083                                                      (unsigned long)-1);
3084                         if (err) /* File system offline */
3085                                 goto out;
3086                 }
3087
3088                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3089                 while (cache) {
3090                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3091                                 btrfs_put_block_group(cache);
3092                                 goto again;
3093                         }
3094
3095                         if (cache->dirty)
3096                                 break;
3097                         cache = next_block_group(root, cache);
3098                 }
3099                 if (!cache) {
3100                         if (last == 0)
3101                                 break;
3102                         last = 0;
3103                         continue;
3104                 }
3105
3106                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3107                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3108                 cache->dirty = 0;
3109                 last = cache->key.objectid + cache->key.offset;
3110
3111                 err = write_one_cache_group(trans, root, path, cache);
3112                 if (err) /* File system offline */
3113                         goto out;
3114
3115                 btrfs_put_block_group(cache);
3116         }
3117
3118         while (1) {
3119                 /*
3120                  * I don't think this is needed since we're just marking our
3121                  * preallocated extent as written, but just in case it can't
3122                  * hurt.
3123                  */
3124                 if (last == 0) {
3125                         err = btrfs_run_delayed_refs(trans, root,
3126                                                      (unsigned long)-1);
3127                         if (err) /* File system offline */
3128                                 goto out;
3129                 }
3130
3131                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3132                 while (cache) {
3133                         /*
3134                          * Really this shouldn't happen, but it could if we
3135                          * couldn't write the entire preallocated extent and
3136                          * splitting the extent resulted in a new block.
3137                          */
3138                         if (cache->dirty) {
3139                                 btrfs_put_block_group(cache);
3140                                 goto again;
3141                         }
3142                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3143                                 break;
3144                         cache = next_block_group(root, cache);
3145                 }
3146                 if (!cache) {
3147                         if (last == 0)
3148                                 break;
3149                         last = 0;
3150                         continue;
3151                 }
3152
3153                 err = btrfs_write_out_cache(root, trans, cache, path);
3154
3155                 /*
3156                  * If we didn't have an error then the cache state is still
3157                  * NEED_WRITE, so we can set it to WRITTEN.
3158                  */
3159                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3160                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3161                 last = cache->key.objectid + cache->key.offset;
3162                 btrfs_put_block_group(cache);
3163         }
3164 out:
3165
3166         btrfs_free_path(path);
3167         return err;
3168 }
3169
3170 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3171 {
3172         struct btrfs_block_group_cache *block_group;
3173         int readonly = 0;
3174
3175         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3176         if (!block_group || block_group->ro)
3177                 readonly = 1;
3178         if (block_group)
3179                 btrfs_put_block_group(block_group);
3180         return readonly;
3181 }
3182
3183 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3184                              u64 total_bytes, u64 bytes_used,
3185                              struct btrfs_space_info **space_info)
3186 {
3187         struct btrfs_space_info *found;
3188         int i;
3189         int factor;
3190
3191         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3192                      BTRFS_BLOCK_GROUP_RAID10))
3193                 factor = 2;
3194         else
3195                 factor = 1;
3196
3197         found = __find_space_info(info, flags);
3198         if (found) {
3199                 spin_lock(&found->lock);
3200                 found->total_bytes += total_bytes;
3201                 found->disk_total += total_bytes * factor;
3202                 found->bytes_used += bytes_used;
3203                 found->disk_used += bytes_used * factor;
3204                 found->full = 0;
3205                 spin_unlock(&found->lock);
3206                 *space_info = found;
3207                 return 0;
3208         }
3209         found = kzalloc(sizeof(*found), GFP_NOFS);
3210         if (!found)
3211                 return -ENOMEM;
3212
3213         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3214                 INIT_LIST_HEAD(&found->block_groups[i]);
3215         init_rwsem(&found->groups_sem);
3216         spin_lock_init(&found->lock);
3217         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3218         found->total_bytes = total_bytes;
3219         found->disk_total = total_bytes * factor;
3220         found->bytes_used = bytes_used;
3221         found->disk_used = bytes_used * factor;
3222         found->bytes_pinned = 0;
3223         found->bytes_reserved = 0;
3224         found->bytes_readonly = 0;
3225         found->bytes_may_use = 0;
3226         found->full = 0;
3227         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3228         found->chunk_alloc = 0;
3229         found->flush = 0;
3230         init_waitqueue_head(&found->wait);
3231         *space_info = found;
3232         list_add_rcu(&found->list, &info->space_info);
3233         if (flags & BTRFS_BLOCK_GROUP_DATA)
3234                 info->data_sinfo = found;
3235         return 0;
3236 }
3237
3238 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3239 {
3240         u64 extra_flags = chunk_to_extended(flags) &
3241                                 BTRFS_EXTENDED_PROFILE_MASK;
3242
3243         if (flags & BTRFS_BLOCK_GROUP_DATA)
3244                 fs_info->avail_data_alloc_bits |= extra_flags;
3245         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3246                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3247         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3248                 fs_info->avail_system_alloc_bits |= extra_flags;
3249 }
3250
3251 /*
3252  * returns target flags in extended format or 0 if restripe for this
3253  * chunk_type is not in progress
3254  *
3255  * should be called with either volume_mutex or balance_lock held
3256  */
3257 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3258 {
3259         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3260         u64 target = 0;
3261
3262         if (!bctl)
3263                 return 0;
3264
3265         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3266             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3267                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3268         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3269                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3270                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3271         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3272                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3273                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3274         }
3275
3276         return target;
3277 }
3278
3279 /*
3280  * @flags: available profiles in extended format (see ctree.h)
3281  *
3282  * Returns reduced profile in chunk format.  If profile changing is in
3283  * progress (either running or paused) picks the target profile (if it's
3284  * already available), otherwise falls back to plain reducing.
3285  */
3286 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3287 {
3288         /*
3289          * we add in the count of missing devices because we want
3290          * to make sure that any RAID levels on a degraded FS
3291          * continue to be honored.
3292          */
3293         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3294                 root->fs_info->fs_devices->missing_devices;
3295         u64 target;
3296
3297         /*
3298          * see if restripe for this chunk_type is in progress, if so
3299          * try to reduce to the target profile
3300          */
3301         spin_lock(&root->fs_info->balance_lock);
3302         target = get_restripe_target(root->fs_info, flags);
3303         if (target) {
3304                 /* pick target profile only if it's already available */
3305                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3306                         spin_unlock(&root->fs_info->balance_lock);
3307                         return extended_to_chunk(target);
3308                 }
3309         }
3310         spin_unlock(&root->fs_info->balance_lock);
3311
3312         if (num_devices == 1)
3313                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3314         if (num_devices < 4)
3315                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3316
3317         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3318             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3319                       BTRFS_BLOCK_GROUP_RAID10))) {
3320                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3321         }
3322
3323         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3324             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3325                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3326         }
3327
3328         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3329             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3330              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3331              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3332                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3333         }
3334
3335         return extended_to_chunk(flags);
3336 }
3337
3338 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3339 {
3340         if (flags & BTRFS_BLOCK_GROUP_DATA)
3341                 flags |= root->fs_info->avail_data_alloc_bits;
3342         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3343                 flags |= root->fs_info->avail_system_alloc_bits;
3344         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3345                 flags |= root->fs_info->avail_metadata_alloc_bits;
3346
3347         return btrfs_reduce_alloc_profile(root, flags);
3348 }
3349
3350 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3351 {
3352         u64 flags;
3353
3354         if (data)
3355                 flags = BTRFS_BLOCK_GROUP_DATA;
3356         else if (root == root->fs_info->chunk_root)
3357                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3358         else
3359                 flags = BTRFS_BLOCK_GROUP_METADATA;
3360
3361         return get_alloc_profile(root, flags);
3362 }
3363
3364 /*
3365  * This will check the space that the inode allocates from to make sure we have
3366  * enough space for bytes.
3367  */
3368 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3369 {
3370         struct btrfs_space_info *data_sinfo;
3371         struct btrfs_root *root = BTRFS_I(inode)->root;
3372         struct btrfs_fs_info *fs_info = root->fs_info;
3373         u64 used;
3374         int ret = 0, committed = 0, alloc_chunk = 1;
3375
3376         /* make sure bytes are sectorsize aligned */
3377         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3378
3379         if (root == root->fs_info->tree_root ||
3380             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3381                 alloc_chunk = 0;
3382                 committed = 1;
3383         }
3384
3385         data_sinfo = fs_info->data_sinfo;
3386         if (!data_sinfo)
3387                 goto alloc;
3388
3389 again:
3390         /* make sure we have enough space to handle the data first */
3391         spin_lock(&data_sinfo->lock);
3392         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3393                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3394                 data_sinfo->bytes_may_use;
3395
3396         if (used + bytes > data_sinfo->total_bytes) {
3397                 struct btrfs_trans_handle *trans;
3398
3399                 /*
3400                  * if we don't have enough free bytes in this space then we need
3401                  * to alloc a new chunk.
3402                  */
3403                 if (!data_sinfo->full && alloc_chunk) {
3404                         u64 alloc_target;
3405
3406                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3407                         spin_unlock(&data_sinfo->lock);
3408 alloc:
3409                         alloc_target = btrfs_get_alloc_profile(root, 1);
3410                         trans = btrfs_join_transaction(root);
3411                         if (IS_ERR(trans))
3412                                 return PTR_ERR(trans);
3413
3414                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3415                                              bytes + 2 * 1024 * 1024,
3416                                              alloc_target,
3417                                              CHUNK_ALLOC_NO_FORCE);
3418                         btrfs_end_transaction(trans, root);
3419                         if (ret < 0) {
3420                                 if (ret != -ENOSPC)
3421                                         return ret;
3422                                 else
3423                                         goto commit_trans;
3424                         }
3425
3426                         if (!data_sinfo)
3427                                 data_sinfo = fs_info->data_sinfo;
3428
3429                         goto again;
3430                 }
3431
3432                 /*
3433                  * If we have less pinned bytes than we want to allocate then
3434                  * don't bother committing the transaction, it won't help us.
