Merge branch 'btrfs-3.0' into for-linus
[linux-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include "compat.h"
27 #include "hash.h"
28 #include "ctree.h"
29 #include "disk-io.h"
30 #include "print-tree.h"
31 #include "transaction.h"
32 #include "volumes.h"
33 #include "locking.h"
34 #include "free-space-cache.h"
35
36 /* control flags for do_chunk_alloc's force field
37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
38  * if we really need one.
39  *
40  * CHUNK_ALLOC_FORCE means it must try to allocate one
41  *
42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
43  * if we have very few chunks already allocated.  This is
44  * used as part of the clustering code to help make sure
45  * we have a good pool of storage to cluster in, without
46  * filling the FS with empty chunks
47  *
48  */
49 enum {
50         CHUNK_ALLOC_NO_FORCE = 0,
51         CHUNK_ALLOC_FORCE = 1,
52         CHUNK_ALLOC_LIMITED = 2,
53 };
54
55 static int update_block_group(struct btrfs_trans_handle *trans,
56                               struct btrfs_root *root,
57                               u64 bytenr, u64 num_bytes, int alloc);
58 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
59                                 struct btrfs_root *root,
60                                 u64 bytenr, u64 num_bytes, u64 parent,
61                                 u64 root_objectid, u64 owner_objectid,
62                                 u64 owner_offset, int refs_to_drop,
63                                 struct btrfs_delayed_extent_op *extra_op);
64 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
65                                     struct extent_buffer *leaf,
66                                     struct btrfs_extent_item *ei);
67 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
68                                       struct btrfs_root *root,
69                                       u64 parent, u64 root_objectid,
70                                       u64 flags, u64 owner, u64 offset,
71                                       struct btrfs_key *ins, int ref_mod);
72 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
73                                      struct btrfs_root *root,
74                                      u64 parent, u64 root_objectid,
75                                      u64 flags, struct btrfs_disk_key *key,
76                                      int level, struct btrfs_key *ins);
77 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
78                           struct btrfs_root *extent_root, u64 alloc_bytes,
79                           u64 flags, int force);
80 static int find_next_key(struct btrfs_path *path, int level,
81                          struct btrfs_key *key);
82 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
83                             int dump_block_groups);
84
85 static noinline int
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
87 {
88         smp_mb();
89         return cache->cached == BTRFS_CACHE_FINISHED;
90 }
91
92 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
93 {
94         return (cache->flags & bits) == bits;
95 }
96
97 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
98 {
99         atomic_inc(&cache->count);
100 }
101
102 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
103 {
104         if (atomic_dec_and_test(&cache->count)) {
105                 WARN_ON(cache->pinned > 0);
106                 WARN_ON(cache->reserved > 0);
107                 WARN_ON(cache->reserved_pinned > 0);
108                 kfree(cache->free_space_ctl);
109                 kfree(cache);
110         }
111 }
112
113 /*
114  * this adds the block group to the fs_info rb tree for the block group
115  * cache
116  */
117 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
118                                 struct btrfs_block_group_cache *block_group)
119 {
120         struct rb_node **p;
121         struct rb_node *parent = NULL;
122         struct btrfs_block_group_cache *cache;
123
124         spin_lock(&info->block_group_cache_lock);
125         p = &info->block_group_cache_tree.rb_node;
126
127         while (*p) {
128                 parent = *p;
129                 cache = rb_entry(parent, struct btrfs_block_group_cache,
130                                  cache_node);
131                 if (block_group->key.objectid < cache->key.objectid) {
132                         p = &(*p)->rb_left;
133                 } else if (block_group->key.objectid > cache->key.objectid) {
134                         p = &(*p)->rb_right;
135                 } else {
136                         spin_unlock(&info->block_group_cache_lock);
137                         return -EEXIST;
138                 }
139         }
140
141         rb_link_node(&block_group->cache_node, parent, p);
142         rb_insert_color(&block_group->cache_node,
143                         &info->block_group_cache_tree);
144         spin_unlock(&info->block_group_cache_lock);
145
146         return 0;
147 }
148
149 /*
150  * This will return the block group at or after bytenr if contains is 0, else
151  * it will return the block group that contains the bytenr
152  */
153 static struct btrfs_block_group_cache *
154 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
155                               int contains)
156 {
157         struct btrfs_block_group_cache *cache, *ret = NULL;
158         struct rb_node *n;
159         u64 end, start;
160
161         spin_lock(&info->block_group_cache_lock);
162         n = info->block_group_cache_tree.rb_node;
163
164         while (n) {
165                 cache = rb_entry(n, struct btrfs_block_group_cache,
166                                  cache_node);
167                 end = cache->key.objectid + cache->key.offset - 1;
168                 start = cache->key.objectid;
169
170                 if (bytenr < start) {
171                         if (!contains && (!ret || start < ret->key.objectid))
172                                 ret = cache;
173                         n = n->rb_left;
174                 } else if (bytenr > start) {
175                         if (contains && bytenr <= end) {
176                                 ret = cache;
177                                 break;
178                         }
179                         n = n->rb_right;
180                 } else {
181                         ret = cache;
182                         break;
183                 }
184         }
185         if (ret)
186                 btrfs_get_block_group(ret);
187         spin_unlock(&info->block_group_cache_lock);
188
189         return ret;
190 }
191
192 static int add_excluded_extent(struct btrfs_root *root,
193                                u64 start, u64 num_bytes)
194 {
195         u64 end = start + num_bytes - 1;
196         set_extent_bits(&root->fs_info->freed_extents[0],
197                         start, end, EXTENT_UPTODATE, GFP_NOFS);
198         set_extent_bits(&root->fs_info->freed_extents[1],
199                         start, end, EXTENT_UPTODATE, GFP_NOFS);
200         return 0;
201 }
202
203 static void free_excluded_extents(struct btrfs_root *root,
204                                   struct btrfs_block_group_cache *cache)
205 {
206         u64 start, end;
207
208         start = cache->key.objectid;
209         end = start + cache->key.offset - 1;
210
211         clear_extent_bits(&root->fs_info->freed_extents[0],
212                           start, end, EXTENT_UPTODATE, GFP_NOFS);
213         clear_extent_bits(&root->fs_info->freed_extents[1],
214                           start, end, EXTENT_UPTODATE, GFP_NOFS);
215 }
216
217 static int exclude_super_stripes(struct btrfs_root *root,
218                                  struct btrfs_block_group_cache *cache)
219 {
220         u64 bytenr;
221         u64 *logical;
222         int stripe_len;
223         int i, nr, ret;
224
225         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
226                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
227                 cache->bytes_super += stripe_len;
228                 ret = add_excluded_extent(root, cache->key.objectid,
229                                           stripe_len);
230                 BUG_ON(ret);
231         }
232
233         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
234                 bytenr = btrfs_sb_offset(i);
235                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
236                                        cache->key.objectid, bytenr,
237                                        0, &logical, &nr, &stripe_len);
238                 BUG_ON(ret);
239
240                 while (nr--) {
241                         cache->bytes_super += stripe_len;
242                         ret = add_excluded_extent(root, logical[nr],
243                                                   stripe_len);
244                         BUG_ON(ret);
245                 }
246
247                 kfree(logical);
248         }
249         return 0;
250 }
251
252 static struct btrfs_caching_control *
253 get_caching_control(struct btrfs_block_group_cache *cache)
254 {
255         struct btrfs_caching_control *ctl;
256
257         spin_lock(&cache->lock);
258         if (cache->cached != BTRFS_CACHE_STARTED) {
259                 spin_unlock(&cache->lock);
260                 return NULL;
261         }
262
263         /* We're loading it the fast way, so we don't have a caching_ctl. */
264         if (!cache->caching_ctl) {
265                 spin_unlock(&cache->lock);
266                 return NULL;
267         }
268
269         ctl = cache->caching_ctl;
270         atomic_inc(&ctl->count);
271         spin_unlock(&cache->lock);
272         return ctl;
273 }
274
275 static void put_caching_control(struct btrfs_caching_control *ctl)
276 {
277         if (atomic_dec_and_test(&ctl->count))
278                 kfree(ctl);
279 }
280
281 /*
282  * this is only called by cache_block_group, since we could have freed extents
283  * we need to check the pinned_extents for any extents that can't be used yet
284  * since their free space will be released as soon as the transaction commits.
285  */
286 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
287                               struct btrfs_fs_info *info, u64 start, u64 end)
288 {
289         u64 extent_start, extent_end, size, total_added = 0;
290         int ret;
291
292         while (start < end) {
293                 ret = find_first_extent_bit(info->pinned_extents, start,
294                                             &extent_start, &extent_end,
295                                             EXTENT_DIRTY | EXTENT_UPTODATE);
296                 if (ret)
297                         break;
298
299                 if (extent_start <= start) {
300                         start = extent_end + 1;
301                 } else if (extent_start > start && extent_start < end) {
302                         size = extent_start - start;
303                         total_added += size;
304                         ret = btrfs_add_free_space(block_group, start,
305                                                    size);
306                         BUG_ON(ret);
307                         start = extent_end + 1;
308                 } else {
309                         break;
310                 }
311         }
312
313         if (start < end) {
314                 size = end - start;
315                 total_added += size;
316                 ret = btrfs_add_free_space(block_group, start, size);
317                 BUG_ON(ret);
318         }
319
320         return total_added;
321 }
322
323 static noinline void caching_thread(struct btrfs_work *work)
324 {
325         struct btrfs_block_group_cache *block_group;
326         struct btrfs_fs_info *fs_info;
327         struct btrfs_caching_control *caching_ctl;
328         struct btrfs_root *extent_root;
329         struct btrfs_path *path;
330         struct extent_buffer *leaf;
331         struct btrfs_key key;
332         u64 total_found = 0;
333         u64 last = 0;
334         u32 nritems;
335         int ret = 0;
336
337         caching_ctl = container_of(work, struct btrfs_caching_control, work);
338         block_group = caching_ctl->block_group;
339         fs_info = block_group->fs_info;
340         extent_root = fs_info->extent_root;
341
342         path = btrfs_alloc_path();
343         if (!path)
344                 goto out;
345
346         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
347
348         /*
349          * We don't want to deadlock with somebody trying to allocate a new
350          * extent for the extent root while also trying to search the extent
351          * root to add free space.  So we skip locking and search the commit
352          * root, since its read-only
353          */
354         path->skip_locking = 1;
355         path->search_commit_root = 1;
356         path->reada = 1;
357
358         key.objectid = last;
359         key.offset = 0;
360         key.type = BTRFS_EXTENT_ITEM_KEY;
361 again:
362         mutex_lock(&caching_ctl->mutex);
363         /* need to make sure the commit_root doesn't disappear */
364         down_read(&fs_info->extent_commit_sem);
365
366         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
367         if (ret < 0)
368                 goto err;
369
370         leaf = path->nodes[0];
371         nritems = btrfs_header_nritems(leaf);
372
373         while (1) {
374                 if (btrfs_fs_closing(fs_info) > 1) {
375                         last = (u64)-1;
376                         break;
377                 }
378
379                 if (path->slots[0] < nritems) {
380                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
381                 } else {
382                         ret = find_next_key(path, 0, &key);
383                         if (ret)
384                                 break;
385
386                         if (need_resched() ||
387                             btrfs_next_leaf(extent_root, path)) {
388                                 caching_ctl->progress = last;
389                                 btrfs_release_path(path);
390                                 up_read(&fs_info->extent_commit_sem);
391                                 mutex_unlock(&caching_ctl->mutex);
392                                 cond_resched();
393                                 goto again;
394                         }
395                         leaf = path->nodes[0];
396                         nritems = btrfs_header_nritems(leaf);
397                         continue;
398                 }
399
400                 if (key.objectid < block_group->key.objectid) {
401                         path->slots[0]++;
402                         continue;
403                 }
404
405                 if (key.objectid >= block_group->key.objectid +
406                     block_group->key.offset)
407                         break;
408
409                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
410                         total_found += add_new_free_space(block_group,
411                                                           fs_info, last,
412                                                           key.objectid);
413                         last = key.objectid + key.offset;
414
415                         if (total_found > (1024 * 1024 * 2)) {
416                                 total_found = 0;
417                                 wake_up(&caching_ctl->wait);
418                         }
419                 }
420                 path->slots[0]++;
421         }
422         ret = 0;
423
424         total_found += add_new_free_space(block_group, fs_info, last,
425                                           block_group->key.objectid +
426                                           block_group->key.offset);
427         caching_ctl->progress = (u64)-1;
428
429         spin_lock(&block_group->lock);
430         block_group->caching_ctl = NULL;
431         block_group->cached = BTRFS_CACHE_FINISHED;
432         spin_unlock(&block_group->lock);
433
434 err:
435         btrfs_free_path(path);
436         up_read(&fs_info->extent_commit_sem);
437
438         free_excluded_extents(extent_root, block_group);
439
440         mutex_unlock(&caching_ctl->mutex);
441 out:
442         wake_up(&caching_ctl->wait);
443
444         put_caching_control(caching_ctl);
445         btrfs_put_block_group(block_group);
446 }
447
448 static int cache_block_group(struct btrfs_block_group_cache *cache,
449                              struct btrfs_trans_handle *trans,
450                              struct btrfs_root *root,
451                              int load_cache_only)
452 {
453         struct btrfs_fs_info *fs_info = cache->fs_info;
454         struct btrfs_caching_control *caching_ctl;
455         int ret = 0;
456
457         smp_mb();
458         if (cache->cached != BTRFS_CACHE_NO)
459                 return 0;
460
461         /*
462          * We can't do the read from on-disk cache during a commit since we need
463          * to have the normal tree locking.  Also if we are currently trying to
464          * allocate blocks for the tree root we can't do the fast caching since
465          * we likely hold important locks.
466          */
467         if (trans && (!trans->transaction->in_commit) &&
468             (root && root != root->fs_info->tree_root)) {
469                 spin_lock(&cache->lock);
470                 if (cache->cached != BTRFS_CACHE_NO) {
471                         spin_unlock(&cache->lock);
472                         return 0;
473                 }
474                 cache->cached = BTRFS_CACHE_STARTED;
475                 spin_unlock(&cache->lock);
476
477                 ret = load_free_space_cache(fs_info, cache);
478
479                 spin_lock(&cache->lock);
480                 if (ret == 1) {
481                         cache->cached = BTRFS_CACHE_FINISHED;
482                         cache->last_byte_to_unpin = (u64)-1;
483                 } else {
484                         cache->cached = BTRFS_CACHE_NO;
485                 }
486                 spin_unlock(&cache->lock);
487                 if (ret == 1) {
488                         free_excluded_extents(fs_info->extent_root, cache);
489                         return 0;
490                 }
491         }
492
493         if (load_cache_only)
494                 return 0;
495
496         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
497         BUG_ON(!caching_ctl);
498
499         INIT_LIST_HEAD(&caching_ctl->list);
500         mutex_init(&caching_ctl->mutex);
501         init_waitqueue_head(&caching_ctl->wait);
502         caching_ctl->block_group = cache;
503         caching_ctl->progress = cache->key.objectid;
504         /* one for caching kthread, one for caching block group list */
505         atomic_set(&caching_ctl->count, 2);
506         caching_ctl->work.func = caching_thread;
507
508         spin_lock(&cache->lock);
509         if (cache->cached != BTRFS_CACHE_NO) {
510                 spin_unlock(&cache->lock);
511                 kfree(caching_ctl);
512                 return 0;
513         }
514         cache->caching_ctl = caching_ctl;
515         cache->cached = BTRFS_CACHE_STARTED;
516         spin_unlock(&cache->lock);
517
518         down_write(&fs_info->extent_commit_sem);
519         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
520         up_write(&fs_info->extent_commit_sem);
521
522         btrfs_get_block_group(cache);
523
524         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
525
526         return ret;
527 }
528
529 /*
530  * return the block group that starts at or after bytenr
531  */
532 static struct btrfs_block_group_cache *
533 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
534 {
535         struct btrfs_block_group_cache *cache;
536
537         cache = block_group_cache_tree_search(info, bytenr, 0);
538
539         return cache;
540 }
541
542 /*
543  * return the block group that contains the given bytenr
544  */
545 struct btrfs_block_group_cache *btrfs_lookup_block_group(
546                                                  struct btrfs_fs_info *info,
547                                                  u64 bytenr)
548 {
549         struct btrfs_block_group_cache *cache;
550
551         cache = block_group_cache_tree_search(info, bytenr, 1);
552
553         return cache;
554 }
555
556 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
557                                                   u64 flags)
558 {
559         struct list_head *head = &info->space_info;
560         struct btrfs_space_info *found;
561
562         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
563                  BTRFS_BLOCK_GROUP_METADATA;
564
565         rcu_read_lock();
566         list_for_each_entry_rcu(found, head, list) {
567                 if (found->flags & flags) {
568                         rcu_read_unlock();
569                         return found;
570                 }
571         }
572         rcu_read_unlock();
573         return NULL;
574 }
575
576 /*
577  * after adding space to the filesystem, we need to clear the full flags
578  * on all the space infos.
579  */
580 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
581 {
582         struct list_head *head = &info->space_info;
583         struct btrfs_space_info *found;
584
585         rcu_read_lock();
586         list_for_each_entry_rcu(found, head, list)
587                 found->full = 0;
588         rcu_read_unlock();
589 }
590
591 static u64 div_factor(u64 num, int factor)
592 {
593         if (factor == 10)
594                 return num;
595         num *= factor;
596         do_div(num, 10);
597         return num;
598 }
599
600 static u64 div_factor_fine(u64 num, int factor)
601 {
602         if (factor == 100)
603                 return num;
604         num *= factor;
605         do_div(num, 100);
606         return num;
607 }
608
609 u64 btrfs_find_block_group(struct btrfs_root *root,
610                            u64 search_start, u64 search_hint, int owner)
611 {
612         struct btrfs_block_group_cache *cache;
613         u64 used;
614         u64 last = max(search_hint, search_start);
615         u64 group_start = 0;
616         int full_search = 0;
617         int factor = 9;
618         int wrapped = 0;
619 again:
620         while (1) {
621                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
622                 if (!cache)
623                         break;
624
625                 spin_lock(&cache->lock);
626                 last = cache->key.objectid + cache->key.offset;
627                 used = btrfs_block_group_used(&cache->item);
628
629                 if ((full_search || !cache->ro) &&
630                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
631                         if (used + cache->pinned + cache->reserved <
632                             div_factor(cache->key.offset, factor)) {
633                                 group_start = cache->key.objectid;
634                                 spin_unlock(&cache->lock);
635                                 btrfs_put_block_group(cache);
636                                 goto found;
637                         }
638                 }
639                 spin_unlock(&cache->lock);
640                 btrfs_put_block_group(cache);
641                 cond_resched();
642         }
643         if (!wrapped) {
644                 last = search_start;
645                 wrapped = 1;
646                 goto again;
647         }
648         if (!full_search && factor < 10) {
649                 last = search_start;
650                 full_search = 1;
651                 factor = 10;
652                 goto again;
653         }
654 found:
655         return group_start;
656 }
657
658 /* simple helper to search for an existing extent at a given offset */
659 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
660 {
661         int ret;
662         struct btrfs_key key;
663         struct btrfs_path *path;
664
665         path = btrfs_alloc_path();
666         if (!path)
667                 return -ENOMEM;
668
669         key.objectid = start;
670         key.offset = len;
671         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
672         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
673                                 0, 0);
674         btrfs_free_path(path);
675         return ret;
676 }
677
678 /*
679  * helper function to lookup reference count and flags of extent.
680  *
681  * the head node for delayed ref is used to store the sum of all the
682  * reference count modifications queued up in the rbtree. the head
683  * node may also store the extent flags to set. This way you can check
684  * to see what the reference count and extent flags would be if all of
685  * the delayed refs are not processed.
