]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/btrfs/extent-tree.c
0f47b3e2010eaae223689a374c88874ca576bdb7
[karo-tx-linux.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 /* control flags for do_chunk_alloc's force field
38  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
39  * if we really need one.
40  *
41  * CHUNK_ALLOC_FORCE means it must try to allocate one
42  *
43  * CHUNK_ALLOC_LIMITED means to only try and allocate one
44  * if we have very few chunks already allocated.  This is
45  * used as part of the clustering code to help make sure
46  * we have a good pool of storage to cluster in, without
47  * filling the FS with empty chunks
48  *
49  */
50 enum {
51         CHUNK_ALLOC_NO_FORCE = 0,
52         CHUNK_ALLOC_FORCE = 1,
53         CHUNK_ALLOC_LIMITED = 2,
54 };
55
56 /*
57  * Control how reservations are dealt with.
58  *
59  * RESERVE_FREE - freeing a reservation.
60  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
61  *   ENOSPC accounting
62  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
63  *   bytes_may_use as the ENOSPC accounting is done elsewhere
64  */
65 enum {
66         RESERVE_FREE = 0,
67         RESERVE_ALLOC = 1,
68         RESERVE_ALLOC_NO_ACCOUNT = 2,
69 };
70
71 static int update_block_group(struct btrfs_trans_handle *trans,
72                               struct btrfs_root *root,
73                               u64 bytenr, u64 num_bytes, int alloc);
74 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
75                                 struct btrfs_root *root,
76                                 u64 bytenr, u64 num_bytes, u64 parent,
77                                 u64 root_objectid, u64 owner_objectid,
78                                 u64 owner_offset, int refs_to_drop,
79                                 struct btrfs_delayed_extent_op *extra_op);
80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
81                                     struct extent_buffer *leaf,
82                                     struct btrfs_extent_item *ei);
83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
84                                       struct btrfs_root *root,
85                                       u64 parent, u64 root_objectid,
86                                       u64 flags, u64 owner, u64 offset,
87                                       struct btrfs_key *ins, int ref_mod);
88 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
89                                      struct btrfs_root *root,
90                                      u64 parent, u64 root_objectid,
91                                      u64 flags, struct btrfs_disk_key *key,
92                                      int level, struct btrfs_key *ins);
93 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
94                           struct btrfs_root *extent_root, u64 alloc_bytes,
95                           u64 flags, int force);
96 static int find_next_key(struct btrfs_path *path, int level,
97                          struct btrfs_key *key);
98 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
99                             int dump_block_groups);
100 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
101                                        u64 num_bytes, int reserve);
102
103 static noinline int
104 block_group_cache_done(struct btrfs_block_group_cache *cache)
105 {
106         smp_mb();
107         return cache->cached == BTRFS_CACHE_FINISHED;
108 }
109
110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
111 {
112         return (cache->flags & bits) == bits;
113 }
114
115 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
116 {
117         atomic_inc(&cache->count);
118 }
119
120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
121 {
122         if (atomic_dec_and_test(&cache->count)) {
123                 WARN_ON(cache->pinned > 0);
124                 WARN_ON(cache->reserved > 0);
125                 kfree(cache->free_space_ctl);
126                 kfree(cache);
127         }
128 }
129
130 /*
131  * this adds the block group to the fs_info rb tree for the block group
132  * cache
133  */
134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
135                                 struct btrfs_block_group_cache *block_group)
136 {
137         struct rb_node **p;
138         struct rb_node *parent = NULL;
139         struct btrfs_block_group_cache *cache;
140
141         spin_lock(&info->block_group_cache_lock);
142         p = &info->block_group_cache_tree.rb_node;
143
144         while (*p) {
145                 parent = *p;
146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
147                                  cache_node);
148                 if (block_group->key.objectid < cache->key.objectid) {
149                         p = &(*p)->rb_left;
150                 } else if (block_group->key.objectid > cache->key.objectid) {
151                         p = &(*p)->rb_right;
152                 } else {
153                         spin_unlock(&info->block_group_cache_lock);
154                         return -EEXIST;
155                 }
156         }
157
158         rb_link_node(&block_group->cache_node, parent, p);
159         rb_insert_color(&block_group->cache_node,
160                         &info->block_group_cache_tree);
161         spin_unlock(&info->block_group_cache_lock);
162
163         return 0;
164 }
165
166 /*
167  * This will return the block group at or after bytenr if contains is 0, else
168  * it will return the block group that contains the bytenr
169  */
170 static struct btrfs_block_group_cache *
171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
172                               int contains)
173 {
174         struct btrfs_block_group_cache *cache, *ret = NULL;
175         struct rb_node *n;
176         u64 end, start;
177
178         spin_lock(&info->block_group_cache_lock);
179         n = info->block_group_cache_tree.rb_node;
180
181         while (n) {
182                 cache = rb_entry(n, struct btrfs_block_group_cache,
183                                  cache_node);
184                 end = cache->key.objectid + cache->key.offset - 1;
185                 start = cache->key.objectid;
186
187                 if (bytenr < start) {
188                         if (!contains && (!ret || start < ret->key.objectid))
189                                 ret = cache;
190                         n = n->rb_left;
191                 } else if (bytenr > start) {
192                         if (contains && bytenr <= end) {
193                                 ret = cache;
194                                 break;
195                         }
196                         n = n->rb_right;
197                 } else {
198                         ret = cache;
199                         break;
200                 }
201         }
202         if (ret)
203                 btrfs_get_block_group(ret);
204         spin_unlock(&info->block_group_cache_lock);
205
206         return ret;
207 }
208
209 static int add_excluded_extent(struct btrfs_root *root,
210                                u64 start, u64 num_bytes)
211 {
212         u64 end = start + num_bytes - 1;
213         set_extent_bits(&root->fs_info->freed_extents[0],
214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
215         set_extent_bits(&root->fs_info->freed_extents[1],
216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
217         return 0;
218 }
219
220 static void free_excluded_extents(struct btrfs_root *root,
221                                   struct btrfs_block_group_cache *cache)
222 {
223         u64 start, end;
224
225         start = cache->key.objectid;
226         end = start + cache->key.offset - 1;
227
228         clear_extent_bits(&root->fs_info->freed_extents[0],
229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
230         clear_extent_bits(&root->fs_info->freed_extents[1],
231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
232 }
233
234 static int exclude_super_stripes(struct btrfs_root *root,
235                                  struct btrfs_block_group_cache *cache)
236 {
237         u64 bytenr;
238         u64 *logical;
239         int stripe_len;
240         int i, nr, ret;
241
242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244                 cache->bytes_super += stripe_len;
245                 ret = add_excluded_extent(root, cache->key.objectid,
246                                           stripe_len);
247                 BUG_ON(ret);
248         }
249
250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
251                 bytenr = btrfs_sb_offset(i);
252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
253                                        cache->key.objectid, bytenr,
254                                        0, &logical, &nr, &stripe_len);
255                 BUG_ON(ret);
256
257                 while (nr--) {
258                         cache->bytes_super += stripe_len;
259                         ret = add_excluded_extent(root, logical[nr],
260                                                   stripe_len);
261                         BUG_ON(ret);
262                 }
263
264                 kfree(logical);
265         }
266         return 0;
267 }
268
269 static struct btrfs_caching_control *
270 get_caching_control(struct btrfs_block_group_cache *cache)
271 {
272         struct btrfs_caching_control *ctl;
273
274         spin_lock(&cache->lock);
275         if (cache->cached != BTRFS_CACHE_STARTED) {
276                 spin_unlock(&cache->lock);
277                 return NULL;
278         }
279
280         /* We're loading it the fast way, so we don't have a caching_ctl. */
281         if (!cache->caching_ctl) {
282                 spin_unlock(&cache->lock);
283                 return NULL;
284         }
285
286         ctl = cache->caching_ctl;
287         atomic_inc(&ctl->count);
288         spin_unlock(&cache->lock);
289         return ctl;
290 }
291
292 static void put_caching_control(struct btrfs_caching_control *ctl)
293 {
294         if (atomic_dec_and_test(&ctl->count))
295                 kfree(ctl);
296 }
297
298 /*
299  * this is only called by cache_block_group, since we could have freed extents
300  * we need to check the pinned_extents for any extents that can't be used yet
301  * since their free space will be released as soon as the transaction commits.
302  */
303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
304                               struct btrfs_fs_info *info, u64 start, u64 end)
305 {
306         u64 extent_start, extent_end, size, total_added = 0;
307         int ret;
308
309         while (start < end) {
310                 ret = find_first_extent_bit(info->pinned_extents, start,
311                                             &extent_start, &extent_end,
312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
313                 if (ret)
314                         break;
315
316                 if (extent_start <= start) {
317                         start = extent_end + 1;
318                 } else if (extent_start > start && extent_start < end) {
319                         size = extent_start - start;
320                         total_added += size;
321                         ret = btrfs_add_free_space(block_group, start,
322                                                    size);
323                         BUG_ON(ret);
324                         start = extent_end + 1;
325                 } else {
326                         break;
327                 }
328         }
329
330         if (start < end) {
331                 size = end - start;
332                 total_added += size;
333                 ret = btrfs_add_free_space(block_group, start, size);
334                 BUG_ON(ret);
335         }
336
337         return total_added;
338 }
339
340 static noinline void caching_thread(struct btrfs_work *work)
341 {
342         struct btrfs_block_group_cache *block_group;
343         struct btrfs_fs_info *fs_info;
344         struct btrfs_caching_control *caching_ctl;
345         struct btrfs_root *extent_root;
346         struct btrfs_path *path;
347         struct extent_buffer *leaf;
348         struct btrfs_key key;
349         u64 total_found = 0;
350         u64 last = 0;
351         u32 nritems;
352         int ret = 0;
353
354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
355         block_group = caching_ctl->block_group;
356         fs_info = block_group->fs_info;
357         extent_root = fs_info->extent_root;
358
359         path = btrfs_alloc_path();
360         if (!path)
361                 goto out;
362
363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
364
365         /*
366          * We don't want to deadlock with somebody trying to allocate a new
367          * extent for the extent root while also trying to search the extent
368          * root to add free space.  So we skip locking and search the commit
369          * root, since its read-only
370          */
371         path->skip_locking = 1;
372         path->search_commit_root = 1;
373         path->reada = 1;
374
375         key.objectid = last;
376         key.offset = 0;
377         key.type = BTRFS_EXTENT_ITEM_KEY;
378 again:
379         mutex_lock(&caching_ctl->mutex);
380         /* need to make sure the commit_root doesn't disappear */
381         down_read(&fs_info->extent_commit_sem);
382
383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
384         if (ret < 0)
385                 goto err;
386
387         leaf = path->nodes[0];
388         nritems = btrfs_header_nritems(leaf);
389
390         while (1) {
391                 if (btrfs_fs_closing(fs_info) > 1) {
392                         last = (u64)-1;
393                         break;
394                 }
395
396                 if (path->slots[0] < nritems) {
397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
398                 } else {
399                         ret = find_next_key(path, 0, &key);
400                         if (ret)
401                                 break;
402
403                         if (need_resched() ||
404                             btrfs_next_leaf(extent_root, path)) {
405                                 caching_ctl->progress = last;
406                                 btrfs_release_path(path);
407                                 up_read(&fs_info->extent_commit_sem);
408                                 mutex_unlock(&caching_ctl->mutex);
409                                 cond_resched();
410                                 goto again;
411                         }
412                         leaf = path->nodes[0];
413                         nritems = btrfs_header_nritems(leaf);
414                         continue;
415                 }
416
417                 if (key.objectid < block_group->key.objectid) {
418                         path->slots[0]++;
419                         continue;
420                 }
421
422                 if (key.objectid >= block_group->key.objectid +
423                     block_group->key.offset)
424                         break;
425
426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
427                         total_found += add_new_free_space(block_group,
428                                                           fs_info, last,
429                                                           key.objectid);
430                         last = key.objectid + key.offset;
431
432                         if (total_found > (1024 * 1024 * 2)) {
433                                 total_found = 0;
434                                 wake_up(&caching_ctl->wait);
435                         }
436                 }
437                 path->slots[0]++;
438         }
439         ret = 0;
440
441         total_found += add_new_free_space(block_group, fs_info, last,
442                                           block_group->key.objectid +
443                                           block_group->key.offset);
444         caching_ctl->progress = (u64)-1;
445
446         spin_lock(&block_group->lock);
447         block_group->caching_ctl = NULL;
448         block_group->cached = BTRFS_CACHE_FINISHED;
449         spin_unlock(&block_group->lock);
450
451 err:
452         btrfs_free_path(path);
453         up_read(&fs_info->extent_commit_sem);
454
455         free_excluded_extents(extent_root, block_group);
456
457         mutex_unlock(&caching_ctl->mutex);
458 out:
459         wake_up(&caching_ctl->wait);
460
461         put_caching_control(caching_ctl);
462         btrfs_put_block_group(block_group);
463 }
464
465 static int cache_block_group(struct btrfs_block_group_cache *cache,
466                              struct btrfs_trans_handle *trans,
467                              struct btrfs_root *root,
468                              int load_cache_only)
469 {
470         struct btrfs_fs_info *fs_info = cache->fs_info;
471         struct btrfs_caching_control *caching_ctl;
472         int ret = 0;
473
474         smp_mb();
475         if (cache->cached != BTRFS_CACHE_NO)
476                 return 0;
477
478         /*
479          * We can't do the read from on-disk cache during a commit since we need
480          * to have the normal tree locking.  Also if we are currently trying to
481          * allocate blocks for the tree root we can't do the fast caching since
482          * we likely hold important locks.
483          */
484         if (trans && (!trans->transaction->in_commit) &&
485             (root && root != root->fs_info->tree_root) &&
486             btrfs_test_opt(root, SPACE_CACHE)) {
487                 spin_lock(&cache->lock);
488                 if (cache->cached != BTRFS_CACHE_NO) {
489                         spin_unlock(&cache->lock);
490                         return 0;
491                 }
492                 cache->cached = BTRFS_CACHE_STARTED;
493                 spin_unlock(&cache->lock);
494
495                 ret = load_free_space_cache(fs_info, cache);
496
497                 spin_lock(&cache->lock);
498                 if (ret == 1) {
499                         cache->cached = BTRFS_CACHE_FINISHED;
500                         cache->last_byte_to_unpin = (u64)-1;
501                 } else {
502                         cache->cached = BTRFS_CACHE_NO;
503                 }
504                 spin_unlock(&cache->lock);
505                 if (ret == 1) {
506                         free_excluded_extents(fs_info->extent_root, cache);
507                         return 0;
508                 }
509         }
510
511         if (load_cache_only)
512                 return 0;
513
514         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
515         BUG_ON(!caching_ctl);
516
517         INIT_LIST_HEAD(&caching_ctl->list);
518         mutex_init(&caching_ctl->mutex);
519         init_waitqueue_head(&caching_ctl->wait);
520         caching_ctl->block_group = cache;
521         caching_ctl->progress = cache->key.objectid;
522         /* one for caching kthread, one for caching block group list */
523         atomic_set(&caching_ctl->count, 2);
524         caching_ctl->work.func = caching_thread;
525
526         spin_lock(&cache->lock);
527         if (cache->cached != BTRFS_CACHE_NO) {
528                 spin_unlock(&cache->lock);
529                 kfree(caching_ctl);
530                 return 0;
531         }
532         cache->caching_ctl = caching_ctl;
533         cache->cached = BTRFS_CACHE_STARTED;
534         spin_unlock(&cache->lock);
535
536         down_write(&fs_info->extent_commit_sem);
537         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
538         up_write(&fs_info->extent_commit_sem);
539
540         btrfs_get_block_group(cache);
541
542         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
543
544         return ret;
545 }
546
547 /*
548  * return the block group that starts at or after bytenr
549  */
550 static struct btrfs_block_group_cache *
551 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
552 {
553         struct btrfs_block_group_cache *cache;
554
555         cache = block_group_cache_tree_search(info, bytenr, 0);
556
557         return cache;
558 }
559
560 /*
561  * return the block group that contains the given bytenr
562  */
563 struct btrfs_block_group_cache *btrfs_lookup_block_group(
564                                                  struct btrfs_fs_info *info,
565                                                  u64 bytenr)
566 {
567         struct btrfs_block_group_cache *cache;
568
569         cache = block_group_cache_tree_search(info, bytenr, 1);
570
571         return cache;
572 }
573
574 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
575                                                   u64 flags)
576 {
577         struct list_head *head = &info->space_info;
578         struct btrfs_space_info *found;
579
580         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
581                  BTRFS_BLOCK_GROUP_METADATA;
582
583         rcu_read_lock();
584         list_for_each_entry_rcu(found, head, list) {
585                 if (found->flags & flags) {
586                         rcu_read_unlock();
587                         return found;
588                 }
589         }
590         rcu_read_unlock();
591         return NULL;
592 }
593
594 /*
595  * after adding space to the filesystem, we need to clear the full flags
596  * on all the space infos.
597  */
598 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
599 {
600         struct list_head *head = &info->space_info;
601         struct btrfs_space_info *found;
602
603         rcu_read_lock();
604         list_for_each_entry_rcu(found, head, list)
605                 found->full = 0;
606         rcu_read_unlock();
607 }
608
609 static u64 div_factor(u64 num, int factor)
610 {
611         if (factor == 10)
612                 return num;
613         num *= factor;
614         do_div(num, 10);
615         return num;
616 }
617
618 static u64 div_factor_fine(u64 num, int factor)
619 {
620         if (factor == 100)
621                 return num;
622         num *= factor;
623         do_div(num, 100);
624         return num;
625 }
626
627 u64 btrfs_find_block_group(struct btrfs_root *root,
628                            u64 search_start, u64 search_hint, int owner)
629 {
630         struct btrfs_block_group_cache *cache;
631         u64 used;
632         u64 last = max(search_hint, search_start);
633         u64 group_start = 0;
634         int full_search = 0;
635         int factor = 9;
636         int wrapped = 0;
637 again:
638         while (1) {
639                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
640                 if (!cache)
641                         break;
642
643                 spin_lock(&cache->lock);
644                 last = cache->key.objectid + cache->key.offset;
645                 used = btrfs_block_group_used(&cache->item);
646
647                 if ((full_search || !cache->ro) &&
648                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
649                         if (used + cache->pinned + cache->reserved <
650                             div_factor(cache->key.offset, factor)) {
651                                 group_start = cache->key.objectid;
652                                 spin_unlock(&cache->lock);
653                                 btrfs_put_block_group(cache);
654                                 goto found;
655                         }
656                 }
657                 spin_unlock(&cache->lock);
658                 btrfs_put_block_group(cache);
659                 cond_resched();
660         }
661         if (!wrapped) {
662                 last = search_start;
663                 wrapped = 1;
664                 goto again;
665         }
666         if (!full_search && factor < 10) {
667                 last = search_start;
668                 full_search = 1;
669                 factor = 10;
670                 goto again;
671         }
672 found:
673         return group_start;
674 }
675
676 /* simple helper to search for an existing extent at a given offset */
677 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
678 {
679         int ret;
680         struct btrfs_key key;
681         struct btrfs_path *path;
682
683         path = btrfs_alloc_path();
684         if (!path)
685                 return -ENOMEM;
686
687         key.objectid = start;
688         key.offset = len;
689         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
690         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
691                                 0, 0);
692         btrfs_free_path(path);
693         return ret;
694 }
695
696 /*
697  * helper function to lookup reference count and flags of extent.
698  *
699  * the head node for delayed ref is used to store the sum of all the
700  * reference count modifications queued up in the rbtree. the head
701  * node may also store the extent flags to set. This way you can check
702  * to see what the reference count and extent flags would be if all of
703  * the delayed refs are not processed.
704  */
705 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
706                              struct btrfs_root *root, u64 bytenr,
707                              u64 num_bytes, u64 *refs, u64 *flags)
708 {
709         struct btrfs_delayed_ref_head *head;
710         struct btrfs_delayed_ref_root *delayed_refs;
711         struct btrfs_path *path;
712         struct btrfs_extent_item *ei;
713         struct extent_buffer *leaf;
714         struct btrfs_key key;
715         u32 item_size;
716         u64 num_refs;
717         u64 extent_flags;
718         int ret;
719
720         path = btrfs_alloc_path();
721         if (!path)
722                 return -ENOMEM;
723
724         key.objectid = bytenr;
725         key.type = BTRFS_EXTENT_ITEM_KEY;
726         key.offset = num_bytes;
727         if (!trans) {
728                 path->skip_locking = 1;
729                 path->search_commit_root = 1;
730         }
731 again:
732         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
733                                 &key, path, 0, 0);
734         if (ret < 0)
735                 goto out_free;
736
737         if (ret == 0) {
738                 leaf = path->nodes[0];
739                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
740                 if (item_size >= sizeof(*ei)) {
741                         ei = btrfs_item_ptr(leaf, path->slots[0],
742                                             struct btrfs_extent_item);
743                         num_refs = btrfs_extent_refs(leaf, ei);
744                         extent_flags = btrfs_extent_flags(leaf, ei);
745                 } else {
746 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
747                         struct btrfs_extent_item_v0 *ei0;
748                         BUG_ON(item_size != sizeof(*ei0));
749                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
750                                              struct btrfs_extent_item_v0);
751                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
752                         /* FIXME: this isn't correct for data */
753                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
754 #else
755                         BUG();
756 #endif
757                 }
758                 BUG_ON(num_refs == 0);
759         } else {
760                 num_refs = 0;
761                 extent_flags = 0;
762                 ret = 0;
763         }
764
765         if (!trans)
766                 goto out;
767
768         delayed_refs = &trans->transaction->delayed_refs;
769         spin_lock(&delayed_refs->lock);
770         head = btrfs_find_delayed_ref_head(trans, bytenr);
771         if (head) {
772                 if (!mutex_trylock(&head->mutex)) {
773                         atomic_inc(&head->node.refs);
774                         spin_unlock(&delayed_refs->lock);
775
776                         btrfs_release_path(path);
777
778                         /*
779                          * Mutex was contended, block until it's released and try
780                          * again
781                          */
782                         mutex_lock(&head->mutex);
783                         mutex_unlock(&head->mutex);
784                         btrfs_put_delayed_ref(&head->node);
785                         goto again;
786                 }
787                 if (head->extent_op && head->extent_op->update_flags)
788                         extent_flags |= head->extent_op->flags_to_set;
789                 else
790                         BUG_ON(num_refs == 0);
791
792                 num_refs += head->node.ref_mod;
793                 mutex_unlock(&head->mutex);
794         }
795         spin_unlock(&delayed_refs->lock);
796 out:
797         WARN_ON(num_refs == 0);
798         if (refs)
799                 *refs = num_refs;
800         if (flags)
801                 *flags = extent_flags;
802 out_free:
803         btrfs_free_path(path);
804         return ret;
805 }
806
807 /*
808  * Back reference rules.  Back refs have three main goals:
809  *
810  * 1) differentiate between all holders of references to an extent so that
811  *    when a reference is dropped we can make sure it was a valid reference
812  *    before freeing the extent.
813  *
814  * 2) Provide enough information to quickly find the holders of an extent
815  *    if we notice a given block is corrupted or bad.
816  *
817  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
818  *    maintenance.  This is actually the same as #2, but with a slightly
819  *    different use case.
820  *
821  * There are two kinds of back refs. The implicit back refs is optimized
822  * for pointers in non-shared tree blocks. For a given pointer in a block,
823  * back refs of this kind provide information about the block's owner tree
824  * and the pointer's key. These information allow us to find the block by
825  * b-tree searching. The full back refs is for pointers in tree blocks not
826  * referenced by their owner trees. The location of tree block is recorded
827  * in the back refs. Actually the full back refs is generic, and can be
828  * used in all cases the implicit back refs is used. The major shortcoming
829  * of the full back refs is its overhead. Every time a tree block gets
830  * COWed, we have to update back refs entry for all pointers in it.
831  *
832  * For a newly allocated tree block, we use implicit back refs for
833  * pointers in it. This means most tree related operations only involve
834  * implicit back refs. For a tree block created in old transaction, the
835  * only way to drop a reference to it is COW it. So we can detect the
836  * event that tree block loses its owner tree's reference and do the
837  * back refs conversion.
838  *
839  * When a tree block is COW'd through a tree, there are four cases:
840  *
841  * The reference count of the block is one and the tree is the block's
842  * owner tree. Nothing to do in this case.
843  *
844  * The reference count of the block is one and the tree is not the
845  * block's owner tree. In this case, full back refs is used for pointers
846  * in the block. Remove these full back refs, add implicit back refs for
847  * every pointers in the new block.
848  *
849  * The reference count of the block is greater than one and the tree is
850  * the block's owner tree. In this case, implicit back refs is used for
851  * pointers in the block. Add full back refs for every pointers in the
852  * block, increase lower level extents' reference counts. The original
853  * implicit back refs are entailed to the new block.
854  *
855  * The reference count of the block is greater than one and the tree is
856  * not the block's owner tree. Add implicit back refs for every pointer in
857  * the new block, increase lower level extents' reference count.
858  *
859  * Back Reference Key composing:
860  *
861  * The key objectid corresponds to the first byte in the extent,
862  * The key type is used to differentiate between types of back refs.
863  * There are different meanings of the key offset for different types
864  * of back refs.
865  *
866  * File extents can be referenced by:
867  *
868  * - multiple snapshots, subvolumes, or different generations in one subvol
869  * - different files inside a single subvolume
870  * - different offsets inside a file (bookend extents in file.c)
871  *
872  * The extent ref structure for the implicit back refs has fields for:
873  *
874  * - Objectid of the subvolume root
875  * - objectid of the file holding the reference
876  * - original offset in the file
877  * - how many bookend extents
878  *
879  * The key offset for the implicit back refs is hash of the first
880  * three fields.
881  *
882  * The extent ref structure for the full back refs has field for:
883  *
884  * - number of pointers in the tree leaf
885  *
886  * The key offset for the implicit back refs is the first byte of
887  * the tree leaf
888  *
889  * When a file extent is allocated, The implicit back refs is used.
890  * the fields are filled in:
891  *
892  *     (root_key.objectid, inode objectid, offset in file, 1)
893  *
894  * When a file extent is removed file truncation, we find the
895  * corresponding implicit back refs and check the following fields:
896  *
897  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
898  *
899  * Btree extents can be referenced by:
900  *
901  * - Different subvolumes
902  *
903  * Both the implicit back refs and the full back refs for tree blocks
904  * only consist of key. The key offset for the implicit back refs is
905  * objectid of block's owner tree. The key offset for the full back refs
906  * is the first byte of parent block.
907  *
908  * When implicit back refs is used, information about the lowest key and
909  * level of the tree block are required. These information are stored in
910  * tree block info structure.
