Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux...
[sfrench/cifs-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/pagemap.h>
20 #include <linux/writeback.h>
21 #include <linux/blkdev.h>
22 #include <linux/sort.h>
23 #include <linux/rcupdate.h>
24 #include <linux/kthread.h>
25 #include <linux/slab.h>
26 #include <linux/ratelimit.h>
27 #include <linux/percpu_counter.h>
28 #include "hash.h"
29 #include "tree-log.h"
30 #include "disk-io.h"
31 #include "print-tree.h"
32 #include "volumes.h"
33 #include "raid56.h"
34 #include "locking.h"
35 #include "free-space-cache.h"
36 #include "math.h"
37 #include "sysfs.h"
38 #include "qgroup.h"
39
40 #undef SCRAMBLE_DELAYED_REFS
41
42 /*
43  * control flags for do_chunk_alloc's force field
44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
45  * if we really need one.
46  *
47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
48  * if we have very few chunks already allocated.  This is
49  * used as part of the clustering code to help make sure
50  * we have a good pool of storage to cluster in, without
51  * filling the FS with empty chunks
52  *
53  * CHUNK_ALLOC_FORCE means it must try to allocate one
54  *
55  */
56 enum {
57         CHUNK_ALLOC_NO_FORCE = 0,
58         CHUNK_ALLOC_LIMITED = 1,
59         CHUNK_ALLOC_FORCE = 2,
60 };
61
62 /*
63  * Control how reservations are dealt with.
64  *
65  * RESERVE_FREE - freeing a reservation.
66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
67  *   ENOSPC accounting
68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
70  */
71 enum {
72         RESERVE_FREE = 0,
73         RESERVE_ALLOC = 1,
74         RESERVE_ALLOC_NO_ACCOUNT = 2,
75 };
76
77 static int update_block_group(struct btrfs_root *root,
78                               u64 bytenr, u64 num_bytes, int alloc);
79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
80                                 struct btrfs_root *root,
81                                 u64 bytenr, u64 num_bytes, u64 parent,
82                                 u64 root_objectid, u64 owner_objectid,
83                                 u64 owner_offset, int refs_to_drop,
84                                 struct btrfs_delayed_extent_op *extra_op,
85                                 int no_quota);
86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
87                                     struct extent_buffer *leaf,
88                                     struct btrfs_extent_item *ei);
89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
90                                       struct btrfs_root *root,
91                                       u64 parent, u64 root_objectid,
92                                       u64 flags, u64 owner, u64 offset,
93                                       struct btrfs_key *ins, int ref_mod);
94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
95                                      struct btrfs_root *root,
96                                      u64 parent, u64 root_objectid,
97                                      u64 flags, struct btrfs_disk_key *key,
98                                      int level, struct btrfs_key *ins,
99                                      int no_quota);
100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
101                           struct btrfs_root *extent_root, u64 flags,
102                           int force);
103 static int find_next_key(struct btrfs_path *path, int level,
104                          struct btrfs_key *key);
105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
106                             int dump_block_groups);
107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
108                                        u64 num_bytes, int reserve,
109                                        int delalloc);
110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
111                                u64 num_bytes);
112 int btrfs_pin_extent(struct btrfs_root *root,
113                      u64 bytenr, u64 num_bytes, int reserved);
114
115 static noinline int
116 block_group_cache_done(struct btrfs_block_group_cache *cache)
117 {
118         smp_mb();
119         return cache->cached == BTRFS_CACHE_FINISHED ||
120                 cache->cached == BTRFS_CACHE_ERROR;
121 }
122
123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
124 {
125         return (cache->flags & bits) == bits;
126 }
127
128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
129 {
130         atomic_inc(&cache->count);
131 }
132
133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
134 {
135         if (atomic_dec_and_test(&cache->count)) {
136                 WARN_ON(cache->pinned > 0);
137                 WARN_ON(cache->reserved > 0);
138                 kfree(cache->free_space_ctl);
139                 kfree(cache);
140         }
141 }
142
143 /*
144  * this adds the block group to the fs_info rb tree for the block group
145  * cache
146  */
147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
148                                 struct btrfs_block_group_cache *block_group)
149 {
150         struct rb_node **p;
151         struct rb_node *parent = NULL;
152         struct btrfs_block_group_cache *cache;
153
154         spin_lock(&info->block_group_cache_lock);
155         p = &info->block_group_cache_tree.rb_node;
156
157         while (*p) {
158                 parent = *p;
159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
160                                  cache_node);
161                 if (block_group->key.objectid < cache->key.objectid) {
162                         p = &(*p)->rb_left;
163                 } else if (block_group->key.objectid > cache->key.objectid) {
164                         p = &(*p)->rb_right;
165                 } else {
166                         spin_unlock(&info->block_group_cache_lock);
167                         return -EEXIST;
168                 }
169         }
170
171         rb_link_node(&block_group->cache_node, parent, p);
172         rb_insert_color(&block_group->cache_node,
173                         &info->block_group_cache_tree);
174
175         if (info->first_logical_byte > block_group->key.objectid)
176                 info->first_logical_byte = block_group->key.objectid;
177
178         spin_unlock(&info->block_group_cache_lock);
179
180         return 0;
181 }
182
183 /*
184  * This will return the block group at or after bytenr if contains is 0, else
185  * it will return the block group that contains the bytenr
186  */
187 static struct btrfs_block_group_cache *
188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
189                               int contains)
190 {
191         struct btrfs_block_group_cache *cache, *ret = NULL;
192         struct rb_node *n;
193         u64 end, start;
194
195         spin_lock(&info->block_group_cache_lock);
196         n = info->block_group_cache_tree.rb_node;
197
198         while (n) {
199                 cache = rb_entry(n, struct btrfs_block_group_cache,
200                                  cache_node);
201                 end = cache->key.objectid + cache->key.offset - 1;
202                 start = cache->key.objectid;
203
204                 if (bytenr < start) {
205                         if (!contains && (!ret || start < ret->key.objectid))
206                                 ret = cache;
207                         n = n->rb_left;
208                 } else if (bytenr > start) {
209                         if (contains && bytenr <= end) {
210                                 ret = cache;
211                                 break;
212                         }
213                         n = n->rb_right;
214                 } else {
215                         ret = cache;
216                         break;
217                 }
218         }
219         if (ret) {
220                 btrfs_get_block_group(ret);
221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
222                         info->first_logical_byte = ret->key.objectid;
223         }
224         spin_unlock(&info->block_group_cache_lock);
225
226         return ret;
227 }
228
229 static int add_excluded_extent(struct btrfs_root *root,
230                                u64 start, u64 num_bytes)
231 {
232         u64 end = start + num_bytes - 1;
233         set_extent_bits(&root->fs_info->freed_extents[0],
234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
235         set_extent_bits(&root->fs_info->freed_extents[1],
236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
237         return 0;
238 }
239
240 static void free_excluded_extents(struct btrfs_root *root,
241                                   struct btrfs_block_group_cache *cache)
242 {
243         u64 start, end;
244
245         start = cache->key.objectid;
246         end = start + cache->key.offset - 1;
247
248         clear_extent_bits(&root->fs_info->freed_extents[0],
249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
250         clear_extent_bits(&root->fs_info->freed_extents[1],
251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
252 }
253
254 static int exclude_super_stripes(struct btrfs_root *root,
255                                  struct btrfs_block_group_cache *cache)
256 {
257         u64 bytenr;
258         u64 *logical;
259         int stripe_len;
260         int i, nr, ret;
261
262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
264                 cache->bytes_super += stripe_len;
265                 ret = add_excluded_extent(root, cache->key.objectid,
266                                           stripe_len);
267                 if (ret)
268                         return ret;
269         }
270
271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
272                 bytenr = btrfs_sb_offset(i);
273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
274                                        cache->key.objectid, bytenr,
275                                        0, &logical, &nr, &stripe_len);
276                 if (ret)
277                         return ret;
278
279                 while (nr--) {
280                         u64 start, len;
281
282                         if (logical[nr] > cache->key.objectid +
283                             cache->key.offset)
284                                 continue;
285
286                         if (logical[nr] + stripe_len <= cache->key.objectid)
287                                 continue;
288
289                         start = logical[nr];
290                         if (start < cache->key.objectid) {
291                                 start = cache->key.objectid;
292                                 len = (logical[nr] + stripe_len) - start;
293                         } else {
294                                 len = min_t(u64, stripe_len,
295                                             cache->key.objectid +
296                                             cache->key.offset - start);
297                         }
298
299                         cache->bytes_super += len;
300                         ret = add_excluded_extent(root, start, len);
301                         if (ret) {
302                                 kfree(logical);
303                                 return ret;
304                         }
305                 }
306
307                 kfree(logical);
308         }
309         return 0;
310 }
311
312 static struct btrfs_caching_control *
313 get_caching_control(struct btrfs_block_group_cache *cache)
314 {
315         struct btrfs_caching_control *ctl;
316
317         spin_lock(&cache->lock);
318         if (cache->cached != BTRFS_CACHE_STARTED) {
319                 spin_unlock(&cache->lock);
320                 return NULL;
321         }
322
323         /* We're loading it the fast way, so we don't have a caching_ctl. */
324         if (!cache->caching_ctl) {
325                 spin_unlock(&cache->lock);
326                 return NULL;
327         }
328
329         ctl = cache->caching_ctl;
330         atomic_inc(&ctl->count);
331         spin_unlock(&cache->lock);
332         return ctl;
333 }
334
335 static void put_caching_control(struct btrfs_caching_control *ctl)
336 {
337         if (atomic_dec_and_test(&ctl->count))
338                 kfree(ctl);
339 }
340
341 /*
342  * this is only called by cache_block_group, since we could have freed extents
343  * we need to check the pinned_extents for any extents that can't be used yet
344  * since their free space will be released as soon as the transaction commits.
