0c482e0d7c434093227d162a23ac56d6fc895943
[sfrench/cifs-2.6.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 "compat.h"
25 #include "hash.h"
26 #include "crc32c.h"
27 #include "ctree.h"
28 #include "disk-io.h"
29 #include "print-tree.h"
30 #include "transaction.h"
31 #include "volumes.h"
32 #include "locking.h"
33 #include "ref-cache.h"
34
35 #define PENDING_EXTENT_INSERT 0
36 #define PENDING_EXTENT_DELETE 1
37 #define PENDING_BACKREF_UPDATE 2
38
39 struct pending_extent_op {
40         int type;
41         u64 bytenr;
42         u64 num_bytes;
43         u64 parent;
44         u64 orig_parent;
45         u64 generation;
46         u64 orig_generation;
47         int level;
48         struct list_head list;
49         int del;
50 };
51
52 static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
53                                          struct btrfs_root *root, u64 parent,
54                                          u64 root_objectid, u64 ref_generation,
55                                          u64 owner, struct btrfs_key *ins,
56                                          int ref_mod);
57 static int update_reserved_extents(struct btrfs_root *root,
58                                    u64 bytenr, u64 num, int reserve);
59 static int update_block_group(struct btrfs_trans_handle *trans,
60                               struct btrfs_root *root,
61                               u64 bytenr, u64 num_bytes, int alloc,
62                               int mark_free);
63 static noinline int __btrfs_free_extent(struct btrfs_trans_handle *trans,
64                                         struct btrfs_root *root,
65                                         u64 bytenr, u64 num_bytes, u64 parent,
66                                         u64 root_objectid, u64 ref_generation,
67                                         u64 owner_objectid, int pin,
68                                         int ref_to_drop);
69
70 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
71                           struct btrfs_root *extent_root, u64 alloc_bytes,
72                           u64 flags, int force);
73
74 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
75 {
76         return (cache->flags & bits) == bits;
77 }
78
79 /*
80  * this adds the block group to the fs_info rb tree for the block group
81  * cache
82  */
83 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
84                                 struct btrfs_block_group_cache *block_group)
85 {
86         struct rb_node **p;
87         struct rb_node *parent = NULL;
88         struct btrfs_block_group_cache *cache;
89
90         spin_lock(&info->block_group_cache_lock);
91         p = &info->block_group_cache_tree.rb_node;
92
93         while (*p) {
94                 parent = *p;
95                 cache = rb_entry(parent, struct btrfs_block_group_cache,
96                                  cache_node);
97                 if (block_group->key.objectid < cache->key.objectid) {
98                         p = &(*p)->rb_left;
99                 } else if (block_group->key.objectid > cache->key.objectid) {
100                         p = &(*p)->rb_right;
101                 } else {
102                         spin_unlock(&info->block_group_cache_lock);
103                         return -EEXIST;
104                 }
105         }
106
107         rb_link_node(&block_group->cache_node, parent, p);
108         rb_insert_color(&block_group->cache_node,
109                         &info->block_group_cache_tree);
110         spin_unlock(&info->block_group_cache_lock);
111
112         return 0;
113 }
114
115 /*
116  * This will return the block group at or after bytenr if contains is 0, else
117  * it will return the block group that contains the bytenr
118  */
119 static struct btrfs_block_group_cache *
120 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
121                               int contains)
122 {
123         struct btrfs_block_group_cache *cache, *ret = NULL;
124         struct rb_node *n;
125         u64 end, start;
126
127         spin_lock(&info->block_group_cache_lock);
128         n = info->block_group_cache_tree.rb_node;
129
130         while (n) {
131                 cache = rb_entry(n, struct btrfs_block_group_cache,
132                                  cache_node);
133                 end = cache->key.objectid + cache->key.offset - 1;
134                 start = cache->key.objectid;
135
136                 if (bytenr < start) {
137                         if (!contains && (!ret || start < ret->key.objectid))
138                                 ret = cache;
139                         n = n->rb_left;
140                 } else if (bytenr > start) {
141                         if (contains && bytenr <= end) {
142                                 ret = cache;
143                                 break;
144                         }
145                         n = n->rb_right;
146                 } else {
147                         ret = cache;
148                         break;
149                 }
150         }
151         if (ret)
152                 atomic_inc(&ret->count);
153         spin_unlock(&info->block_group_cache_lock);
154
155         return ret;
156 }
157
158 /*
159  * this is only called by cache_block_group, since we could have freed extents
160  * we need to check the pinned_extents for any extents that can't be used yet
161  * since their free space will be released as soon as the transaction commits.
162  */
163 static int add_new_free_space(struct btrfs_block_group_cache *block_group,
164                               struct btrfs_fs_info *info, u64 start, u64 end)
165 {
166         u64 extent_start, extent_end, size;
167         int ret;
168
169         mutex_lock(&info->pinned_mutex);
170         while (start < end) {
171                 ret = find_first_extent_bit(&info->pinned_extents, start,
172                                             &extent_start, &extent_end,
173                                             EXTENT_DIRTY);
174                 if (ret)
175                         break;
176
177                 if (extent_start == start) {
178                         start = extent_end + 1;
179                 } else if (extent_start > start && extent_start < end) {
180                         size = extent_start - start;
181                         ret = btrfs_add_free_space(block_group, start,
182                                                    size);
183                         BUG_ON(ret);
184                         start = extent_end + 1;
185                 } else {
186                         break;
187                 }
188         }
189
190         if (start < end) {
191                 size = end - start;
192                 ret = btrfs_add_free_space(block_group, start, size);
193                 BUG_ON(ret);
194         }
195         mutex_unlock(&info->pinned_mutex);
196
197         return 0;
198 }
199
200 static int remove_sb_from_cache(struct btrfs_root *root,
201                                 struct btrfs_block_group_cache *cache)
202 {
203         u64 bytenr;
204         u64 *logical;
205         int stripe_len;
206         int i, nr, ret;
207
208         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
209                 bytenr = btrfs_sb_offset(i);
210                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
211                                        cache->key.objectid, bytenr, 0,
212                                        &logical, &nr, &stripe_len);
213                 BUG_ON(ret);
214                 while (nr--) {
215                         btrfs_remove_free_space(cache, logical[nr],
216                                                 stripe_len);
217                 }
218                 kfree(logical);
219         }
220         return 0;
221 }
222
223 static int cache_block_group(struct btrfs_root *root,
224                              struct btrfs_block_group_cache *block_group)
225 {
226         struct btrfs_path *path;
227         int ret = 0;
228         struct btrfs_key key;
229         struct extent_buffer *leaf;
230         int slot;
231         u64 last;
232
233         if (!block_group)
234                 return 0;
235
236         root = root->fs_info->extent_root;
237
238         if (block_group->cached)
239                 return 0;
240
241         path = btrfs_alloc_path();
242         if (!path)
243                 return -ENOMEM;
244
245         path->reada = 2;
246         /*
247          * we get into deadlocks with paths held by callers of this function.
248          * since the alloc_mutex is protecting things right now, just
249          * skip the locking here
250          */
251         path->skip_locking = 1;
252         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
253         key.objectid = last;
254         key.offset = 0;
255         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
256         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
257         if (ret < 0)
258                 goto err;
259
260         while (1) {
261                 leaf = path->nodes[0];
262                 slot = path->slots[0];
263                 if (slot >= btrfs_header_nritems(leaf)) {
264                         ret = btrfs_next_leaf(root, path);
265                         if (ret < 0)
266                                 goto err;
267                         if (ret == 0)
268                                 continue;
269                         else
270                                 break;
271                 }
272                 btrfs_item_key_to_cpu(leaf, &key, slot);
273                 if (key.objectid < block_group->key.objectid)
274                         goto next;
275
276                 if (key.objectid >= block_group->key.objectid +
277                     block_group->key.offset)
278                         break;
279
280                 if (btrfs_key_type(&key) == BTRFS_EXTENT_ITEM_KEY) {
281                         add_new_free_space(block_group, root->fs_info, last,
282                                            key.objectid);
283
284                         last = key.objectid + key.offset;
285                 }
286 next:
287                 path->slots[0]++;
288         }
289
290         add_new_free_space(block_group, root->fs_info, last,
291                            block_group->key.objectid +
292                            block_group->key.offset);
293
294         remove_sb_from_cache(root, block_group);
295         block_group->cached = 1;
296         ret = 0;
297 err:
298         btrfs_free_path(path);
299         return ret;
300 }
301
302 /*
303  * return the block group that starts at or after bytenr
304  */
305 static struct btrfs_block_group_cache *
306 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
307 {
308         struct btrfs_block_group_cache *cache;
309
310         cache = block_group_cache_tree_search(info, bytenr, 0);
311
312         return cache;
313 }
314
315 /*
316  * return the block group that contains teh given bytenr
317  */
318 struct btrfs_block_group_cache *btrfs_lookup_block_group(
319                                                  struct btrfs_fs_info *info,
320                                                  u64 bytenr)
321 {
322         struct btrfs_block_group_cache *cache;
323
324         cache = block_group_cache_tree_search(info, bytenr, 1);
325
326         return cache;
327 }
328
329 static inline void put_block_group(struct btrfs_block_group_cache *cache)
330 {
331         if (atomic_dec_and_test(&cache->count))
332                 kfree(cache);
333 }
334
335 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
336                                                   u64 flags)
337 {
338         struct list_head *head = &info->space_info;
339         struct btrfs_space_info *found;
340
341         rcu_read_lock();
342         list_for_each_entry_rcu(found, head, list) {
343                 if (found->flags == flags) {
344                         rcu_read_unlock();
345                         return found;
346                 }
347         }
348         rcu_read_unlock();
349         return NULL;
350 }
351
352 /*
353  * after adding space to the filesystem, we need to clear the full flags
354  * on all the space infos.
355  */
356 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
357 {
358         struct list_head *head = &info->space_info;
359         struct btrfs_space_info *found;
360
361         rcu_read_lock();
362         list_for_each_entry_rcu(found, head, list)
363                 found->full = 0;
364         rcu_read_unlock();
365 }
366
367 static u64 div_factor(u64 num, int factor)
368 {
369         if (factor == 10)
370                 return num;
371         num *= factor;
372         do_div(num, 10);
373         return num;
374 }
375
376 u64 btrfs_find_block_group(struct btrfs_root *root,
377                            u64 search_start, u64 search_hint, int owner)
378 {
379         struct btrfs_block_group_cache *cache;
380         u64 used;
381         u64 last = max(search_hint, search_start);
382         u64 group_start = 0;
383         int full_search = 0;
384         int factor = 9;
385         int wrapped = 0;
386 again:
387         while (1) {
388                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
389                 if (!cache)
390                         break;
391
392                 spin_lock(&cache->lock);
393                 last = cache->key.objectid + cache->key.offset;
394                 used = btrfs_block_group_used(&cache->item);
395
396                 if ((full_search || !cache->ro) &&
397                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
398                         if (used + cache->pinned + cache->reserved <
399                             div_factor(cache->key.offset, factor)) {
400                                 group_start = cache->key.objectid;
401                                 spin_unlock(&cache->lock);
402                                 put_block_group(cache);
403                                 goto found;
404                         }
405                 }
406                 spin_unlock(&cache->lock);
407                 put_block_group(cache);
408                 cond_resched();
409         }
410         if (!wrapped) {
411                 last = search_start;
412                 wrapped = 1;
413                 goto again;
414         }
415         if (!full_search && factor < 10) {
416                 last = search_start;
417                 full_search = 1;
418                 factor = 10;
419                 goto again;
420         }
421 found:
422         return group_start;
423 }
424
425 /* simple helper to search for an existing extent at a given offset */
426 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
427 {
428         int ret;
429         struct btrfs_key key;
430         struct btrfs_path *path;
431
432         path = btrfs_alloc_path();
433         BUG_ON(!path);
434         key.objectid = start;
435         key.offset = len;
436         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
437         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
438                                 0, 0);
439         btrfs_free_path(path);
440         return ret;
441 }
442
443 /*
444  * Back reference rules.  Back refs have three main goals:
445  *
446  * 1) differentiate between all holders of references to an extent so that
447  *    when a reference is dropped we can make sure it was a valid reference
448  *    before freeing the extent.
449  *
450  * 2) Provide enough information to quickly find the holders of an extent
451  *    if we notice a given block is corrupted or bad.
452  *
453  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
454  *    maintenance.  This is actually the same as #2, but with a slightly
455  *    different use case.
456  *
457  * File extents can be referenced by:
458  *
459  * - multiple snapshots, subvolumes, or different generations in one subvol
460  * - different files inside a single subvolume
461  * - different offsets inside a file (bookend extents in file.c)
462  *
463  * The extent ref structure has fields for:
464  *
465  * - Objectid of the subvolume root
466  * - Generation number of the tree holding the reference
467  * - objectid of the file holding the reference
468  * - number of references holding by parent node (alway 1 for tree blocks)
469  *
470  * Btree leaf may hold multiple references to a file extent. In most cases,
471  * these references are from same file and the corresponding offsets inside
472  * the file are close together.
473  *
474  * When a file extent is allocated the fields are filled in:
475  *     (root_key.objectid, trans->transid, inode objectid, 1)
476  *
477  * When a leaf is cow'd new references are added for every file extent found
478  * in the leaf.  It looks similar to the create case, but trans->transid will
479  * be different when the block is cow'd.
480  *
481  *     (root_key.objectid, trans->transid, inode objectid,
482  *      number of references in the leaf)
483  *
484  * When a file extent is removed either during snapshot deletion or
485  * file truncation, we find the corresponding back reference and check
486  * the following fields:
487  *
488  *     (btrfs_header_owner(leaf), btrfs_header_generation(leaf),
489  *      inode objectid)
490  *
491  * Btree extents can be referenced by:
492  *
493  * - Different subvolumes
494  * - Different generations of the same subvolume
495  *
496  * When a tree block is created, back references are inserted:
497  *
498  * (root->root_key.objectid, trans->transid, level, 1)
499  *
500  * When a tree block is cow'd, new back references are added for all the
501  * blocks it points to. If the tree block isn't in reference counted root,
502  * the old back references are removed. These new back references are of
503  * the form (trans->transid will have increased since creation):
504  *
505  * (root->root_key.objectid, trans->transid, level, 1)
506  *
507  * When a backref is in deleting, the following fields are checked:
508  *
509  * if backref was for a tree root:
510  *     (btrfs_header_owner(itself), btrfs_header_generation(itself), level)
511  * else
512  *     (btrfs_header_owner(parent), btrfs_header_generation(parent), level)
513  *
514  * Back Reference Key composing:
515  *
516  * The key objectid corresponds to the first byte in the extent, the key
517  * type is set to BTRFS_EXTENT_REF_KEY, and the key offset is the first
518  * byte of parent extent. If a extent is tree root, the key offset is set
519  * to the key objectid.