686  */
687 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
688                              struct btrfs_root *root, u64 bytenr,
689                              u64 num_bytes, u64 *refs, u64 *flags)
690 {
691         struct btrfs_delayed_ref_head *head;
692         struct btrfs_delayed_ref_root *delayed_refs;
693         struct btrfs_path *path;
694         struct btrfs_extent_item *ei;
695         struct extent_buffer *leaf;
696         struct btrfs_key key;
697         u32 item_size;
698         u64 num_refs;
699         u64 extent_flags;
700         int ret;
701
702         path = btrfs_alloc_path();
703         if (!path)
704                 return -ENOMEM;
705
706         key.objectid = bytenr;
707         key.type = BTRFS_EXTENT_ITEM_KEY;
708         key.offset = num_bytes;
709         if (!trans) {
710                 path->skip_locking = 1;
711                 path->search_commit_root = 1;
712         }
713 again:
714         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
715                                 &key, path, 0, 0);
716         if (ret < 0)
717                 goto out_free;
718
719         if (ret == 0) {
720                 leaf = path->nodes[0];
721                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
722                 if (item_size >= sizeof(*ei)) {
723                         ei = btrfs_item_ptr(leaf, path->slots[0],
724                                             struct btrfs_extent_item);
725                         num_refs = btrfs_extent_refs(leaf, ei);
726                         extent_flags = btrfs_extent_flags(leaf, ei);
727                 } else {
728 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
729                         struct btrfs_extent_item_v0 *ei0;
730                         BUG_ON(item_size != sizeof(*ei0));
731                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
732                                              struct btrfs_extent_item_v0);
733                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
734                         /* FIXME: this isn't correct for data */
735                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
736 #else
737                         BUG();
738 #endif
739                 }
740                 BUG_ON(num_refs == 0);
741         } else {
742                 num_refs = 0;
743                 extent_flags = 0;
744                 ret = 0;
745         }
746
747         if (!trans)
748                 goto out;
749
750         delayed_refs = &trans->transaction->delayed_refs;
751         spin_lock(&delayed_refs->lock);
752         head = btrfs_find_delayed_ref_head(trans, bytenr);
753         if (head) {
754                 if (!mutex_trylock(&head->mutex)) {
755                         atomic_inc(&head->node.refs);
756                         spin_unlock(&delayed_refs->lock);
757
758                         btrfs_release_path(path);
759
760                         /*
761                          * Mutex was contended, block until it's released and try
762                          * again
763                          */
764                         mutex_lock(&head->mutex);
765                         mutex_unlock(&head->mutex);
766                         btrfs_put_delayed_ref(&head->node);
767                         goto again;
768                 }
769                 if (head->extent_op && head->extent_op->update_flags)
770                         extent_flags |= head->extent_op->flags_to_set;
771                 else
772                         BUG_ON(num_refs == 0);
773
774                 num_refs += head->node.ref_mod;
775                 mutex_unlock(&head->mutex);
776         }
777         spin_unlock(&delayed_refs->lock);
778 out:
779         WARN_ON(num_refs == 0);
780         if (refs)
781                 *refs = num_refs;
782         if (flags)
783                 *flags = extent_flags;
784 out_free:
785         btrfs_free_path(path);
786         return ret;
787 }
788
789 /*
790  * Back reference rules.  Back refs have three main goals:
791  *
792  * 1) differentiate between all holders of references to an extent so that
793  *    when a reference is dropped we can make sure it was a valid reference
794  *    before freeing the extent.
795  *
796  * 2) Provide enough information to quickly find the holders of an extent
797  *    if we notice a given block is corrupted or bad.
798  *
799  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
800  *    maintenance.  This is actually the same as #2, but with a slightly
801  *    different use case.
802  *
803  * There are two kinds of back refs. The implicit back refs is optimized
804  * for pointers in non-shared tree blocks. For a given pointer in a block,
805  * back refs of this kind provide information about the block's owner tree
806  * and the pointer's key. These information allow us to find the block by
807  * b-tree searching. The full back refs is for pointers in tree blocks not
808  * referenced by their owner trees. The location of tree block is recorded
809  * in the back refs. Actually the full back refs is generic, and can be
810  * used in all cases the implicit back refs is used. The major shortcoming
811  * of the full back refs is its overhead. Every time a tree block gets
812  * COWed, we have to update back refs entry for all pointers in it.
813  *
814  * For a newly allocated tree block, we use implicit back refs for
815  * pointers in it. This means most tree related operations only involve
816  * implicit back refs. For a tree block created in old transaction, the
817  * only way to drop a reference to it is COW it. So we can detect the
818  * event that tree block loses its owner tree's reference and do the
819  * back refs conversion.
820  *
821  * When a tree block is COW'd through a tree, there are four cases:
822  *
823  * The reference count of the block is one and the tree is the block's
824  * owner tree. Nothing to do in this case.
825  *
826  * The reference count of the block is one and the tree is not the
827  * block's owner tree. In this case, full back refs is used for pointers
828  * in the block. Remove these full back refs, add implicit back refs for
829  * every pointers in the new block.
830  *
831  * The reference count of the block is greater than one and the tree is
832  * the block's owner tree. In this case, implicit back refs is used for
833  * pointers in the block. Add full back refs for every pointers in the
834  * block, increase lower level extents' reference counts. The original
835  * implicit back refs are entailed to the new block.
836  *
837  * The reference count of the block is greater than one and the tree is
838  * not the block's owner tree. Add implicit back refs for every pointer in
839  * the new block, increase lower level extents' reference count.
840  *
841  * Back Reference Key composing:
842  *
843  * The key objectid corresponds to the first byte in the extent,
844  * The key type is used to differentiate between types of back refs.
845  * There are different meanings of the key offset for different types
846  * of back refs.
847  *
848  * File extents can be referenced by:
849  *
850  * - multiple snapshots, subvolumes, or different generations in one subvol
851  * - different files inside a single subvolume
852  * - different offsets inside a file (bookend extents in file.c)
853  *
854  * The extent ref structure for the implicit back refs has fields for:
855  *
856  * - Objectid of the subvolume root
857  * - objectid of the file holding the reference
858  * - original offset in the file
859  * - how many bookend extents
860  *
861  * The key offset for the implicit back refs is hash of the first
862  * three fields.
863  *
864  * The extent ref structure for the full back refs has field for:
865  *
866  * - number of pointers in the tree leaf
867  *
868  * The key offset for the implicit back refs is the first byte of
869  * the tree leaf
870  *
871  * When a file extent is allocated, The implicit back refs is used.
872  * the fields are filled in:
873  *
874  *     (root_key.objectid, inode objectid, offset in file, 1)
875  *
876  * When a file extent is removed file truncation, we find the
877  * corresponding implicit back refs and check the following fields:
878  *
879  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
880  *
881  * Btree extents can be referenced by:
882  *
883  * - Different subvolumes
884  *
885  * Both the implicit back refs and the full back refs for tree blocks
886  * only consist of key. The key offset for the implicit back refs is
887  * objectid of block's owner tree. The key offset for the full back refs
888  * is the first byte of parent block.
889  *
890  * When implicit back refs is used, information about the lowest key and
891  * level of the tree block are required. These information are stored in
892  * tree block info structure.
893  */
894
895 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
896 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
897                                   struct btrfs_root *root,
898                                   struct btrfs_path *path,
899                                   u64 owner, u32 extra_size)
900 {
901         struct btrfs_extent_item *item;
902         struct btrfs_extent_item_v0 *ei0;
903         struct btrfs_extent_ref_v0 *ref0;
904         struct btrfs_tree_block_info *bi;
905         struct extent_buffer *leaf;
906         struct btrfs_key key;
907         struct btrfs_key found_key;
908         u32 new_size = sizeof(*item);
909         u64 refs;
910         int ret;
911
912         leaf = path->nodes[0];
913         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
914
915         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
916         ei0 = btrfs_item_ptr(leaf, path->slots[0],
917                              struct btrfs_extent_item_v0);
918         refs = btrfs_extent_refs_v0(leaf, ei0);
919
920         if (owner == (u64)-1) {
921                 while (1) {
922                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
923                                 ret = btrfs_next_leaf(root, path);
924                                 if (ret < 0)
925                                         return ret;
926                                 BUG_ON(ret > 0);
927                                 leaf = path->nodes[0];
928                         }
929                         btrfs_item_key_to_cpu(leaf, &found_key,
930                                               path->slots[0]);
931                         BUG_ON(key.objectid != found_key.objectid);
932                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
933                                 path->slots[0]++;
934                                 continue;
935                         }
936                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
937                                               struct btrfs_extent_ref_v0);
938                         owner = btrfs_ref_objectid_v0(leaf, ref0);
939                         break;
940                 }
941         }
942         btrfs_release_path(path);
943
944         if (owner < BTRFS_FIRST_FREE_OBJECTID)
945                 new_size += sizeof(*bi);
946
947         new_size -= sizeof(*ei0);
948         ret = btrfs_search_slot(trans, root, &key, path,
949                                 new_size + extra_size, 1);
950         if (ret < 0)
951                 return ret;
952         BUG_ON(ret);
953
954         ret = btrfs_extend_item(trans, root, path, new_size);
955
956         leaf = path->nodes[0];
957         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
958         btrfs_set_extent_refs(leaf, item, refs);
959         /* FIXME: get real generation */
960         btrfs_set_extent_generation(leaf, item, 0);
961         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
962                 btrfs_set_extent_flags(leaf, item,
963                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
964                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
965                 bi = (struct btrfs_tree_block_info *)(item + 1);
966                 /* FIXME: get first key of the block */
967                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
968                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
969         } else {
970                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
971         }
972         btrfs_mark_buffer_dirty(leaf);
973         return 0;
974 }
975 #endif
976
977 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
978 {
979         u32 high_crc = ~(u32)0;
980         u32 low_crc = ~(u32)0;
981         __le64 lenum;
982
983         lenum = cpu_to_le64(root_objectid);
984         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
985         lenum = cpu_to_le64(owner);
986         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
987         lenum = cpu_to_le64(offset);
988         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
989
990         return ((u64)high_crc << 31) ^ (u64)low_crc;
991 }
992
993 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
994                                      struct btrfs_extent_data_ref *ref)
995 {
996         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
997                                     btrfs_extent_data_ref_objectid(leaf, ref),
998                                     btrfs_extent_data_ref_offset(leaf, ref));
999 }
1000
1001 static int match_extent_data_ref(struct extent_buffer *leaf,
1002                                  struct btrfs_extent_data_ref *ref,
1003                                  u64 root_objectid, u64 owner, u64 offset)
1004 {
1005         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1006             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1007             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1008                 return 0;
1009         return 1;
1010 }
1011
1012 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1013                                            struct btrfs_root *root,
1014                                            struct btrfs_path *path,
1015                                            u64 bytenr, u64 parent,
1016                                            u64 root_objectid,
1017                                            u64 owner, u64 offset)
1018 {
1019         struct btrfs_key key;
1020         struct btrfs_extent_data_ref *ref;
1021         struct extent_buffer *leaf;
1022         u32 nritems;
1023         int ret;
1024         int recow;
1025         int err = -ENOENT;
1026
1027         key.objectid = bytenr;
1028         if (parent) {
1029                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1030                 key.offset = parent;
1031         } else {
1032                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1033                 key.offset = hash_extent_data_ref(root_objectid,
1034                                                   owner, offset);
1035         }
1036 again:
1037         recow = 0;
1038         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1039         if (ret < 0) {
1040                 err = ret;
1041                 goto fail;
1042         }
1043
1044         if (parent) {
1045                 if (!ret)
1046                         return 0;
1047 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1048                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1049                 btrfs_release_path(path);
1050                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1051                 if (ret < 0) {
1052                         err = ret;
1053                         goto fail;
1054                 }
1055                 if (!ret)
1056                         return 0;
1057 #endif
1058                 goto fail;
1059         }
1060
1061         leaf = path->nodes[0];
1062         nritems = btrfs_header_nritems(leaf);
1063         while (1) {
1064                 if (path->slots[0] >= nritems) {
1065                         ret = btrfs_next_leaf(root, path);
1066                         if (ret < 0)
1067                                 err = ret;
1068                         if (ret)
1069                                 goto fail;
1070
1071                         leaf = path->nodes[0];
1072                         nritems = btrfs_header_nritems(leaf);
1073                         recow = 1;
1074                 }
1075
1076                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1077                 if (key.objectid != bytenr ||
1078                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1079                         goto fail;
1080
1081                 ref = btrfs_item_ptr(leaf, path->slots[0],
1082                                      struct btrfs_extent_data_ref);
1083
1084                 if (match_extent_data_ref(leaf, ref, root_objectid,
1085                                           owner, offset)) {
1086                         if (recow) {
1087                                 btrfs_release_path(path);
1088                                 goto again;
1089                         }
1090                         err = 0;
1091                         break;
1092                 }
1093                 path->slots[0]++;
1094         }
1095 fail:
1096         return err;
1097 }
1098
1099 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1100                                            struct btrfs_root *root,
1101                                            struct btrfs_path *path,
1102                                            u64 bytenr, u64 parent,
1103                                            u64 root_objectid, u64 owner,
1104                                            u64 offset, int refs_to_add)
1105 {
1106         struct btrfs_key key;
1107         struct extent_buffer *leaf;
1108         u32 size;
1109         u32 num_refs;
1110         int ret;
1111
1112         key.objectid = bytenr;
1113         if (parent) {
1114                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1115                 key.offset = parent;
1116                 size = sizeof(struct btrfs_shared_data_ref);
1117         } else {
1118                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1119                 key.offset = hash_extent_data_ref(root_objectid,
1120                                                   owner, offset);
1121                 size = sizeof(struct btrfs_extent_data_ref);
1122         }
1123
1124         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1125         if (ret && ret != -EEXIST)
1126                 goto fail;
1127
1128         leaf = path->nodes[0];
1129         if (parent) {
1130                 struct btrfs_shared_data_ref *ref;
1131                 ref = btrfs_item_ptr(leaf, path->slots[0],
1132                                      struct btrfs_shared_data_ref);
1133                 if (ret == 0) {
1134                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1135                 } else {
1136                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1137                         num_refs += refs_to_add;
1138                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1139                 }
1140         } else {
1141                 struct btrfs_extent_data_ref *ref;
1142                 while (ret == -EEXIST) {
1143                         ref = btrfs_item_ptr(leaf, path->slots[0],
1144                                              struct btrfs_extent_data_ref);
1145                         if (match_extent_data_ref(leaf, ref, root_objectid,
1146                                                   owner, offset))
1147                                 break;
1148                         btrfs_release_path(path);
1149                         key.offset++;
1150                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1151                                                       size);
1152                         if (ret && ret != -EEXIST)
1153                                 goto fail;
1154
1155                         leaf = path->nodes[0];
1156                 }
1157                 ref = btrfs_item_ptr(leaf, path->slots[0],
1158                                      struct btrfs_extent_data_ref);
1159                 if (ret == 0) {
1160                         btrfs_set_extent_data_ref_root(leaf, ref,
1161                                                        root_objectid);
1162                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1163                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1164                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1165                 } else {
1166                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1167                         num_refs += refs_to_add;
1168                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1169                 }
1170         }
1171         btrfs_mark_buffer_dirty(leaf);
1172         ret = 0;
1173 fail:
1174         btrfs_release_path(path);
1175         return ret;
1176 }
1177
1178 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1179                                            struct btrfs_root *root,
1180                                            struct btrfs_path *path,
1181                                            int refs_to_drop)
1182 {
1183         struct btrfs_key key;
1184         struct btrfs_extent_data_ref *ref1 = NULL;
1185         struct btrfs_shared_data_ref *ref2 = NULL;
1186         struct extent_buffer *leaf;
1187         u32 num_refs = 0;
1188         int ret = 0;
1189
1190         leaf = path->nodes[0];
1191         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1192
1193         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1194                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1195                                       struct btrfs_extent_data_ref);
1196                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1197         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1198                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1199                                       struct btrfs_shared_data_ref);
1200                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1201 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1202         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1203                 struct btrfs_extent_ref_v0 *ref0;
1204                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1205                                       struct btrfs_extent_ref_v0);
1206                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1207 #endif
1208         } else {
1209                 BUG();
1210         }
1211
1212         BUG_ON(num_refs < refs_to_drop);
1213         num_refs -= refs_to_drop;
1214
1215         if (num_refs == 0) {
1216                 ret = btrfs_del_item(trans, root, path);
1217         } else {
1218                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1219                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1220                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1221                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1222 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1223                 else {
1224                         struct btrfs_extent_ref_v0 *ref0;
1225                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1226                                         struct btrfs_extent_ref_v0);
1227                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1228                 }
1229 #endif
1230                 btrfs_mark_buffer_dirty(leaf);
1231         }
1232         return ret;
1233 }
1234
1235 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1236                                           struct btrfs_path *path,
1237                                           struct btrfs_extent_inline_ref *iref)
1238 {
1239         struct btrfs_key key;
1240         struct extent_buffer *leaf;
1241         struct btrfs_extent_data_ref *ref1;
1242         struct btrfs_shared_data_ref *ref2;
1243         u32 num_refs = 0;
1244
1245         leaf = path->nodes[0];
1246         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1247         if (iref) {
1248                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1249                     BTRFS_EXTENT_DATA_REF_KEY) {
1250                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1251                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1252                 } else {
1253                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1254                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1255                 }
1256         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1257                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_extent_data_ref);
1259                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1260         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1261                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1262                                       struct btrfs_shared_data_ref);
1263                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1264 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1265         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1266                 struct btrfs_extent_ref_v0 *ref0;
1267                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1268                                       struct btrfs_extent_ref_v0);
1269                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1270 #endif
1271         } else {
1272                 WARN_ON(1);
1273         }
1274         return num_refs;
1275 }
1276
1277 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1278                                           struct btrfs_root *root,
1279                                           struct btrfs_path *path,
1280                                           u64 bytenr, u64 parent,
1281                                           u64 root_objectid)
1282 {
1283         struct btrfs_key key;
1284         int ret;
1285
1286         key.objectid = bytenr;
1287         if (parent) {
1288                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1289                 key.offset = parent;
1290         } else {
1291                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1292                 key.offset = root_objectid;
1293         }
1294
1295         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1296         if (ret > 0)
1297                 ret = -ENOENT;
1298 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1299         if (ret == -ENOENT && parent) {
1300                 btrfs_release_path(path);
1301                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1302                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1303                 if (ret > 0)
1304                         ret = -ENOENT;
1305         }
1306 #endif
1307         return ret;
1308 }
1309
1310 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1311                                           struct btrfs_root *root,
1312                                           struct btrfs_path *path,
1313                                           u64 bytenr, u64 parent,
1314                                           u64 root_objectid)
1315 {
1316         struct btrfs_key key;
1317         int ret;
1318
1319         key.objectid = bytenr;
1320         if (parent) {
1321                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1322                 key.offset = parent;
1323         } else {
1324                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1325                 key.offset = root_objectid;
1326         }
1327
1328         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1329         btrfs_release_path(path);
1330         return ret;
1331 }
1332
1333 static inline int extent_ref_type(u64 parent, u64 owner)
1334 {
1335         int type;
1336         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1337                 if (parent > 0)
1338                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1339                 else
1340                         type = BTRFS_TREE_BLOCK_REF_KEY;
1341         } else {
1342                 if (parent > 0)
1343                         type = BTRFS_SHARED_DATA_REF_KEY;
1344                 else
1345                         type = BTRFS_EXTENT_DATA_REF_KEY;
1346         }
1347         return type;
1348 }
1349
1350 static int find_next_key(struct btrfs_path *path, int level,
1351                          struct btrfs_key *key)
1352
1353 {
1354         for (; level < BTRFS_MAX_LEVEL; level++) {
1355                 if (!path->nodes[level])
1356                         break;
1357                 if (path->slots[level] + 1 >=
1358                     btrfs_header_nritems(path->nodes[level]))
1359                         continue;
1360                 if (level == 0)
1361                         btrfs_item_key_to_cpu(path->nodes[level], key,
1362                                               path->slots[level] + 1);
1363                 else
1364                         btrfs_node_key_to_cpu(path->nodes[level], key,
1365                                               path->slots[level] + 1);
1366                 return 0;
1367         }
1368         return 1;
1369 }
1370
1371 /*
1372  * look for inline back ref. if back ref is found, *ref_ret is set
1373  * to the address of inline back ref, and 0 is returned.
1374  *
1375  * if back ref isn't found, *ref_ret is set to the address where it
1376  * should be inserted, and -ENOENT is returned.
1377  *
1378  * if insert is true and there are too many inline back refs, the path
1379  * points to the extent item, and -EAGAIN is returned.
1380  *
1381  * NOTE: inline back refs are ordered in the same way that back ref
1382  *       items in the tree are ordered.
1383  */
1384 static noinline_for_stack
1385 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1386                                  struct btrfs_root *root,
1387                                  struct btrfs_path *path,
1388                                  struct btrfs_extent_inline_ref **ref_ret,
1389                                  u64 bytenr, u64 num_bytes,
1390                                  u64 parent, u64 root_objectid,
1391                                  u64 owner, u64 offset, int insert)
1392 {
1393         struct btrfs_key key;
1394         struct extent_buffer *leaf;
1395         struct btrfs_extent_item *ei;
1396         struct btrfs_extent_inline_ref *iref;
1397         u64 flags;
1398         u64 item_size;
1399         unsigned long ptr;
1400         unsigned long end;
1401         int extra_size;
1402         int type;
1403         int want;
1404         int ret;
1405         int err = 0;
1406
1407         key.objectid = bytenr;
1408         key.type = BTRFS_EXTENT_ITEM_KEY;
1409         key.offset = num_bytes;
1410
1411         want = extent_ref_type(parent, owner);
1412         if (insert) {
1413                 extra_size = btrfs_extent_inline_ref_size(want);
1414                 path->keep_locks = 1;
1415         } else
1416                 extra_size = -1;
1417         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1418         if (ret < 0) {
1419                 err = ret;
1420                 goto out;
1421         }
1422         BUG_ON(ret);
1423
1424         leaf = path->nodes[0];
1425         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1426 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1427         if (item_size < sizeof(*ei)) {
1428                 if (!insert) {
1429                         err = -ENOENT;
1430                         goto out;
1431                 }
1432                 ret = convert_extent_item_v0(trans, root, path, owner,
1433                                              extra_size);
1434                 if (ret < 0) {
1435                         err = ret;
1436                         goto out;
1437                 }
1438                 leaf = path->nodes[0];
1439                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1440         }
1441 #endif
1442         BUG_ON(item_size < sizeof(*ei));
1443
1444         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1445         flags = btrfs_extent_flags(leaf, ei);
1446
1447         ptr = (unsigned long)(ei + 1);
1448         end = (unsigned long)ei + item_size;
1449
1450         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1451                 ptr += sizeof(struct btrfs_tree_block_info);
1452                 BUG_ON(ptr > end);
1453         } else {
1454                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1455         }
1456
1457         err = -ENOENT;
1458         while (1) {
1459                 if (ptr >= end) {
1460                         WARN_ON(ptr > end);
1461                         break;
1462                 }
1463                 iref = (struct btrfs_extent_inline_ref *)ptr;
1464                 type = btrfs_extent_inline_ref_type(leaf, iref);
1465                 if (want < type)
1466                         break;
1467                 if (want > type) {
1468                         ptr += btrfs_extent_inline_ref_size(type);
1469                         continue;
1470                 }
1471
1472                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1473                         struct btrfs_extent_data_ref *dref;
1474                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1475                         if (match_extent_data_ref(leaf, dref, root_objectid,
1476                                                   owner, offset)) {
1477                                 err = 0;
1478                                 break;
1479                         }
1480                         if (hash_extent_data_ref_item(leaf, dref) <
1481                             hash_extent_data_ref(root_objectid, owner, offset))
1482                                 break;
1483                 } else {
1484                         u64 ref_offset;
1485                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1486                         if (parent > 0) {
1487                                 if (parent == ref_offset) {
1488                                         err = 0;
1489                                         break;
1490                                 }
1491                                 if (ref_offset < parent)
1492                                         break;
1493                         } else {
1494                                 if (root_objectid == ref_offset) {
1495                                         err = 0;
1496                                         break;
1497                                 }
1498                                 if (ref_offset < root_objectid)
1499                                         break;
1500                         }
1501                 }
1502                 ptr += btrfs_extent_inline_ref_size(type);
1503         }
1504         if (err == -ENOENT && insert) {
1505                 if (item_size + extra_size >=
1506                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1507                         err = -EAGAIN;
1508                         goto out;
1509                 }
1510                 /*
1511                  * To add new inline back ref, we have to make sure
1512                  * there is no corresponding back ref item.
