Btrfs: cleanup for duplicated code in find_free_extent
[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 #undef SCRAMBLE_DELAYED_REFS
38
39 /*
40  * control flags for do_chunk_alloc's force field
41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
42  * if we really need one.
43  *
44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
45  * if we have very few chunks already allocated.  This is
46  * used as part of the clustering code to help make sure
47  * we have a good pool of storage to cluster in, without
48  * filling the FS with empty chunks
49  *
50  * CHUNK_ALLOC_FORCE means it must try to allocate one
51  *
52  */
53 enum {
54         CHUNK_ALLOC_NO_FORCE = 0,
55         CHUNK_ALLOC_LIMITED = 1,
56         CHUNK_ALLOC_FORCE = 2,
57 };
58
59 /*
60  * Control how reservations are dealt with.
61  *
62  * RESERVE_FREE - freeing a reservation.
63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
64  *   ENOSPC accounting
65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
67  */
68 enum {
69         RESERVE_FREE = 0,
70         RESERVE_ALLOC = 1,
71         RESERVE_ALLOC_NO_ACCOUNT = 2,
72 };
73
74 static int update_block_group(struct btrfs_trans_handle *trans,
75                               struct btrfs_root *root,
76                               u64 bytenr, u64 num_bytes, int alloc);
77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
78                                 struct btrfs_root *root,
79                                 u64 bytenr, u64 num_bytes, u64 parent,
80                                 u64 root_objectid, u64 owner_objectid,
81                                 u64 owner_offset, int refs_to_drop,
82                                 struct btrfs_delayed_extent_op *extra_op);
83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
84                                     struct extent_buffer *leaf,
85                                     struct btrfs_extent_item *ei);
86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
87                                       struct btrfs_root *root,
88                                       u64 parent, u64 root_objectid,
89                                       u64 flags, u64 owner, u64 offset,
90                                       struct btrfs_key *ins, int ref_mod);
91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
92                                      struct btrfs_root *root,
93                                      u64 parent, u64 root_objectid,
94                                      u64 flags, struct btrfs_disk_key *key,
95                                      int level, struct btrfs_key *ins);
96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
97                           struct btrfs_root *extent_root, u64 flags,
98                           int force);
99 static int find_next_key(struct btrfs_path *path, int level,
100                          struct btrfs_key *key);
101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
102                             int dump_block_groups);
103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
104                                        u64 num_bytes, int reserve);
105
106 static noinline int
107 block_group_cache_done(struct btrfs_block_group_cache *cache)
108 {
109         smp_mb();
110         return cache->cached == BTRFS_CACHE_FINISHED;
111 }
112
113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
114 {
115         return (cache->flags & bits) == bits;
116 }
117
118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
119 {
120         atomic_inc(&cache->count);
121 }
122
123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
124 {
125         if (atomic_dec_and_test(&cache->count)) {
126                 WARN_ON(cache->pinned > 0);
127                 WARN_ON(cache->reserved > 0);
128                 kfree(cache->free_space_ctl);
129                 kfree(cache);
130         }
131 }
132
133 /*
134  * this adds the block group to the fs_info rb tree for the block group
135  * cache
136  */
137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
138                                 struct btrfs_block_group_cache *block_group)
139 {
140         struct rb_node **p;
141         struct rb_node *parent = NULL;
142         struct btrfs_block_group_cache *cache;
143
144         spin_lock(&info->block_group_cache_lock);
145         p = &info->block_group_cache_tree.rb_node;
146
147         while (*p) {
148                 parent = *p;
149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
150                                  cache_node);
151                 if (block_group->key.objectid < cache->key.objectid) {
152                         p = &(*p)->rb_left;
153                 } else if (block_group->key.objectid > cache->key.objectid) {
154                         p = &(*p)->rb_right;
155                 } else {
156                         spin_unlock(&info->block_group_cache_lock);
157                         return -EEXIST;
158                 }
159         }
160
161         rb_link_node(&block_group->cache_node, parent, p);
162         rb_insert_color(&block_group->cache_node,
163                         &info->block_group_cache_tree);
164         spin_unlock(&info->block_group_cache_lock);
165
166         return 0;
167 }
168
169 /*
170  * This will return the block group at or after bytenr if contains is 0, else
171  * it will return the block group that contains the bytenr
172  */
173 static struct btrfs_block_group_cache *
174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
175                               int contains)
176 {
177         struct btrfs_block_group_cache *cache, *ret = NULL;
178         struct rb_node *n;
179         u64 end, start;
180
181         spin_lock(&info->block_group_cache_lock);
182         n = info->block_group_cache_tree.rb_node;
183
184         while (n) {
185                 cache = rb_entry(n, struct btrfs_block_group_cache,
186                                  cache_node);
187                 end = cache->key.objectid + cache->key.offset - 1;
188                 start = cache->key.objectid;
189
190                 if (bytenr < start) {
191                         if (!contains && (!ret || start < ret->key.objectid))
192                                 ret = cache;
193                         n = n->rb_left;
194                 } else if (bytenr > start) {
195                         if (contains && bytenr <= end) {
196                                 ret = cache;
197                                 break;
198                         }
199                         n = n->rb_right;
200                 } else {
201                         ret = cache;
202                         break;
203                 }
204         }
205         if (ret)
206                 btrfs_get_block_group(ret);
207         spin_unlock(&info->block_group_cache_lock);
208
209         return ret;
210 }
211
212 static int add_excluded_extent(struct btrfs_root *root,
213                                u64 start, u64 num_bytes)
214 {
215         u64 end = start + num_bytes - 1;
216         set_extent_bits(&root->fs_info->freed_extents[0],
217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
218         set_extent_bits(&root->fs_info->freed_extents[1],
219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
220         return 0;
221 }
222
223 static void free_excluded_extents(struct btrfs_root *root,
224                                   struct btrfs_block_group_cache *cache)
225 {
226         u64 start, end;
227
228         start = cache->key.objectid;
229         end = start + cache->key.offset - 1;
230
231         clear_extent_bits(&root->fs_info->freed_extents[0],
232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
233         clear_extent_bits(&root->fs_info->freed_extents[1],
234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
235 }
236
237 static int exclude_super_stripes(struct btrfs_root *root,
238                                  struct btrfs_block_group_cache *cache)
239 {
240         u64 bytenr;
241         u64 *logical;
242         int stripe_len;
243         int i, nr, ret;
244
245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
247                 cache->bytes_super += stripe_len;
248                 ret = add_excluded_extent(root, cache->key.objectid,
249                                           stripe_len);
250                 BUG_ON(ret); /* -ENOMEM */
251         }
252
253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
254                 bytenr = btrfs_sb_offset(i);
255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
256                                        cache->key.objectid, bytenr,
257                                        0, &logical, &nr, &stripe_len);
258                 BUG_ON(ret); /* -ENOMEM */
259
260                 while (nr--) {
261                         cache->bytes_super += stripe_len;
262                         ret = add_excluded_extent(root, logical[nr],
263                                                   stripe_len);
264                         BUG_ON(ret); /* -ENOMEM */
265                 }
266
267                 kfree(logical);
268         }
269         return 0;
270 }
271
272 static struct btrfs_caching_control *
273 get_caching_control(struct btrfs_block_group_cache *cache)
274 {
275         struct btrfs_caching_control *ctl;
276
277         spin_lock(&cache->lock);
278         if (cache->cached != BTRFS_CACHE_STARTED) {
279                 spin_unlock(&cache->lock);
280                 return NULL;
281         }
282
283         /* We're loading it the fast way, so we don't have a caching_ctl. */
284         if (!cache->caching_ctl) {
285                 spin_unlock(&cache->lock);
286                 return NULL;
287         }
288
289         ctl = cache->caching_ctl;
290         atomic_inc(&ctl->count);
291         spin_unlock(&cache->lock);
292         return ctl;
293 }
294
295 static void put_caching_control(struct btrfs_caching_control *ctl)
296 {
297         if (atomic_dec_and_test(&ctl->count))
298                 kfree(ctl);
299 }
300
301 /*
302  * this is only called by cache_block_group, since we could have freed extents
303  * we need to check the pinned_extents for any extents that can't be used yet
304  * since their free space will be released as soon as the transaction commits.
305  */
306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
307                               struct btrfs_fs_info *info, u64 start, u64 end)
308 {
309         u64 extent_start, extent_end, size, total_added = 0;
310         int ret;
311
312         while (start < end) {
313                 ret = find_first_extent_bit(info->pinned_extents, start,
314                                             &extent_start, &extent_end,
315                                             EXTENT_DIRTY | EXTENT_UPTODATE);
316                 if (ret)
317                         break;
318
319                 if (extent_start <= start) {
320                         start = extent_end + 1;
321                 } else if (extent_start > start && extent_start < end) {
322                         size = extent_start - start;
323                         total_added += size;
324                         ret = btrfs_add_free_space(block_group, start,
325                                                    size);
326                         BUG_ON(ret); /* -ENOMEM or logic error */
327                         start = extent_end + 1;
328                 } else {
329                         break;
330                 }
331         }
332
333         if (start < end) {
334                 size = end - start;
335                 total_added += size;
336                 ret = btrfs_add_free_space(block_group, start, size);
337                 BUG_ON(ret); /* -ENOMEM or logic error */
338         }
339
340         return total_added;
341 }
342
343 static noinline void caching_thread(struct btrfs_work *work)
344 {
345         struct btrfs_block_group_cache *block_group;
346         struct btrfs_fs_info *fs_info;
347         struct btrfs_caching_control *caching_ctl;
348         struct btrfs_root *extent_root;
349         struct btrfs_path *path;
350         struct extent_buffer *leaf;
351         struct btrfs_key key;
352         u64 total_found = 0;
353         u64 last = 0;
354         u32 nritems;
355         int ret = 0;
356
357         caching_ctl = container_of(work, struct btrfs_caching_control, work);
358         block_group = caching_ctl->block_group;
359         fs_info = block_group->fs_info;
360         extent_root = fs_info->extent_root;
361
362         path = btrfs_alloc_path();
363         if (!path)
364                 goto out;
365
366         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
367
368         /*
369          * We don't want to deadlock with somebody trying to allocate a new
370          * extent for the extent root while also trying to search the extent
371          * root to add free space.  So we skip locking and search the commit
372          * root, since its read-only
373          */
374         path->skip_locking = 1;
375         path->search_commit_root = 1;
376         path->reada = 1;
377
378         key.objectid = last;
379         key.offset = 0;
380         key.type = BTRFS_EXTENT_ITEM_KEY;
381 again:
382         mutex_lock(&caching_ctl->mutex);
383         /* need to make sure the commit_root doesn't disappear */
384         down_read(&fs_info->extent_commit_sem);
385
386         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
387         if (ret < 0)
388                 goto err;
389
390         leaf = path->nodes[0];
391         nritems = btrfs_header_nritems(leaf);
392
393         while (1) {
394                 if (btrfs_fs_closing(fs_info) > 1) {
395                         last = (u64)-1;
396                         break;
397                 }
398
399                 if (path->slots[0] < nritems) {
400                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
401                 } else {
402                         ret = find_next_key(path, 0, &key);
403                         if (ret)
404                                 break;
405
406                         if (need_resched() ||
407                             btrfs_next_leaf(extent_root, path)) {
408                                 caching_ctl->progress = last;
409                                 btrfs_release_path(path);
410                                 up_read(&fs_info->extent_commit_sem);
411                                 mutex_unlock(&caching_ctl->mutex);
412                                 cond_resched();
413                                 goto again;
414                         }
415                         leaf = path->nodes[0];
416                         nritems = btrfs_header_nritems(leaf);
417                         continue;
418                 }
419
420                 if (key.objectid < block_group->key.objectid) {
421                         path->slots[0]++;
422                         continue;
423                 }
424
425                 if (key.objectid >= block_group->key.objectid +
426                     block_group->key.offset)
427                         break;
428
429                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
430                         total_found += add_new_free_space(block_group,
431                                                           fs_info, last,
432                                                           key.objectid);
433                         last = key.objectid + key.offset;
434
435                         if (total_found > (1024 * 1024 * 2)) {
436                                 total_found = 0;
437                                 wake_up(&caching_ctl->wait);
438                         }
439                 }
440                 path->slots[0]++;
441         }
442         ret = 0;
443
444         total_found += add_new_free_space(block_group, fs_info, last,
445                                           block_group->key.objectid +
446                                           block_group->key.offset);
447         caching_ctl->progress = (u64)-1;
448
449         spin_lock(&block_group->lock);
450         block_group->caching_ctl = NULL;
451         block_group->cached = BTRFS_CACHE_FINISHED;
452         spin_unlock(&block_group->lock);
453
454 err:
455         btrfs_free_path(path);
456         up_read(&fs_info->extent_commit_sem);
457
458         free_excluded_extents(extent_root, block_group);
459
460         mutex_unlock(&caching_ctl->mutex);
461 out:
462         wake_up(&caching_ctl->wait);
463
464         put_caching_control(caching_ctl);
465         btrfs_put_block_group(block_group);
466 }
467
468 static int cache_block_group(struct btrfs_block_group_cache *cache,
469                              struct btrfs_trans_handle *trans,
470                              struct btrfs_root *root,
471                              int load_cache_only)
472 {
473         DEFINE_WAIT(wait);
474         struct btrfs_fs_info *fs_info = cache->fs_info;
475         struct btrfs_caching_control *caching_ctl;
476         int ret = 0;
477
478         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
479         if (!caching_ctl)
480                 return -ENOMEM;
481
482         INIT_LIST_HEAD(&caching_ctl->list);
483         mutex_init(&caching_ctl->mutex);
484         init_waitqueue_head(&caching_ctl->wait);
485         caching_ctl->block_group = cache;
486         caching_ctl->progress = cache->key.objectid;
487         atomic_set(&caching_ctl->count, 1);
488         caching_ctl->work.func = caching_thread;
489
490         spin_lock(&cache->lock);
491         /*
492          * This should be a rare occasion, but this could happen I think in the
493          * case where one thread starts to load the space cache info, and then
494          * some other thread starts a transaction commit which tries to do an
495          * allocation while the other thread is still loading the space cache
496          * info.  The previous loop should have kept us from choosing this block
497          * group, but if we've moved to the state where we will wait on caching
498          * block groups we need to first check if we're doing a fast load here,
499          * so we can wait for it to finish, otherwise we could end up allocating
500          * from a block group who's cache gets evicted for one reason or
501          * another.
502          */
503         while (cache->cached == BTRFS_CACHE_FAST) {
504                 struct btrfs_caching_control *ctl;
505
506                 ctl = cache->caching_ctl;
507                 atomic_inc(&ctl->count);
508                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
509                 spin_unlock(&cache->lock);
510
511                 schedule();
512
513                 finish_wait(&ctl->wait, &wait);
514                 put_caching_control(ctl);
515                 spin_lock(&cache->lock);
516         }
517
518         if (cache->cached != BTRFS_CACHE_NO) {
519                 spin_unlock(&cache->lock);
520                 kfree(caching_ctl);
521                 return 0;
522         }
523         WARN_ON(cache->caching_ctl);
524         cache->caching_ctl = caching_ctl;
525         cache->cached = BTRFS_CACHE_FAST;
526         spin_unlock(&cache->lock);
527
528         /*
529          * We can't do the read from on-disk cache during a commit since we need
530          * to have the normal tree locking.  Also if we are currently trying to
531          * allocate blocks for the tree root we can't do the fast caching since
532          * we likely hold important locks.
533          */
534         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
535                 ret = load_free_space_cache(fs_info, cache);
536
537                 spin_lock(&cache->lock);
538                 if (ret == 1) {
539                         cache->caching_ctl = NULL;
540                         cache->cached = BTRFS_CACHE_FINISHED;
541                         cache->last_byte_to_unpin = (u64)-1;
542                 } else {
543                         if (load_cache_only) {
544                                 cache->caching_ctl = NULL;
545                                 cache->cached = BTRFS_CACHE_NO;
546                         } else {
547                                 cache->cached = BTRFS_CACHE_STARTED;
548                         }
549                 }
550                 spin_unlock(&cache->lock);
551                 wake_up(&caching_ctl->wait);
552                 if (ret == 1) {
553                         put_caching_control(caching_ctl);
554                         free_excluded_extents(fs_info->extent_root, cache);
555                         return 0;
556                 }
557         } else {
558                 /*
559                  * We are not going to do the fast caching, set cached to the
560                  * appropriate value and wakeup any waiters.
561                  */
562                 spin_lock(&cache->lock);
563                 if (load_cache_only) {
564                         cache->caching_ctl = NULL;
565                         cache->cached = BTRFS_CACHE_NO;
566                 } else {
567                         cache->cached = BTRFS_CACHE_STARTED;
568                 }
569                 spin_unlock(&cache->lock);
570                 wake_up(&caching_ctl->wait);
571         }
572
573         if (load_cache_only) {
574                 put_caching_control(caching_ctl);
575                 return 0;
576         }
577
578         down_write(&fs_info->extent_commit_sem);
579         atomic_inc(&caching_ctl->count);
580         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
581         up_write(&fs_info->extent_commit_sem);
582
583         btrfs_get_block_group(cache);
584
585         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
586
587         return ret;
588 }
589
590 /*
591  * return the block group that starts at or after bytenr
592  */
593 static struct btrfs_block_group_cache *
594 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
595 {
596         struct btrfs_block_group_cache *cache;
597
598         cache = block_group_cache_tree_search(info, bytenr, 0);
599
600         return cache;
601 }
602
603 /*
604  * return the block group that contains the given bytenr
605  */
606 struct btrfs_block_group_cache *btrfs_lookup_block_group(
607                                                  struct btrfs_fs_info *info,
608                                                  u64 bytenr)
609 {
610         struct btrfs_block_group_cache *cache;
611
612         cache = block_group_cache_tree_search(info, bytenr, 1);
613
614         return cache;
615 }
616
617 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
618                                                   u64 flags)
619 {
620         struct list_head *head = &info->space_info;
621         struct btrfs_space_info *found;
622
623         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
624
625         rcu_read_lock();
626         list_for_each_entry_rcu(found, head, list) {
627                 if (found->flags & flags) {
628                         rcu_read_unlock();
629                         return found;
630                 }
631         }
632         rcu_read_unlock();
633         return NULL;
634 }
635
636 /*
637  * after adding space to the filesystem, we need to clear the full flags
638  * on all the space infos.
639  */
640 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
641 {
642         struct list_head *head = &info->space_info;
643         struct btrfs_space_info *found;
644
645         rcu_read_lock();
646         list_for_each_entry_rcu(found, head, list)
647                 found->full = 0;
648         rcu_read_unlock();
649 }
650
651 static u64 div_factor(u64 num, int factor)
652 {
653         if (factor == 10)
654                 return num;
655         num *= factor;
656         do_div(num, 10);
657         return num;
658 }
659
660 static u64 div_factor_fine(u64 num, int factor)
661 {
662         if (factor == 100)
663                 return num;
664         num *= factor;
665         do_div(num, 100);
666         return num;
667 }
668
669 u64 btrfs_find_block_group(struct btrfs_root *root,
670                            u64 search_start, u64 search_hint, int owner)
671 {
672         struct btrfs_block_group_cache *cache;
673         u64 used;
674         u64 last = max(search_hint, search_start);
675         u64 group_start = 0;
676         int full_search = 0;
677         int factor = 9;
678         int wrapped = 0;
679 again:
680         while (1) {
681                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
682                 if (!cache)
683                         break;
684
685                 spin_lock(&cache->lock);
686                 last = cache->key.objectid + cache->key.offset;
687                 used = btrfs_block_group_used(&cache->item);
688
689                 if ((full_search || !cache->ro) &&
690                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
691                         if (used + cache->pinned + cache->reserved <
692                             div_factor(cache->key.offset, factor)) {
693                                 group_start = cache->key.objectid;
694                                 spin_unlock(&cache->lock);
695                                 btrfs_put_block_group(cache);
696                                 goto found;
697                         }
698                 }
699                 spin_unlock(&cache->lock);
700                 btrfs_put_block_group(cache);
701                 cond_resched();
702         }
703         if (!wrapped) {
704                 last = search_start;
705                 wrapped = 1;
706                 goto again;
707         }
708         if (!full_search && factor < 10) {
709                 last = search_start;
710                 full_search = 1;
711                 factor = 10;
712                 goto again;
713         }
714 found:
715         return group_start;
716 }
717
718 /* simple helper to search for an existing extent at a given offset */
719 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
720 {
721         int ret;
722         struct btrfs_key key;
723         struct btrfs_path *path;
724
725         path = btrfs_alloc_path();
726         if (!path)
727                 return -ENOMEM;
728
729         key.objectid = start;
730         key.offset = len;
731         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
732         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
733                                 0, 0);
734         btrfs_free_path(path);
735         return ret;
736 }
737
738 /*
739  * helper function to lookup reference count and flags of extent.
