Btrfs: add BTRFS_AVAIL_ALLOC_BIT_SINGLE bit
[linux-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <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_TYPE_MASK;
622
623         rcu_read_lock();
624         list_for_each_entry_rcu(found, head, list) {
625                 if (found->flags & flags) {
626                         rcu_read_unlock();
627                         return found;
628                 }
629         }
630         rcu_read_unlock();
631         return NULL;
632 }
633
634 /*
635  * after adding space to the filesystem, we need to clear the full flags
636  * on all the space infos.
637  */
638 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
639 {
640         struct list_head *head = &info->space_info;
641         struct btrfs_space_info *found;
642
643         rcu_read_lock();
644         list_for_each_entry_rcu(found, head, list)
645                 found->full = 0;
646         rcu_read_unlock();
647 }
648
649 static u64 div_factor(u64 num, int factor)
650 {
651         if (factor == 10)
652                 return num;
653         num *= factor;
654         do_div(num, 10);
655         return num;
656 }
657
658 static u64 div_factor_fine(u64 num, int factor)
659 {
660         if (factor == 100)
661                 return num;
662         num *= factor;
663         do_div(num, 100);
664         return num;
665 }
666
667 u64 btrfs_find_block_group(struct btrfs_root *root,
668                            u64 search_start, u64 search_hint, int owner)
669 {
670         struct btrfs_block_group_cache *cache;
671         u64 used;
672         u64 last = max(search_hint, search_start);
673         u64 group_start = 0;
674         int full_search = 0;
675         int factor = 9;
676         int wrapped = 0;
677 again:
678         while (1) {
679                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
680                 if (!cache)
681                         break;
682
683                 spin_lock(&cache->lock);
684                 last = cache->key.objectid + cache->key.offset;
685                 used = btrfs_block_group_used(&cache->item);
686
687                 if ((full_search || !cache->ro) &&
688                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
689                         if (used + cache->pinned + cache->reserved <
690                             div_factor(cache->key.offset, factor)) {
691                                 group_start = cache->key.objectid;
692                                 spin_unlock(&cache->lock);
693                                 btrfs_put_block_group(cache);
694                                 goto found;
695                         }
696                 }
697                 spin_unlock(&cache->lock);
698                 btrfs_put_block_group(cache);
699                 cond_resched();
700         }
701         if (!wrapped) {
702                 last = search_start;
703                 wrapped = 1;
704                 goto again;
705         }
706         if (!full_search && factor < 10) {
707                 last = search_start;
708                 full_search = 1;
709                 factor = 10;
710                 goto again;
711         }
712 found:
713         return group_start;
714 }
715
716 /* simple helper to search for an existing extent at a given offset */
717 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
718 {
719         int ret;
720         struct btrfs_key key;
721         struct btrfs_path *path;
722
723         path = btrfs_alloc_path();
724         if (!path)
725                 return -ENOMEM;
726
727         key.objectid = start;
728         key.offset = len;
729         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
730         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
731                                 0, 0);
732         btrfs_free_path(path);
733         return ret;
734 }
735
736 /*
737  * helper function to lookup reference count and flags of extent.
738  *
739  * the head node for delayed ref is used to store the sum of all the
740  * reference count modifications queued up in the rbtree. the head
741  * node may also store the extent flags to set. This way you can check
742  * to see what the reference count and extent flags would be if all of
743  * the delayed refs are not processed.
744  */
745 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
746                              struct btrfs_root *root, u64 bytenr,
747                              u64 num_bytes, u64 *refs, u64 *flags)
748 {
749         struct btrfs_delayed_ref_head *head;
750         struct btrfs_delayed_ref_root *delayed_refs;
751         struct btrfs_path *path;
752         struct btrfs_extent_item *ei;
753         struct extent_buffer *leaf;
754         struct btrfs_key key;
755         u32 item_size;
756         u64 num_refs;
757         u64 extent_flags;
758         int ret;
759
760         path = btrfs_alloc_path();
761         if (!path)
762                 return -ENOMEM;
763
764         key.objectid = bytenr;
765         key.type = BTRFS_EXTENT_ITEM_KEY;
766         key.offset = num_bytes;
767         if (!trans) {
768                 path->skip_locking = 1;
769                 path->search_commit_root = 1;
770         }
771 again:
772         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
773                                 &key, path, 0, 0);
774         if (ret < 0)
775                 goto out_free;
776
777         if (ret == 0) {
778                 leaf = path->nodes[0];
779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
780                 if (item_size >= sizeof(*ei)) {
781                         ei = btrfs_item_ptr(leaf, path->slots[0],
782                                             struct btrfs_extent_item);
783                         num_refs = btrfs_extent_refs(leaf, ei);
784                         extent_flags = btrfs_extent_flags(leaf, ei);
785                 } else {
786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
787                         struct btrfs_extent_item_v0 *ei0;
788                         BUG_ON(item_size != sizeof(*ei0));
789                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
790                                              struct btrfs_extent_item_v0);
791                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
792                         /* FIXME: this isn't correct for data */
793                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
794 #else
795                         BUG();
796 #endif
797                 }
798                 BUG_ON(num_refs == 0);
799         } else {
800                 num_refs = 0;
801                 extent_flags = 0;
802                 ret = 0;
803         }
804
805         if (!trans)
806                 goto out;
807
808         delayed_refs = &trans->transaction->delayed_refs;
809         spin_lock(&delayed_refs->lock);
810         head = btrfs_find_delayed_ref_head(trans, bytenr);
811         if (head) {
812                 if (!mutex_trylock(&head->mutex)) {
813                         atomic_inc(&head->node.refs);
814                         spin_unlock(&delayed_refs->lock);
815
816                         btrfs_release_path(path);
817
818                         /*
819                          * Mutex was contended, block until it's released and try
820                          * again
821                          */
822                         mutex_lock(&head->mutex);
823                         mutex_unlock(&head->mutex);
824                         btrfs_put_delayed_ref(&head->node);
825                         goto again;
826                 }
827                 if (head->extent_op && head->extent_op->update_flags)
828                         extent_flags |= head->extent_op->flags_to_set;
829                 else
830                         BUG_ON(num_refs == 0);
831
832                 num_refs += head->node.ref_mod;
833                 mutex_unlock(&head->mutex);
834         }
835         spin_unlock(&delayed_refs->lock);
836 out:
837         WARN_ON(num_refs == 0);
838         if (refs)
839                 *refs = num_refs;
840         if (flags)
841                 *flags = extent_flags;
842 out_free:
843         btrfs_free_path(path);
844         return ret;
845 }
846
847 /*
848  * Back reference rules.  Back refs have three main goals:
849  *
850  * 1) differentiate between all holders of references to an extent so that
851  *    when a reference is dropped we can make sure it was a valid reference
852  *    before freeing the extent.
853  *
854  * 2) Provide enough information to quickly find the holders of an extent
855  *    if we notice a given block is corrupted or bad.
856  *
857  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
858  *    maintenance.  This is actually the same as #2, but with a slightly
859  *    different use case.
860  *
861  * There are two kinds of back refs. The implicit back refs is optimized
862  * for pointers in non-shared tree blocks. For a given pointer in a block,
863  * back refs of this kind provide information about the block's owner tree
864  * and the pointer's key. These information allow us to find the block by
865  * b-tree searching. The full back refs is for pointers in tree blocks not
866  * referenced by their owner trees. The location of tree block is recorded
867  * in the back refs. Actually the full back refs is generic, and can be
868  * used in all cases the implicit back refs is used. The major shortcoming
869  * of the full back refs is its overhead. Every time a tree block gets
870  * COWed, we have to update back refs entry for all pointers in it.
871  *
872  * For a newly allocated tree block, we use implicit back refs for
873  * pointers in it. This means most tree related operations only involve
874  * implicit back refs. For a tree block created in old transaction, the
875  * only way to drop a reference to it is COW it. So we can detect the
876  * event that tree block loses its owner tree's reference and do the
877  * back refs conversion.
878  *
879  * When a tree block is COW'd through a tree, there are four cases:
880  *
881  * The reference count of the block is one and the tree is the block's
882  * owner tree. Nothing to do in this case.
883  *
884  * The reference count of the block is one and the tree is not the
885  * block's owner tree. In this case, full back refs is used for pointers
886  * in the block. Remove these full back refs, add implicit back refs for
887  * every pointers in the new block.
888  *
889  * The reference count of the block is greater than one and the tree is
890  * the block's owner tree. In this case, implicit back refs is used for
891  * pointers in the block. Add full back refs for every pointers in the
892  * block, increase lower level extents' reference counts. The original
893  * implicit back refs are entailed to the new block.
894  *
895  * The reference count of the block is greater than one and the tree is
896  * not the block's owner tree. Add implicit back refs for every pointer in
897  * the new block, increase lower level extents' reference count.
898  *
899  * Back Reference Key composing:
900  *
901  * The key objectid corresponds to the first byte in the extent,
902  * The key type is used to differentiate between types of back refs.
903  * There are different meanings of the key offset for different types
904  * of back refs.
905  *
906  * File extents can be referenced by:
907  *
908  * - multiple snapshots, subvolumes, or different generations in one subvol
909  * - different files inside a single subvolume
910  * - different offsets inside a file (bookend extents in file.c)
911  *
912  * The extent ref structure for the implicit back refs has fields for:
913  *
914  * - Objectid of the subvolume root
915  * - objectid of the file holding the reference
916  * - original offset in the file
917  * - how many bookend extents
918  *
919  * The key offset for the implicit back refs is hash of the first
920  * three fields.
921  *
922  * The extent ref structure for the full back refs has field for:
923  *
924  * - number of pointers in the tree leaf
925  *
926  * The key offset for the implicit back refs is the first byte of
927  * the tree leaf
928  *
929  * When a file extent is allocated, The implicit back refs is used.
930  * the fields are filled in:
931  *
932  *     (root_key.objectid, inode objectid, offset in file, 1)
933  *
934  * When a file extent is removed file truncation, we find the
935  * corresponding implicit back refs and check the following fields:
936  *
937  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
938  *
939  * Btree extents can be referenced by:
940  *
941  * - Different subvolumes
942  *
943  * Both the implicit back refs and the full back refs for tree blocks
944  * only consist of key. The key offset for the implicit back refs is
945  * objectid of block's owner tree. The key offset for the full back refs
946  * is the first byte of parent block.
947  *
948  * When implicit back refs is used, information about the lowest key and
949  * level of the tree block are required. These information are stored in
950  * tree block info structure.
951  */
952
953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
954 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
955                                   struct btrfs_root *root,
956                                   struct btrfs_path *path,
957                                   u64 owner, u32 extra_size)
958 {
959         struct btrfs_extent_item *item;
960         struct btrfs_extent_item_v0 *ei0;
961         struct btrfs_extent_ref_v0 *ref0;
962         struct btrfs_tree_block_info *bi;
963         struct extent_buffer *leaf;
964         struct btrfs_key key;
965         struct btrfs_key found_key;
966         u32 new_size = sizeof(*item);
967         u64 refs;
968         int ret;
969
970         leaf = path->nodes[0];
971         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
972
973         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
974         ei0 = btrfs_item_ptr(leaf, path->slots[0],
975                              struct btrfs_extent_item_v0);
976         refs = btrfs_extent_refs_v0(leaf, ei0);
977
978         if (owner == (u64)-1) {
979                 while (1) {
980                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
981                                 ret = btrfs_next_leaf(root, path);
982                                 if (ret < 0)
983                                         return ret;
984                                 BUG_ON(ret > 0);
985                                 leaf = path->nodes[0];
986                         }
987                         btrfs_item_key_to_cpu(leaf, &found_key,
988                                               path->slots[0]);
989                         BUG_ON(key.objectid != found_key.objectid);
990                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
991                                 path->slots[0]++;
992                                 continue;
993                         }
994                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
995                                               struct btrfs_extent_ref_v0);
996                         owner = btrfs_ref_objectid_v0(leaf, ref0);
997                         break;
998                 }
999         }
1000         btrfs_release_path(path);
1001
1002         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1003                 new_size += sizeof(*bi);
1004
1005         new_size -= sizeof(*ei0);
1006         ret = btrfs_search_slot(trans, root, &key, path,
1007                                 new_size + extra_size, 1);
1008         if (ret < 0)
1009                 return ret;
1010         BUG_ON(ret);
1011
1012         ret = btrfs_extend_item(trans, root, path, new_size);
1013
1014         leaf = path->nodes[0];
1015         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1016         btrfs_set_extent_refs(leaf, item, refs);
1017         /* FIXME: get real generation */
1018         btrfs_set_extent_generation(leaf, item, 0);
1019         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1020                 btrfs_set_extent_flags(leaf, item,
1021                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1022                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1023                 bi = (struct btrfs_tree_block_info *)(item + 1);
1024                 /* FIXME: get first key of the block */
1025                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1026                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1027         } else {
1028                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1029         }
1030         btrfs_mark_buffer_dirty(leaf);
1031         return 0;
1032 }
1033 #endif
1034
1035 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1036 {
1037         u32 high_crc = ~(u32)0;
1038         u32 low_crc = ~(u32)0;
1039         __le64 lenum;
1040
1041         lenum = cpu_to_le64(root_objectid);
1042         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1043         lenum = cpu_to_le64(owner);
1044         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(offset);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047
1048         return ((u64)high_crc << 31) ^ (u64)low_crc;
1049 }
1050
1051 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1052                                      struct btrfs_extent_data_ref *ref)
1053 {
1054         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1055                                     btrfs_extent_data_ref_objectid(leaf, ref),
1056                                     btrfs_extent_data_ref_offset(leaf, ref));
1057 }
1058
1059 static int match_extent_data_ref(struct extent_buffer *leaf,
1060                                  struct btrfs_extent_data_ref *ref,
1061                                  u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1064             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1065             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1066                 return 0;
1067         return 1;
1068 }
1069
1070 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1071                                            struct btrfs_root *root,
1072                                            struct btrfs_path *path,
1073                                            u64 bytenr, u64 parent,
1074                                            u64 root_objectid,
1075                                            u64 owner, u64 offset)
1076 {
1077         struct btrfs_key key;
1078         struct btrfs_extent_data_ref *ref;
1079         struct extent_buffer *leaf;
1080         u32 nritems;
1081         int ret;
1082         int recow;
1083         int err = -ENOENT;
1084
1085         key.objectid = bytenr;
1086         if (parent) {
1087                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1088                 key.offset = parent;
1089         } else {
1090                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1091                 key.offset = hash_extent_data_ref(root_objectid,
1092                                                   owner, offset);
1093         }
1094 again:
1095         recow = 0;
1096         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1097         if (ret < 0) {
1098                 err = ret;
1099                 goto fail;
1100         }
1101
1102         if (parent) {
1103                 if (!ret)
1104                         return 0;
1105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1106                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1107                 btrfs_release_path(path);
1108                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109                 if (ret < 0) {
1110                         err = ret;
1111                         goto fail;
1112                 }
1113                 if (!ret)
1114                         return 0;
1115 #endif
1116                 goto fail;
1117         }
1118
1119         leaf = path->nodes[0];
1120         nritems = btrfs_header_nritems(leaf);
1121         while (1) {
1122                 if (path->slots[0] >= nritems) {
1123                         ret = btrfs_next_leaf(root, path);
1124                         if (ret < 0)
1125                                 err = ret;
1126                         if (ret)
1127                                 goto fail;
1128
1129                         leaf = path->nodes[0];
1130                         nritems = btrfs_header_nritems(leaf);
1131                         recow = 1;
1132                 }
1133
1134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135                 if (key.objectid != bytenr ||
1136                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1137                         goto fail;
1138
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141
1142                 if (match_extent_data_ref(leaf, ref, root_objectid,
1143                                           owner, offset)) {
1144                         if (recow) {
1145                                 btrfs_release_path(path);
1146                                 goto again;
1147                         }
1148                         err = 0;
1149                         break;
1150                 }
1151                 path->slots[0]++;
1152         }
1153 fail:
1154         return err;
1155 }
1156
1157 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1158                                            struct btrfs_root *root,
1159                                            struct btrfs_path *path,
1160                                            u64 bytenr, u64 parent,
1161                                            u64 root_objectid, u64 owner,
1162                                            u64 offset, int refs_to_add)
1163 {
1164         struct btrfs_key key;
1165         struct extent_buffer *leaf;
1166         u32 size;
1167         u32 num_refs;
1168         int ret;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174                 size = sizeof(struct btrfs_shared_data_ref);
1175         } else {
1176                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1177                 key.offset = hash_extent_data_ref(root_objectid,
1178                                                   owner, offset);
1179                 size = sizeof(struct btrfs_extent_data_ref);
1180         }
1181
1182         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1183         if (ret && ret != -EEXIST)
1184                 goto fail;
1185
1186         leaf = path->nodes[0];
1187         if (parent) {
1188                 struct btrfs_shared_data_ref *ref;
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_shared_data_ref);
1191                 if (ret == 0) {
1192                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1193                 } else {
1194                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1195                         num_refs += refs_to_add;
1196                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1197                 }
1198         } else {
1199                 struct btrfs_extent_data_ref *ref;
1200                 while (ret == -EEXIST) {
1201                         ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                              struct btrfs_extent_data_ref);
1203                         if (match_extent_data_ref(leaf, ref, root_objectid,
1204                                                   owner, offset))
1205                                 break;
1206                         btrfs_release_path(path);
1207                         key.offset++;
1208                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1209                                                       size);
1210                         if (ret && ret != -EEXIST)
1211                                 goto fail;
1212
1213                         leaf = path->nodes[0];
1214                 }
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_extent_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_extent_data_ref_root(leaf, ref,
1219                                                        root_objectid);
1220                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1221                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1222                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1223                 } else {
1224                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1225                         num_refs += refs_to_add;
1226                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1227                 }
1228         }
1229         btrfs_mark_buffer_dirty(leaf);
1230         ret = 0;
1231 fail:
1232         btrfs_release_path(path);
1233         return ret;
1234 }
1235
1236 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1237                                            struct btrfs_root *root,
1238                                            struct btrfs_path *path,
1239                                            int refs_to_drop)
1240 {
1241         struct btrfs_key key;
1242         struct btrfs_extent_data_ref *ref1 = NULL;
1243         struct btrfs_shared_data_ref *ref2 = NULL;
1244         struct extent_buffer *leaf;
1245         u32 num_refs = 0;
1246         int ret = 0;
1247
1248         leaf = path->nodes[0];
1249         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1250
1251         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1252                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1253                                       struct btrfs_extent_data_ref);
1254                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1255         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1256                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1257                                       struct btrfs_shared_data_ref);
1258                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1259 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1260         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1261                 struct btrfs_extent_ref_v0 *ref0;
1262                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1263                                       struct btrfs_extent_ref_v0);
1264                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1265 #endif
1266         } else {
1267                 BUG();
1268         }
1269
1270         BUG_ON(num_refs < refs_to_drop);
1271         num_refs -= refs_to_drop;
1272
1273         if (num_refs == 0) {
1274                 ret = btrfs_del_item(trans, root, path);
1275         } else {
1276                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1277                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1278                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1279                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281                 else {
1282                         struct btrfs_extent_ref_v0 *ref0;
1283                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                         struct btrfs_extent_ref_v0);
1285                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1286                 }
1287 #endif
1288                 btrfs_mark_buffer_dirty(leaf);
1289         }
1290         return ret;
1291 }
1292
1293 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           struct btrfs_extent_inline_ref *iref)
1296 {
1297         struct btrfs_key key;
1298         struct extent_buffer *leaf;
1299         struct btrfs_extent_data_ref *ref1;
1300         struct btrfs_shared_data_ref *ref2;
1301         u32 num_refs = 0;
1302
1303         leaf = path->nodes[0];
1304         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1305         if (iref) {
1306                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1307                     BTRFS_EXTENT_DATA_REF_KEY) {
1308                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310                 } else {
1311                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1313                 }
1314         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1315                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1316                                       struct btrfs_extent_data_ref);
1317                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1318         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1319                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_shared_data_ref);
1321                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1322 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1323         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1324                 struct btrfs_extent_ref_v0 *ref0;
1325                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1326                                       struct btrfs_extent_ref_v0);
1327                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1328 #endif
1329         } else {
1330                 WARN_ON(1);
1331         }
1332         return num_refs;
1333 }
1334
1335 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1336                                           struct btrfs_root *root,
1337                                           struct btrfs_path *path,
1338                                           u64 bytenr, u64 parent,
1339                                           u64 root_objectid)
1340 {
1341         struct btrfs_key key;
1342         int ret;
1343
1344         key.objectid = bytenr;
1345         if (parent) {
1346                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1347                 key.offset = parent;
1348         } else {
1349                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1350                 key.offset = root_objectid;
1351         }
1352
1353         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1354         if (ret > 0)
1355                 ret = -ENOENT;
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         if (ret == -ENOENT && parent) {
1358                 btrfs_release_path(path);
1359                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1360                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1361                 if (ret > 0)
1362                         ret = -ENOENT;
1363         }
1364 #endif
1365         return ret;
1366 }
1367
1368 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1369                                           struct btrfs_root *root,
1370                                           struct btrfs_path *path,
1371                                           u64 bytenr, u64 parent,
1372                                           u64 root_objectid)
1373 {
1374         struct btrfs_key key;
1375         int ret;
1376
1377         key.objectid = bytenr;
1378         if (parent) {
1379                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1380                 key.offset = parent;
1381         } else {
1382                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1383                 key.offset = root_objectid;
1384         }
1385
1386         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1387         btrfs_release_path(path);
1388         return ret;
1389 }
1390
1391 static inline int extent_ref_type(u64 parent, u64 owner)
1392 {
1393         int type;
1394         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1395                 if (parent > 0)
1396                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1397                 else
1398                         type = BTRFS_TREE_BLOCK_REF_KEY;
1399         } else {
1400                 if (parent > 0)
1401                         type = BTRFS_SHARED_DATA_REF_KEY;
1402                 else
1403                         type = BTRFS_EXTENT_DATA_REF_KEY;
1404         }
1405         return type;
1406 }
1407
1408 static int find_next_key(struct btrfs_path *path, int level,
1409                          struct btrfs_key *key)
1410
1411 {
1412         for (; level < BTRFS_MAX_LEVEL; level++) {
1413                 if (!path->nodes[level])
1414                         break;
1415                 if (path->slots[level] + 1 >=
1416                     btrfs_header_nritems(path->nodes[level]))
1417                         continue;
1418                 if (level == 0)
1419                         btrfs_item_key_to_cpu(path->nodes[level], key,
1420                                               path->slots[level] + 1);
1421                 else
1422                         btrfs_node_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 return 0;
1425         }
1426         return 1;
1427 }
1428
1429 /*
1430  * look for inline back ref. if back ref is found, *ref_ret is set
1431  * to the address of inline back ref, and 0 is returned.