3435                  */
3436                 if (data_sinfo->bytes_pinned < bytes)
3437                         committed = 1;
3438                 spin_unlock(&data_sinfo->lock);
3439
3440                 /* commit the current transaction and try again */
3441 commit_trans:
3442                 if (!committed &&
3443                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3444                         committed = 1;
3445                         trans = btrfs_join_transaction(root);
3446                         if (IS_ERR(trans))
3447                                 return PTR_ERR(trans);
3448                         ret = btrfs_commit_transaction(trans, root);
3449                         if (ret)
3450                                 return ret;
3451                         goto again;
3452                 }
3453
3454                 return -ENOSPC;
3455         }
3456         data_sinfo->bytes_may_use += bytes;
3457         trace_btrfs_space_reservation(root->fs_info, "space_info",
3458                                       data_sinfo->flags, bytes, 1);
3459         spin_unlock(&data_sinfo->lock);
3460
3461         return 0;
3462 }
3463
3464 /*
3465  * Called if we need to clear a data reservation for this inode.
3466  */
3467 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3468 {
3469         struct btrfs_root *root = BTRFS_I(inode)->root;
3470         struct btrfs_space_info *data_sinfo;
3471
3472         /* make sure bytes are sectorsize aligned */
3473         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3474
3475         data_sinfo = root->fs_info->data_sinfo;
3476         spin_lock(&data_sinfo->lock);
3477         data_sinfo->bytes_may_use -= bytes;
3478         trace_btrfs_space_reservation(root->fs_info, "space_info",
3479                                       data_sinfo->flags, bytes, 0);
3480         spin_unlock(&data_sinfo->lock);
3481 }
3482
3483 static void force_metadata_allocation(struct btrfs_fs_info *info)
3484 {
3485         struct list_head *head = &info->space_info;
3486         struct btrfs_space_info *found;
3487
3488         rcu_read_lock();
3489         list_for_each_entry_rcu(found, head, list) {
3490                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3491                         found->force_alloc = CHUNK_ALLOC_FORCE;
3492         }
3493         rcu_read_unlock();
3494 }
3495
3496 static int should_alloc_chunk(struct btrfs_root *root,
3497                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3498                               int force)
3499 {
3500         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3501         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3502         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3503         u64 thresh;
3504
3505         if (force == CHUNK_ALLOC_FORCE)
3506                 return 1;
3507
3508         /*
3509          * We need to take into account the global rsv because for all intents
3510          * and purposes it's used space.  Don't worry about locking the
3511          * global_rsv, it doesn't change except when the transaction commits.
3512          */
3513         num_allocated += global_rsv->size;
3514
3515         /*
3516          * in limited mode, we want to have some free space up to
3517          * about 1% of the FS size.
3518          */
3519         if (force == CHUNK_ALLOC_LIMITED) {
3520                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3521                 thresh = max_t(u64, 64 * 1024 * 1024,
3522                                div_factor_fine(thresh, 1));
3523
3524                 if (num_bytes - num_allocated < thresh)
3525                         return 1;
3526         }
3527         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3528
3529         /* 256MB or 2% of the FS */
3530         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3531         /* system chunks need a much small threshold */
3532         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3533                 thresh = 32 * 1024 * 1024;
3534
3535         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3536                 return 0;
3537         return 1;
3538 }
3539
3540 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3541 {
3542         u64 num_dev;
3543
3544         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3545             type & BTRFS_BLOCK_GROUP_RAID0)
3546                 num_dev = root->fs_info->fs_devices->rw_devices;
3547         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3548                 num_dev = 2;
3549         else
3550                 num_dev = 1;    /* DUP or single */
3551
3552         /* metadata for updaing devices and chunk tree */
3553         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3554 }
3555
3556 static void check_system_chunk(struct btrfs_trans_handle *trans,
3557                                struct btrfs_root *root, u64 type)
3558 {
3559         struct btrfs_space_info *info;
3560         u64 left;
3561         u64 thresh;
3562
3563         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3564         spin_lock(&info->lock);
3565         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3566                 info->bytes_reserved - info->bytes_readonly;
3567         spin_unlock(&info->lock);
3568
3569         thresh = get_system_chunk_thresh(root, type);
3570         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3571                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3572                        left, thresh, type);
3573                 dump_space_info(info, 0, 0);
3574         }
3575
3576         if (left < thresh) {
3577                 u64 flags;
3578
3579                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3580                 btrfs_alloc_chunk(trans, root, flags);
3581         }
3582 }
3583
3584 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3585                           struct btrfs_root *extent_root, u64 alloc_bytes,
3586                           u64 flags, int force)
3587 {
3588         struct btrfs_space_info *space_info;
3589         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3590         int wait_for_alloc = 0;
3591         int ret = 0;
3592
3593         space_info = __find_space_info(extent_root->fs_info, flags);
3594         if (!space_info) {
3595                 ret = update_space_info(extent_root->fs_info, flags,
3596                                         0, 0, &space_info);
3597                 BUG_ON(ret); /* -ENOMEM */
3598         }
3599         BUG_ON(!space_info); /* Logic error */
3600
3601 again:
3602         spin_lock(&space_info->lock);
3603         if (force < space_info->force_alloc)
3604                 force = space_info->force_alloc;
3605         if (space_info->full) {
3606                 spin_unlock(&space_info->lock);
3607                 return 0;
3608         }
3609
3610         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3611                 spin_unlock(&space_info->lock);
3612                 return 0;
3613         } else if (space_info->chunk_alloc) {
3614                 wait_for_alloc = 1;
3615         } else {
3616                 space_info->chunk_alloc = 1;
3617         }
3618
3619         spin_unlock(&space_info->lock);
3620
3621         mutex_lock(&fs_info->chunk_mutex);
3622
3623         /*
3624          * The chunk_mutex is held throughout the entirety of a chunk
3625          * allocation, so once we've acquired the chunk_mutex we know that the
3626          * other guy is done and we need to recheck and see if we should
3627          * allocate.
3628          */
3629         if (wait_for_alloc) {
3630                 mutex_unlock(&fs_info->chunk_mutex);
3631                 wait_for_alloc = 0;
3632                 goto again;
3633         }
3634
3635         /*
3636          * If we have mixed data/metadata chunks we want to make sure we keep
3637          * allocating mixed chunks instead of individual chunks.
3638          */
3639         if (btrfs_mixed_space_info(space_info))
3640                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3641
3642         /*
3643          * if we're doing a data chunk, go ahead and make sure that
3644          * we keep a reasonable number of metadata chunks allocated in the
3645          * FS as well.
3646          */
3647         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3648                 fs_info->data_chunk_allocations++;
3649                 if (!(fs_info->data_chunk_allocations %
3650                       fs_info->metadata_ratio))
3651                         force_metadata_allocation(fs_info);
3652         }
3653
3654         /*
3655          * Check if we have enough space in SYSTEM chunk because we may need
3656          * to update devices.
3657          */
3658         check_system_chunk(trans, extent_root, flags);
3659
3660         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3661         if (ret < 0 && ret != -ENOSPC)
3662                 goto out;
3663
3664         spin_lock(&space_info->lock);
3665         if (ret)
3666                 space_info->full = 1;
3667         else
3668                 ret = 1;
3669
3670         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3671         space_info->chunk_alloc = 0;
3672         spin_unlock(&space_info->lock);
3673 out:
3674         mutex_unlock(&fs_info->chunk_mutex);
3675         return ret;
3676 }
3677
3678 /*
3679  * shrink metadata reservation for delalloc
3680  */
3681 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3682                             bool wait_ordered)
3683 {
3684         struct btrfs_block_rsv *block_rsv;
3685         struct btrfs_space_info *space_info;
3686         struct btrfs_trans_handle *trans;
3687         u64 delalloc_bytes;
3688         u64 max_reclaim;
3689         long time_left;
3690         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3691         int loops = 0;
3692
3693         trans = (struct btrfs_trans_handle *)current->journal_info;
3694         block_rsv = &root->fs_info->delalloc_block_rsv;
3695         space_info = block_rsv->space_info;
3696
3697         smp_mb();
3698         delalloc_bytes = root->fs_info->delalloc_bytes;
3699         if (delalloc_bytes == 0) {
3700                 if (trans)
3701                         return;
3702                 btrfs_wait_ordered_extents(root, 0, 0);
3703                 return;
3704         }
3705
3706         while (delalloc_bytes && loops < 3) {
3707                 max_reclaim = min(delalloc_bytes, to_reclaim);
3708                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3709                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3710                                                WB_REASON_FS_FREE_SPACE);
3711
3712                 spin_lock(&space_info->lock);
3713                 if (space_info->bytes_used + space_info->bytes_reserved +
3714                     space_info->bytes_pinned + space_info->bytes_readonly +
3715                     space_info->bytes_may_use + orig <=
3716                     space_info->total_bytes) {
3717                         spin_unlock(&space_info->lock);
3718                         break;
3719                 }
3720                 spin_unlock(&space_info->lock);
3721
3722                 loops++;
3723                 if (wait_ordered && !trans) {
3724                         btrfs_wait_ordered_extents(root, 0, 0);
3725                 } else {
3726                         time_left = schedule_timeout_killable(1);
3727                         if (time_left)
3728                                 break;
3729                 }
3730                 smp_mb();
3731                 delalloc_bytes = root->fs_info->delalloc_bytes;
3732         }
3733 }
3734
3735 /**
3736  * maybe_commit_transaction - possibly commit the transaction if its ok to
3737  * @root - the root we're allocating for
3738  * @bytes - the number of bytes we want to reserve
3739  * @force - force the commit
3740  *
3741  * This will check to make sure that committing the transaction will actually
3742  * get us somewhere and then commit the transaction if it does.  Otherwise it
3743  * will return -ENOSPC.
3744  */
3745 static int may_commit_transaction(struct btrfs_root *root,
3746                                   struct btrfs_space_info *space_info,
3747                                   u64 bytes, int force)
3748 {
3749         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3750         struct btrfs_trans_handle *trans;
3751
3752         trans = (struct btrfs_trans_handle *)current->journal_info;
3753         if (trans)
3754                 return -EAGAIN;
3755
3756         if (force)
3757                 goto commit;
3758
3759         /* See if there is enough pinned space to make this reservation */
3760         spin_lock(&space_info->lock);
3761         if (space_info->bytes_pinned >= bytes) {
3762                 spin_unlock(&space_info->lock);
3763                 goto commit;
3764         }
3765         spin_unlock(&space_info->lock);
3766
3767         /*
3768          * See if there is some space in the delayed insertion reservation for
3769          * this reservation.