740  *
741  * the head node for delayed ref is used to store the sum of all the
742  * reference count modifications queued up in the rbtree. the head
743  * node may also store the extent flags to set. This way you can check
744  * to see what the reference count and extent flags would be if all of
745  * the delayed refs are not processed.
746  */
747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
748                              struct btrfs_root *root, u64 bytenr,
749                              u64 num_bytes, u64 *refs, u64 *flags)
750 {
751         struct btrfs_delayed_ref_head *head;
752         struct btrfs_delayed_ref_root *delayed_refs;
753         struct btrfs_path *path;
754         struct btrfs_extent_item *ei;
755         struct extent_buffer *leaf;
756         struct btrfs_key key;
757         u32 item_size;
758         u64 num_refs;
759         u64 extent_flags;
760         int ret;
761
762         path = btrfs_alloc_path();
763         if (!path)
764                 return -ENOMEM;
765
766         key.objectid = bytenr;
767         key.type = BTRFS_EXTENT_ITEM_KEY;
768         key.offset = num_bytes;
769         if (!trans) {
770                 path->skip_locking = 1;
771                 path->search_commit_root = 1;
772         }
773 again:
774         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
775                                 &key, path, 0, 0);
776         if (ret < 0)
777                 goto out_free;
778
779         if (ret == 0) {
780                 leaf = path->nodes[0];
781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
782                 if (item_size >= sizeof(*ei)) {
783                         ei = btrfs_item_ptr(leaf, path->slots[0],
784                                             struct btrfs_extent_item);
785                         num_refs = btrfs_extent_refs(leaf, ei);
786                         extent_flags = btrfs_extent_flags(leaf, ei);
787                 } else {
788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
789                         struct btrfs_extent_item_v0 *ei0;
790                         BUG_ON(item_size != sizeof(*ei0));
791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
792                                              struct btrfs_extent_item_v0);
793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
794                         /* FIXME: this isn't correct for data */
795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
796 #else
797                         BUG();
798 #endif
799                 }
800                 BUG_ON(num_refs == 0);
801         } else {
802                 num_refs = 0;
803                 extent_flags = 0;
804                 ret = 0;
805         }
806
807         if (!trans)
808                 goto out;
809
810         delayed_refs = &trans->transaction->delayed_refs;
811         spin_lock(&delayed_refs->lock);
812         head = btrfs_find_delayed_ref_head(trans, bytenr);
813         if (head) {
814                 if (!mutex_trylock(&head->mutex)) {
815                         atomic_inc(&head->node.refs);
816                         spin_unlock(&delayed_refs->lock);
817
818                         btrfs_release_path(path);
819
820                         /*
821                          * Mutex was contended, block until it's released and try
822                          * again
823                          */
824                         mutex_lock(&head->mutex);
825                         mutex_unlock(&head->mutex);
826                         btrfs_put_delayed_ref(&head->node);
827                         goto again;
828                 }
829                 if (head->extent_op && head->extent_op->update_flags)
830                         extent_flags |= head->extent_op->flags_to_set;
831                 else
832                         BUG_ON(num_refs == 0);
833
834                 num_refs += head->node.ref_mod;
835                 mutex_unlock(&head->mutex);
836         }
837         spin_unlock(&delayed_refs->lock);
838 out:
839         WARN_ON(num_refs == 0);
840         if (refs)
841                 *refs = num_refs;
842         if (flags)
843                 *flags = extent_flags;
844 out_free:
845         btrfs_free_path(path);
846         return ret;
847 }
848
849 /*
850  * Back reference rules.  Back refs have three main goals:
851  *
852  * 1) differentiate between all holders of references to an extent so that
853  *    when a reference is dropped we can make sure it was a valid reference
854  *    before freeing the extent.
855  *
856  * 2) Provide enough information to quickly find the holders of an extent
857  *    if we notice a given block is corrupted or bad.
858  *
859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
860  *    maintenance.  This is actually the same as #2, but with a slightly
861  *    different use case.
862  *
863  * There are two kinds of back refs. The implicit back refs is optimized
864  * for pointers in non-shared tree blocks. For a given pointer in a block,
865  * back refs of this kind provide information about the block's owner tree
866  * and the pointer's key. These information allow us to find the block by
867  * b-tree searching. The full back refs is for pointers in tree blocks not
868  * referenced by their owner trees. The location of tree block is recorded
869  * in the back refs. Actually the full back refs is generic, and can be
870  * used in all cases the implicit back refs is used. The major shortcoming
871  * of the full back refs is its overhead. Every time a tree block gets
872  * COWed, we have to update back refs entry for all pointers in it.
873  *
874  * For a newly allocated tree block, we use implicit back refs for
875  * pointers in it. This means most tree related operations only involve
876  * implicit back refs. For a tree block created in old transaction, the
877  * only way to drop a reference to it is COW it. So we can detect the
878  * event that tree block loses its owner tree's reference and do the
879  * back refs conversion.
880  *
881  * When a tree block is COW'd through a tree, there are four cases:
882  *
883  * The reference count of the block is one and the tree is the block's
884  * owner tree. Nothing to do in this case.
885  *
886  * The reference count of the block is one and the tree is not the
887  * block's owner tree. In this case, full back refs is used for pointers
888  * in the block. Remove these full back refs, add implicit back refs for
889  * every pointers in the new block.
890  *
891  * The reference count of the block is greater than one and the tree is
892  * the block's owner tree. In this case, implicit back refs is used for
893  * pointers in the block. Add full back refs for every pointers in the
894  * block, increase lower level extents' reference counts. The original
895  * implicit back refs are entailed to the new block.
896  *
897  * The reference count of the block is greater than one and the tree is
898  * not the block's owner tree. Add implicit back refs for every pointer in
899  * the new block, increase lower level extents' reference count.
900  *
901  * Back Reference Key composing:
902  *
903  * The key objectid corresponds to the first byte in the extent,
904  * The key type is used to differentiate between types of back refs.
905  * There are different meanings of the key offset for different types
906  * of back refs.
907  *
908  * File extents can be referenced by:
909  *
910  * - multiple snapshots, subvolumes, or different generations in one subvol
911  * - different files inside a single subvolume
912  * - different offsets inside a file (bookend extents in file.c)
913  *
914  * The extent ref structure for the implicit back refs has fields for:
915  *
916  * - Objectid of the subvolume root
917  * - objectid of the file holding the reference
918  * - original offset in the file
919  * - how many bookend extents
920  *
921  * The key offset for the implicit back refs is hash of the first
922  * three fields.
923  *
924  * The extent ref structure for the full back refs has field for:
925  *
926  * - number of pointers in the tree leaf
927  *
928  * The key offset for the implicit back refs is the first byte of
929  * the tree leaf
930  *
931  * When a file extent is allocated, The implicit back refs is used.
932  * the fields are filled in:
933  *
934  *     (root_key.objectid, inode objectid, offset in file, 1)
935  *
936  * When a file extent is removed file truncation, we find the
937  * corresponding implicit back refs and check the following fields:
938  *
939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
940  *
941  * Btree extents can be referenced by:
942  *
943  * - Different subvolumes
944  *
945  * Both the implicit back refs and the full back refs for tree blocks
946  * only consist of key. The key offset for the implicit back refs is
947  * objectid of block's owner tree. The key offset for the full back refs
948  * is the first byte of parent block.
949  *
950  * When implicit back refs is used, information about the lowest key and
951  * level of the tree block are required. These information are stored in
952  * tree block info structure.
953  */
954
955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
957                                   struct btrfs_root *root,
958                                   struct btrfs_path *path,
959                                   u64 owner, u32 extra_size)
960 {
961         struct btrfs_extent_item *item;
962         struct btrfs_extent_item_v0 *ei0;
963         struct btrfs_extent_ref_v0 *ref0;
964         struct btrfs_tree_block_info *bi;
965         struct extent_buffer *leaf;
966         struct btrfs_key key;
967         struct btrfs_key found_key;
968         u32 new_size = sizeof(*item);
969         u64 refs;
970         int ret;
971
972         leaf = path->nodes[0];
973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
974
975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
977                              struct btrfs_extent_item_v0);
978         refs = btrfs_extent_refs_v0(leaf, ei0);
979
980         if (owner == (u64)-1) {
981                 while (1) {
982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
983                                 ret = btrfs_next_leaf(root, path);
984                                 if (ret < 0)
985                                         return ret;
986                                 BUG_ON(ret > 0); /* Corruption */
987                                 leaf = path->nodes[0];
988                         }
989                         btrfs_item_key_to_cpu(leaf, &found_key,
990                                               path->slots[0]);
991                         BUG_ON(key.objectid != found_key.objectid);
992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
993                                 path->slots[0]++;
994                                 continue;
995                         }
996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
997                                               struct btrfs_extent_ref_v0);
998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(trans, root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466
1467         key.objectid = bytenr;
1468         key.type = BTRFS_EXTENT_ITEM_KEY;
1469         key.offset = num_bytes;
1470
1471         want = extent_ref_type(parent, owner);
1472         if (insert) {
1473                 extra_size = btrfs_extent_inline_ref_size(want);
1474                 path->keep_locks = 1;
1475         } else
1476                 extra_size = -1;
1477         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1478         if (ret < 0) {
1479                 err = ret;
1480                 goto out;
1481         }
1482         if (ret && !insert) {
1483                 err = -ENOENT;
1484                 goto out;
1485         }
1486         BUG_ON(ret); /* Corruption */
1487
1488         leaf = path->nodes[0];
1489         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1490 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1491         if (item_size < sizeof(*ei)) {
1492                 if (!insert) {
1493                         err = -ENOENT;
1494                         goto out;
1495                 }
1496                 ret = convert_extent_item_v0(trans, root, path, owner,
1497                                              extra_size);
1498                 if (ret < 0) {
1499                         err = ret;
1500                         goto out;
1501                 }
1502                 leaf = path->nodes[0];
1503                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1504         }
1505 #endif
1506         BUG_ON(item_size < sizeof(*ei));
1507
1508         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1509         flags = btrfs_extent_flags(leaf, ei);
1510
1511         ptr = (unsigned long)(ei + 1);
1512         end = (unsigned long)ei + item_size;
1513
1514         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1515                 ptr += sizeof(struct btrfs_tree_block_info);
1516                 BUG_ON(ptr > end);
1517         } else {
1518                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1519         }
1520
1521         err = -ENOENT;
1522         while (1) {
1523                 if (ptr >= end) {
1524                         WARN_ON(ptr > end);
1525                         break;
1526                 }
1527                 iref = (struct btrfs_extent_inline_ref *)ptr;
1528                 type = btrfs_extent_inline_ref_type(leaf, iref);
1529                 if (want < type)
1530                         break;
1531                 if (want > type) {
1532                         ptr += btrfs_extent_inline_ref_size(type);
1533                         continue;
1534                 }
1535
1536                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1537                         struct btrfs_extent_data_ref *dref;
1538                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1539                         if (match_extent_data_ref(leaf, dref, root_objectid,
1540                                                   owner, offset)) {
1541                                 err = 0;
1542                                 break;
1543                         }
1544                         if (hash_extent_data_ref_item(leaf, dref) <
1545                             hash_extent_data_ref(root_objectid, owner, offset))
1546                                 break;
1547                 } else {
1548                         u64 ref_offset;
1549                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1550                         if (parent > 0) {
1551                                 if (parent == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < parent)
1556                                         break;
1557                         } else {
1558                                 if (root_objectid == ref_offset) {
1559                                         err = 0;
1560                                         break;
1561                                 }
1562                                 if (ref_offset < root_objectid)
1563                                         break;
1564                         }
1565                 }
1566                 ptr += btrfs_extent_inline_ref_size(type);
1567         }
1568         if (err == -ENOENT && insert) {
1569                 if (item_size + extra_size >=
1570                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574                 /*
1575                  * To add new inline back ref, we have to make sure
1576                  * there is no corresponding back ref item.
1577                  * For simplicity, we just do not add new inline back
1578                  * ref if there is any kind of item for this block
1579                  */
1580                 if (find_next_key(path, 0, &key) == 0 &&
1581                     key.objectid == bytenr &&
1582                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1583                         err = -EAGAIN;
1584                         goto out;
1585                 }
1586         }
1587         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1588 out:
1589         if (insert) {
1590                 path->keep_locks = 0;
1591                 btrfs_unlock_up_safe(path, 1);
1592         }
1593         return err;
1594 }
1595
1596 /*
1597  * helper to add new inline back ref
1598  */
1599 static noinline_for_stack
1600 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_root *root,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref *iref,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int refs_to_add,
1606                                  struct btrfs_delayed_extent_op *extent_op)
1607 {
1608         struct extent_buffer *leaf;
1609         struct btrfs_extent_item *ei;
1610         unsigned long ptr;
1611         unsigned long end;
1612         unsigned long item_offset;
1613         u64 refs;
1614         int size;
1615         int type;
1616
1617         leaf = path->nodes[0];
1618         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1619         item_offset = (unsigned long)iref - (unsigned long)ei;
1620
1621         type = extent_ref_type(parent, owner);
1622         size = btrfs_extent_inline_ref_size(type);
1623
1624         btrfs_extend_item(trans, root, path, size);
1625
1626         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1627         refs = btrfs_extent_refs(leaf, ei);
1628         refs += refs_to_add;
1629         btrfs_set_extent_refs(leaf, ei, refs);
1630         if (extent_op)
1631                 __run_delayed_extent_op(extent_op, leaf, ei);
1632
1633         ptr = (unsigned long)ei + item_offset;
1634         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1635         if (ptr < end - size)
1636                 memmove_extent_buffer(leaf, ptr + size, ptr,
1637                                       end - size - ptr);
1638
1639         iref = (struct btrfs_extent_inline_ref *)ptr;
1640         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1641         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1642                 struct btrfs_extent_data_ref *dref;
1643                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1644                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1645                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1646                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1647                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1648         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1649                 struct btrfs_shared_data_ref *sref;
1650                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1651                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1655         } else {
1656                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1657         }
1658         btrfs_mark_buffer_dirty(leaf);
1659 }
1660
1661 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1662                                  struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref **ref_ret,
1665                                  u64 bytenr, u64 num_bytes, u64 parent,
1666                                  u64 root_objectid, u64 owner, u64 offset)
1667 {
1668         int ret;
1669
1670         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1671                                            bytenr, num_bytes, parent,
1672                                            root_objectid, owner, offset, 0);
1673         if (ret != -ENOENT)
1674                 return ret;
1675
1676         btrfs_release_path(path);
1677         *ref_ret = NULL;
1678
1679         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1680                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1681                                             root_objectid);
1682         } else {
1683                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1684                                              root_objectid, owner, offset);
1685         }
1686         return ret;
1687 }
1688
1689 /*
1690  * helper to update/remove inline back ref
1691  */
1692 static noinline_for_stack
1693 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1694                                   struct btrfs_root *root,
1695                                   struct btrfs_path *path,
1696                                   struct btrfs_extent_inline_ref *iref,
1697                                   int refs_to_mod,
1698                                   struct btrfs_delayed_extent_op *extent_op)
1699 {
1700         struct extent_buffer *leaf;
1701         struct btrfs_extent_item *ei;
1702         struct btrfs_extent_data_ref *dref = NULL;
1703         struct btrfs_shared_data_ref *sref = NULL;
1704         unsigned long ptr;
1705         unsigned long end;
1706         u32 item_size;
1707         int size;
1708         int type;
1709         u64 refs;
1710
1711         leaf = path->nodes[0];
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1715         refs += refs_to_mod;
1716         btrfs_set_extent_refs(leaf, ei, refs);
1717         if (extent_op)
1718                 __run_delayed_extent_op(extent_op, leaf, ei);
1719
1720         type = btrfs_extent_inline_ref_type(leaf, iref);
1721
1722         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1723                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1724                 refs = btrfs_extent_data_ref_count(leaf, dref);
1725         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1726                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1727                 refs = btrfs_shared_data_ref_count(leaf, sref);
1728         } else {
1729                 refs = 1;
1730                 BUG_ON(refs_to_mod != -1);
1731         }
1732
1733         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1734         refs += refs_to_mod;
1735
1736         if (refs > 0) {
1737                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1738                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1739                 else
1740                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1741         } else {
1742                 size =  btrfs_extent_inline_ref_size(type);
1743                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1744                 ptr = (unsigned long)iref;
1745                 end = (unsigned long)ei + item_size;
1746                 if (ptr + size < end)
1747                         memmove_extent_buffer(leaf, ptr, ptr + size,
1748                                               end - ptr - size);
1749                 item_size -= size;
1750                 btrfs_truncate_item(trans, root, path, item_size, 1);
1751         }
1752         btrfs_mark_buffer_dirty(leaf);
1753 }
1754
1755 static noinline_for_stack
1756 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1757                                  struct btrfs_root *root,
1758                                  struct btrfs_path *path,
1759                                  u64 bytenr, u64 num_bytes, u64 parent,
1760                                  u64 root_objectid, u64 owner,
1761                                  u64 offset, int refs_to_add,
1762                                  struct btrfs_delayed_extent_op *extent_op)
1763 {
1764         struct btrfs_extent_inline_ref *iref;
1765         int ret;
1766
1767         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1768                                            bytenr, num_bytes, parent,
1769                                            root_objectid, owner, offset, 1);
1770         if (ret == 0) {
1771                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1772                 update_inline_extent_backref(trans, root, path, iref,
1773                                              refs_to_add, extent_op);
1774         } else if (ret == -ENOENT) {
1775                 setup_inline_extent_backref(trans, root, path, iref, parent,
1776                                             root_objectid, owner, offset,
1777                                             refs_to_add, extent_op);
1778                 ret = 0;
1779         }
1780         return ret;
1781 }
1782
1783 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1784                                  struct btrfs_root *root,
1785                                  struct btrfs_path *path,
1786                                  u64 bytenr, u64 parent, u64 root_objectid,
1787                                  u64 owner, u64 offset, int refs_to_add)
1788 {
1789         int ret;
1790         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1791                 BUG_ON(refs_to_add != 1);
1792                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1793                                             parent, root_objectid);
1794         } else {
1795                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1796                                              parent, root_objectid,
1797                                              owner, offset, refs_to_add);
1798         }
1799         return ret;
1800 }
1801
1802 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  int refs_to_drop, int is_data)
1807 {
1808         int ret = 0;
1809
1810         BUG_ON(!is_data && refs_to_drop != 1);
1811         if (iref) {
1812                 update_inline_extent_backref(trans, root, path, iref,
1813                                              -refs_to_drop, NULL);
1814         } else if (is_data) {
1815                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1816         } else {
1817                 ret = btrfs_del_item(trans, root, path);
1818         }
1819         return ret;
1820 }
1821
1822 static int btrfs_issue_discard(struct block_device *bdev,
1823                                 u64 start, u64 len)
1824 {
1825         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1826 }
1827
1828 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1829                                 u64 num_bytes, u64 *actual_bytes)
1830 {
1831         int ret;
1832         u64 discarded_bytes = 0;
1833         struct btrfs_bio *bbio = NULL;
1834
1835
1836         /* Tell the block device(s) that the sectors can be discarded */
1837         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1838                               bytenr, &num_bytes, &bbio, 0);
1839         /* Error condition is -ENOMEM */
1840         if (!ret) {
1841                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1842                 int i;
1843
1844
1845                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1846                         if (!stripe->dev->can_discard)
1847                                 continue;
1848
1849                         ret = btrfs_issue_discard(stripe->dev->bdev,
1850                                                   stripe->physical,
1851                                                   stripe->length);
1852                         if (!ret)
1853                                 discarded_bytes += stripe->length;
1854                         else if (ret != -EOPNOTSUPP)
1855                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1856
1857                         /*
1858                          * Just in case we get back EOPNOTSUPP for some reason,
1859                          * just ignore the return value so we don't screw up
1860                          * people calling discard_extent.
