Btrfs: create special free space cache inode
[linux-3.10.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include "compat.h"
27 #include "hash.h"
28 #include "ctree.h"
29 #include "disk-io.h"
30 #include "print-tree.h"
31 #include "transaction.h"
32 #include "volumes.h"
33 #include "locking.h"
34 #include "free-space-cache.h"
35
36 static int update_block_group(struct btrfs_trans_handle *trans,
37                               struct btrfs_root *root,
38                               u64 bytenr, u64 num_bytes, int alloc);
39 static int update_reserved_bytes(struct btrfs_block_group_cache *cache,
40                                  u64 num_bytes, int reserve, int sinfo);
41 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
42                                 struct btrfs_root *root,
43                                 u64 bytenr, u64 num_bytes, u64 parent,
44                                 u64 root_objectid, u64 owner_objectid,
45                                 u64 owner_offset, int refs_to_drop,
46                                 struct btrfs_delayed_extent_op *extra_op);
47 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
48                                     struct extent_buffer *leaf,
49                                     struct btrfs_extent_item *ei);
50 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
51                                       struct btrfs_root *root,
52                                       u64 parent, u64 root_objectid,
53                                       u64 flags, u64 owner, u64 offset,
54                                       struct btrfs_key *ins, int ref_mod);
55 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
56                                      struct btrfs_root *root,
57                                      u64 parent, u64 root_objectid,
58                                      u64 flags, struct btrfs_disk_key *key,
59                                      int level, struct btrfs_key *ins);
60 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
61                           struct btrfs_root *extent_root, u64 alloc_bytes,
62                           u64 flags, int force);
63 static int find_next_key(struct btrfs_path *path, int level,
64                          struct btrfs_key *key);
65 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
66                             int dump_block_groups);
67
68 static noinline int
69 block_group_cache_done(struct btrfs_block_group_cache *cache)
70 {
71         smp_mb();
72         return cache->cached == BTRFS_CACHE_FINISHED;
73 }
74
75 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
76 {
77         return (cache->flags & bits) == bits;
78 }
79
80 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
81 {
82         atomic_inc(&cache->count);
83 }
84
85 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
86 {
87         if (atomic_dec_and_test(&cache->count)) {
88                 WARN_ON(cache->pinned > 0);
89                 WARN_ON(cache->reserved > 0);
90                 WARN_ON(cache->reserved_pinned > 0);
91                 kfree(cache);
92         }
93 }
94
95 /*
96  * this adds the block group to the fs_info rb tree for the block group
97  * cache
98  */
99 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
100                                 struct btrfs_block_group_cache *block_group)
101 {
102         struct rb_node **p;
103         struct rb_node *parent = NULL;
104         struct btrfs_block_group_cache *cache;
105
106         spin_lock(&info->block_group_cache_lock);
107         p = &info->block_group_cache_tree.rb_node;
108
109         while (*p) {
110                 parent = *p;
111                 cache = rb_entry(parent, struct btrfs_block_group_cache,
112                                  cache_node);
113                 if (block_group->key.objectid < cache->key.objectid) {
114                         p = &(*p)->rb_left;
115                 } else if (block_group->key.objectid > cache->key.objectid) {
116                         p = &(*p)->rb_right;
117                 } else {
118                         spin_unlock(&info->block_group_cache_lock);
119                         return -EEXIST;
120                 }
121         }
122
123         rb_link_node(&block_group->cache_node, parent, p);
124         rb_insert_color(&block_group->cache_node,
125                         &info->block_group_cache_tree);
126         spin_unlock(&info->block_group_cache_lock);
127
128         return 0;
129 }
130
131 /*
132  * This will return the block group at or after bytenr if contains is 0, else
133  * it will return the block group that contains the bytenr
134  */
135 static struct btrfs_block_group_cache *
136 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
137                               int contains)
138 {
139         struct btrfs_block_group_cache *cache, *ret = NULL;
140         struct rb_node *n;
141         u64 end, start;
142
143         spin_lock(&info->block_group_cache_lock);
144         n = info->block_group_cache_tree.rb_node;
145
146         while (n) {
147                 cache = rb_entry(n, struct btrfs_block_group_cache,
148                                  cache_node);
149                 end = cache->key.objectid + cache->key.offset - 1;
150                 start = cache->key.objectid;
151
152                 if (bytenr < start) {
153                         if (!contains && (!ret || start < ret->key.objectid))
154                                 ret = cache;
155                         n = n->rb_left;
156                 } else if (bytenr > start) {
157                         if (contains && bytenr <= end) {
158                                 ret = cache;
159                                 break;
160                         }
161                         n = n->rb_right;
162                 } else {
163                         ret = cache;
164                         break;
165                 }
166         }
167         if (ret)
168                 btrfs_get_block_group(ret);
169         spin_unlock(&info->block_group_cache_lock);
170
171         return ret;
172 }
173
174 static int add_excluded_extent(struct btrfs_root *root,
175                                u64 start, u64 num_bytes)
176 {
177         u64 end = start + num_bytes - 1;
178         set_extent_bits(&root->fs_info->freed_extents[0],
179                         start, end, EXTENT_UPTODATE, GFP_NOFS);
180         set_extent_bits(&root->fs_info->freed_extents[1],
181                         start, end, EXTENT_UPTODATE, GFP_NOFS);
182         return 0;
183 }
184
185 static void free_excluded_extents(struct btrfs_root *root,
186                                   struct btrfs_block_group_cache *cache)
187 {
188         u64 start, end;
189
190         start = cache->key.objectid;
191         end = start + cache->key.offset - 1;
192
193         clear_extent_bits(&root->fs_info->freed_extents[0],
194                           start, end, EXTENT_UPTODATE, GFP_NOFS);
195         clear_extent_bits(&root->fs_info->freed_extents[1],
196                           start, end, EXTENT_UPTODATE, GFP_NOFS);
197 }
198
199 static int exclude_super_stripes(struct btrfs_root *root,
200                                  struct btrfs_block_group_cache *cache)
201 {
202         u64 bytenr;
203         u64 *logical;
204         int stripe_len;
205         int i, nr, ret;
206
207         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
208                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
209                 cache->bytes_super += stripe_len;
210                 ret = add_excluded_extent(root, cache->key.objectid,
211                                           stripe_len);
212                 BUG_ON(ret);
213         }
214
215         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
216                 bytenr = btrfs_sb_offset(i);
217                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
218                                        cache->key.objectid, bytenr,
219                                        0, &logical, &nr, &stripe_len);
220                 BUG_ON(ret);
221
222                 while (nr--) {
223                         cache->bytes_super += stripe_len;
224                         ret = add_excluded_extent(root, logical[nr],
225                                                   stripe_len);
226                         BUG_ON(ret);
227                 }
228
229                 kfree(logical);
230         }
231         return 0;
232 }
233
234 static struct btrfs_caching_control *
235 get_caching_control(struct btrfs_block_group_cache *cache)
236 {
237         struct btrfs_caching_control *ctl;
238
239         spin_lock(&cache->lock);
240         if (cache->cached != BTRFS_CACHE_STARTED) {
241                 spin_unlock(&cache->lock);
242                 return NULL;
243         }
244
245         ctl = cache->caching_ctl;
246         atomic_inc(&ctl->count);
247         spin_unlock(&cache->lock);
248         return ctl;
249 }
250
251 static void put_caching_control(struct btrfs_caching_control *ctl)
252 {
253         if (atomic_dec_and_test(&ctl->count))
254                 kfree(ctl);
255 }
256
257 /*
258  * this is only called by cache_block_group, since we could have freed extents
259  * we need to check the pinned_extents for any extents that can't be used yet
260  * since their free space will be released as soon as the transaction commits.
261  */
262 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
263                               struct btrfs_fs_info *info, u64 start, u64 end)
264 {
265         u64 extent_start, extent_end, size, total_added = 0;
266         int ret;
267
268         while (start < end) {
269                 ret = find_first_extent_bit(info->pinned_extents, start,
270                                             &extent_start, &extent_end,
271                                             EXTENT_DIRTY | EXTENT_UPTODATE);
272                 if (ret)
273                         break;
274
275                 if (extent_start <= start) {
276                         start = extent_end + 1;
277                 } else if (extent_start > start && extent_start < end) {
278                         size = extent_start - start;
279                         total_added += size;
280                         ret = btrfs_add_free_space(block_group, start,
281                                                    size);
282                         BUG_ON(ret);
283                         start = extent_end + 1;
284                 } else {
285                         break;
286                 }
287         }
288
289         if (start < end) {
290                 size = end - start;
291                 total_added += size;
292                 ret = btrfs_add_free_space(block_group, start, size);
293                 BUG_ON(ret);
294         }
295
296         return total_added;
297 }
298
299 static int caching_kthread(void *data)
300 {
301         struct btrfs_block_group_cache *block_group = data;
302         struct btrfs_fs_info *fs_info = block_group->fs_info;
303         struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
304         struct btrfs_root *extent_root = fs_info->extent_root;
305         struct btrfs_path *path;
306         struct extent_buffer *leaf;
307         struct btrfs_key key;
308         u64 total_found = 0;
309         u64 last = 0;
310         u32 nritems;
311         int ret = 0;
312
313         path = btrfs_alloc_path();
314         if (!path)
315                 return -ENOMEM;
316
317         exclude_super_stripes(extent_root, block_group);
318         spin_lock(&block_group->space_info->lock);
319         block_group->space_info->bytes_readonly += block_group->bytes_super;
320         spin_unlock(&block_group->space_info->lock);
321
322         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
323
324         /*
325          * We don't want to deadlock with somebody trying to allocate a new
326          * extent for the extent root while also trying to search the extent
327          * root to add free space.  So we skip locking and search the commit
328          * root, since its read-only
329          */
330         path->skip_locking = 1;
331         path->search_commit_root = 1;
332         path->reada = 2;
333
334         key.objectid = last;
335         key.offset = 0;
336         key.type = BTRFS_EXTENT_ITEM_KEY;
337 again:
338         mutex_lock(&caching_ctl->mutex);
339         /* need to make sure the commit_root doesn't disappear */
340         down_read(&fs_info->extent_commit_sem);
341
342         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
343         if (ret < 0)
344                 goto err;
345
346         leaf = path->nodes[0];
347         nritems = btrfs_header_nritems(leaf);
348
349         while (1) {
350                 smp_mb();
351                 if (fs_info->closing > 1) {
352                         last = (u64)-1;
353                         break;
354                 }
355
356                 if (path->slots[0] < nritems) {
357                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
358                 } else {
359                         ret = find_next_key(path, 0, &key);
360                         if (ret)
361                                 break;
362
363                         caching_ctl->progress = last;
364                         btrfs_release_path(extent_root, path);
365                         up_read(&fs_info->extent_commit_sem);
366                         mutex_unlock(&caching_ctl->mutex);
367                         if (btrfs_transaction_in_commit(fs_info))
368                                 schedule_timeout(1);
369                         else
370                                 cond_resched();
371                         goto again;
372                 }
373
374                 if (key.objectid < block_group->key.objectid) {
375                         path->slots[0]++;
376                         continue;
377                 }
378
379                 if (key.objectid >= block_group->key.objectid +
380                     block_group->key.offset)
381                         break;
382
383                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
384                         total_found += add_new_free_space(block_group,
385                                                           fs_info, last,
386                                                           key.objectid);
387                         last = key.objectid + key.offset;
388
389                         if (total_found > (1024 * 1024 * 2)) {
390                                 total_found = 0;
391                                 wake_up(&caching_ctl->wait);
392                         }
393                 }
394                 path->slots[0]++;
395         }
396         ret = 0;
397
398         total_found += add_new_free_space(block_group, fs_info, last,
399                                           block_group->key.objectid +
400                                           block_group->key.offset);
401         caching_ctl->progress = (u64)-1;
402
403         spin_lock(&block_group->lock);
404         block_group->caching_ctl = NULL;
405         block_group->cached = BTRFS_CACHE_FINISHED;
406         spin_unlock(&block_group->lock);
407
408 err:
409         btrfs_free_path(path);
410         up_read(&fs_info->extent_commit_sem);
411
412         free_excluded_extents(extent_root, block_group);
413
414         mutex_unlock(&caching_ctl->mutex);
415         wake_up(&caching_ctl->wait);
416
417         put_caching_control(caching_ctl);
418         atomic_dec(&block_group->space_info->caching_threads);
419         btrfs_put_block_group(block_group);
420
421         return 0;
422 }
423
424 static int cache_block_group(struct btrfs_block_group_cache *cache)
425 {
426         struct btrfs_fs_info *fs_info = cache->fs_info;
427         struct btrfs_caching_control *caching_ctl;
428         struct task_struct *tsk;
429         int ret = 0;
430
431         smp_mb();
432         if (cache->cached != BTRFS_CACHE_NO)
433                 return 0;
434
435         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_KERNEL);
436         BUG_ON(!caching_ctl);
437
438         INIT_LIST_HEAD(&caching_ctl->list);
439         mutex_init(&caching_ctl->mutex);
440         init_waitqueue_head(&caching_ctl->wait);
441         caching_ctl->block_group = cache;
442         caching_ctl->progress = cache->key.objectid;
443         /* one for caching kthread, one for caching block group list */
444         atomic_set(&caching_ctl->count, 2);
445
446         spin_lock(&cache->lock);
447         if (cache->cached != BTRFS_CACHE_NO) {
448                 spin_unlock(&cache->lock);
449                 kfree(caching_ctl);
450                 return 0;
451         }
452         cache->caching_ctl = caching_ctl;
453         cache->cached = BTRFS_CACHE_STARTED;
454         spin_unlock(&cache->lock);
455
456         down_write(&fs_info->extent_commit_sem);
457         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
458         up_write(&fs_info->extent_commit_sem);
459
460         atomic_inc(&cache->space_info->caching_threads);
461         btrfs_get_block_group(cache);
462
463         tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
464                           cache->key.objectid);
465         if (IS_ERR(tsk)) {
466                 ret = PTR_ERR(tsk);
467                 printk(KERN_ERR "error running thread %d\n", ret);
468                 BUG();
469         }
470
471         return ret;
472 }
473
474 /*
475  * return the block group that starts at or after bytenr
476  */
477 static struct btrfs_block_group_cache *
478 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
479 {
480         struct btrfs_block_group_cache *cache;
481
482         cache = block_group_cache_tree_search(info, bytenr, 0);
483
484         return cache;
485 }
486
487 /*
488  * return the block group that contains the given bytenr
489  */
490 struct btrfs_block_group_cache *btrfs_lookup_block_group(
491                                                  struct btrfs_fs_info *info,
492                                                  u64 bytenr)
493 {
494         struct btrfs_block_group_cache *cache;
495
496         cache = block_group_cache_tree_search(info, bytenr, 1);
497
498         return cache;
499 }
500
501 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
502                                                   u64 flags)
503 {
504         struct list_head *head = &info->space_info;
505         struct btrfs_space_info *found;
506
507         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
508                  BTRFS_BLOCK_GROUP_METADATA;
509
510         rcu_read_lock();
511         list_for_each_entry_rcu(found, head, list) {
512                 if (found->flags == flags) {
513                         rcu_read_unlock();
514                         return found;
515                 }
516         }
517         rcu_read_unlock();
518         return NULL;
519 }
520
521 /*
522  * after adding space to the filesystem, we need to clear the full flags
523  * on all the space infos.
524  */
525 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
526 {
527         struct list_head *head = &info->space_info;
528         struct btrfs_space_info *found;
529
530         rcu_read_lock();
531         list_for_each_entry_rcu(found, head, list)
532                 found->full = 0;
533         rcu_read_unlock();
534 }
535
536 static u64 div_factor(u64 num, int factor)
537 {
538         if (factor == 10)
539                 return num;
540         num *= factor;
541         do_div(num, 10);
542         return num;
543 }
544
545 u64 btrfs_find_block_group(struct btrfs_root *root,
546                            u64 search_start, u64 search_hint, int owner)
547 {
548         struct btrfs_block_group_cache *cache;
549         u64 used;
550         u64 last = max(search_hint, search_start);
551         u64 group_start = 0;
552         int full_search = 0;
553         int factor = 9;
554         int wrapped = 0;
555 again:
556         while (1) {
557                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
558                 if (!cache)
559                         break;
560
561                 spin_lock(&cache->lock);
562                 last = cache->key.objectid + cache->key.offset;
563                 used = btrfs_block_group_used(&cache->item);
564
565                 if ((full_search || !cache->ro) &&
566                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
567                         if (used + cache->pinned + cache->reserved <
568                             div_factor(cache->key.offset, factor)) {
569                                 group_start = cache->key.objectid;
570                                 spin_unlock(&cache->lock);
571                                 btrfs_put_block_group(cache);
572                                 goto found;
573                         }
574                 }
575                 spin_unlock(&cache->lock);
576                 btrfs_put_block_group(cache);
577                 cond_resched();
578         }
579         if (!wrapped) {
580                 last = search_start;
581                 wrapped = 1;
582                 goto again;
583         }
584         if (!full_search && factor < 10) {
585                 last = search_start;
586                 full_search = 1;
587                 factor = 10;
588                 goto again;
589         }
590 found:
591         return group_start;
592 }
593
594 /* simple helper to search for an existing extent at a given offset */
595 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
596 {
597         int ret;
598         struct btrfs_key key;
599         struct btrfs_path *path;
600
601         path = btrfs_alloc_path();
602         BUG_ON(!path);
603         key.objectid = start;
604         key.offset = len;
605         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
606         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
607                                 0, 0);
608         btrfs_free_path(path);
609         return ret;
610 }
611
612 /*
613  * helper function to lookup reference count and flags of extent.
614  *
615  * the head node for delayed ref is used to store the sum of all the
616  * reference count modifications queued up in the rbtree. the head
617  * node may also store the extent flags to set. This way you can check
618  * to see what the reference count and extent flags would be if all of
619  * the delayed refs are not processed.
620  */
621 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
622                              struct btrfs_root *root, u64 bytenr,
623                              u64 num_bytes, u64 *refs, u64 *flags)
624 {
625         struct btrfs_delayed_ref_head *head;
626         struct btrfs_delayed_ref_root *delayed_refs;
627         struct btrfs_path *path;
628         struct btrfs_extent_item *ei;
629         struct extent_buffer *leaf;
630         struct btrfs_key key;
631         u32 item_size;
632         u64 num_refs;
633         u64 extent_flags;
634         int ret;
635
636         path = btrfs_alloc_path();
637         if (!path)
638                 return -ENOMEM;
639
640         key.objectid = bytenr;
641         key.type = BTRFS_EXTENT_ITEM_KEY;
642         key.offset = num_bytes;
643         if (!trans) {
644                 path->skip_locking = 1;
645                 path->search_commit_root = 1;
646         }
647 again:
648         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
649                                 &key, path, 0, 0);
650         if (ret < 0)
651                 goto out_free;
652
653         if (ret == 0) {
654                 leaf = path->nodes[0];
655                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
656                 if (item_size >= sizeof(*ei)) {
657                         ei = btrfs_item_ptr(leaf, path->slots[0],
658                                             struct btrfs_extent_item);
659                         num_refs = btrfs_extent_refs(leaf, ei);
660                         extent_flags = btrfs_extent_flags(leaf, ei);
661                 } else {
662 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
663                         struct btrfs_extent_item_v0 *ei0;
664                         BUG_ON(item_size != sizeof(*ei0));
665                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
666                                              struct btrfs_extent_item_v0);
667                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
668                         /* FIXME: this isn't correct for data */
669                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
670 #else
671                         BUG();
672 #endif
673                 }
674                 BUG_ON(num_refs == 0);
675         } else {
676                 num_refs = 0;
677                 extent_flags = 0;
678                 ret = 0;
679         }
680
681         if (!trans)
682                 goto out;
683
684         delayed_refs = &trans->transaction->delayed_refs;
685         spin_lock(&delayed_refs->lock);
686         head = btrfs_find_delayed_ref_head(trans, bytenr);
687         if (head) {
688                 if (!mutex_trylock(&head->mutex)) {
689                         atomic_inc(&head->node.refs);
690                         spin_unlock(&delayed_refs->lock);
691
692                         btrfs_release_path(root->fs_info->extent_root, path);
693
694                         mutex_lock(&head->mutex);
695                         mutex_unlock(&head->mutex);
696                         btrfs_put_delayed_ref(&head->node);
697                         goto again;
698                 }
699                 if (head->extent_op && head->extent_op->update_flags)
700                         extent_flags |= head->extent_op->flags_to_set;
701                 else
702                         BUG_ON(num_refs == 0);
703
704                 num_refs += head->node.ref_mod;
705                 mutex_unlock(&head->mutex);
706         }
707         spin_unlock(&delayed_refs->lock);
708 out:
709         WARN_ON(num_refs == 0);
710         if (refs)
711                 *refs = num_refs;
712         if (flags)
713                 *flags = extent_flags;
714 out_free:
715         btrfs_free_path(path);
716         return ret;
717 }
718
719 /*
720  * Back reference rules.  Back refs have three main goals:
721  *
722  * 1) differentiate between all holders of references to an extent so that
723  *    when a reference is dropped we can make sure it was a valid reference
724  *    before freeing the extent.
725  *
726  * 2) Provide enough information to quickly find the holders of an extent
727  *    if we notice a given block is corrupted or bad.
728  *
729  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
730  *    maintenance.  This is actually the same as #2, but with a slightly
731  *    different use case.
732  *
733  * There are two kinds of back refs. The implicit back refs is optimized
734  * for pointers in non-shared tree blocks. For a given pointer in a block,
735  * back refs of this kind provide information about the block's owner tree
736  * and the pointer's key. These information allow us to find the block by
737  * b-tree searching. The full back refs is for pointers in tree blocks not
738  * referenced by their owner trees. The location of tree block is recorded
739  * in the back refs. Actually the full back refs is generic, and can be
740  * used in all cases the implicit back refs is used. The major shortcoming
741  * of the full back refs is its overhead. Every time a tree block gets
742  * COWed, we have to update back refs entry for all pointers in it.
