Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/cmarinas...
[linux-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <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 /* control flags for do_chunk_alloc's force field
37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
38  * if we really need one.
39  *
40  * CHUNK_ALLOC_FORCE means it must try to allocate one
41  *
42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
43  * if we have very few chunks already allocated.  This is
44  * used as part of the clustering code to help make sure
45  * we have a good pool of storage to cluster in, without
46  * filling the FS with empty chunks
47  *
48  */
49 enum {
50         CHUNK_ALLOC_NO_FORCE = 0,
51         CHUNK_ALLOC_FORCE = 1,
52         CHUNK_ALLOC_LIMITED = 2,
53 };
54
55 static int update_block_group(struct btrfs_trans_handle *trans,
56                               struct btrfs_root *root,
57                               u64 bytenr, u64 num_bytes, int alloc);
58 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
59                                 struct btrfs_root *root,
60                                 u64 bytenr, u64 num_bytes, u64 parent,
61                                 u64 root_objectid, u64 owner_objectid,
62                                 u64 owner_offset, int refs_to_drop,
63                                 struct btrfs_delayed_extent_op *extra_op);
64 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
65                                     struct extent_buffer *leaf,
66                                     struct btrfs_extent_item *ei);
67 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
68                                       struct btrfs_root *root,
69                                       u64 parent, u64 root_objectid,
70                                       u64 flags, u64 owner, u64 offset,
71                                       struct btrfs_key *ins, int ref_mod);
72 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
73                                      struct btrfs_root *root,
74                                      u64 parent, u64 root_objectid,
75                                      u64 flags, struct btrfs_disk_key *key,
76                                      int level, struct btrfs_key *ins);
77 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
78                           struct btrfs_root *extent_root, u64 alloc_bytes,
79                           u64 flags, int force);
80 static int find_next_key(struct btrfs_path *path, int level,
81                          struct btrfs_key *key);
82 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
83                             int dump_block_groups);
84
85 static noinline int
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
87 {
88         smp_mb();
89         return cache->cached == BTRFS_CACHE_FINISHED;
90 }
91
92 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
93 {
94         return (cache->flags & bits) == bits;
95 }
96
97 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
98 {
99         atomic_inc(&cache->count);
100 }
101
102 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
103 {
104         if (atomic_dec_and_test(&cache->count)) {
105                 WARN_ON(cache->pinned > 0);
106                 WARN_ON(cache->reserved > 0);
107                 WARN_ON(cache->reserved_pinned > 0);
108                 kfree(cache);
109         }
110 }
111
112 /*
113  * this adds the block group to the fs_info rb tree for the block group
114  * cache
115  */
116 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
117                                 struct btrfs_block_group_cache *block_group)
118 {
119         struct rb_node **p;
120         struct rb_node *parent = NULL;
121         struct btrfs_block_group_cache *cache;
122
123         spin_lock(&info->block_group_cache_lock);
124         p = &info->block_group_cache_tree.rb_node;
125
126         while (*p) {
127                 parent = *p;
128                 cache = rb_entry(parent, struct btrfs_block_group_cache,
129                                  cache_node);
130                 if (block_group->key.objectid < cache->key.objectid) {
131                         p = &(*p)->rb_left;
132                 } else if (block_group->key.objectid > cache->key.objectid) {
133                         p = &(*p)->rb_right;
134                 } else {
135                         spin_unlock(&info->block_group_cache_lock);
136                         return -EEXIST;
137                 }
138         }
139
140         rb_link_node(&block_group->cache_node, parent, p);
141         rb_insert_color(&block_group->cache_node,
142                         &info->block_group_cache_tree);
143         spin_unlock(&info->block_group_cache_lock);
144
145         return 0;
146 }
147
148 /*
149  * This will return the block group at or after bytenr if contains is 0, else
150  * it will return the block group that contains the bytenr
151  */
152 static struct btrfs_block_group_cache *
153 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
154                               int contains)
155 {
156         struct btrfs_block_group_cache *cache, *ret = NULL;
157         struct rb_node *n;
158         u64 end, start;
159
160         spin_lock(&info->block_group_cache_lock);
161         n = info->block_group_cache_tree.rb_node;
162
163         while (n) {
164                 cache = rb_entry(n, struct btrfs_block_group_cache,
165                                  cache_node);
166                 end = cache->key.objectid + cache->key.offset - 1;
167                 start = cache->key.objectid;
168
169                 if (bytenr < start) {
170                         if (!contains && (!ret || start < ret->key.objectid))
171                                 ret = cache;
172                         n = n->rb_left;
173                 } else if (bytenr > start) {
174                         if (contains && bytenr <= end) {
175                                 ret = cache;
176                                 break;
177                         }
178                         n = n->rb_right;
179                 } else {
180                         ret = cache;
181                         break;
182                 }
183         }
184         if (ret)
185                 btrfs_get_block_group(ret);
186         spin_unlock(&info->block_group_cache_lock);
187
188         return ret;
189 }
190
191 static int add_excluded_extent(struct btrfs_root *root,
192                                u64 start, u64 num_bytes)
193 {
194         u64 end = start + num_bytes - 1;
195         set_extent_bits(&root->fs_info->freed_extents[0],
196                         start, end, EXTENT_UPTODATE, GFP_NOFS);
197         set_extent_bits(&root->fs_info->freed_extents[1],
198                         start, end, EXTENT_UPTODATE, GFP_NOFS);
199         return 0;
200 }
201
202 static void free_excluded_extents(struct btrfs_root *root,
203                                   struct btrfs_block_group_cache *cache)
204 {
205         u64 start, end;
206
207         start = cache->key.objectid;
208         end = start + cache->key.offset - 1;
209
210         clear_extent_bits(&root->fs_info->freed_extents[0],
211                           start, end, EXTENT_UPTODATE, GFP_NOFS);
212         clear_extent_bits(&root->fs_info->freed_extents[1],
213                           start, end, EXTENT_UPTODATE, GFP_NOFS);
214 }
215
216 static int exclude_super_stripes(struct btrfs_root *root,
217                                  struct btrfs_block_group_cache *cache)
218 {
219         u64 bytenr;
220         u64 *logical;
221         int stripe_len;
222         int i, nr, ret;
223
224         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
225                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
226                 cache->bytes_super += stripe_len;
227                 ret = add_excluded_extent(root, cache->key.objectid,
228                                           stripe_len);
229                 BUG_ON(ret);
230         }
231
232         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
233                 bytenr = btrfs_sb_offset(i);
234                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
235                                        cache->key.objectid, bytenr,
236                                        0, &logical, &nr, &stripe_len);
237                 BUG_ON(ret);
238
239                 while (nr--) {
240                         cache->bytes_super += stripe_len;
241                         ret = add_excluded_extent(root, logical[nr],
242                                                   stripe_len);
243                         BUG_ON(ret);
244                 }
245
246                 kfree(logical);
247         }
248         return 0;
249 }
250
251 static struct btrfs_caching_control *
252 get_caching_control(struct btrfs_block_group_cache *cache)
253 {
254         struct btrfs_caching_control *ctl;
255
256         spin_lock(&cache->lock);
257         if (cache->cached != BTRFS_CACHE_STARTED) {
258                 spin_unlock(&cache->lock);
259                 return NULL;
260         }
261
262         /* We're loading it the fast way, so we don't have a caching_ctl. */
263         if (!cache->caching_ctl) {
264                 spin_unlock(&cache->lock);
265                 return NULL;
266         }
267
268         ctl = cache->caching_ctl;
269         atomic_inc(&ctl->count);
270         spin_unlock(&cache->lock);
271         return ctl;
272 }
273
274 static void put_caching_control(struct btrfs_caching_control *ctl)
275 {
276         if (atomic_dec_and_test(&ctl->count))
277                 kfree(ctl);
278 }
279
280 /*
281  * this is only called by cache_block_group, since we could have freed extents
282  * we need to check the pinned_extents for any extents that can't be used yet
283  * since their free space will be released as soon as the transaction commits.
284  */
285 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
286                               struct btrfs_fs_info *info, u64 start, u64 end)
287 {
288         u64 extent_start, extent_end, size, total_added = 0;
289         int ret;
290
291         while (start < end) {
292                 ret = find_first_extent_bit(info->pinned_extents, start,
293                                             &extent_start, &extent_end,
294                                             EXTENT_DIRTY | EXTENT_UPTODATE);
295                 if (ret)
296                         break;
297
298                 if (extent_start <= start) {
299                         start = extent_end + 1;
300                 } else if (extent_start > start && extent_start < end) {
301                         size = extent_start - start;
302                         total_added += size;
303                         ret = btrfs_add_free_space(block_group, start,
304                                                    size);
305                         BUG_ON(ret);
306                         start = extent_end + 1;
307                 } else {
308                         break;
309                 }
310         }
311
312         if (start < end) {
313                 size = end - start;
314                 total_added += size;
315                 ret = btrfs_add_free_space(block_group, start, size);
316                 BUG_ON(ret);
317         }
318
319         return total_added;
320 }
321
322 static int caching_kthread(void *data)
323 {
324         struct btrfs_block_group_cache *block_group = data;
325         struct btrfs_fs_info *fs_info = block_group->fs_info;
326         struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
327         struct btrfs_root *extent_root = fs_info->extent_root;
328         struct btrfs_path *path;
329         struct extent_buffer *leaf;
330         struct btrfs_key key;
331         u64 total_found = 0;
332         u64 last = 0;
333         u32 nritems;
334         int ret = 0;
335
336         path = btrfs_alloc_path();
337         if (!path)
338                 return -ENOMEM;
339
340         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
341
342         /*
343          * We don't want to deadlock with somebody trying to allocate a new
344          * extent for the extent root while also trying to search the extent
345          * root to add free space.  So we skip locking and search the commit
346          * root, since its read-only
347          */
348         path->skip_locking = 1;
349         path->search_commit_root = 1;
350         path->reada = 2;
351
352         key.objectid = last;
353         key.offset = 0;
354         key.type = BTRFS_EXTENT_ITEM_KEY;
355 again:
356         mutex_lock(&caching_ctl->mutex);
357         /* need to make sure the commit_root doesn't disappear */
358         down_read(&fs_info->extent_commit_sem);
359
360         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
361         if (ret < 0)
362                 goto err;
363
364         leaf = path->nodes[0];
365         nritems = btrfs_header_nritems(leaf);
366
367         while (1) {
368                 smp_mb();
369                 if (fs_info->closing > 1) {
370                         last = (u64)-1;
371                         break;
372                 }
373
374                 if (path->slots[0] < nritems) {
375                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
376                 } else {
377                         ret = find_next_key(path, 0, &key);
378                         if (ret)
379                                 break;
380
381                         caching_ctl->progress = last;
382                         btrfs_release_path(extent_root, path);
383                         up_read(&fs_info->extent_commit_sem);
384                         mutex_unlock(&caching_ctl->mutex);
385                         if (btrfs_transaction_in_commit(fs_info))
386                                 schedule_timeout(1);
387                         else
388                                 cond_resched();
389                         goto again;
390                 }
391
392                 if (key.objectid < block_group->key.objectid) {
393                         path->slots[0]++;
394                         continue;
395                 }
396
397                 if (key.objectid >= block_group->key.objectid +
398                     block_group->key.offset)
399                         break;
400
401                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
402                         total_found += add_new_free_space(block_group,
403                                                           fs_info, last,
404                                                           key.objectid);
405                         last = key.objectid + key.offset;
406
407                         if (total_found > (1024 * 1024 * 2)) {
408                                 total_found = 0;
409                                 wake_up(&caching_ctl->wait);
410                         }
411                 }
412                 path->slots[0]++;
413         }
414         ret = 0;
415
416         total_found += add_new_free_space(block_group, fs_info, last,
417                                           block_group->key.objectid +
418                                           block_group->key.offset);
419         caching_ctl->progress = (u64)-1;
420
421         spin_lock(&block_group->lock);
422         block_group->caching_ctl = NULL;
423         block_group->cached = BTRFS_CACHE_FINISHED;
424         spin_unlock(&block_group->lock);
425
426 err:
427         btrfs_free_path(path);
428         up_read(&fs_info->extent_commit_sem);
429
430         free_excluded_extents(extent_root, block_group);
431
432         mutex_unlock(&caching_ctl->mutex);
433         wake_up(&caching_ctl->wait);
434
435         put_caching_control(caching_ctl);
436         atomic_dec(&block_group->space_info->caching_threads);
437         btrfs_put_block_group(block_group);
438
439         return 0;
440 }
441
442 static int cache_block_group(struct btrfs_block_group_cache *cache,
443                              struct btrfs_trans_handle *trans,
444                              struct btrfs_root *root,
445                              int load_cache_only)
446 {
447         struct btrfs_fs_info *fs_info = cache->fs_info;
448         struct btrfs_caching_control *caching_ctl;
449         struct task_struct *tsk;
450         int ret = 0;
451
452         smp_mb();
453         if (cache->cached != BTRFS_CACHE_NO)
454                 return 0;
455
456         /*
457          * We can't do the read from on-disk cache during a commit since we need
458          * to have the normal tree locking.  Also if we are currently trying to
459          * allocate blocks for the tree root we can't do the fast caching since
460          * we likely hold important locks.
461          */
462         if (trans && (!trans->transaction->in_commit) &&
463             (root && root != root->fs_info->tree_root)) {
464                 spin_lock(&cache->lock);
465                 if (cache->cached != BTRFS_CACHE_NO) {
466                         spin_unlock(&cache->lock);
467                         return 0;
468                 }
469                 cache->cached = BTRFS_CACHE_STARTED;
470                 spin_unlock(&cache->lock);
471
472                 ret = load_free_space_cache(fs_info, cache);
473
474                 spin_lock(&cache->lock);
475                 if (ret == 1) {
476                         cache->cached = BTRFS_CACHE_FINISHED;
477                         cache->last_byte_to_unpin = (u64)-1;
478                 } else {
479                         cache->cached = BTRFS_CACHE_NO;
480                 }
481                 spin_unlock(&cache->lock);
482                 if (ret == 1) {
483                         free_excluded_extents(fs_info->extent_root, cache);
484                         return 0;
485                 }
486         }
487
488         if (load_cache_only)
489                 return 0;
490
491         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
492         BUG_ON(!caching_ctl);
493
494         INIT_LIST_HEAD(&caching_ctl->list);
495         mutex_init(&caching_ctl->mutex);
496         init_waitqueue_head(&caching_ctl->wait);
497         caching_ctl->block_group = cache;
498         caching_ctl->progress = cache->key.objectid;
499         /* one for caching kthread, one for caching block group list */
500         atomic_set(&caching_ctl->count, 2);
501
502         spin_lock(&cache->lock);
503         if (cache->cached != BTRFS_CACHE_NO) {
504                 spin_unlock(&cache->lock);
505                 kfree(caching_ctl);
506                 return 0;
507         }
508         cache->caching_ctl = caching_ctl;
509         cache->cached = BTRFS_CACHE_STARTED;
510         spin_unlock(&cache->lock);
511
512         down_write(&fs_info->extent_commit_sem);
513         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
514         up_write(&fs_info->extent_commit_sem);
515
516         atomic_inc(&cache->space_info->caching_threads);
517         btrfs_get_block_group(cache);
518
519         tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
520                           cache->key.objectid);
521         if (IS_ERR(tsk)) {
522                 ret = PTR_ERR(tsk);
523                 printk(KERN_ERR "error running thread %d\n", ret);
524                 BUG();
525         }
526
527         return ret;
528 }
529
530 /*
531  * return the block group that starts at or after bytenr
532  */
533 static struct btrfs_block_group_cache *
534 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
535 {
536         struct btrfs_block_group_cache *cache;
537
538         cache = block_group_cache_tree_search(info, bytenr, 0);
539
540         return cache;
541 }
542
543 /*
544  * return the block group that contains the given bytenr
545  */
546 struct btrfs_block_group_cache *btrfs_lookup_block_group(
547                                                  struct btrfs_fs_info *info,
548                                                  u64 bytenr)
549 {
550         struct btrfs_block_group_cache *cache;
551
552         cache = block_group_cache_tree_search(info, bytenr, 1);
553
554         return cache;
555 }
556
557 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
558                                                   u64 flags)
559 {
560         struct list_head *head = &info->space_info;
561         struct btrfs_space_info *found;
562
563         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
564                  BTRFS_BLOCK_GROUP_METADATA;
565
566         rcu_read_lock();
567         list_for_each_entry_rcu(found, head, list) {
568                 if (found->flags & flags) {
569                         rcu_read_unlock();
570                         return found;
571                 }
572         }
573         rcu_read_unlock();
574         return NULL;
575 }
576
577 /*
578  * after adding space to the filesystem, we need to clear the full flags
579  * on all the space infos.
580  */
581 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
582 {
583         struct list_head *head = &info->space_info;
584         struct btrfs_space_info *found;
585
586         rcu_read_lock();
587         list_for_each_entry_rcu(found, head, list)
588                 found->full = 0;
589         rcu_read_unlock();
590 }
591
592 static u64 div_factor(u64 num, int factor)
593 {
594         if (factor == 10)
595                 return num;
596         num *= factor;
597         do_div(num, 10);
598         return num;
599 }
600
601 static u64 div_factor_fine(u64 num, int factor)
602 {
603         if (factor == 100)
604                 return num;
605         num *= factor;
606         do_div(num, 100);
607         return num;
608 }
609
610 u64 btrfs_find_block_group(struct btrfs_root *root,
611                            u64 search_start, u64 search_hint, int owner)
612 {
613         struct btrfs_block_group_cache *cache;
614         u64 used;
615         u64 last = max(search_hint, search_start);
616         u64 group_start = 0;
617         int full_search = 0;
618         int factor = 9;
619         int wrapped = 0;
620 again:
621         while (1) {
622                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
623                 if (!cache)
624                         break;
625
626                 spin_lock(&cache->lock);
627                 last = cache->key.objectid + cache->key.offset;
628                 used = btrfs_block_group_used(&cache->item);
629
630                 if ((full_search || !cache->ro) &&
631                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
632                         if (used + cache->pinned + cache->reserved <
633                             div_factor(cache->key.offset, factor)) {
634                                 group_start = cache->key.objectid;
635                                 spin_unlock(&cache->lock);
636                                 btrfs_put_block_group(cache);
637                                 goto found;
638                         }
639                 }
640                 spin_unlock(&cache->lock);
641                 btrfs_put_block_group(cache);
642                 cond_resched();
643         }
644         if (!wrapped) {
645                 last = search_start;
646                 wrapped = 1;
647                 goto again;
648         }
649         if (!full_search && factor < 10) {
650                 last = search_start;
651                 full_search = 1;
652                 factor = 10;
653                 goto again;
654         }
655 found:
656         return group_start;
657 }
658
659 /* simple helper to search for an existing extent at a given offset */
660 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
661 {
662         int ret;
663         struct btrfs_key key;
664         struct btrfs_path *path;
665
666         path = btrfs_alloc_path();
667         BUG_ON(!path);
668         key.objectid = start;
669         key.offset = len;
670         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
671         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
672                                 0, 0);
673         btrfs_free_path(path);
674         return ret;
675 }
676
677 /*
678  * helper function to lookup reference count and flags of extent.
679  *
680  * the head node for delayed ref is used to store the sum of all the
681  * reference count modifications queued up in the rbtree. the head
682  * node may also store the extent flags to set. This way you can check
683  * to see what the reference count and extent flags would be if all of
684  * the delayed refs are not processed.
685  */
686 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
687                              struct btrfs_root *root, u64 bytenr,
688                              u64 num_bytes, u64 *refs, u64 *flags)
689 {
690         struct btrfs_delayed_ref_head *head;
691         struct btrfs_delayed_ref_root *delayed_refs;
692         struct btrfs_path *path;
693         struct btrfs_extent_item *ei;
694         struct extent_buffer *leaf;
695         struct btrfs_key key;
696         u32 item_size;
697         u64 num_refs;
698         u64 extent_flags;
699         int ret;
700
701         path = btrfs_alloc_path();
702         if (!path)
703                 return -ENOMEM;
704
705         key.objectid = bytenr;
706         key.type = BTRFS_EXTENT_ITEM_KEY;
707         key.offset = num_bytes;
708         if (!trans) {
709                 path->skip_locking = 1;
710                 path->search_commit_root = 1;
711         }
712 again:
713         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
714                                 &key, path, 0, 0);
715         if (ret < 0)
716                 goto out_free;
717
718         if (ret == 0) {
719                 leaf = path->nodes[0];
720                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
721                 if (item_size >= sizeof(*ei)) {
722                         ei = btrfs_item_ptr(leaf, path->slots[0],
723                                             struct btrfs_extent_item);
724                         num_refs = btrfs_extent_refs(leaf, ei);
725                         extent_flags = btrfs_extent_flags(leaf, ei);
726                 } else {
727 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
728                         struct btrfs_extent_item_v0 *ei0;
729                         BUG_ON(item_size != sizeof(*ei0));
730                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
731                                              struct btrfs_extent_item_v0);
732                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
733                         /* FIXME: this isn't correct for data */
734                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
735 #else
736                         BUG();
737 #endif
738                 }
739                 BUG_ON(num_refs == 0);
740         } else {
741                 num_refs = 0;
742                 extent_flags = 0;
743                 ret = 0;
744         }
745
746         if (!trans)
747                 goto out;
748
749         delayed_refs = &trans->transaction->delayed_refs;
750         spin_lock(&delayed_refs->lock);
751         head = btrfs_find_delayed_ref_head(trans, bytenr);
752         if (head) {
753                 if (!mutex_trylock(&head->mutex)) {
754                         atomic_inc(&head->node.refs);
755                         spin_unlock(&delayed_refs->lock);
756
757                         btrfs_release_path(root->fs_info->extent_root, path);
758
759                         mutex_lock(&head->mutex);
760                         mutex_unlock(&head->mutex);
761                         btrfs_put_delayed_ref(&head->node);
762                         goto again;
763                 }
764                 if (head->extent_op && head->extent_op->update_flags)
765                         extent_flags |= head->extent_op->flags_to_set;
766                 else
767                         BUG_ON(num_refs == 0);
768
769                 num_refs += head->node.ref_mod;
770                 mutex_unlock(&head->mutex);
771         }
772         spin_unlock(&delayed_refs->lock);
773 out:
774         WARN_ON(num_refs == 0);
775         if (refs)
776                 *refs = num_refs;
777         if (flags)
778                 *flags = extent_flags;
779 out_free:
780         btrfs_free_path(path);
781         return ret;
782 }
783
784 /*
785  * Back reference rules.  Back refs have three main goals:
786  *
787  * 1) differentiate between all holders of references to an extent so that
788  *    when a reference is dropped we can make sure it was a valid reference
789  *    before freeing the extent.