3770          */
3771         if (space_info != delayed_rsv->space_info)
3772                 return -ENOSPC;
3773
3774         spin_lock(&space_info->lock);
3775         spin_lock(&delayed_rsv->lock);
3776         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3777                 spin_unlock(&delayed_rsv->lock);
3778                 spin_unlock(&space_info->lock);
3779                 return -ENOSPC;
3780         }
3781         spin_unlock(&delayed_rsv->lock);
3782         spin_unlock(&space_info->lock);
3783
3784 commit:
3785         trans = btrfs_join_transaction(root);
3786         if (IS_ERR(trans))
3787                 return -ENOSPC;
3788
3789         return btrfs_commit_transaction(trans, root);
3790 }
3791
3792 enum flush_state {
3793         FLUSH_DELALLOC          =       1,
3794         FLUSH_DELALLOC_WAIT     =       2,
3795         FLUSH_DELAYED_ITEMS_NR  =       3,
3796         FLUSH_DELAYED_ITEMS     =       4,
3797         COMMIT_TRANS            =       5,
3798 };
3799
3800 static int flush_space(struct btrfs_root *root,
3801                        struct btrfs_space_info *space_info, u64 num_bytes,
3802                        u64 orig_bytes, int state)
3803 {
3804         struct btrfs_trans_handle *trans;
3805         int nr;
3806         int ret = 0;
3807
3808         switch (state) {
3809         case FLUSH_DELALLOC:
3810         case FLUSH_DELALLOC_WAIT:
3811                 shrink_delalloc(root, num_bytes, orig_bytes,
3812                                 state == FLUSH_DELALLOC_WAIT);
3813                 break;
3814         case FLUSH_DELAYED_ITEMS_NR:
3815         case FLUSH_DELAYED_ITEMS:
3816                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3817                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3818
3819                         nr = (int)div64_u64(num_bytes, bytes);
3820                         if (!nr)
3821                                 nr = 1;
3822                         nr *= 2;
3823                 } else {
3824                         nr = -1;
3825                 }
3826                 trans = btrfs_join_transaction(root);
3827                 if (IS_ERR(trans)) {
3828                         ret = PTR_ERR(trans);
3829                         break;
3830                 }
3831                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3832                 btrfs_end_transaction(trans, root);
3833                 break;
3834         case COMMIT_TRANS:
3835                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3836                 break;
3837         default:
3838                 ret = -ENOSPC;
3839                 break;
3840         }
3841
3842         return ret;
3843 }
3844 /**
3845  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3846  * @root - the root we're allocating for
3847  * @block_rsv - the block_rsv we're allocating for
3848  * @orig_bytes - the number of bytes we want
3849  * @flush - wether or not we can flush to make our reservation
3850  *
3851  * This will reserve orgi_bytes number of bytes from the space info associated
3852  * with the block_rsv.  If there is not enough space it will make an attempt to
3853  * flush out space to make room.  It will do this by flushing delalloc if
3854  * possible or committing the transaction.  If flush is 0 then no attempts to
3855  * regain reservations will be made and this will fail if there is not enough
3856  * space already.
3857  */
3858 static int reserve_metadata_bytes(struct btrfs_root *root,
3859                                   struct btrfs_block_rsv *block_rsv,
3860                                   u64 orig_bytes, int flush)
3861 {
3862         struct btrfs_space_info *space_info = block_rsv->space_info;
3863         u64 used;
3864         u64 num_bytes = orig_bytes;
3865         int flush_state = FLUSH_DELALLOC;
3866         int ret = 0;
3867         bool flushing = false;
3868         bool committed = false;
3869
3870 again:
3871         ret = 0;
3872         spin_lock(&space_info->lock);
3873         /*
3874          * We only want to wait if somebody other than us is flushing and we are
3875          * actually alloed to flush.
3876          */
3877         while (flush && !flushing && space_info->flush) {
3878                 spin_unlock(&space_info->lock);
3879                 /*
3880                  * If we have a trans handle we can't wait because the flusher
3881                  * may have to commit the transaction, which would mean we would
3882                  * deadlock since we are waiting for the flusher to finish, but
3883                  * hold the current transaction open.
3884                  */
3885                 if (current->journal_info)
3886                         return -EAGAIN;
3887                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3888                 /* Must have been killed, return */
3889                 if (ret)
3890                         return -EINTR;
3891
3892                 spin_lock(&space_info->lock);
3893         }
3894
3895         ret = -ENOSPC;
3896         used = space_info->bytes_used + space_info->bytes_reserved +
3897                 space_info->bytes_pinned + space_info->bytes_readonly +
3898                 space_info->bytes_may_use;
3899
3900         /*
3901          * The idea here is that we've not already over-reserved the block group
3902          * then we can go ahead and save our reservation first and then start
3903          * flushing if we need to.  Otherwise if we've already overcommitted
3904          * lets start flushing stuff first and then come back and try to make
3905          * our reservation.
3906          */
3907         if (used <= space_info->total_bytes) {
3908                 if (used + orig_bytes <= space_info->total_bytes) {
3909                         space_info->bytes_may_use += orig_bytes;
3910                         trace_btrfs_space_reservation(root->fs_info,
3911                                 "space_info", space_info->flags, orig_bytes, 1);
3912                         ret = 0;
3913                 } else {
3914                         /*
3915                          * Ok set num_bytes to orig_bytes since we aren't
3916                          * overocmmitted, this way we only try and reclaim what
3917                          * we need.
3918                          */
3919                         num_bytes = orig_bytes;
3920                 }
3921         } else {
3922                 /*
3923                  * Ok we're over committed, set num_bytes to the overcommitted
3924                  * amount plus the amount of bytes that we need for this
3925                  * reservation.
3926                  */
3927                 num_bytes = used - space_info->total_bytes +
3928                         (orig_bytes * 2);
3929         }
3930
3931         if (ret) {
3932                 u64 profile = btrfs_get_alloc_profile(root, 0);
3933                 u64 avail;
3934
3935                 /*
3936                  * If we have a lot of space that's pinned, don't bother doing
3937                  * the overcommit dance yet and just commit the transaction.
3938                  */
3939                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3940                 do_div(avail, 10);
3941                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3942                         space_info->flush = 1;
3943                         flushing = true;
3944                         spin_unlock(&space_info->lock);
3945                         ret = may_commit_transaction(root, space_info,
3946                                                      orig_bytes, 1);
3947                         if (ret)
3948                                 goto out;
3949                         committed = true;
3950                         goto again;
3951                 }
3952
3953                 spin_lock(&root->fs_info->free_chunk_lock);
3954                 avail = root->fs_info->free_chunk_space;
3955
3956                 /*
3957                  * If we have dup, raid1 or raid10 then only half of the free
3958                  * space is actually useable.
3959                  */
3960                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3961                                BTRFS_BLOCK_GROUP_RAID1 |
3962                                BTRFS_BLOCK_GROUP_RAID10))
3963                         avail >>= 1;
3964
3965                 /*
3966                  * If we aren't flushing don't let us overcommit too much, say
3967                  * 1/8th of the space.  If we can flush, let it overcommit up to
3968                  * 1/2 of the space.
3969                  */
3970                 if (flush)
3971                         avail >>= 3;
3972                 else
3973                         avail >>= 1;
3974                  spin_unlock(&root->fs_info->free_chunk_lock);
3975
3976                 if (used + num_bytes < space_info->total_bytes + avail) {
3977                         space_info->bytes_may_use += orig_bytes;
3978                         trace_btrfs_space_reservation(root->fs_info,
3979                                 "space_info", space_info->flags, orig_bytes, 1);
3980                         ret = 0;
3981                 }
3982         }
3983
3984         /*
3985          * Couldn't make our reservation, save our place so while we're trying
3986          * to reclaim space we can actually use it instead of somebody else
3987          * stealing it from us.
