Merge branch 'for-4.14/block-postmerge' of git://git.kernel.dk/linux-block
[sfrench/cifs-2.6.git] / fs / btrfs / extent-tree.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 #include <linux/sched.h>
19 #include <linux/sched/signal.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/sort.h>
24 #include <linux/rcupdate.h>
25 #include <linux/kthread.h>
26 #include <linux/slab.h>
27 #include <linux/ratelimit.h>
28 #include <linux/percpu_counter.h>
29 #include "hash.h"
30 #include "tree-log.h"
31 #include "disk-io.h"
32 #include "print-tree.h"
33 #include "volumes.h"
34 #include "raid56.h"
35 #include "locking.h"
36 #include "free-space-cache.h"
37 #include "free-space-tree.h"
38 #include "math.h"
39 #include "sysfs.h"
40 #include "qgroup.h"
41
42 #undef SCRAMBLE_DELAYED_REFS
43
44 /*
45  * control flags for do_chunk_alloc's force field
46  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
47  * if we really need one.
48  *
49  * CHUNK_ALLOC_LIMITED means to only try and allocate one
50  * if we have very few chunks already allocated.  This is
51  * used as part of the clustering code to help make sure
52  * we have a good pool of storage to cluster in, without
53  * filling the FS with empty chunks
54  *
55  * CHUNK_ALLOC_FORCE means it must try to allocate one
56  *
57  */
58 enum {
59         CHUNK_ALLOC_NO_FORCE = 0,
60         CHUNK_ALLOC_LIMITED = 1,
61         CHUNK_ALLOC_FORCE = 2,
62 };
63
64 static int update_block_group(struct btrfs_trans_handle *trans,
65                               struct btrfs_fs_info *fs_info, u64 bytenr,
66                               u64 num_bytes, int alloc);
67 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
68                                struct btrfs_fs_info *fs_info,
69                                 struct btrfs_delayed_ref_node *node, u64 parent,
70                                 u64 root_objectid, u64 owner_objectid,
71                                 u64 owner_offset, int refs_to_drop,
72                                 struct btrfs_delayed_extent_op *extra_op);
73 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
74                                     struct extent_buffer *leaf,
75                                     struct btrfs_extent_item *ei);
76 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
77                                       struct btrfs_fs_info *fs_info,
78                                       u64 parent, u64 root_objectid,
79                                       u64 flags, u64 owner, u64 offset,
80                                       struct btrfs_key *ins, int ref_mod);
81 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
82                                      struct btrfs_fs_info *fs_info,
83                                      u64 parent, u64 root_objectid,
84                                      u64 flags, struct btrfs_disk_key *key,
85                                      int level, struct btrfs_key *ins);
86 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
87                           struct btrfs_fs_info *fs_info, u64 flags,
88                           int force);
89 static int find_next_key(struct btrfs_path *path, int level,
90                          struct btrfs_key *key);
91 static void dump_space_info(struct btrfs_fs_info *fs_info,
92                             struct btrfs_space_info *info, u64 bytes,
93                             int dump_block_groups);
94 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
95                                     u64 ram_bytes, u64 num_bytes, int delalloc);
96 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
97                                      u64 num_bytes, int delalloc);
98 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
99                                u64 num_bytes);
100 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
101                                     struct btrfs_space_info *space_info,
102                                     u64 orig_bytes,
103                                     enum btrfs_reserve_flush_enum flush,
104                                     bool system_chunk);
105 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
106                                      struct btrfs_space_info *space_info,
107                                      u64 num_bytes);
108 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
109                                      struct btrfs_space_info *space_info,
110                                      u64 num_bytes);
111
112 static noinline int
113 block_group_cache_done(struct btrfs_block_group_cache *cache)
114 {
115         smp_mb();
116         return cache->cached == BTRFS_CACHE_FINISHED ||
117                 cache->cached == BTRFS_CACHE_ERROR;
118 }
119
120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
121 {
122         return (cache->flags & bits) == bits;
123 }
124
125 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
126 {
127         atomic_inc(&cache->count);
128 }
129
130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
131 {
132         if (atomic_dec_and_test(&cache->count)) {
133                 WARN_ON(cache->pinned > 0);
134                 WARN_ON(cache->reserved > 0);
135
136                 /*
137                  * If not empty, someone is still holding mutex of
138                  * full_stripe_lock, which can only be released by caller.