790  *
791  * 2) Provide enough information to quickly find the holders of an extent
792  *    if we notice a given block is corrupted or bad.
793  *
794  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
795  *    maintenance.  This is actually the same as #2, but with a slightly
796  *    different use case.
797  *
798  * There are two kinds of back refs. The implicit back refs is optimized
799  * for pointers in non-shared tree blocks. For a given pointer in a block,
800  * back refs of this kind provide information about the block's owner tree
801  * and the pointer's key. These information allow us to find the block by
802  * b-tree searching. The full back refs is for pointers in tree blocks not
803  * referenced by their owner trees. The location of tree block is recorded
804  * in the back refs. Actually the full back refs is generic, and can be
805  * used in all cases the implicit back refs is used. The major shortcoming
806  * of the full back refs is its overhead. Every time a tree block gets
807  * COWed, we have to update back refs entry for all pointers in it.
808  *
809  * For a newly allocated tree block, we use implicit back refs for
810  * pointers in it. This means most tree related operations only involve
811  * implicit back refs. For a tree block created in old transaction, the
812  * only way to drop a reference to it is COW it. So we can detect the
813  * event that tree block loses its owner tree's reference and do the
814  * back refs conversion.
815  *
816  * When a tree block is COW'd through a tree, there are four cases:
817  *
818  * The reference count of the block is one and the tree is the block's
819  * owner tree. Nothing to do in this case.
820  *
821  * The reference count of the block is one and the tree is not the
822  * block's owner tree. In this case, full back refs is used for pointers
823  * in the block. Remove these full back refs, add implicit back refs for
824  * every pointers in the new block.
825  *
826  * The reference count of the block is greater than one and the tree is
827  * the block's owner tree. In this case, implicit back refs is used for
828  * pointers in the block. Add full back refs for every pointers in the
829  * block, increase lower level extents' reference counts. The original
830  * implicit back refs are entailed to the new block.
831  *
832  * The reference count of the block is greater than one and the tree is
833  * not the block's owner tree. Add implicit back refs for every pointer in
834  * the new block, increase lower level extents' reference count.
835  *
836  * Back Reference Key composing:
837  *
838  * The key objectid corresponds to the first byte in the extent,
839  * The key type is used to differentiate between types of back refs.
840  * There are different meanings of the key offset for different types
841  * of back refs.
842  *
843  * File extents can be referenced by:
844  *
845  * - multiple snapshots, subvolumes, or different generations in one subvol
846  * - different files inside a single subvolume
847  * - different offsets inside a file (bookend extents in file.c)
848  *
849  * The extent ref structure for the implicit back refs has fields for:
850  *
851  * - Objectid of the subvolume root
852  * - objectid of the file holding the reference
853  * - original offset in the file
854  * - how many bookend extents
855  *
856  * The key offset for the implicit back refs is hash of the first
857  * three fields.
858  *
859  * The extent ref structure for the full back refs has field for:
860  *
861  * - number of pointers in the tree leaf
862  *
863  * The key offset for the implicit back refs is the first byte of
864  * the tree leaf
865  *
866  * When a file extent is allocated, The implicit back refs is used.
867  * the fields are filled in:
868  *
869  *     (root_key.objectid, inode objectid, offset in file, 1)
870  *
871  * When a file extent is removed file truncation, we find the
872  * corresponding implicit back refs and check the following fields:
873  *
874  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
875  *
876  * Btree extents can be referenced by:
877  *
878  * - Different subvolumes
879  *
880  * Both the implicit back refs and the full back refs for tree blocks
881  * only consist of key. The key offset for the implicit back refs is
882  * objectid of block's owner tree. The key offset for the full back refs
883  * is the first byte of parent block.
884  *
885  * When implicit back refs is used, information about the lowest key and
886  * level of the tree block are required. These information are stored in
887  * tree block info structure.
888  */
889
890 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
891 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
892                                   struct btrfs_root *root,
893                                   struct btrfs_path *path,
894                                   u64 owner, u32 extra_size)
895 {
896         struct btrfs_extent_item *item;
897         struct btrfs_extent_item_v0 *ei0;
898         struct btrfs_extent_ref_v0 *ref0;
899         struct btrfs_tree_block_info *bi;
900         struct extent_buffer *leaf;
901         struct btrfs_key key;
902         struct btrfs_key found_key;
903         u32 new_size = sizeof(*item);
904         u64 refs;
905         int ret;
906
907         leaf = path->nodes[0];
908         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
909
910         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
911         ei0 = btrfs_item_ptr(leaf, path->slots[0],
912                              struct btrfs_extent_item_v0);
913         refs = btrfs_extent_refs_v0(leaf, ei0);
914
915         if (owner == (u64)-1) {
916                 while (1) {
917                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
918                                 ret = btrfs_next_leaf(root, path);
919                                 if (ret < 0)
920                                         return ret;
921                                 BUG_ON(ret > 0);
922                                 leaf = path->nodes[0];
923                         }
924                         btrfs_item_key_to_cpu(leaf, &found_key,
925                                               path->slots[0]);
926                         BUG_ON(key.objectid != found_key.objectid);
927                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
928                                 path->slots[0]++;
929                                 continue;
930                         }
931                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
932                                               struct btrfs_extent_ref_v0);
933                         owner = btrfs_ref_objectid_v0(leaf, ref0);
934                         break;
935                 }
936         }
937         btrfs_release_path(root, path);
938
939         if (owner < BTRFS_FIRST_FREE_OBJECTID)
940                 new_size += sizeof(*bi);
941
942         new_size -= sizeof(*ei0);
943         ret = btrfs_search_slot(trans, root, &key, path,
944                                 new_size + extra_size, 1);
945         if (ret < 0)
946                 return ret;
947         BUG_ON(ret);
948
949         ret = btrfs_extend_item(trans, root, path, new_size);
950         BUG_ON(ret);
951
952         leaf = path->nodes[0];
953         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
954         btrfs_set_extent_refs(leaf, item, refs);
955         /* FIXME: get real generation */
956         btrfs_set_extent_generation(leaf, item, 0);
957         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
958                 btrfs_set_extent_flags(leaf, item,
959                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
960                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
961                 bi = (struct btrfs_tree_block_info *)(item + 1);
962                 /* FIXME: get first key of the block */
963                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
964                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
965         } else {
966                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
967         }
968         btrfs_mark_buffer_dirty(leaf);
969         return 0;
970 }
971 #endif
972
973 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
974 {
975         u32 high_crc = ~(u32)0;
976         u32 low_crc = ~(u32)0;
977         __le64 lenum;
978
979         lenum = cpu_to_le64(root_objectid);
980         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
981         lenum = cpu_to_le64(owner);
982         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
983         lenum = cpu_to_le64(offset);
984         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
985
986         return ((u64)high_crc << 31) ^ (u64)low_crc;
987 }
988
989 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
990                                      struct btrfs_extent_data_ref *ref)
991 {
992         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
993                                     btrfs_extent_data_ref_objectid(leaf, ref),
994                                     btrfs_extent_data_ref_offset(leaf, ref));
995 }
996
997 static int match_extent_data_ref(struct extent_buffer *leaf,
998                                  struct btrfs_extent_data_ref *ref,
999                                  u64 root_objectid, u64 owner, u64 offset)
1000 {
1001         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1002             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1003             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1004                 return 0;
1005         return 1;
1006 }
1007
1008 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1009                                            struct btrfs_root *root,
1010                                            struct btrfs_path *path,
1011                                            u64 bytenr, u64 parent,
1012                                            u64 root_objectid,
1013                                            u64 owner, u64 offset)
1014 {
1015         struct btrfs_key key;
1016         struct btrfs_extent_data_ref *ref;
1017         struct extent_buffer *leaf;
1018         u32 nritems;
1019         int ret;
1020         int recow;
1021         int err = -ENOENT;
1022
1023         key.objectid = bytenr;
1024         if (parent) {
1025                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1026                 key.offset = parent;
1027         } else {
1028                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1029                 key.offset = hash_extent_data_ref(root_objectid,
1030                                                   owner, offset);
1031         }
1032 again:
1033         recow = 0;
1034         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1035         if (ret < 0) {
1036                 err = ret;
1037                 goto fail;
1038         }
1039
1040         if (parent) {
1041                 if (!ret)
1042                         return 0;
1043 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1044                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1045                 btrfs_release_path(root, path);
1046                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1047                 if (ret < 0) {
1048                         err = ret;
1049                         goto fail;
1050                 }
1051                 if (!ret)
1052                         return 0;
1053 #endif
1054                 goto fail;
1055         }
1056
1057         leaf = path->nodes[0];
1058         nritems = btrfs_header_nritems(leaf);
1059         while (1) {
1060                 if (path->slots[0] >= nritems) {
1061                         ret = btrfs_next_leaf(root, path);
1062                         if (ret < 0)
1063                                 err = ret;
1064                         if (ret)
1065                                 goto fail;
1066
1067                         leaf = path->nodes[0];
1068                         nritems = btrfs_header_nritems(leaf);
1069                         recow = 1;
1070                 }
1071
1072                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1073                 if (key.objectid != bytenr ||
1074                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1075                         goto fail;
1076
1077                 ref = btrfs_item_ptr(leaf, path->slots[0],
1078                                      struct btrfs_extent_data_ref);
1079
1080                 if (match_extent_data_ref(leaf, ref, root_objectid,
1081                                           owner, offset)) {
1082                         if (recow) {
1083                                 btrfs_release_path(root, path);
1084                                 goto again;
1085                         }
1086                         err = 0;
1087                         break;
1088                 }
1089                 path->slots[0]++;
1090         }
1091 fail:
1092         return err;
1093 }
1094
1095 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid, u64 owner,
1100                                            u64 offset, int refs_to_add)
1101 {
1102         struct btrfs_key key;
1103         struct extent_buffer *leaf;
1104         u32 size;
1105         u32 num_refs;
1106         int ret;
1107
1108         key.objectid = bytenr;
1109         if (parent) {
1110                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1111                 key.offset = parent;
1112                 size = sizeof(struct btrfs_shared_data_ref);
1113         } else {
1114                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1115                 key.offset = hash_extent_data_ref(root_objectid,
1116                                                   owner, offset);
1117                 size = sizeof(struct btrfs_extent_data_ref);
1118         }
1119
1120         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1121         if (ret && ret != -EEXIST)
1122                 goto fail;
1123
1124         leaf = path->nodes[0];
1125         if (parent) {
1126                 struct btrfs_shared_data_ref *ref;
1127                 ref = btrfs_item_ptr(leaf, path->slots[0],
1128                                      struct btrfs_shared_data_ref);
1129                 if (ret == 0) {
1130                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1131                 } else {
1132                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1133                         num_refs += refs_to_add;
1134                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1135                 }
1136         } else {
1137                 struct btrfs_extent_data_ref *ref;
1138                 while (ret == -EEXIST) {
1139                         ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                              struct btrfs_extent_data_ref);
1141                         if (match_extent_data_ref(leaf, ref, root_objectid,
1142                                                   owner, offset))
1143                                 break;
1144                         btrfs_release_path(root, path);
1145                         key.offset++;
1146                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1147                                                       size);
1148                         if (ret && ret != -EEXIST)
1149                                 goto fail;
1150
1151                         leaf = path->nodes[0];
1152                 }
1153                 ref = btrfs_item_ptr(leaf, path->slots[0],
1154                                      struct btrfs_extent_data_ref);
1155                 if (ret == 0) {
1156                         btrfs_set_extent_data_ref_root(leaf, ref,
1157                                                        root_objectid);
1158                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1159                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1160                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1161                 } else {
1162                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1163                         num_refs += refs_to_add;
1164                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1165                 }
1166         }
1167         btrfs_mark_buffer_dirty(leaf);
1168         ret = 0;
1169 fail:
1170         btrfs_release_path(root, path);
1171         return ret;
1172 }
1173
1174 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1175                                            struct btrfs_root *root,
1176                                            struct btrfs_path *path,
1177                                            int refs_to_drop)
1178 {
1179         struct btrfs_key key;
1180         struct btrfs_extent_data_ref *ref1 = NULL;
1181         struct btrfs_shared_data_ref *ref2 = NULL;
1182         struct extent_buffer *leaf;
1183         u32 num_refs = 0;
1184         int ret = 0;
1185
1186         leaf = path->nodes[0];
1187         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1188
1189         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1190                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1191                                       struct btrfs_extent_data_ref);
1192                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1193         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1194                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1195                                       struct btrfs_shared_data_ref);
1196                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1197 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1198         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1199                 struct btrfs_extent_ref_v0 *ref0;
1200                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1201                                       struct btrfs_extent_ref_v0);
1202                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1203 #endif
1204         } else {
1205                 BUG();
1206         }
1207
1208         BUG_ON(num_refs < refs_to_drop);
1209         num_refs -= refs_to_drop;
1210
1211         if (num_refs == 0) {
1212                 ret = btrfs_del_item(trans, root, path);
1213         } else {
1214                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1215                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1216                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1217                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1218 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1219                 else {
1220                         struct btrfs_extent_ref_v0 *ref0;
1221                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1222                                         struct btrfs_extent_ref_v0);
1223                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1224                 }
1225 #endif
1226                 btrfs_mark_buffer_dirty(leaf);
1227         }
1228         return ret;
1229 }
1230
1231 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1232                                           struct btrfs_path *path,
1233                                           struct btrfs_extent_inline_ref *iref)
1234 {
1235         struct btrfs_key key;
1236         struct extent_buffer *leaf;
1237         struct btrfs_extent_data_ref *ref1;
1238         struct btrfs_shared_data_ref *ref2;
1239         u32 num_refs = 0;
1240
1241         leaf = path->nodes[0];
1242         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1243         if (iref) {
1244                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1245                     BTRFS_EXTENT_DATA_REF_KEY) {
1246                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1247                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1248                 } else {
1249                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1250                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1251                 }
1252         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1253                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1254                                       struct btrfs_extent_data_ref);
1255                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1256         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1257                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_shared_data_ref);
1259                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1260 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1261         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1262                 struct btrfs_extent_ref_v0 *ref0;
1263                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1264                                       struct btrfs_extent_ref_v0);
1265                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1266 #endif
1267         } else {
1268                 WARN_ON(1);
1269         }
1270         return num_refs;
1271 }
1272
1273 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1274                                           struct btrfs_root *root,
1275                                           struct btrfs_path *path,
1276                                           u64 bytenr, u64 parent,
1277                                           u64 root_objectid)
1278 {
1279         struct btrfs_key key;
1280         int ret;
1281
1282         key.objectid = bytenr;
1283         if (parent) {
1284                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1285                 key.offset = parent;
1286         } else {
1287                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1288                 key.offset = root_objectid;
1289         }
1290
1291         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1292         if (ret > 0)
1293                 ret = -ENOENT;
1294 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1295         if (ret == -ENOENT && parent) {
1296                 btrfs_release_path(root, path);
1297                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1298                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1299                 if (ret > 0)
1300                         ret = -ENOENT;
1301         }
1302 #endif
1303         return ret;
1304 }
1305
1306 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1307                                           struct btrfs_root *root,
1308                                           struct btrfs_path *path,
1309                                           u64 bytenr, u64 parent,
1310                                           u64 root_objectid)
1311 {
1312         struct btrfs_key key;
1313         int ret;
1314
1315         key.objectid = bytenr;
1316         if (parent) {
1317                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1318                 key.offset = parent;
1319         } else {
1320                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1321                 key.offset = root_objectid;
1322         }
1323
1324         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1325         btrfs_release_path(root, path);
1326         return ret;
1327 }
1328
1329 static inline int extent_ref_type(u64 parent, u64 owner)
1330 {
1331         int type;
1332         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1333                 if (parent > 0)
1334                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1335                 else
1336                         type = BTRFS_TREE_BLOCK_REF_KEY;
1337         } else {
1338                 if (parent > 0)
1339                         type = BTRFS_SHARED_DATA_REF_KEY;
1340                 else
1341                         type = BTRFS_EXTENT_DATA_REF_KEY;
1342         }
1343         return type;
1344 }
1345
1346 static int find_next_key(struct btrfs_path *path, int level,
1347                          struct btrfs_key *key)
1348
1349 {
1350         for (; level < BTRFS_MAX_LEVEL; level++) {
1351                 if (!path->nodes[level])
1352                         break;
1353                 if (path->slots[level] + 1 >=
1354                     btrfs_header_nritems(path->nodes[level]))
1355                         continue;
1356                 if (level == 0)
1357                         btrfs_item_key_to_cpu(path->nodes[level], key,
1358                                               path->slots[level] + 1);
1359                 else
1360                         btrfs_node_key_to_cpu(path->nodes[level], key,
1361                                               path->slots[level] + 1);
1362                 return 0;
1363         }
1364         return 1;
1365 }
1366
1367 /*
1368  * look for inline back ref. if back ref is found, *ref_ret is set
1369  * to the address of inline back ref, and 0 is returned.
1370  *
1371  * if back ref isn't found, *ref_ret is set to the address where it
1372  * should be inserted, and -ENOENT is returned.
1373  *
1374  * if insert is true and there are too many inline back refs, the path
1375  * points to the extent item, and -EAGAIN is returned.
1376  *
1377  * NOTE: inline back refs are ordered in the same way that back ref
1378  *       items in the tree are ordered.
1379  */
1380 static noinline_for_stack
1381 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1382                                  struct btrfs_root *root,
1383                                  struct btrfs_path *path,
1384                                  struct btrfs_extent_inline_ref **ref_ret,
1385                                  u64 bytenr, u64 num_bytes,
1386                                  u64 parent, u64 root_objectid,
1387                                  u64 owner, u64 offset, int insert)
1388 {
1389         struct btrfs_key key;
1390         struct extent_buffer *leaf;
1391         struct btrfs_extent_item *ei;
1392         struct btrfs_extent_inline_ref *iref;
1393         u64 flags;
1394         u64 item_size;
1395         unsigned long ptr;
1396         unsigned long end;
1397         int extra_size;
1398         int type;
1399         int want;
1400         int ret;
1401         int err = 0;
1402
1403         key.objectid = bytenr;
1404         key.type = BTRFS_EXTENT_ITEM_KEY;
1405         key.offset = num_bytes;
1406
1407         want = extent_ref_type(parent, owner);
1408         if (insert) {
1409                 extra_size = btrfs_extent_inline_ref_size(want);
1410                 path->keep_locks = 1;
1411         } else
1412                 extra_size = -1;
1413         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1414         if (ret < 0) {
1415                 err = ret;
1416                 goto out;
1417         }
1418         BUG_ON(ret);
1419
1420         leaf = path->nodes[0];
1421         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1422 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1423         if (item_size < sizeof(*ei)) {
1424                 if (!insert) {
1425                         err = -ENOENT;
1426                         goto out;
1427                 }
1428                 ret = convert_extent_item_v0(trans, root, path, owner,
1429                                              extra_size);
1430                 if (ret < 0) {
1431                         err = ret;
1432                         goto out;
1433                 }
1434                 leaf = path->nodes[0];
1435                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1436         }
1437 #endif
1438         BUG_ON(item_size < sizeof(*ei));
1439
1440         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1441         flags = btrfs_extent_flags(leaf, ei);
1442
1443         ptr = (unsigned long)(ei + 1);
1444         end = (unsigned long)ei + item_size;
1445
1446         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1447                 ptr += sizeof(struct btrfs_tree_block_info);
1448                 BUG_ON(ptr > end);
1449         } else {
1450                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1451         }
1452
1453         err = -ENOENT;
1454         while (1) {
1455                 if (ptr >= end) {
1456                         WARN_ON(ptr > end);
1457                         break;
1458                 }
1459                 iref = (struct btrfs_extent_inline_ref *)ptr;
1460                 type = btrfs_extent_inline_ref_type(leaf, iref);
1461                 if (want < type)
1462                         break;
1463                 if (want > type) {
1464                         ptr += btrfs_extent_inline_ref_size(type);
1465                         continue;
1466                 }
1467
1468                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1469                         struct btrfs_extent_data_ref *dref;
1470                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1471                         if (match_extent_data_ref(leaf, dref, root_objectid,
1472                                                   owner, offset)) {
1473                                 err = 0;
1474                                 break;
1475                         }
1476                         if (hash_extent_data_ref_item(leaf, dref) <
1477                             hash_extent_data_ref(root_objectid, owner, offset))
1478                                 break;
1479                 } else {
1480                         u64 ref_offset;
1481                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1482                         if (parent > 0) {
1483                                 if (parent == ref_offset) {
1484                                         err = 0;
1485                                         break;
1486                                 }
1487                                 if (ref_offset < parent)
1488                                         break;
1489                         } else {
1490                                 if (root_objectid == ref_offset) {
1491                                         err = 0;
1492                                         break;
1493                                 }
1494                                 if (ref_offset < root_objectid)
1495                                         break;
1496                         }
1497                 }
1498                 ptr += btrfs_extent_inline_ref_size(type);
1499         }
1500         if (err == -ENOENT && insert) {
1501                 if (item_size + extra_size >=
1502                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1503                         err = -EAGAIN;
1504                         goto out;
1505                 }
1506                 /*
1507                  * To add new inline back ref, we have to make sure
1508                  * there is no corresponding back ref item.
