]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - fs/btrfs/extent-tree.c
Btrfs: let allocation start from the right raid type
[~shefty/rdma-dev.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include "compat.h"
28 #include "hash.h"
29 #include "ctree.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "transaction.h"
33 #include "volumes.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36 #include "math.h"
37
38 #undef SCRAMBLE_DELAYED_REFS
39
40 /*
41  * control flags for do_chunk_alloc's force field
42  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
43  * if we really need one.
44  *
45  * CHUNK_ALLOC_LIMITED means to only try and allocate one
46  * if we have very few chunks already allocated.  This is
47  * used as part of the clustering code to help make sure
48  * we have a good pool of storage to cluster in, without
49  * filling the FS with empty chunks
50  *
51  * CHUNK_ALLOC_FORCE means it must try to allocate one
52  *
53  */
54 enum {
55         CHUNK_ALLOC_NO_FORCE = 0,
56         CHUNK_ALLOC_LIMITED = 1,
57         CHUNK_ALLOC_FORCE = 2,
58 };
59
60 /*
61  * Control how reservations are dealt with.
62  *
63  * RESERVE_FREE - freeing a reservation.
64  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
65  *   ENOSPC accounting
66  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
67  *   bytes_may_use as the ENOSPC accounting is done elsewhere
68  */
69 enum {
70         RESERVE_FREE = 0,
71         RESERVE_ALLOC = 1,
72         RESERVE_ALLOC_NO_ACCOUNT = 2,
73 };
74
75 static int update_block_group(struct btrfs_trans_handle *trans,
76                               struct btrfs_root *root,
77                               u64 bytenr, u64 num_bytes, int alloc);
78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
79                                 struct btrfs_root *root,
80                                 u64 bytenr, u64 num_bytes, u64 parent,
81                                 u64 root_objectid, u64 owner_objectid,
82                                 u64 owner_offset, int refs_to_drop,
83                                 struct btrfs_delayed_extent_op *extra_op);
84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
85                                     struct extent_buffer *leaf,
86                                     struct btrfs_extent_item *ei);
87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
88                                       struct btrfs_root *root,
89                                       u64 parent, u64 root_objectid,
90                                       u64 flags, u64 owner, u64 offset,
91                                       struct btrfs_key *ins, int ref_mod);
92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
93                                      struct btrfs_root *root,
94                                      u64 parent, u64 root_objectid,
95                                      u64 flags, struct btrfs_disk_key *key,
96                                      int level, struct btrfs_key *ins);
97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
98                           struct btrfs_root *extent_root, u64 flags,
99                           int force);
100 static int find_next_key(struct btrfs_path *path, int level,
101                          struct btrfs_key *key);
102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
103                             int dump_block_groups);
104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
105                                        u64 num_bytes, int reserve);
106
107 static noinline int
108 block_group_cache_done(struct btrfs_block_group_cache *cache)
109 {
110         smp_mb();
111         return cache->cached == BTRFS_CACHE_FINISHED;
112 }
113
114 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
115 {
116         return (cache->flags & bits) == bits;
117 }
118
119 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
120 {
121         atomic_inc(&cache->count);
122 }
123
124 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
125 {
126         if (atomic_dec_and_test(&cache->count)) {
127                 WARN_ON(cache->pinned > 0);
128                 WARN_ON(cache->reserved > 0);
129                 kfree(cache->free_space_ctl);
130                 kfree(cache);
131         }
132 }
133
134 /*
135  * this adds the block group to the fs_info rb tree for the block group
136  * cache
137  */
138 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
139                                 struct btrfs_block_group_cache *block_group)
140 {
141         struct rb_node **p;
142         struct rb_node *parent = NULL;
143         struct btrfs_block_group_cache *cache;
144
145         spin_lock(&info->block_group_cache_lock);
146         p = &info->block_group_cache_tree.rb_node;
147
148         while (*p) {
149                 parent = *p;
150                 cache = rb_entry(parent, struct btrfs_block_group_cache,
151                                  cache_node);
152                 if (block_group->key.objectid < cache->key.objectid) {
153                         p = &(*p)->rb_left;
154                 } else if (block_group->key.objectid > cache->key.objectid) {
155                         p = &(*p)->rb_right;
156                 } else {
157                         spin_unlock(&info->block_group_cache_lock);
158                         return -EEXIST;
159                 }
160         }
161
162         rb_link_node(&block_group->cache_node, parent, p);
163         rb_insert_color(&block_group->cache_node,
164                         &info->block_group_cache_tree);
165         spin_unlock(&info->block_group_cache_lock);
166
167         return 0;
168 }
169
170 /*
171  * This will return the block group at or after bytenr if contains is 0, else
172  * it will return the block group that contains the bytenr
173  */
174 static struct btrfs_block_group_cache *
175 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
176                               int contains)
177 {
178         struct btrfs_block_group_cache *cache, *ret = NULL;
179         struct rb_node *n;
180         u64 end, start;
181
182         spin_lock(&info->block_group_cache_lock);
183         n = info->block_group_cache_tree.rb_node;
184
185         while (n) {
186                 cache = rb_entry(n, struct btrfs_block_group_cache,
187                                  cache_node);
188                 end = cache->key.objectid + cache->key.offset - 1;
189                 start = cache->key.objectid;
190
191                 if (bytenr < start) {
192                         if (!contains && (!ret || start < ret->key.objectid))
193                                 ret = cache;
194                         n = n->rb_left;
195                 } else if (bytenr > start) {
196                         if (contains && bytenr <= end) {
197                                 ret = cache;
198                                 break;
199                         }
200                         n = n->rb_right;
201                 } else {
202                         ret = cache;
203                         break;
204                 }
205         }
206         if (ret)
207                 btrfs_get_block_group(ret);
208         spin_unlock(&info->block_group_cache_lock);
209
210         return ret;
211 }
212
213 static int add_excluded_extent(struct btrfs_root *root,
214                                u64 start, u64 num_bytes)
215 {
216         u64 end = start + num_bytes - 1;
217         set_extent_bits(&root->fs_info->freed_extents[0],
218                         start, end, EXTENT_UPTODATE, GFP_NOFS);
219         set_extent_bits(&root->fs_info->freed_extents[1],
220                         start, end, EXTENT_UPTODATE, GFP_NOFS);
221         return 0;
222 }
223
224 static void free_excluded_extents(struct btrfs_root *root,
225                                   struct btrfs_block_group_cache *cache)
226 {
227         u64 start, end;
228
229         start = cache->key.objectid;
230         end = start + cache->key.offset - 1;
231
232         clear_extent_bits(&root->fs_info->freed_extents[0],
233                           start, end, EXTENT_UPTODATE, GFP_NOFS);
234         clear_extent_bits(&root->fs_info->freed_extents[1],
235                           start, end, EXTENT_UPTODATE, GFP_NOFS);
236 }
237
238 static int exclude_super_stripes(struct btrfs_root *root,
239                                  struct btrfs_block_group_cache *cache)
240 {
241         u64 bytenr;
242         u64 *logical;
243         int stripe_len;
244         int i, nr, ret;
245
246         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
247                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
248                 cache->bytes_super += stripe_len;
249                 ret = add_excluded_extent(root, cache->key.objectid,
250                                           stripe_len);
251                 BUG_ON(ret); /* -ENOMEM */
252         }
253
254         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
255                 bytenr = btrfs_sb_offset(i);
256                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
257                                        cache->key.objectid, bytenr,
258                                        0, &logical, &nr, &stripe_len);
259                 BUG_ON(ret); /* -ENOMEM */
260
261                 while (nr--) {
262                         cache->bytes_super += stripe_len;
263                         ret = add_excluded_extent(root, logical[nr],
264                                                   stripe_len);
265                         BUG_ON(ret); /* -ENOMEM */
266                 }
267
268                 kfree(logical);
269         }
270         return 0;
271 }
272
273 static struct btrfs_caching_control *
274 get_caching_control(struct btrfs_block_group_cache *cache)
275 {
276         struct btrfs_caching_control *ctl;
277
278         spin_lock(&cache->lock);
279         if (cache->cached != BTRFS_CACHE_STARTED) {
280                 spin_unlock(&cache->lock);
281                 return NULL;
282         }
283
284         /* We're loading it the fast way, so we don't have a caching_ctl. */
285         if (!cache->caching_ctl) {
286                 spin_unlock(&cache->lock);
287                 return NULL;
288         }
289
290         ctl = cache->caching_ctl;
291         atomic_inc(&ctl->count);
292         spin_unlock(&cache->lock);
293         return ctl;
294 }
295
296 static void put_caching_control(struct btrfs_caching_control *ctl)
297 {
298         if (atomic_dec_and_test(&ctl->count))
299                 kfree(ctl);
300 }
301
302 /*
303  * this is only called by cache_block_group, since we could have freed extents
304  * we need to check the pinned_extents for any extents that can't be used yet
305  * since their free space will be released as soon as the transaction commits.