1432  *
1433  * if back ref isn't found, *ref_ret is set to the address where it
1434  * should be inserted, and -ENOENT is returned.
1435  *
1436  * if insert is true and there are too many inline back refs, the path
1437  * points to the extent item, and -EAGAIN is returned.
1438  *
1439  * NOTE: inline back refs are ordered in the same way that back ref
1440  *       items in the tree are ordered.
1441  */
1442 static noinline_for_stack
1443 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1444                                  struct btrfs_root *root,
1445                                  struct btrfs_path *path,
1446                                  struct btrfs_extent_inline_ref **ref_ret,
1447                                  u64 bytenr, u64 num_bytes,
1448                                  u64 parent, u64 root_objectid,
1449                                  u64 owner, u64 offset, int insert)
1450 {
1451         struct btrfs_key key;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *ei;
1454         struct btrfs_extent_inline_ref *iref;
1455         u64 flags;
1456         u64 item_size;
1457         unsigned long ptr;
1458         unsigned long end;
1459         int extra_size;
1460         int type;
1461         int want;
1462         int ret;
1463         int err = 0;
1464
1465         key.objectid = bytenr;
1466         key.type = BTRFS_EXTENT_ITEM_KEY;
1467         key.offset = num_bytes;
1468
1469         want = extent_ref_type(parent, owner);
1470         if (insert) {
1471                 extra_size = btrfs_extent_inline_ref_size(want);
1472                 path->keep_locks = 1;
1473         } else
1474                 extra_size = -1;
1475         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1476         if (ret < 0) {
1477                 err = ret;
1478                 goto out;
1479         }
1480         BUG_ON(ret);
1481
1482         leaf = path->nodes[0];
1483         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1484 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1485         if (item_size < sizeof(*ei)) {
1486                 if (!insert) {
1487                         err = -ENOENT;
1488                         goto out;
1489                 }
1490                 ret = convert_extent_item_v0(trans, root, path, owner,
1491                                              extra_size);
1492                 if (ret < 0) {
1493                         err = ret;
1494                         goto out;
1495                 }
1496                 leaf = path->nodes[0];
1497                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1498         }
1499 #endif
1500         BUG_ON(item_size < sizeof(*ei));
1501
1502         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1503         flags = btrfs_extent_flags(leaf, ei);
1504
1505         ptr = (unsigned long)(ei + 1);
1506         end = (unsigned long)ei + item_size;
1507
1508         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1509                 ptr += sizeof(struct btrfs_tree_block_info);
1510                 BUG_ON(ptr > end);
1511         } else {
1512                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1513         }
1514
1515         err = -ENOENT;
1516         while (1) {
1517                 if (ptr >= end) {
1518                         WARN_ON(ptr > end);
1519                         break;
1520                 }
1521                 iref = (struct btrfs_extent_inline_ref *)ptr;
1522                 type = btrfs_extent_inline_ref_type(leaf, iref);
1523                 if (want < type)
1524                         break;
1525                 if (want > type) {
1526                         ptr += btrfs_extent_inline_ref_size(type);
1527                         continue;
1528                 }
1529
1530                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1531                         struct btrfs_extent_data_ref *dref;
1532                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1533                         if (match_extent_data_ref(leaf, dref, root_objectid,
1534                                                   owner, offset)) {
1535                                 err = 0;
1536                                 break;
1537                         }
1538                         if (hash_extent_data_ref_item(leaf, dref) <
1539                             hash_extent_data_ref(root_objectid, owner, offset))
1540                                 break;
1541                 } else {
1542                         u64 ref_offset;
1543                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1544                         if (parent > 0) {
1545                                 if (parent == ref_offset) {
1546                                         err = 0;
1547                                         break;
1548                                 }
1549                                 if (ref_offset < parent)
1550                                         break;
1551                         } else {
1552                                 if (root_objectid == ref_offset) {
1553                                         err = 0;
1554                                         break;
1555                                 }
1556                                 if (ref_offset < root_objectid)
1557                                         break;
1558                         }
1559                 }
1560                 ptr += btrfs_extent_inline_ref_size(type);
1561         }
1562         if (err == -ENOENT && insert) {
1563                 if (item_size + extra_size >=
1564                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1565                         err = -EAGAIN;
1566                         goto out;
1567                 }
1568                 /*
1569                  * To add new inline back ref, we have to make sure
1570                  * there is no corresponding back ref item.
1571                  * For simplicity, we just do not add new inline back
1572                  * ref if there is any kind of item for this block
1573                  */
1574                 if (find_next_key(path, 0, &key) == 0 &&
1575                     key.objectid == bytenr &&
1576                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1577                         err = -EAGAIN;
1578                         goto out;
1579                 }
1580         }
1581         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1582 out:
1583         if (insert) {
1584                 path->keep_locks = 0;
1585                 btrfs_unlock_up_safe(path, 1);
1586         }
1587         return err;
1588 }
1589
1590 /*
1591  * helper to add new inline back ref
1592  */
1593 static noinline_for_stack
1594 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1595                                 struct btrfs_root *root,
1596                                 struct btrfs_path *path,
1597                                 struct btrfs_extent_inline_ref *iref,
1598                                 u64 parent, u64 root_objectid,
1599                                 u64 owner, u64 offset, int refs_to_add,
1600                                 struct btrfs_delayed_extent_op *extent_op)
1601 {
1602         struct extent_buffer *leaf;
1603         struct btrfs_extent_item *ei;
1604         unsigned long ptr;
1605         unsigned long end;
1606         unsigned long item_offset;
1607         u64 refs;
1608         int size;
1609         int type;
1610         int ret;
1611
1612         leaf = path->nodes[0];
1613         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1614         item_offset = (unsigned long)iref - (unsigned long)ei;
1615
1616         type = extent_ref_type(parent, owner);
1617         size = btrfs_extent_inline_ref_size(type);
1618
1619         ret = btrfs_extend_item(trans, root, path, size);
1620
1621         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1622         refs = btrfs_extent_refs(leaf, ei);
1623         refs += refs_to_add;
1624         btrfs_set_extent_refs(leaf, ei, refs);
1625         if (extent_op)
1626                 __run_delayed_extent_op(extent_op, leaf, ei);
1627
1628         ptr = (unsigned long)ei + item_offset;
1629         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1630         if (ptr < end - size)
1631                 memmove_extent_buffer(leaf, ptr + size, ptr,
1632                                       end - size - ptr);
1633
1634         iref = (struct btrfs_extent_inline_ref *)ptr;
1635         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1636         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1637                 struct btrfs_extent_data_ref *dref;
1638                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1639                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1640                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1641                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1642                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1643         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1644                 struct btrfs_shared_data_ref *sref;
1645                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1646                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1647                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1648         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1649                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1650         } else {
1651                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1652         }
1653         btrfs_mark_buffer_dirty(leaf);
1654         return 0;
1655 }
1656
1657 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1658                                  struct btrfs_root *root,
1659                                  struct btrfs_path *path,
1660                                  struct btrfs_extent_inline_ref **ref_ret,
1661                                  u64 bytenr, u64 num_bytes, u64 parent,
1662                                  u64 root_objectid, u64 owner, u64 offset)
1663 {
1664         int ret;
1665
1666         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1667                                            bytenr, num_bytes, parent,
1668                                            root_objectid, owner, offset, 0);
1669         if (ret != -ENOENT)
1670                 return ret;
1671
1672         btrfs_release_path(path);
1673         *ref_ret = NULL;
1674
1675         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1676                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1677                                             root_objectid);
1678         } else {
1679                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1680                                              root_objectid, owner, offset);
1681         }
1682         return ret;
1683 }
1684
1685 /*
1686  * helper to update/remove inline back ref
1687  */
1688 static noinline_for_stack
1689 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1690                                  struct btrfs_root *root,
1691                                  struct btrfs_path *path,
1692                                  struct btrfs_extent_inline_ref *iref,
1693                                  int refs_to_mod,
1694                                  struct btrfs_delayed_extent_op *extent_op)
1695 {
1696         struct extent_buffer *leaf;
1697         struct btrfs_extent_item *ei;
1698         struct btrfs_extent_data_ref *dref = NULL;
1699         struct btrfs_shared_data_ref *sref = NULL;
1700         unsigned long ptr;
1701         unsigned long end;
1702         u32 item_size;
1703         int size;
1704         int type;
1705         int ret;
1706         u64 refs;
1707
1708         leaf = path->nodes[0];
1709         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1710         refs = btrfs_extent_refs(leaf, ei);
1711         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1712         refs += refs_to_mod;
1713         btrfs_set_extent_refs(leaf, ei, refs);
1714         if (extent_op)
1715                 __run_delayed_extent_op(extent_op, leaf, ei);
1716
1717         type = btrfs_extent_inline_ref_type(leaf, iref);
1718
1719         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1720                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1721                 refs = btrfs_extent_data_ref_count(leaf, dref);
1722         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1723                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1724                 refs = btrfs_shared_data_ref_count(leaf, sref);
1725         } else {
1726                 refs = 1;
1727                 BUG_ON(refs_to_mod != -1);
1728         }
1729
1730         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1731         refs += refs_to_mod;
1732
1733         if (refs > 0) {
1734                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1735                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1736                 else
1737                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1738         } else {
1739                 size =  btrfs_extent_inline_ref_size(type);
1740                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1741                 ptr = (unsigned long)iref;
1742                 end = (unsigned long)ei + item_size;
1743                 if (ptr + size < end)
1744                         memmove_extent_buffer(leaf, ptr, ptr + size,
1745                                               end - ptr - size);
1746                 item_size -= size;
1747                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1748         }
1749         btrfs_mark_buffer_dirty(leaf);
1750         return 0;
1751 }
1752
1753 static noinline_for_stack
1754 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1755                                  struct btrfs_root *root,
1756                                  struct btrfs_path *path,
1757                                  u64 bytenr, u64 num_bytes, u64 parent,
1758                                  u64 root_objectid, u64 owner,
1759                                  u64 offset, int refs_to_add,
1760                                  struct btrfs_delayed_extent_op *extent_op)
1761 {
1762         struct btrfs_extent_inline_ref *iref;
1763         int ret;
1764
1765         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1766                                            bytenr, num_bytes, parent,
1767                                            root_objectid, owner, offset, 1);
1768         if (ret == 0) {
1769                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1770                 ret = update_inline_extent_backref(trans, root, path, iref,
1771                                                    refs_to_add, extent_op);
1772         } else if (ret == -ENOENT) {
1773                 ret = setup_inline_extent_backref(trans, root, path, iref,
1774                                                   parent, root_objectid,
1775                                                   owner, offset, refs_to_add,
1776                                                   extent_op);
1777         }
1778         return ret;
1779 }
1780
1781 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1782                                  struct btrfs_root *root,
1783                                  struct btrfs_path *path,
1784                                  u64 bytenr, u64 parent, u64 root_objectid,
1785                                  u64 owner, u64 offset, int refs_to_add)
1786 {
1787         int ret;
1788         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1789                 BUG_ON(refs_to_add != 1);
1790                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1791                                             parent, root_objectid);
1792         } else {
1793                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1794                                              parent, root_objectid,
1795                                              owner, offset, refs_to_add);
1796         }
1797         return ret;
1798 }
1799
1800 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref *iref,
1804                                  int refs_to_drop, int is_data)
1805 {
1806         int ret;
1807
1808         BUG_ON(!is_data && refs_to_drop != 1);
1809         if (iref) {
1810                 ret = update_inline_extent_backref(trans, root, path, iref,
1811                                                    -refs_to_drop, NULL);
1812         } else if (is_data) {
1813                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1814         } else {
1815                 ret = btrfs_del_item(trans, root, path);
1816         }
1817         return ret;
1818 }
1819
1820 static int btrfs_issue_discard(struct block_device *bdev,
1821                                 u64 start, u64 len)
1822 {
1823         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1824 }
1825
1826 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1827                                 u64 num_bytes, u64 *actual_bytes)
1828 {
1829         int ret;
1830         u64 discarded_bytes = 0;
1831         struct btrfs_bio *bbio = NULL;
1832
1833
1834         /* Tell the block device(s) that the sectors can be discarded */
1835         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1836                               bytenr, &num_bytes, &bbio, 0);
1837         if (!ret) {
1838                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1839                 int i;
1840
1841
1842                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1843                         if (!stripe->dev->can_discard)
1844                                 continue;
1845
1846                         ret = btrfs_issue_discard(stripe->dev->bdev,
1847                                                   stripe->physical,
1848                                                   stripe->length);
1849                         if (!ret)
1850                                 discarded_bytes += stripe->length;
1851                         else if (ret != -EOPNOTSUPP)
1852                                 break;
1853
1854                         /*
1855                          * Just in case we get back EOPNOTSUPP for some reason,
1856                          * just ignore the return value so we don't screw up
1857                          * people calling discard_extent.