1509                  * For simplicity, we just do not add new inline back
1510                  * ref if there is any kind of item for this block
1511                  */
1512                 if (find_next_key(path, 0, &key) == 0 &&
1513                     key.objectid == bytenr &&
1514                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1515                         err = -EAGAIN;
1516                         goto out;
1517                 }
1518         }
1519         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1520 out:
1521         if (insert) {
1522                 path->keep_locks = 0;
1523                 btrfs_unlock_up_safe(path, 1);
1524         }
1525         return err;
1526 }
1527
1528 /*
1529  * helper to add new inline back ref
1530  */
1531 static noinline_for_stack
1532 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1533                                 struct btrfs_root *root,
1534                                 struct btrfs_path *path,
1535                                 struct btrfs_extent_inline_ref *iref,
1536                                 u64 parent, u64 root_objectid,
1537                                 u64 owner, u64 offset, int refs_to_add,
1538                                 struct btrfs_delayed_extent_op *extent_op)
1539 {
1540         struct extent_buffer *leaf;
1541         struct btrfs_extent_item *ei;
1542         unsigned long ptr;
1543         unsigned long end;
1544         unsigned long item_offset;
1545         u64 refs;
1546         int size;
1547         int type;
1548         int ret;
1549
1550         leaf = path->nodes[0];
1551         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1552         item_offset = (unsigned long)iref - (unsigned long)ei;
1553
1554         type = extent_ref_type(parent, owner);
1555         size = btrfs_extent_inline_ref_size(type);
1556
1557         ret = btrfs_extend_item(trans, root, path, size);
1558         BUG_ON(ret);
1559
1560         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1561         refs = btrfs_extent_refs(leaf, ei);
1562         refs += refs_to_add;
1563         btrfs_set_extent_refs(leaf, ei, refs);
1564         if (extent_op)
1565                 __run_delayed_extent_op(extent_op, leaf, ei);
1566
1567         ptr = (unsigned long)ei + item_offset;
1568         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1569         if (ptr < end - size)
1570                 memmove_extent_buffer(leaf, ptr + size, ptr,
1571                                       end - size - ptr);
1572
1573         iref = (struct btrfs_extent_inline_ref *)ptr;
1574         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1575         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1576                 struct btrfs_extent_data_ref *dref;
1577                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1578                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1579                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1580                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1581                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1582         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1583                 struct btrfs_shared_data_ref *sref;
1584                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1585                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1586                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1587         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1588                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1589         } else {
1590                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1591         }
1592         btrfs_mark_buffer_dirty(leaf);
1593         return 0;
1594 }
1595
1596 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1597                                  struct btrfs_root *root,
1598                                  struct btrfs_path *path,
1599                                  struct btrfs_extent_inline_ref **ref_ret,
1600                                  u64 bytenr, u64 num_bytes, u64 parent,
1601                                  u64 root_objectid, u64 owner, u64 offset)
1602 {
1603         int ret;
1604
1605         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1606                                            bytenr, num_bytes, parent,
1607                                            root_objectid, owner, offset, 0);
1608         if (ret != -ENOENT)
1609                 return ret;
1610
1611         btrfs_release_path(root, path);
1612         *ref_ret = NULL;
1613
1614         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1615                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1616                                             root_objectid);
1617         } else {
1618                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1619                                              root_objectid, owner, offset);
1620         }
1621         return ret;
1622 }
1623
1624 /*
1625  * helper to update/remove inline back ref
1626  */
1627 static noinline_for_stack
1628 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1629                                  struct btrfs_root *root,
1630                                  struct btrfs_path *path,
1631                                  struct btrfs_extent_inline_ref *iref,
1632                                  int refs_to_mod,
1633                                  struct btrfs_delayed_extent_op *extent_op)
1634 {
1635         struct extent_buffer *leaf;
1636         struct btrfs_extent_item *ei;
1637         struct btrfs_extent_data_ref *dref = NULL;
1638         struct btrfs_shared_data_ref *sref = NULL;
1639         unsigned long ptr;
1640         unsigned long end;
1641         u32 item_size;
1642         int size;
1643         int type;
1644         int ret;
1645         u64 refs;
1646
1647         leaf = path->nodes[0];
1648         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1649         refs = btrfs_extent_refs(leaf, ei);
1650         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1651         refs += refs_to_mod;
1652         btrfs_set_extent_refs(leaf, ei, refs);
1653         if (extent_op)
1654                 __run_delayed_extent_op(extent_op, leaf, ei);
1655
1656         type = btrfs_extent_inline_ref_type(leaf, iref);
1657
1658         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1659                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1660                 refs = btrfs_extent_data_ref_count(leaf, dref);
1661         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1662                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1663                 refs = btrfs_shared_data_ref_count(leaf, sref);
1664         } else {
1665                 refs = 1;
1666                 BUG_ON(refs_to_mod != -1);
1667         }
1668
1669         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1670         refs += refs_to_mod;
1671
1672         if (refs > 0) {
1673                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1674                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1675                 else
1676                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1677         } else {
1678                 size =  btrfs_extent_inline_ref_size(type);
1679                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1680                 ptr = (unsigned long)iref;
1681                 end = (unsigned long)ei + item_size;
1682                 if (ptr + size < end)
1683                         memmove_extent_buffer(leaf, ptr, ptr + size,
1684                                               end - ptr - size);
1685                 item_size -= size;
1686                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1687                 BUG_ON(ret);
1688         }
1689         btrfs_mark_buffer_dirty(leaf);
1690         return 0;
1691 }
1692
1693 static noinline_for_stack
1694 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1695                                  struct btrfs_root *root,
1696                                  struct btrfs_path *path,
1697                                  u64 bytenr, u64 num_bytes, u64 parent,
1698                                  u64 root_objectid, u64 owner,
1699                                  u64 offset, int refs_to_add,
1700                                  struct btrfs_delayed_extent_op *extent_op)
1701 {
1702         struct btrfs_extent_inline_ref *iref;
1703         int ret;
1704
1705         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1706                                            bytenr, num_bytes, parent,
1707                                            root_objectid, owner, offset, 1);
1708         if (ret == 0) {
1709                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1710                 ret = update_inline_extent_backref(trans, root, path, iref,
1711                                                    refs_to_add, extent_op);
1712         } else if (ret == -ENOENT) {
1713                 ret = setup_inline_extent_backref(trans, root, path, iref,
1714                                                   parent, root_objectid,
1715                                                   owner, offset, refs_to_add,
1716                                                   extent_op);
1717         }
1718         return ret;
1719 }
1720
1721 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1722                                  struct btrfs_root *root,
1723                                  struct btrfs_path *path,
1724                                  u64 bytenr, u64 parent, u64 root_objectid,
1725                                  u64 owner, u64 offset, int refs_to_add)
1726 {
1727         int ret;
1728         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1729                 BUG_ON(refs_to_add != 1);
1730                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1731                                             parent, root_objectid);
1732         } else {
1733                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1734                                              parent, root_objectid,
1735                                              owner, offset, refs_to_add);
1736         }
1737         return ret;
1738 }
1739
1740 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1741                                  struct btrfs_root *root,
1742                                  struct btrfs_path *path,
1743                                  struct btrfs_extent_inline_ref *iref,
1744                                  int refs_to_drop, int is_data)
1745 {
1746         int ret;
1747
1748         BUG_ON(!is_data && refs_to_drop != 1);
1749         if (iref) {
1750                 ret = update_inline_extent_backref(trans, root, path, iref,
1751                                                    -refs_to_drop, NULL);
1752         } else if (is_data) {
1753                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1754         } else {
1755                 ret = btrfs_del_item(trans, root, path);
1756         }
1757         return ret;
1758 }
1759
1760 static int btrfs_issue_discard(struct block_device *bdev,
1761                                 u64 start, u64 len)
1762 {
1763         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1764 }
1765
1766 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1767                                 u64 num_bytes, u64 *actual_bytes)
1768 {
1769         int ret;
1770         u64 discarded_bytes = 0;
1771         struct btrfs_multi_bio *multi = NULL;
1772
1773
1774         /* Tell the block device(s) that the sectors can be discarded */
1775         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1776                               bytenr, &num_bytes, &multi, 0);
1777         if (!ret) {
1778                 struct btrfs_bio_stripe *stripe = multi->stripes;
1779                 int i;
1780
1781
1782                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1783                         ret = btrfs_issue_discard(stripe->dev->bdev,
1784                                                   stripe->physical,
1785                                                   stripe->length);
1786                         if (!ret)
1787                                 discarded_bytes += stripe->length;
1788                         else if (ret != -EOPNOTSUPP)
1789                                 break;
1790                 }
1791                 kfree(multi);
1792         }
1793         if (discarded_bytes && ret == -EOPNOTSUPP)
1794                 ret = 0;
1795
1796         if (actual_bytes)
1797                 *actual_bytes = discarded_bytes;
1798
1799
1800         return ret;
1801 }
1802
1803 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1804                          struct btrfs_root *root,
1805                          u64 bytenr, u64 num_bytes, u64 parent,
1806                          u64 root_objectid, u64 owner, u64 offset)
1807 {
1808         int ret;
1809         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1810                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1811
1812         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1813                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1814                                         parent, root_objectid, (int)owner,
1815                                         BTRFS_ADD_DELAYED_REF, NULL);
1816         } else {
1817                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1818                                         parent, root_objectid, owner, offset,
1819                                         BTRFS_ADD_DELAYED_REF, NULL);
1820         }
1821         return ret;
1822 }
1823
1824 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1825                                   struct btrfs_root *root,
1826                                   u64 bytenr, u64 num_bytes,
1827                                   u64 parent, u64 root_objectid,
1828                                   u64 owner, u64 offset, int refs_to_add,
1829                                   struct btrfs_delayed_extent_op *extent_op)
1830 {
1831         struct btrfs_path *path;
1832         struct extent_buffer *leaf;
1833         struct btrfs_extent_item *item;
1834         u64 refs;
1835         int ret;
1836         int err = 0;
1837
1838         path = btrfs_alloc_path();
1839         if (!path)
1840                 return -ENOMEM;
1841
1842         path->reada = 1;
1843         path->leave_spinning = 1;
1844         /* this will setup the path even if it fails to insert the back ref */
1845         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1846                                            path, bytenr, num_bytes, parent,
1847                                            root_objectid, owner, offset,
1848                                            refs_to_add, extent_op);
1849         if (ret == 0)
1850                 goto out;
1851
1852         if (ret != -EAGAIN) {
1853                 err = ret;
1854                 goto out;
1855         }
1856
1857         leaf = path->nodes[0];
1858         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1859         refs = btrfs_extent_refs(leaf, item);
1860         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1861         if (extent_op)
1862                 __run_delayed_extent_op(extent_op, leaf, item);
1863
1864         btrfs_mark_buffer_dirty(leaf);
1865         btrfs_release_path(root->fs_info->extent_root, path);
1866
1867         path->reada = 1;
1868         path->leave_spinning = 1;
1869
1870         /* now insert the actual backref */
1871         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1872                                     path, bytenr, parent, root_objectid,
1873                                     owner, offset, refs_to_add);
1874         BUG_ON(ret);
1875 out:
1876         btrfs_free_path(path);
1877         return err;
1878 }
1879
1880 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1881                                 struct btrfs_root *root,
1882                                 struct btrfs_delayed_ref_node *node,
1883                                 struct btrfs_delayed_extent_op *extent_op,
1884                                 int insert_reserved)
1885 {
1886         int ret = 0;
1887         struct btrfs_delayed_data_ref *ref;
1888         struct btrfs_key ins;
1889         u64 parent = 0;
1890         u64 ref_root = 0;
1891         u64 flags = 0;
1892
1893         ins.objectid = node->bytenr;
1894         ins.offset = node->num_bytes;
1895         ins.type = BTRFS_EXTENT_ITEM_KEY;
1896
1897         ref = btrfs_delayed_node_to_data_ref(node);
1898         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1899                 parent = ref->parent;
1900         else
1901                 ref_root = ref->root;
1902
1903         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1904                 if (extent_op) {
1905                         BUG_ON(extent_op->update_key);
1906                         flags |= extent_op->flags_to_set;
1907                 }
1908                 ret = alloc_reserved_file_extent(trans, root,
1909                                                  parent, ref_root, flags,
1910                                                  ref->objectid, ref->offset,
1911                                                  &ins, node->ref_mod);
1912         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1913                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1914                                              node->num_bytes, parent,
1915                                              ref_root, ref->objectid,
1916                                              ref->offset, node->ref_mod,
1917                                              extent_op);
1918         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1919                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1920                                           node->num_bytes, parent,
1921                                           ref_root, ref->objectid,
1922                                           ref->offset, node->ref_mod,
1923                                           extent_op);
1924         } else {
1925                 BUG();
1926         }
1927         return ret;
1928 }
1929
1930 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1931                                     struct extent_buffer *leaf,
1932                                     struct btrfs_extent_item *ei)
1933 {
1934         u64 flags = btrfs_extent_flags(leaf, ei);
1935         if (extent_op->update_flags) {
1936                 flags |= extent_op->flags_to_set;
1937                 btrfs_set_extent_flags(leaf, ei, flags);
1938         }
1939
1940         if (extent_op->update_key) {
1941                 struct btrfs_tree_block_info *bi;
1942                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1943                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1944                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1945         }
1946 }
1947
1948 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1949                                  struct btrfs_root *root,
1950                                  struct btrfs_delayed_ref_node *node,
1951                                  struct btrfs_delayed_extent_op *extent_op)
1952 {
1953         struct btrfs_key key;
1954         struct btrfs_path *path;
1955         struct btrfs_extent_item *ei;
1956         struct extent_buffer *leaf;
1957         u32 item_size;
1958         int ret;
1959         int err = 0;
1960
1961         path = btrfs_alloc_path();
1962         if (!path)
1963                 return -ENOMEM;
1964
1965         key.objectid = node->bytenr;
1966         key.type = BTRFS_EXTENT_ITEM_KEY;
1967         key.offset = node->num_bytes;
1968
1969         path->reada = 1;
1970         path->leave_spinning = 1;
1971         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1972                                 path, 0, 1);
1973         if (ret < 0) {
1974                 err = ret;
1975                 goto out;
1976         }
1977         if (ret > 0) {
1978                 err = -EIO;
1979                 goto out;
1980         }
1981
1982         leaf = path->nodes[0];
1983         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1984 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1985         if (item_size < sizeof(*ei)) {
1986                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
1987                                              path, (u64)-1, 0);
1988                 if (ret < 0) {
1989                         err = ret;
1990                         goto out;
1991                 }
1992                 leaf = path->nodes[0];
1993                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1994         }
1995 #endif
1996         BUG_ON(item_size < sizeof(*ei));
1997         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1998         __run_delayed_extent_op(extent_op, leaf, ei);
1999
2000         btrfs_mark_buffer_dirty(leaf);
2001 out:
2002         btrfs_free_path(path);
2003         return err;
2004 }
2005
2006 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2007                                 struct btrfs_root *root,
2008                                 struct btrfs_delayed_ref_node *node,
2009                                 struct btrfs_delayed_extent_op *extent_op,
2010                                 int insert_reserved)
2011 {
2012         int ret = 0;
2013         struct btrfs_delayed_tree_ref *ref;
2014         struct btrfs_key ins;
2015         u64 parent = 0;
2016         u64 ref_root = 0;
2017
2018         ins.objectid = node->bytenr;
2019         ins.offset = node->num_bytes;
2020         ins.type = BTRFS_EXTENT_ITEM_KEY;
2021
2022         ref = btrfs_delayed_node_to_tree_ref(node);
2023         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2024                 parent = ref->parent;
2025         else
2026                 ref_root = ref->root;
2027
2028         BUG_ON(node->ref_mod != 1);
2029         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2030                 BUG_ON(!extent_op || !extent_op->update_flags ||
2031                        !extent_op->update_key);
2032                 ret = alloc_reserved_tree_block(trans, root,
2033                                                 parent, ref_root,
2034                                                 extent_op->flags_to_set,
2035                                                 &extent_op->key,
2036                                                 ref->level, &ins);
2037         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2038                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2039                                              node->num_bytes, parent, ref_root,
2040                                              ref->level, 0, 1, extent_op);
2041         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2042                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2043                                           node->num_bytes, parent, ref_root,
2044                                           ref->level, 0, 1, extent_op);
2045         } else {
2046                 BUG();
2047         }
2048         return ret;
2049 }
2050
2051 /* helper function to actually process a single delayed ref entry */
2052 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2053                                struct btrfs_root *root,
2054                                struct btrfs_delayed_ref_node *node,
2055                                struct btrfs_delayed_extent_op *extent_op,
2056                                int insert_reserved)
2057 {
2058         int ret;
2059         if (btrfs_delayed_ref_is_head(node)) {
2060                 struct btrfs_delayed_ref_head *head;
2061                 /*
2062                  * we've hit the end of the chain and we were supposed
2063                  * to insert this extent into the tree.  But, it got
2064                  * deleted before we ever needed to insert it, so all
2065                  * we have to do is clean up the accounting
2066                  */
2067                 BUG_ON(extent_op);
2068                 head = btrfs_delayed_node_to_head(node);
2069                 if (insert_reserved) {
2070                         btrfs_pin_extent(root, node->bytenr,
2071                                          node->num_bytes, 1);
2072                         if (head->is_data) {
2073                                 ret = btrfs_del_csums(trans, root,
2074                                                       node->bytenr,
2075                                                       node->num_bytes);
2076                                 BUG_ON(ret);
2077                         }
2078                 }
2079                 mutex_unlock(&head->mutex);
2080                 return 0;
2081         }
2082
2083         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2084             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2085                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2086                                            insert_reserved);
2087         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2088                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2089                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2090                                            insert_reserved);
2091         else
2092                 BUG();
2093         return ret;
2094 }
2095
2096 static noinline struct btrfs_delayed_ref_node *
2097 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2098 {
2099         struct rb_node *node;
2100         struct btrfs_delayed_ref_node *ref;
2101         int action = BTRFS_ADD_DELAYED_REF;
2102 again:
2103         /*
2104          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2105          * this prevents ref count from going down to zero when
2106          * there still are pending delayed ref.
2107          */
2108         node = rb_prev(&head->node.rb_node);
2109         while (1) {
2110                 if (!node)
2111                         break;
2112                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2113                                 rb_node);
2114                 if (ref->bytenr != head->node.bytenr)
2115                         break;
2116                 if (ref->action == action)
2117                         return ref;
2118                 node = rb_prev(node);
2119         }
2120         if (action == BTRFS_ADD_DELAYED_REF) {
2121                 action = BTRFS_DROP_DELAYED_REF;
2122                 goto again;
2123         }
2124         return NULL;
2125 }
2126
2127 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2128                                        struct btrfs_root *root,
2129                                        struct list_head *cluster)
2130 {
2131         struct btrfs_delayed_ref_root *delayed_refs;
2132         struct btrfs_delayed_ref_node *ref;
2133         struct btrfs_delayed_ref_head *locked_ref = NULL;
2134         struct btrfs_delayed_extent_op *extent_op;
2135         int ret;
2136         int count = 0;
2137         int must_insert_reserved = 0;
2138
2139         delayed_refs = &trans->transaction->delayed_refs;
2140         while (1) {
2141                 if (!locked_ref) {
2142                         /* pick a new head ref from the cluster list */
2143                         if (list_empty(cluster))
2144                                 break;
2145
2146                         locked_ref = list_entry(cluster->next,
2147                                      struct btrfs_delayed_ref_head, cluster);
2148
2149                         /* grab the lock that says we are going to process
2150                          * all the refs for this head */
2151                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2152
2153                         /*
2154                          * we may have dropped the spin lock to get the head
2155                          * mutex lock, and that might have given someone else
2156                          * time to free the head.  If that's true, it has been
2157                          * removed from our list and we can move on.
2158                          */
2159                         if (ret == -EAGAIN) {
2160                                 locked_ref = NULL;
2161                                 count++;
2162                                 continue;
2163                         }
2164                 }
2165
2166                 /*
2167                  * record the must insert reserved flag before we
2168                  * drop the spin lock.