743  *
744  * For a newly allocated tree block, we use implicit back refs for
745  * pointers in it. This means most tree related operations only involve
746  * implicit back refs. For a tree block created in old transaction, the
747  * only way to drop a reference to it is COW it. So we can detect the
748  * event that tree block loses its owner tree's reference and do the
749  * back refs conversion.
750  *
751  * When a tree block is COW'd through a tree, there are four cases:
752  *
753  * The reference count of the block is one and the tree is the block's
754  * owner tree. Nothing to do in this case.
755  *
756  * The reference count of the block is one and the tree is not the
757  * block's owner tree. In this case, full back refs is used for pointers
758  * in the block. Remove these full back refs, add implicit back refs for
759  * every pointers in the new block.
760  *
761  * The reference count of the block is greater than one and the tree is
762  * the block's owner tree. In this case, implicit back refs is used for
763  * pointers in the block. Add full back refs for every pointers in the
764  * block, increase lower level extents' reference counts. The original
765  * implicit back refs are entailed to the new block.
766  *
767  * The reference count of the block is greater than one and the tree is
768  * not the block's owner tree. Add implicit back refs for every pointer in
769  * the new block, increase lower level extents' reference count.
770  *
771  * Back Reference Key composing:
772  *
773  * The key objectid corresponds to the first byte in the extent,
774  * The key type is used to differentiate between types of back refs.
775  * There are different meanings of the key offset for different types
776  * of back refs.
777  *
778  * File extents can be referenced by:
779  *
780  * - multiple snapshots, subvolumes, or different generations in one subvol
781  * - different files inside a single subvolume
782  * - different offsets inside a file (bookend extents in file.c)
783  *
784  * The extent ref structure for the implicit back refs has fields for:
785  *
786  * - Objectid of the subvolume root
787  * - objectid of the file holding the reference
788  * - original offset in the file
789  * - how many bookend extents
790  *
791  * The key offset for the implicit back refs is hash of the first
792  * three fields.
793  *
794  * The extent ref structure for the full back refs has field for:
795  *
796  * - number of pointers in the tree leaf
797  *
798  * The key offset for the implicit back refs is the first byte of
799  * the tree leaf
800  *
801  * When a file extent is allocated, The implicit back refs is used.
802  * the fields are filled in:
803  *
804  *     (root_key.objectid, inode objectid, offset in file, 1)
805  *
806  * When a file extent is removed file truncation, we find the
807  * corresponding implicit back refs and check the following fields:
808  *
809  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
810  *
811  * Btree extents can be referenced by:
812  *
813  * - Different subvolumes
814  *
815  * Both the implicit back refs and the full back refs for tree blocks
816  * only consist of key. The key offset for the implicit back refs is
817  * objectid of block's owner tree. The key offset for the full back refs
818  * is the first byte of parent block.
819  *
820  * When implicit back refs is used, information about the lowest key and
821  * level of the tree block are required. These information are stored in
822  * tree block info structure.
823  */
824
825 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
826 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
827                                   struct btrfs_root *root,
828                                   struct btrfs_path *path,
829                                   u64 owner, u32 extra_size)
830 {
831         struct btrfs_extent_item *item;
832         struct btrfs_extent_item_v0 *ei0;
833         struct btrfs_extent_ref_v0 *ref0;
834         struct btrfs_tree_block_info *bi;
835         struct extent_buffer *leaf;
836         struct btrfs_key key;
837         struct btrfs_key found_key;
838         u32 new_size = sizeof(*item);
839         u64 refs;
840         int ret;
841
842         leaf = path->nodes[0];
843         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
844
845         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
846         ei0 = btrfs_item_ptr(leaf, path->slots[0],
847                              struct btrfs_extent_item_v0);
848         refs = btrfs_extent_refs_v0(leaf, ei0);
849
850         if (owner == (u64)-1) {
851                 while (1) {
852                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
853                                 ret = btrfs_next_leaf(root, path);
854                                 if (ret < 0)
855                                         return ret;
856                                 BUG_ON(ret > 0);
857                                 leaf = path->nodes[0];
858                         }
859                         btrfs_item_key_to_cpu(leaf, &found_key,
860                                               path->slots[0]);
861                         BUG_ON(key.objectid != found_key.objectid);
862                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
863                                 path->slots[0]++;
864                                 continue;
865                         }
866                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
867                                               struct btrfs_extent_ref_v0);
868                         owner = btrfs_ref_objectid_v0(leaf, ref0);
869                         break;
870                 }
871         }
872         btrfs_release_path(root, path);
873
874         if (owner < BTRFS_FIRST_FREE_OBJECTID)
875                 new_size += sizeof(*bi);
876
877         new_size -= sizeof(*ei0);
878         ret = btrfs_search_slot(trans, root, &key, path,
879                                 new_size + extra_size, 1);
880         if (ret < 0)
881                 return ret;
882         BUG_ON(ret);
883
884         ret = btrfs_extend_item(trans, root, path, new_size);
885         BUG_ON(ret);
886
887         leaf = path->nodes[0];
888         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
889         btrfs_set_extent_refs(leaf, item, refs);
890         /* FIXME: get real generation */
891         btrfs_set_extent_generation(leaf, item, 0);
892         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
893                 btrfs_set_extent_flags(leaf, item,
894                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
895                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
896                 bi = (struct btrfs_tree_block_info *)(item + 1);
897                 /* FIXME: get first key of the block */
898                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
899                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
900         } else {
901                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
902         }
903         btrfs_mark_buffer_dirty(leaf);
904         return 0;
905 }
906 #endif
907
908 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
909 {
910         u32 high_crc = ~(u32)0;
911         u32 low_crc = ~(u32)0;
912         __le64 lenum;
913
914         lenum = cpu_to_le64(root_objectid);
915         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
916         lenum = cpu_to_le64(owner);
917         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
918         lenum = cpu_to_le64(offset);
919         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
920
921         return ((u64)high_crc << 31) ^ (u64)low_crc;
922 }
923
924 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
925                                      struct btrfs_extent_data_ref *ref)
926 {
927         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
928                                     btrfs_extent_data_ref_objectid(leaf, ref),
929                                     btrfs_extent_data_ref_offset(leaf, ref));
930 }
931
932 static int match_extent_data_ref(struct extent_buffer *leaf,
933                                  struct btrfs_extent_data_ref *ref,
934                                  u64 root_objectid, u64 owner, u64 offset)
935 {
936         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
937             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
938             btrfs_extent_data_ref_offset(leaf, ref) != offset)
939                 return 0;
940         return 1;
941 }
942
943 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
944                                            struct btrfs_root *root,
945                                            struct btrfs_path *path,
946                                            u64 bytenr, u64 parent,
947                                            u64 root_objectid,
948                                            u64 owner, u64 offset)
949 {
950         struct btrfs_key key;
951         struct btrfs_extent_data_ref *ref;
952         struct extent_buffer *leaf;
953         u32 nritems;
954         int ret;
955         int recow;
956         int err = -ENOENT;
957
958         key.objectid = bytenr;
959         if (parent) {
960                 key.type = BTRFS_SHARED_DATA_REF_KEY;
961                 key.offset = parent;
962         } else {
963                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
964                 key.offset = hash_extent_data_ref(root_objectid,
965                                                   owner, offset);
966         }
967 again:
968         recow = 0;
969         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
970         if (ret < 0) {
971                 err = ret;
972                 goto fail;
973         }
974
975         if (parent) {
976                 if (!ret)
977                         return 0;
978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
979                 key.type = BTRFS_EXTENT_REF_V0_KEY;
980                 btrfs_release_path(root, path);
981                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
982                 if (ret < 0) {
983                         err = ret;
984                         goto fail;
985                 }
986                 if (!ret)
987                         return 0;
988 #endif
989                 goto fail;
990         }
991
992         leaf = path->nodes[0];
993         nritems = btrfs_header_nritems(leaf);
994         while (1) {
995                 if (path->slots[0] >= nritems) {
996                         ret = btrfs_next_leaf(root, path);
997                         if (ret < 0)
998                                 err = ret;
999                         if (ret)
1000                                 goto fail;
1001
1002                         leaf = path->nodes[0];
1003                         nritems = btrfs_header_nritems(leaf);
1004                         recow = 1;
1005                 }
1006
1007                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1008                 if (key.objectid != bytenr ||
1009                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1010                         goto fail;
1011
1012                 ref = btrfs_item_ptr(leaf, path->slots[0],
1013                                      struct btrfs_extent_data_ref);
1014
1015                 if (match_extent_data_ref(leaf, ref, root_objectid,
1016                                           owner, offset)) {
1017                         if (recow) {
1018                                 btrfs_release_path(root, path);
1019                                 goto again;
1020                         }
1021                         err = 0;
1022                         break;
1023                 }
1024                 path->slots[0]++;
1025         }
1026 fail:
1027         return err;
1028 }
1029
1030 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1031                                            struct btrfs_root *root,
1032                                            struct btrfs_path *path,
1033                                            u64 bytenr, u64 parent,
1034                                            u64 root_objectid, u64 owner,
1035                                            u64 offset, int refs_to_add)
1036 {
1037         struct btrfs_key key;
1038         struct extent_buffer *leaf;
1039         u32 size;
1040         u32 num_refs;
1041         int ret;
1042
1043         key.objectid = bytenr;
1044         if (parent) {
1045                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1046                 key.offset = parent;
1047                 size = sizeof(struct btrfs_shared_data_ref);
1048         } else {
1049                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1050                 key.offset = hash_extent_data_ref(root_objectid,
1051                                                   owner, offset);
1052                 size = sizeof(struct btrfs_extent_data_ref);
1053         }
1054
1055         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1056         if (ret && ret != -EEXIST)
1057                 goto fail;
1058
1059         leaf = path->nodes[0];
1060         if (parent) {
1061                 struct btrfs_shared_data_ref *ref;
1062                 ref = btrfs_item_ptr(leaf, path->slots[0],
1063                                      struct btrfs_shared_data_ref);
1064                 if (ret == 0) {
1065                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1066                 } else {
1067                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1068                         num_refs += refs_to_add;
1069                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1070                 }
1071         } else {
1072                 struct btrfs_extent_data_ref *ref;
1073                 while (ret == -EEXIST) {
1074                         ref = btrfs_item_ptr(leaf, path->slots[0],
1075                                              struct btrfs_extent_data_ref);
1076                         if (match_extent_data_ref(leaf, ref, root_objectid,
1077                                                   owner, offset))
1078                                 break;
1079                         btrfs_release_path(root, path);
1080                         key.offset++;
1081                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1082                                                       size);
1083                         if (ret && ret != -EEXIST)
1084                                 goto fail;
1085
1086                         leaf = path->nodes[0];
1087                 }
1088                 ref = btrfs_item_ptr(leaf, path->slots[0],
1089                                      struct btrfs_extent_data_ref);
1090                 if (ret == 0) {
1091                         btrfs_set_extent_data_ref_root(leaf, ref,
1092                                                        root_objectid);
1093                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1094                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1095                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1096                 } else {
1097                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1098                         num_refs += refs_to_add;
1099                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1100                 }
1101         }
1102         btrfs_mark_buffer_dirty(leaf);
1103         ret = 0;
1104 fail:
1105         btrfs_release_path(root, path);
1106         return ret;
1107 }
1108
1109 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1110                                            struct btrfs_root *root,
1111                                            struct btrfs_path *path,
1112                                            int refs_to_drop)
1113 {
1114         struct btrfs_key key;
1115         struct btrfs_extent_data_ref *ref1 = NULL;
1116         struct btrfs_shared_data_ref *ref2 = NULL;
1117         struct extent_buffer *leaf;
1118         u32 num_refs = 0;
1119         int ret = 0;
1120
1121         leaf = path->nodes[0];
1122         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1123
1124         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1125                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1126                                       struct btrfs_extent_data_ref);
1127                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1128         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1129                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1130                                       struct btrfs_shared_data_ref);
1131                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1132 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1133         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1134                 struct btrfs_extent_ref_v0 *ref0;
1135                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1136                                       struct btrfs_extent_ref_v0);
1137                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1138 #endif
1139         } else {
1140                 BUG();
1141         }
1142
1143         BUG_ON(num_refs < refs_to_drop);
1144         num_refs -= refs_to_drop;
1145
1146         if (num_refs == 0) {
1147                 ret = btrfs_del_item(trans, root, path);
1148         } else {
1149                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1150                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1151                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1152                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1153 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1154                 else {
1155                         struct btrfs_extent_ref_v0 *ref0;
1156                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1157                                         struct btrfs_extent_ref_v0);
1158                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1159                 }
1160 #endif
1161                 btrfs_mark_buffer_dirty(leaf);
1162         }
1163         return ret;
1164 }
1165
1166 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1167                                           struct btrfs_path *path,
1168                                           struct btrfs_extent_inline_ref *iref)
1169 {
1170         struct btrfs_key key;
1171         struct extent_buffer *leaf;
1172         struct btrfs_extent_data_ref *ref1;
1173         struct btrfs_shared_data_ref *ref2;
1174         u32 num_refs = 0;
1175
1176         leaf = path->nodes[0];
1177         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1178         if (iref) {
1179                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1180                     BTRFS_EXTENT_DATA_REF_KEY) {
1181                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1182                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1183                 } else {
1184                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1185                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1186                 }
1187         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1188                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1189                                       struct btrfs_extent_data_ref);
1190                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1191         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1192                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1193                                       struct btrfs_shared_data_ref);
1194                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1195 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1196         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1197                 struct btrfs_extent_ref_v0 *ref0;
1198                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1199                                       struct btrfs_extent_ref_v0);
1200                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1201 #endif
1202         } else {
1203                 WARN_ON(1);
1204         }
1205         return num_refs;
1206 }
1207
1208 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1209                                           struct btrfs_root *root,
1210                                           struct btrfs_path *path,
1211                                           u64 bytenr, u64 parent,
1212                                           u64 root_objectid)
1213 {
1214         struct btrfs_key key;
1215         int ret;
1216
1217         key.objectid = bytenr;
1218         if (parent) {
1219                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1220                 key.offset = parent;
1221         } else {
1222                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1223                 key.offset = root_objectid;
1224         }
1225
1226         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1227         if (ret > 0)
1228                 ret = -ENOENT;
1229 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1230         if (ret == -ENOENT && parent) {
1231                 btrfs_release_path(root, path);
1232                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1233                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1234                 if (ret > 0)
1235                         ret = -ENOENT;
1236         }
1237 #endif
1238         return ret;
1239 }
1240
1241 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1242                                           struct btrfs_root *root,
1243                                           struct btrfs_path *path,
1244                                           u64 bytenr, u64 parent,
1245                                           u64 root_objectid)
1246 {
1247         struct btrfs_key key;
1248         int ret;
1249
1250         key.objectid = bytenr;
1251         if (parent) {
1252                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1253                 key.offset = parent;
1254         } else {
1255                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1256                 key.offset = root_objectid;
1257         }
1258
1259         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1260         btrfs_release_path(root, path);
1261         return ret;
1262 }
1263
1264 static inline int extent_ref_type(u64 parent, u64 owner)
1265 {
1266         int type;
1267         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1268                 if (parent > 0)
1269                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1270                 else
1271                         type = BTRFS_TREE_BLOCK_REF_KEY;
1272         } else {
1273                 if (parent > 0)
1274                         type = BTRFS_SHARED_DATA_REF_KEY;
1275                 else
1276                         type = BTRFS_EXTENT_DATA_REF_KEY;
1277         }
1278         return type;
1279 }
1280
1281 static int find_next_key(struct btrfs_path *path, int level,
1282                          struct btrfs_key *key)
1283
1284 {
1285         for (; level < BTRFS_MAX_LEVEL; level++) {
1286                 if (!path->nodes[level])
1287                         break;
1288                 if (path->slots[level] + 1 >=
1289                     btrfs_header_nritems(path->nodes[level]))
1290                         continue;
1291                 if (level == 0)
1292                         btrfs_item_key_to_cpu(path->nodes[level], key,
1293                                               path->slots[level] + 1);
1294                 else
1295                         btrfs_node_key_to_cpu(path->nodes[level], key,
1296                                               path->slots[level] + 1);
1297                 return 0;
1298         }
1299         return 1;
1300 }
1301
1302 /*
1303  * look for inline back ref. if back ref is found, *ref_ret is set
1304  * to the address of inline back ref, and 0 is returned.
1305  *
1306  * if back ref isn't found, *ref_ret is set to the address where it
1307  * should be inserted, and -ENOENT is returned.
1308  *
1309  * if insert is true and there are too many inline back refs, the path
1310  * points to the extent item, and -EAGAIN is returned.
1311  *
1312  * NOTE: inline back refs are ordered in the same way that back ref
1313  *       items in the tree are ordered.
1314  */
1315 static noinline_for_stack
1316 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1317                                  struct btrfs_root *root,
1318                                  struct btrfs_path *path,
1319                                  struct btrfs_extent_inline_ref **ref_ret,
1320                                  u64 bytenr, u64 num_bytes,
1321                                  u64 parent, u64 root_objectid,
1322                                  u64 owner, u64 offset, int insert)
1323 {
1324         struct btrfs_key key;
1325         struct extent_buffer *leaf;
1326         struct btrfs_extent_item *ei;
1327         struct btrfs_extent_inline_ref *iref;
1328         u64 flags;
1329         u64 item_size;
1330         unsigned long ptr;
1331         unsigned long end;
1332         int extra_size;
1333         int type;
1334         int want;
1335         int ret;
1336         int err = 0;
1337
1338         key.objectid = bytenr;
1339         key.type = BTRFS_EXTENT_ITEM_KEY;
1340         key.offset = num_bytes;
1341
1342         want = extent_ref_type(parent, owner);
1343         if (insert) {
1344                 extra_size = btrfs_extent_inline_ref_size(want);
1345                 path->keep_locks = 1;
1346         } else
1347                 extra_size = -1;
1348         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1349         if (ret < 0) {
1350                 err = ret;
1351                 goto out;
1352         }
1353         BUG_ON(ret);
1354
1355         leaf = path->nodes[0];
1356         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1357 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1358         if (item_size < sizeof(*ei)) {
1359                 if (!insert) {
1360                         err = -ENOENT;
1361                         goto out;
1362                 }
1363                 ret = convert_extent_item_v0(trans, root, path, owner,
1364                                              extra_size);
1365                 if (ret < 0) {
1366                         err = ret;
1367                         goto out;
1368                 }
1369                 leaf = path->nodes[0];
1370                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1371         }
1372 #endif
1373         BUG_ON(item_size < sizeof(*ei));
1374
1375         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1376         flags = btrfs_extent_flags(leaf, ei);
1377
1378         ptr = (unsigned long)(ei + 1);
1379         end = (unsigned long)ei + item_size;
1380
1381         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1382                 ptr += sizeof(struct btrfs_tree_block_info);
1383                 BUG_ON(ptr > end);
1384         } else {
1385                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1386         }
1387
1388         err = -ENOENT;
1389         while (1) {
1390                 if (ptr >= end) {
1391                         WARN_ON(ptr > end);
1392                         break;
1393                 }
1394                 iref = (struct btrfs_extent_inline_ref *)ptr;
1395                 type = btrfs_extent_inline_ref_type(leaf, iref);
1396                 if (want < type)
1397                         break;
1398                 if (want > type) {
1399                         ptr += btrfs_extent_inline_ref_size(type);
1400                         continue;
1401                 }
1402
1403                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1404                         struct btrfs_extent_data_ref *dref;
1405                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1406                         if (match_extent_data_ref(leaf, dref, root_objectid,
1407                                                   owner, offset)) {
1408                                 err = 0;
1409                                 break;
1410                         }
1411                         if (hash_extent_data_ref_item(leaf, dref) <
1412                             hash_extent_data_ref(root_objectid, owner, offset))
1413                                 break;
1414                 } else {
1415                         u64 ref_offset;
1416                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1417                         if (parent > 0) {
1418                                 if (parent == ref_offset) {
1419                                         err = 0;
1420                                         break;
1421                                 }
1422                                 if (ref_offset < parent)
1423                                         break;
1424                         } else {
1425                                 if (root_objectid == ref_offset) {
1426                                         err = 0;
1427                                         break;
1428                                 }
1429                                 if (ref_offset < root_objectid)
1430                                         break;
1431                         }
1432                 }
1433                 ptr += btrfs_extent_inline_ref_size(type);
1434         }
1435         if (err == -ENOENT && insert) {
1436                 if (item_size + extra_size >=
1437                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1438                         err = -EAGAIN;
1439                         goto out;
1440                 }
1441                 /*
1442                  * To add new inline back ref, we have to make sure
1443                  * there is no corresponding back ref item.