345  */
346 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
347                               struct btrfs_fs_info *info, u64 start, u64 end)
348 {
349         u64 extent_start, extent_end, size, total_added = 0;
350         int ret;
351
352         while (start < end) {
353                 ret = find_first_extent_bit(info->pinned_extents, start,
354                                             &extent_start, &extent_end,
355                                             EXTENT_DIRTY | EXTENT_UPTODATE,
356                                             NULL);
357                 if (ret)
358                         break;
359
360                 if (extent_start <= start) {
361                         start = extent_end + 1;
362                 } else if (extent_start > start && extent_start < end) {
363                         size = extent_start - start;
364                         total_added += size;
365                         ret = btrfs_add_free_space(block_group, start,
366                                                    size);
367                         BUG_ON(ret); /* -ENOMEM or logic error */
368                         start = extent_end + 1;
369                 } else {
370                         break;
371                 }
372         }
373
374         if (start < end) {
375                 size = end - start;
376                 total_added += size;
377                 ret = btrfs_add_free_space(block_group, start, size);
378                 BUG_ON(ret); /* -ENOMEM or logic error */
379         }
380
381         return total_added;
382 }
383
384 static noinline void caching_thread(struct btrfs_work *work)
385 {
386         struct btrfs_block_group_cache *block_group;
387         struct btrfs_fs_info *fs_info;
388         struct btrfs_caching_control *caching_ctl;
389         struct btrfs_root *extent_root;
390         struct btrfs_path *path;
391         struct extent_buffer *leaf;
392         struct btrfs_key key;
393         u64 total_found = 0;
394         u64 last = 0;
395         u32 nritems;
396         int ret = -ENOMEM;
397
398         caching_ctl = container_of(work, struct btrfs_caching_control, work);
399         block_group = caching_ctl->block_group;
400         fs_info = block_group->fs_info;
401         extent_root = fs_info->extent_root;
402
403         path = btrfs_alloc_path();
404         if (!path)
405                 goto out;
406
407         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
408
409         /*
410          * We don't want to deadlock with somebody trying to allocate a new
411          * extent for the extent root while also trying to search the extent
412          * root to add free space.  So we skip locking and search the commit
413          * root, since its read-only
414          */
415         path->skip_locking = 1;
416         path->search_commit_root = 1;
417         path->reada = 1;
418
419         key.objectid = last;
420         key.offset = 0;
421         key.type = BTRFS_EXTENT_ITEM_KEY;
422 again:
423         mutex_lock(&caching_ctl->mutex);
424         /* need to make sure the commit_root doesn't disappear */
425         down_read(&fs_info->commit_root_sem);
426
427 next:
428         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
429         if (ret < 0)
430                 goto err;
431
432         leaf = path->nodes[0];
433         nritems = btrfs_header_nritems(leaf);
434
435         while (1) {
436                 if (btrfs_fs_closing(fs_info) > 1) {
437                         last = (u64)-1;
438                         break;
439                 }
440
441                 if (path->slots[0] < nritems) {
442                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
443                 } else {
444                         ret = find_next_key(path, 0, &key);
445                         if (ret)
446                                 break;
447
448                         if (need_resched() ||
449                             rwsem_is_contended(&fs_info->commit_root_sem)) {
450                                 caching_ctl->progress = last;
451                                 btrfs_release_path(path);
452                                 up_read(&fs_info->commit_root_sem);
453                                 mutex_unlock(&caching_ctl->mutex);
454                                 cond_resched();
455                                 goto again;
456                         }
457
458                         ret = btrfs_next_leaf(extent_root, path);
459                         if (ret < 0)
460                                 goto err;
461                         if (ret)
462                                 break;
463                         leaf = path->nodes[0];
464                         nritems = btrfs_header_nritems(leaf);
465                         continue;
466                 }
467
468                 if (key.objectid < last) {
469                         key.objectid = last;
470                         key.offset = 0;
471                         key.type = BTRFS_EXTENT_ITEM_KEY;
472
473                         caching_ctl->progress = last;
474                         btrfs_release_path(path);
475                         goto next;
476                 }
477
478                 if (key.objectid < block_group->key.objectid) {
479                         path->slots[0]++;
480                         continue;
481                 }
482
483                 if (key.objectid >= block_group->key.objectid +
484                     block_group->key.offset)
485                         break;
486
487                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
488                     key.type == BTRFS_METADATA_ITEM_KEY) {
489                         total_found += add_new_free_space(block_group,
490                                                           fs_info, last,
491                                                           key.objectid);
492                         if (key.type == BTRFS_METADATA_ITEM_KEY)
493                                 last = key.objectid +
494                                         fs_info->tree_root->nodesize;
495                         else
496                                 last = key.objectid + key.offset;
497
498                         if (total_found > (1024 * 1024 * 2)) {
499                                 total_found = 0;
500                                 wake_up(&caching_ctl->wait);
501                         }
502                 }
503                 path->slots[0]++;
504         }
505         ret = 0;
506
507         total_found += add_new_free_space(block_group, fs_info, last,
508                                           block_group->key.objectid +
509                                           block_group->key.offset);
510         caching_ctl->progress = (u64)-1;
511
512         spin_lock(&block_group->lock);
513         block_group->caching_ctl = NULL;
514         block_group->cached = BTRFS_CACHE_FINISHED;
515         spin_unlock(&block_group->lock);
516
517 err:
518         btrfs_free_path(path);
519         up_read(&fs_info->commit_root_sem);
520
521         free_excluded_extents(extent_root, block_group);
522
523         mutex_unlock(&caching_ctl->mutex);
524 out:
525         if (ret) {
526                 spin_lock(&block_group->lock);
527                 block_group->caching_ctl = NULL;
528                 block_group->cached = BTRFS_CACHE_ERROR;
529                 spin_unlock(&block_group->lock);
530         }
531         wake_up(&caching_ctl->wait);
532
533         put_caching_control(caching_ctl);
534         btrfs_put_block_group(block_group);
535 }
536
537 static int cache_block_group(struct btrfs_block_group_cache *cache,
538                              int load_cache_only)
539 {
540         DEFINE_WAIT(wait);
541         struct btrfs_fs_info *fs_info = cache->fs_info;
542         struct btrfs_caching_control *caching_ctl;
543         int ret = 0;
544
545         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
546         if (!caching_ctl)
547                 return -ENOMEM;
548
549         INIT_LIST_HEAD(&caching_ctl->list);
550         mutex_init(&caching_ctl->mutex);
551         init_waitqueue_head(&caching_ctl->wait);
552         caching_ctl->block_group = cache;
553         caching_ctl->progress = cache->key.objectid;
554         atomic_set(&caching_ctl->count, 1);
555         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
556                         caching_thread, NULL, NULL);
557
558         spin_lock(&cache->lock);
559         /*
560          * This should be a rare occasion, but this could happen I think in the
561          * case where one thread starts to load the space cache info, and then
562          * some other thread starts a transaction commit which tries to do an
563          * allocation while the other thread is still loading the space cache
564          * info.  The previous loop should have kept us from choosing this block
565          * group, but if we've moved to the state where we will wait on caching
566          * block groups we need to first check if we're doing a fast load here,
567          * so we can wait for it to finish, otherwise we could end up allocating
568          * from a block group who's cache gets evicted for one reason or
569          * another.
570          */
571         while (cache->cached == BTRFS_CACHE_FAST) {
572                 struct btrfs_caching_control *ctl;
573
574                 ctl = cache->caching_ctl;
575                 atomic_inc(&ctl->count);
576                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
577                 spin_unlock(&cache->lock);
578
579                 schedule();
580
581                 finish_wait(&ctl->wait, &wait);
582                 put_caching_control(ctl);
583                 spin_lock(&cache->lock);
584         }
585
586         if (cache->cached != BTRFS_CACHE_NO) {
587                 spin_unlock(&cache->lock);
588                 kfree(caching_ctl);
589                 return 0;
590         }
591         WARN_ON(cache->caching_ctl);
592         cache->caching_ctl = caching_ctl;
593         cache->cached = BTRFS_CACHE_FAST;
594         spin_unlock(&cache->lock);
595
596         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
597                 ret = load_free_space_cache(fs_info, cache);
598
599                 spin_lock(&cache->lock);
600                 if (ret == 1) {
601                         cache->caching_ctl = NULL;
602                         cache->cached = BTRFS_CACHE_FINISHED;
603                         cache->last_byte_to_unpin = (u64)-1;
604                 } else {
605                         if (load_cache_only) {
606                                 cache->caching_ctl = NULL;
607                                 cache->cached = BTRFS_CACHE_NO;
608                         } else {
609                                 cache->cached = BTRFS_CACHE_STARTED;
610                         }
611                 }
612                 spin_unlock(&cache->lock);
613                 wake_up(&caching_ctl->wait);
614                 if (ret == 1) {
615                         put_caching_control(caching_ctl);
616                         free_excluded_extents(fs_info->extent_root, cache);
617                         return 0;
618                 }
619         } else {
620                 /*
621                  * We are not going to do the fast caching, set cached to the
622                  * appropriate value and wakeup any waiters.
623                  */
624                 spin_lock(&cache->lock);
625                 if (load_cache_only) {
626                         cache->caching_ctl = NULL;
627                         cache->cached = BTRFS_CACHE_NO;
628                 } else {
629                         cache->cached = BTRFS_CACHE_STARTED;
630                 }
631                 spin_unlock(&cache->lock);
632                 wake_up(&caching_ctl->wait);
633         }
634
635         if (load_cache_only) {
636                 put_caching_control(caching_ctl);
637                 return 0;
638         }
639
640         down_write(&fs_info->commit_root_sem);
641         atomic_inc(&caching_ctl->count);
642         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
643         up_write(&fs_info->commit_root_sem);
644
645         btrfs_get_block_group(cache);
646
647         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
648
649         return ret;
650 }
651
652 /*
653  * return the block group that starts at or after bytenr
654  */
655 static struct btrfs_block_group_cache *
656 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
657 {
658         struct btrfs_block_group_cache *cache;
659
660         cache = block_group_cache_tree_search(info, bytenr, 0);
661
662         return cache;
663 }
664
665 /*
666  * return the block group that contains the given bytenr
667  */
668 struct btrfs_block_group_cache *btrfs_lookup_block_group(
669                                                  struct btrfs_fs_info *info,
670                                                  u64 bytenr)
671 {
672         struct btrfs_block_group_cache *cache;
673
674         cache = block_group_cache_tree_search(info, bytenr, 1);
675
676         return cache;
677 }
678
679 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
680                                                   u64 flags)
681 {
682         struct list_head *head = &info->space_info;
683         struct btrfs_space_info *found;
684
685         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
686
687         rcu_read_lock();
688         list_for_each_entry_rcu(found, head, list) {
689                 if (found->flags & flags) {
690                         rcu_read_unlock();
691                         return found;
692                 }
693         }
694         rcu_read_unlock();
695         return NULL;
696 }
697
698 /*
699  * after adding space to the filesystem, we need to clear the full flags
700  * on all the space infos.
701  */
702 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
703 {
704         struct list_head *head = &info->space_info;
705         struct btrfs_space_info *found;
706
707         rcu_read_lock();
708         list_for_each_entry_rcu(found, head, list)
709                 found->full = 0;
710         rcu_read_unlock();
711 }
712
713 /* simple helper to search for an existing data extent at a given offset */
714 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
715 {
716         int ret;
717         struct btrfs_key key;
718         struct btrfs_path *path;
719
720         path = btrfs_alloc_path();
721         if (!path)
722                 return -ENOMEM;
723
724         key.objectid = start;
725         key.offset = len;
726         key.type = BTRFS_EXTENT_ITEM_KEY;
727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
728                                 0, 0);
729         btrfs_free_path(path);
730         return ret;
731 }
732
733 /*
734  * helper function to lookup reference count and flags of a tree block.
735  *
736  * the head node for delayed ref is used to store the sum of all the
737  * reference count modifications queued up in the rbtree. the head
738  * node may also store the extent flags to set. This way you can check
739  * to see what the reference count and extent flags would be if all of
740  * the delayed refs are not processed.