1861                          */
1862                         ret = 0;
1863                 }
1864                 kfree(bbio);
1865         }
1866
1867         if (actual_bytes)
1868                 *actual_bytes = discarded_bytes;
1869
1870
1871         return ret;
1872 }
1873
1874 /* Can return -ENOMEM */
1875 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          u64 bytenr, u64 num_bytes, u64 parent,
1878                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1879 {
1880         int ret;
1881         struct btrfs_fs_info *fs_info = root->fs_info;
1882
1883         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1884                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1885
1886         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, (int)owner,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         } else {
1892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1893                                         num_bytes,
1894                                         parent, root_objectid, owner, offset,
1895                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1896         }
1897         return ret;
1898 }
1899
1900 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1901                                   struct btrfs_root *root,
1902                                   u64 bytenr, u64 num_bytes,
1903                                   u64 parent, u64 root_objectid,
1904                                   u64 owner, u64 offset, int refs_to_add,
1905                                   struct btrfs_delayed_extent_op *extent_op)
1906 {
1907         struct btrfs_path *path;
1908         struct extent_buffer *leaf;
1909         struct btrfs_extent_item *item;
1910         u64 refs;
1911         int ret;
1912         int err = 0;
1913
1914         path = btrfs_alloc_path();
1915         if (!path)
1916                 return -ENOMEM;
1917
1918         path->reada = 1;
1919         path->leave_spinning = 1;
1920         /* this will setup the path even if it fails to insert the back ref */
1921         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1922                                            path, bytenr, num_bytes, parent,
1923                                            root_objectid, owner, offset,
1924                                            refs_to_add, extent_op);
1925         if (ret == 0)
1926                 goto out;
1927
1928         if (ret != -EAGAIN) {
1929                 err = ret;
1930                 goto out;
1931         }
1932
1933         leaf = path->nodes[0];
1934         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1935         refs = btrfs_extent_refs(leaf, item);
1936         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1937         if (extent_op)
1938                 __run_delayed_extent_op(extent_op, leaf, item);
1939
1940         btrfs_mark_buffer_dirty(leaf);
1941         btrfs_release_path(path);
1942
1943         path->reada = 1;
1944         path->leave_spinning = 1;
1945
1946         /* now insert the actual backref */
1947         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1948                                     path, bytenr, parent, root_objectid,
1949                                     owner, offset, refs_to_add);
1950         if (ret)
1951                 btrfs_abort_transaction(trans, root, ret);
1952 out:
1953         btrfs_free_path(path);
1954         return err;
1955 }
1956
1957 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1958                                 struct btrfs_root *root,
1959                                 struct btrfs_delayed_ref_node *node,
1960                                 struct btrfs_delayed_extent_op *extent_op,
1961                                 int insert_reserved)
1962 {
1963         int ret = 0;
1964         struct btrfs_delayed_data_ref *ref;
1965         struct btrfs_key ins;
1966         u64 parent = 0;
1967         u64 ref_root = 0;
1968         u64 flags = 0;
1969
1970         ins.objectid = node->bytenr;
1971         ins.offset = node->num_bytes;
1972         ins.type = BTRFS_EXTENT_ITEM_KEY;
1973
1974         ref = btrfs_delayed_node_to_data_ref(node);
1975         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1976                 parent = ref->parent;
1977         else
1978                 ref_root = ref->root;
1979
1980         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1981                 if (extent_op) {
1982                         BUG_ON(extent_op->update_key);
1983                         flags |= extent_op->flags_to_set;
1984                 }
1985                 ret = alloc_reserved_file_extent(trans, root,
1986                                                  parent, ref_root, flags,
1987                                                  ref->objectid, ref->offset,
1988                                                  &ins, node->ref_mod);
1989         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1990                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1991                                              node->num_bytes, parent,
1992                                              ref_root, ref->objectid,
1993                                              ref->offset, node->ref_mod,
1994                                              extent_op);
1995         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1996                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1997                                           node->num_bytes, parent,
1998                                           ref_root, ref->objectid,
1999                                           ref->offset, node->ref_mod,
2000                                           extent_op);
2001         } else {
2002                 BUG();
2003         }
2004         return ret;
2005 }
2006
2007 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2008                                     struct extent_buffer *leaf,
2009                                     struct btrfs_extent_item *ei)
2010 {
2011         u64 flags = btrfs_extent_flags(leaf, ei);
2012         if (extent_op->update_flags) {
2013                 flags |= extent_op->flags_to_set;
2014                 btrfs_set_extent_flags(leaf, ei, flags);
2015         }
2016
2017         if (extent_op->update_key) {
2018                 struct btrfs_tree_block_info *bi;
2019                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2020                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2021                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2022         }
2023 }
2024
2025 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2026                                  struct btrfs_root *root,
2027                                  struct btrfs_delayed_ref_node *node,
2028                                  struct btrfs_delayed_extent_op *extent_op)
2029 {
2030         struct btrfs_key key;
2031         struct btrfs_path *path;
2032         struct btrfs_extent_item *ei;
2033         struct extent_buffer *leaf;
2034         u32 item_size;
2035         int ret;
2036         int err = 0;
2037
2038         if (trans->aborted)
2039                 return 0;
2040
2041         path = btrfs_alloc_path();
2042         if (!path)
2043                 return -ENOMEM;
2044
2045         key.objectid = node->bytenr;
2046         key.type = BTRFS_EXTENT_ITEM_KEY;
2047         key.offset = node->num_bytes;
2048
2049         path->reada = 1;
2050         path->leave_spinning = 1;
2051         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2052                                 path, 0, 1);
2053         if (ret < 0) {
2054                 err = ret;
2055                 goto out;
2056         }
2057         if (ret > 0) {
2058                 err = -EIO;
2059                 goto out;
2060         }
2061
2062         leaf = path->nodes[0];
2063         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2065         if (item_size < sizeof(*ei)) {
2066                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2067                                              path, (u64)-1, 0);
2068                 if (ret < 0) {
2069                         err = ret;
2070                         goto out;
2071                 }
2072                 leaf = path->nodes[0];
2073                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2074         }
2075 #endif
2076         BUG_ON(item_size < sizeof(*ei));
2077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         __run_delayed_extent_op(extent_op, leaf, ei);
2079
2080         btrfs_mark_buffer_dirty(leaf);
2081 out:
2082         btrfs_free_path(path);
2083         return err;
2084 }
2085
2086 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_root *root,
2088                                 struct btrfs_delayed_ref_node *node,
2089                                 struct btrfs_delayed_extent_op *extent_op,
2090                                 int insert_reserved)
2091 {
2092         int ret = 0;
2093         struct btrfs_delayed_tree_ref *ref;
2094         struct btrfs_key ins;
2095         u64 parent = 0;
2096         u64 ref_root = 0;
2097
2098         ins.objectid = node->bytenr;
2099         ins.offset = node->num_bytes;
2100         ins.type = BTRFS_EXTENT_ITEM_KEY;
2101
2102         ref = btrfs_delayed_node_to_tree_ref(node);
2103         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2104                 parent = ref->parent;
2105         else
2106                 ref_root = ref->root;
2107
2108         BUG_ON(node->ref_mod != 1);
2109         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2110                 BUG_ON(!extent_op || !extent_op->update_flags ||
2111                        !extent_op->update_key);
2112                 ret = alloc_reserved_tree_block(trans, root,
2113                                                 parent, ref_root,
2114                                                 extent_op->flags_to_set,
2115                                                 &extent_op->key,
2116                                                 ref->level, &ins);
2117         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2118                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2119                                              node->num_bytes, parent, ref_root,
2120                                              ref->level, 0, 1, extent_op);
2121         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2122                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2123                                           node->num_bytes, parent, ref_root,
2124                                           ref->level, 0, 1, extent_op);
2125         } else {
2126                 BUG();
2127         }
2128         return ret;
2129 }
2130
2131 /* helper function to actually process a single delayed ref entry */
2132 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2133                                struct btrfs_root *root,
2134                                struct btrfs_delayed_ref_node *node,
2135                                struct btrfs_delayed_extent_op *extent_op,
2136                                int insert_reserved)
2137 {
2138         int ret = 0;
2139
2140         if (trans->aborted)
2141                 return 0;
2142
2143         if (btrfs_delayed_ref_is_head(node)) {
2144                 struct btrfs_delayed_ref_head *head;
2145                 /*
2146                  * we've hit the end of the chain and we were supposed
2147                  * to insert this extent into the tree.  But, it got
2148                  * deleted before we ever needed to insert it, so all
2149                  * we have to do is clean up the accounting
2150                  */
2151                 BUG_ON(extent_op);
2152                 head = btrfs_delayed_node_to_head(node);
2153                 if (insert_reserved) {
2154                         btrfs_pin_extent(root, node->bytenr,
2155                                          node->num_bytes, 1);
2156                         if (head->is_data) {
2157                                 ret = btrfs_del_csums(trans, root,
2158                                                       node->bytenr,
2159                                                       node->num_bytes);
2160                         }
2161                 }
2162                 mutex_unlock(&head->mutex);
2163                 return ret;
2164         }
2165
2166         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2167             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2168                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2169                                            insert_reserved);
2170         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2171                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2172                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2173                                            insert_reserved);
2174         else
2175                 BUG();
2176         return ret;
2177 }
2178
2179 static noinline struct btrfs_delayed_ref_node *
2180 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2181 {
2182         struct rb_node *node;
2183         struct btrfs_delayed_ref_node *ref;
2184         int action = BTRFS_ADD_DELAYED_REF;
2185 again:
2186         /*
2187          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2188          * this prevents ref count from going down to zero when
2189          * there still are pending delayed ref.
2190          */
2191         node = rb_prev(&head->node.rb_node);
2192         while (1) {
2193                 if (!node)
2194                         break;
2195                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2196                                 rb_node);
2197                 if (ref->bytenr != head->node.bytenr)
2198                         break;
2199                 if (ref->action == action)
2200                         return ref;
2201                 node = rb_prev(node);
2202         }
2203         if (action == BTRFS_ADD_DELAYED_REF) {
2204                 action = BTRFS_DROP_DELAYED_REF;
2205                 goto again;
2206         }
2207         return NULL;
2208 }
2209
2210 /*
2211  * Returns 0 on success or if called with an already aborted transaction.
2212  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2213  */
2214 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2215                                        struct btrfs_root *root,
2216                                        struct list_head *cluster)
2217 {
2218         struct btrfs_delayed_ref_root *delayed_refs;
2219         struct btrfs_delayed_ref_node *ref;
2220         struct btrfs_delayed_ref_head *locked_ref = NULL;
2221         struct btrfs_delayed_extent_op *extent_op;
2222         struct btrfs_fs_info *fs_info = root->fs_info;
2223         int ret;
2224         int count = 0;
2225         int must_insert_reserved = 0;
2226
2227         delayed_refs = &trans->transaction->delayed_refs;
2228         while (1) {
2229                 if (!locked_ref) {
2230                         /* pick a new head ref from the cluster list */
2231                         if (list_empty(cluster))
2232                                 break;
2233
2234                         locked_ref = list_entry(cluster->next,
2235                                      struct btrfs_delayed_ref_head, cluster);
2236
2237                         /* grab the lock that says we are going to process
2238                          * all the refs for this head */
2239                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2240
2241                         /*
2242                          * we may have dropped the spin lock to get the head
2243                          * mutex lock, and that might have given someone else
2244                          * time to free the head.  If that's true, it has been
2245                          * removed from our list and we can move on.
2246                          */
2247                         if (ret == -EAGAIN) {
2248                                 locked_ref = NULL;
2249                                 count++;
2250                                 continue;
2251                         }
2252                 }
2253
2254                 /*
2255                  * We need to try and merge add/drops of the same ref since we
2256                  * can run into issues with relocate dropping the implicit ref
2257                  * and then it being added back again before the drop can
2258                  * finish.  If we merged anything we need to re-loop so we can
2259                  * get a good ref.
2260                  */
2261                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2262                                          locked_ref);
2263
2264                 /*
2265                  * locked_ref is the head node, so we have to go one
2266                  * node back for any delayed ref updates
2267                  */
2268                 ref = select_delayed_ref(locked_ref);
2269
2270                 if (ref && ref->seq &&
2271                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2272                         /*
2273                          * there are still refs with lower seq numbers in the
2274                          * process of being added. Don't run this ref yet.
2275                          */
2276                         list_del_init(&locked_ref->cluster);
2277                         mutex_unlock(&locked_ref->mutex);
2278                         locked_ref = NULL;
2279                         delayed_refs->num_heads_ready++;
2280                         spin_unlock(&delayed_refs->lock);
2281                         cond_resched();
2282                         spin_lock(&delayed_refs->lock);
2283                         continue;
2284                 }
2285
2286                 /*
2287                  * record the must insert reserved flag before we
2288                  * drop the spin lock.
2289                  */
2290                 must_insert_reserved = locked_ref->must_insert_reserved;
2291                 locked_ref->must_insert_reserved = 0;
2292
2293                 extent_op = locked_ref->extent_op;
2294                 locked_ref->extent_op = NULL;
2295
2296                 if (!ref) {
2297                         /* All delayed refs have been processed, Go ahead
2298                          * and send the head node to run_one_delayed_ref,
2299                          * so that any accounting fixes can happen
2300                          */
2301                         ref = &locked_ref->node;
2302
2303                         if (extent_op && must_insert_reserved) {
2304                                 kfree(extent_op);
2305                                 extent_op = NULL;
2306                         }
2307
2308                         if (extent_op) {
2309                                 spin_unlock(&delayed_refs->lock);
2310
2311                                 ret = run_delayed_extent_op(trans, root,
2312                                                             ref, extent_op);
2313                                 kfree(extent_op);
2314
2315                                 if (ret) {
2316                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2317                                         spin_lock(&delayed_refs->lock);
2318                                         return ret;
2319                                 }
2320
2321                                 goto next;
2322                         }
2323
2324                         list_del_init(&locked_ref->cluster);
2325                         locked_ref = NULL;
2326                 }
2327
2328                 ref->in_tree = 0;
2329                 rb_erase(&ref->rb_node, &delayed_refs->root);
2330                 delayed_refs->num_entries--;
2331                 if (locked_ref) {
2332                         /*
2333                          * when we play the delayed ref, also correct the
2334                          * ref_mod on head
2335                          */
2336                         switch (ref->action) {
2337                         case BTRFS_ADD_DELAYED_REF:
2338                         case BTRFS_ADD_DELAYED_EXTENT:
2339                                 locked_ref->node.ref_mod -= ref->ref_mod;
2340                                 break;
2341                         case BTRFS_DROP_DELAYED_REF:
2342                                 locked_ref->node.ref_mod += ref->ref_mod;
2343                                 break;
2344                         default:
2345                                 WARN_ON(1);
2346                         }
2347                 }
2348                 spin_unlock(&delayed_refs->lock);
2349
2350                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2351                                           must_insert_reserved);
2352
2353                 btrfs_put_delayed_ref(ref);
2354                 kfree(extent_op);
2355                 count++;
2356
2357                 if (ret) {
2358                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2359                         spin_lock(&delayed_refs->lock);
2360                         return ret;
2361                 }
2362
2363 next:
2364                 cond_resched();
2365                 spin_lock(&delayed_refs->lock);
2366         }
2367         return count;
2368 }
2369
2370 #ifdef SCRAMBLE_DELAYED_REFS
2371 /*
2372  * Normally delayed refs get processed in ascending bytenr order. This
2373  * correlates in most cases to the order added. To expose dependencies on this
2374  * order, we start to process the tree in the middle instead of the beginning
2375  */
2376 static u64 find_middle(struct rb_root *root)
2377 {
2378         struct rb_node *n = root->rb_node;
2379         struct btrfs_delayed_ref_node *entry;
2380         int alt = 1;
2381         u64 middle;
2382         u64 first = 0, last = 0;
2383
2384         n = rb_first(root);
2385         if (n) {
2386                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2387                 first = entry->bytenr;
2388         }
2389         n = rb_last(root);
2390         if (n) {
2391                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2392                 last = entry->bytenr;
2393         }
2394         n = root->rb_node;
2395
2396         while (n) {
2397                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2398                 WARN_ON(!entry->in_tree);
2399
2400                 middle = entry->bytenr;
2401
2402                 if (alt)
2403                         n = n->rb_left;
2404                 else
2405                         n = n->rb_right;
2406
2407                 alt = 1 - alt;
2408         }
2409         return middle;
2410 }
2411 #endif
2412
2413 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2414                                          struct btrfs_fs_info *fs_info)
2415 {
2416         struct qgroup_update *qgroup_update;
2417         int ret = 0;
2418
2419         if (list_empty(&trans->qgroup_ref_list) !=
2420             !trans->delayed_ref_elem.seq) {
2421                 /* list without seq or seq without list */
2422                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2423                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2424                         trans->delayed_ref_elem.seq);
2425                 BUG();
2426         }
2427
2428         if (!trans->delayed_ref_elem.seq)
2429                 return 0;
2430
2431         while (!list_empty(&trans->qgroup_ref_list)) {
2432                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2433                                                  struct qgroup_update, list);
2434                 list_del(&qgroup_update->list);
2435                 if (!ret)
2436                         ret = btrfs_qgroup_account_ref(
2437                                         trans, fs_info, qgroup_update->node,
2438                                         qgroup_update->extent_op);
2439                 kfree(qgroup_update);
2440         }
2441
2442         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2443
2444         return ret;
2445 }
2446
2447 /*
2448  * this starts processing the delayed reference count updates and
2449  * extent insertions we have queued up so far.  count can be
2450  * 0, which means to process everything in the tree at the start
2451  * of the run (but not newly added entries), or it can be some target
2452  * number you'd like to process.
2453  *
2454  * Returns 0 on success or if called with an aborted transaction
2455  * Returns <0 on error and aborts the transaction
2456  */
2457 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2458                            struct btrfs_root *root, unsigned long count)
2459 {
2460         struct rb_node *node;
2461         struct btrfs_delayed_ref_root *delayed_refs;
2462         struct btrfs_delayed_ref_node *ref;
2463         struct list_head cluster;
2464         int ret;
2465         u64 delayed_start;
2466         int run_all = count == (unsigned long)-1;
2467         int run_most = 0;
2468         int loops;
2469
2470         /* We'll clean this up in btrfs_cleanup_transaction */
2471         if (trans->aborted)
2472                 return 0;
2473
2474         if (root == root->fs_info->extent_root)
2475                 root = root->fs_info->tree_root;
2476
2477         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2478
2479         delayed_refs = &trans->transaction->delayed_refs;
2480         INIT_LIST_HEAD(&cluster);
2481 again:
2482         loops = 0;
2483         spin_lock(&delayed_refs->lock);
2484
2485 #ifdef SCRAMBLE_DELAYED_REFS
2486         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2487 #endif
2488
2489         if (count == 0) {
2490                 count = delayed_refs->num_entries * 2;
2491                 run_most = 1;
2492         }
2493         while (1) {
2494                 if (!(run_all || run_most) &&
2495                     delayed_refs->num_heads_ready < 64)
2496                         break;
2497
2498                 /*
2499                  * go find something we can process in the rbtree.  We start at
2500                  * the beginning of the tree, and then build a cluster
2501                  * of refs to process starting at the first one we are able to
2502                  * lock
2503                  */
2504                 delayed_start = delayed_refs->run_delayed_start;
2505                 ret = btrfs_find_ref_cluster(trans, &cluster,
2506                                              delayed_refs->run_delayed_start);
2507                 if (ret)
2508                         break;
2509
2510                 ret = run_clustered_refs(trans, root, &cluster);
2511                 if (ret < 0) {
2512                         spin_unlock(&delayed_refs->lock);
2513                         btrfs_abort_transaction(trans, root, ret);
2514                         return ret;
2515                 }
2516
2517                 count -= min_t(unsigned long, ret, count);
2518
2519                 if (count == 0)
2520                         break;
2521
2522                 if (delayed_start >= delayed_refs->run_delayed_start) {
2523                         if (loops == 0) {
2524                                 /*
2525                                  * btrfs_find_ref_cluster looped. let's do one
2526                                  * more cycle. if we don't run any delayed ref
2527                                  * during that cycle (because we can't because
2528                                  * all of them are blocked), bail out.