520  */
521
522 static noinline int lookup_extent_backref(struct btrfs_trans_handle *trans,
523                                           struct btrfs_root *root,
524                                           struct btrfs_path *path,
525                                           u64 bytenr, u64 parent,
526                                           u64 ref_root, u64 ref_generation,
527                                           u64 owner_objectid, int del)
528 {
529         struct btrfs_key key;
530         struct btrfs_extent_ref *ref;
531         struct extent_buffer *leaf;
532         u64 ref_objectid;
533         int ret;
534
535         key.objectid = bytenr;
536         key.type = BTRFS_EXTENT_REF_KEY;
537         key.offset = parent;
538
539         ret = btrfs_search_slot(trans, root, &key, path, del ? -1 : 0, 1);
540         if (ret < 0)
541                 goto out;
542         if (ret > 0) {
543                 ret = -ENOENT;
544                 goto out;
545         }
546
547         leaf = path->nodes[0];
548         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
549         ref_objectid = btrfs_ref_objectid(leaf, ref);
550         if (btrfs_ref_root(leaf, ref) != ref_root ||
551             btrfs_ref_generation(leaf, ref) != ref_generation ||
552             (ref_objectid != owner_objectid &&
553              ref_objectid != BTRFS_MULTIPLE_OBJECTIDS)) {
554                 ret = -EIO;
555                 WARN_ON(1);
556                 goto out;
557         }
558         ret = 0;
559 out:
560         return ret;
561 }
562
563 static noinline int insert_extent_backref(struct btrfs_trans_handle *trans,
564                                           struct btrfs_root *root,
565                                           struct btrfs_path *path,
566                                           u64 bytenr, u64 parent,
567                                           u64 ref_root, u64 ref_generation,
568                                           u64 owner_objectid,
569                                           int refs_to_add)
570 {
571         struct btrfs_key key;
572         struct extent_buffer *leaf;
573         struct btrfs_extent_ref *ref;
574         u32 num_refs;
575         int ret;
576
577         key.objectid = bytenr;
578         key.type = BTRFS_EXTENT_REF_KEY;
579         key.offset = parent;
580
581         ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*ref));
582         if (ret == 0) {
583                 leaf = path->nodes[0];
584                 ref = btrfs_item_ptr(leaf, path->slots[0],
585                                      struct btrfs_extent_ref);
586                 btrfs_set_ref_root(leaf, ref, ref_root);
587                 btrfs_set_ref_generation(leaf, ref, ref_generation);
588                 btrfs_set_ref_objectid(leaf, ref, owner_objectid);
589                 btrfs_set_ref_num_refs(leaf, ref, refs_to_add);
590         } else if (ret == -EEXIST) {
591                 u64 existing_owner;
592
593                 BUG_ON(owner_objectid < BTRFS_FIRST_FREE_OBJECTID);
594                 leaf = path->nodes[0];
595                 ref = btrfs_item_ptr(leaf, path->slots[0],
596                                      struct btrfs_extent_ref);
597                 if (btrfs_ref_root(leaf, ref) != ref_root ||
598                     btrfs_ref_generation(leaf, ref) != ref_generation) {
599                         ret = -EIO;
600                         WARN_ON(1);
601                         goto out;
602                 }
603
604                 num_refs = btrfs_ref_num_refs(leaf, ref);
605                 BUG_ON(num_refs == 0);
606                 btrfs_set_ref_num_refs(leaf, ref, num_refs + refs_to_add);
607
608                 existing_owner = btrfs_ref_objectid(leaf, ref);
609                 if (existing_owner != owner_objectid &&
610                     existing_owner != BTRFS_MULTIPLE_OBJECTIDS) {
611                         btrfs_set_ref_objectid(leaf, ref,
612                                         BTRFS_MULTIPLE_OBJECTIDS);
613                 }
614                 ret = 0;
615         } else {
616                 goto out;
617         }
618         btrfs_unlock_up_safe(path, 1);
619         btrfs_mark_buffer_dirty(path->nodes[0]);
620 out:
621         btrfs_release_path(root, path);
622         return ret;
623 }
624
625 static noinline int remove_extent_backref(struct btrfs_trans_handle *trans,
626                                           struct btrfs_root *root,
627                                           struct btrfs_path *path,
628                                           int refs_to_drop)
629 {
630         struct extent_buffer *leaf;
631         struct btrfs_extent_ref *ref;
632         u32 num_refs;
633         int ret = 0;
634
635         leaf = path->nodes[0];
636         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_ref);
637         num_refs = btrfs_ref_num_refs(leaf, ref);
638         BUG_ON(num_refs < refs_to_drop);
639         num_refs -= refs_to_drop;
640         if (num_refs == 0) {
641                 ret = btrfs_del_item(trans, root, path);
642         } else {
643                 btrfs_set_ref_num_refs(leaf, ref, num_refs);
644                 btrfs_mark_buffer_dirty(leaf);
645         }
646         btrfs_release_path(root, path);
647         return ret;
648 }
649
650 #ifdef BIO_RW_DISCARD
651 static void btrfs_issue_discard(struct block_device *bdev,
652                                 u64 start, u64 len)
653 {
654         blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_KERNEL);
655 }
656 #endif
657
658 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
659                                 u64 num_bytes)
660 {
661 #ifdef BIO_RW_DISCARD
662         int ret;
663         u64 map_length = num_bytes;
664         struct btrfs_multi_bio *multi = NULL;
665
666         /* Tell the block device(s) that the sectors can be discarded */
667         ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
668                               bytenr, &map_length, &multi, 0);
669         if (!ret) {
670                 struct btrfs_bio_stripe *stripe = multi->stripes;
671                 int i;
672
673                 if (map_length > num_bytes)
674                         map_length = num_bytes;
675
676                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
677                         btrfs_issue_discard(stripe->dev->bdev,
678                                             stripe->physical,
679                                             map_length);
680                 }
681                 kfree(multi);
682         }
683
684         return ret;
685 #else
686         return 0;
687 #endif
688 }
689
690 static int __btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
691                                      struct btrfs_root *root, u64 bytenr,
692                                      u64 num_bytes,
693                                      u64 orig_parent, u64 parent,
694                                      u64 orig_root, u64 ref_root,
695                                      u64 orig_generation, u64 ref_generation,
696                                      u64 owner_objectid)
697 {
698         int ret;
699         int pin = owner_objectid < BTRFS_FIRST_FREE_OBJECTID;
700
701         ret = btrfs_update_delayed_ref(trans, bytenr, num_bytes,
702                                        orig_parent, parent, orig_root,
703                                        ref_root, orig_generation,
704                                        ref_generation, owner_objectid, pin);
705         BUG_ON(ret);
706         return ret;
707 }
708
709 int btrfs_update_extent_ref(struct btrfs_trans_handle *trans,
710                             struct btrfs_root *root, u64 bytenr,
711                             u64 num_bytes, u64 orig_parent, u64 parent,
712                             u64 ref_root, u64 ref_generation,
713                             u64 owner_objectid)
714 {
715         int ret;
716         if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
717             owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
718                 return 0;
719
720         ret = __btrfs_update_extent_ref(trans, root, bytenr, num_bytes,
721                                         orig_parent, parent, ref_root,
722                                         ref_root, ref_generation,
723                                         ref_generation, owner_objectid);
724         return ret;
725 }
726 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
727                                   struct btrfs_root *root, u64 bytenr,
728                                   u64 num_bytes,
729                                   u64 orig_parent, u64 parent,
730                                   u64 orig_root, u64 ref_root,
731                                   u64 orig_generation, u64 ref_generation,
732                                   u64 owner_objectid)
733 {
734         int ret;
735
736         ret = btrfs_add_delayed_ref(trans, bytenr, num_bytes, parent, ref_root,
737                                     ref_generation, owner_objectid,
738                                     BTRFS_ADD_DELAYED_REF, 0);
739         BUG_ON(ret);
740         return ret;
741 }
742
743 static noinline_for_stack int add_extent_ref(struct btrfs_trans_handle *trans,
744                           struct btrfs_root *root, u64 bytenr,
745                           u64 num_bytes, u64 parent, u64 ref_root,
746                           u64 ref_generation, u64 owner_objectid,
747                           int refs_to_add)
748 {
749         struct btrfs_path *path;
750         int ret;
751         struct btrfs_key key;
752         struct extent_buffer *l;
753         struct btrfs_extent_item *item;
754         u32 refs;
755
756         path = btrfs_alloc_path();
757         if (!path)
758                 return -ENOMEM;
759
760         path->reada = 1;
761         path->leave_spinning = 1;
762         key.objectid = bytenr;
763         key.type = BTRFS_EXTENT_ITEM_KEY;
764         key.offset = num_bytes;
765
766         /* first find the extent item and update its reference count */
767         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
768                                 path, 0, 1);
769         if (ret < 0) {
770                 btrfs_set_path_blocking(path);
771                 return ret;
772         }
773
774         if (ret > 0) {
775                 WARN_ON(1);
776                 btrfs_free_path(path);
777                 return -EIO;
778         }
779         l = path->nodes[0];
780
781         btrfs_item_key_to_cpu(l, &key, path->slots[0]);
782         if (key.objectid != bytenr) {
783                 btrfs_print_leaf(root->fs_info->extent_root, path->nodes[0]);
784                 printk(KERN_ERR "btrfs wanted %llu found %llu\n",
785                        (unsigned long long)bytenr,
786                        (unsigned long long)key.objectid);
787                 BUG();
788         }
789         BUG_ON(key.type != BTRFS_EXTENT_ITEM_KEY);
790
791         item = btrfs_item_ptr(l, path->slots[0], struct btrfs_extent_item);
792
793         refs = btrfs_extent_refs(l, item);
794         btrfs_set_extent_refs(l, item, refs + refs_to_add);
795         btrfs_unlock_up_safe(path, 1);
796
797         btrfs_mark_buffer_dirty(path->nodes[0]);
798
799         btrfs_release_path(root->fs_info->extent_root, path);
800
801         path->reada = 1;
802         path->leave_spinning = 1;
803
804         /* now insert the actual backref */
805         ret = insert_extent_backref(trans, root->fs_info->extent_root,
806                                     path, bytenr, parent,
807                                     ref_root, ref_generation,
808                                     owner_objectid, refs_to_add);
809         BUG_ON(ret);
810         btrfs_free_path(path);
811         return 0;
812 }
813
814 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
815                          struct btrfs_root *root,
816                          u64 bytenr, u64 num_bytes, u64 parent,
817                          u64 ref_root, u64 ref_generation,
818                          u64 owner_objectid)
819 {
820         int ret;
821         if (ref_root == BTRFS_TREE_LOG_OBJECTID &&
822             owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
823                 return 0;
824
825         ret = __btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0, parent,
826                                      0, ref_root, 0, ref_generation,
827                                      owner_objectid);
828         return ret;
829 }
830
831 static int drop_delayed_ref(struct btrfs_trans_handle *trans,
832                                         struct btrfs_root *root,
833                                         struct btrfs_delayed_ref_node *node)
834 {
835         int ret = 0;
836         struct btrfs_delayed_ref *ref = btrfs_delayed_node_to_ref(node);
837
838         BUG_ON(node->ref_mod == 0);
839         ret = __btrfs_free_extent(trans, root, node->bytenr, node->num_bytes,
840                                   node->parent, ref->root, ref->generation,
841                                   ref->owner_objectid, ref->pin, node->ref_mod);
842
843         return ret;
844 }
845
846 /* helper function to actually process a single delayed ref entry */
847 static noinline int run_one_delayed_ref(struct btrfs_trans_handle *trans,
848                                         struct btrfs_root *root,
849                                         struct btrfs_delayed_ref_node *node,
850                                         int insert_reserved)
851 {
852         int ret;
853         struct btrfs_delayed_ref *ref;
854
855         if (node->parent == (u64)-1) {
856                 struct btrfs_delayed_ref_head *head;
857                 /*
858                  * we've hit the end of the chain and we were supposed
859                  * to insert this extent into the tree.  But, it got
860                  * deleted before we ever needed to insert it, so all
861                  * we have to do is clean up the accounting
862                  */
863                 if (insert_reserved) {
864                         update_reserved_extents(root, node->bytenr,
865                                                 node->num_bytes, 0);
866                 }
867                 head = btrfs_delayed_node_to_head(node);
868                 mutex_unlock(&head->mutex);
869                 return 0;
870         }
871
872         ref = btrfs_delayed_node_to_ref(node);
873         if (ref->action == BTRFS_ADD_DELAYED_REF) {
874                 if (insert_reserved) {
875                         struct btrfs_key ins;
876
877                         ins.objectid = node->bytenr;
878                         ins.offset = node->num_bytes;
879                         ins.type = BTRFS_EXTENT_ITEM_KEY;
880
881                         /* record the full extent allocation */
882                         ret = __btrfs_alloc_reserved_extent(trans, root,
883                                         node->parent, ref->root,
884                                         ref->generation, ref->owner_objectid,
885                                         &ins, node->ref_mod);
886                         update_reserved_extents(root, node->bytenr,
887                                                 node->num_bytes, 0);
888                 } else {
889                         /* just add one backref */
890                         ret = add_extent_ref(trans, root, node->bytenr,
891                                      node->num_bytes,
892                                      node->parent, ref->root, ref->generation,
893                                      ref->owner_objectid, node->ref_mod);
894                 }
895                 BUG_ON(ret);
896         } else if (ref->action == BTRFS_DROP_DELAYED_REF) {
897                 WARN_ON(insert_reserved);
898                 ret = drop_delayed_ref(trans, root, node);
899         }
900         return 0;
901 }
902
903 static noinline struct btrfs_delayed_ref_node *
904 select_delayed_ref(struct btrfs_delayed_ref_head *head)
905 {
906         struct rb_node *node;
907         struct btrfs_delayed_ref_node *ref;
908         int action = BTRFS_ADD_DELAYED_REF;
909 again:
910         /*
911          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
912          * this prevents ref count from going down to zero when
913          * there still are pending delayed ref.
914          */
915         node = rb_prev(&head->node.rb_node);
916         while (1) {
917                 if (!node)
918                         break;
919                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
920                                 rb_node);
921                 if (ref->bytenr != head->node.bytenr)
922                         break;
923                 if (btrfs_delayed_node_to_ref(ref)->action == action)
924                         return ref;
925                 node = rb_prev(node);
926         }
927         if (action == BTRFS_ADD_DELAYED_REF) {
928                 action = BTRFS_DROP_DELAYED_REF;
929                 goto again;
930         }
931         return NULL;
932 }
933
934 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
935                                        struct btrfs_root *root,
936                                        struct list_head *cluster)
937 {
938         struct btrfs_delayed_ref_root *delayed_refs;
939         struct btrfs_delayed_ref_node *ref;
940         struct btrfs_delayed_ref_head *locked_ref = NULL;
941         int ret;
942         int count = 0;
943         int must_insert_reserved = 0;
944
945         delayed_refs = &trans->transaction->delayed_refs;
946         while (1) {
947                 if (!locked_ref) {
948                         /* pick a new head ref from the cluster list */
949                         if (list_empty(cluster))
950                                 break;
951
952                         locked_ref = list_entry(cluster->next,
953                                      struct btrfs_delayed_ref_head, cluster);
954
955                         /* grab the lock that says we are going to process
956                          * all the refs for this head */
957                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
958
959                         /*
960                          * we may have dropped the spin lock to get the head
961                          * mutex lock, and that might have given someone else
962                          * time to free the head.  If that's true, it has been
963                          * removed from our list and we can move on.