911  */
912
913 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
914 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
915                                   struct btrfs_root *root,
916                                   struct btrfs_path *path,
917                                   u64 owner, u32 extra_size)
918 {
919         struct btrfs_extent_item *item;
920         struct btrfs_extent_item_v0 *ei0;
921         struct btrfs_extent_ref_v0 *ref0;
922         struct btrfs_tree_block_info *bi;
923         struct extent_buffer *leaf;
924         struct btrfs_key key;
925         struct btrfs_key found_key;
926         u32 new_size = sizeof(*item);
927         u64 refs;
928         int ret;
929
930         leaf = path->nodes[0];
931         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
932
933         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
934         ei0 = btrfs_item_ptr(leaf, path->slots[0],
935                              struct btrfs_extent_item_v0);
936         refs = btrfs_extent_refs_v0(leaf, ei0);
937
938         if (owner == (u64)-1) {
939                 while (1) {
940                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
941                                 ret = btrfs_next_leaf(root, path);
942                                 if (ret < 0)
943                                         return ret;
944                                 BUG_ON(ret > 0);
945                                 leaf = path->nodes[0];
946                         }
947                         btrfs_item_key_to_cpu(leaf, &found_key,
948                                               path->slots[0]);
949                         BUG_ON(key.objectid != found_key.objectid);
950                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
951                                 path->slots[0]++;
952                                 continue;
953                         }
954                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
955                                               struct btrfs_extent_ref_v0);
956                         owner = btrfs_ref_objectid_v0(leaf, ref0);
957                         break;
958                 }
959         }
960         btrfs_release_path(path);
961
962         if (owner < BTRFS_FIRST_FREE_OBJECTID)
963                 new_size += sizeof(*bi);
964
965         new_size -= sizeof(*ei0);
966         ret = btrfs_search_slot(trans, root, &key, path,
967                                 new_size + extra_size, 1);
968         if (ret < 0)
969                 return ret;
970         BUG_ON(ret);
971
972         ret = btrfs_extend_item(trans, root, path, new_size);
973
974         leaf = path->nodes[0];
975         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
976         btrfs_set_extent_refs(leaf, item, refs);
977         /* FIXME: get real generation */
978         btrfs_set_extent_generation(leaf, item, 0);
979         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
980                 btrfs_set_extent_flags(leaf, item,
981                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
982                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
983                 bi = (struct btrfs_tree_block_info *)(item + 1);
984                 /* FIXME: get first key of the block */
985                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
986                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
987         } else {
988                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
989         }
990         btrfs_mark_buffer_dirty(leaf);
991         return 0;
992 }
993 #endif
994
995 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
996 {
997         u32 high_crc = ~(u32)0;
998         u32 low_crc = ~(u32)0;
999         __le64 lenum;
1000
1001         lenum = cpu_to_le64(root_objectid);
1002         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1003         lenum = cpu_to_le64(owner);
1004         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1005         lenum = cpu_to_le64(offset);
1006         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1007
1008         return ((u64)high_crc << 31) ^ (u64)low_crc;
1009 }
1010
1011 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1012                                      struct btrfs_extent_data_ref *ref)
1013 {
1014         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1015                                     btrfs_extent_data_ref_objectid(leaf, ref),
1016                                     btrfs_extent_data_ref_offset(leaf, ref));
1017 }
1018
1019 static int match_extent_data_ref(struct extent_buffer *leaf,
1020                                  struct btrfs_extent_data_ref *ref,
1021                                  u64 root_objectid, u64 owner, u64 offset)
1022 {
1023         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1024             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1025             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1026                 return 0;
1027         return 1;
1028 }
1029
1030 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1031                                            struct btrfs_root *root,
1032                                            struct btrfs_path *path,
1033                                            u64 bytenr, u64 parent,
1034                                            u64 root_objectid,
1035                                            u64 owner, u64 offset)
1036 {
1037         struct btrfs_key key;
1038         struct btrfs_extent_data_ref *ref;
1039         struct extent_buffer *leaf;
1040         u32 nritems;
1041         int ret;
1042         int recow;
1043         int err = -ENOENT;
1044
1045         key.objectid = bytenr;
1046         if (parent) {
1047                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1048                 key.offset = parent;
1049         } else {
1050                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1051                 key.offset = hash_extent_data_ref(root_objectid,
1052                                                   owner, offset);
1053         }
1054 again:
1055         recow = 0;
1056         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1057         if (ret < 0) {
1058                 err = ret;
1059                 goto fail;
1060         }
1061
1062         if (parent) {
1063                 if (!ret)
1064                         return 0;
1065 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1066                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1067                 btrfs_release_path(path);
1068                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1069                 if (ret < 0) {
1070                         err = ret;
1071                         goto fail;
1072                 }
1073                 if (!ret)
1074                         return 0;
1075 #endif
1076                 goto fail;
1077         }
1078
1079         leaf = path->nodes[0];
1080         nritems = btrfs_header_nritems(leaf);
1081         while (1) {
1082                 if (path->slots[0] >= nritems) {
1083                         ret = btrfs_next_leaf(root, path);
1084                         if (ret < 0)
1085                                 err = ret;
1086                         if (ret)
1087                                 goto fail;
1088
1089                         leaf = path->nodes[0];
1090                         nritems = btrfs_header_nritems(leaf);
1091                         recow = 1;
1092                 }
1093
1094                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1095                 if (key.objectid != bytenr ||
1096                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1097                         goto fail;
1098
1099                 ref = btrfs_item_ptr(leaf, path->slots[0],
1100                                      struct btrfs_extent_data_ref);
1101
1102                 if (match_extent_data_ref(leaf, ref, root_objectid,
1103                                           owner, offset)) {
1104                         if (recow) {
1105                                 btrfs_release_path(path);
1106                                 goto again;
1107                         }
1108                         err = 0;
1109                         break;
1110                 }
1111                 path->slots[0]++;
1112         }
1113 fail:
1114         return err;
1115 }
1116
1117 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1118                                            struct btrfs_root *root,
1119                                            struct btrfs_path *path,
1120                                            u64 bytenr, u64 parent,
1121                                            u64 root_objectid, u64 owner,
1122                                            u64 offset, int refs_to_add)
1123 {
1124         struct btrfs_key key;
1125         struct extent_buffer *leaf;
1126         u32 size;
1127         u32 num_refs;
1128         int ret;
1129
1130         key.objectid = bytenr;
1131         if (parent) {
1132                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1133                 key.offset = parent;
1134                 size = sizeof(struct btrfs_shared_data_ref);
1135         } else {
1136                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1137                 key.offset = hash_extent_data_ref(root_objectid,
1138                                                   owner, offset);
1139                 size = sizeof(struct btrfs_extent_data_ref);
1140         }
1141
1142         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1143         if (ret && ret != -EEXIST)
1144                 goto fail;
1145
1146         leaf = path->nodes[0];
1147         if (parent) {
1148                 struct btrfs_shared_data_ref *ref;
1149                 ref = btrfs_item_ptr(leaf, path->slots[0],
1150                                      struct btrfs_shared_data_ref);
1151                 if (ret == 0) {
1152                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1153                 } else {
1154                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1155                         num_refs += refs_to_add;
1156                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1157                 }
1158         } else {
1159                 struct btrfs_extent_data_ref *ref;
1160                 while (ret == -EEXIST) {
1161                         ref = btrfs_item_ptr(leaf, path->slots[0],
1162                                              struct btrfs_extent_data_ref);
1163                         if (match_extent_data_ref(leaf, ref, root_objectid,
1164                                                   owner, offset))
1165                                 break;
1166                         btrfs_release_path(path);
1167                         key.offset++;
1168                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1169                                                       size);
1170                         if (ret && ret != -EEXIST)
1171                                 goto fail;
1172
1173                         leaf = path->nodes[0];
1174                 }
1175                 ref = btrfs_item_ptr(leaf, path->slots[0],
1176                                      struct btrfs_extent_data_ref);
1177                 if (ret == 0) {
1178                         btrfs_set_extent_data_ref_root(leaf, ref,
1179                                                        root_objectid);
1180                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1181                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1182                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1183                 } else {
1184                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1185                         num_refs += refs_to_add;
1186                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1187                 }
1188         }
1189         btrfs_mark_buffer_dirty(leaf);
1190         ret = 0;
1191 fail:
1192         btrfs_release_path(path);
1193         return ret;
1194 }
1195
1196 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1197                                            struct btrfs_root *root,
1198                                            struct btrfs_path *path,
1199                                            int refs_to_drop)
1200 {
1201         struct btrfs_key key;
1202         struct btrfs_extent_data_ref *ref1 = NULL;
1203         struct btrfs_shared_data_ref *ref2 = NULL;
1204         struct extent_buffer *leaf;
1205         u32 num_refs = 0;
1206         int ret = 0;
1207
1208         leaf = path->nodes[0];
1209         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1210
1211         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1212                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1213                                       struct btrfs_extent_data_ref);
1214                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1215         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1216                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1217                                       struct btrfs_shared_data_ref);
1218                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1219 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1220         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1221                 struct btrfs_extent_ref_v0 *ref0;
1222                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1223                                       struct btrfs_extent_ref_v0);
1224                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1225 #endif
1226         } else {
1227                 BUG();
1228         }
1229
1230         BUG_ON(num_refs < refs_to_drop);
1231         num_refs -= refs_to_drop;
1232
1233         if (num_refs == 0) {
1234                 ret = btrfs_del_item(trans, root, path);
1235         } else {
1236                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1237                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1238                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1239                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1240 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1241                 else {
1242                         struct btrfs_extent_ref_v0 *ref0;
1243                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1244                                         struct btrfs_extent_ref_v0);
1245                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1246                 }
1247 #endif
1248                 btrfs_mark_buffer_dirty(leaf);
1249         }
1250         return ret;
1251 }
1252
1253 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1254                                           struct btrfs_path *path,
1255                                           struct btrfs_extent_inline_ref *iref)
1256 {
1257         struct btrfs_key key;
1258         struct extent_buffer *leaf;
1259         struct btrfs_extent_data_ref *ref1;
1260         struct btrfs_shared_data_ref *ref2;
1261         u32 num_refs = 0;
1262
1263         leaf = path->nodes[0];
1264         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1265         if (iref) {
1266                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1267                     BTRFS_EXTENT_DATA_REF_KEY) {
1268                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1269                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1270                 } else {
1271                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1272                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1273                 }
1274         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1275                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1276                                       struct btrfs_extent_data_ref);
1277                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1278         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1279                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1280                                       struct btrfs_shared_data_ref);
1281                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1284                 struct btrfs_extent_ref_v0 *ref0;
1285                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                       struct btrfs_extent_ref_v0);
1287                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1288 #endif
1289         } else {
1290                 WARN_ON(1);
1291         }
1292         return num_refs;
1293 }
1294
1295 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1296                                           struct btrfs_root *root,
1297                                           struct btrfs_path *path,
1298                                           u64 bytenr, u64 parent,
1299                                           u64 root_objectid)
1300 {
1301         struct btrfs_key key;
1302         int ret;
1303
1304         key.objectid = bytenr;
1305         if (parent) {
1306                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1307                 key.offset = parent;
1308         } else {
1309                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1310                 key.offset = root_objectid;
1311         }
1312
1313         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1314         if (ret > 0)
1315                 ret = -ENOENT;
1316 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1317         if (ret == -ENOENT && parent) {
1318                 btrfs_release_path(path);
1319                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1320                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1321                 if (ret > 0)
1322                         ret = -ENOENT;
1323         }
1324 #endif
1325         return ret;
1326 }
1327
1328 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1329                                           struct btrfs_root *root,
1330                                           struct btrfs_path *path,
1331                                           u64 bytenr, u64 parent,
1332                                           u64 root_objectid)
1333 {
1334         struct btrfs_key key;
1335         int ret;
1336
1337         key.objectid = bytenr;
1338         if (parent) {
1339                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1340                 key.offset = parent;
1341         } else {
1342                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1343                 key.offset = root_objectid;
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1347         btrfs_release_path(path);
1348         return ret;
1349 }
1350
1351 static inline int extent_ref_type(u64 parent, u64 owner)
1352 {
1353         int type;
1354         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1355                 if (parent > 0)
1356                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1357                 else
1358                         type = BTRFS_TREE_BLOCK_REF_KEY;
1359         } else {
1360                 if (parent > 0)
1361                         type = BTRFS_SHARED_DATA_REF_KEY;
1362                 else
1363                         type = BTRFS_EXTENT_DATA_REF_KEY;
1364         }
1365         return type;
1366 }
1367
1368 static int find_next_key(struct btrfs_path *path, int level,
1369                          struct btrfs_key *key)
1370
1371 {
1372         for (; level < BTRFS_MAX_LEVEL; level++) {
1373                 if (!path->nodes[level])
1374                         break;
1375                 if (path->slots[level] + 1 >=
1376                     btrfs_header_nritems(path->nodes[level]))
1377                         continue;
1378                 if (level == 0)
1379                         btrfs_item_key_to_cpu(path->nodes[level], key,
1380                                               path->slots[level] + 1);
1381                 else
1382                         btrfs_node_key_to_cpu(path->nodes[level], key,
1383                                               path->slots[level] + 1);
1384                 return 0;
1385         }
1386         return 1;
1387 }
1388
1389 /*
1390  * look for inline back ref. if back ref is found, *ref_ret is set
1391  * to the address of inline back ref, and 0 is returned.
1392  *
1393  * if back ref isn't found, *ref_ret is set to the address where it
1394  * should be inserted, and -ENOENT is returned.
1395  *
1396  * if insert is true and there are too many inline back refs, the path
1397  * points to the extent item, and -EAGAIN is returned.
1398  *
1399  * NOTE: inline back refs are ordered in the same way that back ref
1400  *       items in the tree are ordered.
1401  */
1402 static noinline_for_stack
1403 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1404                                  struct btrfs_root *root,
1405                                  struct btrfs_path *path,
1406                                  struct btrfs_extent_inline_ref **ref_ret,
1407                                  u64 bytenr, u64 num_bytes,
1408                                  u64 parent, u64 root_objectid,
1409                                  u64 owner, u64 offset, int insert)
1410 {
1411         struct btrfs_key key;
1412         struct extent_buffer *leaf;
1413         struct btrfs_extent_item *ei;
1414         struct btrfs_extent_inline_ref *iref;
1415         u64 flags;
1416         u64 item_size;
1417         unsigned long ptr;
1418         unsigned long end;
1419         int extra_size;
1420         int type;
1421         int want;
1422         int ret;
1423         int err = 0;
1424
1425         key.objectid = bytenr;
1426         key.type = BTRFS_EXTENT_ITEM_KEY;
1427         key.offset = num_bytes;
1428
1429         want = extent_ref_type(parent, owner);
1430         if (insert) {
1431                 extra_size = btrfs_extent_inline_ref_size(want);
1432                 path->keep_locks = 1;
1433         } else
1434                 extra_size = -1;
1435         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1436         if (ret < 0) {
1437                 err = ret;
1438                 goto out;
1439         }
1440         BUG_ON(ret);
1441
1442         leaf = path->nodes[0];
1443         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1444 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1445         if (item_size < sizeof(*ei)) {
1446                 if (!insert) {
1447                         err = -ENOENT;
1448                         goto out;
1449                 }
1450                 ret = convert_extent_item_v0(trans, root, path, owner,
1451                                              extra_size);
1452                 if (ret < 0) {
1453                         err = ret;
1454                         goto out;
1455                 }
1456                 leaf = path->nodes[0];
1457                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1458         }
1459 #endif
1460         BUG_ON(item_size < sizeof(*ei));
1461
1462         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1463         flags = btrfs_extent_flags(leaf, ei);
1464
1465         ptr = (unsigned long)(ei + 1);
1466         end = (unsigned long)ei + item_size;
1467
1468         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1469                 ptr += sizeof(struct btrfs_tree_block_info);
1470                 BUG_ON(ptr > end);
1471         } else {
1472                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1473         }
1474
1475         err = -ENOENT;
1476         while (1) {
1477                 if (ptr >= end) {
1478                         WARN_ON(ptr > end);
1479                         break;
1480                 }
1481                 iref = (struct btrfs_extent_inline_ref *)ptr;
1482                 type = btrfs_extent_inline_ref_type(leaf, iref);
1483                 if (want < type)
1484                         break;
1485                 if (want > type) {
1486                         ptr += btrfs_extent_inline_ref_size(type);
1487                         continue;
1488                 }
1489
1490                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1491                         struct btrfs_extent_data_ref *dref;
1492                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1493                         if (match_extent_data_ref(leaf, dref, root_objectid,
1494                                                   owner, offset)) {
1495                                 err = 0;
1496                                 break;
1497                         }
1498                         if (hash_extent_data_ref_item(leaf, dref) <
1499                             hash_extent_data_ref(root_objectid, owner, offset))
1500                                 break;
1501                 } else {
1502                         u64 ref_offset;
1503                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1504                         if (parent > 0) {
1505                                 if (parent == ref_offset) {
1506                                         err = 0;
1507                                         break;
1508                                 }
1509                                 if (ref_offset < parent)
1510                                         break;
1511                         } else {
1512                                 if (root_objectid == ref_offset) {
1513                                         err = 0;
1514                                         break;
1515                                 }
1516                                 if (ref_offset < root_objectid)
1517                                         break;
1518                         }
1519                 }
1520                 ptr += btrfs_extent_inline_ref_size(type);
1521         }
1522         if (err == -ENOENT && insert) {
1523                 if (item_size + extra_size >=
1524                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1525                         err = -EAGAIN;
1526                         goto out;
1527                 }
1528                 /*
1529                  * To add new inline back ref, we have to make sure
1530                  * there is no corresponding back ref item.
1531                  * For simplicity, we just do not add new inline back
1532                  * ref if there is any kind of item for this block
1533                  */
1534                 if (find_next_key(path, 0, &key) == 0 &&
1535                     key.objectid == bytenr &&
1536                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1537                         err = -EAGAIN;
1538                         goto out;
1539                 }
1540         }
1541         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1542 out:
1543         if (insert) {
1544                 path->keep_locks = 0;
1545                 btrfs_unlock_up_safe(path, 1);
1546         }
1547         return err;
1548 }
1549
1550 /*
1551  * helper to add new inline back ref
1552  */
1553 static noinline_for_stack
1554 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1555                                 struct btrfs_root *root,
1556                                 struct btrfs_path *path,
1557                                 struct btrfs_extent_inline_ref *iref,
1558                                 u64 parent, u64 root_objectid,
1559                                 u64 owner, u64 offset, int refs_to_add,
1560                                 struct btrfs_delayed_extent_op *extent_op)
1561 {
1562         struct extent_buffer *leaf;
1563         struct btrfs_extent_item *ei;
1564         unsigned long ptr;
1565         unsigned long end;
1566         unsigned long item_offset;
1567         u64 refs;
1568         int size;
1569         int type;
1570         int ret;
1571
1572         leaf = path->nodes[0];
1573         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1574         item_offset = (unsigned long)iref - (unsigned long)ei;
1575
1576         type = extent_ref_type(parent, owner);
1577         size = btrfs_extent_inline_ref_size(type);
1578
1579         ret = btrfs_extend_item(trans, root, path, size);
1580
1581         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1582         refs = btrfs_extent_refs(leaf, ei);
1583         refs += refs_to_add;
1584         btrfs_set_extent_refs(leaf, ei, refs);
1585         if (extent_op)
1586                 __run_delayed_extent_op(extent_op, leaf, ei);
1587
1588         ptr = (unsigned long)ei + item_offset;
1589         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1590         if (ptr < end - size)
1591                 memmove_extent_buffer(leaf, ptr + size, ptr,
1592                                       end - size - ptr);
1593
1594         iref = (struct btrfs_extent_inline_ref *)ptr;
1595         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1596         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1597                 struct btrfs_extent_data_ref *dref;
1598                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1599                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1600                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1601                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1602                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1603         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1604                 struct btrfs_shared_data_ref *sref;
1605                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1606                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1607                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1608         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1609                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1610         } else {
1611                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1612         }
1613         btrfs_mark_buffer_dirty(leaf);
1614         return 0;
1615 }
1616
1617 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1618                                  struct btrfs_root *root,
1619                                  struct btrfs_path *path,
1620                                  struct btrfs_extent_inline_ref **ref_ret,
1621                                  u64 bytenr, u64 num_bytes, u64 parent,
1622                                  u64 root_objectid, u64 owner, u64 offset)
1623 {
1624         int ret;
1625
1626         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1627                                            bytenr, num_bytes, parent,
1628                                            root_objectid, owner, offset, 0);
1629         if (ret != -ENOENT)
1630                 return ret;
1631
1632         btrfs_release_path(path);
1633         *ref_ret = NULL;
1634
1635         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1636                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1637                                             root_objectid);
1638         } else {
1639                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1640                                              root_objectid, owner, offset);
1641         }
1642         return ret;
1643 }
1644
1645 /*
1646  * helper to update/remove inline back ref
1647  */
1648 static noinline_for_stack
1649 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1650                                  struct btrfs_root *root,
1651                                  struct btrfs_path *path,
1652                                  struct btrfs_extent_inline_ref *iref,
1653                                  int refs_to_mod,
1654                                  struct btrfs_delayed_extent_op *extent_op)
1655 {
1656         struct extent_buffer *leaf;
1657         struct btrfs_extent_item *ei;
1658         struct btrfs_extent_data_ref *dref = NULL;
1659         struct btrfs_shared_data_ref *sref = NULL;
1660         unsigned long ptr;
1661         unsigned long end;
1662         u32 item_size;
1663         int size;
1664         int type;
1665         int ret;
1666         u64 refs;
1667
1668         leaf = path->nodes[0];
1669         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1670         refs = btrfs_extent_refs(leaf, ei);
1671         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1672         refs += refs_to_mod;
1673         btrfs_set_extent_refs(leaf, ei, refs);
1674         if (extent_op)
1675                 __run_delayed_extent_op(extent_op, leaf, ei);
1676
1677         type = btrfs_extent_inline_ref_type(leaf, iref);
1678
1679         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1680                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1681                 refs = btrfs_extent_data_ref_count(leaf, dref);
1682         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1683                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1684                 refs = btrfs_shared_data_ref_count(leaf, sref);
1685         } else {
1686                 refs = 1;
1687                 BUG_ON(refs_to_mod != -1);
1688         }
1689
1690         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1691         refs += refs_to_mod;
1692
1693         if (refs > 0) {
1694                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1695                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1696                 else
1697                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1698         } else {
1699                 size =  btrfs_extent_inline_ref_size(type);
1700                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701                 ptr = (unsigned long)iref;
1702                 end = (unsigned long)ei + item_size;
1703                 if (ptr + size < end)
1704                         memmove_extent_buffer(leaf, ptr, ptr + size,
1705                                               end - ptr - size);
1706                 item_size -= size;
1707                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1708         }
1709         btrfs_mark_buffer_dirty(leaf);
1710         return 0;
1711 }
1712
1713 static noinline_for_stack
1714 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1715                                  struct btrfs_root *root,
1716                                  struct btrfs_path *path,
1717                                  u64 bytenr, u64 num_bytes, u64 parent,
1718                                  u64 root_objectid, u64 owner,
1719                                  u64 offset, int refs_to_add,
1720                                  struct btrfs_delayed_extent_op *extent_op)
1721 {
1722         struct btrfs_extent_inline_ref *iref;
1723         int ret;
1724
1725         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1726                                            bytenr, num_bytes, parent,
1727                                            root_objectid, owner, offset, 1);
1728         if (ret == 0) {
1729                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1730                 ret = update_inline_extent_backref(trans, root, path, iref,
1731                                                    refs_to_add, extent_op);
1732         } else if (ret == -ENOENT) {
1733                 ret = setup_inline_extent_backref(trans, root, path, iref,
1734                                                   parent, root_objectid,
1735                                                   owner, offset, refs_to_add,
1736                                                   extent_op);
1737         }
1738         return ret;
1739 }
1740
1741 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1742                                  struct btrfs_root *root,
1743                                  struct btrfs_path *path,
1744                                  u64 bytenr, u64 parent, u64 root_objectid,
1745                                  u64 owner, u64 offset, int refs_to_add)
1746 {
1747         int ret;
1748         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1749                 BUG_ON(refs_to_add != 1);
1750                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1751                                             parent, root_objectid);
1752         } else {
1753                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1754                                              parent, root_objectid,
1755                                              owner, offset, refs_to_add);
1756         }
1757         return ret;
1758 }
1759
1760 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1761                                  struct btrfs_root *root,
1762                                  struct btrfs_path *path,
1763                                  struct btrfs_extent_inline_ref *iref,
1764                                  int refs_to_drop, int is_data)
1765 {
1766         int ret;
1767
1768         BUG_ON(!is_data && refs_to_drop != 1);
1769         if (iref) {
1770                 ret = update_inline_extent_backref(trans, root, path, iref,
1771                                                    -refs_to_drop, NULL);
1772         } else if (is_data) {
1773                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1774         } else {
1775                 ret = btrfs_del_item(trans, root, path);
1776         }
1777         return ret;
1778 }
1779
1780 static int btrfs_issue_discard(struct block_device *bdev,
1781                                 u64 start, u64 len)
1782 {
1783         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1784 }
1785
1786 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1787                                 u64 num_bytes, u64 *actual_bytes)
1788 {
1789         int ret;
1790         u64 discarded_bytes = 0;
1791         struct btrfs_bio *bbio = NULL;
1792
1793
1794         /* Tell the block device(s) that the sectors can be discarded */
1795         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1796                               bytenr, &num_bytes, &bbio, 0);
1797         if (!ret) {
1798                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1799                 int i;
1800
1801
1802                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1803                         if (!stripe->dev->can_discard)
1804                                 continue;
1805
1806                         ret = btrfs_issue_discard(stripe->dev->bdev,
1807                                                   stripe->physical,
1808                                                   stripe->length);
1809                         if (!ret)
1810                                 discarded_bytes += stripe->length;
1811                         else if (ret != -EOPNOTSUPP)
1812                                 break;
1813
1814                         /*
1815                          * Just in case we get back EOPNOTSUPP for some reason,
1816                          * just ignore the return value so we don't screw up
1817                          * people calling discard_extent.
1818                          */
1819                         ret = 0;
1820                 }
1821                 kfree(bbio);
1822         }
1823
1824         if (actual_bytes)
1825                 *actual_bytes = discarded_bytes;
1826
1827
1828         return ret;
1829 }
1830
1831 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1832                          struct btrfs_root *root,
1833                          u64 bytenr, u64 num_bytes, u64 parent,
1834                          u64 root_objectid, u64 owner, u64 offset)
1835 {
1836         int ret;
1837         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1838                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1839
1840         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1841                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1842                                         parent, root_objectid, (int)owner,
1843                                         BTRFS_ADD_DELAYED_REF, NULL);
1844         } else {
1845                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1846                                         parent, root_objectid, owner, offset,
1847                                         BTRFS_ADD_DELAYED_REF, NULL);
1848         }
1849         return ret;
1850 }
1851
1852 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1853                                   struct btrfs_root *root,
1854                                   u64 bytenr, u64 num_bytes,
1855                                   u64 parent, u64 root_objectid,
1856                                   u64 owner, u64 offset, int refs_to_add,
1857                                   struct btrfs_delayed_extent_op *extent_op)
1858 {
1859         struct btrfs_path *path;
1860         struct extent_buffer *leaf;
1861         struct btrfs_extent_item *item;
1862         u64 refs;
1863         int ret;
1864         int err = 0;
1865
1866         path = btrfs_alloc_path();
1867         if (!path)
1868                 return -ENOMEM;
1869
1870         path->reada = 1;
1871         path->leave_spinning = 1;
1872         /* this will setup the path even if it fails to insert the back ref */
1873         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1874                                            path, bytenr, num_bytes, parent,
1875                                            root_objectid, owner, offset,
1876                                            refs_to_add, extent_op);
1877         if (ret == 0)
1878                 goto out;
1879
1880         if (ret != -EAGAIN) {
1881                 err = ret;
1882                 goto out;
1883         }
1884
1885         leaf = path->nodes[0];
1886         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1887         refs = btrfs_extent_refs(leaf, item);
1888         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1889         if (extent_op)
1890                 __run_delayed_extent_op(extent_op, leaf, item);
1891
1892         btrfs_mark_buffer_dirty(leaf);
1893         btrfs_release_path(path);
1894
1895         path->reada = 1;
1896         path->leave_spinning = 1;
1897
1898         /* now insert the actual backref */
1899         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1900                                     path, bytenr, parent, root_objectid,
1901                                     owner, offset, refs_to_add);
1902         BUG_ON(ret);
1903 out:
1904         btrfs_free_path(path);
1905         return err;
1906 }
1907
1908 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1909                                 struct btrfs_root *root,
1910                                 struct btrfs_delayed_ref_node *node,
1911                                 struct btrfs_delayed_extent_op *extent_op,
1912                                 int insert_reserved)
1913 {
1914         int ret = 0;
1915         struct btrfs_delayed_data_ref *ref;
1916         struct btrfs_key ins;
1917         u64 parent = 0;
1918         u64 ref_root = 0;
1919         u64 flags = 0;
1920
1921         ins.objectid = node->bytenr;
1922         ins.offset = node->num_bytes;
1923         ins.type = BTRFS_EXTENT_ITEM_KEY;
1924
1925         ref = btrfs_delayed_node_to_data_ref(node);
1926         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1927                 parent = ref->parent;
1928         else
1929                 ref_root = ref->root;
1930
1931         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1932                 if (extent_op) {
1933                         BUG_ON(extent_op->update_key);
1934                         flags |= extent_op->flags_to_set;
1935                 }
1936                 ret = alloc_reserved_file_extent(trans, root,
1937                                                  parent, ref_root, flags,
1938                                                  ref->objectid, ref->offset,
1939                                                  &ins, node->ref_mod);
1940         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1941                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1942                                              node->num_bytes, parent,
1943                                              ref_root, ref->objectid,
1944                                              ref->offset, node->ref_mod,
1945                                              extent_op);
1946         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1947                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1948                                           node->num_bytes, parent,
1949                                           ref_root, ref->objectid,
1950                                           ref->offset, node->ref_mod,
1951                                           extent_op);
1952         } else {
1953                 BUG();
1954         }
1955         return ret;
1956 }
1957
1958 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1959                                     struct extent_buffer *leaf,
1960                                     struct btrfs_extent_item *ei)
1961 {
1962         u64 flags = btrfs_extent_flags(leaf, ei);
1963         if (extent_op->update_flags) {
1964                 flags |= extent_op->flags_to_set;
1965                 btrfs_set_extent_flags(leaf, ei, flags);
1966         }
1967
1968         if (extent_op->update_key) {
1969                 struct btrfs_tree_block_info *bi;
1970                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1971                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1972                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1973         }
1974 }
1975
1976 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1977                                  struct btrfs_root *root,
1978                                  struct btrfs_delayed_ref_node *node,
1979                                  struct btrfs_delayed_extent_op *extent_op)
1980 {
1981         struct btrfs_key key;
1982         struct btrfs_path *path;
1983         struct btrfs_extent_item *ei;
1984         struct extent_buffer *leaf;
1985         u32 item_size;
1986         int ret;
1987         int err = 0;
1988
1989         path = btrfs_alloc_path();
1990         if (!path)
1991                 return -ENOMEM;
1992
1993         key.objectid = node->bytenr;
1994         key.type = BTRFS_EXTENT_ITEM_KEY;
1995         key.offset = node->num_bytes;
1996
1997         path->reada = 1;
1998         path->leave_spinning = 1;
1999         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2000                                 path, 0, 1);
2001         if (ret < 0) {
2002                 err = ret;
2003                 goto out;
2004         }
2005         if (ret > 0) {
2006                 err = -EIO;
2007                 goto out;
2008         }
2009
2010         leaf = path->nodes[0];
2011         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2012 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2013         if (item_size < sizeof(*ei)) {
2014                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2015                                              path, (u64)-1, 0);
2016                 if (ret < 0) {
2017                         err = ret;
2018                         goto out;
2019                 }
2020                 leaf = path->nodes[0];
2021                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2022         }
2023 #endif
2024         BUG_ON(item_size < sizeof(*ei));
2025         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2026         __run_delayed_extent_op(extent_op, leaf, ei);
2027
2028         btrfs_mark_buffer_dirty(leaf);
2029 out:
2030         btrfs_free_path(path);
2031         return err;
2032 }
2033
2034 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2035                                 struct btrfs_root *root,
2036                                 struct btrfs_delayed_ref_node *node,
2037                                 struct btrfs_delayed_extent_op *extent_op,
2038                                 int insert_reserved)
2039 {
2040         int ret = 0;
2041         struct btrfs_delayed_tree_ref *ref;
2042         struct btrfs_key ins;
2043         u64 parent = 0;
2044         u64 ref_root = 0;
2045
2046         ins.objectid = node->bytenr;
2047         ins.offset = node->num_bytes;
2048         ins.type = BTRFS_EXTENT_ITEM_KEY;
2049
2050         ref = btrfs_delayed_node_to_tree_ref(node);
2051         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2052                 parent = ref->parent;
2053         else
2054                 ref_root = ref->root;
2055
2056         BUG_ON(node->ref_mod != 1);
2057         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2058                 BUG_ON(!extent_op || !extent_op->update_flags ||
2059                        !extent_op->update_key);
2060                 ret = alloc_reserved_tree_block(trans, root,
2061                                                 parent, ref_root,
2062                                                 extent_op->flags_to_set,
2063                                                 &extent_op->key,
2064                                                 ref->level, &ins);
2065         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2066                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2067                                              node->num_bytes, parent, ref_root,
2068                                              ref->level, 0, 1, extent_op);
2069         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2070                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2071                                           node->num_bytes, parent, ref_root,
2072                                           ref->level, 0, 1, extent_op);
2073         } else {
2074                 BUG();
2075         }
2076         return ret;
2077 }
2078
2079 /* helper function to actually process a single delayed ref entry */
2080 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2081                                struct btrfs_root *root,
2082                                struct btrfs_delayed_ref_node *node,
2083                                struct btrfs_delayed_extent_op *extent_op,
2084                                int insert_reserved)
2085 {
2086         int ret;
2087         if (btrfs_delayed_ref_is_head(node)) {
2088                 struct btrfs_delayed_ref_head *head;
2089                 /*
2090                  * we've hit the end of the chain and we were supposed
2091                  * to insert this extent into the tree.  But, it got
2092                  * deleted before we ever needed to insert it, so all
2093                  * we have to do is clean up the accounting
2094                  */
2095                 BUG_ON(extent_op);
2096                 head = btrfs_delayed_node_to_head(node);
2097                 if (insert_reserved) {
2098                         btrfs_pin_extent(root, node->bytenr,
2099                                          node->num_bytes, 1);
2100                         if (head->is_data) {
2101                                 ret = btrfs_del_csums(trans, root,
2102                                                       node->bytenr,
2103                                                       node->num_bytes);
2104                                 BUG_ON(ret);
2105                         }
2106                 }
2107                 mutex_unlock(&head->mutex);
2108                 return 0;
2109         }
2110
2111         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2112             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2113                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2114                                            insert_reserved);
2115         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2116                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2117                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2118                                            insert_reserved);
2119         else
2120                 BUG();
2121         return ret;
2122 }
2123
2124 static noinline struct btrfs_delayed_ref_node *
2125 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2126 {
2127         struct rb_node *node;
2128         struct btrfs_delayed_ref_node *ref;
2129         int action = BTRFS_ADD_DELAYED_REF;
2130 again:
2131         /*
2132          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2133          * this prevents ref count from going down to zero when
2134          * there still are pending delayed ref.