2529                                  */
2530                                 loops = 1;
2531                         } else {
2532                                 /*
2533                                  * no runnable refs left, stop trying
2534                                  */
2535                                 BUG_ON(run_all);
2536                                 break;
2537                         }
2538                 }
2539                 if (ret) {
2540                         /* refs were run, let's reset staleness detection */
2541                         loops = 0;
2542                 }
2543         }
2544
2545         if (run_all) {
2546                 if (!list_empty(&trans->new_bgs)) {
2547                         spin_unlock(&delayed_refs->lock);
2548                         btrfs_create_pending_block_groups(trans, root);
2549                         spin_lock(&delayed_refs->lock);
2550                 }
2551
2552                 node = rb_first(&delayed_refs->root);
2553                 if (!node)
2554                         goto out;
2555                 count = (unsigned long)-1;
2556
2557                 while (node) {
2558                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2559                                        rb_node);
2560                         if (btrfs_delayed_ref_is_head(ref)) {
2561                                 struct btrfs_delayed_ref_head *head;
2562
2563                                 head = btrfs_delayed_node_to_head(ref);
2564                                 atomic_inc(&ref->refs);
2565
2566                                 spin_unlock(&delayed_refs->lock);
2567                                 /*
2568                                  * Mutex was contended, block until it's
2569                                  * released and try again
2570                                  */
2571                                 mutex_lock(&head->mutex);
2572                                 mutex_unlock(&head->mutex);
2573
2574                                 btrfs_put_delayed_ref(ref);
2575                                 cond_resched();
2576                                 goto again;
2577                         }
2578                         node = rb_next(node);
2579                 }
2580                 spin_unlock(&delayed_refs->lock);
2581                 schedule_timeout(1);
2582                 goto again;
2583         }
2584 out:
2585         spin_unlock(&delayed_refs->lock);
2586         assert_qgroups_uptodate(trans);
2587         return 0;
2588 }
2589
2590 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2591                                 struct btrfs_root *root,
2592                                 u64 bytenr, u64 num_bytes, u64 flags,
2593                                 int is_data)
2594 {
2595         struct btrfs_delayed_extent_op *extent_op;
2596         int ret;
2597
2598         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2599         if (!extent_op)
2600                 return -ENOMEM;
2601
2602         extent_op->flags_to_set = flags;
2603         extent_op->update_flags = 1;
2604         extent_op->update_key = 0;
2605         extent_op->is_data = is_data ? 1 : 0;
2606
2607         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2608                                           num_bytes, extent_op);
2609         if (ret)
2610                 kfree(extent_op);
2611         return ret;
2612 }
2613
2614 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2615                                       struct btrfs_root *root,
2616                                       struct btrfs_path *path,
2617                                       u64 objectid, u64 offset, u64 bytenr)
2618 {
2619         struct btrfs_delayed_ref_head *head;
2620         struct btrfs_delayed_ref_node *ref;
2621         struct btrfs_delayed_data_ref *data_ref;
2622         struct btrfs_delayed_ref_root *delayed_refs;
2623         struct rb_node *node;
2624         int ret = 0;
2625
2626         ret = -ENOENT;
2627         delayed_refs = &trans->transaction->delayed_refs;
2628         spin_lock(&delayed_refs->lock);
2629         head = btrfs_find_delayed_ref_head(trans, bytenr);
2630         if (!head)
2631                 goto out;
2632
2633         if (!mutex_trylock(&head->mutex)) {
2634                 atomic_inc(&head->node.refs);
2635                 spin_unlock(&delayed_refs->lock);
2636
2637                 btrfs_release_path(path);
2638
2639                 /*
2640                  * Mutex was contended, block until it's released and let
2641                  * caller try again
2642                  */
2643                 mutex_lock(&head->mutex);
2644                 mutex_unlock(&head->mutex);
2645                 btrfs_put_delayed_ref(&head->node);
2646                 return -EAGAIN;
2647         }
2648
2649         node = rb_prev(&head->node.rb_node);
2650         if (!node)
2651                 goto out_unlock;
2652
2653         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2654
2655         if (ref->bytenr != bytenr)
2656                 goto out_unlock;
2657
2658         ret = 1;
2659         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2660                 goto out_unlock;
2661
2662         data_ref = btrfs_delayed_node_to_data_ref(ref);
2663
2664         node = rb_prev(node);
2665         if (node) {
2666                 int seq = ref->seq;
2667
2668                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2669                 if (ref->bytenr == bytenr && ref->seq == seq)
2670                         goto out_unlock;
2671         }
2672
2673         if (data_ref->root != root->root_key.objectid ||
2674             data_ref->objectid != objectid || data_ref->offset != offset)
2675                 goto out_unlock;
2676
2677         ret = 0;
2678 out_unlock:
2679         mutex_unlock(&head->mutex);
2680 out:
2681         spin_unlock(&delayed_refs->lock);
2682         return ret;
2683 }
2684
2685 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2686                                         struct btrfs_root *root,
2687                                         struct btrfs_path *path,
2688                                         u64 objectid, u64 offset, u64 bytenr)
2689 {
2690         struct btrfs_root *extent_root = root->fs_info->extent_root;
2691         struct extent_buffer *leaf;
2692         struct btrfs_extent_data_ref *ref;
2693         struct btrfs_extent_inline_ref *iref;
2694         struct btrfs_extent_item *ei;
2695         struct btrfs_key key;
2696         u32 item_size;
2697         int ret;
2698
2699         key.objectid = bytenr;
2700         key.offset = (u64)-1;
2701         key.type = BTRFS_EXTENT_ITEM_KEY;
2702
2703         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2704         if (ret < 0)
2705                 goto out;
2706         BUG_ON(ret == 0); /* Corruption */
2707
2708         ret = -ENOENT;
2709         if (path->slots[0] == 0)
2710                 goto out;
2711
2712         path->slots[0]--;
2713         leaf = path->nodes[0];
2714         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2715
2716         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2717                 goto out;
2718
2719         ret = 1;
2720         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2721 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2722         if (item_size < sizeof(*ei)) {
2723                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2724                 goto out;
2725         }
2726 #endif
2727         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2728
2729         if (item_size != sizeof(*ei) +
2730             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2731                 goto out;
2732
2733         if (btrfs_extent_generation(leaf, ei) <=
2734             btrfs_root_last_snapshot(&root->root_item))
2735                 goto out;
2736
2737         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2738         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2739             BTRFS_EXTENT_DATA_REF_KEY)
2740                 goto out;
2741
2742         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2743         if (btrfs_extent_refs(leaf, ei) !=
2744             btrfs_extent_data_ref_count(leaf, ref) ||
2745             btrfs_extent_data_ref_root(leaf, ref) !=
2746             root->root_key.objectid ||
2747             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2748             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2749                 goto out;
2750
2751         ret = 0;
2752 out:
2753         return ret;
2754 }
2755
2756 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2757                           struct btrfs_root *root,
2758                           u64 objectid, u64 offset, u64 bytenr)
2759 {
2760         struct btrfs_path *path;
2761         int ret;
2762         int ret2;
2763
2764         path = btrfs_alloc_path();
2765         if (!path)
2766                 return -ENOENT;
2767
2768         do {
2769                 ret = check_committed_ref(trans, root, path, objectid,
2770                                           offset, bytenr);
2771                 if (ret && ret != -ENOENT)
2772                         goto out;
2773
2774                 ret2 = check_delayed_ref(trans, root, path, objectid,
2775                                          offset, bytenr);
2776         } while (ret2 == -EAGAIN);
2777
2778         if (ret2 && ret2 != -ENOENT) {
2779                 ret = ret2;
2780                 goto out;
2781         }
2782
2783         if (ret != -ENOENT || ret2 != -ENOENT)
2784                 ret = 0;
2785 out:
2786         btrfs_free_path(path);
2787         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2788                 WARN_ON(ret > 0);
2789         return ret;
2790 }
2791
2792 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2793                            struct btrfs_root *root,
2794                            struct extent_buffer *buf,
2795                            int full_backref, int inc, int for_cow)
2796 {
2797         u64 bytenr;
2798         u64 num_bytes;
2799         u64 parent;
2800         u64 ref_root;
2801         u32 nritems;
2802         struct btrfs_key key;
2803         struct btrfs_file_extent_item *fi;
2804         int i;
2805         int level;
2806         int ret = 0;
2807         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2808                             u64, u64, u64, u64, u64, u64, int);
2809
2810         ref_root = btrfs_header_owner(buf);
2811         nritems = btrfs_header_nritems(buf);
2812         level = btrfs_header_level(buf);
2813
2814         if (!root->ref_cows && level == 0)
2815                 return 0;
2816
2817         if (inc)
2818                 process_func = btrfs_inc_extent_ref;
2819         else
2820                 process_func = btrfs_free_extent;
2821
2822         if (full_backref)
2823                 parent = buf->start;
2824         else
2825                 parent = 0;
2826
2827         for (i = 0; i < nritems; i++) {
2828                 if (level == 0) {
2829                         btrfs_item_key_to_cpu(buf, &key, i);
2830                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2831                                 continue;
2832                         fi = btrfs_item_ptr(buf, i,
2833                                             struct btrfs_file_extent_item);
2834                         if (btrfs_file_extent_type(buf, fi) ==
2835                             BTRFS_FILE_EXTENT_INLINE)
2836                                 continue;
2837                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2838                         if (bytenr == 0)
2839                                 continue;
2840
2841                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2842                         key.offset -= btrfs_file_extent_offset(buf, fi);
2843                         ret = process_func(trans, root, bytenr, num_bytes,
2844                                            parent, ref_root, key.objectid,
2845                                            key.offset, for_cow);
2846                         if (ret)
2847                                 goto fail;
2848                 } else {
2849                         bytenr = btrfs_node_blockptr(buf, i);
2850                         num_bytes = btrfs_level_size(root, level - 1);
2851                         ret = process_func(trans, root, bytenr, num_bytes,
2852                                            parent, ref_root, level - 1, 0,
2853                                            for_cow);
2854                         if (ret)
2855                                 goto fail;
2856                 }
2857         }
2858         return 0;
2859 fail:
2860         return ret;
2861 }
2862
2863 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2864                   struct extent_buffer *buf, int full_backref, int for_cow)
2865 {
2866         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2867 }
2868
2869 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2870                   struct extent_buffer *buf, int full_backref, int for_cow)
2871 {
2872         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2873 }
2874
2875 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2876                                  struct btrfs_root *root,
2877                                  struct btrfs_path *path,
2878                                  struct btrfs_block_group_cache *cache)
2879 {
2880         int ret;
2881         struct btrfs_root *extent_root = root->fs_info->extent_root;
2882         unsigned long bi;
2883         struct extent_buffer *leaf;
2884
2885         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2886         if (ret < 0)
2887                 goto fail;
2888         BUG_ON(ret); /* Corruption */
2889
2890         leaf = path->nodes[0];
2891         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2892         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2893         btrfs_mark_buffer_dirty(leaf);
2894         btrfs_release_path(path);
2895 fail:
2896         if (ret) {
2897                 btrfs_abort_transaction(trans, root, ret);
2898                 return ret;
2899         }
2900         return 0;
2901
2902 }
2903
2904 static struct btrfs_block_group_cache *
2905 next_block_group(struct btrfs_root *root,
2906                  struct btrfs_block_group_cache *cache)
2907 {
2908         struct rb_node *node;
2909         spin_lock(&root->fs_info->block_group_cache_lock);
2910         node = rb_next(&cache->cache_node);
2911         btrfs_put_block_group(cache);
2912         if (node) {
2913                 cache = rb_entry(node, struct btrfs_block_group_cache,
2914                                  cache_node);
2915                 btrfs_get_block_group(cache);
2916         } else
2917                 cache = NULL;
2918         spin_unlock(&root->fs_info->block_group_cache_lock);
2919         return cache;
2920 }
2921
2922 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2923                             struct btrfs_trans_handle *trans,
2924                             struct btrfs_path *path)
2925 {
2926         struct btrfs_root *root = block_group->fs_info->tree_root;
2927         struct inode *inode = NULL;
2928         u64 alloc_hint = 0;
2929         int dcs = BTRFS_DC_ERROR;
2930         int num_pages = 0;
2931         int retries = 0;
2932         int ret = 0;
2933
2934         /*
2935          * If this block group is smaller than 100 megs don't bother caching the
2936          * block group.
2937          */
2938         if (block_group->key.offset < (100 * 1024 * 1024)) {
2939                 spin_lock(&block_group->lock);
2940                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2941                 spin_unlock(&block_group->lock);
2942                 return 0;
2943         }
2944
2945 again:
2946         inode = lookup_free_space_inode(root, block_group, path);
2947         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2948                 ret = PTR_ERR(inode);
2949                 btrfs_release_path(path);
2950                 goto out;
2951         }
2952
2953         if (IS_ERR(inode)) {
2954                 BUG_ON(retries);
2955                 retries++;
2956
2957                 if (block_group->ro)
2958                         goto out_free;
2959
2960                 ret = create_free_space_inode(root, trans, block_group, path);
2961                 if (ret)
2962                         goto out_free;
2963                 goto again;
2964         }
2965
2966         /* We've already setup this transaction, go ahead and exit */
2967         if (block_group->cache_generation == trans->transid &&
2968             i_size_read(inode)) {
2969                 dcs = BTRFS_DC_SETUP;
2970                 goto out_put;
2971         }
2972
2973         /*
2974          * We want to set the generation to 0, that way if anything goes wrong
2975          * from here on out we know not to trust this cache when we load up next
2976          * time.
2977          */
2978         BTRFS_I(inode)->generation = 0;
2979         ret = btrfs_update_inode(trans, root, inode);
2980         WARN_ON(ret);
2981
2982         if (i_size_read(inode) > 0) {
2983                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2984                                                       inode);
2985                 if (ret)
2986                         goto out_put;
2987         }
2988
2989         spin_lock(&block_group->lock);
2990         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2991             !btrfs_test_opt(root, SPACE_CACHE)) {
2992                 /*
2993                  * don't bother trying to write stuff out _if_
2994                  * a) we're not cached,
2995                  * b) we're with nospace_cache mount option.
2996                  */
2997                 dcs = BTRFS_DC_WRITTEN;
2998                 spin_unlock(&block_group->lock);
2999                 goto out_put;
3000         }
3001         spin_unlock(&block_group->lock);
3002
3003         /*
3004          * Try to preallocate enough space based on how big the block group is.
3005          * Keep in mind this has to include any pinned space which could end up
3006          * taking up quite a bit since it's not folded into the other space
3007          * cache.
3008          */
3009         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3010         if (!num_pages)
3011                 num_pages = 1;
3012
3013         num_pages *= 16;
3014         num_pages *= PAGE_CACHE_SIZE;
3015
3016         ret = btrfs_check_data_free_space(inode, num_pages);
3017         if (ret)
3018                 goto out_put;
3019
3020         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3021                                               num_pages, num_pages,
3022                                               &alloc_hint);
3023         if (!ret)
3024                 dcs = BTRFS_DC_SETUP;
3025         btrfs_free_reserved_data_space(inode, num_pages);
3026
3027 out_put:
3028         iput(inode);
3029 out_free:
3030         btrfs_release_path(path);
3031 out:
3032         spin_lock(&block_group->lock);
3033         if (!ret && dcs == BTRFS_DC_SETUP)
3034                 block_group->cache_generation = trans->transid;
3035         block_group->disk_cache_state = dcs;
3036         spin_unlock(&block_group->lock);
3037
3038         return ret;
3039 }
3040
3041 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3042                                    struct btrfs_root *root)
3043 {
3044         struct btrfs_block_group_cache *cache;
3045         int err = 0;
3046         struct btrfs_path *path;
3047         u64 last = 0;
3048
3049         path = btrfs_alloc_path();
3050         if (!path)
3051                 return -ENOMEM;
3052
3053 again:
3054         while (1) {
3055                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3056                 while (cache) {
3057                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3058                                 break;
3059                         cache = next_block_group(root, cache);
3060                 }
3061                 if (!cache) {
3062                         if (last == 0)
3063                                 break;
3064                         last = 0;
3065                         continue;
3066                 }
3067                 err = cache_save_setup(cache, trans, path);
3068                 last = cache->key.objectid + cache->key.offset;
3069                 btrfs_put_block_group(cache);
3070         }
3071
3072         while (1) {
3073                 if (last == 0) {
3074                         err = btrfs_run_delayed_refs(trans, root,
3075                                                      (unsigned long)-1);
3076                         if (err) /* File system offline */
3077                                 goto out;
3078                 }
3079
3080                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3081                 while (cache) {
3082                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3083                                 btrfs_put_block_group(cache);
3084                                 goto again;
3085                         }
3086
3087                         if (cache->dirty)
3088                                 break;
3089                         cache = next_block_group(root, cache);
3090                 }
3091                 if (!cache) {
3092                         if (last == 0)
3093                                 break;
3094                         last = 0;
3095                         continue;
3096                 }
3097
3098                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3099                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3100                 cache->dirty = 0;
3101                 last = cache->key.objectid + cache->key.offset;
3102
3103                 err = write_one_cache_group(trans, root, path, cache);
3104                 if (err) /* File system offline */
3105                         goto out;
3106
3107                 btrfs_put_block_group(cache);
3108         }
3109
3110         while (1) {
3111                 /*
3112                  * I don't think this is needed since we're just marking our
3113                  * preallocated extent as written, but just in case it can't
3114                  * hurt.
3115                  */
3116                 if (last == 0) {
3117                         err = btrfs_run_delayed_refs(trans, root,
3118                                                      (unsigned long)-1);
3119                         if (err) /* File system offline */
3120                                 goto out;
3121                 }
3122
3123                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3124                 while (cache) {
3125                         /*
3126                          * Really this shouldn't happen, but it could if we
3127                          * couldn't write the entire preallocated extent and
3128                          * splitting the extent resulted in a new block.
3129                          */
3130                         if (cache->dirty) {
3131                                 btrfs_put_block_group(cache);
3132                                 goto again;
3133                         }
3134                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3135                                 break;
3136                         cache = next_block_group(root, cache);
3137                 }
3138                 if (!cache) {
3139                         if (last == 0)
3140                                 break;
3141                         last = 0;
3142                         continue;
3143                 }
3144
3145                 err = btrfs_write_out_cache(root, trans, cache, path);
3146
3147                 /*
3148                  * If we didn't have an error then the cache state is still
3149                  * NEED_WRITE, so we can set it to WRITTEN.