1858                          */
1859                         ret = 0;
1860                 }
1861                 kfree(bbio);
1862         }
1863
1864         if (actual_bytes)
1865                 *actual_bytes = discarded_bytes;
1866
1867
1868         return ret;
1869 }
1870
1871 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1872                          struct btrfs_root *root,
1873                          u64 bytenr, u64 num_bytes, u64 parent,
1874                          u64 root_objectid, u64 owner, u64 offset)
1875 {
1876         int ret;
1877         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1878                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1879
1880         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1881                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1882                                         parent, root_objectid, (int)owner,
1883                                         BTRFS_ADD_DELAYED_REF, NULL);
1884         } else {
1885                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1886                                         parent, root_objectid, owner, offset,
1887                                         BTRFS_ADD_DELAYED_REF, NULL);
1888         }
1889         return ret;
1890 }
1891
1892 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1893                                   struct btrfs_root *root,
1894                                   u64 bytenr, u64 num_bytes,
1895                                   u64 parent, u64 root_objectid,
1896                                   u64 owner, u64 offset, int refs_to_add,
1897                                   struct btrfs_delayed_extent_op *extent_op)
1898 {
1899         struct btrfs_path *path;
1900         struct extent_buffer *leaf;
1901         struct btrfs_extent_item *item;
1902         u64 refs;
1903         int ret;
1904         int err = 0;
1905
1906         path = btrfs_alloc_path();
1907         if (!path)
1908                 return -ENOMEM;
1909
1910         path->reada = 1;
1911         path->leave_spinning = 1;
1912         /* this will setup the path even if it fails to insert the back ref */
1913         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1914                                            path, bytenr, num_bytes, parent,
1915                                            root_objectid, owner, offset,
1916                                            refs_to_add, extent_op);
1917         if (ret == 0)
1918                 goto out;
1919
1920         if (ret != -EAGAIN) {
1921                 err = ret;
1922                 goto out;
1923         }
1924
1925         leaf = path->nodes[0];
1926         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1927         refs = btrfs_extent_refs(leaf, item);
1928         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1929         if (extent_op)
1930                 __run_delayed_extent_op(extent_op, leaf, item);
1931
1932         btrfs_mark_buffer_dirty(leaf);
1933         btrfs_release_path(path);
1934
1935         path->reada = 1;
1936         path->leave_spinning = 1;
1937
1938         /* now insert the actual backref */
1939         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1940                                     path, bytenr, parent, root_objectid,
1941                                     owner, offset, refs_to_add);
1942         BUG_ON(ret);
1943 out:
1944         btrfs_free_path(path);
1945         return err;
1946 }
1947
1948 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1949                                 struct btrfs_root *root,
1950                                 struct btrfs_delayed_ref_node *node,
1951                                 struct btrfs_delayed_extent_op *extent_op,
1952                                 int insert_reserved)
1953 {
1954         int ret = 0;
1955         struct btrfs_delayed_data_ref *ref;
1956         struct btrfs_key ins;
1957         u64 parent = 0;
1958         u64 ref_root = 0;
1959         u64 flags = 0;
1960
1961         ins.objectid = node->bytenr;
1962         ins.offset = node->num_bytes;
1963         ins.type = BTRFS_EXTENT_ITEM_KEY;
1964
1965         ref = btrfs_delayed_node_to_data_ref(node);
1966         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1967                 parent = ref->parent;
1968         else
1969                 ref_root = ref->root;
1970
1971         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1972                 if (extent_op) {
1973                         BUG_ON(extent_op->update_key);
1974                         flags |= extent_op->flags_to_set;
1975                 }
1976                 ret = alloc_reserved_file_extent(trans, root,
1977                                                  parent, ref_root, flags,
1978                                                  ref->objectid, ref->offset,
1979                                                  &ins, node->ref_mod);
1980         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1981                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1982                                              node->num_bytes, parent,
1983                                              ref_root, ref->objectid,
1984                                              ref->offset, node->ref_mod,
1985                                              extent_op);
1986         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1987                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1988                                           node->num_bytes, parent,
1989                                           ref_root, ref->objectid,
1990                                           ref->offset, node->ref_mod,
1991                                           extent_op);
1992         } else {
1993                 BUG();
1994         }
1995         return ret;
1996 }
1997
1998 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1999                                     struct extent_buffer *leaf,
2000                                     struct btrfs_extent_item *ei)
2001 {
2002         u64 flags = btrfs_extent_flags(leaf, ei);
2003         if (extent_op->update_flags) {
2004                 flags |= extent_op->flags_to_set;
2005                 btrfs_set_extent_flags(leaf, ei, flags);
2006         }
2007
2008         if (extent_op->update_key) {
2009                 struct btrfs_tree_block_info *bi;
2010                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2011                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2012                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2013         }
2014 }
2015
2016 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2017                                  struct btrfs_root *root,
2018                                  struct btrfs_delayed_ref_node *node,
2019                                  struct btrfs_delayed_extent_op *extent_op)
2020 {
2021         struct btrfs_key key;
2022         struct btrfs_path *path;
2023         struct btrfs_extent_item *ei;
2024         struct extent_buffer *leaf;
2025         u32 item_size;
2026         int ret;
2027         int err = 0;
2028
2029         path = btrfs_alloc_path();
2030         if (!path)
2031                 return -ENOMEM;
2032
2033         key.objectid = node->bytenr;
2034         key.type = BTRFS_EXTENT_ITEM_KEY;
2035         key.offset = node->num_bytes;
2036
2037         path->reada = 1;
2038         path->leave_spinning = 1;
2039         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2040                                 path, 0, 1);
2041         if (ret < 0) {
2042                 err = ret;
2043                 goto out;
2044         }
2045         if (ret > 0) {
2046                 err = -EIO;
2047                 goto out;
2048         }
2049
2050         leaf = path->nodes[0];
2051         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2052 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2053         if (item_size < sizeof(*ei)) {
2054                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2055                                              path, (u64)-1, 0);
2056                 if (ret < 0) {
2057                         err = ret;
2058                         goto out;
2059                 }
2060                 leaf = path->nodes[0];
2061                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2062         }
2063 #endif
2064         BUG_ON(item_size < sizeof(*ei));
2065         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2066         __run_delayed_extent_op(extent_op, leaf, ei);
2067
2068         btrfs_mark_buffer_dirty(leaf);
2069 out:
2070         btrfs_free_path(path);
2071         return err;
2072 }
2073
2074 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2075                                 struct btrfs_root *root,
2076                                 struct btrfs_delayed_ref_node *node,
2077                                 struct btrfs_delayed_extent_op *extent_op,
2078                                 int insert_reserved)
2079 {
2080         int ret = 0;
2081         struct btrfs_delayed_tree_ref *ref;
2082         struct btrfs_key ins;
2083         u64 parent = 0;
2084         u64 ref_root = 0;
2085
2086         ins.objectid = node->bytenr;
2087         ins.offset = node->num_bytes;
2088         ins.type = BTRFS_EXTENT_ITEM_KEY;
2089
2090         ref = btrfs_delayed_node_to_tree_ref(node);
2091         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2092                 parent = ref->parent;
2093         else
2094                 ref_root = ref->root;
2095
2096         BUG_ON(node->ref_mod != 1);
2097         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2098                 BUG_ON(!extent_op || !extent_op->update_flags ||
2099                        !extent_op->update_key);
2100                 ret = alloc_reserved_tree_block(trans, root,
2101                                                 parent, ref_root,
2102                                                 extent_op->flags_to_set,
2103                                                 &extent_op->key,
2104                                                 ref->level, &ins);
2105         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2106                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2107                                              node->num_bytes, parent, ref_root,
2108                                              ref->level, 0, 1, extent_op);
2109         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2110                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2111                                           node->num_bytes, parent, ref_root,
2112                                           ref->level, 0, 1, extent_op);
2113         } else {
2114                 BUG();
2115         }
2116         return ret;
2117 }
2118
2119 /* helper function to actually process a single delayed ref entry */
2120 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2121                                struct btrfs_root *root,
2122                                struct btrfs_delayed_ref_node *node,
2123                                struct btrfs_delayed_extent_op *extent_op,
2124                                int insert_reserved)
2125 {
2126         int ret;
2127         if (btrfs_delayed_ref_is_head(node)) {
2128                 struct btrfs_delayed_ref_head *head;
2129                 /*
2130                  * we've hit the end of the chain and we were supposed
2131                  * to insert this extent into the tree.  But, it got
2132                  * deleted before we ever needed to insert it, so all
2133                  * we have to do is clean up the accounting
2134                  */
2135                 BUG_ON(extent_op);
2136                 head = btrfs_delayed_node_to_head(node);
2137                 if (insert_reserved) {
2138                         btrfs_pin_extent(root, node->bytenr,
2139                                          node->num_bytes, 1);
2140                         if (head->is_data) {
2141                                 ret = btrfs_del_csums(trans, root,
2142                                                       node->bytenr,
2143                                                       node->num_bytes);
2144                                 BUG_ON(ret);
2145                         }
2146                 }
2147                 mutex_unlock(&head->mutex);
2148                 return 0;
2149         }
2150
2151         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2152             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2153                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2154                                            insert_reserved);
2155         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2156                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2157                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2158                                            insert_reserved);
2159         else
2160                 BUG();
2161         return ret;
2162 }
2163
2164 static noinline struct btrfs_delayed_ref_node *
2165 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2166 {
2167         struct rb_node *node;
2168         struct btrfs_delayed_ref_node *ref;
2169         int action = BTRFS_ADD_DELAYED_REF;
2170 again:
2171         /*
2172          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2173          * this prevents ref count from going down to zero when
2174          * there still are pending delayed ref.
2175          */
2176         node = rb_prev(&head->node.rb_node);
2177         while (1) {
2178                 if (!node)
2179                         break;
2180                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2181                                 rb_node);
2182                 if (ref->bytenr != head->node.bytenr)
2183                         break;
2184                 if (ref->action == action)
2185                         return ref;
2186                 node = rb_prev(node);
2187         }
2188         if (action == BTRFS_ADD_DELAYED_REF) {
2189                 action = BTRFS_DROP_DELAYED_REF;
2190                 goto again;
2191         }
2192         return NULL;
2193 }
2194
2195 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2196                                        struct btrfs_root *root,
2197                                        struct list_head *cluster)
2198 {
2199         struct btrfs_delayed_ref_root *delayed_refs;
2200         struct btrfs_delayed_ref_node *ref;
2201         struct btrfs_delayed_ref_head *locked_ref = NULL;
2202         struct btrfs_delayed_extent_op *extent_op;
2203         int ret;
2204         int count = 0;
2205         int must_insert_reserved = 0;
2206
2207         delayed_refs = &trans->transaction->delayed_refs;
2208         while (1) {
2209                 if (!locked_ref) {
2210                         /* pick a new head ref from the cluster list */
2211                         if (list_empty(cluster))
2212                                 break;
2213
2214                         locked_ref = list_entry(cluster->next,
2215                                      struct btrfs_delayed_ref_head, cluster);
2216
2217                         /* grab the lock that says we are going to process
2218                          * all the refs for this head */
2219                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2220
2221                         /*
2222                          * we may have dropped the spin lock to get the head
2223                          * mutex lock, and that might have given someone else
2224                          * time to free the head.  If that's true, it has been
2225                          * removed from our list and we can move on.
2226                          */
2227                         if (ret == -EAGAIN) {
2228                                 locked_ref = NULL;
2229                                 count++;
2230                                 continue;
2231                         }
2232                 }
2233
2234                 /*
2235                  * record the must insert reserved flag before we
2236                  * drop the spin lock.
2237                  */
2238                 must_insert_reserved = locked_ref->must_insert_reserved;
2239                 locked_ref->must_insert_reserved = 0;
2240
2241                 extent_op = locked_ref->extent_op;
2242                 locked_ref->extent_op = NULL;
2243
2244                 /*
2245                  * locked_ref is the head node, so we have to go one
2246                  * node back for any delayed ref updates
2247                  */
2248                 ref = select_delayed_ref(locked_ref);
2249                 if (!ref) {
2250                         /* All delayed refs have been processed, Go ahead
2251                          * and send the head node to run_one_delayed_ref,
2252                          * so that any accounting fixes can happen
2253                          */
2254                         ref = &locked_ref->node;
2255
2256                         if (extent_op && must_insert_reserved) {
2257                                 kfree(extent_op);
2258                                 extent_op = NULL;
2259                         }
2260
2261                         if (extent_op) {
2262                                 spin_unlock(&delayed_refs->lock);
2263
2264                                 ret = run_delayed_extent_op(trans, root,
2265                                                             ref, extent_op);
2266                                 BUG_ON(ret);
2267                                 kfree(extent_op);
2268
2269                                 cond_resched();
2270                                 spin_lock(&delayed_refs->lock);
2271                                 continue;
2272                         }
2273
2274                         list_del_init(&locked_ref->cluster);
2275                         locked_ref = NULL;
2276                 }
2277
2278                 ref->in_tree = 0;
2279                 rb_erase(&ref->rb_node, &delayed_refs->root);
2280                 delayed_refs->num_entries--;
2281
2282                 spin_unlock(&delayed_refs->lock);
2283
2284                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2285                                           must_insert_reserved);
2286                 BUG_ON(ret);
2287
2288                 btrfs_put_delayed_ref(ref);
2289                 kfree(extent_op);
2290                 count++;
2291
2292                 cond_resched();
2293                 spin_lock(&delayed_refs->lock);
2294         }
2295         return count;
2296 }
2297
2298 /*
2299  * this starts processing the delayed reference count updates and
2300  * extent insertions we have queued up so far.  count can be
2301  * 0, which means to process everything in the tree at the start
2302  * of the run (but not newly added entries), or it can be some target
2303  * number you'd like to process.
2304  */
2305 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2306                            struct btrfs_root *root, unsigned long count)
2307 {
2308         struct rb_node *node;
2309         struct btrfs_delayed_ref_root *delayed_refs;
2310         struct btrfs_delayed_ref_node *ref;
2311         struct list_head cluster;
2312         int ret;
2313         int run_all = count == (unsigned long)-1;
2314         int run_most = 0;
2315
2316         if (root == root->fs_info->extent_root)
2317                 root = root->fs_info->tree_root;
2318
2319         delayed_refs = &trans->transaction->delayed_refs;
2320         INIT_LIST_HEAD(&cluster);
2321 again:
2322         spin_lock(&delayed_refs->lock);
2323         if (count == 0) {
2324                 count = delayed_refs->num_entries * 2;
2325                 run_most = 1;
2326         }
2327         while (1) {
2328                 if (!(run_all || run_most) &&
2329                     delayed_refs->num_heads_ready < 64)
2330                         break;
2331
2332                 /*
2333                  * go find something we can process in the rbtree.  We start at
2334                  * the beginning of the tree, and then build a cluster
2335                  * of refs to process starting at the first one we are able to
2336                  * lock
2337                  */
2338                 ret = btrfs_find_ref_cluster(trans, &cluster,
2339                                              delayed_refs->run_delayed_start);
2340                 if (ret)
2341                         break;
2342
2343                 ret = run_clustered_refs(trans, root, &cluster);
2344                 BUG_ON(ret < 0);
2345
2346                 count -= min_t(unsigned long, ret, count);
2347
2348                 if (count == 0)
2349                         break;
2350         }
2351
2352         if (run_all) {
2353                 node = rb_first(&delayed_refs->root);
2354                 if (!node)
2355                         goto out;
2356                 count = (unsigned long)-1;
2357
2358                 while (node) {
2359                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2360                                        rb_node);
2361                         if (btrfs_delayed_ref_is_head(ref)) {
2362                                 struct btrfs_delayed_ref_head *head;
2363
2364                                 head = btrfs_delayed_node_to_head(ref);
2365                                 atomic_inc(&ref->refs);
2366
2367                                 spin_unlock(&delayed_refs->lock);
2368                                 /*
2369                                  * Mutex was contended, block until it's
2370                                  * released and try again
2371                                  */
2372                                 mutex_lock(&head->mutex);
2373                                 mutex_unlock(&head->mutex);
2374
2375                                 btrfs_put_delayed_ref(ref);
2376                                 cond_resched();
2377                                 goto again;
2378                         }
2379                         node = rb_next(node);
2380                 }
2381                 spin_unlock(&delayed_refs->lock);
2382                 schedule_timeout(1);
2383                 goto again;
2384         }
2385 out:
2386         spin_unlock(&delayed_refs->lock);
2387         return 0;
2388 }
2389
2390 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2391                                 struct btrfs_root *root,
2392                                 u64 bytenr, u64 num_bytes, u64 flags,
2393                                 int is_data)
2394 {
2395         struct btrfs_delayed_extent_op *extent_op;
2396         int ret;
2397
2398         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2399         if (!extent_op)
2400                 return -ENOMEM;
2401
2402         extent_op->flags_to_set = flags;
2403         extent_op->update_flags = 1;
2404         extent_op->update_key = 0;
2405         extent_op->is_data = is_data ? 1 : 0;
2406
2407         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2408         if (ret)
2409                 kfree(extent_op);
2410         return ret;
2411 }
2412
2413 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2414                                       struct btrfs_root *root,
2415                                       struct btrfs_path *path,
2416                                       u64 objectid, u64 offset, u64 bytenr)
2417 {
2418         struct btrfs_delayed_ref_head *head;
2419         struct btrfs_delayed_ref_node *ref;
2420         struct btrfs_delayed_data_ref *data_ref;
2421         struct btrfs_delayed_ref_root *delayed_refs;
2422         struct rb_node *node;
2423         int ret = 0;
2424
2425         ret = -ENOENT;
2426         delayed_refs = &trans->transaction->delayed_refs;
2427         spin_lock(&delayed_refs->lock);
2428         head = btrfs_find_delayed_ref_head(trans, bytenr);
2429         if (!head)
2430                 goto out;
2431
2432         if (!mutex_trylock(&head->mutex)) {
2433                 atomic_inc(&head->node.refs);
2434                 spin_unlock(&delayed_refs->lock);
2435
2436                 btrfs_release_path(path);
2437
2438                 /*
2439                  * Mutex was contended, block until it's released and let
2440                  * caller try again
2441                  */
2442                 mutex_lock(&head->mutex);
2443                 mutex_unlock(&head->mutex);
2444                 btrfs_put_delayed_ref(&head->node);
2445                 return -EAGAIN;
2446         }
2447
2448         node = rb_prev(&head->node.rb_node);
2449         if (!node)
2450                 goto out_unlock;
2451
2452         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2453
2454         if (ref->bytenr != bytenr)
2455                 goto out_unlock;
2456
2457         ret = 1;
2458         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2459                 goto out_unlock;
2460
2461         data_ref = btrfs_delayed_node_to_data_ref(ref);
2462
2463         node = rb_prev(node);
2464         if (node) {
2465                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2466                 if (ref->bytenr == bytenr)
2467                         goto out_unlock;
2468         }
2469
2470         if (data_ref->root != root->root_key.objectid ||
2471             data_ref->objectid != objectid || data_ref->offset != offset)
2472                 goto out_unlock;
2473
2474         ret = 0;
2475 out_unlock:
2476         mutex_unlock(&head->mutex);
2477 out:
2478         spin_unlock(&delayed_refs->lock);
2479         return ret;
2480 }
2481
2482 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2483                                         struct btrfs_root *root,
2484                                         struct btrfs_path *path,
2485                                         u64 objectid, u64 offset, u64 bytenr)
2486 {
2487         struct btrfs_root *extent_root = root->fs_info->extent_root;
2488         struct extent_buffer *leaf;
2489         struct btrfs_extent_data_ref *ref;
2490         struct btrfs_extent_inline_ref *iref;
2491         struct btrfs_extent_item *ei;
2492         struct btrfs_key key;
2493         u32 item_size;
2494         int ret;
2495
2496         key.objectid = bytenr;
2497         key.offset = (u64)-1;
2498         key.type = BTRFS_EXTENT_ITEM_KEY;
2499
2500         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2501         if (ret < 0)
2502                 goto out;
2503         BUG_ON(ret == 0);
2504
2505         ret = -ENOENT;
2506         if (path->slots[0] == 0)
2507                 goto out;
2508
2509         path->slots[0]--;
2510         leaf = path->nodes[0];
2511         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2512
2513         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2514                 goto out;
2515
2516         ret = 1;
2517         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2518 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2519         if (item_size < sizeof(*ei)) {
2520                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2521                 goto out;
2522         }
2523 #endif
2524         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2525
2526         if (item_size != sizeof(*ei) +
2527             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2528                 goto out;
2529
2530         if (btrfs_extent_generation(leaf, ei) <=
2531             btrfs_root_last_snapshot(&root->root_item))
2532                 goto out;
2533
2534         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2535         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2536             BTRFS_EXTENT_DATA_REF_KEY)
2537                 goto out;
2538
2539         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2540         if (btrfs_extent_refs(leaf, ei) !=
2541             btrfs_extent_data_ref_count(leaf, ref) ||
2542             btrfs_extent_data_ref_root(leaf, ref) !=
2543             root->root_key.objectid ||
2544             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2545             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2546                 goto out;
2547
2548         ret = 0;
2549 out:
2550         return ret;
2551 }
2552
2553 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2554                           struct btrfs_root *root,
2555                           u64 objectid, u64 offset, u64 bytenr)
2556 {
2557         struct btrfs_path *path;
2558         int ret;
2559         int ret2;
2560
2561         path = btrfs_alloc_path();
2562         if (!path)
2563                 return -ENOENT;
2564
2565         do {
2566                 ret = check_committed_ref(trans, root, path, objectid,
2567                                           offset, bytenr);
2568                 if (ret && ret != -ENOENT)
2569                         goto out;
2570
2571                 ret2 = check_delayed_ref(trans, root, path, objectid,
2572                                          offset, bytenr);
2573         } while (ret2 == -EAGAIN);
2574
2575         if (ret2 && ret2 != -ENOENT) {
2576                 ret = ret2;
2577                 goto out;
2578         }
2579
2580         if (ret != -ENOENT || ret2 != -ENOENT)
2581                 ret = 0;
2582 out:
2583         btrfs_free_path(path);
2584         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2585                 WARN_ON(ret > 0);
2586         return ret;
2587 }
2588
2589 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2590                            struct btrfs_root *root,
2591                            struct extent_buffer *buf,
2592                            int full_backref, int inc)
2593 {
2594         u64 bytenr;
2595         u64 num_bytes;
2596         u64 parent;
2597         u64 ref_root;
2598         u32 nritems;
2599         struct btrfs_key key;
2600         struct btrfs_file_extent_item *fi;
2601         int i;
2602         int level;
2603         int ret = 0;
2604         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2605                             u64, u64, u64, u64, u64, u64);
2606
2607         ref_root = btrfs_header_owner(buf);
2608         nritems = btrfs_header_nritems(buf);
2609         level = btrfs_header_level(buf);
2610
2611         if (!root->ref_cows && level == 0)
2612                 return 0;
2613
2614         if (inc)
2615                 process_func = btrfs_inc_extent_ref;
2616         else
2617                 process_func = btrfs_free_extent;
2618
2619         if (full_backref)
2620                 parent = buf->start;
2621         else
2622                 parent = 0;
2623
2624         for (i = 0; i < nritems; i++) {
2625                 if (level == 0) {
2626                         btrfs_item_key_to_cpu(buf, &key, i);
2627                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2628                                 continue;
2629                         fi = btrfs_item_ptr(buf, i,
2630                                             struct btrfs_file_extent_item);
2631                         if (btrfs_file_extent_type(buf, fi) ==
2632                             BTRFS_FILE_EXTENT_INLINE)
2633                                 continue;
2634                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2635                         if (bytenr == 0)
2636                                 continue;
2637
2638                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2639                         key.offset -= btrfs_file_extent_offset(buf, fi);
2640                         ret = process_func(trans, root, bytenr, num_bytes,
2641                                            parent, ref_root, key.objectid,
2642                                            key.offset);
2643                         if (ret)
2644                                 goto fail;
2645                 } else {
2646                         bytenr = btrfs_node_blockptr(buf, i);
2647                         num_bytes = btrfs_level_size(root, level - 1);
2648                         ret = process_func(trans, root, bytenr, num_bytes,
2649                                            parent, ref_root, level - 1, 0);
2650                         if (ret)
2651                                 goto fail;
2652                 }
2653         }
2654         return 0;
2655 fail:
2656         BUG();
2657         return ret;
2658 }
2659
2660 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2661                   struct extent_buffer *buf, int full_backref)
2662 {
2663         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2664 }
2665
2666 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2667                   struct extent_buffer *buf, int full_backref)
2668 {
2669         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2670 }
2671
2672 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2673                                  struct btrfs_root *root,
2674                                  struct btrfs_path *path,
2675                                  struct btrfs_block_group_cache *cache)
2676 {
2677         int ret;
2678         struct btrfs_root *extent_root = root->fs_info->extent_root;
2679         unsigned long bi;
2680         struct extent_buffer *leaf;
2681
2682         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2683         if (ret < 0)
2684                 goto fail;
2685         BUG_ON(ret);
2686
2687         leaf = path->nodes[0];
2688         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2689         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2690         btrfs_mark_buffer_dirty(leaf);
2691         btrfs_release_path(path);
2692 fail:
2693         if (ret)
2694                 return ret;
2695         return 0;
2696
2697 }
2698
2699 static struct btrfs_block_group_cache *
2700 next_block_group(struct btrfs_root *root,
2701                  struct btrfs_block_group_cache *cache)
2702 {
2703         struct rb_node *node;
2704         spin_lock(&root->fs_info->block_group_cache_lock);
2705         node = rb_next(&cache->cache_node);
2706         btrfs_put_block_group(cache);
2707         if (node) {
2708                 cache = rb_entry(node, struct btrfs_block_group_cache,
2709                                  cache_node);
2710                 btrfs_get_block_group(cache);
2711         } else
2712                 cache = NULL;
2713         spin_unlock(&root->fs_info->block_group_cache_lock);
2714         return cache;
2715 }
2716
2717 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2718                             struct btrfs_trans_handle *trans,
2719                             struct btrfs_path *path)
2720 {
2721         struct btrfs_root *root = block_group->fs_info->tree_root;
2722         struct inode *inode = NULL;
2723         u64 alloc_hint = 0;
2724         int dcs = BTRFS_DC_ERROR;
2725         int num_pages = 0;
2726         int retries = 0;
2727         int ret = 0;
2728
2729         /*
2730          * If this block group is smaller than 100 megs don't bother caching the
2731          * block group.
