Btrfs: mark delayed refs as for cow
[linux-3.10.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36
37 /* control flags for do_chunk_alloc's force field
38  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
39  * if we really need one.
40  *
41  * CHUNK_ALLOC_FORCE means it must try to allocate one
42  *
43  * CHUNK_ALLOC_LIMITED means to only try and allocate one
44  * if we have very few chunks already allocated.  This is
45  * used as part of the clustering code to help make sure
46  * we have a good pool of storage to cluster in, without
47  * filling the FS with empty chunks
48  *
49  */
50 enum {
51         CHUNK_ALLOC_NO_FORCE = 0,
52         CHUNK_ALLOC_FORCE = 1,
53         CHUNK_ALLOC_LIMITED = 2,
54 };
55
56 /*
57  * Control how reservations are dealt with.
58  *
59  * RESERVE_FREE - freeing a reservation.
60  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
61  *   ENOSPC accounting
62  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
63  *   bytes_may_use as the ENOSPC accounting is done elsewhere
64  */
65 enum {
66         RESERVE_FREE = 0,
67         RESERVE_ALLOC = 1,
68         RESERVE_ALLOC_NO_ACCOUNT = 2,
69 };
70
71 static int update_block_group(struct btrfs_trans_handle *trans,
72                               struct btrfs_root *root,
73                               u64 bytenr, u64 num_bytes, int alloc);
74 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
75                                 struct btrfs_root *root,
76                                 u64 bytenr, u64 num_bytes, u64 parent,
77                                 u64 root_objectid, u64 owner_objectid,
78                                 u64 owner_offset, int refs_to_drop,
79                                 struct btrfs_delayed_extent_op *extra_op);
80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
81                                     struct extent_buffer *leaf,
82                                     struct btrfs_extent_item *ei);
83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
84                                       struct btrfs_root *root,
85                                       u64 parent, u64 root_objectid,
86                                       u64 flags, u64 owner, u64 offset,
87                                       struct btrfs_key *ins, int ref_mod);
88 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
89                                      struct btrfs_root *root,
90                                      u64 parent, u64 root_objectid,
91                                      u64 flags, struct btrfs_disk_key *key,
92                                      int level, struct btrfs_key *ins);
93 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
94                           struct btrfs_root *extent_root, u64 alloc_bytes,
95                           u64 flags, int force);
96 static int find_next_key(struct btrfs_path *path, int level,
97                          struct btrfs_key *key);
98 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
99                             int dump_block_groups);
100 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
101                                        u64 num_bytes, int reserve);
102
103 static noinline int
104 block_group_cache_done(struct btrfs_block_group_cache *cache)
105 {
106         smp_mb();
107         return cache->cached == BTRFS_CACHE_FINISHED;
108 }
109
110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
111 {
112         return (cache->flags & bits) == bits;
113 }
114
115 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
116 {
117         atomic_inc(&cache->count);
118 }
119
120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
121 {
122         if (atomic_dec_and_test(&cache->count)) {
123                 WARN_ON(cache->pinned > 0);
124                 WARN_ON(cache->reserved > 0);
125                 kfree(cache->free_space_ctl);
126                 kfree(cache);
127         }
128 }
129
130 /*
131  * this adds the block group to the fs_info rb tree for the block group
132  * cache
133  */
134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
135                                 struct btrfs_block_group_cache *block_group)
136 {
137         struct rb_node **p;
138         struct rb_node *parent = NULL;
139         struct btrfs_block_group_cache *cache;
140
141         spin_lock(&info->block_group_cache_lock);
142         p = &info->block_group_cache_tree.rb_node;
143
144         while (*p) {
145                 parent = *p;
146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
147                                  cache_node);
148                 if (block_group->key.objectid < cache->key.objectid) {
149                         p = &(*p)->rb_left;
150                 } else if (block_group->key.objectid > cache->key.objectid) {
151                         p = &(*p)->rb_right;
152                 } else {
153                         spin_unlock(&info->block_group_cache_lock);
154                         return -EEXIST;
155                 }
156         }
157
158         rb_link_node(&block_group->cache_node, parent, p);
159         rb_insert_color(&block_group->cache_node,
160                         &info->block_group_cache_tree);
161         spin_unlock(&info->block_group_cache_lock);
162
163         return 0;
164 }
165
166 /*
167  * This will return the block group at or after bytenr if contains is 0, else
168  * it will return the block group that contains the bytenr
169  */
170 static struct btrfs_block_group_cache *
171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
172                               int contains)
173 {
174         struct btrfs_block_group_cache *cache, *ret = NULL;
175         struct rb_node *n;
176         u64 end, start;
177
178         spin_lock(&info->block_group_cache_lock);
179         n = info->block_group_cache_tree.rb_node;
180
181         while (n) {
182                 cache = rb_entry(n, struct btrfs_block_group_cache,
183                                  cache_node);
184                 end = cache->key.objectid + cache->key.offset - 1;
185                 start = cache->key.objectid;
186
187                 if (bytenr < start) {
188                         if (!contains && (!ret || start < ret->key.objectid))
189                                 ret = cache;
190                         n = n->rb_left;
191                 } else if (bytenr > start) {
192                         if (contains && bytenr <= end) {
193                                 ret = cache;
194                                 break;
195                         }
196                         n = n->rb_right;
197                 } else {
198                         ret = cache;
199                         break;
200                 }
201         }
202         if (ret)
203                 btrfs_get_block_group(ret);
204         spin_unlock(&info->block_group_cache_lock);
205
206         return ret;
207 }
208
209 static int add_excluded_extent(struct btrfs_root *root,
210                                u64 start, u64 num_bytes)
211 {
212         u64 end = start + num_bytes - 1;
213         set_extent_bits(&root->fs_info->freed_extents[0],
214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
215         set_extent_bits(&root->fs_info->freed_extents[1],
216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
217         return 0;
218 }
219
220 static void free_excluded_extents(struct btrfs_root *root,
221                                   struct btrfs_block_group_cache *cache)
222 {
223         u64 start, end;
224
225         start = cache->key.objectid;
226         end = start + cache->key.offset - 1;
227
228         clear_extent_bits(&root->fs_info->freed_extents[0],
229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
230         clear_extent_bits(&root->fs_info->freed_extents[1],
231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
232 }
233
234 static int exclude_super_stripes(struct btrfs_root *root,
235                                  struct btrfs_block_group_cache *cache)
236 {
237         u64 bytenr;
238         u64 *logical;
239         int stripe_len;
240         int i, nr, ret;
241
242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244                 cache->bytes_super += stripe_len;
245                 ret = add_excluded_extent(root, cache->key.objectid,
246                                           stripe_len);
247                 BUG_ON(ret);
248         }
249
250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
251                 bytenr = btrfs_sb_offset(i);
252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
253                                        cache->key.objectid, bytenr,
254                                        0, &logical, &nr, &stripe_len);
255                 BUG_ON(ret);
256
257                 while (nr--) {
258                         cache->bytes_super += stripe_len;
259                         ret = add_excluded_extent(root, logical[nr],
260                                                   stripe_len);
261                         BUG_ON(ret);
262                 }
263
264                 kfree(logical);
265         }
266         return 0;
267 }
268
269 static struct btrfs_caching_control *
270 get_caching_control(struct btrfs_block_group_cache *cache)
271 {
272         struct btrfs_caching_control *ctl;
273
274         spin_lock(&cache->lock);
275         if (cache->cached != BTRFS_CACHE_STARTED) {
276                 spin_unlock(&cache->lock);
277                 return NULL;
278         }
279
280         /* We're loading it the fast way, so we don't have a caching_ctl. */
281         if (!cache->caching_ctl) {
282                 spin_unlock(&cache->lock);
283                 return NULL;
284         }
285
286         ctl = cache->caching_ctl;
287         atomic_inc(&ctl->count);
288         spin_unlock(&cache->lock);
289         return ctl;
290 }
291
292 static void put_caching_control(struct btrfs_caching_control *ctl)
293 {
294         if (atomic_dec_and_test(&ctl->count))
295                 kfree(ctl);
296 }
297
298 /*
299  * this is only called by cache_block_group, since we could have freed extents
300  * we need to check the pinned_extents for any extents that can't be used yet
301  * since their free space will be released as soon as the transaction commits.
302  */
303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
304                               struct btrfs_fs_info *info, u64 start, u64 end)
305 {
306         u64 extent_start, extent_end, size, total_added = 0;
307         int ret;
308
309         while (start < end) {
310                 ret = find_first_extent_bit(info->pinned_extents, start,
311                                             &extent_start, &extent_end,
312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
313                 if (ret)
314                         break;
315
316                 if (extent_start <= start) {
317                         start = extent_end + 1;
318                 } else if (extent_start > start && extent_start < end) {
319                         size = extent_start - start;
320                         total_added += size;
321                         ret = btrfs_add_free_space(block_group, start,
322                                                    size);
323                         BUG_ON(ret);
324                         start = extent_end + 1;
325                 } else {
326                         break;
327                 }
328         }
329
330         if (start < end) {
331                 size = end - start;
332                 total_added += size;
333                 ret = btrfs_add_free_space(block_group, start, size);
334                 BUG_ON(ret);
335         }
336
337         return total_added;
338 }
339
340 static noinline void caching_thread(struct btrfs_work *work)
341 {
342         struct btrfs_block_group_cache *block_group;
343         struct btrfs_fs_info *fs_info;
344         struct btrfs_caching_control *caching_ctl;
345         struct btrfs_root *extent_root;
346         struct btrfs_path *path;
347         struct extent_buffer *leaf;
348         struct btrfs_key key;
349         u64 total_found = 0;
350         u64 last = 0;
351         u32 nritems;
352         int ret = 0;
353
354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
355         block_group = caching_ctl->block_group;
356         fs_info = block_group->fs_info;
357         extent_root = fs_info->extent_root;
358
359         path = btrfs_alloc_path();
360         if (!path)
361                 goto out;
362
363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
364
365         /*
366          * We don't want to deadlock with somebody trying to allocate a new
367          * extent for the extent root while also trying to search the extent
368          * root to add free space.  So we skip locking and search the commit
369          * root, since its read-only
370          */
371         path->skip_locking = 1;
372         path->search_commit_root = 1;
373         path->reada = 1;
374
375         key.objectid = last;
376         key.offset = 0;
377         key.type = BTRFS_EXTENT_ITEM_KEY;
378 again:
379         mutex_lock(&caching_ctl->mutex);
380         /* need to make sure the commit_root doesn't disappear */
381         down_read(&fs_info->extent_commit_sem);
382
383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
384         if (ret < 0)
385                 goto err;
386
387         leaf = path->nodes[0];
388         nritems = btrfs_header_nritems(leaf);
389
390         while (1) {
391                 if (btrfs_fs_closing(fs_info) > 1) {
392                         last = (u64)-1;
393                         break;
394                 }
395
396                 if (path->slots[0] < nritems) {
397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
398                 } else {
399                         ret = find_next_key(path, 0, &key);
400                         if (ret)
401                                 break;
402
403                         if (need_resched() ||
404                             btrfs_next_leaf(extent_root, path)) {
405                                 caching_ctl->progress = last;
406                                 btrfs_release_path(path);
407                                 up_read(&fs_info->extent_commit_sem);
408                                 mutex_unlock(&caching_ctl->mutex);
409                                 cond_resched();
410                                 goto again;
411                         }
412                         leaf = path->nodes[0];
413                         nritems = btrfs_header_nritems(leaf);
414                         continue;
415                 }
416
417                 if (key.objectid < block_group->key.objectid) {
418                         path->slots[0]++;
419                         continue;
420                 }
421
422                 if (key.objectid >= block_group->key.objectid +
423                     block_group->key.offset)
424                         break;
425
426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
427                         total_found += add_new_free_space(block_group,
428                                                           fs_info, last,
429                                                           key.objectid);
430                         last = key.objectid + key.offset;
431
432                         if (total_found > (1024 * 1024 * 2)) {
433                                 total_found = 0;
434                                 wake_up(&caching_ctl->wait);
435                         }
436                 }
437                 path->slots[0]++;
438         }
439         ret = 0;
440
441         total_found += add_new_free_space(block_group, fs_info, last,
442                                           block_group->key.objectid +
443                                           block_group->key.offset);
444         caching_ctl->progress = (u64)-1;
445
446         spin_lock(&block_group->lock);
447         block_group->caching_ctl = NULL;
448         block_group->cached = BTRFS_CACHE_FINISHED;
449         spin_unlock(&block_group->lock);
450
451 err:
452         btrfs_free_path(path);
453         up_read(&fs_info->extent_commit_sem);
454
455         free_excluded_extents(extent_root, block_group);
456
457         mutex_unlock(&caching_ctl->mutex);
458 out:
459         wake_up(&caching_ctl->wait);
460
461         put_caching_control(caching_ctl);
462         btrfs_put_block_group(block_group);
463 }
464
465 static int cache_block_group(struct btrfs_block_group_cache *cache,
466                              struct btrfs_trans_handle *trans,
467                              struct btrfs_root *root,
468                              int load_cache_only)
469 {
470         DEFINE_WAIT(wait);
471         struct btrfs_fs_info *fs_info = cache->fs_info;
472         struct btrfs_caching_control *caching_ctl;
473         int ret = 0;
474
475         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
476         BUG_ON(!caching_ctl);
477
478         INIT_LIST_HEAD(&caching_ctl->list);
479         mutex_init(&caching_ctl->mutex);
480         init_waitqueue_head(&caching_ctl->wait);
481         caching_ctl->block_group = cache;
482         caching_ctl->progress = cache->key.objectid;
483         atomic_set(&caching_ctl->count, 1);
484         caching_ctl->work.func = caching_thread;
485
486         spin_lock(&cache->lock);
487         /*
488          * This should be a rare occasion, but this could happen I think in the
489          * case where one thread starts to load the space cache info, and then
490          * some other thread starts a transaction commit which tries to do an
491          * allocation while the other thread is still loading the space cache
492          * info.  The previous loop should have kept us from choosing this block
493          * group, but if we've moved to the state where we will wait on caching
494          * block groups we need to first check if we're doing a fast load here,
495          * so we can wait for it to finish, otherwise we could end up allocating
496          * from a block group who's cache gets evicted for one reason or
497          * another.
498          */
499         while (cache->cached == BTRFS_CACHE_FAST) {
500                 struct btrfs_caching_control *ctl;
501
502                 ctl = cache->caching_ctl;
503                 atomic_inc(&ctl->count);
504                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
505                 spin_unlock(&cache->lock);
506
507                 schedule();
508
509                 finish_wait(&ctl->wait, &wait);
510                 put_caching_control(ctl);
511                 spin_lock(&cache->lock);
512         }
513
514         if (cache->cached != BTRFS_CACHE_NO) {
515                 spin_unlock(&cache->lock);
516                 kfree(caching_ctl);
517                 return 0;
518         }
519         WARN_ON(cache->caching_ctl);
520         cache->caching_ctl = caching_ctl;
521         cache->cached = BTRFS_CACHE_FAST;
522         spin_unlock(&cache->lock);
523
524         /*
525          * We can't do the read from on-disk cache during a commit since we need
526          * to have the normal tree locking.  Also if we are currently trying to
527          * allocate blocks for the tree root we can't do the fast caching since
528          * we likely hold important locks.
529          */
530         if (trans && (!trans->transaction->in_commit) &&
531             (root && root != root->fs_info->tree_root) &&
532             btrfs_test_opt(root, SPACE_CACHE)) {
533                 ret = load_free_space_cache(fs_info, cache);
534
535                 spin_lock(&cache->lock);
536                 if (ret == 1) {
537                         cache->caching_ctl = NULL;
538                         cache->cached = BTRFS_CACHE_FINISHED;
539                         cache->last_byte_to_unpin = (u64)-1;
540                 } else {
541                         if (load_cache_only) {
542                                 cache->caching_ctl = NULL;
543                                 cache->cached = BTRFS_CACHE_NO;
544                         } else {
545                                 cache->cached = BTRFS_CACHE_STARTED;
546                         }
547                 }
548                 spin_unlock(&cache->lock);
549                 wake_up(&caching_ctl->wait);
550                 if (ret == 1) {
551                         put_caching_control(caching_ctl);
552                         free_excluded_extents(fs_info->extent_root, cache);
553                         return 0;
554                 }
555         } else {
556                 /*
557                  * We are not going to do the fast caching, set cached to the
558                  * appropriate value and wakeup any waiters.
559                  */
560                 spin_lock(&cache->lock);
561                 if (load_cache_only) {
562                         cache->caching_ctl = NULL;
563                         cache->cached = BTRFS_CACHE_NO;
564                 } else {
565                         cache->cached = BTRFS_CACHE_STARTED;
566                 }
567                 spin_unlock(&cache->lock);
568                 wake_up(&caching_ctl->wait);
569         }
570
571         if (load_cache_only) {
572                 put_caching_control(caching_ctl);
573                 return 0;
574         }
575
576         down_write(&fs_info->extent_commit_sem);
577         atomic_inc(&caching_ctl->count);
578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
579         up_write(&fs_info->extent_commit_sem);
580
581         btrfs_get_block_group(cache);
582
583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
584
585         return ret;
586 }
587
588 /*
589  * return the block group that starts at or after bytenr
590  */
591 static struct btrfs_block_group_cache *
592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
593 {
594         struct btrfs_block_group_cache *cache;
595
596         cache = block_group_cache_tree_search(info, bytenr, 0);
597
598         return cache;
599 }
600
601 /*
602  * return the block group that contains the given bytenr
603  */
604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
605                                                  struct btrfs_fs_info *info,
606                                                  u64 bytenr)
607 {
608         struct btrfs_block_group_cache *cache;
609
610         cache = block_group_cache_tree_search(info, bytenr, 1);
611
612         return cache;
613 }
614
615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
616                                                   u64 flags)
617 {
618         struct list_head *head = &info->space_info;
619         struct btrfs_space_info *found;
620
621         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
622                  BTRFS_BLOCK_GROUP_METADATA;
623
624         rcu_read_lock();
625         list_for_each_entry_rcu(found, head, list) {
626                 if (found->flags & flags) {
627                         rcu_read_unlock();
628                         return found;
629                 }
630         }
631         rcu_read_unlock();
632         return NULL;
633 }
634
635 /*
636  * after adding space to the filesystem, we need to clear the full flags
637  * on all the space infos.
638  */
639 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
640 {
641         struct list_head *head = &info->space_info;
642         struct btrfs_space_info *found;
643
644         rcu_read_lock();
645         list_for_each_entry_rcu(found, head, list)
646                 found->full = 0;
647         rcu_read_unlock();
648 }
649
650 static u64 div_factor(u64 num, int factor)
651 {
652         if (factor == 10)
653                 return num;
654         num *= factor;
655         do_div(num, 10);
656         return num;
657 }
658
659 static u64 div_factor_fine(u64 num, int factor)
660 {
661         if (factor == 100)
662                 return num;
663         num *= factor;
664         do_div(num, 100);
665         return num;
666 }
667
668 u64 btrfs_find_block_group(struct btrfs_root *root,
669                            u64 search_start, u64 search_hint, int owner)
670 {
671         struct btrfs_block_group_cache *cache;
672         u64 used;
673         u64 last = max(search_hint, search_start);
674         u64 group_start = 0;
675         int full_search = 0;
676         int factor = 9;
677         int wrapped = 0;
678 again:
679         while (1) {
680                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
681                 if (!cache)
682                         break;
683
684                 spin_lock(&cache->lock);
685                 last = cache->key.objectid + cache->key.offset;
686                 used = btrfs_block_group_used(&cache->item);
687
688                 if ((full_search || !cache->ro) &&
689                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
690                         if (used + cache->pinned + cache->reserved <
691                             div_factor(cache->key.offset, factor)) {
692                                 group_start = cache->key.objectid;
693                                 spin_unlock(&cache->lock);
694                                 btrfs_put_block_group(cache);
695                                 goto found;
696                         }
697                 }
698                 spin_unlock(&cache->lock);
699                 btrfs_put_block_group(cache);
700                 cond_resched();
701         }
702         if (!wrapped) {
703                 last = search_start;
704                 wrapped = 1;
705                 goto again;
706         }
707         if (!full_search && factor < 10) {
708                 last = search_start;
709                 full_search = 1;
710                 factor = 10;
711                 goto again;
712         }
713 found:
714         return group_start;
715 }
716
717 /* simple helper to search for an existing extent at a given offset */
718 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
719 {
720         int ret;
721         struct btrfs_key key;
722         struct btrfs_path *path;
723
724         path = btrfs_alloc_path();
725         if (!path)
726                 return -ENOMEM;
727
728         key.objectid = start;
729         key.offset = len;
730         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
731         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
732                                 0, 0);
733         btrfs_free_path(path);
734         return ret;
735 }
736
737 /*
738  * helper function to lookup reference count and flags of extent.
739  *
740  * the head node for delayed ref is used to store the sum of all the
741  * reference count modifications queued up in the rbtree. the head
742  * node may also store the extent flags to set. This way you can check
743  * to see what the reference count and extent flags would be if all of
744  * the delayed refs are not processed.