741  */
742 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
743                              struct btrfs_root *root, u64 bytenr,
744                              u64 offset, int metadata, u64 *refs, u64 *flags)
745 {
746         struct btrfs_delayed_ref_head *head;
747         struct btrfs_delayed_ref_root *delayed_refs;
748         struct btrfs_path *path;
749         struct btrfs_extent_item *ei;
750         struct extent_buffer *leaf;
751         struct btrfs_key key;
752         u32 item_size;
753         u64 num_refs;
754         u64 extent_flags;
755         int ret;
756
757         /*
758          * If we don't have skinny metadata, don't bother doing anything
759          * different
760          */
761         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
762                 offset = root->nodesize;
763                 metadata = 0;
764         }
765
766         path = btrfs_alloc_path();
767         if (!path)
768                 return -ENOMEM;
769
770         if (!trans) {
771                 path->skip_locking = 1;
772                 path->search_commit_root = 1;
773         }
774
775 search_again:
776         key.objectid = bytenr;
777         key.offset = offset;
778         if (metadata)
779                 key.type = BTRFS_METADATA_ITEM_KEY;
780         else
781                 key.type = BTRFS_EXTENT_ITEM_KEY;
782
783         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
784                                 &key, path, 0, 0);
785         if (ret < 0)
786                 goto out_free;
787
788         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
789                 if (path->slots[0]) {
790                         path->slots[0]--;
791                         btrfs_item_key_to_cpu(path->nodes[0], &key,
792                                               path->slots[0]);
793                         if (key.objectid == bytenr &&
794                             key.type == BTRFS_EXTENT_ITEM_KEY &&
795                             key.offset == root->nodesize)
796                                 ret = 0;
797                 }
798         }
799
800         if (ret == 0) {
801                 leaf = path->nodes[0];
802                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
803                 if (item_size >= sizeof(*ei)) {
804                         ei = btrfs_item_ptr(leaf, path->slots[0],
805                                             struct btrfs_extent_item);
806                         num_refs = btrfs_extent_refs(leaf, ei);
807                         extent_flags = btrfs_extent_flags(leaf, ei);
808                 } else {
809 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
810                         struct btrfs_extent_item_v0 *ei0;
811                         BUG_ON(item_size != sizeof(*ei0));
812                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
813                                              struct btrfs_extent_item_v0);
814                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
815                         /* FIXME: this isn't correct for data */
816                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
817 #else
818                         BUG();
819 #endif
820                 }
821                 BUG_ON(num_refs == 0);
822         } else {
823                 num_refs = 0;
824                 extent_flags = 0;
825                 ret = 0;
826         }
827
828         if (!trans)
829                 goto out;
830
831         delayed_refs = &trans->transaction->delayed_refs;
832         spin_lock(&delayed_refs->lock);
833         head = btrfs_find_delayed_ref_head(trans, bytenr);
834         if (head) {
835                 if (!mutex_trylock(&head->mutex)) {
836                         atomic_inc(&head->node.refs);
837                         spin_unlock(&delayed_refs->lock);
838
839                         btrfs_release_path(path);
840
841                         /*
842                          * Mutex was contended, block until it's released and try
843                          * again
844                          */
845                         mutex_lock(&head->mutex);
846                         mutex_unlock(&head->mutex);
847                         btrfs_put_delayed_ref(&head->node);
848                         goto search_again;
849                 }
850                 spin_lock(&head->lock);
851                 if (head->extent_op && head->extent_op->update_flags)
852                         extent_flags |= head->extent_op->flags_to_set;
853                 else
854                         BUG_ON(num_refs == 0);
855
856                 num_refs += head->node.ref_mod;
857                 spin_unlock(&head->lock);
858                 mutex_unlock(&head->mutex);
859         }
860         spin_unlock(&delayed_refs->lock);
861 out:
862         WARN_ON(num_refs == 0);
863         if (refs)
864                 *refs = num_refs;
865         if (flags)
866                 *flags = extent_flags;
867 out_free:
868         btrfs_free_path(path);
869         return ret;
870 }
871
872 /*
873  * Back reference rules.  Back refs have three main goals:
874  *
875  * 1) differentiate between all holders of references to an extent so that
876  *    when a reference is dropped we can make sure it was a valid reference
877  *    before freeing the extent.
878  *
879  * 2) Provide enough information to quickly find the holders of an extent
880  *    if we notice a given block is corrupted or bad.
881  *
882  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
883  *    maintenance.  This is actually the same as #2, but with a slightly
884  *    different use case.
885  *
886  * There are two kinds of back refs. The implicit back refs is optimized
887  * for pointers in non-shared tree blocks. For a given pointer in a block,
888  * back refs of this kind provide information about the block's owner tree
889  * and the pointer's key. These information allow us to find the block by
890  * b-tree searching. The full back refs is for pointers in tree blocks not
891  * referenced by their owner trees. The location of tree block is recorded
892  * in the back refs. Actually the full back refs is generic, and can be
893  * used in all cases the implicit back refs is used. The major shortcoming
894  * of the full back refs is its overhead. Every time a tree block gets
895  * COWed, we have to update back refs entry for all pointers in it.
896  *
897  * For a newly allocated tree block, we use implicit back refs for
898  * pointers in it. This means most tree related operations only involve
899  * implicit back refs. For a tree block created in old transaction, the
900  * only way to drop a reference to it is COW it. So we can detect the
901  * event that tree block loses its owner tree's reference and do the
902  * back refs conversion.
903  *
904  * When a tree block is COW'd through a tree, there are four cases:
905  *
906  * The reference count of the block is one and the tree is the block's
907  * owner tree. Nothing to do in this case.
908  *
909  * The reference count of the block is one and the tree is not the
910  * block's owner tree. In this case, full back refs is used for pointers
911  * in the block. Remove these full back refs, add implicit back refs for
912  * every pointers in the new block.
913  *
914  * The reference count of the block is greater than one and the tree is
915  * the block's owner tree. In this case, implicit back refs is used for
916  * pointers in the block. Add full back refs for every pointers in the
917  * block, increase lower level extents' reference counts. The original
918  * implicit back refs are entailed to the new block.
919  *
920  * The reference count of the block is greater than one and the tree is
921  * not the block's owner tree. Add implicit back refs for every pointer in
922  * the new block, increase lower level extents' reference count.
923  *
924  * Back Reference Key composing:
925  *
926  * The key objectid corresponds to the first byte in the extent,
927  * The key type is used to differentiate between types of back refs.
928  * There are different meanings of the key offset for different types
929  * of back refs.
930  *
931  * File extents can be referenced by:
932  *
933  * - multiple snapshots, subvolumes, or different generations in one subvol
934  * - different files inside a single subvolume
935  * - different offsets inside a file (bookend extents in file.c)
936  *
937  * The extent ref structure for the implicit back refs has fields for:
938  *
939  * - Objectid of the subvolume root
940  * - objectid of the file holding the reference
941  * - original offset in the file
942  * - how many bookend extents
943  *
944  * The key offset for the implicit back refs is hash of the first
945  * three fields.
946  *
947  * The extent ref structure for the full back refs has field for:
948  *
949  * - number of pointers in the tree leaf
950  *
951  * The key offset for the implicit back refs is the first byte of
952  * the tree leaf
953  *
954  * When a file extent is allocated, The implicit back refs is used.
955  * the fields are filled in:
956  *
957  *     (root_key.objectid, inode objectid, offset in file, 1)
958  *
959  * When a file extent is removed file truncation, we find the
960  * corresponding implicit back refs and check the following fields:
961  *
962  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
963  *
964  * Btree extents can be referenced by:
965  *
966  * - Different subvolumes
967  *
968  * Both the implicit back refs and the full back refs for tree blocks
969  * only consist of key. The key offset for the implicit back refs is
970  * objectid of block's owner tree. The key offset for the full back refs
971  * is the first byte of parent block.
972  *
973  * When implicit back refs is used, information about the lowest key and
974  * level of the tree block are required. These information are stored in
975  * tree block info structure.
976  */
977
978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
979 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
980                                   struct btrfs_root *root,
981                                   struct btrfs_path *path,
982                                   u64 owner, u32 extra_size)
983 {
984         struct btrfs_extent_item *item;
985         struct btrfs_extent_item_v0 *ei0;
986         struct btrfs_extent_ref_v0 *ref0;
987         struct btrfs_tree_block_info *bi;
988         struct extent_buffer *leaf;
989         struct btrfs_key key;
990         struct btrfs_key found_key;
991         u32 new_size = sizeof(*item);
992         u64 refs;
993         int ret;
994
995         leaf = path->nodes[0];
996         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
997
998         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
999         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1000                              struct btrfs_extent_item_v0);
1001         refs = btrfs_extent_refs_v0(leaf, ei0);
1002
1003         if (owner == (u64)-1) {
1004                 while (1) {
1005                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1006                                 ret = btrfs_next_leaf(root, path);
1007                                 if (ret < 0)
1008                                         return ret;
1009                                 BUG_ON(ret > 0); /* Corruption */
1010                                 leaf = path->nodes[0];
1011                         }
1012                         btrfs_item_key_to_cpu(leaf, &found_key,
1013                                               path->slots[0]);
1014                         BUG_ON(key.objectid != found_key.objectid);
1015                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1016                                 path->slots[0]++;
1017                                 continue;
1018                         }
1019                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1020                                               struct btrfs_extent_ref_v0);
1021                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1022                         break;
1023                 }
1024         }
1025         btrfs_release_path(path);
1026
1027         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1028                 new_size += sizeof(*bi);
1029
1030         new_size -= sizeof(*ei0);
1031         ret = btrfs_search_slot(trans, root, &key, path,
1032                                 new_size + extra_size, 1);
1033         if (ret < 0)
1034                 return ret;
1035         BUG_ON(ret); /* Corruption */
1036
1037         btrfs_extend_item(root, path, new_size);
1038
1039         leaf = path->nodes[0];
1040         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1041         btrfs_set_extent_refs(leaf, item, refs);
1042         /* FIXME: get real generation */
1043         btrfs_set_extent_generation(leaf, item, 0);
1044         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1045                 btrfs_set_extent_flags(leaf, item,
1046                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1047                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1048                 bi = (struct btrfs_tree_block_info *)(item + 1);
1049                 /* FIXME: get first key of the block */
1050                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1051                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1052         } else {
1053                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1054         }
1055         btrfs_mark_buffer_dirty(leaf);
1056         return 0;
1057 }
1058 #endif
1059
1060 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         u32 high_crc = ~(u32)0;
1063         u32 low_crc = ~(u32)0;
1064         __le64 lenum;
1065
1066         lenum = cpu_to_le64(root_objectid);
1067         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1068         lenum = cpu_to_le64(owner);
1069         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1070         lenum = cpu_to_le64(offset);
1071         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1072
1073         return ((u64)high_crc << 31) ^ (u64)low_crc;
1074 }
1075
1076 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1077                                      struct btrfs_extent_data_ref *ref)
1078 {
1079         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1080                                     btrfs_extent_data_ref_objectid(leaf, ref),
1081                                     btrfs_extent_data_ref_offset(leaf, ref));
1082 }
1083
1084 static int match_extent_data_ref(struct extent_buffer *leaf,
1085                                  struct btrfs_extent_data_ref *ref,
1086                                  u64 root_objectid, u64 owner, u64 offset)
1087 {
1088         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1089             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1090             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1091                 return 0;
1092         return 1;
1093 }
1094
1095 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid,
1100                                            u64 owner, u64 offset)
1101 {
1102         struct btrfs_key key;
1103         struct btrfs_extent_data_ref *ref;
1104         struct extent_buffer *leaf;
1105         u32 nritems;