3150                  */
3151                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3152                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3153                 last = cache->key.objectid + cache->key.offset;
3154                 btrfs_put_block_group(cache);
3155         }
3156 out:
3157
3158         btrfs_free_path(path);
3159         return err;
3160 }
3161
3162 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3163 {
3164         struct btrfs_block_group_cache *block_group;
3165         int readonly = 0;
3166
3167         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3168         if (!block_group || block_group->ro)
3169                 readonly = 1;
3170         if (block_group)
3171                 btrfs_put_block_group(block_group);
3172         return readonly;
3173 }
3174
3175 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3176                              u64 total_bytes, u64 bytes_used,
3177                              struct btrfs_space_info **space_info)
3178 {
3179         struct btrfs_space_info *found;
3180         int i;
3181         int factor;
3182
3183         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3184                      BTRFS_BLOCK_GROUP_RAID10))
3185                 factor = 2;
3186         else
3187                 factor = 1;
3188
3189         found = __find_space_info(info, flags);
3190         if (found) {
3191                 spin_lock(&found->lock);
3192                 found->total_bytes += total_bytes;
3193                 found->disk_total += total_bytes * factor;
3194                 found->bytes_used += bytes_used;
3195                 found->disk_used += bytes_used * factor;
3196                 found->full = 0;
3197                 spin_unlock(&found->lock);
3198                 *space_info = found;
3199                 return 0;
3200         }
3201         found = kzalloc(sizeof(*found), GFP_NOFS);
3202         if (!found)
3203                 return -ENOMEM;
3204
3205         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3206                 INIT_LIST_HEAD(&found->block_groups[i]);
3207         init_rwsem(&found->groups_sem);
3208         spin_lock_init(&found->lock);
3209         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3210         found->total_bytes = total_bytes;
3211         found->disk_total = total_bytes * factor;
3212         found->bytes_used = bytes_used;
3213         found->disk_used = bytes_used * factor;
3214         found->bytes_pinned = 0;
3215         found->bytes_reserved = 0;
3216         found->bytes_readonly = 0;
3217         found->bytes_may_use = 0;
3218         found->full = 0;
3219         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3220         found->chunk_alloc = 0;
3221         found->flush = 0;
3222         init_waitqueue_head(&found->wait);
3223         *space_info = found;
3224         list_add_rcu(&found->list, &info->space_info);
3225         if (flags & BTRFS_BLOCK_GROUP_DATA)
3226                 info->data_sinfo = found;
3227         return 0;
3228 }
3229
3230 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3231 {
3232         u64 extra_flags = chunk_to_extended(flags) &
3233                                 BTRFS_EXTENDED_PROFILE_MASK;
3234
3235         if (flags & BTRFS_BLOCK_GROUP_DATA)
3236                 fs_info->avail_data_alloc_bits |= extra_flags;
3237         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3238                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3239         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3240                 fs_info->avail_system_alloc_bits |= extra_flags;
3241 }
3242
3243 /*
3244  * returns target flags in extended format or 0 if restripe for this
3245  * chunk_type is not in progress
3246  *
3247  * should be called with either volume_mutex or balance_lock held
3248  */
3249 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3250 {
3251         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3252         u64 target = 0;
3253
3254         if (!bctl)
3255                 return 0;
3256
3257         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3258             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3259                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3260         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3261                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3262                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3263         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3264                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3265                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3266         }
3267
3268         return target;
3269 }
3270
3271 /*
3272  * @flags: available profiles in extended format (see ctree.h)
3273  *
3274  * Returns reduced profile in chunk format.  If profile changing is in
3275  * progress (either running or paused) picks the target profile (if it's
3276  * already available), otherwise falls back to plain reducing.
3277  */
3278 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3279 {
3280         /*
3281          * we add in the count of missing devices because we want
3282          * to make sure that any RAID levels on a degraded FS
3283          * continue to be honored.
3284          */
3285         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3286                 root->fs_info->fs_devices->missing_devices;
3287         u64 target;
3288
3289         /*
3290          * see if restripe for this chunk_type is in progress, if so
3291          * try to reduce to the target profile
3292          */
3293         spin_lock(&root->fs_info->balance_lock);
3294         target = get_restripe_target(root->fs_info, flags);
3295         if (target) {
3296                 /* pick target profile only if it's already available */
3297                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3298                         spin_unlock(&root->fs_info->balance_lock);
3299                         return extended_to_chunk(target);
3300                 }
3301         }
3302         spin_unlock(&root->fs_info->balance_lock);
3303
3304         if (num_devices == 1)
3305                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3306         if (num_devices < 4)
3307                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3308
3309         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3310             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3311                       BTRFS_BLOCK_GROUP_RAID10))) {
3312                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3313         }
3314
3315         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3316             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3317                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3318         }
3319
3320         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3321             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3322              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3323              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3324                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3325         }
3326
3327         return extended_to_chunk(flags);
3328 }
3329
3330 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3331 {
3332         if (flags & BTRFS_BLOCK_GROUP_DATA)
3333                 flags |= root->fs_info->avail_data_alloc_bits;
3334         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3335                 flags |= root->fs_info->avail_system_alloc_bits;
3336         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3337                 flags |= root->fs_info->avail_metadata_alloc_bits;
3338
3339         return btrfs_reduce_alloc_profile(root, flags);
3340 }
3341
3342 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3343 {
3344         u64 flags;
3345
3346         if (data)
3347                 flags = BTRFS_BLOCK_GROUP_DATA;
3348         else if (root == root->fs_info->chunk_root)
3349                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3350         else
3351                 flags = BTRFS_BLOCK_GROUP_METADATA;
3352
3353         return get_alloc_profile(root, flags);
3354 }
3355
3356 /*
3357  * This will check the space that the inode allocates from to make sure we have
3358  * enough space for bytes.
3359  */
3360 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3361 {
3362         struct btrfs_space_info *data_sinfo;
3363         struct btrfs_root *root = BTRFS_I(inode)->root;
3364         struct btrfs_fs_info *fs_info = root->fs_info;
3365         u64 used;
3366         int ret = 0, committed = 0, alloc_chunk = 1;
3367
3368         /* make sure bytes are sectorsize aligned */
3369         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3370
3371         if (root == root->fs_info->tree_root ||
3372             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3373                 alloc_chunk = 0;
3374                 committed = 1;
3375         }
3376
3377         data_sinfo = fs_info->data_sinfo;
3378         if (!data_sinfo)
3379                 goto alloc;
3380
3381 again:
3382         /* make sure we have enough space to handle the data first */
3383         spin_lock(&data_sinfo->lock);
3384         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3385                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3386                 data_sinfo->bytes_may_use;
3387
3388         if (used + bytes > data_sinfo->total_bytes) {
3389                 struct btrfs_trans_handle *trans;
3390
3391                 /*
3392                  * if we don't have enough free bytes in this space then we need
3393                  * to alloc a new chunk.
3394                  */
3395                 if (!data_sinfo->full && alloc_chunk) {
3396                         u64 alloc_target;
3397
3398                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3399                         spin_unlock(&data_sinfo->lock);
3400 alloc:
3401                         alloc_target = btrfs_get_alloc_profile(root, 1);
3402                         trans = btrfs_join_transaction(root);
3403                         if (IS_ERR(trans))
3404                                 return PTR_ERR(trans);
3405
3406                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3407                                              alloc_target,
3408                                              CHUNK_ALLOC_NO_FORCE);
3409                         btrfs_end_transaction(trans, root);
3410                         if (ret < 0) {
3411                                 if (ret != -ENOSPC)
3412                                         return ret;
3413                                 else
3414                                         goto commit_trans;
3415                         }
3416
3417                         if (!data_sinfo)
3418                                 data_sinfo = fs_info->data_sinfo;
3419
3420                         goto again;
3421                 }
3422
3423                 /*
3424                  * If we have less pinned bytes than we want to allocate then
3425                  * don't bother committing the transaction, it won't help us.
3426                  */
3427                 if (data_sinfo->bytes_pinned < bytes)
3428                         committed = 1;
3429                 spin_unlock(&data_sinfo->lock);
3430
3431                 /* commit the current transaction and try again */
3432 commit_trans:
3433                 if (!committed &&
3434                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3435                         committed = 1;
3436                         trans = btrfs_join_transaction(root);
3437                         if (IS_ERR(trans))
3438                                 return PTR_ERR(trans);
3439                         ret = btrfs_commit_transaction(trans, root);
3440                         if (ret)
3441                                 return ret;
3442                         goto again;
3443                 }
3444
3445                 return -ENOSPC;
3446         }
3447         data_sinfo->bytes_may_use += bytes;
3448         trace_btrfs_space_reservation(root->fs_info, "space_info",
3449                                       data_sinfo->flags, bytes, 1);
3450         spin_unlock(&data_sinfo->lock);
3451
3452         return 0;
3453 }
3454
3455 /*
3456  * Called if we need to clear a data reservation for this inode.
3457  */
3458 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3459 {
3460         struct btrfs_root *root = BTRFS_I(inode)->root;
3461         struct btrfs_space_info *data_sinfo;
3462
3463         /* make sure bytes are sectorsize aligned */
3464         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3465
3466         data_sinfo = root->fs_info->data_sinfo;
3467         spin_lock(&data_sinfo->lock);
3468         data_sinfo->bytes_may_use -= bytes;
3469         trace_btrfs_space_reservation(root->fs_info, "space_info",
3470                                       data_sinfo->flags, bytes, 0);
3471         spin_unlock(&data_sinfo->lock);
3472 }
3473
3474 static void force_metadata_allocation(struct btrfs_fs_info *info)
3475 {
3476         struct list_head *head = &info->space_info;
3477         struct btrfs_space_info *found;
3478
3479         rcu_read_lock();
3480         list_for_each_entry_rcu(found, head, list) {
3481                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3482                         found->force_alloc = CHUNK_ALLOC_FORCE;
3483         }
3484         rcu_read_unlock();
3485 }
3486
3487 static int should_alloc_chunk(struct btrfs_root *root,
3488                               struct btrfs_space_info *sinfo, int force)
3489 {
3490         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3491         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3492         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3493         u64 thresh;
3494
3495         if (force == CHUNK_ALLOC_FORCE)
3496                 return 1;
3497
3498         /*
3499          * We need to take into account the global rsv because for all intents
3500          * and purposes it's used space.  Don't worry about locking the
3501          * global_rsv, it doesn't change except when the transaction commits.
3502          */
3503         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3504                 num_allocated += global_rsv->size;
3505
3506         /*
3507          * in limited mode, we want to have some free space up to
3508          * about 1% of the FS size.
3509          */
3510         if (force == CHUNK_ALLOC_LIMITED) {
3511                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3512                 thresh = max_t(u64, 64 * 1024 * 1024,
3513                                div_factor_fine(thresh, 1));
3514
3515                 if (num_bytes - num_allocated < thresh)
3516                         return 1;
3517         }
3518
3519         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3520                 return 0;
3521         return 1;
3522 }
3523
3524 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3525 {
3526         u64 num_dev;
3527
3528         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3529             type & BTRFS_BLOCK_GROUP_RAID0)
3530                 num_dev = root->fs_info->fs_devices->rw_devices;
3531         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3532                 num_dev = 2;
3533         else
3534                 num_dev = 1;    /* DUP or single */
3535
3536         /* metadata for updaing devices and chunk tree */
3537         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3538 }
3539
3540 static void check_system_chunk(struct btrfs_trans_handle *trans,
3541                                struct btrfs_root *root, u64 type)
3542 {
3543         struct btrfs_space_info *info;
3544         u64 left;
3545         u64 thresh;
3546
3547         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3548         spin_lock(&info->lock);
3549         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3550                 info->bytes_reserved - info->bytes_readonly;
3551         spin_unlock(&info->lock);
3552
3553         thresh = get_system_chunk_thresh(root, type);
3554         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3555                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3556                        left, thresh, type);
3557                 dump_space_info(info, 0, 0);
3558         }
3559
3560         if (left < thresh) {
3561                 u64 flags;
3562
3563                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3564                 btrfs_alloc_chunk(trans, root, flags);
3565         }
3566 }
3567
3568 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3569                           struct btrfs_root *extent_root, u64 flags, int force)
3570 {
3571         struct btrfs_space_info *space_info;
3572         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3573         int wait_for_alloc = 0;
3574         int ret = 0;
3575
3576         space_info = __find_space_info(extent_root->fs_info, flags);
3577         if (!space_info) {
3578                 ret = update_space_info(extent_root->fs_info, flags,
3579                                         0, 0, &space_info);
3580                 BUG_ON(ret); /* -ENOMEM */
3581         }
3582         BUG_ON(!space_info); /* Logic error */
3583
3584 again:
3585         spin_lock(&space_info->lock);
3586         if (force < space_info->force_alloc)
3587                 force = space_info->force_alloc;
3588         if (space_info->full) {
3589                 spin_unlock(&space_info->lock);
3590                 return 0;
3591         }
3592
3593         if (!should_alloc_chunk(extent_root, space_info, force)) {
3594                 spin_unlock(&space_info->lock);
3595                 return 0;
3596         } else if (space_info->chunk_alloc) {
3597                 wait_for_alloc = 1;
3598         } else {
3599                 space_info->chunk_alloc = 1;
3600         }
3601
3602         spin_unlock(&space_info->lock);
3603
3604         mutex_lock(&fs_info->chunk_mutex);
3605
3606         /*
3607          * The chunk_mutex is held throughout the entirety of a chunk
3608          * allocation, so once we've acquired the chunk_mutex we know that the
3609          * other guy is done and we need to recheck and see if we should
3610          * allocate.
3611          */
3612         if (wait_for_alloc) {
3613                 mutex_unlock(&fs_info->chunk_mutex);
3614                 wait_for_alloc = 0;
3615                 goto again;
3616         }
3617
3618         /*
3619          * If we have mixed data/metadata chunks we want to make sure we keep
3620          * allocating mixed chunks instead of individual chunks.
3621          */
3622         if (btrfs_mixed_space_info(space_info))
3623                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3624
3625         /*
3626          * if we're doing a data chunk, go ahead and make sure that
3627          * we keep a reasonable number of metadata chunks allocated in the
3628          * FS as well.
3629          */
3630         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3631                 fs_info->data_chunk_allocations++;
3632                 if (!(fs_info->data_chunk_allocations %
3633                       fs_info->metadata_ratio))
3634                         force_metadata_allocation(fs_info);
3635         }
3636
3637         /*
3638          * Check if we have enough space in SYSTEM chunk because we may need
3639          * to update devices.
3640          */
3641         check_system_chunk(trans, extent_root, flags);
3642
3643         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3644         if (ret < 0 && ret != -ENOSPC)
3645                 goto out;
3646
3647         spin_lock(&space_info->lock);
3648         if (ret)
3649                 space_info->full = 1;
3650         else
3651                 ret = 1;
3652
3653         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3654         space_info->chunk_alloc = 0;
3655         spin_unlock(&space_info->lock);
3656 out:
3657         mutex_unlock(&fs_info->chunk_mutex);
3658         return ret;
3659 }
3660
3661 static int can_overcommit(struct btrfs_root *root,
3662                           struct btrfs_space_info *space_info, u64 bytes,
3663                           int flush)
3664 {
3665         u64 profile = btrfs_get_alloc_profile(root, 0);
3666         u64 avail;
3667         u64 used;
3668
3669         used = space_info->bytes_used + space_info->bytes_reserved +
3670                 space_info->bytes_pinned + space_info->bytes_readonly +
3671                 space_info->bytes_may_use;
3672
3673         spin_lock(&root->fs_info->free_chunk_lock);
3674         avail = root->fs_info->free_chunk_space;
3675         spin_unlock(&root->fs_info->free_chunk_lock);
3676
3677         /*
3678          * If we have dup, raid1 or raid10 then only half of the free
3679          * space is actually useable.
3680          */
3681         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3682                        BTRFS_BLOCK_GROUP_RAID1 |
3683                        BTRFS_BLOCK_GROUP_RAID10))
3684                 avail >>= 1;
3685
3686         /*
3687          * If we aren't flushing don't let us overcommit too much, say
3688          * 1/8th of the space.  If we can flush, let it overcommit up to
3689          * 1/2 of the space.
3690          */
3691         if (flush)
3692                 avail >>= 3;
3693         else
3694                 avail >>= 1;
3695
3696         if (used + bytes < space_info->total_bytes + avail)
3697                 return 1;
3698         return 0;
3699 }
3700
3701 /*
3702  * shrink metadata reservation for delalloc
3703  */
3704 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3705                             bool wait_ordered)
3706 {
3707         struct btrfs_block_rsv *block_rsv;
3708         struct btrfs_space_info *space_info;
3709         struct btrfs_trans_handle *trans;
3710         u64 delalloc_bytes;
3711         u64 max_reclaim;
3712         long time_left;
3713         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3714         int loops = 0;
3715
3716         trans = (struct btrfs_trans_handle *)current->journal_info;
3717         block_rsv = &root->fs_info->delalloc_block_rsv;
3718         space_info = block_rsv->space_info;
3719
3720         smp_mb();
3721         delalloc_bytes = root->fs_info->delalloc_bytes;
3722         if (delalloc_bytes == 0) {
3723                 if (trans)
3724                         return;
3725                 btrfs_wait_ordered_extents(root, 0, 0);
3726                 return;
3727         }
3728
3729         while (delalloc_bytes && loops < 3) {
3730                 max_reclaim = min(delalloc_bytes, to_reclaim);
3731                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3732                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3733                                                WB_REASON_FS_FREE_SPACE);
3734
3735                 /*
3736                  * We need to wait for the async pages to actually start before
3737                  * we do anything.
3738                  */
3739                 wait_event(root->fs_info->async_submit_wait,
3740                            !atomic_read(&root->fs_info->async_delalloc_pages));
3741
3742                 spin_lock(&space_info->lock);
3743                 if (can_overcommit(root, space_info, orig, !trans)) {
3744                         spin_unlock(&space_info->lock);
3745                         break;
3746                 }
3747                 spin_unlock(&space_info->lock);
3748
3749                 loops++;
3750                 if (wait_ordered && !trans) {
3751                         btrfs_wait_ordered_extents(root, 0, 0);
3752                 } else {
3753                         time_left = schedule_timeout_killable(1);
3754                         if (time_left)
3755                                 break;
3756                 }
3757                 smp_mb();
3758                 delalloc_bytes = root->fs_info->delalloc_bytes;
3759         }
3760 }
3761
3762 /**
3763  * maybe_commit_transaction - possibly commit the transaction if its ok to
3764  * @root - the root we're allocating for
3765  * @bytes - the number of bytes we want to reserve
3766  * @force - force the commit
3767  *
3768  * This will check to make sure that committing the transaction will actually
3769  * get us somewhere and then commit the transaction if it does.  Otherwise it
3770  * will return -ENOSPC.
3771  */
3772 static int may_commit_transaction(struct btrfs_root *root,
3773                                   struct btrfs_space_info *space_info,
3774                                   u64 bytes, int force)
3775 {
3776         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3777         struct btrfs_trans_handle *trans;
3778
3779         trans = (struct btrfs_trans_handle *)current->journal_info;
3780         if (trans)
3781                 return -EAGAIN;
3782
3783         if (force)
3784                 goto commit;
3785
3786         /* See if there is enough pinned space to make this reservation */
3787         spin_lock(&space_info->lock);
3788         if (space_info->bytes_pinned >= bytes) {
3789                 spin_unlock(&space_info->lock);
3790                 goto commit;
3791         }
3792         spin_unlock(&space_info->lock);
3793
3794         /*
3795          * See if there is some space in the delayed insertion reservation for
3796          * this reservation.
