37d8927015d9fa2b25082d04acf409614f3732f9
[sfrench/cifs-2.6.git] / fs / btrfs / extent-tree.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/sched/signal.h>
8 #include <linux/pagemap.h>
9 #include <linux/writeback.h>
10 #include <linux/blkdev.h>
11 #include <linux/sort.h>
12 #include <linux/rcupdate.h>
13 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/ratelimit.h>
16 #include <linux/percpu_counter.h>
17 #include <linux/lockdep.h>
18 #include <linux/crc32c.h>
19 #include "tree-log.h"
20 #include "disk-io.h"
21 #include "print-tree.h"
22 #include "volumes.h"
23 #include "raid56.h"
24 #include "locking.h"
25 #include "free-space-cache.h"
26 #include "free-space-tree.h"
27 #include "math.h"
28 #include "sysfs.h"
29 #include "qgroup.h"
30 #include "ref-verify.h"
31
32 #undef SCRAMBLE_DELAYED_REFS
33
34 /*
35  * control flags for do_chunk_alloc's force field
36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
37  * if we really need one.
38  *
39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
40  * if we have very few chunks already allocated.  This is
41  * used as part of the clustering code to help make sure
42  * we have a good pool of storage to cluster in, without
43  * filling the FS with empty chunks
44  *
45  * CHUNK_ALLOC_FORCE means it must try to allocate one
46  *
47  */
48 enum {
49         CHUNK_ALLOC_NO_FORCE = 0,
50         CHUNK_ALLOC_LIMITED = 1,
51         CHUNK_ALLOC_FORCE = 2,
52 };
53
54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
55                                struct btrfs_delayed_ref_node *node, u64 parent,
56                                u64 root_objectid, u64 owner_objectid,
57                                u64 owner_offset, int refs_to_drop,
58                                struct btrfs_delayed_extent_op *extra_op);
59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
60                                     struct extent_buffer *leaf,
61                                     struct btrfs_extent_item *ei);
62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
63                                       u64 parent, u64 root_objectid,
64                                       u64 flags, u64 owner, u64 offset,
65                                       struct btrfs_key *ins, int ref_mod);
66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
67                                      struct btrfs_delayed_ref_node *node,
68                                      struct btrfs_delayed_extent_op *extent_op);
69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
70                           int force);
71 static int find_next_key(struct btrfs_path *path, int level,
72                          struct btrfs_key *key);
73 static void dump_space_info(struct btrfs_fs_info *fs_info,
74                             struct btrfs_space_info *info, u64 bytes,
75                             int dump_block_groups);
76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
77                                u64 num_bytes);
78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
79                                      struct btrfs_space_info *space_info,
80                                      u64 num_bytes);
81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
82                                      struct btrfs_space_info *space_info,
83                                      u64 num_bytes);
84
85 static noinline int
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
87 {
88         smp_mb();
89         return cache->cached == BTRFS_CACHE_FINISHED ||
90                 cache->cached == BTRFS_CACHE_ERROR;
91 }
92
93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
94 {
95         return (cache->flags & bits) == bits;
96 }
97
98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
99 {
100         atomic_inc(&cache->count);
101 }
102
103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
104 {
105         if (atomic_dec_and_test(&cache->count)) {
106                 WARN_ON(cache->pinned > 0);
107                 WARN_ON(cache->reserved > 0);
108
109                 /*
110                  * If not empty, someone is still holding mutex of
111                  * full_stripe_lock, which can only be released by caller.
112                  * And it will definitely cause use-after-free when caller
113                  * tries to release full stripe lock.
114                  *
115                  * No better way to resolve, but only to warn.
116                  */
117                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
118                 kfree(cache->free_space_ctl);
119                 kfree(cache);
120         }
121 }
122
123 /*
124  * this adds the block group to the fs_info rb tree for the block group
125  * cache
126  */
127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
128                                 struct btrfs_block_group_cache *block_group)
129 {
130         struct rb_node **p;
131         struct rb_node *parent = NULL;
132         struct btrfs_block_group_cache *cache;
133
134         spin_lock(&info->block_group_cache_lock);
135         p = &info->block_group_cache_tree.rb_node;
136
137         while (*p) {
138                 parent = *p;
139                 cache = rb_entry(parent, struct btrfs_block_group_cache,
140                                  cache_node);
141                 if (block_group->key.objectid < cache->key.objectid) {
142                         p = &(*p)->rb_left;
143                 } else if (block_group->key.objectid > cache->key.objectid) {
144                         p = &(*p)->rb_right;
145                 } else {
146                         spin_unlock(&info->block_group_cache_lock);
147                         return -EEXIST;
148                 }
149         }
150
151         rb_link_node(&block_group->cache_node, parent, p);
152         rb_insert_color(&block_group->cache_node,
153                         &info->block_group_cache_tree);
154
155         if (info->first_logical_byte > block_group->key.objectid)
156                 info->first_logical_byte = block_group->key.objectid;
157
158         spin_unlock(&info->block_group_cache_lock);
159
160         return 0;
161 }
162
163 /*
164  * This will return the block group at or after bytenr if contains is 0, else
165  * it will return the block group that contains the bytenr
166  */
167 static struct btrfs_block_group_cache *
168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
169                               int contains)
170 {
171         struct btrfs_block_group_cache *cache, *ret = NULL;
172         struct rb_node *n;
173         u64 end, start;
174
175         spin_lock(&info->block_group_cache_lock);
176         n = info->block_group_cache_tree.rb_node;
177
178         while (n) {
179                 cache = rb_entry(n, struct btrfs_block_group_cache,
180                                  cache_node);
181                 end = cache->key.objectid + cache->key.offset - 1;
182                 start = cache->key.objectid;
183
184                 if (bytenr < start) {
185                         if (!contains && (!ret || start < ret->key.objectid))
186                                 ret = cache;
187                         n = n->rb_left;
188                 } else if (bytenr > start) {
189                         if (contains && bytenr <= end) {
190                                 ret = cache;
191                                 break;
192                         }
193                         n = n->rb_right;
194                 } else {
195                         ret = cache;
196                         break;
197                 }
198         }
199         if (ret) {
200                 btrfs_get_block_group(ret);
201                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
202                         info->first_logical_byte = ret->key.objectid;
203         }
204         spin_unlock(&info->block_group_cache_lock);
205
206         return ret;
207 }
208
209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
210                                u64 start, u64 num_bytes)
211 {
212         u64 end = start + num_bytes - 1;
213         set_extent_bits(&fs_info->freed_extents[0],
214                         start, end, EXTENT_UPTODATE);
215         set_extent_bits(&fs_info->freed_extents[1],
216                         start, end, EXTENT_UPTODATE);
217         return 0;
218 }
219
220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
221 {
222         struct btrfs_fs_info *fs_info = cache->fs_info;
223         u64 start, end;
224
225         start = cache->key.objectid;
226         end = start + cache->key.offset - 1;
227
228         clear_extent_bits(&fs_info->freed_extents[0],
229                           start, end, EXTENT_UPTODATE);
230         clear_extent_bits(&fs_info->freed_extents[1],
231                           start, end, EXTENT_UPTODATE);
232 }
233
234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
235 {
236         struct btrfs_fs_info *fs_info = cache->fs_info;
237         u64 bytenr;
238         u64 *logical;
239         int stripe_len;
240         int i, nr, ret;
241
242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244                 cache->bytes_super += stripe_len;
245                 ret = add_excluded_extent(fs_info, cache->key.objectid,
246                                           stripe_len);
247                 if (ret)
248                         return ret;
249         }
250
251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252                 bytenr = btrfs_sb_offset(i);
253                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
254                                        bytenr, &logical, &nr, &stripe_len);
255                 if (ret)
256                         return ret;
257
258                 while (nr--) {
259                         u64 start, len;
260
261                         if (logical[nr] > cache->key.objectid +
262                             cache->key.offset)
263                                 continue;
264
265                         if (logical[nr] + stripe_len <= cache->key.objectid)
266                                 continue;
267
268                         start = logical[nr];
269                         if (start < cache->key.objectid) {
270                                 start = cache->key.objectid;
271                                 len = (logical[nr] + stripe_len) - start;
272                         } else {
273                                 len = min_t(u64, stripe_len,
274                                             cache->key.objectid +
275                                             cache->key.offset - start);
276                         }
277
278                         cache->bytes_super += len;
279                         ret = add_excluded_extent(fs_info, start, len);
280                         if (ret) {
281                                 kfree(logical);
282                                 return ret;
283                         }
284                 }
285
286                 kfree(logical);
287         }
288         return 0;
289 }
290
291 static struct btrfs_caching_control *
292 get_caching_control(struct btrfs_block_group_cache *cache)
293 {
294         struct btrfs_caching_control *ctl;
295
296         spin_lock(&cache->lock);
297         if (!cache->caching_ctl) {
298                 spin_unlock(&cache->lock);
299                 return NULL;
300         }
301
302         ctl = cache->caching_ctl;
303         refcount_inc(&ctl->count);
304         spin_unlock(&cache->lock);
305         return ctl;
306 }
307
308 static void put_caching_control(struct btrfs_caching_control *ctl)
309 {
310         if (refcount_dec_and_test(&ctl->count))
311                 kfree(ctl);
312 }
313
314 #ifdef CONFIG_BTRFS_DEBUG
315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
316 {
317         struct btrfs_fs_info *fs_info = block_group->fs_info;
318         u64 start = block_group->key.objectid;
319         u64 len = block_group->key.offset;
320         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
321                 fs_info->nodesize : fs_info->sectorsize;
322         u64 step = chunk << 1;
323
324         while (len > chunk) {
325                 btrfs_remove_free_space(block_group, start, chunk);
326                 start += step;
327                 if (len < step)
328                         len = 0;
329                 else
330                         len -= step;
331         }
332 }
333 #endif
334
335 /*
336  * this is only called by cache_block_group, since we could have freed extents
337  * we need to check the pinned_extents for any extents that can't be used yet
338  * since their free space will be released as soon as the transaction commits.
339  */
340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
341                        u64 start, u64 end)
342 {
343         struct btrfs_fs_info *info = block_group->fs_info;
344         u64 extent_start, extent_end, size, total_added = 0;
345         int ret;
346
347         while (start < end) {
348                 ret = find_first_extent_bit(info->pinned_extents, start,
349                                             &extent_start, &extent_end,
350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
351                                             NULL);
352                 if (ret)
353                         break;
354
355                 if (extent_start <= start) {
356                         start = extent_end + 1;
357                 } else if (extent_start > start && extent_start < end) {
358                         size = extent_start - start;
359                         total_added += size;
360                         ret = btrfs_add_free_space(block_group, start,
361                                                    size);
362                         BUG_ON(ret); /* -ENOMEM or logic error */
363                         start = extent_end + 1;
364                 } else {
365                         break;
366                 }
367         }
368
369         if (start < end) {
370                 size = end - start;
371                 total_added += size;
372                 ret = btrfs_add_free_space(block_group, start, size);
373                 BUG_ON(ret); /* -ENOMEM or logic error */
374         }
375
376         return total_added;
377 }
378
379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
380 {
381         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
382         struct btrfs_fs_info *fs_info = block_group->fs_info;
383         struct btrfs_root *extent_root = fs_info->extent_root;
384         struct btrfs_path *path;
385         struct extent_buffer *leaf;
386         struct btrfs_key key;
387         u64 total_found = 0;
388         u64 last = 0;
389         u32 nritems;
390         int ret;
391         bool wakeup = true;
392
393         path = btrfs_alloc_path();
394         if (!path)
395                 return -ENOMEM;
396
397         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
398
399 #ifdef CONFIG_BTRFS_DEBUG
400         /*
401          * If we're fragmenting we don't want to make anybody think we can
402          * allocate from this block group until we've had a chance to fragment
403          * the free space.