1106         int ret;
1107         int recow;
1108         int err = -ENOENT;
1109
1110         key.objectid = bytenr;
1111         if (parent) {
1112                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1113                 key.offset = parent;
1114         } else {
1115                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1116                 key.offset = hash_extent_data_ref(root_objectid,
1117                                                   owner, offset);
1118         }
1119 again:
1120         recow = 0;
1121         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122         if (ret < 0) {
1123                 err = ret;
1124                 goto fail;
1125         }
1126
1127         if (parent) {
1128                 if (!ret)
1129                         return 0;
1130 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1131                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1132                 btrfs_release_path(path);
1133                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134                 if (ret < 0) {
1135                         err = ret;
1136                         goto fail;
1137                 }
1138                 if (!ret)
1139                         return 0;
1140 #endif
1141                 goto fail;
1142         }
1143
1144         leaf = path->nodes[0];
1145         nritems = btrfs_header_nritems(leaf);
1146         while (1) {
1147                 if (path->slots[0] >= nritems) {
1148                         ret = btrfs_next_leaf(root, path);
1149                         if (ret < 0)
1150                                 err = ret;
1151                         if (ret)
1152                                 goto fail;
1153
1154                         leaf = path->nodes[0];
1155                         nritems = btrfs_header_nritems(leaf);
1156                         recow = 1;
1157                 }
1158
1159                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1160                 if (key.objectid != bytenr ||
1161                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1162                         goto fail;
1163
1164                 ref = btrfs_item_ptr(leaf, path->slots[0],
1165                                      struct btrfs_extent_data_ref);
1166
1167                 if (match_extent_data_ref(leaf, ref, root_objectid,
1168                                           owner, offset)) {
1169                         if (recow) {
1170                                 btrfs_release_path(path);
1171                                 goto again;
1172                         }
1173                         err = 0;
1174                         break;
1175                 }
1176                 path->slots[0]++;
1177         }
1178 fail:
1179         return err;
1180 }
1181
1182 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1183                                            struct btrfs_root *root,
1184                                            struct btrfs_path *path,
1185                                            u64 bytenr, u64 parent,
1186                                            u64 root_objectid, u64 owner,
1187                                            u64 offset, int refs_to_add)
1188 {
1189         struct btrfs_key key;
1190         struct extent_buffer *leaf;
1191         u32 size;
1192         u32 num_refs;
1193         int ret;
1194
1195         key.objectid = bytenr;
1196         if (parent) {
1197                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1198                 key.offset = parent;
1199                 size = sizeof(struct btrfs_shared_data_ref);
1200         } else {
1201                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1202                 key.offset = hash_extent_data_ref(root_objectid,
1203                                                   owner, offset);
1204                 size = sizeof(struct btrfs_extent_data_ref);
1205         }
1206
1207         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1208         if (ret && ret != -EEXIST)
1209                 goto fail;
1210
1211         leaf = path->nodes[0];
1212         if (parent) {
1213                 struct btrfs_shared_data_ref *ref;
1214                 ref = btrfs_item_ptr(leaf, path->slots[0],
1215                                      struct btrfs_shared_data_ref);
1216                 if (ret == 0) {
1217                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1218                 } else {
1219                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1220                         num_refs += refs_to_add;
1221                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1222                 }
1223         } else {
1224                 struct btrfs_extent_data_ref *ref;
1225                 while (ret == -EEXIST) {
1226                         ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                              struct btrfs_extent_data_ref);
1228                         if (match_extent_data_ref(leaf, ref, root_objectid,
1229                                                   owner, offset))
1230                                 break;
1231                         btrfs_release_path(path);
1232                         key.offset++;
1233                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1234                                                       size);
1235                         if (ret && ret != -EEXIST)
1236                                 goto fail;
1237
1238                         leaf = path->nodes[0];
1239                 }
1240                 ref = btrfs_item_ptr(leaf, path->slots[0],
1241                                      struct btrfs_extent_data_ref);
1242                 if (ret == 0) {
1243                         btrfs_set_extent_data_ref_root(leaf, ref,
1244                                                        root_objectid);
1245                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1246                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1247                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1248                 } else {
1249                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1250                         num_refs += refs_to_add;
1251                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1252                 }
1253         }
1254         btrfs_mark_buffer_dirty(leaf);
1255         ret = 0;
1256 fail:
1257         btrfs_release_path(path);
1258         return ret;
1259 }
1260
1261 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            int refs_to_drop, int *last_ref)
1265 {
1266         struct btrfs_key key;
1267         struct btrfs_extent_data_ref *ref1 = NULL;
1268         struct btrfs_shared_data_ref *ref2 = NULL;
1269         struct extent_buffer *leaf;
1270         u32 num_refs = 0;
1271         int ret = 0;
1272
1273         leaf = path->nodes[0];
1274         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1275
1276         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1277                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_extent_data_ref);
1279                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1280         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1281                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1282                                       struct btrfs_shared_data_ref);
1283                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1284 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1285         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1286                 struct btrfs_extent_ref_v0 *ref0;
1287                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_ref_v0);
1289                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1290 #endif
1291         } else {
1292                 BUG();
1293         }
1294
1295         BUG_ON(num_refs < refs_to_drop);
1296         num_refs -= refs_to_drop;
1297
1298         if (num_refs == 0) {
1299                 ret = btrfs_del_item(trans, root, path);
1300                 *last_ref = 1;
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1320                                           struct btrfs_path *path,
1321                                           struct btrfs_extent_inline_ref *iref)
1322 {
1323         struct btrfs_key key;
1324         struct extent_buffer *leaf;
1325         struct btrfs_extent_data_ref *ref1;
1326         struct btrfs_shared_data_ref *ref2;
1327         u32 num_refs = 0;
1328
1329         leaf = path->nodes[0];
1330         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1331         if (iref) {
1332                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1333                     BTRFS_EXTENT_DATA_REF_KEY) {
1334                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1335                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1336                 } else {
1337                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1338                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1339                 }
1340         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 WARN_ON(1);
1357         }
1358         return num_refs;
1359 }
1360
1361 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1362                                           struct btrfs_root *root,
1363                                           struct btrfs_path *path,
1364                                           u64 bytenr, u64 parent,
1365                                           u64 root_objectid)
1366 {
1367         struct btrfs_key key;
1368         int ret;
1369
1370         key.objectid = bytenr;
1371         if (parent) {
1372                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1373                 key.offset = parent;
1374         } else {
1375                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1376                 key.offset = root_objectid;
1377         }
1378
1379         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1380         if (ret > 0)
1381                 ret = -ENOENT;
1382 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1383         if (ret == -ENOENT && parent) {
1384                 btrfs_release_path(path);
1385                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1386                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1387                 if (ret > 0)
1388                         ret = -ENOENT;
1389         }
1390 #endif
1391         return ret;
1392 }
1393
1394 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1395                                           struct btrfs_root *root,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_key key;
1401         int ret;
1402
1403         key.objectid = bytenr;
1404         if (parent) {
1405                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1406                 key.offset = parent;
1407         } else {
1408                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1409                 key.offset = root_objectid;
1410         }
1411
1412         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1413         btrfs_release_path(path);
1414         return ret;
1415 }
1416
1417 static inline int extent_ref_type(u64 parent, u64 owner)
1418 {
1419         int type;
1420         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1421                 if (parent > 0)
1422                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1423                 else
1424                         type = BTRFS_TREE_BLOCK_REF_KEY;
1425         } else {
1426                 if (parent > 0)
1427                         type = BTRFS_SHARED_DATA_REF_KEY;
1428                 else
1429                         type = BTRFS_EXTENT_DATA_REF_KEY;
1430         }
1431         return type;
1432 }
1433
1434 static int find_next_key(struct btrfs_path *path, int level,
1435                          struct btrfs_key *key)
1436
1437 {
1438         for (; level < BTRFS_MAX_LEVEL; level++) {
1439                 if (!path->nodes[level])
1440                         break;
1441                 if (path->slots[level] + 1 >=
1442                     btrfs_header_nritems(path->nodes[level]))
1443                         continue;
1444                 if (level == 0)
1445                         btrfs_item_key_to_cpu(path->nodes[level], key,
1446                                               path->slots[level] + 1);
1447                 else
1448                         btrfs_node_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 return 0;
1451         }
1452         return 1;
1453 }
1454
1455 /*
1456  * look for inline back ref. if back ref is found, *ref_ret is set
1457  * to the address of inline back ref, and 0 is returned.
1458  *
1459  * if back ref isn't found, *ref_ret is set to the address where it
1460  * should be inserted, and -ENOENT is returned.
1461  *
1462  * if insert is true and there are too many inline back refs, the path
1463  * points to the extent item, and -EAGAIN is returned.
1464  *
1465  * NOTE: inline back refs are ordered in the same way that back ref
1466  *       items in the tree are ordered.
1467  */
1468 static noinline_for_stack
1469 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1470                                  struct btrfs_root *root,
1471                                  struct btrfs_path *path,
1472                                  struct btrfs_extent_inline_ref **ref_ret,
1473                                  u64 bytenr, u64 num_bytes,
1474                                  u64 parent, u64 root_objectid,
1475                                  u64 owner, u64 offset, int insert)
1476 {
1477         struct btrfs_key key;
1478         struct extent_buffer *leaf;
1479         struct btrfs_extent_item *ei;
1480         struct btrfs_extent_inline_ref *iref;
1481         u64 flags;
1482         u64 item_size;
1483         unsigned long ptr;
1484         unsigned long end;
1485         int extra_size;
1486         int type;
1487         int want;
1488         int ret;
1489         int err = 0;
1490         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1491                                                  SKINNY_METADATA);
1492
1493         key.objectid = bytenr;
1494         key.type = BTRFS_EXTENT_ITEM_KEY;
1495         key.offset = num_bytes;
1496
1497         want = extent_ref_type(parent, owner);
1498         if (insert) {
1499                 extra_size = btrfs_extent_inline_ref_size(want);
1500                 path->keep_locks = 1;
1501         } else
1502                 extra_size = -1;
1503
1504         /*
1505          * Owner is our parent level, so we can just add one to get the level
1506          * for the block we are interested in.
1507          */
1508         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 key.type = BTRFS_METADATA_ITEM_KEY;
1510                 key.offset = owner;
1511         }
1512
1513 again:
1514         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1515         if (ret < 0) {
1516                 err = ret;
1517                 goto out;
1518         }
1519
1520         /*
1521          * We may be a newly converted file system which still has the old fat
1522          * extent entries for metadata, so try and see if we have one of those.