964                          */
965                         if (ret == -EAGAIN) {
966                                 locked_ref = NULL;
967                                 count++;
968                                 continue;
969                         }
970                 }
971
972                 /*
973                  * record the must insert reserved flag before we
974                  * drop the spin lock.
975                  */
976                 must_insert_reserved = locked_ref->must_insert_reserved;
977                 locked_ref->must_insert_reserved = 0;
978
979                 /*
980                  * locked_ref is the head node, so we have to go one
981                  * node back for any delayed ref updates
982                  */
983                 ref = select_delayed_ref(locked_ref);
984                 if (!ref) {
985                         /* All delayed refs have been processed, Go ahead
986                          * and send the head node to run_one_delayed_ref,
987                          * so that any accounting fixes can happen
988                          */
989                         ref = &locked_ref->node;
990                         list_del_init(&locked_ref->cluster);
991                         locked_ref = NULL;
992                 }
993
994                 ref->in_tree = 0;
995                 rb_erase(&ref->rb_node, &delayed_refs->root);
996                 delayed_refs->num_entries--;
997                 spin_unlock(&delayed_refs->lock);
998
999                 ret = run_one_delayed_ref(trans, root, ref,
1000                                           must_insert_reserved);
1001                 BUG_ON(ret);
1002                 btrfs_put_delayed_ref(ref);
1003
1004                 count++;
1005                 cond_resched();
1006                 spin_lock(&delayed_refs->lock);
1007         }
1008         return count;
1009 }
1010
1011 /*
1012  * this starts processing the delayed reference count updates and
1013  * extent insertions we have queued up so far.  count can be
1014  * 0, which means to process everything in the tree at the start
1015  * of the run (but not newly added entries), or it can be some target
1016  * number you'd like to process.
1017  */
1018 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
1019                            struct btrfs_root *root, unsigned long count)
1020 {
1021         struct rb_node *node;
1022         struct btrfs_delayed_ref_root *delayed_refs;
1023         struct btrfs_delayed_ref_node *ref;
1024         struct list_head cluster;
1025         int ret;
1026         int run_all = count == (unsigned long)-1;
1027         int run_most = 0;
1028
1029         if (root == root->fs_info->extent_root)
1030                 root = root->fs_info->tree_root;
1031
1032         delayed_refs = &trans->transaction->delayed_refs;
1033         INIT_LIST_HEAD(&cluster);
1034 again:
1035         spin_lock(&delayed_refs->lock);
1036         if (count == 0) {
1037                 count = delayed_refs->num_entries * 2;
1038                 run_most = 1;
1039         }
1040         while (1) {
1041                 if (!(run_all || run_most) &&
1042                     delayed_refs->num_heads_ready < 64)
1043                         break;
1044
1045                 /*
1046                  * go find something we can process in the rbtree.  We start at
1047                  * the beginning of the tree, and then build a cluster
1048                  * of refs to process starting at the first one we are able to
1049                  * lock
1050                  */
1051                 ret = btrfs_find_ref_cluster(trans, &cluster,
1052                                              delayed_refs->run_delayed_start);
1053                 if (ret)
1054                         break;
1055
1056                 ret = run_clustered_refs(trans, root, &cluster);
1057                 BUG_ON(ret < 0);
1058
1059                 count -= min_t(unsigned long, ret, count);
1060
1061                 if (count == 0)
1062                         break;
1063         }
1064
1065         if (run_all) {
1066                 node = rb_first(&delayed_refs->root);
1067                 if (!node)
1068                         goto out;
1069                 count = (unsigned long)-1;
1070
1071                 while (node) {
1072                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
1073                                        rb_node);
1074                         if (btrfs_delayed_ref_is_head(ref)) {
1075                                 struct btrfs_delayed_ref_head *head;
1076
1077                                 head = btrfs_delayed_node_to_head(ref);
1078                                 atomic_inc(&ref->refs);
1079
1080                                 spin_unlock(&delayed_refs->lock);
1081                                 mutex_lock(&head->mutex);
1082                                 mutex_unlock(&head->mutex);
1083
1084                                 btrfs_put_delayed_ref(ref);
1085                                 cond_resched();
1086                                 goto again;
1087                         }
1088                         node = rb_next(node);
1089                 }
1090                 spin_unlock(&delayed_refs->lock);
1091                 schedule_timeout(1);
1092                 goto again;
1093         }
1094 out:
1095         spin_unlock(&delayed_refs->lock);
1096         return 0;
1097 }
1098
1099 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
1100                           struct btrfs_root *root, u64 objectid, u64 bytenr)
1101 {
1102         struct btrfs_root *extent_root = root->fs_info->extent_root;
1103         struct btrfs_path *path;
1104         struct extent_buffer *leaf;
1105         struct btrfs_extent_ref *ref_item;
1106         struct btrfs_key key;
1107         struct btrfs_key found_key;
1108         u64 ref_root;
1109         u64 last_snapshot;
1110         u32 nritems;
1111         int ret;
1112
1113         key.objectid = bytenr;
1114         key.offset = (u64)-1;
1115         key.type = BTRFS_EXTENT_ITEM_KEY;
1116
1117         path = btrfs_alloc_path();
1118         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1119         if (ret < 0)
1120                 goto out;
1121         BUG_ON(ret == 0);
1122
1123         ret = -ENOENT;
1124         if (path->slots[0] == 0)
1125                 goto out;
1126
1127         path->slots[0]--;
1128         leaf = path->nodes[0];
1129         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1130
1131         if (found_key.objectid != bytenr ||
1132             found_key.type != BTRFS_EXTENT_ITEM_KEY)
1133                 goto out;
1134
1135         last_snapshot = btrfs_root_last_snapshot(&root->root_item);
1136         while (1) {
1137                 leaf = path->nodes[0];
1138                 nritems = btrfs_header_nritems(leaf);
1139                 if (path->slots[0] >= nritems) {
1140                         ret = btrfs_next_leaf(extent_root, path);
1141                         if (ret < 0)
1142                                 goto out;
1143                         if (ret == 0)
1144                                 continue;
1145                         break;
1146                 }
1147                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1148                 if (found_key.objectid != bytenr)
1149                         break;
1150
1151                 if (found_key.type != BTRFS_EXTENT_REF_KEY) {
1152                         path->slots[0]++;
1153                         continue;
1154                 }
1155
1156                 ref_item = btrfs_item_ptr(leaf, path->slots[0],
1157                                           struct btrfs_extent_ref);
1158                 ref_root = btrfs_ref_root(leaf, ref_item);
1159                 if ((ref_root != root->root_key.objectid &&
1160                      ref_root != BTRFS_TREE_LOG_OBJECTID) ||
1161                      objectid != btrfs_ref_objectid(leaf, ref_item)) {
1162                         ret = 1;
1163                         goto out;
1164                 }
1165                 if (btrfs_ref_generation(leaf, ref_item) <= last_snapshot) {
1166                         ret = 1;
1167                         goto out;
1168                 }
1169
1170                 path->slots[0]++;
1171         }
1172         ret = 0;
1173 out:
1174         btrfs_free_path(path);
1175         return ret;
1176 }
1177
1178 int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
1179                     struct extent_buffer *buf, u32 nr_extents)
1180 {
1181         struct btrfs_key key;
1182         struct btrfs_file_extent_item *fi;
1183         u64 root_gen;
1184         u32 nritems;
1185         int i;
1186         int level;
1187         int ret = 0;
1188         int shared = 0;
1189
1190         if (!root->ref_cows)
1191                 return 0;
1192
1193         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
1194                 shared = 0;
1195                 root_gen = root->root_key.offset;
1196         } else {
1197                 shared = 1;
1198                 root_gen = trans->transid - 1;
1199         }
1200
1201         level = btrfs_header_level(buf);
1202         nritems = btrfs_header_nritems(buf);
1203
1204         if (level == 0) {
1205                 struct btrfs_leaf_ref *ref;
1206                 struct btrfs_extent_info *info;
1207
1208                 ref = btrfs_alloc_leaf_ref(root, nr_extents);
1209                 if (!ref) {
1210                         ret = -ENOMEM;
1211                         goto out;
1212                 }
1213
1214                 ref->root_gen = root_gen;
1215                 ref->bytenr = buf->start;
1216                 ref->owner = btrfs_header_owner(buf);
1217                 ref->generation = btrfs_header_generation(buf);
1218                 ref->nritems = nr_extents;
1219                 info = ref->extents;
1220
1221                 for (i = 0; nr_extents > 0 && i < nritems; i++) {
1222                         u64 disk_bytenr;
1223                         btrfs_item_key_to_cpu(buf, &key, i);
1224                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
1225                                 continue;
1226                         fi = btrfs_item_ptr(buf, i,
1227                                             struct btrfs_file_extent_item);
1228                         if (btrfs_file_extent_type(buf, fi) ==
1229                             BTRFS_FILE_EXTENT_INLINE)
1230                                 continue;
1231                         disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
1232                         if (disk_bytenr == 0)
1233                                 continue;
1234
1235                         info->bytenr = disk_bytenr;
1236                         info->num_bytes =
1237                                 btrfs_file_extent_disk_num_bytes(buf, fi);
1238                         info->objectid = key.objectid;
1239                         info->offset = key.offset;
1240                         info++;
1241                 }
1242
1243                 ret = btrfs_add_leaf_ref(root, ref, shared);
1244                 if (ret == -EEXIST && shared) {
1245                         struct btrfs_leaf_ref *old;
1246                         old = btrfs_lookup_leaf_ref(root, ref->bytenr);
1247                         BUG_ON(!old);
1248                         btrfs_remove_leaf_ref(root, old);
1249                         btrfs_free_leaf_ref(root, old);
1250                         ret = btrfs_add_leaf_ref(root, ref, shared);
1251                 }
1252                 WARN_ON(ret);
1253                 btrfs_free_leaf_ref(root, ref);
1254         }
1255 out:
1256         return ret;
1257 }
1258
1259 /* when a block goes through cow, we update the reference counts of
1260  * everything that block points to.  The internal pointers of the block
1261  * can be in just about any order, and it is likely to have clusters of
1262  * things that are close together and clusters of things that are not.
1263  *
1264  * To help reduce the seeks that come with updating all of these reference
1265  * counts, sort them by byte number before actual updates are done.
1266  *
1267  * struct refsort is used to match byte number to slot in the btree block.
1268  * we sort based on the byte number and then use the slot to actually
1269  * find the item.
1270  *
1271  * struct refsort is smaller than strcut btrfs_item and smaller than
1272  * struct btrfs_key_ptr.  Since we're currently limited to the page size
1273  * for a btree block, there's no way for a kmalloc of refsorts for a
1274  * single node to be bigger than a page.
1275  */
1276 struct refsort {
1277         u64 bytenr;
1278         u32 slot;
1279 };
1280
1281 /*
1282  * for passing into sort()
1283  */
1284 static int refsort_cmp(const void *a_void, const void *b_void)
1285 {
1286         const struct refsort *a = a_void;
1287         const struct refsort *b = b_void;
1288
1289         if (a->bytenr < b->bytenr)
1290                 return -1;
1291         if (a->bytenr > b->bytenr)
1292                 return 1;
1293         return 0;
1294 }
1295
1296
1297 noinline int btrfs_inc_ref(struct btrfs_trans_handle *trans,
1298                            struct btrfs_root *root,
1299                            struct extent_buffer *orig_buf,
1300                            struct extent_buffer *buf, u32 *nr_extents)
1301 {
1302         u64 bytenr;
1303         u64 ref_root;
1304         u64 orig_root;
1305         u64 ref_generation;
1306         u64 orig_generation;
1307         struct refsort *sorted;
1308         u32 nritems;
1309         u32 nr_file_extents = 0;
1310         struct btrfs_key key;
1311         struct btrfs_file_extent_item *fi;
1312         int i;
1313         int level;
1314         int ret = 0;
1315         int faili = 0;
1316         int refi = 0;
1317         int slot;
1318         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
1319                             u64, u64, u64, u64, u64, u64, u64, u64, u64);
1320
1321         ref_root = btrfs_header_owner(buf);
1322         ref_generation = btrfs_header_generation(buf);
1323         orig_root = btrfs_header_owner(orig_buf);
1324         orig_generation = btrfs_header_generation(orig_buf);
1325
1326         nritems = btrfs_header_nritems(buf);
1327         level = btrfs_header_level(buf);
1328
1329         sorted = kmalloc(sizeof(struct refsort) * nritems, GFP_NOFS);
1330         BUG_ON(!sorted);
1331
1332         if (root->ref_cows) {
1333                 process_func = __btrfs_inc_extent_ref;
1334         } else {
1335                 if (level == 0 &&
1336                     root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
1337                         goto out;
1338                 if (level != 0 &&
1339                     root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID)
1340                         goto out;
1341                 process_func = __btrfs_update_extent_ref;
1342         }
1343
1344         /*
1345          * we make two passes through the items.  In the first pass we
1346          * only record the byte number and slot.  Then we sort based on
1347          * byte number and do the actual work based on the sorted results
1348          */
1349         for (i = 0; i < nritems; i++) {
1350                 cond_resched();
1351                 if (level == 0) {
1352                         btrfs_item_key_to_cpu(buf, &key, i);
1353                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
1354                                 continue;
1355                         fi = btrfs_item_ptr(buf, i,
1356                                             struct btrfs_file_extent_item);
1357                         if (btrfs_file_extent_type(buf, fi) ==
1358                             BTRFS_FILE_EXTENT_INLINE)
1359                                 continue;
1360                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
1361                         if (bytenr == 0)
1362                                 continue;
1363
1364                         nr_file_extents++;
1365                         sorted[refi].bytenr = bytenr;
1366                         sorted[refi].slot = i;
1367                         refi++;
1368                 } else {
1369                         bytenr = btrfs_node_blockptr(buf, i);
1370                         sorted[refi].bytenr = bytenr;
1371                         sorted[refi].slot = i;
1372                         refi++;
1373                 }
1374         }
1375         /*
1376          * if refi == 0, we didn't actually put anything into the sorted
1377          * array and we're done
1378          */
1379         if (refi == 0)
1380                 goto out;
1381
1382         sort(sorted, refi, sizeof(struct refsort), refsort_cmp, NULL);
1383
1384         for (i = 0; i < refi; i++) {
1385                 cond_resched();
1386                 slot = sorted[i].slot;
1387                 bytenr = sorted[i].bytenr;
1388
1389                 if (level == 0) {
1390                         btrfs_item_key_to_cpu(buf, &key, slot);
1391                         fi = btrfs_item_ptr(buf, slot,
1392                                             struct btrfs_file_extent_item);
1393
1394                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
1395                         if (bytenr == 0)
1396                                 continue;
1397
1398                         ret = process_func(trans, root, bytenr,
1399                                    btrfs_file_extent_disk_num_bytes(buf, fi),
1400                                    orig_buf->start, buf->start,
1401                                    orig_root, ref_root,
1402                                    orig_generation, ref_generation,
1403                                    key.objectid);
1404
1405                         if (ret) {
1406                                 faili = slot;
1407                                 WARN_ON(1);
1408                                 goto fail;
1409                         }
1410                 } else {
1411                         ret = process_func(trans, root, bytenr, buf->len,
1412                                            orig_buf->start, buf->start,
1413                                            orig_root, ref_root,
1414                                            orig_generation, ref_generation,
1415                                            level - 1);