2135          */
2136         node = rb_prev(&head->node.rb_node);
2137         while (1) {
2138                 if (!node)
2139                         break;
2140                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2141                                 rb_node);
2142                 if (ref->bytenr != head->node.bytenr)
2143                         break;
2144                 if (ref->action == action)
2145                         return ref;
2146                 node = rb_prev(node);
2147         }
2148         if (action == BTRFS_ADD_DELAYED_REF) {
2149                 action = BTRFS_DROP_DELAYED_REF;
2150                 goto again;
2151         }
2152         return NULL;
2153 }
2154
2155 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2156                                        struct btrfs_root *root,
2157                                        struct list_head *cluster)
2158 {
2159         struct btrfs_delayed_ref_root *delayed_refs;
2160         struct btrfs_delayed_ref_node *ref;
2161         struct btrfs_delayed_ref_head *locked_ref = NULL;
2162         struct btrfs_delayed_extent_op *extent_op;
2163         int ret;
2164         int count = 0;
2165         int must_insert_reserved = 0;
2166
2167         delayed_refs = &trans->transaction->delayed_refs;
2168         while (1) {
2169                 if (!locked_ref) {
2170                         /* pick a new head ref from the cluster list */
2171                         if (list_empty(cluster))
2172                                 break;
2173
2174                         locked_ref = list_entry(cluster->next,
2175                                      struct btrfs_delayed_ref_head, cluster);
2176
2177                         /* grab the lock that says we are going to process
2178                          * all the refs for this head */
2179                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2180
2181                         /*
2182                          * we may have dropped the spin lock to get the head
2183                          * mutex lock, and that might have given someone else
2184                          * time to free the head.  If that's true, it has been
2185                          * removed from our list and we can move on.
2186                          */
2187                         if (ret == -EAGAIN) {
2188                                 locked_ref = NULL;
2189                                 count++;
2190                                 continue;
2191                         }
2192                 }
2193
2194                 /*
2195                  * record the must insert reserved flag before we
2196                  * drop the spin lock.
2197                  */
2198                 must_insert_reserved = locked_ref->must_insert_reserved;
2199                 locked_ref->must_insert_reserved = 0;
2200
2201                 extent_op = locked_ref->extent_op;
2202                 locked_ref->extent_op = NULL;
2203
2204                 /*
2205                  * locked_ref is the head node, so we have to go one
2206                  * node back for any delayed ref updates
2207                  */
2208                 ref = select_delayed_ref(locked_ref);
2209                 if (!ref) {
2210                         /* All delayed refs have been processed, Go ahead
2211                          * and send the head node to run_one_delayed_ref,
2212                          * so that any accounting fixes can happen
2213                          */
2214                         ref = &locked_ref->node;
2215
2216                         if (extent_op && must_insert_reserved) {
2217                                 kfree(extent_op);
2218                                 extent_op = NULL;
2219                         }
2220
2221                         if (extent_op) {
2222                                 spin_unlock(&delayed_refs->lock);
2223
2224                                 ret = run_delayed_extent_op(trans, root,
2225                                                             ref, extent_op);
2226                                 BUG_ON(ret);
2227                                 kfree(extent_op);
2228
2229                                 cond_resched();
2230                                 spin_lock(&delayed_refs->lock);
2231                                 continue;
2232                         }
2233
2234                         list_del_init(&locked_ref->cluster);
2235                         locked_ref = NULL;
2236                 }
2237
2238                 ref->in_tree = 0;
2239                 rb_erase(&ref->rb_node, &delayed_refs->root);
2240                 delayed_refs->num_entries--;
2241
2242                 spin_unlock(&delayed_refs->lock);
2243
2244                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2245                                           must_insert_reserved);
2246                 BUG_ON(ret);
2247
2248                 btrfs_put_delayed_ref(ref);
2249                 kfree(extent_op);
2250                 count++;
2251
2252                 cond_resched();
2253                 spin_lock(&delayed_refs->lock);
2254         }
2255         return count;
2256 }
2257
2258 /*
2259  * this starts processing the delayed reference count updates and
2260  * extent insertions we have queued up so far.  count can be
2261  * 0, which means to process everything in the tree at the start
2262  * of the run (but not newly added entries), or it can be some target
2263  * number you'd like to process.
2264  */
2265 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2266                            struct btrfs_root *root, unsigned long count)
2267 {
2268         struct rb_node *node;
2269         struct btrfs_delayed_ref_root *delayed_refs;
2270         struct btrfs_delayed_ref_node *ref;
2271         struct list_head cluster;
2272         int ret;
2273         int run_all = count == (unsigned long)-1;
2274         int run_most = 0;
2275
2276         if (root == root->fs_info->extent_root)
2277                 root = root->fs_info->tree_root;
2278
2279         delayed_refs = &trans->transaction->delayed_refs;
2280         INIT_LIST_HEAD(&cluster);
2281 again:
2282         spin_lock(&delayed_refs->lock);
2283         if (count == 0) {
2284                 count = delayed_refs->num_entries * 2;
2285                 run_most = 1;
2286         }
2287         while (1) {
2288                 if (!(run_all || run_most) &&
2289                     delayed_refs->num_heads_ready < 64)
2290                         break;
2291
2292                 /*
2293                  * go find something we can process in the rbtree.  We start at
2294                  * the beginning of the tree, and then build a cluster
2295                  * of refs to process starting at the first one we are able to
2296                  * lock
2297                  */
2298                 ret = btrfs_find_ref_cluster(trans, &cluster,
2299                                              delayed_refs->run_delayed_start);
2300                 if (ret)
2301                         break;
2302
2303                 ret = run_clustered_refs(trans, root, &cluster);
2304                 BUG_ON(ret < 0);
2305
2306                 count -= min_t(unsigned long, ret, count);
2307
2308                 if (count == 0)
2309                         break;
2310         }
2311
2312         if (run_all) {
2313                 node = rb_first(&delayed_refs->root);
2314                 if (!node)
2315                         goto out;
2316                 count = (unsigned long)-1;
2317
2318                 while (node) {
2319                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2320                                        rb_node);
2321                         if (btrfs_delayed_ref_is_head(ref)) {
2322                                 struct btrfs_delayed_ref_head *head;
2323
2324                                 head = btrfs_delayed_node_to_head(ref);
2325                                 atomic_inc(&ref->refs);
2326
2327                                 spin_unlock(&delayed_refs->lock);
2328                                 /*
2329                                  * Mutex was contended, block until it's
2330                                  * released and try again
2331                                  */
2332                                 mutex_lock(&head->mutex);
2333                                 mutex_unlock(&head->mutex);
2334
2335                                 btrfs_put_delayed_ref(ref);
2336                                 cond_resched();
2337                                 goto again;
2338                         }
2339                         node = rb_next(node);
2340                 }
2341                 spin_unlock(&delayed_refs->lock);
2342                 schedule_timeout(1);
2343                 goto again;
2344         }
2345 out:
2346         spin_unlock(&delayed_refs->lock);
2347         return 0;
2348 }
2349
2350 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2351                                 struct btrfs_root *root,
2352                                 u64 bytenr, u64 num_bytes, u64 flags,
2353                                 int is_data)
2354 {
2355         struct btrfs_delayed_extent_op *extent_op;
2356         int ret;
2357
2358         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2359         if (!extent_op)
2360                 return -ENOMEM;
2361
2362         extent_op->flags_to_set = flags;
2363         extent_op->update_flags = 1;
2364         extent_op->update_key = 0;
2365         extent_op->is_data = is_data ? 1 : 0;
2366
2367         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2368         if (ret)
2369                 kfree(extent_op);
2370         return ret;
2371 }
2372
2373 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2374                                       struct btrfs_root *root,
2375                                       struct btrfs_path *path,
2376                                       u64 objectid, u64 offset, u64 bytenr)
2377 {
2378         struct btrfs_delayed_ref_head *head;
2379         struct btrfs_delayed_ref_node *ref;
2380         struct btrfs_delayed_data_ref *data_ref;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct rb_node *node;
2383         int ret = 0;
2384
2385         ret = -ENOENT;
2386         delayed_refs = &trans->transaction->delayed_refs;
2387         spin_lock(&delayed_refs->lock);
2388         head = btrfs_find_delayed_ref_head(trans, bytenr);
2389         if (!head)
2390                 goto out;
2391
2392         if (!mutex_trylock(&head->mutex)) {
2393                 atomic_inc(&head->node.refs);
2394                 spin_unlock(&delayed_refs->lock);
2395
2396                 btrfs_release_path(path);
2397
2398                 /*
2399                  * Mutex was contended, block until it's released and let
2400                  * caller try again
2401                  */
2402                 mutex_lock(&head->mutex);
2403                 mutex_unlock(&head->mutex);
2404                 btrfs_put_delayed_ref(&head->node);
2405                 return -EAGAIN;
2406         }
2407
2408         node = rb_prev(&head->node.rb_node);
2409         if (!node)
2410                 goto out_unlock;
2411
2412         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2413
2414         if (ref->bytenr != bytenr)
2415                 goto out_unlock;
2416
2417         ret = 1;
2418         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2419                 goto out_unlock;
2420
2421         data_ref = btrfs_delayed_node_to_data_ref(ref);
2422
2423         node = rb_prev(node);
2424         if (node) {
2425                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2426                 if (ref->bytenr == bytenr)
2427                         goto out_unlock;
2428         }
2429
2430         if (data_ref->root != root->root_key.objectid ||
2431             data_ref->objectid != objectid || data_ref->offset != offset)
2432                 goto out_unlock;
2433
2434         ret = 0;
2435 out_unlock:
2436         mutex_unlock(&head->mutex);
2437 out:
2438         spin_unlock(&delayed_refs->lock);
2439         return ret;
2440 }
2441
2442 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2443                                         struct btrfs_root *root,
2444                                         struct btrfs_path *path,
2445                                         u64 objectid, u64 offset, u64 bytenr)
2446 {
2447         struct btrfs_root *extent_root = root->fs_info->extent_root;
2448         struct extent_buffer *leaf;
2449         struct btrfs_extent_data_ref *ref;
2450         struct btrfs_extent_inline_ref *iref;
2451         struct btrfs_extent_item *ei;
2452         struct btrfs_key key;
2453         u32 item_size;
2454         int ret;
2455
2456         key.objectid = bytenr;
2457         key.offset = (u64)-1;
2458         key.type = BTRFS_EXTENT_ITEM_KEY;
2459
2460         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2461         if (ret < 0)
2462                 goto out;
2463         BUG_ON(ret == 0);
2464
2465         ret = -ENOENT;
2466         if (path->slots[0] == 0)
2467                 goto out;
2468
2469         path->slots[0]--;
2470         leaf = path->nodes[0];
2471         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2472
2473         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2474                 goto out;
2475
2476         ret = 1;
2477         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2479         if (item_size < sizeof(*ei)) {
2480                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2481                 goto out;
2482         }
2483 #endif
2484         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2485
2486         if (item_size != sizeof(*ei) +
2487             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2488                 goto out;
2489
2490         if (btrfs_extent_generation(leaf, ei) <=
2491             btrfs_root_last_snapshot(&root->root_item))
2492                 goto out;
2493
2494         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2495         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2496             BTRFS_EXTENT_DATA_REF_KEY)
2497                 goto out;
2498
2499         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2500         if (btrfs_extent_refs(leaf, ei) !=
2501             btrfs_extent_data_ref_count(leaf, ref) ||
2502             btrfs_extent_data_ref_root(leaf, ref) !=
2503             root->root_key.objectid ||
2504             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2505             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2506                 goto out;
2507
2508         ret = 0;
2509 out:
2510         return ret;
2511 }
2512
2513 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2514                           struct btrfs_root *root,
2515                           u64 objectid, u64 offset, u64 bytenr)
2516 {
2517         struct btrfs_path *path;
2518         int ret;
2519         int ret2;
2520
2521         path = btrfs_alloc_path();
2522         if (!path)
2523                 return -ENOENT;
2524
2525         do {
2526                 ret = check_committed_ref(trans, root, path, objectid,
2527                                           offset, bytenr);
2528                 if (ret && ret != -ENOENT)
2529                         goto out;
2530
2531                 ret2 = check_delayed_ref(trans, root, path, objectid,
2532                                          offset, bytenr);
2533         } while (ret2 == -EAGAIN);
2534
2535         if (ret2 && ret2 != -ENOENT) {
2536                 ret = ret2;
2537                 goto out;
2538         }
2539
2540         if (ret != -ENOENT || ret2 != -ENOENT)
2541                 ret = 0;
2542 out:
2543         btrfs_free_path(path);
2544         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2545                 WARN_ON(ret > 0);
2546         return ret;
2547 }
2548
2549 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2550                            struct btrfs_root *root,
2551                            struct extent_buffer *buf,
2552                            int full_backref, int inc)
2553 {
2554         u64 bytenr;
2555         u64 num_bytes;
2556         u64 parent;
2557         u64 ref_root;
2558         u32 nritems;
2559         struct btrfs_key key;
2560         struct btrfs_file_extent_item *fi;
2561         int i;
2562         int level;
2563         int ret = 0;
2564         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2565                             u64, u64, u64, u64, u64, u64);
2566
2567         ref_root = btrfs_header_owner(buf);
2568         nritems = btrfs_header_nritems(buf);
2569         level = btrfs_header_level(buf);
2570
2571         if (!root->ref_cows && level == 0)
2572                 return 0;
2573
2574         if (inc)
2575                 process_func = btrfs_inc_extent_ref;
2576         else
2577                 process_func = btrfs_free_extent;
2578
2579         if (full_backref)
2580                 parent = buf->start;
2581         else
2582                 parent = 0;
2583
2584         for (i = 0; i < nritems; i++) {
2585                 if (level == 0) {
2586                         btrfs_item_key_to_cpu(buf, &key, i);
2587                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2588                                 continue;
2589                         fi = btrfs_item_ptr(buf, i,
2590                                             struct btrfs_file_extent_item);
2591                         if (btrfs_file_extent_type(buf, fi) ==
2592                             BTRFS_FILE_EXTENT_INLINE)
2593                                 continue;
2594                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2595                         if (bytenr == 0)
2596                                 continue;
2597
2598                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2599                         key.offset -= btrfs_file_extent_offset(buf, fi);
2600                         ret = process_func(trans, root, bytenr, num_bytes,
2601                                            parent, ref_root, key.objectid,
2602                                            key.offset);
2603                         if (ret)
2604                                 goto fail;
2605                 } else {
2606                         bytenr = btrfs_node_blockptr(buf, i);
2607                         num_bytes = btrfs_level_size(root, level - 1);
2608                         ret = process_func(trans, root, bytenr, num_bytes,
2609                                            parent, ref_root, level - 1, 0);
2610                         if (ret)
2611                                 goto fail;
2612                 }
2613         }
2614         return 0;
2615 fail:
2616         BUG();
2617         return ret;
2618 }
2619
2620 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2621                   struct extent_buffer *buf, int full_backref)
2622 {
2623         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2624 }
2625
2626 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2627                   struct extent_buffer *buf, int full_backref)
2628 {
2629         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2630 }
2631
2632 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2633                                  struct btrfs_root *root,
2634                                  struct btrfs_path *path,
2635                                  struct btrfs_block_group_cache *cache)
2636 {
2637         int ret;
2638         struct btrfs_root *extent_root = root->fs_info->extent_root;
2639         unsigned long bi;
2640         struct extent_buffer *leaf;
2641
2642         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2643         if (ret < 0)
2644                 goto fail;
2645         BUG_ON(ret);
2646
2647         leaf = path->nodes[0];
2648         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2649         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2650         btrfs_mark_buffer_dirty(leaf);
2651         btrfs_release_path(path);
2652 fail:
2653         if (ret)
2654                 return ret;
2655         return 0;
2656
2657 }
2658
2659 static struct btrfs_block_group_cache *
2660 next_block_group(struct btrfs_root *root,
2661                  struct btrfs_block_group_cache *cache)
2662 {
2663         struct rb_node *node;
2664         spin_lock(&root->fs_info->block_group_cache_lock);
2665         node = rb_next(&cache->cache_node);
2666         btrfs_put_block_group(cache);
2667         if (node) {
2668                 cache = rb_entry(node, struct btrfs_block_group_cache,
2669                                  cache_node);
2670                 btrfs_get_block_group(cache);
2671         } else
2672                 cache = NULL;
2673         spin_unlock(&root->fs_info->block_group_cache_lock);
2674         return cache;
2675 }
2676
2677 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2678                             struct btrfs_trans_handle *trans,
2679                             struct btrfs_path *path)
2680 {
2681         struct btrfs_root *root = block_group->fs_info->tree_root;
2682         struct inode *inode = NULL;
2683         u64 alloc_hint = 0;
2684         int dcs = BTRFS_DC_ERROR;
2685         int num_pages = 0;
2686         int retries = 0;
2687         int ret = 0;
2688
2689         /*
2690          * If this block group is smaller than 100 megs don't bother caching the
2691          * block group.
2692          */
2693         if (block_group->key.offset < (100 * 1024 * 1024)) {
2694                 spin_lock(&block_group->lock);
2695                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2696                 spin_unlock(&block_group->lock);
2697                 return 0;
2698         }
2699
2700 again:
2701         inode = lookup_free_space_inode(root, block_group, path);
2702         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2703                 ret = PTR_ERR(inode);
2704                 btrfs_release_path(path);
2705                 goto out;
2706         }
2707
2708         if (IS_ERR(inode)) {
2709                 BUG_ON(retries);
2710                 retries++;
2711
2712                 if (block_group->ro)
2713                         goto out_free;
2714
2715                 ret = create_free_space_inode(root, trans, block_group, path);
2716                 if (ret)
2717                         goto out_free;
2718                 goto again;
2719         }
2720
2721         /* We've already setup this transaction, go ahead and exit */
2722         if (block_group->cache_generation == trans->transid &&
2723             i_size_read(inode)) {
2724                 dcs = BTRFS_DC_SETUP;
2725                 goto out_put;
2726         }
2727
2728         /*
2729          * We want to set the generation to 0, that way if anything goes wrong
2730          * from here on out we know not to trust this cache when we load up next
2731          * time.
2732          */
2733         BTRFS_I(inode)->generation = 0;
2734         ret = btrfs_update_inode(trans, root, inode);
2735         WARN_ON(ret);
2736
2737         if (i_size_read(inode) > 0) {
2738                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2739                                                       inode);
2740                 if (ret)
2741                         goto out_put;
2742         }
2743
2744         spin_lock(&block_group->lock);
2745         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2746                 /* We're not cached, don't bother trying to write stuff out */
2747                 dcs = BTRFS_DC_WRITTEN;
2748                 spin_unlock(&block_group->lock);
2749                 goto out_put;
2750         }
2751         spin_unlock(&block_group->lock);
2752
2753         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2754         if (!num_pages)
2755                 num_pages = 1;
2756
2757         /*
2758          * Just to make absolutely sure we have enough space, we're going to
2759          * preallocate 12 pages worth of space for each block group.  In
2760          * practice we ought to use at most 8, but we need extra space so we can
2761          * add our header and have a terminator between the extents and the
2762          * bitmaps.
2763          */
2764         num_pages *= 16;
2765         num_pages *= PAGE_CACHE_SIZE;
2766
2767         ret = btrfs_check_data_free_space(inode, num_pages);
2768         if (ret)
2769                 goto out_put;
2770
2771         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2772                                               num_pages, num_pages,
2773                                               &alloc_hint);
2774         if (!ret)
2775                 dcs = BTRFS_DC_SETUP;
2776         btrfs_free_reserved_data_space(inode, num_pages);
2777
2778 out_put:
2779         iput(inode);
2780 out_free:
2781         btrfs_release_path(path);
2782 out:
2783         spin_lock(&block_group->lock);
2784         if (!ret)
2785                 block_group->cache_generation = trans->transid;
2786         block_group->disk_cache_state = dcs;
2787         spin_unlock(&block_group->lock);
2788
2789         return ret;
2790 }
2791
2792 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2793                                    struct btrfs_root *root)
2794 {
2795         struct btrfs_block_group_cache *cache;
2796         int err = 0;
2797         struct btrfs_path *path;
2798         u64 last = 0;
2799
2800         path = btrfs_alloc_path();
2801         if (!path)
2802                 return -ENOMEM;
2803
2804 again:
2805         while (1) {
2806                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2807                 while (cache) {
2808                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2809                                 break;
2810                         cache = next_block_group(root, cache);
2811                 }
2812                 if (!cache) {
2813                         if (last == 0)
2814                                 break;
2815                         last = 0;
2816                         continue;
2817                 }
2818                 err = cache_save_setup(cache, trans, path);
2819                 last = cache->key.objectid + cache->key.offset;
2820                 btrfs_put_block_group(cache);
2821         }
2822
2823         while (1) {
2824                 if (last == 0) {
2825                         err = btrfs_run_delayed_refs(trans, root,
2826                                                      (unsigned long)-1);
2827                         BUG_ON(err);
2828                 }
2829
2830                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2831                 while (cache) {
2832                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2833                                 btrfs_put_block_group(cache);
2834                                 goto again;
2835                         }
2836
2837                         if (cache->dirty)
2838                                 break;
2839                         cache = next_block_group(root, cache);
2840                 }
2841                 if (!cache) {
2842                         if (last == 0)
2843                                 break;
2844                         last = 0;
2845                         continue;
2846                 }
2847
2848                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2849                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2850                 cache->dirty = 0;
2851                 last = cache->key.objectid + cache->key.offset;
2852
2853                 err = write_one_cache_group(trans, root, path, cache);
2854                 BUG_ON(err);
2855                 btrfs_put_block_group(cache);
2856         }
2857
2858         while (1) {
2859                 /*
2860                  * I don't think this is needed since we're just marking our
2861                  * preallocated extent as written, but just in case it can't
2862                  * hurt.
2863                  */
2864                 if (last == 0) {
2865                         err = btrfs_run_delayed_refs(trans, root,
2866                                                      (unsigned long)-1);
2867                         BUG_ON(err);
2868                 }
2869
2870                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2871                 while (cache) {
2872                         /*
2873                          * Really this shouldn't happen, but it could if we
2874                          * couldn't write the entire preallocated extent and
2875                          * splitting the extent resulted in a new block.
2876                          */
2877                         if (cache->dirty) {
2878                                 btrfs_put_block_group(cache);
2879                                 goto again;
2880                         }
2881                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2882                                 break;
2883                         cache = next_block_group(root, cache);
2884                 }
2885                 if (!cache) {
2886                         if (last == 0)
2887                                 break;
2888                         last = 0;
2889                         continue;
2890                 }
2891
2892                 btrfs_write_out_cache(root, trans, cache, path);
2893
2894                 /*
2895                  * If we didn't have an error then the cache state is still
2896                  * NEED_WRITE, so we can set it to WRITTEN.
2897                  */
2898                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2899                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2900                 last = cache->key.objectid + cache->key.offset;
2901                 btrfs_put_block_group(cache);
2902         }
2903
2904         btrfs_free_path(path);
2905         return 0;
2906 }
2907
2908 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2909 {
2910         struct btrfs_block_group_cache *block_group;
2911         int readonly = 0;
2912
2913         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2914         if (!block_group || block_group->ro)
2915                 readonly = 1;
2916         if (block_group)
2917                 btrfs_put_block_group(block_group);
2918         return readonly;
2919 }
2920
2921 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2922                              u64 total_bytes, u64 bytes_used,
2923                              struct btrfs_space_info **space_info)
2924 {
2925         struct btrfs_space_info *found;
2926         int i;
2927         int factor;
2928
2929         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2930                      BTRFS_BLOCK_GROUP_RAID10))
2931                 factor = 2;
2932         else
2933                 factor = 1;
2934
2935         found = __find_space_info(info, flags);
2936         if (found) {
2937                 spin_lock(&found->lock);
2938                 found->total_bytes += total_bytes;
2939                 found->disk_total += total_bytes * factor;
2940                 found->bytes_used += bytes_used;
2941                 found->disk_used += bytes_used * factor;
2942                 found->full = 0;
2943                 spin_unlock(&found->lock);
2944                 *space_info = found;
2945                 return 0;
2946         }
2947         found = kzalloc(sizeof(*found), GFP_NOFS);
2948         if (!found)
2949                 return -ENOMEM;
2950
2951         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2952                 INIT_LIST_HEAD(&found->block_groups[i]);
2953         init_rwsem(&found->groups_sem);
2954         spin_lock_init(&found->lock);
2955         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2956                                 BTRFS_BLOCK_GROUP_SYSTEM |
2957                                 BTRFS_BLOCK_GROUP_METADATA);
2958         found->total_bytes = total_bytes;
2959         found->disk_total = total_bytes * factor;
2960         found->bytes_used = bytes_used;
2961         found->disk_used = bytes_used * factor;
2962         found->bytes_pinned = 0;
2963         found->bytes_reserved = 0;
2964         found->bytes_readonly = 0;
2965         found->bytes_may_use = 0;
2966         found->full = 0;
2967         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2968         found->chunk_alloc = 0;
2969         found->flush = 0;
2970         init_waitqueue_head(&found->wait);
2971         *space_info = found;
2972         list_add_rcu(&found->list, &info->space_info);
2973         return 0;
2974 }
2975
2976 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2977 {
2978         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2979                                    BTRFS_BLOCK_GROUP_RAID1 |
2980                                    BTRFS_BLOCK_GROUP_RAID10 |
2981                                    BTRFS_BLOCK_GROUP_DUP);
2982         if (extra_flags) {
2983                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2984                         fs_info->avail_data_alloc_bits |= extra_flags;
2985                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2986                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2987                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2988                         fs_info->avail_system_alloc_bits |= extra_flags;
2989         }
2990 }
2991
2992 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2993 {
2994         /*
2995          * we add in the count of missing devices because we want
2996          * to make sure that any RAID levels on a degraded FS
2997          * continue to be honored.