1444                  * For simplicity, we just do not add new inline back
1445                  * ref if there is any kind of item for this block
1446                  */
1447                 if (find_next_key(path, 0, &key) == 0 &&
1448                     key.objectid == bytenr &&
1449                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1450                         err = -EAGAIN;
1451                         goto out;
1452                 }
1453         }
1454         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1455 out:
1456         if (insert) {
1457                 path->keep_locks = 0;
1458                 btrfs_unlock_up_safe(path, 1);
1459         }
1460         return err;
1461 }
1462
1463 /*
1464  * helper to add new inline back ref
1465  */
1466 static noinline_for_stack
1467 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1468                                 struct btrfs_root *root,
1469                                 struct btrfs_path *path,
1470                                 struct btrfs_extent_inline_ref *iref,
1471                                 u64 parent, u64 root_objectid,
1472                                 u64 owner, u64 offset, int refs_to_add,
1473                                 struct btrfs_delayed_extent_op *extent_op)
1474 {
1475         struct extent_buffer *leaf;
1476         struct btrfs_extent_item *ei;
1477         unsigned long ptr;
1478         unsigned long end;
1479         unsigned long item_offset;
1480         u64 refs;
1481         int size;
1482         int type;
1483         int ret;
1484
1485         leaf = path->nodes[0];
1486         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1487         item_offset = (unsigned long)iref - (unsigned long)ei;
1488
1489         type = extent_ref_type(parent, owner);
1490         size = btrfs_extent_inline_ref_size(type);
1491
1492         ret = btrfs_extend_item(trans, root, path, size);
1493         BUG_ON(ret);
1494
1495         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1496         refs = btrfs_extent_refs(leaf, ei);
1497         refs += refs_to_add;
1498         btrfs_set_extent_refs(leaf, ei, refs);
1499         if (extent_op)
1500                 __run_delayed_extent_op(extent_op, leaf, ei);
1501
1502         ptr = (unsigned long)ei + item_offset;
1503         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1504         if (ptr < end - size)
1505                 memmove_extent_buffer(leaf, ptr + size, ptr,
1506                                       end - size - ptr);
1507
1508         iref = (struct btrfs_extent_inline_ref *)ptr;
1509         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1510         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1511                 struct btrfs_extent_data_ref *dref;
1512                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1513                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1514                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1515                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1516                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1517         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1518                 struct btrfs_shared_data_ref *sref;
1519                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1520                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1521                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1522         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1523                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1524         } else {
1525                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1526         }
1527         btrfs_mark_buffer_dirty(leaf);
1528         return 0;
1529 }
1530
1531 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1532                                  struct btrfs_root *root,
1533                                  struct btrfs_path *path,
1534                                  struct btrfs_extent_inline_ref **ref_ret,
1535                                  u64 bytenr, u64 num_bytes, u64 parent,
1536                                  u64 root_objectid, u64 owner, u64 offset)
1537 {
1538         int ret;
1539
1540         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1541                                            bytenr, num_bytes, parent,
1542                                            root_objectid, owner, offset, 0);
1543         if (ret != -ENOENT)
1544                 return ret;
1545
1546         btrfs_release_path(root, path);
1547         *ref_ret = NULL;
1548
1549         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1550                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1551                                             root_objectid);
1552         } else {
1553                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1554                                              root_objectid, owner, offset);
1555         }
1556         return ret;
1557 }
1558
1559 /*
1560  * helper to update/remove inline back ref
1561  */
1562 static noinline_for_stack
1563 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1564                                  struct btrfs_root *root,
1565                                  struct btrfs_path *path,
1566                                  struct btrfs_extent_inline_ref *iref,
1567                                  int refs_to_mod,
1568                                  struct btrfs_delayed_extent_op *extent_op)
1569 {
1570         struct extent_buffer *leaf;
1571         struct btrfs_extent_item *ei;
1572         struct btrfs_extent_data_ref *dref = NULL;
1573         struct btrfs_shared_data_ref *sref = NULL;
1574         unsigned long ptr;
1575         unsigned long end;
1576         u32 item_size;
1577         int size;
1578         int type;
1579         int ret;
1580         u64 refs;
1581
1582         leaf = path->nodes[0];
1583         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1584         refs = btrfs_extent_refs(leaf, ei);
1585         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1586         refs += refs_to_mod;
1587         btrfs_set_extent_refs(leaf, ei, refs);
1588         if (extent_op)
1589                 __run_delayed_extent_op(extent_op, leaf, ei);
1590
1591         type = btrfs_extent_inline_ref_type(leaf, iref);
1592
1593         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1594                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1595                 refs = btrfs_extent_data_ref_count(leaf, dref);
1596         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1597                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1598                 refs = btrfs_shared_data_ref_count(leaf, sref);
1599         } else {
1600                 refs = 1;
1601                 BUG_ON(refs_to_mod != -1);
1602         }
1603
1604         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1605         refs += refs_to_mod;
1606
1607         if (refs > 0) {
1608                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1609                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1610                 else
1611                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1612         } else {
1613                 size =  btrfs_extent_inline_ref_size(type);
1614                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1615                 ptr = (unsigned long)iref;
1616                 end = (unsigned long)ei + item_size;
1617                 if (ptr + size < end)
1618                         memmove_extent_buffer(leaf, ptr, ptr + size,
1619                                               end - ptr - size);
1620                 item_size -= size;
1621                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1622                 BUG_ON(ret);
1623         }
1624         btrfs_mark_buffer_dirty(leaf);
1625         return 0;
1626 }
1627
1628 static noinline_for_stack
1629 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1630                                  struct btrfs_root *root,
1631                                  struct btrfs_path *path,
1632                                  u64 bytenr, u64 num_bytes, u64 parent,
1633                                  u64 root_objectid, u64 owner,
1634                                  u64 offset, int refs_to_add,
1635                                  struct btrfs_delayed_extent_op *extent_op)
1636 {
1637         struct btrfs_extent_inline_ref *iref;
1638         int ret;
1639
1640         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1641                                            bytenr, num_bytes, parent,
1642                                            root_objectid, owner, offset, 1);
1643         if (ret == 0) {
1644                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1645                 ret = update_inline_extent_backref(trans, root, path, iref,
1646                                                    refs_to_add, extent_op);
1647         } else if (ret == -ENOENT) {
1648                 ret = setup_inline_extent_backref(trans, root, path, iref,
1649                                                   parent, root_objectid,
1650                                                   owner, offset, refs_to_add,
1651                                                   extent_op);
1652         }
1653         return ret;
1654 }
1655
1656 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1657                                  struct btrfs_root *root,
1658                                  struct btrfs_path *path,
1659                                  u64 bytenr, u64 parent, u64 root_objectid,
1660                                  u64 owner, u64 offset, int refs_to_add)
1661 {
1662         int ret;
1663         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1664                 BUG_ON(refs_to_add != 1);
1665                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1666                                             parent, root_objectid);
1667         } else {
1668                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1669                                              parent, root_objectid,
1670                                              owner, offset, refs_to_add);
1671         }
1672         return ret;
1673 }
1674
1675 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1676                                  struct btrfs_root *root,
1677                                  struct btrfs_path *path,
1678                                  struct btrfs_extent_inline_ref *iref,
1679                                  int refs_to_drop, int is_data)
1680 {
1681         int ret;
1682
1683         BUG_ON(!is_data && refs_to_drop != 1);
1684         if (iref) {
1685                 ret = update_inline_extent_backref(trans, root, path, iref,
1686                                                    -refs_to_drop, NULL);
1687         } else if (is_data) {
1688                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1689         } else {
1690                 ret = btrfs_del_item(trans, root, path);
1691         }
1692         return ret;
1693 }
1694
1695 static void btrfs_issue_discard(struct block_device *bdev,
1696                                 u64 start, u64 len)
1697 {
1698         blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_KERNEL,
1699                         BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
1700 }
1701
1702 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1703                                 u64 num_bytes)
1704 {
1705         int ret;
1706         u64 map_length = num_bytes;
1707         struct btrfs_multi_bio *multi = NULL;
1708
1709         if (!btrfs_test_opt(root, DISCARD))
1710                 return 0;
1711
1712         /* Tell the block device(s) that the sectors can be discarded */
1713         ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
1714                               bytenr, &map_length, &multi, 0);
1715         if (!ret) {
1716                 struct btrfs_bio_stripe *stripe = multi->stripes;
1717                 int i;
1718
1719                 if (map_length > num_bytes)
1720                         map_length = num_bytes;
1721
1722                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1723                         btrfs_issue_discard(stripe->dev->bdev,
1724                                             stripe->physical,
1725                                             map_length);
1726                 }
1727                 kfree(multi);
1728         }
1729
1730         return ret;
1731 }
1732
1733 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1734                          struct btrfs_root *root,
1735                          u64 bytenr, u64 num_bytes, u64 parent,
1736                          u64 root_objectid, u64 owner, u64 offset)
1737 {
1738         int ret;
1739         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1740                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1741
1742         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1743                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1744                                         parent, root_objectid, (int)owner,
1745                                         BTRFS_ADD_DELAYED_REF, NULL);
1746         } else {
1747                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1748                                         parent, root_objectid, owner, offset,
1749                                         BTRFS_ADD_DELAYED_REF, NULL);
1750         }
1751         return ret;
1752 }
1753
1754 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1755                                   struct btrfs_root *root,
1756                                   u64 bytenr, u64 num_bytes,
1757                                   u64 parent, u64 root_objectid,
1758                                   u64 owner, u64 offset, int refs_to_add,
1759                                   struct btrfs_delayed_extent_op *extent_op)
1760 {
1761         struct btrfs_path *path;
1762         struct extent_buffer *leaf;
1763         struct btrfs_extent_item *item;
1764         u64 refs;
1765         int ret;
1766         int err = 0;
1767
1768         path = btrfs_alloc_path();
1769         if (!path)
1770                 return -ENOMEM;
1771
1772         path->reada = 1;
1773         path->leave_spinning = 1;
1774         /* this will setup the path even if it fails to insert the back ref */
1775         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1776                                            path, bytenr, num_bytes, parent,
1777                                            root_objectid, owner, offset,
1778                                            refs_to_add, extent_op);
1779         if (ret == 0)
1780                 goto out;
1781
1782         if (ret != -EAGAIN) {
1783                 err = ret;
1784                 goto out;
1785         }
1786
1787         leaf = path->nodes[0];
1788         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1789         refs = btrfs_extent_refs(leaf, item);
1790         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1791         if (extent_op)
1792                 __run_delayed_extent_op(extent_op, leaf, item);
1793
1794         btrfs_mark_buffer_dirty(leaf);
1795         btrfs_release_path(root->fs_info->extent_root, path);
1796
1797         path->reada = 1;
1798         path->leave_spinning = 1;
1799
1800         /* now insert the actual backref */
1801         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1802                                     path, bytenr, parent, root_objectid,
1803                                     owner, offset, refs_to_add);
1804         BUG_ON(ret);
1805 out:
1806         btrfs_free_path(path);
1807         return err;
1808 }
1809
1810 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1811                                 struct btrfs_root *root,
1812                                 struct btrfs_delayed_ref_node *node,
1813                                 struct btrfs_delayed_extent_op *extent_op,
1814                                 int insert_reserved)
1815 {
1816         int ret = 0;
1817         struct btrfs_delayed_data_ref *ref;
1818         struct btrfs_key ins;
1819         u64 parent = 0;
1820         u64 ref_root = 0;
1821         u64 flags = 0;
1822
1823         ins.objectid = node->bytenr;
1824         ins.offset = node->num_bytes;
1825         ins.type = BTRFS_EXTENT_ITEM_KEY;
1826
1827         ref = btrfs_delayed_node_to_data_ref(node);
1828         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1829                 parent = ref->parent;
1830         else
1831                 ref_root = ref->root;
1832
1833         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1834                 if (extent_op) {
1835                         BUG_ON(extent_op->update_key);
1836                         flags |= extent_op->flags_to_set;
1837                 }
1838                 ret = alloc_reserved_file_extent(trans, root,
1839                                                  parent, ref_root, flags,
1840                                                  ref->objectid, ref->offset,
1841                                                  &ins, node->ref_mod);
1842         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1843                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1844                                              node->num_bytes, parent,
1845                                              ref_root, ref->objectid,
1846                                              ref->offset, node->ref_mod,
1847                                              extent_op);
1848         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1849                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1850                                           node->num_bytes, parent,
1851                                           ref_root, ref->objectid,
1852                                           ref->offset, node->ref_mod,
1853                                           extent_op);
1854         } else {
1855                 BUG();
1856         }
1857         return ret;
1858 }
1859
1860 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1861                                     struct extent_buffer *leaf,
1862                                     struct btrfs_extent_item *ei)
1863 {
1864         u64 flags = btrfs_extent_flags(leaf, ei);
1865         if (extent_op->update_flags) {
1866                 flags |= extent_op->flags_to_set;
1867                 btrfs_set_extent_flags(leaf, ei, flags);
1868         }
1869
1870         if (extent_op->update_key) {
1871                 struct btrfs_tree_block_info *bi;
1872                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1873                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1874                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1875         }
1876 }
1877
1878 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1879                                  struct btrfs_root *root,
1880                                  struct btrfs_delayed_ref_node *node,
1881                                  struct btrfs_delayed_extent_op *extent_op)
1882 {
1883         struct btrfs_key key;
1884         struct btrfs_path *path;
1885         struct btrfs_extent_item *ei;
1886         struct extent_buffer *leaf;
1887         u32 item_size;
1888         int ret;
1889         int err = 0;
1890
1891         path = btrfs_alloc_path();
1892         if (!path)
1893                 return -ENOMEM;
1894
1895         key.objectid = node->bytenr;
1896         key.type = BTRFS_EXTENT_ITEM_KEY;
1897         key.offset = node->num_bytes;
1898
1899         path->reada = 1;
1900         path->leave_spinning = 1;
1901         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1902                                 path, 0, 1);
1903         if (ret < 0) {
1904                 err = ret;
1905                 goto out;
1906         }
1907         if (ret > 0) {
1908                 err = -EIO;
1909                 goto out;
1910         }
1911
1912         leaf = path->nodes[0];
1913         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1914 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1915         if (item_size < sizeof(*ei)) {
1916                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
1917                                              path, (u64)-1, 0);
1918                 if (ret < 0) {
1919                         err = ret;
1920                         goto out;
1921                 }
1922                 leaf = path->nodes[0];
1923                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1924         }
1925 #endif
1926         BUG_ON(item_size < sizeof(*ei));
1927         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1928         __run_delayed_extent_op(extent_op, leaf, ei);
1929
1930         btrfs_mark_buffer_dirty(leaf);
1931 out:
1932         btrfs_free_path(path);
1933         return err;
1934 }
1935
1936 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1937                                 struct btrfs_root *root,
1938                                 struct btrfs_delayed_ref_node *node,
1939                                 struct btrfs_delayed_extent_op *extent_op,
1940                                 int insert_reserved)
1941 {
1942         int ret = 0;
1943         struct btrfs_delayed_tree_ref *ref;
1944         struct btrfs_key ins;
1945         u64 parent = 0;
1946         u64 ref_root = 0;
1947
1948         ins.objectid = node->bytenr;
1949         ins.offset = node->num_bytes;
1950         ins.type = BTRFS_EXTENT_ITEM_KEY;
1951
1952         ref = btrfs_delayed_node_to_tree_ref(node);
1953         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
1954                 parent = ref->parent;
1955         else
1956                 ref_root = ref->root;
1957
1958         BUG_ON(node->ref_mod != 1);
1959         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1960                 BUG_ON(!extent_op || !extent_op->update_flags ||
1961                        !extent_op->update_key);
1962                 ret = alloc_reserved_tree_block(trans, root,
1963                                                 parent, ref_root,
1964                                                 extent_op->flags_to_set,
1965                                                 &extent_op->key,
1966                                                 ref->level, &ins);
1967         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1968                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1969                                              node->num_bytes, parent, ref_root,
1970                                              ref->level, 0, 1, extent_op);
1971         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1972                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1973                                           node->num_bytes, parent, ref_root,
1974                                           ref->level, 0, 1, extent_op);
1975         } else {
1976                 BUG();
1977         }
1978         return ret;
1979 }
1980
1981 /* helper function to actually process a single delayed ref entry */
1982 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
1983                                struct btrfs_root *root,
1984                                struct btrfs_delayed_ref_node *node,
1985                                struct btrfs_delayed_extent_op *extent_op,
1986                                int insert_reserved)
1987 {
1988         int ret;
1989         if (btrfs_delayed_ref_is_head(node)) {
1990                 struct btrfs_delayed_ref_head *head;
1991                 /*
1992                  * we've hit the end of the chain and we were supposed
1993                  * to insert this extent into the tree.  But, it got
1994                  * deleted before we ever needed to insert it, so all
1995                  * we have to do is clean up the accounting
1996                  */
1997                 BUG_ON(extent_op);
1998                 head = btrfs_delayed_node_to_head(node);
1999                 if (insert_reserved) {
2000                         btrfs_pin_extent(root, node->bytenr,
2001                                          node->num_bytes, 1);
2002                         if (head->is_data) {
2003                                 ret = btrfs_del_csums(trans, root,
2004                                                       node->bytenr,
2005                                                       node->num_bytes);
2006                                 BUG_ON(ret);
2007                         }
2008                 }
2009                 mutex_unlock(&head->mutex);
2010                 return 0;
2011         }
2012
2013         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2014             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2015                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2016                                            insert_reserved);
2017         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2018                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2019                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2020                                            insert_reserved);
2021         else
2022                 BUG();
2023         return ret;
2024 }
2025
2026 static noinline struct btrfs_delayed_ref_node *
2027 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2028 {
2029         struct rb_node *node;
2030         struct btrfs_delayed_ref_node *ref;
2031         int action = BTRFS_ADD_DELAYED_REF;
2032 again:
2033         /*
2034          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2035          * this prevents ref count from going down to zero when
2036          * there still are pending delayed ref.
2037          */
2038         node = rb_prev(&head->node.rb_node);
2039         while (1) {
2040                 if (!node)
2041                         break;
2042                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2043                                 rb_node);
2044                 if (ref->bytenr != head->node.bytenr)
2045                         break;
2046                 if (ref->action == action)
2047                         return ref;
2048                 node = rb_prev(node);
2049         }
2050         if (action == BTRFS_ADD_DELAYED_REF) {
2051                 action = BTRFS_DROP_DELAYED_REF;
2052                 goto again;
2053         }
2054         return NULL;
2055 }
2056
2057 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2058                                        struct btrfs_root *root,
2059                                        struct list_head *cluster)
2060 {
2061         struct btrfs_delayed_ref_root *delayed_refs;
2062         struct btrfs_delayed_ref_node *ref;
2063         struct btrfs_delayed_ref_head *locked_ref = NULL;
2064         struct btrfs_delayed_extent_op *extent_op;
2065         int ret;
2066         int count = 0;
2067         int must_insert_reserved = 0;
2068
2069         delayed_refs = &trans->transaction->delayed_refs;
2070         while (1) {
2071                 if (!locked_ref) {
2072                         /* pick a new head ref from the cluster list */
2073                         if (list_empty(cluster))
2074                                 break;
2075
2076                         locked_ref = list_entry(cluster->next,
2077                                      struct btrfs_delayed_ref_head, cluster);
2078
2079                         /* grab the lock that says we are going to process
2080                          * all the refs for this head */
2081                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2082
2083                         /*
2084                          * we may have dropped the spin lock to get the head
2085                          * mutex lock, and that might have given someone else
2086                          * time to free the head.  If that's true, it has been
2087                          * removed from our list and we can move on.
2088                          */
2089                         if (ret == -EAGAIN) {
2090                                 locked_ref = NULL;
2091                                 count++;
2092                                 continue;
2093                         }
2094                 }
2095
2096                 /*
2097                  * record the must insert reserved flag before we
2098                  * drop the spin lock.
2099                  */
2100                 must_insert_reserved = locked_ref->must_insert_reserved;
2101                 locked_ref->must_insert_reserved = 0;
2102
2103                 extent_op = locked_ref->extent_op;
2104                 locked_ref->extent_op = NULL;
2105
2106                 /*
2107                  * locked_ref is the head node, so we have to go one
2108                  * node back for any delayed ref updates
2109                  */
2110                 ref = select_delayed_ref(locked_ref);
2111                 if (!ref) {
2112                         /* All delayed refs have been processed, Go ahead
2113                          * and send the head node to run_one_delayed_ref,
2114                          * so that any accounting fixes can happen
2115                          */
2116                         ref = &locked_ref->node;
2117
2118                         if (extent_op && must_insert_reserved) {
2119                                 kfree(extent_op);
2120                                 extent_op = NULL;
2121                         }
2122
2123                         if (extent_op) {
2124                                 spin_unlock(&delayed_refs->lock);
2125
2126                                 ret = run_delayed_extent_op(trans, root,
2127                                                             ref, extent_op);
2128                                 BUG_ON(ret);
2129                                 kfree(extent_op);
2130
2131                                 cond_resched();
2132                                 spin_lock(&delayed_refs->lock);
2133                                 continue;
2134                         }
2135
2136                         list_del_init(&locked_ref->cluster);
2137                         locked_ref = NULL;
2138                 }
2139
2140                 ref->in_tree = 0;
2141                 rb_erase(&ref->rb_node, &delayed_refs->root);
2142                 delayed_refs->num_entries--;
2143
2144                 spin_unlock(&delayed_refs->lock);
2145
2146                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2147                                           must_insert_reserved);
2148                 BUG_ON(ret);
2149
2150                 btrfs_put_delayed_ref(ref);
2151                 kfree(extent_op);
2152                 count++;
2153
2154                 cond_resched();
2155                 spin_lock(&delayed_refs->lock);
2156         }
2157         return count;
2158 }
2159
2160 /*
2161  * this starts processing the delayed reference count updates and
2162  * extent insertions we have queued up so far.  count can be
2163  * 0, which means to process everything in the tree at the start
2164  * of the run (but not newly added entries), or it can be some target
2165  * number you'd like to process.