1416                         if (ret) {
1417                                 faili = slot;
1418                                 WARN_ON(1);
1419                                 goto fail;
1420                         }
1421                 }
1422         }
1423 out:
1424         kfree(sorted);
1425         if (nr_extents) {
1426                 if (level == 0)
1427                         *nr_extents = nr_file_extents;
1428                 else
1429                         *nr_extents = nritems;
1430         }
1431         return 0;
1432 fail:
1433         kfree(sorted);
1434         WARN_ON(1);
1435         return ret;
1436 }
1437
1438 int btrfs_update_ref(struct btrfs_trans_handle *trans,
1439                      struct btrfs_root *root, struct extent_buffer *orig_buf,
1440                      struct extent_buffer *buf, int start_slot, int nr)
1441
1442 {
1443         u64 bytenr;
1444         u64 ref_root;
1445         u64 orig_root;
1446         u64 ref_generation;
1447         u64 orig_generation;
1448         struct btrfs_key key;
1449         struct btrfs_file_extent_item *fi;
1450         int i;
1451         int ret;
1452         int slot;
1453         int level;
1454
1455         BUG_ON(start_slot < 0);
1456         BUG_ON(start_slot + nr > btrfs_header_nritems(buf));
1457
1458         ref_root = btrfs_header_owner(buf);
1459         ref_generation = btrfs_header_generation(buf);
1460         orig_root = btrfs_header_owner(orig_buf);
1461         orig_generation = btrfs_header_generation(orig_buf);
1462         level = btrfs_header_level(buf);
1463
1464         if (!root->ref_cows) {
1465                 if (level == 0 &&
1466                     root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
1467                         return 0;
1468                 if (level != 0 &&
1469                     root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID)
1470                         return 0;
1471         }
1472
1473         for (i = 0, slot = start_slot; i < nr; i++, slot++) {
1474                 cond_resched();
1475                 if (level == 0) {
1476                         btrfs_item_key_to_cpu(buf, &key, slot);
1477                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
1478                                 continue;
1479                         fi = btrfs_item_ptr(buf, slot,
1480                                             struct btrfs_file_extent_item);
1481                         if (btrfs_file_extent_type(buf, fi) ==
1482                             BTRFS_FILE_EXTENT_INLINE)
1483                                 continue;
1484                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
1485                         if (bytenr == 0)
1486                                 continue;
1487                         ret = __btrfs_update_extent_ref(trans, root, bytenr,
1488                                     btrfs_file_extent_disk_num_bytes(buf, fi),
1489                                     orig_buf->start, buf->start,
1490                                     orig_root, ref_root, orig_generation,
1491                                     ref_generation, key.objectid);
1492                         if (ret)
1493                                 goto fail;
1494                 } else {
1495                         bytenr = btrfs_node_blockptr(buf, slot);
1496                         ret = __btrfs_update_extent_ref(trans, root, bytenr,
1497                                             buf->len, orig_buf->start,
1498                                             buf->start, orig_root, ref_root,
1499                                             orig_generation, ref_generation,
1500                                             level - 1);
1501                         if (ret)
1502                                 goto fail;
1503                 }
1504         }
1505         return 0;
1506 fail:
1507         WARN_ON(1);
1508         return -1;
1509 }
1510
1511 static int write_one_cache_group(struct btrfs_trans_handle *trans,
1512                                  struct btrfs_root *root,
1513                                  struct btrfs_path *path,
1514                                  struct btrfs_block_group_cache *cache)
1515 {
1516         int ret;
1517         struct btrfs_root *extent_root = root->fs_info->extent_root;
1518         unsigned long bi;
1519         struct extent_buffer *leaf;
1520
1521         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
1522         if (ret < 0)
1523                 goto fail;
1524         BUG_ON(ret);
1525
1526         leaf = path->nodes[0];
1527         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
1528         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
1529         btrfs_mark_buffer_dirty(leaf);
1530         btrfs_release_path(extent_root, path);
1531 fail:
1532         if (ret)
1533                 return ret;
1534         return 0;
1535
1536 }
1537
1538 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
1539                                    struct btrfs_root *root)
1540 {
1541         struct btrfs_block_group_cache *cache, *entry;
1542         struct rb_node *n;
1543         int err = 0;
1544         int werr = 0;
1545         struct btrfs_path *path;
1546         u64 last = 0;
1547
1548         path = btrfs_alloc_path();
1549         if (!path)
1550                 return -ENOMEM;
1551
1552         while (1) {
1553                 cache = NULL;
1554                 spin_lock(&root->fs_info->block_group_cache_lock);
1555                 for (n = rb_first(&root->fs_info->block_group_cache_tree);
1556                      n; n = rb_next(n)) {
1557                         entry = rb_entry(n, struct btrfs_block_group_cache,
1558                                          cache_node);
1559                         if (entry->dirty) {
1560                                 cache = entry;
1561                                 break;
1562                         }
1563                 }
1564                 spin_unlock(&root->fs_info->block_group_cache_lock);
1565
1566                 if (!cache)
1567                         break;
1568
1569                 cache->dirty = 0;
1570                 last += cache->key.offset;
1571
1572                 err = write_one_cache_group(trans, root,
1573                                             path, cache);
1574                 /*
1575                  * if we fail to write the cache group, we want
1576                  * to keep it marked dirty in hopes that a later
1577                  * write will work
1578                  */
1579                 if (err) {
1580                         werr = err;
1581                         continue;
1582                 }
1583         }
1584         btrfs_free_path(path);
1585         return werr;
1586 }
1587
1588 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
1589 {
1590         struct btrfs_block_group_cache *block_group;
1591         int readonly = 0;
1592
1593         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
1594         if (!block_group || block_group->ro)
1595                 readonly = 1;
1596         if (block_group)
1597                 put_block_group(block_group);
1598         return readonly;
1599 }
1600
1601 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
1602                              u64 total_bytes, u64 bytes_used,
1603                              struct btrfs_space_info **space_info)
1604 {
1605         struct btrfs_space_info *found;
1606
1607         found = __find_space_info(info, flags);
1608         if (found) {
1609                 spin_lock(&found->lock);
1610                 found->total_bytes += total_bytes;
1611                 found->bytes_used += bytes_used;
1612                 found->full = 0;
1613                 spin_unlock(&found->lock);
1614                 *space_info = found;
1615                 return 0;
1616         }
1617         found = kzalloc(sizeof(*found), GFP_NOFS);
1618         if (!found)
1619                 return -ENOMEM;
1620
1621         INIT_LIST_HEAD(&found->block_groups);
1622         init_rwsem(&found->groups_sem);
1623         spin_lock_init(&found->lock);
1624         found->flags = flags;
1625         found->total_bytes = total_bytes;
1626         found->bytes_used = bytes_used;
1627         found->bytes_pinned = 0;
1628         found->bytes_reserved = 0;
1629         found->bytes_readonly = 0;
1630         found->bytes_delalloc = 0;
1631         found->full = 0;
1632         found->force_alloc = 0;
1633         *space_info = found;
1634         list_add_rcu(&found->list, &info->space_info);
1635         return 0;
1636 }
1637
1638 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1639 {
1640         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
1641                                    BTRFS_BLOCK_GROUP_RAID1 |
1642                                    BTRFS_BLOCK_GROUP_RAID10 |
1643                                    BTRFS_BLOCK_GROUP_DUP);
1644         if (extra_flags) {
1645                 if (flags & BTRFS_BLOCK_GROUP_DATA)
1646                         fs_info->avail_data_alloc_bits |= extra_flags;
1647                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1648                         fs_info->avail_metadata_alloc_bits |= extra_flags;
1649                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1650                         fs_info->avail_system_alloc_bits |= extra_flags;
1651         }
1652 }
1653
1654 static void set_block_group_readonly(struct btrfs_block_group_cache *cache)
1655 {
1656         spin_lock(&cache->space_info->lock);
1657         spin_lock(&cache->lock);
1658         if (!cache->ro) {
1659                 cache->space_info->bytes_readonly += cache->key.offset -
1660                                         btrfs_block_group_used(&cache->item);
1661                 cache->ro = 1;
1662         }
1663         spin_unlock(&cache->lock);
1664         spin_unlock(&cache->space_info->lock);
1665 }
1666
1667 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
1668 {
1669         u64 num_devices = root->fs_info->fs_devices->rw_devices;
1670
1671         if (num_devices == 1)
1672                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
1673         if (num_devices < 4)
1674                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
1675
1676         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
1677             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
1678                       BTRFS_BLOCK_GROUP_RAID10))) {
1679                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
1680         }
1681
1682         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
1683             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
1684                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
1685         }
1686
1687         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
1688             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
1689              (flags & BTRFS_BLOCK_GROUP_RAID10) |
1690              (flags & BTRFS_BLOCK_GROUP_DUP)))
1691                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
1692         return flags;
1693 }
1694
1695 static u64 btrfs_get_alloc_profile(struct btrfs_root *root, u64 data)
1696 {
1697         struct btrfs_fs_info *info = root->fs_info;
1698         u64 alloc_profile;
1699
1700         if (data) {
1701                 alloc_profile = info->avail_data_alloc_bits &
1702                         info->data_alloc_profile;
1703                 data = BTRFS_BLOCK_GROUP_DATA | alloc_profile;
1704         } else if (root == root->fs_info->chunk_root) {
1705                 alloc_profile = info->avail_system_alloc_bits &
1706                         info->system_alloc_profile;
1707                 data = BTRFS_BLOCK_GROUP_SYSTEM | alloc_profile;
1708         } else {
1709                 alloc_profile = info->avail_metadata_alloc_bits &
1710                         info->metadata_alloc_profile;
1711                 data = BTRFS_BLOCK_GROUP_METADATA | alloc_profile;
1712         }
1713
1714         return btrfs_reduce_alloc_profile(root, data);
1715 }
1716
1717 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
1718 {
1719         u64 alloc_target;
1720
1721         alloc_target = btrfs_get_alloc_profile(root, 1);
1722         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
1723                                                        alloc_target);
1724 }
1725
1726 /*
1727  * for now this just makes sure we have at least 5% of our metadata space free
1728  * for use.
1729  */
1730 int btrfs_check_metadata_free_space(struct btrfs_root *root)
1731 {
1732         struct btrfs_fs_info *info = root->fs_info;
1733         struct btrfs_space_info *meta_sinfo;
1734         u64 alloc_target, thresh;
1735         int committed = 0, ret;
1736
1737         /* get the space info for where the metadata will live */
1738         alloc_target = btrfs_get_alloc_profile(root, 0);
1739         meta_sinfo = __find_space_info(info, alloc_target);
1740
1741 again:
1742         spin_lock(&meta_sinfo->lock);
1743         if (!meta_sinfo->full)
1744                 thresh = meta_sinfo->total_bytes * 80;
1745         else
1746                 thresh = meta_sinfo->total_bytes * 95;
1747
1748         do_div(thresh, 100);
1749
1750         if (meta_sinfo->bytes_used + meta_sinfo->bytes_reserved +
1751             meta_sinfo->bytes_pinned + meta_sinfo->bytes_readonly > thresh) {
1752                 struct btrfs_trans_handle *trans;
1753                 if (!meta_sinfo->full) {
1754                         meta_sinfo->force_alloc = 1;
1755                         spin_unlock(&meta_sinfo->lock);
1756
1757                         trans = btrfs_start_transaction(root, 1);
1758                         if (!trans)
1759                                 return -ENOMEM;
1760
1761                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
1762                                              2 * 1024 * 1024, alloc_target, 0);
1763                         btrfs_end_transaction(trans, root);
1764                         goto again;
1765                 }
1766                 spin_unlock(&meta_sinfo->lock);
1767
1768                 if (!committed) {
1769                         committed = 1;
1770                         trans = btrfs_join_transaction(root, 1);
1771                         if (!trans)
1772                                 return -ENOMEM;
1773                         ret = btrfs_commit_transaction(trans, root);
1774                         if (ret)
1775                                 return ret;
1776                         goto again;
1777                 }
1778                 return -ENOSPC;
1779         }
1780         spin_unlock(&meta_sinfo->lock);
1781
1782         return 0;
1783 }
1784
1785 /*
1786  * This will check the space that the inode allocates from to make sure we have
1787  * enough space for bytes.
1788  */
1789 int btrfs_check_data_free_space(struct btrfs_root *root, struct inode *inode,
1790                                 u64 bytes)
1791 {
1792         struct btrfs_space_info *data_sinfo;
1793         int ret = 0, committed = 0;
1794
1795         /* make sure bytes are sectorsize aligned */
1796         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
1797
1798         data_sinfo = BTRFS_I(inode)->space_info;
1799 again:
1800         /* make sure we have enough space to handle the data first */
1801         spin_lock(&data_sinfo->lock);
1802         if (data_sinfo->total_bytes - data_sinfo->bytes_used -
1803             data_sinfo->bytes_delalloc - data_sinfo->bytes_reserved -
1804             data_sinfo->bytes_pinned - data_sinfo->bytes_readonly -
1805             data_sinfo->bytes_may_use < bytes) {
1806                 struct btrfs_trans_handle *trans;
1807
1808                 /*
1809                  * if we don't have enough free bytes in this space then we need
1810                  * to alloc a new chunk.
1811                  */
1812                 if (!data_sinfo->full) {
1813                         u64 alloc_target;
1814
1815                         data_sinfo->force_alloc = 1;
1816                         spin_unlock(&data_sinfo->lock);
1817
1818                         alloc_target = btrfs_get_alloc_profile(root, 1);
1819                         trans = btrfs_start_transaction(root, 1);
1820                         if (!trans)
1821                                 return -ENOMEM;
1822
1823                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
1824                                              bytes + 2 * 1024 * 1024,
1825                                              alloc_target, 0);
1826                         btrfs_end_transaction(trans, root);
1827                         if (ret)
1828                                 return ret;
1829                         goto again;
1830                 }
1831                 spin_unlock(&data_sinfo->lock);
1832
1833                 /* commit the current transaction and try again */
1834                 if (!committed) {
1835                         committed = 1;
1836                         trans = btrfs_join_transaction(root, 1);
1837                         if (!trans)
1838                                 return -ENOMEM;
1839                         ret = btrfs_commit_transaction(trans, root);
1840                         if (ret)
1841                                 return ret;
1842                         goto again;
1843                 }
1844
1845                 printk(KERN_ERR "no space left, need %llu, %llu delalloc bytes"
1846                        ", %llu bytes_used, %llu bytes_reserved, "
1847                        "%llu bytes_pinned, %llu bytes_readonly, %llu may use"
1848                        "%llu total\n", bytes, data_sinfo->bytes_delalloc,
1849                        data_sinfo->bytes_used, data_sinfo->bytes_reserved,
1850                        data_sinfo->bytes_pinned, data_sinfo->bytes_readonly,
1851                        data_sinfo->bytes_may_use, data_sinfo->total_bytes);
1852                 return -ENOSPC;
1853         }
1854         data_sinfo->bytes_may_use += bytes;
1855         BTRFS_I(inode)->reserved_bytes += bytes;
1856         spin_unlock(&data_sinfo->lock);
1857
1858         return btrfs_check_metadata_free_space(root);
1859 }
1860
1861 /*
1862  * if there was an error for whatever reason after calling
1863  * btrfs_check_data_free_space, call this so we can cleanup the counters.