2998          */
2999         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3000                 root->fs_info->fs_devices->missing_devices;
3001
3002         if (num_devices == 1)
3003                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3004         if (num_devices < 4)
3005                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3006
3007         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3008             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3009                       BTRFS_BLOCK_GROUP_RAID10))) {
3010                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3011         }
3012
3013         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3014             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3015                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3016         }
3017
3018         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3019             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3020              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3021              (flags & BTRFS_BLOCK_GROUP_DUP)))
3022                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3023         return flags;
3024 }
3025
3026 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3027 {
3028         if (flags & BTRFS_BLOCK_GROUP_DATA)
3029                 flags |= root->fs_info->avail_data_alloc_bits &
3030                          root->fs_info->data_alloc_profile;
3031         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3032                 flags |= root->fs_info->avail_system_alloc_bits &
3033                          root->fs_info->system_alloc_profile;
3034         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3035                 flags |= root->fs_info->avail_metadata_alloc_bits &
3036                          root->fs_info->metadata_alloc_profile;
3037         return btrfs_reduce_alloc_profile(root, flags);
3038 }
3039
3040 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3041 {
3042         u64 flags;
3043
3044         if (data)
3045                 flags = BTRFS_BLOCK_GROUP_DATA;
3046         else if (root == root->fs_info->chunk_root)
3047                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3048         else
3049                 flags = BTRFS_BLOCK_GROUP_METADATA;
3050
3051         return get_alloc_profile(root, flags);
3052 }
3053
3054 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3055 {
3056         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3057                                                        BTRFS_BLOCK_GROUP_DATA);
3058 }
3059
3060 /*
3061  * This will check the space that the inode allocates from to make sure we have
3062  * enough space for bytes.
3063  */
3064 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3065 {
3066         struct btrfs_space_info *data_sinfo;
3067         struct btrfs_root *root = BTRFS_I(inode)->root;
3068         u64 used;
3069         int ret = 0, committed = 0, alloc_chunk = 1;
3070
3071         /* make sure bytes are sectorsize aligned */
3072         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3073
3074         if (root == root->fs_info->tree_root ||
3075             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3076                 alloc_chunk = 0;
3077                 committed = 1;
3078         }
3079
3080         data_sinfo = BTRFS_I(inode)->space_info;
3081         if (!data_sinfo)
3082                 goto alloc;
3083
3084 again:
3085         /* make sure we have enough space to handle the data first */
3086         spin_lock(&data_sinfo->lock);
3087         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3088                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3089                 data_sinfo->bytes_may_use;
3090
3091         if (used + bytes > data_sinfo->total_bytes) {
3092                 struct btrfs_trans_handle *trans;
3093
3094                 /*
3095                  * if we don't have enough free bytes in this space then we need
3096                  * to alloc a new chunk.
3097                  */
3098                 if (!data_sinfo->full && alloc_chunk) {
3099                         u64 alloc_target;
3100
3101                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3102                         spin_unlock(&data_sinfo->lock);
3103 alloc:
3104                         alloc_target = btrfs_get_alloc_profile(root, 1);
3105                         trans = btrfs_join_transaction(root);
3106                         if (IS_ERR(trans))
3107                                 return PTR_ERR(trans);
3108
3109                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3110                                              bytes + 2 * 1024 * 1024,
3111                                              alloc_target,
3112                                              CHUNK_ALLOC_NO_FORCE);
3113                         btrfs_end_transaction(trans, root);
3114                         if (ret < 0) {
3115                                 if (ret != -ENOSPC)
3116                                         return ret;
3117                                 else
3118                                         goto commit_trans;
3119                         }
3120
3121                         if (!data_sinfo) {
3122                                 btrfs_set_inode_space_info(root, inode);
3123                                 data_sinfo = BTRFS_I(inode)->space_info;
3124                         }
3125                         goto again;
3126                 }
3127
3128                 /*
3129                  * If we have less pinned bytes than we want to allocate then
3130                  * don't bother committing the transaction, it won't help us.
3131                  */
3132                 if (data_sinfo->bytes_pinned < bytes)
3133                         committed = 1;
3134                 spin_unlock(&data_sinfo->lock);
3135
3136                 /* commit the current transaction and try again */
3137 commit_trans:
3138                 if (!committed &&
3139                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3140                         committed = 1;
3141                         trans = btrfs_join_transaction(root);
3142                         if (IS_ERR(trans))
3143                                 return PTR_ERR(trans);
3144                         ret = btrfs_commit_transaction(trans, root);
3145                         if (ret)
3146                                 return ret;
3147                         goto again;
3148                 }
3149
3150                 return -ENOSPC;
3151         }
3152         data_sinfo->bytes_may_use += bytes;
3153         spin_unlock(&data_sinfo->lock);
3154
3155         return 0;
3156 }
3157
3158 /*
3159  * Called if we need to clear a data reservation for this inode.
3160  */
3161 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3162 {
3163         struct btrfs_root *root = BTRFS_I(inode)->root;
3164         struct btrfs_space_info *data_sinfo;
3165
3166         /* make sure bytes are sectorsize aligned */
3167         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3168
3169         data_sinfo = BTRFS_I(inode)->space_info;
3170         spin_lock(&data_sinfo->lock);
3171         data_sinfo->bytes_may_use -= bytes;
3172         spin_unlock(&data_sinfo->lock);
3173 }
3174
3175 static void force_metadata_allocation(struct btrfs_fs_info *info)
3176 {
3177         struct list_head *head = &info->space_info;
3178         struct btrfs_space_info *found;
3179
3180         rcu_read_lock();
3181         list_for_each_entry_rcu(found, head, list) {
3182                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3183                         found->force_alloc = CHUNK_ALLOC_FORCE;
3184         }
3185         rcu_read_unlock();
3186 }
3187
3188 static int should_alloc_chunk(struct btrfs_root *root,
3189                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3190                               int force)
3191 {
3192         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3193         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3194         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3195         u64 thresh;
3196
3197         if (force == CHUNK_ALLOC_FORCE)
3198                 return 1;
3199
3200         /*
3201          * We need to take into account the global rsv because for all intents
3202          * and purposes it's used space.  Don't worry about locking the
3203          * global_rsv, it doesn't change except when the transaction commits.
3204          */
3205         num_allocated += global_rsv->size;
3206
3207         /*
3208          * in limited mode, we want to have some free space up to
3209          * about 1% of the FS size.
3210          */
3211         if (force == CHUNK_ALLOC_LIMITED) {
3212                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3213                 thresh = max_t(u64, 64 * 1024 * 1024,
3214                                div_factor_fine(thresh, 1));
3215
3216                 if (num_bytes - num_allocated < thresh)
3217                         return 1;
3218         }
3219
3220         /*
3221          * we have two similar checks here, one based on percentage
3222          * and once based on a hard number of 256MB.  The idea
3223          * is that if we have a good amount of free
3224          * room, don't allocate a chunk.  A good mount is
3225          * less than 80% utilized of the chunks we have allocated,
3226          * or more than 256MB free
3227          */
3228         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3229                 return 0;
3230
3231         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3232                 return 0;
3233
3234         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3235
3236         /* 256MB or 5% of the FS */
3237         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3238
3239         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3240                 return 0;
3241         return 1;
3242 }
3243
3244 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3245                           struct btrfs_root *extent_root, u64 alloc_bytes,
3246                           u64 flags, int force)
3247 {
3248         struct btrfs_space_info *space_info;
3249         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3250         int wait_for_alloc = 0;
3251         int ret = 0;
3252
3253         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3254
3255         space_info = __find_space_info(extent_root->fs_info, flags);
3256         if (!space_info) {
3257                 ret = update_space_info(extent_root->fs_info, flags,
3258                                         0, 0, &space_info);
3259                 BUG_ON(ret);
3260         }
3261         BUG_ON(!space_info);
3262
3263 again:
3264         spin_lock(&space_info->lock);
3265         if (space_info->force_alloc)
3266                 force = space_info->force_alloc;
3267         if (space_info->full) {
3268                 spin_unlock(&space_info->lock);
3269                 return 0;
3270         }
3271
3272         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3273                 spin_unlock(&space_info->lock);
3274                 return 0;
3275         } else if (space_info->chunk_alloc) {
3276                 wait_for_alloc = 1;
3277         } else {
3278                 space_info->chunk_alloc = 1;
3279         }
3280
3281         spin_unlock(&space_info->lock);
3282
3283         mutex_lock(&fs_info->chunk_mutex);
3284
3285         /*
3286          * The chunk_mutex is held throughout the entirety of a chunk
3287          * allocation, so once we've acquired the chunk_mutex we know that the
3288          * other guy is done and we need to recheck and see if we should
3289          * allocate.
3290          */
3291         if (wait_for_alloc) {
3292                 mutex_unlock(&fs_info->chunk_mutex);
3293                 wait_for_alloc = 0;
3294                 goto again;
3295         }
3296
3297         /*
3298          * If we have mixed data/metadata chunks we want to make sure we keep
3299          * allocating mixed chunks instead of individual chunks.
3300          */
3301         if (btrfs_mixed_space_info(space_info))
3302                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3303
3304         /*
3305          * if we're doing a data chunk, go ahead and make sure that
3306          * we keep a reasonable number of metadata chunks allocated in the
3307          * FS as well.
3308          */
3309         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3310                 fs_info->data_chunk_allocations++;
3311                 if (!(fs_info->data_chunk_allocations %
3312                       fs_info->metadata_ratio))
3313                         force_metadata_allocation(fs_info);
3314         }
3315
3316         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3317         if (ret < 0 && ret != -ENOSPC)
3318                 goto out;
3319
3320         spin_lock(&space_info->lock);
3321         if (ret)
3322                 space_info->full = 1;
3323         else
3324                 ret = 1;
3325
3326         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3327         space_info->chunk_alloc = 0;
3328         spin_unlock(&space_info->lock);
3329 out:
3330         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3331         return ret;
3332 }
3333
3334 /*
3335  * shrink metadata reservation for delalloc
3336  */
3337 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3338                            bool wait_ordered)
3339 {
3340         struct btrfs_block_rsv *block_rsv;
3341         struct btrfs_space_info *space_info;
3342         struct btrfs_trans_handle *trans;
3343         u64 reserved;
3344         u64 max_reclaim;
3345         u64 reclaimed = 0;
3346         long time_left;
3347         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3348         int loops = 0;
3349         unsigned long progress;
3350
3351         trans = (struct btrfs_trans_handle *)current->journal_info;
3352         block_rsv = &root->fs_info->delalloc_block_rsv;
3353         space_info = block_rsv->space_info;
3354
3355         smp_mb();
3356         reserved = space_info->bytes_may_use;
3357         progress = space_info->reservation_progress;
3358
3359         if (reserved == 0)
3360                 return 0;
3361
3362         smp_mb();
3363         if (root->fs_info->delalloc_bytes == 0) {
3364                 if (trans)
3365                         return 0;
3366                 btrfs_wait_ordered_extents(root, 0, 0);
3367                 return 0;
3368         }
3369
3370         max_reclaim = min(reserved, to_reclaim);
3371         nr_pages = max_t(unsigned long, nr_pages,
3372                          max_reclaim >> PAGE_CACHE_SHIFT);
3373         while (loops < 1024) {
3374                 /* have the flusher threads jump in and do some IO */
3375                 smp_mb();
3376                 nr_pages = min_t(unsigned long, nr_pages,
3377                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3378                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3379
3380                 spin_lock(&space_info->lock);
3381                 if (reserved > space_info->bytes_may_use)
3382                         reclaimed += reserved - space_info->bytes_may_use;
3383                 reserved = space_info->bytes_may_use;
3384                 spin_unlock(&space_info->lock);
3385
3386                 loops++;
3387
3388                 if (reserved == 0 || reclaimed >= max_reclaim)
3389                         break;
3390
3391                 if (trans && trans->transaction->blocked)
3392                         return -EAGAIN;
3393
3394                 if (wait_ordered && !trans) {
3395                         btrfs_wait_ordered_extents(root, 0, 0);
3396                 } else {
3397                         time_left = schedule_timeout_interruptible(1);
3398
3399                         /* We were interrupted, exit */
3400                         if (time_left)
3401                                 break;
3402                 }
3403
3404                 /* we've kicked the IO a few times, if anything has been freed,
3405                  * exit.  There is no sense in looping here for a long time
3406                  * when we really need to commit the transaction, or there are
3407                  * just too many writers without enough free space
3408                  */
3409
3410                 if (loops > 3) {
3411                         smp_mb();
3412                         if (progress != space_info->reservation_progress)
3413                                 break;
3414                 }
3415
3416         }
3417
3418         return reclaimed >= to_reclaim;
3419 }
3420
3421 /**
3422  * maybe_commit_transaction - possibly commit the transaction if its ok to
3423  * @root - the root we're allocating for
3424  * @bytes - the number of bytes we want to reserve
3425  * @force - force the commit
3426  *
3427  * This will check to make sure that committing the transaction will actually
3428  * get us somewhere and then commit the transaction if it does.  Otherwise it
3429  * will return -ENOSPC.
3430  */
3431 static int may_commit_transaction(struct btrfs_root *root,
3432                                   struct btrfs_space_info *space_info,
3433                                   u64 bytes, int force)
3434 {
3435         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3436         struct btrfs_trans_handle *trans;
3437
3438         trans = (struct btrfs_trans_handle *)current->journal_info;
3439         if (trans)
3440                 return -EAGAIN;
3441
3442         if (force)
3443                 goto commit;
3444
3445         /* See if there is enough pinned space to make this reservation */
3446         spin_lock(&space_info->lock);
3447         if (space_info->bytes_pinned >= bytes) {
3448                 spin_unlock(&space_info->lock);
3449                 goto commit;
3450         }
3451         spin_unlock(&space_info->lock);
3452
3453         /*
3454          * See if there is some space in the delayed insertion reservation for
3455          * this reservation.
3456          */
3457         if (space_info != delayed_rsv->space_info)
3458                 return -ENOSPC;
3459
3460         spin_lock(&delayed_rsv->lock);
3461         if (delayed_rsv->size < bytes) {
3462                 spin_unlock(&delayed_rsv->lock);
3463                 return -ENOSPC;
3464         }
3465         spin_unlock(&delayed_rsv->lock);
3466
3467 commit:
3468         trans = btrfs_join_transaction(root);
3469         if (IS_ERR(trans))
3470                 return -ENOSPC;
3471
3472         return btrfs_commit_transaction(trans, root);
3473 }
3474
3475 /**
3476  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3477  * @root - the root we're allocating for
3478  * @block_rsv - the block_rsv we're allocating for
3479  * @orig_bytes - the number of bytes we want
3480  * @flush - wether or not we can flush to make our reservation
3481  *
3482  * This will reserve orgi_bytes number of bytes from the space info associated
3483  * with the block_rsv.  If there is not enough space it will make an attempt to
3484  * flush out space to make room.  It will do this by flushing delalloc if
3485  * possible or committing the transaction.  If flush is 0 then no attempts to
3486  * regain reservations will be made and this will fail if there is not enough
3487  * space already.
3488  */
3489 static int reserve_metadata_bytes(struct btrfs_root *root,
3490                                   struct btrfs_block_rsv *block_rsv,
3491                                   u64 orig_bytes, int flush)
3492 {
3493         struct btrfs_space_info *space_info = block_rsv->space_info;
3494         u64 used;
3495         u64 num_bytes = orig_bytes;
3496         int retries = 0;
3497         int ret = 0;
3498         bool committed = false;
3499         bool flushing = false;
3500         bool wait_ordered = false;
3501
3502 again:
3503         ret = 0;
3504         spin_lock(&space_info->lock);
3505         /*
3506          * We only want to wait if somebody other than us is flushing and we are
3507          * actually alloed to flush.
3508          */
3509         while (flush && !flushing && space_info->flush) {
3510                 spin_unlock(&space_info->lock);
3511                 /*
3512                  * If we have a trans handle we can't wait because the flusher
3513                  * may have to commit the transaction, which would mean we would
3514                  * deadlock since we are waiting for the flusher to finish, but
3515                  * hold the current transaction open.
3516                  */
3517                 if (current->journal_info)
3518                         return -EAGAIN;
3519                 ret = wait_event_interruptible(space_info->wait,
3520                                                !space_info->flush);
3521                 /* Must have been interrupted, return */
3522                 if (ret)
3523                         return -EINTR;
3524
3525                 spin_lock(&space_info->lock);
3526         }
3527
3528         ret = -ENOSPC;
3529         used = space_info->bytes_used + space_info->bytes_reserved +
3530                 space_info->bytes_pinned + space_info->bytes_readonly +
3531                 space_info->bytes_may_use;
3532
3533         /*
3534          * The idea here is that we've not already over-reserved the block group
3535          * then we can go ahead and save our reservation first and then start
3536          * flushing if we need to.  Otherwise if we've already overcommitted
3537          * lets start flushing stuff first and then come back and try to make
3538          * our reservation.
3539          */
3540         if (used <= space_info->total_bytes) {
3541                 if (used + orig_bytes <= space_info->total_bytes) {
3542                         space_info->bytes_may_use += orig_bytes;
3543                         ret = 0;
3544                 } else {
3545                         /*
3546                          * Ok set num_bytes to orig_bytes since we aren't
3547                          * overocmmitted, this way we only try and reclaim what
3548                          * we need.
3549                          */
3550                         num_bytes = orig_bytes;
3551                 }
3552         } else {
3553                 /*
3554                  * Ok we're over committed, set num_bytes to the overcommitted
3555                  * amount plus the amount of bytes that we need for this
3556                  * reservation.
3557                  */
3558                 wait_ordered = true;
3559                 num_bytes = used - space_info->total_bytes +
3560                         (orig_bytes * (retries + 1));
3561         }
3562
3563         if (ret) {
3564                 u64 profile = btrfs_get_alloc_profile(root, 0);
3565                 u64 avail;
3566
3567                 /*
3568                  * If we have a lot of space that's pinned, don't bother doing
3569                  * the overcommit dance yet and just commit the transaction.
3570                  */
3571                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3572                 do_div(avail, 10);
3573                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3574                         space_info->flush = 1;
3575                         flushing = true;
3576                         spin_unlock(&space_info->lock);
3577                         ret = may_commit_transaction(root, space_info,
3578                                                      orig_bytes, 1);
3579                         if (ret)
3580                                 goto out;
3581                         committed = true;
3582                         goto again;
3583                 }
3584
3585                 spin_lock(&root->fs_info->free_chunk_lock);
3586                 avail = root->fs_info->free_chunk_space;
3587
3588                 /*
3589                  * If we have dup, raid1 or raid10 then only half of the free
3590                  * space is actually useable.
3591                  */
3592                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3593                                BTRFS_BLOCK_GROUP_RAID1 |
3594                                BTRFS_BLOCK_GROUP_RAID10))
3595                         avail >>= 1;
3596
3597                 /*
3598                  * If we aren't flushing don't let us overcommit too much, say
3599                  * 1/8th of the space.  If we can flush, let it overcommit up to
3600                  * 1/2 of the space.
3601                  */
3602                 if (flush)
3603                         avail >>= 3;
3604                 else
3605                         avail >>= 1;
3606                  spin_unlock(&root->fs_info->free_chunk_lock);
3607
3608                 if (used + num_bytes < space_info->total_bytes + avail) {
3609                         space_info->bytes_may_use += orig_bytes;
3610                         ret = 0;
3611                 } else {
3612                         wait_ordered = true;
3613                 }
3614         }
3615
3616         /*
3617          * Couldn't make our reservation, save our place so while we're trying
3618          * to reclaim space we can actually use it instead of somebody else
3619          * stealing it from us.
3620          */
3621         if (ret && flush) {
3622                 flushing = true;
3623                 space_info->flush = 1;
3624         }
3625
3626         spin_unlock(&space_info->lock);
3627
3628         if (!ret || !flush)
3629                 goto out;
3630
3631         /*
3632          * We do synchronous shrinking since we don't actually unreserve
3633          * metadata until after the IO is completed.
3634          */
3635         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3636         if (ret < 0)
3637                 goto out;
3638
3639         ret = 0;
3640
3641         /*
3642          * So if we were overcommitted it's possible that somebody else flushed
3643          * out enough space and we simply didn't have enough space to reclaim,
3644          * so go back around and try again.
3645          */
3646         if (retries < 2) {
3647                 wait_ordered = true;
3648                 retries++;
3649                 goto again;
3650         }
3651
3652         ret = -ENOSPC;
3653         if (committed)
3654                 goto out;
3655
3656         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3657         if (!ret) {
3658                 committed = true;
3659                 goto again;
3660         }
3661
3662 out:
3663         if (flushing) {
3664                 spin_lock(&space_info->lock);
3665                 space_info->flush = 0;
3666                 wake_up_all(&space_info->wait);
3667                 spin_unlock(&space_info->lock);
3668         }
3669         return ret;
3670 }
3671
3672 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3673                                              struct btrfs_root *root)
3674 {
3675         struct btrfs_block_rsv *block_rsv = NULL;
3676
3677         if (root->ref_cows || root == root->fs_info->csum_root)
3678                 block_rsv = trans->block_rsv;
3679
3680         if (!block_rsv)
3681                 block_rsv = root->block_rsv;
3682
3683         if (!block_rsv)
3684                 block_rsv = &root->fs_info->empty_block_rsv;
3685
3686         return block_rsv;
3687 }
3688
3689 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3690                                u64 num_bytes)
3691 {
3692         int ret = -ENOSPC;
3693         spin_lock(&block_rsv->lock);
3694         if (block_rsv->reserved >= num_bytes) {
3695                 block_rsv->reserved -= num_bytes;
3696                 if (block_rsv->reserved < block_rsv->size)
3697                         block_rsv->full = 0;
3698                 ret = 0;
3699         }
3700         spin_unlock(&block_rsv->lock);
3701         return ret;
3702 }
3703
3704 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3705                                 u64 num_bytes, int update_size)
3706 {
3707         spin_lock(&block_rsv->lock);
3708         block_rsv->reserved += num_bytes;
3709         if (update_size)
3710                 block_rsv->size += num_bytes;
3711         else if (block_rsv->reserved >= block_rsv->size)
3712                 block_rsv->full = 1;
3713         spin_unlock(&block_rsv->lock);
3714 }
3715
3716 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3717                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3718 {
3719         struct btrfs_space_info *space_info = block_rsv->space_info;
3720
3721         spin_lock(&block_rsv->lock);
3722         if (num_bytes == (u64)-1)
3723                 num_bytes = block_rsv->size;
3724         block_rsv->size -= num_bytes;
3725         if (block_rsv->reserved >= block_rsv->size) {
3726                 num_bytes = block_rsv->reserved - block_rsv->size;
3727                 block_rsv->reserved = block_rsv->size;
3728                 block_rsv->full = 1;
3729         } else {
3730                 num_bytes = 0;
3731         }
3732         spin_unlock(&block_rsv->lock);
3733
3734         if (num_bytes > 0) {
3735                 if (dest) {
3736                         spin_lock(&dest->lock);
3737                         if (!dest->full) {
3738                                 u64 bytes_to_add;
3739
3740                                 bytes_to_add = dest->size - dest->reserved;
3741                                 bytes_to_add = min(num_bytes, bytes_to_add);
3742                                 dest->reserved += bytes_to_add;
3743                                 if (dest->reserved >= dest->size)
3744                                         dest->full = 1;
3745                                 num_bytes -= bytes_to_add;
3746                         }
3747                         spin_unlock(&dest->lock);
3748                 }
3749                 if (num_bytes) {
3750                         spin_lock(&space_info->lock);
3751                         space_info->bytes_may_use -= num_bytes;
3752                         space_info->reservation_progress++;
3753                         spin_unlock(&space_info->lock);
3754                 }
3755         }
3756 }
3757
3758 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3759                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3760 {
3761         int ret;
3762
3763         ret = block_rsv_use_bytes(src, num_bytes);
3764         if (ret)
3765                 return ret;
3766
3767         block_rsv_add_bytes(dst, num_bytes, 1);
3768         return 0;
3769 }
3770
3771 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3772 {
3773         memset(rsv, 0, sizeof(*rsv));
3774         spin_lock_init(&rsv->lock);
3775 }
3776
3777 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3778 {
3779         struct btrfs_block_rsv *block_rsv;
3780         struct btrfs_fs_info *fs_info = root->fs_info;
3781
3782         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3783         if (!block_rsv)
3784                 return NULL;
3785
3786         btrfs_init_block_rsv(block_rsv);
3787         block_rsv->space_info = __find_space_info(fs_info,
3788                                                   BTRFS_BLOCK_GROUP_METADATA);
3789         return block_rsv;
3790 }
3791
3792 void btrfs_free_block_rsv(struct btrfs_root *root,
3793                           struct btrfs_block_rsv *rsv)
3794 {
3795         btrfs_block_rsv_release(root, rsv, (u64)-1);
3796         kfree(rsv);
3797 }
3798
3799 static inline int __block_rsv_add(struct btrfs_root *root,
3800                                   struct btrfs_block_rsv *block_rsv,
3801                                   u64 num_bytes, int flush)
3802 {
3803         int ret;
3804
3805         if (num_bytes == 0)
3806                 return 0;
3807
3808         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3809         if (!ret) {
3810                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3811                 return 0;
3812         }
3813
3814         return ret;
3815 }
3816
3817 int btrfs_block_rsv_add(struct btrfs_root *root,
3818                         struct btrfs_block_rsv *block_rsv,
3819                         u64 num_bytes)
3820 {
3821         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3822 }
3823
3824 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3825                                 struct btrfs_block_rsv *block_rsv,
3826                                 u64 num_bytes)
3827 {
3828         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3829 }
3830
3831 int btrfs_block_rsv_check(struct btrfs_root *root,
3832                           struct btrfs_block_rsv *block_rsv, int min_factor)
3833 {
3834         u64 num_bytes = 0;
3835         int ret = -ENOSPC;
3836
3837         if (!block_rsv)
3838                 return 0;
3839
3840         spin_lock(&block_rsv->lock);
3841         num_bytes = div_factor(block_rsv->size, min_factor);
3842         if (block_rsv->reserved >= num_bytes)
3843                 ret = 0;
3844         spin_unlock(&block_rsv->lock);
3845
3846         return ret;
3847 }
3848
3849 int btrfs_block_rsv_refill(struct btrfs_root *root,
3850                           struct btrfs_block_rsv *block_rsv,
3851                           u64 min_reserved)
3852 {
3853         u64 num_bytes = 0;
3854         int ret = -ENOSPC;
3855
3856         if (!block_rsv)
3857                 return 0;
3858
3859         spin_lock(&block_rsv->lock);
3860         num_bytes = min_reserved;
3861         if (block_rsv->reserved >= num_bytes)
3862                 ret = 0;
3863         else
3864                 num_bytes -= block_rsv->reserved;
3865         spin_unlock(&block_rsv->lock);
3866
3867         if (!ret)
3868                 return 0;
3869
3870         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3871         if (!ret) {
3872                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3873                 return 0;
3874         }
3875
3876         return ret;
3877 }
3878
3879 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3880                             struct btrfs_block_rsv *dst_rsv,
3881                             u64 num_bytes)
3882 {
3883         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3884 }
3885
3886 void btrfs_block_rsv_release(struct btrfs_root *root,
3887                              struct btrfs_block_rsv *block_rsv,
3888                              u64 num_bytes)
3889 {
3890         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3891         if (global_rsv->full || global_rsv == block_rsv ||
3892             block_rsv->space_info != global_rsv->space_info)
3893                 global_rsv = NULL;
3894         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3895 }
3896
3897 /*
3898  * helper to calculate size of global block reservation.