306  */
307 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
308                               struct btrfs_fs_info *info, u64 start, u64 end)
309 {
310         u64 extent_start, extent_end, size, total_added = 0;
311         int ret;
312
313         while (start < end) {
314                 ret = find_first_extent_bit(info->pinned_extents, start,
315                                             &extent_start, &extent_end,
316                                             EXTENT_DIRTY | EXTENT_UPTODATE,
317                                             NULL);
318                 if (ret)
319                         break;
320
321                 if (extent_start <= start) {
322                         start = extent_end + 1;
323                 } else if (extent_start > start && extent_start < end) {
324                         size = extent_start - start;
325                         total_added += size;
326                         ret = btrfs_add_free_space(block_group, start,
327                                                    size);
328                         BUG_ON(ret); /* -ENOMEM or logic error */
329                         start = extent_end + 1;
330                 } else {
331                         break;
332                 }
333         }
334
335         if (start < end) {
336                 size = end - start;
337                 total_added += size;
338                 ret = btrfs_add_free_space(block_group, start, size);
339                 BUG_ON(ret); /* -ENOMEM or logic error */
340         }
341
342         return total_added;
343 }
344
345 static noinline void caching_thread(struct btrfs_work *work)
346 {
347         struct btrfs_block_group_cache *block_group;
348         struct btrfs_fs_info *fs_info;
349         struct btrfs_caching_control *caching_ctl;
350         struct btrfs_root *extent_root;
351         struct btrfs_path *path;
352         struct extent_buffer *leaf;
353         struct btrfs_key key;
354         u64 total_found = 0;
355         u64 last = 0;
356         u32 nritems;
357         int ret = 0;
358
359         caching_ctl = container_of(work, struct btrfs_caching_control, work);
360         block_group = caching_ctl->block_group;
361         fs_info = block_group->fs_info;
362         extent_root = fs_info->extent_root;
363
364         path = btrfs_alloc_path();
365         if (!path)
366                 goto out;
367
368         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
369
370         /*
371          * We don't want to deadlock with somebody trying to allocate a new
372          * extent for the extent root while also trying to search the extent
373          * root to add free space.  So we skip locking and search the commit
374          * root, since its read-only
375          */
376         path->skip_locking = 1;
377         path->search_commit_root = 1;
378         path->reada = 1;
379
380         key.objectid = last;
381         key.offset = 0;
382         key.type = BTRFS_EXTENT_ITEM_KEY;
383 again:
384         mutex_lock(&caching_ctl->mutex);
385         /* need to make sure the commit_root doesn't disappear */
386         down_read(&fs_info->extent_commit_sem);
387
388         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
389         if (ret < 0)
390                 goto err;
391
392         leaf = path->nodes[0];
393         nritems = btrfs_header_nritems(leaf);
394
395         while (1) {
396                 if (btrfs_fs_closing(fs_info) > 1) {
397                         last = (u64)-1;
398                         break;
399                 }
400
401                 if (path->slots[0] < nritems) {
402                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
403                 } else {
404                         ret = find_next_key(path, 0, &key);
405                         if (ret)
406                                 break;
407
408                         if (need_resched() ||
409                             btrfs_next_leaf(extent_root, path)) {
410                                 caching_ctl->progress = last;
411                                 btrfs_release_path(path);
412                                 up_read(&fs_info->extent_commit_sem);
413                                 mutex_unlock(&caching_ctl->mutex);
414                                 cond_resched();
415                                 goto again;
416                         }
417                         leaf = path->nodes[0];
418                         nritems = btrfs_header_nritems(leaf);
419                         continue;
420                 }
421
422                 if (key.objectid < block_group->key.objectid) {
423                         path->slots[0]++;
424                         continue;
425                 }
426
427                 if (key.objectid >= block_group->key.objectid +
428                     block_group->key.offset)
429                         break;
430
431                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
432                         total_found += add_new_free_space(block_group,
433                                                           fs_info, last,
434                                                           key.objectid);
435                         last = key.objectid + key.offset;
436
437                         if (total_found > (1024 * 1024 * 2)) {
438                                 total_found = 0;
439                                 wake_up(&caching_ctl->wait);
440                         }
441                 }
442                 path->slots[0]++;
443         }
444         ret = 0;
445
446         total_found += add_new_free_space(block_group, fs_info, last,
447                                           block_group->key.objectid +
448                                           block_group->key.offset);
449         caching_ctl->progress = (u64)-1;
450
451         spin_lock(&block_group->lock);
452         block_group->caching_ctl = NULL;
453         block_group->cached = BTRFS_CACHE_FINISHED;
454         spin_unlock(&block_group->lock);
455
456 err:
457         btrfs_free_path(path);
458         up_read(&fs_info->extent_commit_sem);
459
460         free_excluded_extents(extent_root, block_group);
461
462         mutex_unlock(&caching_ctl->mutex);
463 out:
464         wake_up(&caching_ctl->wait);
465
466         put_caching_control(caching_ctl);
467         btrfs_put_block_group(block_group);
468 }
469
470 static int cache_block_group(struct btrfs_block_group_cache *cache,
471                              struct btrfs_trans_handle *trans,
472                              struct btrfs_root *root,
473                              int load_cache_only)
474 {
475         DEFINE_WAIT(wait);
476         struct btrfs_fs_info *fs_info = cache->fs_info;
477         struct btrfs_caching_control *caching_ctl;
478         int ret = 0;
479
480         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
481         if (!caching_ctl)
482                 return -ENOMEM;
483
484         INIT_LIST_HEAD(&caching_ctl->list);
485         mutex_init(&caching_ctl->mutex);
486         init_waitqueue_head(&caching_ctl->wait);
487         caching_ctl->block_group = cache;
488         caching_ctl->progress = cache->key.objectid;
489         atomic_set(&caching_ctl->count, 1);
490         caching_ctl->work.func = caching_thread;
491
492         spin_lock(&cache->lock);
493         /*
494          * This should be a rare occasion, but this could happen I think in the
495          * case where one thread starts to load the space cache info, and then
496          * some other thread starts a transaction commit which tries to do an
497          * allocation while the other thread is still loading the space cache
498          * info.  The previous loop should have kept us from choosing this block
499          * group, but if we've moved to the state where we will wait on caching
500          * block groups we need to first check if we're doing a fast load here,
501          * so we can wait for it to finish, otherwise we could end up allocating
502          * from a block group who's cache gets evicted for one reason or
503          * another.
504          */
505         while (cache->cached == BTRFS_CACHE_FAST) {
506                 struct btrfs_caching_control *ctl;
507
508                 ctl = cache->caching_ctl;
509                 atomic_inc(&ctl->count);
510                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
511                 spin_unlock(&cache->lock);
512
513                 schedule();
514
515                 finish_wait(&ctl->wait, &wait);
516                 put_caching_control(ctl);
517                 spin_lock(&cache->lock);
518         }
519
520         if (cache->cached != BTRFS_CACHE_NO) {
521                 spin_unlock(&cache->lock);
522                 kfree(caching_ctl);
523                 return 0;
524         }
525         WARN_ON(cache->caching_ctl);
526         cache->caching_ctl = caching_ctl;
527         cache->cached = BTRFS_CACHE_FAST;
528         spin_unlock(&cache->lock);
529
530         /*
531          * We can't do the read from on-disk cache during a commit since we need
532          * to have the normal tree locking.  Also if we are currently trying to
533          * allocate blocks for the tree root we can't do the fast caching since
534          * we likely hold important locks.
535          */
536         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
537                 ret = load_free_space_cache(fs_info, cache);
538
539                 spin_lock(&cache->lock);
540                 if (ret == 1) {
541                         cache->caching_ctl = NULL;
542                         cache->cached = BTRFS_CACHE_FINISHED;
543                         cache->last_byte_to_unpin = (u64)-1;
544                 } else {
545                         if (load_cache_only) {
546                                 cache->caching_ctl = NULL;
547                                 cache->cached = BTRFS_CACHE_NO;
548                         } else {
549                                 cache->cached = BTRFS_CACHE_STARTED;
550                         }
551                 }
552                 spin_unlock(&cache->lock);
553                 wake_up(&caching_ctl->wait);
554                 if (ret == 1) {
555                         put_caching_control(caching_ctl);
556                         free_excluded_extents(fs_info->extent_root, cache);
557                         return 0;
558                 }
559         } else {
560                 /*
561                  * We are not going to do the fast caching, set cached to the
562                  * appropriate value and wakeup any waiters.
563                  */
564                 spin_lock(&cache->lock);
565                 if (load_cache_only) {
566                         cache->caching_ctl = NULL;
567                         cache->cached = BTRFS_CACHE_NO;
568                 } else {
569                         cache->cached = BTRFS_CACHE_STARTED;
570                 }
571                 spin_unlock(&cache->lock);
572                 wake_up(&caching_ctl->wait);
573         }
574
575         if (load_cache_only) {
576                 put_caching_control(caching_ctl);
577                 return 0;
578         }
579
580         down_write(&fs_info->extent_commit_sem);
581         atomic_inc(&caching_ctl->count);
582         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
583         up_write(&fs_info->extent_commit_sem);
584
585         btrfs_get_block_group(cache);
586
587         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
588
589         return ret;
590 }
591
592 /*
593  * return the block group that starts at or after bytenr
594  */
595 static struct btrfs_block_group_cache *
596 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
597 {
598         struct btrfs_block_group_cache *cache;
599
600         cache = block_group_cache_tree_search(info, bytenr, 0);
601
602         return cache;
603 }
604
605 /*
606  * return the block group that contains the given bytenr
607  */
608 struct btrfs_block_group_cache *btrfs_lookup_block_group(
609                                                  struct btrfs_fs_info *info,
610                                                  u64 bytenr)
611 {
612         struct btrfs_block_group_cache *cache;
613
614         cache = block_group_cache_tree_search(info, bytenr, 1);
615
616         return cache;
617 }
618
619 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
620                                                   u64 flags)
621 {
622         struct list_head *head = &info->space_info;
623         struct btrfs_space_info *found;
624
625         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
626
627         rcu_read_lock();
628         list_for_each_entry_rcu(found, head, list) {
629                 if (found->flags & flags) {
630                         rcu_read_unlock();
631                         return found;
632                 }
633         }
634         rcu_read_unlock();
635         return NULL;
636 }
637
638 /*
639  * after adding space to the filesystem, we need to clear the full flags
640  * on all the space infos.
641  */
642 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
643 {
644         struct list_head *head = &info->space_info;
645         struct btrfs_space_info *found;
646
647         rcu_read_lock();
648         list_for_each_entry_rcu(found, head, list)
649                 found->full = 0;
650         rcu_read_unlock();
651 }
652
653 u64 btrfs_find_block_group(struct btrfs_root *root,
654                            u64 search_start, u64 search_hint, int owner)
655 {
656         struct btrfs_block_group_cache *cache;
657         u64 used;
658         u64 last = max(search_hint, search_start);
659         u64 group_start = 0;
660         int full_search = 0;
661         int factor = 9;
662         int wrapped = 0;
663 again:
664         while (1) {
665                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
666                 if (!cache)
667                         break;
668
669                 spin_lock(&cache->lock);
670                 last = cache->key.objectid + cache->key.offset;
671                 used = btrfs_block_group_used(&cache->item);
672
673                 if ((full_search || !cache->ro) &&
674                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
675                         if (used + cache->pinned + cache->reserved <
676                             div_factor(cache->key.offset, factor)) {
677                                 group_start = cache->key.objectid;
678                                 spin_unlock(&cache->lock);
679                                 btrfs_put_block_group(cache);
680                                 goto found;
681                         }
682                 }
683                 spin_unlock(&cache->lock);
684                 btrfs_put_block_group(cache);
685                 cond_resched();
686         }
687         if (!wrapped) {
688                 last = search_start;
689                 wrapped = 1;
690                 goto again;
691         }
692         if (!full_search && factor < 10) {
693                 last = search_start;
694                 full_search = 1;
695                 factor = 10;
696                 goto again;
697         }
698 found:
699         return group_start;
700 }
701
702 /* simple helper to search for an existing extent at a given offset */
703 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
704 {
705         int ret;
706         struct btrfs_key key;
707         struct btrfs_path *path;
708
709         path = btrfs_alloc_path();
710         if (!path)
711                 return -ENOMEM;
712
713         key.objectid = start;
714         key.offset = len;
715         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
716         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
717                                 0, 0);
718         btrfs_free_path(path);
719         return ret;
720 }
721
722 /*
723  * helper function to lookup reference count and flags of extent.
724  *
725  * the head node for delayed ref is used to store the sum of all the
726  * reference count modifications queued up in the rbtree. the head
727  * node may also store the extent flags to set. This way you can check
728  * to see what the reference count and extent flags would be if all of
729  * the delayed refs are not processed.
730  */
731 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
732                              struct btrfs_root *root, u64 bytenr,
733                              u64 num_bytes, u64 *refs, u64 *flags)
734 {
735         struct btrfs_delayed_ref_head *head;
736         struct btrfs_delayed_ref_root *delayed_refs;
737         struct btrfs_path *path;
738         struct btrfs_extent_item *ei;
739         struct extent_buffer *leaf;
740         struct btrfs_key key;
741         u32 item_size;
742         u64 num_refs;
743         u64 extent_flags;
744         int ret;
745
746         path = btrfs_alloc_path();
747         if (!path)
748                 return -ENOMEM;
749
750         key.objectid = bytenr;
751         key.type = BTRFS_EXTENT_ITEM_KEY;
752         key.offset = num_bytes;
753         if (!trans) {
754                 path->skip_locking = 1;
755                 path->search_commit_root = 1;
756         }
757 again:
758         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
759                                 &key, path, 0, 0);
760         if (ret < 0)
761                 goto out_free;
762
763         if (ret == 0) {
764                 leaf = path->nodes[0];
765                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
766                 if (item_size >= sizeof(*ei)) {
767                         ei = btrfs_item_ptr(leaf, path->slots[0],
768                                             struct btrfs_extent_item);
769                         num_refs = btrfs_extent_refs(leaf, ei);
770                         extent_flags = btrfs_extent_flags(leaf, ei);
771                 } else {
772 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
773                         struct btrfs_extent_item_v0 *ei0;
774                         BUG_ON(item_size != sizeof(*ei0));
775                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
776                                              struct btrfs_extent_item_v0);
777                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
778                         /* FIXME: this isn't correct for data */
779                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
780 #else
781                         BUG();
782 #endif
783                 }
784                 BUG_ON(num_refs == 0);
785         } else {
786                 num_refs = 0;
787                 extent_flags = 0;
788                 ret = 0;
789         }
790
791         if (!trans)
792                 goto out;
793
794         delayed_refs = &trans->transaction->delayed_refs;
795         spin_lock(&delayed_refs->lock);
796         head = btrfs_find_delayed_ref_head(trans, bytenr);
797         if (head) {
798                 if (!mutex_trylock(&head->mutex)) {
799                         atomic_inc(&head->node.refs);
800                         spin_unlock(&delayed_refs->lock);
801
802                         btrfs_release_path(path);
803
804                         /*
805                          * Mutex was contended, block until it's released and try
806                          * again
807                          */
808                         mutex_lock(&head->mutex);
809                         mutex_unlock(&head->mutex);
810                         btrfs_put_delayed_ref(&head->node);
811                         goto again;
812                 }
813                 if (head->extent_op && head->extent_op->update_flags)
814                         extent_flags |= head->extent_op->flags_to_set;
815                 else
816                         BUG_ON(num_refs == 0);
817
818                 num_refs += head->node.ref_mod;
819                 mutex_unlock(&head->mutex);
820         }
821         spin_unlock(&delayed_refs->lock);
822 out:
823         WARN_ON(num_refs == 0);
824         if (refs)
825                 *refs = num_refs;
826         if (flags)
827                 *flags = extent_flags;
828 out_free:
829         btrfs_free_path(path);
830         return ret;
831 }
832
833 /*
834  * Back reference rules.  Back refs have three main goals:
835  *
836  * 1) differentiate between all holders of references to an extent so that
837  *    when a reference is dropped we can make sure it was a valid reference
838  *    before freeing the extent.