139                  * And it will definitely cause use-after-free when caller
140                  * tries to release full stripe lock.
141                  *
142                  * No better way to resolve, but only to warn.
143                  */
144                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
145                 kfree(cache->free_space_ctl);
146                 kfree(cache);
147         }
148 }
149
150 /*
151  * this adds the block group to the fs_info rb tree for the block group
152  * cache
153  */
154 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
155                                 struct btrfs_block_group_cache *block_group)
156 {
157         struct rb_node **p;
158         struct rb_node *parent = NULL;
159         struct btrfs_block_group_cache *cache;
160
161         spin_lock(&info->block_group_cache_lock);
162         p = &info->block_group_cache_tree.rb_node;
163
164         while (*p) {
165                 parent = *p;
166                 cache = rb_entry(parent, struct btrfs_block_group_cache,
167                                  cache_node);
168                 if (block_group->key.objectid < cache->key.objectid) {
169                         p = &(*p)->rb_left;
170                 } else if (block_group->key.objectid > cache->key.objectid) {
171                         p = &(*p)->rb_right;
172                 } else {
173                         spin_unlock(&info->block_group_cache_lock);
174                         return -EEXIST;
175                 }
176         }
177
178         rb_link_node(&block_group->cache_node, parent, p);
179         rb_insert_color(&block_group->cache_node,
180                         &info->block_group_cache_tree);
181
182         if (info->first_logical_byte > block_group->key.objectid)
183                 info->first_logical_byte = block_group->key.objectid;
184
185         spin_unlock(&info->block_group_cache_lock);
186
187         return 0;
188 }
189
190 /*
191  * This will return the block group at or after bytenr if contains is 0, else
192  * it will return the block group that contains the bytenr
193  */
194 static struct btrfs_block_group_cache *
195 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
196                               int contains)
197 {
198         struct btrfs_block_group_cache *cache, *ret = NULL;
199         struct rb_node *n;
200         u64 end, start;
201
202         spin_lock(&info->block_group_cache_lock);
203         n = info->block_group_cache_tree.rb_node;
204
205         while (n) {
206                 cache = rb_entry(n, struct btrfs_block_group_cache,
207                                  cache_node);
208                 end = cache->key.objectid + cache->key.offset - 1;
209                 start = cache->key.objectid;
210
211                 if (bytenr < start) {
212                         if (!contains && (!ret || start < ret->key.objectid))
213                                 ret = cache;
214                         n = n->rb_left;
215                 } else if (bytenr > start) {
216                         if (contains && bytenr <= end) {
217                                 ret = cache;
218                                 break;
219                         }
220                         n = n->rb_right;
221                 } else {
222                         ret = cache;
223                         break;
224                 }
225         }
226         if (ret) {
227                 btrfs_get_block_group(ret);
228                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
229                         info->first_logical_byte = ret->key.objectid;
230         }
231         spin_unlock(&info->block_group_cache_lock);
232
233         return ret;
234 }
235
236 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
237                                u64 start, u64 num_bytes)
238 {
239         u64 end = start + num_bytes - 1;
240         set_extent_bits(&fs_info->freed_extents[0],
241                         start, end, EXTENT_UPTODATE);
242         set_extent_bits(&fs_info->freed_extents[1],
243                         start, end, EXTENT_UPTODATE);
244         return 0;
245 }
246
247 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
248                                   struct btrfs_block_group_cache *cache)
249 {
250         u64 start, end;
251
252         start = cache->key.objectid;
253         end = start + cache->key.offset - 1;
254
255         clear_extent_bits(&fs_info->freed_extents[0],
256                           start, end, EXTENT_UPTODATE);
257         clear_extent_bits(&fs_info->freed_extents[1],
258                           start, end, EXTENT_UPTODATE);
259 }
260
261 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
262                                  struct btrfs_block_group_cache *cache)
263 {
264         u64 bytenr;
265         u64 *logical;
266         int stripe_len;
267         int i, nr, ret;
268
269         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
270                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
271                 cache->bytes_super += stripe_len;
272                 ret = add_excluded_extent(fs_info, cache->key.objectid,
273                                           stripe_len);
274                 if (ret)
275                         return ret;
276         }
277
278         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
279                 bytenr = btrfs_sb_offset(i);
280                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
281                                        bytenr, 0, &logical, &nr, &stripe_len);