1864  */
1865 void btrfs_free_reserved_data_space(struct btrfs_root *root,
1866                                     struct inode *inode, u64 bytes)
1867 {
1868         struct btrfs_space_info *data_sinfo;
1869
1870         /* make sure bytes are sectorsize aligned */
1871         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
1872
1873         data_sinfo = BTRFS_I(inode)->space_info;
1874         spin_lock(&data_sinfo->lock);
1875         data_sinfo->bytes_may_use -= bytes;
1876         BTRFS_I(inode)->reserved_bytes -= bytes;
1877         spin_unlock(&data_sinfo->lock);
1878 }
1879
1880 /* called when we are adding a delalloc extent to the inode's io_tree */
1881 void btrfs_delalloc_reserve_space(struct btrfs_root *root, struct inode *inode,
1882                                   u64 bytes)
1883 {
1884         struct btrfs_space_info *data_sinfo;
1885
1886         /* get the space info for where this inode will be storing its data */
1887         data_sinfo = BTRFS_I(inode)->space_info;
1888
1889         /* make sure we have enough space to handle the data first */
1890         spin_lock(&data_sinfo->lock);
1891         data_sinfo->bytes_delalloc += bytes;
1892
1893         /*
1894          * we are adding a delalloc extent without calling
1895          * btrfs_check_data_free_space first.  This happens on a weird
1896          * writepage condition, but shouldn't hurt our accounting
1897          */
1898         if (unlikely(bytes > BTRFS_I(inode)->reserved_bytes)) {
1899                 data_sinfo->bytes_may_use -= BTRFS_I(inode)->reserved_bytes;
1900                 BTRFS_I(inode)->reserved_bytes = 0;
1901         } else {
1902                 data_sinfo->bytes_may_use -= bytes;
1903                 BTRFS_I(inode)->reserved_bytes -= bytes;
1904         }
1905
1906         spin_unlock(&data_sinfo->lock);
1907 }
1908
1909 /* called when we are clearing an delalloc extent from the inode's io_tree */
1910 void btrfs_delalloc_free_space(struct btrfs_root *root, struct inode *inode,
1911                               u64 bytes)
1912 {
1913         struct btrfs_space_info *info;
1914
1915         info = BTRFS_I(inode)->space_info;
1916
1917         spin_lock(&info->lock);
1918         info->bytes_delalloc -= bytes;
1919         spin_unlock(&info->lock);
1920 }
1921
1922 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
1923                           struct btrfs_root *extent_root, u64 alloc_bytes,
1924                           u64 flags, int force)
1925 {
1926         struct btrfs_space_info *space_info;
1927         u64 thresh;
1928         int ret = 0;
1929
1930         mutex_lock(&extent_root->fs_info->chunk_mutex);
1931
1932         flags = btrfs_reduce_alloc_profile(extent_root, flags);
1933
1934         space_info = __find_space_info(extent_root->fs_info, flags);
1935         if (!space_info) {
1936                 ret = update_space_info(extent_root->fs_info, flags,
1937                                         0, 0, &space_info);
1938                 BUG_ON(ret);
1939         }
1940         BUG_ON(!space_info);
1941
1942         spin_lock(&space_info->lock);
1943         if (space_info->force_alloc) {
1944                 force = 1;
1945                 space_info->force_alloc = 0;
1946         }
1947         if (space_info->full) {
1948                 spin_unlock(&space_info->lock);
1949                 goto out;
1950         }
1951
1952         thresh = space_info->total_bytes - space_info->bytes_readonly;
1953         thresh = div_factor(thresh, 6);
1954         if (!force &&
1955            (space_info->bytes_used + space_info->bytes_pinned +
1956             space_info->bytes_reserved + alloc_bytes) < thresh) {
1957                 spin_unlock(&space_info->lock);
1958                 goto out;
1959         }
1960         spin_unlock(&space_info->lock);
1961
1962         ret = btrfs_alloc_chunk(trans, extent_root, flags);
1963         if (ret)
1964                 space_info->full = 1;
1965 out:
1966         mutex_unlock(&extent_root->fs_info->chunk_mutex);
1967         return ret;
1968 }
1969
1970 static int update_block_group(struct btrfs_trans_handle *trans,
1971                               struct btrfs_root *root,
1972                               u64 bytenr, u64 num_bytes, int alloc,
1973                               int mark_free)
1974 {
1975         struct btrfs_block_group_cache *cache;
1976         struct btrfs_fs_info *info = root->fs_info;
1977         u64 total = num_bytes;
1978         u64 old_val;
1979         u64 byte_in_group;
1980
1981         while (total) {
1982                 cache = btrfs_lookup_block_group(info, bytenr);
1983                 if (!cache)
1984                         return -1;
1985                 byte_in_group = bytenr - cache->key.objectid;
1986                 WARN_ON(byte_in_group > cache->key.offset);
1987
1988                 spin_lock(&cache->space_info->lock);
1989                 spin_lock(&cache->lock);
1990                 cache->dirty = 1;
1991                 old_val = btrfs_block_group_used(&cache->item);
1992                 num_bytes = min(total, cache->key.offset - byte_in_group);
1993                 if (alloc) {
1994                         old_val += num_bytes;
1995                         cache->space_info->bytes_used += num_bytes;
1996                         if (cache->ro)
1997                                 cache->space_info->bytes_readonly -= num_bytes;
1998                         btrfs_set_block_group_used(&cache->item, old_val);
1999                         spin_unlock(&cache->lock);
2000                         spin_unlock(&cache->space_info->lock);
2001                 } else {
2002                         old_val -= num_bytes;
2003                         cache->space_info->bytes_used -= num_bytes;
2004                         if (cache->ro)
2005                                 cache->space_info->bytes_readonly += num_bytes;
2006                         btrfs_set_block_group_used(&cache->item, old_val);
2007                         spin_unlock(&cache->lock);
2008                         spin_unlock(&cache->space_info->lock);
2009                         if (mark_free) {
2010                                 int ret;
2011
2012                                 ret = btrfs_discard_extent(root, bytenr,
2013                                                            num_bytes);
2014                                 WARN_ON(ret);
2015
2016                                 ret = btrfs_add_free_space(cache, bytenr,
2017                                                            num_bytes);
2018                                 WARN_ON(ret);
2019                         }
2020                 }
2021                 put_block_group(cache);
2022                 total -= num_bytes;
2023                 bytenr += num_bytes;
2024         }
2025         return 0;
2026 }
2027
2028 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
2029 {
2030         struct btrfs_block_group_cache *cache;
2031         u64 bytenr;
2032
2033         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
2034         if (!cache)
2035                 return 0;
2036
2037         bytenr = cache->key.objectid;
2038         put_block_group(cache);
2039
2040         return bytenr;
2041 }
2042
2043 int btrfs_update_pinned_extents(struct btrfs_root *root,
2044                                 u64 bytenr, u64 num, int pin)
2045 {
2046         u64 len;
2047         struct btrfs_block_group_cache *cache;
2048         struct btrfs_fs_info *fs_info = root->fs_info;
2049
2050         WARN_ON(!mutex_is_locked(&root->fs_info->pinned_mutex));
2051         if (pin) {
2052                 set_extent_dirty(&fs_info->pinned_extents,
2053                                 bytenr, bytenr + num - 1, GFP_NOFS);
2054         } else {
2055                 clear_extent_dirty(&fs_info->pinned_extents,
2056                                 bytenr, bytenr + num - 1, GFP_NOFS);
2057         }
2058         mutex_unlock(&root->fs_info->pinned_mutex);
2059
2060         while (num > 0) {
2061                 cache = btrfs_lookup_block_group(fs_info, bytenr);
2062                 BUG_ON(!cache);
2063                 len = min(num, cache->key.offset -
2064                           (bytenr - cache->key.objectid));
2065                 if (pin) {
2066                         spin_lock(&cache->space_info->lock);
2067                         spin_lock(&cache->lock);
2068                         cache->pinned += len;
2069                         cache->space_info->bytes_pinned += len;
2070                         spin_unlock(&cache->lock);
2071                         spin_unlock(&cache->space_info->lock);
2072                         fs_info->total_pinned += len;
2073                 } else {
2074                         spin_lock(&cache->space_info->lock);
2075                         spin_lock(&cache->lock);
2076                         cache->pinned -= len;
2077                         cache->space_info->bytes_pinned -= len;
2078                         spin_unlock(&cache->lock);
2079                         spin_unlock(&cache->space_info->lock);
2080                         fs_info->total_pinned -= len;
2081                         if (cache->cached)
2082                                 btrfs_add_free_space(cache, bytenr, len);
2083                 }
2084                 put_block_group(cache);
2085                 bytenr += len;
2086                 num -= len;
2087         }
2088         return 0;
2089 }
2090
2091 static int update_reserved_extents(struct btrfs_root *root,
2092                                    u64 bytenr, u64 num, int reserve)
2093 {
2094         u64 len;
2095         struct btrfs_block_group_cache *cache;
2096         struct btrfs_fs_info *fs_info = root->fs_info;
2097
2098         while (num > 0) {
2099                 cache = btrfs_lookup_block_group(fs_info, bytenr);
2100                 BUG_ON(!cache);
2101                 len = min(num, cache->key.offset -
2102                           (bytenr - cache->key.objectid));
2103
2104                 spin_lock(&cache->space_info->lock);
2105                 spin_lock(&cache->lock);
2106                 if (reserve) {
2107                         cache->reserved += len;
2108                         cache->space_info->bytes_reserved += len;
2109                 } else {
2110                         cache->reserved -= len;
2111                         cache->space_info->bytes_reserved -= len;
2112                 }
2113                 spin_unlock(&cache->lock);
2114                 spin_unlock(&cache->space_info->lock);
2115                 put_block_group(cache);
2116                 bytenr += len;
2117                 num -= len;
2118         }
2119         return 0;
2120 }
2121
2122 int btrfs_copy_pinned(struct btrfs_root *root, struct extent_io_tree *copy)
2123 {
2124         u64 last = 0;
2125         u64 start;
2126         u64 end;
2127         struct extent_io_tree *pinned_extents = &root->fs_info->pinned_extents;
2128         int ret;
2129
2130         mutex_lock(&root->fs_info->pinned_mutex);
2131         while (1) {
2132                 ret = find_first_extent_bit(pinned_extents, last,
2133                                             &start, &end, EXTENT_DIRTY);
2134                 if (ret)
2135                         break;
2136                 set_extent_dirty(copy, start, end, GFP_NOFS);
2137                 last = end + 1;
2138         }
2139         mutex_unlock(&root->fs_info->pinned_mutex);
2140         return 0;
2141 }
2142
2143 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
2144                                struct btrfs_root *root,
2145                                struct extent_io_tree *unpin)
2146 {
2147         u64 start;
2148         u64 end;
2149         int ret;
2150
2151         while (1) {
2152                 mutex_lock(&root->fs_info->pinned_mutex);
2153                 ret = find_first_extent_bit(unpin, 0, &start, &end,
2154                                             EXTENT_DIRTY);
2155                 if (ret)
2156                         break;
2157
2158                 ret = btrfs_discard_extent(root, start, end + 1 - start);
2159
2160                 /* unlocks the pinned mutex */
2161                 btrfs_update_pinned_extents(root, start, end + 1 - start, 0);
2162                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
2163
2164                 cond_resched();
2165         }
2166         mutex_unlock(&root->fs_info->pinned_mutex);
2167         return ret;
2168 }
2169
2170 static int pin_down_bytes(struct btrfs_trans_handle *trans,
2171                           struct btrfs_root *root,
2172                           struct btrfs_path *path,
2173                           u64 bytenr, u64 num_bytes, int is_data,
2174                           struct extent_buffer **must_clean)
2175 {
2176         int err = 0;
2177         struct extent_buffer *buf;
2178
2179         if (is_data)
2180                 goto pinit;
2181
2182         buf = btrfs_find_tree_block(root, bytenr, num_bytes);
2183         if (!buf)
2184                 goto pinit;
2185
2186         /* we can reuse a block if it hasn't been written
2187          * and it is from this transaction.  We can't
2188          * reuse anything from the tree log root because
2189          * it has tiny sub-transactions.