839  *
840  * 2) Provide enough information to quickly find the holders of an extent
841  *    if we notice a given block is corrupted or bad.
842  *
843  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
844  *    maintenance.  This is actually the same as #2, but with a slightly
845  *    different use case.
846  *
847  * There are two kinds of back refs. The implicit back refs is optimized
848  * for pointers in non-shared tree blocks. For a given pointer in a block,
849  * back refs of this kind provide information about the block's owner tree
850  * and the pointer's key. These information allow us to find the block by
851  * b-tree searching. The full back refs is for pointers in tree blocks not
852  * referenced by their owner trees. The location of tree block is recorded
853  * in the back refs. Actually the full back refs is generic, and can be
854  * used in all cases the implicit back refs is used. The major shortcoming
855  * of the full back refs is its overhead. Every time a tree block gets
856  * COWed, we have to update back refs entry for all pointers in it.
857  *
858  * For a newly allocated tree block, we use implicit back refs for
859  * pointers in it. This means most tree related operations only involve
860  * implicit back refs. For a tree block created in old transaction, the
861  * only way to drop a reference to it is COW it. So we can detect the
862  * event that tree block loses its owner tree's reference and do the
863  * back refs conversion.
864  *
865  * When a tree block is COW'd through a tree, there are four cases:
866  *
867  * The reference count of the block is one and the tree is the block's
868  * owner tree. Nothing to do in this case.
869  *
870  * The reference count of the block is one and the tree is not the
871  * block's owner tree. In this case, full back refs is used for pointers
872  * in the block. Remove these full back refs, add implicit back refs for
873  * every pointers in the new block.
874  *
875  * The reference count of the block is greater than one and the tree is
876  * the block's owner tree. In this case, implicit back refs is used for
877  * pointers in the block. Add full back refs for every pointers in the
878  * block, increase lower level extents' reference counts. The original
879  * implicit back refs are entailed to the new block.
880  *
881  * The reference count of the block is greater than one and the tree is
882  * not the block's owner tree. Add implicit back refs for every pointer in
883  * the new block, increase lower level extents' reference count.
884  *
885  * Back Reference Key composing:
886  *
887  * The key objectid corresponds to the first byte in the extent,
888  * The key type is used to differentiate between types of back refs.
889  * There are different meanings of the key offset for different types
890  * of back refs.
891  *
892  * File extents can be referenced by:
893  *
894  * - multiple snapshots, subvolumes, or different generations in one subvol
895  * - different files inside a single subvolume
896  * - different offsets inside a file (bookend extents in file.c)
897  *
898  * The extent ref structure for the implicit back refs has fields for:
899  *
900  * - Objectid of the subvolume root
901  * - objectid of the file holding the reference
902  * - original offset in the file
903  * - how many bookend extents
904  *
905  * The key offset for the implicit back refs is hash of the first
906  * three fields.
907  *
908  * The extent ref structure for the full back refs has field for:
909  *
910  * - number of pointers in the tree leaf
911  *
912  * The key offset for the implicit back refs is the first byte of
913  * the tree leaf
914  *
915  * When a file extent is allocated, The implicit back refs is used.
916  * the fields are filled in:
917  *
918  *     (root_key.objectid, inode objectid, offset in file, 1)
919  *
920  * When a file extent is removed file truncation, we find the
921  * corresponding implicit back refs and check the following fields:
922  *
923  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
924  *
925  * Btree extents can be referenced by:
926  *
927  * - Different subvolumes
928  *
929  * Both the implicit back refs and the full back refs for tree blocks
930  * only consist of key. The key offset for the implicit back refs is
931  * objectid of block's owner tree. The key offset for the full back refs
932  * is the first byte of parent block.
933  *
934  * When implicit back refs is used, information about the lowest key and
935  * level of the tree block are required. These information are stored in
936  * tree block info structure.
937  */
938
939 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
940 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
941                                   struct btrfs_root *root,
942                                   struct btrfs_path *path,
943                                   u64 owner, u32 extra_size)
944 {
945         struct btrfs_extent_item *item;
946         struct btrfs_extent_item_v0 *ei0;
947         struct btrfs_extent_ref_v0 *ref0;
948         struct btrfs_tree_block_info *bi;
949         struct extent_buffer *leaf;
950         struct btrfs_key key;
951         struct btrfs_key found_key;
952         u32 new_size = sizeof(*item);
953         u64 refs;
954         int ret;
955
956         leaf = path->nodes[0];
957         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
958
959         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
960         ei0 = btrfs_item_ptr(leaf, path->slots[0],
961                              struct btrfs_extent_item_v0);
962         refs = btrfs_extent_refs_v0(leaf, ei0);
963
964         if (owner == (u64)-1) {
965                 while (1) {
966                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
967                                 ret = btrfs_next_leaf(root, path);
968                                 if (ret < 0)
969                                         return ret;
970                                 BUG_ON(ret > 0); /* Corruption */
971                                 leaf = path->nodes[0];
972                         }
973                         btrfs_item_key_to_cpu(leaf, &found_key,
974                                               path->slots[0]);
975                         BUG_ON(key.objectid != found_key.objectid);
976                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
977                                 path->slots[0]++;
978                                 continue;
979                         }
980                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
981                                               struct btrfs_extent_ref_v0);
982                         owner = btrfs_ref_objectid_v0(leaf, ref0);
983                         break;
984                 }
985         }
986         btrfs_release_path(path);
987
988         if (owner < BTRFS_FIRST_FREE_OBJECTID)
989                 new_size += sizeof(*bi);
990
991         new_size -= sizeof(*ei0);
992         ret = btrfs_search_slot(trans, root, &key, path,
993                                 new_size + extra_size, 1);
994         if (ret < 0)
995                 return ret;
996         BUG_ON(ret); /* Corruption */
997
998         btrfs_extend_item(trans, root, path, new_size);
999
1000         leaf = path->nodes[0];
1001         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1002         btrfs_set_extent_refs(leaf, item, refs);
1003         /* FIXME: get real generation */
1004         btrfs_set_extent_generation(leaf, item, 0);
1005         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1006                 btrfs_set_extent_flags(leaf, item,
1007                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1008                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1009                 bi = (struct btrfs_tree_block_info *)(item + 1);
1010                 /* FIXME: get first key of the block */
1011                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1012                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1013         } else {
1014                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1015         }
1016         btrfs_mark_buffer_dirty(leaf);
1017         return 0;
1018 }
1019 #endif
1020
1021 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1022 {
1023         u32 high_crc = ~(u32)0;
1024         u32 low_crc = ~(u32)0;
1025         __le64 lenum;
1026
1027         lenum = cpu_to_le64(root_objectid);
1028         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1029         lenum = cpu_to_le64(owner);
1030         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1031         lenum = cpu_to_le64(offset);
1032         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1033
1034         return ((u64)high_crc << 31) ^ (u64)low_crc;
1035 }
1036
1037 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1038                                      struct btrfs_extent_data_ref *ref)
1039 {
1040         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1041                                     btrfs_extent_data_ref_objectid(leaf, ref),
1042                                     btrfs_extent_data_ref_offset(leaf, ref));
1043 }
1044
1045 static int match_extent_data_ref(struct extent_buffer *leaf,
1046                                  struct btrfs_extent_data_ref *ref,
1047                                  u64 root_objectid, u64 owner, u64 offset)
1048 {
1049         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1050             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1051             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1052                 return 0;
1053         return 1;
1054 }
1055
1056 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1057                                            struct btrfs_root *root,
1058                                            struct btrfs_path *path,
1059                                            u64 bytenr, u64 parent,
1060                                            u64 root_objectid,
1061                                            u64 owner, u64 offset)
1062 {
1063         struct btrfs_key key;
1064         struct btrfs_extent_data_ref *ref;
1065         struct extent_buffer *leaf;
1066         u32 nritems;
1067         int ret;
1068         int recow;
1069         int err = -ENOENT;
1070
1071         key.objectid = bytenr;
1072         if (parent) {
1073                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1074                 key.offset = parent;
1075         } else {
1076                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1077                 key.offset = hash_extent_data_ref(root_objectid,
1078                                                   owner, offset);
1079         }
1080 again:
1081         recow = 0;
1082         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1083         if (ret < 0) {
1084                 err = ret;
1085                 goto fail;
1086         }
1087
1088         if (parent) {
1089                 if (!ret)
1090                         return 0;
1091 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1092                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1093                 btrfs_release_path(path);
1094                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1095                 if (ret < 0) {
1096                         err = ret;
1097                         goto fail;
1098                 }
1099                 if (!ret)
1100                         return 0;
1101 #endif
1102                 goto fail;
1103         }
1104
1105         leaf = path->nodes[0];
1106         nritems = btrfs_header_nritems(leaf);
1107         while (1) {
1108                 if (path->slots[0] >= nritems) {
1109                         ret = btrfs_next_leaf(root, path);
1110                         if (ret < 0)
1111                                 err = ret;
1112                         if (ret)
1113                                 goto fail;
1114
1115                         leaf = path->nodes[0];
1116                         nritems = btrfs_header_nritems(leaf);
1117                         recow = 1;
1118                 }
1119
1120                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1121                 if (key.objectid != bytenr ||
1122                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1123                         goto fail;
1124
1125                 ref = btrfs_item_ptr(leaf, path->slots[0],
1126                                      struct btrfs_extent_data_ref);
1127
1128                 if (match_extent_data_ref(leaf, ref, root_objectid,
1129                                           owner, offset)) {
1130                         if (recow) {
1131                                 btrfs_release_path(path);
1132                                 goto again;
1133                         }
1134                         err = 0;
1135                         break;
1136                 }
1137                 path->slots[0]++;
1138         }
1139 fail:
1140         return err;
1141 }
1142
1143 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1144                                            struct btrfs_root *root,
1145                                            struct btrfs_path *path,
1146                                            u64 bytenr, u64 parent,
1147                                            u64 root_objectid, u64 owner,
1148                                            u64 offset, int refs_to_add)
1149 {
1150         struct btrfs_key key;
1151         struct extent_buffer *leaf;
1152         u32 size;
1153         u32 num_refs;
1154         int ret;
1155
1156         key.objectid = bytenr;
1157         if (parent) {
1158                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1159                 key.offset = parent;
1160                 size = sizeof(struct btrfs_shared_data_ref);
1161         } else {
1162                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1163                 key.offset = hash_extent_data_ref(root_objectid,
1164                                                   owner, offset);
1165                 size = sizeof(struct btrfs_extent_data_ref);
1166         }
1167
1168         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1169         if (ret && ret != -EEXIST)
1170                 goto fail;
1171
1172         leaf = path->nodes[0];
1173         if (parent) {
1174                 struct btrfs_shared_data_ref *ref;
1175                 ref = btrfs_item_ptr(leaf, path->slots[0],
1176                                      struct btrfs_shared_data_ref);
1177                 if (ret == 0) {
1178                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1179                 } else {
1180                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1181                         num_refs += refs_to_add;
1182                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1183                 }
1184         } else {
1185                 struct btrfs_extent_data_ref *ref;
1186                 while (ret == -EEXIST) {
1187                         ref = btrfs_item_ptr(leaf, path->slots[0],
1188                                              struct btrfs_extent_data_ref);
1189                         if (match_extent_data_ref(leaf, ref, root_objectid,
1190                                                   owner, offset))
1191                                 break;
1192                         btrfs_release_path(path);
1193                         key.offset++;