282                 if (ret)
283                         return ret;
284
285                 while (nr--) {
286                         u64 start, len;
287
288                         if (logical[nr] > cache->key.objectid +
289                             cache->key.offset)
290                                 continue;
291
292                         if (logical[nr] + stripe_len <= cache->key.objectid)
293                                 continue;
294
295                         start = logical[nr];
296                         if (start < cache->key.objectid) {
297                                 start = cache->key.objectid;
298                                 len = (logical[nr] + stripe_len) - start;
299                         } else {
300                                 len = min_t(u64, stripe_len,
301                                             cache->key.objectid +
302                                             cache->key.offset - start);
303                         }
304
305                         cache->bytes_super += len;
306                         ret = add_excluded_extent(fs_info, start, len);
307                         if (ret) {
308                                 kfree(logical);
309                                 return ret;
310                         }
311                 }
312
313                 kfree(logical);
314         }
315         return 0;
316 }
317
318 static struct btrfs_caching_control *
319 get_caching_control(struct btrfs_block_group_cache *cache)
320 {
321         struct btrfs_caching_control *ctl;
322
323         spin_lock(&cache->lock);
324         if (!cache->caching_ctl) {
325                 spin_unlock(&cache->lock);
326                 return NULL;
327         }
328
329         ctl = cache->caching_ctl;
330         refcount_inc(&ctl->count);
331         spin_unlock(&cache->lock);
332         return ctl;
333 }
334
335 static void put_caching_control(struct btrfs_caching_control *ctl)
336 {
337         if (refcount_dec_and_test(&ctl->count))
338                 kfree(ctl);
339 }
340
341 #ifdef CONFIG_BTRFS_DEBUG
342 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
343 {
344         struct btrfs_fs_info *fs_info = block_group->fs_info;
345         u64 start = block_group->key.objectid;
346         u64 len = block_group->key.offset;
347         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
348                 fs_info->nodesize : fs_info->sectorsize;
349         u64 step = chunk << 1;
350
351         while (len > chunk) {
352                 btrfs_remove_free_space(block_group, start, chunk);
353                 start += step;
354                 if (len < step)
355                         len = 0;
356                 else
357                         len -= step;
358         }
359 }
360 #endif
361
362 /*
363  * this is only called by cache_block_group, since we could have freed extents
364  * we need to check the pinned_extents for any extents that can't be used yet
365  * since their free space will be released as soon as the transaction commits.
366  */
367 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
368                        struct btrfs_fs_info *info, u64 start, u64 end)
369 {
370         u64 extent_start, extent_end, size, total_added = 0;
371         int ret;
372
373         while (start < end) {
374                 ret = find_first_extent_bit(info->pinned_extents, start,
375                                             &extent_start, &extent_end,
376                                             EXTENT_DIRTY | EXTENT_UPTODATE,
377                                             NULL);
378                 if (ret)
379                         break;
380
381                 if (extent_start <= start) {
382                         start = extent_end + 1;
383                 } else if (extent_start > start && extent_start < end) {
384                         size = extent_start - start;
385                         total_added += size;
386                         ret = btrfs_add_free_space(block_group, start,
387                                                    size);
388                         BUG_ON(ret); /* -ENOMEM or logic error */
389                         start = extent_end + 1;
390                 } else {
391                         break;
392                 }
393         }
394
395         if (start < end) {
396                 size = end - start;
397                 total_added += size;
398                 ret = btrfs_add_free_space(block_group, start, size);
399                 BUG_ON(ret); /* -ENOMEM or logic error */
400         }
401
402         return total_added;
403 }
404
405 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
406 {
407         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
408         struct btrfs_fs_info *fs_info = block_group->fs_info;
409         struct btrfs_root *extent_root = fs_info->extent_root;
410         struct btrfs_path *path;
411         struct extent_buffer *leaf;
412         struct btrfs_key key;
413         u64 total_found = 0;
414         u64 last = 0;
415         u32 nritems;
416         int ret;
417         bool wakeup = true;
418
419         path = btrfs_alloc_path();
420         if (!path)
421                 return -ENOMEM;
422
423         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
424
425 #ifdef CONFIG_BTRFS_DEBUG
426         /*
427          * If we're fragmenting we don't want to make anybody think we can
428          * allocate from this block group until we've had a chance to fragment
429          * the free space.