1523          */
1524         if (ret > 0 && skinny_metadata) {
1525                 skinny_metadata = false;
1526                 if (path->slots[0]) {
1527                         path->slots[0]--;
1528                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1529                                               path->slots[0]);
1530                         if (key.objectid == bytenr &&
1531                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1532                             key.offset == num_bytes)
1533                                 ret = 0;
1534                 }
1535                 if (ret) {
1536                         key.objectid = bytenr;
1537                         key.type = BTRFS_EXTENT_ITEM_KEY;
1538                         key.offset = num_bytes;
1539                         btrfs_release_path(path);
1540                         goto again;
1541                 }
1542         }
1543
1544         if (ret && !insert) {
1545                 err = -ENOENT;
1546                 goto out;
1547         } else if (WARN_ON(ret)) {
1548                 err = -EIO;
1549                 goto out;
1550         }
1551
1552         leaf = path->nodes[0];
1553         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1555         if (item_size < sizeof(*ei)) {
1556                 if (!insert) {
1557                         err = -ENOENT;
1558                         goto out;
1559                 }
1560                 ret = convert_extent_item_v0(trans, root, path, owner,
1561                                              extra_size);
1562                 if (ret < 0) {
1563                         err = ret;
1564                         goto out;
1565                 }
1566                 leaf = path->nodes[0];
1567                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1568         }
1569 #endif
1570         BUG_ON(item_size < sizeof(*ei));
1571
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         flags = btrfs_extent_flags(leaf, ei);
1574
1575         ptr = (unsigned long)(ei + 1);
1576         end = (unsigned long)ei + item_size;
1577
1578         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1579                 ptr += sizeof(struct btrfs_tree_block_info);
1580                 BUG_ON(ptr > end);
1581         }
1582
1583         err = -ENOENT;
1584         while (1) {
1585                 if (ptr >= end) {
1586                         WARN_ON(ptr > end);
1587                         break;
1588                 }
1589                 iref = (struct btrfs_extent_inline_ref *)ptr;
1590                 type = btrfs_extent_inline_ref_type(leaf, iref);
1591                 if (want < type)
1592                         break;
1593                 if (want > type) {
1594                         ptr += btrfs_extent_inline_ref_size(type);
1595                         continue;
1596                 }
1597
1598                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1599                         struct btrfs_extent_data_ref *dref;
1600                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1601                         if (match_extent_data_ref(leaf, dref, root_objectid,
1602                                                   owner, offset)) {
1603                                 err = 0;
1604                                 break;
1605                         }
1606                         if (hash_extent_data_ref_item(leaf, dref) <
1607                             hash_extent_data_ref(root_objectid, owner, offset))
1608                                 break;
1609                 } else {
1610                         u64 ref_offset;
1611                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1612                         if (parent > 0) {
1613                                 if (parent == ref_offset) {
1614                                         err = 0;
1615                                         break;
1616                                 }
1617                                 if (ref_offset < parent)
1618                                         break;
1619                         } else {
1620                                 if (root_objectid == ref_offset) {
1621                                         err = 0;
1622                                         break;
1623                                 }
1624                                 if (ref_offset < root_objectid)
1625                                         break;
1626                         }
1627                 }
1628                 ptr += btrfs_extent_inline_ref_size(type);
1629         }
1630         if (err == -ENOENT && insert) {
1631                 if (item_size + extra_size >=
1632                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1633                         err = -EAGAIN;
1634                         goto out;
1635                 }
1636                 /*
1637                  * To add new inline back ref, we have to make sure
1638                  * there is no corresponding back ref item.
1639                  * For simplicity, we just do not add new inline back
1640                  * ref if there is any kind of item for this block
1641                  */
1642                 if (find_next_key(path, 0, &key) == 0 &&
1643                     key.objectid == bytenr &&
1644                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648         }
1649         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1650 out:
1651         if (insert) {
1652                 path->keep_locks = 0;
1653                 btrfs_unlock_up_safe(path, 1);
1654         }
1655         return err;
1656 }
1657
1658 /*
1659  * helper to add new inline back ref
1660  */
1661 static noinline_for_stack
1662 void setup_inline_extent_backref(struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref *iref,
1665                                  u64 parent, u64 root_objectid,
1666                                  u64 owner, u64 offset, int refs_to_add,
1667                                  struct btrfs_delayed_extent_op *extent_op)
1668 {
1669         struct extent_buffer *leaf;
1670         struct btrfs_extent_item *ei;
1671         unsigned long ptr;
1672         unsigned long end;
1673         unsigned long item_offset;
1674         u64 refs;
1675         int size;
1676         int type;
1677
1678         leaf = path->nodes[0];
1679         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1680         item_offset = (unsigned long)iref - (unsigned long)ei;
1681
1682         type = extent_ref_type(parent, owner);
1683         size = btrfs_extent_inline_ref_size(type);
1684
1685         btrfs_extend_item(root, path, size);
1686
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         refs = btrfs_extent_refs(leaf, ei);
1689         refs += refs_to_add;
1690         btrfs_set_extent_refs(leaf, ei, refs);
1691         if (extent_op)
1692                 __run_delayed_extent_op(extent_op, leaf, ei);
1693
1694         ptr = (unsigned long)ei + item_offset;
1695         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1696         if (ptr < end - size)
1697                 memmove_extent_buffer(leaf, ptr + size, ptr,
1698                                       end - size - ptr);
1699
1700         iref = (struct btrfs_extent_inline_ref *)ptr;
1701         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1702         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1703                 struct btrfs_extent_data_ref *dref;
1704                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1705                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1706                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1707                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1708                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 struct btrfs_shared_data_ref *sref;
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1713                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1714         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1715                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1716         } else {
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1718         }
1719         btrfs_mark_buffer_dirty(leaf);
1720 }
1721
1722 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1723                                  struct btrfs_root *root,
1724                                  struct btrfs_path *path,
1725                                  struct btrfs_extent_inline_ref **ref_ret,
1726                                  u64 bytenr, u64 num_bytes, u64 parent,
1727                                  u64 root_objectid, u64 owner, u64 offset)
1728 {
1729         int ret;
1730
1731         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1732                                            bytenr, num_bytes, parent,
1733                                            root_objectid, owner, offset, 0);
1734         if (ret != -ENOENT)
1735                 return ret;
1736
1737         btrfs_release_path(path);
1738         *ref_ret = NULL;
1739
1740         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1741                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1742                                             root_objectid);
1743         } else {
1744                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1745                                              root_objectid, owner, offset);
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * helper to update/remove inline back ref
1752  */
1753 static noinline_for_stack
1754 void update_inline_extent_backref(struct btrfs_root *root,
1755                                   struct btrfs_path *path,
1756                                   struct btrfs_extent_inline_ref *iref,
1757                                   int refs_to_mod,
1758                                   struct btrfs_delayed_extent_op *extent_op,
1759                                   int *last_ref)
1760 {
1761         struct extent_buffer *leaf;
1762         struct btrfs_extent_item *ei;
1763         struct btrfs_extent_data_ref *dref = NULL;
1764         struct btrfs_shared_data_ref *sref = NULL;
1765         unsigned long ptr;
1766         unsigned long end;
1767         u32 item_size;
1768         int size;
1769         int type;
1770         u64 refs;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         refs = btrfs_extent_refs(leaf, ei);
1775         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1776         refs += refs_to_mod;
1777         btrfs_set_extent_refs(leaf, ei, refs);
1778         if (extent_op)
1779                 __run_delayed_extent_op(extent_op, leaf, ei);
1780
1781         type = btrfs_extent_inline_ref_type(leaf, iref);
1782
1783         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1784                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1785                 refs = btrfs_extent_data_ref_count(leaf, dref);
1786         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1787                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1788                 refs = btrfs_shared_data_ref_count(leaf, sref);
1789         } else {
1790                 refs = 1;
1791                 BUG_ON(refs_to_mod != -1);
1792         }
1793
1794         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1795         refs += refs_to_mod;
1796
1797         if (refs > 0) {
1798                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1799                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1800                 else
1801                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1802         } else {
1803                 *last_ref = 1;
1804                 size =  btrfs_extent_inline_ref_size(type);
1805                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1806                 ptr = (unsigned long)iref;
1807                 end = (unsigned long)ei + item_size;
1808                 if (ptr + size < end)
1809                         memmove_extent_buffer(leaf, ptr, ptr + size,
1810                                               end - ptr - size);
1811                 item_size -= size;
1812                 btrfs_truncate_item(root, path, item_size, 1);
1813         }
1814         btrfs_mark_buffer_dirty(leaf);
1815 }
1816
1817 static noinline_for_stack
1818 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1819                                  struct btrfs_root *root,
1820                                  struct btrfs_path *path,
1821                                  u64 bytenr, u64 num_bytes, u64 parent,
1822                                  u64 root_objectid, u64 owner,
1823                                  u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct btrfs_extent_inline_ref *iref;
1827         int ret;
1828
1829         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1830                                            bytenr, num_bytes, parent,
1831                                            root_objectid, owner, offset, 1);
1832         if (ret == 0) {
1833                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1834                 update_inline_extent_backref(root, path, iref,
1835                                              refs_to_add, extent_op, NULL);
1836         } else if (ret == -ENOENT) {
1837                 setup_inline_extent_backref(root, path, iref, parent,
1838                                             root_objectid, owner, offset,
1839                                             refs_to_add, extent_op);
1840                 ret = 0;
1841         }
1842         return ret;
1843 }
1844
1845 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1846                                  struct btrfs_root *root,
1847                                  struct btrfs_path *path,
1848                                  u64 bytenr, u64 parent, u64 root_objectid,
1849                                  u64 owner, u64 offset, int refs_to_add)
1850 {
1851         int ret;
1852         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1853                 BUG_ON(refs_to_add != 1);
1854                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1855                                             parent, root_objectid);
1856         } else {
1857                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1858                                              parent, root_objectid,
1859                                              owner, offset, refs_to_add);
1860         }
1861         return ret;
1862 }
1863
1864 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1865                                  struct btrfs_root *root,
1866                                  struct btrfs_path *path,
1867                                  struct btrfs_extent_inline_ref *iref,
1868                                  int refs_to_drop, int is_data, int *last_ref)
1869 {
1870         int ret = 0;
1871
1872         BUG_ON(!is_data && refs_to_drop != 1);
1873         if (iref) {
1874                 update_inline_extent_backref(root, path, iref,
1875                                              -refs_to_drop, NULL, last_ref);
1876         } else if (is_data) {
1877                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1878                                              last_ref);
1879         } else {
1880                 *last_ref = 1;
1881                 ret = btrfs_del_item(trans, root, path);
1882         }
1883         return ret;
1884 }
1885
1886 static int btrfs_issue_discard(struct block_device *bdev,
1887                                 u64 start, u64 len)
1888 {
1889         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1890 }
1891
1892 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1893                                 u64 num_bytes, u64 *actual_bytes)
1894 {
1895         int ret;
1896         u64 discarded_bytes = 0;
1897         struct btrfs_bio *bbio = NULL;
1898
1899
1900         /* Tell the block device(s) that the sectors can be discarded */
1901         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1902                               bytenr, &num_bytes, &bbio, 0);
1903         /* Error condition is -ENOMEM */
1904         if (!ret) {
1905                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1906                 int i;
1907
1908
1909                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1910                         if (!stripe->dev->can_discard)
1911                                 continue;
1912
1913                         ret = btrfs_issue_discard(stripe->dev->bdev,
1914                                                   stripe->physical,
1915                                                   stripe->length);
1916                         if (!ret)
1917                                 discarded_bytes += stripe->length;
1918                         else if (ret != -EOPNOTSUPP)
1919                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1920
1921                         /*
1922                          * Just in case we get back EOPNOTSUPP for some reason,
1923                          * just ignore the return value so we don't screw up
1924                          * people calling discard_extent.