3988          */
3989         if (ret && flush) {
3990                 flushing = true;
3991                 space_info->flush = 1;
3992         }
3993
3994         spin_unlock(&space_info->lock);
3995
3996         if (!ret || !flush)
3997                 goto out;
3998
3999         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4000                           flush_state);
4001         flush_state++;
4002         if (!ret)
4003                 goto again;
4004         else if (flush_state <= COMMIT_TRANS)
4005                 goto again;
4006
4007 out:
4008         if (flushing) {
4009                 spin_lock(&space_info->lock);
4010                 space_info->flush = 0;
4011                 wake_up_all(&space_info->wait);
4012                 spin_unlock(&space_info->lock);
4013         }
4014         return ret;
4015 }
4016
4017 static struct btrfs_block_rsv *get_block_rsv(
4018                                         const struct btrfs_trans_handle *trans,
4019                                         const struct btrfs_root *root)
4020 {
4021         struct btrfs_block_rsv *block_rsv = NULL;
4022
4023         if (root->ref_cows)
4024                 block_rsv = trans->block_rsv;
4025
4026         if (root == root->fs_info->csum_root && trans->adding_csums)
4027                 block_rsv = trans->block_rsv;
4028
4029         if (!block_rsv)
4030                 block_rsv = root->block_rsv;
4031
4032         if (!block_rsv)
4033                 block_rsv = &root->fs_info->empty_block_rsv;
4034
4035         return block_rsv;
4036 }
4037
4038 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4039                                u64 num_bytes)
4040 {
4041         int ret = -ENOSPC;
4042         spin_lock(&block_rsv->lock);
4043         if (block_rsv->reserved >= num_bytes) {
4044                 block_rsv->reserved -= num_bytes;
4045                 if (block_rsv->reserved < block_rsv->size)
4046                         block_rsv->full = 0;
4047                 ret = 0;
4048         }
4049         spin_unlock(&block_rsv->lock);
4050         return ret;
4051 }
4052
4053 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4054                                 u64 num_bytes, int update_size)
4055 {
4056         spin_lock(&block_rsv->lock);
4057         block_rsv->reserved += num_bytes;
4058         if (update_size)
4059                 block_rsv->size += num_bytes;
4060         else if (block_rsv->reserved >= block_rsv->size)
4061                 block_rsv->full = 1;
4062         spin_unlock(&block_rsv->lock);
4063 }
4064
4065 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4066                                     struct btrfs_block_rsv *block_rsv,
4067                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4068 {
4069         struct btrfs_space_info *space_info = block_rsv->space_info;
4070
4071         spin_lock(&block_rsv->lock);
4072         if (num_bytes == (u64)-1)
4073                 num_bytes = block_rsv->size;
4074         block_rsv->size -= num_bytes;
4075         if (block_rsv->reserved >= block_rsv->size) {
4076                 num_bytes = block_rsv->reserved - block_rsv->size;
4077                 block_rsv->reserved = block_rsv->size;
4078                 block_rsv->full = 1;
4079         } else {
4080                 num_bytes = 0;
4081         }
4082         spin_unlock(&block_rsv->lock);
4083
4084         if (num_bytes > 0) {
4085                 if (dest) {
4086                         spin_lock(&dest->lock);
4087                         if (!dest->full) {
4088                                 u64 bytes_to_add;
4089
4090                                 bytes_to_add = dest->size - dest->reserved;
4091                                 bytes_to_add = min(num_bytes, bytes_to_add);
4092                                 dest->reserved += bytes_to_add;
4093                                 if (dest->reserved >= dest->size)
4094                                         dest->full = 1;
4095                                 num_bytes -= bytes_to_add;
4096                         }
4097                         spin_unlock(&dest->lock);
4098                 }
4099                 if (num_bytes) {
4100                         spin_lock(&space_info->lock);
4101                         space_info->bytes_may_use -= num_bytes;
4102                         trace_btrfs_space_reservation(fs_info, "space_info",
4103                                         space_info->flags, num_bytes, 0);
4104                         space_info->reservation_progress++;
4105                         spin_unlock(&space_info->lock);
4106                 }
4107         }
4108 }
4109
4110 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4111                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4112 {
4113         int ret;
4114
4115         ret = block_rsv_use_bytes(src, num_bytes);
4116         if (ret)
4117                 return ret;
4118
4119         block_rsv_add_bytes(dst, num_bytes, 1);
4120         return 0;
4121 }
4122
4123 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4124 {
4125         memset(rsv, 0, sizeof(*rsv));
4126         spin_lock_init(&rsv->lock);
4127 }
4128
4129 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4130 {
4131         struct btrfs_block_rsv *block_rsv;
4132         struct btrfs_fs_info *fs_info = root->fs_info;
4133
4134         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4135         if (!block_rsv)
4136                 return NULL;
4137
4138         btrfs_init_block_rsv(block_rsv);
4139         block_rsv->space_info = __find_space_info(fs_info,
4140                                                   BTRFS_BLOCK_GROUP_METADATA);
4141         return block_rsv;
4142 }
4143
4144 void btrfs_free_block_rsv(struct btrfs_root *root,
4145                           struct btrfs_block_rsv *rsv)
4146 {
4147         btrfs_block_rsv_release(root, rsv, (u64)-1);
4148         kfree(rsv);
4149 }
4150
4151 static inline int __block_rsv_add(struct btrfs_root *root,
4152                                   struct btrfs_block_rsv *block_rsv,
4153                                   u64 num_bytes, int flush)
4154 {
4155         int ret;
4156
4157         if (num_bytes == 0)
4158                 return 0;
4159
4160         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4161         if (!ret) {
4162                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4163                 return 0;
4164         }
4165
4166         return ret;
4167 }
4168
4169 int btrfs_block_rsv_add(struct btrfs_root *root,
4170                         struct btrfs_block_rsv *block_rsv,
4171                         u64 num_bytes)
4172 {
4173         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4174 }
4175
4176 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4177                                 struct btrfs_block_rsv *block_rsv,
4178                                 u64 num_bytes)
4179 {
4180         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4181 }
4182
4183 int btrfs_block_rsv_check(struct btrfs_root *root,
4184                           struct btrfs_block_rsv *block_rsv, int min_factor)
4185 {
4186         u64 num_bytes = 0;
4187         int ret = -ENOSPC;
4188
4189         if (!block_rsv)
4190                 return 0;
4191
4192         spin_lock(&block_rsv->lock);
4193         num_bytes = div_factor(block_rsv->size, min_factor);
4194         if (block_rsv->reserved >= num_bytes)
4195                 ret = 0;
4196         spin_unlock(&block_rsv->lock);
4197
4198         return ret;
4199 }
4200
4201 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4202                                            struct btrfs_block_rsv *block_rsv,
4203                                            u64 min_reserved, int flush)
4204 {
4205         u64 num_bytes = 0;
4206         int ret = -ENOSPC;
4207
4208         if (!block_rsv)
4209                 return 0;
4210
4211         spin_lock(&block_rsv->lock);
4212         num_bytes = min_reserved;
4213         if (block_rsv->reserved >= num_bytes)
4214                 ret = 0;
4215         else
4216                 num_bytes -= block_rsv->reserved;
4217         spin_unlock(&block_rsv->lock);
4218
4219         if (!ret)
4220                 return 0;
4221
4222         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4223         if (!ret) {
4224                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4225                 return 0;
4226         }
4227
4228         return ret;
4229 }
4230
4231 int btrfs_block_rsv_refill(struct btrfs_root *root,
4232                            struct btrfs_block_rsv *block_rsv,
4233                            u64 min_reserved)
4234 {
4235         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4236 }
4237
4238 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4239                                    struct btrfs_block_rsv *block_rsv,
4240                                    u64 min_reserved)
4241 {
4242         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4243 }
4244
4245 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4246                             struct btrfs_block_rsv *dst_rsv,
4247                             u64 num_bytes)
4248 {
4249         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4250 }
4251
4252 void btrfs_block_rsv_release(struct btrfs_root *root,
4253                              struct btrfs_block_rsv *block_rsv,
4254                              u64 num_bytes)
4255 {
4256         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4257         if (global_rsv->full || global_rsv == block_rsv ||
4258             block_rsv->space_info != global_rsv->space_info)
4259                 global_rsv = NULL;
4260         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4261                                 num_bytes);
4262 }
4263
4264 /*
4265  * helper to calculate size of global block reservation.
4266  * the desired value is sum of space used by extent tree,
4267  * checksum tree and root tree
4268  */
4269 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4270 {
4271         struct btrfs_space_info *sinfo;
4272         u64 num_bytes;
4273         u64 meta_used;
4274         u64 data_used;
4275         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4276
4277         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4278         spin_lock(&sinfo->lock);
4279         data_used = sinfo->bytes_used;
4280         spin_unlock(&sinfo->lock);
4281
4282         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4283         spin_lock(&sinfo->lock);
4284         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4285                 data_used = 0;
4286         meta_used = sinfo->bytes_used;
4287         spin_unlock(&sinfo->lock);
4288
4289         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4290                     csum_size * 2;
4291         num_bytes += div64_u64(data_used + meta_used, 50);
4292
4293         if (num_bytes * 3 > meta_used)
4294                 num_bytes = div64_u64(meta_used, 3);
4295
4296         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4297 }
4298
4299 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4300 {
4301         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4302         struct btrfs_space_info *sinfo = block_rsv->space_info;
4303         u64 num_bytes;
4304
4305         num_bytes = calc_global_metadata_size(fs_info);
4306
4307         spin_lock(&sinfo->lock);
4308         spin_lock(&block_rsv->lock);
4309
4310         block_rsv->size = num_bytes;
4311
4312         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4313                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4314                     sinfo->bytes_may_use;
4315
4316         if (sinfo->total_bytes > num_bytes) {
4317                 num_bytes = sinfo->total_bytes - num_bytes;
4318                 block_rsv->reserved += num_bytes;
4319                 sinfo->bytes_may_use += num_bytes;
4320                 trace_btrfs_space_reservation(fs_info, "space_info",
4321                                       sinfo->flags, num_bytes, 1);
4322         }
4323
4324         if (block_rsv->reserved >= block_rsv->size) {
4325                 num_bytes = block_rsv->reserved - block_rsv->size;
4326                 sinfo->bytes_may_use -= num_bytes;
4327                 trace_btrfs_space_reservation(fs_info, "space_info",
4328                                       sinfo->flags, num_bytes, 0);
4329                 sinfo->reservation_progress++;
4330                 block_rsv->reserved = block_rsv->size;
4331                 block_rsv->full = 1;
4332         }
4333
4334         spin_unlock(&block_rsv->lock);
4335         spin_unlock(&sinfo->lock);
4336 }
4337
4338 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4339 {
4340         struct btrfs_space_info *space_info;
4341
4342         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4343         fs_info->chunk_block_rsv.space_info = space_info;
4344
4345         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4346         fs_info->global_block_rsv.space_info = space_info;
4347         fs_info->delalloc_block_rsv.space_info = space_info;
4348         fs_info->trans_block_rsv.space_info = space_info;
4349         fs_info->empty_block_rsv.space_info = space_info;
4350         fs_info->delayed_block_rsv.space_info = space_info;
4351
4352         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4353         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4354         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4355         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4356         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4357
4358         update_global_block_rsv(fs_info);
4359 }
4360
4361 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4362 {
4363         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4364                                 (u64)-1);
4365         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4366         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4367         WARN_ON(fs_info->trans_block_rsv.size > 0);
4368         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4369         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4370         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4371         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4372         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4373 }
4374
4375 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4376                                   struct btrfs_root *root)
4377 {
4378         if (!trans->block_rsv)
4379                 return;
4380
4381         if (!trans->bytes_reserved)
4382                 return;
4383
4384         trace_btrfs_space_reservation(root->fs_info, "transaction",
4385                                       trans->transid, trans->bytes_reserved, 0);
4386         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4387         trans->bytes_reserved = 0;
4388 }
4389
4390 /* Can only return 0 or -ENOSPC */
4391 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4392                                   struct inode *inode)
4393 {
4394         struct btrfs_root *root = BTRFS_I(inode)->root;
4395         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4396         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4397
4398         /*
4399          * We need to hold space in order to delete our orphan item once we've
4400          * added it, so this takes the reservation so we can release it later
4401          * when we are truly done with the orphan item.
4402          */
4403         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4404         trace_btrfs_space_reservation(root->fs_info, "orphan",
4405                                       btrfs_ino(inode), num_bytes, 1);
4406         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4407 }
4408
4409 void btrfs_orphan_release_metadata(struct inode *inode)
4410 {
4411         struct btrfs_root *root = BTRFS_I(inode)->root;
4412         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4413         trace_btrfs_space_reservation(root->fs_info, "orphan",
4414                                       btrfs_ino(inode), num_bytes, 0);
4415         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4416 }
4417
4418 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4419                                 struct btrfs_pending_snapshot *pending)
4420 {
4421         struct btrfs_root *root = pending->root;
4422         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4423         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4424         /*
4425          * two for root back/forward refs, two for directory entries
4426          * and one for root of the snapshot.