1194                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1195                                                       size);
1196                         if (ret && ret != -EEXIST)
1197                                 goto fail;
1198
1199                         leaf = path->nodes[0];
1200                 }
1201                 ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                      struct btrfs_extent_data_ref);
1203                 if (ret == 0) {
1204                         btrfs_set_extent_data_ref_root(leaf, ref,
1205                                                        root_objectid);
1206                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1207                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1208                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1209                 } else {
1210                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1211                         num_refs += refs_to_add;
1212                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1213                 }
1214         }
1215         btrfs_mark_buffer_dirty(leaf);
1216         ret = 0;
1217 fail:
1218         btrfs_release_path(path);
1219         return ret;
1220 }
1221
1222 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1223                                            struct btrfs_root *root,
1224                                            struct btrfs_path *path,
1225                                            int refs_to_drop)
1226 {
1227         struct btrfs_key key;
1228         struct btrfs_extent_data_ref *ref1 = NULL;
1229         struct btrfs_shared_data_ref *ref2 = NULL;
1230         struct extent_buffer *leaf;
1231         u32 num_refs = 0;
1232         int ret = 0;
1233
1234         leaf = path->nodes[0];
1235         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1236
1237         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1238                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1239                                       struct btrfs_extent_data_ref);
1240                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1241         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1242                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1243                                       struct btrfs_shared_data_ref);
1244                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1245 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1246         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1247                 struct btrfs_extent_ref_v0 *ref0;
1248                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1249                                       struct btrfs_extent_ref_v0);
1250                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1251 #endif
1252         } else {
1253                 BUG();
1254         }
1255
1256         BUG_ON(num_refs < refs_to_drop);
1257         num_refs -= refs_to_drop;
1258
1259         if (num_refs == 0) {
1260                 ret = btrfs_del_item(trans, root, path);
1261         } else {
1262                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1263                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1264                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1265                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1266 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1267                 else {
1268                         struct btrfs_extent_ref_v0 *ref0;
1269                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1270                                         struct btrfs_extent_ref_v0);
1271                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1272                 }
1273 #endif
1274                 btrfs_mark_buffer_dirty(leaf);
1275         }
1276         return ret;
1277 }
1278
1279 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1280                                           struct btrfs_path *path,
1281                                           struct btrfs_extent_inline_ref *iref)
1282 {
1283         struct btrfs_key key;
1284         struct extent_buffer *leaf;
1285         struct btrfs_extent_data_ref *ref1;
1286         struct btrfs_shared_data_ref *ref2;
1287         u32 num_refs = 0;
1288
1289         leaf = path->nodes[0];
1290         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1291         if (iref) {
1292                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1293                     BTRFS_EXTENT_DATA_REF_KEY) {
1294                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1295                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1296                 } else {
1297                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1298                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1299                 }
1300         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302                                       struct btrfs_extent_data_ref);
1303                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306                                       struct btrfs_shared_data_ref);
1307                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1309         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1310                 struct btrfs_extent_ref_v0 *ref0;
1311                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1312                                       struct btrfs_extent_ref_v0);
1313                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1314 #endif
1315         } else {
1316                 WARN_ON(1);
1317         }
1318         return num_refs;
1319 }
1320
1321 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1322                                           struct btrfs_root *root,
1323                                           struct btrfs_path *path,
1324                                           u64 bytenr, u64 parent,
1325                                           u64 root_objectid)
1326 {
1327         struct btrfs_key key;
1328         int ret;
1329
1330         key.objectid = bytenr;
1331         if (parent) {
1332                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1333                 key.offset = parent;
1334         } else {
1335                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1336                 key.offset = root_objectid;
1337         }
1338
1339         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1340         if (ret > 0)
1341                 ret = -ENOENT;
1342 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1343         if (ret == -ENOENT && parent) {
1344                 btrfs_release_path(path);
1345                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1346                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1347                 if (ret > 0)
1348                         ret = -ENOENT;
1349         }
1350 #endif
1351         return ret;
1352 }
1353
1354 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1355                                           struct btrfs_root *root,
1356                                           struct btrfs_path *path,
1357                                           u64 bytenr, u64 parent,
1358                                           u64 root_objectid)
1359 {
1360         struct btrfs_key key;
1361         int ret;
1362
1363         key.objectid = bytenr;
1364         if (parent) {
1365                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1366                 key.offset = parent;
1367         } else {
1368                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1369                 key.offset = root_objectid;
1370         }
1371
1372         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1373         btrfs_release_path(path);
1374         return ret;
1375 }
1376
1377 static inline int extent_ref_type(u64 parent, u64 owner)
1378 {
1379         int type;
1380         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1381                 if (parent > 0)
1382                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1383                 else
1384                         type = BTRFS_TREE_BLOCK_REF_KEY;
1385         } else {
1386                 if (parent > 0)
1387                         type = BTRFS_SHARED_DATA_REF_KEY;
1388                 else
1389                         type = BTRFS_EXTENT_DATA_REF_KEY;
1390         }
1391         return type;
1392 }
1393
1394 static int find_next_key(struct btrfs_path *path, int level,
1395                          struct btrfs_key *key)
1396
1397 {
1398         for (; level < BTRFS_MAX_LEVEL; level++) {
1399                 if (!path->nodes[level])
1400                         break;
1401                 if (path->slots[level] + 1 >=
1402                     btrfs_header_nritems(path->nodes[level]))
1403                         continue;
1404                 if (level == 0)
1405                         btrfs_item_key_to_cpu(path->nodes[level], key,
1406                                               path->slots[level] + 1);
1407                 else
1408                         btrfs_node_key_to_cpu(path->nodes[level], key,
1409                                               path->slots[level] + 1);
1410                 return 0;
1411         }
1412         return 1;
1413 }
1414
1415 /*
1416  * look for inline back ref. if back ref is found, *ref_ret is set
1417  * to the address of inline back ref, and 0 is returned.
1418  *
1419  * if back ref isn't found, *ref_ret is set to the address where it
1420  * should be inserted, and -ENOENT is returned.
1421  *
1422  * if insert is true and there are too many inline back refs, the path
1423  * points to the extent item, and -EAGAIN is returned.
1424  *
1425  * NOTE: inline back refs are ordered in the same way that back ref
1426  *       items in the tree are ordered.
1427  */
1428 static noinline_for_stack
1429 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1430                                  struct btrfs_root *root,
1431                                  struct btrfs_path *path,
1432                                  struct btrfs_extent_inline_ref **ref_ret,
1433                                  u64 bytenr, u64 num_bytes,
1434                                  u64 parent, u64 root_objectid,
1435                                  u64 owner, u64 offset, int insert)
1436 {
1437         struct btrfs_key key;
1438         struct extent_buffer *leaf;
1439         struct btrfs_extent_item *ei;
1440         struct btrfs_extent_inline_ref *iref;
1441         u64 flags;
1442         u64 item_size;
1443         unsigned long ptr;
1444         unsigned long end;
1445         int extra_size;
1446         int type;
1447         int want;
1448         int ret;
1449         int err = 0;
1450
1451         key.objectid = bytenr;
1452         key.type = BTRFS_EXTENT_ITEM_KEY;
1453         key.offset = num_bytes;
1454
1455         want = extent_ref_type(parent, owner);
1456         if (insert) {
1457                 extra_size = btrfs_extent_inline_ref_size(want);
1458                 path->keep_locks = 1;
1459         } else
1460                 extra_size = -1;
1461         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1462         if (ret < 0) {
1463                 err = ret;
1464                 goto out;
1465         }
1466         if (ret && !insert) {
1467                 err = -ENOENT;
1468                 goto out;
1469         }
1470         BUG_ON(ret); /* Corruption */
1471
1472         leaf = path->nodes[0];
1473         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1474 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1475         if (item_size < sizeof(*ei)) {
1476                 if (!insert) {
1477                         err = -ENOENT;
1478                         goto out;
1479                 }
1480                 ret = convert_extent_item_v0(trans, root, path, owner,
1481                                              extra_size);
1482                 if (ret < 0) {
1483                         err = ret;
1484                         goto out;
1485                 }
1486                 leaf = path->nodes[0];
1487                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1488         }
1489 #endif
1490         BUG_ON(item_size < sizeof(*ei));
1491
1492         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1493         flags = btrfs_extent_flags(leaf, ei);
1494
1495         ptr = (unsigned long)(ei + 1);
1496         end = (unsigned long)ei + item_size;
1497
1498         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1499                 ptr += sizeof(struct btrfs_tree_block_info);
1500                 BUG_ON(ptr > end);
1501         } else {
1502                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1503         }
1504
1505         err = -ENOENT;
1506         while (1) {
1507                 if (ptr >= end) {
1508                         WARN_ON(ptr > end);
1509                         break;
1510                 }
1511                 iref = (struct btrfs_extent_inline_ref *)ptr;
1512                 type = btrfs_extent_inline_ref_type(leaf, iref);
1513                 if (want < type)
1514                         break;
1515                 if (want > type) {
1516                         ptr += btrfs_extent_inline_ref_size(type);
1517                         continue;
1518                 }
1519
1520                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1521                         struct btrfs_extent_data_ref *dref;
1522                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1523                         if (match_extent_data_ref(leaf, dref, root_objectid,
1524                                                   owner, offset)) {
1525                                 err = 0;
1526                                 break;
1527                         }
1528                         if (hash_extent_data_ref_item(leaf, dref) <
1529                             hash_extent_data_ref(root_objectid, owner, offset))
1530                                 break;
1531                 } else {
1532                         u64 ref_offset;
1533                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1534                         if (parent > 0) {
1535                                 if (parent == ref_offset) {
1536                                         err = 0;
1537                                         break;
1538                                 }
1539                                 if (ref_offset < parent)
1540                                         break;
1541                         } else {
1542                                 if (root_objectid == ref_offset) {
1543                                         err = 0;
1544                                         break;
1545                                 }
1546                                 if (ref_offset < root_objectid)
1547                                         break;
1548                         }
1549                 }
1550                 ptr += btrfs_extent_inline_ref_size(type);
1551         }
1552         if (err == -ENOENT && insert) {
1553                 if (item_size + extra_size >=
1554                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1555                         err = -EAGAIN;
1556                         goto out;
1557                 }
1558                 /*
1559                  * To add new inline back ref, we have to make sure
1560                  * there is no corresponding back ref item.