2190          */
2191         if (btrfs_buffer_uptodate(buf, 0) &&
2192             btrfs_try_tree_lock(buf)) {
2193                 u64 header_owner = btrfs_header_owner(buf);
2194                 u64 header_transid = btrfs_header_generation(buf);
2195                 if (header_owner != BTRFS_TREE_LOG_OBJECTID &&
2196                     header_owner != BTRFS_TREE_RELOC_OBJECTID &&
2197                     header_owner != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2198                     header_transid == trans->transid &&
2199                     !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
2200                         *must_clean = buf;
2201                         return 1;
2202                 }
2203                 btrfs_tree_unlock(buf);
2204         }
2205         free_extent_buffer(buf);
2206 pinit:
2207         btrfs_set_path_blocking(path);
2208         mutex_lock(&root->fs_info->pinned_mutex);
2209         /* unlocks the pinned mutex */
2210         btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
2211
2212         BUG_ON(err < 0);
2213         return 0;
2214 }
2215
2216 /*
2217  * remove an extent from the root, returns 0 on success
2218  */
2219 static int __free_extent(struct btrfs_trans_handle *trans,
2220                          struct btrfs_root *root,
2221                          u64 bytenr, u64 num_bytes, u64 parent,
2222                          u64 root_objectid, u64 ref_generation,
2223                          u64 owner_objectid, int pin, int mark_free,
2224                          int refs_to_drop)
2225 {
2226         struct btrfs_path *path;
2227         struct btrfs_key key;
2228         struct btrfs_fs_info *info = root->fs_info;
2229         struct btrfs_root *extent_root = info->extent_root;
2230         struct extent_buffer *leaf;
2231         int ret;
2232         int extent_slot = 0;
2233         int found_extent = 0;
2234         int num_to_del = 1;
2235         struct btrfs_extent_item *ei;
2236         u32 refs;
2237
2238         key.objectid = bytenr;
2239         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
2240         key.offset = num_bytes;
2241         path = btrfs_alloc_path();
2242         if (!path)
2243                 return -ENOMEM;
2244
2245         path->reada = 1;
2246         path->leave_spinning = 1;
2247         ret = lookup_extent_backref(trans, extent_root, path,
2248                                     bytenr, parent, root_objectid,
2249                                     ref_generation, owner_objectid, 1);
2250         if (ret == 0) {
2251                 struct btrfs_key found_key;
2252                 extent_slot = path->slots[0];
2253                 while (extent_slot > 0) {
2254                         extent_slot--;
2255                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2256                                               extent_slot);
2257                         if (found_key.objectid != bytenr)
2258                                 break;
2259                         if (found_key.type == BTRFS_EXTENT_ITEM_KEY &&
2260                             found_key.offset == num_bytes) {
2261                                 found_extent = 1;
2262                                 break;
2263                         }
2264                         if (path->slots[0] - extent_slot > 5)
2265                                 break;
2266                 }
2267                 if (!found_extent) {
2268                         ret = remove_extent_backref(trans, extent_root, path,
2269                                                     refs_to_drop);
2270                         BUG_ON(ret);
2271                         btrfs_release_path(extent_root, path);
2272                         path->leave_spinning = 1;
2273                         ret = btrfs_search_slot(trans, extent_root,
2274                                                 &key, path, -1, 1);
2275                         if (ret) {
2276                                 printk(KERN_ERR "umm, got %d back from search"
2277                                        ", was looking for %llu\n", ret,
2278                                        (unsigned long long)bytenr);
2279                                 btrfs_print_leaf(extent_root, path->nodes[0]);
2280                         }
2281                         BUG_ON(ret);
2282                         extent_slot = path->slots[0];
2283                 }
2284         } else {
2285                 btrfs_print_leaf(extent_root, path->nodes[0]);
2286                 WARN_ON(1);
2287                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
2288                        "parent %llu root %llu gen %llu owner %llu\n",
2289                        (unsigned long long)bytenr,
2290                        (unsigned long long)parent,
2291                        (unsigned long long)root_objectid,
2292                        (unsigned long long)ref_generation,
2293                        (unsigned long long)owner_objectid);
2294         }
2295
2296         leaf = path->nodes[0];
2297         ei = btrfs_item_ptr(leaf, extent_slot,
2298                             struct btrfs_extent_item);
2299         refs = btrfs_extent_refs(leaf, ei);
2300
2301         /*
2302          * we're not allowed to delete the extent item if there
2303          * are other delayed ref updates pending
2304          */
2305
2306         BUG_ON(refs < refs_to_drop);
2307         refs -= refs_to_drop;
2308         btrfs_set_extent_refs(leaf, ei, refs);
2309         btrfs_mark_buffer_dirty(leaf);
2310
2311         if (refs == 0 && found_extent &&
2312             path->slots[0] == extent_slot + 1) {
2313                 struct btrfs_extent_ref *ref;
2314                 ref = btrfs_item_ptr(leaf, path->slots[0],
2315                                      struct btrfs_extent_ref);
2316                 BUG_ON(btrfs_ref_num_refs(leaf, ref) != refs_to_drop);
2317                 /* if the back ref and the extent are next to each other
2318                  * they get deleted below in one shot
2319                  */
2320                 path->slots[0] = extent_slot;
2321                 num_to_del = 2;
2322         } else if (found_extent) {
2323                 /* otherwise delete the extent back ref */
2324                 ret = remove_extent_backref(trans, extent_root, path,
2325                                             refs_to_drop);
2326                 BUG_ON(ret);
2327                 /* if refs are 0, we need to setup the path for deletion */
2328                 if (refs == 0) {
2329                         btrfs_release_path(extent_root, path);
2330                         path->leave_spinning = 1;
2331                         ret = btrfs_search_slot(trans, extent_root, &key, path,
2332                                                 -1, 1);
2333                         BUG_ON(ret);
2334                 }
2335         }
2336
2337         if (refs == 0) {
2338                 u64 super_used;
2339                 u64 root_used;
2340                 struct extent_buffer *must_clean = NULL;
2341
2342                 if (pin) {
2343                         ret = pin_down_bytes(trans, root, path,
2344                                 bytenr, num_bytes,
2345                                 owner_objectid >= BTRFS_FIRST_FREE_OBJECTID,
2346                                 &must_clean);
2347                         if (ret > 0)
2348                                 mark_free = 1;
2349                         BUG_ON(ret < 0);
2350                 }
2351
2352                 /* block accounting for super block */
2353                 spin_lock(&info->delalloc_lock);
2354                 super_used = btrfs_super_bytes_used(&info->super_copy);
2355                 btrfs_set_super_bytes_used(&info->super_copy,
2356                                            super_used - num_bytes);
2357
2358                 /* block accounting for root item */
2359                 root_used = btrfs_root_used(&root->root_item);
2360                 btrfs_set_root_used(&root->root_item,
2361                                            root_used - num_bytes);
2362                 spin_unlock(&info->delalloc_lock);
2363
2364                 /*
2365                  * it is going to be very rare for someone to be waiting
2366                  * on the block we're freeing.  del_items might need to
2367                  * schedule, so rather than get fancy, just force it
2368                  * to blocking here
2369                  */
2370                 if (must_clean)
2371                         btrfs_set_lock_blocking(must_clean);
2372
2373                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
2374                                       num_to_del);
2375                 BUG_ON(ret);
2376                 btrfs_release_path(extent_root, path);
2377
2378                 if (must_clean) {
2379                         clean_tree_block(NULL, root, must_clean);
2380                         btrfs_tree_unlock(must_clean);
2381                         free_extent_buffer(must_clean);
2382                 }
2383
2384                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
2385                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
2386                         BUG_ON(ret);
2387                 }
2388
2389                 ret = update_block_group(trans, root, bytenr, num_bytes, 0,
2390                                          mark_free);
2391                 BUG_ON(ret);
2392         }
2393         btrfs_free_path(path);
2394         return ret;
2395 }
2396
2397 /*
2398  * remove an extent from the root, returns 0 on success
2399  */
2400 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
2401                                         struct btrfs_root *root,
2402                                         u64 bytenr, u64 num_bytes, u64 parent,
2403                                         u64 root_objectid, u64 ref_generation,
2404                                         u64 owner_objectid, int pin,
2405                                         int refs_to_drop)
2406 {
2407         WARN_ON(num_bytes < root->sectorsize);
2408
2409         /*
2410          * if metadata always pin
2411          * if data pin when any transaction has committed this
2412          */
2413         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID ||
2414             ref_generation != trans->transid)
2415                 pin = 1;
2416
2417         if (ref_generation != trans->transid)
2418                 pin = 1;
2419
2420         return __free_extent(trans, root, bytenr, num_bytes, parent,
2421                             root_objectid, ref_generation,
2422                             owner_objectid, pin, pin == 0, refs_to_drop);
2423 }
2424
2425 /*
2426  * when we free an extent, it is possible (and likely) that we free the last
2427  * delayed ref for that extent as well.  This searches the delayed ref tree for
2428  * a given extent, and if there are no other delayed refs to be processed, it
2429  * removes it from the tree.
2430  */
2431 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
2432                                       struct btrfs_root *root, u64 bytenr)
2433 {
2434         struct btrfs_delayed_ref_head *head;
2435         struct btrfs_delayed_ref_root *delayed_refs;
2436         struct btrfs_delayed_ref_node *ref;
2437         struct rb_node *node;
2438         int ret;
2439
2440         delayed_refs = &trans->transaction->delayed_refs;
2441         spin_lock(&delayed_refs->lock);
2442         head = btrfs_find_delayed_ref_head(trans, bytenr);
2443         if (!head)
2444                 goto out;
2445
2446         node = rb_prev(&head->node.rb_node);
2447         if (!node)
2448                 goto out;
2449
2450         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2451
2452         /* there are still entries for this ref, we can't drop it */
2453         if (ref->bytenr == bytenr)
2454                 goto out;
2455
2456         /*
2457          * waiting for the lock here would deadlock.  If someone else has it
2458          * locked they are already in the process of dropping it anyway
2459          */
2460         if (!mutex_trylock(&head->mutex))
2461                 goto out;
2462
2463         /*
2464          * at this point we have a head with no other entries.  Go
2465          * ahead and process it.
2466          */
2467         head->node.in_tree = 0;
2468         rb_erase(&head->node.rb_node, &delayed_refs->root);
2469
2470         delayed_refs->num_entries--;
2471
2472         /*
2473          * we don't take a ref on the node because we're removing it from the
2474          * tree, so we just steal the ref the tree was holding.
2475          */
2476         delayed_refs->num_heads--;
2477         if (list_empty(&head->cluster))
2478                 delayed_refs->num_heads_ready--;
2479
2480         list_del_init(&head->cluster);
2481         spin_unlock(&delayed_refs->lock);
2482
2483         ret = run_one_delayed_ref(trans, root->fs_info->tree_root,
2484                                   &head->node, head->must_insert_reserved);
2485         BUG_ON(ret);
2486         btrfs_put_delayed_ref(&head->node);
2487         return 0;
2488 out:
2489         spin_unlock(&delayed_refs->lock);
2490         return 0;
2491 }
2492
2493 int btrfs_free_extent(struct btrfs_trans_handle *trans,
2494                       struct btrfs_root *root,
2495                       u64 bytenr, u64 num_bytes, u64 parent,
2496                       u64 root_objectid, u64 ref_generation,
2497                       u64 owner_objectid, int pin)
2498 {
2499         int ret;
2500
2501         /*
2502          * tree log blocks never actually go into the extent allocation
2503          * tree, just update pinning info and exit early.
2504          *
2505          * data extents referenced by the tree log do need to have
2506          * their reference counts bumped.
2507          */
2508         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID &&
2509             owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
2510                 mutex_lock(&root->fs_info->pinned_mutex);
2511
2512                 /* unlocks the pinned mutex */
2513                 btrfs_update_pinned_extents(root, bytenr, num_bytes, 1);
2514                 update_reserved_extents(root, bytenr, num_bytes, 0);
2515                 ret = 0;
2516         } else {
2517                 ret = btrfs_add_delayed_ref(trans, bytenr, num_bytes, parent,
2518                                        root_objectid, ref_generation,
2519                                        owner_objectid,
2520                                        BTRFS_DROP_DELAYED_REF, 1);
2521                 BUG_ON(ret);
2522                 ret = check_ref_cleanup(trans, root, bytenr);
2523                 BUG_ON(ret);
2524         }
2525         return ret;
2526 }
2527
2528 static u64 stripe_align(struct btrfs_root *root, u64 val)
2529 {
2530         u64 mask = ((u64)root->stripesize - 1);
2531         u64 ret = (val + mask) & ~mask;
2532         return ret;
2533 }
2534
2535 /*
2536  * walks the btree of allocated extents and find a hole of a given size.
2537  * The key ins is changed to record the hole:
2538  * ins->objectid == block start
2539  * ins->flags = BTRFS_EXTENT_ITEM_KEY
2540  * ins->offset == number of blocks
2541  * Any available blocks before search_start are skipped.
2542  */
2543 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
2544                                      struct btrfs_root *orig_root,
2545                                      u64 num_bytes, u64 empty_size,
2546                                      u64 search_start, u64 search_end,
2547                                      u64 hint_byte, struct btrfs_key *ins,
2548                                      u64 exclude_start, u64 exclude_nr,
2549                                      int data)
2550 {
2551         int ret = 0;
2552         struct btrfs_root *root = orig_root->fs_info->extent_root;
2553         u64 total_needed = num_bytes;
2554         u64 *last_ptr = NULL;
2555         u64 last_wanted = 0;
2556         struct btrfs_block_group_cache *block_group = NULL;
2557         int chunk_alloc_done = 0;
2558         int empty_cluster = 2 * 1024 * 1024;
2559         int allowed_chunk_alloc = 0;
2560         struct list_head *head = NULL, *cur = NULL;
2561         int loop = 0;
2562         int extra_loop = 0;
2563         struct btrfs_space_info *space_info;
2564
2565         WARN_ON(num_bytes < root->sectorsize);
2566         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
2567         ins->objectid = 0;
2568         ins->offset = 0;
2569
2570         if (orig_root->ref_cows || empty_size)
2571                 allowed_chunk_alloc = 1;
2572
2573         if (data & BTRFS_BLOCK_GROUP_METADATA) {
2574                 last_ptr = &root->fs_info->last_alloc;
2575                 if (!btrfs_test_opt(root, SSD))
2576                         empty_cluster = 64 * 1024;
2577         }
2578
2579         if ((data & BTRFS_BLOCK_GROUP_DATA) && btrfs_test_opt(root, SSD))
2580                 last_ptr = &root->fs_info->last_data_alloc;
2581
2582         if (last_ptr) {
2583                 if (*last_ptr) {
2584                         hint_byte = *last_ptr;
2585                         last_wanted = *last_ptr;
2586                 } else
2587                         empty_size += empty_cluster;
2588         } else {
2589                 empty_cluster = 0;
2590         }
2591         search_start = max(search_start, first_logical_byte(root, 0));
2592         search_start = max(search_start, hint_byte);
2593
2594         if (last_wanted && search_start != last_wanted) {
2595                 last_wanted = 0;
2596                 empty_size += empty_cluster;
2597         }
2598
2599         total_needed += empty_size;
2600         block_group = btrfs_lookup_block_group(root->fs_info, search_start);
2601         if (!block_group)
2602                 block_group = btrfs_lookup_first_block_group(root->fs_info,
2603                                                              search_start);
2604         space_info = __find_space_info(root->fs_info, data);
2605
2606         down_read(&space_info->groups_sem);
2607         while (1) {
2608                 struct btrfs_free_space *free_space;
2609                 /*
2610                  * the only way this happens if our hint points to a block
2611                  * group thats not of the proper type, while looping this
2612                  * should never happen
2613                  */
2614                 if (empty_size)
2615                         extra_loop = 1;
2616
2617                 if (!block_group)
2618                         goto new_group_no_lock;
2619
2620                 if (unlikely(!block_group->cached)) {
2621                         mutex_lock(&block_group->cache_mutex);
2622                         ret = cache_block_group(root, block_group);
2623                         mutex_unlock(&block_group->cache_mutex);
2624                         if (ret)
2625                                 break;
2626                 }
2627
2628                 mutex_lock(&block_group->alloc_mutex);
2629                 if (unlikely(!block_group_bits(block_group, data)))
2630                         goto new_group;
2631
2632                 if (unlikely(block_group->ro))
2633                         goto new_group;
2634
2635                 free_space = btrfs_find_free_space(block_group, search_start,
2636                                                    total_needed);
2637                 if (free_space) {
2638                         u64 start = block_group->key.objectid;
2639                         u64 end = block_group->key.objectid +
2640                                 block_group->key.offset;
2641
2642                         search_start = stripe_align(root, free_space->offset);
2643
2644                         /* move on to the next group */
2645                         if (search_start + num_bytes >= search_end)
2646                                 goto new_group;
2647
2648                         /* move on to the next group */
2649                         if (search_start + num_bytes > end)
2650                                 goto new_group;
2651
2652                         if (last_wanted && search_start != last_wanted) {
2653                                 total_needed += empty_cluster;
2654                                 empty_size += empty_cluster;
2655                                 last_wanted = 0;
2656                                 /*
2657                                  * if search_start is still in this block group
2658                                  * then we just re-search this block group
2659                                  */
2660                                 if (search_start >= start &&
2661                                     search_start < end) {
2662                                         mutex_unlock(&block_group->alloc_mutex);
2663                                         continue;
2664                                 }
2665
2666                                 /* else we go to the next block group */
2667                                 goto new_group;
2668                         }
2669
2670                         if (exclude_nr > 0 &&
2671                             (search_start + num_bytes > exclude_start &&
2672                              search_start < exclude_start + exclude_nr)) {
2673                                 search_start = exclude_start + exclude_nr;
2674                                 /*
2675                                  * if search_start is still in this block group
2676                                  * then we just re-search this block group
2677                                  */
2678                                 if (search_start >= start &&
2679                                     search_start < end) {
2680                                         mutex_unlock(&block_group->alloc_mutex);
2681                                         last_wanted = 0;
2682                                         continue;
2683                                 }
2684
2685                                 /* else we go to the next block group */
2686                                 goto new_group;
2687                         }
2688
2689                         ins->objectid = search_start;
2690                         ins->offset = num_bytes;
2691
2692                         btrfs_remove_free_space_lock(block_group, search_start,
2693                                                      num_bytes);
2694                         /* we are all good, lets return */
2695                         mutex_unlock(&block_group->alloc_mutex);
2696                         break;
2697                 }
2698 new_group:
2699                 mutex_unlock(&block_group->alloc_mutex);
2700                 put_block_group(block_group);
2701                 block_group = NULL;
2702 new_group_no_lock:
2703                 /* don't try to compare new allocations against the
2704                  * last allocation any more
2705                  */
2706                 last_wanted = 0;
2707
2708                 /*
2709                  * Here's how this works.