2166  */
2167 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2168                            struct btrfs_root *root, unsigned long count)
2169 {
2170         struct rb_node *node;
2171         struct btrfs_delayed_ref_root *delayed_refs;
2172         struct btrfs_delayed_ref_node *ref;
2173         struct list_head cluster;
2174         int ret;
2175         int run_all = count == (unsigned long)-1;
2176         int run_most = 0;
2177
2178         if (root == root->fs_info->extent_root)
2179                 root = root->fs_info->tree_root;
2180
2181         delayed_refs = &trans->transaction->delayed_refs;
2182         INIT_LIST_HEAD(&cluster);
2183 again:
2184         spin_lock(&delayed_refs->lock);
2185         if (count == 0) {
2186                 count = delayed_refs->num_entries * 2;
2187                 run_most = 1;
2188         }
2189         while (1) {
2190                 if (!(run_all || run_most) &&
2191                     delayed_refs->num_heads_ready < 64)
2192                         break;
2193
2194                 /*
2195                  * go find something we can process in the rbtree.  We start at
2196                  * the beginning of the tree, and then build a cluster
2197                  * of refs to process starting at the first one we are able to
2198                  * lock
2199                  */
2200                 ret = btrfs_find_ref_cluster(trans, &cluster,
2201                                              delayed_refs->run_delayed_start);
2202                 if (ret)
2203                         break;
2204
2205                 ret = run_clustered_refs(trans, root, &cluster);
2206                 BUG_ON(ret < 0);
2207
2208                 count -= min_t(unsigned long, ret, count);
2209
2210                 if (count == 0)
2211                         break;
2212         }
2213
2214         if (run_all) {
2215                 node = rb_first(&delayed_refs->root);
2216                 if (!node)
2217                         goto out;
2218                 count = (unsigned long)-1;
2219
2220                 while (node) {
2221                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2222                                        rb_node);
2223                         if (btrfs_delayed_ref_is_head(ref)) {
2224                                 struct btrfs_delayed_ref_head *head;
2225
2226                                 head = btrfs_delayed_node_to_head(ref);
2227                                 atomic_inc(&ref->refs);
2228
2229                                 spin_unlock(&delayed_refs->lock);
2230                                 mutex_lock(&head->mutex);
2231                                 mutex_unlock(&head->mutex);
2232
2233                                 btrfs_put_delayed_ref(ref);
2234                                 cond_resched();
2235                                 goto again;
2236                         }
2237                         node = rb_next(node);
2238                 }
2239                 spin_unlock(&delayed_refs->lock);
2240                 schedule_timeout(1);
2241                 goto again;
2242         }
2243 out:
2244         spin_unlock(&delayed_refs->lock);
2245         return 0;
2246 }
2247
2248 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2249                                 struct btrfs_root *root,
2250                                 u64 bytenr, u64 num_bytes, u64 flags,
2251                                 int is_data)
2252 {
2253         struct btrfs_delayed_extent_op *extent_op;
2254         int ret;
2255
2256         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2257         if (!extent_op)
2258                 return -ENOMEM;
2259
2260         extent_op->flags_to_set = flags;
2261         extent_op->update_flags = 1;
2262         extent_op->update_key = 0;
2263         extent_op->is_data = is_data ? 1 : 0;
2264
2265         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2266         if (ret)
2267                 kfree(extent_op);
2268         return ret;
2269 }
2270
2271 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2272                                       struct btrfs_root *root,
2273                                       struct btrfs_path *path,
2274                                       u64 objectid, u64 offset, u64 bytenr)
2275 {
2276         struct btrfs_delayed_ref_head *head;
2277         struct btrfs_delayed_ref_node *ref;
2278         struct btrfs_delayed_data_ref *data_ref;
2279         struct btrfs_delayed_ref_root *delayed_refs;
2280         struct rb_node *node;
2281         int ret = 0;
2282
2283         ret = -ENOENT;
2284         delayed_refs = &trans->transaction->delayed_refs;
2285         spin_lock(&delayed_refs->lock);
2286         head = btrfs_find_delayed_ref_head(trans, bytenr);
2287         if (!head)
2288                 goto out;
2289
2290         if (!mutex_trylock(&head->mutex)) {
2291                 atomic_inc(&head->node.refs);
2292                 spin_unlock(&delayed_refs->lock);
2293
2294                 btrfs_release_path(root->fs_info->extent_root, path);
2295
2296                 mutex_lock(&head->mutex);
2297                 mutex_unlock(&head->mutex);
2298                 btrfs_put_delayed_ref(&head->node);
2299                 return -EAGAIN;
2300         }
2301
2302         node = rb_prev(&head->node.rb_node);
2303         if (!node)
2304                 goto out_unlock;
2305
2306         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2307
2308         if (ref->bytenr != bytenr)
2309                 goto out_unlock;
2310
2311         ret = 1;
2312         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2313                 goto out_unlock;
2314
2315         data_ref = btrfs_delayed_node_to_data_ref(ref);
2316
2317         node = rb_prev(node);
2318         if (node) {
2319                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2320                 if (ref->bytenr == bytenr)
2321                         goto out_unlock;
2322         }
2323
2324         if (data_ref->root != root->root_key.objectid ||
2325             data_ref->objectid != objectid || data_ref->offset != offset)
2326                 goto out_unlock;
2327
2328         ret = 0;
2329 out_unlock:
2330         mutex_unlock(&head->mutex);
2331 out:
2332         spin_unlock(&delayed_refs->lock);
2333         return ret;
2334 }
2335
2336 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2337                                         struct btrfs_root *root,
2338                                         struct btrfs_path *path,
2339                                         u64 objectid, u64 offset, u64 bytenr)
2340 {
2341         struct btrfs_root *extent_root = root->fs_info->extent_root;
2342         struct extent_buffer *leaf;
2343         struct btrfs_extent_data_ref *ref;
2344         struct btrfs_extent_inline_ref *iref;
2345         struct btrfs_extent_item *ei;
2346         struct btrfs_key key;
2347         u32 item_size;
2348         int ret;
2349
2350         key.objectid = bytenr;
2351         key.offset = (u64)-1;
2352         key.type = BTRFS_EXTENT_ITEM_KEY;
2353
2354         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2355         if (ret < 0)
2356                 goto out;
2357         BUG_ON(ret == 0);
2358
2359         ret = -ENOENT;
2360         if (path->slots[0] == 0)
2361                 goto out;
2362
2363         path->slots[0]--;
2364         leaf = path->nodes[0];
2365         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2366
2367         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2368                 goto out;
2369
2370         ret = 1;
2371         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2372 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2373         if (item_size < sizeof(*ei)) {
2374                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2375                 goto out;
2376         }
2377 #endif
2378         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2379
2380         if (item_size != sizeof(*ei) +
2381             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2382                 goto out;
2383
2384         if (btrfs_extent_generation(leaf, ei) <=
2385             btrfs_root_last_snapshot(&root->root_item))
2386                 goto out;
2387
2388         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2389         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2390             BTRFS_EXTENT_DATA_REF_KEY)
2391                 goto out;
2392
2393         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2394         if (btrfs_extent_refs(leaf, ei) !=
2395             btrfs_extent_data_ref_count(leaf, ref) ||
2396             btrfs_extent_data_ref_root(leaf, ref) !=
2397             root->root_key.objectid ||
2398             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2399             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2400                 goto out;
2401
2402         ret = 0;
2403 out:
2404         return ret;
2405 }
2406
2407 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2408                           struct btrfs_root *root,
2409                           u64 objectid, u64 offset, u64 bytenr)
2410 {
2411         struct btrfs_path *path;
2412         int ret;
2413         int ret2;
2414
2415         path = btrfs_alloc_path();
2416         if (!path)
2417                 return -ENOENT;
2418
2419         do {
2420                 ret = check_committed_ref(trans, root, path, objectid,
2421                                           offset, bytenr);
2422                 if (ret && ret != -ENOENT)
2423                         goto out;
2424
2425                 ret2 = check_delayed_ref(trans, root, path, objectid,
2426                                          offset, bytenr);
2427         } while (ret2 == -EAGAIN);
2428
2429         if (ret2 && ret2 != -ENOENT) {
2430                 ret = ret2;
2431                 goto out;
2432         }
2433
2434         if (ret != -ENOENT || ret2 != -ENOENT)
2435                 ret = 0;
2436 out:
2437         btrfs_free_path(path);
2438         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2439                 WARN_ON(ret > 0);
2440         return ret;
2441 }
2442
2443 #if 0
2444 int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2445                     struct extent_buffer *buf, u32 nr_extents)
2446 {
2447         struct btrfs_key key;
2448         struct btrfs_file_extent_item *fi;
2449         u64 root_gen;
2450         u32 nritems;
2451         int i;
2452         int level;
2453         int ret = 0;
2454         int shared = 0;
2455
2456         if (!root->ref_cows)
2457                 return 0;
2458
2459         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
2460                 shared = 0;
2461                 root_gen = root->root_key.offset;
2462         } else {
2463                 shared = 1;
2464                 root_gen = trans->transid - 1;
2465         }
2466
2467         level = btrfs_header_level(buf);
2468         nritems = btrfs_header_nritems(buf);
2469
2470         if (level == 0) {
2471                 struct btrfs_leaf_ref *ref;
2472                 struct btrfs_extent_info *info;
2473
2474                 ref = btrfs_alloc_leaf_ref(root, nr_extents);
2475                 if (!ref) {
2476                         ret = -ENOMEM;
2477                         goto out;
2478                 }
2479
2480                 ref->root_gen = root_gen;
2481                 ref->bytenr = buf->start;
2482                 ref->owner = btrfs_header_owner(buf);
2483                 ref->generation = btrfs_header_generation(buf);
2484                 ref->nritems = nr_extents;
2485                 info = ref->extents;
2486
2487                 for (i = 0; nr_extents > 0 && i < nritems; i++) {
2488                         u64 disk_bytenr;
2489                         btrfs_item_key_to_cpu(buf, &key, i);
2490                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2491                                 continue;
2492                         fi = btrfs_item_ptr(buf, i,
2493                                             struct btrfs_file_extent_item);
2494                         if (btrfs_file_extent_type(buf, fi) ==
2495                             BTRFS_FILE_EXTENT_INLINE)
2496                                 continue;
2497                         disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2498                         if (disk_bytenr == 0)
2499                                 continue;
2500
2501                         info->bytenr = disk_bytenr;
2502                         info->num_bytes =
2503                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2504                         info->objectid = key.objectid;
2505                         info->offset = key.offset;
2506                         info++;
2507                 }
2508
2509                 ret = btrfs_add_leaf_ref(root, ref, shared);
2510                 if (ret == -EEXIST && shared) {
2511                         struct btrfs_leaf_ref *old;
2512                         old = btrfs_lookup_leaf_ref(root, ref->bytenr);
2513                         BUG_ON(!old);
2514                         btrfs_remove_leaf_ref(root, old);
2515                         btrfs_free_leaf_ref(root, old);
2516                         ret = btrfs_add_leaf_ref(root, ref, shared);
2517                 }
2518                 WARN_ON(ret);
2519                 btrfs_free_leaf_ref(root, ref);
2520         }
2521 out:
2522         return ret;
2523 }
2524
2525 /* when a block goes through cow, we update the reference counts of
2526  * everything that block points to.  The internal pointers of the block
2527  * can be in just about any order, and it is likely to have clusters of
2528  * things that are close together and clusters of things that are not.
2529  *
2530  * To help reduce the seeks that come with updating all of these reference
2531  * counts, sort them by byte number before actual updates are done.
2532  *
2533  * struct refsort is used to match byte number to slot in the btree block.
2534  * we sort based on the byte number and then use the slot to actually
2535  * find the item.
2536  *
2537  * struct refsort is smaller than strcut btrfs_item and smaller than
2538  * struct btrfs_key_ptr.  Since we're currently limited to the page size
2539  * for a btree block, there's no way for a kmalloc of refsorts for a
2540  * single node to be bigger than a page.
2541  */
2542 struct refsort {
2543         u64 bytenr;
2544         u32 slot;
2545 };
2546
2547 /*
2548  * for passing into sort()
2549  */
2550 static int refsort_cmp(const void *a_void, const void *b_void)
2551 {
2552         const struct refsort *a = a_void;
2553         const struct refsort *b = b_void;
2554
2555         if (a->bytenr < b->bytenr)
2556                 return -1;
2557         if (a->bytenr > b->bytenr)
2558                 return 1;
2559         return 0;
2560 }
2561 #endif
2562
2563 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2564                            struct btrfs_root *root,
2565                            struct extent_buffer *buf,
2566                            int full_backref, int inc)
2567 {
2568         u64 bytenr;
2569         u64 num_bytes;
2570         u64 parent;
2571         u64 ref_root;
2572         u32 nritems;
2573         struct btrfs_key key;
2574         struct btrfs_file_extent_item *fi;
2575         int i;
2576         int level;
2577         int ret = 0;
2578         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2579                             u64, u64, u64, u64, u64, u64);
2580
2581         ref_root = btrfs_header_owner(buf);
2582         nritems = btrfs_header_nritems(buf);
2583         level = btrfs_header_level(buf);
2584
2585         if (!root->ref_cows && level == 0)
2586                 return 0;
2587
2588         if (inc)
2589                 process_func = btrfs_inc_extent_ref;
2590         else
2591                 process_func = btrfs_free_extent;
2592
2593         if (full_backref)
2594                 parent = buf->start;
2595         else
2596                 parent = 0;
2597
2598         for (i = 0; i < nritems; i++) {
2599                 if (level == 0) {
2600                         btrfs_item_key_to_cpu(buf, &key, i);
2601                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2602                                 continue;
2603                         fi = btrfs_item_ptr(buf, i,
2604                                             struct btrfs_file_extent_item);
2605                         if (btrfs_file_extent_type(buf, fi) ==
2606                             BTRFS_FILE_EXTENT_INLINE)
2607                                 continue;
2608                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2609                         if (bytenr == 0)
2610                                 continue;
2611
2612                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2613                         key.offset -= btrfs_file_extent_offset(buf, fi);
2614                         ret = process_func(trans, root, bytenr, num_bytes,
2615                                            parent, ref_root, key.objectid,
2616                                            key.offset);
2617                         if (ret)
2618                                 goto fail;
2619                 } else {
2620                         bytenr = btrfs_node_blockptr(buf, i);
2621                         num_bytes = btrfs_level_size(root, level - 1);
2622                         ret = process_func(trans, root, bytenr, num_bytes,
2623                                            parent, ref_root, level - 1, 0);
2624                         if (ret)
2625                                 goto fail;
2626                 }
2627         }
2628         return 0;
2629 fail:
2630         BUG();
2631         return ret;
2632 }
2633
2634 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2635                   struct extent_buffer *buf, int full_backref)
2636 {
2637         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2638 }
2639
2640 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2641                   struct extent_buffer *buf, int full_backref)
2642 {
2643         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2644 }
2645
2646 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2647                                  struct btrfs_root *root,
2648                                  struct btrfs_path *path,
2649                                  struct btrfs_block_group_cache *cache)
2650 {
2651         int ret;
2652         struct btrfs_root *extent_root = root->fs_info->extent_root;
2653         unsigned long bi;
2654         struct extent_buffer *leaf;
2655
2656         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2657         if (ret < 0)
2658                 goto fail;
2659         BUG_ON(ret);
2660
2661         leaf = path->nodes[0];
2662         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2663         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2664         btrfs_mark_buffer_dirty(leaf);
2665         btrfs_release_path(extent_root, path);
2666 fail:
2667         if (ret)
2668                 return ret;
2669         return 0;
2670
2671 }
2672
2673 static struct btrfs_block_group_cache *
2674 next_block_group(struct btrfs_root *root,
2675                  struct btrfs_block_group_cache *cache)
2676 {
2677         struct rb_node *node;
2678         spin_lock(&root->fs_info->block_group_cache_lock);
2679         node = rb_next(&cache->cache_node);
2680         btrfs_put_block_group(cache);
2681         if (node) {
2682                 cache = rb_entry(node, struct btrfs_block_group_cache,
2683                                  cache_node);
2684                 btrfs_get_block_group(cache);
2685         } else
2686                 cache = NULL;
2687         spin_unlock(&root->fs_info->block_group_cache_lock);
2688         return cache;
2689 }
2690
2691 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2692                             struct btrfs_trans_handle *trans,
2693                             struct btrfs_path *path)
2694 {
2695         struct btrfs_root *root = block_group->fs_info->tree_root;
2696         struct inode *inode = NULL;
2697         u64 alloc_hint = 0;
2698         int num_pages = 0;
2699         int retries = 0;
2700         int ret = 0;
2701
2702         /*
2703          * If this block group is smaller than 100 megs don't bother caching the
2704          * block group.
2705          */
2706         if (block_group->key.offset < (100 * 1024 * 1024)) {
2707                 spin_lock(&block_group->lock);
2708                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2709                 spin_unlock(&block_group->lock);
2710                 return 0;
2711         }
2712
2713 again:
2714         inode = lookup_free_space_inode(root, block_group, path);
2715         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2716                 ret = PTR_ERR(inode);
2717                 btrfs_release_path(root, path);
2718                 goto out;
2719         }
2720
2721         if (IS_ERR(inode)) {
2722                 BUG_ON(retries);
2723                 retries++;
2724
2725                 if (block_group->ro)
2726                         goto out_free;
2727
2728                 ret = create_free_space_inode(root, trans, block_group, path);
2729                 if (ret)
2730                         goto out_free;
2731                 goto again;
2732         }
2733
2734         /*
2735          * We want to set the generation to 0, that way if anything goes wrong
2736          * from here on out we know not to trust this cache when we load up next
2737          * time.
2738          */
2739         BTRFS_I(inode)->generation = 0;
2740         ret = btrfs_update_inode(trans, root, inode);
2741         WARN_ON(ret);
2742
2743         if (i_size_read(inode) > 0) {
2744                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2745                                                       inode);
2746                 if (ret)
2747                         goto out_put;
2748         }
2749
2750         spin_lock(&block_group->lock);
2751         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2752                 spin_unlock(&block_group->lock);
2753                 goto out_put;
2754         }
2755         spin_unlock(&block_group->lock);
2756
2757         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2758         if (!num_pages)
2759                 num_pages = 1;
2760
2761         /*
2762          * Just to make absolutely sure we have enough space, we're going to
2763          * preallocate 12 pages worth of space for each block group.  In
2764          * practice we ought to use at most 8, but we need extra space so we can
2765          * add our header and have a terminator between the extents and the
2766          * bitmaps.
2767          */
2768         num_pages *= 16;
2769         num_pages *= PAGE_CACHE_SIZE;
2770
2771         ret = btrfs_check_data_free_space(inode, num_pages);
2772         if (ret)
2773                 goto out_put;
2774
2775         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2776                                               num_pages, num_pages,
2777                                               &alloc_hint);
2778         btrfs_free_reserved_data_space(inode, num_pages);
2779 out_put:
2780         iput(inode);
2781 out_free:
2782         btrfs_release_path(root, path);
2783 out:
2784         spin_lock(&block_group->lock);
2785         if (ret)
2786                 block_group->disk_cache_state = BTRFS_DC_ERROR;
2787         else
2788                 block_group->disk_cache_state = BTRFS_DC_SETUP;
2789         spin_unlock(&block_group->lock);
2790
2791         return ret;
2792 }
2793
2794 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2795                                    struct btrfs_root *root)
2796 {
2797         struct btrfs_block_group_cache *cache;
2798         int err = 0;
2799         struct btrfs_path *path;
2800         u64 last = 0;
2801
2802         path = btrfs_alloc_path();
2803         if (!path)
2804                 return -ENOMEM;
2805
2806 again:
2807         while (1) {
2808                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2809                 while (cache) {
2810                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2811                                 break;
2812                         cache = next_block_group(root, cache);
2813                 }
2814                 if (!cache) {
2815                         if (last == 0)
2816                                 break;
2817                         last = 0;
2818                         continue;
2819                 }
2820                 err = cache_save_setup(cache, trans, path);
2821                 last = cache->key.objectid + cache->key.offset;
2822                 btrfs_put_block_group(cache);
2823         }
2824
2825         while (1) {
2826                 if (last == 0) {
2827                         err = btrfs_run_delayed_refs(trans, root,
2828                                                      (unsigned long)-1);
2829                         BUG_ON(err);
2830                 }
2831
2832                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2833                 while (cache) {
2834                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2835                                 btrfs_put_block_group(cache);
2836                                 goto again;
2837                         }
2838
2839                         if (cache->dirty)
2840                                 break;
2841                         cache = next_block_group(root, cache);
2842                 }
2843                 if (!cache) {
2844                         if (last == 0)
2845                                 break;
2846                         last = 0;
2847                         continue;
2848                 }
2849
2850                 cache->dirty = 0;
2851                 last = cache->key.objectid + cache->key.offset;
2852
2853                 err = write_one_cache_group(trans, root, path, cache);
2854                 BUG_ON(err);
2855                 btrfs_put_block_group(cache);
2856         }
2857
2858         btrfs_free_path(path);
2859         return 0;
2860 }
2861
2862 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2863 {
2864         struct btrfs_block_group_cache *block_group;
2865         int readonly = 0;
2866
2867         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2868         if (!block_group || block_group->ro)
2869                 readonly = 1;
2870         if (block_group)
2871                 btrfs_put_block_group(block_group);
2872         return readonly;
2873 }
2874
2875 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2876                              u64 total_bytes, u64 bytes_used,
2877                              struct btrfs_space_info **space_info)
2878 {
2879         struct btrfs_space_info *found;
2880         int i;
2881         int factor;
2882
2883         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2884                      BTRFS_BLOCK_GROUP_RAID10))
2885                 factor = 2;
2886         else
2887                 factor = 1;
2888
2889         found = __find_space_info(info, flags);
2890         if (found) {
2891                 spin_lock(&found->lock);
2892                 found->total_bytes += total_bytes;
2893                 found->bytes_used += bytes_used;
2894                 found->disk_used += bytes_used * factor;
2895                 found->full = 0;
2896                 spin_unlock(&found->lock);
2897                 *space_info = found;
2898                 return 0;
2899         }
2900         found = kzalloc(sizeof(*found), GFP_NOFS);
2901         if (!found)
2902                 return -ENOMEM;
2903
2904         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2905                 INIT_LIST_HEAD(&found->block_groups[i]);
2906         init_rwsem(&found->groups_sem);
2907         spin_lock_init(&found->lock);
2908         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2909                                 BTRFS_BLOCK_GROUP_SYSTEM |
2910                                 BTRFS_BLOCK_GROUP_METADATA);
2911         found->total_bytes = total_bytes;
2912         found->bytes_used = bytes_used;
2913         found->disk_used = bytes_used * factor;
2914         found->bytes_pinned = 0;
2915         found->bytes_reserved = 0;
2916         found->bytes_readonly = 0;
2917         found->bytes_may_use = 0;
2918         found->full = 0;
2919         found->force_alloc = 0;
2920         *space_info = found;
2921         list_add_rcu(&found->list, &info->space_info);
2922         atomic_set(&found->caching_threads, 0);
2923         return 0;
2924 }
2925
2926 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2927 {
2928         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2929                                    BTRFS_BLOCK_GROUP_RAID1 |
2930                                    BTRFS_BLOCK_GROUP_RAID10 |
2931                                    BTRFS_BLOCK_GROUP_DUP);
2932         if (extra_flags) {
2933                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2934                         fs_info->avail_data_alloc_bits |= extra_flags;
2935                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2936                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2937                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2938                         fs_info->avail_system_alloc_bits |= extra_flags;
2939         }
2940 }
2941
2942 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2943 {
2944         u64 num_devices = root->fs_info->fs_devices->rw_devices;
2945
2946         if (num_devices == 1)
2947                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
2948         if (num_devices < 4)
2949                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
2950
2951         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
2952             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
2953                       BTRFS_BLOCK_GROUP_RAID10))) {
2954                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
2955         }
2956
2957         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
2958             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
2959                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
2960         }
2961
2962         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
2963             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
2964              (flags & BTRFS_BLOCK_GROUP_RAID10) |
2965              (flags & BTRFS_BLOCK_GROUP_DUP)))
2966                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
2967         return flags;
2968 }
2969
2970 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
2971 {
2972         if (flags & BTRFS_BLOCK_GROUP_DATA)
2973                 flags |= root->fs_info->avail_data_alloc_bits &
2974                          root->fs_info->data_alloc_profile;
2975         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2976                 flags |= root->fs_info->avail_system_alloc_bits &
2977                          root->fs_info->system_alloc_profile;
2978         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
2979                 flags |= root->fs_info->avail_metadata_alloc_bits &
2980                          root->fs_info->metadata_alloc_profile;
2981         return btrfs_reduce_alloc_profile(root, flags);
2982 }
2983
2984 static u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
2985 {
2986         u64 flags;
2987
2988         if (data)
2989                 flags = BTRFS_BLOCK_GROUP_DATA;
2990         else if (root == root->fs_info->chunk_root)
2991                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
2992         else
2993                 flags = BTRFS_BLOCK_GROUP_METADATA;
2994
2995         return get_alloc_profile(root, flags);
2996 }
2997
2998 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
2999 {
3000         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3001                                                        BTRFS_BLOCK_GROUP_DATA);
3002 }
3003
3004 /*
3005  * This will check the space that the inode allocates from to make sure we have
3006  * enough space for bytes.