2169                  */
2170                 must_insert_reserved = locked_ref->must_insert_reserved;
2171                 locked_ref->must_insert_reserved = 0;
2172
2173                 extent_op = locked_ref->extent_op;
2174                 locked_ref->extent_op = NULL;
2175
2176                 /*
2177                  * locked_ref is the head node, so we have to go one
2178                  * node back for any delayed ref updates
2179                  */
2180                 ref = select_delayed_ref(locked_ref);
2181                 if (!ref) {
2182                         /* All delayed refs have been processed, Go ahead
2183                          * and send the head node to run_one_delayed_ref,
2184                          * so that any accounting fixes can happen
2185                          */
2186                         ref = &locked_ref->node;
2187
2188                         if (extent_op && must_insert_reserved) {
2189                                 kfree(extent_op);
2190                                 extent_op = NULL;
2191                         }
2192
2193                         if (extent_op) {
2194                                 spin_unlock(&delayed_refs->lock);
2195
2196                                 ret = run_delayed_extent_op(trans, root,
2197                                                             ref, extent_op);
2198                                 BUG_ON(ret);
2199                                 kfree(extent_op);
2200
2201                                 cond_resched();
2202                                 spin_lock(&delayed_refs->lock);
2203                                 continue;
2204                         }
2205
2206                         list_del_init(&locked_ref->cluster);
2207                         locked_ref = NULL;
2208                 }
2209
2210                 ref->in_tree = 0;
2211                 rb_erase(&ref->rb_node, &delayed_refs->root);
2212                 delayed_refs->num_entries--;
2213
2214                 spin_unlock(&delayed_refs->lock);
2215
2216                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2217                                           must_insert_reserved);
2218                 BUG_ON(ret);
2219
2220                 btrfs_put_delayed_ref(ref);
2221                 kfree(extent_op);
2222                 count++;
2223
2224                 cond_resched();
2225                 spin_lock(&delayed_refs->lock);
2226         }
2227         return count;
2228 }
2229
2230 /*
2231  * this starts processing the delayed reference count updates and
2232  * extent insertions we have queued up so far.  count can be
2233  * 0, which means to process everything in the tree at the start
2234  * of the run (but not newly added entries), or it can be some target
2235  * number you'd like to process.
2236  */
2237 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2238                            struct btrfs_root *root, unsigned long count)
2239 {
2240         struct rb_node *node;
2241         struct btrfs_delayed_ref_root *delayed_refs;
2242         struct btrfs_delayed_ref_node *ref;
2243         struct list_head cluster;
2244         int ret;
2245         int run_all = count == (unsigned long)-1;
2246         int run_most = 0;
2247
2248         if (root == root->fs_info->extent_root)
2249                 root = root->fs_info->tree_root;
2250
2251         delayed_refs = &trans->transaction->delayed_refs;
2252         INIT_LIST_HEAD(&cluster);
2253 again:
2254         spin_lock(&delayed_refs->lock);
2255         if (count == 0) {
2256                 count = delayed_refs->num_entries * 2;
2257                 run_most = 1;
2258         }
2259         while (1) {
2260                 if (!(run_all || run_most) &&
2261                     delayed_refs->num_heads_ready < 64)
2262                         break;
2263
2264                 /*
2265                  * go find something we can process in the rbtree.  We start at
2266                  * the beginning of the tree, and then build a cluster
2267                  * of refs to process starting at the first one we are able to
2268                  * lock
2269                  */
2270                 ret = btrfs_find_ref_cluster(trans, &cluster,
2271                                              delayed_refs->run_delayed_start);
2272                 if (ret)
2273                         break;
2274
2275                 ret = run_clustered_refs(trans, root, &cluster);
2276                 BUG_ON(ret < 0);
2277
2278                 count -= min_t(unsigned long, ret, count);
2279
2280                 if (count == 0)
2281                         break;
2282         }
2283
2284         if (run_all) {
2285                 node = rb_first(&delayed_refs->root);
2286                 if (!node)
2287                         goto out;
2288                 count = (unsigned long)-1;
2289
2290                 while (node) {
2291                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2292                                        rb_node);
2293                         if (btrfs_delayed_ref_is_head(ref)) {
2294                                 struct btrfs_delayed_ref_head *head;
2295
2296                                 head = btrfs_delayed_node_to_head(ref);
2297                                 atomic_inc(&ref->refs);
2298
2299                                 spin_unlock(&delayed_refs->lock);
2300                                 mutex_lock(&head->mutex);
2301                                 mutex_unlock(&head->mutex);
2302
2303                                 btrfs_put_delayed_ref(ref);
2304                                 cond_resched();
2305                                 goto again;
2306                         }
2307                         node = rb_next(node);
2308                 }
2309                 spin_unlock(&delayed_refs->lock);
2310                 schedule_timeout(1);
2311                 goto again;
2312         }
2313 out:
2314         spin_unlock(&delayed_refs->lock);
2315         return 0;
2316 }
2317
2318 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2319                                 struct btrfs_root *root,
2320                                 u64 bytenr, u64 num_bytes, u64 flags,
2321                                 int is_data)
2322 {
2323         struct btrfs_delayed_extent_op *extent_op;
2324         int ret;
2325
2326         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2327         if (!extent_op)
2328                 return -ENOMEM;
2329
2330         extent_op->flags_to_set = flags;
2331         extent_op->update_flags = 1;
2332         extent_op->update_key = 0;
2333         extent_op->is_data = is_data ? 1 : 0;
2334
2335         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2336         if (ret)
2337                 kfree(extent_op);
2338         return ret;
2339 }
2340
2341 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2342                                       struct btrfs_root *root,
2343                                       struct btrfs_path *path,
2344                                       u64 objectid, u64 offset, u64 bytenr)
2345 {
2346         struct btrfs_delayed_ref_head *head;
2347         struct btrfs_delayed_ref_node *ref;
2348         struct btrfs_delayed_data_ref *data_ref;
2349         struct btrfs_delayed_ref_root *delayed_refs;
2350         struct rb_node *node;
2351         int ret = 0;
2352
2353         ret = -ENOENT;
2354         delayed_refs = &trans->transaction->delayed_refs;
2355         spin_lock(&delayed_refs->lock);
2356         head = btrfs_find_delayed_ref_head(trans, bytenr);
2357         if (!head)
2358                 goto out;
2359
2360         if (!mutex_trylock(&head->mutex)) {
2361                 atomic_inc(&head->node.refs);
2362                 spin_unlock(&delayed_refs->lock);
2363
2364                 btrfs_release_path(root->fs_info->extent_root, path);
2365
2366                 mutex_lock(&head->mutex);
2367                 mutex_unlock(&head->mutex);
2368                 btrfs_put_delayed_ref(&head->node);
2369                 return -EAGAIN;
2370         }
2371
2372         node = rb_prev(&head->node.rb_node);
2373         if (!node)
2374                 goto out_unlock;
2375
2376         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2377
2378         if (ref->bytenr != bytenr)
2379                 goto out_unlock;
2380
2381         ret = 1;
2382         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2383                 goto out_unlock;
2384
2385         data_ref = btrfs_delayed_node_to_data_ref(ref);
2386
2387         node = rb_prev(node);
2388         if (node) {
2389                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2390                 if (ref->bytenr == bytenr)
2391                         goto out_unlock;
2392         }
2393
2394         if (data_ref->root != root->root_key.objectid ||
2395             data_ref->objectid != objectid || data_ref->offset != offset)
2396                 goto out_unlock;
2397
2398         ret = 0;
2399 out_unlock:
2400         mutex_unlock(&head->mutex);
2401 out:
2402         spin_unlock(&delayed_refs->lock);
2403         return ret;
2404 }
2405
2406 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2407                                         struct btrfs_root *root,
2408                                         struct btrfs_path *path,
2409                                         u64 objectid, u64 offset, u64 bytenr)
2410 {
2411         struct btrfs_root *extent_root = root->fs_info->extent_root;
2412         struct extent_buffer *leaf;
2413         struct btrfs_extent_data_ref *ref;
2414         struct btrfs_extent_inline_ref *iref;
2415         struct btrfs_extent_item *ei;
2416         struct btrfs_key key;
2417         u32 item_size;
2418         int ret;
2419
2420         key.objectid = bytenr;
2421         key.offset = (u64)-1;
2422         key.type = BTRFS_EXTENT_ITEM_KEY;
2423
2424         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2425         if (ret < 0)
2426                 goto out;
2427         BUG_ON(ret == 0);
2428
2429         ret = -ENOENT;
2430         if (path->slots[0] == 0)
2431                 goto out;
2432
2433         path->slots[0]--;
2434         leaf = path->nodes[0];
2435         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2436
2437         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2438                 goto out;
2439
2440         ret = 1;
2441         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2442 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2443         if (item_size < sizeof(*ei)) {
2444                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2445                 goto out;
2446         }
2447 #endif
2448         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2449
2450         if (item_size != sizeof(*ei) +
2451             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2452                 goto out;
2453
2454         if (btrfs_extent_generation(leaf, ei) <=
2455             btrfs_root_last_snapshot(&root->root_item))
2456                 goto out;
2457
2458         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2459         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2460             BTRFS_EXTENT_DATA_REF_KEY)
2461                 goto out;
2462
2463         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2464         if (btrfs_extent_refs(leaf, ei) !=
2465             btrfs_extent_data_ref_count(leaf, ref) ||
2466             btrfs_extent_data_ref_root(leaf, ref) !=
2467             root->root_key.objectid ||
2468             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2469             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2470                 goto out;
2471
2472         ret = 0;
2473 out:
2474         return ret;
2475 }
2476
2477 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2478                           struct btrfs_root *root,
2479                           u64 objectid, u64 offset, u64 bytenr)
2480 {
2481         struct btrfs_path *path;
2482         int ret;
2483         int ret2;
2484
2485         path = btrfs_alloc_path();
2486         if (!path)
2487                 return -ENOENT;
2488
2489         do {
2490                 ret = check_committed_ref(trans, root, path, objectid,
2491                                           offset, bytenr);
2492                 if (ret && ret != -ENOENT)
2493                         goto out;
2494
2495                 ret2 = check_delayed_ref(trans, root, path, objectid,
2496                                          offset, bytenr);
2497         } while (ret2 == -EAGAIN);
2498
2499         if (ret2 && ret2 != -ENOENT) {
2500                 ret = ret2;
2501                 goto out;
2502         }
2503
2504         if (ret != -ENOENT || ret2 != -ENOENT)
2505                 ret = 0;
2506 out:
2507         btrfs_free_path(path);
2508         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2509                 WARN_ON(ret > 0);
2510         return ret;
2511 }
2512
2513 #if 0
2514 int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2515                     struct extent_buffer *buf, u32 nr_extents)
2516 {
2517         struct btrfs_key key;
2518         struct btrfs_file_extent_item *fi;
2519         u64 root_gen;
2520         u32 nritems;
2521         int i;
2522         int level;
2523         int ret = 0;
2524         int shared = 0;
2525
2526         if (!root->ref_cows)
2527                 return 0;
2528
2529         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
2530                 shared = 0;
2531                 root_gen = root->root_key.offset;
2532         } else {
2533                 shared = 1;
2534                 root_gen = trans->transid - 1;
2535         }
2536
2537         level = btrfs_header_level(buf);
2538         nritems = btrfs_header_nritems(buf);
2539
2540         if (level == 0) {
2541                 struct btrfs_leaf_ref *ref;
2542                 struct btrfs_extent_info *info;
2543
2544                 ref = btrfs_alloc_leaf_ref(root, nr_extents);
2545                 if (!ref) {
2546                         ret = -ENOMEM;
2547                         goto out;
2548                 }
2549
2550                 ref->root_gen = root_gen;
2551                 ref->bytenr = buf->start;
2552                 ref->owner = btrfs_header_owner(buf);
2553                 ref->generation = btrfs_header_generation(buf);
2554                 ref->nritems = nr_extents;
2555                 info = ref->extents;
2556
2557                 for (i = 0; nr_extents > 0 && i < nritems; i++) {
2558                         u64 disk_bytenr;
2559                         btrfs_item_key_to_cpu(buf, &key, i);
2560                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2561                                 continue;
2562                         fi = btrfs_item_ptr(buf, i,
2563                                             struct btrfs_file_extent_item);
2564                         if (btrfs_file_extent_type(buf, fi) ==
2565                             BTRFS_FILE_EXTENT_INLINE)
2566                                 continue;
2567                         disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2568                         if (disk_bytenr == 0)
2569                                 continue;
2570
2571                         info->bytenr = disk_bytenr;
2572                         info->num_bytes =
2573                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2574                         info->objectid = key.objectid;
2575                         info->offset = key.offset;
2576                         info++;
2577                 }
2578
2579                 ret = btrfs_add_leaf_ref(root, ref, shared);
2580                 if (ret == -EEXIST && shared) {
2581                         struct btrfs_leaf_ref *old;
2582                         old = btrfs_lookup_leaf_ref(root, ref->bytenr);
2583                         BUG_ON(!old);
2584                         btrfs_remove_leaf_ref(root, old);
2585                         btrfs_free_leaf_ref(root, old);
2586                         ret = btrfs_add_leaf_ref(root, ref, shared);
2587                 }
2588                 WARN_ON(ret);
2589                 btrfs_free_leaf_ref(root, ref);
2590         }
2591 out:
2592         return ret;
2593 }
2594
2595 /* when a block goes through cow, we update the reference counts of
2596  * everything that block points to.  The internal pointers of the block
2597  * can be in just about any order, and it is likely to have clusters of
2598  * things that are close together and clusters of things that are not.
2599  *
2600  * To help reduce the seeks that come with updating all of these reference
2601  * counts, sort them by byte number before actual updates are done.
2602  *
2603  * struct refsort is used to match byte number to slot in the btree block.
2604  * we sort based on the byte number and then use the slot to actually
2605  * find the item.
2606  *
2607  * struct refsort is smaller than strcut btrfs_item and smaller than
2608  * struct btrfs_key_ptr.  Since we're currently limited to the page size
2609  * for a btree block, there's no way for a kmalloc of refsorts for a
2610  * single node to be bigger than a page.
2611  */
2612 struct refsort {
2613         u64 bytenr;
2614         u32 slot;
2615 };
2616
2617 /*
2618  * for passing into sort()
2619  */
2620 static int refsort_cmp(const void *a_void, const void *b_void)
2621 {
2622         const struct refsort *a = a_void;
2623         const struct refsort *b = b_void;
2624
2625         if (a->bytenr < b->bytenr)
2626                 return -1;
2627         if (a->bytenr > b->bytenr)
2628                 return 1;
2629         return 0;
2630 }
2631 #endif
2632
2633 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2634                            struct btrfs_root *root,
2635                            struct extent_buffer *buf,
2636                            int full_backref, int inc)
2637 {
2638         u64 bytenr;
2639         u64 num_bytes;
2640         u64 parent;
2641         u64 ref_root;
2642         u32 nritems;
2643         struct btrfs_key key;
2644         struct btrfs_file_extent_item *fi;
2645         int i;
2646         int level;
2647         int ret = 0;
2648         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2649                             u64, u64, u64, u64, u64, u64);
2650
2651         ref_root = btrfs_header_owner(buf);
2652         nritems = btrfs_header_nritems(buf);
2653         level = btrfs_header_level(buf);
2654
2655         if (!root->ref_cows && level == 0)
2656                 return 0;
2657
2658         if (inc)
2659                 process_func = btrfs_inc_extent_ref;
2660         else
2661                 process_func = btrfs_free_extent;
2662
2663         if (full_backref)
2664                 parent = buf->start;
2665         else
2666                 parent = 0;
2667
2668         for (i = 0; i < nritems; i++) {
2669                 if (level == 0) {
2670                         btrfs_item_key_to_cpu(buf, &key, i);
2671                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2672                                 continue;
2673                         fi = btrfs_item_ptr(buf, i,
2674                                             struct btrfs_file_extent_item);
2675                         if (btrfs_file_extent_type(buf, fi) ==
2676                             BTRFS_FILE_EXTENT_INLINE)
2677                                 continue;
2678                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2679                         if (bytenr == 0)
2680                                 continue;
2681
2682                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2683                         key.offset -= btrfs_file_extent_offset(buf, fi);
2684                         ret = process_func(trans, root, bytenr, num_bytes,
2685                                            parent, ref_root, key.objectid,
2686                                            key.offset);
2687                         if (ret)
2688                                 goto fail;
2689                 } else {
2690                         bytenr = btrfs_node_blockptr(buf, i);
2691                         num_bytes = btrfs_level_size(root, level - 1);
2692                         ret = process_func(trans, root, bytenr, num_bytes,
2693                                            parent, ref_root, level - 1, 0);
2694                         if (ret)
2695                                 goto fail;
2696                 }
2697         }
2698         return 0;
2699 fail:
2700         BUG();
2701         return ret;
2702 }
2703
2704 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2705                   struct extent_buffer *buf, int full_backref)
2706 {
2707         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2708 }
2709
2710 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2711                   struct extent_buffer *buf, int full_backref)
2712 {
2713         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2714 }
2715
2716 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2717                                  struct btrfs_root *root,
2718                                  struct btrfs_path *path,
2719                                  struct btrfs_block_group_cache *cache)
2720 {
2721         int ret;
2722         struct btrfs_root *extent_root = root->fs_info->extent_root;
2723         unsigned long bi;
2724         struct extent_buffer *leaf;
2725
2726         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2727         if (ret < 0)
2728                 goto fail;
2729         BUG_ON(ret);
2730
2731         leaf = path->nodes[0];
2732         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2733         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2734         btrfs_mark_buffer_dirty(leaf);
2735         btrfs_release_path(extent_root, path);
2736 fail:
2737         if (ret)
2738                 return ret;
2739         return 0;
2740
2741 }
2742
2743 static struct btrfs_block_group_cache *
2744 next_block_group(struct btrfs_root *root,
2745                  struct btrfs_block_group_cache *cache)
2746 {
2747         struct rb_node *node;
2748         spin_lock(&root->fs_info->block_group_cache_lock);
2749         node = rb_next(&cache->cache_node);
2750         btrfs_put_block_group(cache);
2751         if (node) {
2752                 cache = rb_entry(node, struct btrfs_block_group_cache,
2753                                  cache_node);
2754                 btrfs_get_block_group(cache);
2755         } else
2756                 cache = NULL;
2757         spin_unlock(&root->fs_info->block_group_cache_lock);
2758         return cache;
2759 }
2760
2761 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2762                             struct btrfs_trans_handle *trans,
2763                             struct btrfs_path *path)
2764 {
2765         struct btrfs_root *root = block_group->fs_info->tree_root;
2766         struct inode *inode = NULL;
2767         u64 alloc_hint = 0;
2768         int dcs = BTRFS_DC_ERROR;
2769         int num_pages = 0;
2770         int retries = 0;
2771         int ret = 0;
2772
2773         /*
2774          * If this block group is smaller than 100 megs don't bother caching the
2775          * block group.
2776          */
2777         if (block_group->key.offset < (100 * 1024 * 1024)) {
2778                 spin_lock(&block_group->lock);
2779                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2780                 spin_unlock(&block_group->lock);
2781                 return 0;
2782         }
2783
2784 again:
2785         inode = lookup_free_space_inode(root, block_group, path);
2786         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2787                 ret = PTR_ERR(inode);
2788                 btrfs_release_path(root, path);
2789                 goto out;
2790         }
2791
2792         if (IS_ERR(inode)) {
2793                 BUG_ON(retries);
2794                 retries++;
2795
2796                 if (block_group->ro)
2797                         goto out_free;
2798
2799                 ret = create_free_space_inode(root, trans, block_group, path);
2800                 if (ret)
2801                         goto out_free;
2802                 goto again;
2803         }
2804
2805         /*
2806          * We want to set the generation to 0, that way if anything goes wrong
2807          * from here on out we know not to trust this cache when we load up next
2808          * time.
2809          */
2810         BTRFS_I(inode)->generation = 0;
2811         ret = btrfs_update_inode(trans, root, inode);
2812         WARN_ON(ret);
2813
2814         if (i_size_read(inode) > 0) {
2815                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2816                                                       inode);
2817                 if (ret)
2818                         goto out_put;
2819         }
2820
2821         spin_lock(&block_group->lock);
2822         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2823                 /* We're not cached, don't bother trying to write stuff out */
2824                 dcs = BTRFS_DC_WRITTEN;
2825                 spin_unlock(&block_group->lock);
2826                 goto out_put;
2827         }
2828         spin_unlock(&block_group->lock);
2829
2830         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2831         if (!num_pages)
2832                 num_pages = 1;
2833
2834         /*
2835          * Just to make absolutely sure we have enough space, we're going to
2836          * preallocate 12 pages worth of space for each block group.  In
2837          * practice we ought to use at most 8, but we need extra space so we can
2838          * add our header and have a terminator between the extents and the
2839          * bitmaps.
2840          */
2841         num_pages *= 16;
2842         num_pages *= PAGE_CACHE_SIZE;
2843
2844         ret = btrfs_check_data_free_space(inode, num_pages);
2845         if (ret)
2846                 goto out_put;
2847
2848         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2849                                               num_pages, num_pages,
2850                                               &alloc_hint);
2851         if (!ret)
2852                 dcs = BTRFS_DC_SETUP;
2853         btrfs_free_reserved_data_space(inode, num_pages);
2854 out_put:
2855         iput(inode);
2856 out_free:
2857         btrfs_release_path(root, path);
2858 out:
2859         spin_lock(&block_group->lock);
2860         block_group->disk_cache_state = dcs;
2861         spin_unlock(&block_group->lock);
2862
2863         return ret;
2864 }
2865
2866 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2867                                    struct btrfs_root *root)
2868 {
2869         struct btrfs_block_group_cache *cache;
2870         int err = 0;
2871         struct btrfs_path *path;
2872         u64 last = 0;
2873
2874         path = btrfs_alloc_path();
2875         if (!path)
2876                 return -ENOMEM;
2877
2878 again:
2879         while (1) {
2880                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2881                 while (cache) {
2882                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2883                                 break;
2884                         cache = next_block_group(root, cache);
2885                 }
2886                 if (!cache) {
2887                         if (last == 0)
2888                                 break;
2889                         last = 0;
2890                         continue;
2891                 }
2892                 err = cache_save_setup(cache, trans, path);
2893                 last = cache->key.objectid + cache->key.offset;
2894                 btrfs_put_block_group(cache);
2895         }
2896
2897         while (1) {
2898                 if (last == 0) {
2899                         err = btrfs_run_delayed_refs(trans, root,
2900                                                      (unsigned long)-1);
2901                         BUG_ON(err);
2902                 }
2903
2904                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2905                 while (cache) {
2906                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2907                                 btrfs_put_block_group(cache);
2908                                 goto again;
2909                         }
2910
2911                         if (cache->dirty)
2912                                 break;
2913                         cache = next_block_group(root, cache);
2914                 }
2915                 if (!cache) {
2916                         if (last == 0)
2917                                 break;
2918                         last = 0;
2919                         continue;
2920                 }
2921
2922                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2923                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2924                 cache->dirty = 0;
2925                 last = cache->key.objectid + cache->key.offset;
2926
2927                 err = write_one_cache_group(trans, root, path, cache);
2928                 BUG_ON(err);
2929                 btrfs_put_block_group(cache);
2930         }
2931
2932         while (1) {
2933                 /*
2934                  * I don't think this is needed since we're just marking our
2935                  * preallocated extent as written, but just in case it can't
2936                  * hurt.