430          */
431         if (btrfs_should_fragment_free_space(block_group))
432                 wakeup = false;
433 #endif
434         /*
435          * We don't want to deadlock with somebody trying to allocate a new
436          * extent for the extent root while also trying to search the extent
437          * root to add free space.  So we skip locking and search the commit
438          * root, since its read-only
439          */
440         path->skip_locking = 1;
441         path->search_commit_root = 1;
442         path->reada = READA_FORWARD;
443
444         key.objectid = last;
445         key.offset = 0;
446         key.type = BTRFS_EXTENT_ITEM_KEY;
447
448 next:
449         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
450         if (ret < 0)
451                 goto out;
452
453         leaf = path->nodes[0];
454         nritems = btrfs_header_nritems(leaf);
455
456         while (1) {
457                 if (btrfs_fs_closing(fs_info) > 1) {
458                         last = (u64)-1;
459                         break;
460                 }
461
462                 if (path->slots[0] < nritems) {
463                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
464                 } else {
465                         ret = find_next_key(path, 0, &key);
466                         if (ret)
467                                 break;
468
469                         if (need_resched() ||
470                             rwsem_is_contended(&fs_info->commit_root_sem)) {
471                                 if (wakeup)
472                                         caching_ctl->progress = last;
473                                 btrfs_release_path(path);
474                                 up_read(&fs_info->commit_root_sem);
475                                 mutex_unlock(&caching_ctl->mutex);
476                                 cond_resched();
477                                 mutex_lock(&caching_ctl->mutex);
478                                 down_read(&fs_info->commit_root_sem);
479                                 goto next;
480                         }
481
482                         ret = btrfs_next_leaf(extent_root, path);
483                         if (ret < 0)
484                                 goto out;
485                         if (ret)
486                                 break;
487                         leaf = path->nodes[0];
488                         nritems = btrfs_header_nritems(leaf);
489                         continue;
490                 }
491
492                 if (key.objectid < last) {
493                         key.objectid = last;
494                         key.offset = 0;
495                         key.type = BTRFS_EXTENT_ITEM_KEY;
496
497                         if (wakeup)
498                                 caching_ctl->progress = last;
499                         btrfs_release_path(path);
500                         goto next;
501                 }
502
503                 if (key.objectid < block_group->key.objectid) {
504                         path->slots[0]++;
505                         continue;
506                 }
507
508                 if (key.objectid >= block_group->key.objectid +
509                     block_group->key.offset)
510                         break;
511
512                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
513                     key.type == BTRFS_METADATA_ITEM_KEY) {
514                         total_found += add_new_free_space(block_group,
515                                                           fs_info, last,
516                                                           key.objectid);
517                         if (key.type == BTRFS_METADATA_ITEM_KEY)
518                                 last = key.objectid +
519                                         fs_info->nodesize;
520                         else
521                                 last = key.objectid + key.offset;
522
523                         if (total_found > CACHING_CTL_WAKE_UP) {
524                                 total_found = 0;
525                                 if (wakeup)
526                                         wake_up(&caching_ctl->wait);