2710                  * loop == 0: we were searching a block group via a hint
2711                  *              and didn't find anything, so we start at
2712                  *              the head of the block groups and keep searching
2713                  * loop == 1: we're searching through all of the block groups
2714                  *              if we hit the head again we have searched
2715                  *              all of the block groups for this space and we
2716                  *              need to try and allocate, if we cant error out.
2717                  * loop == 2: we allocated more space and are looping through
2718                  *              all of the block groups again.
2719                  */
2720                 if (loop == 0) {
2721                         head = &space_info->block_groups;
2722                         cur = head->next;
2723                         loop++;
2724                 } else if (loop == 1 && cur == head) {
2725                         int keep_going;
2726
2727                         /* at this point we give up on the empty_size
2728                          * allocations and just try to allocate the min
2729                          * space.
2730                          *
2731                          * The extra_loop field was set if an empty_size
2732                          * allocation was attempted above, and if this
2733                          * is try we need to try the loop again without
2734                          * the additional empty_size.
2735                          */
2736                         total_needed -= empty_size;
2737                         empty_size = 0;
2738                         keep_going = extra_loop;
2739                         loop++;
2740
2741                         if (allowed_chunk_alloc && !chunk_alloc_done) {
2742                                 up_read(&space_info->groups_sem);
2743                                 ret = do_chunk_alloc(trans, root, num_bytes +
2744                                                      2 * 1024 * 1024, data, 1);
2745                                 down_read(&space_info->groups_sem);
2746                                 if (ret < 0)
2747                                         goto loop_check;
2748                                 head = &space_info->block_groups;
2749                                 /*
2750                                  * we've allocated a new chunk, keep
2751                                  * trying
2752                                  */
2753                                 keep_going = 1;
2754                                 chunk_alloc_done = 1;
2755                         } else if (!allowed_chunk_alloc) {
2756                                 space_info->force_alloc = 1;
2757                         }
2758 loop_check:
2759                         if (keep_going) {
2760                                 cur = head->next;
2761                                 extra_loop = 0;
2762                         } else {
2763                                 break;
2764                         }
2765                 } else if (cur == head) {
2766                         break;
2767                 }
2768
2769                 block_group = list_entry(cur, struct btrfs_block_group_cache,
2770                                          list);
2771                 atomic_inc(&block_group->count);
2772
2773                 search_start = block_group->key.objectid;
2774                 cur = cur->next;
2775         }
2776
2777         /* we found what we needed */
2778         if (ins->objectid) {
2779                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
2780                         trans->block_group = block_group->key.objectid;
2781
2782                 if (last_ptr)
2783                         *last_ptr = ins->objectid + ins->offset;
2784                 ret = 0;
2785         } else if (!ret) {
2786                 printk(KERN_ERR "btrfs searching for %llu bytes, "
2787                        "num_bytes %llu, loop %d, allowed_alloc %d\n",
2788                        (unsigned long long)total_needed,
2789                        (unsigned long long)num_bytes,
2790                        loop, allowed_chunk_alloc);
2791                 ret = -ENOSPC;
2792         }
2793         if (block_group)
2794                 put_block_group(block_group);
2795
2796         up_read(&space_info->groups_sem);
2797         return ret;
2798 }
2799
2800 static void dump_space_info(struct btrfs_space_info *info, u64 bytes)
2801 {
2802         struct btrfs_block_group_cache *cache;
2803
2804         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
2805                (unsigned long long)(info->total_bytes - info->bytes_used -
2806                                     info->bytes_pinned - info->bytes_reserved),
2807                (info->full) ? "" : "not ");
2808         printk(KERN_INFO "space_info total=%llu, pinned=%llu, delalloc=%llu,"
2809                " may_use=%llu, used=%llu\n", info->total_bytes,
2810                info->bytes_pinned, info->bytes_delalloc, info->bytes_may_use,
2811                info->bytes_used);
2812
2813         down_read(&info->groups_sem);
2814         list_for_each_entry(cache, &info->block_groups, list) {
2815                 spin_lock(&cache->lock);
2816                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
2817                        "%llu pinned %llu reserved\n",
2818                        (unsigned long long)cache->key.objectid,
2819                        (unsigned long long)cache->key.offset,
2820                        (unsigned long long)btrfs_block_group_used(&cache->item),
2821                        (unsigned long long)cache->pinned,
2822                        (unsigned long long)cache->reserved);
2823                 btrfs_dump_free_space(cache, bytes);
2824                 spin_unlock(&cache->lock);
2825         }
2826         up_read(&info->groups_sem);
2827 }
2828
2829 static int __btrfs_reserve_extent(struct btrfs_trans_handle *trans,
2830                                   struct btrfs_root *root,
2831                                   u64 num_bytes, u64 min_alloc_size,
2832                                   u64 empty_size, u64 hint_byte,
2833                                   u64 search_end, struct btrfs_key *ins,
2834                                   u64 data)
2835 {
2836         int ret;
2837         u64 search_start = 0;
2838         struct btrfs_fs_info *info = root->fs_info;
2839
2840         data = btrfs_get_alloc_profile(root, data);
2841 again:
2842         /*
2843          * the only place that sets empty_size is btrfs_realloc_node, which
2844          * is not called recursively on allocations
2845          */
2846         if (empty_size || root->ref_cows) {
2847                 if (!(data & BTRFS_BLOCK_GROUP_METADATA)) {
2848                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
2849                                      2 * 1024 * 1024,
2850                                      BTRFS_BLOCK_GROUP_METADATA |
2851                                      (info->metadata_alloc_profile &
2852                                       info->avail_metadata_alloc_bits), 0);
2853                 }
2854                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
2855                                      num_bytes + 2 * 1024 * 1024, data, 0);
2856         }
2857
2858         WARN_ON(num_bytes < root->sectorsize);
2859         ret = find_free_extent(trans, root, num_bytes, empty_size,
2860                                search_start, search_end, hint_byte, ins,
2861                                trans->alloc_exclude_start,
2862                                trans->alloc_exclude_nr, data);
2863
2864         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
2865                 num_bytes = num_bytes >> 1;
2866                 num_bytes = num_bytes & ~(root->sectorsize - 1);
2867                 num_bytes = max(num_bytes, min_alloc_size);
2868                 do_chunk_alloc(trans, root->fs_info->extent_root,
2869                                num_bytes, data, 1);
2870                 goto again;
2871         }
2872         if (ret) {
2873                 struct btrfs_space_info *sinfo;
2874
2875                 sinfo = __find_space_info(root->fs_info, data);
2876                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
2877                        "wanted %llu\n", (unsigned long long)data,
2878                        (unsigned long long)num_bytes);
2879                 dump_space_info(sinfo, num_bytes);
2880                 BUG();
2881         }
2882
2883         return ret;
2884 }
2885
2886 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
2887 {
2888         struct btrfs_block_group_cache *cache;
2889         int ret = 0;
2890
2891         cache = btrfs_lookup_block_group(root->fs_info, start);
2892         if (!cache) {
2893                 printk(KERN_ERR "Unable to find block group for %llu\n",
2894                        (unsigned long long)start);
2895                 return -ENOSPC;
2896         }
2897
2898         ret = btrfs_discard_extent(root, start, len);
2899
2900         btrfs_add_free_space(cache, start, len);
2901         put_block_group(cache);
2902         update_reserved_extents(root, start, len, 0);
2903
2904         return ret;
2905 }
2906
2907 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
2908                                   struct btrfs_root *root,
2909                                   u64 num_bytes, u64 min_alloc_size,
2910                                   u64 empty_size, u64 hint_byte,
2911                                   u64 search_end, struct btrfs_key *ins,
2912                                   u64 data)
2913 {
2914         int ret;
2915         ret = __btrfs_reserve_extent(trans, root, num_bytes, min_alloc_size,
2916                                      empty_size, hint_byte, search_end, ins,
2917                                      data);
2918         update_reserved_extents(root, ins->objectid, ins->offset, 1);
2919         return ret;
2920 }
2921
2922 static int __btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
2923                                          struct btrfs_root *root, u64 parent,
2924                                          u64 root_objectid, u64 ref_generation,
2925                                          u64 owner, struct btrfs_key *ins,
2926                                          int ref_mod)
2927 {
2928         int ret;
2929         u64 super_used;
2930         u64 root_used;
2931         u64 num_bytes = ins->offset;
2932         u32 sizes[2];
2933         struct btrfs_fs_info *info = root->fs_info;
2934         struct btrfs_root *extent_root = info->extent_root;
2935         struct btrfs_extent_item *extent_item;
2936         struct btrfs_extent_ref *ref;
2937         struct btrfs_path *path;
2938         struct btrfs_key keys[2];
2939
2940         if (parent == 0)
2941                 parent = ins->objectid;
2942
2943         /* block accounting for super block */
2944         spin_lock(&info->delalloc_lock);
2945         super_used = btrfs_super_bytes_used(&info->super_copy);
2946         btrfs_set_super_bytes_used(&info->super_copy, super_used + num_bytes);
2947
2948         /* block accounting for root item */
2949         root_used = btrfs_root_used(&root->root_item);
2950         btrfs_set_root_used(&root->root_item, root_used + num_bytes);
2951         spin_unlock(&info->delalloc_lock);
2952
2953         memcpy(&keys[0], ins, sizeof(*ins));
2954         keys[1].objectid = ins->objectid;
2955         keys[1].type = BTRFS_EXTENT_REF_KEY;
2956         keys[1].offset = parent;
2957         sizes[0] = sizeof(*extent_item);
2958         sizes[1] = sizeof(*ref);
2959
2960         path = btrfs_alloc_path();
2961         BUG_ON(!path);
2962
2963         path->leave_spinning = 1;
2964         ret = btrfs_insert_empty_items(trans, extent_root, path, keys,
2965                                        sizes, 2);
2966         BUG_ON(ret);
2967
2968         extent_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2969                                      struct btrfs_extent_item);
2970         btrfs_set_extent_refs(path->nodes[0], extent_item, ref_mod);
2971         ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
2972                              struct btrfs_extent_ref);
2973
2974         btrfs_set_ref_root(path->nodes[0], ref, root_objectid);
2975         btrfs_set_ref_generation(path->nodes[0], ref, ref_generation);
2976         btrfs_set_ref_objectid(path->nodes[0], ref, owner);
2977         btrfs_set_ref_num_refs(path->nodes[0], ref, ref_mod);
2978
2979         btrfs_mark_buffer_dirty(path->nodes[0]);
2980
2981         trans->alloc_exclude_start = 0;
2982         trans->alloc_exclude_nr = 0;
2983         btrfs_free_path(path);
2984
2985         if (ret)
2986                 goto out;
2987
2988         ret = update_block_group(trans, root, ins->objectid,
2989                                  ins->offset, 1, 0);
2990         if (ret) {
2991                 printk(KERN_ERR "btrfs update block group failed for %llu "
2992                        "%llu\n", (unsigned long long)ins->objectid,
2993                        (unsigned long long)ins->offset);
2994                 BUG();
2995         }
2996 out:
2997         return ret;
2998 }
2999
3000 int btrfs_alloc_reserved_extent(struct btrfs_trans_handle *trans,
3001                                 struct btrfs_root *root, u64 parent,
3002                                 u64 root_objectid, u64 ref_generation,
3003                                 u64 owner, struct btrfs_key *ins)
3004 {
3005         int ret;
3006
3007         if (root_objectid == BTRFS_TREE_LOG_OBJECTID)
3008                 return 0;
3009
3010         ret = btrfs_add_delayed_ref(trans, ins->objectid,
3011                                     ins->offset, parent, root_objectid,
3012                                     ref_generation, owner,
3013                                     BTRFS_ADD_DELAYED_EXTENT, 0);
3014         BUG_ON(ret);
3015         return ret;
3016 }
3017
3018 /*
3019  * this is used by the tree logging recovery code.  It records that
3020  * an extent has been allocated and makes sure to clear the free
3021  * space cache bits as well
3022  */
3023 int btrfs_alloc_logged_extent(struct btrfs_trans_handle *trans,
3024                                 struct btrfs_root *root, u64 parent,
3025                                 u64 root_objectid, u64 ref_generation,
3026                                 u64 owner, struct btrfs_key *ins)
3027 {
3028         int ret;
3029         struct btrfs_block_group_cache *block_group;
3030
3031         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
3032         mutex_lock(&block_group->cache_mutex);
3033         cache_block_group(root, block_group);
3034         mutex_unlock(&block_group->cache_mutex);
3035
3036         ret = btrfs_remove_free_space(block_group, ins->objectid,
3037                                       ins->offset);
3038         BUG_ON(ret);
3039         put_block_group(block_group);
3040         ret = __btrfs_alloc_reserved_extent(trans, root, parent, root_objectid,
3041                                             ref_generation, owner, ins, 1);
3042         return ret;
3043 }
3044
3045 /*
3046  * finds a free extent and does all the dirty work required for allocation
3047  * returns the key for the extent through ins, and a tree buffer for
3048  * the first block of the extent through buf.
3049  *
3050  * returns 0 if everything worked, non-zero otherwise.
3051  */
3052 int btrfs_alloc_extent(struct btrfs_trans_handle *trans,
3053                        struct btrfs_root *root,
3054                        u64 num_bytes, u64 parent, u64 min_alloc_size,
3055                        u64 root_objectid, u64 ref_generation,
3056                        u64 owner_objectid, u64 empty_size, u64 hint_byte,
3057                        u64 search_end, struct btrfs_key *ins, u64 data)
3058 {
3059         int ret;
3060         ret = __btrfs_reserve_extent(trans, root, num_bytes,
3061                                      min_alloc_size, empty_size, hint_byte,
3062                                      search_end, ins, data);
3063         BUG_ON(ret);
3064         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
3065                 ret = btrfs_add_delayed_ref(trans, ins->objectid,
3066                                             ins->offset, parent, root_objectid,
3067                                             ref_generation, owner_objectid,
3068                                             BTRFS_ADD_DELAYED_EXTENT, 0);
3069                 BUG_ON(ret);
3070         }
3071         update_reserved_extents(root, ins->objectid, ins->offset, 1);
3072         return ret;
3073 }
3074
3075 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
3076                                             struct btrfs_root *root,
3077                                             u64 bytenr, u32 blocksize,
3078                                             int level)
3079 {
3080         struct extent_buffer *buf;
3081
3082         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
3083         if (!buf)
3084                 return ERR_PTR(-ENOMEM);
3085         btrfs_set_header_generation(buf, trans->transid);
3086         btrfs_set_buffer_lockdep_class(buf, level);
3087         btrfs_tree_lock(buf);
3088         clean_tree_block(trans, root, buf);
3089
3090         btrfs_set_lock_blocking(buf);
3091         btrfs_set_buffer_uptodate(buf);
3092
3093         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
3094                 set_extent_dirty(&root->dirty_log_pages, buf->start,
3095                          buf->start + buf->len - 1, GFP_NOFS);
3096         } else {
3097                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
3098                          buf->start + buf->len - 1, GFP_NOFS);
3099         }
3100         trans->blocks_used++;
3101         /* this returns a buffer locked for blocking */
3102         return buf;
3103 }
3104
3105 /*
3106  * helper function to allocate a block for a given tree
3107  * returns the tree buffer or NULL.