404          */
405         if (btrfs_should_fragment_free_space(block_group))
406                 wakeup = false;
407 #endif
408         /*
409          * We don't want to deadlock with somebody trying to allocate a new
410          * extent for the extent root while also trying to search the extent
411          * root to add free space.  So we skip locking and search the commit
412          * root, since its read-only
413          */
414         path->skip_locking = 1;
415         path->search_commit_root = 1;
416         path->reada = READA_FORWARD;
417
418         key.objectid = last;
419         key.offset = 0;
420         key.type = BTRFS_EXTENT_ITEM_KEY;
421
422 next:
423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
424         if (ret < 0)
425                 goto out;
426
427         leaf = path->nodes[0];
428         nritems = btrfs_header_nritems(leaf);
429
430         while (1) {
431                 if (btrfs_fs_closing(fs_info) > 1) {
432                         last = (u64)-1;
433                         break;
434                 }
435
436                 if (path->slots[0] < nritems) {
437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
438                 } else {
439                         ret = find_next_key(path, 0, &key);
440                         if (ret)
441                                 break;
442
443                         if (need_resched() ||
444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
445                                 if (wakeup)
446                                         caching_ctl->progress = last;
447                                 btrfs_release_path(path);
448                                 up_read(&fs_info->commit_root_sem);
449                                 mutex_unlock(&caching_ctl->mutex);
450                                 cond_resched();
451                                 mutex_lock(&caching_ctl->mutex);
452                                 down_read(&fs_info->commit_root_sem);
453                                 goto next;
454                         }
455
456                         ret = btrfs_next_leaf(extent_root, path);
457                         if (ret < 0)
458                                 goto out;
459                         if (ret)
460                                 break;
461                         leaf = path->nodes[0];
462                         nritems = btrfs_header_nritems(leaf);
463                         continue;
464                 }
465
466                 if (key.objectid < last) {
467                         key.objectid = last;
468                         key.offset = 0;
469                         key.type = BTRFS_EXTENT_ITEM_KEY;
470
471                         if (wakeup)
472                                 caching_ctl->progress = last;
473                         btrfs_release_path(path);
474                         goto next;
475                 }
476
477                 if (key.objectid < block_group->key.objectid) {
478                         path->slots[0]++;
479                         continue;
480                 }
481
482                 if (key.objectid >= block_group->key.objectid +
483                     block_group->key.offset)
484                         break;
485
486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
487                     key.type == BTRFS_METADATA_ITEM_KEY) {
488                         total_found += add_new_free_space(block_group, last,
489                                                           key.objectid);
490                         if (key.type == BTRFS_METADATA_ITEM_KEY)
491                                 last = key.objectid +
492                                         fs_info->nodesize;
493                         else
494                                 last = key.objectid + key.offset;
495
496                         if (total_found > CACHING_CTL_WAKE_UP) {
497                                 total_found = 0;
498                                 if (wakeup)
499                                         wake_up(&caching_ctl->wait);
500                         }
501                 }
502                 path->slots[0]++;
503         }
504         ret = 0;
505
506         total_found += add_new_free_space(block_group, last,
507                                           block_group->key.objectid +
508                                           block_group->key.offset);
509         caching_ctl->progress = (u64)-1;
510
511 out:
512         btrfs_free_path(path);
513         return ret;
514 }
515
516 static noinline void caching_thread(struct btrfs_work *work)
517 {
518         struct btrfs_block_group_cache *block_group;
519         struct btrfs_fs_info *fs_info;
520         struct btrfs_caching_control *caching_ctl;
521         int ret;
522
523         caching_ctl = container_of(work, struct btrfs_caching_control, work);
524         block_group = caching_ctl->block_group;
525         fs_info = block_group->fs_info;
526
527         mutex_lock(&caching_ctl->mutex);
528         down_read(&fs_info->commit_root_sem);
529
530         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
531                 ret = load_free_space_tree(caching_ctl);
532         else
533                 ret = load_extent_tree_free(caching_ctl);
534
535         spin_lock(&block_group->lock);
536         block_group->caching_ctl = NULL;
537         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
538         spin_unlock(&block_group->lock);
539
540 #ifdef CONFIG_BTRFS_DEBUG
541         if (btrfs_should_fragment_free_space(block_group)) {
542                 u64 bytes_used;
543
544                 spin_lock(&block_group->space_info->lock);
545                 spin_lock(&block_group->lock);
546                 bytes_used = block_group->key.offset -
547                         btrfs_block_group_used(&block_group->item);
548                 block_group->space_info->bytes_used += bytes_used >> 1;
549                 spin_unlock(&block_group->lock);
550                 spin_unlock(&block_group->space_info->lock);
551                 fragment_free_space(block_group);
552         }
553 #endif
554
555         caching_ctl->progress = (u64)-1;
556
557         up_read(&fs_info->commit_root_sem);
558         free_excluded_extents(block_group);
559         mutex_unlock(&caching_ctl->mutex);
560
561         wake_up(&caching_ctl->wait);
562
563         put_caching_control(caching_ctl);
564         btrfs_put_block_group(block_group);
565 }
566
567 static int cache_block_group(struct btrfs_block_group_cache *cache,
568                              int load_cache_only)
569 {
570         DEFINE_WAIT(wait);
571         struct btrfs_fs_info *fs_info = cache->fs_info;
572         struct btrfs_caching_control *caching_ctl;
573         int ret = 0;
574
575         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
576         if (!caching_ctl)
577                 return -ENOMEM;
578
579         INIT_LIST_HEAD(&caching_ctl->list);
580         mutex_init(&caching_ctl->mutex);
581         init_waitqueue_head(&caching_ctl->wait);
582         caching_ctl->block_group = cache;
583         caching_ctl->progress = cache->key.objectid;
584         refcount_set(&caching_ctl->count, 1);
585         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
586                         caching_thread, NULL, NULL);
587
588         spin_lock(&cache->lock);
589         /*
590          * This should be a rare occasion, but this could happen I think in the
591          * case where one thread starts to load the space cache info, and then
592          * some other thread starts a transaction commit which tries to do an
593          * allocation while the other thread is still loading the space cache
594          * info.  The previous loop should have kept us from choosing this block
595          * group, but if we've moved to the state where we will wait on caching
596          * block groups we need to first check if we're doing a fast load here,
597          * so we can wait for it to finish, otherwise we could end up allocating
598          * from a block group who's cache gets evicted for one reason or
599          * another.
600          */
601         while (cache->cached == BTRFS_CACHE_FAST) {
602                 struct btrfs_caching_control *ctl;
603
604                 ctl = cache->caching_ctl;
605                 refcount_inc(&ctl->count);
606                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
607                 spin_unlock(&cache->lock);
608
609                 schedule();
610
611                 finish_wait(&ctl->wait, &wait);
612                 put_caching_control(ctl);
613                 spin_lock(&cache->lock);
614         }
615
616         if (cache->cached != BTRFS_CACHE_NO) {
617                 spin_unlock(&cache->lock);
618                 kfree(caching_ctl);
619                 return 0;
620         }
621         WARN_ON(cache->caching_ctl);
622         cache->caching_ctl = caching_ctl;
623         cache->cached = BTRFS_CACHE_FAST;
624         spin_unlock(&cache->lock);
625
626         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
627                 mutex_lock(&caching_ctl->mutex);
628                 ret = load_free_space_cache(fs_info, cache);
629
630                 spin_lock(&cache->lock);
631                 if (ret == 1) {
632                         cache->caching_ctl = NULL;
633                         cache->cached = BTRFS_CACHE_FINISHED;
634                         cache->last_byte_to_unpin = (u64)-1;
635                         caching_ctl->progress = (u64)-1;
636                 } else {
637                         if (load_cache_only) {
638                                 cache->caching_ctl = NULL;
639                                 cache->cached = BTRFS_CACHE_NO;
640                         } else {
641                                 cache->cached = BTRFS_CACHE_STARTED;
642                                 cache->has_caching_ctl = 1;
643                         }
644                 }
645                 spin_unlock(&cache->lock);
646 #ifdef CONFIG_BTRFS_DEBUG
647                 if (ret == 1 &&
648                     btrfs_should_fragment_free_space(cache)) {
649                         u64 bytes_used;
650
651                         spin_lock(&cache->space_info->lock);
652                         spin_lock(&cache->lock);
653                         bytes_used = cache->key.offset -
654                                 btrfs_block_group_used(&cache->item);
655                         cache->space_info->bytes_used += bytes_used >> 1;
656                         spin_unlock(&cache->lock);
657                         spin_unlock(&cache->space_info->lock);
658                         fragment_free_space(cache);
659                 }
660 #endif
661                 mutex_unlock(&caching_ctl->mutex);
662
663                 wake_up(&caching_ctl->wait);
664                 if (ret == 1) {
665                         put_caching_control(caching_ctl);
666                         free_excluded_extents(cache);
667                         return 0;
668                 }
669         } else {
670                 /*
671                  * We're either using the free space tree or no caching at all.
672                  * Set cached to the appropriate value and wakeup any waiters.
673                  */
674                 spin_lock(&cache->lock);
675                 if (load_cache_only) {
676                         cache->caching_ctl = NULL;
677                         cache->cached = BTRFS_CACHE_NO;
678                 } else {
679                         cache->cached = BTRFS_CACHE_STARTED;
680                         cache->has_caching_ctl = 1;
681                 }
682                 spin_unlock(&cache->lock);
683                 wake_up(&caching_ctl->wait);
684         }
685
686         if (load_cache_only) {
687                 put_caching_control(caching_ctl);
688                 return 0;
689         }
690
691         down_write(&fs_info->commit_root_sem);
692         refcount_inc(&caching_ctl->count);
693         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
694         up_write(&fs_info->commit_root_sem);
695
696         btrfs_get_block_group(cache);
697
698         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
699
700         return ret;
701 }
702
703 /*
704  * return the block group that starts at or after bytenr
705  */
706 static struct btrfs_block_group_cache *
707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
708 {
709         return block_group_cache_tree_search(info, bytenr, 0);
710 }
711
712 /*
713  * return the block group that contains the given bytenr
714  */
715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
716                                                  struct btrfs_fs_info *info,
717                                                  u64 bytenr)
718 {
719         return block_group_cache_tree_search(info, bytenr, 1);
720 }
721
722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
723                                                   u64 flags)
724 {
725         struct list_head *head = &info->space_info;
726         struct btrfs_space_info *found;
727
728         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
729
730         rcu_read_lock();
731         list_for_each_entry_rcu(found, head, list) {
732                 if (found->flags & flags) {
733                         rcu_read_unlock();
734                         return found;
735                 }
736         }
737         rcu_read_unlock();
738         return NULL;
739 }
740
741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
742                              bool metadata, u64 root_objectid)
743 {
744         struct btrfs_space_info *space_info;
745         u64 flags;
746
747         if (metadata) {
748                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
749                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
750                 else
751                         flags = BTRFS_BLOCK_GROUP_METADATA;
752         } else {
753                 flags = BTRFS_BLOCK_GROUP_DATA;
754         }
755
756         space_info = __find_space_info(fs_info, flags);
757         ASSERT(space_info);
758         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
759 }
760
761 /*
762  * after adding space to the filesystem, we need to clear the full flags
763  * on all the space infos.
764  */
765 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
766 {
767         struct list_head *head = &info->space_info;
768         struct btrfs_space_info *found;
769
770         rcu_read_lock();
771         list_for_each_entry_rcu(found, head, list)
772                 found->full = 0;
773         rcu_read_unlock();
774 }
775
776 /* simple helper to search for an existing data extent at a given offset */
777 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
778 {
779         int ret;
780         struct btrfs_key key;
781         struct btrfs_path *path;
782
783         path = btrfs_alloc_path();
784         if (!path)
785                 return -ENOMEM;
786
787         key.objectid = start;
788         key.offset = len;
789         key.type = BTRFS_EXTENT_ITEM_KEY;
790         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
791         btrfs_free_path(path);
792         return ret;
793 }
794
795 /*
796  * helper function to lookup reference count and flags of a tree block.
797  *
798  * the head node for delayed ref is used to store the sum of all the
799  * reference count modifications queued up in the rbtree. the head
800  * node may also store the extent flags to set. This way you can check
801  * to see what the reference count and extent flags would be if all of
802  * the delayed refs are not processed.
803  */
804 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
805                              struct btrfs_fs_info *fs_info, u64 bytenr,
806                              u64 offset, int metadata, u64 *refs, u64 *flags)
807 {
808         struct btrfs_delayed_ref_head *head;
809         struct btrfs_delayed_ref_root *delayed_refs;
810         struct btrfs_path *path;
811         struct btrfs_extent_item *ei;
812         struct extent_buffer *leaf;
813         struct btrfs_key key;
814         u32 item_size;
815         u64 num_refs;
816         u64 extent_flags;
817         int ret;
818
819         /*
820          * If we don't have skinny metadata, don't bother doing anything
821          * different
822          */
823         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
824                 offset = fs_info->nodesize;
825                 metadata = 0;
826         }
827
828         path = btrfs_alloc_path();
829         if (!path)
830                 return -ENOMEM;
831
832         if (!trans) {
833                 path->skip_locking = 1;
834                 path->search_commit_root = 1;
835         }
836
837 search_again:
838         key.objectid = bytenr;
839         key.offset = offset;
840         if (metadata)
841                 key.type = BTRFS_METADATA_ITEM_KEY;
842         else
843                 key.type = BTRFS_EXTENT_ITEM_KEY;
844
845         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
846         if (ret < 0)
847                 goto out_free;
848
849         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
850                 if (path->slots[0]) {
851                         path->slots[0]--;
852                         btrfs_item_key_to_cpu(path->nodes[0], &key,
853                                               path->slots[0]);
854                         if (key.objectid == bytenr &&
855                             key.type == BTRFS_EXTENT_ITEM_KEY &&
856                             key.offset == fs_info->nodesize)
857                                 ret = 0;
858                 }
859         }
860
861         if (ret == 0) {
862                 leaf = path->nodes[0];
863                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
864                 if (item_size >= sizeof(*ei)) {
865                         ei = btrfs_item_ptr(leaf, path->slots[0],
866                                             struct btrfs_extent_item);
867                         num_refs = btrfs_extent_refs(leaf, ei);
868                         extent_flags = btrfs_extent_flags(leaf, ei);
869                 } else {
870 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
871                         struct btrfs_extent_item_v0 *ei0;
872                         BUG_ON(item_size != sizeof(*ei0));
873                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
874                                              struct btrfs_extent_item_v0);
875                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
876                         /* FIXME: this isn't correct for data */
877                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
878 #else
879                         BUG();
880 #endif
881                 }
882                 BUG_ON(num_refs == 0);
883         } else {
884                 num_refs = 0;
885                 extent_flags = 0;
886                 ret = 0;
887         }
888
889         if (!trans)
890                 goto out;
891
892         delayed_refs = &trans->transaction->delayed_refs;
893         spin_lock(&delayed_refs->lock);
894         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
895         if (head) {
896                 if (!mutex_trylock(&head->mutex)) {
897                         refcount_inc(&head->refs);
898                         spin_unlock(&delayed_refs->lock);
899
900                         btrfs_release_path(path);
901
902                         /*
903                          * Mutex was contended, block until it's released and try
904                          * again
905                          */
906                         mutex_lock(&head->mutex);
907                         mutex_unlock(&head->mutex);
908                         btrfs_put_delayed_ref_head(head);
909                         goto search_again;
910                 }
911                 spin_lock(&head->lock);
912                 if (head->extent_op && head->extent_op->update_flags)
913                         extent_flags |= head->extent_op->flags_to_set;
914                 else
915                         BUG_ON(num_refs == 0);
916
917                 num_refs += head->ref_mod;
918                 spin_unlock(&head->lock);
919                 mutex_unlock(&head->mutex);
920         }
921         spin_unlock(&delayed_refs->lock);
922 out:
923         WARN_ON(num_refs == 0);
924         if (refs)
925                 *refs = num_refs;
926         if (flags)
927                 *flags = extent_flags;
928 out_free:
929         btrfs_free_path(path);
930         return ret;
931 }
932
933 /*
934  * Back reference rules.  Back refs have three main goals:
935  *
936  * 1) differentiate between all holders of references to an extent so that
937  *    when a reference is dropped we can make sure it was a valid reference
938  *    before freeing the extent.