2732          */
2733         if (block_group->key.offset < (100 * 1024 * 1024)) {
2734                 spin_lock(&block_group->lock);
2735                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2736                 spin_unlock(&block_group->lock);
2737                 return 0;
2738         }
2739
2740 again:
2741         inode = lookup_free_space_inode(root, block_group, path);
2742         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2743                 ret = PTR_ERR(inode);
2744                 btrfs_release_path(path);
2745                 goto out;
2746         }
2747
2748         if (IS_ERR(inode)) {
2749                 BUG_ON(retries);
2750                 retries++;
2751
2752                 if (block_group->ro)
2753                         goto out_free;
2754
2755                 ret = create_free_space_inode(root, trans, block_group, path);
2756                 if (ret)
2757                         goto out_free;
2758                 goto again;
2759         }
2760
2761         /* We've already setup this transaction, go ahead and exit */
2762         if (block_group->cache_generation == trans->transid &&
2763             i_size_read(inode)) {
2764                 dcs = BTRFS_DC_SETUP;
2765                 goto out_put;
2766         }
2767
2768         /*
2769          * We want to set the generation to 0, that way if anything goes wrong
2770          * from here on out we know not to trust this cache when we load up next
2771          * time.
2772          */
2773         BTRFS_I(inode)->generation = 0;
2774         ret = btrfs_update_inode(trans, root, inode);
2775         WARN_ON(ret);
2776
2777         if (i_size_read(inode) > 0) {
2778                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2779                                                       inode);
2780                 if (ret)
2781                         goto out_put;
2782         }
2783
2784         spin_lock(&block_group->lock);
2785         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2786                 /* We're not cached, don't bother trying to write stuff out */
2787                 dcs = BTRFS_DC_WRITTEN;
2788                 spin_unlock(&block_group->lock);
2789                 goto out_put;
2790         }
2791         spin_unlock(&block_group->lock);
2792
2793         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2794         if (!num_pages)
2795                 num_pages = 1;
2796
2797         /*
2798          * Just to make absolutely sure we have enough space, we're going to
2799          * preallocate 12 pages worth of space for each block group.  In
2800          * practice we ought to use at most 8, but we need extra space so we can
2801          * add our header and have a terminator between the extents and the
2802          * bitmaps.
2803          */
2804         num_pages *= 16;
2805         num_pages *= PAGE_CACHE_SIZE;
2806
2807         ret = btrfs_check_data_free_space(inode, num_pages);
2808         if (ret)
2809                 goto out_put;
2810
2811         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2812                                               num_pages, num_pages,
2813                                               &alloc_hint);
2814         if (!ret)
2815                 dcs = BTRFS_DC_SETUP;
2816         btrfs_free_reserved_data_space(inode, num_pages);
2817
2818 out_put:
2819         iput(inode);
2820 out_free:
2821         btrfs_release_path(path);
2822 out:
2823         spin_lock(&block_group->lock);
2824         if (!ret && dcs == BTRFS_DC_SETUP)
2825                 block_group->cache_generation = trans->transid;
2826         block_group->disk_cache_state = dcs;
2827         spin_unlock(&block_group->lock);
2828
2829         return ret;
2830 }
2831
2832 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2833                                    struct btrfs_root *root)
2834 {
2835         struct btrfs_block_group_cache *cache;
2836         int err = 0;
2837         struct btrfs_path *path;
2838         u64 last = 0;
2839
2840         path = btrfs_alloc_path();
2841         if (!path)
2842                 return -ENOMEM;
2843
2844 again:
2845         while (1) {
2846                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2847                 while (cache) {
2848                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2849                                 break;
2850                         cache = next_block_group(root, cache);
2851                 }
2852                 if (!cache) {
2853                         if (last == 0)
2854                                 break;
2855                         last = 0;
2856                         continue;
2857                 }
2858                 err = cache_save_setup(cache, trans, path);
2859                 last = cache->key.objectid + cache->key.offset;
2860                 btrfs_put_block_group(cache);
2861         }
2862
2863         while (1) {
2864                 if (last == 0) {
2865                         err = btrfs_run_delayed_refs(trans, root,
2866                                                      (unsigned long)-1);
2867                         BUG_ON(err);
2868                 }
2869
2870                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2871                 while (cache) {
2872                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2873                                 btrfs_put_block_group(cache);
2874                                 goto again;
2875                         }
2876
2877                         if (cache->dirty)
2878                                 break;
2879                         cache = next_block_group(root, cache);
2880                 }
2881                 if (!cache) {
2882                         if (last == 0)
2883                                 break;
2884                         last = 0;
2885                         continue;
2886                 }
2887
2888                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2889                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2890                 cache->dirty = 0;
2891                 last = cache->key.objectid + cache->key.offset;
2892
2893                 err = write_one_cache_group(trans, root, path, cache);
2894                 BUG_ON(err);
2895                 btrfs_put_block_group(cache);
2896         }
2897
2898         while (1) {
2899                 /*
2900                  * I don't think this is needed since we're just marking our
2901                  * preallocated extent as written, but just in case it can't
2902                  * hurt.
2903                  */
2904                 if (last == 0) {
2905                         err = btrfs_run_delayed_refs(trans, root,
2906                                                      (unsigned long)-1);
2907                         BUG_ON(err);
2908                 }
2909
2910                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2911                 while (cache) {
2912                         /*
2913                          * Really this shouldn't happen, but it could if we
2914                          * couldn't write the entire preallocated extent and
2915                          * splitting the extent resulted in a new block.
2916                          */
2917                         if (cache->dirty) {
2918                                 btrfs_put_block_group(cache);
2919                                 goto again;
2920                         }
2921                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2922                                 break;
2923                         cache = next_block_group(root, cache);
2924                 }
2925                 if (!cache) {
2926                         if (last == 0)
2927                                 break;
2928                         last = 0;
2929                         continue;
2930                 }
2931
2932                 btrfs_write_out_cache(root, trans, cache, path);
2933
2934                 /*
2935                  * If we didn't have an error then the cache state is still
2936                  * NEED_WRITE, so we can set it to WRITTEN.
2937                  */
2938                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2939                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2940                 last = cache->key.objectid + cache->key.offset;
2941                 btrfs_put_block_group(cache);
2942         }
2943
2944         btrfs_free_path(path);
2945         return 0;
2946 }
2947
2948 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2949 {
2950         struct btrfs_block_group_cache *block_group;
2951         int readonly = 0;
2952
2953         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2954         if (!block_group || block_group->ro)
2955                 readonly = 1;
2956         if (block_group)
2957                 btrfs_put_block_group(block_group);
2958         return readonly;
2959 }
2960
2961 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2962                              u64 total_bytes, u64 bytes_used,
2963                              struct btrfs_space_info **space_info)
2964 {
2965         struct btrfs_space_info *found;
2966         int i;
2967         int factor;
2968
2969         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2970                      BTRFS_BLOCK_GROUP_RAID10))
2971                 factor = 2;
2972         else
2973                 factor = 1;
2974
2975         found = __find_space_info(info, flags);
2976         if (found) {
2977                 spin_lock(&found->lock);
2978                 found->total_bytes += total_bytes;
2979                 found->disk_total += total_bytes * factor;
2980                 found->bytes_used += bytes_used;
2981                 found->disk_used += bytes_used * factor;
2982                 found->full = 0;
2983                 spin_unlock(&found->lock);
2984                 *space_info = found;
2985                 return 0;
2986         }
2987         found = kzalloc(sizeof(*found), GFP_NOFS);
2988         if (!found)
2989                 return -ENOMEM;
2990
2991         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2992                 INIT_LIST_HEAD(&found->block_groups[i]);
2993         init_rwsem(&found->groups_sem);
2994         spin_lock_init(&found->lock);
2995         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
2996         found->total_bytes = total_bytes;
2997         found->disk_total = total_bytes * factor;
2998         found->bytes_used = bytes_used;
2999         found->disk_used = bytes_used * factor;
3000         found->bytes_pinned = 0;
3001         found->bytes_reserved = 0;
3002         found->bytes_readonly = 0;
3003         found->bytes_may_use = 0;
3004         found->full = 0;
3005         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3006         found->chunk_alloc = 0;
3007         found->flush = 0;
3008         init_waitqueue_head(&found->wait);
3009         *space_info = found;
3010         list_add_rcu(&found->list, &info->space_info);
3011         return 0;
3012 }
3013
3014 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3015 {
3016         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3017
3018         /* chunk -> extended profile */
3019         if (extra_flags == 0)
3020                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3021
3022         if (flags & BTRFS_BLOCK_GROUP_DATA)
3023                 fs_info->avail_data_alloc_bits |= extra_flags;
3024         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3025                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3026         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3027                 fs_info->avail_system_alloc_bits |= extra_flags;
3028 }
3029
3030 /*
3031  * @flags: available profiles in extended format (see ctree.h)
3032  *
3033  * Returns reduced profile in chunk format.
3034  */
3035 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3036 {
3037         /*
3038          * we add in the count of missing devices because we want
3039          * to make sure that any RAID levels on a degraded FS
3040          * continue to be honored.
3041          */
3042         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3043                 root->fs_info->fs_devices->missing_devices;
3044
3045         if (num_devices == 1)
3046                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3047         if (num_devices < 4)
3048                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3049
3050         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3051             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3052                       BTRFS_BLOCK_GROUP_RAID10))) {
3053                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3054         }
3055
3056         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3057             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3058                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3059         }
3060
3061         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3062             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3063              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3064              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3065                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3066         }
3067
3068         /* extended -> chunk profile */
3069         flags &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
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         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3078                 flags |= root->fs_info->avail_system_alloc_bits;
3079         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3080                 flags |= root->fs_info->avail_metadata_alloc_bits;
3081
3082         return btrfs_reduce_alloc_profile(root, flags);
3083 }
3084
3085 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3086 {
3087         u64 flags;
3088
3089         if (data)
3090                 flags = BTRFS_BLOCK_GROUP_DATA;
3091         else if (root == root->fs_info->chunk_root)
3092                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3093         else
3094                 flags = BTRFS_BLOCK_GROUP_METADATA;
3095
3096         return get_alloc_profile(root, flags);
3097 }
3098
3099 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3100 {
3101         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3102                                                        BTRFS_BLOCK_GROUP_DATA);
3103 }
3104
3105 /*
3106  * This will check the space that the inode allocates from to make sure we have
3107  * enough space for bytes.
3108  */
3109 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3110 {
3111         struct btrfs_space_info *data_sinfo;
3112         struct btrfs_root *root = BTRFS_I(inode)->root;
3113         u64 used;
3114         int ret = 0, committed = 0, alloc_chunk = 1;
3115
3116         /* make sure bytes are sectorsize aligned */
3117         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3118
3119         if (root == root->fs_info->tree_root ||
3120             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3121                 alloc_chunk = 0;
3122                 committed = 1;
3123         }
3124
3125         data_sinfo = BTRFS_I(inode)->space_info;
3126         if (!data_sinfo)
3127                 goto alloc;
3128
3129 again:
3130         /* make sure we have enough space to handle the data first */
3131         spin_lock(&data_sinfo->lock);
3132         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3133                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3134                 data_sinfo->bytes_may_use;
3135
3136         if (used + bytes > data_sinfo->total_bytes) {
3137                 struct btrfs_trans_handle *trans;
3138
3139                 /*
3140                  * if we don't have enough free bytes in this space then we need
3141                  * to alloc a new chunk.
3142                  */
3143                 if (!data_sinfo->full && alloc_chunk) {
3144                         u64 alloc_target;
3145
3146                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3147                         spin_unlock(&data_sinfo->lock);
3148 alloc:
3149                         alloc_target = btrfs_get_alloc_profile(root, 1);
3150                         trans = btrfs_join_transaction(root);
3151                         if (IS_ERR(trans))
3152                                 return PTR_ERR(trans);
3153
3154                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3155                                              bytes + 2 * 1024 * 1024,
3156                                              alloc_target,
3157                                              CHUNK_ALLOC_NO_FORCE);
3158                         btrfs_end_transaction(trans, root);
3159                         if (ret < 0) {
3160                                 if (ret != -ENOSPC)
3161                                         return ret;
3162                                 else
3163                                         goto commit_trans;
3164                         }
3165
3166                         if (!data_sinfo) {
3167                                 btrfs_set_inode_space_info(root, inode);
3168                                 data_sinfo = BTRFS_I(inode)->space_info;
3169                         }
3170                         goto again;
3171                 }
3172
3173                 /*
3174                  * If we have less pinned bytes than we want to allocate then
3175                  * don't bother committing the transaction, it won't help us.
3176                  */
3177                 if (data_sinfo->bytes_pinned < bytes)
3178                         committed = 1;
3179                 spin_unlock(&data_sinfo->lock);
3180
3181                 /* commit the current transaction and try again */
3182 commit_trans:
3183                 if (!committed &&
3184                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3185                         committed = 1;
3186                         trans = btrfs_join_transaction(root);
3187                         if (IS_ERR(trans))
3188                                 return PTR_ERR(trans);
3189                         ret = btrfs_commit_transaction(trans, root);
3190                         if (ret)
3191                                 return ret;
3192                         goto again;
3193                 }
3194
3195                 return -ENOSPC;
3196         }
3197         data_sinfo->bytes_may_use += bytes;
3198         spin_unlock(&data_sinfo->lock);
3199
3200         return 0;
3201 }
3202
3203 /*
3204  * Called if we need to clear a data reservation for this inode.
3205  */
3206 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3207 {
3208         struct btrfs_root *root = BTRFS_I(inode)->root;
3209         struct btrfs_space_info *data_sinfo;
3210
3211         /* make sure bytes are sectorsize aligned */
3212         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3213
3214         data_sinfo = BTRFS_I(inode)->space_info;
3215         spin_lock(&data_sinfo->lock);
3216         data_sinfo->bytes_may_use -= bytes;
3217         spin_unlock(&data_sinfo->lock);
3218 }
3219
3220 static void force_metadata_allocation(struct btrfs_fs_info *info)
3221 {
3222         struct list_head *head = &info->space_info;
3223         struct btrfs_space_info *found;
3224
3225         rcu_read_lock();
3226         list_for_each_entry_rcu(found, head, list) {
3227                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3228                         found->force_alloc = CHUNK_ALLOC_FORCE;
3229         }
3230         rcu_read_unlock();
3231 }
3232
3233 static int should_alloc_chunk(struct btrfs_root *root,
3234                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3235                               int force)
3236 {
3237         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3238         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3239         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3240         u64 thresh;
3241
3242         if (force == CHUNK_ALLOC_FORCE)
3243                 return 1;
3244
3245         /*
3246          * We need to take into account the global rsv because for all intents
3247          * and purposes it's used space.  Don't worry about locking the
3248          * global_rsv, it doesn't change except when the transaction commits.