527                         }
528                 }
529                 path->slots[0]++;
530         }
531         ret = 0;
532
533         total_found += add_new_free_space(block_group, fs_info, last,
534                                           block_group->key.objectid +
535                                           block_group->key.offset);
536         caching_ctl->progress = (u64)-1;
537
538 out:
539         btrfs_free_path(path);
540         return ret;
541 }
542
543 static noinline void caching_thread(struct btrfs_work *work)
544 {
545         struct btrfs_block_group_cache *block_group;
546         struct btrfs_fs_info *fs_info;
547         struct btrfs_caching_control *caching_ctl;
548         struct btrfs_root *extent_root;
549         int ret;
550
551         caching_ctl = container_of(work, struct btrfs_caching_control, work);
552         block_group = caching_ctl->block_group;
553         fs_info = block_group->fs_info;
554         extent_root = fs_info->extent_root;
555
556         mutex_lock(&caching_ctl->mutex);
557         down_read(&fs_info->commit_root_sem);
558
559         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
560                 ret = load_free_space_tree(caching_ctl);
561         else
562                 ret = load_extent_tree_free(caching_ctl);
563
564         spin_lock(&block_group->lock);
565         block_group->caching_ctl = NULL;
566         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
567         spin_unlock(&block_group->lock);
568
569 #ifdef CONFIG_BTRFS_DEBUG
570         if (btrfs_should_fragment_free_space(block_group)) {
571                 u64 bytes_used;
572
573                 spin_lock(&block_group->space_info->lock);
574                 spin_lock(&block_group->lock);
575                 bytes_used = block_group->key.offset -
576                         btrfs_block_group_used(&block_group->item);
577                 block_group->space_info->bytes_used += bytes_used >> 1;
578                 spin_unlock(&block_group->lock);
579                 spin_unlock(&block_group->space_info->lock);
580                 fragment_free_space(block_group);
581         }
582 #endif
583
584         caching_ctl->progress = (u64)-1;
585
586         up_read(&fs_info->commit_root_sem);
587         free_excluded_extents(fs_info, block_group);
588         mutex_unlock(&caching_ctl->mutex);
589
590         wake_up(&caching_ctl->wait);
591
592         put_caching_control(caching_ctl);
593         btrfs_put_block_group(block_group);
594 }
595
596 static int cache_block_group(struct btrfs_block_group_cache *cache,
597                              int load_cache_only)
598 {
599         DEFINE_WAIT(wait);
600         struct btrfs_fs_info *fs_info = cache->fs_info;
601         struct btrfs_caching_control *caching_ctl;
602         int ret = 0;
603
604         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
605         if (!caching_ctl)
606                 return -ENOMEM;
607
608         INIT_LIST_HEAD(&caching_ctl->list);
609         mutex_init(&caching_ctl->mutex);
610         init_waitqueue_head(&caching_ctl->wait);
611         caching_ctl->block_group = cache;
612         caching_ctl->progress = cache->key.objectid;
613         refcount_set(&caching_ctl->count, 1);
614         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
615                         caching_thread, NULL, NULL);
616
617         spin_lock(&cache->lock);
618         /*
619          * This should be a rare occasion, but this could happen I think in the
620          * case where one thread starts to load the space cache info, and then
621          * some other thread starts a transaction commit which tries to do an
622          * allocation while the other thread is still loading the space cache
623          * info.  The previous loop should have kept us from choosing this block
624          * group, but if we've moved to the state where we will wait on caching
625          * block groups we need to first check if we're doing a fast load here,
626          * so we can wait for it to finish, otherwise we could end up allocating
627          * from a block group who's cache gets evicted for one reason or
628          * another.