745  */
746 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
747                              struct btrfs_root *root, u64 bytenr,
748                              u64 num_bytes, u64 *refs, u64 *flags)
749 {
750         struct btrfs_delayed_ref_head *head;
751         struct btrfs_delayed_ref_root *delayed_refs;
752         struct btrfs_path *path;
753         struct btrfs_extent_item *ei;
754         struct extent_buffer *leaf;
755         struct btrfs_key key;
756         u32 item_size;
757         u64 num_refs;
758         u64 extent_flags;
759         int ret;
760
761         path = btrfs_alloc_path();
762         if (!path)
763                 return -ENOMEM;
764
765         key.objectid = bytenr;
766         key.type = BTRFS_EXTENT_ITEM_KEY;
767         key.offset = num_bytes;
768         if (!trans) {
769                 path->skip_locking = 1;
770                 path->search_commit_root = 1;
771         }
772 again:
773         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
774                                 &key, path, 0, 0);
775         if (ret < 0)
776                 goto out_free;
777
778         if (ret == 0) {
779                 leaf = path->nodes[0];
780                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
781                 if (item_size >= sizeof(*ei)) {
782                         ei = btrfs_item_ptr(leaf, path->slots[0],
783                                             struct btrfs_extent_item);
784                         num_refs = btrfs_extent_refs(leaf, ei);
785                         extent_flags = btrfs_extent_flags(leaf, ei);
786                 } else {
787 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
788                         struct btrfs_extent_item_v0 *ei0;
789                         BUG_ON(item_size != sizeof(*ei0));
790                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
791                                              struct btrfs_extent_item_v0);
792                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
793                         /* FIXME: this isn't correct for data */
794                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
795 #else
796                         BUG();
797 #endif
798                 }
799                 BUG_ON(num_refs == 0);
800         } else {
801                 num_refs = 0;
802                 extent_flags = 0;
803                 ret = 0;
804         }
805
806         if (!trans)
807                 goto out;
808
809         delayed_refs = &trans->transaction->delayed_refs;
810         spin_lock(&delayed_refs->lock);
811         head = btrfs_find_delayed_ref_head(trans, bytenr);
812         if (head) {
813                 if (!mutex_trylock(&head->mutex)) {
814                         atomic_inc(&head->node.refs);
815                         spin_unlock(&delayed_refs->lock);
816
817                         btrfs_release_path(path);
818
819                         /*
820                          * Mutex was contended, block until it's released and try
821                          * again
822                          */
823                         mutex_lock(&head->mutex);
824                         mutex_unlock(&head->mutex);
825                         btrfs_put_delayed_ref(&head->node);
826                         goto again;
827                 }
828                 if (head->extent_op && head->extent_op->update_flags)
829                         extent_flags |= head->extent_op->flags_to_set;
830                 else
831                         BUG_ON(num_refs == 0);
832
833                 num_refs += head->node.ref_mod;
834                 mutex_unlock(&head->mutex);
835         }
836         spin_unlock(&delayed_refs->lock);
837 out:
838         WARN_ON(num_refs == 0);
839         if (refs)
840                 *refs = num_refs;
841         if (flags)
842                 *flags = extent_flags;
843 out_free:
844         btrfs_free_path(path);
845         return ret;
846 }
847
848 /*
849  * Back reference rules.  Back refs have three main goals:
850  *
851  * 1) differentiate between all holders of references to an extent so that
852  *    when a reference is dropped we can make sure it was a valid reference
853  *    before freeing the extent.
854  *
855  * 2) Provide enough information to quickly find the holders of an extent
856  *    if we notice a given block is corrupted or bad.
857  *
858  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
859  *    maintenance.  This is actually the same as #2, but with a slightly
860  *    different use case.
861  *
862  * There are two kinds of back refs. The implicit back refs is optimized
863  * for pointers in non-shared tree blocks. For a given pointer in a block,
864  * back refs of this kind provide information about the block's owner tree
865  * and the pointer's key. These information allow us to find the block by
866  * b-tree searching. The full back refs is for pointers in tree blocks not
867  * referenced by their owner trees. The location of tree block is recorded
868  * in the back refs. Actually the full back refs is generic, and can be
869  * used in all cases the implicit back refs is used. The major shortcoming
870  * of the full back refs is its overhead. Every time a tree block gets
871  * COWed, we have to update back refs entry for all pointers in it.
872  *
873  * For a newly allocated tree block, we use implicit back refs for
874  * pointers in it. This means most tree related operations only involve
875  * implicit back refs. For a tree block created in old transaction, the
876  * only way to drop a reference to it is COW it. So we can detect the
877  * event that tree block loses its owner tree's reference and do the
878  * back refs conversion.
879  *
880  * When a tree block is COW'd through a tree, there are four cases:
881  *
882  * The reference count of the block is one and the tree is the block's
883  * owner tree. Nothing to do in this case.
884  *
885  * The reference count of the block is one and the tree is not the
886  * block's owner tree. In this case, full back refs is used for pointers
887  * in the block. Remove these full back refs, add implicit back refs for
888  * every pointers in the new block.
889  *
890  * The reference count of the block is greater than one and the tree is
891  * the block's owner tree. In this case, implicit back refs is used for
892  * pointers in the block. Add full back refs for every pointers in the
893  * block, increase lower level extents' reference counts. The original
894  * implicit back refs are entailed to the new block.
895  *
896  * The reference count of the block is greater than one and the tree is
897  * not the block's owner tree. Add implicit back refs for every pointer in
898  * the new block, increase lower level extents' reference count.
899  *
900  * Back Reference Key composing:
901  *
902  * The key objectid corresponds to the first byte in the extent,
903  * The key type is used to differentiate between types of back refs.
904  * There are different meanings of the key offset for different types
905  * of back refs.
906  *
907  * File extents can be referenced by:
908  *
909  * - multiple snapshots, subvolumes, or different generations in one subvol
910  * - different files inside a single subvolume
911  * - different offsets inside a file (bookend extents in file.c)
912  *
913  * The extent ref structure for the implicit back refs has fields for:
914  *
915  * - Objectid of the subvolume root
916  * - objectid of the file holding the reference
917  * - original offset in the file
918  * - how many bookend extents
919  *
920  * The key offset for the implicit back refs is hash of the first
921  * three fields.
922  *
923  * The extent ref structure for the full back refs has field for:
924  *
925  * - number of pointers in the tree leaf
926  *
927  * The key offset for the implicit back refs is the first byte of
928  * the tree leaf
929  *
930  * When a file extent is allocated, The implicit back refs is used.
931  * the fields are filled in:
932  *
933  *     (root_key.objectid, inode objectid, offset in file, 1)
934  *
935  * When a file extent is removed file truncation, we find the
936  * corresponding implicit back refs and check the following fields:
937  *
938  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
939  *
940  * Btree extents can be referenced by:
941  *
942  * - Different subvolumes
943  *
944  * Both the implicit back refs and the full back refs for tree blocks
945  * only consist of key. The key offset for the implicit back refs is
946  * objectid of block's owner tree. The key offset for the full back refs
947  * is the first byte of parent block.
948  *
949  * When implicit back refs is used, information about the lowest key and
950  * level of the tree block are required. These information are stored in
951  * tree block info structure.
952  */
953
954 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
955 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
956                                   struct btrfs_root *root,
957                                   struct btrfs_path *path,
958                                   u64 owner, u32 extra_size)
959 {
960         struct btrfs_extent_item *item;
961         struct btrfs_extent_item_v0 *ei0;
962         struct btrfs_extent_ref_v0 *ref0;
963         struct btrfs_tree_block_info *bi;
964         struct extent_buffer *leaf;
965         struct btrfs_key key;
966         struct btrfs_key found_key;
967         u32 new_size = sizeof(*item);
968         u64 refs;
969         int ret;
970
971         leaf = path->nodes[0];
972         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
973
974         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
975         ei0 = btrfs_item_ptr(leaf, path->slots[0],
976                              struct btrfs_extent_item_v0);
977         refs = btrfs_extent_refs_v0(leaf, ei0);
978
979         if (owner == (u64)-1) {
980                 while (1) {
981                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
982                                 ret = btrfs_next_leaf(root, path);
983                                 if (ret < 0)
984                                         return ret;
985                                 BUG_ON(ret > 0);
986                                 leaf = path->nodes[0];
987                         }
988                         btrfs_item_key_to_cpu(leaf, &found_key,
989                                               path->slots[0]);
990                         BUG_ON(key.objectid != found_key.objectid);
991                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
992                                 path->slots[0]++;
993                                 continue;
994                         }
995                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
996                                               struct btrfs_extent_ref_v0);
997                         owner = btrfs_ref_objectid_v0(leaf, ref0);
998                         break;
999                 }
1000         }
1001         btrfs_release_path(path);
1002
1003         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1004                 new_size += sizeof(*bi);
1005
1006         new_size -= sizeof(*ei0);
1007         ret = btrfs_search_slot(trans, root, &key, path,
1008                                 new_size + extra_size, 1);
1009         if (ret < 0)
1010                 return ret;
1011         BUG_ON(ret);
1012
1013         ret = btrfs_extend_item(trans, root, path, new_size);
1014
1015         leaf = path->nodes[0];
1016         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1017         btrfs_set_extent_refs(leaf, item, refs);
1018         /* FIXME: get real generation */
1019         btrfs_set_extent_generation(leaf, item, 0);
1020         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1021                 btrfs_set_extent_flags(leaf, item,
1022                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1023                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1024                 bi = (struct btrfs_tree_block_info *)(item + 1);
1025                 /* FIXME: get first key of the block */
1026                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1027                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1028         } else {
1029                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1030         }
1031         btrfs_mark_buffer_dirty(leaf);
1032         return 0;
1033 }
1034 #endif
1035
1036 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1037 {
1038         u32 high_crc = ~(u32)0;
1039         u32 low_crc = ~(u32)0;
1040         __le64 lenum;
1041
1042         lenum = cpu_to_le64(root_objectid);
1043         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1044         lenum = cpu_to_le64(owner);
1045         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1046         lenum = cpu_to_le64(offset);
1047         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1048
1049         return ((u64)high_crc << 31) ^ (u64)low_crc;
1050 }
1051
1052 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1053                                      struct btrfs_extent_data_ref *ref)
1054 {
1055         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1056                                     btrfs_extent_data_ref_objectid(leaf, ref),
1057                                     btrfs_extent_data_ref_offset(leaf, ref));
1058 }
1059
1060 static int match_extent_data_ref(struct extent_buffer *leaf,
1061                                  struct btrfs_extent_data_ref *ref,
1062                                  u64 root_objectid, u64 owner, u64 offset)
1063 {
1064         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1065             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1066             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1067                 return 0;
1068         return 1;
1069 }
1070
1071 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1072                                            struct btrfs_root *root,
1073                                            struct btrfs_path *path,
1074                                            u64 bytenr, u64 parent,
1075                                            u64 root_objectid,
1076                                            u64 owner, u64 offset)
1077 {
1078         struct btrfs_key key;
1079         struct btrfs_extent_data_ref *ref;
1080         struct extent_buffer *leaf;
1081         u32 nritems;
1082         int ret;
1083         int recow;
1084         int err = -ENOENT;
1085
1086         key.objectid = bytenr;
1087         if (parent) {
1088                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1089                 key.offset = parent;
1090         } else {
1091                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1092                 key.offset = hash_extent_data_ref(root_objectid,
1093                                                   owner, offset);
1094         }
1095 again:
1096         recow = 0;
1097         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1098         if (ret < 0) {
1099                 err = ret;
1100                 goto fail;
1101         }
1102
1103         if (parent) {
1104                 if (!ret)
1105                         return 0;
1106 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1107                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1108                 btrfs_release_path(path);
1109                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1110                 if (ret < 0) {
1111                         err = ret;
1112                         goto fail;
1113                 }
1114                 if (!ret)
1115                         return 0;
1116 #endif
1117                 goto fail;
1118         }
1119
1120         leaf = path->nodes[0];
1121         nritems = btrfs_header_nritems(leaf);
1122         while (1) {
1123                 if (path->slots[0] >= nritems) {
1124                         ret = btrfs_next_leaf(root, path);
1125                         if (ret < 0)
1126                                 err = ret;
1127                         if (ret)
1128                                 goto fail;
1129
1130                         leaf = path->nodes[0];
1131                         nritems = btrfs_header_nritems(leaf);
1132                         recow = 1;
1133                 }
1134
1135                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1136                 if (key.objectid != bytenr ||
1137                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1138                         goto fail;
1139
1140                 ref = btrfs_item_ptr(leaf, path->slots[0],
1141                                      struct btrfs_extent_data_ref);
1142
1143                 if (match_extent_data_ref(leaf, ref, root_objectid,
1144                                           owner, offset)) {
1145                         if (recow) {
1146                                 btrfs_release_path(path);
1147                                 goto again;
1148                         }
1149                         err = 0;
1150                         break;
1151                 }
1152                 path->slots[0]++;
1153         }
1154 fail:
1155         return err;
1156 }
1157
1158 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1159                                            struct btrfs_root *root,
1160                                            struct btrfs_path *path,
1161                                            u64 bytenr, u64 parent,
1162                                            u64 root_objectid, u64 owner,
1163                                            u64 offset, int refs_to_add)
1164 {
1165         struct btrfs_key key;
1166         struct extent_buffer *leaf;
1167         u32 size;
1168         u32 num_refs;
1169         int ret;
1170
1171         key.objectid = bytenr;
1172         if (parent) {
1173                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1174                 key.offset = parent;
1175                 size = sizeof(struct btrfs_shared_data_ref);
1176         } else {
1177                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1178                 key.offset = hash_extent_data_ref(root_objectid,
1179                                                   owner, offset);
1180                 size = sizeof(struct btrfs_extent_data_ref);
1181         }
1182
1183         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1184         if (ret && ret != -EEXIST)
1185                 goto fail;
1186
1187         leaf = path->nodes[0];
1188         if (parent) {
1189                 struct btrfs_shared_data_ref *ref;
1190                 ref = btrfs_item_ptr(leaf, path->slots[0],
1191                                      struct btrfs_shared_data_ref);
1192                 if (ret == 0) {
1193                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1194                 } else {
1195                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1196                         num_refs += refs_to_add;
1197                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1198                 }
1199         } else {
1200                 struct btrfs_extent_data_ref *ref;
1201                 while (ret == -EEXIST) {
1202                         ref = btrfs_item_ptr(leaf, path->slots[0],
1203                                              struct btrfs_extent_data_ref);
1204                         if (match_extent_data_ref(leaf, ref, root_objectid,
1205                                                   owner, offset))
1206                                 break;
1207                         btrfs_release_path(path);
1208                         key.offset++;
1209                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1210                                                       size);
1211                         if (ret && ret != -EEXIST)
1212                                 goto fail;
1213
1214                         leaf = path->nodes[0];
1215                 }
1216                 ref = btrfs_item_ptr(leaf, path->slots[0],
1217                                      struct btrfs_extent_data_ref);
1218                 if (ret == 0) {
1219                         btrfs_set_extent_data_ref_root(leaf, ref,
1220                                                        root_objectid);
1221                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1222                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1223                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1224                 } else {
1225                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1226                         num_refs += refs_to_add;
1227                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1228                 }
1229         }
1230         btrfs_mark_buffer_dirty(leaf);
1231         ret = 0;
1232 fail:
1233         btrfs_release_path(path);
1234         return ret;
1235 }
1236
1237 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1238                                            struct btrfs_root *root,
1239                                            struct btrfs_path *path,
1240                                            int refs_to_drop)
1241 {
1242         struct btrfs_key key;
1243         struct btrfs_extent_data_ref *ref1 = NULL;
1244         struct btrfs_shared_data_ref *ref2 = NULL;
1245         struct extent_buffer *leaf;
1246         u32 num_refs = 0;
1247         int ret = 0;
1248
1249         leaf = path->nodes[0];
1250         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1251
1252         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1253                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1254                                       struct btrfs_extent_data_ref);
1255                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1256         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1257                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_shared_data_ref);
1259                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1260 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1261         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1262                 struct btrfs_extent_ref_v0 *ref0;
1263                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1264                                       struct btrfs_extent_ref_v0);
1265                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1266 #endif
1267         } else {
1268                 BUG();
1269         }
1270
1271         BUG_ON(num_refs < refs_to_drop);
1272         num_refs -= refs_to_drop;
1273
1274         if (num_refs == 0) {
1275                 ret = btrfs_del_item(trans, root, path);
1276         } else {
1277                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1278                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1279                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1280                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1281 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1282                 else {
1283                         struct btrfs_extent_ref_v0 *ref0;
1284                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1285                                         struct btrfs_extent_ref_v0);
1286                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1287                 }
1288 #endif
1289                 btrfs_mark_buffer_dirty(leaf);
1290         }
1291         return ret;
1292 }
1293
1294 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1295                                           struct btrfs_path *path,
1296                                           struct btrfs_extent_inline_ref *iref)
1297 {
1298         struct btrfs_key key;
1299         struct extent_buffer *leaf;
1300         struct btrfs_extent_data_ref *ref1;
1301         struct btrfs_shared_data_ref *ref2;
1302         u32 num_refs = 0;
1303
1304         leaf = path->nodes[0];
1305         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1306         if (iref) {
1307                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1308                     BTRFS_EXTENT_DATA_REF_KEY) {
1309                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1310                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1311                 } else {
1312                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1313                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1314                 }
1315         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1316                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1317                                       struct btrfs_extent_data_ref);
1318                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1319         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1320                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1321                                       struct btrfs_shared_data_ref);
1322                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1323 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1324         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1325                 struct btrfs_extent_ref_v0 *ref0;
1326                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1327                                       struct btrfs_extent_ref_v0);
1328                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1329 #endif
1330         } else {
1331                 WARN_ON(1);
1332         }
1333         return num_refs;
1334 }
1335
1336 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1337                                           struct btrfs_root *root,
1338                                           struct btrfs_path *path,
1339                                           u64 bytenr, u64 parent,
1340                                           u64 root_objectid)
1341 {
1342         struct btrfs_key key;
1343         int ret;
1344
1345         key.objectid = bytenr;
1346         if (parent) {
1347                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1348                 key.offset = parent;
1349         } else {
1350                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1351                 key.offset = root_objectid;
1352         }
1353
1354         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1355         if (ret > 0)
1356                 ret = -ENOENT;
1357 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1358         if (ret == -ENOENT && parent) {
1359                 btrfs_release_path(path);
1360                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1361                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1362                 if (ret > 0)
1363                         ret = -ENOENT;
1364         }
1365 #endif
1366         return ret;
1367 }
1368
1369 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1370                                           struct btrfs_root *root,
1371                                           struct btrfs_path *path,
1372                                           u64 bytenr, u64 parent,
1373                                           u64 root_objectid)
1374 {
1375         struct btrfs_key key;
1376         int ret;
1377
1378         key.objectid = bytenr;
1379         if (parent) {
1380                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1381                 key.offset = parent;
1382         } else {
1383                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1384                 key.offset = root_objectid;
1385         }
1386
1387         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1388         btrfs_release_path(path);
1389         return ret;
1390 }
1391
1392 static inline int extent_ref_type(u64 parent, u64 owner)
1393 {
1394         int type;
1395         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1396                 if (parent > 0)
1397                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1398                 else
1399                         type = BTRFS_TREE_BLOCK_REF_KEY;
1400         } else {
1401                 if (parent > 0)
1402                         type = BTRFS_SHARED_DATA_REF_KEY;
1403                 else
1404                         type = BTRFS_EXTENT_DATA_REF_KEY;
1405         }
1406         return type;
1407 }
1408
1409 static int find_next_key(struct btrfs_path *path, int level,
1410                          struct btrfs_key *key)
1411
1412 {
1413         for (; level < BTRFS_MAX_LEVEL; level++) {
1414                 if (!path->nodes[level])
1415                         break;
1416                 if (path->slots[level] + 1 >=
1417                     btrfs_header_nritems(path->nodes[level]))
1418                         continue;
1419                 if (level == 0)
1420                         btrfs_item_key_to_cpu(path->nodes[level], key,
1421                                               path->slots[level] + 1);
1422                 else
1423                         btrfs_node_key_to_cpu(path->nodes[level], key,
1424                                               path->slots[level] + 1);
1425                 return 0;
1426         }
1427         return 1;
1428 }
1429
1430 /*
1431  * look for inline back ref. if back ref is found, *ref_ret is set
1432  * to the address of inline back ref, and 0 is returned.
1433  *
1434  * if back ref isn't found, *ref_ret is set to the address where it
1435  * should be inserted, and -ENOENT is returned.
1436  *
1437  * if insert is true and there are too many inline back refs, the path
1438  * points to the extent item, and -EAGAIN is returned.
1439  *
1440  * NOTE: inline back refs are ordered in the same way that back ref
1441  *       items in the tree are ordered.