2937                  */
2938                 if (last == 0) {
2939                         err = btrfs_run_delayed_refs(trans, root,
2940                                                      (unsigned long)-1);
2941                         BUG_ON(err);
2942                 }
2943
2944                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2945                 while (cache) {
2946                         /*
2947                          * Really this shouldn't happen, but it could if we
2948                          * couldn't write the entire preallocated extent and
2949                          * splitting the extent resulted in a new block.
2950                          */
2951                         if (cache->dirty) {
2952                                 btrfs_put_block_group(cache);
2953                                 goto again;
2954                         }
2955                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2956                                 break;
2957                         cache = next_block_group(root, cache);
2958                 }
2959                 if (!cache) {
2960                         if (last == 0)
2961                                 break;
2962                         last = 0;
2963                         continue;
2964                 }
2965
2966                 btrfs_write_out_cache(root, trans, cache, path);
2967
2968                 /*
2969                  * If we didn't have an error then the cache state is still
2970                  * NEED_WRITE, so we can set it to WRITTEN.
2971                  */
2972                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2973                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2974                 last = cache->key.objectid + cache->key.offset;
2975                 btrfs_put_block_group(cache);
2976         }
2977
2978         btrfs_free_path(path);
2979         return 0;
2980 }
2981
2982 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2983 {
2984         struct btrfs_block_group_cache *block_group;
2985         int readonly = 0;
2986
2987         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2988         if (!block_group || block_group->ro)
2989                 readonly = 1;
2990         if (block_group)
2991                 btrfs_put_block_group(block_group);
2992         return readonly;
2993 }
2994
2995 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2996                              u64 total_bytes, u64 bytes_used,
2997                              struct btrfs_space_info **space_info)
2998 {
2999         struct btrfs_space_info *found;
3000         int i;
3001         int factor;
3002
3003         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3004                      BTRFS_BLOCK_GROUP_RAID10))
3005                 factor = 2;
3006         else
3007                 factor = 1;
3008
3009         found = __find_space_info(info, flags);
3010         if (found) {
3011                 spin_lock(&found->lock);
3012                 found->total_bytes += total_bytes;
3013                 found->disk_total += total_bytes * factor;
3014                 found->bytes_used += bytes_used;
3015                 found->disk_used += bytes_used * factor;
3016                 found->full = 0;
3017                 spin_unlock(&found->lock);
3018                 *space_info = found;
3019                 return 0;
3020         }
3021         found = kzalloc(sizeof(*found), GFP_NOFS);
3022         if (!found)
3023                 return -ENOMEM;
3024
3025         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3026                 INIT_LIST_HEAD(&found->block_groups[i]);
3027         init_rwsem(&found->groups_sem);
3028         spin_lock_init(&found->lock);
3029         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
3030                                 BTRFS_BLOCK_GROUP_SYSTEM |
3031                                 BTRFS_BLOCK_GROUP_METADATA);
3032         found->total_bytes = total_bytes;
3033         found->disk_total = total_bytes * factor;
3034         found->bytes_used = bytes_used;
3035         found->disk_used = bytes_used * factor;
3036         found->bytes_pinned = 0;
3037         found->bytes_reserved = 0;
3038         found->bytes_readonly = 0;
3039         found->bytes_may_use = 0;
3040         found->full = 0;
3041         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3042         found->chunk_alloc = 0;
3043         *space_info = found;
3044         list_add_rcu(&found->list, &info->space_info);
3045         atomic_set(&found->caching_threads, 0);
3046         return 0;
3047 }
3048
3049 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3050 {
3051         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
3052                                    BTRFS_BLOCK_GROUP_RAID1 |
3053                                    BTRFS_BLOCK_GROUP_RAID10 |
3054                                    BTRFS_BLOCK_GROUP_DUP);
3055         if (extra_flags) {
3056                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3057                         fs_info->avail_data_alloc_bits |= extra_flags;
3058                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
3059                         fs_info->avail_metadata_alloc_bits |= extra_flags;
3060                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3061                         fs_info->avail_system_alloc_bits |= extra_flags;
3062         }
3063 }
3064
3065 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3066 {
3067         /*
3068          * we add in the count of missing devices because we want
3069          * to make sure that any RAID levels on a degraded FS
3070          * continue to be honored.
3071          */
3072         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3073                 root->fs_info->fs_devices->missing_devices;
3074
3075         if (num_devices == 1)
3076                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3077         if (num_devices < 4)
3078                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3079
3080         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3081             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3082                       BTRFS_BLOCK_GROUP_RAID10))) {
3083                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3084         }
3085
3086         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3087             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3088                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3089         }
3090
3091         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3092             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3093              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3094              (flags & BTRFS_BLOCK_GROUP_DUP)))
3095                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3096         return flags;
3097 }
3098
3099 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3100 {
3101         if (flags & BTRFS_BLOCK_GROUP_DATA)
3102                 flags |= root->fs_info->avail_data_alloc_bits &
3103                          root->fs_info->data_alloc_profile;
3104         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3105                 flags |= root->fs_info->avail_system_alloc_bits &
3106                          root->fs_info->system_alloc_profile;
3107         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3108                 flags |= root->fs_info->avail_metadata_alloc_bits &
3109                          root->fs_info->metadata_alloc_profile;
3110         return btrfs_reduce_alloc_profile(root, flags);
3111 }
3112
3113 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3114 {
3115         u64 flags;
3116
3117         if (data)
3118                 flags = BTRFS_BLOCK_GROUP_DATA;
3119         else if (root == root->fs_info->chunk_root)
3120                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3121         else
3122                 flags = BTRFS_BLOCK_GROUP_METADATA;
3123
3124         return get_alloc_profile(root, flags);
3125 }
3126
3127 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3128 {
3129         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3130                                                        BTRFS_BLOCK_GROUP_DATA);
3131 }
3132
3133 /*
3134  * This will check the space that the inode allocates from to make sure we have
3135  * enough space for bytes.
3136  */
3137 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3138 {
3139         struct btrfs_space_info *data_sinfo;
3140         struct btrfs_root *root = BTRFS_I(inode)->root;
3141         u64 used;
3142         int ret = 0, committed = 0, alloc_chunk = 1;
3143
3144         /* make sure bytes are sectorsize aligned */
3145         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3146
3147         if (root == root->fs_info->tree_root) {
3148                 alloc_chunk = 0;
3149                 committed = 1;
3150         }
3151
3152         data_sinfo = BTRFS_I(inode)->space_info;
3153         if (!data_sinfo)
3154                 goto alloc;
3155
3156 again:
3157         /* make sure we have enough space to handle the data first */
3158         spin_lock(&data_sinfo->lock);
3159         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3160                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3161                 data_sinfo->bytes_may_use;
3162
3163         if (used + bytes > data_sinfo->total_bytes) {
3164                 struct btrfs_trans_handle *trans;
3165
3166                 /*
3167                  * if we don't have enough free bytes in this space then we need
3168                  * to alloc a new chunk.
3169                  */
3170                 if (!data_sinfo->full && alloc_chunk) {
3171                         u64 alloc_target;
3172
3173                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3174                         spin_unlock(&data_sinfo->lock);
3175 alloc:
3176                         alloc_target = btrfs_get_alloc_profile(root, 1);
3177                         trans = btrfs_join_transaction(root, 1);
3178                         if (IS_ERR(trans))
3179                                 return PTR_ERR(trans);
3180
3181                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3182                                              bytes + 2 * 1024 * 1024,
3183                                              alloc_target,
3184                                              CHUNK_ALLOC_NO_FORCE);
3185                         btrfs_end_transaction(trans, root);
3186                         if (ret < 0) {
3187                                 if (ret != -ENOSPC)
3188                                         return ret;
3189                                 else
3190                                         goto commit_trans;
3191                         }
3192
3193                         if (!data_sinfo) {
3194                                 btrfs_set_inode_space_info(root, inode);
3195                                 data_sinfo = BTRFS_I(inode)->space_info;
3196                         }
3197                         goto again;
3198                 }
3199                 spin_unlock(&data_sinfo->lock);
3200
3201                 /* commit the current transaction and try again */
3202 commit_trans:
3203                 if (!committed && !root->fs_info->open_ioctl_trans) {
3204                         committed = 1;
3205                         trans = btrfs_join_transaction(root, 1);
3206                         if (IS_ERR(trans))
3207                                 return PTR_ERR(trans);
3208                         ret = btrfs_commit_transaction(trans, root);
3209                         if (ret)
3210                                 return ret;
3211                         goto again;
3212                 }
3213
3214 #if 0 /* I hope we never need this code again, just in case */
3215                 printk(KERN_ERR "no space left, need %llu, %llu bytes_used, "
3216                        "%llu bytes_reserved, " "%llu bytes_pinned, "
3217                        "%llu bytes_readonly, %llu may use %llu total\n",
3218                        (unsigned long long)bytes,
3219                        (unsigned long long)data_sinfo->bytes_used,
3220                        (unsigned long long)data_sinfo->bytes_reserved,
3221                        (unsigned long long)data_sinfo->bytes_pinned,
3222                        (unsigned long long)data_sinfo->bytes_readonly,
3223                        (unsigned long long)data_sinfo->bytes_may_use,
3224                        (unsigned long long)data_sinfo->total_bytes);
3225 #endif
3226                 return -ENOSPC;
3227         }
3228         data_sinfo->bytes_may_use += bytes;
3229         BTRFS_I(inode)->reserved_bytes += bytes;
3230         spin_unlock(&data_sinfo->lock);
3231
3232         return 0;
3233 }
3234
3235 /*
3236  * called when we are clearing an delalloc extent from the
3237  * inode's io_tree or there was an error for whatever reason
3238  * after calling btrfs_check_data_free_space
3239  */
3240 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3241 {
3242         struct btrfs_root *root = BTRFS_I(inode)->root;
3243         struct btrfs_space_info *data_sinfo;
3244
3245         /* make sure bytes are sectorsize aligned */
3246         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3247
3248         data_sinfo = BTRFS_I(inode)->space_info;
3249         spin_lock(&data_sinfo->lock);
3250         data_sinfo->bytes_may_use -= bytes;
3251         BTRFS_I(inode)->reserved_bytes -= bytes;
3252         spin_unlock(&data_sinfo->lock);
3253 }
3254
3255 static void force_metadata_allocation(struct btrfs_fs_info *info)
3256 {
3257         struct list_head *head = &info->space_info;
3258         struct btrfs_space_info *found;
3259
3260         rcu_read_lock();
3261         list_for_each_entry_rcu(found, head, list) {
3262                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3263                         found->force_alloc = CHUNK_ALLOC_FORCE;
3264         }
3265         rcu_read_unlock();
3266 }
3267
3268 static int should_alloc_chunk(struct btrfs_root *root,
3269                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3270                               int force)
3271 {
3272         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3273         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3274         u64 thresh;
3275
3276         if (force == CHUNK_ALLOC_FORCE)
3277                 return 1;
3278
3279         /*
3280          * in limited mode, we want to have some free space up to
3281          * about 1% of the FS size.
3282          */
3283         if (force == CHUNK_ALLOC_LIMITED) {
3284                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3285                 thresh = max_t(u64, 64 * 1024 * 1024,
3286                                div_factor_fine(thresh, 1));
3287
3288                 if (num_bytes - num_allocated < thresh)
3289                         return 1;
3290         }
3291
3292         /*
3293          * we have two similar checks here, one based on percentage
3294          * and once based on a hard number of 256MB.  The idea
3295          * is that if we have a good amount of free
3296          * room, don't allocate a chunk.  A good mount is
3297          * less than 80% utilized of the chunks we have allocated,
3298          * or more than 256MB free
3299          */
3300         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3301                 return 0;
3302
3303         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3304                 return 0;
3305
3306         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3307
3308         /* 256MB or 5% of the FS */
3309         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3310
3311         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3312                 return 0;
3313         return 1;
3314 }
3315
3316 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3317                           struct btrfs_root *extent_root, u64 alloc_bytes,
3318                           u64 flags, int force)
3319 {
3320         struct btrfs_space_info *space_info;
3321         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3322         int wait_for_alloc = 0;
3323         int ret = 0;
3324
3325         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3326
3327         space_info = __find_space_info(extent_root->fs_info, flags);
3328         if (!space_info) {
3329                 ret = update_space_info(extent_root->fs_info, flags,
3330                                         0, 0, &space_info);
3331                 BUG_ON(ret);
3332         }
3333         BUG_ON(!space_info);
3334
3335 again:
3336         spin_lock(&space_info->lock);
3337         if (space_info->force_alloc)
3338                 force = space_info->force_alloc;
3339         if (space_info->full) {
3340                 spin_unlock(&space_info->lock);
3341                 return 0;
3342         }
3343
3344         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3345                 spin_unlock(&space_info->lock);
3346                 return 0;
3347         } else if (space_info->chunk_alloc) {
3348                 wait_for_alloc = 1;
3349         } else {
3350                 space_info->chunk_alloc = 1;
3351         }
3352
3353         spin_unlock(&space_info->lock);
3354
3355         mutex_lock(&fs_info->chunk_mutex);
3356
3357         /*
3358          * The chunk_mutex is held throughout the entirety of a chunk
3359          * allocation, so once we've acquired the chunk_mutex we know that the
3360          * other guy is done and we need to recheck and see if we should
3361          * allocate.
3362          */
3363         if (wait_for_alloc) {
3364                 mutex_unlock(&fs_info->chunk_mutex);
3365                 wait_for_alloc = 0;
3366                 goto again;
3367         }
3368
3369         /*
3370          * If we have mixed data/metadata chunks we want to make sure we keep
3371          * allocating mixed chunks instead of individual chunks.
3372          */
3373         if (btrfs_mixed_space_info(space_info))
3374                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3375
3376         /*
3377          * if we're doing a data chunk, go ahead and make sure that
3378          * we keep a reasonable number of metadata chunks allocated in the
3379          * FS as well.
3380          */
3381         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3382                 fs_info->data_chunk_allocations++;
3383                 if (!(fs_info->data_chunk_allocations %
3384                       fs_info->metadata_ratio))
3385                         force_metadata_allocation(fs_info);
3386         }
3387
3388         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3389         spin_lock(&space_info->lock);
3390         if (ret)
3391                 space_info->full = 1;
3392         else
3393                 ret = 1;
3394
3395         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3396         space_info->chunk_alloc = 0;
3397         spin_unlock(&space_info->lock);
3398         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3399         return ret;
3400 }
3401
3402 /*
3403  * shrink metadata reservation for delalloc
3404  */
3405 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3406                            struct btrfs_root *root, u64 to_reclaim, int sync)
3407 {
3408         struct btrfs_block_rsv *block_rsv;
3409         struct btrfs_space_info *space_info;
3410         u64 reserved;
3411         u64 max_reclaim;
3412         u64 reclaimed = 0;
3413         long time_left;
3414         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3415         int loops = 0;
3416         unsigned long progress;
3417
3418         block_rsv = &root->fs_info->delalloc_block_rsv;
3419         space_info = block_rsv->space_info;
3420
3421         smp_mb();
3422         reserved = space_info->bytes_reserved;
3423         progress = space_info->reservation_progress;
3424
3425         if (reserved == 0)
3426                 return 0;
3427
3428         max_reclaim = min(reserved, to_reclaim);
3429
3430         while (loops < 1024) {
3431                 /* have the flusher threads jump in and do some IO */
3432                 smp_mb();
3433                 nr_pages = min_t(unsigned long, nr_pages,
3434                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3435                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3436
3437                 spin_lock(&space_info->lock);
3438                 if (reserved > space_info->bytes_reserved)
3439                         reclaimed += reserved - space_info->bytes_reserved;
3440                 reserved = space_info->bytes_reserved;
3441                 spin_unlock(&space_info->lock);
3442
3443                 loops++;
3444
3445                 if (reserved == 0 || reclaimed >= max_reclaim)
3446                         break;
3447
3448                 if (trans && trans->transaction->blocked)
3449                         return -EAGAIN;
3450
3451                 time_left = schedule_timeout_interruptible(1);
3452
3453                 /* We were interrupted, exit */
3454                 if (time_left)
3455                         break;
3456
3457                 /* we've kicked the IO a few times, if anything has been freed,
3458                  * exit.  There is no sense in looping here for a long time
3459                  * when we really need to commit the transaction, or there are
3460                  * just too many writers without enough free space
3461                  */
3462
3463                 if (loops > 3) {
3464                         smp_mb();
3465                         if (progress != space_info->reservation_progress)
3466                                 break;
3467                 }
3468
3469         }
3470         return reclaimed >= to_reclaim;
3471 }
3472
3473 /*
3474  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3475  * idea is if this is the first call (retries == 0) then we will add to our
3476  * reserved count if we can't make the allocation in order to hold our place
3477  * while we go and try and free up space.  That way for retries > 1 we don't try
3478  * and add space, we just check to see if the amount of unused space is >= the
3479  * total space, meaning that our reservation is valid.
3480  *
3481  * However if we don't intend to retry this reservation, pass -1 as retries so
3482  * that it short circuits this logic.
3483  */
3484 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3485                                   struct btrfs_root *root,
3486                                   struct btrfs_block_rsv *block_rsv,
3487                                   u64 orig_bytes, int flush)
3488 {
3489         struct btrfs_space_info *space_info = block_rsv->space_info;
3490         u64 unused;
3491         u64 num_bytes = orig_bytes;
3492         int retries = 0;
3493         int ret = 0;
3494         bool reserved = false;
3495         bool committed = false;
3496
3497 again:
3498         ret = -ENOSPC;
3499         if (reserved)
3500                 num_bytes = 0;
3501
3502         spin_lock(&space_info->lock);
3503         unused = space_info->bytes_used + space_info->bytes_reserved +
3504                  space_info->bytes_pinned + space_info->bytes_readonly +
3505                  space_info->bytes_may_use;
3506
3507         /*
3508          * The idea here is that we've not already over-reserved the block group
3509          * then we can go ahead and save our reservation first and then start
3510          * flushing if we need to.  Otherwise if we've already overcommitted
3511          * lets start flushing stuff first and then come back and try to make
3512          * our reservation.
3513          */
3514         if (unused <= space_info->total_bytes) {
3515                 unused = space_info->total_bytes - unused;
3516                 if (unused >= num_bytes) {
3517                         if (!reserved)
3518                                 space_info->bytes_reserved += orig_bytes;
3519                         ret = 0;
3520                 } else {
3521                         /*
3522                          * Ok set num_bytes to orig_bytes since we aren't
3523                          * overocmmitted, this way we only try and reclaim what
3524                          * we need.
3525                          */
3526                         num_bytes = orig_bytes;
3527                 }
3528         } else {
3529                 /*
3530                  * Ok we're over committed, set num_bytes to the overcommitted
3531                  * amount plus the amount of bytes that we need for this
3532                  * reservation.
3533                  */
3534                 num_bytes = unused - space_info->total_bytes +
3535                         (orig_bytes * (retries + 1));
3536         }
3537
3538         /*
3539          * Couldn't make our reservation, save our place so while we're trying
3540          * to reclaim space we can actually use it instead of somebody else
3541          * stealing it from us.
3542          */
3543         if (ret && !reserved) {
3544                 space_info->bytes_reserved += orig_bytes;
3545                 reserved = true;
3546         }
3547
3548         spin_unlock(&space_info->lock);
3549
3550         if (!ret)
3551                 return 0;
3552
3553         if (!flush)
3554                 goto out;
3555
3556         /*
3557          * We do synchronous shrinking since we don't actually unreserve
3558          * metadata until after the IO is completed.
3559          */
3560         ret = shrink_delalloc(trans, root, num_bytes, 1);
3561         if (ret > 0)
3562                 return 0;
3563         else if (ret < 0)
3564                 goto out;
3565
3566         /*
3567          * So if we were overcommitted it's possible that somebody else flushed
3568          * out enough space and we simply didn't have enough space to reclaim,
3569          * so go back around and try again.
3570          */
3571         if (retries < 2) {
3572                 retries++;
3573                 goto again;
3574         }
3575
3576         spin_lock(&space_info->lock);
3577         /*
3578          * Not enough space to be reclaimed, don't bother committing the
3579          * transaction.