1925                          */
1926                         ret = 0;
1927                 }
1928                 kfree(bbio);
1929         }
1930
1931         if (actual_bytes)
1932                 *actual_bytes = discarded_bytes;
1933
1934
1935         if (ret == -EOPNOTSUPP)
1936                 ret = 0;
1937         return ret;
1938 }
1939
1940 /* Can return -ENOMEM */
1941 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1942                          struct btrfs_root *root,
1943                          u64 bytenr, u64 num_bytes, u64 parent,
1944                          u64 root_objectid, u64 owner, u64 offset,
1945                          int no_quota)
1946 {
1947         int ret;
1948         struct btrfs_fs_info *fs_info = root->fs_info;
1949
1950         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1951                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1952
1953         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1954                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1955                                         num_bytes,
1956                                         parent, root_objectid, (int)owner,
1957                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1958         } else {
1959                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1960                                         num_bytes,
1961                                         parent, root_objectid, owner, offset,
1962                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1963         }
1964         return ret;
1965 }
1966
1967 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1968                                   struct btrfs_root *root,
1969                                   u64 bytenr, u64 num_bytes,
1970                                   u64 parent, u64 root_objectid,
1971                                   u64 owner, u64 offset, int refs_to_add,
1972                                   int no_quota,
1973                                   struct btrfs_delayed_extent_op *extent_op)
1974 {
1975         struct btrfs_fs_info *fs_info = root->fs_info;
1976         struct btrfs_path *path;
1977         struct extent_buffer *leaf;
1978         struct btrfs_extent_item *item;
1979         struct btrfs_key key;
1980         u64 refs;
1981         int ret;
1982         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1983
1984         path = btrfs_alloc_path();
1985         if (!path)
1986                 return -ENOMEM;
1987
1988         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
1989                 no_quota = 1;
1990
1991         path->reada = 1;
1992         path->leave_spinning = 1;
1993         /* this will setup the path even if it fails to insert the back ref */
1994         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
1995                                            bytenr, num_bytes, parent,
1996                                            root_objectid, owner, offset,
1997                                            refs_to_add, extent_op);
1998         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
1999                 goto out;
2000         /*
2001          * Ok we were able to insert an inline extent and it appears to be a new
2002          * reference, deal with the qgroup accounting.
2003          */
2004         if (!ret && !no_quota) {
2005                 ASSERT(root->fs_info->quota_enabled);
2006                 leaf = path->nodes[0];
2007                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2008                 item = btrfs_item_ptr(leaf, path->slots[0],
2009                                       struct btrfs_extent_item);
2010                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2011                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2012                 btrfs_release_path(path);
2013
2014                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2015                                               bytenr, num_bytes, type, 0);
2016                 goto out;
2017         }
2018
2019         /*
2020          * Ok we had -EAGAIN which means we didn't have space to insert and
2021          * inline extent ref, so just update the reference count and add a
2022          * normal backref.
2023          */
2024         leaf = path->nodes[0];
2025         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2026         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2027         refs = btrfs_extent_refs(leaf, item);
2028         if (refs)
2029                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2030         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2031         if (extent_op)
2032                 __run_delayed_extent_op(extent_op, leaf, item);
2033
2034         btrfs_mark_buffer_dirty(leaf);
2035         btrfs_release_path(path);
2036
2037         if (!no_quota) {
2038                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2039                                               bytenr, num_bytes, type, 0);
2040                 if (ret)
2041                         goto out;
2042         }
2043
2044         path->reada = 1;
2045         path->leave_spinning = 1;
2046         /* now insert the actual backref */
2047         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2048                                     path, bytenr, parent, root_objectid,
2049                                     owner, offset, refs_to_add);
2050         if (ret)
2051                 btrfs_abort_transaction(trans, root, ret);
2052 out:
2053         btrfs_free_path(path);
2054         return ret;
2055 }
2056
2057 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2058                                 struct btrfs_root *root,
2059                                 struct btrfs_delayed_ref_node *node,
2060                                 struct btrfs_delayed_extent_op *extent_op,
2061                                 int insert_reserved)
2062 {
2063         int ret = 0;
2064         struct btrfs_delayed_data_ref *ref;
2065         struct btrfs_key ins;
2066         u64 parent = 0;
2067         u64 ref_root = 0;
2068         u64 flags = 0;
2069
2070         ins.objectid = node->bytenr;
2071         ins.offset = node->num_bytes;
2072         ins.type = BTRFS_EXTENT_ITEM_KEY;
2073
2074         ref = btrfs_delayed_node_to_data_ref(node);
2075         trace_run_delayed_data_ref(node, ref, node->action);
2076
2077         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2078                 parent = ref->parent;
2079         ref_root = ref->root;
2080
2081         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2082                 if (extent_op)
2083                         flags |= extent_op->flags_to_set;
2084                 ret = alloc_reserved_file_extent(trans, root,
2085                                                  parent, ref_root, flags,
2086                                                  ref->objectid, ref->offset,
2087                                                  &ins, node->ref_mod);
2088         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2089                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2090                                              node->num_bytes, parent,
2091                                              ref_root, ref->objectid,
2092                                              ref->offset, node->ref_mod,
2093                                              node->no_quota, extent_op);
2094         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2095                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2096                                           node->num_bytes, parent,
2097                                           ref_root, ref->objectid,
2098                                           ref->offset, node->ref_mod,
2099                                           extent_op, node->no_quota);
2100         } else {
2101                 BUG();
2102         }
2103         return ret;
2104 }
2105
2106 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2107                                     struct extent_buffer *leaf,
2108                                     struct btrfs_extent_item *ei)
2109 {
2110         u64 flags = btrfs_extent_flags(leaf, ei);
2111         if (extent_op->update_flags) {
2112                 flags |= extent_op->flags_to_set;
2113                 btrfs_set_extent_flags(leaf, ei, flags);
2114         }
2115
2116         if (extent_op->update_key) {
2117                 struct btrfs_tree_block_info *bi;
2118                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2119                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2120                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2121         }
2122 }
2123
2124 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2125                                  struct btrfs_root *root,
2126                                  struct btrfs_delayed_ref_node *node,
2127                                  struct btrfs_delayed_extent_op *extent_op)
2128 {
2129         struct btrfs_key key;
2130         struct btrfs_path *path;
2131         struct btrfs_extent_item *ei;
2132         struct extent_buffer *leaf;
2133         u32 item_size;
2134         int ret;
2135         int err = 0;
2136         int metadata = !extent_op->is_data;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2142                 metadata = 0;
2143
2144         path = btrfs_alloc_path();
2145         if (!path)
2146                 return -ENOMEM;
2147
2148         key.objectid = node->bytenr;
2149
2150         if (metadata) {
2151                 key.type = BTRFS_METADATA_ITEM_KEY;
2152                 key.offset = extent_op->level;
2153         } else {
2154                 key.type = BTRFS_EXTENT_ITEM_KEY;
2155                 key.offset = node->num_bytes;
2156         }
2157
2158 again:
2159         path->reada = 1;
2160         path->leave_spinning = 1;
2161         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2162                                 path, 0, 1);
2163         if (ret < 0) {
2164                 err = ret;
2165                 goto out;
2166         }
2167         if (ret > 0) {
2168                 if (metadata) {
2169                         if (path->slots[0] > 0) {
2170                                 path->slots[0]--;
2171                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2172                                                       path->slots[0]);
2173                                 if (key.objectid == node->bytenr &&
2174                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2175                                     key.offset == node->num_bytes)
2176                                         ret = 0;
2177                         }
2178                         if (ret > 0) {
2179                                 btrfs_release_path(path);
2180                                 metadata = 0;
2181
2182                                 key.objectid = node->bytenr;
2183                                 key.offset = node->num_bytes;
2184                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2185                                 goto again;
2186                         }
2187                 } else {
2188                         err = -EIO;
2189                         goto out;
2190                 }
2191         }
2192
2193         leaf = path->nodes[0];
2194         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2195 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2196         if (item_size < sizeof(*ei)) {
2197                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2198                                              path, (u64)-1, 0);
2199                 if (ret < 0) {
2200                         err = ret;
2201                         goto out;
2202                 }
2203                 leaf = path->nodes[0];
2204                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2205         }
2206 #endif
2207         BUG_ON(item_size < sizeof(*ei));
2208         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2209         __run_delayed_extent_op(extent_op, leaf, ei);
2210
2211         btrfs_mark_buffer_dirty(leaf);
2212 out:
2213         btrfs_free_path(path);
2214         return err;
2215 }
2216
2217 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2218                                 struct btrfs_root *root,
2219                                 struct btrfs_delayed_ref_node *node,
2220                                 struct btrfs_delayed_extent_op *extent_op,
2221                                 int insert_reserved)
2222 {
2223         int ret = 0;
2224         struct btrfs_delayed_tree_ref *ref;
2225         struct btrfs_key ins;
2226         u64 parent = 0;
2227         u64 ref_root = 0;
2228         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2229                                                  SKINNY_METADATA);
2230
2231         ref = btrfs_delayed_node_to_tree_ref(node);
2232         trace_run_delayed_tree_ref(node, ref, node->action);
2233
2234         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2235                 parent = ref->parent;
2236         ref_root = ref->root;
2237
2238         ins.objectid = node->bytenr;
2239         if (skinny_metadata) {
2240                 ins.offset = ref->level;
2241                 ins.type = BTRFS_METADATA_ITEM_KEY;
2242         } else {
2243                 ins.offset = node->num_bytes;
2244                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2245         }
2246
2247         BUG_ON(node->ref_mod != 1);
2248         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2249                 BUG_ON(!extent_op || !extent_op->update_flags);
2250                 ret = alloc_reserved_tree_block(trans, root,
2251                                                 parent, ref_root,
2252                                                 extent_op->flags_to_set,
2253                                                 &extent_op->key,
2254                                                 ref->level, &ins,
2255                                                 node->no_quota);
2256         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2257                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2258                                              node->num_bytes, parent, ref_root,
2259                                              ref->level, 0, 1, node->no_quota,
2260                                              extent_op);
2261         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2262                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2263                                           node->num_bytes, parent, ref_root,
2264                                           ref->level, 0, 1, extent_op,
2265                                           node->no_quota);
2266         } else {
2267                 BUG();
2268         }
2269         return ret;
2270 }
2271
2272 /* helper function to actually process a single delayed ref entry */
2273 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2274                                struct btrfs_root *root,
2275                                struct btrfs_delayed_ref_node *node,
2276                                struct btrfs_delayed_extent_op *extent_op,
2277                                int insert_reserved)
2278 {
2279         int ret = 0;
2280
2281         if (trans->aborted) {
2282                 if (insert_reserved)
2283                         btrfs_pin_extent(root, node->bytenr,
2284                                          node->num_bytes, 1);
2285                 return 0;
2286         }
2287
2288         if (btrfs_delayed_ref_is_head(node)) {
2289                 struct btrfs_delayed_ref_head *head;
2290                 /*
2291                  * we've hit the end of the chain and we were supposed
2292                  * to insert this extent into the tree.  But, it got
2293                  * deleted before we ever needed to insert it, so all
2294                  * we have to do is clean up the accounting
2295                  */
2296                 BUG_ON(extent_op);
2297                 head = btrfs_delayed_node_to_head(node);
2298                 trace_run_delayed_ref_head(node, head, node->action);
2299
2300                 if (insert_reserved) {
2301                         btrfs_pin_extent(root, node->bytenr,
2302                                          node->num_bytes, 1);
2303                         if (head->is_data) {
2304                                 ret = btrfs_del_csums(trans, root,
2305                                                       node->bytenr,
2306                                                       node->num_bytes);
2307                         }
2308                 }
2309                 return ret;
2310         }
2311
2312         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2313             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2314                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2315                                            insert_reserved);
2316         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2317                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2318                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2319                                            insert_reserved);
2320         else
2321                 BUG();
2322         return ret;
2323 }
2324
2325 static noinline struct btrfs_delayed_ref_node *
2326 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2327 {
2328         struct rb_node *node;
2329         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2330
2331         /*
2332          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2333          * this prevents ref count from going down to zero when
2334          * there still are pending delayed ref.