939  *
940  * 2) Provide enough information to quickly find the holders of an extent
941  *    if we notice a given block is corrupted or bad.
942  *
943  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
944  *    maintenance.  This is actually the same as #2, but with a slightly
945  *    different use case.
946  *
947  * There are two kinds of back refs. The implicit back refs is optimized
948  * for pointers in non-shared tree blocks. For a given pointer in a block,
949  * back refs of this kind provide information about the block's owner tree
950  * and the pointer's key. These information allow us to find the block by
951  * b-tree searching. The full back refs is for pointers in tree blocks not
952  * referenced by their owner trees. The location of tree block is recorded
953  * in the back refs. Actually the full back refs is generic, and can be
954  * used in all cases the implicit back refs is used. The major shortcoming
955  * of the full back refs is its overhead. Every time a tree block gets
956  * COWed, we have to update back refs entry for all pointers in it.
957  *
958  * For a newly allocated tree block, we use implicit back refs for
959  * pointers in it. This means most tree related operations only involve
960  * implicit back refs. For a tree block created in old transaction, the
961  * only way to drop a reference to it is COW it. So we can detect the
962  * event that tree block loses its owner tree's reference and do the
963  * back refs conversion.
964  *
965  * When a tree block is COWed through a tree, there are four cases:
966  *
967  * The reference count of the block is one and the tree is the block's
968  * owner tree. Nothing to do in this case.
969  *
970  * The reference count of the block is one and the tree is not the
971  * block's owner tree. In this case, full back refs is used for pointers
972  * in the block. Remove these full back refs, add implicit back refs for
973  * every pointers in the new block.
974  *
975  * The reference count of the block is greater than one and the tree is
976  * the block's owner tree. In this case, implicit back refs is used for
977  * pointers in the block. Add full back refs for every pointers in the
978  * block, increase lower level extents' reference counts. The original
979  * implicit back refs are entailed to the new block.
980  *
981  * The reference count of the block is greater than one and the tree is
982  * not the block's owner tree. Add implicit back refs for every pointer in
983  * the new block, increase lower level extents' reference count.
984  *
985  * Back Reference Key composing:
986  *
987  * The key objectid corresponds to the first byte in the extent,
988  * The key type is used to differentiate between types of back refs.
989  * There are different meanings of the key offset for different types
990  * of back refs.
991  *
992  * File extents can be referenced by:
993  *
994  * - multiple snapshots, subvolumes, or different generations in one subvol
995  * - different files inside a single subvolume
996  * - different offsets inside a file (bookend extents in file.c)
997  *
998  * The extent ref structure for the implicit back refs has fields for:
999  *
1000  * - Objectid of the subvolume root
1001  * - objectid of the file holding the reference
1002  * - original offset in the file
1003  * - how many bookend extents
1004  *
1005  * The key offset for the implicit back refs is hash of the first
1006  * three fields.
1007  *
1008  * The extent ref structure for the full back refs has field for:
1009  *
1010  * - number of pointers in the tree leaf
1011  *
1012  * The key offset for the implicit back refs is the first byte of
1013  * the tree leaf
1014  *
1015  * When a file extent is allocated, The implicit back refs is used.
1016  * the fields are filled in:
1017  *
1018  *     (root_key.objectid, inode objectid, offset in file, 1)
1019  *
1020  * When a file extent is removed file truncation, we find the
1021  * corresponding implicit back refs and check the following fields:
1022  *
1023  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1024  *
1025  * Btree extents can be referenced by:
1026  *
1027  * - Different subvolumes
1028  *
1029  * Both the implicit back refs and the full back refs for tree blocks
1030  * only consist of key. The key offset for the implicit back refs is
1031  * objectid of block's owner tree. The key offset for the full back refs
1032  * is the first byte of parent block.
1033  *
1034  * When implicit back refs is used, information about the lowest key and
1035  * level of the tree block are required. These information are stored in
1036  * tree block info structure.
1037  */
1038
1039 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1040 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1041                                   struct btrfs_fs_info *fs_info,
1042                                   struct btrfs_path *path,
1043                                   u64 owner, u32 extra_size)
1044 {
1045         struct btrfs_root *root = fs_info->extent_root;
1046         struct btrfs_extent_item *item;
1047         struct btrfs_extent_item_v0 *ei0;
1048         struct btrfs_extent_ref_v0 *ref0;
1049         struct btrfs_tree_block_info *bi;
1050         struct extent_buffer *leaf;
1051         struct btrfs_key key;
1052         struct btrfs_key found_key;
1053         u32 new_size = sizeof(*item);
1054         u64 refs;
1055         int ret;
1056
1057         leaf = path->nodes[0];
1058         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1059
1060         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1061         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1062                              struct btrfs_extent_item_v0);
1063         refs = btrfs_extent_refs_v0(leaf, ei0);
1064
1065         if (owner == (u64)-1) {
1066                 while (1) {
1067                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1068                                 ret = btrfs_next_leaf(root, path);
1069                                 if (ret < 0)
1070                                         return ret;
1071                                 BUG_ON(ret > 0); /* Corruption */
1072                                 leaf = path->nodes[0];
1073                         }
1074                         btrfs_item_key_to_cpu(leaf, &found_key,
1075                                               path->slots[0]);
1076                         BUG_ON(key.objectid != found_key.objectid);
1077                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1078                                 path->slots[0]++;
1079                                 continue;
1080                         }
1081                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1082                                               struct btrfs_extent_ref_v0);
1083                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1084                         break;
1085                 }
1086         }
1087         btrfs_release_path(path);
1088
1089         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1090                 new_size += sizeof(*bi);
1091
1092         new_size -= sizeof(*ei0);
1093         ret = btrfs_search_slot(trans, root, &key, path,
1094                                 new_size + extra_size, 1);
1095         if (ret < 0)
1096                 return ret;
1097         BUG_ON(ret); /* Corruption */
1098
1099         btrfs_extend_item(fs_info, path, new_size);
1100
1101         leaf = path->nodes[0];
1102         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1103         btrfs_set_extent_refs(leaf, item, refs);
1104         /* FIXME: get real generation */
1105         btrfs_set_extent_generation(leaf, item, 0);
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1107                 btrfs_set_extent_flags(leaf, item,
1108                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1109                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1110                 bi = (struct btrfs_tree_block_info *)(item + 1);
1111                 /* FIXME: get first key of the block */
1112                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1113                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1114         } else {
1115                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1116         }
1117         btrfs_mark_buffer_dirty(leaf);
1118         return 0;
1119 }
1120 #endif
1121
1122 /*
1123  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1124  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1125  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1126  */
1127 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1128                                      struct btrfs_extent_inline_ref *iref,
1129                                      enum btrfs_inline_ref_type is_data)
1130 {
1131         int type = btrfs_extent_inline_ref_type(eb, iref);
1132         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1133
1134         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1135             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1136             type == BTRFS_SHARED_DATA_REF_KEY ||
1137             type == BTRFS_EXTENT_DATA_REF_KEY) {
1138                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1139                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1140                                 return type;
1141                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1142                                 ASSERT(eb->fs_info);
1143                                 /*
1144                                  * Every shared one has parent tree
1145                                  * block, which must be aligned to
1146                                  * nodesize.
1147                                  */
1148                                 if (offset &&
1149                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1150                                         return type;
1151                         }
1152                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1153                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1154                                 return type;
1155                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1156                                 ASSERT(eb->fs_info);
1157                                 /*
1158                                  * Every shared one has parent tree
1159                                  * block, which must be aligned to
1160                                  * nodesize.
1161                                  */
1162                                 if (offset &&
1163                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1164                                         return type;
1165                         }
1166                 } else {
1167                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1168                         return type;
1169                 }
1170         }
1171
1172         btrfs_print_leaf((struct extent_buffer *)eb);
1173         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1174                   eb->start, type);
1175         WARN_ON(1);
1176
1177         return BTRFS_REF_TYPE_INVALID;
1178 }
1179
1180 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1181 {
1182         u32 high_crc = ~(u32)0;
1183         u32 low_crc = ~(u32)0;
1184         __le64 lenum;
1185
1186         lenum = cpu_to_le64(root_objectid);
1187         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1188         lenum = cpu_to_le64(owner);
1189         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1190         lenum = cpu_to_le64(offset);
1191         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1192
1193         return ((u64)high_crc << 31) ^ (u64)low_crc;
1194 }
1195
1196 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1197                                      struct btrfs_extent_data_ref *ref)
1198 {
1199         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1200                                     btrfs_extent_data_ref_objectid(leaf, ref),
1201                                     btrfs_extent_data_ref_offset(leaf, ref));
1202 }
1203
1204 static int match_extent_data_ref(struct extent_buffer *leaf,
1205                                  struct btrfs_extent_data_ref *ref,
1206                                  u64 root_objectid, u64 owner, u64 offset)
1207 {
1208         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1209             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1210             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1211                 return 0;
1212         return 1;
1213 }
1214
1215 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1216                                            struct btrfs_path *path,
1217                                            u64 bytenr, u64 parent,
1218                                            u64 root_objectid,
1219                                            u64 owner, u64 offset)
1220 {
1221         struct btrfs_root *root = trans->fs_info->extent_root;
1222         struct btrfs_key key;
1223         struct btrfs_extent_data_ref *ref;
1224         struct extent_buffer *leaf;
1225         u32 nritems;
1226         int ret;
1227         int recow;
1228         int err = -ENOENT;
1229
1230         key.objectid = bytenr;
1231         if (parent) {
1232                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1233                 key.offset = parent;
1234         } else {
1235                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1236                 key.offset = hash_extent_data_ref(root_objectid,
1237                                                   owner, offset);
1238         }
1239 again:
1240         recow = 0;
1241         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1242         if (ret < 0) {
1243                 err = ret;
1244                 goto fail;
1245         }
1246
1247         if (parent) {
1248                 if (!ret)
1249                         return 0;
1250 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1251                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1252                 btrfs_release_path(path);
1253                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1254                 if (ret < 0) {
1255                         err = ret;
1256                         goto fail;
1257                 }
1258                 if (!ret)
1259                         return 0;
1260 #endif
1261                 goto fail;
1262         }
1263
1264         leaf = path->nodes[0];
1265         nritems = btrfs_header_nritems(leaf);
1266         while (1) {
1267                 if (path->slots[0] >= nritems) {
1268                         ret = btrfs_next_leaf(root, path);
1269                         if (ret < 0)
1270                                 err = ret;
1271                         if (ret)
1272                                 goto fail;
1273
1274                         leaf = path->nodes[0];
1275                         nritems = btrfs_header_nritems(leaf);
1276                         recow = 1;
1277                 }
1278
1279                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1280                 if (key.objectid != bytenr ||
1281                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1282                         goto fail;
1283
1284                 ref = btrfs_item_ptr(leaf, path->slots[0],
1285                                      struct btrfs_extent_data_ref);
1286
1287                 if (match_extent_data_ref(leaf, ref, root_objectid,
1288                                           owner, offset)) {
1289                         if (recow) {
1290                                 btrfs_release_path(path);
1291                                 goto again;
1292                         }
1293                         err = 0;
1294                         break;
1295                 }
1296                 path->slots[0]++;
1297         }
1298 fail:
1299         return err;
1300 }
1301
1302 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1303                                            struct btrfs_path *path,
1304                                            u64 bytenr, u64 parent,
1305                                            u64 root_objectid, u64 owner,
1306                                            u64 offset, int refs_to_add)
1307 {
1308         struct btrfs_root *root = trans->fs_info->extent_root;
1309         struct btrfs_key key;
1310         struct extent_buffer *leaf;
1311         u32 size;
1312         u32 num_refs;
1313         int ret;
1314
1315         key.objectid = bytenr;
1316         if (parent) {
1317                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1318                 key.offset = parent;
1319                 size = sizeof(struct btrfs_shared_data_ref);
1320         } else {
1321                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1322                 key.offset = hash_extent_data_ref(root_objectid,
1323                                                   owner, offset);
1324                 size = sizeof(struct btrfs_extent_data_ref);
1325         }
1326
1327         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1328         if (ret && ret != -EEXIST)
1329                 goto fail;
1330
1331         leaf = path->nodes[0];
1332         if (parent) {
1333                 struct btrfs_shared_data_ref *ref;
1334                 ref = btrfs_item_ptr(leaf, path->slots[0],
1335                                      struct btrfs_shared_data_ref);
1336                 if (ret == 0) {
1337                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1338                 } else {
1339                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1340                         num_refs += refs_to_add;
1341                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1342                 }
1343         } else {
1344                 struct btrfs_extent_data_ref *ref;
1345                 while (ret == -EEXIST) {
1346                         ref = btrfs_item_ptr(leaf, path->slots[0],
1347                                              struct btrfs_extent_data_ref);
1348                         if (match_extent_data_ref(leaf, ref, root_objectid,
1349                                                   owner, offset))
1350                                 break;
1351                         btrfs_release_path(path);
1352                         key.offset++;
1353                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1354                                                       size);
1355                         if (ret && ret != -EEXIST)
1356                                 goto fail;
1357
1358                         leaf = path->nodes[0];
1359                 }
1360                 ref = btrfs_item_ptr(leaf, path->slots[0],
1361                                      struct btrfs_extent_data_ref);
1362                 if (ret == 0) {
1363                         btrfs_set_extent_data_ref_root(leaf, ref,
1364                                                        root_objectid);
1365                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1366                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1367                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1368                 } else {
1369                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1370                         num_refs += refs_to_add;
1371                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1372                 }
1373         }
1374         btrfs_mark_buffer_dirty(leaf);
1375         ret = 0;
1376 fail:
1377         btrfs_release_path(path);
1378         return ret;
1379 }
1380
1381 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1382                                            struct btrfs_path *path,
1383                                            int refs_to_drop, int *last_ref)
1384 {
1385         struct btrfs_key key;
1386         struct btrfs_extent_data_ref *ref1 = NULL;
1387         struct btrfs_shared_data_ref *ref2 = NULL;
1388         struct extent_buffer *leaf;
1389         u32 num_refs = 0;
1390         int ret = 0;
1391
1392         leaf = path->nodes[0];
1393         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1394
1395         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1396                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1397                                       struct btrfs_extent_data_ref);
1398                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1399         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1400                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1401                                       struct btrfs_shared_data_ref);
1402                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1403 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1404         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1405                 struct btrfs_extent_ref_v0 *ref0;
1406                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1407                                       struct btrfs_extent_ref_v0);
1408                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1409 #endif
1410         } else {
1411                 BUG();
1412         }
1413
1414         BUG_ON(num_refs < refs_to_drop);
1415         num_refs -= refs_to_drop;
1416
1417         if (num_refs == 0) {
1418                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1419                 *last_ref = 1;
1420         } else {
1421                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1422                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1423                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1424                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1425 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1426                 else {
1427                         struct btrfs_extent_ref_v0 *ref0;
1428                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1429                                         struct btrfs_extent_ref_v0);
1430                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1431                 }
1432 #endif
1433                 btrfs_mark_buffer_dirty(leaf);
1434         }
1435         return ret;
1436 }
1437
1438 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1439                                           struct btrfs_extent_inline_ref *iref)
1440 {
1441         struct btrfs_key key;
1442         struct extent_buffer *leaf;
1443         struct btrfs_extent_data_ref *ref1;
1444         struct btrfs_shared_data_ref *ref2;
1445         u32 num_refs = 0;
1446         int type;
1447
1448         leaf = path->nodes[0];
1449         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1450         if (iref) {
1451                 /*
1452                  * If type is invalid, we should have bailed out earlier than
1453                  * this call.