3007  */
3008 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3009 {
3010         struct btrfs_space_info *data_sinfo;
3011         struct btrfs_root *root = BTRFS_I(inode)->root;
3012         u64 used;
3013         int ret = 0, committed = 0, alloc_chunk = 1;
3014
3015         /* make sure bytes are sectorsize aligned */
3016         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3017
3018         if (root == root->fs_info->tree_root) {
3019                 alloc_chunk = 0;
3020                 committed = 1;
3021         }
3022
3023         data_sinfo = BTRFS_I(inode)->space_info;
3024         if (!data_sinfo)
3025                 goto alloc;
3026
3027 again:
3028         /* make sure we have enough space to handle the data first */
3029         spin_lock(&data_sinfo->lock);
3030         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3031                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3032                 data_sinfo->bytes_may_use;
3033
3034         if (used + bytes > data_sinfo->total_bytes) {
3035                 struct btrfs_trans_handle *trans;
3036
3037                 /*
3038                  * if we don't have enough free bytes in this space then we need
3039                  * to alloc a new chunk.
3040                  */
3041                 if (!data_sinfo->full && alloc_chunk) {
3042                         u64 alloc_target;
3043
3044                         data_sinfo->force_alloc = 1;
3045                         spin_unlock(&data_sinfo->lock);
3046 alloc:
3047                         alloc_target = btrfs_get_alloc_profile(root, 1);
3048                         trans = btrfs_join_transaction(root, 1);
3049                         if (IS_ERR(trans))
3050                                 return PTR_ERR(trans);
3051
3052                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3053                                              bytes + 2 * 1024 * 1024,
3054                                              alloc_target, 0);
3055                         btrfs_end_transaction(trans, root);
3056                         if (ret < 0)
3057                                 return ret;
3058
3059                         if (!data_sinfo) {
3060                                 btrfs_set_inode_space_info(root, inode);
3061                                 data_sinfo = BTRFS_I(inode)->space_info;
3062                         }
3063                         goto again;
3064                 }
3065                 spin_unlock(&data_sinfo->lock);
3066
3067                 /* commit the current transaction and try again */
3068                 if (!committed && !root->fs_info->open_ioctl_trans) {
3069                         committed = 1;
3070                         trans = btrfs_join_transaction(root, 1);
3071                         if (IS_ERR(trans))
3072                                 return PTR_ERR(trans);
3073                         ret = btrfs_commit_transaction(trans, root);
3074                         if (ret)
3075                                 return ret;
3076                         goto again;
3077                 }
3078
3079 #if 0 /* I hope we never need this code again, just in case */
3080                 printk(KERN_ERR "no space left, need %llu, %llu bytes_used, "
3081                        "%llu bytes_reserved, " "%llu bytes_pinned, "
3082                        "%llu bytes_readonly, %llu may use %llu total\n",
3083                        (unsigned long long)bytes,
3084                        (unsigned long long)data_sinfo->bytes_used,
3085                        (unsigned long long)data_sinfo->bytes_reserved,
3086                        (unsigned long long)data_sinfo->bytes_pinned,
3087                        (unsigned long long)data_sinfo->bytes_readonly,
3088                        (unsigned long long)data_sinfo->bytes_may_use,
3089                        (unsigned long long)data_sinfo->total_bytes);
3090 #endif
3091                 return -ENOSPC;
3092         }
3093         data_sinfo->bytes_may_use += bytes;
3094         BTRFS_I(inode)->reserved_bytes += bytes;
3095         spin_unlock(&data_sinfo->lock);
3096
3097         return 0;
3098 }
3099
3100 /*
3101  * called when we are clearing an delalloc extent from the
3102  * inode's io_tree or there was an error for whatever reason
3103  * after calling btrfs_check_data_free_space
3104  */
3105 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3106 {
3107         struct btrfs_root *root = BTRFS_I(inode)->root;
3108         struct btrfs_space_info *data_sinfo;
3109
3110         /* make sure bytes are sectorsize aligned */
3111         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3112
3113         data_sinfo = BTRFS_I(inode)->space_info;
3114         spin_lock(&data_sinfo->lock);
3115         data_sinfo->bytes_may_use -= bytes;
3116         BTRFS_I(inode)->reserved_bytes -= bytes;
3117         spin_unlock(&data_sinfo->lock);
3118 }
3119
3120 static void force_metadata_allocation(struct btrfs_fs_info *info)
3121 {
3122         struct list_head *head = &info->space_info;
3123         struct btrfs_space_info *found;
3124
3125         rcu_read_lock();
3126         list_for_each_entry_rcu(found, head, list) {
3127                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3128                         found->force_alloc = 1;
3129         }
3130         rcu_read_unlock();
3131 }
3132
3133 static int should_alloc_chunk(struct btrfs_space_info *sinfo,
3134                               u64 alloc_bytes)
3135 {
3136         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3137
3138         if (sinfo->bytes_used + sinfo->bytes_reserved +
3139             alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3140                 return 0;
3141
3142         if (sinfo->bytes_used + sinfo->bytes_reserved +
3143             alloc_bytes < div_factor(num_bytes, 8))
3144                 return 0;
3145
3146         return 1;
3147 }
3148
3149 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3150                           struct btrfs_root *extent_root, u64 alloc_bytes,
3151                           u64 flags, int force)
3152 {
3153         struct btrfs_space_info *space_info;
3154         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3155         int ret = 0;
3156
3157         mutex_lock(&fs_info->chunk_mutex);
3158
3159         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3160
3161         space_info = __find_space_info(extent_root->fs_info, flags);
3162         if (!space_info) {
3163                 ret = update_space_info(extent_root->fs_info, flags,
3164                                         0, 0, &space_info);
3165                 BUG_ON(ret);
3166         }
3167         BUG_ON(!space_info);
3168
3169         spin_lock(&space_info->lock);
3170         if (space_info->force_alloc)
3171                 force = 1;
3172         if (space_info->full) {
3173                 spin_unlock(&space_info->lock);
3174                 goto out;
3175         }
3176
3177         if (!force && !should_alloc_chunk(space_info, alloc_bytes)) {
3178                 spin_unlock(&space_info->lock);
3179                 goto out;
3180         }
3181         spin_unlock(&space_info->lock);
3182
3183         /*
3184          * if we're doing a data chunk, go ahead and make sure that
3185          * we keep a reasonable number of metadata chunks allocated in the
3186          * FS as well.
3187          */
3188         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3189                 fs_info->data_chunk_allocations++;
3190                 if (!(fs_info->data_chunk_allocations %
3191                       fs_info->metadata_ratio))
3192                         force_metadata_allocation(fs_info);
3193         }
3194
3195         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3196         spin_lock(&space_info->lock);
3197         if (ret)
3198                 space_info->full = 1;
3199         else
3200                 ret = 1;
3201         space_info->force_alloc = 0;
3202         spin_unlock(&space_info->lock);
3203 out:
3204         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3205         return ret;
3206 }
3207
3208 static int maybe_allocate_chunk(struct btrfs_trans_handle *trans,
3209                                 struct btrfs_root *root,
3210                                 struct btrfs_space_info *sinfo, u64 num_bytes)
3211 {
3212         int ret;
3213         int end_trans = 0;
3214
3215         if (sinfo->full)
3216                 return 0;
3217
3218         spin_lock(&sinfo->lock);
3219         ret = should_alloc_chunk(sinfo, num_bytes + 2 * 1024 * 1024);
3220         spin_unlock(&sinfo->lock);
3221         if (!ret)
3222                 return 0;
3223
3224         if (!trans) {
3225                 trans = btrfs_join_transaction(root, 1);
3226                 BUG_ON(IS_ERR(trans));
3227                 end_trans = 1;
3228         }
3229
3230         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3231                              num_bytes + 2 * 1024 * 1024,
3232                              get_alloc_profile(root, sinfo->flags), 0);
3233
3234         if (end_trans)
3235                 btrfs_end_transaction(trans, root);
3236
3237         return ret == 1 ? 1 : 0;
3238 }
3239
3240 /*
3241  * shrink metadata reservation for delalloc
3242  */
3243 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3244                            struct btrfs_root *root, u64 to_reclaim)
3245 {
3246         struct btrfs_block_rsv *block_rsv;
3247         u64 reserved;
3248         u64 max_reclaim;
3249         u64 reclaimed = 0;
3250         int pause = 1;
3251         int ret;
3252
3253         block_rsv = &root->fs_info->delalloc_block_rsv;
3254         spin_lock(&block_rsv->lock);
3255         reserved = block_rsv->reserved;
3256         spin_unlock(&block_rsv->lock);
3257
3258         if (reserved == 0)
3259                 return 0;
3260
3261         max_reclaim = min(reserved, to_reclaim);
3262
3263         while (1) {
3264                 ret = btrfs_start_one_delalloc_inode(root, trans ? 1 : 0);
3265                 if (!ret) {
3266                         __set_current_state(TASK_INTERRUPTIBLE);
3267                         schedule_timeout(pause);
3268                         pause <<= 1;
3269                         if (pause > HZ / 10)
3270                                 pause = HZ / 10;
3271                 } else {
3272                         pause = 1;
3273                 }
3274
3275                 spin_lock(&block_rsv->lock);
3276                 if (reserved > block_rsv->reserved)
3277                         reclaimed = reserved - block_rsv->reserved;
3278                 reserved = block_rsv->reserved;
3279                 spin_unlock(&block_rsv->lock);
3280
3281                 if (reserved == 0 || reclaimed >= max_reclaim)
3282                         break;
3283
3284                 if (trans && trans->transaction->blocked)
3285                         return -EAGAIN;
3286         }
3287         return reclaimed >= to_reclaim;
3288 }
3289
3290 static int should_retry_reserve(struct btrfs_trans_handle *trans,
3291                                 struct btrfs_root *root,
3292                                 struct btrfs_block_rsv *block_rsv,
3293                                 u64 num_bytes, int *retries)
3294 {
3295         struct btrfs_space_info *space_info = block_rsv->space_info;
3296         int ret;
3297
3298         if ((*retries) > 2)
3299                 return -ENOSPC;
3300
3301         ret = maybe_allocate_chunk(trans, root, space_info, num_bytes);
3302         if (ret)
3303                 return 1;
3304
3305         if (trans && trans->transaction->in_commit)
3306                 return -ENOSPC;
3307
3308         ret = shrink_delalloc(trans, root, num_bytes);
3309         if (ret)
3310                 return ret;
3311
3312         spin_lock(&space_info->lock);
3313         if (space_info->bytes_pinned < num_bytes)
3314                 ret = 1;
3315         spin_unlock(&space_info->lock);
3316         if (ret)
3317                 return -ENOSPC;
3318
3319         (*retries)++;
3320
3321         if (trans)
3322                 return -EAGAIN;
3323
3324         trans = btrfs_join_transaction(root, 1);
3325         BUG_ON(IS_ERR(trans));
3326         ret = btrfs_commit_transaction(trans, root);
3327         BUG_ON(ret);
3328
3329         return 1;
3330 }
3331
3332 static int reserve_metadata_bytes(struct btrfs_block_rsv *block_rsv,
3333                                   u64 num_bytes)
3334 {
3335         struct btrfs_space_info *space_info = block_rsv->space_info;
3336         u64 unused;
3337         int ret = -ENOSPC;
3338
3339         spin_lock(&space_info->lock);
3340         unused = space_info->bytes_used + space_info->bytes_reserved +
3341                  space_info->bytes_pinned + space_info->bytes_readonly;
3342
3343         if (unused < space_info->total_bytes)
3344                 unused = space_info->total_bytes - unused;
3345         else
3346                 unused = 0;
3347
3348         if (unused >= num_bytes) {
3349                 if (block_rsv->priority >= 10) {
3350                         space_info->bytes_reserved += num_bytes;
3351                         ret = 0;
3352                 } else {
3353                         if ((unused + block_rsv->reserved) *
3354                             block_rsv->priority >=
3355                             (num_bytes + block_rsv->reserved) * 10) {
3356                                 space_info->bytes_reserved += num_bytes;
3357                                 ret = 0;
3358                         }
3359                 }
3360         }
3361         spin_unlock(&space_info->lock);
3362
3363         return ret;
3364 }
3365
3366 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3367                                              struct btrfs_root *root)
3368 {
3369         struct btrfs_block_rsv *block_rsv;
3370         if (root->ref_cows)
3371                 block_rsv = trans->block_rsv;
3372         else
3373                 block_rsv = root->block_rsv;
3374
3375         if (!block_rsv)
3376                 block_rsv = &root->fs_info->empty_block_rsv;
3377
3378         return block_rsv;
3379 }
3380
3381 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3382                                u64 num_bytes)
3383 {
3384         int ret = -ENOSPC;
3385         spin_lock(&block_rsv->lock);
3386         if (block_rsv->reserved >= num_bytes) {
3387                 block_rsv->reserved -= num_bytes;
3388                 if (block_rsv->reserved < block_rsv->size)
3389                         block_rsv->full = 0;
3390                 ret = 0;
3391         }
3392         spin_unlock(&block_rsv->lock);
3393         return ret;
3394 }
3395
3396 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3397                                 u64 num_bytes, int update_size)
3398 {
3399         spin_lock(&block_rsv->lock);
3400         block_rsv->reserved += num_bytes;
3401         if (update_size)
3402                 block_rsv->size += num_bytes;
3403         else if (block_rsv->reserved >= block_rsv->size)
3404                 block_rsv->full = 1;
3405         spin_unlock(&block_rsv->lock);
3406 }
3407
3408 void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3409                              struct btrfs_block_rsv *dest, u64 num_bytes)
3410 {
3411         struct btrfs_space_info *space_info = block_rsv->space_info;
3412
3413         spin_lock(&block_rsv->lock);
3414         if (num_bytes == (u64)-1)
3415                 num_bytes = block_rsv->size;
3416         block_rsv->size -= num_bytes;
3417         if (block_rsv->reserved >= block_rsv->size) {
3418                 num_bytes = block_rsv->reserved - block_rsv->size;
3419                 block_rsv->reserved = block_rsv->size;
3420                 block_rsv->full = 1;
3421         } else {
3422                 num_bytes = 0;
3423         }
3424         spin_unlock(&block_rsv->lock);
3425
3426         if (num_bytes > 0) {
3427                 if (dest) {
3428                         block_rsv_add_bytes(dest, num_bytes, 0);
3429                 } else {
3430                         spin_lock(&space_info->lock);
3431                         space_info->bytes_reserved -= num_bytes;
3432                         spin_unlock(&space_info->lock);
3433                 }
3434         }
3435 }
3436
3437 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3438                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3439 {
3440         int ret;
3441
3442         ret = block_rsv_use_bytes(src, num_bytes);
3443         if (ret)
3444                 return ret;
3445
3446         block_rsv_add_bytes(dst, num_bytes, 1);
3447         return 0;
3448 }
3449
3450 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3451 {
3452         memset(rsv, 0, sizeof(*rsv));
3453         spin_lock_init(&rsv->lock);
3454         atomic_set(&rsv->usage, 1);
3455         rsv->priority = 6;
3456         INIT_LIST_HEAD(&rsv->list);
3457 }
3458
3459 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3460 {
3461         struct btrfs_block_rsv *block_rsv;
3462         struct btrfs_fs_info *fs_info = root->fs_info;
3463         u64 alloc_target;
3464
3465         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3466         if (!block_rsv)
3467                 return NULL;
3468
3469         btrfs_init_block_rsv(block_rsv);
3470
3471         alloc_target = btrfs_get_alloc_profile(root, 0);
3472         block_rsv->space_info = __find_space_info(fs_info,
3473                                                   BTRFS_BLOCK_GROUP_METADATA);
3474
3475         return block_rsv;
3476 }
3477
3478 void btrfs_free_block_rsv(struct btrfs_root *root,
3479                           struct btrfs_block_rsv *rsv)
3480 {
3481         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3482                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3483                 if (!rsv->durable)
3484                         kfree(rsv);
3485         }
3486 }
3487
3488 /*
3489  * make the block_rsv struct be able to capture freed space.
3490  * the captured space will re-add to the the block_rsv struct
3491  * after transaction commit
3492  */
3493 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3494                                  struct btrfs_block_rsv *block_rsv)
3495 {
3496         block_rsv->durable = 1;
3497         mutex_lock(&fs_info->durable_block_rsv_mutex);
3498         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3499         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3500 }
3501
3502 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3503                         struct btrfs_root *root,
3504                         struct btrfs_block_rsv *block_rsv,
3505                         u64 num_bytes, int *retries)
3506 {
3507         int ret;
3508
3509         if (num_bytes == 0)
3510                 return 0;
3511 again:
3512         ret = reserve_metadata_bytes(block_rsv, num_bytes);
3513         if (!ret) {
3514                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3515                 return 0;
3516         }
3517
3518         ret = should_retry_reserve(trans, root, block_rsv, num_bytes, retries);
3519         if (ret > 0)
3520                 goto again;
3521
3522         return ret;
3523 }
3524
3525 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3526                           struct btrfs_root *root,
3527                           struct btrfs_block_rsv *block_rsv,
3528                           u64 min_reserved, int min_factor)
3529 {
3530         u64 num_bytes = 0;
3531         int commit_trans = 0;
3532         int ret = -ENOSPC;
3533
3534         if (!block_rsv)
3535                 return 0;
3536
3537         spin_lock(&block_rsv->lock);
3538         if (min_factor > 0)
3539                 num_bytes = div_factor(block_rsv->size, min_factor);
3540         if (min_reserved > num_bytes)
3541                 num_bytes = min_reserved;
3542
3543         if (block_rsv->reserved >= num_bytes) {
3544                 ret = 0;
3545         } else {
3546                 num_bytes -= block_rsv->reserved;
3547                 if (block_rsv->durable &&
3548                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3549                         commit_trans = 1;
3550         }
3551         spin_unlock(&block_rsv->lock);
3552         if (!ret)
3553                 return 0;
3554
3555         if (block_rsv->refill_used) {
3556                 ret = reserve_metadata_bytes(block_rsv, num_bytes);
3557                 if (!ret) {
3558                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3559                         return 0;
3560                 }
3561         }
3562
3563         if (commit_trans) {
3564                 if (trans)
3565                         return -EAGAIN;
3566
3567                 trans = btrfs_join_transaction(root, 1);
3568                 BUG_ON(IS_ERR(trans));
3569                 ret = btrfs_commit_transaction(trans, root);
3570                 return 0;
3571         }
3572
3573         WARN_ON(1);
3574         printk(KERN_INFO"block_rsv size %llu reserved %llu freed %llu %llu\n",
3575                 block_rsv->size, block_rsv->reserved,
3576                 block_rsv->freed[0], block_rsv->freed[1]);
3577
3578         return -ENOSPC;
3579 }
3580
3581 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3582                             struct btrfs_block_rsv *dst_rsv,
3583                             u64 num_bytes)
3584 {
3585         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3586 }
3587
3588 void btrfs_block_rsv_release(struct btrfs_root *root,
3589                              struct btrfs_block_rsv *block_rsv,
3590                              u64 num_bytes)
3591 {
3592         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3593         if (global_rsv->full || global_rsv == block_rsv ||
3594             block_rsv->space_info != global_rsv->space_info)
3595                 global_rsv = NULL;
3596         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3597 }
3598
3599 /*
3600  * helper to calculate size of global block reservation.
3601  * the desired value is sum of space used by extent tree,
3602  * checksum tree and root tree
3603  */
3604 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3605 {
3606         struct btrfs_space_info *sinfo;
3607         u64 num_bytes;
3608         u64 meta_used;
3609         u64 data_used;
3610         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3611 #if 0
3612         /*
3613          * per tree used space accounting can be inaccuracy, so we
3614          * can't rely on it.