1442  */
1443 static noinline_for_stack
1444 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1445                                  struct btrfs_root *root,
1446                                  struct btrfs_path *path,
1447                                  struct btrfs_extent_inline_ref **ref_ret,
1448                                  u64 bytenr, u64 num_bytes,
1449                                  u64 parent, u64 root_objectid,
1450                                  u64 owner, u64 offset, int insert)
1451 {
1452         struct btrfs_key key;
1453         struct extent_buffer *leaf;
1454         struct btrfs_extent_item *ei;
1455         struct btrfs_extent_inline_ref *iref;
1456         u64 flags;
1457         u64 item_size;
1458         unsigned long ptr;
1459         unsigned long end;
1460         int extra_size;
1461         int type;
1462         int want;
1463         int ret;
1464         int err = 0;
1465
1466         key.objectid = bytenr;
1467         key.type = BTRFS_EXTENT_ITEM_KEY;
1468         key.offset = num_bytes;
1469
1470         want = extent_ref_type(parent, owner);
1471         if (insert) {
1472                 extra_size = btrfs_extent_inline_ref_size(want);
1473                 path->keep_locks = 1;
1474         } else
1475                 extra_size = -1;
1476         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1477         if (ret < 0) {
1478                 err = ret;
1479                 goto out;
1480         }
1481         BUG_ON(ret);
1482
1483         leaf = path->nodes[0];
1484         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1485 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1486         if (item_size < sizeof(*ei)) {
1487                 if (!insert) {
1488                         err = -ENOENT;
1489                         goto out;
1490                 }
1491                 ret = convert_extent_item_v0(trans, root, path, owner,
1492                                              extra_size);
1493                 if (ret < 0) {
1494                         err = ret;
1495                         goto out;
1496                 }
1497                 leaf = path->nodes[0];
1498                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1499         }
1500 #endif
1501         BUG_ON(item_size < sizeof(*ei));
1502
1503         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1504         flags = btrfs_extent_flags(leaf, ei);
1505
1506         ptr = (unsigned long)(ei + 1);
1507         end = (unsigned long)ei + item_size;
1508
1509         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1510                 ptr += sizeof(struct btrfs_tree_block_info);
1511                 BUG_ON(ptr > end);
1512         } else {
1513                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1514         }
1515
1516         err = -ENOENT;
1517         while (1) {
1518                 if (ptr >= end) {
1519                         WARN_ON(ptr > end);
1520                         break;
1521                 }
1522                 iref = (struct btrfs_extent_inline_ref *)ptr;
1523                 type = btrfs_extent_inline_ref_type(leaf, iref);
1524                 if (want < type)
1525                         break;
1526                 if (want > type) {
1527                         ptr += btrfs_extent_inline_ref_size(type);
1528                         continue;
1529                 }
1530
1531                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1532                         struct btrfs_extent_data_ref *dref;
1533                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1534                         if (match_extent_data_ref(leaf, dref, root_objectid,
1535                                                   owner, offset)) {
1536                                 err = 0;
1537                                 break;
1538                         }
1539                         if (hash_extent_data_ref_item(leaf, dref) <
1540                             hash_extent_data_ref(root_objectid, owner, offset))
1541                                 break;
1542                 } else {
1543                         u64 ref_offset;
1544                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1545                         if (parent > 0) {
1546                                 if (parent == ref_offset) {
1547                                         err = 0;
1548                                         break;
1549                                 }
1550                                 if (ref_offset < parent)
1551                                         break;
1552                         } else {
1553                                 if (root_objectid == ref_offset) {
1554                                         err = 0;
1555                                         break;
1556                                 }
1557                                 if (ref_offset < root_objectid)
1558                                         break;
1559                         }
1560                 }
1561                 ptr += btrfs_extent_inline_ref_size(type);
1562         }
1563         if (err == -ENOENT && insert) {
1564                 if (item_size + extra_size >=
1565                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1566                         err = -EAGAIN;
1567                         goto out;
1568                 }
1569                 /*
1570                  * To add new inline back ref, we have to make sure
1571                  * there is no corresponding back ref item.
1572                  * For simplicity, we just do not add new inline back
1573                  * ref if there is any kind of item for this block
1574                  */
1575                 if (find_next_key(path, 0, &key) == 0 &&
1576                     key.objectid == bytenr &&
1577                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1578                         err = -EAGAIN;
1579                         goto out;
1580                 }
1581         }
1582         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1583 out:
1584         if (insert) {
1585                 path->keep_locks = 0;
1586                 btrfs_unlock_up_safe(path, 1);
1587         }
1588         return err;
1589 }
1590
1591 /*
1592  * helper to add new inline back ref
1593  */
1594 static noinline_for_stack
1595 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1596                                 struct btrfs_root *root,
1597                                 struct btrfs_path *path,
1598                                 struct btrfs_extent_inline_ref *iref,
1599                                 u64 parent, u64 root_objectid,
1600                                 u64 owner, u64 offset, int refs_to_add,
1601                                 struct btrfs_delayed_extent_op *extent_op)
1602 {
1603         struct extent_buffer *leaf;
1604         struct btrfs_extent_item *ei;
1605         unsigned long ptr;
1606         unsigned long end;
1607         unsigned long item_offset;
1608         u64 refs;
1609         int size;
1610         int type;
1611         int ret;
1612
1613         leaf = path->nodes[0];
1614         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1615         item_offset = (unsigned long)iref - (unsigned long)ei;
1616
1617         type = extent_ref_type(parent, owner);
1618         size = btrfs_extent_inline_ref_size(type);
1619
1620         ret = btrfs_extend_item(trans, root, path, size);
1621
1622         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1623         refs = btrfs_extent_refs(leaf, ei);
1624         refs += refs_to_add;
1625         btrfs_set_extent_refs(leaf, ei, refs);
1626         if (extent_op)
1627                 __run_delayed_extent_op(extent_op, leaf, ei);
1628
1629         ptr = (unsigned long)ei + item_offset;
1630         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1631         if (ptr < end - size)
1632                 memmove_extent_buffer(leaf, ptr + size, ptr,
1633                                       end - size - ptr);
1634
1635         iref = (struct btrfs_extent_inline_ref *)ptr;
1636         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1637         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1638                 struct btrfs_extent_data_ref *dref;
1639                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1640                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1641                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1642                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1643                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1644         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1645                 struct btrfs_shared_data_ref *sref;
1646                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1647                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1648                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1649         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1653         }
1654         btrfs_mark_buffer_dirty(leaf);
1655         return 0;
1656 }
1657
1658 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1659                                  struct btrfs_root *root,
1660                                  struct btrfs_path *path,
1661                                  struct btrfs_extent_inline_ref **ref_ret,
1662                                  u64 bytenr, u64 num_bytes, u64 parent,
1663                                  u64 root_objectid, u64 owner, u64 offset)
1664 {
1665         int ret;
1666
1667         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1668                                            bytenr, num_bytes, parent,
1669                                            root_objectid, owner, offset, 0);
1670         if (ret != -ENOENT)
1671                 return ret;
1672
1673         btrfs_release_path(path);
1674         *ref_ret = NULL;
1675
1676         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1677                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1678                                             root_objectid);
1679         } else {
1680                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1681                                              root_objectid, owner, offset);
1682         }
1683         return ret;
1684 }
1685
1686 /*
1687  * helper to update/remove inline back ref
1688  */
1689 static noinline_for_stack
1690 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1691                                  struct btrfs_root *root,
1692                                  struct btrfs_path *path,
1693                                  struct btrfs_extent_inline_ref *iref,
1694                                  int refs_to_mod,
1695                                  struct btrfs_delayed_extent_op *extent_op)
1696 {
1697         struct extent_buffer *leaf;
1698         struct btrfs_extent_item *ei;
1699         struct btrfs_extent_data_ref *dref = NULL;
1700         struct btrfs_shared_data_ref *sref = NULL;
1701         unsigned long ptr;
1702         unsigned long end;
1703         u32 item_size;
1704         int size;
1705         int type;
1706         int ret;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751         return 0;
1752 }
1753
1754 static noinline_for_stack
1755 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1756                                  struct btrfs_root *root,
1757                                  struct btrfs_path *path,
1758                                  u64 bytenr, u64 num_bytes, u64 parent,
1759                                  u64 root_objectid, u64 owner,
1760                                  u64 offset, int refs_to_add,
1761                                  struct btrfs_delayed_extent_op *extent_op)
1762 {
1763         struct btrfs_extent_inline_ref *iref;
1764         int ret;
1765
1766         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1767                                            bytenr, num_bytes, parent,
1768                                            root_objectid, owner, offset, 1);
1769         if (ret == 0) {
1770                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1771                 ret = update_inline_extent_backref(trans, root, path, iref,
1772                                                    refs_to_add, extent_op);
1773         } else if (ret == -ENOENT) {
1774                 ret = setup_inline_extent_backref(trans, root, path, iref,
1775                                                   parent, root_objectid,
1776                                                   owner, offset, refs_to_add,
1777                                                   extent_op);
1778         }
1779         return ret;
1780 }
1781
1782 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1783                                  struct btrfs_root *root,
1784                                  struct btrfs_path *path,
1785                                  u64 bytenr, u64 parent, u64 root_objectid,
1786                                  u64 owner, u64 offset, int refs_to_add)
1787 {
1788         int ret;
1789         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1790                 BUG_ON(refs_to_add != 1);
1791                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1792                                             parent, root_objectid);
1793         } else {
1794                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1795                                              parent, root_objectid,
1796                                              owner, offset, refs_to_add);
1797         }
1798         return ret;
1799 }
1800
1801 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1802                                  struct btrfs_root *root,
1803                                  struct btrfs_path *path,
1804                                  struct btrfs_extent_inline_ref *iref,
1805                                  int refs_to_drop, int is_data)
1806 {
1807         int ret;
1808
1809         BUG_ON(!is_data && refs_to_drop != 1);
1810         if (iref) {
1811                 ret = update_inline_extent_backref(trans, root, path, iref,
1812                                                    -refs_to_drop, NULL);
1813         } else if (is_data) {
1814                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1815         } else {
1816                 ret = btrfs_del_item(trans, root, path);
1817         }
1818         return ret;
1819 }
1820
1821 static int btrfs_issue_discard(struct block_device *bdev,
1822                                 u64 start, u64 len)
1823 {
1824         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1825 }
1826
1827 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1828                                 u64 num_bytes, u64 *actual_bytes)
1829 {
1830         int ret;
1831         u64 discarded_bytes = 0;
1832         struct btrfs_bio *bbio = NULL;
1833
1834
1835         /* Tell the block device(s) that the sectors can be discarded */
1836         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1837                               bytenr, &num_bytes, &bbio, 0);
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break;
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1873                          struct btrfs_root *root,
1874                          u64 bytenr, u64 num_bytes, u64 parent,
1875                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1876 {
1877         int ret;
1878         struct btrfs_fs_info *fs_info = root->fs_info;
1879
1880         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1881                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1882
1883         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1884                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1885                                         num_bytes,
1886                                         parent, root_objectid, (int)owner,
1887                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1888         } else {
1889                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1890                                         num_bytes,
1891                                         parent, root_objectid, owner, offset,
1892                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1893         }
1894         return ret;
1895 }
1896
1897 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1898                                   struct btrfs_root *root,
1899                                   u64 bytenr, u64 num_bytes,
1900                                   u64 parent, u64 root_objectid,
1901                                   u64 owner, u64 offset, int refs_to_add,
1902                                   struct btrfs_delayed_extent_op *extent_op)
1903 {
1904         struct btrfs_path *path;
1905         struct extent_buffer *leaf;
1906         struct btrfs_extent_item *item;
1907         u64 refs;
1908         int ret;
1909         int err = 0;
1910
1911         path = btrfs_alloc_path();
1912         if (!path)
1913                 return -ENOMEM;
1914
1915         path->reada = 1;
1916         path->leave_spinning = 1;
1917         /* this will setup the path even if it fails to insert the back ref */
1918         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1919                                            path, bytenr, num_bytes, parent,
1920                                            root_objectid, owner, offset,
1921                                            refs_to_add, extent_op);
1922         if (ret == 0)
1923                 goto out;
1924
1925         if (ret != -EAGAIN) {
1926                 err = ret;
1927                 goto out;
1928         }
1929
1930         leaf = path->nodes[0];
1931         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1932         refs = btrfs_extent_refs(leaf, item);
1933         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1934         if (extent_op)
1935                 __run_delayed_extent_op(extent_op, leaf, item);
1936
1937         btrfs_mark_buffer_dirty(leaf);
1938         btrfs_release_path(path);
1939
1940         path->reada = 1;
1941         path->leave_spinning = 1;
1942
1943         /* now insert the actual backref */
1944         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1945                                     path, bytenr, parent, root_objectid,
1946                                     owner, offset, refs_to_add);
1947         BUG_ON(ret);
1948 out:
1949         btrfs_free_path(path);
1950         return err;
1951 }
1952
1953 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1954                                 struct btrfs_root *root,
1955                                 struct btrfs_delayed_ref_node *node,
1956                                 struct btrfs_delayed_extent_op *extent_op,
1957                                 int insert_reserved)
1958 {
1959         int ret = 0;
1960         struct btrfs_delayed_data_ref *ref;
1961         struct btrfs_key ins;
1962         u64 parent = 0;
1963         u64 ref_root = 0;
1964         u64 flags = 0;
1965
1966         ins.objectid = node->bytenr;
1967         ins.offset = node->num_bytes;
1968         ins.type = BTRFS_EXTENT_ITEM_KEY;
1969
1970         ref = btrfs_delayed_node_to_data_ref(node);
1971         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1972                 parent = ref->parent;
1973         else
1974                 ref_root = ref->root;
1975
1976         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1977                 if (extent_op) {
1978                         BUG_ON(extent_op->update_key);
1979                         flags |= extent_op->flags_to_set;
1980                 }
1981                 ret = alloc_reserved_file_extent(trans, root,
1982                                                  parent, ref_root, flags,
1983                                                  ref->objectid, ref->offset,
1984                                                  &ins, node->ref_mod);
1985         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1986                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1987                                              node->num_bytes, parent,
1988                                              ref_root, ref->objectid,
1989                                              ref->offset, node->ref_mod,
1990                                              extent_op);
1991         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1992                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1993                                           node->num_bytes, parent,
1994                                           ref_root, ref->objectid,
1995                                           ref->offset, node->ref_mod,
1996                                           extent_op);
1997         } else {
1998                 BUG();
1999         }
2000         return ret;
2001 }
2002
2003 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2004                                     struct extent_buffer *leaf,
2005                                     struct btrfs_extent_item *ei)
2006 {
2007         u64 flags = btrfs_extent_flags(leaf, ei);
2008         if (extent_op->update_flags) {
2009                 flags |= extent_op->flags_to_set;
2010                 btrfs_set_extent_flags(leaf, ei, flags);
2011         }
2012
2013         if (extent_op->update_key) {
2014                 struct btrfs_tree_block_info *bi;
2015                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2016                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2017                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2018         }
2019 }
2020
2021 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2022                                  struct btrfs_root *root,
2023                                  struct btrfs_delayed_ref_node *node,
2024                                  struct btrfs_delayed_extent_op *extent_op)
2025 {
2026         struct btrfs_key key;
2027         struct btrfs_path *path;
2028         struct btrfs_extent_item *ei;
2029         struct extent_buffer *leaf;
2030         u32 item_size;
2031         int ret;
2032         int err = 0;
2033
2034         path = btrfs_alloc_path();
2035         if (!path)
2036                 return -ENOMEM;
2037
2038         key.objectid = node->bytenr;
2039         key.type = BTRFS_EXTENT_ITEM_KEY;
2040         key.offset = node->num_bytes;
2041
2042         path->reada = 1;
2043         path->leave_spinning = 1;
2044         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2045                                 path, 0, 1);
2046         if (ret < 0) {
2047                 err = ret;
2048                 goto out;
2049         }
2050         if (ret > 0) {
2051                 err = -EIO;
2052                 goto out;
2053         }
2054
2055         leaf = path->nodes[0];
2056         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2057 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2058         if (item_size < sizeof(*ei)) {
2059                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2060                                              path, (u64)-1, 0);
2061                 if (ret < 0) {
2062                         err = ret;
2063                         goto out;
2064                 }
2065                 leaf = path->nodes[0];
2066                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2067         }
2068 #endif
2069         BUG_ON(item_size < sizeof(*ei));
2070         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2071         __run_delayed_extent_op(extent_op, leaf, ei);
2072
2073         btrfs_mark_buffer_dirty(leaf);
2074 out:
2075         btrfs_free_path(path);
2076         return err;
2077 }
2078
2079 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2080                                 struct btrfs_root *root,
2081                                 struct btrfs_delayed_ref_node *node,
2082                                 struct btrfs_delayed_extent_op *extent_op,
2083                                 int insert_reserved)
2084 {
2085         int ret = 0;
2086         struct btrfs_delayed_tree_ref *ref;
2087         struct btrfs_key ins;
2088         u64 parent = 0;
2089         u64 ref_root = 0;
2090
2091         ins.objectid = node->bytenr;
2092         ins.offset = node->num_bytes;
2093         ins.type = BTRFS_EXTENT_ITEM_KEY;
2094
2095         ref = btrfs_delayed_node_to_tree_ref(node);
2096         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2097                 parent = ref->parent;
2098         else
2099                 ref_root = ref->root;
2100
2101         BUG_ON(node->ref_mod != 1);
2102         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2103                 BUG_ON(!extent_op || !extent_op->update_flags ||
2104                        !extent_op->update_key);
2105                 ret = alloc_reserved_tree_block(trans, root,
2106                                                 parent, ref_root,
2107                                                 extent_op->flags_to_set,
2108                                                 &extent_op->key,
2109                                                 ref->level, &ins);
2110         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2111                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2112                                              node->num_bytes, parent, ref_root,
2113                                              ref->level, 0, 1, extent_op);
2114         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2115                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2116                                           node->num_bytes, parent, ref_root,
2117                                           ref->level, 0, 1, extent_op);
2118         } else {
2119                 BUG();
2120         }
2121         return ret;
2122 }
2123
2124 /* helper function to actually process a single delayed ref entry */
2125 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2126                                struct btrfs_root *root,
2127                                struct btrfs_delayed_ref_node *node,
2128                                struct btrfs_delayed_extent_op *extent_op,
2129                                int insert_reserved)
2130 {
2131         int ret;
2132         if (btrfs_delayed_ref_is_head(node)) {
2133                 struct btrfs_delayed_ref_head *head;
2134                 /*
2135                  * we've hit the end of the chain and we were supposed
2136                  * to insert this extent into the tree.  But, it got
2137                  * deleted before we ever needed to insert it, so all
2138                  * we have to do is clean up the accounting
2139                  */
2140                 BUG_ON(extent_op);
2141                 head = btrfs_delayed_node_to_head(node);
2142                 if (insert_reserved) {
2143                         btrfs_pin_extent(root, node->bytenr,
2144                                          node->num_bytes, 1);
2145                         if (head->is_data) {
2146                                 ret = btrfs_del_csums(trans, root,
2147                                                       node->bytenr,
2148                                                       node->num_bytes);
2149                                 BUG_ON(ret);
2150                         }
2151                 }
2152                 mutex_unlock(&head->mutex);
2153                 return 0;
2154         }
2155
2156         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2157             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2158                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2159                                            insert_reserved);
2160         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2161                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2162                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2163                                            insert_reserved);
2164         else
2165                 BUG();
2166         return ret;
2167 }
2168
2169 static noinline struct btrfs_delayed_ref_node *
2170 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2171 {
2172         struct rb_node *node;
2173         struct btrfs_delayed_ref_node *ref;
2174         int action = BTRFS_ADD_DELAYED_REF;
2175 again:
2176         /*
2177          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2178          * this prevents ref count from going down to zero when
2179          * there still are pending delayed ref.
2180          */
2181         node = rb_prev(&head->node.rb_node);
2182         while (1) {
2183                 if (!node)
2184                         break;
2185                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2186                                 rb_node);
2187                 if (ref->bytenr != head->node.bytenr)
2188                         break;
2189                 if (ref->action == action)
2190                         return ref;
2191                 node = rb_prev(node);
2192         }
2193         if (action == BTRFS_ADD_DELAYED_REF) {
2194                 action = BTRFS_DROP_DELAYED_REF;
2195                 goto again;
2196         }
2197         return NULL;
2198 }
2199
2200 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2201                                        struct btrfs_root *root,
2202                                        struct list_head *cluster)
2203 {
2204         struct btrfs_delayed_ref_root *delayed_refs;
2205         struct btrfs_delayed_ref_node *ref;
2206         struct btrfs_delayed_ref_head *locked_ref = NULL;
2207         struct btrfs_delayed_extent_op *extent_op;
2208         int ret;
2209         int count = 0;
2210         int must_insert_reserved = 0;
2211
2212         delayed_refs = &trans->transaction->delayed_refs;
2213         while (1) {
2214                 if (!locked_ref) {
2215                         /* pick a new head ref from the cluster list */
2216                         if (list_empty(cluster))
2217                                 break;
2218
2219                         locked_ref = list_entry(cluster->next,
2220                                      struct btrfs_delayed_ref_head, cluster);
2221
2222                         /* grab the lock that says we are going to process
2223                          * all the refs for this head */
2224                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2225
2226                         /*
2227                          * we may have dropped the spin lock to get the head
2228                          * mutex lock, and that might have given someone else
2229                          * time to free the head.  If that's true, it has been
2230                          * removed from our list and we can move on.