1454                  */
1455                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1456                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1457                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1458                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1459                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1460                 } else {
1461                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1462                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1463                 }
1464         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1465                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1466                                       struct btrfs_extent_data_ref);
1467                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1468         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1469                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1470                                       struct btrfs_shared_data_ref);
1471                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1472 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1473         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1474                 struct btrfs_extent_ref_v0 *ref0;
1475                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1476                                       struct btrfs_extent_ref_v0);
1477                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1478 #endif
1479         } else {
1480                 WARN_ON(1);
1481         }
1482         return num_refs;
1483 }
1484
1485 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1486                                           struct btrfs_path *path,
1487                                           u64 bytenr, u64 parent,
1488                                           u64 root_objectid)
1489 {
1490         struct btrfs_root *root = trans->fs_info->extent_root;
1491         struct btrfs_key key;
1492         int ret;
1493
1494         key.objectid = bytenr;
1495         if (parent) {
1496                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1497                 key.offset = parent;
1498         } else {
1499                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1500                 key.offset = root_objectid;
1501         }
1502
1503         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1504         if (ret > 0)
1505                 ret = -ENOENT;
1506 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1507         if (ret == -ENOENT && parent) {
1508                 btrfs_release_path(path);
1509                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1510                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1511                 if (ret > 0)
1512                         ret = -ENOENT;
1513         }
1514 #endif
1515         return ret;
1516 }
1517
1518 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1519                                           struct btrfs_path *path,
1520                                           u64 bytenr, u64 parent,
1521                                           u64 root_objectid)
1522 {
1523         struct btrfs_key key;
1524         int ret;
1525
1526         key.objectid = bytenr;
1527         if (parent) {
1528                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1529                 key.offset = parent;
1530         } else {
1531                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1532                 key.offset = root_objectid;
1533         }
1534
1535         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1536                                       path, &key, 0);
1537         btrfs_release_path(path);
1538         return ret;
1539 }
1540
1541 static inline int extent_ref_type(u64 parent, u64 owner)
1542 {
1543         int type;
1544         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1545                 if (parent > 0)
1546                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1547                 else
1548                         type = BTRFS_TREE_BLOCK_REF_KEY;
1549         } else {
1550                 if (parent > 0)
1551                         type = BTRFS_SHARED_DATA_REF_KEY;
1552                 else
1553                         type = BTRFS_EXTENT_DATA_REF_KEY;
1554         }
1555         return type;
1556 }
1557
1558 static int find_next_key(struct btrfs_path *path, int level,
1559                          struct btrfs_key *key)
1560
1561 {
1562         for (; level < BTRFS_MAX_LEVEL; level++) {
1563                 if (!path->nodes[level])
1564                         break;
1565                 if (path->slots[level] + 1 >=
1566                     btrfs_header_nritems(path->nodes[level]))
1567                         continue;
1568                 if (level == 0)
1569                         btrfs_item_key_to_cpu(path->nodes[level], key,
1570                                               path->slots[level] + 1);
1571                 else
1572                         btrfs_node_key_to_cpu(path->nodes[level], key,
1573                                               path->slots[level] + 1);
1574                 return 0;
1575         }
1576         return 1;
1577 }
1578
1579 /*
1580  * look for inline back ref. if back ref is found, *ref_ret is set
1581  * to the address of inline back ref, and 0 is returned.
1582  *
1583  * if back ref isn't found, *ref_ret is set to the address where it
1584  * should be inserted, and -ENOENT is returned.
1585  *
1586  * if insert is true and there are too many inline back refs, the path
1587  * points to the extent item, and -EAGAIN is returned.
1588  *
1589  * NOTE: inline back refs are ordered in the same way that back ref
1590  *       items in the tree are ordered.
1591  */
1592 static noinline_for_stack
1593 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1594                                  struct btrfs_path *path,
1595                                  struct btrfs_extent_inline_ref **ref_ret,
1596                                  u64 bytenr, u64 num_bytes,
1597                                  u64 parent, u64 root_objectid,
1598                                  u64 owner, u64 offset, int insert)
1599 {
1600         struct btrfs_fs_info *fs_info = trans->fs_info;
1601         struct btrfs_root *root = fs_info->extent_root;
1602         struct btrfs_key key;
1603         struct extent_buffer *leaf;
1604         struct btrfs_extent_item *ei;
1605         struct btrfs_extent_inline_ref *iref;
1606         u64 flags;
1607         u64 item_size;
1608         unsigned long ptr;
1609         unsigned long end;
1610         int extra_size;
1611         int type;
1612         int want;
1613         int ret;
1614         int err = 0;
1615         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1616         int needed;
1617
1618         key.objectid = bytenr;
1619         key.type = BTRFS_EXTENT_ITEM_KEY;
1620         key.offset = num_bytes;
1621
1622         want = extent_ref_type(parent, owner);
1623         if (insert) {
1624                 extra_size = btrfs_extent_inline_ref_size(want);
1625                 path->keep_locks = 1;
1626         } else
1627                 extra_size = -1;
1628
1629         /*
1630          * Owner is our level, so we can just add one to get the level for the
1631          * block we are interested in.
1632          */
1633         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1634                 key.type = BTRFS_METADATA_ITEM_KEY;
1635                 key.offset = owner;
1636         }
1637
1638 again:
1639         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1640         if (ret < 0) {
1641                 err = ret;
1642                 goto out;
1643         }
1644
1645         /*
1646          * We may be a newly converted file system which still has the old fat
1647          * extent entries for metadata, so try and see if we have one of those.
1648          */
1649         if (ret > 0 && skinny_metadata) {
1650                 skinny_metadata = false;
1651                 if (path->slots[0]) {
1652                         path->slots[0]--;
1653                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1654                                               path->slots[0]);
1655                         if (key.objectid == bytenr &&
1656                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1657                             key.offset == num_bytes)
1658                                 ret = 0;
1659                 }
1660                 if (ret) {
1661                         key.objectid = bytenr;
1662                         key.type = BTRFS_EXTENT_ITEM_KEY;
1663                         key.offset = num_bytes;
1664                         btrfs_release_path(path);
1665                         goto again;
1666                 }
1667         }
1668
1669         if (ret && !insert) {
1670                 err = -ENOENT;
1671                 goto out;
1672         } else if (WARN_ON(ret)) {
1673                 err = -EIO;
1674                 goto out;
1675         }
1676
1677         leaf = path->nodes[0];
1678         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1679 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1680         if (item_size < sizeof(*ei)) {
1681                 if (!insert) {
1682                         err = -ENOENT;
1683                         goto out;
1684                 }
1685                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1686                                              extra_size);
1687                 if (ret < 0) {
1688                         err = ret;
1689                         goto out;
1690                 }
1691                 leaf = path->nodes[0];
1692                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1693         }
1694 #endif
1695         BUG_ON(item_size < sizeof(*ei));
1696
1697         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1698         flags = btrfs_extent_flags(leaf, ei);
1699
1700         ptr = (unsigned long)(ei + 1);
1701         end = (unsigned long)ei + item_size;
1702
1703         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1704                 ptr += sizeof(struct btrfs_tree_block_info);
1705                 BUG_ON(ptr > end);
1706         }
1707
1708         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1709                 needed = BTRFS_REF_TYPE_DATA;
1710         else
1711                 needed = BTRFS_REF_TYPE_BLOCK;
1712
1713         err = -ENOENT;
1714         while (1) {
1715                 if (ptr >= end) {
1716                         WARN_ON(ptr > end);
1717                         break;
1718                 }
1719                 iref = (struct btrfs_extent_inline_ref *)ptr;
1720                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1721                 if (type == BTRFS_REF_TYPE_INVALID) {
1722                         err = -EINVAL;
1723                         goto out;
1724                 }
1725
1726                 if (want < type)
1727                         break;
1728                 if (want > type) {
1729                         ptr += btrfs_extent_inline_ref_size(type);
1730                         continue;
1731                 }
1732
1733                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1734                         struct btrfs_extent_data_ref *dref;
1735                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1736                         if (match_extent_data_ref(leaf, dref, root_objectid,
1737                                                   owner, offset)) {
1738                                 err = 0;
1739                                 break;
1740                         }
1741                         if (hash_extent_data_ref_item(leaf, dref) <
1742                             hash_extent_data_ref(root_objectid, owner, offset))
1743                                 break;
1744                 } else {
1745                         u64 ref_offset;
1746                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1747                         if (parent > 0) {
1748                                 if (parent == ref_offset) {
1749                                         err = 0;
1750                                         break;
1751                                 }
1752                                 if (ref_offset < parent)
1753                                         break;
1754                         } else {
1755                                 if (root_objectid == ref_offset) {
1756                                         err = 0;
1757                                         break;
1758                                 }
1759                                 if (ref_offset < root_objectid)
1760                                         break;
1761                         }
1762                 }
1763                 ptr += btrfs_extent_inline_ref_size(type);
1764         }
1765         if (err == -ENOENT && insert) {
1766                 if (item_size + extra_size >=
1767                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1768                         err = -EAGAIN;
1769                         goto out;
1770                 }
1771                 /*
1772                  * To add new inline back ref, we have to make sure
1773                  * there is no corresponding back ref item.