3615          */
3616         spin_lock(&fs_info->extent_root->accounting_lock);
3617         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item);
3618         spin_unlock(&fs_info->extent_root->accounting_lock);
3619
3620         spin_lock(&fs_info->csum_root->accounting_lock);
3621         num_bytes += btrfs_root_used(&fs_info->csum_root->root_item);
3622         spin_unlock(&fs_info->csum_root->accounting_lock);
3623
3624         spin_lock(&fs_info->tree_root->accounting_lock);
3625         num_bytes += btrfs_root_used(&fs_info->tree_root->root_item);
3626         spin_unlock(&fs_info->tree_root->accounting_lock);
3627 #endif
3628         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3629         spin_lock(&sinfo->lock);
3630         data_used = sinfo->bytes_used;
3631         spin_unlock(&sinfo->lock);
3632
3633         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3634         spin_lock(&sinfo->lock);
3635         meta_used = sinfo->bytes_used;
3636         spin_unlock(&sinfo->lock);
3637
3638         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3639                     csum_size * 2;
3640         num_bytes += div64_u64(data_used + meta_used, 50);
3641
3642         if (num_bytes * 3 > meta_used)
3643                 num_bytes = div64_u64(meta_used, 3);
3644
3645         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3646 }
3647
3648 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3649 {
3650         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3651         struct btrfs_space_info *sinfo = block_rsv->space_info;
3652         u64 num_bytes;
3653
3654         num_bytes = calc_global_metadata_size(fs_info);
3655
3656         spin_lock(&block_rsv->lock);
3657         spin_lock(&sinfo->lock);
3658
3659         block_rsv->size = num_bytes;
3660
3661         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3662                     sinfo->bytes_reserved + sinfo->bytes_readonly;
3663
3664         if (sinfo->total_bytes > num_bytes) {
3665                 num_bytes = sinfo->total_bytes - num_bytes;
3666                 block_rsv->reserved += num_bytes;
3667                 sinfo->bytes_reserved += num_bytes;
3668         }
3669
3670         if (block_rsv->reserved >= block_rsv->size) {
3671                 num_bytes = block_rsv->reserved - block_rsv->size;
3672                 sinfo->bytes_reserved -= num_bytes;
3673                 block_rsv->reserved = block_rsv->size;
3674                 block_rsv->full = 1;
3675         }
3676 #if 0
3677         printk(KERN_INFO"global block rsv size %llu reserved %llu\n",
3678                 block_rsv->size, block_rsv->reserved);
3679 #endif
3680         spin_unlock(&sinfo->lock);
3681         spin_unlock(&block_rsv->lock);
3682 }
3683
3684 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3685 {
3686         struct btrfs_space_info *space_info;
3687
3688         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3689         fs_info->chunk_block_rsv.space_info = space_info;
3690         fs_info->chunk_block_rsv.priority = 10;
3691
3692         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3693         fs_info->global_block_rsv.space_info = space_info;
3694         fs_info->global_block_rsv.priority = 10;
3695         fs_info->global_block_rsv.refill_used = 1;
3696         fs_info->delalloc_block_rsv.space_info = space_info;
3697         fs_info->trans_block_rsv.space_info = space_info;
3698         fs_info->empty_block_rsv.space_info = space_info;
3699         fs_info->empty_block_rsv.priority = 10;
3700
3701         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3702         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3703         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3704         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3705         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3706
3707         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3708
3709         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3710
3711         update_global_block_rsv(fs_info);
3712 }
3713
3714 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3715 {
3716         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3717         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3718         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3719         WARN_ON(fs_info->trans_block_rsv.size > 0);
3720         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3721         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3722         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3723 }
3724
3725 static u64 calc_trans_metadata_size(struct btrfs_root *root, int num_items)
3726 {
3727         return (root->leafsize + root->nodesize * (BTRFS_MAX_LEVEL - 1)) *
3728                 3 * num_items;
3729 }
3730
3731 int btrfs_trans_reserve_metadata(struct btrfs_trans_handle *trans,
3732                                  struct btrfs_root *root,
3733                                  int num_items, int *retries)
3734 {
3735         u64 num_bytes;
3736         int ret;
3737
3738         if (num_items == 0 || root->fs_info->chunk_root == root)
3739                 return 0;
3740
3741         num_bytes = calc_trans_metadata_size(root, num_items);
3742         ret = btrfs_block_rsv_add(trans, root, &root->fs_info->trans_block_rsv,
3743                                   num_bytes, retries);
3744         if (!ret) {
3745                 trans->bytes_reserved += num_bytes;
3746                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3747         }
3748         return ret;
3749 }
3750
3751 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3752                                   struct btrfs_root *root)
3753 {
3754         if (!trans->bytes_reserved)
3755                 return;
3756
3757         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3758         btrfs_block_rsv_release(root, trans->block_rsv,
3759                                 trans->bytes_reserved);
3760         trans->bytes_reserved = 0;
3761 }
3762
3763 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3764                                   struct inode *inode)
3765 {
3766         struct btrfs_root *root = BTRFS_I(inode)->root;
3767         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3768         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3769
3770         /*
3771          * one for deleting orphan item, one for updating inode and
3772          * two for calling btrfs_truncate_inode_items.
3773          *
3774          * btrfs_truncate_inode_items is a delete operation, it frees
3775          * more space than it uses in most cases. So two units of
3776          * metadata space should be enough for calling it many times.
3777          * If all of the metadata space is used, we can commit
3778          * transaction and use space it freed.
3779          */
3780         u64 num_bytes = calc_trans_metadata_size(root, 4);
3781         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3782 }
3783
3784 void btrfs_orphan_release_metadata(struct inode *inode)
3785 {
3786         struct btrfs_root *root = BTRFS_I(inode)->root;
3787         u64 num_bytes = calc_trans_metadata_size(root, 4);
3788         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3789 }
3790
3791 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3792                                 struct btrfs_pending_snapshot *pending)
3793 {
3794         struct btrfs_root *root = pending->root;
3795         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3796         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3797         /*
3798          * two for root back/forward refs, two for directory entries
3799          * and one for root of the snapshot.
3800          */
3801         u64 num_bytes = calc_trans_metadata_size(root, 5);
3802         dst_rsv->space_info = src_rsv->space_info;
3803         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3804 }
3805
3806 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
3807 {
3808         return num_bytes >>= 3;
3809 }
3810
3811 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
3812 {
3813         struct btrfs_root *root = BTRFS_I(inode)->root;
3814         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
3815         u64 to_reserve;
3816         int nr_extents;
3817         int retries = 0;
3818         int ret;
3819
3820         if (btrfs_transaction_in_commit(root->fs_info))
3821                 schedule_timeout(1);
3822
3823         num_bytes = ALIGN(num_bytes, root->sectorsize);
3824 again:
3825         spin_lock(&BTRFS_I(inode)->accounting_lock);
3826         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents) + 1;
3827         if (nr_extents > BTRFS_I(inode)->reserved_extents) {
3828                 nr_extents -= BTRFS_I(inode)->reserved_extents;
3829                 to_reserve = calc_trans_metadata_size(root, nr_extents);
3830         } else {
3831                 nr_extents = 0;
3832                 to_reserve = 0;
3833         }
3834
3835         to_reserve += calc_csum_metadata_size(inode, num_bytes);
3836         ret = reserve_metadata_bytes(block_rsv, to_reserve);
3837         if (ret) {
3838                 spin_unlock(&BTRFS_I(inode)->accounting_lock);
3839                 ret = should_retry_reserve(NULL, root, block_rsv, to_reserve,
3840                                            &retries);
3841                 if (ret > 0)
3842                         goto again;
3843                 return ret;
3844         }
3845
3846         BTRFS_I(inode)->reserved_extents += nr_extents;
3847         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
3848         spin_unlock(&BTRFS_I(inode)->accounting_lock);
3849
3850         block_rsv_add_bytes(block_rsv, to_reserve, 1);
3851
3852         if (block_rsv->size > 512 * 1024 * 1024)
3853                 shrink_delalloc(NULL, root, to_reserve);
3854
3855         return 0;
3856 }
3857
3858 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
3859 {
3860         struct btrfs_root *root = BTRFS_I(inode)->root;
3861         u64 to_free;
3862         int nr_extents;
3863
3864         num_bytes = ALIGN(num_bytes, root->sectorsize);
3865         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
3866
3867         spin_lock(&BTRFS_I(inode)->accounting_lock);
3868         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents);
3869         if (nr_extents < BTRFS_I(inode)->reserved_extents) {
3870                 nr_extents = BTRFS_I(inode)->reserved_extents - nr_extents;
3871                 BTRFS_I(inode)->reserved_extents -= nr_extents;
3872         } else {
3873                 nr_extents = 0;
3874         }
3875         spin_unlock(&BTRFS_I(inode)->accounting_lock);
3876
3877         to_free = calc_csum_metadata_size(inode, num_bytes);
3878         if (nr_extents > 0)
3879                 to_free += calc_trans_metadata_size(root, nr_extents);
3880
3881         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
3882                                 to_free);
3883 }
3884
3885 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
3886 {
3887         int ret;
3888
3889         ret = btrfs_check_data_free_space(inode, num_bytes);
3890         if (ret)
3891                 return ret;
3892
3893         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
3894         if (ret) {
3895                 btrfs_free_reserved_data_space(inode, num_bytes);
3896                 return ret;
3897         }
3898
3899         return 0;
3900 }
3901
3902 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
3903 {
3904         btrfs_delalloc_release_metadata(inode, num_bytes);
3905         btrfs_free_reserved_data_space(inode, num_bytes);
3906 }
3907
3908 static int update_block_group(struct btrfs_trans_handle *trans,
3909                               struct btrfs_root *root,
3910                               u64 bytenr, u64 num_bytes, int alloc)
3911 {
3912         struct btrfs_block_group_cache *cache = NULL;
3913         struct btrfs_fs_info *info = root->fs_info;
3914         u64 total = num_bytes;
3915         u64 old_val;
3916         u64 byte_in_group;
3917         int factor;
3918
3919         /* block accounting for super block */
3920         spin_lock(&info->delalloc_lock);
3921         old_val = btrfs_super_bytes_used(&info->super_copy);
3922         if (alloc)
3923                 old_val += num_bytes;
3924         else
3925                 old_val -= num_bytes;
3926         btrfs_set_super_bytes_used(&info->super_copy, old_val);
3927         spin_unlock(&info->delalloc_lock);
3928
3929         while (total) {
3930                 cache = btrfs_lookup_block_group(info, bytenr);
3931                 if (!cache)
3932                         return -1;
3933                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
3934                                     BTRFS_BLOCK_GROUP_RAID1 |
3935                                     BTRFS_BLOCK_GROUP_RAID10))
3936                         factor = 2;
3937                 else
3938                         factor = 1;
3939
3940                 byte_in_group = bytenr - cache->key.objectid;
3941                 WARN_ON(byte_in_group > cache->key.offset);
3942
3943                 spin_lock(&cache->space_info->lock);
3944                 spin_lock(&cache->lock);
3945
3946                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
3947                     cache->disk_cache_state < BTRFS_DC_CLEAR)
3948                         cache->disk_cache_state = BTRFS_DC_CLEAR;
3949
3950                 cache->dirty = 1;
3951                 old_val = btrfs_block_group_used(&cache->item);
3952                 num_bytes = min(total, cache->key.offset - byte_in_group);
3953                 if (alloc) {
3954                         old_val += num_bytes;
3955                         btrfs_set_block_group_used(&cache->item, old_val);
3956                         cache->reserved -= num_bytes;
3957                         cache->space_info->bytes_reserved -= num_bytes;
3958                         cache->space_info->bytes_used += num_bytes;
3959                         cache->space_info->disk_used += num_bytes * factor;
3960                         spin_unlock(&cache->lock);
3961                         spin_unlock(&cache->space_info->lock);
3962                 } else {
3963                         old_val -= num_bytes;
3964                         btrfs_set_block_group_used(&cache->item, old_val);
3965                         cache->pinned += num_bytes;
3966                         cache->space_info->bytes_pinned += num_bytes;
3967                         cache->space_info->bytes_used -= num_bytes;
3968                         cache->space_info->disk_used -= num_bytes * factor;
3969                         spin_unlock(&cache->lock);
3970                         spin_unlock(&cache->space_info->lock);
3971
3972                         set_extent_dirty(info->pinned_extents,
3973                                          bytenr, bytenr + num_bytes - 1,
3974                                          GFP_NOFS | __GFP_NOFAIL);
3975                 }
3976                 btrfs_put_block_group(cache);
3977                 total -= num_bytes;
3978                 bytenr += num_bytes;
3979         }
3980         return 0;
3981 }
3982
3983 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
3984 {
3985         struct btrfs_block_group_cache *cache;
3986         u64 bytenr;
3987
3988         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
3989         if (!cache)
3990                 return 0;
3991
3992         bytenr = cache->key.objectid;
3993         btrfs_put_block_group(cache);
3994
3995         return bytenr;
3996 }
3997
3998 static int pin_down_extent(struct btrfs_root *root,
3999                            struct btrfs_block_group_cache *cache,
4000                            u64 bytenr, u64 num_bytes, int reserved)
4001 {
4002         spin_lock(&cache->space_info->lock);
4003         spin_lock(&cache->lock);
4004         cache->pinned += num_bytes;
4005         cache->space_info->bytes_pinned += num_bytes;
4006         if (reserved) {
4007                 cache->reserved -= num_bytes;
4008                 cache->space_info->bytes_reserved -= num_bytes;
4009         }
4010         spin_unlock(&cache->lock);
4011         spin_unlock(&cache->space_info->lock);
4012
4013         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4014                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4015         return 0;
4016 }
4017
4018 /*
4019  * this function must be called within transaction
4020  */
4021 int btrfs_pin_extent(struct btrfs_root *root,
4022                      u64 bytenr, u64 num_bytes, int reserved)
4023 {
4024         struct btrfs_block_group_cache *cache;
4025
4026         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4027         BUG_ON(!cache);
4028
4029         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4030
4031         btrfs_put_block_group(cache);
4032         return 0;
4033 }
4034
4035 /*
4036  * update size of reserved extents. this function may return -EAGAIN
4037  * if 'reserve' is true or 'sinfo' is false.
4038  */
4039 static int update_reserved_bytes(struct btrfs_block_group_cache *cache,
4040                                  u64 num_bytes, int reserve, int sinfo)
4041 {
4042         int ret = 0;
4043         if (sinfo) {
4044                 struct btrfs_space_info *space_info = cache->space_info;
4045                 spin_lock(&space_info->lock);
4046                 spin_lock(&cache->lock);
4047                 if (reserve) {
4048                         if (cache->ro) {
4049                                 ret = -EAGAIN;
4050                         } else {
4051                                 cache->reserved += num_bytes;
4052                                 space_info->bytes_reserved += num_bytes;
4053                         }
4054                 } else {
4055                         if (cache->ro)
4056                                 space_info->bytes_readonly += num_bytes;
4057                         cache->reserved -= num_bytes;
4058                         space_info->bytes_reserved -= num_bytes;
4059                 }
4060                 spin_unlock(&cache->lock);
4061                 spin_unlock(&space_info->lock);
4062         } else {
4063                 spin_lock(&cache->lock);
4064                 if (cache->ro) {
4065                         ret = -EAGAIN;
4066                 } else {
4067                         if (reserve)
4068                                 cache->reserved += num_bytes;
4069                         else
4070                                 cache->reserved -= num_bytes;
4071                 }
4072                 spin_unlock(&cache->lock);
4073         }
4074         return ret;
4075 }
4076
4077 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4078                                 struct btrfs_root *root)
4079 {
4080         struct btrfs_fs_info *fs_info = root->fs_info;
4081         struct btrfs_caching_control *next;
4082         struct btrfs_caching_control *caching_ctl;
4083         struct btrfs_block_group_cache *cache;
4084
4085         down_write(&fs_info->extent_commit_sem);
4086
4087         list_for_each_entry_safe(caching_ctl, next,
4088                                  &fs_info->caching_block_groups, list) {
4089                 cache = caching_ctl->block_group;
4090                 if (block_group_cache_done(cache)) {
4091                         cache->last_byte_to_unpin = (u64)-1;
4092                         list_del_init(&caching_ctl->list);
4093                         put_caching_control(caching_ctl);
4094                 } else {
4095                         cache->last_byte_to_unpin = caching_ctl->progress;
4096                 }
4097         }
4098
4099         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4100                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4101         else
4102                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4103
4104         up_write(&fs_info->extent_commit_sem);
4105
4106         update_global_block_rsv(fs_info);
4107         return 0;
4108 }
4109
4110 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4111 {
4112         struct btrfs_fs_info *fs_info = root->fs_info;
4113         struct btrfs_block_group_cache *cache = NULL;
4114         u64 len;
4115
4116         while (start <= end) {
4117                 if (!cache ||
4118                     start >= cache->key.objectid + cache->key.offset) {
4119                         if (cache)
4120                                 btrfs_put_block_group(cache);
4121                         cache = btrfs_lookup_block_group(fs_info, start);
4122                         BUG_ON(!cache);
4123                 }
4124
4125                 len = cache->key.objectid + cache->key.offset - start;
4126                 len = min(len, end + 1 - start);
4127
4128                 if (start < cache->last_byte_to_unpin) {
4129                         len = min(len, cache->last_byte_to_unpin - start);
4130                         btrfs_add_free_space(cache, start, len);
4131                 }
4132
4133                 start += len;
4134
4135                 spin_lock(&cache->space_info->lock);
4136                 spin_lock(&cache->lock);
4137                 cache->pinned -= len;
4138                 cache->space_info->bytes_pinned -= len;
4139                 if (cache->ro) {
4140                         cache->space_info->bytes_readonly += len;
4141                 } else if (cache->reserved_pinned > 0) {
4142                         len = min(len, cache->reserved_pinned);
4143                         cache->reserved_pinned -= len;
4144                         cache->space_info->bytes_reserved += len;
4145                 }
4146                 spin_unlock(&cache->lock);
4147                 spin_unlock(&cache->space_info->lock);
4148         }
4149
4150         if (cache)
4151                 btrfs_put_block_group(cache);
4152         return 0;
4153 }
4154
4155 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4156                                struct btrfs_root *root)
4157 {
4158         struct btrfs_fs_info *fs_info = root->fs_info;
4159         struct extent_io_tree *unpin;
4160         struct btrfs_block_rsv *block_rsv;
4161         struct btrfs_block_rsv *next_rsv;
4162         u64 start;
4163         u64 end;
4164         int idx;
4165         int ret;
4166
4167         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4168                 unpin = &fs_info->freed_extents[1];
4169         else
4170                 unpin = &fs_info->freed_extents[0];
4171
4172         while (1) {
4173                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4174                                             EXTENT_DIRTY);
4175                 if (ret)
4176                         break;
4177
4178                 ret = btrfs_discard_extent(root, start, end + 1 - start);
4179
4180                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4181                 unpin_extent_range(root, start, end);
4182                 cond_resched();
4183         }
4184
4185         mutex_lock(&fs_info->durable_block_rsv_mutex);
4186         list_for_each_entry_safe(block_rsv, next_rsv,
4187                                  &fs_info->durable_block_rsv_list, list) {
4188
4189                 idx = trans->transid & 0x1;
4190                 if (block_rsv->freed[idx] > 0) {
4191                         block_rsv_add_bytes(block_rsv,
4192                                             block_rsv->freed[idx], 0);
4193                         block_rsv->freed[idx] = 0;
4194                 }
4195                 if (atomic_read(&block_rsv->usage) == 0) {
4196                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4197
4198                         if (block_rsv->freed[0] == 0 &&
4199                             block_rsv->freed[1] == 0) {
4200                                 list_del_init(&block_rsv->list);
4201                                 kfree(block_rsv);
4202                         }
4203                 } else {
4204                         btrfs_block_rsv_release(root, block_rsv, 0);
4205                 }
4206         }
4207         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4208
4209         return 0;
4210 }
4211
4212 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4213                                 struct btrfs_root *root,
4214                                 u64 bytenr, u64 num_bytes, u64 parent,
4215                                 u64 root_objectid, u64 owner_objectid,
4216                                 u64 owner_offset, int refs_to_drop,
4217                                 struct btrfs_delayed_extent_op *extent_op)
4218 {
4219         struct btrfs_key key;
4220         struct btrfs_path *path;
4221         struct btrfs_fs_info *info = root->fs_info;
4222         struct btrfs_root *extent_root = info->extent_root;
4223         struct extent_buffer *leaf;
4224         struct btrfs_extent_item *ei;
4225         struct btrfs_extent_inline_ref *iref;
4226         int ret;
4227         int is_data;
4228         int extent_slot = 0;
4229         int found_extent = 0;
4230         int num_to_del = 1;
4231         u32 item_size;
4232         u64 refs;
4233
4234         path = btrfs_alloc_path();
4235         if (!path)
4236                 return -ENOMEM;
4237
4238         path->reada = 1;
4239         path->leave_spinning = 1;
4240
4241         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4242         BUG_ON(!is_data && refs_to_drop != 1);
4243
4244         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4245                                     bytenr, num_bytes, parent,
4246                                     root_objectid, owner_objectid,
4247                                     owner_offset);
4248         if (ret == 0) {
4249                 extent_slot = path->slots[0];
4250                 while (extent_slot >= 0) {
4251                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4252                                               extent_slot);
4253                         if (key.objectid != bytenr)
4254                                 break;
4255                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4256                             key.offset == num_bytes) {
4257                                 found_extent = 1;
4258                                 break;
4259                         }
4260                         if (path->slots[0] - extent_slot > 5)
4261                                 break;
4262                         extent_slot--;
4263                 }
4264 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4265                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4266                 if (found_extent && item_size < sizeof(*ei))
4267                         found_extent = 0;
4268 #endif
4269                 if (!found_extent) {
4270                         BUG_ON(iref);
4271                         ret = remove_extent_backref(trans, extent_root, path,
4272                                                     NULL, refs_to_drop,
4273                                                     is_data);
4274                         BUG_ON(ret);
4275                         btrfs_release_path(extent_root, path);
4276                         path->leave_spinning = 1;
4277
4278                         key.objectid = bytenr;
4279                         key.type = BTRFS_EXTENT_ITEM_KEY;
4280                         key.offset = num_bytes;
4281
4282                         ret = btrfs_search_slot(trans, extent_root,
4283                                                 &key, path, -1, 1);
4284                         if (ret) {
4285                                 printk(KERN_ERR "umm, got %d back from search"
4286                                        ", was looking for %llu\n", ret,
4287                                        (unsigned long long)bytenr);
4288                                 btrfs_print_leaf(extent_root, path->nodes[0]);
4289                         }
4290                         BUG_ON(ret);
4291                         extent_slot = path->slots[0];
4292                 }
4293         } else {
4294                 btrfs_print_leaf(extent_root, path->nodes[0]);
4295                 WARN_ON(1);
4296                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4297                        "parent %llu root %llu  owner %llu offset %llu\n",
4298                        (unsigned long long)bytenr,
4299                        (unsigned long long)parent,
4300                        (unsigned long long)root_objectid,
4301                        (unsigned long long)owner_objectid,
4302                        (unsigned long long)owner_offset);
4303         }
4304
4305         leaf = path->nodes[0];
4306         item_size = btrfs_item_size_nr(leaf, extent_slot);
4307 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4308         if (item_size < sizeof(*ei)) {
4309                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4310                 ret = convert_extent_item_v0(trans, extent_root, path,
4311                                              owner_objectid, 0);
4312                 BUG_ON(ret < 0);
4313
4314                 btrfs_release_path(extent_root, path);
4315                 path->leave_spinning = 1;
4316
4317                 key.objectid = bytenr;
4318                 key.type = BTRFS_EXTENT_ITEM_KEY;
4319                 key.offset = num_bytes;
4320
4321                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4322                                         -1, 1);
4323                 if (ret) {
4324                         printk(KERN_ERR "umm, got %d back from search"
4325                                ", was looking for %llu\n", ret,
4326                                (unsigned long long)bytenr);
4327                         btrfs_print_leaf(extent_root, path->nodes[0]);
4328                 }
4329                 BUG_ON(ret);
4330                 extent_slot = path->slots[0];
4331                 leaf = path->nodes[0];
4332                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4333         }
4334 #endif
4335         BUG_ON(item_size < sizeof(*ei));
4336         ei = btrfs_item_ptr(leaf, extent_slot,
4337                             struct btrfs_extent_item);
4338         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4339                 struct btrfs_tree_block_info *bi;
4340                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4341                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4342                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4343         }
4344
4345         refs = btrfs_extent_refs(leaf, ei);
4346         BUG_ON(refs < refs_to_drop);
4347         refs -= refs_to_drop;
4348
4349         if (refs > 0) {
4350                 if (extent_op)
4351                         __run_delayed_extent_op(extent_op, leaf, ei);
4352                 /*
4353                  * In the case of inline back ref, reference count will
4354                  * be updated by remove_extent_backref
4355                  */
4356                 if (iref) {
4357                         BUG_ON(!found_extent);
4358                 } else {
4359                         btrfs_set_extent_refs(leaf, ei, refs);
4360                         btrfs_mark_buffer_dirty(leaf);
4361                 }
4362                 if (found_extent) {
4363                         ret = remove_extent_backref(trans, extent_root, path,
4364                                                     iref, refs_to_drop,
4365                                                     is_data);
4366                         BUG_ON(ret);
4367                 }
4368         } else {
4369                 if (found_extent) {
4370                         BUG_ON(is_data && refs_to_drop !=
4371                                extent_data_ref_count(root, path, iref));
4372                         if (iref) {
4373                                 BUG_ON(path->slots[0] != extent_slot);
4374                         } else {
4375                                 BUG_ON(path->slots[0] != extent_slot + 1);
4376                                 path->slots[0] = extent_slot;
4377                                 num_to_del = 2;
4378                         }
4379                 }
4380
4381                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4382                                       num_to_del);
4383                 BUG_ON(ret);
4384                 btrfs_release_path(extent_root, path);
4385
4386                 if (is_data) {
4387                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4388                         BUG_ON(ret);
4389                 } else {
4390                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4391                              bytenr >> PAGE_CACHE_SHIFT,
4392                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4393                 }
4394
4395                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4396                 BUG_ON(ret);
4397         }
4398         btrfs_free_path(path);
4399         return ret;
4400 }
4401
4402 /*
4403  * when we free an block, it is possible (and likely) that we free the last
4404  * delayed ref for that extent as well.  This searches the delayed ref tree for
4405  * a given extent, and if there are no other delayed refs to be processed, it
4406  * removes it from the tree.