3108  */
3109 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
3110                                              struct btrfs_root *root,
3111                                              u32 blocksize, u64 parent,
3112                                              u64 root_objectid,
3113                                              u64 ref_generation,
3114                                              int level,
3115                                              u64 hint,
3116                                              u64 empty_size)
3117 {
3118         struct btrfs_key ins;
3119         int ret;
3120         struct extent_buffer *buf;
3121
3122         ret = btrfs_alloc_extent(trans, root, blocksize, parent, blocksize,
3123                                  root_objectid, ref_generation, level,
3124                                  empty_size, hint, (u64)-1, &ins, 0);
3125         if (ret) {
3126                 BUG_ON(ret > 0);
3127                 return ERR_PTR(ret);
3128         }
3129
3130         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
3131                                     blocksize, level);
3132         return buf;
3133 }
3134
3135 int btrfs_drop_leaf_ref(struct btrfs_trans_handle *trans,
3136                         struct btrfs_root *root, struct extent_buffer *leaf)
3137 {
3138         u64 leaf_owner;
3139         u64 leaf_generation;
3140         struct refsort *sorted;
3141         struct btrfs_key key;
3142         struct btrfs_file_extent_item *fi;
3143         int i;
3144         int nritems;
3145         int ret;
3146         int refi = 0;
3147         int slot;
3148
3149         BUG_ON(!btrfs_is_leaf(leaf));
3150         nritems = btrfs_header_nritems(leaf);
3151         leaf_owner = btrfs_header_owner(leaf);
3152         leaf_generation = btrfs_header_generation(leaf);
3153
3154         sorted = kmalloc(sizeof(*sorted) * nritems, GFP_NOFS);
3155         /* we do this loop twice.  The first time we build a list
3156          * of the extents we have a reference on, then we sort the list
3157          * by bytenr.  The second time around we actually do the
3158          * extent freeing.
3159          */
3160         for (i = 0; i < nritems; i++) {
3161                 u64 disk_bytenr;
3162                 cond_resched();
3163
3164                 btrfs_item_key_to_cpu(leaf, &key, i);
3165
3166                 /* only extents have references, skip everything else */
3167                 if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3168                         continue;
3169
3170                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
3171
3172                 /* inline extents live in the btree, they don't have refs */
3173                 if (btrfs_file_extent_type(leaf, fi) ==
3174                     BTRFS_FILE_EXTENT_INLINE)
3175                         continue;
3176
3177                 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
3178
3179                 /* holes don't have refs */
3180                 if (disk_bytenr == 0)
3181                         continue;
3182
3183                 sorted[refi].bytenr = disk_bytenr;
3184                 sorted[refi].slot = i;
3185                 refi++;
3186         }
3187
3188         if (refi == 0)
3189                 goto out;
3190
3191         sort(sorted, refi, sizeof(struct refsort), refsort_cmp, NULL);
3192
3193         for (i = 0; i < refi; i++) {
3194                 u64 disk_bytenr;
3195
3196                 disk_bytenr = sorted[i].bytenr;
3197                 slot = sorted[i].slot;
3198
3199                 cond_resched();
3200
3201                 btrfs_item_key_to_cpu(leaf, &key, slot);
3202                 if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3203                         continue;
3204
3205                 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
3206
3207                 ret = btrfs_free_extent(trans, root, disk_bytenr,
3208                                 btrfs_file_extent_disk_num_bytes(leaf, fi),
3209                                 leaf->start, leaf_owner, leaf_generation,
3210                                 key.objectid, 0);
3211                 BUG_ON(ret);
3212
3213                 atomic_inc(&root->fs_info->throttle_gen);
3214                 wake_up(&root->fs_info->transaction_throttle);
3215                 cond_resched();
3216         }
3217 out:
3218         kfree(sorted);
3219         return 0;
3220 }
3221
3222 static noinline int cache_drop_leaf_ref(struct btrfs_trans_handle *trans,
3223                                         struct btrfs_root *root,
3224                                         struct btrfs_leaf_ref *ref)
3225 {
3226         int i;
3227         int ret;
3228         struct btrfs_extent_info *info;
3229         struct refsort *sorted;
3230
3231         if (ref->nritems == 0)
3232                 return 0;
3233
3234         sorted = kmalloc(sizeof(*sorted) * ref->nritems, GFP_NOFS);
3235         for (i = 0; i < ref->nritems; i++) {
3236                 sorted[i].bytenr = ref->extents[i].bytenr;
3237                 sorted[i].slot = i;
3238         }
3239         sort(sorted, ref->nritems, sizeof(struct refsort), refsort_cmp, NULL);
3240
3241         /*
3242          * the items in the ref were sorted when the ref was inserted
3243          * into the ref cache, so this is already in order
3244          */
3245         for (i = 0; i < ref->nritems; i++) {
3246                 info = ref->extents + sorted[i].slot;
3247                 ret = btrfs_free_extent(trans, root, info->bytenr,
3248                                           info->num_bytes, ref->bytenr,
3249                                           ref->owner, ref->generation,
3250                                           info->objectid, 0);
3251
3252                 atomic_inc(&root->fs_info->throttle_gen);
3253                 wake_up(&root->fs_info->transaction_throttle);
3254                 cond_resched();
3255
3256                 BUG_ON(ret);
3257                 info++;
3258         }
3259
3260         kfree(sorted);
3261         return 0;
3262 }
3263
3264 static int drop_snap_lookup_refcount(struct btrfs_trans_handle *trans,
3265                                      struct btrfs_root *root, u64 start,
3266                                      u64 len, u32 *refs)
3267 {
3268         int ret;
3269
3270         ret = btrfs_lookup_extent_ref(trans, root, start, len, refs);
3271         BUG_ON(ret);
3272
3273 #if 0 /* some debugging code in case we see problems here */
3274         /* if the refs count is one, it won't get increased again.  But
3275          * if the ref count is > 1, someone may be decreasing it at
3276          * the same time we are.
3277          */
3278         if (*refs != 1) {
3279                 struct extent_buffer *eb = NULL;
3280                 eb = btrfs_find_create_tree_block(root, start, len);
3281                 if (eb)
3282                         btrfs_tree_lock(eb);
3283
3284                 mutex_lock(&root->fs_info->alloc_mutex);
3285                 ret = lookup_extent_ref(NULL, root, start, len, refs);
3286                 BUG_ON(ret);
3287                 mutex_unlock(&root->fs_info->alloc_mutex);
3288
3289                 if (eb) {
3290                         btrfs_tree_unlock(eb);
3291                         free_extent_buffer(eb);
3292                 }
3293                 if (*refs == 1) {
3294                         printk(KERN_ERR "btrfs block %llu went down to one "
3295                                "during drop_snap\n", (unsigned long long)start);
3296                 }
3297
3298         }
3299 #endif
3300
3301         cond_resched();
3302         return ret;
3303 }
3304
3305 /*
3306  * this is used while deleting old snapshots, and it drops the refs
3307  * on a whole subtree starting from a level 1 node.
3308  *
3309  * The idea is to sort all the leaf pointers, and then drop the
3310  * ref on all the leaves in order.  Most of the time the leaves
3311  * will have ref cache entries, so no leaf IOs will be required to
3312  * find the extents they have references on.
3313  *
3314  * For each leaf, any references it has are also dropped in order
3315  *
3316  * This ends up dropping the references in something close to optimal
3317  * order for reading and modifying the extent allocation tree.
3318  */
3319 static noinline int drop_level_one_refs(struct btrfs_trans_handle *trans,
3320                                         struct btrfs_root *root,
3321                                         struct btrfs_path *path)
3322 {
3323         u64 bytenr;
3324         u64 root_owner;
3325         u64 root_gen;
3326         struct extent_buffer *eb = path->nodes[1];
3327         struct extent_buffer *leaf;
3328         struct btrfs_leaf_ref *ref;
3329         struct refsort *sorted = NULL;
3330         int nritems = btrfs_header_nritems(eb);
3331         int ret;
3332         int i;
3333         int refi = 0;
3334         int slot = path->slots[1];
3335         u32 blocksize = btrfs_level_size(root, 0);
3336         u32 refs;
3337
3338         if (nritems == 0)
3339                 goto out;
3340
3341         root_owner = btrfs_header_owner(eb);
3342         root_gen = btrfs_header_generation(eb);
3343         sorted = kmalloc(sizeof(*sorted) * nritems, GFP_NOFS);
3344
3345         /*
3346          * step one, sort all the leaf pointers so we don't scribble
3347          * randomly into the extent allocation tree
3348          */
3349         for (i = slot; i < nritems; i++) {
3350                 sorted[refi].bytenr = btrfs_node_blockptr(eb, i);
3351                 sorted[refi].slot = i;
3352                 refi++;
3353         }
3354
3355         /*
3356          * nritems won't be zero, but if we're picking up drop_snapshot
3357          * after a crash, slot might be > 0, so double check things
3358          * just in case.
3359          */
3360         if (refi == 0)
3361                 goto out;
3362
3363         sort(sorted, refi, sizeof(struct refsort), refsort_cmp, NULL);
3364
3365         /*
3366          * the first loop frees everything the leaves point to
3367          */
3368         for (i = 0; i < refi; i++) {
3369                 u64 ptr_gen;
3370
3371                 bytenr = sorted[i].bytenr;
3372
3373                 /*
3374                  * check the reference count on this leaf.  If it is > 1
3375                  * we just decrement it below and don't update any
3376                  * of the refs the leaf points to.
3377                  */
3378                 ret = drop_snap_lookup_refcount(trans, root, bytenr,
3379                                                 blocksize, &refs);
3380                 BUG_ON(ret);
3381                 if (refs != 1)
3382                         continue;
3383
3384                 ptr_gen = btrfs_node_ptr_generation(eb, sorted[i].slot);
3385
3386                 /*
3387                  * the leaf only had one reference, which means the
3388                  * only thing pointing to this leaf is the snapshot
3389                  * we're deleting.  It isn't possible for the reference
3390                  * count to increase again later
3391                  *
3392                  * The reference cache is checked for the leaf,
3393                  * and if found we'll be able to drop any refs held by
3394                  * the leaf without needing to read it in.
3395                  */
3396                 ref = btrfs_lookup_leaf_ref(root, bytenr);
3397                 if (ref && ref->generation != ptr_gen) {
3398                         btrfs_free_leaf_ref(root, ref);
3399                         ref = NULL;
3400                 }
3401                 if (ref) {
3402                         ret = cache_drop_leaf_ref(trans, root, ref);
3403                         BUG_ON(ret);
3404                         btrfs_remove_leaf_ref(root, ref);
3405                         btrfs_free_leaf_ref(root, ref);
3406                 } else {
3407                         /*
3408                          * the leaf wasn't in the reference cache, so
3409                          * we have to read it.
3410                          */
3411                         leaf = read_tree_block(root, bytenr, blocksize,
3412                                                ptr_gen);
3413                         ret = btrfs_drop_leaf_ref(trans, root, leaf);
3414                         BUG_ON(ret);
3415                         free_extent_buffer(leaf);
3416                 }
3417                 atomic_inc(&root->fs_info->throttle_gen);
3418                 wake_up(&root->fs_info->transaction_throttle);
3419                 cond_resched();
3420         }
3421
3422         /*
3423          * run through the loop again to free the refs on the leaves.
3424          * This is faster than doing it in the loop above because
3425          * the leaves are likely to be clustered together.  We end up
3426          * working in nice chunks on the extent allocation tree.
3427          */
3428         for (i = 0; i < refi; i++) {
3429                 bytenr = sorted[i].bytenr;
3430                 ret = btrfs_free_extent(trans, root, bytenr,
3431                                         blocksize, eb->start,
3432                                         root_owner, root_gen, 0, 1);
3433                 BUG_ON(ret);
3434
3435                 atomic_inc(&root->fs_info->throttle_gen);
3436                 wake_up(&root->fs_info->transaction_throttle);
3437                 cond_resched();
3438         }
3439 out:
3440         kfree(sorted);
3441
3442         /*
3443          * update the path to show we've processed the entire level 1
3444          * node.  This will get saved into the root's drop_snapshot_progress
3445          * field so these drops are not repeated again if this transaction
3446          * commits.
3447          */
3448         path->slots[1] = nritems;
3449         return 0;
3450 }
3451
3452 /*
3453  * helper function for drop_snapshot, this walks down the tree dropping ref
3454  * counts as it goes.
3455  */
3456 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
3457                                    struct btrfs_root *root,
3458                                    struct btrfs_path *path, int *level)
3459 {
3460         u64 root_owner;
3461         u64 root_gen;
3462         u64 bytenr;
3463         u64 ptr_gen;
3464         struct extent_buffer *next;
3465         struct extent_buffer *cur;
3466         struct extent_buffer *parent;
3467         u32 blocksize;
3468         int ret;
3469         u32 refs;
3470
3471         WARN_ON(*level < 0);
3472         WARN_ON(*level >= BTRFS_MAX_LEVEL);
3473         ret = drop_snap_lookup_refcount(trans, root, path->nodes[*level]->start,
3474                                 path->nodes[*level]->len, &refs);
3475         BUG_ON(ret);
3476         if (refs > 1)
3477                 goto out;
3478
3479         /*
3480          * walk down to the last node level and free all the leaves
3481          */
3482         while (*level >= 0) {
3483                 WARN_ON(*level < 0);
3484                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
3485                 cur = path->nodes[*level];
3486
3487                 if (btrfs_header_level(cur) != *level)
3488                         WARN_ON(1);
3489
3490                 if (path->slots[*level] >=
3491                     btrfs_header_nritems(cur))
3492                         break;
3493
3494                 /* the new code goes down to level 1 and does all the
3495                  * leaves pointed to that node in bulk.  So, this check
3496                  * for level 0 will always be false.
3497                  *
3498                  * But, the disk format allows the drop_snapshot_progress
3499                  * field in the root to leave things in a state where
3500                  * a leaf will need cleaning up here.  If someone crashes
3501                  * with the old code and then boots with the new code,
3502                  * we might find a leaf here.
3503                  */
3504                 if (*level == 0) {
3505                         ret = btrfs_drop_leaf_ref(trans, root, cur);
3506                         BUG_ON(ret);
3507                         break;
3508                 }
3509
3510                 /*
3511                  * once we get to level one, process the whole node
3512                  * at once, including everything below it.
3513                  */
3514                 if (*level == 1) {
3515                         ret = drop_level_one_refs(trans, root, path);
3516                         BUG_ON(ret);
3517                         break;
3518                 }