1774                  * For simplicity, we just do not add new inline back
1775                  * ref if there is any kind of item for this block
1776                  */
1777                 if (find_next_key(path, 0, &key) == 0 &&
1778                     key.objectid == bytenr &&
1779                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1780                         err = -EAGAIN;
1781                         goto out;
1782                 }
1783         }
1784         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1785 out:
1786         if (insert) {
1787                 path->keep_locks = 0;
1788                 btrfs_unlock_up_safe(path, 1);
1789         }
1790         return err;
1791 }
1792
1793 /*
1794  * helper to add new inline back ref
1795  */
1796 static noinline_for_stack
1797 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1798                                  struct btrfs_path *path,
1799                                  struct btrfs_extent_inline_ref *iref,
1800                                  u64 parent, u64 root_objectid,
1801                                  u64 owner, u64 offset, int refs_to_add,
1802                                  struct btrfs_delayed_extent_op *extent_op)
1803 {
1804         struct extent_buffer *leaf;
1805         struct btrfs_extent_item *ei;
1806         unsigned long ptr;
1807         unsigned long end;
1808         unsigned long item_offset;
1809         u64 refs;
1810         int size;
1811         int type;
1812
1813         leaf = path->nodes[0];
1814         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1815         item_offset = (unsigned long)iref - (unsigned long)ei;
1816
1817         type = extent_ref_type(parent, owner);
1818         size = btrfs_extent_inline_ref_size(type);
1819
1820         btrfs_extend_item(fs_info, path, size);
1821
1822         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1823         refs = btrfs_extent_refs(leaf, ei);
1824         refs += refs_to_add;
1825         btrfs_set_extent_refs(leaf, ei, refs);
1826         if (extent_op)
1827                 __run_delayed_extent_op(extent_op, leaf, ei);
1828
1829         ptr = (unsigned long)ei + item_offset;
1830         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1831         if (ptr < end - size)
1832                 memmove_extent_buffer(leaf, ptr + size, ptr,
1833                                       end - size - ptr);
1834
1835         iref = (struct btrfs_extent_inline_ref *)ptr;
1836         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1837         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1838                 struct btrfs_extent_data_ref *dref;
1839                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1840                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1841                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1842                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1843                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1844         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1845                 struct btrfs_shared_data_ref *sref;
1846                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1847                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1848                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1849         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1850                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1851         } else {
1852                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1853         }
1854         btrfs_mark_buffer_dirty(leaf);
1855 }
1856
1857 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1858                                  struct btrfs_path *path,
1859                                  struct btrfs_extent_inline_ref **ref_ret,
1860                                  u64 bytenr, u64 num_bytes, u64 parent,
1861                                  u64 root_objectid, u64 owner, u64 offset)
1862 {
1863         int ret;
1864
1865         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1866                                            num_bytes, parent, root_objectid,
1867                                            owner, offset, 0);
1868         if (ret != -ENOENT)
1869                 return ret;
1870
1871         btrfs_release_path(path);
1872         *ref_ret = NULL;
1873
1874         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1875                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1876                                             root_objectid);
1877         } else {
1878                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1879                                              root_objectid, owner, offset);
1880         }
1881         return ret;
1882 }
1883
1884 /*
1885  * helper to update/remove inline back ref
1886  */
1887 static noinline_for_stack
1888 void update_inline_extent_backref(struct btrfs_path *path,
1889                                   struct btrfs_extent_inline_ref *iref,
1890                                   int refs_to_mod,
1891                                   struct btrfs_delayed_extent_op *extent_op,
1892                                   int *last_ref)
1893 {
1894         struct extent_buffer *leaf = path->nodes[0];
1895         struct btrfs_fs_info *fs_info = leaf->fs_info;
1896         struct btrfs_extent_item *ei;
1897         struct btrfs_extent_data_ref *dref = NULL;
1898         struct btrfs_shared_data_ref *sref = NULL;
1899         unsigned long ptr;
1900         unsigned long end;
1901         u32 item_size;
1902         int size;
1903         int type;
1904         u64 refs;
1905
1906         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1907         refs = btrfs_extent_refs(leaf, ei);
1908         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1909         refs += refs_to_mod;
1910         btrfs_set_extent_refs(leaf, ei, refs);
1911         if (extent_op)
1912                 __run_delayed_extent_op(extent_op, leaf, ei);
1913
1914         /*
1915          * If type is invalid, we should have bailed out after
1916          * lookup_inline_extent_backref().
1917          */
1918         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1919         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1920
1921         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1922                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1923                 refs = btrfs_extent_data_ref_count(leaf, dref);
1924         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1925                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1926                 refs = btrfs_shared_data_ref_count(leaf, sref);
1927         } else {
1928                 refs = 1;
1929                 BUG_ON(refs_to_mod != -1);
1930         }
1931
1932         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1933         refs += refs_to_mod;
1934
1935         if (refs > 0) {
1936                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1937                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1938                 else
1939                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1940         } else {
1941                 *last_ref = 1;
1942                 size =  btrfs_extent_inline_ref_size(type);
1943                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1944                 ptr = (unsigned long)iref;
1945                 end = (unsigned long)ei + item_size;
1946                 if (ptr + size < end)
1947                         memmove_extent_buffer(leaf, ptr, ptr + size,
1948                                               end - ptr - size);
1949                 item_size -= size;
1950                 btrfs_truncate_item(fs_info, path, item_size, 1);
1951         }
1952         btrfs_mark_buffer_dirty(leaf);
1953 }
1954
1955 static noinline_for_stack
1956 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1957                                  struct btrfs_path *path,
1958                                  u64 bytenr, u64 num_bytes, u64 parent,
1959                                  u64 root_objectid, u64 owner,
1960                                  u64 offset, int refs_to_add,
1961                                  struct btrfs_delayed_extent_op *extent_op)
1962 {
1963         struct btrfs_extent_inline_ref *iref;
1964         int ret;
1965
1966         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1967                                            num_bytes, parent, root_objectid,
1968                                            owner, offset, 1);
1969         if (ret == 0) {
1970                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1971                 update_inline_extent_backref(path, iref, refs_to_add,
1972                                              extent_op, NULL);
1973         } else if (ret == -ENOENT) {
1974                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1975                                             root_objectid, owner, offset,
1976                                             refs_to_add, extent_op);
1977                 ret = 0;
1978         }
1979         return ret;
1980 }
1981
1982 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1983                                  struct btrfs_path *path,
1984                                  u64 bytenr, u64 parent, u64 root_objectid,
1985                                  u64 owner, u64 offset, int refs_to_add)
1986 {
1987         int ret;
1988         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1989                 BUG_ON(refs_to_add != 1);
1990                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1991                                             root_objectid);
1992         } else {
1993                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1994                                              root_objectid, owner, offset,
1995                                              refs_to_add);
1996         }
1997         return ret;
1998 }
1999
2000 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2001                                  struct btrfs_path *path,
2002                                  struct btrfs_extent_inline_ref *iref,
2003                                  int refs_to_drop, int is_data, int *last_ref)
2004 {
2005         int ret = 0;
2006
2007         BUG_ON(!is_data && refs_to_drop != 1);
2008         if (iref) {
2009                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
2010                                              last_ref);
2011         } else if (is_data) {
2012                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
2013                                              last_ref);
2014         } else {
2015                 *last_ref = 1;
2016                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
2017         }
2018         return ret;
2019 }
2020
2021 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2022 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2023                                u64 *discarded_bytes)
2024 {
2025         int j, ret = 0;
2026         u64 bytes_left, end;
2027         u64 aligned_start = ALIGN(start, 1 << 9);
2028
2029         if (WARN_ON(start != aligned_start)) {
2030                 len -= aligned_start - start;
2031                 len = round_down(len, 1 << 9);
2032                 start = aligned_start;
2033         }
2034
2035         *discarded_bytes = 0;
2036
2037         if (!len)
2038                 return 0;
2039
2040         end = start + len;
2041         bytes_left = len;
2042
2043         /* Skip any superblocks on this device. */
2044         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2045                 u64 sb_start = btrfs_sb_offset(j);
2046                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2047                 u64 size = sb_start - start;
2048
2049                 if (!in_range(sb_start, start, bytes_left) &&
2050                     !in_range(sb_end, start, bytes_left) &&
2051                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2052                         continue;
2053
2054                 /*
2055                  * Superblock spans beginning of range.  Adjust start and
2056                  * try again.
2057                  */
2058                 if (sb_start <= start) {
2059                         start += sb_end - start;
2060                         if (start > end) {
2061                                 bytes_left = 0;
2062                                 break;
2063                         }
2064                         bytes_left = end - start;
2065                         continue;
2066                 }
2067
2068                 if (size) {
2069                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2070                                                    GFP_NOFS, 0);
2071                         if (!ret)
2072                                 *discarded_bytes += size;
2073                         else if (ret != -EOPNOTSUPP)
2074                                 return ret;
2075                 }
2076
2077                 start = sb_end;
2078                 if (start > end) {
2079                         bytes_left = 0;
2080                         break;
2081                 }
2082                 bytes_left = end - start;
2083         }
2084
2085         if (bytes_left) {
2086                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2087                                            GFP_NOFS, 0);
2088                 if (!ret)
2089                         *discarded_bytes += bytes_left;
2090         }
2091         return ret;
2092 }
2093
2094 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2095                          u64 num_bytes, u64 *actual_bytes)
2096 {
2097         int ret;
2098         u64 discarded_bytes = 0;
2099         struct btrfs_bio *bbio = NULL;
2100
2101
2102         /*
2103          * Avoid races with device replace and make sure our bbio has devices
2104          * associated to its stripes that don't go away while we are discarding.
2105          */
2106         btrfs_bio_counter_inc_blocked(fs_info);
2107         /* Tell the block device(s) that the sectors can be discarded */
2108         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2109                               &bbio, 0);
2110         /* Error condition is -ENOMEM */
2111         if (!ret) {
2112                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2113                 int i;
2114
2115
2116                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2117                         u64 bytes;
2118                         struct request_queue *req_q;
2119
2120                         if (!stripe->dev->bdev) {
2121                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2122                                 continue;
2123                         }
2124                         req_q = bdev_get_queue(stripe->dev->bdev);
2125                         if (!blk_queue_discard(req_q))
2126                                 continue;
2127
2128                         ret = btrfs_issue_discard(stripe->dev->bdev,
2129                                                   stripe->physical,
2130                                                   stripe->length,
2131                                                   &bytes);
2132                         if (!ret)
2133                                 discarded_bytes += bytes;
2134                         else if (ret != -EOPNOTSUPP)
2135                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2136
2137                         /*
2138                          * Just in case we get back EOPNOTSUPP for some reason,
2139                          * just ignore the return value so we don't screw up
2140                          * people calling discard_extent.
2141                          */
2142                         ret = 0;
2143                 }
2144                 btrfs_put_bbio(bbio);
2145         }
2146         btrfs_bio_counter_dec(fs_info);
2147
2148         if (actual_bytes)
2149                 *actual_bytes = discarded_bytes;
2150
2151
2152         if (ret == -EOPNOTSUPP)
2153                 ret = 0;
2154         return ret;
2155 }
2156
2157 /* Can return -ENOMEM */
2158 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2159                          struct btrfs_root *root,
2160                          u64 bytenr, u64 num_bytes, u64 parent,
2161                          u64 root_objectid, u64 owner, u64 offset)
2162 {
2163         struct btrfs_fs_info *fs_info = root->fs_info;
2164         int old_ref_mod, new_ref_mod;
2165         int ret;
2166
2167         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2168                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2169
2170         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2171                            owner, offset, BTRFS_ADD_DELAYED_REF);
2172
2173         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2174                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2175                                                  num_bytes, parent,
2176                                                  root_objectid, (int)owner,
2177                                                  BTRFS_ADD_DELAYED_REF, NULL,
2178                                                  &old_ref_mod, &new_ref_mod);
2179         } else {
2180                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2181                                                  num_bytes, parent,
2182                                                  root_objectid, owner, offset,
2183                                                  0, BTRFS_ADD_DELAYED_REF,
2184                                                  &old_ref_mod, &new_ref_mod);
2185         }
2186
2187         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2188                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2189
2190                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2191         }
2192
2193         return ret;
2194 }
2195
2196 /*
2197  * __btrfs_inc_extent_ref - insert backreference for a given extent
2198  *
2199  * @trans:          Handle of transaction
2200  *
2201  * @node:           The delayed ref node used to get the bytenr/length for
2202  *                  extent whose references are incremented.
2203  *
2204  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2205  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2206  *                  bytenr of the parent block. Since new extents are always
2207  *                  created with indirect references, this will only be the case
2208  *                  when relocating a shared extent. In that case, root_objectid
2209  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2210  *                  be 0
2211  *
2212  * @root_objectid:  The id of the root where this modification has originated,
2213  *                  this can be either one of the well-known metadata trees or
2214  *                  the subvolume id which references this extent.
2215  *
2216  * @owner:          For data extents it is the inode number of the owning file.
2217  *                  For metadata extents this parameter holds the level in the
2218  *                  tree of the extent.
2219  *
2220  * @offset:         For metadata extents the offset is ignored and is currently
2221  *                  always passed as 0. For data extents it is the fileoffset
2222  *                  this extent belongs to.
2223  *
2224  * @refs_to_add     Number of references to add
2225  *
2226  * @extent_op       Pointer to a structure, holding information necessary when
2227  *                  updating a tree block's flags
2228  *
2229  */
2230 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2231                                   struct btrfs_delayed_ref_node *node,
2232                                   u64 parent, u64 root_objectid,
2233                                   u64 owner, u64 offset, int refs_to_add,
2234                                   struct btrfs_delayed_extent_op *extent_op)
2235 {
2236         struct btrfs_path *path;
2237         struct extent_buffer *leaf;
2238         struct btrfs_extent_item *item;
2239         struct btrfs_key key;
2240         u64 bytenr = node->bytenr;
2241         u64 num_bytes = node->num_bytes;
2242         u64 refs;
2243         int ret;
2244
2245         path = btrfs_alloc_path();
2246         if (!path)
2247                 return -ENOMEM;
2248
2249         path->reada = READA_FORWARD;
2250         path->leave_spinning = 1;
2251         /* this will setup the path even if it fails to insert the back ref */
2252         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2253                                            parent, root_objectid, owner,
2254                                            offset, refs_to_add, extent_op);
2255         if ((ret < 0 && ret != -EAGAIN) || !ret)
2256                 goto out;
2257
2258         /*
2259          * Ok we had -EAGAIN which means we didn't have space to insert and
2260          * inline extent ref, so just update the reference count and add a
2261          * normal backref.