3797          */
3798         if (space_info != delayed_rsv->space_info)
3799                 return -ENOSPC;
3800
3801         spin_lock(&space_info->lock);
3802         spin_lock(&delayed_rsv->lock);
3803         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3804                 spin_unlock(&delayed_rsv->lock);
3805                 spin_unlock(&space_info->lock);
3806                 return -ENOSPC;
3807         }
3808         spin_unlock(&delayed_rsv->lock);
3809         spin_unlock(&space_info->lock);
3810
3811 commit:
3812         trans = btrfs_join_transaction(root);
3813         if (IS_ERR(trans))
3814                 return -ENOSPC;
3815
3816         return btrfs_commit_transaction(trans, root);
3817 }
3818
3819 enum flush_state {
3820         FLUSH_DELALLOC          =       1,
3821         FLUSH_DELALLOC_WAIT     =       2,
3822         FLUSH_DELAYED_ITEMS_NR  =       3,
3823         FLUSH_DELAYED_ITEMS     =       4,
3824         ALLOC_CHUNK             =       5,
3825         COMMIT_TRANS            =       6,
3826 };
3827
3828 static int flush_space(struct btrfs_root *root,
3829                        struct btrfs_space_info *space_info, u64 num_bytes,
3830                        u64 orig_bytes, int state)
3831 {
3832         struct btrfs_trans_handle *trans;
3833         int nr;
3834         int ret = 0;
3835
3836         switch (state) {
3837         case FLUSH_DELALLOC:
3838         case FLUSH_DELALLOC_WAIT:
3839                 shrink_delalloc(root, num_bytes, orig_bytes,
3840                                 state == FLUSH_DELALLOC_WAIT);
3841                 break;
3842         case FLUSH_DELAYED_ITEMS_NR:
3843         case FLUSH_DELAYED_ITEMS:
3844                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3845                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3846
3847                         nr = (int)div64_u64(num_bytes, bytes);
3848                         if (!nr)
3849                                 nr = 1;
3850                         nr *= 2;
3851                 } else {
3852                         nr = -1;
3853                 }
3854                 trans = btrfs_join_transaction(root);
3855                 if (IS_ERR(trans)) {
3856                         ret = PTR_ERR(trans);
3857                         break;
3858                 }
3859                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3860                 btrfs_end_transaction(trans, root);
3861                 break;
3862         case ALLOC_CHUNK:
3863                 trans = btrfs_join_transaction(root);
3864                 if (IS_ERR(trans)) {
3865                         ret = PTR_ERR(trans);
3866                         break;
3867                 }
3868                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3869                                      btrfs_get_alloc_profile(root, 0),
3870                                      CHUNK_ALLOC_NO_FORCE);
3871                 btrfs_end_transaction(trans, root);
3872                 if (ret == -ENOSPC)
3873                         ret = 0;
3874                 break;
3875         case COMMIT_TRANS:
3876                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3877                 break;
3878         default:
3879                 ret = -ENOSPC;
3880                 break;
3881         }
3882
3883         return ret;
3884 }
3885 /**
3886  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3887  * @root - the root we're allocating for
3888  * @block_rsv - the block_rsv we're allocating for
3889  * @orig_bytes - the number of bytes we want
3890  * @flush - wether or not we can flush to make our reservation
3891  *
3892  * This will reserve orgi_bytes number of bytes from the space info associated
3893  * with the block_rsv.  If there is not enough space it will make an attempt to
3894  * flush out space to make room.  It will do this by flushing delalloc if
3895  * possible or committing the transaction.  If flush is 0 then no attempts to
3896  * regain reservations will be made and this will fail if there is not enough
3897  * space already.
3898  */
3899 static int reserve_metadata_bytes(struct btrfs_root *root,
3900                                   struct btrfs_block_rsv *block_rsv,
3901                                   u64 orig_bytes, int flush)
3902 {
3903         struct btrfs_space_info *space_info = block_rsv->space_info;
3904         u64 used;
3905         u64 num_bytes = orig_bytes;
3906         int flush_state = FLUSH_DELALLOC;
3907         int ret = 0;
3908         bool flushing = false;
3909         bool committed = false;
3910
3911 again:
3912         ret = 0;
3913         spin_lock(&space_info->lock);
3914         /*
3915          * We only want to wait if somebody other than us is flushing and we are
3916          * actually alloed to flush.
3917          */
3918         while (flush && !flushing && space_info->flush) {
3919                 spin_unlock(&space_info->lock);
3920                 /*
3921                  * If we have a trans handle we can't wait because the flusher
3922                  * may have to commit the transaction, which would mean we would
3923                  * deadlock since we are waiting for the flusher to finish, but
3924                  * hold the current transaction open.
3925                  */
3926                 if (current->journal_info)
3927                         return -EAGAIN;
3928                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3929                 /* Must have been killed, return */
3930                 if (ret)
3931                         return -EINTR;
3932
3933                 spin_lock(&space_info->lock);
3934         }
3935
3936         ret = -ENOSPC;
3937         used = space_info->bytes_used + space_info->bytes_reserved +
3938                 space_info->bytes_pinned + space_info->bytes_readonly +
3939                 space_info->bytes_may_use;
3940
3941         /*
3942          * The idea here is that we've not already over-reserved the block group
3943          * then we can go ahead and save our reservation first and then start
3944          * flushing if we need to.  Otherwise if we've already overcommitted
3945          * lets start flushing stuff first and then come back and try to make
3946          * our reservation.
3947          */
3948         if (used <= space_info->total_bytes) {
3949                 if (used + orig_bytes <= space_info->total_bytes) {
3950                         space_info->bytes_may_use += orig_bytes;
3951                         trace_btrfs_space_reservation(root->fs_info,
3952                                 "space_info", space_info->flags, orig_bytes, 1);
3953                         ret = 0;
3954                 } else {
3955                         /*
3956                          * Ok set num_bytes to orig_bytes since we aren't
3957                          * overocmmitted, this way we only try and reclaim what
3958                          * we need.
3959                          */
3960                         num_bytes = orig_bytes;
3961                 }
3962         } else {
3963                 /*
3964                  * Ok we're over committed, set num_bytes to the overcommitted
3965                  * amount plus the amount of bytes that we need for this
3966                  * reservation.
3967                  */
3968                 num_bytes = used - space_info->total_bytes +
3969                         (orig_bytes * 2);
3970         }
3971
3972         if (ret) {
3973                 u64 avail;
3974
3975                 /*
3976                  * If we have a lot of space that's pinned, don't bother doing
3977                  * the overcommit dance yet and just commit the transaction.
3978                  */
3979                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3980                 do_div(avail, 10);
3981                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3982                         space_info->flush = 1;
3983                         flushing = true;
3984                         spin_unlock(&space_info->lock);
3985                         ret = may_commit_transaction(root, space_info,
3986                                                      orig_bytes, 1);
3987                         if (ret)
3988                                 goto out;
3989                         committed = true;
3990                         goto again;
3991                 }
3992
3993                 if (can_overcommit(root, space_info, orig_bytes, flush)) {
3994                         space_info->bytes_may_use += orig_bytes;
3995                         trace_btrfs_space_reservation(root->fs_info,
3996                                 "space_info", space_info->flags, orig_bytes, 1);
3997                         ret = 0;
3998                 }
3999         }
4000
4001         /*
4002          * Couldn't make our reservation, save our place so while we're trying
4003          * to reclaim space we can actually use it instead of somebody else
4004          * stealing it from us.
4005          */
4006         if (ret && flush) {
4007                 flushing = true;
4008                 space_info->flush = 1;
4009         }
4010
4011         spin_unlock(&space_info->lock);
4012
4013         if (!ret || !flush)
4014                 goto out;
4015
4016         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4017                           flush_state);
4018         flush_state++;
4019         if (!ret)
4020                 goto again;
4021         else if (flush_state <= COMMIT_TRANS)
4022                 goto again;
4023
4024 out:
4025         if (flushing) {
4026                 spin_lock(&space_info->lock);
4027                 space_info->flush = 0;
4028                 wake_up_all(&space_info->wait);
4029                 spin_unlock(&space_info->lock);
4030         }
4031         return ret;
4032 }
4033
4034 static struct btrfs_block_rsv *get_block_rsv(
4035                                         const struct btrfs_trans_handle *trans,
4036                                         const struct btrfs_root *root)
4037 {
4038         struct btrfs_block_rsv *block_rsv = NULL;
4039
4040         if (root->ref_cows)
4041                 block_rsv = trans->block_rsv;
4042
4043         if (root == root->fs_info->csum_root && trans->adding_csums)
4044                 block_rsv = trans->block_rsv;
4045
4046         if (!block_rsv)
4047                 block_rsv = root->block_rsv;
4048
4049         if (!block_rsv)
4050                 block_rsv = &root->fs_info->empty_block_rsv;
4051
4052         return block_rsv;
4053 }
4054
4055 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4056                                u64 num_bytes)
4057 {
4058         int ret = -ENOSPC;
4059         spin_lock(&block_rsv->lock);
4060         if (block_rsv->reserved >= num_bytes) {
4061                 block_rsv->reserved -= num_bytes;
4062                 if (block_rsv->reserved < block_rsv->size)
4063                         block_rsv->full = 0;
4064                 ret = 0;
4065         }
4066         spin_unlock(&block_rsv->lock);
4067         return ret;
4068 }
4069
4070 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4071                                 u64 num_bytes, int update_size)
4072 {
4073         spin_lock(&block_rsv->lock);
4074         block_rsv->reserved += num_bytes;
4075         if (update_size)
4076                 block_rsv->size += num_bytes;
4077         else if (block_rsv->reserved >= block_rsv->size)
4078                 block_rsv->full = 1;
4079         spin_unlock(&block_rsv->lock);
4080 }
4081
4082 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4083                                     struct btrfs_block_rsv *block_rsv,
4084                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4085 {
4086         struct btrfs_space_info *space_info = block_rsv->space_info;
4087
4088         spin_lock(&block_rsv->lock);
4089         if (num_bytes == (u64)-1)
4090                 num_bytes = block_rsv->size;
4091         block_rsv->size -= num_bytes;
4092         if (block_rsv->reserved >= block_rsv->size) {
4093                 num_bytes = block_rsv->reserved - block_rsv->size;
4094                 block_rsv->reserved = block_rsv->size;
4095                 block_rsv->full = 1;
4096         } else {
4097                 num_bytes = 0;
4098         }
4099         spin_unlock(&block_rsv->lock);
4100
4101         if (num_bytes > 0) {
4102                 if (dest) {
4103                         spin_lock(&dest->lock);
4104                         if (!dest->full) {
4105                                 u64 bytes_to_add;
4106
4107                                 bytes_to_add = dest->size - dest->reserved;
4108                                 bytes_to_add = min(num_bytes, bytes_to_add);
4109                                 dest->reserved += bytes_to_add;
4110                                 if (dest->reserved >= dest->size)
4111                                         dest->full = 1;
4112                                 num_bytes -= bytes_to_add;
4113                         }
4114                         spin_unlock(&dest->lock);
4115                 }
4116                 if (num_bytes) {
4117                         spin_lock(&space_info->lock);
4118                         space_info->bytes_may_use -= num_bytes;
4119                         trace_btrfs_space_reservation(fs_info, "space_info",
4120                                         space_info->flags, num_bytes, 0);
4121                         space_info->reservation_progress++;
4122                         spin_unlock(&space_info->lock);
4123                 }
4124         }
4125 }
4126
4127 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4128                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4129 {
4130         int ret;
4131
4132         ret = block_rsv_use_bytes(src, num_bytes);
4133         if (ret)
4134                 return ret;
4135
4136         block_rsv_add_bytes(dst, num_bytes, 1);
4137         return 0;
4138 }
4139
4140 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4141 {
4142         memset(rsv, 0, sizeof(*rsv));
4143         spin_lock_init(&rsv->lock);
4144         rsv->type = type;
4145 }
4146
4147 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4148                                               unsigned short type)
4149 {
4150         struct btrfs_block_rsv *block_rsv;
4151         struct btrfs_fs_info *fs_info = root->fs_info;
4152
4153         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4154         if (!block_rsv)
4155                 return NULL;
4156
4157         btrfs_init_block_rsv(block_rsv, type);
4158         block_rsv->space_info = __find_space_info(fs_info,
4159                                                   BTRFS_BLOCK_GROUP_METADATA);
4160         return block_rsv;
4161 }
4162
4163 void btrfs_free_block_rsv(struct btrfs_root *root,
4164                           struct btrfs_block_rsv *rsv)
4165 {
4166         if (!rsv)
4167                 return;
4168         btrfs_block_rsv_release(root, rsv, (u64)-1);
4169         kfree(rsv);
4170 }
4171
4172 static inline int __block_rsv_add(struct btrfs_root *root,
4173                                   struct btrfs_block_rsv *block_rsv,
4174                                   u64 num_bytes, int flush)
4175 {
4176         int ret;
4177
4178         if (num_bytes == 0)
4179                 return 0;
4180
4181         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4182         if (!ret) {
4183                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4184                 return 0;
4185         }
4186
4187         return ret;
4188 }
4189
4190 int btrfs_block_rsv_add(struct btrfs_root *root,
4191                         struct btrfs_block_rsv *block_rsv,
4192                         u64 num_bytes)
4193 {
4194         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4195 }
4196
4197 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4198                                 struct btrfs_block_rsv *block_rsv,
4199                                 u64 num_bytes)
4200 {
4201         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4202 }
4203
4204 int btrfs_block_rsv_check(struct btrfs_root *root,
4205                           struct btrfs_block_rsv *block_rsv, int min_factor)
4206 {
4207         u64 num_bytes = 0;
4208         int ret = -ENOSPC;
4209
4210         if (!block_rsv)
4211                 return 0;
4212
4213         spin_lock(&block_rsv->lock);
4214         num_bytes = div_factor(block_rsv->size, min_factor);
4215         if (block_rsv->reserved >= num_bytes)
4216                 ret = 0;
4217         spin_unlock(&block_rsv->lock);
4218
4219         return ret;
4220 }
4221
4222 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4223                                            struct btrfs_block_rsv *block_rsv,
4224                                            u64 min_reserved, int flush)
4225 {
4226         u64 num_bytes = 0;
4227         int ret = -ENOSPC;
4228
4229         if (!block_rsv)
4230                 return 0;
4231
4232         spin_lock(&block_rsv->lock);
4233         num_bytes = min_reserved;
4234         if (block_rsv->reserved >= num_bytes)
4235                 ret = 0;
4236         else
4237                 num_bytes -= block_rsv->reserved;
4238         spin_unlock(&block_rsv->lock);
4239
4240         if (!ret)
4241                 return 0;
4242
4243         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4244         if (!ret) {
4245                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4246                 return 0;
4247         }
4248
4249         return ret;
4250 }
4251
4252 int btrfs_block_rsv_refill(struct btrfs_root *root,
4253                            struct btrfs_block_rsv *block_rsv,
4254                            u64 min_reserved)
4255 {
4256         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4257 }
4258
4259 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4260                                    struct btrfs_block_rsv *block_rsv,
4261                                    u64 min_reserved)
4262 {
4263         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4264 }
4265
4266 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4267                             struct btrfs_block_rsv *dst_rsv,
4268                             u64 num_bytes)
4269 {
4270         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4271 }
4272
4273 void btrfs_block_rsv_release(struct btrfs_root *root,
4274                              struct btrfs_block_rsv *block_rsv,
4275                              u64 num_bytes)
4276 {
4277         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4278         if (global_rsv->full || global_rsv == block_rsv ||
4279             block_rsv->space_info != global_rsv->space_info)
4280                 global_rsv = NULL;
4281         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4282                                 num_bytes);
4283 }
4284
4285 /*
4286  * helper to calculate size of global block reservation.
4287  * the desired value is sum of space used by extent tree,
4288  * checksum tree and root tree
4289  */
4290 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4291 {
4292         struct btrfs_space_info *sinfo;
4293         u64 num_bytes;
4294         u64 meta_used;
4295         u64 data_used;
4296         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4297
4298         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4299         spin_lock(&sinfo->lock);
4300         data_used = sinfo->bytes_used;
4301         spin_unlock(&sinfo->lock);
4302
4303         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4304         spin_lock(&sinfo->lock);
4305         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4306                 data_used = 0;
4307         meta_used = sinfo->bytes_used;
4308         spin_unlock(&sinfo->lock);
4309
4310         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4311                     csum_size * 2;
4312         num_bytes += div64_u64(data_used + meta_used, 50);
4313
4314         if (num_bytes * 3 > meta_used)
4315                 num_bytes = div64_u64(meta_used, 3);
4316
4317         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4318 }
4319
4320 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4321 {
4322         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4323         struct btrfs_space_info *sinfo = block_rsv->space_info;
4324         u64 num_bytes;
4325
4326         num_bytes = calc_global_metadata_size(fs_info);
4327
4328         spin_lock(&sinfo->lock);
4329         spin_lock(&block_rsv->lock);
4330
4331         block_rsv->size = num_bytes;
4332
4333         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4334                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4335                     sinfo->bytes_may_use;
4336
4337         if (sinfo->total_bytes > num_bytes) {
4338                 num_bytes = sinfo->total_bytes - num_bytes;
4339                 block_rsv->reserved += num_bytes;
4340                 sinfo->bytes_may_use += num_bytes;
4341                 trace_btrfs_space_reservation(fs_info, "space_info",
4342                                       sinfo->flags, num_bytes, 1);
4343         }
4344
4345         if (block_rsv->reserved >= block_rsv->size) {
4346                 num_bytes = block_rsv->reserved - block_rsv->size;
4347                 sinfo->bytes_may_use -= num_bytes;
4348                 trace_btrfs_space_reservation(fs_info, "space_info",
4349                                       sinfo->flags, num_bytes, 0);
4350                 sinfo->reservation_progress++;
4351                 block_rsv->reserved = block_rsv->size;
4352                 block_rsv->full = 1;
4353         }
4354
4355         spin_unlock(&block_rsv->lock);
4356         spin_unlock(&sinfo->lock);
4357 }
4358
4359 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4360 {
4361         struct btrfs_space_info *space_info;
4362
4363         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4364         fs_info->chunk_block_rsv.space_info = space_info;
4365
4366         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4367         fs_info->global_block_rsv.space_info = space_info;
4368         fs_info->delalloc_block_rsv.space_info = space_info;
4369         fs_info->trans_block_rsv.space_info = space_info;
4370         fs_info->empty_block_rsv.space_info = space_info;
4371         fs_info->delayed_block_rsv.space_info = space_info;
4372
4373         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4374         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4375         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4376         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4377         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4378
4379         update_global_block_rsv(fs_info);
4380 }
4381
4382 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4383 {
4384         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4385                                 (u64)-1);
4386         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4387         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4388         WARN_ON(fs_info->trans_block_rsv.size > 0);
4389         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4390         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4391         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4392         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4393         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4394 }
4395
4396 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4397                                   struct btrfs_root *root)
4398 {
4399         if (!trans->block_rsv)
4400                 return;
4401
4402         if (!trans->bytes_reserved)
4403                 return;
4404
4405         trace_btrfs_space_reservation(root->fs_info, "transaction",
4406                                       trans->transid, trans->bytes_reserved, 0);
4407         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4408         trans->bytes_reserved = 0;
4409 }
4410
4411 /* Can only return 0 or -ENOSPC */
4412 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4413                                   struct inode *inode)
4414 {
4415         struct btrfs_root *root = BTRFS_I(inode)->root;
4416         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4417         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4418
4419         /*
4420          * We need to hold space in order to delete our orphan item once we've
4421          * added it, so this takes the reservation so we can release it later
4422          * when we are truly done with the orphan item.
4423          */
4424         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4425         trace_btrfs_space_reservation(root->fs_info, "orphan",
4426                                       btrfs_ino(inode), num_bytes, 1);
4427         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4428 }
4429
4430 void btrfs_orphan_release_metadata(struct inode *inode)
4431 {
4432         struct btrfs_root *root = BTRFS_I(inode)->root;
4433         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4434         trace_btrfs_space_reservation(root->fs_info, "orphan",
4435                                       btrfs_ino(inode), num_bytes, 0);
4436         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4437 }
4438
4439 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4440                                 struct btrfs_pending_snapshot *pending)
4441 {
4442         struct btrfs_root *root = pending->root;
4443         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4444         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4445         /*
4446          * two for root back/forward refs, two for directory entries,
4447          * one for root of the snapshot and one for parent inode.