1513                  * For simplicity, we just do not add new inline back
1514                  * ref if there is any kind of item for this block
1515                  */
1516                 if (find_next_key(path, 0, &key) == 0 &&
1517                     key.objectid == bytenr &&
1518                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1519                         err = -EAGAIN;
1520                         goto out;
1521                 }
1522         }
1523         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1524 out:
1525         if (insert) {
1526                 path->keep_locks = 0;
1527                 btrfs_unlock_up_safe(path, 1);
1528         }
1529         return err;
1530 }
1531
1532 /*
1533  * helper to add new inline back ref
1534  */
1535 static noinline_for_stack
1536 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1537                                 struct btrfs_root *root,
1538                                 struct btrfs_path *path,
1539                                 struct btrfs_extent_inline_ref *iref,
1540                                 u64 parent, u64 root_objectid,
1541                                 u64 owner, u64 offset, int refs_to_add,
1542                                 struct btrfs_delayed_extent_op *extent_op)
1543 {
1544         struct extent_buffer *leaf;
1545         struct btrfs_extent_item *ei;
1546         unsigned long ptr;
1547         unsigned long end;
1548         unsigned long item_offset;
1549         u64 refs;
1550         int size;
1551         int type;
1552         int ret;
1553
1554         leaf = path->nodes[0];
1555         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1556         item_offset = (unsigned long)iref - (unsigned long)ei;
1557
1558         type = extent_ref_type(parent, owner);
1559         size = btrfs_extent_inline_ref_size(type);
1560
1561         ret = btrfs_extend_item(trans, root, path, size);
1562
1563         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1564         refs = btrfs_extent_refs(leaf, ei);
1565         refs += refs_to_add;
1566         btrfs_set_extent_refs(leaf, ei, refs);
1567         if (extent_op)
1568                 __run_delayed_extent_op(extent_op, leaf, ei);
1569
1570         ptr = (unsigned long)ei + item_offset;
1571         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1572         if (ptr < end - size)
1573                 memmove_extent_buffer(leaf, ptr + size, ptr,
1574                                       end - size - ptr);
1575
1576         iref = (struct btrfs_extent_inline_ref *)ptr;
1577         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1578         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1579                 struct btrfs_extent_data_ref *dref;
1580                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1581                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1582                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1583                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1584                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1585         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1586                 struct btrfs_shared_data_ref *sref;
1587                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1588                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1589                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1590         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1591                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1592         } else {
1593                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1594         }
1595         btrfs_mark_buffer_dirty(leaf);
1596         return 0;
1597 }
1598
1599 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1600                                  struct btrfs_root *root,
1601                                  struct btrfs_path *path,
1602                                  struct btrfs_extent_inline_ref **ref_ret,
1603                                  u64 bytenr, u64 num_bytes, u64 parent,
1604                                  u64 root_objectid, u64 owner, u64 offset)
1605 {
1606         int ret;
1607
1608         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1609                                            bytenr, num_bytes, parent,
1610                                            root_objectid, owner, offset, 0);
1611         if (ret != -ENOENT)
1612                 return ret;
1613
1614         btrfs_release_path(path);
1615         *ref_ret = NULL;
1616
1617         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1618                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1619                                             root_objectid);
1620         } else {
1621                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1622                                              root_objectid, owner, offset);
1623         }
1624         return ret;
1625 }
1626
1627 /*
1628  * helper to update/remove inline back ref
1629  */
1630 static noinline_for_stack
1631 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1632                                  struct btrfs_root *root,
1633                                  struct btrfs_path *path,
1634                                  struct btrfs_extent_inline_ref *iref,
1635                                  int refs_to_mod,
1636                                  struct btrfs_delayed_extent_op *extent_op)
1637 {
1638         struct extent_buffer *leaf;
1639         struct btrfs_extent_item *ei;
1640         struct btrfs_extent_data_ref *dref = NULL;
1641         struct btrfs_shared_data_ref *sref = NULL;
1642         unsigned long ptr;
1643         unsigned long end;
1644         u32 item_size;
1645         int size;
1646         int type;
1647         int ret;
1648         u64 refs;
1649
1650         leaf = path->nodes[0];
1651         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1652         refs = btrfs_extent_refs(leaf, ei);
1653         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1654         refs += refs_to_mod;
1655         btrfs_set_extent_refs(leaf, ei, refs);
1656         if (extent_op)
1657                 __run_delayed_extent_op(extent_op, leaf, ei);
1658
1659         type = btrfs_extent_inline_ref_type(leaf, iref);
1660
1661         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1662                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1663                 refs = btrfs_extent_data_ref_count(leaf, dref);
1664         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1665                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1666                 refs = btrfs_shared_data_ref_count(leaf, sref);
1667         } else {
1668                 refs = 1;
1669                 BUG_ON(refs_to_mod != -1);
1670         }
1671
1672         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1673         refs += refs_to_mod;
1674
1675         if (refs > 0) {
1676                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1677                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1678                 else
1679                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1680         } else {
1681                 size =  btrfs_extent_inline_ref_size(type);
1682                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1683                 ptr = (unsigned long)iref;
1684                 end = (unsigned long)ei + item_size;
1685                 if (ptr + size < end)
1686                         memmove_extent_buffer(leaf, ptr, ptr + size,
1687                                               end - ptr - size);
1688                 item_size -= size;
1689                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1690         }
1691         btrfs_mark_buffer_dirty(leaf);
1692         return 0;
1693 }
1694
1695 static noinline_for_stack
1696 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1697                                  struct btrfs_root *root,
1698                                  struct btrfs_path *path,
1699                                  u64 bytenr, u64 num_bytes, u64 parent,
1700                                  u64 root_objectid, u64 owner,
1701                                  u64 offset, int refs_to_add,
1702                                  struct btrfs_delayed_extent_op *extent_op)
1703 {
1704         struct btrfs_extent_inline_ref *iref;
1705         int ret;
1706
1707         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1708                                            bytenr, num_bytes, parent,
1709                                            root_objectid, owner, offset, 1);
1710         if (ret == 0) {
1711                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1712                 ret = update_inline_extent_backref(trans, root, path, iref,
1713                                                    refs_to_add, extent_op);
1714         } else if (ret == -ENOENT) {
1715                 ret = setup_inline_extent_backref(trans, root, path, iref,
1716                                                   parent, root_objectid,
1717                                                   owner, offset, refs_to_add,
1718                                                   extent_op);
1719         }
1720         return ret;
1721 }
1722
1723 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1724                                  struct btrfs_root *root,
1725                                  struct btrfs_path *path,
1726                                  u64 bytenr, u64 parent, u64 root_objectid,
1727                                  u64 owner, u64 offset, int refs_to_add)
1728 {
1729         int ret;
1730         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1731                 BUG_ON(refs_to_add != 1);
1732                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1733                                             parent, root_objectid);
1734         } else {
1735                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1736                                              parent, root_objectid,
1737                                              owner, offset, refs_to_add);
1738         }
1739         return ret;
1740 }
1741
1742 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1743                                  struct btrfs_root *root,
1744                                  struct btrfs_path *path,
1745                                  struct btrfs_extent_inline_ref *iref,
1746                                  int refs_to_drop, int is_data)
1747 {
1748         int ret;
1749
1750         BUG_ON(!is_data && refs_to_drop != 1);
1751         if (iref) {
1752                 ret = update_inline_extent_backref(trans, root, path, iref,
1753                                                    -refs_to_drop, NULL);
1754         } else if (is_data) {
1755                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1756         } else {
1757                 ret = btrfs_del_item(trans, root, path);
1758         }
1759         return ret;
1760 }
1761
1762 static int btrfs_issue_discard(struct block_device *bdev,
1763                                 u64 start, u64 len)
1764 {
1765         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1766 }
1767
1768 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1769                                 u64 num_bytes, u64 *actual_bytes)
1770 {
1771         int ret;
1772         u64 discarded_bytes = 0;
1773         struct btrfs_multi_bio *multi = NULL;
1774
1775
1776         /* Tell the block device(s) that the sectors can be discarded */
1777         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1778                               bytenr, &num_bytes, &multi, 0);
1779         if (!ret) {
1780                 struct btrfs_bio_stripe *stripe = multi->stripes;
1781                 int i;
1782
1783
1784                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1785                         if (!stripe->dev->can_discard)
1786                                 continue;
1787
1788                         ret = btrfs_issue_discard(stripe->dev->bdev,
1789                                                   stripe->physical,
1790                                                   stripe->length);
1791                         if (!ret)
1792                                 discarded_bytes += stripe->length;
1793                         else if (ret != -EOPNOTSUPP)
1794                                 break;
1795
1796                         /*
1797                          * Just in case we get back EOPNOTSUPP for some reason,
1798                          * just ignore the return value so we don't screw up
1799                          * people calling discard_extent.
1800                          */
1801                         ret = 0;
1802                 }
1803                 kfree(multi);
1804         }
1805
1806         if (actual_bytes)
1807                 *actual_bytes = discarded_bytes;
1808
1809
1810         return ret;
1811 }
1812
1813 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1814                          struct btrfs_root *root,
1815                          u64 bytenr, u64 num_bytes, u64 parent,
1816                          u64 root_objectid, u64 owner, u64 offset)
1817 {
1818         int ret;
1819         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1820                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1821
1822         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1823                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1824                                         parent, root_objectid, (int)owner,
1825                                         BTRFS_ADD_DELAYED_REF, NULL);
1826         } else {
1827                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1828                                         parent, root_objectid, owner, offset,
1829                                         BTRFS_ADD_DELAYED_REF, NULL);
1830         }
1831         return ret;
1832 }
1833
1834 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1835                                   struct btrfs_root *root,
1836                                   u64 bytenr, u64 num_bytes,
1837                                   u64 parent, u64 root_objectid,
1838                                   u64 owner, u64 offset, int refs_to_add,
1839                                   struct btrfs_delayed_extent_op *extent_op)
1840 {
1841         struct btrfs_path *path;
1842         struct extent_buffer *leaf;
1843         struct btrfs_extent_item *item;
1844         u64 refs;
1845         int ret;
1846         int err = 0;
1847
1848         path = btrfs_alloc_path();
1849         if (!path)
1850                 return -ENOMEM;
1851
1852         path->reada = 1;
1853         path->leave_spinning = 1;
1854         /* this will setup the path even if it fails to insert the back ref */
1855         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1856                                            path, bytenr, num_bytes, parent,
1857                                            root_objectid, owner, offset,
1858                                            refs_to_add, extent_op);
1859         if (ret == 0)
1860                 goto out;
1861
1862         if (ret != -EAGAIN) {
1863                 err = ret;
1864                 goto out;
1865         }
1866
1867         leaf = path->nodes[0];
1868         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1869         refs = btrfs_extent_refs(leaf, item);
1870         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1871         if (extent_op)
1872                 __run_delayed_extent_op(extent_op, leaf, item);
1873
1874         btrfs_mark_buffer_dirty(leaf);
1875         btrfs_release_path(path);
1876
1877         path->reada = 1;
1878         path->leave_spinning = 1;
1879
1880         /* now insert the actual backref */
1881         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1882                                     path, bytenr, parent, root_objectid,
1883                                     owner, offset, refs_to_add);
1884         BUG_ON(ret);
1885 out:
1886         btrfs_free_path(path);
1887         return err;
1888 }
1889
1890 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1891                                 struct btrfs_root *root,
1892                                 struct btrfs_delayed_ref_node *node,
1893                                 struct btrfs_delayed_extent_op *extent_op,
1894                                 int insert_reserved)
1895 {
1896         int ret = 0;
1897         struct btrfs_delayed_data_ref *ref;
1898         struct btrfs_key ins;
1899         u64 parent = 0;
1900         u64 ref_root = 0;
1901         u64 flags = 0;
1902
1903         ins.objectid = node->bytenr;
1904         ins.offset = node->num_bytes;
1905         ins.type = BTRFS_EXTENT_ITEM_KEY;
1906
1907         ref = btrfs_delayed_node_to_data_ref(node);
1908         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1909                 parent = ref->parent;
1910         else
1911                 ref_root = ref->root;
1912
1913         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1914                 if (extent_op) {
1915                         BUG_ON(extent_op->update_key);
1916                         flags |= extent_op->flags_to_set;
1917                 }
1918                 ret = alloc_reserved_file_extent(trans, root,
1919                                                  parent, ref_root, flags,
1920                                                  ref->objectid, ref->offset,
1921                                                  &ins, node->ref_mod);
1922         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1923                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1924                                              node->num_bytes, parent,
1925                                              ref_root, ref->objectid,
1926                                              ref->offset, node->ref_mod,
1927                                              extent_op);
1928         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1929                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1930                                           node->num_bytes, parent,
1931                                           ref_root, ref->objectid,
1932                                           ref->offset, node->ref_mod,
1933                                           extent_op);
1934         } else {
1935                 BUG();
1936         }
1937         return ret;
1938 }
1939
1940 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1941                                     struct extent_buffer *leaf,
1942                                     struct btrfs_extent_item *ei)
1943 {
1944         u64 flags = btrfs_extent_flags(leaf, ei);
1945         if (extent_op->update_flags) {
1946                 flags |= extent_op->flags_to_set;
1947                 btrfs_set_extent_flags(leaf, ei, flags);
1948         }
1949
1950         if (extent_op->update_key) {
1951                 struct btrfs_tree_block_info *bi;
1952                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1953                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1954                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1955         }
1956 }
1957
1958 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1959                                  struct btrfs_root *root,
1960                                  struct btrfs_delayed_ref_node *node,
1961                                  struct btrfs_delayed_extent_op *extent_op)
1962 {
1963         struct btrfs_key key;
1964         struct btrfs_path *path;
1965         struct btrfs_extent_item *ei;
1966         struct extent_buffer *leaf;
1967         u32 item_size;
1968         int ret;
1969         int err = 0;
1970
1971         path = btrfs_alloc_path();
1972         if (!path)
1973                 return -ENOMEM;
1974
1975         key.objectid = node->bytenr;
1976         key.type = BTRFS_EXTENT_ITEM_KEY;
1977         key.offset = node->num_bytes;
1978
1979         path->reada = 1;
1980         path->leave_spinning = 1;
1981         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1982                                 path, 0, 1);
1983         if (ret < 0) {
1984                 err = ret;
1985                 goto out;
1986         }
1987         if (ret > 0) {
1988                 err = -EIO;
1989                 goto out;
1990         }
1991
1992         leaf = path->nodes[0];
1993         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1994 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1995         if (item_size < sizeof(*ei)) {
1996                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
1997                                              path, (u64)-1, 0);
1998                 if (ret < 0) {
1999                         err = ret;
2000                         goto out;
2001                 }
2002                 leaf = path->nodes[0];
2003                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2004         }
2005 #endif
2006         BUG_ON(item_size < sizeof(*ei));
2007         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2008         __run_delayed_extent_op(extent_op, leaf, ei);
2009
2010         btrfs_mark_buffer_dirty(leaf);
2011 out:
2012         btrfs_free_path(path);
2013         return err;
2014 }
2015
2016 static int run_delayed_tree_ref(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                                 int insert_reserved)
2021 {
2022         int ret = 0;
2023         struct btrfs_delayed_tree_ref *ref;
2024         struct btrfs_key ins;
2025         u64 parent = 0;
2026         u64 ref_root = 0;
2027
2028         ins.objectid = node->bytenr;
2029         ins.offset = node->num_bytes;
2030         ins.type = BTRFS_EXTENT_ITEM_KEY;
2031
2032         ref = btrfs_delayed_node_to_tree_ref(node);
2033         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2034                 parent = ref->parent;
2035         else
2036                 ref_root = ref->root;
2037
2038         BUG_ON(node->ref_mod != 1);
2039         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2040                 BUG_ON(!extent_op || !extent_op->update_flags ||
2041                        !extent_op->update_key);
2042                 ret = alloc_reserved_tree_block(trans, root,
2043                                                 parent, ref_root,
2044                                                 extent_op->flags_to_set,
2045                                                 &extent_op->key,
2046                                                 ref->level, &ins);
2047         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2048                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2049                                              node->num_bytes, parent, ref_root,
2050                                              ref->level, 0, 1, extent_op);
2051         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2052                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2053                                           node->num_bytes, parent, ref_root,
2054                                           ref->level, 0, 1, extent_op);
2055         } else {
2056                 BUG();
2057         }
2058         return ret;
2059 }
2060
2061 /* helper function to actually process a single delayed ref entry */
2062 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2063                                struct btrfs_root *root,
2064                                struct btrfs_delayed_ref_node *node,
2065                                struct btrfs_delayed_extent_op *extent_op,
2066                                int insert_reserved)
2067 {
2068         int ret;
2069         if (btrfs_delayed_ref_is_head(node)) {
2070                 struct btrfs_delayed_ref_head *head;
2071                 /*
2072                  * we've hit the end of the chain and we were supposed
2073                  * to insert this extent into the tree.  But, it got
2074                  * deleted before we ever needed to insert it, so all
2075                  * we have to do is clean up the accounting
2076                  */
2077                 BUG_ON(extent_op);
2078                 head = btrfs_delayed_node_to_head(node);
2079                 if (insert_reserved) {
2080                         btrfs_pin_extent(root, node->bytenr,
2081                                          node->num_bytes, 1);
2082                         if (head->is_data) {
2083                                 ret = btrfs_del_csums(trans, root,
2084                                                       node->bytenr,
2085                                                       node->num_bytes);
2086                                 BUG_ON(ret);
2087                         }
2088                 }
2089                 mutex_unlock(&head->mutex);
2090                 return 0;
2091         }
2092
2093         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2094             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2095                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2096                                            insert_reserved);
2097         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2098                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2099                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2100                                            insert_reserved);
2101         else
2102                 BUG();
2103         return ret;
2104 }
2105
2106 static noinline struct btrfs_delayed_ref_node *
2107 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2108 {
2109         struct rb_node *node;
2110         struct btrfs_delayed_ref_node *ref;
2111         int action = BTRFS_ADD_DELAYED_REF;
2112 again:
2113         /*
2114          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2115          * this prevents ref count from going down to zero when
2116          * there still are pending delayed ref.
2117          */
2118         node = rb_prev(&head->node.rb_node);
2119         while (1) {
2120                 if (!node)
2121                         break;
2122                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2123                                 rb_node);
2124                 if (ref->bytenr != head->node.bytenr)
2125                         break;
2126                 if (ref->action == action)
2127                         return ref;
2128                 node = rb_prev(node);
2129         }
2130         if (action == BTRFS_ADD_DELAYED_REF) {
2131                 action = BTRFS_DROP_DELAYED_REF;
2132                 goto again;
2133         }
2134         return NULL;
2135 }
2136
2137 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2138                                        struct btrfs_root *root,
2139                                        struct list_head *cluster)
2140 {
2141         struct btrfs_delayed_ref_root *delayed_refs;
2142         struct btrfs_delayed_ref_node *ref;
2143         struct btrfs_delayed_ref_head *locked_ref = NULL;
2144         struct btrfs_delayed_extent_op *extent_op;
2145         int ret;
2146         int count = 0;
2147         int must_insert_reserved = 0;
2148
2149         delayed_refs = &trans->transaction->delayed_refs;
2150         while (1) {
2151                 if (!locked_ref) {
2152                         /* pick a new head ref from the cluster list */
2153                         if (list_empty(cluster))
2154                                 break;
2155
2156                         locked_ref = list_entry(cluster->next,
2157                                      struct btrfs_delayed_ref_head, cluster);
2158
2159                         /* grab the lock that says we are going to process
2160                          * all the refs for this head */
2161                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2162
2163                         /*
2164                          * we may have dropped the spin lock to get the head
2165                          * mutex lock, and that might have given someone else
2166                          * time to free the head.  If that's true, it has been
2167                          * removed from our list and we can move on.