3580          */
3581         if (space_info->bytes_pinned < orig_bytes)
3582                 ret = -ENOSPC;
3583         spin_unlock(&space_info->lock);
3584         if (ret)
3585                 goto out;
3586
3587         ret = -EAGAIN;
3588         if (trans || committed)
3589                 goto out;
3590
3591         ret = -ENOSPC;
3592         trans = btrfs_join_transaction(root, 1);
3593         if (IS_ERR(trans))
3594                 goto out;
3595         ret = btrfs_commit_transaction(trans, root);
3596         if (!ret) {
3597                 trans = NULL;
3598                 committed = true;
3599                 goto again;
3600         }
3601
3602 out:
3603         if (reserved) {
3604                 spin_lock(&space_info->lock);
3605                 space_info->bytes_reserved -= orig_bytes;
3606                 spin_unlock(&space_info->lock);
3607         }
3608
3609         return ret;
3610 }
3611
3612 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3613                                              struct btrfs_root *root)
3614 {
3615         struct btrfs_block_rsv *block_rsv;
3616         if (root->ref_cows)
3617                 block_rsv = trans->block_rsv;
3618         else
3619                 block_rsv = root->block_rsv;
3620
3621         if (!block_rsv)
3622                 block_rsv = &root->fs_info->empty_block_rsv;
3623
3624         return block_rsv;
3625 }
3626
3627 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3628                                u64 num_bytes)
3629 {
3630         int ret = -ENOSPC;
3631         spin_lock(&block_rsv->lock);
3632         if (block_rsv->reserved >= num_bytes) {
3633                 block_rsv->reserved -= num_bytes;
3634                 if (block_rsv->reserved < block_rsv->size)
3635                         block_rsv->full = 0;
3636                 ret = 0;
3637         }
3638         spin_unlock(&block_rsv->lock);
3639         return ret;
3640 }
3641
3642 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3643                                 u64 num_bytes, int update_size)
3644 {
3645         spin_lock(&block_rsv->lock);
3646         block_rsv->reserved += num_bytes;
3647         if (update_size)
3648                 block_rsv->size += num_bytes;
3649         else if (block_rsv->reserved >= block_rsv->size)
3650                 block_rsv->full = 1;
3651         spin_unlock(&block_rsv->lock);
3652 }
3653
3654 void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3655                              struct btrfs_block_rsv *dest, u64 num_bytes)
3656 {
3657         struct btrfs_space_info *space_info = block_rsv->space_info;
3658
3659         spin_lock(&block_rsv->lock);
3660         if (num_bytes == (u64)-1)
3661                 num_bytes = block_rsv->size;
3662         block_rsv->size -= num_bytes;
3663         if (block_rsv->reserved >= block_rsv->size) {
3664                 num_bytes = block_rsv->reserved - block_rsv->size;
3665                 block_rsv->reserved = block_rsv->size;
3666                 block_rsv->full = 1;
3667         } else {
3668                 num_bytes = 0;
3669         }
3670         spin_unlock(&block_rsv->lock);
3671
3672         if (num_bytes > 0) {
3673                 if (dest) {
3674                         spin_lock(&dest->lock);
3675                         if (!dest->full) {
3676                                 u64 bytes_to_add;
3677
3678                                 bytes_to_add = dest->size - dest->reserved;
3679                                 bytes_to_add = min(num_bytes, bytes_to_add);
3680                                 dest->reserved += bytes_to_add;
3681                                 if (dest->reserved >= dest->size)
3682                                         dest->full = 1;
3683                                 num_bytes -= bytes_to_add;
3684                         }
3685                         spin_unlock(&dest->lock);
3686                 }
3687                 if (num_bytes) {
3688                         spin_lock(&space_info->lock);
3689                         space_info->bytes_reserved -= num_bytes;
3690                         space_info->reservation_progress++;
3691                         spin_unlock(&space_info->lock);
3692                 }
3693         }
3694 }
3695
3696 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3697                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3698 {
3699         int ret;
3700
3701         ret = block_rsv_use_bytes(src, num_bytes);
3702         if (ret)
3703                 return ret;
3704
3705         block_rsv_add_bytes(dst, num_bytes, 1);
3706         return 0;
3707 }
3708
3709 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3710 {
3711         memset(rsv, 0, sizeof(*rsv));
3712         spin_lock_init(&rsv->lock);
3713         atomic_set(&rsv->usage, 1);
3714         rsv->priority = 6;
3715         INIT_LIST_HEAD(&rsv->list);
3716 }
3717
3718 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3719 {
3720         struct btrfs_block_rsv *block_rsv;
3721         struct btrfs_fs_info *fs_info = root->fs_info;
3722
3723         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3724         if (!block_rsv)
3725                 return NULL;
3726
3727         btrfs_init_block_rsv(block_rsv);
3728         block_rsv->space_info = __find_space_info(fs_info,
3729                                                   BTRFS_BLOCK_GROUP_METADATA);
3730         return block_rsv;
3731 }
3732
3733 void btrfs_free_block_rsv(struct btrfs_root *root,
3734                           struct btrfs_block_rsv *rsv)
3735 {
3736         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3737                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3738                 if (!rsv->durable)
3739                         kfree(rsv);
3740         }
3741 }
3742
3743 /*
3744  * make the block_rsv struct be able to capture freed space.
3745  * the captured space will re-add to the the block_rsv struct
3746  * after transaction commit
3747  */
3748 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3749                                  struct btrfs_block_rsv *block_rsv)
3750 {
3751         block_rsv->durable = 1;
3752         mutex_lock(&fs_info->durable_block_rsv_mutex);
3753         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3754         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3755 }
3756
3757 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3758                         struct btrfs_root *root,
3759                         struct btrfs_block_rsv *block_rsv,
3760                         u64 num_bytes)
3761 {
3762         int ret;
3763
3764         if (num_bytes == 0)
3765                 return 0;
3766
3767         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3768         if (!ret) {
3769                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3770                 return 0;
3771         }
3772
3773         return ret;
3774 }
3775
3776 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3777                           struct btrfs_root *root,
3778                           struct btrfs_block_rsv *block_rsv,
3779                           u64 min_reserved, int min_factor)
3780 {
3781         u64 num_bytes = 0;
3782         int commit_trans = 0;
3783         int ret = -ENOSPC;
3784
3785         if (!block_rsv)
3786                 return 0;
3787
3788         spin_lock(&block_rsv->lock);
3789         if (min_factor > 0)
3790                 num_bytes = div_factor(block_rsv->size, min_factor);
3791         if (min_reserved > num_bytes)
3792                 num_bytes = min_reserved;
3793
3794         if (block_rsv->reserved >= num_bytes) {
3795                 ret = 0;
3796         } else {
3797                 num_bytes -= block_rsv->reserved;
3798                 if (block_rsv->durable &&
3799                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3800                         commit_trans = 1;
3801         }
3802         spin_unlock(&block_rsv->lock);
3803         if (!ret)
3804                 return 0;
3805
3806         if (block_rsv->refill_used) {
3807                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3808                                              num_bytes, 0);
3809                 if (!ret) {
3810                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3811                         return 0;
3812                 }
3813         }
3814
3815         if (commit_trans) {
3816                 if (trans)
3817                         return -EAGAIN;
3818
3819                 trans = btrfs_join_transaction(root, 1);
3820                 BUG_ON(IS_ERR(trans));
3821                 ret = btrfs_commit_transaction(trans, root);
3822                 return 0;
3823         }
3824
3825         return -ENOSPC;
3826 }
3827
3828 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3829                             struct btrfs_block_rsv *dst_rsv,
3830                             u64 num_bytes)
3831 {
3832         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3833 }
3834
3835 void btrfs_block_rsv_release(struct btrfs_root *root,
3836                              struct btrfs_block_rsv *block_rsv,
3837                              u64 num_bytes)
3838 {
3839         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3840         if (global_rsv->full || global_rsv == block_rsv ||
3841             block_rsv->space_info != global_rsv->space_info)
3842                 global_rsv = NULL;
3843         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3844 }
3845
3846 /*
3847  * helper to calculate size of global block reservation.
3848  * the desired value is sum of space used by extent tree,
3849  * checksum tree and root tree
3850  */
3851 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3852 {
3853         struct btrfs_space_info *sinfo;
3854         u64 num_bytes;
3855         u64 meta_used;
3856         u64 data_used;
3857         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3858 #if 0
3859         /*
3860          * per tree used space accounting can be inaccuracy, so we
3861          * can't rely on it.
3862          */
3863         spin_lock(&fs_info->extent_root->accounting_lock);
3864         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item);
3865         spin_unlock(&fs_info->extent_root->accounting_lock);
3866
3867         spin_lock(&fs_info->csum_root->accounting_lock);
3868         num_bytes += btrfs_root_used(&fs_info->csum_root->root_item);
3869         spin_unlock(&fs_info->csum_root->accounting_lock);
3870
3871         spin_lock(&fs_info->tree_root->accounting_lock);
3872         num_bytes += btrfs_root_used(&fs_info->tree_root->root_item);
3873         spin_unlock(&fs_info->tree_root->accounting_lock);
3874 #endif
3875         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3876         spin_lock(&sinfo->lock);
3877         data_used = sinfo->bytes_used;
3878         spin_unlock(&sinfo->lock);
3879
3880         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3881         spin_lock(&sinfo->lock);
3882         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3883                 data_used = 0;
3884         meta_used = sinfo->bytes_used;
3885         spin_unlock(&sinfo->lock);
3886
3887         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3888                     csum_size * 2;
3889         num_bytes += div64_u64(data_used + meta_used, 50);
3890
3891         if (num_bytes * 3 > meta_used)
3892                 num_bytes = div64_u64(meta_used, 3);
3893
3894         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3895 }
3896
3897 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3898 {
3899         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3900         struct btrfs_space_info *sinfo = block_rsv->space_info;
3901         u64 num_bytes;
3902
3903         num_bytes = calc_global_metadata_size(fs_info);
3904
3905         spin_lock(&block_rsv->lock);
3906         spin_lock(&sinfo->lock);
3907
3908         block_rsv->size = num_bytes;
3909
3910         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3911                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3912                     sinfo->bytes_may_use;
3913
3914         if (sinfo->total_bytes > num_bytes) {
3915                 num_bytes = sinfo->total_bytes - num_bytes;
3916                 block_rsv->reserved += num_bytes;
3917                 sinfo->bytes_reserved += num_bytes;
3918         }
3919
3920         if (block_rsv->reserved >= block_rsv->size) {
3921                 num_bytes = block_rsv->reserved - block_rsv->size;
3922                 sinfo->bytes_reserved -= num_bytes;
3923                 sinfo->reservation_progress++;
3924                 block_rsv->reserved = block_rsv->size;
3925                 block_rsv->full = 1;
3926         }
3927 #if 0
3928         printk(KERN_INFO"global block rsv size %llu reserved %llu\n",
3929                 block_rsv->size, block_rsv->reserved);
3930 #endif
3931         spin_unlock(&sinfo->lock);
3932         spin_unlock(&block_rsv->lock);
3933 }
3934
3935 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3936 {
3937         struct btrfs_space_info *space_info;
3938
3939         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3940         fs_info->chunk_block_rsv.space_info = space_info;
3941         fs_info->chunk_block_rsv.priority = 10;
3942
3943         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3944         fs_info->global_block_rsv.space_info = space_info;
3945         fs_info->global_block_rsv.priority = 10;
3946         fs_info->global_block_rsv.refill_used = 1;
3947         fs_info->delalloc_block_rsv.space_info = space_info;
3948         fs_info->trans_block_rsv.space_info = space_info;
3949         fs_info->empty_block_rsv.space_info = space_info;
3950         fs_info->empty_block_rsv.priority = 10;
3951
3952         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3953         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3954         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3955         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3956         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3957
3958         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3959
3960         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3961
3962         update_global_block_rsv(fs_info);
3963 }
3964
3965 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3966 {
3967         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3968         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3969         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3970         WARN_ON(fs_info->trans_block_rsv.size > 0);
3971         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3972         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3973         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3974 }
3975
3976 static u64 calc_trans_metadata_size(struct btrfs_root *root, int num_items)
3977 {
3978         return (root->leafsize + root->nodesize * (BTRFS_MAX_LEVEL - 1)) *
3979                 3 * num_items;
3980 }
3981
3982 int btrfs_trans_reserve_metadata(struct btrfs_trans_handle *trans,
3983                                  struct btrfs_root *root,
3984                                  int num_items)
3985 {
3986         u64 num_bytes;
3987         int ret;
3988
3989         if (num_items == 0 || root->fs_info->chunk_root == root)
3990                 return 0;
3991
3992         num_bytes = calc_trans_metadata_size(root, num_items);
3993         ret = btrfs_block_rsv_add(trans, root, &root->fs_info->trans_block_rsv,
3994                                   num_bytes);
3995         if (!ret) {
3996                 trans->bytes_reserved += num_bytes;
3997                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3998         }
3999         return ret;
4000 }
4001
4002 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4003                                   struct btrfs_root *root)
4004 {
4005         if (!trans->bytes_reserved)
4006                 return;
4007
4008         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
4009         btrfs_block_rsv_release(root, trans->block_rsv,
4010                                 trans->bytes_reserved);
4011         trans->bytes_reserved = 0;
4012 }
4013
4014 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4015                                   struct inode *inode)
4016 {
4017         struct btrfs_root *root = BTRFS_I(inode)->root;
4018         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4019         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4020
4021         /*
4022          * one for deleting orphan item, one for updating inode and
4023          * two for calling btrfs_truncate_inode_items.
4024          *
4025          * btrfs_truncate_inode_items is a delete operation, it frees
4026          * more space than it uses in most cases. So two units of
4027          * metadata space should be enough for calling it many times.
4028          * If all of the metadata space is used, we can commit
4029          * transaction and use space it freed.
4030          */
4031         u64 num_bytes = calc_trans_metadata_size(root, 4);
4032         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4033 }
4034
4035 void btrfs_orphan_release_metadata(struct inode *inode)
4036 {
4037         struct btrfs_root *root = BTRFS_I(inode)->root;
4038         u64 num_bytes = calc_trans_metadata_size(root, 4);
4039         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4040 }
4041
4042 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4043                                 struct btrfs_pending_snapshot *pending)
4044 {
4045         struct btrfs_root *root = pending->root;
4046         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4047         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4048         /*
4049          * two for root back/forward refs, two for directory entries
4050          * and one for root of the snapshot.
4051          */
4052         u64 num_bytes = calc_trans_metadata_size(root, 5);
4053         dst_rsv->space_info = src_rsv->space_info;
4054         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4055 }
4056
4057 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
4058 {
4059         return num_bytes >>= 3;
4060 }
4061
4062 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4063 {
4064         struct btrfs_root *root = BTRFS_I(inode)->root;
4065         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4066         u64 to_reserve;
4067         int nr_extents;
4068         int reserved_extents;
4069         int ret;
4070
4071         if (btrfs_transaction_in_commit(root->fs_info))
4072                 schedule_timeout(1);
4073
4074         num_bytes = ALIGN(num_bytes, root->sectorsize);
4075
4076         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents) + 1;
4077         reserved_extents = atomic_read(&BTRFS_I(inode)->reserved_extents);
4078
4079         if (nr_extents > reserved_extents) {
4080                 nr_extents -= reserved_extents;
4081                 to_reserve = calc_trans_metadata_size(root, nr_extents);
4082         } else {
4083                 nr_extents = 0;
4084                 to_reserve = 0;
4085         }
4086
4087         to_reserve += calc_csum_metadata_size(inode, num_bytes);
4088         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
4089         if (ret)
4090                 return ret;
4091
4092         atomic_add(nr_extents, &BTRFS_I(inode)->reserved_extents);
4093         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
4094
4095         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4096
4097         if (block_rsv->size > 512 * 1024 * 1024)
4098                 shrink_delalloc(NULL, root, to_reserve, 0);
4099
4100         return 0;
4101 }
4102
4103 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4104 {
4105         struct btrfs_root *root = BTRFS_I(inode)->root;
4106         u64 to_free;
4107         int nr_extents;
4108         int reserved_extents;
4109
4110         num_bytes = ALIGN(num_bytes, root->sectorsize);
4111         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
4112         WARN_ON(atomic_read(&BTRFS_I(inode)->outstanding_extents) < 0);
4113
4114         reserved_extents = atomic_read(&BTRFS_I(inode)->reserved_extents);
4115         do {
4116                 int old, new;
4117
4118                 nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents);
4119                 if (nr_extents >= reserved_extents) {
4120                         nr_extents = 0;
4121                         break;
4122                 }
4123                 old = reserved_extents;
4124                 nr_extents = reserved_extents - nr_extents;
4125                 new = reserved_extents - nr_extents;
4126                 old = atomic_cmpxchg(&BTRFS_I(inode)->reserved_extents,
4127                                      reserved_extents, new);
4128                 if (likely(old == reserved_extents))
4129                         break;
4130                 reserved_extents = old;
4131         } while (1);
4132
4133         to_free = calc_csum_metadata_size(inode, num_bytes);
4134         if (nr_extents > 0)
4135                 to_free += calc_trans_metadata_size(root, nr_extents);
4136
4137         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4138                                 to_free);
4139 }
4140
4141 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4142 {
4143         int ret;
4144
4145         ret = btrfs_check_data_free_space(inode, num_bytes);
4146         if (ret)
4147                 return ret;
4148
4149         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4150         if (ret) {
4151                 btrfs_free_reserved_data_space(inode, num_bytes);
4152                 return ret;
4153         }
4154
4155         return 0;
4156 }
4157
4158 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4159 {
4160         btrfs_delalloc_release_metadata(inode, num_bytes);
4161         btrfs_free_reserved_data_space(inode, num_bytes);
4162 }
4163
4164 static int update_block_group(struct btrfs_trans_handle *trans,
4165                               struct btrfs_root *root,
4166                               u64 bytenr, u64 num_bytes, int alloc)
4167 {
4168         struct btrfs_block_group_cache *cache = NULL;
4169         struct btrfs_fs_info *info = root->fs_info;
4170         u64 total = num_bytes;
4171         u64 old_val;
4172         u64 byte_in_group;
4173         int factor;
4174
4175         /* block accounting for super block */
4176         spin_lock(&info->delalloc_lock);
4177         old_val = btrfs_super_bytes_used(&info->super_copy);
4178         if (alloc)
4179                 old_val += num_bytes;
4180         else
4181                 old_val -= num_bytes;
4182         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4183         spin_unlock(&info->delalloc_lock);
4184
4185         while (total) {
4186                 cache = btrfs_lookup_block_group(info, bytenr);
4187                 if (!cache)
4188                         return -1;
4189                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4190                                     BTRFS_BLOCK_GROUP_RAID1 |
4191                                     BTRFS_BLOCK_GROUP_RAID10))
4192                         factor = 2;
4193                 else
4194                         factor = 1;
4195                 /*
4196                  * If this block group has free space cache written out, we
4197                  * need to make sure to load it if we are removing space.  This
4198                  * is because we need the unpinning stage to actually add the
4199                  * space back to the block group, otherwise we will leak space.
4200                  */
4201                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4202                         cache_block_group(cache, trans, NULL, 1);
4203
4204                 byte_in_group = bytenr - cache->key.objectid;
4205                 WARN_ON(byte_in_group > cache->key.offset);
4206
4207                 spin_lock(&cache->space_info->lock);
4208                 spin_lock(&cache->lock);
4209
4210                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4211                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4212                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4213
4214                 cache->dirty = 1;
4215                 old_val = btrfs_block_group_used(&cache->item);
4216                 num_bytes = min(total, cache->key.offset - byte_in_group);
4217                 if (alloc) {
4218                         old_val += num_bytes;
4219                         btrfs_set_block_group_used(&cache->item, old_val);
4220                         cache->reserved -= num_bytes;
4221                         cache->space_info->bytes_reserved -= num_bytes;
4222                         cache->space_info->reservation_progress++;
4223                         cache->space_info->bytes_used += num_bytes;
4224                         cache->space_info->disk_used += num_bytes * factor;
4225                         spin_unlock(&cache->lock);
4226                         spin_unlock(&cache->space_info->lock);
4227                 } else {
4228                         old_val -= num_bytes;
4229                         btrfs_set_block_group_used(&cache->item, old_val);
4230                         cache->pinned += num_bytes;
4231                         cache->space_info->bytes_pinned += num_bytes;
4232                         cache->space_info->bytes_used -= num_bytes;
4233                         cache->space_info->disk_used -= num_bytes * factor;
4234                         spin_unlock(&cache->lock);
4235                         spin_unlock(&cache->space_info->lock);
4236
4237                         set_extent_dirty(info->pinned_extents,
4238                                          bytenr, bytenr + num_bytes - 1,
4239                                          GFP_NOFS | __GFP_NOFAIL);
4240                 }
4241                 btrfs_put_block_group(cache);
4242                 total -= num_bytes;
4243                 bytenr += num_bytes;
4244         }
4245         return 0;
4246 }
4247
4248 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4249 {
4250         struct btrfs_block_group_cache *cache;
4251         u64 bytenr;
4252
4253         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4254         if (!cache)
4255                 return 0;
4256
4257         bytenr = cache->key.objectid;
4258         btrfs_put_block_group(cache);
4259
4260         return bytenr;
4261 }
4262
4263 static int pin_down_extent(struct btrfs_root *root,
4264                            struct btrfs_block_group_cache *cache,
4265                            u64 bytenr, u64 num_bytes, int reserved)
4266 {
4267         spin_lock(&cache->space_info->lock);
4268         spin_lock(&cache->lock);
4269         cache->pinned += num_bytes;
4270         cache->space_info->bytes_pinned += num_bytes;
4271         if (reserved) {
4272                 cache->reserved -= num_bytes;
4273                 cache->space_info->bytes_reserved -= num_bytes;
4274                 cache->space_info->reservation_progress++;
4275         }
4276         spin_unlock(&cache->lock);
4277         spin_unlock(&cache->space_info->lock);
4278
4279         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4280                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4281         return 0;
4282 }
4283
4284 /*
4285  * this function must be called within transaction
4286  */
4287 int btrfs_pin_extent(struct btrfs_root *root,
4288                      u64 bytenr, u64 num_bytes, int reserved)
4289 {
4290         struct btrfs_block_group_cache *cache;
4291
4292         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4293         BUG_ON(!cache);
4294
4295         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4296
4297         btrfs_put_block_group(cache);
4298         return 0;
4299 }
4300
4301 /*
4302  * update size of reserved extents. this function may return -EAGAIN
4303  * if 'reserve' is true or 'sinfo' is false.