2231                          */
2232                         if (ret == -EAGAIN) {
2233                                 locked_ref = NULL;
2234                                 count++;
2235                                 continue;
2236                         }
2237                 }
2238
2239                 /*
2240                  * record the must insert reserved flag before we
2241                  * drop the spin lock.
2242                  */
2243                 must_insert_reserved = locked_ref->must_insert_reserved;
2244                 locked_ref->must_insert_reserved = 0;
2245
2246                 extent_op = locked_ref->extent_op;
2247                 locked_ref->extent_op = NULL;
2248
2249                 /*
2250                  * locked_ref is the head node, so we have to go one
2251                  * node back for any delayed ref updates
2252                  */
2253                 ref = select_delayed_ref(locked_ref);
2254                 if (!ref) {
2255                         /* All delayed refs have been processed, Go ahead
2256                          * and send the head node to run_one_delayed_ref,
2257                          * so that any accounting fixes can happen
2258                          */
2259                         ref = &locked_ref->node;
2260
2261                         if (extent_op && must_insert_reserved) {
2262                                 kfree(extent_op);
2263                                 extent_op = NULL;
2264                         }
2265
2266                         if (extent_op) {
2267                                 spin_unlock(&delayed_refs->lock);
2268
2269                                 ret = run_delayed_extent_op(trans, root,
2270                                                             ref, extent_op);
2271                                 BUG_ON(ret);
2272                                 kfree(extent_op);
2273
2274                                 cond_resched();
2275                                 spin_lock(&delayed_refs->lock);
2276                                 continue;
2277                         }
2278
2279                         list_del_init(&locked_ref->cluster);
2280                         locked_ref = NULL;
2281                 }
2282
2283                 ref->in_tree = 0;
2284                 rb_erase(&ref->rb_node, &delayed_refs->root);
2285                 delayed_refs->num_entries--;
2286
2287                 spin_unlock(&delayed_refs->lock);
2288
2289                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2290                                           must_insert_reserved);
2291                 BUG_ON(ret);
2292
2293                 btrfs_put_delayed_ref(ref);
2294                 kfree(extent_op);
2295                 count++;
2296
2297                 cond_resched();
2298                 spin_lock(&delayed_refs->lock);
2299         }
2300         return count;
2301 }
2302
2303 /*
2304  * this starts processing the delayed reference count updates and
2305  * extent insertions we have queued up so far.  count can be
2306  * 0, which means to process everything in the tree at the start
2307  * of the run (but not newly added entries), or it can be some target
2308  * number you'd like to process.
2309  */
2310 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2311                            struct btrfs_root *root, unsigned long count)
2312 {
2313         struct rb_node *node;
2314         struct btrfs_delayed_ref_root *delayed_refs;
2315         struct btrfs_delayed_ref_node *ref;
2316         struct list_head cluster;
2317         int ret;
2318         int run_all = count == (unsigned long)-1;
2319         int run_most = 0;
2320
2321         if (root == root->fs_info->extent_root)
2322                 root = root->fs_info->tree_root;
2323
2324         delayed_refs = &trans->transaction->delayed_refs;
2325         INIT_LIST_HEAD(&cluster);
2326 again:
2327         spin_lock(&delayed_refs->lock);
2328         if (count == 0) {
2329                 count = delayed_refs->num_entries * 2;
2330                 run_most = 1;
2331         }
2332         while (1) {
2333                 if (!(run_all || run_most) &&
2334                     delayed_refs->num_heads_ready < 64)
2335                         break;
2336
2337                 /*
2338                  * go find something we can process in the rbtree.  We start at
2339                  * the beginning of the tree, and then build a cluster
2340                  * of refs to process starting at the first one we are able to
2341                  * lock
2342                  */
2343                 ret = btrfs_find_ref_cluster(trans, &cluster,
2344                                              delayed_refs->run_delayed_start);
2345                 if (ret)
2346                         break;
2347
2348                 ret = run_clustered_refs(trans, root, &cluster);
2349                 BUG_ON(ret < 0);
2350
2351                 count -= min_t(unsigned long, ret, count);
2352
2353                 if (count == 0)
2354                         break;
2355         }
2356
2357         if (run_all) {
2358                 node = rb_first(&delayed_refs->root);
2359                 if (!node)
2360                         goto out;
2361                 count = (unsigned long)-1;
2362
2363                 while (node) {
2364                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2365                                        rb_node);
2366                         if (btrfs_delayed_ref_is_head(ref)) {
2367                                 struct btrfs_delayed_ref_head *head;
2368
2369                                 head = btrfs_delayed_node_to_head(ref);
2370                                 atomic_inc(&ref->refs);
2371
2372                                 spin_unlock(&delayed_refs->lock);
2373                                 /*
2374                                  * Mutex was contended, block until it's
2375                                  * released and try again
2376                                  */
2377                                 mutex_lock(&head->mutex);
2378                                 mutex_unlock(&head->mutex);
2379
2380                                 btrfs_put_delayed_ref(ref);
2381                                 cond_resched();
2382                                 goto again;
2383                         }
2384                         node = rb_next(node);
2385                 }
2386                 spin_unlock(&delayed_refs->lock);
2387                 schedule_timeout(1);
2388                 goto again;
2389         }
2390 out:
2391         spin_unlock(&delayed_refs->lock);
2392         return 0;
2393 }
2394
2395 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2396                                 struct btrfs_root *root,
2397                                 u64 bytenr, u64 num_bytes, u64 flags,
2398                                 int is_data)
2399 {
2400         struct btrfs_delayed_extent_op *extent_op;
2401         int ret;
2402
2403         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2404         if (!extent_op)
2405                 return -ENOMEM;
2406
2407         extent_op->flags_to_set = flags;
2408         extent_op->update_flags = 1;
2409         extent_op->update_key = 0;
2410         extent_op->is_data = is_data ? 1 : 0;
2411
2412         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2413                                           num_bytes, extent_op);
2414         if (ret)
2415                 kfree(extent_op);
2416         return ret;
2417 }
2418
2419 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2420                                       struct btrfs_root *root,
2421                                       struct btrfs_path *path,
2422                                       u64 objectid, u64 offset, u64 bytenr)
2423 {
2424         struct btrfs_delayed_ref_head *head;
2425         struct btrfs_delayed_ref_node *ref;
2426         struct btrfs_delayed_data_ref *data_ref;
2427         struct btrfs_delayed_ref_root *delayed_refs;
2428         struct rb_node *node;
2429         int ret = 0;
2430
2431         ret = -ENOENT;
2432         delayed_refs = &trans->transaction->delayed_refs;
2433         spin_lock(&delayed_refs->lock);
2434         head = btrfs_find_delayed_ref_head(trans, bytenr);
2435         if (!head)
2436                 goto out;
2437
2438         if (!mutex_trylock(&head->mutex)) {
2439                 atomic_inc(&head->node.refs);
2440                 spin_unlock(&delayed_refs->lock);
2441
2442                 btrfs_release_path(path);
2443
2444                 /*
2445                  * Mutex was contended, block until it's released and let
2446                  * caller try again
2447                  */
2448                 mutex_lock(&head->mutex);
2449                 mutex_unlock(&head->mutex);
2450                 btrfs_put_delayed_ref(&head->node);
2451                 return -EAGAIN;
2452         }
2453
2454         node = rb_prev(&head->node.rb_node);
2455         if (!node)
2456                 goto out_unlock;
2457
2458         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2459
2460         if (ref->bytenr != bytenr)
2461                 goto out_unlock;
2462
2463         ret = 1;
2464         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2465                 goto out_unlock;
2466
2467         data_ref = btrfs_delayed_node_to_data_ref(ref);
2468
2469         node = rb_prev(node);
2470         if (node) {
2471                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2472                 if (ref->bytenr == bytenr)
2473                         goto out_unlock;
2474         }
2475
2476         if (data_ref->root != root->root_key.objectid ||
2477             data_ref->objectid != objectid || data_ref->offset != offset)
2478                 goto out_unlock;
2479
2480         ret = 0;
2481 out_unlock:
2482         mutex_unlock(&head->mutex);
2483 out:
2484         spin_unlock(&delayed_refs->lock);
2485         return ret;
2486 }
2487
2488 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2489                                         struct btrfs_root *root,
2490                                         struct btrfs_path *path,
2491                                         u64 objectid, u64 offset, u64 bytenr)
2492 {
2493         struct btrfs_root *extent_root = root->fs_info->extent_root;
2494         struct extent_buffer *leaf;
2495         struct btrfs_extent_data_ref *ref;
2496         struct btrfs_extent_inline_ref *iref;
2497         struct btrfs_extent_item *ei;
2498         struct btrfs_key key;
2499         u32 item_size;
2500         int ret;
2501
2502         key.objectid = bytenr;
2503         key.offset = (u64)-1;
2504         key.type = BTRFS_EXTENT_ITEM_KEY;
2505
2506         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2507         if (ret < 0)
2508                 goto out;
2509         BUG_ON(ret == 0);
2510
2511         ret = -ENOENT;
2512         if (path->slots[0] == 0)
2513                 goto out;
2514
2515         path->slots[0]--;
2516         leaf = path->nodes[0];
2517         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2518
2519         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2520                 goto out;
2521
2522         ret = 1;
2523         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2524 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2525         if (item_size < sizeof(*ei)) {
2526                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2527                 goto out;
2528         }
2529 #endif
2530         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2531
2532         if (item_size != sizeof(*ei) +
2533             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2534                 goto out;
2535
2536         if (btrfs_extent_generation(leaf, ei) <=
2537             btrfs_root_last_snapshot(&root->root_item))
2538                 goto out;
2539
2540         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2541         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2542             BTRFS_EXTENT_DATA_REF_KEY)
2543                 goto out;
2544
2545         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2546         if (btrfs_extent_refs(leaf, ei) !=
2547             btrfs_extent_data_ref_count(leaf, ref) ||
2548             btrfs_extent_data_ref_root(leaf, ref) !=
2549             root->root_key.objectid ||
2550             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2551             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2552                 goto out;
2553
2554         ret = 0;
2555 out:
2556         return ret;
2557 }
2558
2559 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2560                           struct btrfs_root *root,
2561                           u64 objectid, u64 offset, u64 bytenr)
2562 {
2563         struct btrfs_path *path;
2564         int ret;
2565         int ret2;
2566
2567         path = btrfs_alloc_path();
2568         if (!path)
2569                 return -ENOENT;
2570
2571         do {
2572                 ret = check_committed_ref(trans, root, path, objectid,
2573                                           offset, bytenr);
2574                 if (ret && ret != -ENOENT)
2575                         goto out;
2576
2577                 ret2 = check_delayed_ref(trans, root, path, objectid,
2578                                          offset, bytenr);
2579         } while (ret2 == -EAGAIN);
2580
2581         if (ret2 && ret2 != -ENOENT) {
2582                 ret = ret2;
2583                 goto out;
2584         }
2585
2586         if (ret != -ENOENT || ret2 != -ENOENT)
2587                 ret = 0;
2588 out:
2589         btrfs_free_path(path);
2590         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2591                 WARN_ON(ret > 0);
2592         return ret;
2593 }
2594
2595 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2596                            struct btrfs_root *root,
2597                            struct extent_buffer *buf,
2598                            int full_backref, int inc, int for_cow)
2599 {
2600         u64 bytenr;
2601         u64 num_bytes;
2602         u64 parent;
2603         u64 ref_root;
2604         u32 nritems;
2605         struct btrfs_key key;
2606         struct btrfs_file_extent_item *fi;
2607         int i;
2608         int level;
2609         int ret = 0;
2610         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2611                             u64, u64, u64, u64, u64, u64, int);
2612
2613         ref_root = btrfs_header_owner(buf);
2614         nritems = btrfs_header_nritems(buf);
2615         level = btrfs_header_level(buf);
2616
2617         if (!root->ref_cows && level == 0)
2618                 return 0;
2619
2620         if (inc)
2621                 process_func = btrfs_inc_extent_ref;
2622         else
2623                 process_func = btrfs_free_extent;
2624
2625         if (full_backref)
2626                 parent = buf->start;
2627         else
2628                 parent = 0;
2629
2630         for (i = 0; i < nritems; i++) {
2631                 if (level == 0) {
2632                         btrfs_item_key_to_cpu(buf, &key, i);
2633                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2634                                 continue;
2635                         fi = btrfs_item_ptr(buf, i,
2636                                             struct btrfs_file_extent_item);
2637                         if (btrfs_file_extent_type(buf, fi) ==
2638                             BTRFS_FILE_EXTENT_INLINE)
2639                                 continue;
2640                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2641                         if (bytenr == 0)
2642                                 continue;
2643
2644                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2645                         key.offset -= btrfs_file_extent_offset(buf, fi);
2646                         ret = process_func(trans, root, bytenr, num_bytes,
2647                                            parent, ref_root, key.objectid,
2648                                            key.offset, for_cow);
2649                         if (ret)
2650                                 goto fail;
2651                 } else {
2652                         bytenr = btrfs_node_blockptr(buf, i);
2653                         num_bytes = btrfs_level_size(root, level - 1);
2654                         ret = process_func(trans, root, bytenr, num_bytes,
2655                                            parent, ref_root, level - 1, 0,
2656                                            for_cow);
2657                         if (ret)
2658                                 goto fail;
2659                 }
2660         }
2661         return 0;
2662 fail:
2663         BUG();
2664         return ret;
2665 }
2666
2667 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2668                   struct extent_buffer *buf, int full_backref, int for_cow)
2669 {
2670         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2671 }
2672
2673 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2674                   struct extent_buffer *buf, int full_backref, int for_cow)
2675 {
2676         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2677 }
2678
2679 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2680                                  struct btrfs_root *root,
2681                                  struct btrfs_path *path,
2682                                  struct btrfs_block_group_cache *cache)
2683 {
2684         int ret;
2685         struct btrfs_root *extent_root = root->fs_info->extent_root;
2686         unsigned long bi;
2687         struct extent_buffer *leaf;
2688
2689         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2690         if (ret < 0)
2691                 goto fail;
2692         BUG_ON(ret);
2693
2694         leaf = path->nodes[0];
2695         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2696         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2697         btrfs_mark_buffer_dirty(leaf);
2698         btrfs_release_path(path);
2699 fail:
2700         if (ret)
2701                 return ret;
2702         return 0;
2703
2704 }
2705
2706 static struct btrfs_block_group_cache *
2707 next_block_group(struct btrfs_root *root,
2708                  struct btrfs_block_group_cache *cache)
2709 {
2710         struct rb_node *node;
2711         spin_lock(&root->fs_info->block_group_cache_lock);
2712         node = rb_next(&cache->cache_node);
2713         btrfs_put_block_group(cache);
2714         if (node) {
2715                 cache = rb_entry(node, struct btrfs_block_group_cache,
2716                                  cache_node);
2717                 btrfs_get_block_group(cache);
2718         } else
2719                 cache = NULL;
2720         spin_unlock(&root->fs_info->block_group_cache_lock);
2721         return cache;
2722 }
2723
2724 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2725                             struct btrfs_trans_handle *trans,
2726                             struct btrfs_path *path)
2727 {
2728         struct btrfs_root *root = block_group->fs_info->tree_root;
2729         struct inode *inode = NULL;
2730         u64 alloc_hint = 0;
2731         int dcs = BTRFS_DC_ERROR;
2732         int num_pages = 0;
2733         int retries = 0;
2734         int ret = 0;
2735
2736         /*
2737          * If this block group is smaller than 100 megs don't bother caching the
2738          * block group.
2739          */
2740         if (block_group->key.offset < (100 * 1024 * 1024)) {
2741                 spin_lock(&block_group->lock);
2742                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2743                 spin_unlock(&block_group->lock);
2744                 return 0;
2745         }
2746
2747 again:
2748         inode = lookup_free_space_inode(root, block_group, path);
2749         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2750                 ret = PTR_ERR(inode);
2751                 btrfs_release_path(path);
2752                 goto out;
2753         }
2754
2755         if (IS_ERR(inode)) {
2756                 BUG_ON(retries);
2757                 retries++;
2758
2759                 if (block_group->ro)
2760                         goto out_free;
2761
2762                 ret = create_free_space_inode(root, trans, block_group, path);
2763                 if (ret)
2764                         goto out_free;
2765                 goto again;
2766         }
2767
2768         /* We've already setup this transaction, go ahead and exit */
2769         if (block_group->cache_generation == trans->transid &&
2770             i_size_read(inode)) {
2771                 dcs = BTRFS_DC_SETUP;
2772                 goto out_put;
2773         }
2774
2775         /*
2776          * We want to set the generation to 0, that way if anything goes wrong
2777          * from here on out we know not to trust this cache when we load up next
2778          * time.
2779          */
2780         BTRFS_I(inode)->generation = 0;
2781         ret = btrfs_update_inode(trans, root, inode);
2782         WARN_ON(ret);
2783
2784         if (i_size_read(inode) > 0) {
2785                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2786                                                       inode);
2787                 if (ret)
2788                         goto out_put;
2789         }
2790
2791         spin_lock(&block_group->lock);
2792         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2793                 /* We're not cached, don't bother trying to write stuff out */
2794                 dcs = BTRFS_DC_WRITTEN;
2795                 spin_unlock(&block_group->lock);
2796                 goto out_put;
2797         }
2798         spin_unlock(&block_group->lock);
2799
2800         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2801         if (!num_pages)
2802                 num_pages = 1;
2803
2804         /*
2805          * Just to make absolutely sure we have enough space, we're going to
2806          * preallocate 12 pages worth of space for each block group.  In
2807          * practice we ought to use at most 8, but we need extra space so we can
2808          * add our header and have a terminator between the extents and the
2809          * bitmaps.
2810          */
2811         num_pages *= 16;
2812         num_pages *= PAGE_CACHE_SIZE;
2813
2814         ret = btrfs_check_data_free_space(inode, num_pages);
2815         if (ret)
2816                 goto out_put;
2817
2818         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2819                                               num_pages, num_pages,
2820                                               &alloc_hint);
2821         if (!ret)
2822                 dcs = BTRFS_DC_SETUP;
2823         btrfs_free_reserved_data_space(inode, num_pages);
2824
2825 out_put:
2826         iput(inode);
2827 out_free:
2828         btrfs_release_path(path);
2829 out:
2830         spin_lock(&block_group->lock);
2831         if (!ret)
2832                 block_group->cache_generation = trans->transid;
2833         block_group->disk_cache_state = dcs;
2834         spin_unlock(&block_group->lock);
2835
2836         return ret;
2837 }
2838
2839 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2840                                    struct btrfs_root *root)
2841 {
2842         struct btrfs_block_group_cache *cache;
2843         int err = 0;
2844         struct btrfs_path *path;
2845         u64 last = 0;
2846
2847         path = btrfs_alloc_path();
2848         if (!path)
2849                 return -ENOMEM;
2850
2851 again:
2852         while (1) {
2853                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2854                 while (cache) {
2855                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2856                                 break;
2857                         cache = next_block_group(root, cache);
2858                 }
2859                 if (!cache) {
2860                         if (last == 0)
2861                                 break;
2862                         last = 0;
2863                         continue;
2864                 }
2865                 err = cache_save_setup(cache, trans, path);
2866                 last = cache->key.objectid + cache->key.offset;
2867                 btrfs_put_block_group(cache);
2868         }
2869
2870         while (1) {
2871                 if (last == 0) {
2872                         err = btrfs_run_delayed_refs(trans, root,
2873                                                      (unsigned long)-1);
2874                         BUG_ON(err);
2875                 }
2876
2877                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2878                 while (cache) {
2879                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2880                                 btrfs_put_block_group(cache);
2881                                 goto again;
2882                         }
2883
2884                         if (cache->dirty)
2885                                 break;
2886                         cache = next_block_group(root, cache);
2887                 }
2888                 if (!cache) {
2889                         if (last == 0)
2890                                 break;
2891                         last = 0;
2892                         continue;
2893                 }
2894
2895                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2896                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2897                 cache->dirty = 0;
2898                 last = cache->key.objectid + cache->key.offset;
2899
2900                 err = write_one_cache_group(trans, root, path, cache);
2901                 BUG_ON(err);
2902                 btrfs_put_block_group(cache);
2903         }
2904
2905         while (1) {
2906                 /*
2907                  * I don't think this is needed since we're just marking our
2908                  * preallocated extent as written, but just in case it can't
2909                  * hurt.
2910                  */
2911                 if (last == 0) {
2912                         err = btrfs_run_delayed_refs(trans, root,
2913                                                      (unsigned long)-1);
2914                         BUG_ON(err);
2915                 }
2916
2917                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2918                 while (cache) {
2919                         /*
2920                          * Really this shouldn't happen, but it could if we
2921                          * couldn't write the entire preallocated extent and
2922                          * splitting the extent resulted in a new block.