3249          */
3250         num_allocated += global_rsv->size;
3251
3252         /*
3253          * in limited mode, we want to have some free space up to
3254          * about 1% of the FS size.
3255          */
3256         if (force == CHUNK_ALLOC_LIMITED) {
3257                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3258                 thresh = max_t(u64, 64 * 1024 * 1024,
3259                                div_factor_fine(thresh, 1));
3260
3261                 if (num_bytes - num_allocated < thresh)
3262                         return 1;
3263         }
3264
3265         /*
3266          * we have two similar checks here, one based on percentage
3267          * and once based on a hard number of 256MB.  The idea
3268          * is that if we have a good amount of free
3269          * room, don't allocate a chunk.  A good mount is
3270          * less than 80% utilized of the chunks we have allocated,
3271          * or more than 256MB free
3272          */
3273         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3274                 return 0;
3275
3276         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3277                 return 0;
3278
3279         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3280
3281         /* 256MB or 5% of the FS */
3282         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3283
3284         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3285                 return 0;
3286         return 1;
3287 }
3288
3289 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3290                           struct btrfs_root *extent_root, u64 alloc_bytes,
3291                           u64 flags, int force)
3292 {
3293         struct btrfs_space_info *space_info;
3294         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3295         int wait_for_alloc = 0;
3296         int ret = 0;
3297
3298         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3299
3300         space_info = __find_space_info(extent_root->fs_info, flags);
3301         if (!space_info) {
3302                 ret = update_space_info(extent_root->fs_info, flags,
3303                                         0, 0, &space_info);
3304                 BUG_ON(ret);
3305         }
3306         BUG_ON(!space_info);
3307
3308 again:
3309         spin_lock(&space_info->lock);
3310         if (space_info->force_alloc)
3311                 force = space_info->force_alloc;
3312         if (space_info->full) {
3313                 spin_unlock(&space_info->lock);
3314                 return 0;
3315         }
3316
3317         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3318                 spin_unlock(&space_info->lock);
3319                 return 0;
3320         } else if (space_info->chunk_alloc) {
3321                 wait_for_alloc = 1;
3322         } else {
3323                 space_info->chunk_alloc = 1;
3324         }
3325
3326         spin_unlock(&space_info->lock);
3327
3328         mutex_lock(&fs_info->chunk_mutex);
3329
3330         /*
3331          * The chunk_mutex is held throughout the entirety of a chunk
3332          * allocation, so once we've acquired the chunk_mutex we know that the
3333          * other guy is done and we need to recheck and see if we should
3334          * allocate.
3335          */
3336         if (wait_for_alloc) {
3337                 mutex_unlock(&fs_info->chunk_mutex);
3338                 wait_for_alloc = 0;
3339                 goto again;
3340         }
3341
3342         /*
3343          * If we have mixed data/metadata chunks we want to make sure we keep
3344          * allocating mixed chunks instead of individual chunks.
3345          */
3346         if (btrfs_mixed_space_info(space_info))
3347                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3348
3349         /*
3350          * if we're doing a data chunk, go ahead and make sure that
3351          * we keep a reasonable number of metadata chunks allocated in the
3352          * FS as well.
3353          */
3354         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3355                 fs_info->data_chunk_allocations++;
3356                 if (!(fs_info->data_chunk_allocations %
3357                       fs_info->metadata_ratio))
3358                         force_metadata_allocation(fs_info);
3359         }
3360
3361         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3362         if (ret < 0 && ret != -ENOSPC)
3363                 goto out;
3364
3365         spin_lock(&space_info->lock);
3366         if (ret)
3367                 space_info->full = 1;
3368         else
3369                 ret = 1;
3370
3371         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3372         space_info->chunk_alloc = 0;
3373         spin_unlock(&space_info->lock);
3374 out:
3375         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3376         return ret;
3377 }
3378
3379 /*
3380  * shrink metadata reservation for delalloc
3381  */
3382 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3383                            bool wait_ordered)
3384 {
3385         struct btrfs_block_rsv *block_rsv;
3386         struct btrfs_space_info *space_info;
3387         struct btrfs_trans_handle *trans;
3388         u64 reserved;
3389         u64 max_reclaim;
3390         u64 reclaimed = 0;
3391         long time_left;
3392         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3393         int loops = 0;
3394         unsigned long progress;
3395
3396         trans = (struct btrfs_trans_handle *)current->journal_info;
3397         block_rsv = &root->fs_info->delalloc_block_rsv;
3398         space_info = block_rsv->space_info;
3399
3400         smp_mb();
3401         reserved = space_info->bytes_may_use;
3402         progress = space_info->reservation_progress;
3403
3404         if (reserved == 0)
3405                 return 0;
3406
3407         smp_mb();
3408         if (root->fs_info->delalloc_bytes == 0) {
3409                 if (trans)
3410                         return 0;
3411                 btrfs_wait_ordered_extents(root, 0, 0);
3412                 return 0;
3413         }
3414
3415         max_reclaim = min(reserved, to_reclaim);
3416         nr_pages = max_t(unsigned long, nr_pages,
3417                          max_reclaim >> PAGE_CACHE_SHIFT);
3418         while (loops < 1024) {
3419                 /* have the flusher threads jump in and do some IO */
3420                 smp_mb();
3421                 nr_pages = min_t(unsigned long, nr_pages,
3422                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3423                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3424
3425                 spin_lock(&space_info->lock);
3426                 if (reserved > space_info->bytes_may_use)
3427                         reclaimed += reserved - space_info->bytes_may_use;
3428                 reserved = space_info->bytes_may_use;
3429                 spin_unlock(&space_info->lock);
3430
3431                 loops++;
3432
3433                 if (reserved == 0 || reclaimed >= max_reclaim)
3434                         break;
3435
3436                 if (trans && trans->transaction->blocked)
3437                         return -EAGAIN;
3438
3439                 if (wait_ordered && !trans) {
3440                         btrfs_wait_ordered_extents(root, 0, 0);
3441                 } else {
3442                         time_left = schedule_timeout_interruptible(1);
3443
3444                         /* We were interrupted, exit */
3445                         if (time_left)
3446                                 break;
3447                 }
3448
3449                 /* we've kicked the IO a few times, if anything has been freed,
3450                  * exit.  There is no sense in looping here for a long time
3451                  * when we really need to commit the transaction, or there are
3452                  * just too many writers without enough free space
3453                  */
3454
3455                 if (loops > 3) {
3456                         smp_mb();
3457                         if (progress != space_info->reservation_progress)
3458                                 break;
3459                 }
3460
3461         }
3462
3463         return reclaimed >= to_reclaim;
3464 }
3465
3466 /**
3467  * maybe_commit_transaction - possibly commit the transaction if its ok to
3468  * @root - the root we're allocating for
3469  * @bytes - the number of bytes we want to reserve
3470  * @force - force the commit
3471  *
3472  * This will check to make sure that committing the transaction will actually
3473  * get us somewhere and then commit the transaction if it does.  Otherwise it
3474  * will return -ENOSPC.
3475  */
3476 static int may_commit_transaction(struct btrfs_root *root,
3477                                   struct btrfs_space_info *space_info,
3478                                   u64 bytes, int force)
3479 {
3480         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3481         struct btrfs_trans_handle *trans;
3482
3483         trans = (struct btrfs_trans_handle *)current->journal_info;
3484         if (trans)
3485                 return -EAGAIN;
3486
3487         if (force)
3488                 goto commit;
3489
3490         /* See if there is enough pinned space to make this reservation */
3491         spin_lock(&space_info->lock);
3492         if (space_info->bytes_pinned >= bytes) {
3493                 spin_unlock(&space_info->lock);
3494                 goto commit;
3495         }
3496         spin_unlock(&space_info->lock);
3497
3498         /*
3499          * See if there is some space in the delayed insertion reservation for
3500          * this reservation.
3501          */
3502         if (space_info != delayed_rsv->space_info)
3503                 return -ENOSPC;
3504
3505         spin_lock(&delayed_rsv->lock);
3506         if (delayed_rsv->size < bytes) {
3507                 spin_unlock(&delayed_rsv->lock);
3508                 return -ENOSPC;
3509         }
3510         spin_unlock(&delayed_rsv->lock);
3511
3512 commit:
3513         trans = btrfs_join_transaction(root);
3514         if (IS_ERR(trans))
3515                 return -ENOSPC;
3516
3517         return btrfs_commit_transaction(trans, root);
3518 }
3519
3520 /**
3521  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3522  * @root - the root we're allocating for
3523  * @block_rsv - the block_rsv we're allocating for
3524  * @orig_bytes - the number of bytes we want
3525  * @flush - wether or not we can flush to make our reservation
3526  *
3527  * This will reserve orgi_bytes number of bytes from the space info associated
3528  * with the block_rsv.  If there is not enough space it will make an attempt to
3529  * flush out space to make room.  It will do this by flushing delalloc if
3530  * possible or committing the transaction.  If flush is 0 then no attempts to
3531  * regain reservations will be made and this will fail if there is not enough
3532  * space already.
3533  */
3534 static int reserve_metadata_bytes(struct btrfs_root *root,
3535                                   struct btrfs_block_rsv *block_rsv,
3536                                   u64 orig_bytes, int flush)
3537 {
3538         struct btrfs_space_info *space_info = block_rsv->space_info;
3539         u64 used;
3540         u64 num_bytes = orig_bytes;
3541         int retries = 0;
3542         int ret = 0;
3543         bool committed = false;
3544         bool flushing = false;
3545         bool wait_ordered = false;
3546
3547 again:
3548         ret = 0;
3549         spin_lock(&space_info->lock);
3550         /*
3551          * We only want to wait if somebody other than us is flushing and we are
3552          * actually alloed to flush.
3553          */
3554         while (flush && !flushing && space_info->flush) {
3555                 spin_unlock(&space_info->lock);
3556                 /*
3557                  * If we have a trans handle we can't wait because the flusher
3558                  * may have to commit the transaction, which would mean we would
3559                  * deadlock since we are waiting for the flusher to finish, but
3560                  * hold the current transaction open.
3561                  */
3562                 if (current->journal_info)
3563                         return -EAGAIN;
3564                 ret = wait_event_interruptible(space_info->wait,
3565                                                !space_info->flush);
3566                 /* Must have been interrupted, return */
3567                 if (ret)
3568                         return -EINTR;
3569
3570                 spin_lock(&space_info->lock);
3571         }
3572
3573         ret = -ENOSPC;
3574         used = space_info->bytes_used + space_info->bytes_reserved +
3575                 space_info->bytes_pinned + space_info->bytes_readonly +
3576                 space_info->bytes_may_use;
3577
3578         /*
3579          * The idea here is that we've not already over-reserved the block group
3580          * then we can go ahead and save our reservation first and then start
3581          * flushing if we need to.  Otherwise if we've already overcommitted
3582          * lets start flushing stuff first and then come back and try to make
3583          * our reservation.
3584          */
3585         if (used <= space_info->total_bytes) {
3586                 if (used + orig_bytes <= space_info->total_bytes) {
3587                         space_info->bytes_may_use += orig_bytes;
3588                         ret = 0;
3589                 } else {
3590                         /*
3591                          * Ok set num_bytes to orig_bytes since we aren't
3592                          * overocmmitted, this way we only try and reclaim what
3593                          * we need.
3594                          */
3595                         num_bytes = orig_bytes;
3596                 }
3597         } else {
3598                 /*
3599                  * Ok we're over committed, set num_bytes to the overcommitted
3600                  * amount plus the amount of bytes that we need for this
3601                  * reservation.
3602                  */
3603                 wait_ordered = true;
3604                 num_bytes = used - space_info->total_bytes +
3605                         (orig_bytes * (retries + 1));
3606         }
3607
3608         if (ret) {
3609                 u64 profile = btrfs_get_alloc_profile(root, 0);
3610                 u64 avail;
3611
3612                 /*
3613                  * If we have a lot of space that's pinned, don't bother doing
3614                  * the overcommit dance yet and just commit the transaction.
3615                  */
3616                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3617                 do_div(avail, 10);
3618                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3619                         space_info->flush = 1;
3620                         flushing = true;
3621                         spin_unlock(&space_info->lock);
3622                         ret = may_commit_transaction(root, space_info,
3623                                                      orig_bytes, 1);
3624                         if (ret)
3625                                 goto out;
3626                         committed = true;
3627                         goto again;
3628                 }
3629
3630                 spin_lock(&root->fs_info->free_chunk_lock);
3631                 avail = root->fs_info->free_chunk_space;
3632
3633                 /*
3634                  * If we have dup, raid1 or raid10 then only half of the free
3635                  * space is actually useable.
3636                  */
3637                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3638                                BTRFS_BLOCK_GROUP_RAID1 |
3639                                BTRFS_BLOCK_GROUP_RAID10))
3640                         avail >>= 1;
3641
3642                 /*
3643                  * If we aren't flushing don't let us overcommit too much, say
3644                  * 1/8th of the space.  If we can flush, let it overcommit up to
3645                  * 1/2 of the space.
3646                  */
3647                 if (flush)
3648                         avail >>= 3;
3649                 else
3650                         avail >>= 1;
3651                  spin_unlock(&root->fs_info->free_chunk_lock);
3652
3653                 if (used + num_bytes < space_info->total_bytes + avail) {
3654                         space_info->bytes_may_use += orig_bytes;
3655                         ret = 0;
3656                 } else {
3657                         wait_ordered = true;
3658                 }
3659         }
3660
3661         /*
3662          * Couldn't make our reservation, save our place so while we're trying
3663          * to reclaim space we can actually use it instead of somebody else
3664          * stealing it from us.
3665          */
3666         if (ret && flush) {
3667                 flushing = true;
3668                 space_info->flush = 1;
3669         }
3670
3671         spin_unlock(&space_info->lock);
3672
3673         if (!ret || !flush)
3674                 goto out;
3675
3676         /*
3677          * We do synchronous shrinking since we don't actually unreserve
3678          * metadata until after the IO is completed.
3679          */
3680         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3681         if (ret < 0)
3682                 goto out;
3683
3684         ret = 0;
3685
3686         /*
3687          * So if we were overcommitted it's possible that somebody else flushed
3688          * out enough space and we simply didn't have enough space to reclaim,
3689          * so go back around and try again.
3690          */
3691         if (retries < 2) {
3692                 wait_ordered = true;
3693                 retries++;
3694                 goto again;
3695         }
3696
3697         ret = -ENOSPC;
3698         if (committed)
3699                 goto out;
3700
3701         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3702         if (!ret) {
3703                 committed = true;
3704                 goto again;
3705         }
3706
3707 out:
3708         if (flushing) {
3709                 spin_lock(&space_info->lock);
3710                 space_info->flush = 0;
3711                 wake_up_all(&space_info->wait);
3712                 spin_unlock(&space_info->lock);
3713         }
3714         return ret;
3715 }
3716
3717 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3718                                              struct btrfs_root *root)
3719 {
3720         struct btrfs_block_rsv *block_rsv = NULL;
3721
3722         if (root->ref_cows || root == root->fs_info->csum_root)
3723                 block_rsv = trans->block_rsv;
3724
3725         if (!block_rsv)
3726                 block_rsv = root->block_rsv;
3727
3728         if (!block_rsv)
3729                 block_rsv = &root->fs_info->empty_block_rsv;
3730
3731         return block_rsv;
3732 }
3733
3734 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3735                                u64 num_bytes)
3736 {
3737         int ret = -ENOSPC;
3738         spin_lock(&block_rsv->lock);
3739         if (block_rsv->reserved >= num_bytes) {
3740                 block_rsv->reserved -= num_bytes;
3741                 if (block_rsv->reserved < block_rsv->size)
3742                         block_rsv->full = 0;
3743                 ret = 0;
3744         }
3745         spin_unlock(&block_rsv->lock);
3746         return ret;
3747 }
3748
3749 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3750                                 u64 num_bytes, int update_size)
3751 {
3752         spin_lock(&block_rsv->lock);
3753         block_rsv->reserved += num_bytes;
3754         if (update_size)
3755                 block_rsv->size += num_bytes;
3756         else if (block_rsv->reserved >= block_rsv->size)
3757                 block_rsv->full = 1;
3758         spin_unlock(&block_rsv->lock);
3759 }
3760
3761 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3762                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3763 {
3764         struct btrfs_space_info *space_info = block_rsv->space_info;
3765
3766         spin_lock(&block_rsv->lock);
3767         if (num_bytes == (u64)-1)
3768                 num_bytes = block_rsv->size;
3769         block_rsv->size -= num_bytes;
3770         if (block_rsv->reserved >= block_rsv->size) {
3771                 num_bytes = block_rsv->reserved - block_rsv->size;
3772                 block_rsv->reserved = block_rsv->size;
3773                 block_rsv->full = 1;
3774         } else {
3775                 num_bytes = 0;
3776         }
3777         spin_unlock(&block_rsv->lock);
3778
3779         if (num_bytes > 0) {
3780                 if (dest) {
3781                         spin_lock(&dest->lock);
3782                         if (!dest->full) {
3783                                 u64 bytes_to_add;
3784
3785                                 bytes_to_add = dest->size - dest->reserved;
3786                                 bytes_to_add = min(num_bytes, bytes_to_add);
3787                                 dest->reserved += bytes_to_add;
3788                                 if (dest->reserved >= dest->size)
3789                                         dest->full = 1;
3790                                 num_bytes -= bytes_to_add;
3791                         }
3792                         spin_unlock(&dest->lock);
3793                 }
3794                 if (num_bytes) {
3795                         spin_lock(&space_info->lock);
3796                         space_info->bytes_may_use -= num_bytes;
3797                         space_info->reservation_progress++;
3798                         spin_unlock(&space_info->lock);
3799                 }
3800         }
3801 }
3802
3803 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3804                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3805 {
3806         int ret;
3807
3808         ret = block_rsv_use_bytes(src, num_bytes);
3809         if (ret)
3810                 return ret;
3811
3812         block_rsv_add_bytes(dst, num_bytes, 1);
3813         return 0;
3814 }
3815
3816 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3817 {
3818         memset(rsv, 0, sizeof(*rsv));
3819         spin_lock_init(&rsv->lock);
3820 }
3821
3822 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3823 {
3824         struct btrfs_block_rsv *block_rsv;
3825         struct btrfs_fs_info *fs_info = root->fs_info;
3826
3827         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3828         if (!block_rsv)
3829                 return NULL;
3830
3831         btrfs_init_block_rsv(block_rsv);
3832         block_rsv->space_info = __find_space_info(fs_info,
3833                                                   BTRFS_BLOCK_GROUP_METADATA);
3834         return block_rsv;
3835 }
3836
3837 void btrfs_free_block_rsv(struct btrfs_root *root,
3838                           struct btrfs_block_rsv *rsv)
3839 {
3840         btrfs_block_rsv_release(root, rsv, (u64)-1);
3841         kfree(rsv);
3842 }
3843
3844 static inline int __block_rsv_add(struct btrfs_root *root,
3845                                   struct btrfs_block_rsv *block_rsv,
3846                                   u64 num_bytes, int flush)
3847 {
3848         int ret;
3849
3850         if (num_bytes == 0)
3851                 return 0;
3852
3853         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3854         if (!ret) {
3855                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3856                 return 0;
3857         }
3858
3859         return ret;
3860 }
3861
3862 int btrfs_block_rsv_add(struct btrfs_root *root,
3863                         struct btrfs_block_rsv *block_rsv,
3864                         u64 num_bytes)
3865 {
3866         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3867 }
3868
3869 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3870                                 struct btrfs_block_rsv *block_rsv,
3871                                 u64 num_bytes)
3872 {
3873         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3874 }
3875
3876 int btrfs_block_rsv_check(struct btrfs_root *root,
3877                           struct btrfs_block_rsv *block_rsv, int min_factor)
3878 {
3879         u64 num_bytes = 0;
3880         int ret = -ENOSPC;
3881
3882         if (!block_rsv)
3883                 return 0;
3884
3885         spin_lock(&block_rsv->lock);
3886         num_bytes = div_factor(block_rsv->size, min_factor);
3887         if (block_rsv->reserved >= num_bytes)
3888                 ret = 0;
3889         spin_unlock(&block_rsv->lock);
3890
3891         return ret;
3892 }
3893
3894 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
3895                                            struct btrfs_block_rsv *block_rsv,
3896                                            u64 min_reserved, int flush)
3897 {
3898         u64 num_bytes = 0;
3899         int ret = -ENOSPC;
3900
3901         if (!block_rsv)
3902                 return 0;
3903
3904         spin_lock(&block_rsv->lock);
3905         num_bytes = min_reserved;
3906         if (block_rsv->reserved >= num_bytes)
3907                 ret = 0;
3908         else
3909                 num_bytes -= block_rsv->reserved;
3910         spin_unlock(&block_rsv->lock);
3911
3912         if (!ret)
3913                 return 0;
3914
3915         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3916         if (!ret) {
3917                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3918                 return 0;
3919         }
3920
3921         return ret;
3922 }
3923
3924 int btrfs_block_rsv_refill(struct btrfs_root *root,
3925                            struct btrfs_block_rsv *block_rsv,
3926                            u64 min_reserved)
3927 {
3928         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
3929 }
3930
3931 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
3932                                    struct btrfs_block_rsv *block_rsv,
3933                                    u64 min_reserved)
3934 {
3935         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
3936 }
3937
3938 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3939                             struct btrfs_block_rsv *dst_rsv,
3940                             u64 num_bytes)
3941 {
3942         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3943 }
3944
3945 void btrfs_block_rsv_release(struct btrfs_root *root,
3946                              struct btrfs_block_rsv *block_rsv,
3947                              u64 num_bytes)
3948 {
3949         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3950         if (global_rsv->full || global_rsv == block_rsv ||
3951             block_rsv->space_info != global_rsv->space_info)
3952                 global_rsv = NULL;
3953         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3954 }
3955
3956 /*
3957  * helper to calculate size of global block reservation.