2262          */
2263         leaf = path->nodes[0];
2264         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2265         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2266         refs = btrfs_extent_refs(leaf, item);
2267         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2268         if (extent_op)
2269                 __run_delayed_extent_op(extent_op, leaf, item);
2270
2271         btrfs_mark_buffer_dirty(leaf);
2272         btrfs_release_path(path);
2273
2274         path->reada = READA_FORWARD;
2275         path->leave_spinning = 1;
2276         /* now insert the actual backref */
2277         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2278                                     owner, offset, refs_to_add);
2279         if (ret)
2280                 btrfs_abort_transaction(trans, ret);
2281 out:
2282         btrfs_free_path(path);
2283         return ret;
2284 }
2285
2286 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2287                                 struct btrfs_delayed_ref_node *node,
2288                                 struct btrfs_delayed_extent_op *extent_op,
2289                                 int insert_reserved)
2290 {
2291         int ret = 0;
2292         struct btrfs_delayed_data_ref *ref;
2293         struct btrfs_key ins;
2294         u64 parent = 0;
2295         u64 ref_root = 0;
2296         u64 flags = 0;
2297
2298         ins.objectid = node->bytenr;
2299         ins.offset = node->num_bytes;
2300         ins.type = BTRFS_EXTENT_ITEM_KEY;
2301
2302         ref = btrfs_delayed_node_to_data_ref(node);
2303         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2304
2305         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2306                 parent = ref->parent;
2307         ref_root = ref->root;
2308
2309         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2310                 if (extent_op)
2311                         flags |= extent_op->flags_to_set;
2312                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2313                                                  flags, ref->objectid,
2314                                                  ref->offset, &ins,
2315                                                  node->ref_mod);
2316         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2317                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2318                                              ref->objectid, ref->offset,
2319                                              node->ref_mod, extent_op);
2320         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2321                 ret = __btrfs_free_extent(trans, node, parent,
2322                                           ref_root, ref->objectid,
2323                                           ref->offset, node->ref_mod,
2324                                           extent_op);
2325         } else {
2326                 BUG();
2327         }
2328         return ret;
2329 }
2330
2331 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2332                                     struct extent_buffer *leaf,
2333                                     struct btrfs_extent_item *ei)
2334 {
2335         u64 flags = btrfs_extent_flags(leaf, ei);
2336         if (extent_op->update_flags) {
2337                 flags |= extent_op->flags_to_set;
2338                 btrfs_set_extent_flags(leaf, ei, flags);
2339         }
2340
2341         if (extent_op->update_key) {
2342                 struct btrfs_tree_block_info *bi;
2343                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2344                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2345                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2346         }
2347 }
2348
2349 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2350                                  struct btrfs_delayed_ref_head *head,
2351                                  struct btrfs_delayed_extent_op *extent_op)
2352 {
2353         struct btrfs_fs_info *fs_info = trans->fs_info;
2354         struct btrfs_key key;
2355         struct btrfs_path *path;
2356         struct btrfs_extent_item *ei;
2357         struct extent_buffer *leaf;
2358         u32 item_size;
2359         int ret;
2360         int err = 0;
2361         int metadata = !extent_op->is_data;
2362
2363         if (trans->aborted)
2364                 return 0;
2365
2366         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2367                 metadata = 0;
2368
2369         path = btrfs_alloc_path();
2370         if (!path)
2371                 return -ENOMEM;
2372
2373         key.objectid = head->bytenr;
2374
2375         if (metadata) {
2376                 key.type = BTRFS_METADATA_ITEM_KEY;
2377                 key.offset = extent_op->level;
2378         } else {
2379                 key.type = BTRFS_EXTENT_ITEM_KEY;
2380                 key.offset = head->num_bytes;
2381         }
2382
2383 again:
2384         path->reada = READA_FORWARD;
2385         path->leave_spinning = 1;
2386         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2387         if (ret < 0) {
2388                 err = ret;
2389                 goto out;
2390         }
2391         if (ret > 0) {
2392                 if (metadata) {
2393                         if (path->slots[0] > 0) {
2394                                 path->slots[0]--;
2395                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2396                                                       path->slots[0]);
2397                                 if (key.objectid == head->bytenr &&
2398                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2399                                     key.offset == head->num_bytes)
2400                                         ret = 0;
2401                         }
2402                         if (ret > 0) {
2403                                 btrfs_release_path(path);
2404                                 metadata = 0;
2405
2406                                 key.objectid = head->bytenr;
2407                                 key.offset = head->num_bytes;
2408                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2409                                 goto again;
2410                         }
2411                 } else {
2412                         err = -EIO;
2413                         goto out;
2414                 }
2415         }
2416
2417         leaf = path->nodes[0];
2418         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2419 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2420         if (item_size < sizeof(*ei)) {
2421                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2422                 if (ret < 0) {
2423                         err = ret;
2424                         goto out;
2425                 }
2426                 leaf = path->nodes[0];
2427                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2428         }
2429 #endif
2430         BUG_ON(item_size < sizeof(*ei));
2431         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2432         __run_delayed_extent_op(extent_op, leaf, ei);
2433
2434         btrfs_mark_buffer_dirty(leaf);
2435 out:
2436         btrfs_free_path(path);
2437         return err;
2438 }
2439
2440 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2441                                 struct btrfs_delayed_ref_node *node,
2442                                 struct btrfs_delayed_extent_op *extent_op,
2443                                 int insert_reserved)
2444 {
2445         int ret = 0;
2446         struct btrfs_delayed_tree_ref *ref;
2447         u64 parent = 0;
2448         u64 ref_root = 0;
2449
2450         ref = btrfs_delayed_node_to_tree_ref(node);
2451         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2452
2453         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2454                 parent = ref->parent;
2455         ref_root = ref->root;
2456
2457         if (node->ref_mod != 1) {
2458                 btrfs_err(trans->fs_info,
2459         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2460                           node->bytenr, node->ref_mod, node->action, ref_root,
2461                           parent);
2462                 return -EIO;
2463         }
2464         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2465                 BUG_ON(!extent_op || !extent_op->update_flags);
2466                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2467         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2468                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2469                                              ref->level, 0, 1, extent_op);
2470         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2471                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2472                                           ref->level, 0, 1, extent_op);
2473         } else {
2474                 BUG();
2475         }
2476         return ret;
2477 }
2478
2479 /* helper function to actually process a single delayed ref entry */
2480 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2481                                struct btrfs_delayed_ref_node *node,
2482                                struct btrfs_delayed_extent_op *extent_op,
2483                                int insert_reserved)
2484 {
2485         int ret = 0;
2486
2487         if (trans->aborted) {
2488                 if (insert_reserved)
2489                         btrfs_pin_extent(trans->fs_info, node->bytenr,
2490                                          node->num_bytes, 1);
2491                 return 0;
2492         }
2493
2494         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2495             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2496                 ret = run_delayed_tree_ref(trans, node, extent_op,
2497                                            insert_reserved);
2498         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2499                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2500                 ret = run_delayed_data_ref(trans, node, extent_op,
2501                                            insert_reserved);
2502         else
2503                 BUG();
2504         return ret;
2505 }
2506
2507 static inline struct btrfs_delayed_ref_node *
2508 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2509 {
2510         struct btrfs_delayed_ref_node *ref;
2511
2512         if (RB_EMPTY_ROOT(&head->ref_tree))
2513                 return NULL;
2514
2515         /*
2516          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2517          * This is to prevent a ref count from going down to zero, which deletes
2518          * the extent item from the extent tree, when there still are references
2519          * to add, which would fail because they would not find the extent item.
2520          */
2521         if (!list_empty(&head->ref_add_list))
2522                 return list_first_entry(&head->ref_add_list,
2523                                 struct btrfs_delayed_ref_node, add_list);
2524
2525         ref = rb_entry(rb_first(&head->ref_tree),
2526                        struct btrfs_delayed_ref_node, ref_node);
2527         ASSERT(list_empty(&ref->add_list));
2528         return ref;
2529 }
2530
2531 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2532                                       struct btrfs_delayed_ref_head *head)
2533 {
2534         spin_lock(&delayed_refs->lock);
2535         head->processing = 0;
2536         delayed_refs->num_heads_ready++;
2537         spin_unlock(&delayed_refs->lock);
2538         btrfs_delayed_ref_unlock(head);
2539 }
2540
2541 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2542                              struct btrfs_delayed_ref_head *head)
2543 {
2544         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2545         int ret;
2546
2547         if (!extent_op)
2548                 return 0;
2549         head->extent_op = NULL;
2550         if (head->must_insert_reserved) {
2551                 btrfs_free_delayed_extent_op(extent_op);
2552                 return 0;
2553         }
2554         spin_unlock(&head->lock);
2555         ret = run_delayed_extent_op(trans, head, extent_op);
2556         btrfs_free_delayed_extent_op(extent_op);
2557         return ret ? ret : 1;
2558 }
2559
2560 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2561                             struct btrfs_delayed_ref_head *head)
2562 {
2563
2564         struct btrfs_fs_info *fs_info = trans->fs_info;
2565         struct btrfs_delayed_ref_root *delayed_refs;
2566         int ret;
2567
2568         delayed_refs = &trans->transaction->delayed_refs;
2569
2570         ret = cleanup_extent_op(trans, head);
2571         if (ret < 0) {
2572                 unselect_delayed_ref_head(delayed_refs, head);
2573                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2574                 return ret;
2575         } else if (ret) {
2576                 return ret;
2577         }
2578
2579         /*
2580          * Need to drop our head ref lock and re-acquire the delayed ref lock
2581          * and then re-check to make sure nobody got added.
2582          */
2583         spin_unlock(&head->lock);
2584         spin_lock(&delayed_refs->lock);
2585         spin_lock(&head->lock);
2586         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2587                 spin_unlock(&head->lock);
2588                 spin_unlock(&delayed_refs->lock);
2589                 return 1;
2590         }
2591         delayed_refs->num_heads--;
2592         rb_erase(&head->href_node, &delayed_refs->href_root);
2593         RB_CLEAR_NODE(&head->href_node);
2594         spin_unlock(&head->lock);
2595         spin_unlock(&delayed_refs->lock);
2596         atomic_dec(&delayed_refs->num_entries);
2597
2598         trace_run_delayed_ref_head(fs_info, head, 0);
2599
2600         if (head->total_ref_mod < 0) {
2601                 struct btrfs_space_info *space_info;
2602                 u64 flags;
2603
2604                 if (head->is_data)
2605                         flags = BTRFS_BLOCK_GROUP_DATA;
2606                 else if (head->is_system)
2607                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2608                 else
2609                         flags = BTRFS_BLOCK_GROUP_METADATA;
2610                 space_info = __find_space_info(fs_info, flags);
2611                 ASSERT(space_info);
2612                 percpu_counter_add(&space_info->total_bytes_pinned,
2613                                    -head->num_bytes);
2614
2615                 if (head->is_data) {
2616                         spin_lock(&delayed_refs->lock);
2617                         delayed_refs->pending_csums -= head->num_bytes;
2618                         spin_unlock(&delayed_refs->lock);
2619                 }
2620         }
2621
2622         if (head->must_insert_reserved) {
2623                 btrfs_pin_extent(fs_info, head->bytenr,
2624                                  head->num_bytes, 1);
2625                 if (head->is_data) {
2626                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2627                                               head->num_bytes);
2628                 }
2629         }
2630
2631         /* Also free its reserved qgroup space */
2632         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2633                                       head->qgroup_reserved);
2634         btrfs_delayed_ref_unlock(head);
2635         btrfs_put_delayed_ref_head(head);
2636         return 0;
2637 }
2638
2639 /*
2640  * Returns 0 on success or if called with an already aborted transaction.
2641  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2642  */
2643 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2644                                              unsigned long nr)
2645 {
2646         struct btrfs_fs_info *fs_info = trans->fs_info;
2647         struct btrfs_delayed_ref_root *delayed_refs;
2648         struct btrfs_delayed_ref_node *ref;
2649         struct btrfs_delayed_ref_head *locked_ref = NULL;
2650         struct btrfs_delayed_extent_op *extent_op;
2651         ktime_t start = ktime_get();
2652         int ret;
2653         unsigned long count = 0;
2654         unsigned long actual_count = 0;
2655         int must_insert_reserved = 0;
2656
2657         delayed_refs = &trans->transaction->delayed_refs;
2658         while (1) {
2659                 if (!locked_ref) {
2660                         if (count >= nr)
2661                                 break;
2662
2663                         spin_lock(&delayed_refs->lock);
2664                         locked_ref = btrfs_select_ref_head(trans);
2665                         if (!locked_ref) {
2666                                 spin_unlock(&delayed_refs->lock);
2667                                 break;
2668                         }
2669
2670                         /* grab the lock that says we are going to process
2671                          * all the refs for this head */
2672                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2673                         spin_unlock(&delayed_refs->lock);
2674                         /*
2675                          * we may have dropped the spin lock to get the head
2676                          * mutex lock, and that might have given someone else
2677                          * time to free the head.  If that's true, it has been
2678                          * removed from our list and we can move on.