2335          */
2336         node = rb_first(&head->ref_root);
2337         while (node) {
2338                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2339                                 rb_node);
2340                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2341                         return ref;
2342                 else if (last == NULL)
2343                         last = ref;
2344                 node = rb_next(node);
2345         }
2346         return last;
2347 }
2348
2349 /*
2350  * Returns 0 on success or if called with an already aborted transaction.
2351  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2352  */
2353 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2354                                              struct btrfs_root *root,
2355                                              unsigned long nr)
2356 {
2357         struct btrfs_delayed_ref_root *delayed_refs;
2358         struct btrfs_delayed_ref_node *ref;
2359         struct btrfs_delayed_ref_head *locked_ref = NULL;
2360         struct btrfs_delayed_extent_op *extent_op;
2361         struct btrfs_fs_info *fs_info = root->fs_info;
2362         ktime_t start = ktime_get();
2363         int ret;
2364         unsigned long count = 0;
2365         unsigned long actual_count = 0;
2366         int must_insert_reserved = 0;
2367
2368         delayed_refs = &trans->transaction->delayed_refs;
2369         while (1) {
2370                 if (!locked_ref) {
2371                         if (count >= nr)
2372                                 break;
2373
2374                         spin_lock(&delayed_refs->lock);
2375                         locked_ref = btrfs_select_ref_head(trans);
2376                         if (!locked_ref) {
2377                                 spin_unlock(&delayed_refs->lock);
2378                                 break;
2379                         }
2380
2381                         /* grab the lock that says we are going to process
2382                          * all the refs for this head */
2383                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2384                         spin_unlock(&delayed_refs->lock);
2385                         /*
2386                          * we may have dropped the spin lock to get the head
2387                          * mutex lock, and that might have given someone else
2388                          * time to free the head.  If that's true, it has been
2389                          * removed from our list and we can move on.
2390                          */
2391                         if (ret == -EAGAIN) {
2392                                 locked_ref = NULL;
2393                                 count++;
2394                                 continue;
2395                         }
2396                 }
2397
2398                 /*
2399                  * We need to try and merge add/drops of the same ref since we
2400                  * can run into issues with relocate dropping the implicit ref
2401                  * and then it being added back again before the drop can
2402                  * finish.  If we merged anything we need to re-loop so we can
2403                  * get a good ref.
2404                  */
2405                 spin_lock(&locked_ref->lock);
2406                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2407                                          locked_ref);
2408
2409                 /*
2410                  * locked_ref is the head node, so we have to go one
2411                  * node back for any delayed ref updates
2412                  */
2413                 ref = select_delayed_ref(locked_ref);
2414
2415                 if (ref && ref->seq &&
2416                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2417                         spin_unlock(&locked_ref->lock);
2418                         btrfs_delayed_ref_unlock(locked_ref);
2419                         spin_lock(&delayed_refs->lock);
2420                         locked_ref->processing = 0;
2421                         delayed_refs->num_heads_ready++;
2422                         spin_unlock(&delayed_refs->lock);
2423                         locked_ref = NULL;
2424                         cond_resched();
2425                         count++;
2426                         continue;
2427                 }
2428
2429                 /*
2430                  * record the must insert reserved flag before we
2431                  * drop the spin lock.
2432                  */
2433                 must_insert_reserved = locked_ref->must_insert_reserved;
2434                 locked_ref->must_insert_reserved = 0;
2435
2436                 extent_op = locked_ref->extent_op;
2437                 locked_ref->extent_op = NULL;
2438
2439                 if (!ref) {
2440
2441
2442                         /* All delayed refs have been processed, Go ahead
2443                          * and send the head node to run_one_delayed_ref,
2444                          * so that any accounting fixes can happen
2445                          */
2446                         ref = &locked_ref->node;
2447
2448                         if (extent_op && must_insert_reserved) {
2449                                 btrfs_free_delayed_extent_op(extent_op);
2450                                 extent_op = NULL;
2451                         }
2452
2453                         if (extent_op) {
2454                                 spin_unlock(&locked_ref->lock);
2455                                 ret = run_delayed_extent_op(trans, root,
2456                                                             ref, extent_op);
2457                                 btrfs_free_delayed_extent_op(extent_op);
2458
2459                                 if (ret) {
2460                                         /*
2461                                          * Need to reset must_insert_reserved if
2462                                          * there was an error so the abort stuff
2463                                          * can cleanup the reserved space
2464                                          * properly.
2465                                          */
2466                                         if (must_insert_reserved)
2467                                                 locked_ref->must_insert_reserved = 1;
2468                                         locked_ref->processing = 0;
2469                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2470                                         btrfs_delayed_ref_unlock(locked_ref);
2471                                         return ret;
2472                                 }
2473                                 continue;
2474                         }
2475
2476                         /*
2477                          * Need to drop our head ref lock and re-aqcuire the
2478                          * delayed ref lock and then re-check to make sure
2479                          * nobody got added.
2480                          */
2481                         spin_unlock(&locked_ref->lock);
2482                         spin_lock(&delayed_refs->lock);
2483                         spin_lock(&locked_ref->lock);
2484                         if (rb_first(&locked_ref->ref_root) ||
2485                             locked_ref->extent_op) {
2486                                 spin_unlock(&locked_ref->lock);
2487                                 spin_unlock(&delayed_refs->lock);
2488                                 continue;
2489                         }
2490                         ref->in_tree = 0;
2491                         delayed_refs->num_heads--;
2492                         rb_erase(&locked_ref->href_node,
2493                                  &delayed_refs->href_root);
2494                         spin_unlock(&delayed_refs->lock);
2495                 } else {
2496                         actual_count++;
2497                         ref->in_tree = 0;
2498                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2499                 }
2500                 atomic_dec(&delayed_refs->num_entries);
2501
2502                 if (!btrfs_delayed_ref_is_head(ref)) {
2503                         /*
2504                          * when we play the delayed ref, also correct the
2505                          * ref_mod on head
2506                          */
2507                         switch (ref->action) {
2508                         case BTRFS_ADD_DELAYED_REF:
2509                         case BTRFS_ADD_DELAYED_EXTENT:
2510                                 locked_ref->node.ref_mod -= ref->ref_mod;
2511                                 break;
2512                         case BTRFS_DROP_DELAYED_REF:
2513                                 locked_ref->node.ref_mod += ref->ref_mod;
2514                                 break;
2515                         default:
2516                                 WARN_ON(1);
2517                         }
2518                 }
2519                 spin_unlock(&locked_ref->lock);
2520
2521                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2522                                           must_insert_reserved);
2523
2524                 btrfs_free_delayed_extent_op(extent_op);
2525                 if (ret) {
2526                         locked_ref->processing = 0;
2527                         btrfs_delayed_ref_unlock(locked_ref);
2528                         btrfs_put_delayed_ref(ref);
2529                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2530                         return ret;
2531                 }
2532
2533                 /*
2534                  * If this node is a head, that means all the refs in this head
2535                  * have been dealt with, and we will pick the next head to deal
2536                  * with, so we must unlock the head and drop it from the cluster
2537                  * list before we release it.
2538                  */
2539                 if (btrfs_delayed_ref_is_head(ref)) {
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         locked_ref = NULL;
2542                 }
2543                 btrfs_put_delayed_ref(ref);
2544                 count++;
2545                 cond_resched();
2546         }
2547
2548         /*
2549          * We don't want to include ref heads since we can have empty ref heads
2550          * and those will drastically skew our runtime down since we just do
2551          * accounting, no actual extent tree updates.
2552          */
2553         if (actual_count > 0) {
2554                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2555                 u64 avg;
2556
2557                 /*
2558                  * We weigh the current average higher than our current runtime
2559                  * to avoid large swings in the average.
2560                  */
2561                 spin_lock(&delayed_refs->lock);
2562                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2563                 avg = div64_u64(avg, 4);
2564                 fs_info->avg_delayed_ref_runtime = avg;
2565                 spin_unlock(&delayed_refs->lock);
2566         }
2567         return 0;
2568 }
2569
2570 #ifdef SCRAMBLE_DELAYED_REFS
2571 /*
2572  * Normally delayed refs get processed in ascending bytenr order. This
2573  * correlates in most cases to the order added. To expose dependencies on this
2574  * order, we start to process the tree in the middle instead of the beginning
2575  */
2576 static u64 find_middle(struct rb_root *root)
2577 {
2578         struct rb_node *n = root->rb_node;
2579         struct btrfs_delayed_ref_node *entry;
2580         int alt = 1;
2581         u64 middle;
2582         u64 first = 0, last = 0;
2583
2584         n = rb_first(root);
2585         if (n) {
2586                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2587                 first = entry->bytenr;
2588         }
2589         n = rb_last(root);
2590         if (n) {
2591                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2592                 last = entry->bytenr;
2593         }
2594         n = root->rb_node;
2595
2596         while (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 WARN_ON(!entry->in_tree);
2599
2600                 middle = entry->bytenr;
2601
2602                 if (alt)
2603                         n = n->rb_left;
2604                 else
2605                         n = n->rb_right;
2606
2607                 alt = 1 - alt;
2608         }
2609         return middle;
2610 }
2611 #endif
2612
2613 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2614 {
2615         u64 num_bytes;
2616
2617         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2618                              sizeof(struct btrfs_extent_inline_ref));
2619         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2620                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2621
2622         /*
2623          * We don't ever fill up leaves all the way so multiply by 2 just to be
2624          * closer to what we're really going to want to ouse.
2625          */
2626         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2627 }
2628
2629 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2630                                        struct btrfs_root *root)
2631 {
2632         struct btrfs_block_rsv *global_rsv;
2633         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2634         u64 num_bytes;
2635         int ret = 0;
2636
2637         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2638         num_heads = heads_to_leaves(root, num_heads);
2639         if (num_heads > 1)
2640                 num_bytes += (num_heads - 1) * root->nodesize;
2641         num_bytes <<= 1;
2642         global_rsv = &root->fs_info->global_block_rsv;
2643
2644         /*
2645          * If we can't allocate any more chunks lets make sure we have _lots_ of
2646          * wiggle room since running delayed refs can create more delayed refs.