3958  * the desired value is sum of space used by extent tree,
3959  * checksum tree and root tree
3960  */
3961 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3962 {
3963         struct btrfs_space_info *sinfo;
3964         u64 num_bytes;
3965         u64 meta_used;
3966         u64 data_used;
3967         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3968
3969         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3970         spin_lock(&sinfo->lock);
3971         data_used = sinfo->bytes_used;
3972         spin_unlock(&sinfo->lock);
3973
3974         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3975         spin_lock(&sinfo->lock);
3976         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3977                 data_used = 0;
3978         meta_used = sinfo->bytes_used;
3979         spin_unlock(&sinfo->lock);
3980
3981         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3982                     csum_size * 2;
3983         num_bytes += div64_u64(data_used + meta_used, 50);
3984
3985         if (num_bytes * 3 > meta_used)
3986                 num_bytes = div64_u64(meta_used, 3);
3987
3988         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3989 }
3990
3991 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3992 {
3993         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3994         struct btrfs_space_info *sinfo = block_rsv->space_info;
3995         u64 num_bytes;
3996
3997         num_bytes = calc_global_metadata_size(fs_info);
3998
3999         spin_lock(&block_rsv->lock);
4000         spin_lock(&sinfo->lock);
4001
4002         block_rsv->size = num_bytes;
4003
4004         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4005                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4006                     sinfo->bytes_may_use;
4007
4008         if (sinfo->total_bytes > num_bytes) {
4009                 num_bytes = sinfo->total_bytes - num_bytes;
4010                 block_rsv->reserved += num_bytes;
4011                 sinfo->bytes_may_use += num_bytes;
4012         }
4013
4014         if (block_rsv->reserved >= block_rsv->size) {
4015                 num_bytes = block_rsv->reserved - block_rsv->size;
4016                 sinfo->bytes_may_use -= num_bytes;
4017                 sinfo->reservation_progress++;
4018                 block_rsv->reserved = block_rsv->size;
4019                 block_rsv->full = 1;
4020         }
4021
4022         spin_unlock(&sinfo->lock);
4023         spin_unlock(&block_rsv->lock);
4024 }
4025
4026 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4027 {
4028         struct btrfs_space_info *space_info;
4029
4030         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4031         fs_info->chunk_block_rsv.space_info = space_info;
4032
4033         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4034         fs_info->global_block_rsv.space_info = space_info;
4035         fs_info->delalloc_block_rsv.space_info = space_info;
4036         fs_info->trans_block_rsv.space_info = space_info;
4037         fs_info->empty_block_rsv.space_info = space_info;
4038         fs_info->delayed_block_rsv.space_info = space_info;
4039
4040         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4041         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4042         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4043         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4044         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4045
4046         update_global_block_rsv(fs_info);
4047 }
4048
4049 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4050 {
4051         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
4052         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4053         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4054         WARN_ON(fs_info->trans_block_rsv.size > 0);
4055         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4056         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4057         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4058         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4059         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4060 }
4061
4062 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4063                                   struct btrfs_root *root)
4064 {
4065         if (!trans->bytes_reserved)
4066                 return;
4067
4068         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4069         trans->bytes_reserved = 0;
4070 }
4071
4072 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4073                                   struct inode *inode)
4074 {
4075         struct btrfs_root *root = BTRFS_I(inode)->root;
4076         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4077         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4078
4079         /*
4080          * We need to hold space in order to delete our orphan item once we've
4081          * added it, so this takes the reservation so we can release it later
4082          * when we are truly done with the orphan item.
4083          */
4084         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4085         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4086 }
4087
4088 void btrfs_orphan_release_metadata(struct inode *inode)
4089 {
4090         struct btrfs_root *root = BTRFS_I(inode)->root;
4091         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4092         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4093 }
4094
4095 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4096                                 struct btrfs_pending_snapshot *pending)
4097 {
4098         struct btrfs_root *root = pending->root;
4099         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4100         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4101         /*
4102          * two for root back/forward refs, two for directory entries
4103          * and one for root of the snapshot.
4104          */
4105         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4106         dst_rsv->space_info = src_rsv->space_info;
4107         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4108 }
4109
4110 /**
4111  * drop_outstanding_extent - drop an outstanding extent
4112  * @inode: the inode we're dropping the extent for
4113  *
4114  * This is called when we are freeing up an outstanding extent, either called
4115  * after an error or after an extent is written.  This will return the number of
4116  * reserved extents that need to be freed.  This must be called with
4117  * BTRFS_I(inode)->lock held.
4118  */
4119 static unsigned drop_outstanding_extent(struct inode *inode)
4120 {
4121         unsigned drop_inode_space = 0;
4122         unsigned dropped_extents = 0;
4123
4124         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4125         BTRFS_I(inode)->outstanding_extents--;
4126
4127         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4128             BTRFS_I(inode)->delalloc_meta_reserved) {
4129                 drop_inode_space = 1;
4130                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4131         }
4132
4133         /*
4134          * If we have more or the same amount of outsanding extents than we have
4135          * reserved then we need to leave the reserved extents count alone.
4136          */
4137         if (BTRFS_I(inode)->outstanding_extents >=
4138             BTRFS_I(inode)->reserved_extents)
4139                 return drop_inode_space;
4140
4141         dropped_extents = BTRFS_I(inode)->reserved_extents -
4142                 BTRFS_I(inode)->outstanding_extents;
4143         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4144         return dropped_extents + drop_inode_space;
4145 }
4146
4147 /**
4148  * calc_csum_metadata_size - return the amount of metada space that must be
4149  *      reserved/free'd for the given bytes.
4150  * @inode: the inode we're manipulating
4151  * @num_bytes: the number of bytes in question
4152  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4153  *
4154  * This adjusts the number of csum_bytes in the inode and then returns the
4155  * correct amount of metadata that must either be reserved or freed.  We
4156  * calculate how many checksums we can fit into one leaf and then divide the
4157  * number of bytes that will need to be checksumed by this value to figure out
4158  * how many checksums will be required.  If we are adding bytes then the number
4159  * may go up and we will return the number of additional bytes that must be
4160  * reserved.  If it is going down we will return the number of bytes that must
4161  * be freed.
4162  *
4163  * This must be called with BTRFS_I(inode)->lock held.
4164  */
4165 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4166                                    int reserve)
4167 {
4168         struct btrfs_root *root = BTRFS_I(inode)->root;
4169         u64 csum_size;
4170         int num_csums_per_leaf;
4171         int num_csums;
4172         int old_csums;
4173
4174         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4175             BTRFS_I(inode)->csum_bytes == 0)
4176                 return 0;
4177
4178         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4179         if (reserve)
4180                 BTRFS_I(inode)->csum_bytes += num_bytes;
4181         else
4182                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4183         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4184         num_csums_per_leaf = (int)div64_u64(csum_size,
4185                                             sizeof(struct btrfs_csum_item) +
4186                                             sizeof(struct btrfs_disk_key));
4187         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4188         num_csums = num_csums + num_csums_per_leaf - 1;
4189         num_csums = num_csums / num_csums_per_leaf;
4190
4191         old_csums = old_csums + num_csums_per_leaf - 1;
4192         old_csums = old_csums / num_csums_per_leaf;
4193
4194         /* No change, no need to reserve more */
4195         if (old_csums == num_csums)
4196                 return 0;
4197
4198         if (reserve)
4199                 return btrfs_calc_trans_metadata_size(root,
4200                                                       num_csums - old_csums);
4201
4202         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4203 }
4204
4205 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4206 {
4207         struct btrfs_root *root = BTRFS_I(inode)->root;
4208         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4209         u64 to_reserve = 0;
4210         u64 csum_bytes;
4211         unsigned nr_extents = 0;
4212         int extra_reserve = 0;
4213         int flush = 1;
4214         int ret;
4215
4216         /* Need to be holding the i_mutex here if we aren't free space cache */
4217         if (btrfs_is_free_space_inode(root, inode))
4218                 flush = 0;
4219         else
4220                 WARN_ON(!mutex_is_locked(&inode->i_mutex));
4221
4222         if (flush && btrfs_transaction_in_commit(root->fs_info))
4223                 schedule_timeout(1);
4224
4225         num_bytes = ALIGN(num_bytes, root->sectorsize);
4226
4227         spin_lock(&BTRFS_I(inode)->lock);
4228         BTRFS_I(inode)->outstanding_extents++;
4229
4230         if (BTRFS_I(inode)->outstanding_extents >
4231             BTRFS_I(inode)->reserved_extents)
4232                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4233                         BTRFS_I(inode)->reserved_extents;
4234
4235         /*
4236          * Add an item to reserve for updating the inode when we complete the
4237          * delalloc io.
4238          */
4239         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4240                 nr_extents++;
4241                 extra_reserve = 1;
4242         }
4243
4244         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4245         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4246         csum_bytes = BTRFS_I(inode)->csum_bytes;
4247         spin_unlock(&BTRFS_I(inode)->lock);
4248
4249         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4250         if (ret) {
4251                 u64 to_free = 0;
4252                 unsigned dropped;
4253
4254                 spin_lock(&BTRFS_I(inode)->lock);
4255                 dropped = drop_outstanding_extent(inode);
4256                 /*
4257                  * If the inodes csum_bytes is the same as the original
4258                  * csum_bytes then we know we haven't raced with any free()ers
4259                  * so we can just reduce our inodes csum bytes and carry on.
4260                  * Otherwise we have to do the normal free thing to account for
4261                  * the case that the free side didn't free up its reserve
4262                  * because of this outstanding reservation.
4263                  */
4264                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4265                         calc_csum_metadata_size(inode, num_bytes, 0);
4266                 else
4267                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4268                 spin_unlock(&BTRFS_I(inode)->lock);
4269                 if (dropped)
4270                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4271
4272                 if (to_free)
4273                         btrfs_block_rsv_release(root, block_rsv, to_free);
4274                 return ret;
4275         }
4276
4277         spin_lock(&BTRFS_I(inode)->lock);
4278         if (extra_reserve) {
4279                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4280                 nr_extents--;
4281         }
4282         BTRFS_I(inode)->reserved_extents += nr_extents;
4283         spin_unlock(&BTRFS_I(inode)->lock);
4284
4285         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4286
4287         return 0;
4288 }
4289
4290 /**
4291  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4292  * @inode: the inode to release the reservation for
4293  * @num_bytes: the number of bytes we're releasing
4294  *
4295  * This will release the metadata reservation for an inode.  This can be called
4296  * once we complete IO for a given set of bytes to release their metadata
4297  * reservations.
4298  */
4299 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4300 {
4301         struct btrfs_root *root = BTRFS_I(inode)->root;
4302         u64 to_free = 0;
4303         unsigned dropped;
4304
4305         num_bytes = ALIGN(num_bytes, root->sectorsize);
4306         spin_lock(&BTRFS_I(inode)->lock);
4307         dropped = drop_outstanding_extent(inode);
4308
4309         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4310         spin_unlock(&BTRFS_I(inode)->lock);
4311         if (dropped > 0)
4312                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4313
4314         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4315                                 to_free);
4316 }
4317
4318 /**
4319  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4320  * @inode: inode we're writing to
4321  * @num_bytes: the number of bytes we want to allocate
4322  *
4323  * This will do the following things
4324  *
4325  * o reserve space in the data space info for num_bytes
4326  * o reserve space in the metadata space info based on number of outstanding
4327  *   extents and how much csums will be needed
4328  * o add to the inodes ->delalloc_bytes
4329  * o add it to the fs_info's delalloc inodes list.
4330  *
4331  * This will return 0 for success and -ENOSPC if there is no space left.
4332  */
4333 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4334 {
4335         int ret;
4336
4337         ret = btrfs_check_data_free_space(inode, num_bytes);
4338         if (ret)
4339                 return ret;
4340
4341         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4342         if (ret) {
4343                 btrfs_free_reserved_data_space(inode, num_bytes);
4344                 return ret;
4345         }
4346
4347         return 0;
4348 }
4349
4350 /**
4351  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4352  * @inode: inode we're releasing space for
4353  * @num_bytes: the number of bytes we want to free up
4354  *
4355  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4356  * called in the case that we don't need the metadata AND data reservations
4357  * anymore.  So if there is an error or we insert an inline extent.
4358  *
4359  * This function will release the metadata space that was not used and will
4360  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4361  * list if there are no delalloc bytes left.
4362  */
4363 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4364 {
4365         btrfs_delalloc_release_metadata(inode, num_bytes);
4366         btrfs_free_reserved_data_space(inode, num_bytes);
4367 }
4368
4369 static int update_block_group(struct btrfs_trans_handle *trans,
4370                               struct btrfs_root *root,
4371                               u64 bytenr, u64 num_bytes, int alloc)
4372 {
4373         struct btrfs_block_group_cache *cache = NULL;
4374         struct btrfs_fs_info *info = root->fs_info;
4375         u64 total = num_bytes;
4376         u64 old_val;
4377         u64 byte_in_group;
4378         int factor;
4379
4380         /* block accounting for super block */
4381         spin_lock(&info->delalloc_lock);
4382         old_val = btrfs_super_bytes_used(info->super_copy);
4383         if (alloc)
4384                 old_val += num_bytes;
4385         else
4386                 old_val -= num_bytes;
4387         btrfs_set_super_bytes_used(info->super_copy, old_val);
4388         spin_unlock(&info->delalloc_lock);
4389
4390         while (total) {
4391                 cache = btrfs_lookup_block_group(info, bytenr);
4392                 if (!cache)
4393                         return -1;
4394                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4395                                     BTRFS_BLOCK_GROUP_RAID1 |
4396                                     BTRFS_BLOCK_GROUP_RAID10))
4397                         factor = 2;
4398                 else
4399                         factor = 1;
4400                 /*
4401                  * If this block group has free space cache written out, we
4402                  * need to make sure to load it if we are removing space.  This
4403                  * is because we need the unpinning stage to actually add the
4404                  * space back to the block group, otherwise we will leak space.