2679                          */
2680                         if (ret == -EAGAIN) {
2681                                 locked_ref = NULL;
2682                                 count++;
2683                                 continue;
2684                         }
2685                 }
2686
2687                 /*
2688                  * We need to try and merge add/drops of the same ref since we
2689                  * can run into issues with relocate dropping the implicit ref
2690                  * and then it being added back again before the drop can
2691                  * finish.  If we merged anything we need to re-loop so we can
2692                  * get a good ref.
2693                  * Or we can get node references of the same type that weren't
2694                  * merged when created due to bumps in the tree mod seq, and
2695                  * we need to merge them to prevent adding an inline extent
2696                  * backref before dropping it (triggering a BUG_ON at
2697                  * insert_inline_extent_backref()).
2698                  */
2699                 spin_lock(&locked_ref->lock);
2700                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2701
2702                 ref = select_delayed_ref(locked_ref);
2703
2704                 if (ref && ref->seq &&
2705                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2706                         spin_unlock(&locked_ref->lock);
2707                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2708                         locked_ref = NULL;
2709                         cond_resched();
2710                         count++;
2711                         continue;
2712                 }
2713
2714                 /*
2715                  * We're done processing refs in this ref_head, clean everything
2716                  * up and move on to the next ref_head.
2717                  */
2718                 if (!ref) {
2719                         ret = cleanup_ref_head(trans, locked_ref);
2720                         if (ret > 0 ) {
2721                                 /* We dropped our lock, we need to loop. */
2722                                 ret = 0;
2723                                 continue;
2724                         } else if (ret) {
2725                                 return ret;
2726                         }
2727                         locked_ref = NULL;
2728                         count++;
2729                         continue;
2730                 }
2731
2732                 actual_count++;
2733                 ref->in_tree = 0;
2734                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2735                 RB_CLEAR_NODE(&ref->ref_node);
2736                 if (!list_empty(&ref->add_list))
2737                         list_del(&ref->add_list);
2738                 /*
2739                  * When we play the delayed ref, also correct the ref_mod on
2740                  * head
2741                  */
2742                 switch (ref->action) {
2743                 case BTRFS_ADD_DELAYED_REF:
2744                 case BTRFS_ADD_DELAYED_EXTENT:
2745                         locked_ref->ref_mod -= ref->ref_mod;
2746                         break;
2747                 case BTRFS_DROP_DELAYED_REF:
2748                         locked_ref->ref_mod += ref->ref_mod;
2749                         break;
2750                 default:
2751                         WARN_ON(1);
2752                 }
2753                 atomic_dec(&delayed_refs->num_entries);
2754
2755                 /*
2756                  * Record the must-insert_reserved flag before we drop the spin
2757                  * lock.
2758                  */
2759                 must_insert_reserved = locked_ref->must_insert_reserved;
2760                 locked_ref->must_insert_reserved = 0;
2761
2762                 extent_op = locked_ref->extent_op;
2763                 locked_ref->extent_op = NULL;
2764                 spin_unlock(&locked_ref->lock);
2765
2766                 ret = run_one_delayed_ref(trans, ref, extent_op,
2767                                           must_insert_reserved);
2768
2769                 btrfs_free_delayed_extent_op(extent_op);
2770                 if (ret) {
2771                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2772                         btrfs_put_delayed_ref(ref);
2773                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2774                                     ret);
2775                         return ret;
2776                 }
2777
2778                 btrfs_put_delayed_ref(ref);
2779                 count++;
2780                 cond_resched();
2781         }
2782
2783         /*
2784          * We don't want to include ref heads since we can have empty ref heads
2785          * and those will drastically skew our runtime down since we just do
2786          * accounting, no actual extent tree updates.
2787          */
2788         if (actual_count > 0) {
2789                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2790                 u64 avg;
2791
2792                 /*
2793                  * We weigh the current average higher than our current runtime
2794                  * to avoid large swings in the average.
2795                  */
2796                 spin_lock(&delayed_refs->lock);
2797                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2798                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2799                 spin_unlock(&delayed_refs->lock);
2800         }
2801         return 0;
2802 }
2803
2804 #ifdef SCRAMBLE_DELAYED_REFS
2805 /*
2806  * Normally delayed refs get processed in ascending bytenr order. This
2807  * correlates in most cases to the order added. To expose dependencies on this
2808  * order, we start to process the tree in the middle instead of the beginning
2809  */
2810 static u64 find_middle(struct rb_root *root)
2811 {
2812         struct rb_node *n = root->rb_node;
2813         struct btrfs_delayed_ref_node *entry;
2814         int alt = 1;
2815         u64 middle;
2816         u64 first = 0, last = 0;
2817
2818         n = rb_first(root);
2819         if (n) {
2820                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2821                 first = entry->bytenr;
2822         }
2823         n = rb_last(root);
2824         if (n) {
2825                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2826                 last = entry->bytenr;
2827         }
2828         n = root->rb_node;
2829
2830         while (n) {
2831                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2832                 WARN_ON(!entry->in_tree);
2833
2834                 middle = entry->bytenr;
2835
2836                 if (alt)
2837                         n = n->rb_left;
2838                 else
2839                         n = n->rb_right;
2840
2841                 alt = 1 - alt;
2842         }
2843         return middle;
2844 }
2845 #endif
2846
2847 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2848 {
2849         u64 num_bytes;
2850
2851         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2852                              sizeof(struct btrfs_extent_inline_ref));
2853         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2854                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2855
2856         /*
2857          * We don't ever fill up leaves all the way so multiply by 2 just to be
2858          * closer to what we're really going to want to use.
2859          */
2860         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2861 }
2862
2863 /*
2864  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2865  * would require to store the csums for that many bytes.
2866  */
2867 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2868 {
2869         u64 csum_size;
2870         u64 num_csums_per_leaf;
2871         u64 num_csums;
2872
2873         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2874         num_csums_per_leaf = div64_u64(csum_size,
2875                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2876         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2877         num_csums += num_csums_per_leaf - 1;
2878         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2879         return num_csums;
2880 }
2881
2882 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2883                                        struct btrfs_fs_info *fs_info)
2884 {
2885         struct btrfs_block_rsv *global_rsv;
2886         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2887         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2888         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2889         u64 num_bytes, num_dirty_bgs_bytes;
2890         int ret = 0;
2891
2892         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2893         num_heads = heads_to_leaves(fs_info, num_heads);
2894         if (num_heads > 1)
2895                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2896         num_bytes <<= 1;
2897         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2898                                                         fs_info->nodesize;
2899         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2900                                                              num_dirty_bgs);
2901         global_rsv = &fs_info->global_block_rsv;
2902
2903         /*
2904          * If we can't allocate any more chunks lets make sure we have _lots_ of
2905          * wiggle room since running delayed refs can create more delayed refs.
2906          */
2907         if (global_rsv->space_info->full) {
2908                 num_dirty_bgs_bytes <<= 1;
2909                 num_bytes <<= 1;
2910         }
2911
2912         spin_lock(&global_rsv->lock);
2913         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2914                 ret = 1;
2915         spin_unlock(&global_rsv->lock);
2916         return ret;
2917 }
2918
2919 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2920                                        struct btrfs_fs_info *fs_info)
2921 {
2922         u64 num_entries =
2923                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2924         u64 avg_runtime;
2925         u64 val;
2926
2927         smp_mb();
2928         avg_runtime = fs_info->avg_delayed_ref_runtime;
2929         val = num_entries * avg_runtime;
2930         if (val >= NSEC_PER_SEC)
2931                 return 1;
2932         if (val >= NSEC_PER_SEC / 2)
2933                 return 2;
2934
2935         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2936 }
2937
2938 struct async_delayed_refs {
2939         struct btrfs_root *root;
2940         u64 transid;
2941         int count;
2942         int error;
2943         int sync;
2944         struct completion wait;
2945         struct btrfs_work work;
2946 };
2947
2948 static inline struct async_delayed_refs *
2949 to_async_delayed_refs(struct btrfs_work *work)
2950 {
2951         return container_of(work, struct async_delayed_refs, work);
2952 }
2953
2954 static void delayed_ref_async_start(struct btrfs_work *work)
2955 {
2956         struct async_delayed_refs *async = to_async_delayed_refs(work);
2957         struct btrfs_trans_handle *trans;
2958         struct btrfs_fs_info *fs_info = async->root->fs_info;
2959         int ret;
2960
2961         /* if the commit is already started, we don't need to wait here */
2962         if (btrfs_transaction_blocked(fs_info))
2963                 goto done;
2964
2965         trans = btrfs_join_transaction(async->root);
2966         if (IS_ERR(trans)) {
2967                 async->error = PTR_ERR(trans);
2968                 goto done;
2969         }
2970
2971         /*
2972          * trans->sync means that when we call end_transaction, we won't
2973          * wait on delayed refs
2974          */
2975         trans->sync = true;
2976
2977         /* Don't bother flushing if we got into a different transaction */
2978         if (trans->transid > async->transid)
2979                 goto end;
2980
2981         ret = btrfs_run_delayed_refs(trans, async->count);
2982         if (ret)
2983                 async->error = ret;
2984 end:
2985         ret = btrfs_end_transaction(trans);
2986         if (ret && !async->error)
2987                 async->error = ret;
2988 done:
2989         if (async->sync)
2990                 complete(&async->wait);
2991         else
2992                 kfree(async);
2993 }
2994
2995 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2996                                  unsigned long count, u64 transid, int wait)
2997 {
2998         struct async_delayed_refs *async;
2999         int ret;
3000
3001         async = kmalloc(sizeof(*async), GFP_NOFS);
3002         if (!async)
3003                 return -ENOMEM;
3004
3005         async->root = fs_info->tree_root;
3006         async->count = count;
3007         async->error = 0;
3008         async->transid = transid;
3009         if (wait)
3010                 async->sync = 1;
3011         else
3012                 async->sync = 0;
3013         init_completion(&async->wait);
3014
3015         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3016                         delayed_ref_async_start, NULL, NULL);
3017
3018         btrfs_queue_work(fs_info->extent_workers, &async->work);
3019
3020         if (wait) {
3021                 wait_for_completion(&async->wait);
3022                 ret = async->error;
3023                 kfree(async);
3024                 return ret;
3025         }
3026         return 0;
3027 }
3028
3029 /*
3030  * this starts processing the delayed reference count updates and
3031  * extent insertions we have queued up so far.  count can be
3032  * 0, which means to process everything in the tree at the start
3033  * of the run (but not newly added entries), or it can be some target
3034  * number you'd like to process.
3035  *
3036  * Returns 0 on success or if called with an aborted transaction
3037  * Returns <0 on error and aborts the transaction
3038  */
3039 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3040                            unsigned long count)
3041 {
3042         struct btrfs_fs_info *fs_info = trans->fs_info;
3043         struct rb_node *node;
3044         struct btrfs_delayed_ref_root *delayed_refs;
3045         struct btrfs_delayed_ref_head *head;
3046         int ret;
3047         int run_all = count == (unsigned long)-1;
3048         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3049
3050         /* We'll clean this up in btrfs_cleanup_transaction */
3051         if (trans->aborted)
3052                 return 0;
3053
3054         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3055                 return 0;
3056
3057         delayed_refs = &trans->transaction->delayed_refs;
3058         if (count == 0)
3059                 count = atomic_read(&delayed_refs->num_entries) * 2;
3060
3061 again:
3062 #ifdef SCRAMBLE_DELAYED_REFS
3063         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3064 #endif
3065         trans->can_flush_pending_bgs = false;
3066         ret = __btrfs_run_delayed_refs(trans, count);
3067         if (ret < 0) {
3068                 btrfs_abort_transaction(trans, ret);
3069                 return ret;
3070         }
3071
3072         if (run_all) {
3073                 if (!list_empty(&trans->new_bgs))
3074                         btrfs_create_pending_block_groups(trans);
3075
3076                 spin_lock(&delayed_refs->lock);
3077                 node = rb_first(&delayed_refs->href_root);
3078                 if (!node) {
3079                         spin_unlock(&delayed_refs->lock);
3080                         goto out;
3081                 }
3082                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3083                                 href_node);
3084                 refcount_inc(&head->refs);
3085                 spin_unlock(&delayed_refs->lock);
3086
3087                 /* Mutex was contended, block until it's released and retry. */
3088                 mutex_lock(&head->mutex);
3089                 mutex_unlock(&head->mutex);
3090
3091                 btrfs_put_delayed_ref_head(head);
3092                 cond_resched();
3093                 goto again;
3094         }
3095 out:
3096         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3097         return 0;
3098 }
3099
3100 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3101                                 struct btrfs_fs_info *fs_info,
3102                                 u64 bytenr, u64 num_bytes, u64 flags,
3103                                 int level, int is_data)
3104 {
3105         struct btrfs_delayed_extent_op *extent_op;
3106         int ret;
3107
3108         extent_op = btrfs_alloc_delayed_extent_op();
3109         if (!extent_op)
3110                 return -ENOMEM;
3111
3112         extent_op->flags_to_set = flags;
3113         extent_op->update_flags = true;
3114         extent_op->update_key = false;
3115         extent_op->is_data = is_data ? true : false;
3116         extent_op->level = level;
3117
3118         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3119                                           num_bytes, extent_op);
3120         if (ret)
3121                 btrfs_free_delayed_extent_op(extent_op);
3122         return ret;
3123 }
3124
3125 static noinline int check_delayed_ref(struct btrfs_root *root,
3126                                       struct btrfs_path *path,
3127                                       u64 objectid, u64 offset, u64 bytenr)
3128 {
3129         struct btrfs_delayed_ref_head *head;
3130         struct btrfs_delayed_ref_node *ref;
3131         struct btrfs_delayed_data_ref *data_ref;
3132         struct btrfs_delayed_ref_root *delayed_refs;
3133         struct btrfs_transaction *cur_trans;
3134         struct rb_node *node;
3135         int ret = 0;
3136
3137         spin_lock(&root->fs_info->trans_lock);
3138         cur_trans = root->fs_info->running_transaction;
3139         if (cur_trans)
3140                 refcount_inc(&cur_trans->use_count);
3141         spin_unlock(&root->fs_info->trans_lock);
3142         if (!cur_trans)
3143                 return 0;
3144
3145         delayed_refs = &cur_trans->delayed_refs;
3146         spin_lock(&delayed_refs->lock);
3147         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3148         if (!head) {
3149                 spin_unlock(&delayed_refs->lock);
3150                 btrfs_put_transaction(cur_trans);
3151                 return 0;
3152         }
3153
3154         if (!mutex_trylock(&head->mutex)) {
3155                 refcount_inc(&head->refs);
3156                 spin_unlock(&delayed_refs->lock);
3157
3158                 btrfs_release_path(path);
3159
3160                 /*
3161                  * Mutex was contended, block until it's released and let
3162                  * caller try again
3163                  */
3164                 mutex_lock(&head->mutex);
3165                 mutex_unlock(&head->mutex);
3166                 btrfs_put_delayed_ref_head(head);
3167                 btrfs_put_transaction(cur_trans);
3168                 return -EAGAIN;
3169         }
3170         spin_unlock(&delayed_refs->lock);
3171
3172         spin_lock(&head->lock);
3173         /*
3174          * XXX: We should replace this with a proper search function in the
3175          * future.