4427          */
4428         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4429         dst_rsv->space_info = src_rsv->space_info;
4430         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4431 }
4432
4433 /**
4434  * drop_outstanding_extent - drop an outstanding extent
4435  * @inode: the inode we're dropping the extent for
4436  *
4437  * This is called when we are freeing up an outstanding extent, either called
4438  * after an error or after an extent is written.  This will return the number of
4439  * reserved extents that need to be freed.  This must be called with
4440  * BTRFS_I(inode)->lock held.
4441  */
4442 static unsigned drop_outstanding_extent(struct inode *inode)
4443 {
4444         unsigned drop_inode_space = 0;
4445         unsigned dropped_extents = 0;
4446
4447         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4448         BTRFS_I(inode)->outstanding_extents--;
4449
4450         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4451             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4452                                &BTRFS_I(inode)->runtime_flags))
4453                 drop_inode_space = 1;
4454
4455         /*
4456          * If we have more or the same amount of outsanding extents than we have
4457          * reserved then we need to leave the reserved extents count alone.
4458          */
4459         if (BTRFS_I(inode)->outstanding_extents >=
4460             BTRFS_I(inode)->reserved_extents)
4461                 return drop_inode_space;
4462
4463         dropped_extents = BTRFS_I(inode)->reserved_extents -
4464                 BTRFS_I(inode)->outstanding_extents;
4465         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4466         return dropped_extents + drop_inode_space;
4467 }
4468
4469 /**
4470  * calc_csum_metadata_size - return the amount of metada space that must be
4471  *      reserved/free'd for the given bytes.
4472  * @inode: the inode we're manipulating
4473  * @num_bytes: the number of bytes in question
4474  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4475  *
4476  * This adjusts the number of csum_bytes in the inode and then returns the
4477  * correct amount of metadata that must either be reserved or freed.  We
4478  * calculate how many checksums we can fit into one leaf and then divide the
4479  * number of bytes that will need to be checksumed by this value to figure out
4480  * how many checksums will be required.  If we are adding bytes then the number
4481  * may go up and we will return the number of additional bytes that must be
4482  * reserved.  If it is going down we will return the number of bytes that must
4483  * be freed.
4484  *
4485  * This must be called with BTRFS_I(inode)->lock held.
4486  */
4487 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4488                                    int reserve)
4489 {
4490         struct btrfs_root *root = BTRFS_I(inode)->root;
4491         u64 csum_size;
4492         int num_csums_per_leaf;
4493         int num_csums;
4494         int old_csums;
4495
4496         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4497             BTRFS_I(inode)->csum_bytes == 0)
4498                 return 0;
4499
4500         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4501         if (reserve)
4502                 BTRFS_I(inode)->csum_bytes += num_bytes;
4503         else
4504                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4505         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4506         num_csums_per_leaf = (int)div64_u64(csum_size,
4507                                             sizeof(struct btrfs_csum_item) +
4508                                             sizeof(struct btrfs_disk_key));
4509         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4510         num_csums = num_csums + num_csums_per_leaf - 1;
4511         num_csums = num_csums / num_csums_per_leaf;
4512
4513         old_csums = old_csums + num_csums_per_leaf - 1;
4514         old_csums = old_csums / num_csums_per_leaf;
4515
4516         /* No change, no need to reserve more */
4517         if (old_csums == num_csums)
4518                 return 0;
4519
4520         if (reserve)
4521                 return btrfs_calc_trans_metadata_size(root,
4522                                                       num_csums - old_csums);
4523
4524         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4525 }
4526
4527 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4528 {
4529         struct btrfs_root *root = BTRFS_I(inode)->root;
4530         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4531         u64 to_reserve = 0;
4532         u64 csum_bytes;
4533         unsigned nr_extents = 0;
4534         int extra_reserve = 0;
4535         int flush = 1;
4536         int ret;
4537
4538         /* Need to be holding the i_mutex here if we aren't free space cache */
4539         if (btrfs_is_free_space_inode(inode))
4540                 flush = 0;
4541
4542         if (flush && btrfs_transaction_in_commit(root->fs_info))
4543                 schedule_timeout(1);
4544
4545         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4546         num_bytes = ALIGN(num_bytes, root->sectorsize);
4547
4548         spin_lock(&BTRFS_I(inode)->lock);
4549         BTRFS_I(inode)->outstanding_extents++;
4550
4551         if (BTRFS_I(inode)->outstanding_extents >
4552             BTRFS_I(inode)->reserved_extents)
4553                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4554                         BTRFS_I(inode)->reserved_extents;
4555
4556         /*
4557          * Add an item to reserve for updating the inode when we complete the
4558          * delalloc io.
4559          */
4560         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4561                       &BTRFS_I(inode)->runtime_flags)) {
4562                 nr_extents++;
4563                 extra_reserve = 1;
4564         }
4565
4566         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4567         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4568         csum_bytes = BTRFS_I(inode)->csum_bytes;
4569         spin_unlock(&BTRFS_I(inode)->lock);
4570
4571         if (root->fs_info->quota_enabled) {
4572                 ret = btrfs_qgroup_reserve(root, num_bytes +
4573                                            nr_extents * root->leafsize);
4574                 if (ret)
4575                         return ret;
4576         }
4577
4578         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4579         if (ret) {
4580                 u64 to_free = 0;
4581                 unsigned dropped;
4582
4583                 spin_lock(&BTRFS_I(inode)->lock);
4584                 dropped = drop_outstanding_extent(inode);
4585                 /*
4586                  * If the inodes csum_bytes is the same as the original
4587                  * csum_bytes then we know we haven't raced with any free()ers
4588                  * so we can just reduce our inodes csum bytes and carry on.
4589                  * Otherwise we have to do the normal free thing to account for
4590                  * the case that the free side didn't free up its reserve
4591                  * because of this outstanding reservation.
4592                  */
4593                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4594                         calc_csum_metadata_size(inode, num_bytes, 0);
4595                 else
4596                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4597                 spin_unlock(&BTRFS_I(inode)->lock);
4598                 if (dropped)
4599                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4600
4601                 if (to_free) {
4602                         btrfs_block_rsv_release(root, block_rsv, to_free);
4603                         trace_btrfs_space_reservation(root->fs_info,
4604                                                       "delalloc",
4605                                                       btrfs_ino(inode),
4606                                                       to_free, 0);
4607                 }
4608                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4609                 return ret;
4610         }
4611
4612         spin_lock(&BTRFS_I(inode)->lock);
4613         if (extra_reserve) {
4614                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4615                         &BTRFS_I(inode)->runtime_flags);
4616                 nr_extents--;
4617         }
4618         BTRFS_I(inode)->reserved_extents += nr_extents;
4619         spin_unlock(&BTRFS_I(inode)->lock);
4620         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4621
4622         if (to_reserve)
4623                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4624                                               btrfs_ino(inode), to_reserve, 1);
4625         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4626
4627         return 0;
4628 }
4629
4630 /**
4631  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4632  * @inode: the inode to release the reservation for
4633  * @num_bytes: the number of bytes we're releasing
4634  *
4635  * This will release the metadata reservation for an inode.  This can be called
4636  * once we complete IO for a given set of bytes to release their metadata
4637  * reservations.
4638  */
4639 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4640 {
4641         struct btrfs_root *root = BTRFS_I(inode)->root;
4642         u64 to_free = 0;
4643         unsigned dropped;
4644
4645         num_bytes = ALIGN(num_bytes, root->sectorsize);
4646         spin_lock(&BTRFS_I(inode)->lock);
4647         dropped = drop_outstanding_extent(inode);
4648
4649         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4650         spin_unlock(&BTRFS_I(inode)->lock);
4651         if (dropped > 0)
4652                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4653
4654         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4655                                       btrfs_ino(inode), to_free, 0);
4656         if (root->fs_info->quota_enabled) {
4657                 btrfs_qgroup_free(root, num_bytes +
4658                                         dropped * root->leafsize);
4659         }
4660
4661         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4662                                 to_free);
4663 }
4664
4665 /**
4666  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4667  * @inode: inode we're writing to
4668  * @num_bytes: the number of bytes we want to allocate
4669  *
4670  * This will do the following things
4671  *
4672  * o reserve space in the data space info for num_bytes
4673  * o reserve space in the metadata space info based on number of outstanding
4674  *   extents and how much csums will be needed
4675  * o add to the inodes ->delalloc_bytes
4676  * o add it to the fs_info's delalloc inodes list.
4677  *
4678  * This will return 0 for success and -ENOSPC if there is no space left.
4679  */
4680 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4681 {
4682         int ret;
4683
4684         ret = btrfs_check_data_free_space(inode, num_bytes);
4685         if (ret)
4686                 return ret;
4687
4688         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4689         if (ret) {
4690                 btrfs_free_reserved_data_space(inode, num_bytes);
4691                 return ret;
4692         }
4693
4694         return 0;
4695 }
4696
4697 /**
4698  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4699  * @inode: inode we're releasing space for
4700  * @num_bytes: the number of bytes we want to free up
4701  *
4702  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4703  * called in the case that we don't need the metadata AND data reservations
4704  * anymore.  So if there is an error or we insert an inline extent.
4705  *
4706  * This function will release the metadata space that was not used and will
4707  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4708  * list if there are no delalloc bytes left.
4709  */
4710 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4711 {
4712         btrfs_delalloc_release_metadata(inode, num_bytes);
4713         btrfs_free_reserved_data_space(inode, num_bytes);
4714 }
4715
4716 static int update_block_group(struct btrfs_trans_handle *trans,
4717                               struct btrfs_root *root,
4718                               u64 bytenr, u64 num_bytes, int alloc)
4719 {
4720         struct btrfs_block_group_cache *cache = NULL;
4721         struct btrfs_fs_info *info = root->fs_info;
4722         u64 total = num_bytes;
4723         u64 old_val;
4724         u64 byte_in_group;
4725         int factor;
4726
4727         /* block accounting for super block */
4728         spin_lock(&info->delalloc_lock);
4729         old_val = btrfs_super_bytes_used(info->super_copy);
4730         if (alloc)
4731                 old_val += num_bytes;
4732         else
4733                 old_val -= num_bytes;
4734         btrfs_set_super_bytes_used(info->super_copy, old_val);
4735         spin_unlock(&info->delalloc_lock);
4736
4737         while (total) {
4738                 cache = btrfs_lookup_block_group(info, bytenr);
4739                 if (!cache)
4740                         return -ENOENT;
4741                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4742                                     BTRFS_BLOCK_GROUP_RAID1 |
4743                                     BTRFS_BLOCK_GROUP_RAID10))
4744                         factor = 2;
4745                 else
4746                         factor = 1;
4747                 /*
4748                  * If this block group has free space cache written out, we
4749                  * need to make sure to load it if we are removing space.  This
4750                  * is because we need the unpinning stage to actually add the
4751                  * space back to the block group, otherwise we will leak space.