4304  */
4305 int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4306                                 u64 num_bytes, int reserve, int sinfo)
4307 {
4308         int ret = 0;
4309         if (sinfo) {
4310                 struct btrfs_space_info *space_info = cache->space_info;
4311                 spin_lock(&space_info->lock);
4312                 spin_lock(&cache->lock);
4313                 if (reserve) {
4314                         if (cache->ro) {
4315                                 ret = -EAGAIN;
4316                         } else {
4317                                 cache->reserved += num_bytes;
4318                                 space_info->bytes_reserved += num_bytes;
4319                         }
4320                 } else {
4321                         if (cache->ro)
4322                                 space_info->bytes_readonly += num_bytes;
4323                         cache->reserved -= num_bytes;
4324                         space_info->bytes_reserved -= num_bytes;
4325                         space_info->reservation_progress++;
4326                 }
4327                 spin_unlock(&cache->lock);
4328                 spin_unlock(&space_info->lock);
4329         } else {
4330                 spin_lock(&cache->lock);
4331                 if (cache->ro) {
4332                         ret = -EAGAIN;
4333                 } else {
4334                         if (reserve)
4335                                 cache->reserved += num_bytes;
4336                         else
4337                                 cache->reserved -= num_bytes;
4338                 }
4339                 spin_unlock(&cache->lock);
4340         }
4341         return ret;
4342 }
4343
4344 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4345                                 struct btrfs_root *root)
4346 {
4347         struct btrfs_fs_info *fs_info = root->fs_info;
4348         struct btrfs_caching_control *next;
4349         struct btrfs_caching_control *caching_ctl;
4350         struct btrfs_block_group_cache *cache;
4351
4352         down_write(&fs_info->extent_commit_sem);
4353
4354         list_for_each_entry_safe(caching_ctl, next,
4355                                  &fs_info->caching_block_groups, list) {
4356                 cache = caching_ctl->block_group;
4357                 if (block_group_cache_done(cache)) {
4358                         cache->last_byte_to_unpin = (u64)-1;
4359                         list_del_init(&caching_ctl->list);
4360                         put_caching_control(caching_ctl);
4361                 } else {
4362                         cache->last_byte_to_unpin = caching_ctl->progress;
4363                 }
4364         }
4365
4366         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4367                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4368         else
4369                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4370
4371         up_write(&fs_info->extent_commit_sem);
4372
4373         update_global_block_rsv(fs_info);
4374         return 0;
4375 }
4376
4377 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4378 {
4379         struct btrfs_fs_info *fs_info = root->fs_info;
4380         struct btrfs_block_group_cache *cache = NULL;
4381         u64 len;
4382
4383         while (start <= end) {
4384                 if (!cache ||
4385                     start >= cache->key.objectid + cache->key.offset) {
4386                         if (cache)
4387                                 btrfs_put_block_group(cache);
4388                         cache = btrfs_lookup_block_group(fs_info, start);
4389                         BUG_ON(!cache);
4390                 }
4391
4392                 len = cache->key.objectid + cache->key.offset - start;
4393                 len = min(len, end + 1 - start);
4394
4395                 if (start < cache->last_byte_to_unpin) {
4396                         len = min(len, cache->last_byte_to_unpin - start);
4397                         btrfs_add_free_space(cache, start, len);
4398                 }
4399
4400                 start += len;
4401
4402                 spin_lock(&cache->space_info->lock);
4403                 spin_lock(&cache->lock);
4404                 cache->pinned -= len;
4405                 cache->space_info->bytes_pinned -= len;
4406                 if (cache->ro) {
4407                         cache->space_info->bytes_readonly += len;
4408                 } else if (cache->reserved_pinned > 0) {
4409                         len = min(len, cache->reserved_pinned);
4410                         cache->reserved_pinned -= len;
4411                         cache->space_info->bytes_reserved += len;
4412                 }
4413                 spin_unlock(&cache->lock);
4414                 spin_unlock(&cache->space_info->lock);
4415         }
4416
4417         if (cache)
4418                 btrfs_put_block_group(cache);
4419         return 0;
4420 }
4421
4422 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4423                                struct btrfs_root *root)
4424 {
4425         struct btrfs_fs_info *fs_info = root->fs_info;
4426         struct extent_io_tree *unpin;
4427         struct btrfs_block_rsv *block_rsv;
4428         struct btrfs_block_rsv *next_rsv;
4429         u64 start;
4430         u64 end;
4431         int idx;
4432         int ret;
4433
4434         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4435                 unpin = &fs_info->freed_extents[1];
4436         else
4437                 unpin = &fs_info->freed_extents[0];
4438
4439         while (1) {
4440                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4441                                             EXTENT_DIRTY);
4442                 if (ret)
4443                         break;
4444
4445                 if (btrfs_test_opt(root, DISCARD))
4446                         ret = btrfs_discard_extent(root, start,
4447                                                    end + 1 - start, NULL);
4448
4449                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4450                 unpin_extent_range(root, start, end);
4451                 cond_resched();
4452         }
4453
4454         mutex_lock(&fs_info->durable_block_rsv_mutex);
4455         list_for_each_entry_safe(block_rsv, next_rsv,
4456                                  &fs_info->durable_block_rsv_list, list) {
4457
4458                 idx = trans->transid & 0x1;
4459                 if (block_rsv->freed[idx] > 0) {
4460                         block_rsv_add_bytes(block_rsv,
4461                                             block_rsv->freed[idx], 0);
4462                         block_rsv->freed[idx] = 0;
4463                 }
4464                 if (atomic_read(&block_rsv->usage) == 0) {
4465                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4466
4467                         if (block_rsv->freed[0] == 0 &&
4468                             block_rsv->freed[1] == 0) {
4469                                 list_del_init(&block_rsv->list);
4470                                 kfree(block_rsv);
4471                         }
4472                 } else {
4473                         btrfs_block_rsv_release(root, block_rsv, 0);
4474                 }
4475         }
4476         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4477
4478         return 0;
4479 }
4480
4481 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4482                                 struct btrfs_root *root,
4483                                 u64 bytenr, u64 num_bytes, u64 parent,
4484                                 u64 root_objectid, u64 owner_objectid,
4485                                 u64 owner_offset, int refs_to_drop,
4486                                 struct btrfs_delayed_extent_op *extent_op)
4487 {
4488         struct btrfs_key key;
4489         struct btrfs_path *path;
4490         struct btrfs_fs_info *info = root->fs_info;
4491         struct btrfs_root *extent_root = info->extent_root;
4492         struct extent_buffer *leaf;
4493         struct btrfs_extent_item *ei;
4494         struct btrfs_extent_inline_ref *iref;
4495         int ret;
4496         int is_data;
4497         int extent_slot = 0;
4498         int found_extent = 0;
4499         int num_to_del = 1;
4500         u32 item_size;
4501         u64 refs;
4502
4503         path = btrfs_alloc_path();
4504         if (!path)
4505                 return -ENOMEM;
4506
4507         path->reada = 1;
4508         path->leave_spinning = 1;
4509
4510         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4511         BUG_ON(!is_data && refs_to_drop != 1);
4512
4513         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4514                                     bytenr, num_bytes, parent,
4515                                     root_objectid, owner_objectid,
4516                                     owner_offset);
4517         if (ret == 0) {
4518                 extent_slot = path->slots[0];
4519                 while (extent_slot >= 0) {
4520                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4521                                               extent_slot);
4522                         if (key.objectid != bytenr)
4523                                 break;
4524                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4525                             key.offset == num_bytes) {
4526                                 found_extent = 1;
4527                                 break;
4528                         }
4529                         if (path->slots[0] - extent_slot > 5)
4530                                 break;
4531                         extent_slot--;
4532                 }
4533 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4534                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4535                 if (found_extent && item_size < sizeof(*ei))
4536                         found_extent = 0;
4537 #endif
4538                 if (!found_extent) {
4539                         BUG_ON(iref);
4540                         ret = remove_extent_backref(trans, extent_root, path,
4541                                                     NULL, refs_to_drop,
4542                                                     is_data);
4543                         BUG_ON(ret);
4544                         btrfs_release_path(extent_root, path);
4545                         path->leave_spinning = 1;
4546
4547                         key.objectid = bytenr;
4548                         key.type = BTRFS_EXTENT_ITEM_KEY;
4549                         key.offset = num_bytes;
4550
4551                         ret = btrfs_search_slot(trans, extent_root,
4552                                                 &key, path, -1, 1);
4553                         if (ret) {
4554                                 printk(KERN_ERR "umm, got %d back from search"
4555                                        ", was looking for %llu\n", ret,
4556                                        (unsigned long long)bytenr);
4557                                 btrfs_print_leaf(extent_root, path->nodes[0]);
4558                         }
4559                         BUG_ON(ret);
4560                         extent_slot = path->slots[0];
4561                 }
4562         } else {
4563                 btrfs_print_leaf(extent_root, path->nodes[0]);
4564                 WARN_ON(1);
4565                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4566                        "parent %llu root %llu  owner %llu offset %llu\n",
4567                        (unsigned long long)bytenr,
4568                        (unsigned long long)parent,
4569                        (unsigned long long)root_objectid,
4570                        (unsigned long long)owner_objectid,
4571                        (unsigned long long)owner_offset);
4572         }
4573
4574         leaf = path->nodes[0];
4575         item_size = btrfs_item_size_nr(leaf, extent_slot);
4576 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4577         if (item_size < sizeof(*ei)) {
4578                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4579                 ret = convert_extent_item_v0(trans, extent_root, path,
4580                                              owner_objectid, 0);
4581                 BUG_ON(ret < 0);
4582
4583                 btrfs_release_path(extent_root, path);
4584                 path->leave_spinning = 1;
4585
4586                 key.objectid = bytenr;
4587                 key.type = BTRFS_EXTENT_ITEM_KEY;
4588                 key.offset = num_bytes;
4589
4590                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4591                                         -1, 1);
4592                 if (ret) {
4593                         printk(KERN_ERR "umm, got %d back from search"
4594                                ", was looking for %llu\n", ret,
4595                                (unsigned long long)bytenr);
4596                         btrfs_print_leaf(extent_root, path->nodes[0]);
4597                 }
4598                 BUG_ON(ret);
4599                 extent_slot = path->slots[0];
4600                 leaf = path->nodes[0];
4601                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4602         }
4603 #endif
4604         BUG_ON(item_size < sizeof(*ei));
4605         ei = btrfs_item_ptr(leaf, extent_slot,
4606                             struct btrfs_extent_item);
4607         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4608                 struct btrfs_tree_block_info *bi;
4609                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4610                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4611                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4612         }
4613
4614         refs = btrfs_extent_refs(leaf, ei);
4615         BUG_ON(refs < refs_to_drop);
4616         refs -= refs_to_drop;
4617
4618         if (refs > 0) {
4619                 if (extent_op)
4620                         __run_delayed_extent_op(extent_op, leaf, ei);
4621                 /*
4622                  * In the case of inline back ref, reference count will
4623                  * be updated by remove_extent_backref
4624                  */
4625                 if (iref) {
4626                         BUG_ON(!found_extent);
4627                 } else {
4628                         btrfs_set_extent_refs(leaf, ei, refs);
4629                         btrfs_mark_buffer_dirty(leaf);
4630                 }
4631                 if (found_extent) {
4632                         ret = remove_extent_backref(trans, extent_root, path,
4633                                                     iref, refs_to_drop,
4634                                                     is_data);
4635                         BUG_ON(ret);
4636                 }
4637         } else {
4638                 if (found_extent) {
4639                         BUG_ON(is_data && refs_to_drop !=
4640                                extent_data_ref_count(root, path, iref));
4641                         if (iref) {
4642                                 BUG_ON(path->slots[0] != extent_slot);
4643                         } else {
4644                                 BUG_ON(path->slots[0] != extent_slot + 1);
4645                                 path->slots[0] = extent_slot;
4646                                 num_to_del = 2;
4647                         }
4648                 }
4649
4650                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4651                                       num_to_del);
4652                 BUG_ON(ret);
4653                 btrfs_release_path(extent_root, path);
4654
4655                 if (is_data) {
4656                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4657                         BUG_ON(ret);
4658                 } else {
4659                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4660                              bytenr >> PAGE_CACHE_SHIFT,
4661                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4662                 }
4663
4664                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4665                 BUG_ON(ret);
4666         }
4667         btrfs_free_path(path);
4668         return ret;
4669 }
4670
4671 /*
4672  * when we free an block, it is possible (and likely) that we free the last
4673  * delayed ref for that extent as well.  This searches the delayed ref tree for
4674  * a given extent, and if there are no other delayed refs to be processed, it
4675  * removes it from the tree.
4676  */
4677 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4678                                       struct btrfs_root *root, u64 bytenr)
4679 {
4680         struct btrfs_delayed_ref_head *head;
4681         struct btrfs_delayed_ref_root *delayed_refs;
4682         struct btrfs_delayed_ref_node *ref;
4683         struct rb_node *node;
4684         int ret = 0;
4685
4686         delayed_refs = &trans->transaction->delayed_refs;
4687         spin_lock(&delayed_refs->lock);
4688         head = btrfs_find_delayed_ref_head(trans, bytenr);
4689         if (!head)
4690                 goto out;
4691
4692         node = rb_prev(&head->node.rb_node);
4693         if (!node)
4694                 goto out;
4695
4696         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4697
4698         /* there are still entries for this ref, we can't drop it */
4699         if (ref->bytenr == bytenr)
4700                 goto out;
4701
4702         if (head->extent_op) {
4703                 if (!head->must_insert_reserved)
4704                         goto out;
4705                 kfree(head->extent_op);
4706                 head->extent_op = NULL;
4707         }
4708
4709         /*
4710          * waiting for the lock here would deadlock.  If someone else has it
4711          * locked they are already in the process of dropping it anyway
4712          */
4713         if (!mutex_trylock(&head->mutex))
4714                 goto out;
4715
4716         /*
4717          * at this point we have a head with no other entries.  Go
4718          * ahead and process it.
4719          */
4720         head->node.in_tree = 0;
4721         rb_erase(&head->node.rb_node, &delayed_refs->root);
4722
4723         delayed_refs->num_entries--;
4724
4725         /*
4726          * we don't take a ref on the node because we're removing it from the
4727          * tree, so we just steal the ref the tree was holding.
4728          */
4729         delayed_refs->num_heads--;
4730         if (list_empty(&head->cluster))
4731                 delayed_refs->num_heads_ready--;
4732
4733         list_del_init(&head->cluster);
4734         spin_unlock(&delayed_refs->lock);
4735
4736         BUG_ON(head->extent_op);
4737         if (head->must_insert_reserved)
4738                 ret = 1;
4739
4740         mutex_unlock(&head->mutex);
4741         btrfs_put_delayed_ref(&head->node);
4742         return ret;
4743 out:
4744         spin_unlock(&delayed_refs->lock);
4745         return 0;
4746 }
4747
4748 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4749                            struct btrfs_root *root,
4750                            struct extent_buffer *buf,
4751                            u64 parent, int last_ref)
4752 {
4753         struct btrfs_block_rsv *block_rsv;
4754         struct btrfs_block_group_cache *cache = NULL;
4755         int ret;
4756
4757         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4758                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4759                                                 parent, root->root_key.objectid,
4760                                                 btrfs_header_level(buf),
4761                                                 BTRFS_DROP_DELAYED_REF, NULL);
4762                 BUG_ON(ret);
4763         }
4764
4765         if (!last_ref)
4766                 return;
4767
4768         block_rsv = get_block_rsv(trans, root);
4769         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4770         if (block_rsv->space_info != cache->space_info)
4771                 goto out;
4772
4773         if (btrfs_header_generation(buf) == trans->transid) {
4774                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4775                         ret = check_ref_cleanup(trans, root, buf->start);
4776                         if (!ret)
4777                                 goto pin;
4778                 }
4779
4780                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4781                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4782                         goto pin;
4783                 }
4784
4785                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4786
4787                 btrfs_add_free_space(cache, buf->start, buf->len);
4788                 ret = btrfs_update_reserved_bytes(cache, buf->len, 0, 0);
4789                 if (ret == -EAGAIN) {
4790                         /* block group became read-only */
4791                         btrfs_update_reserved_bytes(cache, buf->len, 0, 1);
4792                         goto out;
4793                 }
4794
4795                 ret = 1;
4796                 spin_lock(&block_rsv->lock);
4797                 if (block_rsv->reserved < block_rsv->size) {
4798                         block_rsv->reserved += buf->len;
4799                         ret = 0;
4800                 }
4801                 spin_unlock(&block_rsv->lock);
4802
4803                 if (ret) {
4804                         spin_lock(&cache->space_info->lock);
4805                         cache->space_info->bytes_reserved -= buf->len;
4806                         cache->space_info->reservation_progress++;
4807                         spin_unlock(&cache->space_info->lock);
4808                 }
4809                 goto out;
4810         }
4811 pin:
4812         if (block_rsv->durable && !cache->ro) {
4813                 ret = 0;
4814                 spin_lock(&cache->lock);
4815                 if (!cache->ro) {
4816                         cache->reserved_pinned += buf->len;
4817                         ret = 1;
4818                 }
4819                 spin_unlock(&cache->lock);
4820
4821                 if (ret) {
4822                         spin_lock(&block_rsv->lock);
4823                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4824                         spin_unlock(&block_rsv->lock);
4825                 }
4826         }
4827 out:
4828         /*
4829          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4830          * anymore.
4831          */
4832         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4833         btrfs_put_block_group(cache);
4834 }
4835
4836 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4837                       struct btrfs_root *root,
4838                       u64 bytenr, u64 num_bytes, u64 parent,
4839                       u64 root_objectid, u64 owner, u64 offset)
4840 {
4841         int ret;
4842
4843         /*
4844          * tree log blocks never actually go into the extent allocation
4845          * tree, just update pinning info and exit early.
4846          */
4847         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4848                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4849                 /* unlocks the pinned mutex */
4850                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4851                 ret = 0;
4852         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4853                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4854                                         parent, root_objectid, (int)owner,
4855                                         BTRFS_DROP_DELAYED_REF, NULL);
4856                 BUG_ON(ret);
4857         } else {
4858                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4859                                         parent, root_objectid, owner,
4860                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4861                 BUG_ON(ret);
4862         }
4863         return ret;
4864 }
4865
4866 static u64 stripe_align(struct btrfs_root *root, u64 val)
4867 {
4868         u64 mask = ((u64)root->stripesize - 1);
4869         u64 ret = (val + mask) & ~mask;
4870         return ret;
4871 }
4872
4873 /*
4874  * when we wait for progress in the block group caching, its because
4875  * our allocation attempt failed at least once.  So, we must sleep
4876  * and let some progress happen before we try again.
4877  *
4878  * This function will sleep at least once waiting for new free space to
4879  * show up, and then it will check the block group free space numbers
4880  * for our min num_bytes.  Another option is to have it go ahead
4881  * and look in the rbtree for a free extent of a given size, but this
4882  * is a good start.
4883  */
4884 static noinline int
4885 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4886                                 u64 num_bytes)
4887 {
4888         struct btrfs_caching_control *caching_ctl;
4889         DEFINE_WAIT(wait);
4890
4891         caching_ctl = get_caching_control(cache);
4892         if (!caching_ctl)
4893                 return 0;
4894
4895         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4896                    (cache->free_space >= num_bytes));
4897
4898         put_caching_control(caching_ctl);
4899         return 0;
4900 }
4901
4902 static noinline int
4903 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4904 {
4905         struct btrfs_caching_control *caching_ctl;
4906         DEFINE_WAIT(wait);
4907
4908         caching_ctl = get_caching_control(cache);
4909         if (!caching_ctl)
4910                 return 0;
4911
4912         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4913
4914         put_caching_control(caching_ctl);
4915         return 0;
4916 }
4917
4918 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4919 {
4920         int index;
4921         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4922                 index = 0;
4923         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4924                 index = 1;
4925         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4926                 index = 2;
4927         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4928                 index = 3;
4929         else
4930                 index = 4;
4931         return index;
4932 }
4933
4934 enum btrfs_loop_type {
4935         LOOP_FIND_IDEAL = 0,
4936         LOOP_CACHING_NOWAIT = 1,
4937         LOOP_CACHING_WAIT = 2,
4938         LOOP_ALLOC_CHUNK = 3,
4939         LOOP_NO_EMPTY_SIZE = 4,
4940 };
4941
4942 /*
4943  * walks the btree of allocated extents and find a hole of a given size.
4944  * The key ins is changed to record the hole:
4945  * ins->objectid == block start
4946  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4947  * ins->offset == number of blocks
4948  * Any available blocks before search_start are skipped.