4407  */
4408 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4409                                       struct btrfs_root *root, u64 bytenr)
4410 {
4411         struct btrfs_delayed_ref_head *head;
4412         struct btrfs_delayed_ref_root *delayed_refs;
4413         struct btrfs_delayed_ref_node *ref;
4414         struct rb_node *node;
4415         int ret = 0;
4416
4417         delayed_refs = &trans->transaction->delayed_refs;
4418         spin_lock(&delayed_refs->lock);
4419         head = btrfs_find_delayed_ref_head(trans, bytenr);
4420         if (!head)
4421                 goto out;
4422
4423         node = rb_prev(&head->node.rb_node);
4424         if (!node)
4425                 goto out;
4426
4427         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4428
4429         /* there are still entries for this ref, we can't drop it */
4430         if (ref->bytenr == bytenr)
4431                 goto out;
4432
4433         if (head->extent_op) {
4434                 if (!head->must_insert_reserved)
4435                         goto out;
4436                 kfree(head->extent_op);
4437                 head->extent_op = NULL;
4438         }
4439
4440         /*
4441          * waiting for the lock here would deadlock.  If someone else has it
4442          * locked they are already in the process of dropping it anyway
4443          */
4444         if (!mutex_trylock(&head->mutex))
4445                 goto out;
4446
4447         /*
4448          * at this point we have a head with no other entries.  Go
4449          * ahead and process it.
4450          */
4451         head->node.in_tree = 0;
4452         rb_erase(&head->node.rb_node, &delayed_refs->root);
4453
4454         delayed_refs->num_entries--;
4455
4456         /*
4457          * we don't take a ref on the node because we're removing it from the
4458          * tree, so we just steal the ref the tree was holding.
4459          */
4460         delayed_refs->num_heads--;
4461         if (list_empty(&head->cluster))
4462                 delayed_refs->num_heads_ready--;
4463
4464         list_del_init(&head->cluster);
4465         spin_unlock(&delayed_refs->lock);
4466
4467         BUG_ON(head->extent_op);
4468         if (head->must_insert_reserved)
4469                 ret = 1;
4470
4471         mutex_unlock(&head->mutex);
4472         btrfs_put_delayed_ref(&head->node);
4473         return ret;
4474 out:
4475         spin_unlock(&delayed_refs->lock);
4476         return 0;
4477 }
4478
4479 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4480                            struct btrfs_root *root,
4481                            struct extent_buffer *buf,
4482                            u64 parent, int last_ref)
4483 {
4484         struct btrfs_block_rsv *block_rsv;
4485         struct btrfs_block_group_cache *cache = NULL;
4486         int ret;
4487
4488         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4489                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4490                                                 parent, root->root_key.objectid,
4491                                                 btrfs_header_level(buf),
4492                                                 BTRFS_DROP_DELAYED_REF, NULL);
4493                 BUG_ON(ret);
4494         }
4495
4496         if (!last_ref)
4497                 return;
4498
4499         block_rsv = get_block_rsv(trans, root);
4500         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4501         if (block_rsv->space_info != cache->space_info)
4502                 goto out;
4503
4504         if (btrfs_header_generation(buf) == trans->transid) {
4505                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4506                         ret = check_ref_cleanup(trans, root, buf->start);
4507                         if (!ret)
4508                                 goto pin;
4509                 }
4510
4511                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4512                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4513                         goto pin;
4514                 }
4515
4516                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4517
4518                 btrfs_add_free_space(cache, buf->start, buf->len);
4519                 ret = update_reserved_bytes(cache, buf->len, 0, 0);
4520                 if (ret == -EAGAIN) {
4521                         /* block group became read-only */
4522                         update_reserved_bytes(cache, buf->len, 0, 1);
4523                         goto out;
4524                 }
4525
4526                 ret = 1;
4527                 spin_lock(&block_rsv->lock);
4528                 if (block_rsv->reserved < block_rsv->size) {
4529                         block_rsv->reserved += buf->len;
4530                         ret = 0;
4531                 }
4532                 spin_unlock(&block_rsv->lock);
4533
4534                 if (ret) {
4535                         spin_lock(&cache->space_info->lock);
4536                         cache->space_info->bytes_reserved -= buf->len;
4537                         spin_unlock(&cache->space_info->lock);
4538                 }
4539                 goto out;
4540         }
4541 pin:
4542         if (block_rsv->durable && !cache->ro) {
4543                 ret = 0;
4544                 spin_lock(&cache->lock);
4545                 if (!cache->ro) {
4546                         cache->reserved_pinned += buf->len;
4547                         ret = 1;
4548                 }
4549                 spin_unlock(&cache->lock);
4550
4551                 if (ret) {
4552                         spin_lock(&block_rsv->lock);
4553                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4554                         spin_unlock(&block_rsv->lock);
4555                 }
4556         }
4557 out:
4558         btrfs_put_block_group(cache);
4559 }
4560
4561 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4562                       struct btrfs_root *root,
4563                       u64 bytenr, u64 num_bytes, u64 parent,
4564                       u64 root_objectid, u64 owner, u64 offset)
4565 {
4566         int ret;
4567
4568         /*
4569          * tree log blocks never actually go into the extent allocation
4570          * tree, just update pinning info and exit early.
4571          */
4572         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4573                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4574                 /* unlocks the pinned mutex */
4575                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4576                 ret = 0;
4577         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4578                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4579                                         parent, root_objectid, (int)owner,
4580                                         BTRFS_DROP_DELAYED_REF, NULL);
4581                 BUG_ON(ret);
4582         } else {
4583                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4584                                         parent, root_objectid, owner,
4585                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4586                 BUG_ON(ret);
4587         }
4588         return ret;
4589 }
4590
4591 static u64 stripe_align(struct btrfs_root *root, u64 val)
4592 {
4593         u64 mask = ((u64)root->stripesize - 1);
4594         u64 ret = (val + mask) & ~mask;
4595         return ret;
4596 }
4597
4598 /*
4599  * when we wait for progress in the block group caching, its because
4600  * our allocation attempt failed at least once.  So, we must sleep
4601  * and let some progress happen before we try again.
4602  *
4603  * This function will sleep at least once waiting for new free space to
4604  * show up, and then it will check the block group free space numbers
4605  * for our min num_bytes.  Another option is to have it go ahead
4606  * and look in the rbtree for a free extent of a given size, but this
4607  * is a good start.
4608  */
4609 static noinline int
4610 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4611                                 u64 num_bytes)
4612 {
4613         struct btrfs_caching_control *caching_ctl;
4614         DEFINE_WAIT(wait);
4615
4616         caching_ctl = get_caching_control(cache);
4617         if (!caching_ctl)
4618                 return 0;
4619
4620         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4621                    (cache->free_space >= num_bytes));
4622
4623         put_caching_control(caching_ctl);
4624         return 0;
4625 }
4626
4627 static noinline int
4628 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4629 {
4630         struct btrfs_caching_control *caching_ctl;
4631         DEFINE_WAIT(wait);
4632
4633         caching_ctl = get_caching_control(cache);
4634         if (!caching_ctl)
4635                 return 0;
4636
4637         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4638
4639         put_caching_control(caching_ctl);
4640         return 0;
4641 }
4642
4643 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4644 {
4645         int index;
4646         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4647                 index = 0;
4648         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4649                 index = 1;
4650         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4651                 index = 2;
4652         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4653                 index = 3;
4654         else
4655                 index = 4;
4656         return index;
4657 }
4658
4659 enum btrfs_loop_type {
4660         LOOP_FIND_IDEAL = 0,
4661         LOOP_CACHING_NOWAIT = 1,
4662         LOOP_CACHING_WAIT = 2,
4663         LOOP_ALLOC_CHUNK = 3,
4664         LOOP_NO_EMPTY_SIZE = 4,
4665 };
4666
4667 /*
4668  * walks the btree of allocated extents and find a hole of a given size.
4669  * The key ins is changed to record the hole:
4670  * ins->objectid == block start
4671  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4672  * ins->offset == number of blocks
4673  * Any available blocks before search_start are skipped.
4674  */
4675 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4676                                      struct btrfs_root *orig_root,
4677                                      u64 num_bytes, u64 empty_size,
4678                                      u64 search_start, u64 search_end,
4679                                      u64 hint_byte, struct btrfs_key *ins,
4680                                      int data)
4681 {
4682         int ret = 0;
4683         struct btrfs_root *root = orig_root->fs_info->extent_root;
4684         struct btrfs_free_cluster *last_ptr = NULL;
4685         struct btrfs_block_group_cache *block_group = NULL;
4686         int empty_cluster = 2 * 1024 * 1024;
4687         int allowed_chunk_alloc = 0;
4688         int done_chunk_alloc = 0;
4689         struct btrfs_space_info *space_info;
4690         int last_ptr_loop = 0;
4691         int loop = 0;
4692         int index = 0;
4693         bool found_uncached_bg = false;
4694         bool failed_cluster_refill = false;
4695         bool failed_alloc = false;
4696         u64 ideal_cache_percent = 0;
4697         u64 ideal_cache_offset = 0;
4698
4699         WARN_ON(num_bytes < root->sectorsize);
4700         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4701         ins->objectid = 0;
4702         ins->offset = 0;
4703
4704         space_info = __find_space_info(root->fs_info, data);
4705         if (!space_info) {
4706                 printk(KERN_ERR "No space info for %d\n", data);
4707                 return -ENOSPC;
4708         }
4709
4710         if (orig_root->ref_cows || empty_size)
4711                 allowed_chunk_alloc = 1;
4712
4713         if (data & BTRFS_BLOCK_GROUP_METADATA) {
4714                 last_ptr = &root->fs_info->meta_alloc_cluster;
4715                 if (!btrfs_test_opt(root, SSD))
4716                         empty_cluster = 64 * 1024;
4717         }
4718
4719         if ((data & BTRFS_BLOCK_GROUP_DATA) && btrfs_test_opt(root, SSD)) {
4720                 last_ptr = &root->fs_info->data_alloc_cluster;
4721         }
4722
4723         if (last_ptr) {
4724                 spin_lock(&last_ptr->lock);
4725                 if (last_ptr->block_group)
4726                         hint_byte = last_ptr->window_start;
4727                 spin_unlock(&last_ptr->lock);
4728         }
4729
4730         search_start = max(search_start, first_logical_byte(root, 0));
4731         search_start = max(search_start, hint_byte);
4732
4733         if (!last_ptr)
4734                 empty_cluster = 0;
4735
4736         if (search_start == hint_byte) {
4737 ideal_cache:
4738                 block_group = btrfs_lookup_block_group(root->fs_info,
4739                                                        search_start);
4740                 /*
4741                  * we don't want to use the block group if it doesn't match our
4742                  * allocation bits, or if its not cached.
4743                  *
4744                  * However if we are re-searching with an ideal block group
4745                  * picked out then we don't care that the block group is cached.
4746                  */
4747                 if (block_group && block_group_bits(block_group, data) &&
4748                     (block_group->cached != BTRFS_CACHE_NO ||
4749                      search_start == ideal_cache_offset)) {
4750                         down_read(&space_info->groups_sem);
4751                         if (list_empty(&block_group->list) ||
4752                             block_group->ro) {
4753                                 /*
4754                                  * someone is removing this block group,
4755                                  * we can't jump into the have_block_group
4756                                  * target because our list pointers are not
4757                                  * valid
4758                                  */
4759                                 btrfs_put_block_group(block_group);
4760                                 up_read(&space_info->groups_sem);
4761                         } else {
4762                                 index = get_block_group_index(block_group);
4763                                 goto have_block_group;
4764                         }
4765                 } else if (block_group) {
4766                         btrfs_put_block_group(block_group);
4767                 }
4768         }
4769 search:
4770         down_read(&space_info->groups_sem);
4771         list_for_each_entry(block_group, &space_info->block_groups[index],
4772                             list) {
4773                 u64 offset;
4774                 int cached;
4775
4776                 btrfs_get_block_group(block_group);
4777                 search_start = block_group->key.objectid;
4778
4779 have_block_group:
4780                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
4781                         u64 free_percent;
4782
4783                         free_percent = btrfs_block_group_used(&block_group->item);
4784                         free_percent *= 100;
4785                         free_percent = div64_u64(free_percent,
4786                                                  block_group->key.offset);
4787                         free_percent = 100 - free_percent;
4788                         if (free_percent > ideal_cache_percent &&
4789                             likely(!block_group->ro)) {
4790                                 ideal_cache_offset = block_group->key.objectid;
4791                                 ideal_cache_percent = free_percent;
4792                         }
4793
4794                         /*
4795                          * We only want to start kthread caching if we are at
4796                          * the point where we will wait for caching to make
4797                          * progress, or if our ideal search is over and we've
4798                          * found somebody to start caching.
4799                          */
4800                         if (loop > LOOP_CACHING_NOWAIT ||
4801                             (loop > LOOP_FIND_IDEAL &&
4802                              atomic_read(&space_info->caching_threads) < 2)) {
4803                                 ret = cache_block_group(block_group);
4804                                 BUG_ON(ret);
4805                         }
4806                         found_uncached_bg = true;
4807
4808                         /*
4809                          * If loop is set for cached only, try the next block
4810                          * group.
4811                          */
4812                         if (loop == LOOP_FIND_IDEAL)
4813                                 goto loop;
4814                 }
4815
4816                 cached = block_group_cache_done(block_group);
4817                 if (unlikely(!cached))
4818                         found_uncached_bg = true;
4819
4820                 if (unlikely(block_group->ro))
4821                         goto loop;
4822
4823                 /*
4824                  * Ok we want to try and use the cluster allocator, so lets look
4825                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
4826                  * have tried the cluster allocator plenty of times at this
4827                  * point and not have found anything, so we are likely way too
4828                  * fragmented for the clustering stuff to find anything, so lets
4829                  * just skip it and let the allocator find whatever block it can
4830                  * find
4831                  */
4832                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
4833                         /*
4834                          * the refill lock keeps out other
4835                          * people trying to start a new cluster
4836                          */
4837                         spin_lock(&last_ptr->refill_lock);
4838                         if (last_ptr->block_group &&
4839                             (last_ptr->block_group->ro ||
4840                             !block_group_bits(last_ptr->block_group, data))) {
4841                                 offset = 0;
4842                                 goto refill_cluster;
4843                         }
4844
4845                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
4846                                                  num_bytes, search_start);
4847                         if (offset) {
4848                                 /* we have a block, we're done */
4849                                 spin_unlock(&last_ptr->refill_lock);
4850                                 goto checks;
4851                         }
4852
4853                         spin_lock(&last_ptr->lock);
4854                         /*
4855                          * whoops, this cluster doesn't actually point to
4856                          * this block group.  Get a ref on the block
4857                          * group is does point to and try again
4858                          */
4859                         if (!last_ptr_loop && last_ptr->block_group &&
4860                             last_ptr->block_group != block_group) {
4861
4862                                 btrfs_put_block_group(block_group);
4863                                 block_group = last_ptr->block_group;
4864                                 btrfs_get_block_group(block_group);
4865                                 spin_unlock(&last_ptr->lock);
4866                                 spin_unlock(&last_ptr->refill_lock);
4867
4868                                 last_ptr_loop = 1;
4869                                 search_start = block_group->key.objectid;
4870                                 /*
4871                                  * we know this block group is properly
4872                                  * in the list because
4873                                  * btrfs_remove_block_group, drops the
4874                                  * cluster before it removes the block
4875                                  * group from the list
4876                                  */
4877                                 goto have_block_group;
4878                         }
4879                         spin_unlock(&last_ptr->lock);
4880 refill_cluster:
4881                         /*
4882                          * this cluster didn't work out, free it and
4883                          * start over
4884                          */
4885                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
4886
4887                         last_ptr_loop = 0;
4888
4889                         /* allocate a cluster in this block group */
4890                         ret = btrfs_find_space_cluster(trans, root,
4891                                                block_group, last_ptr,
4892                                                offset, num_bytes,
4893                                                empty_cluster + empty_size);
4894                         if (ret == 0) {
4895                                 /*
4896                                  * now pull our allocation out of this
4897                                  * cluster
4898                                  */
4899                                 offset = btrfs_alloc_from_cluster(block_group,
4900                                                   last_ptr, num_bytes,
4901                                                   search_start);
4902                                 if (offset) {
4903                                         /* we found one, proceed */
4904                                         spin_unlock(&last_ptr->refill_lock);
4905                                         goto checks;
4906                                 }
4907                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
4908                                    && !failed_cluster_refill) {
4909                                 spin_unlock(&last_ptr->refill_lock);
4910
4911                                 failed_cluster_refill = true;
4912                                 wait_block_group_cache_progress(block_group,
4913                                        num_bytes + empty_cluster + empty_size);
4914                                 goto have_block_group;
4915                         }
4916
4917                         /*
4918                          * at this point we either didn't find a cluster
4919                          * or we weren't able to allocate a block from our
4920                          * cluster.  Free the cluster we've been trying
4921                          * to use, and go to the next block group
4922                          */
4923                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
4924                         spin_unlock(&last_ptr->refill_lock);
4925                         goto loop;
4926                 }
4927
4928                 offset = btrfs_find_space_for_alloc(block_group, search_start,
4929                                                     num_bytes, empty_size);
4930                 /*
4931                  * If we didn't find a chunk, and we haven't failed on this
4932                  * block group before, and this block group is in the middle of
4933                  * caching and we are ok with waiting, then go ahead and wait
4934                  * for progress to be made, and set failed_alloc to true.
4935                  *
4936                  * If failed_alloc is true then we've already waited on this
4937                  * block group once and should move on to the next block group.