2923                          */
2924                         if (cache->dirty) {
2925                                 btrfs_put_block_group(cache);
2926                                 goto again;
2927                         }
2928                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2929                                 break;
2930                         cache = next_block_group(root, cache);
2931                 }
2932                 if (!cache) {
2933                         if (last == 0)
2934                                 break;
2935                         last = 0;
2936                         continue;
2937                 }
2938
2939                 btrfs_write_out_cache(root, trans, cache, path);
2940
2941                 /*
2942                  * If we didn't have an error then the cache state is still
2943                  * NEED_WRITE, so we can set it to WRITTEN.
2944                  */
2945                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2946                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2947                 last = cache->key.objectid + cache->key.offset;
2948                 btrfs_put_block_group(cache);
2949         }
2950
2951         btrfs_free_path(path);
2952         return 0;
2953 }
2954
2955 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2956 {
2957         struct btrfs_block_group_cache *block_group;
2958         int readonly = 0;
2959
2960         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2961         if (!block_group || block_group->ro)
2962                 readonly = 1;
2963         if (block_group)
2964                 btrfs_put_block_group(block_group);
2965         return readonly;
2966 }
2967
2968 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2969                              u64 total_bytes, u64 bytes_used,
2970                              struct btrfs_space_info **space_info)
2971 {
2972         struct btrfs_space_info *found;
2973         int i;
2974         int factor;
2975
2976         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2977                      BTRFS_BLOCK_GROUP_RAID10))
2978                 factor = 2;
2979         else
2980                 factor = 1;
2981
2982         found = __find_space_info(info, flags);
2983         if (found) {
2984                 spin_lock(&found->lock);
2985                 found->total_bytes += total_bytes;
2986                 found->disk_total += total_bytes * factor;
2987                 found->bytes_used += bytes_used;
2988                 found->disk_used += bytes_used * factor;
2989                 found->full = 0;
2990                 spin_unlock(&found->lock);
2991                 *space_info = found;
2992                 return 0;
2993         }
2994         found = kzalloc(sizeof(*found), GFP_NOFS);
2995         if (!found)
2996                 return -ENOMEM;
2997
2998         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2999                 INIT_LIST_HEAD(&found->block_groups[i]);
3000         init_rwsem(&found->groups_sem);
3001         spin_lock_init(&found->lock);
3002         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
3003                                 BTRFS_BLOCK_GROUP_SYSTEM |
3004                                 BTRFS_BLOCK_GROUP_METADATA);
3005         found->total_bytes = total_bytes;
3006         found->disk_total = total_bytes * factor;
3007         found->bytes_used = bytes_used;
3008         found->disk_used = bytes_used * factor;
3009         found->bytes_pinned = 0;
3010         found->bytes_reserved = 0;
3011         found->bytes_readonly = 0;
3012         found->bytes_may_use = 0;
3013         found->full = 0;
3014         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3015         found->chunk_alloc = 0;
3016         found->flush = 0;
3017         init_waitqueue_head(&found->wait);
3018         *space_info = found;
3019         list_add_rcu(&found->list, &info->space_info);
3020         return 0;
3021 }
3022
3023 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3024 {
3025         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
3026                                    BTRFS_BLOCK_GROUP_RAID1 |
3027                                    BTRFS_BLOCK_GROUP_RAID10 |
3028                                    BTRFS_BLOCK_GROUP_DUP);
3029         if (extra_flags) {
3030                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3031                         fs_info->avail_data_alloc_bits |= extra_flags;
3032                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
3033                         fs_info->avail_metadata_alloc_bits |= extra_flags;
3034                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3035                         fs_info->avail_system_alloc_bits |= extra_flags;
3036         }
3037 }
3038
3039 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3040 {
3041         /*
3042          * we add in the count of missing devices because we want
3043          * to make sure that any RAID levels on a degraded FS
3044          * continue to be honored.
3045          */
3046         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3047                 root->fs_info->fs_devices->missing_devices;
3048
3049         if (num_devices == 1)
3050                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3051         if (num_devices < 4)
3052                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3053
3054         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3055             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3056                       BTRFS_BLOCK_GROUP_RAID10))) {
3057                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3058         }
3059
3060         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3061             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3062                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3063         }
3064
3065         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3066             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3067              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3068              (flags & BTRFS_BLOCK_GROUP_DUP)))
3069                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3070         return flags;
3071 }
3072
3073 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3074 {
3075         if (flags & BTRFS_BLOCK_GROUP_DATA)
3076                 flags |= root->fs_info->avail_data_alloc_bits &
3077                          root->fs_info->data_alloc_profile;
3078         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3079                 flags |= root->fs_info->avail_system_alloc_bits &
3080                          root->fs_info->system_alloc_profile;
3081         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3082                 flags |= root->fs_info->avail_metadata_alloc_bits &
3083                          root->fs_info->metadata_alloc_profile;
3084         return btrfs_reduce_alloc_profile(root, flags);
3085 }
3086
3087 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3088 {
3089         u64 flags;
3090
3091         if (data)
3092                 flags = BTRFS_BLOCK_GROUP_DATA;
3093         else if (root == root->fs_info->chunk_root)
3094                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3095         else
3096                 flags = BTRFS_BLOCK_GROUP_METADATA;
3097
3098         return get_alloc_profile(root, flags);
3099 }
3100
3101 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3102 {
3103         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3104                                                        BTRFS_BLOCK_GROUP_DATA);
3105 }
3106
3107 /*
3108  * This will check the space that the inode allocates from to make sure we have
3109  * enough space for bytes.
3110  */
3111 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3112 {
3113         struct btrfs_space_info *data_sinfo;
3114         struct btrfs_root *root = BTRFS_I(inode)->root;
3115         u64 used;
3116         int ret = 0, committed = 0, alloc_chunk = 1;
3117
3118         /* make sure bytes are sectorsize aligned */
3119         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3120
3121         if (root == root->fs_info->tree_root ||
3122             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3123                 alloc_chunk = 0;
3124                 committed = 1;
3125         }
3126
3127         data_sinfo = BTRFS_I(inode)->space_info;
3128         if (!data_sinfo)
3129                 goto alloc;
3130
3131 again:
3132         /* make sure we have enough space to handle the data first */
3133         spin_lock(&data_sinfo->lock);
3134         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3135                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3136                 data_sinfo->bytes_may_use;
3137
3138         if (used + bytes > data_sinfo->total_bytes) {
3139                 struct btrfs_trans_handle *trans;
3140
3141                 /*
3142                  * if we don't have enough free bytes in this space then we need
3143                  * to alloc a new chunk.
3144                  */
3145                 if (!data_sinfo->full && alloc_chunk) {
3146                         u64 alloc_target;
3147
3148                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3149                         spin_unlock(&data_sinfo->lock);
3150 alloc:
3151                         alloc_target = btrfs_get_alloc_profile(root, 1);
3152                         trans = btrfs_join_transaction(root);
3153                         if (IS_ERR(trans))
3154                                 return PTR_ERR(trans);
3155
3156                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3157                                              bytes + 2 * 1024 * 1024,
3158                                              alloc_target,
3159                                              CHUNK_ALLOC_NO_FORCE);
3160                         btrfs_end_transaction(trans, root);
3161                         if (ret < 0) {
3162                                 if (ret != -ENOSPC)
3163                                         return ret;
3164                                 else
3165                                         goto commit_trans;
3166                         }
3167
3168                         if (!data_sinfo) {
3169                                 btrfs_set_inode_space_info(root, inode);
3170                                 data_sinfo = BTRFS_I(inode)->space_info;
3171                         }
3172                         goto again;
3173                 }
3174
3175                 /*
3176                  * If we have less pinned bytes than we want to allocate then
3177                  * don't bother committing the transaction, it won't help us.
3178                  */
3179                 if (data_sinfo->bytes_pinned < bytes)
3180                         committed = 1;
3181                 spin_unlock(&data_sinfo->lock);
3182
3183                 /* commit the current transaction and try again */
3184 commit_trans:
3185                 if (!committed &&
3186                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3187                         committed = 1;
3188                         trans = btrfs_join_transaction(root);
3189                         if (IS_ERR(trans))
3190                                 return PTR_ERR(trans);
3191                         ret = btrfs_commit_transaction(trans, root);
3192                         if (ret)
3193                                 return ret;
3194                         goto again;
3195                 }
3196
3197                 return -ENOSPC;
3198         }
3199         data_sinfo->bytes_may_use += bytes;
3200         spin_unlock(&data_sinfo->lock);
3201
3202         return 0;
3203 }
3204
3205 /*
3206  * Called if we need to clear a data reservation for this inode.
3207  */
3208 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3209 {
3210         struct btrfs_root *root = BTRFS_I(inode)->root;
3211         struct btrfs_space_info *data_sinfo;
3212
3213         /* make sure bytes are sectorsize aligned */
3214         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3215
3216         data_sinfo = BTRFS_I(inode)->space_info;
3217         spin_lock(&data_sinfo->lock);
3218         data_sinfo->bytes_may_use -= bytes;
3219         spin_unlock(&data_sinfo->lock);
3220 }
3221
3222 static void force_metadata_allocation(struct btrfs_fs_info *info)
3223 {
3224         struct list_head *head = &info->space_info;
3225         struct btrfs_space_info *found;
3226
3227         rcu_read_lock();
3228         list_for_each_entry_rcu(found, head, list) {
3229                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3230                         found->force_alloc = CHUNK_ALLOC_FORCE;
3231         }
3232         rcu_read_unlock();
3233 }
3234
3235 static int should_alloc_chunk(struct btrfs_root *root,
3236                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3237                               int force)
3238 {
3239         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3240         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3241         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3242         u64 thresh;
3243
3244         if (force == CHUNK_ALLOC_FORCE)
3245                 return 1;
3246
3247         /*
3248          * We need to take into account the global rsv because for all intents
3249          * and purposes it's used space.  Don't worry about locking the
3250          * global_rsv, it doesn't change except when the transaction commits.
3251          */
3252         num_allocated += global_rsv->size;
3253
3254         /*
3255          * in limited mode, we want to have some free space up to
3256          * about 1% of the FS size.
3257          */
3258         if (force == CHUNK_ALLOC_LIMITED) {
3259                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3260                 thresh = max_t(u64, 64 * 1024 * 1024,
3261                                div_factor_fine(thresh, 1));
3262
3263                 if (num_bytes - num_allocated < thresh)
3264                         return 1;
3265         }
3266
3267         /*
3268          * we have two similar checks here, one based on percentage
3269          * and once based on a hard number of 256MB.  The idea
3270          * is that if we have a good amount of free
3271          * room, don't allocate a chunk.  A good mount is
3272          * less than 80% utilized of the chunks we have allocated,
3273          * or more than 256MB free
3274          */
3275         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3276                 return 0;
3277
3278         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3279                 return 0;
3280
3281         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3282
3283         /* 256MB or 5% of the FS */
3284         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3285
3286         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3287                 return 0;
3288         return 1;
3289 }
3290
3291 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3292                           struct btrfs_root *extent_root, u64 alloc_bytes,
3293                           u64 flags, int force)
3294 {
3295         struct btrfs_space_info *space_info;
3296         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3297         int wait_for_alloc = 0;
3298         int ret = 0;
3299
3300         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3301
3302         space_info = __find_space_info(extent_root->fs_info, flags);
3303         if (!space_info) {
3304                 ret = update_space_info(extent_root->fs_info, flags,
3305                                         0, 0, &space_info);
3306                 BUG_ON(ret);
3307         }
3308         BUG_ON(!space_info);
3309
3310 again:
3311         spin_lock(&space_info->lock);
3312         if (space_info->force_alloc)
3313                 force = space_info->force_alloc;
3314         if (space_info->full) {
3315                 spin_unlock(&space_info->lock);
3316                 return 0;
3317         }
3318
3319         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3320                 spin_unlock(&space_info->lock);
3321                 return 0;
3322         } else if (space_info->chunk_alloc) {
3323                 wait_for_alloc = 1;
3324         } else {
3325                 space_info->chunk_alloc = 1;
3326         }
3327
3328         spin_unlock(&space_info->lock);
3329
3330         mutex_lock(&fs_info->chunk_mutex);
3331
3332         /*
3333          * The chunk_mutex is held throughout the entirety of a chunk
3334          * allocation, so once we've acquired the chunk_mutex we know that the
3335          * other guy is done and we need to recheck and see if we should
3336          * allocate.
3337          */
3338         if (wait_for_alloc) {
3339                 mutex_unlock(&fs_info->chunk_mutex);
3340                 wait_for_alloc = 0;
3341                 goto again;
3342         }
3343
3344         /*
3345          * If we have mixed data/metadata chunks we want to make sure we keep
3346          * allocating mixed chunks instead of individual chunks.
3347          */
3348         if (btrfs_mixed_space_info(space_info))
3349                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3350
3351         /*
3352          * if we're doing a data chunk, go ahead and make sure that
3353          * we keep a reasonable number of metadata chunks allocated in the
3354          * FS as well.
3355          */
3356         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3357                 fs_info->data_chunk_allocations++;
3358                 if (!(fs_info->data_chunk_allocations %
3359                       fs_info->metadata_ratio))
3360                         force_metadata_allocation(fs_info);
3361         }
3362
3363         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3364         if (ret < 0 && ret != -ENOSPC)
3365                 goto out;
3366
3367         spin_lock(&space_info->lock);
3368         if (ret)
3369                 space_info->full = 1;
3370         else
3371                 ret = 1;
3372
3373         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3374         space_info->chunk_alloc = 0;
3375         spin_unlock(&space_info->lock);
3376 out:
3377         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3378         return ret;
3379 }
3380
3381 /*
3382  * shrink metadata reservation for delalloc
3383  */
3384 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3385                            bool wait_ordered)
3386 {
3387         struct btrfs_block_rsv *block_rsv;
3388         struct btrfs_space_info *space_info;
3389         struct btrfs_trans_handle *trans;
3390         u64 reserved;
3391         u64 max_reclaim;
3392         u64 reclaimed = 0;
3393         long time_left;
3394         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3395         int loops = 0;
3396         unsigned long progress;
3397
3398         trans = (struct btrfs_trans_handle *)current->journal_info;
3399         block_rsv = &root->fs_info->delalloc_block_rsv;
3400         space_info = block_rsv->space_info;
3401
3402         smp_mb();
3403         reserved = space_info->bytes_may_use;
3404         progress = space_info->reservation_progress;
3405
3406         if (reserved == 0)
3407                 return 0;
3408
3409         smp_mb();
3410         if (root->fs_info->delalloc_bytes == 0) {
3411                 if (trans)
3412                         return 0;
3413                 btrfs_wait_ordered_extents(root, 0, 0);
3414                 return 0;
3415         }
3416
3417         max_reclaim = min(reserved, to_reclaim);
3418         nr_pages = max_t(unsigned long, nr_pages,
3419                          max_reclaim >> PAGE_CACHE_SHIFT);
3420         while (loops < 1024) {
3421                 /* have the flusher threads jump in and do some IO */
3422                 smp_mb();
3423                 nr_pages = min_t(unsigned long, nr_pages,
3424                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3425                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3426
3427                 spin_lock(&space_info->lock);
3428                 if (reserved > space_info->bytes_may_use)
3429                         reclaimed += reserved - space_info->bytes_may_use;
3430                 reserved = space_info->bytes_may_use;
3431                 spin_unlock(&space_info->lock);
3432
3433                 loops++;
3434
3435                 if (reserved == 0 || reclaimed >= max_reclaim)
3436                         break;
3437
3438                 if (trans && trans->transaction->blocked)
3439                         return -EAGAIN;
3440
3441                 if (wait_ordered && !trans) {
3442                         btrfs_wait_ordered_extents(root, 0, 0);
3443                 } else {
3444                         time_left = schedule_timeout_interruptible(1);
3445
3446                         /* We were interrupted, exit */
3447                         if (time_left)
3448                                 break;
3449                 }
3450
3451                 /* we've kicked the IO a few times, if anything has been freed,
3452                  * exit.  There is no sense in looping here for a long time
3453                  * when we really need to commit the transaction, or there are
3454                  * just too many writers without enough free space
3455                  */
3456
3457                 if (loops > 3) {
3458                         smp_mb();
3459                         if (progress != space_info->reservation_progress)
3460                                 break;
3461                 }
3462
3463         }
3464
3465         return reclaimed >= to_reclaim;
3466 }
3467
3468 /**
3469  * maybe_commit_transaction - possibly commit the transaction if its ok to
3470  * @root - the root we're allocating for
3471  * @bytes - the number of bytes we want to reserve
3472  * @force - force the commit
3473  *
3474  * This will check to make sure that committing the transaction will actually
3475  * get us somewhere and then commit the transaction if it does.  Otherwise it
3476  * will return -ENOSPC.
3477  */
3478 static int may_commit_transaction(struct btrfs_root *root,
3479                                   struct btrfs_space_info *space_info,
3480                                   u64 bytes, int force)
3481 {
3482         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3483         struct btrfs_trans_handle *trans;
3484
3485         trans = (struct btrfs_trans_handle *)current->journal_info;
3486         if (trans)
3487                 return -EAGAIN;
3488
3489         if (force)
3490                 goto commit;
3491
3492         /* See if there is enough pinned space to make this reservation */
3493         spin_lock(&space_info->lock);
3494         if (space_info->bytes_pinned >= bytes) {
3495                 spin_unlock(&space_info->lock);
3496                 goto commit;
3497         }
3498         spin_unlock(&space_info->lock);
3499
3500         /*
3501          * See if there is some space in the delayed insertion reservation for
3502          * this reservation.
3503          */
3504         if (space_info != delayed_rsv->space_info)
3505                 return -ENOSPC;
3506
3507         spin_lock(&delayed_rsv->lock);
3508         if (delayed_rsv->size < bytes) {
3509                 spin_unlock(&delayed_rsv->lock);
3510                 return -ENOSPC;
3511         }
3512         spin_unlock(&delayed_rsv->lock);
3513
3514 commit:
3515         trans = btrfs_join_transaction(root);
3516         if (IS_ERR(trans))
3517                 return -ENOSPC;
3518
3519         return btrfs_commit_transaction(trans, root);
3520 }
3521
3522 /**
3523  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3524  * @root - the root we're allocating for
3525  * @block_rsv - the block_rsv we're allocating for
3526  * @orig_bytes - the number of bytes we want
3527  * @flush - wether or not we can flush to make our reservation
3528  *
3529  * This will reserve orgi_bytes number of bytes from the space info associated
3530  * with the block_rsv.  If there is not enough space it will make an attempt to
3531  * flush out space to make room.  It will do this by flushing delalloc if
3532  * possible or committing the transaction.  If flush is 0 then no attempts to
3533  * regain reservations will be made and this will fail if there is not enough
3534  * space already.
3535  */
3536 static int reserve_metadata_bytes(struct btrfs_root *root,
3537                                   struct btrfs_block_rsv *block_rsv,
3538                                   u64 orig_bytes, int flush)
3539 {
3540         struct btrfs_space_info *space_info = block_rsv->space_info;
3541         u64 used;
3542         u64 num_bytes = orig_bytes;
3543         int retries = 0;
3544         int ret = 0;
3545         bool committed = false;
3546         bool flushing = false;
3547         bool wait_ordered = false;
3548
3549 again:
3550         ret = 0;
3551         spin_lock(&space_info->lock);
3552         /*
3553          * We only want to wait if somebody other than us is flushing and we are
3554          * actually alloed to flush.
3555          */
3556         while (flush && !flushing && space_info->flush) {
3557                 spin_unlock(&space_info->lock);
3558                 /*
3559                  * If we have a trans handle we can't wait because the flusher
3560                  * may have to commit the transaction, which would mean we would
3561                  * deadlock since we are waiting for the flusher to finish, but
3562                  * hold the current transaction open.
3563                  */
3564                 if (current->journal_info)
3565                         return -EAGAIN;
3566                 ret = wait_event_interruptible(space_info->wait,
3567                                                !space_info->flush);
3568                 /* Must have been interrupted, return */
3569                 if (ret)
3570                         return -EINTR;
3571
3572                 spin_lock(&space_info->lock);
3573         }
3574
3575         ret = -ENOSPC;
3576         used = space_info->bytes_used + space_info->bytes_reserved +
3577                 space_info->bytes_pinned + space_info->bytes_readonly +
3578                 space_info->bytes_may_use;
3579
3580         /*
3581          * The idea here is that we've not already over-reserved the block group
3582          * then we can go ahead and save our reservation first and then start
3583          * flushing if we need to.  Otherwise if we've already overcommitted
3584          * lets start flushing stuff first and then come back and try to make
3585          * our reservation.
3586          */
3587         if (used <= space_info->total_bytes) {
3588                 if (used + orig_bytes <= space_info->total_bytes) {
3589                         space_info->bytes_may_use += orig_bytes;
3590                         ret = 0;
3591                 } else {
3592                         /*
3593                          * Ok set num_bytes to orig_bytes since we aren't
3594                          * overocmmitted, this way we only try and reclaim what
3595                          * we need.
3596                          */
3597                         num_bytes = orig_bytes;
3598                 }
3599         } else {
3600                 /*
3601                  * Ok we're over committed, set num_bytes to the overcommitted
3602                  * amount plus the amount of bytes that we need for this
3603                  * reservation.