4949  */
4950 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4951                                      struct btrfs_root *orig_root,
4952                                      u64 num_bytes, u64 empty_size,
4953                                      u64 search_start, u64 search_end,
4954                                      u64 hint_byte, struct btrfs_key *ins,
4955                                      int data)
4956 {
4957         int ret = 0;
4958         struct btrfs_root *root = orig_root->fs_info->extent_root;
4959         struct btrfs_free_cluster *last_ptr = NULL;
4960         struct btrfs_block_group_cache *block_group = NULL;
4961         int empty_cluster = 2 * 1024 * 1024;
4962         int allowed_chunk_alloc = 0;
4963         int done_chunk_alloc = 0;
4964         struct btrfs_space_info *space_info;
4965         int last_ptr_loop = 0;
4966         int loop = 0;
4967         int index = 0;
4968         bool found_uncached_bg = false;
4969         bool failed_cluster_refill = false;
4970         bool failed_alloc = false;
4971         bool use_cluster = true;
4972         u64 ideal_cache_percent = 0;
4973         u64 ideal_cache_offset = 0;
4974
4975         WARN_ON(num_bytes < root->sectorsize);
4976         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4977         ins->objectid = 0;
4978         ins->offset = 0;
4979
4980         space_info = __find_space_info(root->fs_info, data);
4981         if (!space_info) {
4982                 printk(KERN_ERR "No space info for %d\n", data);
4983                 return -ENOSPC;
4984         }
4985
4986         /*
4987          * If the space info is for both data and metadata it means we have a
4988          * small filesystem and we can't use the clustering stuff.
4989          */
4990         if (btrfs_mixed_space_info(space_info))
4991                 use_cluster = false;
4992
4993         if (orig_root->ref_cows || empty_size)
4994                 allowed_chunk_alloc = 1;
4995
4996         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4997                 last_ptr = &root->fs_info->meta_alloc_cluster;
4998                 if (!btrfs_test_opt(root, SSD))
4999                         empty_cluster = 64 * 1024;
5000         }
5001
5002         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5003             btrfs_test_opt(root, SSD)) {
5004                 last_ptr = &root->fs_info->data_alloc_cluster;
5005         }
5006
5007         if (last_ptr) {
5008                 spin_lock(&last_ptr->lock);
5009                 if (last_ptr->block_group)
5010                         hint_byte = last_ptr->window_start;
5011                 spin_unlock(&last_ptr->lock);
5012         }
5013
5014         search_start = max(search_start, first_logical_byte(root, 0));
5015         search_start = max(search_start, hint_byte);
5016
5017         if (!last_ptr)
5018                 empty_cluster = 0;
5019
5020         if (search_start == hint_byte) {
5021 ideal_cache:
5022                 block_group = btrfs_lookup_block_group(root->fs_info,
5023                                                        search_start);
5024                 /*
5025                  * we don't want to use the block group if it doesn't match our
5026                  * allocation bits, or if its not cached.
5027                  *
5028                  * However if we are re-searching with an ideal block group
5029                  * picked out then we don't care that the block group is cached.
5030                  */
5031                 if (block_group && block_group_bits(block_group, data) &&
5032                     (block_group->cached != BTRFS_CACHE_NO ||
5033                      search_start == ideal_cache_offset)) {
5034                         down_read(&space_info->groups_sem);
5035                         if (list_empty(&block_group->list) ||
5036                             block_group->ro) {
5037                                 /*
5038                                  * someone is removing this block group,
5039                                  * we can't jump into the have_block_group
5040                                  * target because our list pointers are not
5041                                  * valid
5042                                  */
5043                                 btrfs_put_block_group(block_group);
5044                                 up_read(&space_info->groups_sem);
5045                         } else {
5046                                 index = get_block_group_index(block_group);
5047                                 goto have_block_group;
5048                         }
5049                 } else if (block_group) {
5050                         btrfs_put_block_group(block_group);
5051                 }
5052         }
5053 search:
5054         down_read(&space_info->groups_sem);
5055         list_for_each_entry(block_group, &space_info->block_groups[index],
5056                             list) {
5057                 u64 offset;
5058                 int cached;
5059
5060                 btrfs_get_block_group(block_group);
5061                 search_start = block_group->key.objectid;
5062
5063                 /*
5064                  * this can happen if we end up cycling through all the
5065                  * raid types, but we want to make sure we only allocate
5066                  * for the proper type.
5067                  */
5068                 if (!block_group_bits(block_group, data)) {
5069                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5070                                 BTRFS_BLOCK_GROUP_RAID1 |
5071                                 BTRFS_BLOCK_GROUP_RAID10;
5072
5073                         /*
5074                          * if they asked for extra copies and this block group
5075                          * doesn't provide them, bail.  This does allow us to
5076                          * fill raid0 from raid1.
5077                          */
5078                         if ((data & extra) && !(block_group->flags & extra))
5079                                 goto loop;
5080                 }
5081
5082 have_block_group:
5083                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5084                         u64 free_percent;
5085
5086                         ret = cache_block_group(block_group, trans,
5087                                                 orig_root, 1);
5088                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5089                                 goto have_block_group;
5090
5091                         free_percent = btrfs_block_group_used(&block_group->item);
5092                         free_percent *= 100;
5093                         free_percent = div64_u64(free_percent,
5094                                                  block_group->key.offset);
5095                         free_percent = 100 - free_percent;
5096                         if (free_percent > ideal_cache_percent &&
5097                             likely(!block_group->ro)) {
5098                                 ideal_cache_offset = block_group->key.objectid;
5099                                 ideal_cache_percent = free_percent;
5100                         }
5101
5102                         /*
5103                          * We only want to start kthread caching if we are at
5104                          * the point where we will wait for caching to make
5105                          * progress, or if our ideal search is over and we've
5106                          * found somebody to start caching.
5107                          */
5108                         if (loop > LOOP_CACHING_NOWAIT ||
5109                             (loop > LOOP_FIND_IDEAL &&
5110                              atomic_read(&space_info->caching_threads) < 2)) {
5111                                 ret = cache_block_group(block_group, trans,
5112                                                         orig_root, 0);
5113                                 BUG_ON(ret);
5114                         }
5115                         found_uncached_bg = true;
5116
5117                         /*
5118                          * If loop is set for cached only, try the next block
5119                          * group.
5120                          */
5121                         if (loop == LOOP_FIND_IDEAL)
5122                                 goto loop;
5123                 }
5124
5125                 cached = block_group_cache_done(block_group);
5126                 if (unlikely(!cached))
5127                         found_uncached_bg = true;
5128
5129                 if (unlikely(block_group->ro))
5130                         goto loop;
5131
5132                 /*
5133                  * Ok we want to try and use the cluster allocator, so lets look
5134                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5135                  * have tried the cluster allocator plenty of times at this
5136                  * point and not have found anything, so we are likely way too
5137                  * fragmented for the clustering stuff to find anything, so lets
5138                  * just skip it and let the allocator find whatever block it can
5139                  * find
5140                  */
5141                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5142                         /*
5143                          * the refill lock keeps out other
5144                          * people trying to start a new cluster
5145                          */
5146                         spin_lock(&last_ptr->refill_lock);
5147                         if (last_ptr->block_group &&
5148                             (last_ptr->block_group->ro ||
5149                             !block_group_bits(last_ptr->block_group, data))) {
5150                                 offset = 0;
5151                                 goto refill_cluster;
5152                         }
5153
5154                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5155                                                  num_bytes, search_start);
5156                         if (offset) {
5157                                 /* we have a block, we're done */
5158                                 spin_unlock(&last_ptr->refill_lock);
5159                                 goto checks;
5160                         }
5161
5162                         spin_lock(&last_ptr->lock);
5163                         /*
5164                          * whoops, this cluster doesn't actually point to
5165                          * this block group.  Get a ref on the block
5166                          * group is does point to and try again
5167                          */
5168                         if (!last_ptr_loop && last_ptr->block_group &&
5169                             last_ptr->block_group != block_group) {
5170
5171                                 btrfs_put_block_group(block_group);
5172                                 block_group = last_ptr->block_group;
5173                                 btrfs_get_block_group(block_group);
5174                                 spin_unlock(&last_ptr->lock);
5175                                 spin_unlock(&last_ptr->refill_lock);
5176
5177                                 last_ptr_loop = 1;
5178                                 search_start = block_group->key.objectid;
5179                                 /*
5180                                  * we know this block group is properly
5181                                  * in the list because
5182                                  * btrfs_remove_block_group, drops the
5183                                  * cluster before it removes the block
5184                                  * group from the list
5185                                  */
5186                                 goto have_block_group;
5187                         }
5188                         spin_unlock(&last_ptr->lock);
5189 refill_cluster:
5190                         /*
5191                          * this cluster didn't work out, free it and
5192                          * start over
5193                          */
5194                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5195
5196                         last_ptr_loop = 0;
5197
5198                         /* allocate a cluster in this block group */
5199                         ret = btrfs_find_space_cluster(trans, root,
5200                                                block_group, last_ptr,
5201                                                offset, num_bytes,
5202                                                empty_cluster + empty_size);
5203                         if (ret == 0) {
5204                                 /*
5205                                  * now pull our allocation out of this
5206                                  * cluster
5207                                  */
5208                                 offset = btrfs_alloc_from_cluster(block_group,
5209                                                   last_ptr, num_bytes,
5210                                                   search_start);
5211                                 if (offset) {
5212                                         /* we found one, proceed */
5213                                         spin_unlock(&last_ptr->refill_lock);
5214                                         goto checks;
5215                                 }
5216                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5217                                    && !failed_cluster_refill) {
5218                                 spin_unlock(&last_ptr->refill_lock);
5219
5220                                 failed_cluster_refill = true;
5221                                 wait_block_group_cache_progress(block_group,
5222                                        num_bytes + empty_cluster + empty_size);
5223                                 goto have_block_group;
5224                         }
5225
5226                         /*
5227                          * at this point we either didn't find a cluster
5228                          * or we weren't able to allocate a block from our
5229                          * cluster.  Free the cluster we've been trying
5230                          * to use, and go to the next block group
5231                          */
5232                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5233                         spin_unlock(&last_ptr->refill_lock);
5234                         goto loop;
5235                 }
5236
5237                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5238                                                     num_bytes, empty_size);
5239                 /*
5240                  * If we didn't find a chunk, and we haven't failed on this
5241                  * block group before, and this block group is in the middle of
5242                  * caching and we are ok with waiting, then go ahead and wait
5243                  * for progress to be made, and set failed_alloc to true.
5244                  *
5245                  * If failed_alloc is true then we've already waited on this
5246                  * block group once and should move on to the next block group.
5247                  */
5248                 if (!offset && !failed_alloc && !cached &&
5249                     loop > LOOP_CACHING_NOWAIT) {
5250                         wait_block_group_cache_progress(block_group,
5251                                                 num_bytes + empty_size);
5252                         failed_alloc = true;
5253                         goto have_block_group;
5254                 } else if (!offset) {
5255                         goto loop;
5256                 }
5257 checks:
5258                 search_start = stripe_align(root, offset);
5259                 /* move on to the next group */
5260                 if (search_start + num_bytes >= search_end) {
5261                         btrfs_add_free_space(block_group, offset, num_bytes);
5262                         goto loop;
5263                 }
5264
5265                 /* move on to the next group */
5266                 if (search_start + num_bytes >
5267                     block_group->key.objectid + block_group->key.offset) {
5268                         btrfs_add_free_space(block_group, offset, num_bytes);
5269                         goto loop;
5270                 }
5271
5272                 ins->objectid = search_start;
5273                 ins->offset = num_bytes;
5274
5275                 if (offset < search_start)
5276                         btrfs_add_free_space(block_group, offset,
5277                                              search_start - offset);
5278                 BUG_ON(offset > search_start);
5279
5280                 ret = btrfs_update_reserved_bytes(block_group, num_bytes, 1,
5281                                             (data & BTRFS_BLOCK_GROUP_DATA));
5282                 if (ret == -EAGAIN) {
5283                         btrfs_add_free_space(block_group, offset, num_bytes);
5284                         goto loop;
5285                 }
5286
5287                 /* we are all good, lets return */
5288                 ins->objectid = search_start;
5289                 ins->offset = num_bytes;
5290
5291                 if (offset < search_start)
5292                         btrfs_add_free_space(block_group, offset,
5293                                              search_start - offset);
5294                 BUG_ON(offset > search_start);
5295                 break;
5296 loop:
5297                 failed_cluster_refill = false;
5298                 failed_alloc = false;
5299                 BUG_ON(index != get_block_group_index(block_group));
5300                 btrfs_put_block_group(block_group);
5301         }
5302         up_read(&space_info->groups_sem);
5303
5304         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5305                 goto search;
5306
5307         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5308          *                      for them to make caching progress.  Also
5309          *                      determine the best possible bg to cache
5310          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5311          *                      caching kthreads as we move along
5312          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5313          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5314          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5315          *                      again
5316          */
5317         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
5318             (found_uncached_bg || empty_size || empty_cluster ||
5319              allowed_chunk_alloc)) {
5320                 index = 0;
5321                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5322                         found_uncached_bg = false;
5323                         loop++;
5324                         if (!ideal_cache_percent &&
5325                             atomic_read(&space_info->caching_threads))
5326                                 goto search;
5327
5328                         /*
5329                          * 1 of the following 2 things have happened so far
5330                          *
5331                          * 1) We found an ideal block group for caching that
5332                          * is mostly full and will cache quickly, so we might
5333                          * as well wait for it.
5334                          *
5335                          * 2) We searched for cached only and we didn't find
5336                          * anything, and we didn't start any caching kthreads
5337                          * either, so chances are we will loop through and
5338                          * start a couple caching kthreads, and then come back
5339                          * around and just wait for them.  This will be slower
5340                          * because we will have 2 caching kthreads reading at
5341                          * the same time when we could have just started one
5342                          * and waited for it to get far enough to give us an
5343                          * allocation, so go ahead and go to the wait caching
5344                          * loop.
5345                          */
5346                         loop = LOOP_CACHING_WAIT;
5347                         search_start = ideal_cache_offset;
5348                         ideal_cache_percent = 0;
5349                         goto ideal_cache;
5350                 } else if (loop == LOOP_FIND_IDEAL) {
5351                         /*
5352                          * Didn't find a uncached bg, wait on anything we find
5353                          * next.
5354                          */
5355                         loop = LOOP_CACHING_WAIT;
5356                         goto search;
5357                 }
5358
5359                 if (loop < LOOP_CACHING_WAIT) {
5360                         loop++;
5361                         goto search;
5362                 }
5363
5364                 if (loop == LOOP_ALLOC_CHUNK) {
5365                         empty_size = 0;
5366                         empty_cluster = 0;
5367                 }
5368
5369                 if (allowed_chunk_alloc) {
5370                         ret = do_chunk_alloc(trans, root, num_bytes +
5371                                              2 * 1024 * 1024, data,
5372                                              CHUNK_ALLOC_LIMITED);
5373                         allowed_chunk_alloc = 0;
5374                         done_chunk_alloc = 1;
5375                 } else if (!done_chunk_alloc &&
5376                            space_info->force_alloc == CHUNK_ALLOC_NO_FORCE) {
5377                         space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5378                 }
5379
5380                 if (loop < LOOP_NO_EMPTY_SIZE) {
5381                         loop++;
5382                         goto search;
5383                 }
5384                 ret = -ENOSPC;
5385         } else if (!ins->objectid) {
5386                 ret = -ENOSPC;
5387         }
5388
5389         /* we found what we needed */
5390         if (ins->objectid) {
5391                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
5392                         trans->block_group = block_group->key.objectid;
5393
5394                 btrfs_put_block_group(block_group);
5395                 ret = 0;
5396         }
5397
5398         return ret;
5399 }
5400
5401 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5402                             int dump_block_groups)
5403 {
5404         struct btrfs_block_group_cache *cache;
5405         int index = 0;
5406
5407         spin_lock(&info->lock);
5408         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5409                (unsigned long long)(info->total_bytes - info->bytes_used -
5410                                     info->bytes_pinned - info->bytes_reserved -
5411                                     info->bytes_readonly),
5412                (info->full) ? "" : "not ");
5413         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5414                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5415                (unsigned long long)info->total_bytes,
5416                (unsigned long long)info->bytes_used,
5417                (unsigned long long)info->bytes_pinned,
5418                (unsigned long long)info->bytes_reserved,
5419                (unsigned long long)info->bytes_may_use,
5420                (unsigned long long)info->bytes_readonly);
5421         spin_unlock(&info->lock);
5422
5423         if (!dump_block_groups)
5424                 return;
5425
5426         down_read(&info->groups_sem);
5427 again:
5428         list_for_each_entry(cache, &info->block_groups[index], list) {
5429                 spin_lock(&cache->lock);
5430                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5431                        "%llu pinned %llu reserved\n",
5432                        (unsigned long long)cache->key.objectid,
5433                        (unsigned long long)cache->key.offset,
5434                        (unsigned long long)btrfs_block_group_used(&cache->item),
5435                        (unsigned long long)cache->pinned,
5436                        (unsigned long long)cache->reserved);
5437                 btrfs_dump_free_space(cache, bytes);
5438                 spin_unlock(&cache->lock);
5439         }
5440         if (++index < BTRFS_NR_RAID_TYPES)
5441                 goto again;
5442         up_read(&info->groups_sem);
5443 }
5444
5445 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5446                          struct btrfs_root *root,
5447                          u64 num_bytes, u64 min_alloc_size,
5448                          u64 empty_size, u64 hint_byte,
5449                          u64 search_end, struct btrfs_key *ins,
5450                          u64 data)
5451 {
5452         int ret;
5453         u64 search_start = 0;
5454
5455         data = btrfs_get_alloc_profile(root, data);
5456 again:
5457         /*
5458          * the only place that sets empty_size is btrfs_realloc_node, which
5459          * is not called recursively on allocations
5460          */
5461         if (empty_size || root->ref_cows)
5462                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5463                                      num_bytes + 2 * 1024 * 1024, data,
5464                                      CHUNK_ALLOC_NO_FORCE);
5465
5466         WARN_ON(num_bytes < root->sectorsize);
5467         ret = find_free_extent(trans, root, num_bytes, empty_size,
5468                                search_start, search_end, hint_byte,
5469                                ins, data);
5470
5471         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5472                 num_bytes = num_bytes >> 1;
5473                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5474                 num_bytes = max(num_bytes, min_alloc_size);
5475                 do_chunk_alloc(trans, root->fs_info->extent_root,
5476                                num_bytes, data, CHUNK_ALLOC_FORCE);
5477                 goto again;
5478         }
5479         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5480                 struct btrfs_space_info *sinfo;
5481
5482                 sinfo = __find_space_info(root->fs_info, data);
5483                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5484                        "wanted %llu\n", (unsigned long long)data,
5485                        (unsigned long long)num_bytes);
5486                 dump_space_info(sinfo, num_bytes, 1);
5487         }
5488
5489         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5490
5491         return ret;
5492 }
5493
5494 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5495 {
5496         struct btrfs_block_group_cache *cache;
5497         int ret = 0;
5498
5499         cache = btrfs_lookup_block_group(root->fs_info, start);
5500         if (!cache) {
5501                 printk(KERN_ERR "Unable to find block group for %llu\n",
5502                        (unsigned long long)start);
5503                 return -ENOSPC;
5504         }
5505
5506         if (btrfs_test_opt(root, DISCARD))
5507                 ret = btrfs_discard_extent(root, start, len, NULL);
5508
5509         btrfs_add_free_space(cache, start, len);
5510         btrfs_update_reserved_bytes(cache, len, 0, 1);
5511         btrfs_put_block_group(cache);
5512
5513         trace_btrfs_reserved_extent_free(root, start, len);
5514
5515         return ret;
5516 }
5517
5518 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5519                                       struct btrfs_root *root,
5520                                       u64 parent, u64 root_objectid,
5521                                       u64 flags, u64 owner, u64 offset,
5522                                       struct btrfs_key *ins, int ref_mod)
5523 {
5524         int ret;
5525         struct btrfs_fs_info *fs_info = root->fs_info;
5526         struct btrfs_extent_item *extent_item;
5527         struct btrfs_extent_inline_ref *iref;
5528         struct btrfs_path *path;
5529         struct extent_buffer *leaf;
5530         int type;
5531         u32 size;
5532
5533         if (parent > 0)
5534                 type = BTRFS_SHARED_DATA_REF_KEY;
5535         else
5536                 type = BTRFS_EXTENT_DATA_REF_KEY;
5537
5538         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5539
5540         path = btrfs_alloc_path();
5541         if (!path)
5542                 return -ENOMEM;
5543
5544         path->leave_spinning = 1;
5545         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5546                                       ins, size);
5547         BUG_ON(ret);
5548
5549         leaf = path->nodes[0];
5550         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5551                                      struct btrfs_extent_item);
5552         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5553         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5554         btrfs_set_extent_flags(leaf, extent_item,
5555                                flags | BTRFS_EXTENT_FLAG_DATA);
5556
5557         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5558         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5559         if (parent > 0) {
5560                 struct btrfs_shared_data_ref *ref;
5561                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5562                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5563                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5564         } else {
5565                 struct btrfs_extent_data_ref *ref;
5566                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5567                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5568                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5569                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5570                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5571         }
5572
5573         btrfs_mark_buffer_dirty(path->nodes[0]);
5574         btrfs_free_path(path);
5575
5576         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5577         if (ret) {
5578                 printk(KERN_ERR "btrfs update block group failed for %llu "
5579                        "%llu\n", (unsigned long long)ins->objectid,
5580                        (unsigned long long)ins->offset);
5581                 BUG();
5582         }
5583         return ret;
5584 }
5585
5586 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5587                                      struct btrfs_root *root,
5588                                      u64 parent, u64 root_objectid,
5589                                      u64 flags, struct btrfs_disk_key *key,
5590                                      int level, struct btrfs_key *ins)
5591 {
5592         int ret;
5593         struct btrfs_fs_info *fs_info = root->fs_info;
5594         struct btrfs_extent_item *extent_item;
5595         struct btrfs_tree_block_info *block_info;
5596         struct btrfs_extent_inline_ref *iref;
5597         struct btrfs_path *path;
5598         struct extent_buffer *leaf;
5599         u32 size&