2168                          */
2169                         if (ret == -EAGAIN) {
2170                                 locked_ref = NULL;
2171                                 count++;
2172                                 continue;
2173                         }
2174                 }
2175
2176                 /*
2177                  * record the must insert reserved flag before we
2178                  * drop the spin lock.
2179                  */
2180                 must_insert_reserved = locked_ref->must_insert_reserved;
2181                 locked_ref->must_insert_reserved = 0;
2182
2183                 extent_op = locked_ref->extent_op;
2184                 locked_ref->extent_op = NULL;
2185
2186                 /*
2187                  * locked_ref is the head node, so we have to go one
2188                  * node back for any delayed ref updates
2189                  */
2190                 ref = select_delayed_ref(locked_ref);
2191                 if (!ref) {
2192                         /* All delayed refs have been processed, Go ahead
2193                          * and send the head node to run_one_delayed_ref,
2194                          * so that any accounting fixes can happen
2195                          */
2196                         ref = &locked_ref->node;
2197
2198                         if (extent_op && must_insert_reserved) {
2199                                 kfree(extent_op);
2200                                 extent_op = NULL;
2201                         }
2202
2203                         if (extent_op) {
2204                                 spin_unlock(&delayed_refs->lock);
2205
2206                                 ret = run_delayed_extent_op(trans, root,
2207                                                             ref, extent_op);
2208                                 BUG_ON(ret);
2209                                 kfree(extent_op);
2210
2211                                 cond_resched();
2212                                 spin_lock(&delayed_refs->lock);
2213                                 continue;
2214                         }
2215
2216                         list_del_init(&locked_ref->cluster);
2217                         locked_ref = NULL;
2218                 }
2219
2220                 ref->in_tree = 0;
2221                 rb_erase(&ref->rb_node, &delayed_refs->root);
2222                 delayed_refs->num_entries--;
2223
2224                 spin_unlock(&delayed_refs->lock);
2225
2226                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2227                                           must_insert_reserved);
2228                 BUG_ON(ret);
2229
2230                 btrfs_put_delayed_ref(ref);
2231                 kfree(extent_op);
2232                 count++;
2233
2234                 cond_resched();
2235                 spin_lock(&delayed_refs->lock);
2236         }
2237         return count;
2238 }
2239
2240 /*
2241  * this starts processing the delayed reference count updates and
2242  * extent insertions we have queued up so far.  count can be
2243  * 0, which means to process everything in the tree at the start
2244  * of the run (but not newly added entries), or it can be some target
2245  * number you'd like to process.
2246  */
2247 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2248                            struct btrfs_root *root, unsigned long count)
2249 {
2250         struct rb_node *node;
2251         struct btrfs_delayed_ref_root *delayed_refs;
2252         struct btrfs_delayed_ref_node *ref;
2253         struct list_head cluster;
2254         int ret;
2255         int run_all = count == (unsigned long)-1;
2256         int run_most = 0;
2257
2258         if (root == root->fs_info->extent_root)
2259                 root = root->fs_info->tree_root;
2260
2261         delayed_refs = &trans->transaction->delayed_refs;
2262         INIT_LIST_HEAD(&cluster);
2263 again:
2264         spin_lock(&delayed_refs->lock);
2265         if (count == 0) {
2266                 count = delayed_refs->num_entries * 2;
2267                 run_most = 1;
2268         }
2269         while (1) {
2270                 if (!(run_all || run_most) &&
2271                     delayed_refs->num_heads_ready < 64)
2272                         break;
2273
2274                 /*
2275                  * go find something we can process in the rbtree.  We start at
2276                  * the beginning of the tree, and then build a cluster
2277                  * of refs to process starting at the first one we are able to
2278                  * lock
2279                  */
2280                 ret = btrfs_find_ref_cluster(trans, &cluster,
2281                                              delayed_refs->run_delayed_start);
2282                 if (ret)
2283                         break;
2284
2285                 ret = run_clustered_refs(trans, root, &cluster);
2286                 BUG_ON(ret < 0);
2287
2288                 count -= min_t(unsigned long, ret, count);
2289
2290                 if (count == 0)
2291                         break;
2292         }
2293
2294         if (run_all) {
2295                 node = rb_first(&delayed_refs->root);
2296                 if (!node)
2297                         goto out;
2298                 count = (unsigned long)-1;
2299
2300                 while (node) {
2301                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2302                                        rb_node);
2303                         if (btrfs_delayed_ref_is_head(ref)) {
2304                                 struct btrfs_delayed_ref_head *head;
2305
2306                                 head = btrfs_delayed_node_to_head(ref);
2307                                 atomic_inc(&ref->refs);
2308
2309                                 spin_unlock(&delayed_refs->lock);
2310                                 /*
2311                                  * Mutex was contended, block until it's
2312                                  * released and try again
2313                                  */
2314                                 mutex_lock(&head->mutex);
2315                                 mutex_unlock(&head->mutex);
2316
2317                                 btrfs_put_delayed_ref(ref);
2318                                 cond_resched();
2319                                 goto again;
2320                         }
2321                         node = rb_next(node);
2322                 }
2323                 spin_unlock(&delayed_refs->lock);
2324                 schedule_timeout(1);
2325                 goto again;
2326         }
2327 out:
2328         spin_unlock(&delayed_refs->lock);
2329         return 0;
2330 }
2331
2332 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2333                                 struct btrfs_root *root,
2334                                 u64 bytenr, u64 num_bytes, u64 flags,
2335                                 int is_data)
2336 {
2337         struct btrfs_delayed_extent_op *extent_op;
2338         int ret;
2339
2340         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2341         if (!extent_op)
2342                 return -ENOMEM;
2343
2344         extent_op->flags_to_set = flags;
2345         extent_op->update_flags = 1;
2346         extent_op->update_key = 0;
2347         extent_op->is_data = is_data ? 1 : 0;
2348
2349         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2350         if (ret)
2351                 kfree(extent_op);
2352         return ret;
2353 }
2354
2355 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2356                                       struct btrfs_root *root,
2357                                       struct btrfs_path *path,
2358                                       u64 objectid, u64 offset, u64 bytenr)
2359 {
2360         struct btrfs_delayed_ref_head *head;
2361         struct btrfs_delayed_ref_node *ref;
2362         struct btrfs_delayed_data_ref *data_ref;
2363         struct btrfs_delayed_ref_root *delayed_refs;
2364         struct rb_node *node;
2365         int ret = 0;
2366
2367         ret = -ENOENT;
2368         delayed_refs = &trans->transaction->delayed_refs;
2369         spin_lock(&delayed_refs->lock);
2370         head = btrfs_find_delayed_ref_head(trans, bytenr);
2371         if (!head)
2372                 goto out;
2373
2374         if (!mutex_trylock(&head->mutex)) {
2375                 atomic_inc(&head->node.refs);
2376                 spin_unlock(&delayed_refs->lock);
2377
2378                 btrfs_release_path(path);
2379
2380                 /*
2381                  * Mutex was contended, block until it's released and let
2382                  * caller try again
2383                  */
2384                 mutex_lock(&head->mutex);
2385                 mutex_unlock(&head->mutex);
2386                 btrfs_put_delayed_ref(&head->node);
2387                 return -EAGAIN;
2388         }
2389
2390         node = rb_prev(&head->node.rb_node);
2391         if (!node)
2392                 goto out_unlock;
2393
2394         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2395
2396         if (ref->bytenr != bytenr)
2397                 goto out_unlock;
2398
2399         ret = 1;
2400         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2401                 goto out_unlock;
2402
2403         data_ref = btrfs_delayed_node_to_data_ref(ref);
2404
2405         node = rb_prev(node);
2406         if (node) {
2407                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2408                 if (ref->bytenr == bytenr)
2409                         goto out_unlock;
2410         }
2411
2412         if (data_ref->root != root->root_key.objectid ||
2413             data_ref->objectid != objectid || data_ref->offset != offset)
2414                 goto out_unlock;
2415
2416         ret = 0;
2417 out_unlock:
2418         mutex_unlock(&head->mutex);
2419 out:
2420         spin_unlock(&delayed_refs->lock);
2421         return ret;
2422 }
2423
2424 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2425                                         struct btrfs_root *root,
2426                                         struct btrfs_path *path,
2427                                         u64 objectid, u64 offset, u64 bytenr)
2428 {
2429         struct btrfs_root *extent_root = root->fs_info->extent_root;
2430         struct extent_buffer *leaf;
2431         struct btrfs_extent_data_ref *ref;
2432         struct btrfs_extent_inline_ref *iref;
2433         struct btrfs_extent_item *ei;
2434         struct btrfs_key key;
2435         u32 item_size;
2436         int ret;
2437
2438         key.objectid = bytenr;
2439         key.offset = (u64)-1;
2440         key.type = BTRFS_EXTENT_ITEM_KEY;
2441
2442         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2443         if (ret < 0)
2444                 goto out;
2445         BUG_ON(ret == 0);
2446
2447         ret = -ENOENT;
2448         if (path->slots[0] == 0)
2449                 goto out;
2450
2451         path->slots[0]--;
2452         leaf = path->nodes[0];
2453         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2454
2455         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2456                 goto out;
2457
2458         ret = 1;
2459         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2460 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2461         if (item_size < sizeof(*ei)) {
2462                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2463                 goto out;
2464         }
2465 #endif
2466         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2467
2468         if (item_size != sizeof(*ei) +
2469             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2470                 goto out;
2471
2472         if (btrfs_extent_generation(leaf, ei) <=
2473             btrfs_root_last_snapshot(&root->root_item))
2474                 goto out;
2475
2476         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2477         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2478             BTRFS_EXTENT_DATA_REF_KEY)
2479                 goto out;
2480
2481         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2482         if (btrfs_extent_refs(leaf, ei) !=
2483             btrfs_extent_data_ref_count(leaf, ref) ||
2484             btrfs_extent_data_ref_root(leaf, ref) !=
2485             root->root_key.objectid ||
2486             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2487             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2488                 goto out;
2489
2490         ret = 0;
2491 out:
2492         return ret;
2493 }
2494
2495 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2496                           struct btrfs_root *root,
2497                           u64 objectid, u64 offset, u64 bytenr)
2498 {
2499         struct btrfs_path *path;
2500         int ret;
2501         int ret2;
2502
2503         path = btrfs_alloc_path();
2504         if (!path)
2505                 return -ENOENT;
2506
2507         do {
2508                 ret = check_committed_ref(trans, root, path, objectid,
2509                                           offset, bytenr);
2510                 if (ret && ret != -ENOENT)
2511                         goto out;
2512
2513                 ret2 = check_delayed_ref(trans, root, path, objectid,
2514                                          offset, bytenr);
2515         } while (ret2 == -EAGAIN);
2516
2517         if (ret2 && ret2 != -ENOENT) {
2518                 ret = ret2;
2519                 goto out;
2520         }
2521
2522         if (ret != -ENOENT || ret2 != -ENOENT)
2523                 ret = 0;
2524 out:
2525         btrfs_free_path(path);
2526         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2527                 WARN_ON(ret > 0);
2528         return ret;
2529 }
2530
2531 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2532                            struct btrfs_root *root,
2533                            struct extent_buffer *buf,
2534                            int full_backref, int inc)
2535 {
2536         u64 bytenr;
2537         u64 num_bytes;
2538         u64 parent;
2539         u64 ref_root;
2540         u32 nritems;
2541         struct btrfs_key key;
2542         struct btrfs_file_extent_item *fi;
2543         int i;
2544         int level;
2545         int ret = 0;
2546         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2547                             u64, u64, u64, u64, u64, u64);
2548
2549         ref_root = btrfs_header_owner(buf);
2550         nritems = btrfs_header_nritems(buf);
2551         level = btrfs_header_level(buf);
2552
2553         if (!root->ref_cows && level == 0)
2554                 return 0;
2555
2556         if (inc)
2557                 process_func = btrfs_inc_extent_ref;
2558         else
2559                 process_func = btrfs_free_extent;
2560
2561         if (full_backref)
2562                 parent = buf->start;
2563         else
2564                 parent = 0;
2565
2566         for (i = 0; i < nritems; i++) {
2567                 if (level == 0) {
2568                         btrfs_item_key_to_cpu(buf, &key, i);
2569                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2570                                 continue;
2571                         fi = btrfs_item_ptr(buf, i,
2572                                             struct btrfs_file_extent_item);
2573                         if (btrfs_file_extent_type(buf, fi) ==
2574                             BTRFS_FILE_EXTENT_INLINE)
2575                                 continue;
2576                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2577                         if (bytenr == 0)
2578                                 continue;
2579
2580                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2581                         key.offset -= btrfs_file_extent_offset(buf, fi);
2582                         ret = process_func(trans, root, bytenr, num_bytes,
2583                                            parent, ref_root, key.objectid,
2584                                            key.offset);
2585                         if (ret)
2586                                 goto fail;
2587                 } else {
2588                         bytenr = btrfs_node_blockptr(buf, i);
2589                         num_bytes = btrfs_level_size(root, level - 1);
2590                         ret = process_func(trans, root, bytenr, num_bytes,
2591                                            parent, ref_root, level - 1, 0);
2592                         if (ret)
2593                                 goto fail;
2594                 }
2595         }
2596         return 0;
2597 fail:
2598         BUG();
2599         return ret;
2600 }
2601
2602 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2603                   struct extent_buffer *buf, int full_backref)
2604 {
2605         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2606 }
2607
2608 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2609                   struct extent_buffer *buf, int full_backref)
2610 {
2611         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2612 }
2613
2614 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2615                                  struct btrfs_root *root,
2616                                  struct btrfs_path *path,
2617                                  struct btrfs_block_group_cache *cache)
2618 {
2619         int ret;
2620         struct btrfs_root *extent_root = root->fs_info->extent_root;
2621         unsigned long bi;
2622         struct extent_buffer *leaf;
2623
2624         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2625         if (ret < 0)
2626                 goto fail;
2627         BUG_ON(ret);
2628
2629         leaf = path->nodes[0];
2630         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2631         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2632         btrfs_mark_buffer_dirty(leaf);
2633         btrfs_release_path(path);
2634 fail:
2635         if (ret)
2636                 return ret;
2637         return 0;
2638
2639 }
2640
2641 static struct btrfs_block_group_cache *
2642 next_block_group(struct btrfs_root *root,
2643                  struct btrfs_block_group_cache *cache)
2644 {
2645         struct rb_node *node;
2646         spin_lock(&root->fs_info->block_group_cache_lock);
2647         node = rb_next(&cache->cache_node);
2648         btrfs_put_block_group(cache);
2649         if (node) {
2650                 cache = rb_entry(node, struct btrfs_block_group_cache,
2651                                  cache_node);
2652                 btrfs_get_block_group(cache);
2653         } else
2654                 cache = NULL;
2655         spin_unlock(&root->fs_info->block_group_cache_lock);
2656         return cache;
2657 }
2658
2659 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2660                             struct btrfs_trans_handle *trans,
2661                             struct btrfs_path *path)
2662 {
2663         struct btrfs_root *root = block_group->fs_info->tree_root;
2664         struct inode *inode = NULL;
2665         u64 alloc_hint = 0;
2666         int dcs = BTRFS_DC_ERROR;
2667         int num_pages = 0;
2668         int retries = 0;
2669         int ret = 0;
2670
2671         /*
2672          * If this block group is smaller than 100 megs don't bother caching the
2673          * block group.
2674          */
2675         if (block_group->key.offset < (100 * 1024 * 1024)) {
2676                 spin_lock(&block_group->lock);
2677                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2678                 spin_unlock(&block_group->lock);
2679                 return 0;
2680         }
2681
2682 again:
2683         inode = lookup_free_space_inode(root, block_group, path);
2684         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2685                 ret = PTR_ERR(inode);
2686                 btrfs_release_path(path);
2687                 goto out;
2688         }
2689
2690         if (IS_ERR(inode)) {
2691                 BUG_ON(retries);
2692                 retries++;
2693
2694                 if (block_group->ro)
2695                         goto out_free;
2696
2697                 ret = create_free_space_inode(root, trans, block_group, path);
2698                 if (ret)
2699                         goto out_free;
2700                 goto again;
2701         }
2702
2703         /*
2704          * We want to set the generation to 0, that way if anything goes wrong
2705          * from here on out we know not to trust this cache when we load up next
2706          * time.
2707          */
2708         BTRFS_I(inode)->generation = 0;
2709         ret = btrfs_update_inode(trans, root, inode);
2710         WARN_ON(ret);
2711
2712         if (i_size_read(inode) > 0) {
2713                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2714                                                       inode);
2715                 if (ret)
2716                         goto out_put;
2717         }
2718
2719         spin_lock(&block_group->lock);
2720         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2721                 /* We're not cached, don't bother trying to write stuff out */
2722                 dcs = BTRFS_DC_WRITTEN;
2723                 spin_unlock(&block_group->lock);
2724                 goto out_put;
2725         }
2726         spin_unlock(&block_group->lock);
2727
2728         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2729         if (!num_pages)
2730                 num_pages = 1;
2731
2732         /*
2733          * Just to make absolutely sure we have enough space, we're going to
2734          * preallocate 12 pages worth of space for each block group.  In
2735          * practice we ought to use at most 8, but we need extra space so we can
2736          * add our header and have a terminator between the extents and the
2737          * bitmaps.
2738          */
2739         num_pages *= 16;
2740         num_pages *= PAGE_CACHE_SIZE;
2741
2742         ret = btrfs_check_data_free_space(inode, num_pages);
2743         if (ret)
2744                 goto out_put;
2745
2746         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2747                                               num_pages, num_pages,
2748                                               &alloc_hint);
2749         if (!ret)
2750                 dcs = BTRFS_DC_SETUP;
2751         btrfs_free_reserved_data_space(inode, num_pages);
2752 out_put:
2753         iput(inode);
2754 out_free:
2755         btrfs_release_path(path);
2756 out:
2757         spin_lock(&block_group->lock);
2758         block_group->disk_cache_state = dcs;
2759         spin_unlock(&block_group->lock);
2760
2761         return ret;
2762 }
2763
2764 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2765                                    struct btrfs_root *root)
2766 {
2767         struct btrfs_block_group_cache *cache;
2768         int err = 0;
2769         struct btrfs_path *path;
2770         u64 last = 0;
2771
2772         path = btrfs_alloc_path();
2773         if (!path)
2774                 return -ENOMEM;
2775
2776 again:
2777         while (1) {
2778                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2779                 while (cache) {
2780                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2781                                 break;
2782                         cache = next_block_group(root, cache);
2783                 }
2784                 if (!cache) {
2785                         if (last == 0)
2786                                 break;
2787                         last = 0;
2788                         continue;
2789                 }
2790                 err = cache_save_setup(cache, trans, path);
2791                 last = cache->key.objectid + cache->key.offset;
2792                 btrfs_put_block_group(cache);
2793         }
2794
2795         while (1) {
2796                 if (last == 0) {
2797                         err = btrfs_run_delayed_refs(trans, root,
2798                                                      (unsigned long)-1);
2799                         BUG_ON(err);
2800                 }
2801
2802                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2803                 while (cache) {
2804                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2805                                 btrfs_put_block_group(cache);
2806                                 goto again;
2807                         }
2808
2809                         if (cache->dirty)
2810                                 break;
2811                         cache = next_block_group(root, cache);
2812                 }
2813                 if (!cache) {
2814                         if (last == 0)
2815                                 break;
2816                         last = 0;
2817                         continue;
2818                 }
2819
2820                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2821                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2822                 cache->dirty = 0;
2823                 last = cache->key.objectid + cache->key.offset;
2824
2825                 err = write_one_cache_group(trans, root, path, cache);
2826                 BUG_ON(err);
2827                 btrfs_put_block_group(cache);
2828         }
2829
2830         while (1) {
2831                 /*
2832                  * I don't think this is needed since we're just marking our
2833                  * preallocated extent as written, but just in case it can't
2834                  * hurt.
2835                  */
2836                 if (last == 0) {
2837                         err = btrfs_run_delayed_refs(trans, root,
2838                                                      (unsigned long)-1);
2839                         BUG_ON(err);
2840                 }
2841
2842                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2843                 while (cache) {
2844                         /*
2845                          * Really this shouldn't happen, but it could if we
2846                          * couldn't write the entire preallocated extent and
2847                          * splitting the extent resulted in a new block.
2848                          */
2849                         if (cache->dirty) {
2850                                 btrfs_put_block_group(cache);
2851                                 goto again;
2852                         }
2853                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2854                                 break;
2855                         cache = next_block_group(root, cache);
2856                 }
2857                 if (!cache) {
2858                         if (last == 0)
2859                                 break;
2860                         last = 0;
2861                         continue;
2862                 }
2863
2864                 btrfs_write_out_cache(root, trans, cache, path);
2865
2866                 /*
2867                  * If we didn't have an error then the cache state is still
2868                  * NEED_WRITE, so we can set it to WRITTEN.