4752                  */
4753                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4754                         cache_block_group(cache, trans, NULL, 1);
4755
4756                 byte_in_group = bytenr - cache->key.objectid;
4757                 WARN_ON(byte_in_group > cache->key.offset);
4758
4759                 spin_lock(&cache->space_info->lock);
4760                 spin_lock(&cache->lock);
4761
4762                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4763                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4764                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4765
4766                 cache->dirty = 1;
4767                 old_val = btrfs_block_group_used(&cache->item);
4768                 num_bytes = min(total, cache->key.offset - byte_in_group);
4769                 if (alloc) {
4770                         old_val += num_bytes;
4771                         btrfs_set_block_group_used(&cache->item, old_val);
4772                         cache->reserved -= num_bytes;
4773                         cache->space_info->bytes_reserved -= num_bytes;
4774                         cache->space_info->bytes_used += num_bytes;
4775                         cache->space_info->disk_used += num_bytes * factor;
4776                         spin_unlock(&cache->lock);
4777                         spin_unlock(&cache->space_info->lock);
4778                 } else {
4779                         old_val -= num_bytes;
4780                         btrfs_set_block_group_used(&cache->item, old_val);
4781                         cache->pinned += num_bytes;
4782                         cache->space_info->bytes_pinned += num_bytes;
4783                         cache->space_info->bytes_used -= num_bytes;
4784                         cache->space_info->disk_used -= num_bytes * factor;
4785                         spin_unlock(&cache->lock);
4786                         spin_unlock(&cache->space_info->lock);
4787
4788                         set_extent_dirty(info->pinned_extents,
4789                                          bytenr, bytenr + num_bytes - 1,
4790                                          GFP_NOFS | __GFP_NOFAIL);
4791                 }
4792                 btrfs_put_block_group(cache);
4793                 total -= num_bytes;
4794                 bytenr += num_bytes;
4795         }
4796         return 0;
4797 }
4798
4799 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4800 {
4801         struct btrfs_block_group_cache *cache;
4802         u64 bytenr;
4803
4804         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4805         if (!cache)
4806                 return 0;
4807
4808         bytenr = cache->key.objectid;
4809         btrfs_put_block_group(cache);
4810
4811         return bytenr;
4812 }
4813
4814 static int pin_down_extent(struct btrfs_root *root,
4815                            struct btrfs_block_group_cache *cache,
4816                            u64 bytenr, u64 num_bytes, int reserved)
4817 {
4818         spin_lock(&cache->space_info->lock);
4819         spin_lock(&cache->lock);
4820         cache->pinned += num_bytes;
4821         cache->space_info->bytes_pinned += num_bytes;
4822         if (reserved) {
4823                 cache->reserved -= num_bytes;
4824                 cache->space_info->bytes_reserved -= num_bytes;
4825         }
4826         spin_unlock(&cache->lock);
4827         spin_unlock(&cache->space_info->lock);
4828
4829         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4830                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4831         return 0;
4832 }
4833
4834 /*
4835  * this function must be called within transaction
4836  */
4837 int btrfs_pin_extent(struct btrfs_root *root,
4838                      u64 bytenr, u64 num_bytes, int reserved)
4839 {
4840         struct btrfs_block_group_cache *cache;
4841
4842         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4843         BUG_ON(!cache); /* Logic error */
4844
4845         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4846
4847         btrfs_put_block_group(cache);
4848         return 0;
4849 }
4850
4851 /*
4852  * this function must be called within transaction
4853  */
4854 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4855                                     struct btrfs_root *root,
4856                                     u64 bytenr, u64 num_bytes)
4857 {
4858         struct btrfs_block_group_cache *cache;
4859
4860         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4861         BUG_ON(!cache); /* Logic error */
4862
4863         /*
4864          * pull in the free space cache (if any) so that our pin
4865          * removes the free space from the cache.  We have load_only set
4866          * to one because the slow code to read in the free extents does check
4867          * the pinned extents.
4868          */
4869         cache_block_group(cache, trans, root, 1);
4870
4871         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4872
4873         /* remove us from the free space cache (if we're there at all) */
4874         btrfs_remove_free_space(cache, bytenr, num_bytes);
4875         btrfs_put_block_group(cache);
4876         return 0;
4877 }
4878
4879 /**
4880  * btrfs_update_reserved_bytes - update the block_group and space info counters
4881  * @cache:      The cache we are manipulating
4882  * @num_bytes:  The number of bytes in question
4883  * @reserve:    One of the reservation enums
4884  *
4885  * This is called by the allocator when it reserves space, or by somebody who is
4886  * freeing space that was never actually used on disk.  For example if you
4887  * reserve some space for a new leaf in transaction A and before transaction A
4888  * commits you free that leaf, you call this with reserve set to 0 in order to
4889  * clear the reservation.
4890  *
4891  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4892  * ENOSPC accounting.  For data we handle the reservation through clearing the
4893  * delalloc bits in the io_tree.  We have to do this since we could end up
4894  * allocating less disk space for the amount of data we have reserved in the
4895  * case of compression.
4896  *
4897  * If this is a reservation and the block group has become read only we cannot
4898  * make the reservation and return -EAGAIN, otherwise this function always
4899  * succeeds.
4900  */
4901 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4902                                        u64 num_bytes, int reserve)
4903 {
4904         struct btrfs_space_info *space_info = cache->space_info;
4905         int ret = 0;
4906
4907         spin_lock(&space_info->lock);
4908         spin_lock(&cache->lock);
4909         if (reserve != RESERVE_FREE) {
4910                 if (cache->ro) {
4911                         ret = -EAGAIN;
4912                 } else {
4913                         cache->reserved += num_bytes;
4914                         space_info->bytes_reserved += num_bytes;
4915                         if (reserve == RESERVE_ALLOC) {
4916                                 trace_btrfs_space_reservation(cache->fs_info,
4917                                                 "space_info", space_info->flags,
4918                                                 num_bytes, 0);
4919                                 space_info->bytes_may_use -= num_bytes;
4920                         }
4921                 }
4922         } else {
4923                 if (cache->ro)
4924                         space_info->bytes_readonly += num_bytes;
4925                 cache->reserved -= num_bytes;
4926                 space_info->bytes_reserved -= num_bytes;
4927                 space_info->reservation_progress++;
4928         }
4929         spin_unlock(&cache->lock);
4930         spin_unlock(&space_info->lock);
4931         return ret;
4932 }
4933
4934 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4935                                 struct btrfs_root *root)
4936 {
4937         struct btrfs_fs_info *fs_info = root->fs_info;
4938         struct btrfs_caching_control *next;
4939         struct btrfs_caching_control *caching_ctl;
4940         struct btrfs_block_group_cache *cache;
4941
4942         down_write(&fs_info->extent_commit_sem);
4943
4944         list_for_each_entry_safe(caching_ctl, next,
4945                                  &fs_info->caching_block_groups, list) {
4946                 cache = caching_ctl->block_group;
4947                 if (block_group_cache_done(cache)) {
4948                         cache->last_byte_to_unpin = (u64)-1;
4949                         list_del_init(&caching_ctl->list);
4950                         put_caching_control(caching_ctl);
4951                 } else {
4952                         cache->last_byte_to_unpin = caching_ctl->progress;
4953                 }
4954         }
4955
4956         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4957                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4958         else
4959                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4960
4961         up_write(&fs_info->extent_commit_sem);
4962
4963         update_global_block_rsv(fs_info);
4964 }
4965
4966 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4967 {
4968         struct btrfs_fs_info *fs_info = root->fs_info;
4969         struct btrfs_block_group_cache *cache = NULL;
4970         u64 len;
4971
4972         while (start <= end) {
4973                 if (!cache ||
4974                     start >= cache->key.objectid + cache->key.offset) {
4975                         if (cache)
4976                                 btrfs_put_block_group(cache);
4977                         cache = btrfs_lookup_block_group(fs_info, start);
4978                         BUG_ON(!cache); /* Logic error */
4979                 }
4980
4981                 len = cache->key.objectid + cache->key.offset - start;
4982                 len = min(len, end + 1 - start);
4983
4984                 if (start < cache->last_byte_to_unpin) {
4985                         len = min(len, cache->last_byte_to_unpin - start);
4986                         btrfs_add_free_space(cache, start, len);
4987                 }
4988
4989                 start += len;
4990
4991                 spin_lock(&cache->space_info->lock);
4992                 spin_lock(&cache->lock);
4993                 cache->pinned -= len;
4994                 cache->space_info->bytes_pinned -= len;
4995                 if (cache->ro)
4996                         cache->space_info->bytes_readonly += len;
4997                 spin_unlock(&cache->lock);
4998                 spin_unlock(&cache->space_info->lock);
4999         }
5000
5001         if (cache)
5002                 btrfs_put_block_group(cache);
5003         return 0;
5004 }
5005
5006 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5007                                struct btrfs_root *root)
5008 {
5009         struct btrfs_fs_info *fs_info = root->fs_info;
5010         struct extent_io_tree *unpin;
5011         u64 start;
5012         u64 end;
5013         int ret;
5014
5015         if (trans->aborted)
5016                 return 0;
5017
5018         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5019                 unpin = &fs_info->freed_extents[1];
5020         else
5021                 unpin = &fs_info->freed_extents[0];
5022
5023         while (1) {
5024                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5025                                             EXTENT_DIRTY);
5026                 if (ret)
5027                         break;
5028
5029                 if (btrfs_test_opt(root, DISCARD))
5030                         ret = btrfs_discard_extent(root, start,
5031                                                    end + 1 - start, NULL);
5032
5033                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5034                 unpin_extent_range(root, start, end);
5035                 cond_resched();
5036         }
5037
5038         return 0;
5039 }
5040
5041 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5042                                 struct btrfs_root *root,
5043                                 u64 bytenr, u64 num_bytes, u64 parent,
5044                                 u64 root_objectid, u64 owner_objectid,
5045                                 u64 owner_offset, int refs_to_drop,
5046                                 struct btrfs_delayed_extent_op *extent_op)
5047 {
5048         struct btrfs_key key;
5049         struct btrfs_path *path;
5050         struct btrfs_fs_info *info = root->fs_info;
5051         struct btrfs_root *extent_root = info->extent_root;
5052         struct extent_buffer *leaf;
5053         struct btrfs_extent_item *ei;
5054         struct btrfs_extent_inline_ref *iref;
5055         int ret;
5056         int is_data;
5057         int extent_slot = 0;
5058         int found_extent = 0;
5059         int num_to_del = 1;
5060         u32 item_size;
5061         u64 refs;
5062
5063         path = btrfs_alloc_path();