4405                  */
4406                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4407                         cache_block_group(cache, trans, NULL, 1);
4408
4409                 byte_in_group = bytenr - cache->key.objectid;
4410                 WARN_ON(byte_in_group > cache->key.offset);
4411
4412                 spin_lock(&cache->space_info->lock);
4413                 spin_lock(&cache->lock);
4414
4415                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4416                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4417                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4418
4419                 cache->dirty = 1;
4420                 old_val = btrfs_block_group_used(&cache->item);
4421                 num_bytes = min(total, cache->key.offset - byte_in_group);
4422                 if (alloc) {
4423                         old_val += num_bytes;
4424                         btrfs_set_block_group_used(&cache->item, old_val);
4425                         cache->reserved -= num_bytes;
4426                         cache->space_info->bytes_reserved -= num_bytes;
4427                         cache->space_info->bytes_used += num_bytes;
4428                         cache->space_info->disk_used += num_bytes * factor;
4429                         spin_unlock(&cache->lock);
4430                         spin_unlock(&cache->space_info->lock);
4431                 } else {
4432                         old_val -= num_bytes;
4433                         btrfs_set_block_group_used(&cache->item, old_val);
4434                         cache->pinned += num_bytes;
4435                         cache->space_info->bytes_pinned += num_bytes;
4436                         cache->space_info->bytes_used -= num_bytes;
4437                         cache->space_info->disk_used -= num_bytes * factor;
4438                         spin_unlock(&cache->lock);
4439                         spin_unlock(&cache->space_info->lock);
4440
4441                         set_extent_dirty(info->pinned_extents,
4442                                          bytenr, bytenr + num_bytes - 1,
4443                                          GFP_NOFS | __GFP_NOFAIL);
4444                 }
4445                 btrfs_put_block_group(cache);
4446                 total -= num_bytes;
4447                 bytenr += num_bytes;
4448         }
4449         return 0;
4450 }
4451
4452 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4453 {
4454         struct btrfs_block_group_cache *cache;
4455         u64 bytenr;
4456
4457         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4458         if (!cache)
4459                 return 0;
4460
4461         bytenr = cache->key.objectid;
4462         btrfs_put_block_group(cache);
4463
4464         return bytenr;
4465 }
4466
4467 static int pin_down_extent(struct btrfs_root *root,
4468                            struct btrfs_block_group_cache *cache,
4469                            u64 bytenr, u64 num_bytes, int reserved)
4470 {
4471         spin_lock(&cache->space_info->lock);
4472         spin_lock(&cache->lock);
4473         cache->pinned += num_bytes;
4474         cache->space_info->bytes_pinned += num_bytes;
4475         if (reserved) {
4476                 cache->reserved -= num_bytes;
4477                 cache->space_info->bytes_reserved -= num_bytes;
4478         }
4479         spin_unlock(&cache->lock);
4480         spin_unlock(&cache->space_info->lock);
4481
4482         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4483                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4484         return 0;
4485 }
4486
4487 /*
4488  * this function must be called within transaction
4489  */
4490 int btrfs_pin_extent(struct btrfs_root *root,
4491                      u64 bytenr, u64 num_bytes, int reserved)
4492 {
4493         struct btrfs_block_group_cache *cache;
4494
4495         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4496         BUG_ON(!cache);
4497
4498         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4499
4500         btrfs_put_block_group(cache);
4501         return 0;
4502 }
4503
4504 /*
4505  * this function must be called within transaction
4506  */
4507 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4508                                     struct btrfs_root *root,
4509                                     u64 bytenr, u64 num_bytes)
4510 {
4511         struct btrfs_block_group_cache *cache;
4512
4513         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4514         BUG_ON(!cache);
4515
4516         /*
4517          * pull in the free space cache (if any) so that our pin
4518          * removes the free space from the cache.  We have load_only set
4519          * to one because the slow code to read in the free extents does check
4520          * the pinned extents.
4521          */
4522         cache_block_group(cache, trans, root, 1);
4523
4524         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4525
4526         /* remove us from the free space cache (if we're there at all) */
4527         btrfs_remove_free_space(cache, bytenr, num_bytes);
4528         btrfs_put_block_group(cache);
4529         return 0;
4530 }
4531
4532 /**
4533  * btrfs_update_reserved_bytes - update the block_group and space info counters
4534  * @cache:      The cache we are manipulating
4535  * @num_bytes:  The number of bytes in question
4536  * @reserve:    One of the reservation enums
4537  *
4538  * This is called by the allocator when it reserves space, or by somebody who is
4539  * freeing space that was never actually used on disk.  For example if you
4540  * reserve some space for a new leaf in transaction A and before transaction A
4541  * commits you free that leaf, you call this with reserve set to 0 in order to
4542  * clear the reservation.
4543  *
4544  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4545  * ENOSPC accounting.  For data we handle the reservation through clearing the
4546  * delalloc bits in the io_tree.  We have to do this since we could end up
4547  * allocating less disk space for the amount of data we have reserved in the
4548  * case of compression.
4549  *
4550  * If this is a reservation and the block group has become read only we cannot
4551  * make the reservation and return -EAGAIN, otherwise this function always
4552  * succeeds.
4553  */
4554 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4555                                        u64 num_bytes, int reserve)
4556 {
4557         struct btrfs_space_info *space_info = cache->space_info;
4558         int ret = 0;
4559         spin_lock(&space_info->lock);
4560         spin_lock(&cache->lock);
4561         if (reserve != RESERVE_FREE) {
4562                 if (cache->ro) {
4563                         ret = -EAGAIN;
4564                 } else {
4565                         cache->reserved += num_bytes;
4566                         space_info->bytes_reserved += num_bytes;
4567                         if (reserve == RESERVE_ALLOC) {
4568                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4569                                 space_info->bytes_may_use -= num_bytes;
4570                         }
4571                 }
4572         } else {
4573                 if (cache->ro)
4574                         space_info->bytes_readonly += num_bytes;
4575                 cache->reserved -= num_bytes;
4576                 space_info->bytes_reserved -= num_bytes;
4577                 space_info->reservation_progress++;
4578         }
4579         spin_unlock(&cache->lock);
4580         spin_unlock(&space_info->lock);
4581         return ret;
4582 }
4583
4584 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4585                                 struct btrfs_root *root)
4586 {
4587         struct btrfs_fs_info *fs_info = root->fs_info;
4588         struct btrfs_caching_control *next;
4589         struct btrfs_caching_control *caching_ctl;
4590         struct btrfs_block_group_cache *cache;
4591
4592         down_write(&fs_info->extent_commit_sem);
4593
4594         list_for_each_entry_safe(caching_ctl, next,
4595                                  &fs_info->caching_block_groups, list) {
4596                 cache = caching_ctl->block_group;
4597                 if (block_group_cache_done(cache)) {
4598                         cache->last_byte_to_unpin = (u64)-1;
4599                         list_del_init(&caching_ctl->list);
4600                         put_caching_control(caching_ctl);
4601                 } else {
4602                         cache->last_byte_to_unpin = caching_ctl->progress;
4603                 }
4604         }
4605
4606         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4607                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4608         else
4609                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4610
4611         up_write(&fs_info->extent_commit_sem);
4612
4613         update_global_block_rsv(fs_info);
4614         return 0;
4615 }
4616
4617 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4618 {
4619         struct btrfs_fs_info *fs_info = root->fs_info;
4620         struct btrfs_block_group_cache *cache = NULL;
4621         u64 len;
4622
4623         while (start <= end) {
4624                 if (!cache ||
4625                     start >= cache->key.objectid + cache->key.offset) {
4626                         if (cache)
4627                                 btrfs_put_block_group(cache);
4628                         cache = btrfs_lookup_block_group(fs_info, start);
4629                         BUG_ON(!cache);
4630                 }
4631
4632                 len = cache->key.objectid + cache->key.offset - start;
4633                 len = min(len, end + 1 - start);
4634
4635                 if (start < cache->last_byte_to_unpin) {
4636                         len = min(len, cache->last_byte_to_unpin - start);
4637                         btrfs_add_free_space(cache, start, len);
4638                 }
4639
4640                 start += len;
4641
4642                 spin_lock(&cache->space_info->lock);
4643                 spin_lock(&cache->lock);
4644                 cache->pinned -= len;
4645                 cache->space_info->bytes_pinned -= len;
4646                 if (cache->ro)
4647                         cache->space_info->bytes_readonly += len;
4648                 spin_unlock(&cache->lock);
4649                 spin_unlock(&cache->space_info->lock);
4650         }
4651
4652         if (cache)
4653                 btrfs_put_block_group(cache);
4654         return 0;
4655 }
4656
4657 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4658                                struct btrfs_root *root)
4659 {
4660         struct btrfs_fs_info *fs_info = root->fs_info;
4661         struct extent_io_tree *unpin;
4662         u64 start;
4663         u64 end;
4664         int ret;
4665
4666         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4667                 unpin = &fs_info->freed_extents[1];
4668         else
4669                 unpin = &fs_info->freed_extents[0];
4670
4671         while (1) {
4672                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4673                                             EXTENT_DIRTY);
4674                 if (ret)
4675                         break;
4676
4677                 if (btrfs_test_opt(root, DISCARD))
4678                         ret = btrfs_discard_extent(root, start,
4679                                                    end + 1 - start, NULL);
4680
4681                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4682                 unpin_extent_range(root, start, end);
4683                 cond_resched();
4684         }
4685
4686         return 0;
4687 }
4688
4689 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4690                                 struct btrfs_root *root,
4691                                 u64 bytenr, u64 num_bytes, u64 parent,
4692                                 u64 root_objectid, u64 owner_objectid,
4693                                 u64 owner_offset, int refs_to_drop,
4694                                 struct btrfs_delayed_extent_op *extent_op)
4695 {
4696         struct btrfs_key key;
4697         struct btrfs_path *path;
4698         struct btrfs_fs_info *info = root->fs_info;
4699         struct btrfs_root *extent_root = info->extent_root;
4700         struct extent_buffer *leaf;
4701         struct btrfs_extent_item *ei;
4702         struct btrfs_extent_inline_ref *iref;
4703         int ret;
4704         int is_data;
4705         int extent_slot = 0;
4706         int found_extent = 0;
4707         int num_to_del = 1;
4708         u32 item_size;
4709         u64 refs;
4710
4711         path = btrfs_alloc_path();
4712         if (!path)
4713                 return -ENOMEM;
4714
4715         path->reada = 1;
4716         path->leave_spinning = 1;
4717
4718         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4719         BUG_ON(!is_data && refs_to_drop != 1);
4720
4721         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4722                                     bytenr, num_bytes, parent,
4723                                     root_objectid, owner_objectid,
4724                                     owner_offset);
4725         if (ret == 0) {
4726                 extent_slot = path->slots[0];
4727                 while (extent_slot >= 0) {
4728                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4729                                               extent_slot);
4730                         if (key.objectid != bytenr)
4731                                 break;
4732                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4733                             key.offset == num_bytes) {
4734                                 found_extent = 1;
4735                                 break;
4736                         }
4737                         if (path->slots[0] - extent_slot > 5)
4738                                 break;
4739                         extent_slot--;
4740                 }
4741 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4742                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4743                 if (found_extent && item_size < sizeof(*ei))
4744                         found_extent = 0;
4745 #endif
4746                 if (!found_extent) {
4747                         BUG_ON(iref);
4748                         ret = remove_extent_backref(trans, extent_root, path,
4749                                                     NULL, refs_to_drop,
4750                                                     is_data);
4751                         BUG_ON(ret);
4752                         btrfs_release_path(path);
4753                         path->leave_spinning = 1;
4754
4755                         key.objectid = bytenr;
4756                         key.type = BTRFS_EXTENT_ITEM_KEY;
4757                         key.offset = num_bytes;
4758
4759                         ret = btrfs_search_slot(trans, extent_root,
4760                                                 &key, path, -1, 1);
4761                         if (ret) {
4762                                 printk(KERN_ERR "umm, got %d back from search"
4763                                        ", was looking for %llu\n", ret,
4764                                        (unsigned long long)bytenr);
4765                                 if (ret > 0)
4766                                         btrfs_print_leaf(extent_root,
4767                                                          path->nodes[0]);
4768                         }
4769                         BUG_ON(ret);
4770                         extent_slot = path->slots[0];
4771                 }
4772         } else {
4773                 btrfs_print_leaf(extent_root, path->nodes[0]);
4774                 WARN_ON(1);
4775                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4776                        "parent %llu root %llu  owner %llu offset %llu\n",
4777                        (unsigned long long)bytenr,
4778                        (unsigned long long)parent,
4779                        (unsigned long long)root_objectid,
4780                        (unsigned long long)owner_objectid,
4781                        (unsigned long long)owner_offset);
4782         }
4783
4784         leaf = path->nodes[0];
4785         item_size = btrfs_item_size_nr(leaf, extent_slot);
4786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4787         if (item_size < sizeof(*ei)) {
4788                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4789                 ret = convert_extent_item_v0(trans, extent_root, path,
4790                                              owner_objectid, 0);
4791                 BUG_ON(ret < 0);
4792
4793                 btrfs_release_path(path);
4794                 path->leave_spinning = 1;
4795
4796                 key.objectid = bytenr;
4797                 key.type = BTRFS_EXTENT_ITEM_KEY;
4798                 key.offset = num_bytes;
4799
4800                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4801                                         -1, 1);
4802                 if (ret) {
4803                         printk(KERN_ERR "umm, got %d back from search"
4804                                ", was looking for %llu\n", ret,
4805                                (unsigned long long)bytenr);
4806                         btrfs_print_leaf(extent_root, path->nodes[0]);
4807                 }
4808                 BUG_ON(ret);
4809                 extent_slot = path->slots[0];
4810                 leaf = path->nodes[0];
4811                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4812         }
4813 #endif
4814         BUG_ON(item_size < sizeof(*ei));
4815         ei = btrfs_item_ptr(leaf, extent_slot,
4816                             struct btrfs_extent_item);
4817         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4818                 struct btrfs_tree_block_info *bi;
4819                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4820                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4821                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4822         }
4823
4824         refs = btrfs_extent_refs(leaf, ei);
4825         BUG_ON(refs < refs_to_drop);
4826         refs -= refs_to_drop;
4827
4828         if (refs > 0) {
4829                 if (extent_op)
4830                         __run_delayed_extent_op(extent_op, leaf, ei);
4831                 /*
4832                  * In the case of inline back ref, reference count will
4833                  * be updated by remove_extent_backref
4834                  */
4835                 if (iref) {
4836                         BUG_ON(!found_extent);
4837                 } else {
4838                         btrfs_set_extent_refs(leaf, ei, refs);
4839                         btrfs_mark_buffer_dirty(leaf);
4840                 }
4841                 if (found_extent) {
4842                         ret = remove_extent_backref(trans, extent_root, path,
4843                                                     iref, refs_to_drop,
4844                                                     is_data);
4845                         BUG_ON(ret);
4846                 }
4847         } else {
4848                 if (found_extent) {
4849                         BUG_ON(is_data && refs_to_drop !=
4850                                extent_data_ref_count(root, path, iref));
4851                         if (iref) {
4852                                 BUG_ON(path->slots[0] != extent_slot);
4853                         } else {
4854                                 BUG_ON(path->slots[0] != extent_slot + 1);
4855                                 path->slots[0] = extent_slot;
4856                                 num_to_del = 2;
4857                         }
4858                 }
4859
4860                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4861                                       num_to_del);
4862                 BUG_ON(ret);
4863                 btrfs_release_path(path);
4864
4865                 if (is_data) {
4866                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4867                         BUG_ON(ret);
4868                 } else {
4869                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4870                              bytenr >> PAGE_CACHE_SHIFT,
4871                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4872                 }
4873
4874                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4875                 BUG_ON(ret);
4876         }
4877         btrfs_free_path(path);
4878         return ret;
4879 }
4880
4881 /*
4882  * when we free an block, it is possible (and likely) that we free the last
4883  * delayed ref for that extent as well.  This searches the delayed ref tree for
4884  * a given extent, and if there are no other delayed refs to be processed, it
4885  * removes it from the tree.
4886  */
4887 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4888                                       struct btrfs_root *root, u64 bytenr)
4889 {
4890         struct btrfs_delayed_ref_head *head;
4891         struct btrfs_delayed_ref_root *delayed_refs;
4892         struct btrfs_delayed_ref_node *ref;
4893         struct rb_node *node;
4894         int ret = 0;
4895
4896         delayed_refs = &trans->transaction->delayed_refs;
4897         spin_lock(&delayed_refs->lock);
4898         head = btrfs_find_delayed_ref_head(trans, bytenr);
4899         if (!head)
4900                 goto out;
4901
4902         node = rb_prev(&head->node.rb_node);
4903         if (!node)
4904                 goto out;
4905
4906         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4907
4908         /* there are still entries for this ref, we can't drop it */
4909         if (ref->bytenr == bytenr)
4910                 goto out;
4911
4912         if (head->extent_op) {
4913                 if (!head->must_insert_reserved)
4914                         goto out;
4915                 kfree(head->extent_op);
4916                 head->extent_op = NULL;
4917         }
4918
4919         /*
4920          * waiting for the lock here would deadlock.  If someone else has it
4921          * locked they are already in the process of dropping it anyway
4922          */
4923         if (!mutex_trylock(&head->mutex))
4924                 goto out;
4925
4926         /*
4927          * at this point we have a head with no other entries.  Go
4928          * ahead and process it.
4929          */
4930         head->node.in_tree = 0;
4931         rb_erase(&head->node.rb_node, &delayed_refs->root);
4932
4933         delayed_refs->num_entries--;
4934
4935         /*
4936          * we don't take a ref on the node because we're removing it from the
4937          * tree, so we just steal the ref the tree was holding.
4938          */
4939         delayed_refs->num_heads--;
4940         if (list_empty(&head->cluster))
4941                 delayed_refs->num_heads_ready--;
4942
4943         list_del_init(&head->cluster);
4944         spin_unlock(&delayed_refs->lock);
4945
4946         BUG_ON(head->extent_op);
4947         if (head->must_insert_reserved)
4948                 ret = 1;
4949
4950         mutex_unlock(&head->mutex);
4951         btrfs_put_delayed_ref(&head->node);
4952         return ret;
4953 out:
4954         spin_unlock(&delayed_refs->lock);
4955         return 0;
4956 }
4957
4958 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4959                            struct btrfs_root *root,
4960                            struct extent_buffer *buf,
4961                            u64 parent, int last_ref)
4962 {
4963         struct btrfs_block_group_cache *cache = NULL;
4964         int ret;
4965
4966         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4967                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4968                                                 parent, root->root_key.objectid,
4969                                                 btrfs_header_level(buf),
4970                                                 BTRFS_DROP_DELAYED_REF, NULL);
4971                 BUG_ON(ret);
4972         }
4973
4974         if (!last_ref)
4975                 return;
4976
4977         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4978
4979         if (btrfs_header_generation(buf) == trans->transid) {
4980                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4981                         ret = check_ref_cleanup(trans, root, buf->start);
4982                         if (!ret)
4983                                 goto out;
4984                 }
4985
4986                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4987                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4988                         goto out;
4989                 }
4990
4991                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4992
4993                 btrfs_add_free_space(cache, buf->start, buf->len);
4994                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4995         }
4996 out:
4997         /*
4998          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4999          * anymore.
5000          */
5001         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5002         btrfs_put_block_group(cache);
5003 }
5004
5005 int btrfs_free_extent(struct btrfs_trans_handle *trans,
5006                       struct btrfs_root *root,
5007                       u64 bytenr, u64 num_bytes, u64 parent,
5008                       u64 root_objectid, u64 owner, u64 offset)
5009 {
5010         int ret;
5011
5012         /*
5013          * tree log blocks never actually go into the extent allocation
5014          * tree, just update pinning info and exit early.
5015          */
5016         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5017                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5018                 /* unlocks the pinned mutex */
5019                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5020                 ret = 0;
5021         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5022                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
5023                                         parent, root_objectid, (int)owner,
5024                                         BTRFS_DROP_DELAYED_REF, NULL);
5025                 BUG_ON(ret);
5026         } else {
5027                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
5028                                         parent, root_objectid, owner,
5029                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
5030                 BUG_ON(ret);
5031         }
5032         return ret;
5033 }
5034
5035 static u64 stripe_align(struct btrfs_root *root, u64 val)
5036 {
5037         u64 mask = ((u64)root->stripesize - 1);
5038         u64 ret = (val + mask) & ~mask;
5039         return ret;
5040 }
5041
5042 /*
5043  * when we wait for progress in the block group caching, its because
5044  * our allocation attempt failed at least once.  So, we must sleep
5045  * and let some progress happen before we try again.
5046  *
5047  * This function will sleep at least once waiting for new free space to
5048  * show up, and then it will check the block group free space numbers
5049  * for our min num_bytes.  Another option is to have it go ahead
5050  * and look in the rbtree for a free extent of a given size, but this
5051  * is a good start.