629          */
630         while (cache->cached == BTRFS_CACHE_FAST) {
631                 struct btrfs_caching_control *ctl;
632
633                 ctl = cache->caching_ctl;
634                 refcount_inc(&ctl->count);
635                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
636                 spin_unlock(&cache->lock);
637
638                 schedule();
639
640                 finish_wait(&ctl->wait, &wait);
641                 put_caching_control(ctl);
642                 spin_lock(&cache->lock);
643         }
644
645         if (cache->cached != BTRFS_CACHE_NO) {
646                 spin_unlock(&cache->lock);
647                 kfree(caching_ctl);
648                 return 0;
649         }
650         WARN_ON(cache->caching_ctl);
651         cache->caching_ctl = caching_ctl;
652         cache->cached = BTRFS_CACHE_FAST;
653         spin_unlock(&cache->lock);
654
655         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
656                 mutex_lock(&caching_ctl->mutex);
657                 ret = load_free_space_cache(fs_info, cache);
658
659                 spin_lock(&cache->lock);
660                 if (ret == 1) {
661                         cache->caching_ctl = NULL;
662                         cache->cached = BTRFS_CACHE_FINISHED;
663                         cache->last_byte_to_unpin = (u64)-1;
664                         caching_ctl->progress = (u64)-1;
665                 } else {
666                         if (load_cache_only) {
667                                 cache->caching_ctl = NULL;
668                                 cache->cached = BTRFS_CACHE_NO;
669                         } else {
670                                 cache->cached = BTRFS_CACHE_STARTED;
671                                 cache->has_caching_ctl = 1;
672                         }
673                 }
674                 spin_unlock(&cache->lock);
675 #ifdef CONFIG_BTRFS_DEBUG
676                 if (ret == 1 &&
677                     btrfs_should_fragment_free_space(cache)) {
678                         u64 bytes_used;
679
680                         spin_lock(&cache->space_info->lock);
681                         spin_lock(&cache->lock);
682                         bytes_used = cache->key.offset -
683                                 btrfs_block_group_used(&cache->item);
684                         cache->space_info->bytes_used += bytes_used >> 1;
685                         spin_unlock(&cache->lock);
686                         spin_unlock(&cache->space_info->lock);
687                         fragment_free_space(cache);
688                 }
689 #endif
690                 mutex_unlock(&caching_ctl->mutex);
691
692                 wake_up(&caching_ctl->wait);
693                 if (ret == 1) {
694                         put_caching_control(caching_ctl);
695                         free_excluded_extents(fs_info, cache);
696                         return 0;
697                 }
698         } else {
699                 /*
700                  * We're either using the free space tree or no caching at all.
701                  * Set cached to the appropriate value and wakeup any waiters.
702                  */
703                 spin_lock(&cache->lock);
704                 if (load_cache_only) {
705                         cache->caching_ctl = NULL;
706                         cache->cached = BTRFS_CACHE_NO;
707                 } else {
708                         cache->cached = BTRFS_CACHE_STARTED;
709                         cache->has_caching_ctl = 1;
710                 }
711                 spin_unlock(&cache->lock);
712                 wake_up(&caching_ctl->wait);
713         }
714
715         if (load_cache_only) {
716                 put_caching_control(caching_ctl);
717                 return 0;
718         }
719
720         down_write(&fs_info->commit_root_sem);
721         refcount_inc(&caching_ctl->count);
722         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
723         up_write(&fs_info->commit_root_sem);
724
725         btrfs_get_block_group(cache);
726
727         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
728
729         return ret;
730 }
731
732 /*
733  * return the block group that starts at or after bytenr
734  */
735 static struct btrfs_block_group_cache *
736 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
737 {
738         return block_group_cache_tree_search(info, bytenr, 0);
739 }
740
741 /*
742  * return the block group that contains the given bytenr
743  */
744 struct btrfs_block_group_cache *btrfs_lookup_block_group(
745                                                  struct btrfs_fs_info *info,
746                                                  u64 bytenr)
747 {
748         return block_group_cache_tree_search(info, bytenr, 1);
749 }
750
751 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
752                                                   u64 flags)
753 {
754         struct list_head *head = &info->space_info;
755         struct btrfs_space_info *found;
756
757         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
758
759         rcu_read_lock();
760         list_for_each_entry_rcu(found, head, list) {
761                 if (found->flags & flags) {
762                         rcu_read_unlock();
763                         return found;
764                 }
765         }
766         rcu_read_unlock();
767         return NULL;
768 }
769
770 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
771                              u64 owner, u64 root_objectid)
772 {
773         struct btrfs_space_info *space_info;
774         u64 flags;
775
776         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
777                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
778                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
779                 else
780                         flags = BTRFS_BLOCK_GROUP_METADATA;
781         } else {
782                 flags = BTRFS_BLOCK_GROUP_DATA;
783         }
784
785         space_info = __find_space_info(fs_info, flags);
786         ASSERT(space_info);
787         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
788 }
789
790 /*
791  * after adding space to the filesystem, we need to clear the full flags
792  * on all the space infos.