3604                  */
3605                 wait_ordered = true;
3606                 num_bytes = used - space_info->total_bytes +
3607                         (orig_bytes * (retries + 1));
3608         }
3609
3610         if (ret) {
3611                 u64 profile = btrfs_get_alloc_profile(root, 0);
3612                 u64 avail;
3613
3614                 /*
3615                  * If we have a lot of space that's pinned, don't bother doing
3616                  * the overcommit dance yet and just commit the transaction.
3617                  */
3618                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3619                 do_div(avail, 10);
3620                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3621                         space_info->flush = 1;
3622                         flushing = true;
3623                         spin_unlock(&space_info->lock);
3624                         ret = may_commit_transaction(root, space_info,
3625                                                      orig_bytes, 1);
3626                         if (ret)
3627                                 goto out;
3628                         committed = true;
3629                         goto again;
3630                 }
3631
3632                 spin_lock(&root->fs_info->free_chunk_lock);
3633                 avail = root->fs_info->free_chunk_space;
3634
3635                 /*
3636                  * If we have dup, raid1 or raid10 then only half of the free
3637                  * space is actually useable.
3638                  */
3639                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3640                                BTRFS_BLOCK_GROUP_RAID1 |
3641                                BTRFS_BLOCK_GROUP_RAID10))
3642                         avail >>= 1;
3643
3644                 /*
3645                  * If we aren't flushing don't let us overcommit too much, say
3646                  * 1/8th of the space.  If we can flush, let it overcommit up to
3647                  * 1/2 of the space.
3648                  */
3649                 if (flush)
3650                         avail >>= 3;
3651                 else
3652                         avail >>= 1;
3653                  spin_unlock(&root->fs_info->free_chunk_lock);
3654
3655                 if (used + num_bytes < space_info->total_bytes + avail) {
3656                         space_info->bytes_may_use += orig_bytes;
3657                         ret = 0;
3658                 } else {
3659                         wait_ordered = true;
3660                 }
3661         }
3662
3663         /*
3664          * Couldn't make our reservation, save our place so while we're trying
3665          * to reclaim space we can actually use it instead of somebody else
3666          * stealing it from us.
3667          */
3668         if (ret && flush) {
3669                 flushing = true;
3670                 space_info->flush = 1;
3671         }
3672
3673         spin_unlock(&space_info->lock);
3674
3675         if (!ret || !flush)
3676                 goto out;
3677
3678         /*
3679          * We do synchronous shrinking since we don't actually unreserve
3680          * metadata until after the IO is completed.
3681          */
3682         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3683         if (ret < 0)
3684                 goto out;
3685
3686         ret = 0;
3687
3688         /*
3689          * So if we were overcommitted it's possible that somebody else flushed
3690          * out enough space and we simply didn't have enough space to reclaim,
3691          * so go back around and try again.
3692          */
3693         if (retries < 2) {
3694                 wait_ordered = true;
3695                 retries++;
3696                 goto again;
3697         }
3698
3699         ret = -ENOSPC;
3700         if (committed)
3701                 goto out;
3702
3703         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3704         if (!ret) {
3705                 committed = true;
3706                 goto again;
3707         }
3708
3709 out:
3710         if (flushing) {
3711                 spin_lock(&space_info->lock);
3712                 space_info->flush = 0;
3713                 wake_up_all(&space_info->wait);
3714                 spin_unlock(&space_info->lock);
3715         }
3716         return ret;
3717 }
3718
3719 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3720                                              struct btrfs_root *root)
3721 {
3722         struct btrfs_block_rsv *block_rsv = NULL;
3723
3724         if (root->ref_cows || root == root->fs_info->csum_root)
3725                 block_rsv = trans->block_rsv;
3726
3727         if (!block_rsv)
3728                 block_rsv = root->block_rsv;
3729
3730         if (!block_rsv)
3731                 block_rsv = &root->fs_info->empty_block_rsv;
3732
3733         return block_rsv;
3734 }
3735
3736 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3737                                u64 num_bytes)
3738 {
3739         int ret = -ENOSPC;
3740         spin_lock(&block_rsv->lock);
3741         if (block_rsv->reserved >= num_bytes) {
3742                 block_rsv->reserved -= num_bytes;
3743                 if (block_rsv->reserved < block_rsv->size)
3744                         block_rsv->full = 0;
3745                 ret = 0;
3746         }
3747         spin_unlock(&block_rsv->lock);
3748         return ret;
3749 }
3750
3751 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3752                                 u64 num_bytes, int update_size)
3753 {
3754         spin_lock(&block_rsv->lock);
3755         block_rsv->reserved += num_bytes;
3756         if (update_size)
3757                 block_rsv->size += num_bytes;
3758         else if (block_rsv->reserved >= block_rsv->size)
3759                 block_rsv->full = 1;
3760         spin_unlock(&block_rsv->lock);
3761 }
3762
3763 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3764                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3765 {
3766         struct btrfs_space_info *space_info = block_rsv->space_info;
3767
3768         spin_lock(&block_rsv->lock);
3769         if (num_bytes == (u64)-1)
3770                 num_bytes = block_rsv->size;
3771         block_rsv->size -= num_bytes;
3772         if (block_rsv->reserved >= block_rsv->size) {
3773                 num_bytes = block_rsv->reserved - block_rsv->size;
3774                 block_rsv->reserved = block_rsv->size;
3775                 block_rsv->full = 1;
3776         } else {
3777                 num_bytes = 0;
3778         }
3779         spin_unlock(&block_rsv->lock);
3780
3781         if (num_bytes > 0) {
3782                 if (dest) {
3783                         spin_lock(&dest->lock);
3784                         if (!dest->full) {
3785                                 u64 bytes_to_add;
3786
3787                                 bytes_to_add = dest->size - dest->reserved;
3788                                 bytes_to_add = min(num_bytes, bytes_to_add);
3789                                 dest->reserved += bytes_to_add;
3790                                 if (dest->reserved >= dest->size)
3791                                         dest->full = 1;
3792                                 num_bytes -= bytes_to_add;
3793                         }
3794                         spin_unlock(&dest->lock);
3795                 }
3796                 if (num_bytes) {
3797                         spin_lock(&space_info->lock);
3798                         space_info->bytes_may_use -= num_bytes;
3799                         space_info->reservation_progress++;
3800                         spin_unlock(&space_info->lock);
3801                 }
3802         }
3803 }
3804
3805 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3806                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3807 {
3808         int ret;
3809
3810         ret = block_rsv_use_bytes(src, num_bytes);
3811         if (ret)
3812                 return ret;
3813
3814         block_rsv_add_bytes(dst, num_bytes, 1);
3815         return 0;
3816 }
3817
3818 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3819 {
3820         memset(rsv, 0, sizeof(*rsv));
3821         spin_lock_init(&rsv->lock);
3822 }
3823
3824 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3825 {
3826         struct btrfs_block_rsv *block_rsv;
3827         struct btrfs_fs_info *fs_info = root->fs_info;
3828
3829         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3830         if (!block_rsv)
3831                 return NULL;
3832
3833         btrfs_init_block_rsv(block_rsv);
3834         block_rsv->space_info = __find_space_info(fs_info,
3835                                                   BTRFS_BLOCK_GROUP_METADATA);
3836         return block_rsv;
3837 }
3838
3839 void btrfs_free_block_rsv(struct btrfs_root *root,
3840                           struct btrfs_block_rsv *rsv)
3841 {
3842         btrfs_block_rsv_release(root, rsv, (u64)-1);
3843         kfree(rsv);
3844 }
3845
3846 static inline int __block_rsv_add(struct btrfs_root *root,
3847                                   struct btrfs_block_rsv *block_rsv,
3848                                   u64 num_bytes, int flush)
3849 {
3850         int ret;
3851
3852         if (num_bytes == 0)
3853                 return 0;
3854
3855         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3856         if (!ret) {
3857                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3858                 return 0;
3859         }
3860
3861         return ret;
3862 }
3863
3864 int btrfs_block_rsv_add(struct btrfs_root *root,
3865                         struct btrfs_block_rsv *block_rsv,
3866                         u64 num_bytes)
3867 {
3868         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3869 }
3870
3871 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3872                                 struct btrfs_block_rsv *block_rsv,
3873                                 u64 num_bytes)
3874 {
3875         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3876 }
3877
3878 int btrfs_block_rsv_check(struct btrfs_root *root,
3879                           struct btrfs_block_rsv *block_rsv, int min_factor)
3880 {
3881         u64 num_bytes = 0;
3882         int ret = -ENOSPC;
3883
3884         if (!block_rsv)
3885                 return 0;
3886
3887         spin_lock(&block_rsv->lock);
3888         num_bytes = div_factor(block_rsv->size, min_factor);
3889         if (block_rsv->reserved >= num_bytes)
3890                 ret = 0;
3891         spin_unlock(&block_rsv->lock);
3892
3893         return ret;
3894 }
3895
3896 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
3897                                            struct btrfs_block_rsv *block_rsv,
3898                                            u64 min_reserved, int flush)
3899 {
3900         u64 num_bytes = 0;
3901         int ret = -ENOSPC;
3902
3903         if (!block_rsv)
3904                 return 0;
3905
3906         spin_lock(&block_rsv->lock);
3907         num_bytes = min_reserved;
3908         if (block_rsv->reserved >= num_bytes)
3909                 ret = 0;
3910         else
3911                 num_bytes -= block_rsv->reserved;
3912         spin_unlock(&block_rsv->lock);
3913
3914         if (!ret)
3915                 return 0;
3916
3917         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3918         if (!ret) {
3919                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3920                 return 0;
3921         }
3922
3923         return ret;
3924 }
3925
3926 int btrfs_block_rsv_refill(struct btrfs_root *root,
3927                            struct btrfs_block_rsv *block_rsv,
3928                            u64 min_reserved)
3929 {
3930         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
3931 }
3932
3933 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
3934                                    struct btrfs_block_rsv *block_rsv,
3935                                    u64 min_reserved)
3936 {
3937         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
3938 }
3939
3940 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3941                             struct btrfs_block_rsv *dst_rsv,
3942                             u64 num_bytes)
3943 {
3944         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3945 }
3946
3947 void btrfs_block_rsv_release(struct btrfs_root *root,
3948                              struct btrfs_block_rsv *block_rsv,
3949                              u64 num_bytes)
3950 {
3951         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3952         if (global_rsv->full || global_rsv == block_rsv ||
3953             block_rsv->space_info != global_rsv->space_info)
3954                 global_rsv = NULL;
3955         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3956 }
3957
3958 /*
3959  * helper to calculate size of global block reservation.
3960  * the desired value is sum of space used by extent tree,
3961  * checksum tree and root tree
3962  */
3963 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3964 {
3965         struct btrfs_space_info *sinfo;
3966         u64 num_bytes;
3967         u64 meta_used;
3968         u64 data_used;
3969         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3970
3971         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3972         spin_lock(&sinfo->lock);
3973         data_used = sinfo->bytes_used;
3974         spin_unlock(&sinfo->lock);
3975
3976         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3977         spin_lock(&sinfo->lock);
3978         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3979                 data_used = 0;
3980         meta_used = sinfo->bytes_used;
3981         spin_unlock(&sinfo->lock);
3982
3983         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3984                     csum_size * 2;
3985         num_bytes += div64_u64(data_used + meta_used, 50);
3986
3987         if (num_bytes * 3 > meta_used)
3988                 num_bytes = div64_u64(meta_used, 3);
3989
3990         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3991 }
3992
3993 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3994 {
3995         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3996         struct btrfs_space_info *sinfo = block_rsv->space_info;
3997         u64 num_bytes;
3998
3999         num_bytes = calc_global_metadata_size(fs_info);
4000
4001         spin_lock(&block_rsv->lock);
4002         spin_lock(&sinfo->lock);
4003
4004         block_rsv->size = num_bytes;
4005
4006         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4007                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4008                     sinfo->bytes_may_use;
4009
4010         if (sinfo->total_bytes > num_bytes) {
4011                 num_bytes = sinfo->total_bytes - num_bytes;
4012                 block_rsv->reserved += num_bytes;
4013                 sinfo->bytes_may_use += num_bytes;
4014         }
4015
4016         if (block_rsv->reserved >= block_rsv->size) {
4017                 num_bytes = block_rsv->reserved - block_rsv->size;
4018                 sinfo->bytes_may_use -= num_bytes;
4019                 sinfo->reservation_progress++;
4020                 block_rsv->reserved = block_rsv->size;
4021                 block_rsv->full = 1;
4022         }
4023
4024         spin_unlock(&sinfo->lock);
4025         spin_unlock(&block_rsv->lock);
4026 }
4027
4028 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4029 {
4030         struct btrfs_space_info *space_info;
4031
4032         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4033         fs_info->chunk_block_rsv.space_info = space_info;
4034
4035         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4036         fs_info->global_block_rsv.space_info = space_info;
4037         fs_info->delalloc_block_rsv.space_info = space_info;
4038         fs_info->trans_block_rsv.space_info = space_info;
4039         fs_info->empty_block_rsv.space_info = space_info;
4040         fs_info->delayed_block_rsv.space_info = space_info;
4041
4042         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4043         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4044         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4045         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4046         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4047
4048         update_global_block_rsv(fs_info);
4049 }
4050
4051 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4052 {
4053         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
4054         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4055         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4056         WARN_ON(fs_info->trans_block_rsv.size > 0);
4057         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4058         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4059         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4060         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4061         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4062 }
4063
4064 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4065                                   struct btrfs_root *root)
4066 {
4067         if (!trans->bytes_reserved)
4068                 return;
4069
4070         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4071         trans->bytes_reserved = 0;
4072 }
4073
4074 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4075                                   struct inode *inode)
4076 {
4077         struct btrfs_root *root = BTRFS_I(inode)->root;
4078         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4079         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4080
4081         /*
4082          * We need to hold space in order to delete our orphan item once we've
4083          * added it, so this takes the reservation so we can release it later
4084          * when we are truly done with the orphan item.
4085          */
4086         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4087         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4088 }
4089
4090 void btrfs_orphan_release_metadata(struct inode *inode)
4091 {
4092         struct btrfs_root *root = BTRFS_I(inode)->root;
4093         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4094         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4095 }
4096
4097 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4098                                 struct btrfs_pending_snapshot *pending)
4099 {
4100         struct btrfs_root *root = pending->root;
4101         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4102         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4103         /*
4104          * two for root back/forward refs, two for directory entries
4105          * and one for root of the snapshot.
4106          */
4107         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4108         dst_rsv->space_info = src_rsv->space_info;
4109         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4110 }
4111
4112 /**
4113  * drop_outstanding_extent - drop an outstanding extent
4114  * @inode: the inode we're dropping the extent for
4115  *
4116  * This is called when we are freeing up an outstanding extent, either called
4117  * after an error or after an extent is written.  This will return the number of
4118  * reserved extents that need to be freed.  This must be called with
4119  * BTRFS_I(inode)->lock held.
4120  */
4121 static unsigned drop_outstanding_extent(struct inode *inode)
4122 {
4123         unsigned drop_inode_space = 0;
4124         unsigned dropped_extents = 0;
4125
4126         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4127         BTRFS_I(inode)->outstanding_extents--;
4128
4129         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4130             BTRFS_I(inode)->delalloc_meta_reserved) {
4131                 drop_inode_space = 1;
4132                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4133         }
4134
4135         /*
4136          * If we have more or the same amount of outsanding extents than we have
4137          * reserved then we need to leave the reserved extents count alone.
4138          */
4139         if (BTRFS_I(inode)->outstanding_extents >=
4140             BTRFS_I(inode)->reserved_extents)
4141                 return drop_inode_space;
4142
4143         dropped_extents = BTRFS_I(inode)->reserved_extents -
4144                 BTRFS_I(inode)->outstanding_extents;
4145         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4146         return dropped_extents + drop_inode_space;
4147 }
4148
4149 /**
4150  * calc_csum_metadata_size - return the amount of metada space that must be
4151  *      reserved/free'd for the given bytes.
4152  * @inode: the inode we're manipulating
4153  * @num_bytes: the number of bytes in question
4154  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4155  *
4156  * This adjusts the number of csum_bytes in the inode and then returns the
4157  * correct amount of metadata that must either be reserved or freed.  We
4158  * calculate how many checksums we can fit into one leaf and then divide the
4159  * number of bytes that will need to be checksumed by this value to figure out
4160  * how many checksums will be required.  If we are adding bytes then the number
4161  * may go up and we will return the number of additional bytes that must be
4162  * reserved.  If it is going down we will return the number of bytes that must
4163  * be freed.
4164  *
4165  * This must be called with BTRFS_I(inode)->lock held.
4166  */
4167 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4168                                    int reserve)
4169 {
4170         struct btrfs_root *root = BTRFS_I(inode)->root;
4171         u64 csum_size;
4172         int num_csums_per_leaf;
4173         int num_csums;
4174         int old_csums;
4175
4176         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4177             BTRFS_I(inode)->csum_bytes == 0)
4178                 return 0;
4179
4180         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4181         if (reserve)
4182                 BTRFS_I(inode)->csum_bytes += num_bytes;
4183         else
4184                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4185         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4186         num_csums_per_leaf = (int)div64_u64(csum_size,
4187                                             sizeof(struct btrfs_csum_item) +
4188                                             sizeof(struct btrfs_disk_key));
4189         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4190         num_csums = num_csums + num_csums_per_leaf - 1;
4191         num_csums = num_csums / num_csums_per_leaf;
4192
4193         old_csums = old_csums + num_csums_per_leaf - 1;
4194         old_csums = old_csums / num_csums_per_leaf;
4195
4196         /* No change, no need to reserve more */
4197         if (old_csums == num_csums)
4198                 return 0;
4199
4200         if (reserve)
4201                 return btrfs_calc_trans_metadata_size(root,
4202                                                       num_csums - old_csums);
4203
4204         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4205 }
4206
4207 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4208 {
4209         struct btrfs_root *root = BTRFS_I(inode)->root;
4210         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4211         u64 to_reserve = 0;
4212         unsigned nr_extents = 0;
4213         int flush = 1;
4214         int ret;
4215
4216         if (btrfs_is_free_space_inode(root, inode))
4217                 flush = 0;
4218
4219         if (flush && btrfs_transaction_in_commit(root->fs_info))
4220                 schedule_timeout(1);
4221
4222         num_bytes = ALIGN(num_bytes, root->sectorsize);
4223
4224         spin_lock(&BTRFS_I(inode)->lock);
4225         BTRFS_I(inode)->outstanding_extents++;
4226
4227         if (BTRFS_I(inode)->outstanding_extents >
4228             BTRFS_I(inode)->reserved_extents) {
4229                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4230                         BTRFS_I(inode)->reserved_extents;
4231                 BTRFS_I(inode)->reserved_extents += nr_extents;
4232         }
4233
4234         /*
4235          * Add an item to reserve for updating the inode when we complete the
4236          * delalloc io.
4237          */
4238         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4239                 nr_extents++;
4240                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4241         }
4242
4243         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4244         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4245         spin_unlock(&BTRFS_I(inode)->lock);
4246
4247         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4248         if (ret) {
4249                 u64 to_free = 0;
4250                 unsigned dropped;
4251
4252                 spin_lock(&BTRFS_I(inode)->lock);
4253                 dropped = drop_outstanding_extent(inode);
4254                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4255                 spin_unlock(&BTRFS_I(inode)->lock);
4256                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4257
4258                 /*
4259                  * Somebody could have come in and twiddled with the
4260                  * reservation, so if we have to free more than we would have
4261                  * reserved from this reservation go ahead and release those
4262                  * bytes.
4263                  */
4264                 to_free -= to_reserve;
4265                 if (to_free)
4266                         btrfs_block_rsv_release(root, block_rsv, to_free);
4267                 return ret;
4268         }
4269
4270         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4271
4272         return 0;
4273 }
4274
4275 /**
4276  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4277  * @inode: the inode to release the reservation for
4278  * @num_bytes: the number of bytes we're releasing
4279  *
4280  * This will release the metadata reservation for an inode.  This can be called
4281  * once we complete IO for a given set of bytes to release their metadata
4282  * reservations.
4283  */
4284 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4285 {
4286         struct btrfs_root *root = BTRFS_I(inode)->root;
4287         u64 to_free = 0;
4288         unsigned dropped;
4289
4290         num_bytes = ALIGN(num_bytes, root->sectorsize);
4291         spin_lock(&BTRFS_I(inode)->lock);
4292         dropped = drop_outstanding_extent(inode);
4293
4294         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4295         spin_unlock(&BTRFS_I(inode)->lock);
4296         if (dropped > 0)
4297                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4298
4299         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4300                                 to_free);
4301 }
4302
4303 /**
4304  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4305  * @inode: inode we're writing to
4306  * @num_bytes: the number of bytes we want to allocate
4307  *
4308  * This will do the following things
4309  *
4310  * o reserve space in the data space info for num_bytes
4311  * o reserve space in the metadata space info based on number of outstanding
4312  *   extents and how much csums will be needed
4313  * o add to the inodes ->delalloc_bytes
4314  * o add it to the fs_info's delalloc inodes list.