5052  */
5053 static noinline int
5054 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5055                                 u64 num_bytes)
5056 {
5057         struct btrfs_caching_control *caching_ctl;
5058         DEFINE_WAIT(wait);
5059
5060         caching_ctl = get_caching_control(cache);
5061         if (!caching_ctl)
5062                 return 0;
5063
5064         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5065                    (cache->free_space_ctl->free_space >= num_bytes));
5066
5067         put_caching_control(caching_ctl);
5068         return 0;
5069 }
5070
5071 static noinline int
5072 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5073 {
5074         struct btrfs_caching_control *caching_ctl;
5075         DEFINE_WAIT(wait);
5076
5077         caching_ctl = get_caching_control(cache);
5078         if (!caching_ctl)
5079                 return 0;
5080
5081         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5082
5083         put_caching_control(caching_ctl);
5084         return 0;
5085 }
5086
5087 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5088 {
5089         int index;
5090         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5091                 index = 0;
5092         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5093                 index = 1;
5094         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5095                 index = 2;
5096         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5097                 index = 3;
5098         else
5099                 index = 4;
5100         return index;
5101 }
5102
5103 enum btrfs_loop_type {
5104         LOOP_FIND_IDEAL = 0,
5105         LOOP_CACHING_NOWAIT = 1,
5106         LOOP_CACHING_WAIT = 2,
5107         LOOP_ALLOC_CHUNK = 3,
5108         LOOP_NO_EMPTY_SIZE = 4,
5109 };
5110
5111 /*
5112  * walks the btree of allocated extents and find a hole of a given size.
5113  * The key ins is changed to record the hole:
5114  * ins->objectid == block start
5115  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5116  * ins->offset == number of blocks
5117  * Any available blocks before search_start are skipped.
5118  */
5119 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5120                                      struct btrfs_root *orig_root,
5121                                      u64 num_bytes, u64 empty_size,
5122                                      u64 search_start, u64 search_end,
5123                                      u64 hint_byte, struct btrfs_key *ins,
5124                                      u64 data)
5125 {
5126         int ret = 0;
5127         struct btrfs_root *root = orig_root->fs_info->extent_root;
5128         struct btrfs_free_cluster *last_ptr = NULL;
5129         struct btrfs_block_group_cache *block_group = NULL;
5130         struct btrfs_block_group_cache *used_block_group;
5131         int empty_cluster = 2 * 1024 * 1024;
5132         int allowed_chunk_alloc = 0;
5133         int done_chunk_alloc = 0;
5134         struct btrfs_space_info *space_info;
5135         int loop = 0;
5136         int index = 0;
5137         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5138                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5139         bool found_uncached_bg = false;
5140         bool failed_cluster_refill = false;
5141         bool failed_alloc = false;
5142         bool use_cluster = true;
5143         bool have_caching_bg = false;
5144         u64 ideal_cache_percent = 0;
5145         u64 ideal_cache_offset = 0;
5146
5147         WARN_ON(num_bytes < root->sectorsize);
5148         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5149         ins->objectid = 0;
5150         ins->offset = 0;
5151
5152         space_info = __find_space_info(root->fs_info, data);
5153         if (!space_info) {
5154                 printk(KERN_ERR "No space info for %llu\n", data);
5155                 return -ENOSPC;
5156         }
5157
5158         /*
5159          * If the space info is for both data and metadata it means we have a
5160          * small filesystem and we can't use the clustering stuff.
5161          */
5162         if (btrfs_mixed_space_info(space_info))
5163                 use_cluster = false;
5164
5165         if (orig_root->ref_cows || empty_size)
5166                 allowed_chunk_alloc = 1;
5167
5168         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5169                 last_ptr = &root->fs_info->meta_alloc_cluster;
5170                 if (!btrfs_test_opt(root, SSD))
5171                         empty_cluster = 64 * 1024;
5172         }
5173
5174         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5175             btrfs_test_opt(root, SSD)) {
5176                 last_ptr = &root->fs_info->data_alloc_cluster;
5177         }
5178
5179         if (last_ptr) {
5180                 spin_lock(&last_ptr->lock);
5181                 if (last_ptr->block_group)
5182                         hint_byte = last_ptr->window_start;
5183                 spin_unlock(&last_ptr->lock);
5184         }
5185
5186         search_start = max(search_start, first_logical_byte(root, 0));
5187         search_start = max(search_start, hint_byte);
5188
5189         if (!last_ptr)
5190                 empty_cluster = 0;
5191
5192         if (search_start == hint_byte) {
5193 ideal_cache:
5194                 block_group = btrfs_lookup_block_group(root->fs_info,
5195                                                        search_start);
5196                 used_block_group = block_group;
5197                 /*
5198                  * we don't want to use the block group if it doesn't match our
5199                  * allocation bits, or if its not cached.
5200                  *
5201                  * However if we are re-searching with an ideal block group
5202                  * picked out then we don't care that the block group is cached.
5203                  */
5204                 if (block_group && block_group_bits(block_group, data) &&
5205                     (block_group->cached != BTRFS_CACHE_NO ||
5206                      search_start == ideal_cache_offset)) {
5207                         down_read(&space_info->groups_sem);
5208                         if (list_empty(&block_group->list) ||
5209                             block_group->ro) {
5210                                 /*
5211                                  * someone is removing this block group,
5212                                  * we can't jump into the have_block_group
5213                                  * target because our list pointers are not
5214                                  * valid
5215                                  */
5216                                 btrfs_put_block_group(block_group);
5217                                 up_read(&space_info->groups_sem);
5218                         } else {
5219                                 index = get_block_group_index(block_group);
5220                                 goto have_block_group;
5221                         }
5222                 } else if (block_group) {
5223                         btrfs_put_block_group(block_group);
5224                 }
5225         }
5226 search:
5227         have_caching_bg = false;
5228         down_read(&space_info->groups_sem);
5229         list_for_each_entry(block_group, &space_info->block_groups[index],
5230                             list) {
5231                 u64 offset;
5232                 int cached;
5233
5234                 used_block_group = block_group;
5235                 btrfs_get_block_group(block_group);
5236                 search_start = block_group->key.objectid;
5237
5238                 /*
5239                  * this can happen if we end up cycling through all the
5240                  * raid types, but we want to make sure we only allocate
5241                  * for the proper type.
5242                  */
5243                 if (!block_group_bits(block_group, data)) {
5244                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5245                                 BTRFS_BLOCK_GROUP_RAID1 |
5246                                 BTRFS_BLOCK_GROUP_RAID10;
5247
5248                         /*
5249                          * if they asked for extra copies and this block group
5250                          * doesn't provide them, bail.  This does allow us to
5251                          * fill raid0 from raid1.
5252                          */
5253                         if ((data & extra) && !(block_group->flags & extra))
5254                                 goto loop;
5255                 }
5256
5257 have_block_group:
5258                 cached = block_group_cache_done(block_group);
5259                 if (unlikely(!cached)) {
5260                         u64 free_percent;
5261
5262                         found_uncached_bg = true;
5263                         ret = cache_block_group(block_group, trans,
5264                                                 orig_root, 1);
5265                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5266                                 goto alloc;
5267
5268                         free_percent = btrfs_block_group_used(&block_group->item);
5269                         free_percent *= 100;
5270                         free_percent = div64_u64(free_percent,
5271                                                  block_group->key.offset);
5272                         free_percent = 100 - free_percent;
5273                         if (free_percent > ideal_cache_percent &&
5274                             likely(!block_group->ro)) {
5275                                 ideal_cache_offset = block_group->key.objectid;
5276                                 ideal_cache_percent = free_percent;
5277                         }
5278
5279                         /*
5280                          * The caching workers are limited to 2 threads, so we
5281                          * can queue as much work as we care to.
5282                          */
5283                         if (loop > LOOP_FIND_IDEAL) {
5284                                 ret = cache_block_group(block_group, trans,
5285                                                         orig_root, 0);
5286                                 BUG_ON(ret);
5287                         }
5288
5289                         /*
5290                          * If loop is set for cached only, try the next block
5291                          * group.
5292                          */
5293                         if (loop == LOOP_FIND_IDEAL)
5294                                 goto loop;
5295                 }
5296
5297 alloc:
5298                 if (unlikely(block_group->ro))
5299                         goto loop;
5300
5301                 spin_lock(&block_group->free_space_ctl->tree_lock);
5302                 if (cached &&
5303                     block_group->free_space_ctl->free_space <
5304                     num_bytes + empty_cluster + empty_size) {
5305                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5306                         goto loop;
5307                 }
5308                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5309
5310                 /*
5311                  * Ok we want to try and use the cluster allocator, so
5312                  * lets look there
5313                  */
5314                 if (last_ptr) {
5315                         /*
5316                          * the refill lock keeps out other
5317                          * people trying to start a new cluster
5318                          */
5319                         spin_lock(&last_ptr->refill_lock);
5320                         used_block_group = last_ptr->block_group;
5321                         if (used_block_group != block_group &&
5322                             (!used_block_group ||
5323                              used_block_group->ro ||
5324                              !block_group_bits(used_block_group, data))) {
5325                                 used_block_group = block_group;
5326                                 goto refill_cluster;
5327                         }
5328
5329                         if (used_block_group != block_group)
5330                                 btrfs_get_block_group(used_block_group);
5331
5332                         offset = btrfs_alloc_from_cluster(used_block_group,
5333                           last_ptr, num_bytes, used_block_group->key.objectid);
5334                         if (offset) {
5335                                 /* we have a block, we're done */
5336                                 spin_unlock(&last_ptr->refill_lock);
5337                                 goto checks;
5338                         }
5339
5340                         WARN_ON(last_ptr->block_group != used_block_group);
5341                         if (used_block_group != block_group) {
5342                                 btrfs_put_block_group(used_block_group);
5343                                 used_block_group = block_group;
5344                         }
5345 refill_cluster:
5346                         BUG_ON(used_block_group != block_group);
5347                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5348                          * set up a new clusters, so lets just skip it
5349                          * and let the allocator find whatever block
5350                          * it can find.  If we reach this point, we
5351                          * will have tried the cluster allocator
5352                          * plenty of times and not have found
5353                          * anything, so we are likely way too
5354                          * fragmented for the clustering stuff to find
5355                          * anything.  */
5356                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5357                                 spin_unlock(&last_ptr->refill_lock);
5358                                 goto unclustered_alloc;
5359                         }
5360
5361                         /*
5362                          * this cluster didn't work out, free it and
5363                          * start over
5364                          */
5365                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5366
5367                         /* allocate a cluster in this block group */
5368                         ret = btrfs_find_space_cluster(trans, root,
5369                                                block_group, last_ptr,
5370                                                search_start, num_bytes,
5371                                                empty_cluster + empty_size);
5372                         if (ret == 0) {
5373                                 /*
5374                                  * now pull our allocation out of this
5375                                  * cluster
5376                                  */
5377                                 offset = btrfs_alloc_from_cluster(block_group,
5378                                                   last_ptr, num_bytes,
5379                                                   search_start);
5380                                 if (offset) {
5381                                         /* we found one, proceed */
5382                                         spin_unlock(&last_ptr->refill_lock);
5383                                         goto checks;
5384                                 }
5385                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5386                                    && !failed_cluster_refill) {
5387                                 spin_unlock(&last_ptr->refill_lock);
5388
5389                                 failed_cluster_refill = true;
5390                                 wait_block_group_cache_progress(block_group,
5391                                        num_bytes + empty_cluster + empty_size);
5392                                 goto have_block_group;
5393                         }
5394
5395                         /*
5396                          * at this point we either didn't find a cluster
5397                          * or we weren't able to allocate a block from our
5398                          * cluster.  Free the cluster we've been trying
5399                          * to use, and go to the next block group
5400                          */
5401                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5402                         spin_unlock(&last_ptr->refill_lock);
5403                         goto loop;
5404                 }
5405
5406 unclustered_alloc:
5407                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5408                                                     num_bytes, empty_size);
5409                 /*
5410                  * If we didn't find a chunk, and we haven't failed on this
5411                  * block group before, and this block group is in the middle of
5412                  * caching and we are ok with waiting, then go ahead and wait
5413                  * for progress to be made, and set failed_alloc to true.
5414                  *
5415                  * If failed_alloc is true then we've already waited on this
5416                  * block group once and should move on to the next block group.
5417                  */
5418                 if (!offset && !failed_alloc && !cached &&
5419                     loop > LOOP_CACHING_NOWAIT) {
5420                         wait_block_group_cache_progress(block_group,
5421                                                 num_bytes + empty_size);
5422                         failed_alloc = true;
5423                         goto have_block_group;
5424                 } else if (!offset) {
5425                         if (!cached)
5426                                 have_caching_bg = true;
5427                         goto loop;
5428                 }
5429 checks:
5430                 search_start = stripe_align(root, offset);
5431                 /* move on to the next group */
5432                 if (search_start + num_bytes >= search_end) {
5433                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5434                         goto loop;
5435                 }
5436
5437                 /* move on to the next group */
5438                 if (search_start + num_bytes >
5439                     used_block_group->key.objectid + used_block_group->key.offset) {
5440                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5441                         goto loop;
5442                 }
5443
5444                 ins->objectid = search_start;
5445                 ins->offset = num_bytes;
5446
5447                 if (offset < search_start)
5448                         btrfs_add_free_space(used_block_group, offset,
5449                                              search_start - offset);
5450                 BUG_ON(offset > search_start);
5451
5452                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5453                                                   alloc_type);
5454                 if (ret == -EAGAIN) {
5455                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5456                         goto loop;
5457                 }
5458
5459                 /* we are all good, lets return */
5460                 ins->objectid = search_start;
5461                 ins->offset = num_bytes;
5462
5463                 if (offset < search_start)
5464                         btrfs_add_free_space(used_block_group, offset,
5465                                              search_start - offset);
5466                 BUG_ON(offset > search_start);
5467                 if (used_block_group != block_group)
5468                         btrfs_put_block_group(used_block_group);
5469                 btrfs_put_block_group(block_group);
5470                 break;
5471 loop:
5472                 failed_cluster_refill = false;
5473                 failed_alloc = false;
5474                 BUG_ON(index != get_block_group_index(block_group));
5475                 if (used_block_group != block_group)
5476                         btrfs_put_block_group(used_block_group);
5477                 btrfs_put_block_group(block_group);
5478         }
5479         up_read(&space_info->groups_sem);
5480
5481         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5482                 goto search;
5483
5484         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5485                 goto search;
5486
5487         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5488          *                      for them to make caching progress.  Also
5489          *                      determine the best possible bg to cache
5490          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5491          *                      caching kthreads as we move along
5492          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5493          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5494          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5495          *                      again
5496          */
5497         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5498                 index = 0;
5499                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5500                         found_uncached_bg = false;
5501                         loop++;
5502                         if (!ideal_cache_percent)
5503                                 goto search;
5504
5505                         /*
5506                          * 1 of the following 2 things have happened so far
5507                          *
5508                          * 1) We found an ideal block group for caching that
5509                          * is mostly full and will cache quickly, so we might
5510                          * as well wait for it.
5511                          *
5512                          * 2) We searched for cached only and we didn't find
5513                          * anything, and we didn't start any caching kthreads
5514                          * either, so chances are we will loop through and
5515                          * start a couple caching kthreads, and then come back
5516                          * around and just wait for them.  This will be slower
5517                          * because we will have 2 caching kthreads reading at
5518                          * the same time when we could have just started one
5519                          * and waited for it to get far enough to give us an
5520                          * allocation, so go ahead and go to the wait caching
5521                          * loop.
5522                          */
5523                         loop = LOOP_CACHING_WAIT;
5524                         search_start = ideal_cache_offset;
5525                         ideal_cache_percent = 0;
5526                         goto ideal_cache;
5527                 } else if (loop == LOOP_FIND_IDEAL) {
5528                         /*
5529                          * Didn't find a uncached bg, wait on anything we find
5530                          * next.
5531                          */
5532                         loop = LOOP_CACHING_WAIT;
5533                         goto search;
5534                 }
5535
5536                 loop++;
5537
5538                 if (loop == LOOP_ALLOC_CHUNK) {
5539                        if (allowed_chunk_alloc) {
5540                                 ret = do_chunk_alloc(trans, root, num_bytes +
5541                                                      2 * 1024 * 1024, data,
5542                                                      CHUNK_ALLOC_LIMITED);
5543                                 allowed_chunk_alloc = 0;
5544                                 if (ret == 1)
5545                                         done_chunk_alloc = 1;
5546                         } else if (!done_chunk_alloc &&
5547                                    space_info->force_alloc ==
5548                                    CHUNK_ALLOC_NO_FORCE) {
5549                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5550                         }
5551
5552                        /*
5553                         * We didn't allocate a chunk, go ahead and drop the
5554                         * empty size and loop again.
5555                         */
5556                        if (!done_chunk_alloc)
5557                                loop = LOOP_NO_EMPTY_SIZE;
5558                 }
5559
5560                 if (loop == LOOP_NO_EMPTY_SIZE) {
5561                         empty_size = 0;
5562                         empty_cluster = 0;
5563                 }
5564
5565                 goto search;
5566         } else if (!ins->objectid) {
5567                 ret = -ENOSPC;
5568         } else if (ins->objectid) {
5569                 ret = 0;
5570         }
5571
5572         return ret;
5573 }
5574
5575 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5576                             int dump_block_groups)
5577 {
5578         struct btrfs_block_group_cache *cache;
5579         int index = 0;
5580
5581         spin_lock(&info->lock);
5582         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5583                (unsigned long long)info->flags,
5584                (unsigned long long)(info->total_bytes - info->bytes_used -
5585                                     info->bytes_pinned - info->bytes_reserved -
5586                                     info->bytes_readonly),
5587                (info->full) ? "" : "not ");
5588         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5589                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5590                (unsigned long long)info->total_bytes,
5591                (unsigned long long)info->bytes_used,
5592                (unsigned long long)info->bytes_pinned,
5593                (unsigned long long)info->bytes_reserved,
5594                (unsigned long long)info->bytes_may_use,
5595                (unsigned long long)info->bytes_readonly);
5596         spin_unlock(&info->lock);
5597
5598         if (!dump_block_groups)
5599                 return;
5600
5601         down_read(&info->groups_sem);
5602 again:
5603         list_for_each_entry(cache, &info->block_groups[index], list) {
5604                 spin_lock(&cache->lock);
5605                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5606                        "%llu pinned %llu reserved\n",
5607                        (unsigned long long)cache->key.objectid,
5608                        (unsigned long long)cache->key.offset,
5609                        (unsigned long long)btrfs_block_group_used(&cache->item),
5610                        (unsigned long long)cache->pinned,
5611                        (unsigned long long)cache->reserved);
5612                 btrfs_dump_free_space(cache, bytes);
5613                 spin_unlock(&cache->lock);
5614         }
5615         if (++index < BTRFS_NR_RAID_TYPES)
5616                 goto again;
5617         up_read(&info->groups_sem);
5618 }
5619
5620 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5621                          struct btrfs_root *root,
5622                          u64 num_bytes, u64 min_alloc_size,
5623                          u64 empty_size, u64 hint_byte,
5624       &