5064         if (!path)
5065                 return -ENOMEM;
5066
5067         path->reada = 1;
5068         path->leave_spinning = 1;
5069
5070         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5071         BUG_ON(!is_data && refs_to_drop != 1);
5072
5073         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5074                                     bytenr, num_bytes, parent,
5075                                     root_objectid, owner_objectid,
5076                                     owner_offset);
5077         if (ret == 0) {
5078                 extent_slot = path->slots[0];
5079                 while (extent_slot >= 0) {
5080                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5081                                               extent_slot);
5082                         if (key.objectid != bytenr)
5083                                 break;
5084                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5085                             key.offset == num_bytes) {
5086                                 found_extent = 1;
5087                                 break;
5088                         }
5089                         if (path->slots[0] - extent_slot > 5)
5090                                 break;
5091                         extent_slot--;
5092                 }
5093 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5094                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5095                 if (found_extent && item_size < sizeof(*ei))
5096                         found_extent = 0;
5097 #endif
5098                 if (!found_extent) {
5099                         BUG_ON(iref);
5100                         ret = remove_extent_backref(trans, extent_root, path,
5101                                                     NULL, refs_to_drop,
5102                                                     is_data);
5103                         if (ret)
5104                                 goto abort;
5105                         btrfs_release_path(path);
5106                         path->leave_spinning = 1;
5107
5108                         key.objectid = bytenr;
5109                         key.type = BTRFS_EXTENT_ITEM_KEY;
5110                         key.offset = num_bytes;
5111
5112                         ret = btrfs_search_slot(trans, extent_root,
5113                                                 &key, path, -1, 1);
5114                         if (ret) {
5115                                 printk(KERN_ERR "umm, got %d back from search"
5116                                        ", was looking for %llu\n", ret,
5117                                        (unsigned long long)bytenr);
5118                                 if (ret > 0)
5119                                         btrfs_print_leaf(extent_root,
5120                                                          path->nodes[0]);
5121                         }
5122                         if (ret < 0)
5123                                 goto abort;
5124                         extent_slot = path->slots[0];
5125                 }
5126         } else if (ret == -ENOENT) {
5127                 btrfs_print_leaf(extent_root, path->nodes[0]);
5128                 WARN_ON(1);
5129                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5130                        "parent %llu root %llu  owner %llu offset %llu\n",
5131                        (unsigned long long)bytenr,
5132                        (unsigned long long)parent,
5133                        (unsigned long long)root_objectid,
5134                        (unsigned long long)owner_objectid,
5135                        (unsigned long long)owner_offset);
5136         } else {
5137                 goto abort;
5138         }
5139
5140         leaf = path->nodes[0];
5141         item_size = btrfs_item_size_nr(leaf, extent_slot);
5142 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5143         if (item_size < sizeof(*ei)) {
5144                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5145                 ret = convert_extent_item_v0(trans, extent_root, path,
5146                                              owner_objectid, 0);
5147                 if (ret < 0)
5148                         goto abort;
5149
5150                 btrfs_release_path(path);
5151                 path->leave_spinning = 1;
5152
5153                 key.objectid = bytenr;
5154                 key.type = BTRFS_EXTENT_ITEM_KEY;
5155                 key.offset = num_bytes;
5156
5157                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5158                                         -1, 1);
5159                 if (ret) {
5160                         printk(KERN_ERR "umm, got %d back from search"
5161                                ", was looking for %llu\n", ret,
5162                                (unsigned long long)bytenr);
5163                         btrfs_print_leaf(extent_root, path->nodes[0]);
5164                 }
5165                 if (ret < 0)
5166                         goto abort;
5167                 extent_slot = path->slots[0];
5168                 leaf = path->nodes[0];
5169                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5170         }
5171 #endif
5172         BUG_ON(item_size < sizeof(*ei));
5173         ei = btrfs_item_ptr(leaf, extent_slot,
5174                             struct btrfs_extent_item);
5175         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5176                 struct btrfs_tree_block_info *bi;
5177                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5178                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5179                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5180         }
5181
5182         refs = btrfs_extent_refs(leaf, ei);
5183         BUG_ON(refs < refs_to_drop);
5184         refs -= refs_to_drop;
5185
5186         if (refs > 0) {
5187                 if (extent_op)
5188                         __run_delayed_extent_op(extent_op, leaf, ei);
5189                 /*
5190                  * In the case of inline back ref, reference count will
5191                  * be updated by remove_extent_backref
5192                  */
5193                 if (iref) {
5194                         BUG_ON(!found_extent);
5195                 } else {
5196                         btrfs_set_extent_refs(leaf, ei, refs);
5197                         btrfs_mark_buffer_dirty(leaf);
5198                 }
5199                 if (found_extent) {
5200                         ret = remove_extent_backref(trans, extent_root, path,
5201                                                     iref, refs_to_drop,
5202                                                     is_data);
5203                         if (ret)
5204                                 goto abort;
5205                 }
5206         } else {
5207                 if (found_extent) {
5208                         BUG_ON(is_data && refs_to_drop !=
5209                                extent_data_ref_count(root, path, iref));
5210                         if (iref) {
5211                                 BUG_ON(path->slots[0] != extent_slot);
5212                         } else {
5213                                 BUG_ON(path->slots[0] != extent_slot + 1);
5214                                 path->slots[0] = extent_slot;
5215                                 num_to_del = 2;
5216                         }
5217                 }
5218
5219                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5220                                       num_to_del);
5221                 if (ret)
5222                         goto abort;
5223                 btrfs_release_path(path);
5224
5225                 if (is_data) {
5226                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5227                         if (ret)
5228                                 goto abort;
5229                 }
5230
5231                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5232                 if (ret)
5233                         goto abort;
5234         }
5235 out:
5236         btrfs_free_path(path);
5237         return ret;
5238
5239 abort:
5240         btrfs_abort_transaction(trans, extent_root, ret);
5241         goto out;
5242 }
5243
5244 /*
5245  * when we free an block, it is possible (and likely) that we free the last
5246  * delayed ref for that extent as well.  This searches the delayed ref tree for
5247  * a given extent, and if there are no other delayed refs to be processed, it
5248  * removes it from the tree.
5249  */
5250 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5251                                       struct btrfs_root *root, u64 bytenr)
5252 {
5253         struct btrfs_delayed_ref_head *head;
5254         struct btrfs_delayed_ref_root *delayed_refs;
5255         struct btrfs_delayed_ref_node *ref;
5256         struct rb_node *node;
5257         int ret = 0;
5258
5259         delayed_refs = &trans->transaction->delayed_refs;
5260         spin_lock(&delayed_refs->lock);
5261         head = btrfs_find_delayed_ref_head(trans, bytenr);
5262         if (!head)
5263                 goto out;
5264
5265         node = rb_prev(&head->node.rb_node);
5266         if (!node)
5267                 goto out;
5268
5269         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5270
5271         /* there are still entries for this ref, we can't drop it */
5272         if (ref->bytenr == bytenr)
5273                 goto out;
5274
5275         if (head->extent_op) {
5276                 if (!head->must_insert_reserved)
5277                         goto out;
5278                 kfree(head->extent_op);
5279                 head->extent_op = NULL;
5280         }
5281
5282         /*
5283          * waiting for the lock here would deadlock.  If someone else has it
5284          * locked they are already in the process of dropping it anyway
5285          */
5286         if (!mutex_trylock(&head->mutex))
5287                 goto out;
5288
5289         /*
5290          * at this point we have a head with no other entries.  Go
5291          * ahead and process it.
5292          */
5293         head->node.in_tree = 0;
5294         rb_erase(&head->node.rb_node, &delayed_refs->root);
5295
5296         delayed_refs->num_entries--;
5297         if (waitqueue_active(&root->fs_info->tree_mod_seq_wait))
5298                 wake_up(&root->fs_info->tree_mod_seq_wait);
5299
5300         /*
5301          * we don't take a ref on the node because we're removing it from the
5302          * tree, so we just steal the ref the tree was holding.
5303          */
5304         delayed_refs->num_heads--;
5305         if (list_empty(&head->cluster))
5306                 delayed_refs->num_heads_ready--;
5307
5308         list_del_init(&head->cluster);
5309         spin_unlock(&delayed_refs->lock);
5310
5311         BUG_ON(head->extent_op);
5312         if (head->must_insert_reserved)
5313                 ret = 1;
5314
5315         mutex_unlock(&head->mutex);
5316         btrfs_put_delayed_ref(&head->node);
5317         return ret;
5318 out:
5319         spin_unlock(&delayed_refs->lock);
5320         return 0;
5321 }
5322
5323 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5324                            struct btrfs_root *root,
5325                            struct extent_buffer *buf,
5326                            u64 parent, int last_ref)
5327 {
5328         struct btrfs_block_group_cache *cache = NULL;
5329         int ret;
5330
5331         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5332                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5333                                         buf->start, buf->len,
5334                                         parent, root->root_key.objectid,
5335                                         btrfs_header_level(buf),
5336                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5337                 BUG_ON(ret); /* -ENOMEM */
5338         }
5339
5340         if (!last_ref)
5341                 return;
5342
5343         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5344
5345         if (btrfs_header_generation(buf) == trans->transid) {
5346                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5347                         ret = check_ref_cleanup(trans, root, buf->start);
5348                         if (!ret)
5349                                 goto out;
5350                 }
5351
5352                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5353                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5354                         goto out;
5355                 }
5356
5357                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5358
5359                 btrfs_add_free_space(cache, buf->start, buf->len);
5360                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5361         }
5362 out:
5363         /*
5364          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5365          * anymore.