2869                  */
2870                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2871                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2872                 last = cache->key.objectid + cache->key.offset;
2873                 btrfs_put_block_group(cache);
2874         }
2875
2876         btrfs_free_path(path);
2877         return 0;
2878 }
2879
2880 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2881 {
2882         struct btrfs_block_group_cache *block_group;
2883         int readonly = 0;
2884
2885         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2886         if (!block_group || block_group->ro)
2887                 readonly = 1;
2888         if (block_group)
2889                 btrfs_put_block_group(block_group);
2890         return readonly;
2891 }
2892
2893 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2894                              u64 total_bytes, u64 bytes_used,
2895                              struct btrfs_space_info **space_info)
2896 {
2897         struct btrfs_space_info *found;
2898         int i;
2899         int factor;
2900
2901         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2902                      BTRFS_BLOCK_GROUP_RAID10))
2903                 factor = 2;
2904         else
2905                 factor = 1;
2906
2907         found = __find_space_info(info, flags);
2908         if (found) {
2909                 spin_lock(&found->lock);
2910                 found->total_bytes += total_bytes;
2911                 found->disk_total += total_bytes * factor;
2912                 found->bytes_used += bytes_used;
2913                 found->disk_used += bytes_used * factor;
2914                 found->full = 0;
2915                 spin_unlock(&found->lock);
2916                 *space_info = found;
2917                 return 0;
2918         }
2919         found = kzalloc(sizeof(*found), GFP_NOFS);
2920         if (!found)
2921                 return -ENOMEM;
2922
2923         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2924                 INIT_LIST_HEAD(&found->block_groups[i]);
2925         init_rwsem(&found->groups_sem);
2926         spin_lock_init(&found->lock);
2927         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2928                                 BTRFS_BLOCK_GROUP_SYSTEM |
2929                                 BTRFS_BLOCK_GROUP_METADATA);
2930         found->total_bytes = total_bytes;
2931         found->disk_total = total_bytes * factor;
2932         found->bytes_used = bytes_used;
2933         found->disk_used = bytes_used * factor;
2934         found->bytes_pinned = 0;
2935         found->bytes_reserved = 0;
2936         found->bytes_readonly = 0;
2937         found->bytes_may_use = 0;
2938         found->full = 0;
2939         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2940         found->chunk_alloc = 0;
2941         found->flush = 0;
2942         init_waitqueue_head(&found->wait);
2943         *space_info = found;
2944         list_add_rcu(&found->list, &info->space_info);
2945         return 0;
2946 }
2947
2948 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2949 {
2950         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2951                                    BTRFS_BLOCK_GROUP_RAID1 |
2952                                    BTRFS_BLOCK_GROUP_RAID10 |
2953                                    BTRFS_BLOCK_GROUP_DUP);
2954         if (extra_flags) {
2955                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2956                         fs_info->avail_data_alloc_bits |= extra_flags;
2957                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2958                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2959                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2960                         fs_info->avail_system_alloc_bits |= extra_flags;
2961         }
2962 }
2963
2964 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2965 {
2966         /*
2967          * we add in the count of missing devices because we want
2968          * to make sure that any RAID levels on a degraded FS
2969          * continue to be honored.
2970          */
2971         u64 num_devices = root->fs_info->fs_devices->rw_devices +
2972                 root->fs_info->fs_devices->missing_devices;
2973
2974         if (num_devices == 1)
2975                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
2976         if (num_devices < 4)
2977                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
2978
2979         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
2980             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
2981                       BTRFS_BLOCK_GROUP_RAID10))) {
2982                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
2983         }
2984
2985         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
2986             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
2987                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
2988         }
2989
2990         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
2991             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
2992              (flags & BTRFS_BLOCK_GROUP_RAID10) |
2993              (flags & BTRFS_BLOCK_GROUP_DUP)))
2994                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
2995         return flags;
2996 }
2997
2998 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
2999 {
3000         if (flags & BTRFS_BLOCK_GROUP_DATA)
3001                 flags |= root->fs_info->avail_data_alloc_bits &
3002                          root->fs_info->data_alloc_profile;
3003         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3004                 flags |= root->fs_info->avail_system_alloc_bits &
3005                          root->fs_info->system_alloc_profile;
3006         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3007                 flags |= root->fs_info->avail_metadata_alloc_bits &
3008                          root->fs_info->metadata_alloc_profile;
3009         return btrfs_reduce_alloc_profile(root, flags);
3010 }
3011
3012 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3013 {
3014         u64 flags;
3015
3016         if (data)
3017                 flags = BTRFS_BLOCK_GROUP_DATA;
3018         else if (root == root->fs_info->chunk_root)
3019                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3020         else
3021                 flags = BTRFS_BLOCK_GROUP_METADATA;
3022
3023         return get_alloc_profile(root, flags);
3024 }
3025
3026 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3027 {
3028         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3029                                                        BTRFS_BLOCK_GROUP_DATA);
3030 }
3031
3032 /*
3033  * This will check the space that the inode allocates from to make sure we have
3034  * enough space for bytes.
3035  */
3036 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3037 {
3038         struct btrfs_space_info *data_sinfo;
3039         struct btrfs_root *root = BTRFS_I(inode)->root;
3040         u64 used;
3041         int ret = 0, committed = 0, alloc_chunk = 1;
3042
3043         /* make sure bytes are sectorsize aligned */
3044         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3045
3046         if (root == root->fs_info->tree_root ||
3047             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3048                 alloc_chunk = 0;
3049                 committed = 1;
3050         }
3051
3052         data_sinfo = BTRFS_I(inode)->space_info;
3053         if (!data_sinfo)
3054                 goto alloc;
3055
3056 again:
3057         /* make sure we have enough space to handle the data first */
3058         spin_lock(&data_sinfo->lock);
3059         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3060                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3061                 data_sinfo->bytes_may_use;
3062
3063         if (used + bytes > data_sinfo->total_bytes) {
3064                 struct btrfs_trans_handle *trans;
3065
3066                 /*
3067                  * if we don't have enough free bytes in this space then we need
3068                  * to alloc a new chunk.
3069                  */
3070                 if (!data_sinfo->full && alloc_chunk) {
3071                         u64 alloc_target;
3072
3073                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3074                         spin_unlock(&data_sinfo->lock);
3075 alloc:
3076                         alloc_target = btrfs_get_alloc_profile(root, 1);
3077                         trans = btrfs_join_transaction(root);
3078                         if (IS_ERR(trans))
3079                                 return PTR_ERR(trans);
3080
3081                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3082                                              bytes + 2 * 1024 * 1024,
3083                                              alloc_target,
3084                                              CHUNK_ALLOC_NO_FORCE);
3085                         btrfs_end_transaction(trans, root);
3086                         if (ret < 0) {
3087                                 if (ret != -ENOSPC)
3088                                         return ret;
3089                                 else
3090                                         goto commit_trans;
3091                         }
3092
3093                         if (!data_sinfo) {
3094                                 btrfs_set_inode_space_info(root, inode);
3095                                 data_sinfo = BTRFS_I(inode)->space_info;
3096                         }
3097                         goto again;
3098                 }
3099
3100                 /*
3101                  * If we have less pinned bytes than we want to allocate then
3102                  * don't bother committing the transaction, it won't help us.
3103                  */
3104                 if (data_sinfo->bytes_pinned < bytes)
3105                         committed = 1;
3106                 spin_unlock(&data_sinfo->lock);
3107
3108                 /* commit the current transaction and try again */
3109 commit_trans:
3110                 if (!committed &&
3111                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3112                         committed = 1;
3113                         trans = btrfs_join_transaction(root);
3114                         if (IS_ERR(trans))
3115                                 return PTR_ERR(trans);
3116                         ret = btrfs_commit_transaction(trans, root);
3117                         if (ret)
3118                                 return ret;
3119                         goto again;
3120                 }
3121
3122                 return -ENOSPC;
3123         }
3124         data_sinfo->bytes_may_use += bytes;
3125         BTRFS_I(inode)->reserved_bytes += bytes;
3126         spin_unlock(&data_sinfo->lock);
3127
3128         return 0;
3129 }
3130
3131 /*
3132  * called when we are clearing an delalloc extent from the
3133  * inode's io_tree or there was an error for whatever reason
3134  * after calling btrfs_check_data_free_space
3135  */
3136 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3137 {
3138         struct btrfs_root *root = BTRFS_I(inode)->root;
3139         struct btrfs_space_info *data_sinfo;
3140
3141         /* make sure bytes are sectorsize aligned */
3142         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3143
3144         data_sinfo = BTRFS_I(inode)->space_info;
3145         spin_lock(&data_sinfo->lock);
3146         data_sinfo->bytes_may_use -= bytes;
3147         BTRFS_I(inode)->reserved_bytes -= bytes;
3148         spin_unlock(&data_sinfo->lock);
3149 }
3150
3151 static void force_metadata_allocation(struct btrfs_fs_info *info)
3152 {
3153         struct list_head *head = &info->space_info;
3154         struct btrfs_space_info *found;
3155
3156         rcu_read_lock();
3157         list_for_each_entry_rcu(found, head, list) {
3158                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3159                         found->force_alloc = CHUNK_ALLOC_FORCE;
3160         }
3161         rcu_read_unlock();
3162 }
3163
3164 static int should_alloc_chunk(struct btrfs_root *root,
3165                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3166                               int force)
3167 {
3168         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3169         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3170         u64 thresh;
3171
3172         if (force == CHUNK_ALLOC_FORCE)
3173                 return 1;
3174
3175         /*
3176          * in limited mode, we want to have some free space up to
3177          * about 1% of the FS size.
3178          */
3179         if (force == CHUNK_ALLOC_LIMITED) {
3180                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3181                 thresh = max_t(u64, 64 * 1024 * 1024,
3182                                div_factor_fine(thresh, 1));
3183
3184                 if (num_bytes - num_allocated < thresh)
3185                         return 1;
3186         }
3187
3188         /*
3189          * we have two similar checks here, one based on percentage
3190          * and once based on a hard number of 256MB.  The idea
3191          * is that if we have a good amount of free
3192          * room, don't allocate a chunk.  A good mount is
3193          * less than 80% utilized of the chunks we have allocated,
3194          * or more than 256MB free
3195          */
3196         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3197                 return 0;
3198
3199         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3200                 return 0;
3201
3202         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3203
3204         /* 256MB or 5% of the FS */
3205         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3206
3207         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3208                 return 0;
3209         return 1;
3210 }
3211
3212 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3213                           struct btrfs_root *extent_root, u64 alloc_bytes,
3214                           u64 flags, int force)
3215 {
3216         struct btrfs_space_info *space_info;
3217         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3218         int wait_for_alloc = 0;
3219         int ret = 0;
3220
3221         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3222
3223         space_info = __find_space_info(extent_root->fs_info, flags);
3224         if (!space_info) {
3225                 ret = update_space_info(extent_root->fs_info, flags,
3226                                         0, 0, &space_info);
3227                 BUG_ON(ret);
3228         }
3229         BUG_ON(!space_info);
3230
3231 again:
3232         spin_lock(&space_info->lock);
3233         if (space_info->force_alloc)
3234                 force = space_info->force_alloc;
3235         if (space_info->full) {
3236                 spin_unlock(&space_info->lock);
3237                 return 0;
3238         }
3239
3240         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3241                 spin_unlock(&space_info->lock);
3242                 return 0;
3243         } else if (space_info->chunk_alloc) {
3244                 wait_for_alloc = 1;
3245         } else {
3246                 space_info->chunk_alloc = 1;
3247         }
3248
3249         spin_unlock(&space_info->lock);
3250
3251         mutex_lock(&fs_info->chunk_mutex);
3252
3253         /*
3254          * The chunk_mutex is held throughout the entirety of a chunk
3255          * allocation, so once we've acquired the chunk_mutex we know that the
3256          * other guy is done and we need to recheck and see if we should
3257          * allocate.
3258          */
3259         if (wait_for_alloc) {
3260                 mutex_unlock(&fs_info->chunk_mutex);
3261                 wait_for_alloc = 0;
3262                 goto again;
3263         }
3264
3265         /*
3266          * If we have mixed data/metadata chunks we want to make sure we keep
3267          * allocating mixed chunks instead of individual chunks.
3268          */
3269         if (btrfs_mixed_space_info(space_info))
3270                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3271
3272         /*
3273          * if we're doing a data chunk, go ahead and make sure that
3274          * we keep a reasonable number of metadata chunks allocated in the
3275          * FS as well.
3276          */
3277         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3278                 fs_info->data_chunk_allocations++;
3279                 if (!(fs_info->data_chunk_allocations %
3280                       fs_info->metadata_ratio))
3281                         force_metadata_allocation(fs_info);
3282         }
3283
3284         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3285         if (ret < 0 && ret != -ENOSPC)
3286                 goto out;
3287
3288         spin_lock(&space_info->lock);
3289         if (ret)
3290                 space_info->full = 1;
3291         else
3292                 ret = 1;
3293
3294         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3295         space_info->chunk_alloc = 0;
3296         spin_unlock(&space_info->lock);
3297 out:
3298         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3299         return ret;
3300 }
3301
3302 /*
3303  * shrink metadata reservation for delalloc
3304  */
3305 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3306                            struct btrfs_root *root, u64 to_reclaim, int sync)
3307 {
3308         struct btrfs_block_rsv *block_rsv;
3309         struct btrfs_space_info *space_info;
3310         u64 reserved;
3311         u64 max_reclaim;
3312         u64 reclaimed = 0;
3313         long time_left;
3314         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3315         int loops = 0;
3316         unsigned long progress;
3317
3318         block_rsv = &root->fs_info->delalloc_block_rsv;
3319         space_info = block_rsv->space_info;
3320
3321         smp_mb();
3322         reserved = space_info->bytes_reserved;
3323         progress = space_info->reservation_progress;
3324
3325         if (reserved == 0)
3326                 return 0;
3327
3328         smp_mb();
3329         if (root->fs_info->delalloc_bytes == 0) {
3330                 if (trans)
3331                         return 0;
3332                 btrfs_wait_ordered_extents(root, 0, 0);
3333                 return 0;
3334         }
3335
3336         max_reclaim = min(reserved, to_reclaim);
3337
3338         while (loops < 1024) {
3339                 /* have the flusher threads jump in and do some IO */
3340                 smp_mb();
3341                 nr_pages = min_t(unsigned long, nr_pages,
3342                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3343                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3344
3345                 spin_lock(&space_info->lock);
3346                 if (reserved > space_info->bytes_reserved)
3347                         reclaimed += reserved - space_info->bytes_reserved;
3348                 reserved = space_info->bytes_reserved;
3349                 spin_unlock(&space_info->lock);
3350
3351                 loops++;
3352
3353                 if (reserved == 0 || reclaimed >= max_reclaim)
3354                         break;
3355
3356                 if (trans && trans->transaction->blocked)
3357                         return -EAGAIN;
3358
3359                 time_left = schedule_timeout_interruptible(1);
3360
3361                 /* We were interrupted, exit */
3362                 if (time_left)
3363                         break;
3364
3365                 /* we've kicked the IO a few times, if anything has been freed,
3366                  * exit.  There is no sense in looping here for a long time
3367                  * when we really need to commit the transaction, or there are
3368                  * just too many writers without enough free space
3369                  */
3370
3371                 if (loops > 3) {
3372                         smp_mb();
3373                         if (progress != space_info->reservation_progress)
3374                                 break;
3375                 }
3376
3377         }
3378         if (reclaimed >= to_reclaim && !trans)
3379                 btrfs_wait_ordered_extents(root, 0, 0);
3380         return reclaimed >= to_reclaim;
3381 }
3382
3383 /*
3384  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3385  * idea is if this is the first call (retries == 0) then we will add to our
3386  * reserved count if we can't make the allocation in order to hold our place
3387  * while we go and try and free up space.  That way for retries > 1 we don't try
3388  * and add space, we just check to see if the amount of unused space is >= the
3389  * total space, meaning that our reservation is valid.
3390  *
3391  * However if we don't intend to retry this reservation, pass -1 as retries so
3392  * that it short circuits this logic.
3393  */
3394 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3395                                   struct btrfs_root *root,
3396                                   struct btrfs_block_rsv *block_rsv,
3397                                   u64 orig_bytes, int flush)
3398 {
3399         struct btrfs_space_info *space_info = block_rsv->space_info;
3400         u64 unused;
3401         u64 num_bytes = orig_bytes;
3402         int retries = 0;
3403         int ret = 0;
3404         bool committed = false;
3405         bool flushing = false;
3406
3407 again:
3408         ret = 0;
3409         spin_lock(&space_info->lock);
3410         /*
3411          * We only want to wait if somebody other than us is flushing and we are
3412          * actually alloed to flush.
3413          */
3414         while (flush && !flushing && space_info->flush) {
3415                 spin_unlock(&space_info->lock);
3416                 /*
3417                  * If we have a trans handle we can't wait because the flusher
3418                  * may have to commit the transaction, which would mean we would
3419                  * deadlock since we are waiting for the flusher to finish, but
3420                  * hold the current transaction open.
3421                  */
3422                 if (trans)
3423                         return -EAGAIN;
3424                 ret = wait_event_interruptible(space_info->wait,
3425                                                !space_info->flush);
3426                 /* Must have been interrupted, return */
3427                 if (ret)
3428                         return -EINTR;
3429
3430                 spin_lock(&space_info->lock);
3431         }
3432
3433         ret = -ENOSPC;
3434         unused = space_info->bytes_used + space_info->bytes_reserved +
3435                  space_info->bytes_pinned + space_info->bytes_readonly +
3436                  space_info->bytes_may_use;
3437
3438         /*
3439          * The idea here is that we've not already over-reserved the block group
3440          * then we can go ahead and save our reservation first and then start
3441          * flushing if we need to.  Otherwise if we've already overcommitted
3442          * lets start flushing stuff first and then come back and try to make
3443          * our reservation.
3444          */
3445         if (unused <= space_info->total_bytes) {
3446                 unused = space_info->total_bytes - unused;
3447                 if (unused >= num_bytes) {
3448                         space_info->bytes_reserved += orig_bytes;
3449                         ret = 0;
3450                 } else {
3451                         /*
3452                          * Ok set num_bytes to orig_bytes since we aren't
3453                          * overocmmitted, this way we only try and reclaim what
3454                          * we need.
3455                          */
3456                         num_bytes = orig_bytes;
3457                 }
3458         } else {
3459                 /*
3460                  * Ok we're over committed, set num_bytes to the overcommitted
3461                  * amount plus the amount of bytes that we need for this
3462                  * reservation.
3463                  */
3464                 num_bytes = unused - space_info->total_bytes +
3465                         (orig_bytes * (retries + 1));
3466         }
3467
3468         /*
3469          * Couldn't make our reservation, save our place so while we're trying
3470          * to reclaim space we can actually use it instead of somebody else
3471          * stealing it from us.
3472          */
3473         if (ret && flush) {
3474                 flushing = true;
3475                 space_info->flush = 1;
3476         }
3477
3478         spin_unlock(&space_info->lock);
3479
3480         if (!ret || !flush)
3481                 goto out;
3482
3483         /*
3484          * We do synchronous shrinking since we don't actually unreserve
3485          * metadata until after the IO is completed.
3486          */
3487         ret = shrink_delalloc(trans, root, num_bytes, 1);
3488         if (ret < 0)
3489                 goto out;
3490
3491         ret = 0;
3492
3493         /*
3494          * So if we were overcommitted it's possible that somebody else flushed
3495          * out enough space and we simply didn't have enough space to reclaim,
3496          * so go back around and try again.
3497          */
3498         if (retries < 2) {
3499                 retries++;
3500                 goto again;
3501         }
3502
3503         /*
3504          * Not enough space to be reclaimed, don't bother committing the
3505          * transaction.