3176          */
3177         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3178                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3179                 /* If it's a shared ref we know a cross reference exists */
3180                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3181                         ret = 1;
3182                         break;
3183                 }
3184
3185                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3186
3187                 /*
3188                  * If our ref doesn't match the one we're currently looking at
3189                  * then we have a cross reference.
3190                  */
3191                 if (data_ref->root != root->root_key.objectid ||
3192                     data_ref->objectid != objectid ||
3193                     data_ref->offset != offset) {
3194                         ret = 1;
3195                         break;
3196                 }
3197         }
3198         spin_unlock(&head->lock);
3199         mutex_unlock(&head->mutex);
3200         btrfs_put_transaction(cur_trans);
3201         return ret;
3202 }
3203
3204 static noinline int check_committed_ref(struct btrfs_root *root,
3205                                         struct btrfs_path *path,
3206                                         u64 objectid, u64 offset, u64 bytenr)
3207 {
3208         struct btrfs_fs_info *fs_info = root->fs_info;
3209         struct btrfs_root *extent_root = fs_info->extent_root;
3210         struct extent_buffer *leaf;
3211         struct btrfs_extent_data_ref *ref;
3212         struct btrfs_extent_inline_ref *iref;
3213         struct btrfs_extent_item *ei;
3214         struct btrfs_key key;
3215         u32 item_size;
3216         int type;
3217         int ret;
3218
3219         key.objectid = bytenr;
3220         key.offset = (u64)-1;
3221         key.type = BTRFS_EXTENT_ITEM_KEY;
3222
3223         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3224         if (ret < 0)
3225                 goto out;
3226         BUG_ON(ret == 0); /* Corruption */
3227
3228         ret = -ENOENT;
3229         if (path->slots[0] == 0)
3230                 goto out;
3231
3232         path->slots[0]--;
3233         leaf = path->nodes[0];
3234         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3235
3236         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3237                 goto out;
3238
3239         ret = 1;
3240         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3241 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3242         if (item_size < sizeof(*ei)) {
3243                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3244                 goto out;
3245         }
3246 #endif
3247         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3248
3249         if (item_size != sizeof(*ei) +
3250             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3251                 goto out;
3252
3253         if (btrfs_extent_generation(leaf, ei) <=
3254             btrfs_root_last_snapshot(&root->root_item))
3255                 goto out;
3256
3257         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3258
3259         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3260         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3261                 goto out;
3262
3263         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3264         if (btrfs_extent_refs(leaf, ei) !=
3265             btrfs_extent_data_ref_count(leaf, ref) ||
3266             btrfs_extent_data_ref_root(leaf, ref) !=
3267             root->root_key.objectid ||
3268             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3269             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3270                 goto out;
3271
3272         ret = 0;
3273 out:
3274         return ret;
3275 }
3276
3277 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3278                           u64 bytenr)
3279 {
3280         struct btrfs_path *path;
3281         int ret;
3282         int ret2;
3283
3284         path = btrfs_alloc_path();
3285         if (!path)
3286                 return -ENOMEM;
3287
3288         do {
3289                 ret = check_committed_ref(root, path, objectid,
3290                                           offset, bytenr);
3291                 if (ret && ret != -ENOENT)
3292                         goto out;
3293
3294                 ret2 = check_delayed_ref(root, path, objectid,
3295                                          offset, bytenr);
3296         } while (ret2 == -EAGAIN);
3297
3298         if (ret2 && ret2 != -ENOENT) {
3299                 ret = ret2;
3300                 goto out;
3301         }
3302
3303         if (ret != -ENOENT || ret2 != -ENOENT)
3304                 ret = 0;
3305 out:
3306         btrfs_free_path(path);
3307         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3308                 WARN_ON(ret > 0);
3309         return ret;
3310 }
3311
3312 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3313                            struct btrfs_root *root,
3314                            struct extent_buffer *buf,
3315                            int full_backref, int inc)
3316 {
3317         struct btrfs_fs_info *fs_info = root->fs_info;
3318         u64 bytenr;
3319         u64 num_bytes;
3320         u64 parent;
3321         u64 ref_root;
3322         u32 nritems;
3323         struct btrfs_key key;
3324         struct btrfs_file_extent_item *fi;
3325         int i;
3326         int level;
3327         int ret = 0;
3328         int (*process_func)(struct btrfs_trans_handle *,
3329                             struct btrfs_root *,
3330                             u64, u64, u64, u64, u64, u64);
3331
3332
3333         if (btrfs_is_testing(fs_info))
3334                 return 0;
3335
3336         ref_root = btrfs_header_owner(buf);
3337         nritems = btrfs_header_nritems(buf);
3338         level = btrfs_header_level(buf);
3339
3340         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3341                 return 0;
3342
3343         if (inc)
3344                 process_func = btrfs_inc_extent_ref;
3345         else
3346                 process_func = btrfs_free_extent;
3347
3348         if (full_backref)
3349                 parent = buf->start;
3350         else
3351                 parent = 0;
3352
3353         for (i = 0; i < nritems; i++) {
3354                 if (level == 0) {
3355                         btrfs_item_key_to_cpu(buf, &key, i);
3356                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3357                                 continue;
3358                         fi = btrfs_item_ptr(buf, i,
3359                                             struct btrfs_file_extent_item);
3360                         if (btrfs_file_extent_type(buf, fi) ==
3361                             BTRFS_FILE_EXTENT_INLINE)
3362                                 continue;
3363                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3364                         if (bytenr == 0)
3365                                 continue;
3366
3367                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3368                         key.offset -= btrfs_file_extent_offset(buf, fi);
3369                         ret = process_func(trans, root, bytenr, num_bytes,
3370                                            parent, ref_root, key.objectid,
3371                                            key.offset);
3372                         if (ret)
3373                                 goto fail;
3374                 } else {
3375                         bytenr = btrfs_node_blockptr(buf, i);
3376                         num_bytes = fs_info->nodesize;
3377                         ret = process_func(trans, root, bytenr, num_bytes,
3378                                            parent, ref_root, level - 1, 0);
3379                         if (ret)
3380                                 goto fail;
3381                 }
3382         }
3383         return 0;
3384 fail:
3385         return ret;
3386 }
3387
3388 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3389                   struct extent_buffer *buf, int full_backref)
3390 {
3391         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3392 }
3393
3394 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3395                   struct extent_buffer *buf, int full_backref)
3396 {
3397         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3398 }
3399
3400 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3401                                  struct btrfs_fs_info *fs_info,
3402                                  struct btrfs_path *path,
3403                                  struct btrfs_block_group_cache *cache)
3404 {
3405         int ret;
3406         struct btrfs_root *extent_root = fs_info->extent_root;
3407         unsigned long bi;
3408         struct extent_buffer *leaf;
3409
3410         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3411         if (ret) {
3412                 if (ret > 0)
3413                         ret = -ENOENT;
3414                 goto fail;
3415         }
3416
3417         leaf = path->nodes[0];
3418         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3419         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3420         btrfs_mark_buffer_dirty(leaf);
3421 fail:
3422         btrfs_release_path(path);
3423         return ret;
3424
3425 }
3426
3427 static struct btrfs_block_group_cache *
3428 next_block_group(struct btrfs_fs_info *fs_info,
3429                  struct btrfs_block_group_cache *cache)
3430 {
3431         struct rb_node *node;
3432
3433         spin_lock(&fs_info->block_group_cache_lock);
3434
3435         /* If our block group was removed, we need a full search. */
3436         if (RB_EMPTY_NODE(&cache->cache_node)) {
3437                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3438
3439                 spin_unlock(&fs_info->block_group_cache_lock);
3440                 btrfs_put_block_group(cache);
3441                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3442         }
3443         node = rb_next(&cache->cache_node);
3444         btrfs_put_block_group(cache);
3445         if (node) {
3446                 cache = rb_entry(node, struct btrfs_block_group_cache,
3447                                  cache_node);
3448                 btrfs_get_block_group(cache);
3449         } else
3450                 cache = NULL;
3451         spin_unlock(&fs_info->block_group_cache_lock);
3452         return cache;
3453 }
3454
3455 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3456                             struct btrfs_trans_handle *trans,
3457                             struct btrfs_path *path)
3458 {
3459         struct btrfs_fs_info *fs_info = block_group->fs_info;
3460         struct btrfs_root *root = fs_info->tree_root;
3461         struct inode *inode = NULL;
3462         struct extent_changeset *data_reserved = NULL;
3463         u64 alloc_hint = 0;
3464         int dcs = BTRFS_DC_ERROR;
3465         u64 num_pages = 0;
3466         int retries = 0;
3467         int ret = 0;
3468
3469         /*
3470          * If this block group is smaller than 100 megs don't bother caching the
3471          * block group.
3472          */
3473         if (block_group->key.offset < (100 * SZ_1M)) {
3474                 spin_lock(&block_group->lock);
3475                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3476                 spin_unlock(&block_group->lock);
3477                 return 0;
3478         }
3479
3480         if (trans->aborted)
3481                 return 0;
3482 again:
3483         inode = lookup_free_space_inode(fs_info, block_group, path);
3484         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3485                 ret = PTR_ERR(inode);
3486                 btrfs_release_path(path);
3487                 goto out;
3488         }
3489
3490         if (IS_ERR(inode)) {
3491                 BUG_ON(retries);
3492                 retries++;
3493
3494                 if (block_group->ro)
3495                         goto out_free;
3496
3497                 ret = create_free_space_inode(fs_info, trans, block_group,
3498                                               path);
3499                 if (ret)
3500                         goto out_free;
3501                 goto again;
3502         }
3503
3504         /*
3505          * We want to set the generation to 0, that way if anything goes wrong
3506          * from here on out we know not to trust this cache when we load up next
3507          * time.
3508          */
3509         BTRFS_I(inode)->generation = 0;
3510         ret = btrfs_update_inode(trans, root, inode);
3511         if (ret) {
3512                 /*
3513                  * So theoretically we could recover from this, simply set the
3514                  * super cache generation to 0 so we know to invalidate the
3515                  * cache, but then we'd have to keep track of the block groups
3516                  * that fail this way so we know we _have_ to reset this cache
3517                  * before the next commit or risk reading stale cache.  So to
3518                  * limit our exposure to horrible edge cases lets just abort the
3519                  * transaction, this only happens in really bad situations
3520                  * anyway.
3521                  */
3522                 btrfs_abort_transaction(trans, ret);
3523                 goto out_put;
3524         }
3525         WARN_ON(ret);
3526
3527         /* We've already setup this transaction, go ahead and exit */
3528         if (block_group->cache_generation == trans->transid &&
3529             i_size_read(inode)) {
3530                 dcs = BTRFS_DC_SETUP;
3531                 goto out_put;
3532         }
3533
3534         if (i_size_read(inode) > 0) {
3535                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3536                                         &fs_info->global_block_rsv);
3537                 if (ret)
3538                         goto out_put;
3539
3540                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3541                 if (ret)
3542                         goto out_put;
3543         }
3544
3545         spin_lock(&block_group->lock);
3546         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3547             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3548                 /*
3549                  * don't bother trying to write stuff out _if_
3550                  * a) we're not cached,
3551                  * b) we're with nospace_cache mount option,
3552                  * c) we're with v2 space_cache (FREE_SPACE_TREE).
3553                  */
3554                 dcs = BTRFS_DC_WRITTEN;