1561                  * For simplicity, we just do not add new inline back
1562                  * ref if there is any kind of item for this block
1563                  */
1564                 if (find_next_key(path, 0, &key) == 0 &&
1565                     key.objectid == bytenr &&
1566                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1567                         err = -EAGAIN;
1568                         goto out;
1569                 }
1570         }
1571         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1572 out:
1573         if (insert) {
1574                 path->keep_locks = 0;
1575                 btrfs_unlock_up_safe(path, 1);
1576         }
1577         return err;
1578 }
1579
1580 /*
1581  * helper to add new inline back ref
1582  */
1583 static noinline_for_stack
1584 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1585                                  struct btrfs_root *root,
1586                                  struct btrfs_path *path,
1587                                  struct btrfs_extent_inline_ref *iref,
1588                                  u64 parent, u64 root_objectid,
1589                                  u64 owner, u64 offset, int refs_to_add,
1590                                  struct btrfs_delayed_extent_op *extent_op)
1591 {
1592         struct extent_buffer *leaf;
1593         struct btrfs_extent_item *ei;
1594         unsigned long ptr;
1595         unsigned long end;
1596         unsigned long item_offset;
1597         u64 refs;
1598         int size;
1599         int type;
1600
1601         leaf = path->nodes[0];
1602         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1603         item_offset = (unsigned long)iref - (unsigned long)ei;
1604
1605         type = extent_ref_type(parent, owner);
1606         size = btrfs_extent_inline_ref_size(type);
1607
1608         btrfs_extend_item(trans, root, path, size);
1609
1610         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1611         refs = btrfs_extent_refs(leaf, ei);
1612         refs += refs_to_add;
1613         btrfs_set_extent_refs(leaf, ei, refs);
1614         if (extent_op)
1615                 __run_delayed_extent_op(extent_op, leaf, ei);
1616
1617         ptr = (unsigned long)ei + item_offset;
1618         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1619         if (ptr < end - size)
1620                 memmove_extent_buffer(leaf, ptr + size, ptr,
1621                                       end - size - ptr);
1622
1623         iref = (struct btrfs_extent_inline_ref *)ptr;
1624         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1625         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1626                 struct btrfs_extent_data_ref *dref;
1627                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1628                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1629                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1630                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1631                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1632         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1633                 struct btrfs_shared_data_ref *sref;
1634                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1635                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1636                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1637         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1638                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1639         } else {
1640                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1641         }
1642         btrfs_mark_buffer_dirty(leaf);
1643 }
1644
1645 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1646                                  struct btrfs_root *root,
1647                                  struct btrfs_path *path,
1648                                  struct btrfs_extent_inline_ref **ref_ret,
1649                                  u64 bytenr, u64 num_bytes, u64 parent,
1650                                  u64 root_objectid, u64 owner, u64 offset)
1651 {
1652         int ret;
1653
1654         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1655                                            bytenr, num_bytes, parent,
1656                                            root_objectid, owner, offset, 0);
1657         if (ret != -ENOENT)
1658                 return ret;
1659
1660         btrfs_release_path(path);
1661         *ref_ret = NULL;
1662
1663         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1664                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1665                                             root_objectid);
1666         } else {
1667                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1668                                              root_objectid, owner, offset);
1669         }
1670         return ret;
1671 }
1672
1673 /*
1674  * helper to update/remove inline back ref
1675  */
1676 static noinline_for_stack
1677 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1678                                   struct btrfs_root *root,
1679                                   struct btrfs_path *path,
1680                                   struct btrfs_extent_inline_ref *iref,
1681                                   int refs_to_mod,
1682                                   struct btrfs_delayed_extent_op *extent_op)
1683 {
1684         struct extent_buffer *leaf;
1685         struct btrfs_extent_item *ei;
1686         struct btrfs_extent_data_ref *dref = NULL;
1687         struct btrfs_shared_data_ref *sref = NULL;
1688         unsigned long ptr;
1689         unsigned long end;
1690         u32 item_size;
1691         int size;
1692         int type;
1693         u64 refs;
1694
1695         leaf = path->nodes[0];
1696         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1697         refs = btrfs_extent_refs(leaf, ei);
1698         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1699         refs += refs_to_mod;
1700         btrfs_set_extent_refs(leaf, ei, refs);
1701         if (extent_op)
1702                 __run_delayed_extent_op(extent_op, leaf, ei);
1703
1704         type = btrfs_extent_inline_ref_type(leaf, iref);
1705
1706         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1707                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1708                 refs = btrfs_extent_data_ref_count(leaf, dref);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1711                 refs = btrfs_shared_data_ref_count(leaf, sref);
1712         } else {
1713                 refs = 1;
1714                 BUG_ON(refs_to_mod != -1);
1715         }
1716
1717         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1718         refs += refs_to_mod;
1719
1720         if (refs > 0) {
1721                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1722                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1723                 else
1724                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1725         } else {
1726                 size =  btrfs_extent_inline_ref_size(type);
1727                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1728                 ptr = (unsigned long)iref;
1729                 end = (unsigned long)ei + item_size;
1730                 if (ptr + size < end)
1731                         memmove_extent_buffer(leaf, ptr, ptr + size,
1732                                               end - ptr - size);
1733                 item_size -= size;
1734                 btrfs_truncate_item(trans, root, path, item_size, 1);
1735         }
1736         btrfs_mark_buffer_dirty(leaf);
1737 }
1738
1739 static noinline_for_stack
1740 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1741                                  struct btrfs_root *root,
1742                                  struct btrfs_path *path,
1743                                  u64 bytenr, u64 num_bytes, u64 parent,
1744                                  u64 root_objectid, u64 owner,
1745                                  u64 offset, int refs_to_add,
1746                                  struct btrfs_delayed_extent_op *extent_op)
1747 {
1748         struct btrfs_extent_inline_ref *iref;
1749         int ret;
1750
1751         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1752                                            bytenr, num_bytes, parent,
1753                                            root_objectid, owner, offset, 1);
1754         if (ret == 0) {
1755                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1756                 update_inline_extent_backref(trans, root, path, iref,
1757                                              refs_to_add, extent_op);
1758         } else if (ret == -ENOENT) {
1759                 setup_inline_extent_backref(trans, root, path, iref, parent,
1760                                             root_objectid, owner, offset,
1761                                             refs_to_add, extent_op);
1762                 ret = 0;
1763         }
1764         return ret;
1765 }
1766
1767 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1768                                  struct btrfs_root *root,
1769                                  struct btrfs_path *path,
1770                                  u64 bytenr, u64 parent, u64 root_objectid,
1771                                  u64 owner, u64 offset, int refs_to_add)
1772 {
1773         int ret;
1774         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1775                 BUG_ON(refs_to_add != 1);
1776                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1777                                             parent, root_objectid);
1778         } else {
1779                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1780                                              parent, root_objectid,
1781                                              owner, offset, refs_to_add);
1782         }
1783         return ret;
1784 }
1785
1786 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1787                                  struct btrfs_root *root,
1788                                  struct btrfs_path *path,
1789                                  struct btrfs_extent_inline_ref *iref,
1790                                  int refs_to_drop, int is_data)
1791 {
1792         int ret = 0;
1793
1794         BUG_ON(!is_data && refs_to_drop != 1);
1795         if (iref) {
1796                 update_inline_extent_backref(trans, root, path, iref,
1797                                              -refs_to_drop, NULL);
1798         } else if (is_data) {
1799                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1800         } else {
1801                 ret = btrfs_del_item(trans, root, path);
1802         }
1803         return ret;
1804 }
1805
1806 static int btrfs_issue_discard(struct block_device *bdev,
1807                                 u64 start, u64 len)
1808 {
1809         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1810 }
1811
1812 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1813                                 u64 num_bytes, u64 *actual_bytes)
1814 {
1815         int ret;
1816         u64 discarded_bytes = 0;
1817         struct btrfs_bio *bbio = NULL;
1818
1819
1820         /* Tell the block device(s) that the sectors can be discarded */
1821         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1822                               bytenr, &num_bytes, &bbio, 0);
1823         /* Error condition is -ENOMEM */
1824         if (!ret) {
1825                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1826                 int i;
1827
1828
1829                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1830                         if (!stripe->dev->can_discard)
1831                                 continue;
1832
1833                         ret = btrfs_issue_discard(stripe->dev->bdev,
1834                                                   stripe->physical,
1835                                                   stripe->length);
1836                         if (!ret)
1837                                 discarded_bytes += stripe->length;
1838                         else if (ret != -EOPNOTSUPP)
1839                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1840
1841                         /*
1842                          * Just in case we get back EOPNOTSUPP for some reason,
1843                          * just ignore the return value so we don't screw up
1844                          * people calling discard_extent.
1845                          */
1846                         ret = 0;
1847                 }
1848                 kfree(bbio);
1849         }
1850
1851         if (actual_bytes)
1852                 *actual_bytes = discarded_bytes;
1853
1854
1855         return ret;
1856 }
1857
1858 /* Can return -ENOMEM */
1859 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1860                          struct btrfs_root *root,
1861                          u64 bytenr, u64 num_bytes, u64 parent,
1862                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1863 {
1864         int ret;
1865         struct btrfs_fs_info *fs_info = root->fs_info;
1866
1867         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1868                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1869
1870         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1871                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1872                                         num_bytes,
1873                                         parent, root_objectid, (int)owner,
1874                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1875         } else {
1876                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1877                                         num_bytes,
1878                                         parent, root_objectid, owner, offset,
1879                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1880         }
1881         return ret;
1882 }
1883
1884 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1885                                   struct btrfs_root *root,
1886                                   u64 bytenr, u64 num_bytes,
1887                                   u64 parent, u64 root_objectid,
1888                                   u64 owner, u64 offset, int refs_to_add,
1889                                   struct btrfs_delayed_extent_op *extent_op)
1890 {
1891         struct btrfs_path *path;
1892         struct extent_buffer *leaf;
1893         struct btrfs_extent_item *item;
1894         u64 refs;
1895         int ret;
1896         int err = 0;
1897
1898         path = btrfs_alloc_path();
1899         if (!path)
1900                 return -ENOMEM;
1901
1902         path->reada = 1;
1903         path->leave_spinning = 1;
1904         /* this will setup the path even if it fails to insert the back ref */
1905         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1906                                            path, bytenr, num_bytes, parent,
1907                                            root_objectid, owner, offset,
1908                                            refs_to_add, extent_op);
1909         if (ret == 0)
1910                 goto out;
1911
1912         if (ret != -EAGAIN) {
1913                 err = ret;
1914                 goto out;
1915         }
1916
1917         leaf = path->nodes[0];
1918         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1919         refs = btrfs_extent_refs(leaf, item);
1920         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1921         if (extent_op)
1922                 __run_delayed_extent_op(extent_op, leaf, item);
1923
1924         btrfs_mark_buffer_dirty(leaf);
1925         btrfs_release_path(path);
1926
1927         path->reada = 1;
1928         path->leave_spinning = 1;
1929
1930         /* now insert the actual backref */
1931         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1932                                     path, bytenr, parent, root_objectid,
1933                                     owner, offset, refs_to_add);
1934         if (ret)
1935                 btrfs_abort_transaction(trans, root, ret);
1936 out:
1937         btrfs_free_path(path);
1938         return err;
1939 }
1940
1941 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1942                                 struct btrfs_root *root,
1943                                 struct btrfs_delayed_ref_node *node,
1944                                 struct btrfs_delayed_extent_op *extent_op,
1945                                 int insert_reserved)
1946 {
1947         int ret = 0;
1948         struct btrfs_delayed_data_ref *ref;
1949         struct btrfs_key ins;
1950         u64 parent = 0;
1951         u64 ref_root = 0;
1952         u64 flags = 0;
1953
1954         ins.objectid = node->bytenr;
1955         ins.offset = node->num_bytes;
1956         ins.type = BTRFS_EXTENT_ITEM_KEY;
1957
1958         ref = btrfs_delayed_node_to_data_ref(node);
1959         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1960                 parent = ref->parent;
1961         else
1962                 ref_root = ref->root;
1963
1964         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1965                 if (extent_op) {
1966                         BUG_ON(extent_op->update_key);
1967                         flags |= extent_op->flags_to_set;
1968                 }
1969                 ret = alloc_reserved_file_extent(trans, root,
1970                                                  parent, ref_root, flags,
1971                                                  ref->objectid, ref->offset,
1972                                                  &ins, node->ref_mod);
1973         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1974                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1975                                              node->num_bytes, parent,
1976                                              ref_root, ref->objectid,
1977                                              ref->offset, node->ref_mod,
1978                                              extent_op);
1979         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1980                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1981                                           node->num_bytes, parent,
1982                                           ref_root, ref->objectid,
1983                                           ref->offset, node->ref_mod,
1984                                           extent_op);
1985         } else {
1986                 BUG();
1987         }
1988         return ret;
1989 }
1990
1991 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1992                                     struct extent_buffer *leaf,
1993                                     struct btrfs_extent_item *ei)
1994 {
