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