4315  *
4316  * This will return 0 for success and -ENOSPC if there is no space left.
4317  */
4318 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4319 {
4320         int ret;
4321
4322         ret = btrfs_check_data_free_space(inode, num_bytes);
4323         if (ret)
4324                 return ret;
4325
4326         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4327         if (ret) {
4328                 btrfs_free_reserved_data_space(inode, num_bytes);
4329                 return ret;
4330         }
4331
4332         return 0;
4333 }
4334
4335 /**
4336  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4337  * @inode: inode we're releasing space for
4338  * @num_bytes: the number of bytes we want to free up
4339  *
4340  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4341  * called in the case that we don't need the metadata AND data reservations
4342  * anymore.  So if there is an error or we insert an inline extent.
4343  *
4344  * This function will release the metadata space that was not used and will
4345  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4346  * list if there are no delalloc bytes left.
4347  */
4348 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4349 {
4350         btrfs_delalloc_release_metadata(inode, num_bytes);
4351         btrfs_free_reserved_data_space(inode, num_bytes);
4352 }
4353
4354 static int update_block_group(struct btrfs_trans_handle *trans,
4355                               struct btrfs_root *root,
4356                               u64 bytenr, u64 num_bytes, int alloc)
4357 {
4358         struct btrfs_block_group_cache *cache = NULL;
4359         struct btrfs_fs_info *info = root->fs_info;
4360         u64 total = num_bytes;
4361         u64 old_val;
4362         u64 byte_in_group;
4363         int factor;
4364
4365         /* block accounting for super block */
4366         spin_lock(&info->delalloc_lock);
4367         old_val = btrfs_super_bytes_used(info->super_copy);
4368         if (alloc)
4369                 old_val += num_bytes;
4370         else
4371                 old_val -= num_bytes;
4372         btrfs_set_super_bytes_used(info->super_copy, old_val);
4373         spin_unlock(&info->delalloc_lock);
4374
4375         while (total) {
4376                 cache = btrfs_lookup_block_group(info, bytenr);
4377                 if (!cache)
4378                         return -1;
4379                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4380                                     BTRFS_BLOCK_GROUP_RAID1 |
4381                                     BTRFS_BLOCK_GROUP_RAID10))
4382                         factor = 2;
4383                 else
4384                         factor = 1;
4385                 /*
4386                  * If this block group has free space cache written out, we
4387                  * need to make sure to load it if we are removing space.  This
4388                  * is because we need the unpinning stage to actually add the
4389                  * space back to the block group, otherwise we will leak space.
4390                  */
4391                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4392                         cache_block_group(cache, trans, NULL, 1);
4393
4394                 byte_in_group = bytenr - cache->key.objectid;
4395                 WARN_ON(byte_in_group > cache->key.offset);
4396
4397                 spin_lock(&cache->space_info->lock);
4398                 spin_lock(&cache->lock);
4399
4400                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4401                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4402                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4403
4404                 cache->dirty = 1;
4405                 old_val = btrfs_block_group_used(&cache->item);
4406                 num_bytes = min(total, cache->key.offset - byte_in_group);
4407                 if (alloc) {
4408                         old_val += num_bytes;
4409                         btrfs_set_block_group_used(&cache->item, old_val);
4410                         cache->reserved -= num_bytes;
4411                         cache->space_info->bytes_reserved -= num_bytes;
4412                         cache->space_info->bytes_used += num_bytes;
4413                         cache->space_info->disk_used += num_bytes * factor;
4414                         spin_unlock(&cache->lock);
4415                         spin_unlock(&cache->space_info->lock);
4416                 } else {
4417                         old_val -= num_bytes;
4418                         btrfs_set_block_group_used(&cache->item, old_val);
4419                         cache->pinned += num_bytes;
4420                         cache->space_info->bytes_pinned += num_bytes;
4421                         cache->space_info->bytes_used -= num_bytes;
4422                         cache->space_info->disk_used -= num_bytes * factor;
4423                         spin_unlock(&cache->lock);
4424                         spin_unlock(&cache->space_info->lock);
4425
4426                         set_extent_dirty(info->pinned_extents,
4427                                          bytenr, bytenr + num_bytes - 1,
4428                                          GFP_NOFS | __GFP_NOFAIL);
4429                 }
4430                 btrfs_put_block_group(cache);
4431                 total -= num_bytes;
4432                 bytenr += num_bytes;
4433         }
4434         return 0;
4435 }
4436
4437 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4438 {
4439         struct btrfs_block_group_cache *cache;
4440         u64 bytenr;
4441
4442         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4443         if (!cache)
4444                 return 0;
4445
4446         bytenr = cache->key.objectid;
4447         btrfs_put_block_group(cache);
4448
4449         return bytenr;
4450 }
4451
4452 static int pin_down_extent(struct btrfs_root *root,
4453                            struct btrfs_block_group_cache *cache,
4454                            u64 bytenr, u64 num_bytes, int reserved)
4455 {
4456         spin_lock(&cache->space_info->lock);
4457         spin_lock(&cache->lock);
4458         cache->pinned += num_bytes;
4459         cache->space_info->bytes_pinned += num_bytes;
4460         if (reserved) {
4461                 cache->reserved -= num_bytes;
4462                 cache->space_info->bytes_reserved -= num_bytes;
4463         }
4464         spin_unlock(&cache->lock);
4465         spin_unlock(&cache->space_info->lock);
4466
4467         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4468                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4469         return 0;
4470 }
4471
4472 /*
4473  * this function must be called within transaction
4474  */
4475 int btrfs_pin_extent(struct btrfs_root *root,
4476                      u64 bytenr, u64 num_bytes, int reserved)
4477 {
4478         struct btrfs_block_group_cache *cache;
4479
4480         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4481         BUG_ON(!cache);
4482
4483         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4484
4485         btrfs_put_block_group(cache);
4486         return 0;
4487 }
4488
4489 /*
4490  * this function must be called within transaction
4491  */
4492 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4493                                     struct btrfs_root *root,
4494                                     u64 bytenr, u64 num_bytes)
4495 {
4496         struct btrfs_block_group_cache *cache;
4497
4498         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4499         BUG_ON(!cache);
4500
4501         /*
4502          * pull in the free space cache (if any) so that our pin
4503          * removes the free space from the cache.  We have load_only set
4504          * to one because the slow code to read in the free extents does check
4505          * the pinned extents.
4506          */
4507         cache_block_group(cache, trans, root, 1);
4508
4509         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4510
4511         /* remove us from the free space cache (if we're there at all) */
4512         btrfs_remove_free_space(cache, bytenr, num_bytes);
4513         btrfs_put_block_group(cache);
4514         return 0;
4515 }
4516
4517 /**
4518  * btrfs_update_reserved_bytes - update the block_group and space info counters
4519  * @cache:      The cache we are manipulating
4520  * @num_bytes:  The number of bytes in question
4521  * @reserve:    One of the reservation enums
4522  *
4523  * This is called by the allocator when it reserves space, or by somebody who is
4524  * freeing space that was never actually used on disk.  For example if you
4525  * reserve some space for a new leaf in transaction A and before transaction A
4526  * commits you free that leaf, you call this with reserve set to 0 in order to
4527  * clear the reservation.
4528  *
4529  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4530  * ENOSPC accounting.  For data we handle the reservation through clearing the
4531  * delalloc bits in the io_tree.  We have to do this since we could end up
4532  * allocating less disk space for the amount of data we have reserved in the
4533  * case of compression.
4534  *
4535  * If this is a reservation and the block group has become read only we cannot
4536  * make the reservation and return -EAGAIN, otherwise this function always
4537  * succeeds.
4538  */
4539 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4540                                        u64 num_bytes, int reserve)
4541 {
4542         struct btrfs_space_info *space_info = cache->space_info;
4543         int ret = 0;
4544         spin_lock(&space_info->lock);
4545         spin_lock(&cache->lock);
4546         if (reserve != RESERVE_FREE) {
4547                 if (cache->ro) {
4548                         ret = -EAGAIN;
4549                 } else {
4550                         cache->reserved += num_bytes;
4551                         space_info->bytes_reserved += num_bytes;
4552                         if (reserve == RESERVE_ALLOC) {
4553                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4554                                 space_info->bytes_may_use -= num_bytes;
4555                         }
4556                 }
4557         } else {
4558                 if (cache->ro)
4559                         space_info->bytes_readonly += num_bytes;
4560                 cache->reserved -= num_bytes;
4561                 space_info->bytes_reserved -= num_bytes;
4562                 space_info->reservation_progress++;
4563         }
4564         spin_unlock(&cache->lock);
4565         spin_unlock(&space_info->lock);
4566         return ret;
4567 }
4568
4569 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4570                                 struct btrfs_root *root)
4571 {
4572         struct btrfs_fs_info *fs_info = root->fs_info;
4573         struct btrfs_caching_control *next;
4574         struct btrfs_caching_control *caching_ctl;
4575         struct btrfs_block_group_cache *cache;
4576
4577         down_write(&fs_info->extent_commit_sem);
4578
4579         list_for_each_entry_safe(caching_ctl, next,
4580                                  &fs_info->caching_block_groups, list) {
4581                 cache = caching_ctl->block_group;
4582                 if (block_group_cache_done(cache)) {
4583                         cache->last_byte_to_unpin = (u64)-1;
4584                         list_del_init(&caching_ctl->list);
4585                         put_caching_control(caching_ctl);
4586                 } else {
4587                         cache->last_byte_to_unpin = caching_ctl->progress;
4588                 }
4589         }
4590
4591         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4592                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4593         else
4594                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4595
4596         up_write(&fs_info->extent_commit_sem);
4597
4598         update_global_block_rsv(fs_info);
4599         return 0;
4600 }
4601
4602 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4603 {
4604         struct btrfs_fs_info *fs_info = root->fs_info;
4605         struct btrfs_block_group_cache *cache = NULL;
4606         u64 len;
4607
4608         while (start <= end) {
4609                 if (!cache ||
4610                     start >= cache->key.objectid + cache->key.offset) {
4611                         if (cache)
4612                                 btrfs_put_block_group(cache);
4613                         cache = btrfs_lookup_block_group(fs_info, start);
4614                         BUG_ON(!cache);
4615                 }
4616
4617                 len = cache->key.objectid + cache->key.offset - start;
4618                 len = min(len, end + 1 - start);
4619
4620                 if (start < cache->last_byte_to_unpin) {
4621                         len = min(len, cache->last_byte_to_unpin - start);
4622                         btrfs_add_free_space(cache, start, len);
4623                 }
4624
4625                 start += len;
4626
4627                 spin_lock(&cache->space_info->lock);
4628                 spin_lock(&cache->lock);
4629                 cache->pinned -= len;
4630                 cache->space_info->bytes_pinned -= len;
4631                 if (cache->ro)
4632                         cache->space_info->bytes_readonly += len;
4633                 spin_unlock(&cache->lock);
4634                 spin_unlock(&cache->space_info->lock);
4635         }
4636
4637         if (cache)
4638                 btrfs_put_block_group(cache);
4639         return 0;
4640 }
4641
4642 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4643                                struct btrfs_root *root)
4644 {
4645         struct btrfs_fs_info *fs_info = root->fs_info;
4646         struct extent_io_tree *unpin;
4647         u64 start;
4648         u64 end;
4649         int ret;
4650
4651         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4652                 unpin = &fs_info->freed_extents[1];
4653         else
4654                 unpin = &fs_info->freed_extents[0];
4655
4656         while (1) {
4657                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4658                                             EXTENT_DIRTY);
4659                 if (ret)
4660                         break;
4661
4662                 if (btrfs_test_opt(root, DISCARD))
4663                         ret = btrfs_discard_extent(root, start,
4664                                                    end + 1 - start, NULL);
4665
4666                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4667                 unpin_extent_range(root, start, end);
4668                 cond_resched();
4669         }
4670
4671         return 0;
4672 }
4673
4674 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4675                                 struct btrfs_root *root,
4676                                 u64 bytenr, u64 num_bytes, u64 parent,
4677                                 u64 root_objectid, u64 owner_objectid,
4678                                 u64 owner_offset, int refs_to_drop,
4679                                 struct btrfs_delayed_extent_op *extent_op)
4680 {
4681         struct btrfs_key key;
4682         struct btrfs_path *path;
4683         struct btrfs_fs_info *info = root->fs_info;
4684         struct btrfs_root *extent_root = info->extent_root;
4685         struct extent_buffer *leaf;
4686         struct btrfs_extent_item *ei;
4687         struct btrfs_extent_inline_ref *iref;
4688         int ret;
4689         int is_data;
4690         int extent_slot = 0;
4691         int found_extent = 0;
4692         int num_to_del = 1;
4693         u32 item_size;
4694         u64 refs;
4695
4696         path = btrfs_alloc_path();
4697         if (!path)
4698                 return -ENOMEM;
4699
4700         path->reada = 1;
4701         path->leave_spinning = 1;
4702
4703         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4704         BUG_ON(!is_data && refs_to_drop != 1);
4705
4706         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4707                                     bytenr, num_bytes, parent,
4708                                     root_objectid, owner_objectid,
4709                                     owner_offset);
4710         if (ret == 0) {
4711                 extent_slot = path->slots[0];
4712                 while (extent_slot >= 0) {
4713                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4714                                               extent_slot);
4715                         if (key.objectid != bytenr)
4716                                 break;
4717                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4718                             key.offset == num_bytes) {
4719                                 found_extent = 1;
4720                                 break;
4721                         }
4722                         if (path->slots[0] - extent_slot > 5)
4723                                 break;
4724                         extent_slot--;
4725                 }
4726 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4727                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4728                 if (found_extent && item_size < sizeof(*ei))
4729                         found_extent = 0;
4730 #endif
4731                 if (!found_extent) {
4732                         BUG_ON(iref);
4733                         ret = remove_extent_backref(trans, extent_root, path,
4734                                                     NULL, refs_to_drop,
4735                                                     is_data);
4736                         BUG_ON(ret);
4737                         btrfs_release_path(path);
4738                         path->leave_spinning = 1;
4739
4740                         key.objectid = bytenr;
4741                         key.type = BTRFS_EXTENT_ITEM_KEY;
4742                         key.offset = num_bytes;
4743
4744                         ret = btrfs_search_slot(trans, extent_root,
4745                                                 &key, path, -1, 1);
4746                         if (ret) {
4747                                 printk(KERN_ERR "umm, got %d back from search"
4748                                        ", was looking for %llu\n", ret,
4749                                        (unsigned long long)bytenr);
4750                                 if (ret > 0)
4751                                         btrfs_print_leaf(extent_root,
4752                                                          path->nodes[0]);
4753                         }
4754                         BUG_ON(ret);
4755                         extent_slot = path->slots[0];
4756                 }
4757         } else {
4758                 btrfs_print_leaf(extent_root, path->nodes[0]);
4759                 WARN_ON(1);
4760                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4761                        "parent %llu root %llu  owner %llu offset %llu\n",
4762                        (unsigned long long)bytenr,
4763                        (unsigned long long)parent,
4764                        (unsigned long long)root_objectid,
4765                        (unsigned long long)owner_objectid,
4766                        (unsigned long long)owner_offset);
4767         }
4768
4769         leaf = path->nodes[0];
4770         item_size = btrfs_item_size_nr(leaf, extent_slot);
4771 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4772         if (item_size < sizeof(*ei)) {
4773                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4774                 ret = convert_extent_item_v0(trans, extent_root, path,
4775                                              owner_objectid, 0);
4776                 BUG_ON(ret < 0);
4777
4778                 btrfs_release_path(path);
4779                 path->leave_spinning = 1;
4780
4781                 key.objectid = bytenr;
4782                 key.type = BTRFS_EXTENT_ITEM_KEY;
4783                 key.offset = num_bytes;
4784
4785                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4786                                         -1, 1);
4787                 if (ret) {
4788                         printk(KERN_ERR "umm, got %d back from search"
4789                                ", was looking for %llu\n", ret,
4790                                (unsigned long long)bytenr);
4791                         btrfs_print_leaf(extent_root, path->nodes[0]);
4792                 }
4793                 BUG_ON(ret);
4794                 extent_slot = path->slots[0];
4795                 leaf = path->nodes[0];
4796                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4797         }
4798 #endif
4799         BUG_ON(item_size < sizeof(*ei));
4800         ei = btrfs_item_ptr(leaf, extent_slot,
4801                             struct btrfs_extent_item);
4802         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4803                 struct btrfs_tree_block_info *bi;
4804                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4805                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4806                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4807         }
4808
4809         refs = btrfs_extent_refs(leaf, ei);
4810         BUG_ON(refs < refs_to_drop);
4811         refs -= refs_to_drop;
4812
4813         if (refs > 0) {
4814                 if (extent_op)
4815                         __run_delayed_extent_op(extent_op, leaf, ei);
4816                 /*
4817                  * In the case of inline back ref, reference count will
4818                  * be updated by remove_extent_backref
4819                  */
4820                 if (iref) {
4821                         BUG_ON(!found_extent);
4822                 } else {
4823                         btrfs_set_extent_refs(leaf, ei, refs);
4824                         btrfs_mark_buffer_dirty(leaf);
4825                 }
4826                 if (found_extent) {
4827                         ret = remove_extent_backref(trans, extent_root, path,
4828                                                     iref, refs_to_drop,
4829                                                     is_data);
4830                         BUG_ON(ret);
4831                 }
4832         } else {
4833                 if (found_extent) {
4834                         BUG_ON(is_data && refs_to_drop !=
4835                                extent_data_ref_count(root, path, iref));
4836                         if (iref) {
4837                                 BUG_ON(path->slots[0] != extent_slot);
4838                         } else {
4839                                 BUG_ON(path->slots[0] != extent_slot + 1);
4840                                 path->slots[0] = extent_slot;
4841                                 num_to_del = 2;
4842                         }
4843                 }
4844
4845                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4846                                       num_to_del);
4847                 BUG_ON(ret);
4848                 btrfs_release_path(path);
4849
4850                 if (is_data) {
4851                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4852                         BUG_ON(ret);
4853                 } else {
4854                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4855                              bytenr >> PAGE_CACHE_SHIFT,
4856                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4857                 }
4858
4859                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4860                 BUG_ON(ret);
4861         }
4862         btrfs_free_path(path);
4863         return ret;
4864 }
4865
4866 /*
4867  * when we free an block, it is possible (and likely) that we free the last
4868  * delayed ref for that extent as well.  This searches the delayed ref tree for
4869  * a given extent, and if there are no other delayed refs to be processed, it
4870  * removes it from the tree.
4871  */
4872 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4873                                       struct btrfs_root *root, u64 bytenr)
4874 {
4875         struct btrfs_delayed_ref_head *head;
4876         struct btrfs_delayed_ref_root *delayed_refs;
4877         struct btrfs_delayed_ref_node *ref;
4878         struct rb_node *node;
4879         int ret = 0;
4880
4881         delayed_refs = &trans->transaction->delayed_refs;
4882         spin_lock(&delayed_refs->lock);
4883         head = btrfs_find_delayed_ref_head(trans, bytenr);
4884         if (!head)
4885                 goto out;
4886
4887         node = rb_prev(&head->node.rb_node);
4888         if (!node)
4889                 goto out;
4890
4891         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4892
4893         /* there are still entries for this ref, we can't drop it */
4894         if (ref->bytenr == bytenr)
4895                 goto out;
4896
4897         if (head->extent_op) {
4898                 if (!head->must_insert_reserved)
4899                         goto out;
4900                 kfree(head->extent_op);
4901                 head->extent_op = NULL;
4902         }
4903
4904         /*
4905          * waiting for the lock here would deadlock.  If someone else has it
4906          * locked they are already in the process of dropping it anyway
4907          */
4908         if (!mutex_trylock(&head->mutex))
4909                 goto out;
4910
4911         /*
4912          * at this point we have a head with no other entries.  Go
4913          * ahead and process it.
4914          */
4915         head->node.in_tree = 0;
4916         rb_erase(&head->node.rb_node, &delayed_refs->root);
4917
4918         delayed_refs->num_entries--;
4919
4920         /*
4921          * we don't take a ref on the node because we're removing it from the
4922          * tree, so we just steal the ref the tree was holding.
4923          */
4924         delayed_refs->num_heads--;
4925         if (list_empty(&head->cluster))
4926                 delayed_refs->num_heads_ready--;
4927
4928         list_del_init(&head->cluster);
4929         spin_unlock(&delayed_refs->lock);
4930
4931         BUG_ON(head->extent_op);
4932         if (head->must_insert_reserved)
4933                 ret = 1;
4934
4935         mutex_unlock(&head->mutex);
4936         btrfs_put_delayed_ref(&head->node);
4937         return ret;
4938 out:
4939         spin_unlock(&delayed_refs->lock);
4940         return 0;
4941 }
4942
4943 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4944                            struct btrfs_root *root,
4945                            struct extent_buffer *buf,
4946                            u64 parent, int last_ref, int for_cow)
4947 {
4948         struct btrfs_block_group_cache *cache = NULL;
4949         int ret;
4950
4951         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4952                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
4953                                         buf->start, buf->len,
4954                                         parent, root->root_key.objectid,
4955                                         btrfs_header_level(buf),
4956                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
4957                 BUG_ON(ret);
4958         }
4959
4960         if (!last_ref)
4961                 return;
4962
4963         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4964
4965         if (btrfs_header_generation(buf) == trans->transid) {
4966                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4967                         ret = check_ref_cleanup(trans, root, buf->start);
4968                         if (!ret)
4969                                 goto out;
4970                 }
4971
4972                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4973                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4974                         goto out;
4975                 }
4976
4977                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4978
4979                 btrfs_add_free_space(cache, buf->start, buf->len);
4980                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4981         }
4982 out:
4983         /*
4984          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4985          * anymore.