1995         u64 flags = btrfs_extent_flags(leaf, ei);
1996         if (extent_op->update_flags) {
1997                 flags |= extent_op->flags_to_set;
1998                 btrfs_set_extent_flags(leaf, ei, flags);
1999         }
2000
2001         if (extent_op->update_key) {
2002                 struct btrfs_tree_block_info *bi;
2003                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2004                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2005                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2006         }
2007 }
2008
2009 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2010                                  struct btrfs_root *root,
2011                                  struct btrfs_delayed_ref_node *node,
2012                                  struct btrfs_delayed_extent_op *extent_op)
2013 {
2014         struct btrfs_key key;
2015         struct btrfs_path *path;
2016         struct btrfs_extent_item *ei;
2017         struct extent_buffer *leaf;
2018         u32 item_size;
2019         int ret;
2020         int err = 0;
2021
2022         if (trans->aborted)
2023                 return 0;
2024
2025         path = btrfs_alloc_path();
2026         if (!path)
2027                 return -ENOMEM;
2028
2029         key.objectid = node->bytenr;
2030         key.type = BTRFS_EXTENT_ITEM_KEY;
2031         key.offset = node->num_bytes;
2032
2033         path->reada = 1;
2034         path->leave_spinning = 1;
2035         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2036                                 path, 0, 1);
2037         if (ret < 0) {
2038                 err = ret;
2039                 goto out;
2040         }
2041         if (ret > 0) {
2042                 err = -EIO;
2043                 goto out;
2044         }
2045
2046         leaf = path->nodes[0];
2047         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2048 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2049         if (item_size < sizeof(*ei)) {
2050                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2051                                              path, (u64)-1, 0);
2052                 if (ret < 0) {
2053                         err = ret;
2054                         goto out;
2055                 }
2056                 leaf = path->nodes[0];
2057                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2058         }
2059 #endif
2060         BUG_ON(item_size < sizeof(*ei));
2061         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2062         __run_delayed_extent_op(extent_op, leaf, ei);
2063
2064         btrfs_mark_buffer_dirty(leaf);
2065 out:
2066         btrfs_free_path(path);
2067         return err;
2068 }
2069
2070 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2071                                 struct btrfs_root *root,
2072                                 struct btrfs_delayed_ref_node *node,
2073                                 struct btrfs_delayed_extent_op *extent_op,
2074                                 int insert_reserved)
2075 {
2076         int ret = 0;
2077         struct btrfs_delayed_tree_ref *ref;
2078         struct btrfs_key ins;
2079         u64 parent = 0;
2080         u64 ref_root = 0;
2081
2082         ins.objectid = node->bytenr;
2083         ins.offset = node->num_bytes;
2084         ins.type = BTRFS_EXTENT_ITEM_KEY;
2085
2086         ref = btrfs_delayed_node_to_tree_ref(node);
2087         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2088                 parent = ref->parent;
2089         else
2090                 ref_root = ref->root;
2091
2092         BUG_ON(node->ref_mod != 1);
2093         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2094                 BUG_ON(!extent_op || !extent_op->update_flags ||
2095                        !extent_op->update_key);
2096                 ret = alloc_reserved_tree_block(trans, root,
2097                                                 parent, ref_root,
2098                                                 extent_op->flags_to_set,
2099                                                 &extent_op->key,
2100                                                 ref->level, &ins);
2101         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2102                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2103                                              node->num_bytes, parent, ref_root,
2104                                              ref->level, 0, 1, extent_op);
2105         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2106                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2107                                           node->num_bytes, parent, ref_root,
2108                                           ref->level, 0, 1, extent_op);
2109         } else {
2110                 BUG();
2111         }
2112         return ret;
2113 }
2114
2115 /* helper function to actually process a single delayed ref entry */
2116 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2117                                struct btrfs_root *root,
2118                                struct btrfs_delayed_ref_node *node,
2119                                struct btrfs_delayed_extent_op *extent_op,
2120                                int insert_reserved)
2121 {
2122         int ret = 0;
2123
2124         if (trans->aborted)
2125                 return 0;
2126
2127         if (btrfs_delayed_ref_is_head(node)) {
2128                 struct btrfs_delayed_ref_head *head;
2129                 /*
2130                  * we've hit the end of the chain and we were supposed
2131                  * to insert this extent into the tree.  But, it got
2132                  * deleted before we ever needed to insert it, so all
2133                  * we have to do is clean up the accounting
2134                  */
2135                 BUG_ON(extent_op);
2136                 head = btrfs_delayed_node_to_head(node);
2137                 if (insert_reserved) {
2138                         btrfs_pin_extent(root, node->bytenr,
2139                                          node->num_bytes, 1);
2140                         if (head->is_data) {
2141                                 ret = btrfs_del_csums(trans, root,
2142                                                       node->bytenr,
2143                                                       node->num_bytes);
2144                         }
2145                 }
2146                 mutex_unlock(&head->mutex);
2147                 return ret;
2148         }
2149
2150         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2151             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2152                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2153                                            insert_reserved);
2154         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2155                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2156                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2157                                            insert_reserved);
2158         else
2159                 BUG();
2160         return ret;
2161 }
2162
2163 static noinline struct btrfs_delayed_ref_node *
2164 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2165 {
2166         struct rb_node *node;
2167         struct btrfs_delayed_ref_node *ref;
2168         int action = BTRFS_ADD_DELAYED_REF;
2169 again:
2170         /*
2171          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2172          * this prevents ref count from going down to zero when
2173          * there still are pending delayed ref.
2174          */
2175         node = rb_prev(&head->node.rb_node);
2176         while (1) {
2177                 if (!node)
2178                         break;
2179                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2180                                 rb_node);
2181                 if (ref->bytenr != head->node.bytenr)
2182                         break;
2183                 if (ref->action == action)
2184                         return ref;
2185                 node = rb_prev(node);
2186         }
2187         if (action == BTRFS_ADD_DELAYED_REF) {
2188                 action = BTRFS_DROP_DELAYED_REF;
2189                 goto again;
2190         }
2191         return NULL;
2192 }
2193
2194 /*
2195  * Returns 0 on success or if called with an already aborted transaction.
2196  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2197  */
2198 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2199                                        struct btrfs_root *root,
2200                                        struct list_head *cluster)
2201 {
2202         struct btrfs_delayed_ref_root *delayed_refs;
2203         struct btrfs_delayed_ref_node *ref;
2204         struct btrfs_delayed_ref_head *locked_ref = NULL;
2205         struct btrfs_delayed_extent_op *extent_op;
2206         struct btrfs_fs_info *fs_info = root->fs_info;
2207         int ret;
2208         int count = 0;
2209         int must_insert_reserved = 0;
2210
2211         delayed_refs = &trans->transaction->delayed_refs;
2212         while (1) {
2213                 if (!locked_ref) {
2214                         /* pick a new head ref from the cluster list */
2215                         if (list_empty(cluster))
2216                                 break;
2217
2218                         locked_ref = list_entry(cluster->next,
2219                                      struct btrfs_delayed_ref_head, cluster);
2220
2221                         /* grab the lock that says we are going to process
2222                          * all the refs for this head */
2223                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2224
2225                         /*
2226                          * we may have dropped the spin lock to get the head
2227                          * mutex lock, and that might have given someone else
2228                          * time to free the head.  If that's true, it has been
2229                          * removed from our list and we can move on.
2230                          */
2231                         if (ret == -EAGAIN) {
2232                                 locked_ref = NULL;
2233                                 count++;
2234                                 continue;
2235                         }
2236                 }
2237
2238                 /*
2239                  * We need to try and merge add/drops of the same ref since we
2240                  * can run into issues with relocate dropping the implicit ref
2241                  * and then it being added back again before the drop can
2242                  * finish.  If we merged anything we need to re-loop so we can
2243                  * get a good ref.
2244                  */
2245                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2246                                          locked_ref);
2247
2248                 /*
2249                  * locked_ref is the head node, so we have to go one
2250                  * node back for any delayed ref updates
2251                  */
2252                 ref = select_delayed_ref(locked_ref);
2253
2254                 if (ref && ref->seq &&
2255                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2256                         /*
2257                          * there are still refs with lower seq numbers in the
2258                          * process of being added. Don't run this ref yet.
2259                          */
2260                         list_del_init(&locked_ref->cluster);
2261                         mutex_unlock(&locked_ref->mutex);
2262                         locked_ref = NULL;
2263                         delayed_refs->num_heads_ready++;
2264                         spin_unlock(&delayed_refs->lock);
2265                         cond_resched();
2266                         spin_lock(&delayed_refs->lock);
2267                         continue;
2268                 }
2269
2270                 /*
2271                  * record the must insert reserved flag before we
2272                  * drop the spin lock.
2273                  */
2274                 must_insert_reserved = locked_ref->must_insert_reserved;
2275                 locked_ref->must_insert_reserved = 0;
2276
2277                 extent_op = locked_ref->extent_op;
2278                 locked_ref->extent_op = NULL;
2279
2280                 if (!ref) {
2281                         /* All delayed refs have been processed, Go ahead
2282                          * and send the head node to run_one_delayed_ref,
2283                          * so that any accounting fixes can happen
2284                          */
2285                         ref = &locked_ref->node;
2286
2287                         if (extent_op && must_insert_reserved) {
2288                                 kfree(extent_op);
2289                                 extent_op = NULL;
2290                         }
2291
2292                         if (extent_op) {
2293                                 spin_unlock(&delayed_refs->lock);
2294
2295                                 ret = run_delayed_extent_op(trans, root,
2296                                                             ref, extent_op);
2297                                 kfree(extent_op);
2298
2299                                 if (ret) {
2300                                         list_del_init(&locked_ref->cluster);
2301                                         mutex_unlock(&locked_ref->mutex);
2302
2303                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2304                                         spin_lock(&delayed_refs->lock);
2305                                         return ret;
2306                                 }
2307
2308                                 goto next;
2309                         }
2310
2311                         list_del_init(&locked_ref->cluster);
2312                         locked_ref = NULL;
2313                 }
2314
2315                 ref->in_tree = 0;
2316                 rb_erase(&ref->rb_node, &delayed_refs->root);
2317                 delayed_refs->num_entries--;
2318                 if (locked_ref) {
2319                         /*
2320                          * when we play the delayed ref, also correct the
2321                          * ref_mod on head
2322                          */
2323                         switch (ref->action) {
2324                         case BTRFS_ADD_DELAYED_REF:
2325                         case BTRFS_ADD_DELAYED_EXTENT:
2326                                 locked_ref->node.ref_mod -= ref->ref_mod;
2327                                 break;
2328                         case BTRFS_DROP_DELAYED_REF:
2329                                 locked_ref->node.ref_mod += ref->ref_mod;
2330                                 break;
2331                         default:
2332                                 WARN_ON(1);
2333                         }
2334                 }
2335                 spin_unlock(&delayed_refs->lock);
2336
2337                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2338                                           must_insert_reserved);
2339
2340                 btrfs_put_delayed_ref(ref);
2341                 kfree(extent_op);
2342                 count++;
2343
2344                 if (ret) {
2345                         if (locked_ref) {
2346                                 list_del_init(&locked_ref->cluster);
2347                                 mutex_unlock(&locked_ref->mutex);
2348                         }
2349                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2350                         spin_lock(&delayed_refs->lock);
2351                         return ret;
2352                 }
2353
2354 next:
2355                 cond_resched();
2356                 spin_lock(&delayed_refs->lock);
2357         }
2358         return count;
2359 }
2360
2361 #ifdef SCRAMBLE_DELAYED_REFS
2362 /*
2363  * Normally delayed refs get processed in ascending bytenr order. This
2364  * correlates in most cases to the order added. To expose dependencies on this
2365  * order, we start to process the tree in the middle instead of the beginning
2366  */
2367 static u64 find_middle(struct rb_root *root)
2368 {
2369         struct rb_node *n = root->rb_node;
2370         struct btrfs_delayed_ref_node *entry;
2371         int alt = 1;
2372         u64 middle;
2373         u64 first = 0, last = 0;
2374
2375         n = rb_first(root);
2376         if (n) {
2377                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2378                 first = entry->bytenr;
2379         }
2380         n = rb_last(root);
2381         if (n) {
2382                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2383                 last = entry->bytenr;
2384         }
2385         n = root->rb_node;
2386
2387         while (n) {
2388                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2389                 WARN_ON(!entry->in_tree);
2390
2391                 middle = entry->bytenr;
2392
2393                 if (alt)
2394                         n = n->rb_left;
2395                 else
2396                         n = n->rb_right;
2397
2398                 alt = 1 - alt;
2399         }
2400         return middle;
2401 }
2402 #endif
2403
2404 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2405                                          struct btrfs_fs_info *fs_info)
2406 {
2407         struct qgroup_update *qgroup_update;
2408         int ret = 0;
2409
2410         if (list_empty(&trans->qgroup_ref_list) !=
2411             !trans->delayed_ref_elem.seq) {
2412                 /* list without seq or seq without list */
2413                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2414                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2415                         trans->delayed_ref_elem.seq);
2416                 BUG();
2417         }
2418
2419         if (!trans->delayed_ref_elem.seq)
2420                 return 0;
2421
2422         while (!list_empty(&trans->qgroup_ref_list)) {
2423                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2424                                                  struct qgroup_update, list);
2425                 list_del(&qgroup_update->list);
2426                 if (!ret)
2427                         ret = btrfs_qgroup_account_ref(
2428                                         trans, fs_info, qgroup_update->node,
2429                                         qgroup_update->extent_op);
2430                 kfree(qgroup_update);
2431         }
2432
2433         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2434
2435         return ret;
2436 }
2437
2438 /*
2439  * this starts processing the delayed reference count updates and
2440  * extent insertions we have queued up so far.  count can be
2441  * 0, which means to process everything in the tree at the start
2442  * of the run (but not newly added entries), or it can be some target
2443  * number you'd like to process.