4448          */
4449         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 6);
4450         dst_rsv->space_info = src_rsv->space_info;
4451         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4452 }
4453
4454 /**
4455  * drop_outstanding_extent - drop an outstanding extent
4456  * @inode: the inode we're dropping the extent for
4457  *
4458  * This is called when we are freeing up an outstanding extent, either called
4459  * after an error or after an extent is written.  This will return the number of
4460  * reserved extents that need to be freed.  This must be called with
4461  * BTRFS_I(inode)->lock held.
4462  */
4463 static unsigned drop_outstanding_extent(struct inode *inode)
4464 {
4465         unsigned drop_inode_space = 0;
4466         unsigned dropped_extents = 0;
4467
4468         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4469         BTRFS_I(inode)->outstanding_extents--;
4470
4471         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4472             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4473                                &BTRFS_I(inode)->runtime_flags))
4474                 drop_inode_space = 1;
4475
4476         /*
4477          * If we have more or the same amount of outsanding extents than we have
4478          * reserved then we need to leave the reserved extents count alone.
4479          */
4480         if (BTRFS_I(inode)->outstanding_extents >=
4481             BTRFS_I(inode)->reserved_extents)
4482                 return drop_inode_space;
4483
4484         dropped_extents = BTRFS_I(inode)->reserved_extents -
4485                 BTRFS_I(inode)->outstanding_extents;
4486         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4487         return dropped_extents + drop_inode_space;
4488 }
4489
4490 /**
4491  * calc_csum_metadata_size - return the amount of metada space that must be
4492  *      reserved/free'd for the given bytes.
4493  * @inode: the inode we're manipulating
4494  * @num_bytes: the number of bytes in question
4495  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4496  *
4497  * This adjusts the number of csum_bytes in the inode and then returns the
4498  * correct amount of metadata that must either be reserved or freed.  We
4499  * calculate how many checksums we can fit into one leaf and then divide the
4500  * number of bytes that will need to be checksumed by this value to figure out
4501  * how many checksums will be required.  If we are adding bytes then the number
4502  * may go up and we will return the number of additional bytes that must be
4503  * reserved.  If it is going down we will return the number of bytes that must
4504  * be freed.
4505  *
4506  * This must be called with BTRFS_I(inode)->lock held.
4507  */
4508 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4509                                    int reserve)
4510 {
4511         struct btrfs_root *root = BTRFS_I(inode)->root;
4512         u64 csum_size;
4513         int num_csums_per_leaf;
4514         int num_csums;
4515         int old_csums;
4516
4517         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4518             BTRFS_I(inode)->csum_bytes == 0)
4519                 return 0;
4520
4521         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4522         if (reserve)
4523                 BTRFS_I(inode)->csum_bytes += num_bytes;
4524         else
4525                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4526         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4527         num_csums_per_leaf = (int)div64_u64(csum_size,
4528                                             sizeof(struct btrfs_csum_item) +
4529                                             sizeof(struct btrfs_disk_key));
4530         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4531         num_csums = num_csums + num_csums_per_leaf - 1;
4532         num_csums = num_csums / num_csums_per_leaf;
4533
4534         old_csums = old_csums + num_csums_per_leaf - 1;
4535         old_csums = old_csums / num_csums_per_leaf;
4536
4537         /* No change, no need to reserve more */
4538         if (old_csums == num_csums)
4539                 return 0;
4540
4541         if (reserve)
4542                 return btrfs_calc_trans_metadata_size(root,
4543                                                       num_csums - old_csums);
4544
4545         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4546 }
4547
4548 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4549 {
4550         struct btrfs_root *root = BTRFS_I(inode)->root;
4551         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4552         u64 to_reserve = 0;
4553         u64 csum_bytes;
4554         unsigned nr_extents = 0;
4555         int extra_reserve = 0;
4556         int flush = 1;
4557         int ret;
4558
4559         /* Need to be holding the i_mutex here if we aren't free space cache */
4560         if (btrfs_is_free_space_inode(inode))
4561                 flush = 0;
4562
4563         if (flush && btrfs_transaction_in_commit(root->fs_info))
4564                 schedule_timeout(1);
4565
4566         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4567         num_bytes = ALIGN(num_bytes, root->sectorsize);
4568
4569         spin_lock(&BTRFS_I(inode)->lock);
4570         BTRFS_I(inode)->outstanding_extents++;
4571
4572         if (BTRFS_I(inode)->outstanding_extents >
4573             BTRFS_I(inode)->reserved_extents)
4574                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4575                         BTRFS_I(inode)->reserved_extents;
4576
4577         /*
4578          * Add an item to reserve for updating the inode when we complete the
4579          * delalloc io.
4580          */
4581         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4582                       &BTRFS_I(inode)->runtime_flags)) {
4583                 nr_extents++;
4584                 extra_reserve = 1;
4585         }
4586
4587         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4588         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4589         csum_bytes = BTRFS_I(inode)->csum_bytes;
4590         spin_unlock(&BTRFS_I(inode)->lock);
4591
4592         if (root->fs_info->quota_enabled) {
4593                 ret = btrfs_qgroup_reserve(root, num_bytes +
4594                                            nr_extents * root->leafsize);
4595                 if (ret) {
4596                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4597                         return ret;
4598                 }
4599         }
4600
4601         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4602         if (ret) {
4603                 u64 to_free = 0;
4604                 unsigned dropped;
4605
4606                 spin_lock(&BTRFS_I(inode)->lock);
4607                 dropped = drop_outstanding_extent(inode);
4608                 /*
4609                  * If the inodes csum_bytes is the same as the original
4610                  * csum_bytes then we know we haven't raced with any free()ers
4611                  * so we can just reduce our inodes csum bytes and carry on.
4612                  * Otherwise we have to do the normal free thing to account for
4613                  * the case that the free side didn't free up its reserve
4614                  * because of this outstanding reservation.
4615                  */
4616                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4617                         calc_csum_metadata_size(inode, num_bytes, 0);
4618                 else
4619                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4620                 spin_unlock(&BTRFS_I(inode)->lock);
4621                 if (dropped)
4622                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4623
4624                 if (to_free) {
4625                         btrfs_block_rsv_release(root, block_rsv, to_free);
4626                         trace_btrfs_space_reservation(root->fs_info,
4627                                                       "delalloc",
4628                                                       btrfs_ino(inode),
4629                                                       to_free, 0);
4630                 }
4631                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4632                 return ret;
4633         }
4634
4635         spin_lock(&BTRFS_I(inode)->lock);
4636         if (extra_reserve) {
4637                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4638                         &BTRFS_I(inode)->runtime_flags);
4639                 nr_extents--;
4640         }
4641         BTRFS_I(inode)->reserved_extents += nr_extents;
4642         spin_unlock(&BTRFS_I(inode)->lock);
4643         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4644
4645         if (to_reserve)
4646                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4647                                               btrfs_ino(inode), to_reserve, 1);
4648         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4649
4650         return 0;
4651 }
4652
4653 /**
4654  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4655  * @inode: the inode to release the reservation for
4656  * @num_bytes: the number of bytes we're releasing
4657  *
4658  * This will release the metadata reservation for an inode.  This can be called
4659  * once we complete IO for a given set of bytes to release their metadata
4660  * reservations.
4661  */
4662 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4663 {
4664         struct btrfs_root *root = BTRFS_I(inode)->root;
4665         u64 to_free = 0;
4666         unsigned dropped;
4667
4668         num_bytes = ALIGN(num_bytes, root->sectorsize);
4669         spin_lock(&BTRFS_I(inode)->lock);
4670         dropped = drop_outstanding_extent(inode);
4671
4672         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4673         spin_unlock(&BTRFS_I(inode)->lock);
4674         if (dropped > 0)
4675                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4676
4677         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4678                                       btrfs_ino(inode), to_free, 0);
4679         if (root->fs_info->quota_enabled) {
4680                 btrfs_qgroup_free(root, num_bytes +
4681                                         dropped * root->leafsize);
4682         }
4683
4684         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4685                                 to_free);
4686 }
4687
4688 /**
4689  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4690  * @inode: inode we're writing to
4691  * @num_bytes: the number of bytes we want to allocate
4692  *
4693  * This will do the following things
4694  *
4695  * o reserve space in the data space info for num_bytes
4696  * o reserve space in the metadata space info based on number of outstanding
4697  *   extents and how much csums will be needed
4698  * o add to the inodes ->delalloc_bytes
4699  * o add it to the fs_info's delalloc inodes list.
4700  *
4701  * This will return 0 for success and -ENOSPC if there is no space left.
4702  */
4703 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4704 {
4705         int ret;
4706
4707         ret = btrfs_check_data_free_space(inode, num_bytes);
4708         if (ret)
4709                 return ret;
4710
4711         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4712         if (ret) {
4713                 btrfs_free_reserved_data_space(inode, num_bytes);
4714                 return ret;
4715         }
4716
4717         return 0;
4718 }
4719
4720 /**
4721  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4722  * @inode: inode we're releasing space for
4723  * @num_bytes: the number of bytes we want to free up
4724  *
4725  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4726  * called in the case that we don't need the metadata AND data reservations
4727  * anymore.  So if there is an error or we insert an inline extent.
4728  *
4729  * This function will release the metadata space that was not used and will
4730  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4731  * list if there are no delalloc bytes left.
4732  */
4733 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4734 {
4735         btrfs_delalloc_release_metadata(inode, num_bytes);
4736         btrfs_free_reserved_data_space(inode, num_bytes);
4737 }
4738
4739 static int update_block_group(struct btrfs_trans_handle *trans,
4740                               struct btrfs_root *root,
4741                               u64 bytenr, u64 num_bytes, int alloc)
4742 {
4743         struct btrfs_block_group_cache *cache = NULL;
4744         struct btrfs_fs_info *info = root->fs_info;
4745         u64 total = num_bytes;
4746         u64 old_val;
4747         u64 byte_in_group;
4748         int factor;
4749
4750         /* block accounting for super block */
4751         spin_lock(&info->delalloc_lock);
4752         old_val = btrfs_super_bytes_used(info->super_copy);
4753         if (alloc)
4754                 old_val += num_bytes;
4755         else
4756                 old_val -= num_bytes;
4757         btrfs_set_super_bytes_used(info->super_copy, old_val);
4758         spin_unlock(&info->delalloc_lock);
4759
4760         while (total) {
4761                 cache = btrfs_lookup_block_group(info, bytenr);
4762                 if (!cache)
4763                         return -ENOENT;
4764                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4765                                     BTRFS_BLOCK_GROUP_RAID1 |
4766                                     BTRFS_BLOCK_GROUP_RAID10))
4767                         factor = 2;
4768                 else
4769                         factor = 1;
4770                 /*
4771                  * If this block group has free space cache written out, we
4772                  * need to make sure to load it if we are removing space.  This
4773                  * is because we need the unpinning stage to actually add the
4774                  * space back to the block group, otherwise we will leak space.
4775                  */
4776                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4777                         cache_block_group(cache, trans, NULL, 1);
4778
4779                 byte_in_group = bytenr - cache->key.objectid;
4780                 WARN_ON(byte_in_group > cache->key.offset);
4781
4782                 spin_lock(&cache->space_info->lock);
4783                 spin_lock(&cache->lock);
4784
4785                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4786                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4787                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4788
4789                 cache->dirty = 1;
4790                 old_val = btrfs_block_group_used(&cache->item);
4791                 num_bytes = min(total, cache->key.offset - byte_in_group);
4792                 if (alloc) {
4793                         old_val += num_bytes;
4794                         btrfs_set_block_group_used(&cache->item, old_val);
4795                         cache->reserved -= num_bytes;
4796                         cache->space_info->bytes_reserved -= num_bytes;
4797                         cache->space_info->bytes_used += num_bytes;
4798                         cache->space_info->disk_used += num_bytes * factor;
4799                         spin_unlock(&cache->lock);
4800                         spin_unlock(&cache->space_info->lock);
4801                 } else {
4802                         old_val -= num_bytes;
4803                         btrfs_set_block_group_used(&cache->item, old_val);
4804                         cache->pinned += num_bytes;
4805                         cache->space_info->bytes_pinned += num_bytes;
4806                         cache->space_info->bytes_used -= num_bytes;
4807                         cache->space_info->disk_used -= num_bytes * factor;
4808                         spin_unlock(&cache->lock);
4809                         spin_unlock(&cache->space_info->lock);
4810
4811                         set_extent_dirty(info->pinned_extents,
4812                                          bytenr, bytenr + num_bytes - 1,
4813                                          GFP_NOFS | __GFP_NOFAIL);
4814                 }
4815                 btrfs_put_block_group(cache);
4816                 total -= num_bytes;
4817                 bytenr += num_bytes;
4818         }
4819         return 0;
4820 }
4821
4822 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4823 {
4824         struct btrfs_block_group_cache *cache;
4825         u64 bytenr;
4826
4827         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4828         if (!cache)
4829                 return 0;
4830
4831         bytenr = cache->key.objectid;
4832         btrfs_put_block_group(cache);
4833
4834         return bytenr;
4835 }
4836
4837 static int pin_down_extent(struct btrfs_root *root,
4838                            struct btrfs_block_group_cache *cache,
4839                            u64 bytenr, u64 num_bytes, int reserved)
4840 {
4841         spin_lock(&cache->space_info->lock);
4842         spin_lock(&cache->lock);
4843         cache->pinned += num_bytes;
4844         cache->space_info->bytes_pinned += num_bytes;
4845         if (reserved) {
4846                 cache->reserved -= num_bytes;
4847                 cache->space_info->bytes_reserved -= num_bytes;
4848         }
4849         spin_unlock(&cache->lock);
4850         spin_unlock(&cache->space_info->lock);
4851
4852         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4853                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4854         return 0;
4855 }
4856
4857 /*
4858  * this function must be called within transaction
4859  */
4860 int btrfs_pin_extent(struct btrfs_root *root,
4861                      u64 bytenr, u64 num_bytes, int reserved)
4862 {
4863         struct btrfs_block_group_cache *cache;
4864
4865         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4866         BUG_ON(!cache); /* Logic error */
4867
4868         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4869
4870         btrfs_put_block_group(cache);
4871         return 0;
4872 }
4873
4874 /*
4875  * this function must be called within transaction
4876  */
4877 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4878                                     struct btrfs_root *root,
4879                                     u64 bytenr, u64 num_bytes)
4880 {
4881         struct btrfs_block_group_cache *cache;
4882
4883         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4884         BUG_ON(!cache); /* Logic error */
4885
4886         /*
4887          * pull in the free space cache (if any) so that our pin
4888          * removes the free space from the cache.  We have load_only set
4889          * to one because the slow code to read in the free extents does check
4890          * the pinned extents.
4891          */
4892         cache_block_group(cache, trans, root, 1);
4893
4894         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4895
4896         /* remove us from the free space cache (if we're there at all) */
4897         btrfs_remove_free_space(cache, bytenr, num_bytes);
4898         btrfs_put_block_group(cache);
4899         return 0;
4900 }
4901
4902 /**
4903  * btrfs_update_reserved_bytes - update the block_group and space info counters
4904  * @cache:      The cache we are manipulating
4905  * @num_bytes:  The number of bytes in question
4906  * @reserve:    One of the reservation enums
4907  *
4908  * This is called by the allocator when it reserves space, or by somebody who is
4909  * freeing space that was never actually used on disk.  For example if you
4910  * reserve some space for a new leaf in transaction A and before transaction A
4911  * commits you free that leaf, you call this with reserve set to 0 in order to
4912  * clear the reservation.
4913  *
4914  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4915  * ENOSPC accounting.  For data we handle the reservation through clearing the
4916  * delalloc bits in the io_tree.  We have to do this since we could end up
4917  * allocating less disk space for the amount of data we have reserved in the
4918  * case of compression.
4919  *
4920  * If this is a reservation and the block group has become read only we cannot
4921  * make the reservation and return -EAGAIN, otherwise this function always
4922  * succeeds.