793  */
794 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
795 {
796         struct list_head *head = &info->space_info;
797         struct btrfs_space_info *found;
798
799         rcu_read_lock();
800         list_for_each_entry_rcu(found, head, list)
801                 found->full = 0;
802         rcu_read_unlock();
803 }
804
805 /* simple helper to search for an existing data extent at a given offset */
806 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
807 {
808         int ret;
809         struct btrfs_key key;
810         struct btrfs_path *path;
811
812         path = btrfs_alloc_path();
813         if (!path)
814                 return -ENOMEM;
815
816         key.objectid = start;
817         key.offset = len;
818         key.type = BTRFS_EXTENT_ITEM_KEY;
819         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
820         btrfs_free_path(path);
821         return ret;
822 }
823
824 /*
825  * helper function to lookup reference count and flags of a tree block.
826  *
827  * the head node for delayed ref is used to store the sum of all the
828  * reference count modifications queued up in the rbtree. the head
829  * node may also store the extent flags to set. This way you can check
830  * to see what the reference count and extent flags would be if all of
831  * the delayed refs are not processed.
832  */
833 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
834                              struct btrfs_fs_info *fs_info, u64 bytenr,
835                              u64 offset, int metadata, u64 *refs, u64 *flags)
836 {
837         struct btrfs_delayed_ref_head *head;
838         struct btrfs_delayed_ref_root *delayed_refs;
839         struct btrfs_path *path;
840         struct btrfs_extent_item *ei;
841         struct extent_buffer *leaf;
842         struct btrfs_key key;
843         u32 item_size;
844         u64 num_refs;
845         u64 extent_flags;
846         int ret;
847
848         /*
849          * If we don't have skinny metadata, don't bother doing anything
850          * different
851          */
852         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
853                 offset = fs_info->nodesize;
854                 metadata = 0;
855         }
856
857         path = btrfs_alloc_path();
858         if (!path)
859                 return -ENOMEM;
860
861         if (!trans) {
862                 path->skip_locking = 1;
863                 path->search_commit_root = 1;
864         }
865
866 search_again:
867         key.objectid = bytenr;
868         key.offset = offset;
869         if (metadata)
870                 key.type = BTRFS_METADATA_ITEM_KEY;
871         else
872                 key.type = BTRFS_EXTENT_ITEM_KEY;
873
874         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
875         if (ret < 0)
876                 goto out_free;
877
878         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
879                 if (path->slots[0]) {
880                         path->slots[0]--;
881                         btrfs_item_key_to_cpu(path->nodes[0], &key,
882                                               path->slots[0]);
883                         if (key.objectid == bytenr &&
884                             key.type == BTRFS_EXTENT_ITEM_KEY &&
885                             key.offset == fs_info->nodesize)
886                                 ret = 0;
887                 }
888         }
889
890         if (ret == 0) {
891                 leaf = path->nodes[0];
892                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
893                 if (item_size >= sizeof(*ei)) {
894                         ei = btrfs_item_ptr(leaf, path->slots[0],
895                                             struct btrfs_extent_item);
896                         num_refs = btrfs_extent_refs(leaf, ei);
897                         extent_flags = btrfs_extent_flags(leaf, ei);
898                 } else {
899 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
900                         struct btrfs_extent_item_v0 *ei0;
901                         BUG_ON(item_size != sizeof(*ei0));
902                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
903                                              struct btrfs_extent_item_v0);
904                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
905                         /* FIXME: this isn't correct for data */
906                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
907 #else
908                         BUG();
909 #endif
910                 }
911                 BUG_ON(num_refs == 0);
912         } else {
913                 num_refs = 0;
914                 extent_flags = 0;