2444  *
2445  * Returns 0 on success or if called with an aborted transaction
2446  * Returns <0 on error and aborts the transaction
2447  */
2448 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2449                            struct btrfs_root *root, unsigned long count)
2450 {
2451         struct rb_node *node;
2452         struct btrfs_delayed_ref_root *delayed_refs;
2453         struct btrfs_delayed_ref_node *ref;
2454         struct list_head cluster;
2455         int ret;
2456         u64 delayed_start;
2457         int run_all = count == (unsigned long)-1;
2458         int run_most = 0;
2459         int loops;
2460
2461         /* We'll clean this up in btrfs_cleanup_transaction */
2462         if (trans->aborted)
2463                 return 0;
2464
2465         if (root == root->fs_info->extent_root)
2466                 root = root->fs_info->tree_root;
2467
2468         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2469
2470         delayed_refs = &trans->transaction->delayed_refs;
2471         INIT_LIST_HEAD(&cluster);
2472 again:
2473         loops = 0;
2474         spin_lock(&delayed_refs->lock);
2475
2476 #ifdef SCRAMBLE_DELAYED_REFS
2477         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2478 #endif
2479
2480         if (count == 0) {
2481                 count = delayed_refs->num_entries * 2;
2482                 run_most = 1;
2483         }
2484         while (1) {
2485                 if (!(run_all || run_most) &&
2486                     delayed_refs->num_heads_ready < 64)
2487                         break;
2488
2489                 /*
2490                  * go find something we can process in the rbtree.  We start at
2491                  * the beginning of the tree, and then build a cluster
2492                  * of refs to process starting at the first one we are able to
2493                  * lock
2494                  */
2495                 delayed_start = delayed_refs->run_delayed_start;
2496                 ret = btrfs_find_ref_cluster(trans, &cluster,
2497                                              delayed_refs->run_delayed_start);
2498                 if (ret)
2499                         break;
2500
2501                 ret = run_clustered_refs(trans, root, &cluster);
2502                 if (ret < 0) {
2503                         spin_unlock(&delayed_refs->lock);
2504                         btrfs_abort_transaction(trans, root, ret);
2505                         return ret;
2506                 }
2507
2508                 count -= min_t(unsigned long, ret, count);
2509
2510                 if (count == 0)
2511                         break;
2512
2513                 if (delayed_start >= delayed_refs->run_delayed_start) {
2514                         if (loops == 0) {
2515                                 /*
2516                                  * btrfs_find_ref_cluster looped. let's do one
2517                                  * more cycle. if we don't run any delayed ref
2518                                  * during that cycle (because we can't because
2519                                  * all of them are blocked), bail out.
2520                                  */
2521                                 loops = 1;
2522                         } else {
2523                                 /*
2524                                  * no runnable refs left, stop trying
2525                                  */
2526                                 BUG_ON(run_all);
2527                                 break;
2528                         }
2529                 }
2530                 if (ret) {
2531                         /* refs were run, let's reset staleness detection */
2532                         loops = 0;
2533                 }
2534         }
2535
2536         if (run_all) {
2537                 if (!list_empty(&trans->new_bgs)) {
2538                         spin_unlock(&delayed_refs->lock);
2539                         btrfs_create_pending_block_groups(trans, root);
2540                         spin_lock(&delayed_refs->lock);
2541                 }
2542
2543                 node = rb_first(&delayed_refs->root);
2544                 if (!node)
2545                         goto out;
2546                 count = (unsigned long)-1;
2547
2548                 while (node) {
2549                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2550                                        rb_node);
2551                         if (btrfs_delayed_ref_is_head(ref)) {
2552                                 struct btrfs_delayed_ref_head *head;
2553
2554                                 head = btrfs_delayed_node_to_head(ref);
2555                                 atomic_inc(&ref->refs);
2556
2557                                 spin_unlock(&delayed_refs->lock);
2558                                 /*
2559                                  * Mutex was contended, block until it's
2560                                  * released and try again
2561                                  */
2562                                 mutex_lock(&head->mutex);
2563                                 mutex_unlock(&head->mutex);
2564
2565                                 btrfs_put_delayed_ref(ref);
2566                                 cond_resched();
2567                                 goto again;
2568                         }
2569                         node = rb_next(node);
2570                 }
2571                 spin_unlock(&delayed_refs->lock);
2572                 schedule_timeout(1);
2573                 goto again;
2574         }
2575 out:
2576         spin_unlock(&delayed_refs->lock);
2577         assert_qgroups_uptodate(trans);
2578         return 0;
2579 }
2580
2581 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2582                                 struct btrfs_root *root,
2583                                 u64 bytenr, u64 num_bytes, u64 flags,
2584                                 int is_data)
2585 {
2586         struct btrfs_delayed_extent_op *extent_op;
2587         int ret;
2588
2589         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2590         if (!extent_op)
2591                 return -ENOMEM;
2592
2593         extent_op->flags_to_set = flags;
2594         extent_op->update_flags = 1;
2595         extent_op->update_key = 0;
2596         extent_op->is_data = is_data ? 1 : 0;
2597
2598         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2599                                           num_bytes, extent_op);
2600         if (ret)
2601                 kfree(extent_op);
2602         return ret;
2603 }
2604
2605 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2606                                       struct btrfs_root *root,
2607                                       struct btrfs_path *path,
2608                                       u64 objectid, u64 offset, u64 bytenr)
2609 {
2610         struct btrfs_delayed_ref_head *head;
2611         struct btrfs_delayed_ref_node *ref;
2612         struct btrfs_delayed_data_ref *data_ref;
2613         struct btrfs_delayed_ref_root *delayed_refs;
2614         struct rb_node *node;
2615         int ret = 0;
2616
2617         ret = -ENOENT;
2618         delayed_refs = &trans->transaction->delayed_refs;
2619         spin_lock(&delayed_refs->lock);
2620         head = btrfs_find_delayed_ref_head(trans, bytenr);
2621         if (!head)
2622                 goto out;
2623
2624         if (!mutex_trylock(&head->mutex)) {
2625                 atomic_inc(&head->node.refs);
2626                 spin_unlock(&delayed_refs->lock);
2627
2628                 btrfs_release_path(path);
2629
2630                 /*
2631                  * Mutex was contended, block until it's released and let
2632                  * caller try again
2633                  */
2634                 mutex_lock(&head->mutex);
2635                 mutex_unlock(&head->mutex);
2636                 btrfs_put_delayed_ref(&head->node);
2637                 return -EAGAIN;
2638         }
2639
2640         node = rb_prev(&head->node.rb_node);
2641         if (!node)
2642                 goto out_unlock;
2643
2644         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2645
2646         if (ref->bytenr != bytenr)
2647                 goto out_unlock;
2648
2649         ret = 1;
2650         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2651                 goto out_unlock;
2652
2653         data_ref = btrfs_delayed_node_to_data_ref(ref);
2654
2655         node = rb_prev(node);
2656         if (node) {
2657                 int seq = ref->seq;
2658
2659                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2660                 if (ref->bytenr == bytenr && ref->seq == seq)
2661                         goto out_unlock;
2662         }
2663
2664         if (data_ref->root != root->root_key.objectid ||
2665             data_ref->objectid != objectid || data_ref->offset != offset)
2666                 goto out_unlock;
2667
2668         ret = 0;
2669 out_unlock:
2670         mutex_unlock(&head->mutex);
2671 out:
2672         spin_unlock(&delayed_refs->lock);
2673         return ret;
2674 }
2675
2676 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2677                                         struct btrfs_root *root,
2678                                         struct btrfs_path *path,
2679                                         u64 objectid, u64 offset, u64 bytenr)
2680 {
2681         struct btrfs_root *extent_root = root->fs_info->extent_root;
2682         struct extent_buffer *leaf;
2683         struct btrfs_extent_data_ref *ref;
2684         struct btrfs_extent_inline_ref *iref;
2685         struct btrfs_extent_item *ei;
2686         struct btrfs_key key;
2687         u32 item_size;
2688         int ret;
2689
2690         key.objectid = bytenr;
2691         key.offset = (u64)-1;
2692         key.type = BTRFS_EXTENT_ITEM_KEY;
2693
2694         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2695         if (ret < 0)
2696                 goto out;
2697         BUG_ON(ret == 0); /* Corruption */
2698
2699         ret = -ENOENT;
2700         if (path->slots[0] == 0)
2701                 goto out;
2702
2703         path->slots[0]--;
2704         leaf = path->nodes[0];
2705         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2706
2707         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2708                 goto out;
2709
2710         ret = 1;
2711         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2712 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2713         if (item_size < sizeof(*ei)) {
2714                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2715                 goto out;
2716         }
2717 #endif
2718         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2719
2720         if (item_size != sizeof(*ei) +
2721             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2722                 goto out;
2723
2724         if (btrfs_extent_generation(leaf, ei) <=
2725             btrfs_root_last_snapshot(&root->root_item))
2726                 goto out;
2727
2728         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2729         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2730             BTRFS_EXTENT_DATA_REF_KEY)
2731                 goto out;
2732
2733         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2734         if (btrfs_extent_refs(leaf, ei) !=
2735             btrfs_extent_data_ref_count(leaf, ref) ||
2736             btrfs_extent_data_ref_root(leaf, ref) !=
2737             root->root_key.objectid ||
2738             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2739             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2740                 goto out;
2741
2742         ret = 0;
2743 out:
2744         return ret;
2745 }
2746
2747 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2748                           struct btrfs_root *root,
2749                           u64 objectid, u64 offset, u64 bytenr)
2750 {
2751         struct btrfs_path *path;
2752         int ret;
2753         int ret2;
2754
2755         path = btrfs_alloc_path();
2756         if (!path)
2757                 return -ENOENT;
2758
2759         do {
2760                 ret = check_committed_ref(trans, root, path, objectid,
2761                                           offset, bytenr);
2762                 if (ret && ret != -ENOENT)
2763                         goto out;
2764
2765                 ret2 = check_delayed_ref(trans, root, path, objectid,
2766                                          offset, bytenr);
2767         } while (ret2 == -EAGAIN);
2768
2769         if (ret2 && ret2 != -ENOENT) {
2770                 ret = ret2;
2771                 goto out;
2772         }
2773
2774         if (ret != -ENOENT || ret2 != -ENOENT)
2775                 ret = 0;
2776 out:
2777         btrfs_free_path(path);
2778         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2779                 WARN_ON(ret > 0);
2780         return ret;
2781 }
2782
2783 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2784                            struct btrfs_root *root,
2785                            struct extent_buffer *buf,
2786                            int full_backref, int inc, int for_cow)
2787 {
2788         u64 bytenr;
2789         u64 num_bytes;
2790         u64 parent;
2791         u64 ref_root;
2792         u32 nritems;
2793         struct btrfs_key key;
2794         struct btrfs_file_extent_item *fi;
2795         int i;
2796         int level;
2797         int ret = 0;
2798         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2799                             u64, u64, u64, u64, u64, u64, int);
2800
2801         ref_root = btrfs_header_owner(buf);
2802         nritems = btrfs_header_nritems(buf);
2803         level = btrfs_header_level(buf);
2804
2805         if (!root->ref_cows && level == 0)
2806                 return 0;
2807
2808         if (inc)
2809                 process_func = btrfs_inc_extent_ref;
2810         else