3899  * the desired value is sum of space used by extent tree,
3900  * checksum tree and root tree
3901  */
3902 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3903 {
3904         struct btrfs_space_info *sinfo;
3905         u64 num_bytes;
3906         u64 meta_used;
3907         u64 data_used;
3908         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3909
3910         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3911         spin_lock(&sinfo->lock);
3912         data_used = sinfo->bytes_used;
3913         spin_unlock(&sinfo->lock);
3914
3915         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3916         spin_lock(&sinfo->lock);
3917         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3918                 data_used = 0;
3919         meta_used = sinfo->bytes_used;
3920         spin_unlock(&sinfo->lock);
3921
3922         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3923                     csum_size * 2;
3924         num_bytes += div64_u64(data_used + meta_used, 50);
3925
3926         if (num_bytes * 3 > meta_used)
3927                 num_bytes = div64_u64(meta_used, 3);
3928
3929         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3930 }
3931
3932 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3933 {
3934         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3935         struct btrfs_space_info *sinfo = block_rsv->space_info;
3936         u64 num_bytes;
3937
3938         num_bytes = calc_global_metadata_size(fs_info);
3939
3940         spin_lock(&block_rsv->lock);
3941         spin_lock(&sinfo->lock);
3942
3943         block_rsv->size = num_bytes;
3944
3945         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3946                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3947                     sinfo->bytes_may_use;
3948
3949         if (sinfo->total_bytes > num_bytes) {
3950                 num_bytes = sinfo->total_bytes - num_bytes;
3951                 block_rsv->reserved += num_bytes;
3952                 sinfo->bytes_may_use += num_bytes;
3953         }
3954
3955         if (block_rsv->reserved >= block_rsv->size) {
3956                 num_bytes = block_rsv->reserved - block_rsv->size;
3957                 sinfo->bytes_may_use -= num_bytes;
3958                 sinfo->reservation_progress++;
3959                 block_rsv->reserved = block_rsv->size;
3960                 block_rsv->full = 1;
3961         }
3962
3963         spin_unlock(&sinfo->lock);
3964         spin_unlock(&block_rsv->lock);
3965 }
3966
3967 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3968 {
3969         struct btrfs_space_info *space_info;
3970
3971         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3972         fs_info->chunk_block_rsv.space_info = space_info;
3973
3974         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3975         fs_info->global_block_rsv.space_info = space_info;
3976         fs_info->delalloc_block_rsv.space_info = space_info;
3977         fs_info->trans_block_rsv.space_info = space_info;
3978         fs_info->empty_block_rsv.space_info = space_info;
3979         fs_info->delayed_block_rsv.space_info = space_info;
3980
3981         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3982         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3983         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3984         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3985         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3986
3987         update_global_block_rsv(fs_info);
3988 }
3989
3990 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3991 {
3992         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3993         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3994         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3995         WARN_ON(fs_info->trans_block_rsv.size > 0);
3996         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3997         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3998         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3999         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4000         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4001 }
4002
4003 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4004                                   struct btrfs_root *root)
4005 {
4006         if (!trans->bytes_reserved)
4007                 return;
4008
4009         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4010         trans->bytes_reserved = 0;
4011 }
4012
4013 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4014                                   struct inode *inode)
4015 {
4016         struct btrfs_root *root = BTRFS_I(inode)->root;
4017         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4018         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4019
4020         /*
4021          * We need to hold space in order to delete our orphan item once we've
4022          * added it, so this takes the reservation so we can release it later
4023          * when we are truly done with the orphan item.
4024          */
4025         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4026         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4027 }
4028
4029 void btrfs_orphan_release_metadata(struct inode *inode)
4030 {
4031         struct btrfs_root *root = BTRFS_I(inode)->root;
4032         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4033         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4034 }
4035
4036 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4037                                 struct btrfs_pending_snapshot *pending)
4038 {
4039         struct btrfs_root *root = pending->root;
4040         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4041         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4042         /*
4043          * two for root back/forward refs, two for directory entries
4044          * and one for root of the snapshot.
4045          */
4046         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4047         dst_rsv->space_info = src_rsv->space_info;
4048         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4049 }
4050
4051 /**
4052  * drop_outstanding_extent - drop an outstanding extent
4053  * @inode: the inode we're dropping the extent for
4054  *
4055  * This is called when we are freeing up an outstanding extent, either called
4056  * after an error or after an extent is written.  This will return the number of
4057  * reserved extents that need to be freed.  This must be called with
4058  * BTRFS_I(inode)->lock held.
4059  */
4060 static unsigned drop_outstanding_extent(struct inode *inode)
4061 {
4062         unsigned drop_inode_space = 0;
4063         unsigned dropped_extents = 0;
4064
4065         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4066         BTRFS_I(inode)->outstanding_extents--;
4067
4068         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4069             BTRFS_I(inode)->delalloc_meta_reserved) {
4070                 drop_inode_space = 1;
4071                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4072         }
4073
4074         /*
4075          * If we have more or the same amount of outsanding extents than we have
4076          * reserved then we need to leave the reserved extents count alone.
4077          */
4078         if (BTRFS_I(inode)->outstanding_extents >=
4079             BTRFS_I(inode)->reserved_extents)
4080                 return drop_inode_space;
4081
4082         dropped_extents = BTRFS_I(inode)->reserved_extents -
4083                 BTRFS_I(inode)->outstanding_extents;
4084         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4085         return dropped_extents + drop_inode_space;
4086 }
4087
4088 /**
4089  * calc_csum_metadata_size - return the amount of metada space that must be
4090  *      reserved/free'd for the given bytes.
4091  * @inode: the inode we're manipulating
4092  * @num_bytes: the number of bytes in question
4093  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4094  *
4095  * This adjusts the number of csum_bytes in the inode and then returns the
4096  * correct amount of metadata that must either be reserved or freed.  We
4097  * calculate how many checksums we can fit into one leaf and then divide the
4098  * number of bytes that will need to be checksumed by this value to figure out
4099  * how many checksums will be required.  If we are adding bytes then the number
4100  * may go up and we will return the number of additional bytes that must be
4101  * reserved.  If it is going down we will return the number of bytes that must
4102  * be freed.
4103  *
4104  * This must be called with BTRFS_I(inode)->lock held.
4105  */
4106 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4107                                    int reserve)
4108 {
4109         struct btrfs_root *root = BTRFS_I(inode)->root;
4110         u64 csum_size;
4111         int num_csums_per_leaf;
4112         int num_csums;
4113         int old_csums;
4114
4115         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4116             BTRFS_I(inode)->csum_bytes == 0)
4117                 return 0;
4118
4119         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4120         if (reserve)
4121                 BTRFS_I(inode)->csum_bytes += num_bytes;
4122         else
4123                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4124         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4125         num_csums_per_leaf = (int)div64_u64(csum_size,
4126                                             sizeof(struct btrfs_csum_item) +
4127                                             sizeof(struct btrfs_disk_key));
4128         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4129         num_csums = num_csums + num_csums_per_leaf - 1;
4130         num_csums = num_csums / num_csums_per_leaf;
4131
4132         old_csums = old_csums + num_csums_per_leaf - 1;
4133         old_csums = old_csums / num_csums_per_leaf;
4134
4135         /* No change, no need to reserve more */
4136         if (old_csums == num_csums)
4137                 return 0;
4138
4139         if (reserve)
4140                 return btrfs_calc_trans_metadata_size(root,
4141                                                       num_csums - old_csums);
4142
4143         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4144 }
4145
4146 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4147 {
4148         struct btrfs_root *root = BTRFS_I(inode)->root;
4149         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4150         u64 to_reserve = 0;
4151         unsigned nr_extents = 0;
4152         int flush = 1;
4153         int ret;
4154
4155         if (btrfs_is_free_space_inode(root, inode))
4156                 flush = 0;
4157
4158         if (flush && btrfs_transaction_in_commit(root->fs_info))
4159                 schedule_timeout(1);
4160
4161         num_bytes = ALIGN(num_bytes, root->sectorsize);
4162
4163         spin_lock(&BTRFS_I(inode)->lock);
4164         BTRFS_I(inode)->outstanding_extents++;
4165
4166         if (BTRFS_I(inode)->outstanding_extents >
4167             BTRFS_I(inode)->reserved_extents) {
4168                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4169                         BTRFS_I(inode)->reserved_extents;
4170                 BTRFS_I(inode)->reserved_extents += nr_extents;
4171         }
4172
4173         /*
4174          * Add an item to reserve for updating the inode when we complete the
4175          * delalloc io.
4176          */
4177         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4178                 nr_extents++;
4179                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4180         }
4181
4182         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4183         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4184         spin_unlock(&BTRFS_I(inode)->lock);
4185
4186         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4187         if (ret) {
4188                 u64 to_free = 0;
4189                 unsigned dropped;
4190
4191                 spin_lock(&BTRFS_I(inode)->lock);
4192                 dropped = drop_outstanding_extent(inode);
4193                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4194                 spin_unlock(&BTRFS_I(inode)->lock);
4195                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4196
4197                 /*
4198                  * Somebody could have come in and twiddled with the
4199                  * reservation, so if we have to free more than we would have
4200                  * reserved from this reservation go ahead and release those
4201                  * bytes.
4202                  */
4203                 to_free -= to_reserve;
4204                 if (to_free)
4205                         btrfs_block_rsv_release(root, block_rsv, to_free);
4206                 return ret;
4207         }
4208
4209         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4210
4211         return 0;
4212 }
4213
4214 /**
4215  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4216  * @inode: the inode to release the reservation for
4217  * @num_bytes: the number of bytes we're releasing
4218  *
4219  * This will release the metadata reservation for an inode.  This can be called
4220  * once we complete IO for a given set of bytes to release their metadata
4221  * reservations.
4222  */
4223 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4224 {
4225         struct btrfs_root *root = BTRFS_I(inode)->root;
4226         u64 to_free = 0;
4227         unsigned dropped;
4228
4229         num_bytes = ALIGN(num_bytes, root->sectorsize);
4230         spin_lock(&BTRFS_I(inode)->lock);
4231         dropped = drop_outstanding_extent(inode);
4232
4233         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4234         spin_unlock(&BTRFS_I(inode)->lock);
4235         if (dropped > 0)
4236                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4237
4238         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4239                                 to_free);
4240 }
4241
4242 /**
4243  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4244  * @inode: inode we're writing to
4245  * @num_bytes: the number of bytes we want to allocate
4246  *
4247  * This will do the following things
4248  *
4249  * o reserve space in the data space info for num_bytes
4250  * o reserve space in the metadata space info based on number of outstanding
4251  *   extents and how much csums will be needed
4252  * o add to the inodes ->delalloc_bytes
4253  * o add it to the fs_info's delalloc inodes list.
4254  *
4255  * This will return 0 for success and -ENOSPC if there is no space left.
4256  */
4257 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4258 {
4259         int ret;
4260
4261         ret = btrfs_check_data_free_space(inode, num_bytes);
4262         if (ret)
4263                 return ret;
4264
4265         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4266         if (ret) {
4267                 btrfs_free_reserved_data_space(inode, num_bytes);
4268                 return ret;
4269         }
4270
4271         return 0;
4272 }
4273
4274 /**
4275  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4276  * @inode: inode we're releasing space for
4277  * @num_bytes: the number of bytes we want to free up
4278  *
4279  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4280  * called in the case that we don't need the metadata AND data reservations
4281  * anymore.  So if there is an error or we insert an inline extent.
4282  *
4283  * This function will release the metadata space that was not used and will
4284  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4285  * list if there are no delalloc bytes left.
4286  */
4287 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4288 {
4289         btrfs_delalloc_release_metadata(inode, num_bytes);
4290         btrfs_free_reserved_data_space(inode, num_bytes);
4291 }
4292
4293 static int update_block_group(struct btrfs_trans_handle *trans,
4294                               struct btrfs_root *root,
4295                               u64 bytenr, u64 num_bytes, int alloc)
4296 {
4297         struct btrfs_block_group_cache *cache = NULL;
4298         struct btrfs_fs_info *info = root->fs_info;
4299         u64 total = num_bytes;
4300         u64 old_val;
4301         u64 byte_in_group;
4302         int factor;
4303
4304         /* block accounting for super block */
4305         spin_lock(&info->delalloc_lock);
4306         old_val = btrfs_super_bytes_used(info->super_copy);
4307         if (alloc)
4308                 old_val += num_bytes;
4309         else
4310                 old_val -= num_bytes;
4311         btrfs_set_super_bytes_used(info->super_copy, old_val);
4312         spin_unlock(&info->delalloc_lock);
4313
4314         while (total) {
4315                 cache = btrfs_lookup_block_group(info, bytenr);
4316                 if (!cache)
4317                         return -1;
4318                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4319                                     BTRFS_BLOCK_GROUP_RAID1 |
4320                                     BTRFS_BLOCK_GROUP_RAID10))
4321                         factor = 2;
4322                 else
4323                         factor = 1;
4324                 /*
4325                  * If this block group has free space cache written out, we
4326                  * need to make sure to load it if we are removing space.  This
4327                  * is because we need the unpinning stage to actually add the
4328                  * space back to the block group, otherwise we will leak space.
4329                  */
4330                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4331                         cache_block_group(cache, trans, NULL, 1);
4332
4333                 byte_in_group = bytenr - cache->key.objectid;
4334                 WARN_ON(byte_in_group > cache->key.offset);
4335
4336                 spin_lock(&cache->space_info->lock);
4337                 spin_lock(&cache->lock);
4338
4339                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4340                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4341                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4342
4343                 cache->dirty = 1;
4344                 old_val = btrfs_block_group_used(&cache->item);
4345                 num_bytes = min(total, cache->key.offset - byte_in_group);
4346                 if (alloc) {
4347                         old_val += num_bytes;
4348                         btrfs_set_block_group_used(&cache->item, old_val);
4349                         cache->reserved -= num_bytes;
4350                         cache->space_info->bytes_reserved -= num_bytes;
4351                         cache->space_info->bytes_used += num_bytes;
4352                         cache->space_info->disk_used += num_bytes * factor;
4353                         spin_unlock(&cache->lock);
4354                         spin_unlock(&cache->space_info->lock);
4355                 } else {
4356                         old_val -= num_bytes;
4357                         btrfs_set_block_group_used(&cache->item, old_val);
4358                         cache->pinned += num_bytes;
4359                         cache->space_info->bytes_pinned += num_bytes;
4360                         cache->space_info->bytes_used -= num_bytes;
4361                         cache->space_info->disk_used -= num_bytes * factor;
4362                         spin_unlock(&cache->lock);
4363                         spin_unlock(&cache->space_info->lock);
4364
4365                         set_extent_dirty(info->pinned_extents,
4366                                          bytenr, bytenr + num_bytes - 1,
4367                                          GFP_NOFS | __GFP_NOFAIL);
4368                 }
4369                 btrfs_put_block_group(cache);
4370                 total -= num_bytes;
4371                 bytenr += num_bytes;
4372         }
4373         return 0;
4374 }
4375
4376 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4377 {
4378         struct btrfs_block_group_cache *cache;
4379         u64 bytenr;
4380
4381         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4382         if (!cache)
4383                 return 0;
4384
4385         bytenr = cache->key.objectid;
4386         btrfs_put_block_group(cache);
4387
4388         return bytenr;
4389 }
4390
4391 static int pin_down_extent(struct btrfs_root *root,
4392                            struct btrfs_block_group_cache *cache,
4393                            u64 bytenr, u64 num_bytes, int reserved)
4394 {
4395         spin_lock(&cache->space_info->lock);
4396         spin_lock(&cache->lock);
4397         cache->pinned += num_bytes;
4398         cache->space_info->bytes_pinned += num_bytes;
4399         if (reserved) {
4400                 cache->reserved -= num_bytes;
4401                 cache->space_info->bytes_reserved -= num_bytes;
4402         }
4403         spin_unlock(&cache->lock);
4404         spin_unlock(&cache->space_info->lock);
4405
4406         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4407                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4408         return 0;
4409 }
4410
4411 /*
4412  * this function must be called within transaction
4413  */
4414 int btrfs_pin_extent(struct btrfs_root *root,
4415                      u64 bytenr, u64 num_bytes, int reserved)
4416 {
4417         struct btrfs_block_group_cache *cache;
4418
4419         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4420         BUG_ON(!cache);
4421
4422         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4423
4424         btrfs_put_block_group(cache);
4425         return 0;
4426 }
4427
4428 /*
4429  * this function must be called within transaction
4430  */
4431 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4432                                     struct btrfs_root *root,
4433                                     u64 bytenr, u64 num_bytes)
4434 {
4435         struct btrfs_block_group_cache *cache;
4436
4437         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4438         BUG_ON(!cache);
4439
4440         /*
4441          * pull in the free space cache (if any) so that our pin
4442          * removes the free space from the cache.  We have load_only set
4443          * to one because the slow code to read in the free extents does check
4444          * the pinned extents.
4445          */
4446         cache_block_group(cache, trans, root, 1);
4447
4448         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4449
4450         /* remove us from the free space cache (if we're there at all) */
4451         btrfs_remove_free_space(cache, bytenr, num_bytes);
4452         btrfs_put_block_group(cache);
4453         return 0;
4454 }
4455
4456 /**
4457  * btrfs_update_reserved_bytes - update the block_group and space info counters
4458  * @cache:      The cache we are manipulating
4459  * @num_bytes:  The number of bytes in question
4460  * @reserve:    One of the reservation enums
4461  *
4462  * This is called by the allocator when it reserves space, or by somebody who is
4463  * freeing space that was never actually used on disk.  For example if you
4464  * reserve some space for a new leaf in transaction A and before transaction A
4465  * commits you free that leaf, you call this with reserve set to 0 in order to
4466  * clear the reservation.
4467  *
4468  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4469  * ENOSPC accounting.  For data we handle the reservation through clearing the
4470  * delalloc bits in the io_tree.  We have to do this since we could end up
4471  * allocating less disk space for the amount of data we have reserved in the
4472  * case of compression.
4473  *
4474  * If this is a reservation and the block group has become read only we cannot
4475  * make the reservation and return -EAGAIN, otherwise this function always
4476  * succeeds.
4477  */
4478 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4479                                        u64 num_bytes, int reserve)
4480 {
4481         struct btrfs_space_info *space_info = cache->space_info;
4482         int ret = 0;
4483         spin_lock(&space_info->lock);
4484         spin_lock(&cache->lock);
4485         if (reserve != RESERVE_FREE) {
4486                 if (cache->ro) {
4487                         ret = -EAGAIN;
4488                 } else {
4489                         cache->reserved += num_bytes;
4490                         space_info->bytes_reserved += num_bytes;
4491                         if (reserve == RESERVE_ALLOC) {
4492                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4493                                 space_info->bytes_may_use -= num_bytes;
4494                         }
4495                 }
4496         } else {
4497                 if (cache->ro)
4498                         space_info->bytes_readonly += num_bytes;
4499                 cache->reserved -= num_bytes;
4500                 space_info->bytes_reserved -= num_bytes;
4501                 space_info->reservation_progress++;
4502         }
4503         spin_unlock(&cache->lock);
4504         spin_unlock(&space_info->lock);
4505         return ret;
4506 }
4507
4508 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4509                                 struct btrfs_root *root)
4510 {
4511         struct btrfs_fs_info *fs_info = root->fs_info;
4512         struct btrfs_caching_control *next;
4513         struct btrfs_caching_control *caching_ctl;
4514         struct btrfs_block_group_cache *cache;
4515
4516         down_write(&fs_info->extent_commit_sem);
4517
4518         list_for_each_entry_safe(caching_ctl, next,
4519                                  &fs_info->caching_block_groups, list) {
4520                 cache = caching_ctl->block_group;
4521                 if (block_group_cache_done(cache)) {
4522                         cache->last_byte_to_unpin = (u64)-1;
4523                         list_del_init(&caching_ctl->list);
4524                         put_caching_control(caching_ctl);
4525                 } else {
4526                         cache->last_byte_to_unpin = caching_ctl->progress;
4527                 }
4528         }
4529
4530         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4531                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4532         else
4533                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4534
4535         up_write(&fs_info->extent_commit_sem);
4536
4537         update_global_block_rsv(fs_info);
4538         return 0;
4539 }
4540
4541 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4542 {
4543         struct btrfs_fs_info *fs_info = root->fs_info;
4544         struct btrfs_block_group_cache *cache = NULL;
4545         u64 len;
4546
4547         while (start <= end) {
4548                 if (!cache ||
4549                     start >= cache->key.objectid + cache->key.offset) {
4550                         if (cache)
4551                                 btrfs_put_block_group(cache);
4552                         cache = btrfs_lookup_block_group(fs_info, start);
4553                         BUG_ON(!cache);
4554                 }
4555
4556                 len = cache->key.objectid + cache->key.offset - start;
4557                 len = min(len, end + 1 - start);
4558
4559                 if (start < cache->last_byte_to_unpin) {
4560                         len = min(len, cache->last_byte_to_unpin - start);
4561                         btrfs_add_free_space(cache, start, len);
4562                 }
4563
4564                 start += len;
4565
4566                 spin_lock(&cache->space_info->lock);
4567                 spin_lock(&cache->lock);
4568                 cache->pinned -= len;
4569                 cache->space_info->bytes_pinned -= len;
4570                 if (cache->ro)
4571                         cache->space_info->bytes_readonly += len;
4572                 spin_unlock(&cache->lock);
4573                 spin_unlock(&cache->space_info->lock);
4574         }
4575
4576         if (cache)
4577                 btrfs_put_block_group(cache);
4578         return 0;
4579 }
4580
4581 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4582                                struct btrfs_root *root)
4583 {
4584         struct btrfs_fs_info *fs_info = root->fs_info;
4585         struct extent_io_tree *unpin;
4586         u64 start;
4587         u64 end;
4588         int ret;
4589
4590         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4591                 unpin = &fs_info->freed_extents[1];
4592         else
4593                 unpin = &fs_info->freed_extents[0];
4594
4595         while (1) {
4596                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4597                                             EXTENT_DIRTY);
4598                 if (ret)
4599                         break;
4600
4601                 if (btrfs_test_opt(root, DISCARD))
4602                         ret = btrfs_discard_extent(root, start,
4603                                                    end + 1 - start, NULL);
4604
4605                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4606                 unpin_extent_range(root, start, end);
4607                 cond_resched();
4608         }
4609
4610         return 0;
4611 }
4612
4613 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4614                                 struct btrfs_root *root,
4615                                 u64 bytenr, u64 num_bytes, u64 parent,
4616                                 u64 root_objectid, u64 owner_objectid,
4617                                 u64 owner_offset, int refs_to_drop,
4618                                 struct btrfs_delayed_extent_op *extent_op)
4619 {
4620         struct btrfs_key key;
4621         struct btrfs_path *path;
4622         struct btrfs_fs_info *info = root->fs_info;
4623         struct btrfs_root *extent_root = info->extent_root;
4624         struct extent_buffer *leaf;
4625         struct btrfs_extent_item *ei;
4626         struct btrfs_extent_inline_ref *iref;
4627         int ret;
4628         int is_data;
4629         int extent_slot = 0;
4630         int found_extent = 0;
4631         int num_to_del = 1;
4632         u32 item_size;
4633         u64 refs;
4634
4635         path = btrfs_alloc_path();
4636         if (!path)
4637                 return -ENOMEM;
4638
4639         path->reada = 1;
4640         path->leave_spinning = 1;
4641
4642         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4643         BUG_ON(!is_data && refs_to_drop != 1);
4644
4645         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4646                                     bytenr, num_bytes, parent,
4647                                     root_objectid, owner_objectid,
4648                                     owner_offset);
4649         if (ret == 0) {
4650                 extent_slot = path->slots[0];
4651                 while (extent_slot >= 0) {
4652                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4653                                               extent_slot);
4654                         if (key.objectid != bytenr)
4655                                 break;
4656                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4657                             key.offset == num_bytes) {
4658                                 found_extent = 1;
4659                                 break;
4660                         }
4661                         if (path->slots[0] - extent_slot > 5)
4662                                 break;
4663                         extent_slot--;
4664                 }
4665 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4666                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4667                 if (found_extent && item_size < sizeof(*ei))
4668                         found_extent = 0;
4669 #endif
4670                 if (!found_extent) {
4671                         BUG_ON(iref);
4672                         ret = remove_extent_backref(trans, extent_root, path,
4673                                                     NULL, refs_to_drop,
4674                                                     is_data);
4675                         BUG_ON(ret);
4676                         btrfs_release_path(path);
4677                         path->leave_spinning = 1;
4678
4679                         key.objectid = bytenr;
4680                         key.type = BTRFS_EXTENT_ITEM_KEY;
4681                         key.offset = num_bytes;
4682
4683                         ret = btrfs_search_slot(trans, extent_root,
4684                                                 &key, path, -1, 1);
4685                         if (ret) {
4686                                 printk(KERN_ERR "umm, got %d back from search"
4687                                        ", was looking for %llu\n", ret,
4688                                        (unsigned long long)bytenr);
4689                                 if (ret > 0)
4690                                         btrfs_print_leaf(extent_root,
4691                                                          path->nodes[0]);
4692                         }
4693                         BUG_ON(ret);
4694                         extent_slot = path->slots[0];
4695                 }
4696         } else {
4697                 btrfs_print_leaf(extent_root, path->nodes[0]);
4698                 WARN_ON(1);
4699                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4700                        "parent %llu root %llu  owner %llu offset %llu\n",
4701                        (unsigned long long)bytenr,
4702                        (unsigned long long)parent,
4703                        (unsigned long long)root_objectid,
4704                        (unsigned long long)owner_objectid,
4705                        (unsigned long long)owner_offset);
4706         }
4707
4708         leaf = path->nodes[0];
4709         item_size = btrfs_item_size_nr(leaf, extent_slot);
4710 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4711         if (item_size < sizeof(*ei)) {
4712                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4713                 ret = convert_extent_item_v0(trans, extent_root, path,
4714                                              owner_objectid, 0);
4715                 BUG_ON(ret < 0);
4716
4717                 btrfs_release_path(path);
4718                 path->leave_spinning = 1;
4719
4720                 key.objectid = bytenr;
4721                 key.type = BTRFS_EXTENT_ITEM_KEY;
4722                 key.offset = num_bytes;
4723
4724                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4725                                         -1, 1);
4726                 if (ret) {
4727                         printk(KERN_ERR "umm, got %d back from search"
4728                                ", was looking for %llu\n", ret,
4729                                (unsigned long long)bytenr);
4730                         btrfs_print_leaf(extent_root, path->nodes[0]);
4731                 }
4732                 BUG_ON(ret);
4733                 extent_slot = path->slots[0];
4734                 leaf = path->nodes[0];
4735                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4736         }
4737 #endif
4738         BUG_ON(item_size < sizeof(*ei));
4739         ei = btrfs_item_ptr(leaf, extent_slot,
4740                             struct btrfs_extent_item);
4741         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4742                 struct btrfs_tree_block_info *bi;
4743                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4744                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4745                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4746         }
4747
4748         refs = btrfs_extent_refs(leaf, ei);
4749         BUG_ON(refs < refs_to_drop);
4750         refs -= refs_to_drop;
4751
4752         if (refs > 0) {
4753                 if (extent_op)
4754                         __run_delayed_extent_op(extent_op, leaf, ei);
4755                 /*
4756                  * In the case of inline back ref, reference count will
4757                  * be updated by remove_extent_backref
4758                  */
4759                 if (iref) {
4760                         BUG_ON(!found_extent);
4761                 } else {
4762                         btrfs_set_extent_refs(leaf, ei, refs);
4763                         btrfs_mark_buffer_dirty(leaf);
4764                 }
4765                 if (found_extent) {
4766                         ret = remove_extent_backref(trans, extent_root, path,
4767                                                     iref, refs_to_drop,
4768                                                     is_data);
4769                         BUG_ON(ret);
4770                 }
4771         } else {
4772                 if (found_extent) {
4773                         BUG_ON(is_data && refs_to_drop !=
4774                                extent_data_ref_count(root, path, iref));
4775                         if (iref) {
4776                                 BUG_ON(path->slots[0] != extent_slot);
4777                         } else {
4778                                 BUG_ON(path->slots[0] != extent_slot + 1);
4779                                 path->slots[0] = extent_slot;
4780                                 num_to_del = 2;
4781                         }
4782                 }
4783
4784                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4785                                       num_to_del);
4786                 BUG_ON(ret);
4787                 btrfs_release_path(path);
4788
4789                 if (is_data) {
4790                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4791                         BUG_ON(ret);
4792                 } else {
4793                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4794                              bytenr >> PAGE_CACHE_SHIFT,
4795                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4796                 }
4797
4798                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4799                 BUG_ON(ret);
4800         }
4801         btrfs_free_path(path);
4802         return ret;
4803 }
4804
4805 /*
4806  * when we free an block, it is possible (and likely) that we free the last
4807  * delayed ref for that extent as well.  This searches the delayed ref tree for
4808  * a given extent, and if there are no other delayed refs to be processed, it
4809  * removes it from the tree.