4923  */
4924 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4925                                        u64 num_bytes, int reserve)
4926 {
4927         struct btrfs_space_info *space_info = cache->space_info;
4928         int ret = 0;
4929
4930         spin_lock(&space_info->lock);
4931         spin_lock(&cache->lock);
4932         if (reserve != RESERVE_FREE) {
4933                 if (cache->ro) {
4934                         ret = -EAGAIN;
4935                 } else {
4936                         cache->reserved += num_bytes;
4937                         space_info->bytes_reserved += num_bytes;
4938                         if (reserve == RESERVE_ALLOC) {
4939                                 trace_btrfs_space_reservation(cache->fs_info,
4940                                                 "space_info", space_info->flags,
4941                                                 num_bytes, 0);
4942                                 space_info->bytes_may_use -= num_bytes;
4943                         }
4944                 }
4945         } else {
4946                 if (cache->ro)
4947                         space_info->bytes_readonly += num_bytes;
4948                 cache->reserved -= num_bytes;
4949                 space_info->bytes_reserved -= num_bytes;
4950                 space_info->reservation_progress++;
4951         }
4952         spin_unlock(&cache->lock);
4953         spin_unlock(&space_info->lock);
4954         return ret;
4955 }
4956
4957 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4958                                 struct btrfs_root *root)
4959 {
4960         struct btrfs_fs_info *fs_info = root->fs_info;
4961         struct btrfs_caching_control *next;
4962         struct btrfs_caching_control *caching_ctl;
4963         struct btrfs_block_group_cache *cache;
4964
4965         down_write(&fs_info->extent_commit_sem);
4966
4967         list_for_each_entry_safe(caching_ctl, next,
4968                                  &fs_info->caching_block_groups, list) {
4969                 cache = caching_ctl->block_group;
4970                 if (block_group_cache_done(cache)) {
4971                         cache->last_byte_to_unpin = (u64)-1;
4972                         list_del_init(&caching_ctl->list);
4973                         put_caching_control(caching_ctl);
4974                 } else {
4975                         cache->last_byte_to_unpin = caching_ctl->progress;
4976                 }
4977         }
4978
4979         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4980                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4981         else
4982                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4983
4984         up_write(&fs_info->extent_commit_sem);
4985
4986         update_global_block_rsv(fs_info);
4987 }
4988
4989 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4990 {
4991         struct btrfs_fs_info *fs_info = root->fs_info;
4992         struct btrfs_block_group_cache *cache = NULL;
4993         u64 len;
4994
4995         while (start <= end) {
4996                 if (!cache ||
4997                     start >= cache->key.objectid + cache->key.offset) {
4998                         if (cache)
4999                                 btrfs_put_block_group(cache);
5000                         cache = btrfs_lookup_block_group(fs_info, start);
5001                         BUG_ON(!cache); /* Logic error */
5002                 }
5003
5004                 len = cache->key.objectid + cache->key.offset - start;
5005                 len = min(len, end + 1 - start);
5006
5007                 if (start < cache->last_byte_to_unpin) {
5008                         len = min(len, cache->last_byte_to_unpin - start);
5009                         btrfs_add_free_space(cache, start, len);
5010                 }
5011
5012                 start += len;
5013
5014                 spin_lock(&cache->space_info->lock);
5015                 spin_lock(&cache->lock);
5016                 cache->pinned -= len;
5017                 cache->space_info->bytes_pinned -= len;
5018                 if (cache->ro)
5019                         cache->space_info->bytes_readonly += len;
5020                 spin_unlock(&cache->lock);
5021                 spin_unlock(&cache->space_info->lock);
5022         }
5023
5024         if (cache)
5025                 btrfs_put_block_group(cache);
5026         return 0;
5027 }
5028
5029 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5030                                struct btrfs_root *root)
5031 {
5032         struct btrfs_fs_info *fs_info = root->fs_info;
5033         struct extent_io_tree *unpin;
5034         u64 start;
5035         u64 end;
5036         int ret;
5037
5038         if (trans->aborted)
5039                 return 0;
5040
5041         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5042                 unpin = &fs_info->freed_extents[1];
5043         else
5044                 unpin = &fs_info->freed_extents[0];
5045
5046         while (1) {
5047                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5048                                             EXTENT_DIRTY);
5049                 if (ret)
5050                         break;
5051
5052                 if (btrfs_test_opt(root, DISCARD))
5053                         ret = btrfs_discard_extent(root, start,
5054                                                    end + 1 - start, NULL);
5055
5056                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5057                 unpin_extent_range(root, start, end);
5058                 cond_resched();
5059         }
5060
5061         return 0;
5062 }
5063
5064 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5065                                 struct btrfs_root *root,
5066                                 u64 bytenr, u64 num_bytes, u64 parent,
5067                                 u64 root_objectid, u64 owner_objectid,
5068                                 u64 owner_offset, int refs_to_drop,
5069                                 struct btrfs_delayed_extent_op *extent_op)
5070 {
5071         struct btrfs_key key;
5072         struct btrfs_path *path;
5073         struct btrfs_fs_info *info = root->fs_info;
5074         struct btrfs_root *extent_root = info->extent_root;
5075         struct extent_buffer *leaf;
5076         struct btrfs_extent_item *ei;
5077         struct btrfs_extent_inline_ref *iref;
5078         int ret;
5079         int is_data;
5080         int extent_slot = 0;
5081         int found_extent = 0;
5082         int num_to_del = 1;
5083         u32 item_size;
5084         u64 refs;
5085
5086         path = btrfs_alloc_path();
5087         if (!path)
5088                 return -ENOMEM;
5089
5090         path->reada = 1;
5091         path->leave_spinning = 1;
5092
5093         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5094         BUG_ON(!is_data && refs_to_drop != 1);
5095
5096         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5097                                     bytenr, num_bytes, parent,
5098                                     root_objectid, owner_objectid,
5099                                     owner_offset);
5100         if (ret == 0) {
5101                 extent_slot = path->slots[0];
5102                 while (extent_slot >= 0) {
5103                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5104                                               extent_slot);
5105                         if (key.objectid != bytenr)
5106                                 break;
5107                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5108                             key.offset == num_bytes) {
5109                                 found_extent = 1;
5110                                 break;
5111                         }
5112                         if (path->slots[0] - extent_slot > 5)
5113                                 break;
5114                         extent_slot--;
5115                 }
5116 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5117                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5118                 if (found_extent && item_size < sizeof(*ei))
5119                         found_extent = 0;
5120 #endif
5121                 if (!found_extent) {
5122                         BUG_ON(iref);
5123                         ret = remove_extent_backref(trans, extent_root, path,
5124                                                     NULL, refs_to_drop,
5125                                                     is_data);
5126                         if (ret)
5127                                 goto abort;
5128                         btrfs_release_path(path);
5129                         path->leave_spinning = 1;
5130
5131                         key.objectid = bytenr;
5132                         key.type = BTRFS_EXTENT_ITEM_KEY;
5133                         key.offset = num_bytes;
5134
5135                         ret = btrfs_search_slot(trans, extent_root,
5136                                                 &key, path, -1, 1);
5137                         if (ret) {
5138                                 printk(KERN_ERR "umm, got %d back from search"
5139                                        ", was looking for %llu\n", ret,
5140                                        (unsigned long long)bytenr);
5141                                 if (ret > 0)
5142                                         btrfs_print_leaf(extent_root,
5143                                                          path->nodes[0]);
5144                         }
5145                         if (ret < 0)
5146                                 goto abort;
5147                         extent_slot = path->slots[0];
5148                 }
5149         } else if (ret == -ENOENT) {
5150                 btrfs_print_leaf(extent_root, path->nodes[0]);
5151                 WARN_ON(1);
5152                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5153                        "parent %llu root %llu  owner %llu offset %llu\n",
5154                        (unsigned long long)bytenr,
5155                        (unsigned long long)parent,
5156                        (unsigned long long)root_objectid,
5157                        (unsigned long long)owner_objectid,
5158                        (unsigned long long)owner_offset);
5159         } else {
5160                 goto abort;
5161         }
5162
5163         leaf = path->nodes[0];
5164         item_size = btrfs_item_size_nr(leaf, extent_slot);
5165 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5166         if (item_size < sizeof(*ei)) {
5167                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5168                 ret = convert_extent_item_v0(trans, extent_root, path,
5169                                              owner_objectid, 0);
5170                 if (ret < 0)
5171                         goto abort;
5172
5173                 btrfs_release_path(path);
5174                 path->leave_spinning = 1;
5175
5176                 key.objectid = bytenr;
5177                 key.type = BTRFS_EXTENT_ITEM_KEY;
5178                 key.offset = num_bytes;
5179
5180                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5181                                         -1, 1);
5182                 if (ret) {
5183                         printk(KERN_ERR "umm, got %d back from search"
5184                                ", was looking for %llu\n", ret,
5185                                (unsigned long long)bytenr);
5186                         btrfs_print_leaf(extent_root, path->nodes[0]);
5187                 }
5188                 if (ret < 0)
5189                         goto abort;
5190                 extent_slot = path->slots[0];
5191                 leaf = path->nodes[0];
5192                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5193         }
5194 #endif
5195         BUG_ON(item_size < sizeof(*ei));
5196         ei = btrfs_item_ptr(leaf, extent_slot,
5197                             struct btrfs_extent_item);
5198         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5199                 struct btrfs_tree_block_info *bi;
5200                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5201                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5202                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5203         }
5204
5205         refs = btrfs_extent_refs(leaf, ei);
5206         BUG_ON(refs < refs_to_drop);
5207         refs -= refs_to_drop;
5208
5209         if (refs > 0) {
5210                 if (extent_op)
5211                         __run_delayed_extent_op(extent_op, leaf, ei);
5212                 /*
5213                  * In the case of inline back ref, reference count will
5214                  * be updated by remove_extent_backref
5215                  */
5216                 if (iref) {
5217                         BUG_ON(!found_extent);
5218                 } else {
5219                         btrfs_set_extent_refs(leaf, ei, refs);
5220                         btrfs_mark_buffer_dirty(leaf);
5221                 }
5222                 if (found_extent) {
5223                         ret = remove_extent_backref(trans, extent_root, path,
5224                                                     iref, refs_to_drop,
5225                                                     is_data);
5226                         if (ret)
5227                                 goto abort;
5228                 }
5229         } else {
5230                 if (found_extent) {
5231                         BUG_ON(is_data && refs_to_drop !=
5232                                extent_data_ref_count(root, path, iref));
5233                         if (iref) {
5234                                 BUG_ON(path->slots[0] != extent_slot);
5235                         } else {
5236                                 BUG_ON(path->slots[0] != extent_slot + 1);
5237                                 path->slots[0] = extent_slot;
5238                                 num_to_del = 2;
5239                         }
5240                 }
5241
5242                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5243                                       num_to_del);
5244                 if (ret)
5245                         goto abort;
5246                 btrfs_release_path(path);
5247
5248                 if (is_data) {
5249                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5250                         if (ret)
5251                                 goto abort;
5252                 }
5253
5254                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5255                 if (ret)
5256                         goto abort;
5257         }
5258 out:
5259         btrfs_free_path(path);
5260         return ret;
5261
5262 abort:
5263         btrfs_abort_transaction(trans, extent_root, ret);
5264         goto out;
5265 }
5266
5267 /*
5268  * when we free an block, it is possible (and likely) that we free the last
5269  * delayed ref for that extent as well.  This searches the delayed ref tree for
5270  * a given extent, and if there are no other delayed refs to be processed, it
5271  * removes it from the tree.
5272  */
5273 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5274                                       struct btrfs_root *root, u64 bytenr)
5275 {
5276         struct btrfs_delayed_ref_head *head;
5277         struct btrfs_delayed_ref_root *delayed_refs;
5278         struct btrfs_delayed_ref_node *ref;
5279         struct rb_node *node;
5280         int ret = 0;
5281
5282         delayed_refs = &trans->transaction->delayed_refs;
5283         spin_lock(&delayed_refs->lock);
5284         head = btrfs_find_delayed_ref_head(trans, bytenr);
5285         if (!head)
5286                 goto out;
5287
5288         node = rb_prev(&head->node.rb_node);
5289         if (!node)
5290                 goto out;
5291
5292         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5293
5294         /* there are still entries for this ref, we can't drop it */
5295         if (ref->bytenr == bytenr)
5296                 goto out;
5297
5298         if (head->extent_op) {
5299                 if (!head->must_insert_reserved)
5300                         goto out;
5301                 kfree(head->extent_op);
5302                 head->extent_op = NULL;
5303         }
5304
5305         /*
5306          * waiting for the lock here would deadlock.  If someone else has it
5307          * locked they are already in the process of dropping it anyway
5308          */
5309         if (!mutex_trylock(&head->mutex))
5310                 goto out;
5311
5312         /*
5313          * at this point we have a head with no other entries.  Go
5314          * ahead and process it.
5315          */
5316         head->node.in_tree = 0;
5317         rb_erase(&head->node.rb_node, &delayed_refs->root);
5318
5319         delayed_refs->num_entries--;
5320
5321         /*
5322          * we don't take a ref on the node because we're removing it from the
5323          * tree, so we just steal the ref the tree was holding.
5324          */
5325         delayed_refs->num_heads--;
5326         if (list_empty(&head->cluster))
5327                 delayed_refs->num_heads_ready--;
5328
5329         list_del_init(&head->cluster);
5330         spin_unlock(&delayed_refs->lock);
5331
5332         BUG_ON(head->extent_op);
5333         if (head->must_insert_reserved)
5334                 ret = 1;
5335
5336         mutex_unlock(&head->mutex);
5337         btrfs_put_delayed_ref(&head->node);
5338         return ret;
5339 out:
5340         spin_unlock(&delayed_refs->lock);
5341         return 0;
5342 }
5343
5344 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5345                            struct btrfs_root *root,
5346                            struct extent_buffer *buf,
5347                            u64 parent, int last_ref)
5348 {
5349         struct btrfs_block_group_cache *cache = NULL;
5350         int ret;
5351
5352         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5353                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5354                                         buf->start, buf->len,
5355                                         parent, root->root_key.objectid,
5356                                         btrfs_header_level(buf),
5357                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5358                 BUG_ON(ret); /* -ENOMEM */
5359         }
5360
5361         if (!last_ref)
5362                 return;
5363
5364         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5365
5366         if (btrfs_header_generation(buf) == trans->transid) {
5367                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5368                         ret = check_ref_cleanup(trans, root, buf->start);
5369                         if (!ret)
5370                                 goto out;
5371                 }
5372
5373                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5374                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5375                         goto out;
5376                 }
5377
5378                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5379
5380                 btrfs_add_free_space(cache, buf->start, buf->len);
5381                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5382         }
5383 out:
5384         /*
5385          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5386          * anymore.
5387          */
5388         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5389         btrfs_put_block_group(cache);
5390 }
5391
5392 /* Can return -ENOMEM */
5393 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5394                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5395                       u64 owner, u64 offset, int for_cow)
5396 {
5397         int ret;
5398         struct btrfs_fs_info *fs_info = root->fs_info;
5399
5400         /*
5401          * tree log blocks never actually go into the extent allocation
5402          * tree, just update pinning info and exit early.
5403          */
5404         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5405                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5406                 /* unlocks the pinned mutex */
5407                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5408                 ret = 0;
5409         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5410                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5411                                         num_bytes,
5412                                         parent, root_objectid, (int)owner,
5413                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5414         } else {
5415                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5416                                                 num_bytes,
5417                                                 parent, root_objectid, owner,
5418                                                 offset, BTRFS_DROP_DELAYED_REF,
5419                                                 NULL, for_cow);
5420         }
5421         return ret;
5422 }
5423
5424 static u64 stripe_align(struct btrfs_root *root, u64 val)
5425 {
5426         u64 mask = ((u64)root->stripesize - 1);
5427         u64 ret = (val + mask) & ~mask;
5428         return ret;
5429 }
5430
5431 /*
5432  * when we wait for progress in the block group caching, its because
5433  * our allocation attempt failed at least once.  So, we must sleep
5434  * and let some progress happen before we try again.
5435  *
5436  * This function will sleep at least once waiting for new free space to
5437  * show up, and then it will check the block group free space numbers
5438  * for our min num_bytes.  Another option is to have it go ahead
5439  * and look in the rbtree for a free extent of a given size, but this
5440  * is a good start.
5441  */
5442 static noinline int
5443 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5444                                 u64 num_bytes)
5445 {
5446         struct btrfs_caching_control *caching_ctl;
5447         DEFINE_WAIT(wait);
5448
5449         caching_ctl = get_caching_control(cache);
5450         if (!caching_ctl)
5451                 return 0;
5452
5453         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5454                    (cache->free_space_ctl->free_space >= num_bytes));
5455
5456         put_caching_control(caching_ctl);
5457         return 0;
5458 }
5459
5460 static noinline int
5461 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5462 {
5463         struct btrfs_caching_control *caching_ctl;
5464         DEFINE_WAIT(wait);
5465
5466         caching_ctl = get_caching_control(cache);
5467         if (!caching_ctl)
5468                 return 0;
5469
5470         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5471
5472         put_caching_control(caching_ctl);
5473         return 0;
5474 }
5475
5476 static int __get_block_group_index(u64 flags)
5477 {
5478         int index;
5479
5480         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5481                 index = 0;
5482         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5483                 index = 1;
5484         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5485                 index = 2;
5486         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5487                 index = 3;
5488         else
5489                 index = 4;
5490
5491         return index;
5492 }
5493
5494 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5495 {
5496         return __get_block_group_index(cache->flags);
5497 }
5498
5499 enum btrfs_loop_type {
5500         LOOP_CACHING_NOWAIT = 0,
5501         LOOP_CACHING_WAIT = 1,
5502         LOOP_ALLOC_CHUNK = 2,
5503         LOOP_NO_EMPTY_SIZE = 3,
5504 };
5505
5506 /*
5507  * walks the btree of allocated extents and find a hole of a given size.
5508  * The key ins is changed to record the hole:
5509  * ins->objectid == block start
5510  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5511  * ins->offset == number of blocks
5512  * Any available blocks before search_start are skipped.
5513  */
5514 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5515                                      struct btrfs_root *orig_root,
5516                                      u64 num_bytes, u64 empty_size,
5517                                      u64 hint_byte, struct btrfs_key *ins,
5518                                      u64 data)
5519 {
5520         int ret = 0;
5521         struct btrfs_root *root = orig_root->fs_info->extent_root;
5522         struct btrfs_free_cluster *last_ptr = NULL;
5523         struct btrfs_block_group_cache *block_group = NULL;
5524         struct btrfs_block_group_cache *used_block_group;
5525         u64 search_start = 0;
5526         int empty_cluster = 2 * 1024 * 1024;
5527         struct btrfs_space_info *space_info;
5528         int loop = 0;
5529         int index = 0;
5530         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5531                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5532         bool found_uncached_bg = false;
5533         bool failed_cluster_refill = false;
5534         bool failed_alloc = false;
5535         bool use_cluster = true;
5536         bool have_caching_bg = false;
5537
5538         WARN_ON(num_bytes < root->sectorsize);
5539         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5540         ins->objectid = 0;
5541         ins->offset = 0;
5542
5543         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5544
5545         space_info = __find_space_info(root->fs_info, data);
5546         if (!space_info) {
5547                 printk(KERN_ERR "No space info for %llu\n", data);
5548                 return -ENOSPC;
5549         }
5550
5551         /*
5552          * If the space info is for both data and metadata it means we have a
5553          * small filesystem and we can't use the clustering stuff.
5554          */
5555         if (btrfs_mixed_space_info(space_info))
5556                 use_cluster = false;
5557
5558         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5559                 last_ptr = &root->fs_info->meta_alloc_cluster;
5560                 if (!btrfs_test_opt(root, SSD))
5561                         empty_cluster = 64 * 1024;
5562         }
5563
5564         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5565             btrfs_test_opt(root, SSD)) {
5566                 last_ptr = &root->fs_info->data_alloc_cluster;
5567         }
5568
5569         if (last_ptr) {
5570                 spin_lock(&last_ptr->lock);
5571                 if (last_ptr->block_group)
5572                         hint_byte = last_ptr->window_start;
5573                 spin_unlock(&last_ptr->lock);
5574         }
5575
5576         search_start = max(search_start, first_logical_byte(root, 0));
5577         search_start = max(search_start, hint_byte);
5578
5579         if (!last_ptr)
5580                 empty_cluster = 0;
5581
5582         if (search_start == hint_byte) {
5583                 block_group = btrfs_lookup_block_group(root->fs_info,
5584                                                        search_start);
5585                 used_block_group = block_group;
5586                 /*
5587                  * we don't want to use the block group if it doesn't match our
5588                  * allocation bits, or if its not cached.
5589                  *
5590                  * However if we are re-searching with an ideal block group
5591                  * picked out then we don't care that the block group is cached.
5592                  */
5593                 if (block_group && block_group_bits(block_group, data) &&
5594                     block_group->cached != BTRFS_CACHE_NO) {
5595                         down_read(&space_info->groups_sem);
5596                         if (list_empty(&block_group->list) ||
5597                             block_group->ro) {
5598                                 /*
5599                                  * someone is removing this block group,
5600                                  * we can't jump into the have_block_group
5601                                  * target because our list pointers are not
5602                                  * valid
5603                                  */
5604                                 btrfs_put_block_group(block_group);
5605                                 up_read(&space_info->groups_sem);
5606                         } else {
5607                                 index = get_block_group_index(block_group);
5608                                 goto have_block_group;
5609                         }
5610                 } else if (block_group) {
5611                         btrfs_put_block_group(block_group);
5612                 }
5613         }
5614 search:
5615         have_caching_bg = false;
5616         down_read(&space_info->groups_sem);
5617         list_for_each_entry(block_group, &space_info->block_groups[index],
5618                             list) {
5619                 u64 offset;
5620                 int cached;
5621
5622                 used_block_group = block_group;
5623                 btrfs_get_block_group(block_group);
5624                 search_start = block_group->key.objectid;
5625
5626                 /*
5627                  * this can happen if we end up cycling through all the
5628                  * raid types, but we want to make sure we only allocate
5629                  * for the proper type.
5630                  */
5631                 if (!block_group_bits(block_group, data)) {
5632                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5633                                 BTRFS_BLOCK_GROUP_RAID1 |
5634                                 BTRFS_BLOCK_GROUP_RAID10;
5635
5636                         /*
5637                          * if they asked for extra copies and this block group
5638                          * doesn't provide them, bail.  This does allow us to
5639                          * fill raid0 from raid1.
5640                          */
5641                         if ((data & extra) && !(block_group->flags & extra))
5642                                 goto loop;
5643                 }
5644
5645 have_block_group:
5646                 cached = block_group_cache_done(block_group);
5647                 if (unlikely(!cached)) {
5648                         found_uncached_bg = true;
5649                         ret = cache_block_group(block_group, trans,
5650                                                 orig_root, 0);
5651                         BUG_ON(ret < 0);
5652                         ret = 0;
5653                 }
5654
5655                 if (unlikely(block_group->ro))
5656                         goto loop;
5657
5658                 /*
5659                  * Ok we want to try and use the cluster allocator, so
5660                  * lets look there
5661                  */
5662                 if (last_ptr) {
5663                         /*
5664                          * the refill lock keeps out other
5665                          * people trying to start a new cluster
5666                          */
5667                         spin_lock(&last_ptr->refill_lock);
5668                         used_block_group = last_ptr->block_group;
5669                         if (used_block_group != block_group &&
5670                             (!used_block_group ||
5671                              used_block_group->ro ||
5672                              !block_group_bits(used_block_group, data))) {
5673                                 used_block_group = block_group;
5674                                 goto refill_cluster;
5675                         }
5676
5677                         if (used_block_group != block_group)
5678                                 btrfs_get_block_group(used_block_group);
5679
5680                         offset = btrfs_alloc_from_cluster(used_block_group,
5681                           last_ptr, num_bytes, used_block_group->key.objectid);
5682                         if (offset) {
5683                                 /* we have a block, we're done */
5684