4986          */
4987         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4988         btrfs_put_block_group(cache);
4989 }
4990
4991 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
4992                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
4993                       u64 owner, u64 offset, int for_cow)
4994 {
4995         int ret;
4996         struct btrfs_fs_info *fs_info = root->fs_info;
4997
4998         /*
4999          * tree log blocks never actually go into the extent allocation
5000          * tree, just update pinning info and exit early.
5001          */
5002         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5003                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5004                 /* unlocks the pinned mutex */
5005                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5006                 ret = 0;
5007         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5008                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5009                                         num_bytes,
5010                                         parent, root_objectid, (int)owner,
5011                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5012                 BUG_ON(ret);
5013         } else {
5014                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5015                                                 num_bytes,
5016                                                 parent, root_objectid, owner,
5017                                                 offset, BTRFS_DROP_DELAYED_REF,
5018                                                 NULL, for_cow);
5019                 BUG_ON(ret);
5020         }
5021         return ret;
5022 }
5023
5024 static u64 stripe_align(struct btrfs_root *root, u64 val)
5025 {
5026         u64 mask = ((u64)root->stripesize - 1);
5027         u64 ret = (val + mask) & ~mask;
5028         return ret;
5029 }
5030
5031 /*
5032  * when we wait for progress in the block group caching, its because
5033  * our allocation attempt failed at least once.  So, we must sleep
5034  * and let some progress happen before we try again.
5035  *
5036  * This function will sleep at least once waiting for new free space to
5037  * show up, and then it will check the block group free space numbers
5038  * for our min num_bytes.  Another option is to have it go ahead
5039  * and look in the rbtree for a free extent of a given size, but this
5040  * is a good start.
5041  */
5042 static noinline int
5043 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5044                                 u64 num_bytes)
5045 {
5046         struct btrfs_caching_control *caching_ctl;
5047         DEFINE_WAIT(wait);
5048
5049         caching_ctl = get_caching_control(cache);
5050         if (!caching_ctl)
5051                 return 0;
5052
5053         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5054                    (cache->free_space_ctl->free_space >= num_bytes));
5055
5056         put_caching_control(caching_ctl);
5057         return 0;
5058 }
5059
5060 static noinline int
5061 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5062 {
5063         struct btrfs_caching_control *caching_ctl;
5064         DEFINE_WAIT(wait);
5065
5066         caching_ctl = get_caching_control(cache);
5067         if (!caching_ctl)
5068                 return 0;
5069
5070         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5071
5072         put_caching_control(caching_ctl);
5073         return 0;
5074 }
5075
5076 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5077 {
5078         int index;
5079         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5080                 index = 0;
5081         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5082                 index = 1;
5083         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5084                 index = 2;
5085         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5086                 index = 3;
5087         else
5088                 index = 4;
5089         return index;
5090 }
5091
5092 enum btrfs_loop_type {
5093         LOOP_FIND_IDEAL = 0,
5094         LOOP_CACHING_NOWAIT = 1,
5095         LOOP_CACHING_WAIT = 2,
5096         LOOP_ALLOC_CHUNK = 3,
5097         LOOP_NO_EMPTY_SIZE = 4,
5098 };
5099
5100 /*
5101  * walks the btree of allocated extents and find a hole of a given size.
5102  * The key ins is changed to record the hole:
5103  * ins->objectid == block start
5104  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5105  * ins->offset == number of blocks
5106  * Any available blocks before search_start are skipped.
5107  */
5108 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5109                                      struct btrfs_root *orig_root,
5110                                      u64 num_bytes, u64 empty_size,
5111                                      u64 search_start, u64 search_end,
5112                                      u64 hint_byte, struct btrfs_key *ins,
5113                                      u64 data)
5114 {
5115         int ret = 0;
5116         struct btrfs_root *root = orig_root->fs_info->extent_root;
5117         struct btrfs_free_cluster *last_ptr = NULL;
5118         struct btrfs_block_group_cache *block_group = NULL;
5119         int empty_cluster = 2 * 1024 * 1024;
5120         int allowed_chunk_alloc = 0;
5121         int done_chunk_alloc = 0;
5122         struct btrfs_space_info *space_info;
5123         int last_ptr_loop = 0;
5124         int loop = 0;
5125         int index = 0;
5126         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5127                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5128         bool found_uncached_bg = false;
5129         bool failed_cluster_refill = false;
5130         bool failed_alloc = false;
5131         bool use_cluster = true;
5132         bool have_caching_bg = false;
5133         u64 ideal_cache_percent = 0;
5134         u64 ideal_cache_offset = 0;
5135
5136         WARN_ON(num_bytes < root->sectorsize);
5137         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5138         ins->objectid = 0;
5139         ins->offset = 0;
5140
5141         space_info = __find_space_info(root->fs_info, data);
5142         if (!space_info) {
5143                 printk(KERN_ERR "No space info for %llu\n", data);
5144                 return -ENOSPC;
5145         }
5146
5147         /*
5148          * If the space info is for both data and metadata it means we have a
5149          * small filesystem and we can't use the clustering stuff.
5150          */
5151         if (btrfs_mixed_space_info(space_info))
5152                 use_cluster = false;
5153
5154         if (orig_root->ref_cows || empty_size)
5155                 allowed_chunk_alloc = 1;
5156
5157         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5158                 last_ptr = &root->fs_info->meta_alloc_cluster;
5159                 if (!btrfs_test_opt(root, SSD))
5160                         empty_cluster = 64 * 1024;
5161         }
5162
5163         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5164             btrfs_test_opt(root, SSD)) {
5165                 last_ptr = &root->fs_info->data_alloc_cluster;
5166         }
5167
5168         if (last_ptr) {
5169                 spin_lock(&last_ptr->lock);
5170                 if (last_ptr->block_group)
5171                         hint_byte = last_ptr->window_start;
5172                 spin_unlock(&last_ptr->lock);
5173         }
5174
5175         search_start = max(search_start, first_logical_byte(root, 0));
5176         search_start = max(search_start, hint_byte);
5177
5178         if (!last_ptr)
5179                 empty_cluster = 0;
5180
5181         if (search_start == hint_byte) {
5182 ideal_cache:
5183                 block_group = btrfs_lookup_block_group(root->fs_info,
5184                                                        search_start);
5185                 /*
5186                  * we don't want to use the block group if it doesn't match our
5187                  * allocation bits, or if its not cached.
5188                  *
5189                  * However if we are re-searching with an ideal block group
5190                  * picked out then we don't care that the block group is cached.
5191                  */
5192                 if (block_group && block_group_bits(block_group, data) &&
5193                     (block_group->cached != BTRFS_CACHE_NO ||
5194                      search_start == ideal_cache_offset)) {
5195                         down_read(&space_info->groups_sem);
5196                         if (list_empty(&block_group->list) ||
5197                             block_group->ro) {
5198                                 /*
5199                                  * someone is removing this block group,
5200                                  * we can't jump into the have_block_group
5201                                  * target because our list pointers are not
5202                                  * valid
5203                                  */
5204                                 btrfs_put_block_group(block_group);
5205                                 up_read(&space_info->groups_sem);
5206                         } else {
5207                                 index = get_block_group_index(block_group);
5208                                 goto have_block_group;
5209                         }
5210                 } else if (block_group) {
5211                         btrfs_put_block_group(block_group);
5212                 }
5213         }
5214 search:
5215         have_caching_bg = false;
5216         down_read(&space_info->groups_sem);
5217         list_for_each_entry(block_group, &space_info->block_groups[index],
5218                             list) {
5219                 u64 offset;
5220                 int cached;
5221
5222                 btrfs_get_block_group(block_group);
5223                 search_start = block_group->key.objectid;
5224
5225                 /*
5226                  * this can happen if we end up cycling through all the
5227                  * raid types, but we want to make sure we only allocate
5228                  * for the proper type.
5229                  */
5230                 if (!block_group_bits(block_group, data)) {
5231                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5232                                 BTRFS_BLOCK_GROUP_RAID1 |
5233                                 BTRFS_BLOCK_GROUP_RAID10;
5234
5235                         /*
5236                          * if they asked for extra copies and this block group
5237                          * doesn't provide them, bail.  This does allow us to
5238                          * fill raid0 from raid1.
5239                          */
5240                         if ((data & extra) && !(block_group->flags & extra))
5241                                 goto loop;
5242                 }
5243
5244 have_block_group:
5245                 cached = block_group_cache_done(block_group);
5246                 if (unlikely(!cached)) {
5247                         u64 free_percent;
5248
5249                         found_uncached_bg = true;
5250                         ret = cache_block_group(block_group, trans,
5251                                                 orig_root, 1);
5252                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5253                                 goto alloc;
5254
5255                         free_percent = btrfs_block_group_used(&block_group->item);
5256                         free_percent *= 100;
5257                         free_percent = div64_u64(free_percent,
5258                                                  block_group->key.offset);
5259                         free_percent = 100 - free_percent;
5260                         if (free_percent > ideal_cache_percent &&
5261                             likely(!block_group->ro)) {
5262                                 ideal_cache_offset = block_group->key.objectid;
5263                                 ideal_cache_percent = free_percent;
5264                         }
5265
5266                         /*
5267                          * The caching workers are limited to 2 threads, so we
5268                          * can queue as much work as we care to.
5269                          */
5270                         if (loop > LOOP_FIND_IDEAL) {
5271                                 ret = cache_block_group(block_group, trans,
5272                                                         orig_root, 0);
5273                                 BUG_ON(ret);
5274                         }
5275
5276                         /*
5277                          * If loop is set for cached only, try the next block
5278                          * group.
5279                          */
5280                         if (loop == LOOP_FIND_IDEAL)
5281                                 goto loop;
5282                 }
5283
5284 alloc:
5285                 if (unlikely(block_group->ro))
5286                         goto loop;
5287
5288                 spin_lock(&block_group->free_space_ctl->tree_lock);
5289                 if (cached &&
5290                     block_group->free_space_ctl->free_space <
5291                     num_bytes + empty_cluster + empty_size) {
5292                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5293                         goto loop;
5294                 }
5295                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5296
5297                 /*
5298                  * Ok we want to try and use the cluster allocator, so lets look
5299                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5300                  * have tried the cluster allocator plenty of times at this
5301                  * point and not have found anything, so we are likely way too
5302                  * fragmented for the clustering stuff to find anything, so lets
5303                  * just skip it and let the allocator find whatever block it can
5304                  * find
5305                  */
5306                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5307                         /*
5308                          * the refill lock keeps out other
5309                          * people trying to start a new cluster
5310                          */
5311                         spin_lock(&last_ptr->refill_lock);
5312                         if (!last_ptr->block_group ||
5313                             last_ptr->block_group->ro ||
5314                             !block_group_bits(last_ptr->block_group, data))
5315                                 goto refill_cluster;
5316
5317                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5318                                                  num_bytes, search_start);
5319                         if (offset) {
5320                                 /* we have a block, we're done */
5321                                 spin_unlock(&last_ptr->refill_lock);
5322                                 goto checks;
5323                         }
5324
5325                         spin_lock(&last_ptr->lock);
5326                         /*
5327                          * whoops, this cluster doesn't actually point to
5328                          * this block group.  Get a ref on the block
5329                          * group is does point to and try again
5330                          */
5331                         if (!last_ptr_loop && last_ptr->block_group &&
5332                             last_ptr->block_group != block_group &&
5333                             index <=
5334                                  get_block_group_index(last_ptr->block_group)) {
5335
5336                                 btrfs_put_block_group(block_group);
5337                                 block_group = last_ptr->block_group;
5338                                 btrfs_get_block_group(block_group);
5339                                 spin_unlock(&last_ptr->lock);
5340                                 spin_unlock(&last_ptr->refill_lock);
5341
5342                                 last_ptr_loop = 1;
5343                                 search_start = block_group->key.objectid;
5344                                 /*
5345                                  * we know this block group is properly
5346                                  * in the list because
5347                                  * btrfs_remove_block_group, drops the
5348                                  * cluster before it removes the block
5349                                  * group from the list
5350                                  */
5351                                 goto have_block_group;
5352                         }
5353                         spin_unlock(&last_ptr->lock);
5354 refill_cluster:
5355                         /*
5356                          * this cluster didn't work out, free it and
5357                          * start over
5358                          */
5359                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5360
5361                         last_ptr_loop = 0;
5362
5363                         /* allocate a cluster in this block group */
5364                         ret = btrfs_find_space_cluster(trans, root,
5365                                                block_group, last_ptr,
5366                                                search_start, num_bytes,
5367                                                empty_cluster + empty_size);
5368                         if (ret == 0) {
5369                                 /*
5370                                  * now pull our allocation out of this
5371                                  * cluster
5372                                  */
5373                                 offset = btrfs_alloc_from_cluster(block_group,
5374                                                   last_ptr, num_bytes,
5375                                                   search_start);
5376                                 if (offset) {
5377                                         /* we found one, proceed */
5378                                         spin_unlock(&last_ptr->refill_lock);
5379                                         goto checks;
5380                                 }
5381                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5382                                    && !failed_cluster_refill) {
5383                                 spin_unlock(&last_ptr->refill_lock);
5384
5385                                 failed_cluster_refill = true;
5386                                 wait_block_group_cache_progress(block_group,
5387                                        num_bytes + empty_cluster + empty_size);
5388                                 goto have_block_group;
5389                         }
5390
5391                         /*
5392                          * at this point we either didn't find a cluster
5393                          * or we weren't able to allocate a block from our
5394                          * cluster.  Free the cluster we've been trying
5395                          * to use, and go to the next block group
5396                          */
5397                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5398                         spin_unlock(&last_ptr->refill_lock);
5399                         goto loop;
5400                 }
5401
5402                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5403                                                     num_bytes, empty_size);
5404                 /*
5405                  * If we didn't find a chunk, and we haven't failed on this
5406                  * block group before, and this block group is in the middle of
5407                  * caching and we are ok with waiting, then go ahead and wait
5408                  * for progress to be made, and set failed_alloc to true.
5409                  *
5410                  * If failed_alloc is true then we've already waited on this
5411                  * block group once and should move on to the next block group.
5412                  */
5413                 if (!offset && !failed_alloc && !cached &&
5414                     loop > LOOP_CACHING_NOWAIT) {
5415                         wait_block_group_cache_progress(block_group,
5416                                                 num_bytes + empty_size);
5417                         failed_alloc = true;
5418                         goto have_block_group;
5419                 } else if (!offset) {
5420                         if (!cached)
5421                                 have_caching_bg = true;
5422                         goto loop;
5423                 }
5424 checks:
5425                 search_start = stripe_align(root, offset);
5426                 /* move on to the next group */
5427                 if (search_start + num_bytes >= search_end) {
5428                         btrfs_add_free_space(block_group, offset, num_bytes);
5429                         goto loop;
5430                 }
5431
5432                 /* move on to the next group */
5433                 if (search_start + num_bytes >
5434                     block_group->key.objectid + block_group->key.offset) {
5435                         btrfs_add_free_space(block_group, offset, num_bytes);
5436                         goto loop;
5437                 }
5438
5439                 ins->objectid = search_start;
5440                 ins->offset = num_bytes;
5441
5442                 if (offset < search_start)
5443                         btrfs_add_free_space(block_group, offset,
5444                                              search_start - offset);
5445                 BUG_ON(offset > search_start);
5446
5447                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5448                                                   alloc_type);
5449                 if (ret == -EAGAIN) {
5450                         btrfs_add_free_space(block_group, offset, num_bytes);
5451                         goto loop;
5452                 }
5453
5454                 /* we are all good, lets return */
5455                 ins->objectid = search_start;
5456                 ins->offset = num_bytes;
5457
5458                 if (offset < search_start)
5459                         btrfs_add_free_space(block_group, offset,
5460                                              search_start - offset);
5461                 BUG_ON(offset > search_start);
5462                 btrfs_put_block_group(block_group);
5463                 break;
5464 loop:
5465                 failed_cluster_refill = false;
5466                 failed_alloc = false;
5467                 BUG_ON(index != get_block_group_index(block_group));
5468                 btrfs_put_block_group(block_group);
5469         }
5470         up_read(&space_info->groups_sem);
5471
5472         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5473                 goto search;
5474
5475         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5476                 goto search;
5477
5478         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5479          *                      for them to make caching progress.  Also
5480          *                      determine the best possible bg to cache
5481          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5482          *                      caching kthreads as we move along
5483          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5484          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5485          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5486          *                      again
5487          */
5488         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5489                 index = 0;
5490                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5491                         found_uncached_bg = false;
5492                         loop++;
5493                         if (!ideal_cache_percent)
5494                                 goto search;
5495
5496                         /*
5497                          * 1 of the following 2 things have happened so far
5498                          *
5499                          * 1) We found an ideal block group for caching that
5500                          * is mostly full and will cache quickly, so we might
5501                          * as well wait for it.
5502                          *
5503                          * 2) We searched for cached only and we didn't find
5504                          * anything, and we didn't start any caching kthreads
5505                          * either, so chances are we will loop through and
5506                          * start a couple caching kthreads, and then come back
5507                          * around and just wait for them.  This will be slower
5508                          * because we will have 2 caching kthreads reading at
5509                          * the same time when we could have just started one
5510                          * and waited for it to get far enough to give us an
5511                          * allocation, so go ahead and go to the wait caching
5512                          * loop.
5513                          */
5514                         loop = LOOP_CACHING_WAIT;
5515                         search_start = ideal_cache_offset;
5516                         ideal_cache_percent = 0;
5517                         goto ideal_cache;
5518                 } else if (loop == LOOP_FIND_IDEAL) {
5519                         /*
5520                          * Didn't find a uncached bg, wait on anything we find
5521                          * next.
5522                          */
5523                         loop = LOOP_CACHING_WAIT;
5524                         goto search;
5525                 }
5526
5527                 loop++;
5528
5529                 if (loop == LOOP_ALLOC_CHUNK) {
5530                        if (allowed_chunk_alloc) {
5531                                 ret = do_chunk_alloc(trans, root, num_bytes +
5532                                                      2 * 1024 * 1024, data,
5533                                                      CHUNK_ALLOC_LIMITED);
5534                                 allowed_chunk_alloc = 0;
5535                                 if (ret == 1)
5536                                         done_chunk_alloc = 1;
5537                         } else if (!done_chunk_alloc &&
5538                                    space_info->force_alloc ==
5539                                    CHUNK_ALLOC_NO_FORCE) {
5540                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5541                         }
5542
5543                        /*
5544                         * We didn't allocate a chunk, go ahead and drop the
5545                         * empty size and loop again.
5546                         */
5547                        if (!done_chunk_alloc)
5548                                loop = LOOP_NO_EMPTY_SIZE;
5549                 }
5550
5551                 if (loop == LOOP_NO_EMPTY_SIZE) {
5552                         empty_size = 0;
5553                         empty_cluster = 0;
5554                 }
5555
5556                 goto search;
5557         } else if (!ins->objectid) {
5558                 ret = -ENOSPC;
5559         } else if (ins->objectid) {
5560                 ret = 0;
5561         }
5562
5563         return ret;
5564 }
5565
5566 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5567                             int dump_block_groups)
5568 {
5569         struct btrfs_block_group_cache *cache;
5570         int index = 0;
5571
5572         spin_lock(&info->lock);
5573         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5574                (unsigned long long)info->flags,
5575                (unsigned long long)(info->total_bytes - info->bytes_used -
5576                                     info->bytes_pinned - info->bytes_reserved -
5577                                     info->bytes_readonly),
5578                (info->full) ? "" : "not ");
5579         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5580                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5581                (unsigned long long)info->total_bytes,
5582                (unsigned long long)info->bytes_used,
5583                (unsigned long long)info->bytes_pinned,
5584                (unsigned long long)info->bytes_reserved,
5585                (unsigned long long)info->bytes_may_use,
5586                (unsigned long long)info->bytes_readonly);
5587         spin_unlock(&info->lock);
5588
5589         if (!dump_block_groups)
5590                 return;
5591
5592         down_read(&info->groups_sem);
5593 again:
5594         list_for_each_entry(cache, &info->block_groups[index], list) {
5595                 spin_lock(&cache->lock);
5596                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5597                        "%llu pinned %llu reserved\n",
5598                        (unsigned long long)cache->key.objectid,
5599                        (unsigned long long)cache->key.offset,
5600                        (unsigned long long)btrfs_block_group_used(&cache->item),
5601                        (unsigned long long)cache->pinned,
5602                        (unsigned long long)cache->reserved);
5603  &nbs