4938                  */
4939                 if (!offset && !failed_alloc && !cached &&
4940                     loop > LOOP_CACHING_NOWAIT) {
4941                         wait_block_group_cache_progress(block_group,
4942                                                 num_bytes + empty_size);
4943                         failed_alloc = true;
4944                         goto have_block_group;
4945                 } else if (!offset) {
4946                         goto loop;
4947                 }
4948 checks:
4949                 search_start = stripe_align(root, offset);
4950                 /* move on to the next group */
4951                 if (search_start + num_bytes >= search_end) {
4952                         btrfs_add_free_space(block_group, offset, num_bytes);
4953                         goto loop;
4954                 }
4955
4956                 /* move on to the next group */
4957                 if (search_start + num_bytes >
4958                     block_group->key.objectid + block_group->key.offset) {
4959                         btrfs_add_free_space(block_group, offset, num_bytes);
4960                         goto loop;
4961                 }
4962
4963                 ins->objectid = search_start;
4964                 ins->offset = num_bytes;
4965
4966                 if (offset < search_start)
4967                         btrfs_add_free_space(block_group, offset,
4968                                              search_start - offset);
4969                 BUG_ON(offset > search_start);
4970
4971                 ret = update_reserved_bytes(block_group, num_bytes, 1,
4972                                             (data & BTRFS_BLOCK_GROUP_DATA));
4973                 if (ret == -EAGAIN) {
4974                         btrfs_add_free_space(block_group, offset, num_bytes);
4975                         goto loop;
4976                 }
4977
4978                 /* we are all good, lets return */
4979                 ins->objectid = search_start;
4980                 ins->offset = num_bytes;
4981
4982                 if (offset < search_start)
4983                         btrfs_add_free_space(block_group, offset,
4984                                              search_start - offset);
4985                 BUG_ON(offset > search_start);
4986                 break;
4987 loop:
4988                 failed_cluster_refill = false;
4989                 failed_alloc = false;
4990                 BUG_ON(index != get_block_group_index(block_group));
4991                 btrfs_put_block_group(block_group);
4992         }
4993         up_read(&space_info->groups_sem);
4994
4995         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
4996                 goto search;
4997
4998         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
4999          *                      for them to make caching progress.  Also
5000          *                      determine the best possible bg to cache
5001          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5002          *                      caching kthreads as we move along
5003          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5004          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5005          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5006          *                      again
5007          */
5008         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
5009             (found_uncached_bg || empty_size || empty_cluster ||
5010              allowed_chunk_alloc)) {
5011                 index = 0;
5012                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5013                         found_uncached_bg = false;
5014                         loop++;
5015                         if (!ideal_cache_percent &&
5016                             atomic_read(&space_info->caching_threads))
5017                                 goto search;
5018
5019                         /*
5020                          * 1 of the following 2 things have happened so far
5021                          *
5022                          * 1) We found an ideal block group for caching that
5023                          * is mostly full and will cache quickly, so we might
5024                          * as well wait for it.
5025                          *
5026                          * 2) We searched for cached only and we didn't find
5027                          * anything, and we didn't start any caching kthreads
5028                          * either, so chances are we will loop through and
5029                          * start a couple caching kthreads, and then come back
5030                          * around and just wait for them.  This will be slower
5031                          * because we will have 2 caching kthreads reading at
5032                          * the same time when we could have just started one
5033                          * and waited for it to get far enough to give us an
5034                          * allocation, so go ahead and go to the wait caching
5035                          * loop.
5036                          */
5037                         loop = LOOP_CACHING_WAIT;
5038                         search_start = ideal_cache_offset;
5039                         ideal_cache_percent = 0;
5040                         goto ideal_cache;
5041                 } else if (loop == LOOP_FIND_IDEAL) {
5042                         /*
5043                          * Didn't find a uncached bg, wait on anything we find
5044                          * next.
5045                          */
5046                         loop = LOOP_CACHING_WAIT;
5047                         goto search;
5048                 }
5049
5050                 if (loop < LOOP_CACHING_WAIT) {
5051                         loop++;
5052                         goto search;
5053                 }
5054
5055                 if (loop == LOOP_ALLOC_CHUNK) {
5056                         empty_size = 0;
5057                         empty_cluster = 0;
5058                 }
5059
5060                 if (allowed_chunk_alloc) {
5061                         ret = do_chunk_alloc(trans, root, num_bytes +
5062                                              2 * 1024 * 1024, data, 1);
5063                         allowed_chunk_alloc = 0;
5064                         done_chunk_alloc = 1;
5065                 } else if (!done_chunk_alloc) {
5066                         space_info->force_alloc = 1;
5067                 }
5068
5069                 if (loop < LOOP_NO_EMPTY_SIZE) {
5070                         loop++;
5071                         goto search;
5072                 }
5073                 ret = -ENOSPC;
5074         } else if (!ins->objectid) {
5075                 ret = -ENOSPC;
5076         }
5077
5078         /* we found what we needed */
5079         if (ins->objectid) {
5080                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
5081                         trans->block_group = block_group->key.objectid;
5082
5083                 btrfs_put_block_group(block_group);
5084                 ret = 0;
5085         }
5086
5087         return ret;
5088 }
5089
5090 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5091                             int dump_block_groups)
5092 {
5093         struct btrfs_block_group_cache *cache;
5094         int index = 0;
5095
5096         spin_lock(&info->lock);
5097         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5098                (unsigned long long)(info->total_bytes - info->bytes_used -
5099                                     info->bytes_pinned - info->bytes_reserved -
5100                                     info->bytes_readonly),
5101                (info->full) ? "" : "not ");
5102         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5103                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5104                (unsigned long long)info->total_bytes,
5105                (unsigned long long)info->bytes_used,
5106                (unsigned long long)info->bytes_pinned,
5107                (unsigned long long)info->bytes_reserved,
5108                (unsigned long long)info->bytes_may_use,
5109                (unsigned long long)info->bytes_readonly);
5110         spin_unlock(&info->lock);
5111
5112         if (!dump_block_groups)
5113                 return;
5114
5115         down_read(&info->groups_sem);
5116 again:
5117         list_for_each_entry(cache, &info->block_groups[index], list) {
5118                 spin_lock(&cache->lock);
5119                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5120                        "%llu pinned %llu reserved\n",
5121                        (unsigned long long)cache->key.objectid,
5122                        (unsigned long long)cache->key.offset,
5123                        (unsigned long long)btrfs_block_group_used(&cache->item),
5124                        (unsigned long long)cache->pinned,
5125                        (unsigned long long)cache->reserved);
5126                 btrfs_dump_free_space(cache, bytes);
5127                 spin_unlock(&cache->lock);
5128         }
5129         if (++index < BTRFS_NR_RAID_TYPES)
5130                 goto again;
5131         up_read(&info->groups_sem);
5132 }
5133
5134 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5135                          struct btrfs_root *root,
5136                          u64 num_bytes, u64 min_alloc_size,
5137                          u64 empty_size, u64 hint_byte,
5138                          u64 search_end, struct btrfs_key *ins,
5139                          u64 data)
5140 {
5141         int ret;
5142         u64 search_start = 0;
5143
5144         data = btrfs_get_alloc_profile(root, data);
5145 again:
5146         /*
5147          * the only place that sets empty_size is btrfs_realloc_node, which
5148          * is not called recursively on allocations
5149          */
5150         if (empty_size || root->ref_cows)
5151                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5152                                      num_bytes + 2 * 1024 * 1024, data, 0);
5153
5154         WARN_ON(num_bytes < root->sectorsize);
5155         ret = find_free_extent(trans, root, num_bytes, empty_size,
5156                                search_start, search_end, hint_byte,
5157                                ins, data);
5158
5159         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5160                 num_bytes = num_bytes >> 1;
5161                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5162                 num_bytes = max(num_bytes, min_alloc_size);
5163                 do_chunk_alloc(trans, root->fs_info->extent_root,
5164                                num_bytes, data, 1);
5165                 goto again;
5166         }
5167         if (ret == -ENOSPC) {
5168                 struct btrfs_space_info *sinfo;
5169
5170                 sinfo = __find_space_info(root->fs_info, data);
5171                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5172                        "wanted %llu\n", (unsigned long long)data,
5173                        (unsigned long long)num_bytes);
5174                 dump_space_info(sinfo, num_bytes, 1);
5175         }
5176
5177         return ret;
5178 }
5179
5180 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5181 {
5182         struct btrfs_block_group_cache *cache;
5183         int ret = 0;
5184
5185         cache = btrfs_lookup_block_group(root->fs_info, start);
5186         if (!cache) {
5187                 printk(KERN_ERR "Unable to find block group for %llu\n",
5188                        (unsigned long long)start);
5189                 return -ENOSPC;
5190         }
5191
5192         ret = btrfs_discard_extent(root, start, len);
5193
5194         btrfs_add_free_space(cache, start, len);
5195         update_reserved_bytes(cache, len, 0, 1);
5196         btrfs_put_block_group(cache);
5197
5198         return ret;
5199 }
5200
5201 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5202                                       struct btrfs_root *root,
5203                                       u64 parent, u64 root_objectid,
5204                                       u64 flags, u64 owner, u64 offset,
5205                                       struct btrfs_key *ins, int ref_mod)
5206 {
5207         int ret;
5208         struct btrfs_fs_info *fs_info = root->fs_info;
5209         struct btrfs_extent_item *extent_item;
5210         struct btrfs_extent_inline_ref *iref;
5211         struct btrfs_path *path;
5212         struct extent_buffer *leaf;
5213         int type;
5214         u32 size;
5215
5216         if (parent > 0)
5217                 type = BTRFS_SHARED_DATA_REF_KEY;
5218         else
5219                 type = BTRFS_EXTENT_DATA_REF_KEY;
5220
5221         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5222
5223         path = btrfs_alloc_path();
5224         BUG_ON(!path);
5225
5226         path->leave_spinning = 1;
5227         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5228                                       ins, size);
5229         BUG_ON(ret);
5230
5231         leaf = path->nodes[0];
5232         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5233                                      struct btrfs_extent_item);
5234         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5235         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5236         btrfs_set_extent_flags(leaf, extent_item,
5237                                flags | BTRFS_EXTENT_FLAG_DATA);
5238
5239         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5240         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5241         if (parent > 0) {
5242                 struct btrfs_shared_data_ref *ref;
5243                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5244                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5245                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5246         } else {
5247                 struct btrfs_extent_data_ref *ref;
5248                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5249                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5250                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5251                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5252                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5253         }
5254
5255         btrfs_mark_buffer_dirty(path->nodes[0]);
5256         btrfs_free_path(path);
5257
5258         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5259         if (ret) {
5260                 printk(KERN_ERR "btrfs update block group failed for %llu "
5261                        "%llu\n", (unsigned long long)ins->objectid,
5262                        (unsigned long long)ins->offset);
5263                 BUG();
5264         }
5265         return ret;
5266 }
5267
5268 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5269                                      struct btrfs_root *root,
5270                                      u64 parent, u64 root_objectid,
5271                                      u64 flags, struct btrfs_disk_key *key,
5272                                      int level, struct btrfs_key *ins)
5273 {
5274         int ret;
5275         struct btrfs_fs_info *fs_info = root->fs_info;
5276         struct btrfs_extent_item *extent_item;
5277         struct btrfs_tree_block_info *block_info;
5278         struct btrfs_extent_inline_ref *iref;
5279         struct btrfs_path *path;
5280         struct extent_buffer *leaf;
5281         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5282
5283         path = btrfs_alloc_path();
5284         BUG_ON(!path);
5285
5286         path->leave_spinning = 1;
5287         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5288                                       ins, size);
5289         BUG_ON(ret);
5290
5291         leaf = path->nodes[0];
5292         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5293                                      struct btrfs_extent_item);
5294         btrfs_set_extent_refs(leaf, extent_item, 1);
5295         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5296         btrfs_set_extent_flags(leaf, extent_item,
5297                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5298         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5299
5300         btrfs_set_tree_block_key(leaf, block_info, key);
5301         btrfs_set_tree_block_level(leaf, block_info, level);
5302
5303         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5304         if (parent > 0) {
5305                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5306                 btrfs_set_extent_inline_ref_type(leaf, iref,
5307                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5308                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5309         } else {
5310                 btrfs_set_extent_inline_ref_type(leaf, iref,
5311                                                  BTRFS_TREE_BLOCK_REF_KEY);
5312                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5313         }
5314
5315         btrfs_mark_buffer_dirty(leaf);
5316         btrfs_free_path(path);
5317
5318         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5319         if (ret) {
5320                 printk(KERN_ERR "btrfs update block group failed for %llu "
5321                        "%llu\n", (unsigned long long)ins->objectid,
5322                        (unsigned long long)ins->offset);
5323                 BUG();
5324         }
5325         return ret;
5326 }
5327
5328 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5329                                      struct btrfs_root *root,
5330                                      u64 root_objectid, u64 owner,
5331                                      u64 offset, struct btrfs_key *ins)
5332 {
5333         int ret;
5334
5335         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5336
5337         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5338                                          0, root_objectid, owner, offset,
5339                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5340         return ret;
5341 }
5342
5343 /*
5344  * this is used by the tree logging recovery code.  It records that
5345  * an extent has been allocated and makes sure to clear the free
5346  * space cache bits as well
5347  */
5348 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5349                                    struct btrfs_root *root,
5350                                    u64 root_objectid, u64 owner, u64 offset,
5351                                    struct btrfs_key *ins)
5352 {
5353         int ret;
5354         struct btrfs_block_group_cache *block_group;
5355         struct btrfs_caching_control *caching_ctl;
5356         u64 start = ins->objectid;
5357         u64 num_bytes = ins->offset;
5358
5359         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5360         cache_block_group(block_group);
5361         caching_ctl = get_caching_control(block_group);
5362
5363         if (!caching_ctl) {
5364                 BUG_ON(!block_group_cache_done(block_group));
5365                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5366                 BUG_ON(ret);
5367         } else {
5368                 mutex_lock(&caching_ctl->mutex);
5369
5370                 if (start >= caching_ctl->progress) {
5371                         ret = add_excluded_extent(root, start, num_bytes);
5372                         BUG_ON(ret);
5373                 } else if (start + num_bytes <= caching_ctl->progress) {
5374                         ret = btrfs_remove_free_space(block_group,
5375                                                       start, num_bytes);
5376                         BUG_ON(ret);
5377                 } else {
5378                         num_bytes = caching_ctl->progress - start;
5379                         ret = btrfs_remove_free_space(block_group,
5380                                                       start, num_bytes);
5381                         BUG_ON(ret);
5382
5383                         start = caching_ctl->progress;
5384                         num_bytes = ins->objectid + ins->offset -
5385                                     caching_ctl->progress;
5386                         ret = add_excluded_extent(root, start, num_bytes);
5387                         BUG_ON(ret);
5388                 }
5389
5390                 mutex_unlock(&caching_ctl->mutex);
5391                 put_caching_control(caching_ctl);
5392         }
5393
5394         ret = update_reserved_bytes(block_group, ins->offset, 1, 1);
5395         BUG_ON(ret);
5396         btrfs_put_block_group(block_group);
5397         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5398                                          0, owner, offset, ins, 1);
5399         return ret;
5400 }
5401
5402 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5403                                             struct btrfs_root *root,
5404                                             u64 bytenr, u32 blocksize,
5405                                             int level)
5406 {
5407         struct extent_buffer *buf;
5408
5409         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5410         if (!buf)
5411                 return ERR_PTR(-ENOMEM);
5412         btrfs_set_header_generation(buf, trans->transid);
5413         btrfs_set_buffer_lockdep_class(buf, level);
5414         btrfs_tree_lock(buf);
5415         clean_tree_block(trans, root, buf);
5416
5417         btrfs_set_lock_blocking(buf);
5418         btrfs_set_buffer_uptodate(buf);
5419
5420         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5421                 /*
5422                  * we allow two log transactions at a time, use different
5423                  * EXENT bit to differentiate dirty pages.
5424                  */
5425                 if (root->log_transid % 2 == 0)
5426                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5427                                         buf->start + buf->len - 1, GFP_NOFS);
5428                 else
5429                         set_extent_new(&root->dirty_log_pages, buf->start,
5430                                         buf->start + buf->len - 1, GFP_NOFS);
5431         } else {
5432                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5433                          buf->start + buf->len - 1, GFP_NOFS);
5434         }
5435         trans->blocks_used++;
5436         /* this returns a buffer locked for blocking */
5437         return buf;
5438 }
5439
5440 static struct btrfs_block_rsv *
5441 use_block_rsv(struct btrfs_trans_handle *trans,
5442               struct btrfs_root *root, u32 blocksize)
5443 {
5444         struct btrfs_block_rsv *block_rsv;
5445         int ret;
5446
5447         block_rsv = get_block_rsv(trans, root);
5448
5449         if (block_rsv->size == 0) {
5450                 ret = reserve_metadata_bytes(block_rsv, blocksize);
5451                 if (ret)
5452                         return ERR_PTR(ret);
5453                 return block_rsv;
5454         }
5455
5456         ret = block_rsv_use_bytes(block_rsv, blocksize);
5457         if (!ret)
5458                 return block_rsv;
5459
5460         WARN_ON(1);
5461         printk(KERN_INFO"block_rsv size %llu reserved %llu freed %llu %llu\n",
5462                 block_rsv->size, block_rsv->reserved,
5463                 block_rsv->freed[0], block_rsv->freed[1]);
5464
5465         return ERR_PTR(-ENOSPC);
5466 }
5467
5468 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5469 {
5470         block_rsv_add_bytes(block_rsv, blocksize, 0);
5471         block_rsv_release_bytes(block_rsv, NULL, 0);
5472 }
5473
5474 /*
5475  * finds a free extent and does all the dirty work required for allocation
5476  * returns the key for the extent through ins, and a tree buffer for
5477  * the first block of the extent through buf.
5478  *
5479  * returns the tree buffer or NULL.
5480  */
5481 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5482                                         struct btrfs_root *root, u32 blocksize,
5483                                         u64 parent, u64 root_objectid,
5484                                         struct btrfs_disk_key *key, int level,
5485                                         u64 hint, u64 empty_size)
5486 {
5487         struct btrfs_key ins;
5488         struct btrfs_block_rsv *block_rsv;
5489         struct extent_buffer *buf;
5490         u64 flags = 0;
5491         int ret;
5492
5493
5494         block_rsv = use_block_rsv(trans, root, blocksize);
5495         if (IS_ERR(block_rsv))
5496                 return ERR_CAST(block_rsv);
5497
5498         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5499                                    empty_size, hint, (u64)-1, &ins, 0);
5500         if (ret) {
5501                 unuse_block_rsv(block_rsv, blocksize);
5502                 return ERR_PTR(ret);
5503         }
5504
5505         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5506                                     blocksize, level);
5507         BUG_ON(IS_ERR(buf));
5508
5509         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5510                 if (parent == 0)
5511                         parent = ins.objectid;
5512                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5513         } else
5514                 BUG_ON(parent > 0);
5515
5516         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5517                 struct btrfs_delayed_extent_op *extent_op;
5518                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5519                 BUG_ON(!extent_op);
5520                 if (key)
5521                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5522                 else
5523                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5524                 extent_op->flags_to_set = flags;
5525                 extent_op->update_key = 1;
5526                 extent_op->update_flags = 1;
5527                 extent_op->is_data = 0;
5528
5529                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5530                                         ins.offset, parent, root_objectid,
5531                                         level, BTRFS_ADD_DELAYED_EXTENT,
5532                                         extent_op);
5533                 BUG_ON(ret);
5534         }
5535         return buf;
5536 }
5537
5538 struct walk_control {
5539         u64 refs[BTRFS_MAX_LEVEL];
5540         u64 flags[BTRFS_MAX_LEVEL];
5541         struct btrfs_key update_progress;
5542         int stage;
5543         int level;
5544         int shared_level;
5545         int update_ref;
5546         int keep_locks;
5547         int reada_slot;
5548         int reada_count;
5549 };
5550
5551 #define DROP_REFERENCE  1
5552 #define UPDATE_BACKREF  2
5553
5554 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5555                                      struct btrfs_root *root,
5556                                      struct walk_control *wc,
5557                                      struct btrfs_path *path)
5558 {
5559         u64 bytenr;
5560         u64 generation;
5561         u64 refs;
5562         u64 flags;
5563         u64 last = 0;
5564         u32 nritems;
5565         u32 blocksize;
5566         struct btrfs_key key;
5567         struct extent_buffer *eb;
5568         int ret;
5569         int slot;
5570         int nread = 0;
5571
5572         if (path->slots[wc->level] < wc->reada_slot) {
5573                 wc->reada_count = wc->reada_count * 2 / 3;
5574                 wc->reada_count = max(wc->reada_count, 2);
5575         } else {
5576                 wc->reada_count = wc->reada_count * 3 / 2;
5577                 wc->reada_count = min_t(int, wc->reada_count,
5578                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5579         }
5580
5581         eb = path->nodes[wc->level];
5582         nritems = btrfs_header_nritems(eb);
5583         blocksize = btrfs_level_size(root, wc->level - 1);
5584
5585         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5586                 if (nread >= wc->reada_count)
5587                         break;
5588
5589                 cond_resched();
5590                 bytenr = btrfs_node_blockptr(eb, slot);
5591                 generation = btrfs_node_ptr_generation(eb, slot);
5592
5593                 if (slot == path->slots[wc->level])
5594                         goto reada;
5595
5596                 if (wc->stage == UPDATE_BACKREF &&
5597                     generation <= root->root_key.offset)
5598                         continue;
5599
5600                 /* We don't lock the tree block, it's OK to be racy here */
5601                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5602                                                &refs, &flags);
5603                 BUG_ON(ret);
5604                 BUG_ON(refs == 0);
5605
5606                 if (wc->stage == DROP_REFERENCE) {
5607                         if (refs == 1)
5608                                 goto reada;
5609
5610                         if (wc->level == 1 &&
5611                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5612                                 continue;
5613                         if (!wc->update_ref ||
5614                             generation <= root->root_key.offset)
5615                                 continue;
5616                         btrfs_node_key_to_cpu(eb, &key, slot);
5617                         ret = btrfs_comp_cpu_keys(&key,
5618                                                   &wc->update_progress);
5619                         if (ret < 0)
5620                                 continue;
5621                 } else {
5622                         if (wc->level == 1 &&
5623                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5624                                 continue;
5625                 }
5626 reada:
5627                 ret = readahead_tree_block(root, bytenr, blocksize,
5628                                            generation);
5629