3506          */
3507         spin_lock(&space_info->lock);
3508         if (space_info->bytes_pinned < orig_bytes)
3509                 ret = -ENOSPC;
3510         spin_unlock(&space_info->lock);
3511         if (ret)
3512                 goto out;
3513
3514         ret = -EAGAIN;
3515         if (trans)
3516                 goto out;
3517
3518         ret = -ENOSPC;
3519         if (committed)
3520                 goto out;
3521
3522         trans = btrfs_join_transaction(root);
3523         if (IS_ERR(trans))
3524                 goto out;
3525         ret = btrfs_commit_transaction(trans, root);
3526         if (!ret) {
3527                 trans = NULL;
3528                 committed = true;
3529                 goto again;
3530         }
3531
3532 out:
3533         if (flushing) {
3534                 spin_lock(&space_info->lock);
3535                 space_info->flush = 0;
3536                 wake_up_all(&space_info->wait);
3537                 spin_unlock(&space_info->lock);
3538         }
3539         return ret;
3540 }
3541
3542 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3543                                              struct btrfs_root *root)
3544 {
3545         struct btrfs_block_rsv *block_rsv;
3546         if (root->ref_cows)
3547                 block_rsv = trans->block_rsv;
3548         else
3549                 block_rsv = root->block_rsv;
3550
3551         if (!block_rsv)
3552                 block_rsv = &root->fs_info->empty_block_rsv;
3553
3554         return block_rsv;
3555 }
3556
3557 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3558                                u64 num_bytes)
3559 {
3560         int ret = -ENOSPC;
3561         spin_lock(&block_rsv->lock);
3562         if (block_rsv->reserved >= num_bytes) {
3563                 block_rsv->reserved -= num_bytes;
3564                 if (block_rsv->reserved < block_rsv->size)
3565                         block_rsv->full = 0;
3566                 ret = 0;
3567         }
3568         spin_unlock(&block_rsv->lock);
3569         return ret;
3570 }
3571
3572 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3573                                 u64 num_bytes, int update_size)
3574 {
3575         spin_lock(&block_rsv->lock);
3576         block_rsv->reserved += num_bytes;
3577         if (update_size)
3578                 block_rsv->size += num_bytes;
3579         else if (block_rsv->reserved >= block_rsv->size)
3580                 block_rsv->full = 1;
3581         spin_unlock(&block_rsv->lock);
3582 }
3583
3584 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3585                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3586 {
3587         struct btrfs_space_info *space_info = block_rsv->space_info;
3588
3589         spin_lock(&block_rsv->lock);
3590         if (num_bytes == (u64)-1)
3591                 num_bytes = block_rsv->size;
3592         block_rsv->size -= num_bytes;
3593         if (block_rsv->reserved >= block_rsv->size) {
3594                 num_bytes = block_rsv->reserved - block_rsv->size;
3595                 block_rsv->reserved = block_rsv->size;
3596                 block_rsv->full = 1;
3597         } else {
3598                 num_bytes = 0;
3599         }
3600         spin_unlock(&block_rsv->lock);
3601
3602         if (num_bytes > 0) {
3603                 if (dest) {
3604                         spin_lock(&dest->lock);
3605                         if (!dest->full) {
3606                                 u64 bytes_to_add;
3607
3608                                 bytes_to_add = dest->size - dest->reserved;
3609                                 bytes_to_add = min(num_bytes, bytes_to_add);
3610                                 dest->reserved += bytes_to_add;
3611                                 if (dest->reserved >= dest->size)
3612                                         dest->full = 1;
3613                                 num_bytes -= bytes_to_add;
3614                         }
3615                         spin_unlock(&dest->lock);
3616                 }
3617                 if (num_bytes) {
3618                         spin_lock(&space_info->lock);
3619                         space_info->bytes_reserved -= num_bytes;
3620                         space_info->reservation_progress++;
3621                         spin_unlock(&space_info->lock);
3622                 }
3623         }
3624 }
3625
3626 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3627                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3628 {
3629         int ret;
3630
3631         ret = block_rsv_use_bytes(src, num_bytes);
3632         if (ret)
3633                 return ret;
3634
3635         block_rsv_add_bytes(dst, num_bytes, 1);
3636         return 0;
3637 }
3638
3639 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3640 {
3641         memset(rsv, 0, sizeof(*rsv));
3642         spin_lock_init(&rsv->lock);
3643         atomic_set(&rsv->usage, 1);
3644         rsv->priority = 6;
3645         INIT_LIST_HEAD(&rsv->list);
3646 }
3647
3648 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3649 {
3650         struct btrfs_block_rsv *block_rsv;
3651         struct btrfs_fs_info *fs_info = root->fs_info;
3652
3653         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3654         if (!block_rsv)
3655                 return NULL;
3656
3657         btrfs_init_block_rsv(block_rsv);
3658         block_rsv->space_info = __find_space_info(fs_info,
3659                                                   BTRFS_BLOCK_GROUP_METADATA);
3660         return block_rsv;
3661 }
3662
3663 void btrfs_free_block_rsv(struct btrfs_root *root,
3664                           struct btrfs_block_rsv *rsv)
3665 {
3666         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3667                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3668                 if (!rsv->durable)
3669                         kfree(rsv);
3670         }
3671 }
3672
3673 /*
3674  * make the block_rsv struct be able to capture freed space.
3675  * the captured space will re-add to the the block_rsv struct
3676  * after transaction commit
3677  */
3678 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3679                                  struct btrfs_block_rsv *block_rsv)
3680 {
3681         block_rsv->durable = 1;
3682         mutex_lock(&fs_info->durable_block_rsv_mutex);
3683         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3684         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3685 }
3686
3687 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3688                         struct btrfs_root *root,
3689                         struct btrfs_block_rsv *block_rsv,
3690                         u64 num_bytes)
3691 {
3692         int ret;
3693
3694         if (num_bytes == 0)
3695                 return 0;
3696
3697         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3698         if (!ret) {
3699                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3700                 return 0;
3701         }
3702
3703         return ret;
3704 }
3705
3706 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3707                           struct btrfs_root *root,
3708                           struct btrfs_block_rsv *block_rsv,
3709                           u64 min_reserved, int min_factor)
3710 {
3711         u64 num_bytes = 0;
3712         int commit_trans = 0;
3713         int ret = -ENOSPC;
3714
3715         if (!block_rsv)
3716                 return 0;
3717
3718         spin_lock(&block_rsv->lock);
3719         if (min_factor > 0)
3720                 num_bytes = div_factor(block_rsv->size, min_factor);
3721         if (min_reserved > num_bytes)
3722                 num_bytes = min_reserved;
3723
3724         if (block_rsv->reserved >= num_bytes) {
3725                 ret = 0;
3726         } else {
3727                 num_bytes -= block_rsv->reserved;
3728                 if (block_rsv->durable &&
3729                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3730                         commit_trans = 1;
3731         }
3732         spin_unlock(&block_rsv->lock);
3733         if (!ret)
3734                 return 0;
3735
3736         if (block_rsv->refill_used) {
3737                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3738                                              num_bytes, 0);
3739                 if (!ret) {
3740                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3741                         return 0;
3742                 }
3743         }
3744
3745         if (commit_trans) {
3746                 if (trans)
3747                         return -EAGAIN;
3748                 trans = btrfs_join_transaction(root);
3749                 BUG_ON(IS_ERR(trans));
3750                 ret = btrfs_commit_transaction(trans, root);
3751                 return 0;
3752         }
3753
3754         return -ENOSPC;
3755 }
3756
3757 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3758                             struct btrfs_block_rsv *dst_rsv,
3759                             u64 num_bytes)
3760 {
3761         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3762 }
3763
3764 void btrfs_block_rsv_release(struct btrfs_root *root,
3765                              struct btrfs_block_rsv *block_rsv,
3766                              u64 num_bytes)
3767 {
3768         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3769         if (global_rsv->full || global_rsv == block_rsv ||
3770             block_rsv->space_info != global_rsv->space_info)
3771                 global_rsv = NULL;
3772         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3773 }
3774
3775 /*
3776  * helper to calculate size of global block reservation.
3777  * the desired value is sum of space used by extent tree,
3778  * checksum tree and root tree
3779  */
3780 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3781 {
3782         struct btrfs_space_info *sinfo;
3783         u64 num_bytes;
3784         u64 meta_used;
3785         u64 data_used;
3786         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3787
3788         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3789         spin_lock(&sinfo->lock);
3790         data_used = sinfo->bytes_used;
3791         spin_unlock(&sinfo->lock);
3792
3793         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3794         spin_lock(&sinfo->lock);
3795         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3796                 data_used = 0;
3797         meta_used = sinfo->bytes_used;
3798         spin_unlock(&sinfo->lock);
3799
3800         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3801                     csum_size * 2;
3802         num_bytes += div64_u64(data_used + meta_used, 50);
3803
3804         if (num_bytes * 3 > meta_used)
3805                 num_bytes = div64_u64(meta_used, 3);
3806
3807         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3808 }
3809
3810 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3811 {
3812         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3813         struct btrfs_space_info *sinfo = block_rsv->space_info;
3814         u64 num_bytes;
3815
3816         num_bytes = calc_global_metadata_size(fs_info);
3817
3818         spin_lock(&block_rsv->lock);
3819         spin_lock(&sinfo->lock);
3820
3821         block_rsv->size = num_bytes;
3822
3823         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3824                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3825                     sinfo->bytes_may_use;
3826
3827         if (sinfo->total_bytes > num_bytes) {
3828                 num_bytes = sinfo->total_bytes - num_bytes;
3829                 block_rsv->reserved += num_bytes;
3830                 sinfo->bytes_reserved += num_bytes;
3831         }
3832
3833         if (block_rsv->reserved >= block_rsv->size) {
3834                 num_bytes = block_rsv->reserved - block_rsv->size;
3835                 sinfo->bytes_reserved -= num_bytes;
3836                 sinfo->reservation_progress++;
3837                 block_rsv->reserved = block_rsv->size;
3838                 block_rsv->full = 1;
3839         }
3840
3841         spin_unlock(&sinfo->lock);
3842         spin_unlock(&block_rsv->lock);
3843 }
3844
3845 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3846 {
3847         struct btrfs_space_info *space_info;
3848
3849         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3850         fs_info->chunk_block_rsv.space_info = space_info;
3851         fs_info->chunk_block_rsv.priority = 10;
3852
3853         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3854         fs_info->global_block_rsv.space_info = space_info;
3855         fs_info->global_block_rsv.priority = 10;
3856         fs_info->global_block_rsv.refill_used = 1;
3857         fs_info->delalloc_block_rsv.space_info = space_info;
3858         fs_info->trans_block_rsv.space_info = space_info;
3859         fs_info->empty_block_rsv.space_info = space_info;
3860         fs_info->empty_block_rsv.priority = 10;
3861
3862         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3863         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3864         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3865         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3866         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3867
3868         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3869
3870         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3871
3872         update_global_block_rsv(fs_info);
3873 }
3874
3875 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3876 {
3877         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3878         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3879         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3880         WARN_ON(fs_info->trans_block_rsv.size > 0);
3881         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3882         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3883         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3884 }
3885
3886 int btrfs_truncate_reserve_metadata(struct btrfs_trans_handle *trans,
3887                                     struct btrfs_root *root,
3888                                     struct btrfs_block_rsv *rsv)
3889 {
3890         struct btrfs_block_rsv *trans_rsv = &root->fs_info->trans_block_rsv;
3891         u64 num_bytes;
3892         int ret;
3893
3894         /*
3895          * Truncate should be freeing data, but give us 2 items just in case it
3896          * needs to use some space.  We may want to be smarter about this in the
3897          * future.
3898          */
3899         num_bytes = btrfs_calc_trans_metadata_size(root, 2);
3900
3901         /* We already have enough bytes, just return */
3902         if (rsv->reserved >= num_bytes)
3903                 return 0;
3904
3905         num_bytes -= rsv->reserved;
3906
3907         /*
3908          * You should have reserved enough space before hand to do this, so this
3909          * should not fail.
3910          */
3911         ret = block_rsv_migrate_bytes(trans_rsv, rsv, num_bytes);
3912         BUG_ON(ret);
3913
3914         return 0;
3915 }
3916
3917 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3918                                   struct btrfs_root *root)
3919 {
3920         if (!trans->bytes_reserved)
3921                 return;
3922
3923         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3924         btrfs_block_rsv_release(root, trans->block_rsv,
3925                                 trans->bytes_reserved);
3926         trans->bytes_reserved = 0;
3927 }
3928
3929 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3930                                   struct inode *inode)
3931 {
3932         struct btrfs_root *root = BTRFS_I(inode)->root;
3933         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3934         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3935
3936         /*
3937          * We need to hold space in order to delete our orphan item once we've
3938          * added it, so this takes the reservation so we can release it later
3939          * when we are truly done with the orphan item.
3940          */
3941         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3942         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3943 }
3944
3945 void btrfs_orphan_release_metadata(struct inode *inode)
3946 {
3947         struct btrfs_root *root = BTRFS_I(inode)->root;
3948         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3949         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3950 }
3951
3952 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3953                                 struct btrfs_pending_snapshot *pending)
3954 {
3955         struct btrfs_root *root = pending->root;
3956         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3957         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3958         /*
3959          * two for root back/forward refs, two for directory entries
3960          * and one for root of the snapshot.
3961          */
3962         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3963         dst_rsv->space_info = src_rsv->space_info;
3964         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3965 }
3966
3967 static unsigned drop_outstanding_extent(struct inode *inode)
3968 {
3969         unsigned dropped_extents = 0;
3970
3971         spin_lock(&BTRFS_I(inode)->lock);
3972         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
3973         BTRFS_I(inode)->outstanding_extents--;
3974
3975         /*
3976          * If we have more or the same amount of outsanding extents than we have
3977          * reserved then we need to leave the reserved extents count alone.
3978          */
3979         if (BTRFS_I(inode)->outstanding_extents >=
3980             BTRFS_I(inode)->reserved_extents)
3981                 goto out;
3982
3983         dropped_extents = BTRFS_I(inode)->reserved_extents -
3984                 BTRFS_I(inode)->outstanding_extents;
3985         BTRFS_I(inode)->reserved_extents -= dropped_extents;
3986 out:
3987         spin_unlock(&BTRFS_I(inode)->lock);
3988         return dropped_extents;
3989 }
3990
3991 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
3992 {
3993         return num_bytes >>= 3;
3994 }
3995
3996 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
3997 {
3998         struct btrfs_root *root = BTRFS_I(inode)->root;
3999         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4000         u64 to_reserve = 0;
4001         unsigned nr_extents = 0;
4002         int ret;
4003
4004         if (btrfs_transaction_in_commit(root->fs_info))
4005                 schedule_timeout(1);
4006
4007         num_bytes = ALIGN(num_bytes, root->sectorsize);
4008
4009         spin_lock(&BTRFS_I(inode)->lock);
4010         BTRFS_I(inode)->outstanding_extents++;
4011
4012         if (BTRFS_I(inode)->outstanding_extents >
4013             BTRFS_I(inode)->reserved_extents) {
4014                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4015                         BTRFS_I(inode)->reserved_extents;
4016                 BTRFS_I(inode)->reserved_extents += nr_extents;
4017
4018                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4019         }
4020         spin_unlock(&BTRFS_I(inode)->lock);
4021
4022         to_reserve += calc_csum_metadata_size(inode, num_bytes);
4023         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
4024         if (ret) {
4025                 unsigned dropped;
4026                 /*
4027                  * We don't need the return value since our reservation failed,
4028                  * we just need to clean up our counter.
4029                  */
4030                 dropped = drop_outstanding_extent(inode);
4031                 WARN_ON(dropped > 1);
4032                 return ret;
4033         }
4034
4035         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4036
4037         return 0;
4038 }
4039
4040 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4041 {
4042         struct btrfs_root *root = BTRFS_I(inode)->root;
4043         u64 to_free = 0;
4044         unsigned dropped;
4045
4046         num_bytes = ALIGN(num_bytes, root->sectorsize);
4047         dropped = drop_outstanding_extent(inode);
4048
4049         to_free = calc_csum_metadata_size(inode, num_bytes);
4050         if (dropped > 0)
4051                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4052
4053         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4054                                 to_free);
4055 }
4056
4057 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4058 {
4059         int ret;
4060
4061         ret = btrfs_check_data_free_space(inode, num_bytes);
4062         if (ret)
4063                 return ret;
4064
4065         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4066         if (ret) {
4067                 btrfs_free_reserved_data_space(inode, num_bytes);
4068                 return ret;
4069         }
4070
4071         return 0;
4072 }
4073
4074 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4075 {
4076         btrfs_delalloc_release_metadata(inode, num_bytes);
4077         btrfs_free_reserved_data_space(inode, num_bytes);
4078 }
4079
4080 static int update_block_group(struct btrfs_trans_handle *trans,
4081                               struct btrfs_root *root,
4082                               u64 bytenr, u64 num_bytes, int alloc)
4083 {
4084         struct btrfs_block_group_cache *cache = NULL;
4085         struct btrfs_fs_info *info = root->fs_info;
4086         u64 total = num_bytes;
4087         u64 old_val;
4088         u64 byte_in_group;
4089         int factor;
4090
4091         /* block accounting for super block */
4092         spin_lock(&info->delalloc_lock);
4093         old_val = btrfs_super_bytes_used(&info->super_copy);
4094         if (alloc)
4095                 old_val += num_bytes;
4096         else
4097                 old_val -= num_bytes;
4098         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4099         spin_unlock(&info->delalloc_lock);
4100
4101         while (total) {
4102                 cache = btrfs_lookup_block_group(info, bytenr);
4103                 if (!cache)
4104                         return -1;
4105                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4106                                     BTRFS_BLOCK_GROUP_RAID1 |
4107                                     BTRFS_BLOCK_GROUP_RAID10))
4108                         factor = 2;
4109                 else
4110                         factor = 1;
4111                 /*
4112                  * If this block group has free space cache written out, we
4113                  * need to make sure to load it if we are removing space.  This
4114                  * is because we need the unpinning stage to actually add the
4115                  * space back to the block group, otherwise we will leak space.
4116                  */
4117                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4118                         cache_block_group(cache, trans, NULL, 1);
4119
4120                 byte_in_group = bytenr - cache->key.objectid;
4121                 WARN_ON(byte_in_group > cache->key.offset);
4122
4123                 spin_lock(&cache->space_info->lock);
4124                 spin_lock(&cache->lock);
4125
4126                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4127                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4128                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4129
4130                 cache->dirty = 1;
4131                 old_val = btrfs_block_group_used(&cache->item);
4132                 num_bytes = min(total, cache->key.offset - byte_in_group);
4133                 if (alloc) {
4134                         old_val += num_bytes;
4135                         btrfs_set_block_group_used(&cache->item, old_val);
4136                         cache->reserved -= num_bytes;
4137                         cache->space_info->bytes_reserved -= num_bytes;
4138                         cache->space_info->reservation_progress++;
4139                         cache->space_info->bytes_used += num_bytes;
4140                         cache->space_info->disk_used += num_bytes * factor;
4141                         spin_unlock(&cache->lock);
4142                         spin_unlock(&cache->space_info->lock);
4143                 } else {
4144                         old_val -= num_bytes;
4145                         btrfs_set_block_group_used(&cache->item, old_val);
4146                         cache->pinned += num_bytes;
4147                         cache->space_info->bytes_pinned += num_bytes;
4148                         cache->space_info->bytes_used -= num_bytes;
4149                         cache->space_info->disk_used -= num_bytes * factor;
4150                         spin_unlock(&cache->lock);
4151                         spin_unlock(&cache->space_info->lock);
4152
4153                         set_extent_dirty(info->pinned_extents,
4154                                          bytenr, bytenr + num_bytes - 1,
4155                                          GFP_NOFS | __GFP_NOFAIL);
4156                 }
4157                 btrfs_put_block_group(cache);
4158                 total -= num_bytes;
4159                 bytenr += num_bytes;
4160         }
4161         return 0;
4162 }
4163
4164 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4165 {
4166         struct btrfs_block_group_cache *cache;
4167         u64 bytenr;
4168
4169         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4170         if (!cache)
4171                 return 0;
4172
4173         bytenr = cache->key.objectid;
4174         btrfs_put_block_group(cache);
4175
4176         return bytenr;
4177 }
4178
4179 static int pin_down_extent(struct btrfs_root *root,
4180                            struct btrfs_block_group_cache *cache,
4181                            u64 bytenr, u64 num_bytes, int reserved)
4182 {
4183         spin_lock(&cache->space_info->lock);
4184         spin_lock(&cache->lock);
4185         cache->pinned += num_bytes;
4186         cache->space_info->bytes_pinned += num_bytes;
4187         if (reserved) {
4188                 cache->reserved -= num_bytes;
4189                 cache->space_info->bytes_reserved -= num_bytes;
4190                 cache->space_info->reservation_progress++;
4191         }
4192         spin_unlock(&cache->lock);
4193         spin_unlock(&cache->space_info->lock);
4194
4195         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4196                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4197         return 0;
4198 }
4199
4200 /*
4201  * this function must be called within transaction
4202  */
4203 int btrfs_pin_extent(struct btrfs_root *root,
4204                      u64 bytenr, u64 num_bytes, int reserved)
4205 {
4206         struct btrfs_block_group_cache *cache;
4207
4208         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4209         BUG_ON(!cache);
4210
4211         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4212
4213         btrfs_put_block_group(cache);
4214         return 0;
4215 }
4216
4217 /*
4218  * update size of reserved extents. this function may return -EAGAIN
4219  * if 'reserve' is true or 'sinfo' is false.