5366          */
5367         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5368         btrfs_put_block_group(cache);
5369 }
5370
5371 /* Can return -ENOMEM */
5372 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5373                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5374                       u64 owner, u64 offset, int for_cow)
5375 {
5376         int ret;
5377         struct btrfs_fs_info *fs_info = root->fs_info;
5378
5379         /*
5380          * tree log blocks never actually go into the extent allocation
5381          * tree, just update pinning info and exit early.
5382          */
5383         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5384                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5385                 /* unlocks the pinned mutex */
5386                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5387                 ret = 0;
5388         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5389                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5390                                         num_bytes,
5391                                         parent, root_objectid, (int)owner,
5392                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5393         } else {
5394                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5395                                                 num_bytes,
5396                                                 parent, root_objectid, owner,
5397                                                 offset, BTRFS_DROP_DELAYED_REF,
5398                                                 NULL, for_cow);
5399         }
5400         return ret;
5401 }
5402
5403 static u64 stripe_align(struct btrfs_root *root, u64 val)
5404 {
5405         u64 mask = ((u64)root->stripesize - 1);
5406         u64 ret = (val + mask) & ~mask;
5407         return ret;
5408 }
5409
5410 /*
5411  * when we wait for progress in the block group caching, its because
5412  * our allocation attempt failed at least once.  So, we must sleep
5413  * and let some progress happen before we try again.
5414  *
5415  * This function will sleep at least once waiting for new free space to
5416  * show up, and then it will check the block group free space numbers
5417  * for our min num_bytes.  Another option is to have it go ahead
5418  * and look in the rbtree for a free extent of a given size, but this
5419  * is a good start.
5420  */
5421 static noinline int
5422 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5423                                 u64 num_bytes)
5424 {
5425         struct btrfs_caching_control *caching_ctl;
5426         DEFINE_WAIT(wait);
5427
5428         caching_ctl = get_caching_control(cache);
5429         if (!caching_ctl)
5430                 return 0;
5431
5432         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5433                    (cache->free_space_ctl->free_space >= num_bytes));
5434
5435         put_caching_control(caching_ctl);
5436         return 0;
5437 }
5438
5439 static noinline int
5440 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5441 {
5442         struct btrfs_caching_control *caching_ctl;
5443         DEFINE_WAIT(wait);
5444
5445         caching_ctl = get_caching_control(cache);
5446         if (!caching_ctl)
5447                 return 0;
5448
5449         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5450
5451         put_caching_control(caching_ctl);
5452         return 0;
5453 }
5454
5455 static int __get_block_group_index(u64 flags)
5456 {
5457         int index;
5458
5459         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5460                 index = 0;
5461         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5462                 index = 1;
5463         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5464                 index = 2;
5465         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5466                 index = 3;
5467         else
5468                 index = 4;
5469
5470         return index;
5471 }
5472
5473 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5474 {
5475         return __get_block_group_index(cache->flags);
5476 }
5477
5478 enum btrfs_loop_type {
5479         LOOP_CACHING_NOWAIT = 0,
5480         LOOP_CACHING_WAIT = 1,
5481         LOOP_ALLOC_CHUNK = 2,
5482         LOOP_NO_EMPTY_SIZE = 3,
5483 };
5484
5485 /*
5486  * walks the btree of allocated extents and find a hole of a given size.
5487  * The key ins is changed to record the hole:
5488  * ins->objectid == block start
5489  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5490  * ins->offset == number of blocks
5491  * Any available blocks before search_start are skipped.
5492  */
5493 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5494                                      struct btrfs_root *orig_root,
5495                                      u64 num_bytes, u64 empty_size,
5496                                      u64 hint_byte, struct btrfs_key *ins,
5497                                      u64 data)
5498 {
5499         int ret = 0;
5500         struct btrfs_root *root = orig_root->fs_info->extent_root;
5501         struct btrfs_free_cluster *last_ptr = NULL;
5502         struct btrfs_block_group_cache *block_group = NULL;
5503         struct btrfs_block_group_cache *used_block_group;
5504         u64 search_start = 0;
5505         int empty_cluster = 2 * 1024 * 1024;
5506         int allowed_chunk_alloc = 0;
5507         int done_chunk_alloc = 0;
5508         struct btrfs_space_info *space_info;
5509         int loop = 0;
5510         int index = 0;
5511         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5512                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5513         bool found_uncached_bg = false;
5514         bool failed_cluster_refill = false;
5515         bool failed_alloc = false;
5516         bool use_cluster = true;
5517         bool have_caching_bg = false;
5518
5519         WARN_ON(num_bytes < root->sectorsize);
5520         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5521         ins->objectid = 0;
5522         ins->offset = 0;
5523
5524         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5525
5526         space_info = __find_space_info(root->fs_info, data);
5527         if (!space_info) {
5528                 printk(KERN_ERR "No space info for %llu\n", data);
5529                 return -ENOSPC;
5530         }
5531
5532         /*
5533          * If the space info is for both data and metadata it means we have a
5534          * small filesystem and we can't use the clustering stuff.
5535          */
5536         if (btrfs_mixed_space_info(space_info))
5537                 use_cluster = false;
5538
5539         if (orig_root->ref_cows || empty_size)
5540                 allowed_chunk_alloc = 1;
5541
5542         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5543                 last_ptr = &root->fs_info->meta_alloc_cluster;
5544                 if (!btrfs_test_opt(root, SSD))
5545                         empty_cluster = 64 * 1024;
5546         }
5547
5548         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5549             btrfs_test_opt(root, SSD)) {
5550                 last_ptr = &root->fs_info->data_alloc_cluster;
5551         }
5552
5553         if (last_ptr) {
5554                 spin_lock(&last_ptr->lock);
5555                 if (last_ptr->block_group)
5556                         hint_byte = last_ptr->window_start;
5557                 spin_unlock(&last_ptr->lock);
5558         }
5559
5560         search_start = max(search_start, first_logical_byte(root, 0));
5561         search_start = max(search_start, hint_byte);
5562
5563         if (!last_ptr)
5564                 empty_cluster = 0;
5565
5566         if (search_start == hint_byte) {
5567                 block_group = btrfs_lookup_block_group(root->fs_info,
5568                                                        search_start);
5569                 used_block_group = block_group;
5570                 /*
5571                  * we don't want to use the block group if it doesn't match our
5572                  * allocation bits, or if its not cached.
5573                  *
5574                  * However if we are re-searching with an ideal block group
5575                  * picked out then we don't care that the block group is cached.
5576                  */
5577                 if (block_group && block_group_bits(block_group, data) &&
5578                     block_group->cached != BTRFS_CACHE_NO) {
5579                         down_read(&space_info->groups_sem);
5580                         if (list_empty(&block_group->list) ||
5581                             block_group->ro) {
5582                                 /*
5583                                  * someone is removing this block group,
5584                                  * we can't jump into the have_block_group
5585                                  * target because our list pointers are not
5586                                  * valid
5587                                  */
5588                                 btrfs_put_block_group(block_group);
5589                                 up_read(&space_info->groups_sem);
5590                         } else {
5591                                 index = get_block_group_index(block_group);
5592                                 goto have_block_group;
5593                         }
5594                 } else if (block_group) {
5595                         btrfs_put_block_group(block_group);
5596                 }
5597         }
5598 search:
5599         have_caching_bg = false;
5600         down_read(&space_info->groups_sem);
5601         list_for_each_entry(block_group, &space_info->block_groups[index],
5602                             list) {
5603                 u64 offset;
5604                 int cached;
5605
5606                 used_block_group = block_group;
5607                 btrfs_get_block_group(block_group);
5608                 search_start = block_group->key.objectid;
5609
5610                 /*
5611                  * this can happen if we end up cycling through all the
5612                  * raid types, but we want to make sure we only allocate
5613                  * for the proper type.
5614                  */
5615                 if (!block_group_bits(block_group, data)) {
5616                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5617                                 BTRFS_BLOCK_GROUP_RAID1 |
5618                                 BTRFS_BLOCK_GROUP_RAID10;
5619
5620                         /*
5621                          * if they asked for extra copies and this block group
5622                          * doesn't provide them, bail.  This does allow us to
5623                          * fill raid0 from raid1.
5624                          */
5625                         if ((data & extra) && !(block_group->flags & extra))
5626                                 goto loop;
5627                 }
5628
5629 have_block_group:
5630                 cached = block_group_cache_done(block_group);
5631                 if (unlikely(!cached)) {
5632                         found_uncached_bg = true;
5633                         ret = cache_block_group(block_group, trans,
5634                                                 orig_root, 0);
5635                         BUG_ON(ret < 0);
5636                         ret = 0;
5637                 }
5638
5639                 if (unlikely(block_group->ro))
5640                         goto loop;
5641
5642                 /*
5643                  * Ok we want to try and use the cluster allocator, so
5644                  * lets look there
5645                  */
5646                 if (last_ptr) {
5647                         /*
5648                          * the refill lock keeps out other
5649                          * people trying to start a new cluster
5650                          */
5651                         spin_lock(&last_ptr->refill_lock);
5652                         used_block_group = last_ptr->block_group;
5653                         if (used_block_group != block_group &&
5654                             (!used_block_group ||
5655                              used_block_group->ro ||
5656                              !block_group_bits(used_block_group, data))) {
5657                                 used_block_group = block_group;
5658                                 goto refill_cluster;
5659                         }
5660
5661                         if (used_block_group != block_group)
5662                                 btrfs_get_block_group(used_block_group);
5663
5664                         offset = btrfs_alloc_from_cluster(used_block_group,
5665                           last_ptr, num_bytes, used_block_group->key.objectid);
5666                         if (offset) {
5667                                 /* we have a block, we're done */
5668                                 spin_unlock(&last_ptr->refill_lock);
5669                                 trace_btrfs_reserve_extent_cluster(root,
5670                                         block_group, search_start, num_bytes);
5671                                 goto checks;
5672                         }
5673
5674                         WARN_ON(last_ptr->block_group != used_block_group);