4810  */
4811 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4812                                       struct btrfs_root *root, u64 bytenr)
4813 {
4814         struct btrfs_delayed_ref_head *head;
4815         struct btrfs_delayed_ref_root *delayed_refs;
4816         struct btrfs_delayed_ref_node *ref;
4817         struct rb_node *node;
4818         int ret = 0;
4819
4820         delayed_refs = &trans->transaction->delayed_refs;
4821         spin_lock(&delayed_refs->lock);
4822         head = btrfs_find_delayed_ref_head(trans, bytenr);
4823         if (!head)
4824                 goto out;
4825
4826         node = rb_prev(&head->node.rb_node);
4827         if (!node)
4828                 goto out;
4829
4830         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4831
4832         /* there are still entries for this ref, we can't drop it */
4833         if (ref->bytenr == bytenr)
4834                 goto out;
4835
4836         if (head->extent_op) {
4837                 if (!head->must_insert_reserved)
4838                         goto out;
4839                 kfree(head->extent_op);
4840                 head->extent_op = NULL;
4841         }
4842
4843         /*
4844          * waiting for the lock here would deadlock.  If someone else has it
4845          * locked they are already in the process of dropping it anyway
4846          */
4847         if (!mutex_trylock(&head->mutex))
4848                 goto out;
4849
4850         /*
4851          * at this point we have a head with no other entries.  Go
4852          * ahead and process it.
4853          */
4854         head->node.in_tree = 0;
4855         rb_erase(&head->node.rb_node, &delayed_refs->root);
4856
4857         delayed_refs->num_entries--;
4858
4859         /*
4860          * we don't take a ref on the node because we're removing it from the
4861          * tree, so we just steal the ref the tree was holding.
4862          */
4863         delayed_refs->num_heads--;
4864         if (list_empty(&head->cluster))
4865                 delayed_refs->num_heads_ready--;
4866
4867         list_del_init(&head->cluster);
4868         spin_unlock(&delayed_refs->lock);
4869
4870         BUG_ON(head->extent_op);
4871         if (head->must_insert_reserved)
4872                 ret = 1;
4873
4874         mutex_unlock(&head->mutex);
4875         btrfs_put_delayed_ref(&head->node);
4876         return ret;
4877 out:
4878         spin_unlock(&delayed_refs->lock);
4879         return 0;
4880 }
4881
4882 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4883                            struct btrfs_root *root,
4884                            struct extent_buffer *buf,
4885                            u64 parent, int last_ref)
4886 {
4887         struct btrfs_block_group_cache *cache = NULL;
4888         int ret;
4889
4890         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4891                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4892                                                 parent, root->root_key.objectid,
4893                                                 btrfs_header_level(buf),
4894                                                 BTRFS_DROP_DELAYED_REF, NULL);
4895                 BUG_ON(ret);
4896         }
4897
4898         if (!last_ref)
4899                 return;
4900
4901         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4902
4903         if (btrfs_header_generation(buf) == trans->transid) {
4904                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4905                         ret = check_ref_cleanup(trans, root, buf->start);
4906                         if (!ret)
4907                                 goto out;
4908                 }
4909
4910                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4911                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4912                         goto out;
4913                 }
4914
4915                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4916
4917                 btrfs_add_free_space(cache, buf->start, buf->len);
4918                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4919         }
4920 out:
4921         /*
4922          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4923          * anymore.
4924          */
4925         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4926         btrfs_put_block_group(cache);
4927 }
4928
4929 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4930                       struct btrfs_root *root,
4931                       u64 bytenr, u64 num_bytes, u64 parent,
4932                       u64 root_objectid, u64 owner, u64 offset)
4933 {
4934         int ret;
4935
4936         /*
4937          * tree log blocks never actually go into the extent allocation
4938          * tree, just update pinning info and exit early.
4939          */
4940         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4941                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4942                 /* unlocks the pinned mutex */
4943                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4944                 ret = 0;
4945         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4946                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4947                                         parent, root_objectid, (int)owner,
4948                                         BTRFS_DROP_DELAYED_REF, NULL);
4949                 BUG_ON(ret);
4950         } else {
4951                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4952                                         parent, root_objectid, owner,
4953                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4954                 BUG_ON(ret);
4955         }
4956         return ret;
4957 }
4958
4959 static u64 stripe_align(struct btrfs_root *root, u64 val)
4960 {
4961         u64 mask = ((u64)root->stripesize - 1);
4962         u64 ret = (val + mask) & ~mask;
4963         return ret;
4964 }
4965
4966 /*
4967  * when we wait for progress in the block group caching, its because
4968  * our allocation attempt failed at least once.  So, we must sleep
4969  * and let some progress happen before we try again.
4970  *
4971  * This function will sleep at least once waiting for new free space to
4972  * show up, and then it will check the block group free space numbers
4973  * for our min num_bytes.  Another option is to have it go ahead
4974  * and look in the rbtree for a free extent of a given size, but this
4975  * is a good start.
4976  */
4977 static noinline int
4978 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4979                                 u64 num_bytes)
4980 {
4981         struct btrfs_caching_control *caching_ctl;
4982         DEFINE_WAIT(wait);
4983
4984         caching_ctl = get_caching_control(cache);
4985         if (!caching_ctl)
4986                 return 0;
4987
4988         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4989                    (cache->free_space_ctl->free_space >= num_bytes));
4990
4991         put_caching_control(caching_ctl);
4992         return 0;
4993 }
4994
4995 static noinline int
4996 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4997 {
4998         struct btrfs_caching_control *caching_ctl;
4999         DEFINE_WAIT(wait);
5000
5001         caching_ctl = get_caching_control(cache);
5002         if (!caching_ctl)
5003                 return 0;
5004
5005         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5006
5007         put_caching_control(caching_ctl);
5008         return 0;
5009 }
5010
5011 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5012 {
5013         int index;
5014         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5015                 index = 0;
5016         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5017                 index = 1;
5018         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5019                 index = 2;
5020         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5021                 index = 3;
5022         else
5023                 index = 4;
5024         return index;
5025 }
5026
5027 enum btrfs_loop_type {
5028         LOOP_FIND_IDEAL = 0,
5029         LOOP_CACHING_NOWAIT = 1,
5030         LOOP_CACHING_WAIT = 2,
5031         LOOP_ALLOC_CHUNK = 3,
5032         LOOP_NO_EMPTY_SIZE = 4,
5033 };
5034
5035 /*
5036  * walks the btree of allocated extents and find a hole of a given size.
5037  * The key ins is changed to record the hole:
5038  * ins->objectid == block start
5039  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5040  * ins->offset == number of blocks
5041  * Any available blocks before search_start are skipped.
5042  */
5043 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5044                                      struct btrfs_root *orig_root,
5045                                      u64 num_bytes, u64 empty_size,
5046                                      u64 search_start, u64 search_end,
5047                                      u64 hint_byte, struct btrfs_key *ins,
5048                                      u64 data)
5049 {
5050         int ret = 0;
5051         struct btrfs_root *root = orig_root->fs_info->extent_root;
5052         struct btrfs_free_cluster *last_ptr = NULL;
5053         struct btrfs_block_group_cache *block_group = NULL;
5054         int empty_cluster = 2 * 1024 * 1024;
5055         int allowed_chunk_alloc = 0;
5056         int done_chunk_alloc = 0;
5057         struct btrfs_space_info *space_info;
5058         int last_ptr_loop = 0;
5059         int loop = 0;
5060         int index = 0;
5061         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5062                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5063         bool found_uncached_bg = false;
5064         bool failed_cluster_refill = false;
5065         bool failed_alloc = false;
5066         bool use_cluster = true;
5067         bool have_caching_bg = false;
5068         u64 ideal_cache_percent = 0;
5069         u64 ideal_cache_offset = 0;
5070
5071         WARN_ON(num_bytes < root->sectorsize);
5072         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5073         ins->objectid = 0;
5074         ins->offset = 0;
5075
5076         space_info = __find_space_info(root->fs_info, data);
5077         if (!space_info) {
5078                 printk(KERN_ERR "No space info for %llu\n", data);
5079                 return -ENOSPC;
5080         }
5081
5082         /*
5083          * If the space info is for both data and metadata it means we have a
5084          * small filesystem and we can't use the clustering stuff.
5085          */
5086         if (btrfs_mixed_space_info(space_info))
5087                 use_cluster = false;
5088
5089         if (orig_root->ref_cows || empty_size)
5090                 allowed_chunk_alloc = 1;
5091
5092         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5093                 last_ptr = &root->fs_info->meta_alloc_cluster;
5094                 if (!btrfs_test_opt(root, SSD))
5095                         empty_cluster = 64 * 1024;
5096         }
5097
5098         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5099             btrfs_test_opt(root, SSD)) {
5100                 last_ptr = &root->fs_info->data_alloc_cluster;
5101         }
5102
5103         if (last_ptr) {
5104                 spin_lock(&last_ptr->lock);
5105                 if (last_ptr->block_group)
5106                         hint_byte = last_ptr->window_start;
5107                 spin_unlock(&last_ptr->lock);
5108         }
5109
5110         search_start = max(search_start, first_logical_byte(root, 0));
5111         search_start = max(search_start, hint_byte);
5112
5113         if (!last_ptr)
5114                 empty_cluster = 0;
5115
5116         if (search_start == hint_byte) {
5117 ideal_cache:
5118                 block_group = btrfs_lookup_block_group(root->fs_info,
5119                                                        search_start);
5120                 /*
5121                  * we don't want to use the block group if it doesn't match our
5122                  * allocation bits, or if its not cached.
5123                  *
5124                  * However if we are re-searching with an ideal block group
5125                  * picked out then we don't care that the block group is cached.
5126                  */
5127                 if (block_group && block_group_bits(block_group, data) &&
5128                     (block_group->cached != BTRFS_CACHE_NO ||
5129                      search_start == ideal_cache_offset)) {
5130                         down_read(&space_info->groups_sem);
5131                         if (list_empty(&block_group->list) ||
5132                             block_group->ro) {
5133                                 /*
5134                                  * someone is removing this block group,
5135                                  * we can't jump into the have_block_group
5136                                  * target because our list pointers are not
5137                                  * valid
5138                                  */
5139                                 btrfs_put_block_group(block_group);
5140                                 up_read(&space_info->groups_sem);
5141                         } else {
5142                                 index = get_block_group_index(block_group);
5143                                 goto have_block_group;
5144                         }
5145                 } else if (block_group) {
5146                         btrfs_put_block_group(block_group);
5147                 }
5148         }
5149 search:
5150         have_caching_bg = false;
5151         down_read(&space_info->groups_sem);
5152         list_for_each_entry(block_group, &space_info->block_groups[index],
5153                             list) {
5154                 u64 offset;
5155                 int cached;
5156
5157                 btrfs_get_block_group(block_group);
5158                 search_start = block_group->key.objectid;
5159
5160                 /*
5161                  * this can happen if we end up cycling through all the
5162                  * raid types, but we want to make sure we only allocate
5163                  * for the proper type.
5164                  */
5165                 if (!block_group_bits(block_group, data)) {
5166                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5167                                 BTRFS_BLOCK_GROUP_RAID1 |
5168                                 BTRFS_BLOCK_GROUP_RAID10;
5169
5170                         /*
5171                          * if they asked for extra copies and this block group
5172                          * doesn't provide them, bail.  This does allow us to
5173                          * fill raid0 from raid1.
5174                          */
5175                         if ((data & extra) && !(block_group->flags & extra))
5176                                 goto loop;
5177                 }
5178
5179 have_block_group:
5180                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5181                         u64 free_percent;
5182
5183                         ret = cache_block_group(block_group, trans,
5184                                                 orig_root, 1);
5185                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5186                                 goto have_block_group;
5187
5188                         free_percent = btrfs_block_group_used(&block_group->item);
5189                         free_percent *= 100;
5190                         free_percent = div64_u64(free_percent,
5191                                                  block_group->key.offset);
5192                         free_percent = 100 - free_percent;
5193                         if (free_percent > ideal_cache_percent &&
5194                             likely(!block_group->ro)) {
5195                                 ideal_cache_offset = block_group->key.objectid;
5196                                 ideal_cache_percent = free_percent;
5197                         }
5198
5199                         /*
5200                          * The caching workers are limited to 2 threads, so we
5201                          * can queue as much work as we care to.
5202                          */
5203                         if (loop > LOOP_FIND_IDEAL) {
5204                                 ret = cache_block_group(block_group, trans,
5205                                                         orig_root, 0);
5206                                 BUG_ON(ret);
5207                         }
5208                         found_uncached_bg = true;
5209
5210                         /*
5211                          * If loop is set for cached only, try the next block
5212                          * group.
5213                          */
5214                         if (loop == LOOP_FIND_IDEAL)
5215                                 goto loop;
5216                 }
5217
5218                 cached = block_group_cache_done(block_group);
5219                 if (unlikely(!cached))
5220                         found_uncached_bg = true;
5221
5222                 if (unlikely(block_group->ro))
5223                         goto loop;
5224
5225                 spin_lock(&block_group->free_space_ctl->tree_lock);
5226                 if (cached &&
5227                     block_group->free_space_ctl->free_space <
5228                     num_bytes + empty_size) {
5229                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5230                         goto loop;
5231                 }
5232                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5233
5234                 /*
5235                  * Ok we want to try and use the cluster allocator, so lets look
5236                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5237                  * have tried the cluster allocator plenty of times at this
5238                  * point and not have found anything, so we are likely way too
5239                  * fragmented for the clustering stuff to find anything, so lets
5240                  * just skip it and let the allocator find whatever block it can
5241                  * find
5242                  */
5243                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5244                         /*
5245                          * the refill lock keeps out other
5246                          * people trying to start a new cluster
5247                          */
5248                         spin_lock(&last_ptr->refill_lock);
5249                         if (last_ptr->block_group &&
5250                             (last_ptr->block_group->ro ||
5251                             !block_group_bits(last_ptr->block_group, data))) {
5252                                 offset = 0;
5253                                 goto refill_cluster;
5254                         }
5255
5256                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5257                                                  num_bytes, search_start);
5258                         if (offset) {
5259                                 /* we have a block, we're done */
5260                                 spin_unlock(&last_ptr->refill_lock);
5261                                 goto checks;
5262                         }
5263
5264                         spin_lock(&last_ptr->lock);
5265                         /*
5266                          * whoops, this cluster doesn't actually point to
5267                          * this block group.  Get a ref on the block
5268                          * group is does point to and try again
5269                          */
5270                         if (!last_ptr_loop && last_ptr->block_group &&
5271                             last_ptr->block_group != block_group &&
5272                             index <=
5273                                  get_block_group_index(last_ptr->block_group)) {
5274
5275                                 btrfs_put_block_group(block_group);
5276                                 block_group = last_ptr->block_group;
5277                                 btrfs_get_block_group(block_group);
5278                                 spin_unlock(&last_ptr->lock);
5279                                 spin_unlock(&last_ptr->refill_lock);
5280
5281                                 last_ptr_loop = 1;
5282                                 search_start = block_group->key.objectid;
5283                                 /*
5284                                  * we know this block group is properly
5285                                  * in the list because
5286                                  * btrfs_remove_block_group, drops the
5287                                  * cluster before it removes the block
5288                                  * group from the list
5289                                  */
5290                                 goto have_block_group;
5291                         }
5292                         spin_unlock(&last_ptr->lock);
5293 refill_cluster:
5294                         /*
5295                          * this cluster didn't work out, free it and
5296                          * start over
5297                          */
5298                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5299
5300                         last_ptr_loop = 0;
5301
5302                         /* allocate a cluster in this block group */
5303                         ret = btrfs_find_space_cluster(trans, root,
5304                                                block_group, last_ptr,
5305                                                offset, num_bytes,
5306                                                empty_cluster + empty_size);
5307                         if (ret == 0) {
5308                                 /*
5309                                  * now pull our allocation out of this
5310                                  * cluster
5311                                  */
5312                                 offset = btrfs_alloc_from_cluster(block_group,
5313                                                   last_ptr, num_bytes,
5314                                                   search_start);
5315                                 if (offset) {
5316                                         /* we found one, proceed */
5317                                         spin_unlock(&last_ptr->refill_lock);
5318                                         goto checks;
5319                                 }
5320                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5321                                    && !failed_cluster_refill) {
5322                                 spin_unlock(&last_ptr->refill_lock);
5323
5324                                 failed_cluster_refill = true;
5325                                 wait_block_group_cache_progress(block_group,
5326                                        num_bytes + empty_cluster + empty_size);
5327                                 goto have_block_group;
5328                         }
5329
5330                         /*
5331                          * at this point we either didn't find a cluster
5332                          * or we weren't able to allocate a block from our
5333                          * cluster.  Free the cluster we've been trying
5334                          * to use, and go to the next block group
5335                          */
5336                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5337                         spin_unlock(&last_ptr->refill_lock);
5338                         goto loop;
5339                 }
5340
5341                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5342                                                     num_bytes, empty_size);
5343                 /*
5344                  * If we didn't find a chunk, and we haven't failed on this
5345                  * block group before, and this block group is in the middle of
5346                  * caching and we are ok with waiting, then go ahead and wait
5347                  * for progress to be made, and set failed_alloc to true.
5348                  *
5349                  * If failed_alloc is true then we've already waited on this
5350                  * block group once and should move on to the next block group.
5351                  */
5352                 if (!offset && !failed_alloc && !cached &&
5353                     loop > LOOP_CACHING_NOWAIT) {
5354                         wait_block_group_cache_progress(block_group,
5355                                                 num_bytes + empty_size);
5356                         failed_alloc = true;
5357                         goto have_block_group;
5358                 } else if (!offset) {
5359                         if (!cached)
5360                                 have_caching_bg = true;
5361                         goto loop;
5362                 }
5363 checks:
5364                 search_start = stripe_align(root, offset);
5365                 /* move on to the next group */
5366                 if (search_start + num_bytes >= search_end) {
5367                         btrfs_add_free_space(block_group, offset, num_bytes);
5368                         goto loop;
5369                 }
5370
5371                 /* move on to the next group */
5372                 if (search_start + num_bytes >
5373                     block_group->key.objectid + block_group->key.offset) {
5374                         btrfs_add_free_space(block_group, offset, num_bytes);
5375                         goto loop;
5376                 }
5377
5378                 ins->objectid = search_start;
5379                 ins->offset = num_bytes;
5380
5381                 if (offset < search_start)
5382                         btrfs_add_free_space(block_group, offset,
5383                                              search_start - offset);
5384                 BUG_ON(offset > search_start);
5385
5386                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5387                                                   alloc_type);
5388                 if (ret == -EAGAIN) {
5389                         btrfs_add_free_space(block_group, offset, num_bytes);
5390                         goto loop;
5391                 }
5392
5393                 /* we are all good, lets return */
5394                 ins->objectid = search_start;
5395                 ins->offset = num_bytes;
5396
5397                 if (offset < search_start)
5398                         btrfs_add_free_space(block_group, offset,
5399                                              search_start - offset);
5400                 BUG_ON(offset > search_start);
5401                 btrfs_put_block_group(block_group);
5402                 break;
5403 loop:
5404                 failed_cluster_refill = false;
5405                 failed_alloc = false;
5406                 BUG_ON(index != get_block_group_index(block_group));
5407                 btrfs_put_block_group(block_group);
5408         }
5409         up_read(&space_info->groups_sem);
5410
5411         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5412                 goto search;
5413
5414         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5415                 goto search;
5416
5417         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5418          *                      for them to make caching progress.  Also
5419          *                      determine the best possible bg to cache
5420          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5421          *                      caching kthreads as we move along
5422          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5423          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5424          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5425          *                      again
5426          */
5427         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5428                 index = 0;
5429                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5430                         found_uncached_bg = false;
5431                         loop++;
5432                         if (!ideal_cache_percent)
5433                                 goto search;
5434
5435                         /*
5436                          * 1 of the following 2 things have happened so far
5437                          *
5438                          * 1) We found an ideal block group for caching that
5439                          * is mostly full and will cache quickly, so we might
5440                          * as well wait for it.
5441                          *
5442                          * 2) We searched for cached only and we didn't find
5443                          * anything, and we didn't start any caching kthreads
5444                          * either, so chances are we will loop through and
5445                          * start a couple caching kthreads, and then come back
5446                          * around and just wait for them.  This will be slower
5447                          * because we will have 2 caching kthreads reading at
5448                          * the same time when we could have just started one
5449                          * and waited for it to get far enough to give us an
5450                          * allocation, so go ahead and go to the wait caching
5451                          * loop.
5452                          */
5453                         loop = LOOP_CACHING_WAIT;
5454                         search_start = ideal_cache_offset;
5455                         ideal_cache_percent = 0;
5456                         goto ideal_cache;
5457                 } else if (loop == LOOP_FIND_IDEAL) {
5458                         /*
5459                          * Didn't find a uncached bg, wait on anything we find
5460                          * next.
5461                          */
5462                         loop = LOOP_CACHING_WAIT;
5463                         goto search;
5464                 }
5465
5466                 loop++;
5467
5468                 if (loop == LOOP_ALLOC_CHUNK) {
5469                        if (allowed_chunk_alloc) {
5470                                 ret = do_chunk_alloc(trans, root, num_bytes +
5471                                                      2 * 1024 * 1024, data,
5472                                                      CHUNK_ALLOC_LIMITED);
5473                                 allowed_chunk_alloc = 0;
5474                                 if (ret == 1)
5475                                         done_chunk_alloc = 1;
5476                         } else if (!done_chunk_alloc &&
5477                                    space_info->force_alloc ==
5478                                    CHUNK_ALLOC_NO_FORCE) {
5479                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5480                         }
5481
5482                        /*
5483                         * We didn't allocate a chunk, go ahead and drop the
5484                         * empty size and loop again.
5485                         */
5486                        if (!done_chunk_alloc)
5487                                loop = LOOP_NO_EMPTY_SIZE;
5488                 }
5489
5490                 if (loop == LOOP_NO_EMPTY_SIZE) {
5491                         empty_size = 0;
5492                         empty_cluster = 0;
5493                 }
5494
5495                 goto search;
5496         } else if (!ins->objectid) {
5497                 ret = -ENOSPC;
5498         } else if (ins->objectid) {
5499                 ret = 0;
5500         }
5501
5502         return ret;
5503 }
5504
5505 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5506                             int dump_block_groups)
5507 {
5508         struct btrfs_block_group_cache *cache;
5509         int index = 0;
5510
5511         spin_lock(&info->lock);
5512         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5513                (unsigned long long)info->flags,
5514                (unsigned long long)(info->total_bytes - info->bytes_used -
5515                                     info->bytes_pinned - info->bytes_reserved -
5516                                     info->bytes_readonly),
5517                (info->full) ? "" : "not ");
5518         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5519                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5520                (unsigned long long)info->total_bytes,
5521                (unsigned long long)info->bytes_used,
5522                (unsigned long long)info->bytes_pinned,
5523                (unsigned long long)info->bytes_reserved,
5524                (unsigned long long)info->bytes_may_use,
5525                (unsigned long long)info->bytes_readonly);
5526         spin_unlock(&info->lock);
5527
5528         if (!dump_block_groups)
5529                 return;
5530
5531         down_read(&info->groups_sem);
5532 again:
5533         list_for_each_entry(cache, &info->block_groups[index], list) {
5534                 spin_lock(&cache->lock);
5535                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5536                        "%llu pinned %llu reserved\n",
5537                        (unsigned long long)cache->key.objectid,
5538                        (unsigned long long)cache->key.offset,
5539                        (unsigned long long)btrfs_block_group_used(&cache->item),
5540                        (unsigned long long)cache->pinned,
5541                        (unsigned long long)cache->reserved);
5542                 btrfs_dump_free_space(cache, bytes);
5543                 spin_unlock(&cache->lock);
5544         }
5545         if (++index < BTRFS_NR_RAID_TYPES)
5546                 goto again;
5547         up_read(&info->groups_sem);
5548 }
5549
5550 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5551                          struct btrfs_root *root,
5552                          u64 num_bytes, u64 min_alloc_size,
5553                          u64 empty_size, u64 hint_byte,
5554                          u64 search_end, struct btrfs_key *ins,
5555                          u64 data)
5556 {
5557         int ret;
5558         u64 search_start = 0;
5559
5560         data = btrfs_get_alloc_profile(root, data);
5561 again:
5562         /*
5563          * the only place that sets empty_size is btrfs_realloc_node, which
5564          * is not called recursively on allocations
5565          */
5566         if (empty_size || root->ref_cows)
5567                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5568                                      num_bytes + 2 * 1024 * 1024, data,
5569                                      CHUNK_ALLOC_NO_FORCE);
5570
5571         WARN_ON(num_bytes < root->sectorsize);
5572         ret = find_free_extent(trans, root, num_bytes, empty_size,
5573                                search_start, search_end, hint_byte,
5574                                ins, data);
5575
5576         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5577                 num_bytes = num_bytes >> 1;
5578                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5579                 num_bytes = max(num_bytes, min_alloc_size);
5580                 do_chunk_alloc(trans, root->fs_info->extent_root,
5581                                num_bytes, data, CHUNK_ALLOC_FORCE);
5582                 goto again;
5583         }
5584         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5585                 struct btrfs_space_info *sinfo;
5586
5587                 sinfo = __find_space_info(root->fs_info, data);
5588                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5589                        "wanted %llu\n", (unsigned long long)data,
5590                        (unsigned long long)num_bytes);
5591                 dump_space_info(sinfo, num_bytes, 1);
5592         }
5593
5594         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5595
5596         return ret;
5597 }
5598
5599 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5600                                         u64 start, u64 len, int pin)
5601 {
5602         struct btrfs_block_group_cache *cache;
5603         int ret = 0;
5604
5605         cache = btrfs_lookup_block_group(root->fs_info, start);
5606         if (!cache) {
5607                 printk(KERN_ERR "Unable to find block group for %llu\n",
5608                        (unsigned long long)start);
5609                 return -ENOSPC;
5610         }
5611
5612         if (btrfs_test_opt(root, DISCARD))
5613                 ret = btrfs_discard_extent(root, start, len, NULL);
5614
5615         if (pin)
5616                 pin_down_extent(root, cache, start, len, 1);
5617         else {
5618                 btrfs_add_free_space(cache, start, len);
5619                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5620         }
5621         btrfs_put_block_group(cache);
5622
5623         trace_btrfs_reserved_extent_free(root, start, len);
5624
5625         return ret;
5626 }
5627
5628 int btrfs_free_reserved_extent(struct btrfs_root *root,
5629                                         u64 start, u64 len)
5630 {
5631         return __btrfs_free_reserved_extent(root, start, len, 0);
5632 }
5633
5634 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5635                                        u64 start, u64 len)
5636 {
5637         return __btrfs_free_reserved_extent(root, start, len, 1);
5638 }
5639
5640 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5641                                       struct btrfs_root *root,
5642                                       u64 parent, u64 root_objectid,
5643                                       u64 flags, u64 owner, u64 offset,
5644                                       struct btrfs_key *ins, int ref_mod)
5645 {
5646         int ret;
5647         struct btrfs_fs_info *fs_info = root->fs_info;
5648         struct btrfs_extent_item *extent_item;
5649         struct btrfs_extent_inline_ref *iref;
5650         struct btrfs_path *path;
5651         struct extent_buffer *leaf;
5652         int type;
5653         u32 size;
5654
5655         if (parent > 0)
5656                 type = BTRFS_SHARED_DATA_REF_KEY;
5657         else
5658                 type = BTRFS_EXTENT_DATA_REF_KEY;
5659
5660         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5661
5662         path = btrfs_alloc_path();
5663         if (!path)
5664                 return -ENOMEM;
5665
5666         path->leave_spinning = 1;
5667         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5668                                       ins, size);
5669         BUG_ON(ret);
5670
5671         leaf = path->nodes[0];
5672         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5673                                      struct btrfs_extent_item);
5674         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5675         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5676         btrfs_set_extent_flags(leaf, extent_item,
5677                                flags | BTRFS_EXTENT_FLAG_DATA);
5678
5679         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5680         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5681         if (parent > 0) {
5682                 struct btrfs_shared_data_ref *ref;
5683                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5684                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5685                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5686         } else {
5687                 struct btrfs_extent_data_ref *ref;
5688                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5689                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5690                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5691                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5692                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5693         }
5694
5695         btrfs_mark_buffer_dirty(path->nodes[0]);
5696         btrfs_free_path(path);
5697
5698         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5699         if (ret) {
5700                 printk(KERN_ERR "btrfs update block group failed for %llu "
5701                        "%llu\n", (unsigned long long)ins->objectid,
5702                        (unsigned long long)ins->offset);
5703                 BUG();
5704         }
5705         return ret;
5706 }
5707
5708 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5709                                      struct btrfs_root *root,
5710                                      u64 parent, u64 root_objectid,
5711                                      u64 flags, struct btrfs_disk_key *key,
5712                                      int level, struct btrfs_key *ins)
5713 {
5714         int ret;
5715         struct btrfs_fs_info *fs_info = root->fs_info;
5716         struct btrfs_extent_item *extent_item;
5717         struct btrfs_tree_block_info *block_info;
5718         struct btrfs_extent_inline_ref *iref;
5719         struct btrfs_path *path;
5720         struct extent_buffer *leaf;
5721         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5722
5723         path = btrfs_alloc_path();
5724         if (!path)
5725                 return -ENOMEM;
5726
5727         path->leave_spinning = 1;
5728         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5729                                       ins, size);
5730         BUG_ON(ret);
5731
5732         leaf = path->nodes[0];
5733         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5734                                      struct btrfs_extent_item);
5735         btrfs_set_extent_refs(leaf, extent_item, 1);
5736         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5737         btrfs_set_extent_flags(leaf, extent_item,
5738                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5739         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5740
5741         btrfs_set_tree_block_key(leaf, block_info, key);
5742         btrfs_set_tree_block_level(leaf, block_info, level);
5743
5744         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5745         if (parent > 0) {
5746                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5747                 btrfs_set_extent_inline_ref_type(leaf, iref,
5748                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5749                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5750         } else {
5751                 btrfs_set_extent_inline_ref_type(leaf, iref,
5752                                                  BTRFS_TREE_BLOCK_REF_KEY);
5753                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5754         }
5755
5756         btrfs_mark_buffer_dirty(leaf);
5757         btrfs_free_path(path);
5758
5759         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5760         if (ret) {
5761                 printk(KERN_ERR "btrfs update block group failed for %llu "
5762                        "%llu\n", (unsigned long long)ins->objectid,
5763                        (unsigned long long)ins->offset);
5764                 BUG();
5765         }
5766         return ret;
5767 }
5768
5769 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5770                                      struct btrfs_root *root,
5771                                      u64 root_objectid, u64 owner,
5772                                      u64 offset, struct btrfs_key *ins)
5773 {
5774         int ret;
5775
5776         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5777
5778         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5779                                          0, root_objectid, owner, offset,
5780                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5781         return ret;
5782 }
5783
5784 /*
5785  * this is used by the tree logging recovery code.  It records that
5786  * an extent has been allocated and makes sure to clear the free
5787  * space cache bits as well
5788  */
5789 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5790                                    struct btrfs_root *root,
5791                                    u64 root_objectid, u64 owner, u64 offset,
5792                                    struct btrfs_key *ins)
5793 {
5794         int ret;
5795         struct btrfs_block_group_cache *block_group;
5796         struct btrfs_caching_control *caching_ctl;
5797         u64 start = ins->objectid;
5798         u64 num_bytes = ins->offset;
5799
5800         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5801         cache_block_group(block_group, trans, NULL, 0);
5802         caching_ctl = get_caching_control(block_group);
5803
5804         if (!caching_ctl) {
5805                 BUG_ON(!block_group_cache_done(block_group));
5806                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5807                 BUG_ON(ret);
5808         } else {
5809                 mutex_lock(&caching_ctl->mutex);
5810
5811                 if (start >= caching_ctl->progress) {
5812                         ret = add_excluded_extent(root, start, num_bytes);
5813                         BUG_ON(ret);
5814                 } else if (start + num_bytes <= caching_ctl->progress) {
5815                         ret = btrfs_remove_free_space(block_group,
5816                                                       start, num_bytes);
5817                         BUG_ON(ret);
5818                 } else {
5819                         num_bytes = caching_ctl->progress - start;
5820                         ret = btrfs_remove_free_space(block_group,
5821                                                       start, num_bytes);
5822                         BUG_ON(ret);
5823
5824                         start = caching_ctl->progress;
5825                         num_bytes = ins->objectid + ins->offset -
5826                                     caching_ctl->progress;
5827                         ret = add_excluded_extent(root, start, num_bytes);
5828                         BUG_ON(ret);
5829                 }
5830
5831                 mutex_unlock(&caching_ctl->mutex);
5832                 put_caching_control(caching_ctl);
5833         }
5834
5835         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5836                                           RESERVE_ALLOC_NO_ACCOUNT);
5837         BUG_ON(ret);
5838         btrfs_put_block_group(block_group);
5839         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5840                                          0, owner, offset, ins, 1);
5841         return ret;
5842 }
5843
5844 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5845                                             struct btrfs_root *root,
5846                                             u64 bytenr, u32 blocksize,
5847                                             int level)
5848 {
5849         struct extent_buffer *buf;
5850
5851         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5852         if (!buf)
5853                 return ERR_PTR(-ENOMEM);
5854         btrfs_set_header_generation(buf, trans->transid);
5855         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5856         btrfs_tree_lock(buf);
5857         clean_tree_block(trans, root, buf);
5858
5859         btrfs_set_lock_blocking(buf);
5860         btrfs_set_buffer_uptodate(buf);
5861
5862         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5863                 /*
5864                  * we allow two log transactions at a time, use different
5865                  * EXENT bit to differentiate dirty pages.