4220  */
4221 int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4222                                 u64 num_bytes, int reserve, int sinfo)
4223 {
4224         int ret = 0;
4225         if (sinfo) {
4226                 struct btrfs_space_info *space_info = cache->space_info;
4227                 spin_lock(&space_info->lock);
4228                 spin_lock(&cache->lock);
4229                 if (reserve) {
4230                         if (cache->ro) {
4231                                 ret = -EAGAIN;
4232                         } else {
4233                                 cache->reserved += num_bytes;
4234                                 space_info->bytes_reserved += num_bytes;
4235                         }
4236                 } else {
4237                         if (cache->ro)
4238                                 space_info->bytes_readonly += num_bytes;
4239                         cache->reserved -= num_bytes;
4240                         space_info->bytes_reserved -= num_bytes;
4241                         space_info->reservation_progress++;
4242                 }
4243                 spin_unlock(&cache->lock);
4244                 spin_unlock(&space_info->lock);
4245         } else {
4246                 spin_lock(&cache->lock);
4247                 if (cache->ro) {
4248                         ret = -EAGAIN;
4249                 } else {
4250                         if (reserve)
4251                                 cache->reserved += num_bytes;
4252                         else
4253                                 cache->reserved -= num_bytes;
4254                 }
4255                 spin_unlock(&cache->lock);
4256         }
4257         return ret;
4258 }
4259
4260 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4261                                 struct btrfs_root *root)
4262 {
4263         struct btrfs_fs_info *fs_info = root->fs_info;
4264         struct btrfs_caching_control *next;
4265         struct btrfs_caching_control *caching_ctl;
4266         struct btrfs_block_group_cache *cache;
4267
4268         down_write(&fs_info->extent_commit_sem);
4269
4270         list_for_each_entry_safe(caching_ctl, next,
4271                                  &fs_info->caching_block_groups, list) {
4272                 cache = caching_ctl->block_group;
4273                 if (block_group_cache_done(cache)) {
4274                         cache->last_byte_to_unpin = (u64)-1;
4275                         list_del_init(&caching_ctl->list);
4276                         put_caching_control(caching_ctl);
4277                 } else {
4278                         cache->last_byte_to_unpin = caching_ctl->progress;
4279                 }
4280         }
4281
4282         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4283                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4284         else
4285                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4286
4287         up_write(&fs_info->extent_commit_sem);
4288
4289         update_global_block_rsv(fs_info);
4290         return 0;
4291 }
4292
4293 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4294 {
4295         struct btrfs_fs_info *fs_info = root->fs_info;
4296         struct btrfs_block_group_cache *cache = NULL;
4297         u64 len;
4298
4299         while (start <= end) {
4300                 if (!cache ||
4301                     start >= cache->key.objectid + cache->key.offset) {
4302                         if (cache)
4303                                 btrfs_put_block_group(cache);
4304                         cache = btrfs_lookup_block_group(fs_info, start);
4305                         BUG_ON(!cache);
4306                 }
4307
4308                 len = cache->key.objectid + cache->key.offset - start;
4309                 len = min(len, end + 1 - start);
4310
4311                 if (start < cache->last_byte_to_unpin) {
4312                         len = min(len, cache->last_byte_to_unpin - start);
4313                         btrfs_add_free_space(cache, start, len);
4314                 }
4315
4316                 start += len;
4317
4318                 spin_lock(&cache->space_info->lock);
4319                 spin_lock(&cache->lock);
4320                 cache->pinned -= len;
4321                 cache->space_info->bytes_pinned -= len;
4322                 if (cache->ro) {
4323                         cache->space_info->bytes_readonly += len;
4324                 } else if (cache->reserved_pinned > 0) {
4325                         len = min(len, cache->reserved_pinned);
4326                         cache->reserved_pinned -= len;
4327                         cache->space_info->bytes_reserved += len;
4328                 }
4329                 spin_unlock(&cache->lock);
4330                 spin_unlock(&cache->space_info->lock);
4331         }
4332
4333         if (cache)
4334                 btrfs_put_block_group(cache);
4335         return 0;
4336 }
4337
4338 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4339                                struct btrfs_root *root)
4340 {
4341         struct btrfs_fs_info *fs_info = root->fs_info;
4342         struct extent_io_tree *unpin;
4343         struct btrfs_block_rsv *block_rsv;
4344         struct btrfs_block_rsv *next_rsv;
4345         u64 start;
4346         u64 end;
4347         int idx;
4348         int ret;
4349
4350         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4351                 unpin = &fs_info->freed_extents[1];
4352         else
4353                 unpin = &fs_info->freed_extents[0];
4354
4355         while (1) {
4356                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4357                                             EXTENT_DIRTY);
4358                 if (ret)
4359                         break;
4360
4361                 if (btrfs_test_opt(root, DISCARD))
4362                         ret = btrfs_discard_extent(root, start,
4363                                                    end + 1 - start, NULL);
4364
4365                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4366                 unpin_extent_range(root, start, end);
4367                 cond_resched();
4368         }
4369
4370         mutex_lock(&fs_info->durable_block_rsv_mutex);
4371         list_for_each_entry_safe(block_rsv, next_rsv,
4372                                  &fs_info->durable_block_rsv_list, list) {
4373
4374                 idx = trans->transid & 0x1;
4375                 if (block_rsv->freed[idx] > 0) {
4376                         block_rsv_add_bytes(block_rsv,
4377                                             block_rsv->freed[idx], 0);
4378                         block_rsv->freed[idx] = 0;
4379                 }
4380                 if (atomic_read(&block_rsv->usage) == 0) {
4381                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4382
4383                         if (block_rsv->freed[0] == 0 &&
4384                             block_rsv->freed[1] == 0) {
4385                                 list_del_init(&block_rsv->list);
4386                                 kfree(block_rsv);
4387                         }
4388                 } else {
4389                         btrfs_block_rsv_release(root, block_rsv, 0);
4390                 }
4391         }
4392         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4393
4394         return 0;
4395 }
4396
4397 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4398                                 struct btrfs_root *root,
4399                                 u64 bytenr, u64 num_bytes, u64 parent,
4400                                 u64 root_objectid, u64 owner_objectid,
4401                                 u64 owner_offset, int refs_to_drop,
4402                                 struct btrfs_delayed_extent_op *extent_op)
4403 {
4404         struct btrfs_key key;
4405         struct btrfs_path *path;
4406         struct btrfs_fs_info *info = root->fs_info;
4407         struct btrfs_root *extent_root = info->extent_root;
4408         struct extent_buffer *leaf;
4409         struct btrfs_extent_item *ei;
4410         struct btrfs_extent_inline_ref *iref;
4411         int ret;
4412         int is_data;
4413         int extent_slot = 0;
4414         int found_extent = 0;
4415         int num_to_del = 1;
4416         u32 item_size;
4417         u64 refs;
4418
4419         path = btrfs_alloc_path();
4420         if (!path)
4421                 return -ENOMEM;
4422
4423         path->reada = 1;
4424         path->leave_spinning = 1;
4425
4426         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4427         BUG_ON(!is_data && refs_to_drop != 1);
4428
4429         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4430                                     bytenr, num_bytes, parent,
4431                                     root_objectid, owner_objectid,
4432                                     owner_offset);
4433         if (ret == 0) {
4434                 extent_slot = path->slots[0];
4435                 while (extent_slot >= 0) {
4436                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4437                                               extent_slot);
4438                         if (key.objectid != bytenr)
4439                                 break;
4440                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4441                             key.offset == num_bytes) {
4442                                 found_extent = 1;
4443                                 break;
4444                         }
4445                         if (path->slots[0] - extent_slot > 5)
4446                                 break;
4447                         extent_slot--;
4448                 }
4449 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4450                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4451                 if (found_extent && item_size < sizeof(*ei))
4452                         found_extent = 0;
4453 #endif
4454                 if (!found_extent) {
4455                         BUG_ON(iref);
4456                         ret = remove_extent_backref(trans, extent_root, path,
4457                                                     NULL, refs_to_drop,
4458                                                     is_data);
4459                         BUG_ON(ret);
4460                         btrfs_release_path(path);
4461                         path->leave_spinning = 1;
4462
4463                         key.objectid = bytenr;
4464                         key.type = BTRFS_EXTENT_ITEM_KEY;
4465                         key.offset = num_bytes;
4466
4467                         ret = btrfs_search_slot(trans, extent_root,
4468                                                 &key, path, -1, 1);
4469                         if (ret) {
4470                                 printk(KERN_ERR "umm, got %d back from search"
4471                                        ", was looking for %llu\n", ret,
4472                                        (unsigned long long)bytenr);
4473                                 if (ret > 0)
4474                                         btrfs_print_leaf(extent_root,
4475                                                          path->nodes[0]);
4476                         }
4477                         BUG_ON(ret);
4478                         extent_slot = path->slots[0];
4479                 }
4480         } else {
4481                 btrfs_print_leaf(extent_root, path->nodes[0]);
4482                 WARN_ON(1);
4483                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4484                        "parent %llu root %llu  owner %llu offset %llu\n",
4485                        (unsigned long long)bytenr,
4486                        (unsigned long long)parent,
4487                        (unsigned long long)root_objectid,
4488                        (unsigned long long)owner_objectid,
4489                        (unsigned long long)owner_offset);
4490         }
4491
4492         leaf = path->nodes[0];
4493         item_size = btrfs_item_size_nr(leaf, extent_slot);
4494 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4495         if (item_size < sizeof(*ei)) {
4496                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4497                 ret = convert_extent_item_v0(trans, extent_root, path,
4498                                              owner_objectid, 0);
4499                 BUG_ON(ret < 0);
4500
4501                 btrfs_release_path(path);
4502                 path->leave_spinning = 1;
4503
4504                 key.objectid = bytenr;
4505                 key.type = BTRFS_EXTENT_ITEM_KEY;
4506                 key.offset = num_bytes;
4507
4508                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4509                                         -1, 1);
4510                 if (ret) {
4511                         printk(KERN_ERR "umm, got %d back from search"
4512                                ", was looking for %llu\n", ret,
4513                                (unsigned long long)bytenr);
4514                         btrfs_print_leaf(extent_root, path->nodes[0]);
4515                 }
4516                 BUG_ON(ret);
4517                 extent_slot = path->slots[0];
4518                 leaf = path->nodes[0];
4519                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4520         }
4521 #endif
4522         BUG_ON(item_size < sizeof(*ei));
4523         ei = btrfs_item_ptr(leaf, extent_slot,
4524                             struct btrfs_extent_item);
4525         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4526                 struct btrfs_tree_block_info *bi;
4527                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4528                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4529                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4530         }
4531
4532         refs = btrfs_extent_refs(leaf, ei);
4533         BUG_ON(refs < refs_to_drop);
4534         refs -= refs_to_drop;
4535
4536         if (refs > 0) {
4537                 if (extent_op)
4538                         __run_delayed_extent_op(extent_op, leaf, ei);
4539                 /*
4540                  * In the case of inline back ref, reference count will
4541                  * be updated by remove_extent_backref
4542                  */
4543                 if (iref) {
4544                         BUG_ON(!found_extent);
4545                 } else {
4546                         btrfs_set_extent_refs(leaf, ei, refs);
4547                         btrfs_mark_buffer_dirty(leaf);
4548                 }
4549                 if (found_extent) {
4550                         ret = remove_extent_backref(trans, extent_root, path,
4551                                                     iref, refs_to_drop,
4552                                                     is_data);
4553                         BUG_ON(ret);
4554                 }
4555         } else {
4556                 if (found_extent) {
4557                         BUG_ON(is_data && refs_to_drop !=
4558                                extent_data_ref_count(root, path, iref));
4559                         if (iref) {
4560                                 BUG_ON(path->slots[0] != extent_slot);
4561                         } else {
4562                                 BUG_ON(path->slots[0] != extent_slot + 1);
4563                                 path->slots[0] = extent_slot;
4564                                 num_to_del = 2;
4565                         }
4566                 }
4567
4568                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4569                                       num_to_del);
4570                 BUG_ON(ret);
4571                 btrfs_release_path(path);
4572
4573                 if (is_data) {
4574                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4575                         BUG_ON(ret);
4576                 } else {
4577                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4578                              bytenr >> PAGE_CACHE_SHIFT,
4579                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4580                 }
4581
4582                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4583                 BUG_ON(ret);
4584         }
4585         btrfs_free_path(path);
4586         return ret;
4587 }
4588
4589 /*
4590  * when we free an block, it is possible (and likely) that we free the last
4591  * delayed ref for that extent as well.  This searches the delayed ref tree for
4592  * a given extent, and if there are no other delayed refs to be processed, it
4593  * removes it from the tree.
4594  */
4595 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4596                                       struct btrfs_root *root, u64 bytenr)
4597 {
4598         struct btrfs_delayed_ref_head *head;
4599         struct btrfs_delayed_ref_root *delayed_refs;
4600         struct btrfs_delayed_ref_node *ref;
4601         struct rb_node *node;
4602         int ret = 0;
4603
4604         delayed_refs = &trans->transaction->delayed_refs;
4605         spin_lock(&delayed_refs->lock);
4606         head = btrfs_find_delayed_ref_head(trans, bytenr);
4607         if (!head)
4608                 goto out;
4609
4610         node = rb_prev(&head->node.rb_node);
4611         if (!node)
4612                 goto out;
4613
4614         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4615
4616         /* there are still entries for this ref, we can't drop it */
4617         if (ref->bytenr == bytenr)
4618                 goto out;
4619
4620         if (head->extent_op) {
4621                 if (!head->must_insert_reserved)
4622                         goto out;
4623                 kfree(head->extent_op);
4624                 head->extent_op = NULL;
4625         }
4626
4627         /*
4628          * waiting for the lock here would deadlock.  If someone else has it
4629          * locked they are already in the process of dropping it anyway
4630          */
4631         if (!mutex_trylock(&head->mutex))
4632                 goto out;
4633
4634         /*
4635          * at this point we have a head with no other entries.  Go
4636          * ahead and process it.
4637          */
4638         head->node.in_tree = 0;
4639         rb_erase(&head->node.rb_node, &delayed_refs->root);
4640
4641         delayed_refs->num_entries--;
4642
4643         /*
4644          * we don't take a ref on the node because we're removing it from the
4645          * tree, so we just steal the ref the tree was holding.
4646          */
4647         delayed_refs->num_heads--;
4648         if (list_empty(&head->cluster))
4649                 delayed_refs->num_heads_ready--;
4650
4651         list_del_init(&head->cluster);
4652         spin_unlock(&delayed_refs->lock);
4653
4654         BUG_ON(head->extent_op);
4655         if (head->must_insert_reserved)
4656                 ret = 1;
4657
4658         mutex_unlock(&head->mutex);
4659         btrfs_put_delayed_ref(&head->node);
4660         return ret;
4661 out:
4662         spin_unlock(&delayed_refs->lock);
4663         return 0;
4664 }
4665
4666 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4667                            struct btrfs_root *root,
4668                            struct extent_buffer *buf,
4669                            u64 parent, int last_ref)
4670 {
4671         struct btrfs_block_rsv *block_rsv;
4672         struct btrfs_block_group_cache *cache = NULL;
4673         int ret;
4674
4675         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4676                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4677                                                 parent, root->root_key.objectid,
4678                                                 btrfs_header_level(buf),
4679                                                 BTRFS_DROP_DELAYED_REF, NULL);
4680                 BUG_ON(ret);
4681         }
4682
4683         if (!last_ref)
4684                 return;
4685
4686         block_rsv = get_block_rsv(trans, root);
4687         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4688         if (block_rsv->space_info != cache->space_info)
4689                 goto out;
4690
4691         if (btrfs_header_generation(buf) == trans->transid) {
4692                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4693                         ret = check_ref_cleanup(trans, root, buf->start);
4694                         if (!ret)
4695                                 goto pin;
4696                 }
4697
4698                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4699                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4700                         goto pin;
4701                 }
4702
4703                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4704
4705                 btrfs_add_free_space(cache, buf->start, buf->len);
4706                 ret = btrfs_update_reserved_bytes(cache, buf->len, 0, 0);
4707                 if (ret == -EAGAIN) {
4708                         /* block group became read-only */
4709                         btrfs_update_reserved_bytes(cache, buf->len, 0, 1);
4710                         goto out;
4711                 }
4712
4713                 ret = 1;
4714                 spin_lock(&block_rsv->lock);
4715                 if (block_rsv->reserved < block_rsv->size) {
4716                         block_rsv->reserved += buf->len;
4717                         ret = 0;
4718                 }
4719                 spin_unlock(&block_rsv->lock);
4720
4721                 if (ret) {
4722                         spin_lock(&cache->space_info->lock);
4723                         cache->space_info->bytes_reserved -= buf->len;
4724                         cache->space_info->reservation_progress++;
4725                         spin_unlock(&cache->space_info->lock);
4726                 }
4727                 goto out;
4728         }
4729 pin:
4730         if (block_rsv->durable && !cache->ro) {
4731                 ret = 0;
4732                 spin_lock(&cache->lock);
4733                 if (!cache->ro) {
4734                         cache->reserved_pinned += buf->len;
4735                         ret = 1;
4736                 }
4737                 spin_unlock(&cache->lock);
4738
4739                 if (ret) {
4740                         spin_lock(&block_rsv->lock);
4741                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4742                         spin_unlock(&block_rsv->lock);
4743                 }
4744         }
4745 out:
4746         /*
4747          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4748          * anymore.
4749          */
4750         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4751         btrfs_put_block_group(cache);
4752 }
4753
4754 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4755                       struct btrfs_root *root,
4756                       u64 bytenr, u64 num_bytes, u64 parent,
4757                       u64 root_objectid, u64 owner, u64 offset)
4758 {
4759         int ret;
4760
4761         /*
4762          * tree log blocks never actually go into the extent allocation
4763          * tree, just update pinning info and exit early.
4764          */
4765         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4766                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4767                 /* unlocks the pinned mutex */
4768                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4769                 ret = 0;
4770         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4771                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4772                                         parent, root_objectid, (int)owner,
4773                                         BTRFS_DROP_DELAYED_REF, NULL);
4774                 BUG_ON(ret);
4775         } else {
4776                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4777                                         parent, root_objectid, owner,
4778                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4779                 BUG_ON(ret);
4780         }
4781         return ret;
4782 }
4783
4784 static u64 stripe_align(struct btrfs_root *root, u64 val)
4785 {
4786         u64 mask = ((u64)root->stripesize - 1);
4787         u64 ret = (val + mask) & ~mask;
4788         return ret;
4789 }
4790
4791 /*
4792  * when we wait for progress in the block group caching, its because
4793  * our allocation attempt failed at least once.  So, we must sleep
4794  * and let some progress happen before we try again.
4795  *
4796  * This function will sleep at least once waiting for new free space to
4797  * show up, and then it will check the block group free space numbers
4798  * for our min num_bytes.  Another option is to have it go ahead
4799  * and look in the rbtree for a free extent of a given size, but this
4800  * is a good start.
4801  */
4802 static noinline int
4803 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4804                                 u64 num_bytes)
4805 {
4806         struct btrfs_caching_control *caching_ctl;
4807         DEFINE_WAIT(wait);
4808
4809         caching_ctl = get_caching_control(cache);
4810         if (!caching_ctl)
4811                 return 0;
4812
4813         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4814                    (cache->free_space_ctl->free_space >= num_bytes));
4815
4816         put_caching_control(caching_ctl);
4817         return 0;
4818 }
4819
4820 static noinline int
4821 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4822 {
4823         struct btrfs_caching_control *caching_ctl;
4824         DEFINE_WAIT(wait);
4825
4826         caching_ctl = get_caching_control(cache);
4827         if (!caching_ctl)
4828                 return 0;
4829
4830         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4831
4832         put_caching_control(caching_ctl);
4833         return 0;
4834 }
4835
4836 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4837 {
4838         int index;
4839         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4840                 index = 0;
4841         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4842                 index = 1;
4843         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4844                 index = 2;
4845         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4846                 index = 3;
4847         else
4848                 index = 4;
4849         return index;
4850 }
4851
4852 enum btrfs_loop_type {
4853         LOOP_FIND_IDEAL = 0,
4854         LOOP_CACHING_NOWAIT = 1,
4855         LOOP_CACHING_WAIT = 2,
4856         LOOP_ALLOC_CHUNK = 3,
4857         LOOP_NO_EMPTY_SIZE = 4,
4858 };
4859
4860 /*
4861  * walks the btree of allocated extents and find a hole of a given size.
4862  * The key ins is changed to record the hole:
4863  * ins->objectid == block start
4864  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4865  * ins->offset == number of blocks
4866  * Any available blocks before search_start are skipped.
4867  */
4868 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4869                                      struct btrfs_root *orig_root,
4870                                      u64 num_bytes, u64 empty_size,
4871                                      u64 search_start, u64 search_end,
4872                                      u64 hint_byte, struct btrfs_key *ins,
4873                                      u64 data)
4874 {
4875         int ret = 0;
4876         struct btrfs_root *root = orig_root->fs_info->extent_root;
4877         struct btrfs_free_cluster *last_ptr = NULL;
4878         struct btrfs_block_group_cache *block_group = NULL;
4879         int empty_cluster = 2 * 1024 * 1024;
4880         int allowed_chunk_alloc = 0;
4881         int done_chunk_alloc = 0;
4882         struct btrfs_space_info *space_info;
4883         int last_ptr_loop = 0;
4884         int loop = 0;
4885         int index = 0;
4886         bool found_uncached_bg = false;
4887         bool failed_cluster_refill = false;
4888         bool failed_alloc = false;
4889         bool use_cluster = true;
4890         u64 ideal_cache_percent = 0;
4891         u64 ideal_cache_offset = 0;
4892
4893         WARN_ON(num_bytes < root->sectorsize);
4894         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4895         ins->objectid = 0;
4896         ins->offset = 0;
4897
4898         space_info = __find_space_info(root->fs_info, data);
4899         if (!space_info) {
4900                 printk(KERN_ERR "No space info for %llu\n", data);
4901                 return -ENOSPC;
4902         }
4903
4904         /*
4905          * If the space info is for both data and metadata it means we have a
4906          * small filesystem and we can't use the clustering stuff.
4907          */
4908         if (btrfs_mixed_space_info(space_info))
4909                 use_cluster = false;
4910
4911         if (orig_root->ref_cows || empty_size)
4912                 allowed_chunk_alloc = 1;
4913
4914         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4915                 last_ptr = &root->fs_info->meta_alloc_cluster;
4916                 if (!btrfs_test_opt(root, SSD))
4917                         empty_cluster = 64 * 1024;
4918         }
4919
4920         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4921             btrfs_test_opt(root, SSD)) {
4922                 last_ptr = &root->fs_info->data_alloc_cluster;
4923         }
4924
4925         if (last_ptr) {
4926                 spin_lock(&last_ptr->lock);
4927                 if (last_ptr->block_group)
4928                         hint_byte = last_ptr->window_start;
4929                 spin_unlock(&last_ptr->lock);
4930         }
4931
4932         search_start = max(search_start, first_logical_byte(root, 0));
4933         search_start = max(search_start, hint_byte);
4934
4935         if (!last_ptr)
4936                 empty_cluster = 0;
4937
4938         if (search_start == hint_byte) {
4939 ideal_cache:
4940                 block_group = btrfs_lookup_block_group(root->fs_info,
4941                                                        search_start);
4942                 /*
4943                  * we don't want to use the block group if it doesn't match our
4944                  * allocation bits, or if its not cached.