2647          */
2648         if (global_rsv->space_info->full)
2649                 num_bytes <<= 1;
2650
2651         spin_lock(&global_rsv->lock);
2652         if (global_rsv->reserved <= num_bytes)
2653                 ret = 1;
2654         spin_unlock(&global_rsv->lock);
2655         return ret;
2656 }
2657
2658 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2659                                        struct btrfs_root *root)
2660 {
2661         struct btrfs_fs_info *fs_info = root->fs_info;
2662         u64 num_entries =
2663                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2664         u64 avg_runtime;
2665         u64 val;
2666
2667         smp_mb();
2668         avg_runtime = fs_info->avg_delayed_ref_runtime;
2669         val = num_entries * avg_runtime;
2670         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2671                 return 1;
2672         if (val >= NSEC_PER_SEC / 2)
2673                 return 2;
2674
2675         return btrfs_check_space_for_delayed_refs(trans, root);
2676 }
2677
2678 struct async_delayed_refs {
2679         struct btrfs_root *root;
2680         int count;
2681         int error;
2682         int sync;
2683         struct completion wait;
2684         struct btrfs_work work;
2685 };
2686
2687 static void delayed_ref_async_start(struct btrfs_work *work)
2688 {
2689         struct async_delayed_refs *async;
2690         struct btrfs_trans_handle *trans;
2691         int ret;
2692
2693         async = container_of(work, struct async_delayed_refs, work);
2694
2695         trans = btrfs_join_transaction(async->root);
2696         if (IS_ERR(trans)) {
2697                 async->error = PTR_ERR(trans);
2698                 goto done;
2699         }
2700
2701         /*
2702          * trans->sync means that when we call end_transaciton, we won't
2703          * wait on delayed refs
2704          */
2705         trans->sync = true;
2706         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2707         if (ret)
2708                 async->error = ret;
2709
2710         ret = btrfs_end_transaction(trans, async->root);
2711         if (ret && !async->error)
2712                 async->error = ret;
2713 done:
2714         if (async->sync)
2715                 complete(&async->wait);
2716         else
2717                 kfree(async);
2718 }
2719
2720 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2721                                  unsigned long count, int wait)
2722 {
2723         struct async_delayed_refs *async;
2724         int ret;
2725
2726         async = kmalloc(sizeof(*async), GFP_NOFS);
2727         if (!async)
2728                 return -ENOMEM;
2729
2730         async->root = root->fs_info->tree_root;
2731         async->count = count;
2732         async->error = 0;
2733         if (wait)
2734                 async->sync = 1;
2735         else
2736                 async->sync = 0;
2737         init_completion(&async->wait);
2738
2739         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2740                         delayed_ref_async_start, NULL, NULL);
2741
2742         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2743
2744         if (wait) {
2745                 wait_for_completion(&async->wait);
2746                 ret = async->error;
2747                 kfree(async);
2748                 return ret;
2749         }
2750         return 0;
2751 }
2752
2753 /*
2754  * this starts processing the delayed reference count updates and
2755  * extent insertions we have queued up so far.  count can be
2756  * 0, which means to process everything in the tree at the start
2757  * of the run (but not newly added entries), or it can be some target
2758  * number you'd like to process.
2759  *
2760  * Returns 0 on success or if called with an aborted transaction
2761  * Returns <0 on error and aborts the transaction
2762  */
2763 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2764                            struct btrfs_root *root, unsigned long count)
2765 {
2766         struct rb_node *node;
2767         struct btrfs_delayed_ref_root *delayed_refs;
2768         struct btrfs_delayed_ref_head *head;
2769         int ret;
2770         int run_all = count == (unsigned long)-1;
2771         int run_most = 0;
2772
2773         /* We'll clean this up in btrfs_cleanup_transaction */
2774         if (trans->aborted)
2775                 return 0;
2776
2777         if (root == root->fs_info->extent_root)
2778                 root = root->fs_info->tree_root;
2779
2780         delayed_refs = &trans->transaction->delayed_refs;
2781         if (count == 0) {
2782                 count = atomic_read(&delayed_refs->num_entries) * 2;
2783                 run_most = 1;
2784         }
2785
2786 again:
2787 #ifdef SCRAMBLE_DELAYED_REFS
2788         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2789 #endif
2790         ret = __btrfs_run_delayed_refs(trans, root, count);
2791         if (ret < 0) {
2792                 btrfs_abort_transaction(trans, root, ret);
2793                 return ret;
2794         }
2795
2796         if (run_all) {
2797                 if (!list_empty(&trans->new_bgs))
2798                         btrfs_create_pending_block_groups(trans, root);
2799
2800                 spin_lock(&delayed_refs->lock);
2801                 node = rb_first(&delayed_refs->href_root);
2802                 if (!node) {
2803                         spin_unlock(&delayed_refs->lock);
2804                         goto out;
2805                 }
2806                 count = (unsigned long)-1;
2807
2808                 while (node) {
2809                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2810                                         href_node);
2811                         if (btrfs_delayed_ref_is_head(&head->node)) {
2812                                 struct btrfs_delayed_ref_node *ref;
2813
2814                                 ref = &head->node;
2815                                 atomic_inc(&ref->refs);
2816
2817                                 spin_unlock(&delayed_refs->lock);
2818                                 /*
2819                                  * Mutex was contended, block until it's
2820                                  * released and try again
2821                                  */
2822                                 mutex_lock(&head->mutex);
2823                                 mutex_unlock(&head->mutex);
2824
2825                                 btrfs_put_delayed_ref(ref);
2826                                 cond_resched();
2827                                 goto again;
2828                         } else {
2829                                 WARN_ON(1);
2830                         }
2831                         node = rb_next(node);
2832                 }
2833                 spin_unlock(&delayed_refs->lock);
2834                 cond_resched();
2835                 goto again;
2836         }
2837 out:
2838         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2839         if (ret)
2840                 return ret;
2841         assert_qgroups_uptodate(trans);
2842         return 0;
2843 }
2844
2845 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2846                                 struct btrfs_root *root,
2847                                 u64 bytenr, u64 num_bytes, u64 flags,
2848                                 int level, int is_data)
2849 {
2850         struct btrfs_delayed_extent_op *extent_op;
2851         int ret;
2852
2853         extent_op = btrfs_alloc_delayed_extent_op();
2854         if (!extent_op)
2855                 return -ENOMEM;
2856
2857         extent_op->flags_to_set = flags;
2858         extent_op->update_flags = 1;
2859         extent_op->update_key = 0;
2860         extent_op->is_data = is_data ? 1 : 0;
2861         extent_op->level = level;
2862
2863         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2864                                           num_bytes, extent_op);
2865         if (ret)
2866                 btrfs_free_delayed_extent_op(extent_op);
2867         return ret;
2868 }
2869
2870 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2871                                       struct btrfs_root *root,
2872                                       struct btrfs_path *path,
2873                                       u64 objectid, u64 offset, u64 bytenr)
2874 {
2875         struct btrfs_delayed_ref_head *head;
2876         struct btrfs_delayed_ref_node *ref;
2877         struct btrfs_delayed_data_ref *data_ref;
2878         struct btrfs_delayed_ref_root *delayed_refs;
2879         struct rb_node *node;
2880         int ret = 0;
2881
2882         delayed_refs = &trans->transaction->delayed_refs;
2883         spin_lock(&delayed_refs->lock);
2884         head = btrfs_find_delayed_ref_head(trans, bytenr);
2885         if (!head) {
2886                 spin_unlock(&delayed_refs->lock);
2887                 return 0;
2888         }
2889
2890         if (!mutex_trylock(&head->mutex)) {
2891                 atomic_inc(&head->node.refs);
2892                 spin_unlock(&delayed_refs->lock);
2893
2894                 btrfs_release_path(path);
2895
2896                 /*
2897                  * Mutex was contended, block until it's released and let
2898                  * caller try again
2899                  */
2900                 mutex_lock(&head->mutex);
2901                 mutex_unlock(&head->mutex);
2902                 btrfs_put_delayed_ref(&head->node);
2903                 return -EAGAIN;
2904         }
2905         spin_unlock(&delayed_refs->lock);
2906
2907         spin_lock(&head->lock);
2908         node = rb_first(&head->ref_root);
2909         while (node) {
2910                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2911                 node = rb_next(node);
2912
2913                 /* If it's a shared ref we know a cross reference exists */
2914                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2915                         ret = 1;
2916                         break;
2917                 }
2918
2919                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2920
2921                 /*
2922                  * If our ref doesn't match the one we're currently looking at
2923                  * then we have a cross reference.
2924                  */
2925                 if (data_ref->root != root->root_key.objectid ||
2926                     data_ref->objectid != objectid ||
2927                     data_ref->offset != offset) {
2928                         ret = 1;
2929                         break;
2930                 }
2931         }
2932         spin_unlock(&head->lock);
2933         mutex_unlock(&head->mutex);
2934         return ret;
2935 }
2936
2937 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2938                                         struct btrfs_root *root,
2939                                         struct btrfs_path *path,
2940                                         u64 objectid, u64 offset, u64 bytenr)
2941 {
2942         struct btrfs_root *extent_root = root->fs_info->extent_root;
2943         struct extent_buffer *leaf;
2944         struct btrfs_extent_data_ref *ref;
2945         struct btrfs_extent_inline_ref *iref;
2946         struct btrfs_extent_item *ei;
2947         struct btrfs_key key;
2948         u32 item_size;
2949         int ret;
2950
2951         key.objectid = bytenr;
2952         key.offset = (u64)-1;
2953         key.type = BTRFS_EXTENT_ITEM_KEY;
2954
2955         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2956         if (ret < 0)
2957                 goto out;
2958         BUG_ON(ret == 0); /* Corruption */
2959
2960         ret = -ENOENT;
2961         if (path->slots[0] == 0)
2962                 goto out;
2963
2964         path->slots[0]--;
2965         leaf = path->nodes[0];
2966         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2967
2968         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2969                 goto out;
2970
2971         ret = 1;
2972         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2973 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2974         if (item_size < sizeof(*ei)) {
2975                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2976                 goto out;
2977         }
2978 #endif
2979         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2980
2981         if (item_size != sizeof(*ei) +
2982             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2983                 goto out;
2984
2985         if (btrfs_extent_generation(leaf, ei) <=
2986             btrfs_root_last_snapshot(&root->root_item))
2987                 goto out;
2988
2989         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2990         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2991             BTRFS_EXTENT_DATA_REF_KEY)
2992                 goto out;
2993
2994         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2995         if (btrfs_extent_refs(leaf, ei) !=
2996             btrfs_extent_data_ref_count(leaf, ref) ||
2997             btrfs_extent_data_ref_root(leaf, ref) !=
2998             root->root_key.objectid ||
2999             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3000             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3001                 goto out;
3002
3003         ret = 0;
3004 out:
3005         return ret;
3006 }
3007
3008 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3009                           struct btrfs_root *root,
3010                           u64 objectid, u64 offset, u64 bytenr)
3011 {
3012         struct btrfs_path *path;
3013         int ret;
3014         int ret2;
3015
3016         path = btrfs_alloc_path();
3017         if (!path)
3018                 return -ENOENT;
3019
3020         do {
3021                 ret = check_committed_ref(trans, root, path, objectid,
3022                                           offset, bytenr);
3023                 if (ret && ret != -ENOENT)
3024                         goto out;
3025
3026                 ret2 = check_delayed_ref(trans, root, path, objectid,
3027                                          offset, bytenr);
3028         } while (ret2 == -EAGAIN);
3029
3030         if (ret2 && ret2 != -ENOENT) {
3031                 ret = ret2;
3032                 goto out;
3033         }
3034
3035         if (ret != -ENOENT || ret2 != -ENOENT)
3036                 ret = 0;
3037 out:
3038         btrfs_free_path(path);
3039         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3040                 WARN_ON(ret > 0);
3041         return ret;
3042 }
3043
3044 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3045                            struct btrfs_root *root,
3046                            struct extent_buffer *buf,
3047                            int full_backref, int inc)
3048 {
3049         u64 bytenr;
3050         u64 num_bytes;