5866                  */
5867                 if (root->log_transid % 2 == 0)
5868                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5869                                         buf->start + buf->len - 1, GFP_NOFS);
5870                 else
5871                         set_extent_new(&root->dirty_log_pages, buf->start,
5872                                         buf->start + buf->len - 1, GFP_NOFS);
5873         } else {
5874                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5875                          buf->start + buf->len - 1, GFP_NOFS);
5876         }
5877         trans->blocks_used++;
5878         /* this returns a buffer locked for blocking */
5879         return buf;
5880 }
5881
5882 static struct btrfs_block_rsv *
5883 use_block_rsv(struct btrfs_trans_handle *trans,
5884               struct btrfs_root *root, u32 blocksize)
5885 {
5886         struct btrfs_block_rsv *block_rsv;
5887         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5888         int ret;
5889
5890         block_rsv = get_block_rsv(trans, root);
5891
5892         if (block_rsv->size == 0) {
5893                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5894                 /*
5895                  * If we couldn't reserve metadata bytes try and use some from
5896                  * the global reserve.
5897                  */
5898                 if (ret && block_rsv != global_rsv) {
5899                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5900                         if (!ret)
5901                                 return global_rsv;
5902                         return ERR_PTR(ret);
5903                 } else if (ret) {
5904                         return ERR_PTR(ret);
5905                 }
5906                 return block_rsv;
5907         }
5908
5909         ret = block_rsv_use_bytes(block_rsv, blocksize);
5910         if (!ret)
5911                 return block_rsv;
5912         if (ret) {
5913                 static DEFINE_RATELIMIT_STATE(_rs,
5914                                 DEFAULT_RATELIMIT_INTERVAL,
5915                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
5916                 if (__ratelimit(&_rs)) {
5917                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
5918                         WARN_ON(1);
5919                 }
5920                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5921                 if (!ret) {
5922                         return block_rsv;
5923                 } else if (ret && block_rsv != global_rsv) {
5924                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5925                         if (!ret)
5926                                 return global_rsv;
5927                 }
5928         }
5929
5930         return ERR_PTR(-ENOSPC);
5931 }
5932
5933 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5934 {
5935         block_rsv_add_bytes(block_rsv, blocksize, 0);
5936         block_rsv_release_bytes(block_rsv, NULL, 0);
5937 }
5938
5939 /*
5940  * finds a free extent and does all the dirty work required for allocation
5941  * returns the key for the extent through ins, and a tree buffer for
5942  * the first block of the extent through buf.
5943  *
5944  * returns the tree buffer or NULL.
5945  */
5946 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5947                                         struct btrfs_root *root, u32 blocksize,
5948                                         u64 parent, u64 root_objectid,
5949                                         struct btrfs_disk_key *key, int level,
5950                                         u64 hint, u64 empty_size)
5951 {
5952         struct btrfs_key ins;
5953         struct btrfs_block_rsv *block_rsv;
5954         struct extent_buffer *buf;
5955         u64 flags = 0;
5956         int ret;
5957
5958
5959         block_rsv = use_block_rsv(trans, root, blocksize);
5960         if (IS_ERR(block_rsv))
5961                 return ERR_CAST(block_rsv);
5962
5963         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5964                                    empty_size, hint, (u64)-1, &ins, 0);
5965         if (ret) {
5966                 unuse_block_rsv(block_rsv, blocksize);
5967                 return ERR_PTR(ret);
5968         }
5969
5970         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5971                                     blocksize, level);
5972         BUG_ON(IS_ERR(buf));
5973
5974         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5975                 if (parent == 0)
5976                         parent = ins.objectid;
5977                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5978         } else
5979                 BUG_ON(parent > 0);
5980
5981         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5982                 struct btrfs_delayed_extent_op *extent_op;
5983                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5984                 BUG_ON(!extent_op);
5985                 if (key)
5986                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5987                 else
5988                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5989                 extent_op->flags_to_set = flags;
5990                 extent_op->update_key = 1;
5991                 extent_op->update_flags = 1;
5992                 extent_op->is_data = 0;
5993
5994                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5995                                         ins.offset, parent, root_objectid,
5996                                         level, BTRFS_ADD_DELAYED_EXTENT,
5997                                         extent_op);
5998                 BUG_ON(ret);
5999         }
6000         return buf;
6001 }
6002
6003 struct walk_control {
6004         u64 refs[BTRFS_MAX_LEVEL];
6005         u64 flags[BTRFS_MAX_LEVEL];
6006         struct btrfs_key update_progress;
6007         int stage;
6008         int level;
6009         int shared_level;
6010         int update_ref;
6011         int keep_locks;
6012         int reada_slot;
6013         int reada_count;
6014 };
6015
6016 #define DROP_REFERENCE  1
6017 #define UPDATE_BACKREF  2
6018
6019 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6020                                      struct btrfs_root *root,
6021                                      struct walk_control *wc,
6022                                      struct btrfs_path *path)
6023 {
6024         u64 bytenr;
6025         u64 generation;
6026         u64 refs;
6027         u64 flags;
6028         u32 nritems;
6029         u32 blocksize;
6030         struct btrfs_key key;
6031         struct extent_buffer *eb;
6032         int ret;
6033         int slot;
6034         int nread = 0;
6035
6036         if (path->slots[wc->level] < wc->reada_slot) {
6037                 wc->reada_count = wc->reada_count * 2 / 3;
6038                 wc->reada_count = max(wc->reada_count, 2);
6039         } else {
6040                 wc->reada_count = wc->reada_count * 3 / 2;
6041                 wc->reada_count = min_t(int, wc->reada_count,
6042                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6043         }
6044
6045         eb = path->nodes[wc->level];
6046         nritems = btrfs_header_nritems(eb);
6047         blocksize = btrfs_level_size(root, wc->level - 1);
6048
6049         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6050                 if (nread >= wc->reada_count)
6051                         break;
6052
6053                 cond_resched();
6054                 bytenr = btrfs_node_blockptr(eb, slot);
6055                 generation = btrfs_node_ptr_generation(eb, slot);
6056
6057                 if (slot == path->slots[wc->level])
6058                         goto reada;
6059
6060                 if (wc->stage == UPDATE_BACKREF &&
6061                     generation <= root->root_key.offset)
6062                         continue;
6063
6064                 /* We don't lock the tree block, it's OK to be racy here */
6065                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6066                                                &refs, &flags);
6067                 BUG_ON(ret);
6068                 BUG_ON(refs == 0);
6069
6070                 if (wc->stage == DROP_REFERENCE) {
6071                         if (refs == 1)
6072                                 goto reada;
6073
6074                         if (wc->level == 1 &&
6075                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6076                                 continue;
6077                         if (!wc->update_ref ||
6078                             generation <= root->root_key.offset)
6079                                 continue;
6080                         btrfs_node_key_to_cpu(eb, &key, slot);
6081                         ret = btrfs_comp_cpu_keys(&key,
6082                                                   &wc->update_progress);
6083                         if (ret < 0)
6084                                 continue;
6085                 } else {
6086                         if (wc->level == 1 &&
6087                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6088                                 continue;
6089                 }
6090 reada:
6091                 ret = readahead_tree_block(root, bytenr, blocksize,
6092                                            generation);
6093                 if (ret)
6094                         break;
6095                 nread++;
6096         }
6097         wc->reada_slot = slot;
6098 }
6099
6100 /*
6101  * hepler to process tree block while walking down the tree.
6102  *
6103  * when wc->stage == UPDATE_BACKREF, this function updates
6104  * back refs for pointers in the block.
6105  *
6106  * NOTE: return value 1 means we should stop walking down.
6107  */
6108 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6109                                    struct btrfs_root *root,
6110                                    struct btrfs_path *path,
6111                                    struct walk_control *wc, int lookup_info)
6112 {
6113         int level = wc->level;
6114         struct extent_buffer *eb = path->nodes[level];
6115         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6116         int ret;
6117
6118         if (wc->stage == UPDATE_BACKREF &&
6119             btrfs_header_owner(eb) != root->root_key.objectid)
6120                 return 1;
6121
6122         /*
6123          * when reference count of tree block is 1, it won't increase
6124          * again. once full backref flag is set, we never clear it.
6125          */
6126         if (lookup_info &&
6127             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6128              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6129                 BUG_ON(!path->locks[level]);
6130                 ret = btrfs_lookup_extent_info(trans, root,
6131                                                eb->start, eb->len,
6132                                                &wc->refs[level],
6133                                                &wc->flags[level]);
6134                 BUG_ON(ret);
6135                 BUG_ON(wc->refs[level] == 0);
6136         }
6137
6138         if (wc->stage == DROP_REFERENCE) {
6139                 if (wc->refs[level] > 1)
6140                         return 1;
6141
6142                 if (path->locks[level] && !wc->keep_locks) {
6143                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6144                         path->locks[level] = 0;
6145                 }
6146                 return 0;
6147         }
6148
6149         /* wc->stage == UPDATE_BACKREF */
6150         if (!(wc->flags[level] & flag)) {
6151                 BUG_ON(!path->locks[level]);
6152                 ret = btrfs_inc_ref(trans, root, eb, 1);
6153                 BUG_ON(ret);
6154                 ret = btrfs_dec_ref(trans, root, eb, 0);
6155                 BUG_ON(ret);
6156                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6157                                                   eb->len, flag, 0);
6158                 BUG_ON(ret);
6159                 wc->flags[level] |= flag;
6160         }
6161
6162         /*
6163          * the block is shared by multiple trees, so it's not good to
6164          * keep the tree lock
6165          */
6166         if (path->locks[level] && level > 0) {
6167                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6168                 path->locks[level] = 0;
6169         }
6170         return 0;
6171 }
6172
6173 /*
6174  * hepler to process tree block pointer.
6175  *
6176  * when wc->stage == DROP_REFERENCE, this function checks
6177  * reference count of the block pointed to. if the block
6178  * is shared and we need update back refs for the subtree
6179  * rooted at the block, this function changes wc->stage to
6180  * UPDATE_BACKREF. if the block is shared and there is no
6181  * need to update back, this function drops the reference
6182  * to the block.
6183  *
6184  * NOTE: return value 1 means we should stop walking down.
6185  */
6186 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6187                                  struct btrfs_root *root,
6188                                  struct btrfs_path *path,
6189                                  struct walk_control *wc, int *lookup_info)
6190 {
6191         u64 bytenr;
6192         u64 generation;
6193         u64 parent;
6194         u32 blocksize;
6195         struct btrfs_key key;
6196         struct extent_buffer *next;
6197         int level = wc->level;
6198         int reada = 0;
6199         int ret = 0;
6200
6201         generation = btrfs_node_ptr_generation(path->nodes[level],
6202                                                path->slots[level]);
6203         /*
6204          * if the lower level block was created before the snapshot
6205          * was created, we know there is no need to update back refs
6206          * for the subtree
6207          */
6208         if (wc->stage == UPDATE_BACKREF &&
6209             generation <= root->root_key.offset) {
6210                 *lookup_info = 1;
6211                 return 1;
6212         }
6213
6214         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6215         blocksize = btrfs_level_size(root, level - 1);
6216
6217         next = btrfs_find_tree_block(root, bytenr, blocksize);
6218         if (!next) {
6219                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6220                 if (!next)
6221                         return -ENOMEM;
6222                 reada = 1;
6223         }
6224         btrfs_tree_lock(next);
6225         btrfs_set_lock_blocking(next);
6226
6227         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6228                                        &wc->refs[level - 1],
6229                                        &wc->flags[level - 1]);
6230         BUG_ON(ret);
6231         BUG_ON(wc->refs[level - 1] == 0);
6232         *lookup_info = 0;
6233
6234         if (wc->stage == DROP_REFERENCE) {
6235                 if (wc->refs[level - 1] > 1) {
6236                         if (level == 1 &&
6237                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6238                                 goto skip;
6239
6240                         if (!wc->update_ref ||
6241                             generation <= root->root_key.offset)
6242                                 goto skip;
6243
6244                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6245                                               path->slots[level]);
6246                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6247                         if (ret < 0)
6248                                 goto skip;
6249
6250                         wc->stage = UPDATE_BACKREF;
6251                         wc->shared_level = level - 1;
6252                 }
6253         } else {
6254                 if (level == 1 &&
6255                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6256                         goto skip;
6257         }
6258
6259         if (!btrfs_buffer_uptodate(next, generation)) {
6260                 btrfs_tree_unlock(next);
6261                 free_extent_buffer(next);
6262                 next = NULL;
6263                 *lookup_info = 1;
6264         }
6265
6266         if (!next) {
6267                 if (reada && level == 1)
6268                         reada_walk_down(trans, root, wc, path);
6269                 next = read_tree_block(root, bytenr, blocksize, generation);
6270                 if (!next)
6271                         return -EIO;
6272                 btrfs_tree_lock(next);
6273                 btrfs_set_lock_blocking(next);
6274         }
6275
6276         level--;
6277         BUG_ON(level != btrfs_header_level(next));
6278         path->nodes[level] = next;
6279         path->slots[level] = 0;
6280         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6281         wc->level = level;
6282         if (wc->level == 1)
6283                 wc->reada_slot = 0;
6284         return 0;
6285 skip:
6286         wc->refs[level - 1] = 0;
6287         wc->flags[level - 1] = 0;
6288         if (wc->stage == DROP_REFERENCE) {
6289                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6290                         parent = path->nodes[level]->start;
6291                 } else {
6292                         BUG_ON(root->root_key.objectid !=
6293                                btrfs_header_owner(path->nodes[level]));
6294                         parent = 0;
6295                 }
6296
6297                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6298                                         root->root_key.objectid, level - 1, 0);
6299                 BUG_ON(ret);
6300         }
6301         btrfs_tree_unlock(next);
6302         free_extent_buffer(next);
6303         *lookup_info = 1;
6304         return 1;
6305 }
6306
6307 /*
6308  * hepler to process tree block while walking up the tree.
6309  *
6310  * when wc->stage == DROP_REFERENCE, this function drops
6311  * reference count on the block.
6312  *
6313  * when wc->stage == UPDATE_BACKREF, this function changes
6314  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6315  * to UPDATE_BACKREF previously while processing the block.
6316  *
6317  * NOTE: return value 1 means we should stop walking up.
6318  */
6319 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6320                                  struct btrfs_root *root,
6321                                  struct btrfs_path *path,
6322                                  struct walk_control *wc)
6323 {
6324         int ret;
6325         int level = wc->level;
6326         struct extent_buffer *eb = path->nodes[level];
6327         u64 parent = 0;
6328
6329         if (wc->stage == UPDATE_BACKREF) {
6330                 BUG_ON(wc->shared_level < level);
6331                 if (level < wc->shared_level)
6332                         goto out;
6333
6334                 ret = find_next_key(path, level + 1, &wc->update_progress);
6335                 if (ret > 0)
6336                         wc->update_ref = 0;
6337
6338                 wc->stage = DROP_REFERENCE;
6339                 wc->shared_level = -1;
6340                 path->slots[level] = 0;
6341
6342                 /*
6343                  * check reference count again if the block isn't locked.
6344                  * we should start walking down the tree again if reference
6345                  * count is one.
6346                  */
6347                 if (!path->locks[level]) {
6348                         BUG_ON(level == 0);
6349                         btrfs_tree_lock(eb);
6350                         btrfs_set_lock_blocking(eb);
6351                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6352
6353                         ret = btrfs_lookup_extent_info(trans, root,
6354                                                        eb->start, eb->len,
6355                                                        &wc->refs[level],
6356                                                        &wc->flags[level]);
6357                         BUG_ON(ret);
6358                         BUG_ON(wc->refs[level] == 0);
6359                         if (wc->refs[level] == 1) {
6360                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6361                                 return 1;
6362                         }
6363                 }
6364         }
6365
6366         /* wc->stage == DROP_REFERENCE */
6367         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6368
6369         if (wc->refs[level] == 1) {
6370                 if (level == 0) {
6371                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6372                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6373                         else
6374                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6375                         BUG_ON(ret);
6376                 }
6377                 /* make block locked assertion in clean_tree_block happy */
6378                 if (!path->locks[level] &&
6379                     btrfs_header_generation(eb) == trans->transid) {
6380                         btrfs_tree_lock(eb);
6381                         btrfs_set_lock_blocking(eb);
6382                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6383                 }
6384                 clean_tree_block(trans, root, eb);
6385         }
6386
6387         if (eb == root->node) {
6388                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6389                         parent = eb->start;
6390                 else
6391                         BUG_ON(root->root_key.objectid !=
6392                                btrfs_header_owner(eb));
6393         } else {
6394                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6395                         parent = path->nodes[level + 1]->start;
6396                 else
6397                         BUG_ON(root->root_key.objectid !=
6398                                btrfs_header_owner(path->nodes[level + 1]));
6399         }
6400
6401         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6402 out:
6403         wc->refs[level] = 0;
6404         wc->flags[level] = 0;
6405         return 0;
6406 }
6407
6408 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6409                                    struct btrfs_root *root,
6410                                    struct btrfs_path *path,
6411                                    struct walk_control *wc)
6412 {
6413         int level = wc->level;
6414         int lookup_info = 1;
6415         int ret;
6416
6417         while (level >= 0) {
6418                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6419                 if (ret > 0)
6420                         break;
6421
6422                 if (level == 0)
6423                         break;
6424
6425                 if (path->slots[level] >=
6426                     btrfs_header_nritems(path->nodes[level]))
6427                         break;
6428
6429                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6430                 if (ret > 0) {
6431                         path->slots[level]++;
6432                         continue;
6433                 } else if (ret < 0)
6434                         return ret;
6435                 level = wc->level;
6436         }
6437         return 0;
6438 }
6439
6440 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6441                                  struct btrfs_root *root,
6442                                  struct btrfs_path *path,
6443                                  struct walk_control *wc, int max_level)
6444 {
6445         int level = wc->level;
6446         int ret;
6447
6448         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6449         while (level < max_level && path->nodes[level]) {
6450                 wc->level = level;
6451                 if (path->slots[level] + 1 <
6452                     btrfs_header_nritems(path->nodes[level])) {
6453                         path->slots[level]++;
6454                         return 0;
6455                 } else {
6456                         ret = walk_up_proc(trans, root, path, wc);
6457                         if (ret > 0)
6458                                 return 0;
6459
6460                         if (path->locks[level]) {
6461                                 btrfs_tree_unlock_rw(path->nodes[level],
6462                                                      path->locks[level]);
6463                                 path->locks[level] = 0;
6464                         }
6465                         free_extent_buffer(path->nodes[level]);
6466                         path->nodes[level] = NULL;
6467                         level++;
6468                 }
6469         }
6470         return 1;
6471 }
6472
6473 /*
6474  * drop a subvolume tree.
6475  *
6476  * this function traverses the tree freeing any blocks that only
6477  * referenced by the tree.
6478  *
6479  * when a shared tree block is found. this function decreases its
6480  * reference count by one. if update_ref is true, this function
6481  * also make sure backrefs for the shared block and all lower level
6482  * blocks are properly updated.
6483  */
6484 void btrfs_drop_snapshot(struct btrfs_root *root,
6485                          struct btrfs_block_rsv *block_rsv, int update_ref)
6486 {
6487         struct btrfs_path *path;
6488         struct btrfs_trans_handle *trans;
6489         struct btrfs_root *tree_root = root->fs_info->tree_root;
6490         struct btrfs_root_item *root_item = &root->root_item;
6491         struct walk_control *wc;
6492         struct btrfs_key key;
6493         int err = 0;
6494         int ret;
6495         int level;
6496
6497         path = btrfs_alloc_path();
6498         if (!path) {
6499                 err = -ENOMEM;
6500                 goto out;
6501         }
6502
6503         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6504         if (!wc) {
6505                 btrfs_free_path(path);
6506                 err = -ENOMEM;
6507                 goto out;
6508         }
6509
6510         trans = btrfs_start_transaction(tree_root, 0);
6511         BUG_ON(IS_ERR(trans));
6512
6513         if (block_rsv)
6514                 trans->block_rsv = block_rsv;
6515
6516         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6517                 level = btrfs_header_level(root->node);
6518                 path->nodes[level] = btrfs_lock_root_node(root);
6519                 btrfs_set_lock_blocking(path->nodes[level]);
6520                 path->slots[level] = 0;
6521                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6522                 memset(&wc->update_progress, 0,
6523                        sizeof(wc->update_progress));
6524         } else {
6525                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6526                 memcpy(&wc->update_progress, &key,
6527                        sizeof(wc->update_progress));
6528
6529                 level = root_item->drop_level;
6530                 BUG_ON(level == 0);
6531                 path->lowest_level = level;
6532                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6533                 path->lowest_level = 0;
6534                 if (ret < 0) {
6535                         err = ret;
6536                         goto out_free;
6537                 }
6538                 WARN_ON(ret > 0);
6539
6540                 /*
6541                  * unlock our path, this is safe because only this
6542                  * function is allowed to delete this snapshot
6543                  */
6544                 btrfs_unlock_up_safe(path, 0);
6545
6546                 level = btrfs_header_level(root->node);
6547                 while (1) {
6548                         btrfs_tree_lock(path->nodes[level]);
6549                         btrfs_set_lock_blocking(path->nodes[level]);
6550
6551                         ret = btrfs_lookup_extent_info(trans, root,
6552                                                 path->nodes[level]->start,
6553                                                 path->nodes[level]->len,
6554                                                 &wc->refs[level],
6555                                                 &wc->flags[level]);
6556                         BUG_ON(ret);
6557                         BUG_ON(wc->refs[level] == 0);
6558
6559                         if (level == root_item->drop_level)
6560                                 break;
6561
6562                         btrfs_tree_unlock(path->nodes[level]);
6563                         WARN_ON(wc->refs[level] != 1);
6564                         level--;
6565                 }
6566         }
6567
6568         wc->level = level;
6569         wc->shared_level = -1;
6570         wc->stage = DROP_REFERENCE;
6571         wc->update_ref = update_ref;
6572         wc->keep_locks = 0;
6573         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6574
6575         while (1) {
6576                 ret = walk_down_tree(trans, root, path, wc);
6577                 if (ret < 0) {
6578                         err = ret;
6579                         break;
6580                 }
6581
6582                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6583                 if (ret < 0) {
6584                         err = ret;
6585                         break;
6586                 }
6587
6588                 if (ret > 0) {
6589                         BUG_ON(wc->stage != DROP_REFERENCE);
6590                         break;
6591                 }
6592
6593                 if (wc->stage == DROP_REFERENCE) {
6594                         level = wc->level;
6595                         btrfs_node_key(path->nodes[level],
6596                                        &root_item->drop_progress,
6597                                        path->slots[level]);
6598                         root_item->drop_level = level;
6599                 }
6600
6601                 BUG_ON(wc->level == 0);
6602                 if (btrfs_should_end_transaction(trans, tree_root)) {
6603                         ret = btrfs_update_root(trans, tree_root,
6604                                                 &root->root_key,
6605                                                 root_item);
6606                         BUG_ON(ret);
6607
6608                         btrfs_end_transaction_throttle(trans, tree_root);
6609                         trans = btrfs_start_transaction(tree_root, 0);
6610                         BUG_ON(IS_ERR(trans));
6611                         if (block_rsv)
6612                                 trans->block_rsv = block_rsv;
6613                 }
6614         }
6615         btrfs_release_path(path);
6616         BUG_ON(err);
6617
6618         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6619         BUG_ON(ret);
6620
6621         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6622                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6623                                            NULL, NULL);
6624                 BUG_ON(ret < 0);
6625                 if (ret > 0) {
6626                         /* if we fail to delete the orphan item this time
6627                          * around, it'll get picked up the next time.
6628                          *
6629                          * The most common failure here is just -ENOENT.
6630                          */
6631                         btrfs_del_orphan_item(trans, tree_root,
6632                                               root->root_key.objectid);
6633                 }
6634         }
6635
6636         if (root->in_radix) {
6637                 btrfs_free_fs_root(tree_root->fs_info, root);
6638         } else {
6639                 free_extent_buffer(root->node);
6640                 free_extent_buffer(root->commit_root);
6641                 kfree(root);
6642         }
6643 out_free:
6644         btrfs_end_transaction_throttle(trans, tree_root);
6645         kfree(wc);
6646         btrfs_free_path(path);
6647 out:
6648         if (err)
6649                 btrfs_std_error(root->fs_info, err);
6650         return;
6651 }
6652
6653 /*
6654  * drop subtree rooted at tree block 'node'.