4945                  *
4946                  * However if we are re-searching with an ideal block group
4947                  * picked out then we don't care that the block group is cached.
4948                  */
4949                 if (block_group && block_group_bits(block_group, data) &&
4950                     (block_group->cached != BTRFS_CACHE_NO ||
4951                      search_start == ideal_cache_offset)) {
4952                         down_read(&space_info->groups_sem);
4953                         if (list_empty(&block_group->list) ||
4954                             block_group->ro) {
4955                                 /*
4956                                  * someone is removing this block group,
4957                                  * we can't jump into the have_block_group
4958                                  * target because our list pointers are not
4959                                  * valid
4960                                  */
4961                                 btrfs_put_block_group(block_group);
4962                                 up_read(&space_info->groups_sem);
4963                         } else {
4964                                 index = get_block_group_index(block_group);
4965                                 goto have_block_group;
4966                         }
4967                 } else if (block_group) {
4968                         btrfs_put_block_group(block_group);
4969                 }
4970         }
4971 search:
4972         down_read(&space_info->groups_sem);
4973         list_for_each_entry(block_group, &space_info->block_groups[index],
4974                             list) {
4975                 u64 offset;
4976                 int cached;
4977
4978                 btrfs_get_block_group(block_group);
4979                 search_start = block_group->key.objectid;
4980
4981                 /*
4982                  * this can happen if we end up cycling through all the
4983                  * raid types, but we want to make sure we only allocate
4984                  * for the proper type.
4985                  */
4986                 if (!block_group_bits(block_group, data)) {
4987                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
4988                                 BTRFS_BLOCK_GROUP_RAID1 |
4989                                 BTRFS_BLOCK_GROUP_RAID10;
4990
4991                         /*
4992                          * if they asked for extra copies and this block group
4993                          * doesn't provide them, bail.  This does allow us to
4994                          * fill raid0 from raid1.
4995                          */
4996                         if ((data & extra) && !(block_group->flags & extra))
4997                                 goto loop;
4998                 }
4999
5000 have_block_group:
5001                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5002                         u64 free_percent;
5003
5004                         ret = cache_block_group(block_group, trans,
5005                                                 orig_root, 1);
5006                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5007                                 goto have_block_group;
5008
5009                         free_percent = btrfs_block_group_used(&block_group->item);
5010                         free_percent *= 100;
5011                         free_percent = div64_u64(free_percent,
5012                                                  block_group->key.offset);
5013                         free_percent = 100 - free_percent;
5014                         if (free_percent > ideal_cache_percent &&
5015                             likely(!block_group->ro)) {
5016                                 ideal_cache_offset = block_group->key.objectid;
5017                                 ideal_cache_percent = free_percent;
5018                         }
5019
5020                         /*
5021                          * The caching workers are limited to 2 threads, so we
5022                          * can queue as much work as we care to.
5023                          */
5024                         if (loop > LOOP_FIND_IDEAL) {
5025                                 ret = cache_block_group(block_group, trans,
5026                                                         orig_root, 0);
5027                                 BUG_ON(ret);
5028                         }
5029                         found_uncached_bg = true;
5030
5031                         /*
5032                          * If loop is set for cached only, try the next block
5033                          * group.
5034                          */
5035                         if (loop == LOOP_FIND_IDEAL)
5036                                 goto loop;
5037                 }
5038
5039                 cached = block_group_cache_done(block_group);
5040                 if (unlikely(!cached))
5041                         found_uncached_bg = true;
5042
5043                 if (unlikely(block_group->ro))
5044                         goto loop;
5045
5046                 spin_lock(&block_group->free_space_ctl->tree_lock);
5047                 if (cached &&
5048                     block_group->free_space_ctl->free_space <
5049                     num_bytes + empty_size) {
5050                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5051                         goto loop;
5052                 }
5053                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5054
5055                 /*
5056                  * Ok we want to try and use the cluster allocator, so lets look
5057                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5058                  * have tried the cluster allocator plenty of times at this
5059                  * point and not have found anything, so we are likely way too
5060                  * fragmented for the clustering stuff to find anything, so lets
5061                  * just skip it and let the allocator find whatever block it can
5062                  * find
5063                  */
5064                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5065                         /*
5066                          * the refill lock keeps out other
5067                          * people trying to start a new cluster
5068                          */
5069                         spin_lock(&last_ptr->refill_lock);
5070                         if (last_ptr->block_group &&
5071                             (last_ptr->block_group->ro ||
5072                             !block_group_bits(last_ptr->block_group, data))) {
5073                                 offset = 0;
5074                                 goto refill_cluster;
5075                         }
5076
5077                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5078                                                  num_bytes, search_start);
5079                         if (offset) {
5080                                 /* we have a block, we're done */
5081                                 spin_unlock(&last_ptr->refill_lock);
5082                                 goto checks;
5083                         }
5084
5085                         spin_lock(&last_ptr->lock);
5086                         /*
5087                          * whoops, this cluster doesn't actually point to
5088                          * this block group.  Get a ref on the block
5089                          * group is does point to and try again
5090                          */
5091                         if (!last_ptr_loop && last_ptr->block_group &&
5092                             last_ptr->block_group != block_group &&
5093                             index <=
5094                                  get_block_group_index(last_ptr->block_group)) {
5095
5096                                 btrfs_put_block_group(block_group);
5097                                 block_group = last_ptr->block_group;
5098                                 btrfs_get_block_group(block_group);
5099                                 spin_unlock(&last_ptr->lock);
5100                                 spin_unlock(&last_ptr->refill_lock);
5101
5102                                 last_ptr_loop = 1;
5103                                 search_start = block_group->key.objectid;
5104                                 /*
5105                                  * we know this block group is properly
5106                                  * in the list because
5107                                  * btrfs_remove_block_group, drops the
5108                                  * cluster before it removes the block
5109                                  * group from the list
5110                                  */
5111                                 goto have_block_group;
5112                         }
5113                         spin_unlock(&last_ptr->lock);
5114 refill_cluster:
5115                         /*
5116                          * this cluster didn't work out, free it and
5117                          * start over
5118                          */
5119                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5120
5121                         last_ptr_loop = 0;
5122
5123                         /* allocate a cluster in this block group */
5124                         ret = btrfs_find_space_cluster(trans, root,
5125                                                block_group, last_ptr,
5126                                                offset, num_bytes,
5127                                                empty_cluster + empty_size);
5128                         if (ret == 0) {
5129                                 /*
5130                                  * now pull our allocation out of this
5131                                  * cluster
5132                                  */
5133                                 offset = btrfs_alloc_from_cluster(block_group,
5134                                                   last_ptr, num_bytes,
5135                                                   search_start);
5136                                 if (offset) {
5137                                         /* we found one, proceed */
5138                                         spin_unlock(&last_ptr->refill_lock);
5139                                         goto checks;
5140                                 }
5141                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5142                                    && !failed_cluster_refill) {
5143                                 spin_unlock(&last_ptr->refill_lock);
5144
5145                                 failed_cluster_refill = true;
5146                                 wait_block_group_cache_progress(block_group,
5147                                        num_bytes + empty_cluster + empty_size);
5148                                 goto have_block_group;
5149                         }
5150
5151                         /*
5152                          * at this point we either didn't find a cluster
5153                          * or we weren't able to allocate a block from our
5154                          * cluster.  Free the cluster we've been trying
5155                          * to use, and go to the next block group
5156                          */
5157                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5158                         spin_unlock(&last_ptr->refill_lock);
5159                         goto loop;
5160                 }
5161
5162                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5163                                                     num_bytes, empty_size);
5164                 /*
5165                  * If we didn't find a chunk, and we haven't failed on this
5166                  * block group before, and this block group is in the middle of
5167                  * caching and we are ok with waiting, then go ahead and wait
5168                  * for progress to be made, and set failed_alloc to true.
5169                  *
5170                  * If failed_alloc is true then we've already waited on this
5171                  * block group once and should move on to the next block group.
5172                  */
5173                 if (!offset && !failed_alloc && !cached &&
5174                     loop > LOOP_CACHING_NOWAIT) {
5175                         wait_block_group_cache_progress(block_group,
5176                                                 num_bytes + empty_size);
5177                         failed_alloc = true;
5178                         goto have_block_group;
5179                 } else if (!offset) {
5180                         goto loop;
5181                 }
5182 checks:
5183                 search_start = stripe_align(root, offset);
5184                 /* move on to the next group */
5185                 if (search_start + num_bytes >= search_end) {
5186                         btrfs_add_free_space(block_group, offset, num_bytes);
5187                         goto loop;
5188                 }
5189
5190                 /* move on to the next group */
5191                 if (search_start + num_bytes >
5192                     block_group->key.objectid + block_group->key.offset) {
5193                         btrfs_add_free_space(block_group, offset, num_bytes);
5194                         goto loop;
5195                 }
5196
5197                 ins->objectid = search_start;
5198                 ins->offset = num_bytes;
5199
5200                 if (offset < search_start)
5201                         btrfs_add_free_space(block_group, offset,
5202                                              search_start - offset);
5203                 BUG_ON(offset > search_start);
5204
5205                 ret = btrfs_update_reserved_bytes(block_group, num_bytes, 1,
5206                                             (data & BTRFS_BLOCK_GROUP_DATA));
5207                 if (ret == -EAGAIN) {
5208                         btrfs_add_free_space(block_group, offset, num_bytes);
5209                         goto loop;
5210                 }
5211
5212                 /* we are all good, lets return */
5213                 ins->objectid = search_start;
5214                 ins->offset = num_bytes;
5215
5216                 if (offset < search_start)
5217                         btrfs_add_free_space(block_group, offset,
5218                                              search_start - offset);
5219                 BUG_ON(offset > search_start);
5220                 btrfs_put_block_group(block_group);
5221                 break;
5222 loop:
5223                 failed_cluster_refill = false;
5224                 failed_alloc = false;
5225                 BUG_ON(index != get_block_group_index(block_group));
5226                 btrfs_put_block_group(block_group);
5227         }
5228         up_read(&space_info->groups_sem);
5229
5230         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5231                 goto search;
5232
5233         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5234          *                      for them to make caching progress.  Also
5235          *                      determine the best possible bg to cache
5236          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5237          *                      caching kthreads as we move along
5238          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5239          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5240          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5241          *                      again
5242          */
5243         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5244                 index = 0;
5245                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5246                         found_uncached_bg = false;
5247                         loop++;
5248                         if (!ideal_cache_percent)
5249                                 goto search;
5250
5251                         /*
5252                          * 1 of the following 2 things have happened so far
5253                          *
5254                          * 1) We found an ideal block group for caching that
5255                          * is mostly full and will cache quickly, so we might
5256                          * as well wait for it.
5257                          *
5258                          * 2) We searched for cached only and we didn't find
5259                          * anything, and we didn't start any caching kthreads
5260                          * either, so chances are we will loop through and
5261                          * start a couple caching kthreads, and then come back
5262                          * around and just wait for them.  This will be slower
5263                          * because we will have 2 caching kthreads reading at
5264                          * the same time when we could have just started one
5265                          * and waited for it to get far enough to give us an
5266                          * allocation, so go ahead and go to the wait caching
5267                          * loop.
5268                          */
5269                         loop = LOOP_CACHING_WAIT;
5270                         search_start = ideal_cache_offset;
5271                         ideal_cache_percent = 0;
5272                         goto ideal_cache;
5273                 } else if (loop == LOOP_FIND_IDEAL) {
5274                         /*
5275                          * Didn't find a uncached bg, wait on anything we find
5276                          * next.
5277                          */
5278                         loop = LOOP_CACHING_WAIT;
5279                         goto search;
5280                 }
5281
5282                 loop++;
5283
5284                 if (loop == LOOP_ALLOC_CHUNK) {
5285                        if (allowed_chunk_alloc) {
5286                                 ret = do_chunk_alloc(trans, root, num_bytes +
5287                                                      2 * 1024 * 1024, data,
5288                                                      CHUNK_ALLOC_LIMITED);
5289                                 allowed_chunk_alloc = 0;
5290                                 if (ret == 1)
5291                                         done_chunk_alloc = 1;
5292                         } else if (!done_chunk_alloc &&
5293                                    space_info->force_alloc ==
5294                                    CHUNK_ALLOC_NO_FORCE) {
5295                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5296                         }
5297
5298                        /*
5299                         * We didn't allocate a chunk, go ahead and drop the
5300                         * empty size and loop again.
5301                         */
5302                        if (!done_chunk_alloc)
5303                                loop = LOOP_NO_EMPTY_SIZE;
5304                 }
5305
5306                 if (loop == LOOP_NO_EMPTY_SIZE) {
5307                         empty_size = 0;
5308                         empty_cluster = 0;
5309                 }
5310
5311                 goto search;
5312         } else if (!ins->objectid) {
5313                 ret = -ENOSPC;
5314         } else if (ins->objectid) {
5315                 ret = 0;
5316         }
5317
5318         return ret;
5319 }
5320
5321 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5322                             int dump_block_groups)
5323 {
5324         struct btrfs_block_group_cache *cache;
5325         int index = 0;
5326
5327         spin_lock(&info->lock);
5328         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5329                (unsigned long long)(info->total_bytes - info->bytes_used -
5330                                     info->bytes_pinned - info->bytes_reserved -
5331                                     info->bytes_readonly),
5332                (info->full) ? "" : "not ");
5333         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5334                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5335                (unsigned long long)info->total_bytes,
5336                (unsigned long long)info->bytes_used,
5337                (unsigned long long)info->bytes_pinned,
5338                (unsigned long long)info->bytes_reserved,
5339                (unsigned long long)info->bytes_may_use,
5340                (unsigned long long)info->bytes_readonly);
5341         spin_unlock(&info->lock);
5342
5343         if (!dump_block_groups)
5344                 return;
5345
5346         down_read(&info->groups_sem);
5347 again:
5348         list_for_each_entry(cache, &info->block_groups[index], list) {
5349                 spin_lock(&cache->lock);
5350                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5351                        "%llu pinned %llu reserved\n",
5352                        (unsigned long long)cache->key.objectid,
5353                        (unsigned long long)cache->key.offset,
5354                        (unsigned long long)btrfs_block_group_used(&cache->item),
5355                        (unsigned long long)cache->pinned,
5356                        (unsigned long long)cache->reserved);
5357                 btrfs_dump_free_space(cache, bytes);
5358                 spin_unlock(&cache->lock);
5359         }
5360         if (++index < BTRFS_NR_RAID_TYPES)
5361                 goto again;
5362         up_read(&info->groups_sem);
5363 }
5364
5365 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5366                          struct btrfs_root *root,
5367                          u64 num_bytes, u64 min_alloc_size,
5368                          u64 empty_size, u64 hint_byte,
5369                          u64 search_end, struct btrfs_key *ins,
5370                          u64 data)
5371 {
5372         int ret;
5373         u64 search_start = 0;
5374
5375         data = btrfs_get_alloc_profile(root, data);
5376 again:
5377         /*
5378          * the only place that sets empty_size is btrfs_realloc_node, which
5379          * is not called recursively on allocations
5380          */
5381         if (empty_size || root->ref_cows)
5382                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5383                                      num_bytes + 2 * 1024 * 1024, data,
5384                                      CHUNK_ALLOC_NO_FORCE);
5385
5386         WARN_ON(num_bytes < root->sectorsize);
5387         ret = find_free_extent(trans, root, num_bytes, empty_size,
5388                                search_start, search_end, hint_byte,
5389                                ins, data);
5390
5391         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5392                 num_bytes = num_bytes >> 1;
5393                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5394                 num_bytes = max(num_bytes, min_alloc_size);
5395                 do_chunk_alloc(trans, root->fs_info->extent_root,
5396                                num_bytes, data, CHUNK_ALLOC_FORCE);
5397                 goto again;
5398         }
5399         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5400                 struct btrfs_space_info *sinfo;
5401
5402                 sinfo = __find_space_info(root->fs_info, data);
5403                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5404                        "wanted %llu\n", (unsigned long long)data,
5405                        (unsigned long long)num_bytes);
5406                 dump_space_info(sinfo, num_bytes, 1);
5407         }
5408
5409         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5410
5411         return ret;
5412 }
5413
5414 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5415 {
5416         struct btrfs_block_group_cache *cache;
5417         int ret = 0;
5418
5419         cache = btrfs_lookup_block_group(root->fs_info, start);
5420         if (!cache) {
5421                 printk(KERN_ERR "Unable to find block group for %llu\n",
5422                        (unsigned long long)start);
5423                 return -ENOSPC;
5424         }
5425
5426         if (btrfs_test_opt(root, DISCARD))
5427                 ret = btrfs_discard_extent(root, start, len, NULL);
5428
5429         btrfs_add_free_space(cache, start, len);
5430         btrfs_update_reserved_bytes(cache, len, 0, 1);
5431         btrfs_put_block_group(cache);
5432
5433         trace_btrfs_reserved_extent_free(root, start, len);
5434
5435         return ret;
5436 }
5437
5438 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5439                                       struct btrfs_root *root,
5440                                       u64 parent, u64 root_objectid,
5441                                       u64 flags, u64 owner, u64 offset,
5442                                       struct btrfs_key *ins, int ref_mod)
5443 {
5444         int ret;
5445         struct btrfs_fs_info *fs_info = root->fs_info;
5446         struct btrfs_extent_item *extent_item;
5447         struct btrfs_extent_inline_ref *iref;
5448         struct btrfs_path *path;
5449         struct extent_buffer *leaf;
5450         int type;
5451         u32 size;
5452
5453         if (parent > 0)
5454                 type = BTRFS_SHARED_DATA_REF_KEY;
5455         else
5456                 type = BTRFS_EXTENT_DATA_REF_KEY;
5457
5458         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5459
5460         path = btrfs_alloc_path();
5461         if (!path)
5462                 return -ENOMEM;
5463
5464         path->leave_spinning = 1;
5465         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5466                                       ins, size);
5467         BUG_ON(ret);
5468
5469         leaf = path->nodes[0];
5470         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5471                                      struct btrfs_extent_item);
5472         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5473         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5474         btrfs_set_extent_flags(leaf, extent_item,
5475                                flags | BTRFS_EXTENT_FLAG_DATA);
5476
5477         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5478         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5479         if (parent > 0) {
5480                 struct btrfs_shared_data_ref *ref;
5481                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5482                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5483                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5484         } else {
5485                 struct btrfs_extent_data_ref *ref;
5486                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5487                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5488                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5489                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5490                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5491         }
5492
5493         btrfs_mark_buffer_dirty(path->nodes[0]);
5494         btrfs_free_path(path);
5495
5496         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5497         if (ret) {
5498                 printk(KERN_ERR "btrfs update block group failed for %llu "
5499                        "%llu\n", (unsigned long long)ins->objectid,
5500                        (unsigned long long)ins->offset);
5501                 BUG();
5502         }
5503         return ret;
5504 }
5505
5506 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5507                                      struct btrfs_root *root,
5508                                      u64 parent, u64 root_objectid,
5509                                      u64 flags, struct btrfs_disk_key *key,
5510                                      int level, struct btrfs_key *ins)
5511 {
5512         int ret;
5513         struct btrfs_fs_info *fs_info = root->fs_info;
5514         struct btrfs_extent_item *extent_item;
5515         struct btrfs_tree_block_info *block_info;
5516         struct btrfs_extent_inline_ref *iref;
5517         struct btrfs_path *path;
5518         struct extent_buffer *leaf;
5519         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5520
5521         path = btrfs_alloc_path();
5522         if (!path)
5523                 return -ENOMEM;
5524
5525         path->leave_spinning = 1;
5526         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5527                                       ins, size);
5528         BUG_ON(ret);
5529
5530         leaf = path->nodes[0];
5531         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5532                                      struct btrfs_extent_item);
5533         btrfs_set_extent_refs(leaf, extent_item, 1);
5534         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5535         btrfs_set_extent_flags(leaf, extent_item,
5536                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5537         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5538
5539         btrfs_set_tree_block_key(leaf, block_info, key);
5540         btrfs_set_tree_block_level(leaf, block_info, level);
5541
5542         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5543         if (parent > 0) {
5544                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5545                 btrfs_set_extent_inline_ref_type(leaf, iref,
5546                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5547                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5548         } else {
5549                 btrfs_set_extent_inline_ref_type(leaf, iref,
5550                                                  BTRFS_TREE_BLOCK_REF_KEY);
5551                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5552         }
5553
5554         btrfs_mark_buffer_dirty(leaf);
5555         btrfs_free_path(path);
5556
5557         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5558         if (ret) {
5559                 printk(KERN_ERR "btrfs update block group failed for %llu "
5560                        "%llu\n", (unsigned long long)ins->objectid,
5561                        (unsigned long long)ins->offset);
5562                 BUG();
5563         }
5564         return ret;
5565 }
5566
5567 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5568                                      struct btrfs_root *root,
5569                                      u64 root_objectid, u64 owner,
5570                                      u64 offset, struct btrfs_key *ins)
5571 {
5572         int ret;
5573
5574         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5575
5576         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5577                                          0, root_objectid, owner, offset,
5578                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5579         return ret;
5580 }
5581
5582 /*
5583  * this is used by the tree logging recovery code.  It records that
5584  * an extent has been allocated and makes sure to clear the free
5585  * space cache bits as well
5586  */
5587 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,