2811                 process_func = btrfs_free_extent;
2812
2813         if (full_backref)
2814                 parent = buf->start;
2815         else
2816                 parent = 0;
2817
2818         for (i = 0; i < nritems; i++) {
2819                 if (level == 0) {
2820                         btrfs_item_key_to_cpu(buf, &key, i);
2821                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2822                                 continue;
2823                         fi = btrfs_item_ptr(buf, i,
2824                                             struct btrfs_file_extent_item);
2825                         if (btrfs_file_extent_type(buf, fi) ==
2826                             BTRFS_FILE_EXTENT_INLINE)
2827                                 continue;
2828                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2829                         if (bytenr == 0)
2830                                 continue;
2831
2832                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2833                         key.offset -= btrfs_file_extent_offset(buf, fi);
2834                         ret = process_func(trans, root, bytenr, num_bytes,
2835                                            parent, ref_root, key.objectid,
2836                                            key.offset, for_cow);
2837                         if (ret)
2838                                 goto fail;
2839                 } else {
2840                         bytenr = btrfs_node_blockptr(buf, i);
2841                         num_bytes = btrfs_level_size(root, level - 1);
2842                         ret = process_func(trans, root, bytenr, num_bytes,
2843                                            parent, ref_root, level - 1, 0,
2844                                            for_cow);
2845                         if (ret)
2846                                 goto fail;
2847                 }
2848         }
2849         return 0;
2850 fail:
2851         return ret;
2852 }
2853
2854 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2855                   struct extent_buffer *buf, int full_backref, int for_cow)
2856 {
2857         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2858 }
2859
2860 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2861                   struct extent_buffer *buf, int full_backref, int for_cow)
2862 {
2863         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2864 }
2865
2866 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2867                                  struct btrfs_root *root,
2868                                  struct btrfs_path *path,
2869                                  struct btrfs_block_group_cache *cache)
2870 {
2871         int ret;
2872         struct btrfs_root *extent_root = root->fs_info->extent_root;
2873         unsigned long bi;
2874         struct extent_buffer *leaf;
2875
2876         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2877         if (ret < 0)
2878                 goto fail;
2879         BUG_ON(ret); /* Corruption */
2880
2881         leaf = path->nodes[0];
2882         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2883         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2884         btrfs_mark_buffer_dirty(leaf);
2885         btrfs_release_path(path);
2886 fail:
2887         if (ret) {
2888                 btrfs_abort_transaction(trans, root, ret);
2889                 return ret;
2890         }
2891         return 0;
2892
2893 }
2894
2895 static struct btrfs_block_group_cache *
2896 next_block_group(struct btrfs_root *root,
2897                  struct btrfs_block_group_cache *cache)
2898 {
2899         struct rb_node *node;
2900         spin_lock(&root->fs_info->block_group_cache_lock);
2901         node = rb_next(&cache->cache_node);
2902         btrfs_put_block_group(cache);
2903         if (node) {
2904                 cache = rb_entry(node, struct btrfs_block_group_cache,
2905                                  cache_node);
2906                 btrfs_get_block_group(cache);
2907         } else
2908                 cache = NULL;
2909         spin_unlock(&root->fs_info->block_group_cache_lock);
2910         return cache;
2911 }
2912
2913 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2914                             struct btrfs_trans_handle *trans,
2915                             struct btrfs_path *path)
2916 {
2917         struct btrfs_root *root = block_group->fs_info->tree_root;
2918         struct inode *inode = NULL;
2919         u64 alloc_hint = 0;
2920         int dcs = BTRFS_DC_ERROR;
2921         int num_pages = 0;
2922         int retries = 0;
2923         int ret = 0;
2924
2925         /*
2926          * If this block group is smaller than 100 megs don't bother caching the
2927          * block group.
2928          */
2929         if (block_group->key.offset < (100 * 1024 * 1024)) {
2930                 spin_lock(&block_group->lock);
2931                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2932                 spin_unlock(&block_group->lock);
2933                 return 0;
2934         }
2935
2936 again:
2937         inode = lookup_free_space_inode(root, block_group, path);
2938         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2939                 ret = PTR_ERR(inode);
2940                 btrfs_release_path(path);
2941                 goto out;
2942         }
2943
2944         if (IS_ERR(inode)) {
2945                 BUG_ON(retries);
2946                 retries++;
2947
2948                 if (block_group->ro)
2949                         goto out_free;
2950
2951                 ret = create_free_space_inode(root, trans, block_group, path);
2952                 if (ret)
2953                         goto out_free;
2954                 goto again;
2955         }
2956
2957         /* We've already setup this transaction, go ahead and exit */
2958         if (block_group->cache_generation == trans->transid &&
2959             i_size_read(inode)) {
2960                 dcs = BTRFS_DC_SETUP;
2961                 goto out_put;
2962         }
2963
2964         /*
2965          * We want to set the generation to 0, that way if anything goes wrong
2966          * from here on out we know not to trust this cache when we load up next
2967          * time.
2968          */
2969         BTRFS_I(inode)->generation = 0;
2970         ret = btrfs_update_inode(trans, root, inode);
2971         WARN_ON(ret);
2972
2973         if (i_size_read(inode) > 0) {
2974                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2975                                                       inode);
2976                 if (ret)
2977                         goto out_put;
2978         }
2979
2980         spin_lock(&block_group->lock);
2981         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2982             !btrfs_test_opt(root, SPACE_CACHE)) {
2983                 /*
2984                  * don't bother trying to write stuff out _if_
2985                  * a) we're not cached,
2986                  * b) we're with nospace_cache mount option.
2987                  */
2988                 dcs = BTRFS_DC_WRITTEN;
2989                 spin_unlock(&block_group->lock);
2990                 goto out_put;
2991         }
2992         spin_unlock(&block_group->lock);
2993
2994         /*
2995          * Try to preallocate enough space based on how big the block group is.
2996          * Keep in mind this has to include any pinned space which could end up
2997          * taking up quite a bit since it's not folded into the other space
2998          * cache.
2999          */
3000         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3001         if (!num_pages)
3002                 num_pages = 1;
3003
3004         num_pages *= 16;
3005         num_pages *= PAGE_CACHE_SIZE;
3006
3007         ret = btrfs_check_data_free_space(inode, num_pages);
3008         if (ret)
3009                 goto out_put;
3010
3011         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3012                                               num_pages, num_pages,
3013                                               &alloc_hint);
3014         if (!ret)
3015                 dcs = BTRFS_DC_SETUP;
3016         btrfs_free_reserved_data_space(inode, num_pages);
3017
3018 out_put:
3019         iput(inode);
3020 out_free:
3021         btrfs_release_path(path);
3022 out:
3023         spin_lock(&block_group->lock);
3024         if (!ret && dcs == BTRFS_DC_SETUP)
3025                 block_group->cache_generation = trans->transid;
3026         block_group->disk_cache_state = dcs;
3027         spin_unlock(&block_group->lock);
3028
3029         return ret;
3030 }
3031
3032 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3033                                    struct btrfs_root *root)
3034 {
3035         struct btrfs_block_group_cache *cache;
3036         int err = 0;
3037         struct btrfs_path *path;
3038         u64 last = 0;
3039
3040         path = btrfs_alloc_path();
3041         if (!path)
3042                 return -ENOMEM;
3043
3044 again:
3045         while (1) {
3046                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3047                 while (cache) {
3048                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3049                                 break;
3050                         cache = next_block_group(root, cache);
3051                 }
3052                 if (!cache) {
3053                         if (last == 0)
3054                                 break;
3055                         last = 0;
3056                         continue;
3057                 }
3058                 err = cache_save_setup(cache, trans, path);
3059                 last = cache->key.objectid + cache->key.offset;
3060                 btrfs_put_block_group(cache);
3061         }
3062
3063         while (1) {
3064                 if (last == 0) {
3065                         err = btrfs_run_delayed_refs(trans, root,
3066                                                      (unsigned long)-1);
3067                         if (err) /* File system offline */
3068                                 goto out;
3069                 }
3070
3071                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3072                 while (cache) {
3073                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3074                                 btrfs_put_block_group(cache);
3075                                 goto again;
3076                         }
3077
3078                         if (cache->dirty)
3079                                 break;
3080                         cache = next_block_group(root, cache);
3081                 }
3082                 if (!cache) {
3083                         if (last == 0)
3084                                 break;
3085                         last = 0;
3086                         continue;
3087                 }
3088
3089                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3090                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3091                 cache->dirty = 0;
3092                 last = cache->key.objectid + cache->key.offset;
3093
3094                 err = write_one_cache_group(trans, root, path, cache);
3095                 if (err) /* File system offline */
3096                         goto out;
3097
3098                 btrfs_put_block_group(cache);
3099         }
3100
3101         while (1) {
3102                 /*
3103                  * I don't think this is needed since we're just marking our
3104                  * preallocated extent as written, but just in case it can't
3105                  * hurt.
3106                  */
3107                 if (last == 0) {
3108                         err = btrfs_run_delayed_refs(trans, root,
3109                                                      (unsigned long)-1);
3110                         if (err) /* File system offline */
3111                                 goto out;
3112                 }
3113
3114                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3115                 while (cache) {
3116                         /*
3117                          * Really this shouldn't happen, but it could if we
3118                          * couldn't write the entire preallocated extent and
3119                          * splitting the extent resulted in a new block.
3120                          */
3121                         if (cache->dirty) {
3122                                 btrfs_put_block_group(cache);
3123                                 goto again;
3124                         }
3125                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3126                                 break;
3127                         cache = next_block_group(root, cache);
3128                 }
3129                 if (!cache) {
3130                         if (last == 0)
3131                                 break;
3132                         last = 0;
3133                         continue;
3134                 }
3135
3136                 err = btrfs_write_out_cache(root, trans, cache, path);
3137
3138                 /*
3139                  * If we didn't have an error then the cache state is still
3140                  * NEED_WRITE, so we can set it to WRITTEN.
3141                  */
3142                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3143                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3144                 last = cache->key.objectid + cache->key.offset;
3145                 btrfs_put_block_group(cache);
3146         }
3147 out:
3148
3149         btrfs_free_path(path);
3150         return err;
3151 }
3152
3153 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3154 {
3155         struct btrfs_block_group_cache *block_group;
3156         int readonly = 0;
3157
3158         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3159         if (!block_group || block_group->ro)
3160          &n