6655  *
6656  * NOTE: this function will unlock and release tree block 'node'
6657  */
6658 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6659                         struct btrfs_root *root,
6660                         struct extent_buffer *node,
6661                         struct extent_buffer *parent)
6662 {
6663         struct btrfs_path *path;
6664         struct walk_control *wc;
6665         int level;
6666         int parent_level;
6667         int ret = 0;
6668         int wret;
6669
6670         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6671
6672         path = btrfs_alloc_path();
6673         if (!path)
6674                 return -ENOMEM;
6675
6676         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6677         if (!wc) {
6678                 btrfs_free_path(path);
6679                 return -ENOMEM;
6680         }
6681
6682         btrfs_assert_tree_locked(parent);
6683         parent_level = btrfs_header_level(parent);
6684         extent_buffer_get(parent);
6685         path->nodes[parent_level] = parent;
6686         path->slots[parent_level] = btrfs_header_nritems(parent);
6687
6688         btrfs_assert_tree_locked(node);
6689         level = btrfs_header_level(node);
6690         path->nodes[level] = node;
6691         path->slots[level] = 0;
6692         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6693
6694         wc->refs[parent_level] = 1;
6695         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6696         wc->level = level;
6697         wc->shared_level = -1;
6698         wc->stage = DROP_REFERENCE;
6699         wc->update_ref = 0;
6700         wc->keep_locks = 1;
6701         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6702
6703         while (1) {
6704                 wret = walk_down_tree(trans, root, path, wc);
6705                 if (wret < 0) {
6706                         ret = wret;
6707                         break;
6708                 }
6709
6710                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6711                 if (wret < 0)
6712                         ret = wret;
6713                 if (wret != 0)
6714                         break;
6715         }
6716
6717         kfree(wc);
6718         btrfs_free_path(path);
6719         return ret;
6720 }
6721
6722 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6723 {
6724         u64 num_devices;
6725         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6726                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6727
6728         /*
6729          * we add in the count of missing devices because we want
6730          * to make sure that any RAID levels on a degraded FS
6731          * continue to be honored.
6732          */
6733         num_devices = root->fs_info->fs_devices->rw_devices +
6734                 root->fs_info->fs_devices->missing_devices;
6735
6736         if (num_devices == 1) {
6737                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6738                 stripped = flags & ~stripped;
6739
6740                 /* turn raid0 into single device chunks */
6741                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6742                         return stripped;
6743
6744                 /* turn mirroring into duplication */
6745                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6746                              BTRFS_BLOCK_GROUP_RAID10))
6747                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6748                 return flags;
6749         } else {
6750                 /* they already had raid on here, just return */
6751                 if (flags & stripped)
6752                         return flags;
6753
6754                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6755                 stripped = flags & ~stripped;
6756
6757                 /* switch duplicated blocks with raid1 */
6758                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6759                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6760
6761                 /* turn single device chunks into raid0 */
6762                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6763         }
6764         return flags;
6765 }
6766
6767 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6768 {
6769         struct btrfs_space_info *sinfo = cache->space_info;
6770         u64 num_bytes;
6771         u64 min_allocable_bytes;
6772         int ret = -ENOSPC;
6773
6774
6775         /*
6776          * We need some metadata space and system metadata space for
6777          * allocating chunks in some corner cases until we force to set
6778          * it to be readonly.
6779          */
6780         if ((sinfo->flags &
6781              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6782             !force)
6783                 min_allocable_bytes = 1 * 1024 * 1024;
6784         else
6785                 min_allocable_bytes = 0;
6786
6787         spin_lock(&sinfo->lock);
6788         spin_lock(&cache->lock);
6789
6790         if (cache->ro) {
6791                 ret = 0;
6792                 goto out;
6793         }
6794
6795         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6796                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6797
6798         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6799             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6800             min_allocable_bytes <= sinfo->total_bytes) {
6801                 sinfo->bytes_readonly += num_bytes;
6802                 cache->ro = 1;
6803                 ret = 0;
6804         }
6805 out:
6806         spin_unlock(&cache->lock);
6807         spin_unlock(&sinfo->lock);
6808         return ret;
6809 }
6810
6811 int btrfs_set_block_group_ro(struct btrfs_root *root,
6812                              struct btrfs_block_group_cache *cache)
6813
6814 {
6815         struct btrfs_trans_handle *trans;
6816         u64 alloc_flags;
6817         int ret;
6818
6819         BUG_ON(cache->ro);
6820
6821         trans = btrfs_join_transaction(root);
6822         BUG_ON(IS_ERR(trans));
6823
6824         alloc_flags = update_block_group_flags(root, cache->flags);
6825         if (alloc_flags != cache->flags)
6826                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6827                                CHUNK_ALLOC_FORCE);
6828
6829         ret = set_block_group_ro(cache, 0);
6830         if (!ret)
6831                 goto out;
6832         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6833         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6834                              CHUNK_ALLOC_FORCE);
6835         if (ret < 0)
6836                 goto out;
6837         ret = set_block_group_ro(cache, 0);
6838 out:
6839         btrfs_end_transaction(trans, root);
6840         return ret;
6841 }
6842
6843 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6844                             struct btrfs_root *root, u64 type)
6845 {
6846         u64 alloc_flags = get_alloc_profile(root, type);
6847         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6848                               CHUNK_ALLOC_FORCE);
6849 }
6850
6851 /*
6852  * helper to account the unused space of all the readonly block group in the
6853  * list. takes mirrors into account.
6854  */
6855 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6856 {
6857         struct btrfs_block_group_cache *block_group;
6858         u64 free_bytes = 0;
6859         int factor;
6860
6861         list_for_each_entry(block_group, groups_list, list) {
6862                 spin_lock(&block_group->lock);
6863
6864                 if (!block_group->ro) {
6865                         spin_unlock(&block_group->lock);
6866                         continue;
6867                 }
6868
6869                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6870                                           BTRFS_BLOCK_GROUP_RAID10 |
6871                                           BTRFS_BLOCK_GROUP_DUP))
6872                         factor = 2;
6873                 else
6874                         factor = 1;
6875
6876                 free_bytes += (block_group->key.offset -
6877                                btrfs_block_group_used(&block_group->item)) *
6878                                factor;
6879
6880                 spin_unlock(&block_group->lock);
6881         }
6882
6883         return free_bytes;
6884 }
6885
6886 /*
6887  * helper to account the unused space of all the readonly block group in the
6888  * space_info. takes mirrors into account.
6889  */
6890 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6891 {
6892         int i;
6893         u64 free_bytes = 0;
6894
6895         spin_lock(&sinfo->lock);
6896
6897         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6898                 if (!list_empty(&sinfo->block_groups[i]))
6899                         free_bytes += __btrfs_get_ro_block_group_free_space(
6900                                                 &sinfo->block_groups[i]);
6901
6902         spin_unlock(&sinfo->lock);
6903
6904         return free_bytes;
6905 }
6906
6907 int btrfs_set_block_group_rw(struct btrfs_root *root,
6908                               struct btrfs_block_group_cache *cache)
6909 {
6910         struct btrfs_space_info *sinfo = cache->space_info;
6911         u64 num_bytes;
6912
6913         BUG_ON(!cache->ro);
6914
6915         spin_lock(&sinfo->lock);
6916         spin_lock(&cache->lock);
6917         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6918                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6919         sinfo->bytes_readonly -= num_bytes;
6920         cache->ro = 0;
6921         spin_unlock(&cache->lock);
6922         spin_unlock(&sinfo->lock);
6923         return 0;
6924 }
6925
6926 /*
6927  * checks to see if its even possible to relocate this block group.
6928  *
6929  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6930  * ok to go ahead and try.
6931  */
6932 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6933 {
6934         struct btrfs_block_group_cache *block_group;
6935         struct btrfs_space_info *space_info;
6936         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6937         struct btrfs_device *device;
6938         u64 min_free;
6939         u64 dev_min = 1;
6940         u64 dev_nr = 0;
6941         int index;
6942         int full = 0;
6943         int ret = 0;
6944
6945         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6946
6947         /* odd, couldn't find the block group, leave it alone */
6948         if (!block_group)
6949                 return -1;
6950
6951         min_free = btrfs_block_group_used(&block_group->item);
6952
6953         /* no bytes used, we're good */
6954         if (!min_free)
6955                 goto out;
6956
6957         space_info = block_group->space_info;
6958         spin_lock(&space_info->lock);
6959
6960         full = space_info->full;
6961
6962         /*
6963          * if this is the last block group we have in this space, we can't
6964          * relocate it unless we're able to allocate a new chunk below.
6965          *
6966          * Otherwise, we need to make sure we have room in the space to handle
6967          * all of the extents from this block group.  If we can, we're good
6968          */
6969         if ((space_info->total_bytes != block_group->key.offset) &&
6970             (space_info->bytes_used + space_info->bytes_reserved +
6971              space_info->bytes_pinned + space_info->bytes_readonly +
6972              min_free < space_info->total_bytes)) {
6973                 spin_unlock(&space_info->lock);
6974                 goto out;
6975         }
6976         spin_unlock(&space_info->lock);
6977
6978         /*
6979          * ok we don't have enough space, but maybe we have free space on our
6980          * devices to allocate new chunks for relocation, so loop through our
6981          * alloc devices and guess if we have enough space.  However, if we
6982          * were marked as full, then we know there aren't enough chunks, and we
6983          * can just return.
6984          */
6985         ret = -1;
6986         if (full)
6987                 goto out;
6988
6989         /*
6990          * index:
6991          *      0: raid10
6992          *      1: raid1
6993          *      2: dup
6994          *      3: raid0
6995          *      4: single
6996          */
6997         index = get_block_group_index(block_group);
6998         if (index == 0) {
6999                 dev_min = 4;
7000                 /* Divide by 2 */
7001                 min_free >>= 1;
7002         } else if (index == 1) {
7003                 dev_min = 2;
7004         } else if (index == 2) {
7005                 /* Multiply by 2 */
7006                 min_free <<= 1;
7007         } else if (index == 3) {
7008                 dev_min = fs_devices->rw_devices;
7009                 do_div(min_free, dev_min);
7010         }
7011
7012         mutex_lock(&root->fs_info->chunk_mutex);
7013         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7014                 u64 dev_offset;
7015
7016                 /*
7017                  * check to make sure we can actually find a chunk with enough
7018                  * space to fit our block group in.
7019                  */
7020                 if (device->total_bytes > device->bytes_used + min_free) {
7021                         ret = find_free_dev_extent(NULL, device, min_free,
7022                                                    &dev_offset, NULL);
7023                         if (!ret)
7024                                 dev_nr++;
7025
7026                         if (dev_nr >= dev_min)
7027                                 break;
7028
7029                         ret = -1;
7030                 }
7031         }
7032         mutex_unlock(&root->fs_info->chunk_mutex);
7033 out:
7034         btrfs_put_block_group(block_group);
7035         return ret;
7036 }
7037
7038 static int find_first_block_group(struct btrfs_root *root,
7039                 struct btrfs_path *path, struct btrfs_key *key)
7040 {
7041         int ret = 0;
7042         struct btrfs_key found_key;
7043         struct extent_buffer *leaf;
7044         int slot;
7045
7046         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7047         if (ret < 0)
7048                 goto out;
7049
7050         while (1) {
7051                 slot = path->slots[0];
7052                 leaf = path->nodes[0];
7053                 if (slot >= btrfs_header_nritems(leaf)) {
7054                         ret = btrfs_next_leaf(root, path);
7055                         if (ret == 0)
7056                                 continue;
7057                         if (ret < 0)
7058                                 goto out;
7059                         break;
7060                 }
7061                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7062
7063                 if (found_key.objectid >= key->objectid &&
7064                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7065                         ret = 0;
7066                         goto out;
7067                 }
7068                 path->slots[0]++;
7069         }
7070 out:
7071         return ret;
7072 }
7073
7074 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7075 {
7076         struct btrfs_block_group_cache *block_group;
7077         u64 last = 0;
7078
7079         while (1) {
7080                 struct inode *inode;
7081
7082                 block_group = btrfs_lookup_first_block_group(info, last);
7083                 while (block_group) {
7084                         spin_lock(&block_group->lock);
7085                         if (block_group->iref)
7086                                 break;
7087                         spin_unlock(&block_group->lock);
7088                         block_group = next_block_group(info->tree_root,
7089                                                        block_group);
7090                 }
7091                 if (!block_group) {
7092                         if (last == 0)
7093                                 break;
7094                         last = 0;
7095                         continue;
7096                 }
7097
7098                 inode = block_group->inode;
7099                 block_group->iref = 0;
7100                 block_group->inode = NULL;
7101                 spin_unlock(&block_group->lock);
7102                 iput(inode);
7103                 last = block_group->key.objectid + block_group->key.offset;
7104                 btrfs_put_block_group(block_group);
7105         }
7106 }
7107
7108 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7109 {
7110         struct btrfs_block_group_cache *block_group;
7111         struct btrfs_space_info *space_info;
7112         struct btrfs_caching_control *caching_ctl;
7113         struct rb_node *n;
7114
7115         down_write(&info->extent_commit_sem);
7116         while (!list_empty(&info->caching_block_groups)) {
7117                 caching_ctl = list_entry(info->caching_block_groups.next,
7118                                          struct btrfs_caching_control, list);
7119                 list_del(&caching_ctl->list);
7120                 put_caching_control(caching_ctl);
7121         }
7122         up_write(&info->extent_commit_sem);
7123
7124         spin_lock(&info->block_group_cache_lock);
7125         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7126                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7127                                        cache_node);
7128                 rb_erase(&block_group->cache_node,
7129                          &info->block_group_cache_tree);
7130                 spin_unlock(&info->block_group_cache_lock);
7131
7132                 down_write(&block_group->space_info->groups_sem);
7133                 list_del(&block_group->list);
7134                 up_write(&block_group->space_info->groups_sem);
7135
7136                 if (block_group->cached == BTRFS_CACHE_STARTED)
7137                         wait_block_group_cache_done(block_group);
7138
7139                 /*
7140                  * We haven't cached this block group, which means we could
7141                  * possibly have excluded extents on this block group.
7142                  */
7143                 if (block_group->cached == BTRFS_CACHE_NO)
7144                         free_excluded_extents(info->extent_root, block_group);
7145
7146                 btrfs_remove_free_space_cache(block_group);
7147                 btrfs_put_block_group(block_group);
7148
7149                 spin_lock(&info->block_group_cache_lock);
7150         }
7151         spin_unlock(&info->block_group_cache_lock);
7152
7153         /* now that all the block groups are freed, go through and
7154          * free all the space_info structs.  This is only called during
7155          * the final stages of unmount, and so we know nobody is
7156          * using them.  We call synchronize_rcu() once before we start,
7157          * just to be on the safe side.
7158          */
7159         synchronize_rcu();
7160
7161         release_global_block_rsv(info);
7162
7163         while(!list_empty(&info->space_info)) {
7164                 space_info = list_entry(info->space_info.next,
7165                                         struct btrfs_space_info,
7166                                         list);
7167                 if (space_info->bytes_pinned > 0 ||
7168                     space_info->bytes_reserved > 0 ||
7169                     space_info->bytes_may_use > 0) {
7170                         WARN_ON(1);
7171                         dump_space_info(space_info, 0, 0);
7172                 }
7173                 list_del(&space_info->list);
7174                 kfree(space_info);
7175         }
7176         return 0;
7177 }
7178
7179 static void __link_block_group(struct btrfs_space_info *space_info,
7180                                struct btrfs_block_group_cache *cache)
7181 {
7182         int index = get_block_group_index(cache);
7183
7184         down_write(&space_info->groups_sem);
7185         list_add_tail(&cache->list, &space_info->block_groups[index]);
7186         up_write(&space_info->groups_sem);
7187 }
7188
7189 int btrfs_read_block_groups(struct btrfs_root *root)
7190 {
7191         struct btrfs_path *path;
7192         int ret;
7193         struct btrfs_block_group_cache *cache;
7194         struct btrfs_fs_info *info = root->fs_info;
7195         struct btrfs_space_info *space_info;
7196         struct btrfs_key key;
7197         struct btrfs_key found_key;
7198         struct extent_buffer *leaf;
7199         int need_clear = 0;
7200         u64 cache_gen;
7201
7202         root = info->extent_root;
7203         key.objectid = 0;
7204         key.offset = 0;
7205         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7206         path = btrfs_alloc_path();
7207         if (!path)
7208                 return -ENOMEM;
7209         path->reada = 1;
7210
7211         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7212         if (btrfs_test_opt(root, SPACE_CACHE) &&
7213             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7214                 need_clear = 1;
7215         if (btrfs_test_opt(root, CLEAR_CACHE))
7216                 need_clear = 1;
7217
7218         while (1) {
7219                 ret = find_first_block_group(root, path, &key);
7220                 if (ret > 0)
7221                         break;
7222                 if (ret != 0)
7223                         goto error;
7224                 leaf = path->nodes[0];
7225                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7226                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7227                 if (!cache) {
7228                         ret = -ENOMEM;
7229                         goto error;
7230                 }
7231                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7232                                                 GFP_NOFS);
7233                 if (!cache->free_space_ctl) {
7234                         kfree(cache);
7235                         ret = -ENOMEM;
7236                         goto error;
7237                 }
7238
7239                 atomic_set(&cache->count, 1);
7240                 spin_lock_init(&cache->lock);
7241                 cache->fs_info = info;
7242                 INIT_LIST_HEAD(&cache->list);
7243                 INIT_LIST_HEAD(&cache->cluster_list);
7244
7245                 if (need_clear)
7246                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7247
7248                 read_extent_buffer(leaf, &cache->item,
7249                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7250                                    sizeof(cache->item));
7251                 memcpy(&cache->key, &found_key, sizeof(found_key));
7252
7253                 key.objectid = found_key.objectid + found_key.offset;
7254                 btrfs_release_path(path);
7255                 cache->flags = btrfs_block_group_flags(&cache->item);
7256                 cache->sectorsize = root->sectorsize;
7257
7258                 btrfs_init_free_space_ctl(cache);
7259
7260                 /*
7261                  * We need to exclude the super stripes now so that the space
7262                  * info has super bytes accounted for, otherwise we'll think
7263                  * we have more space than we actually do.
7264                  */
7265                 exclude_super_stripes(root, cache);
7266
7267                 /*
7268                  * check for two cases, either we are full, and therefore
7269                  * don't need to bother with the caching work since we won't
7270                  * find any space, or we are empty, and we can just add all
7271                  * the space in and be done with it.  This saves us _alot_ of
7272                  * time, particularly in the full case.
7273                  */
7274                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7275                         cache->last_byte_to_unpin = (u64)-1;
7276                         cache->cached = BTRFS_CACHE_FINISHED;
7277                         free_excluded_extents(root, cache);
7278                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7279                         cache->last_byte_to_unpin = (u64)-1;
7280                         cache->cached = BTRFS_CACHE_FINISHED;
7281                         add_new_free_space(cache, root->fs_info,
7282                                            found_key.objectid,
7283                                            found_key.objectid +
7284                                            found_key.offset);
7285                         free_excluded_extents(root, cache);
7286                 }
7287
7288                 ret = update_space_info(info, cache->flags, found_key.offset,
7289                                         btrfs_block_group_used(&cache->item),
7290                                         &space_info);
7291                 BUG_ON(ret);
7292                 cache->space_info = space_info;
7293                 spin_lock(&cache->space_info->lock);
7294                 cache->space_info->bytes_readonly += cache->bytes_super;
7295                 spin_unlock(&cache->space_info->lock);
7296
7297                 __link_block_group(space_info, cache);
7298
7299                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7300                 BUG_ON(ret);
7301
7302                 set_avail_alloc_bits(root->fs_info, cache->flags);
7303                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7304                         set_block_group_ro(cache, 1);
7305         }
7306
7307         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7308                 if (!(get_alloc_profile(root, space_info->flags) &
7309                       (BTRFS_BLOCK_GROUP_RAID10 |
7310                        BTRFS_BLOCK_GROUP_RAID1 |
7311                        BTRFS_BLOCK_GROUP_DUP)))
7312                         continue;
7313                 /*
7314                  * avoid allocating from un-mirrored block group if there are
7315                  * mirrored block groups.
7316                  */
7317                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7318                         set_block_group_ro(cache, 1);
7319                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7320                         set_block_group_ro(cache, 1);
7321         }
7322
7323         init_global_block_rsv(info);
7324         ret = 0;
7325 error:
7326         btrfs_free_path(path);
7327         return ret;
7328 }
7329
7330 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7331                            struct btrfs_root *root, u64 bytes_used,
7332                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7333                            u64 size)
7334 {
7335         int ret;
7336         struct btrfs_root *extent_root;
7337         struct btrfs_block_group_cache *cache;
7338
7339         extent_root = root->fs_info->extent_root;
7340
7341         root->fs_info->last_trans_log_full_commit = trans->transid;
7342
7343         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7344         if (!cache)
7345                 return -ENOMEM;
7346         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7347                                         GFP_NOFS);
7348         if (!cache->free_space_ctl) {
7349                 kfree(cache);
7350                 return -ENOMEM;
7351         }
7352
7353         cache->key.objectid = chunk_offset;
7354         cache->key.offset = size;
7355         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7356         cache->sectorsize = root->sectorsize;
7357         cache->fs_info = root->fs_info;
7358
7359         atomic_set(&cache->count, 1);
7360         spin_lock_init(&cache->lock);
7361         INIT_LIST_HEAD(&cache->list);
7362         INIT_LIST_HEAD(&cache->cluster_list);
7363
7364         btrfs_init_free_space_ctl(cache);
7365
7366         btrfs_set_block_group_used(&cache->item, bytes_used);
7367         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7368         cache->flags = type;
7369         btrfs_set_block_group_flags(&cache->item, type);
7370
7371         cache->last_byte_to_unpin = (u64)-1;
7372         cache->cached = BTRFS_CACHE_FINISHED;
7373         exclude_super_stripes(root, cache);
7374
7375         add_new_free_space(cache, root->fs_info, chunk_offset,
7376                            chunk_offset + size);
7377
7378         free_excluded_extents(root, cache);
7379
7380         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7381                                 &cache->space_info);
7382         BUG_ON(ret);
7383
7384         spin_lock(&cache->space_info->lock);
7385         cache->space_info->bytes_readonly += cache->bytes_super;
7386         spin_unlock(&cache->space_info->lock);
7387
7388         __link_block_group(cache->space_info, cache);
7389
7390         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7391         BUG_ON(ret);
7392
7393         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7394                                 sizeof(cache->item));
7395         BUG_ON(ret);
7396
7397         set_avail_alloc_bits(extent_root->fs_info, type);
7398
7399         return 0;
7400 }
7401
7402 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7403                              struct btrfs_root *root, u64 group_start)
7404 {
7405         struct btrfs_path *path;
7406         struct btrfs_block_group_cache *block_group;
7407         struct btrfs_free_cluster *cluster;
7408         struct btrfs_root *tree_root = root->fs_info->tree_root;
7409         struct btrfs_key key;
7410         struct inode *inode;
7411         int ret;
7412         int factor;
7413
7414         root = root->fs_info->extent_root;
7415
7416         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7417         BUG_ON(!block_group);
7418         BUG_ON(!block_group->ro);
7419
7420         /*
7421          * Free the reserved super bytes from this block group before
7422          * remove it.
7423          */
7424         free_excluded_extents(root, block_group);
7425
7426         memcpy(&key, &block_group->key, sizeof(key));
7427         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7428                                   BTRFS_BLOCK_GROUP_RAID1 |
7429                                   BTRFS_BLOCK_GROUP_RAID10))
7430                 factor = 2;
7431         else
7432                 factor = 1;
7433
7434         /* make sure this block group isn't part of an allocation cluster */
7435         cluster = &root->fs_info->data_alloc_cluster;
7436         spin_lock(&cluster->refill_lock);
7437         btrfs_return_cluster_to_free_space(block_group, cluster);
7438         spin_unlock(&cluster->refill_lock);
7439
7440         /*
7441          * make sure this block group isn't part of a metadata
7442          * allocation cluster
7443          */
7444         cluster = &root->fs_info->meta_alloc_cluster;
7445         spin_lock(&cluster->refill_lock);
7446         btrfs_return_cluster_to_free_space(block_group, cluster);
7447         spin_unlock(&cluster->refill_lock);
7448
7449         path = btrfs_alloc_path();
7450         if (!path) {
7451                 ret = -ENOMEM;
7452                 goto out;
7453         }
7454
7455         inode = lookup_free_space_inode(tree_root, block_group, path);
7456         if (!IS_ERR(inode)) {
7457                 ret = btrfs_orphan_add(trans, inode);
7458                 BUG_ON(ret);
7459                 clear_nlink(inode);
7460                 /* One for the block groups ref */
7461                 spin_lock(&block_group->lock);
7462                 if (block_group->iref) {
7463                         block_group->iref = 0;
7464                         block_group->inode = NULL;
7465                         spin_unlock(&block_group->lock);
7466                         iput(inode);
7467                 } else {
7468                         spin_unlock(&block_group->lock);
7469                 }
7470                 /* One for our lookup ref */
7471                 btrfs_add_delayed_iput(inode);
7472         }
7473
7474         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7475         key.offset = block_group->key.objectid;
7476         key.type = 0;
7477
7478         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7479         if (ret < 0)
7480                 goto out;
7481         if (ret > 0)
7482                 btrfs_release_path(path);
7483         if (ret == 0) {
7484                 ret = btrfs_del_item(trans, tree_root, path);
7485                 if (ret)
7486                         goto out;
7487                 btrfs_release_path(path);
7488         }
7489
7490         spin_lock(&root->fs_info->block_group_cache_lock);
7491         rb_erase(&block_group->cache_node,
7492                  &root->fs_info->block_group_cache_tree);
7493         spin_unlock(&root->fs_info->block_group_cache_lock);
7494
7495         down_write(&block_group->space_info->groups_sem);
7496         /*
7497          * we must use list_del_init so people can check to see if they
7498          * are still on the list after taking the semaphore
7499          */
7500         list_del_init(&block_group->list);
7501         up_write(&block_group->space_info->groups_sem);
7502
7503         if (block_group->cached == BTRFS_CACHE_STARTED)
7504                 wait_block_group_cache_done(block_group);
7505
7506         btrfs_remove_free_space_cache(block_group);
7507
7508         spin_lock(&block_group->space_info->lock);
7509         block_group->space_info->total_bytes -= block_group->key.offset;
7510         block_group->space_info->bytes_readonly -= block_group->key.offset;
7511         block_group->space_info->disk_total -= block_group->key.offset * factor;
7512         spin_unlock(&block_group->space_info->lock);
7513
7514         memcpy(&key, &block_group->key, sizeof(key));
7515
7516         btrfs_clear_space_info_full(root->fs_info);
7517
7518         btrfs_put_block_group(block_group);
7519         btrfs_put_block_group(block_group);
7520
7521         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7522         if (ret > 0)
7523                 ret = -EIO;
7524         if (ret < 0)
7525                 goto out;
7526
7527         ret = btrfs_del_item(trans, root, path);
7528 out:
7529         btrfs_free_path(path);
7530         return ret;
7531 }
7532
7533 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7534 {
7535         struct btrfs_space_info *space_info;
7536         struct btrfs_super_block *disk_super;
7537         u64 features;
7538         u64 flags;
7539         int mixed = 0;
7540         int ret;
7541
7542         disk_super = fs_info->super_copy;
7543         if (!btrfs_super_root(disk_super))
7544                 return 1;
7545
7546         features = btrfs_super_incompat_flags(disk_super);
7547         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7548                 mixed = 1;
7549
7550         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7551         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7552         if (ret)
7553                 goto out;
7554
7555         if (mixed) {
7556                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7557                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7558         } else {
7559                 flags = BTRFS_BLOCK_GROUP_METADATA;
7560                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7561                 if (ret)
7562                         goto out;
7563
7564                 flags = BTRFS_BLOCK_GROUP_DATA;
7565                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7566         }
7567 out:
7568         return ret;
7569 }
7570
7571 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7572 {
7573         return unpin_extent_range(root, start, end);
7574 }
7575
7576 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7577                                u64 num_bytes, u64 *actual_bytes)
7578 {
7579         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7580 }
7581
7582 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7583 {
7584         struct btrfs_fs_info *fs_info = root->fs_info;
7585         struct btrfs_block_group_cache *cache = NULL;
7586         u64 group_trimmed;
7587         u64 start;
7588         u64 end;
7589         u64 trimmed = 0;
7590         int ret = 0;
7591
7592         cache = btrfs_lookup_block_group(fs_info, range->start);
7593
7594         while (cache) {
7595                 if (cache->key.objectid >= (range->start + range->len)) {
7596                         btrfs_put_block_group(cache);
7597                         break;
7598                 }
7599
7600                 start = max(range->start, cache->key.objectid);
7601                 end = min(range->start + range->len,
7602                                 cache->key.objectid + cache->key.offset);
7603
7604                 if (end - start >= range->minlen) {
7605                         if (!block_group_cache_done(cache)) {
7606                                 ret = cache_block_group(cache, NULL, root, 0);
7607                                 if (!ret)
7608                                         wait_block_group_cache_done(cache);
7609                         }
7610                         ret = btrfs_trim_block_group(cache,
7611                                                      &group_trimmed,
7612                                                      start,
7613                                                      end,
7614                                                      range->minlen);
7615
7616                         trimmed += group_trimmed;
7617                         if (ret) {
7618                                 btrfs_put_block_group(cache);
7619                                 break;
7620                         }
7621                 }
7622
7623                 cache = next_block_group(fs_info->tree_root, cache);
7624         }
7625
7626         range->len = trimmed;
7627         return ret;
7628 }