915                 ret = 0;
916         }
917
918         if (!trans)
919                 goto out;
920
921         delayed_refs = &trans->transaction->delayed_refs;
922         spin_lock(&delayed_refs->lock);
923         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
924         if (head) {
925                 if (!mutex_trylock(&head->mutex)) {
926                         refcount_inc(&head->node.refs);
927                         spin_unlock(&delayed_refs->lock);
928
929                         btrfs_release_path(path);
930
931                         /*
932                          * Mutex was contended, block until it's released and try
933                          * again
934                          */
935                         mutex_lock(&head->mutex);
936                         mutex_unlock(&head->mutex);
937                         btrfs_put_delayed_ref(&head->node);
938                         goto search_again;
939                 }
940                 spin_lock(&head->lock);
941                 if (head->extent_op && head->extent_op->update_flags)
942                         extent_flags |= head->extent_op->flags_to_set;
943                 else
944                         BUG_ON(num_refs == 0);
945
946                 num_refs += head->node.ref_mod;
947                 spin_unlock(&head->lock);
948                 mutex_unlock(&head->mutex);
949         }
950         spin_unlock(&delayed_refs->lock);
951 out:
952         WARN_ON(num_refs == 0);
953         if (refs)
954                 *refs = num_refs;
955         if (flags)
956                 *flags = extent_flags;
957 out_free:
958         btrfs_free_path(path);
959         return ret;
960 }
961
962 /*
963  * Back reference rules.  Back refs have three main goals:
964  *
965  * 1) differentiate between all holders of references to an extent so that
966  *    when a reference is dropped we can make sure it was a valid reference
967  *    before freeing the extent.
968  *
969  * 2) Provide enough information to quickly find the holders of an extent
970  *    if we notice a given block is corrupted or bad.
971  *
972  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
973  *    maintenance.  This is actually the same as #2, but with a slightly
974  *    different use case.
975  *
976  * There are two kinds of back refs. The implicit back refs is optimized
977  * for pointers in non-shared tree blocks. For a given pointer in a block,
978  * back refs of this kind provide information about the block's owner tree
979  * and the pointer's key. These information allow us to find the block by
980  * b-tree searching. The full back refs is for pointers in tree blocks not
981  * referenced by their owner trees. The location of tree block is recorded
982  * in the back refs. Actually the full back refs is generic, and can be
983  * used in all cases the implicit back refs is used. The major shortcoming
984  * of the full back refs is its overhead. Every time a tree block gets
985  * COWed, we have to update back refs entry for all pointers in it.
986  *
987  * For a newly allocated tree block, we use implicit back refs for
988  * pointers in it. This means most tree related operations only involve
989  * implicit back refs. For a tree block created in old transaction, the
990  * only way to drop a reference to it is COW it. So we can detect the
991  * event that tree block loses its owner tree's reference and do the
992  * back refs conversion.
993  *
994  * When a tree block is COWed through a tree, there are four cases:
995  *
996  * The reference count of the block is one and the tree is the block's
997  * owner tree. Nothing to do in this case.
998  *
999  * The reference count of the block is one and the tree is not the
1000  * block's owner tree. In this case, full back refs is used for pointers
1001  * in the block. Remove these full back refs, add implicit back refs for
1002  * every pointers in the new block.
1003  *
1004  * The reference count of the block is greater than one and the tree is
1005  * the block's owner tree. In this case, implicit back refs is used for
1006  * pointers in the block. Add full back refs for every pointers in the
1007  * block, increase lower level extents' reference counts. The original
1008  * implicit back refs are entailed to the new block.
1009  *
1010  * The reference count of the block is greater than one and the tree is
1011  * not the block's owner tree. Add implicit back refs for every pointer in
1012  * the new block, increase lower level extents' reference count.
1013  *
1014  * Back Reference Key composing:
1015  *
1016  * The key objectid corresponds to the first byte in the extent,
1017  * The key type is used to differentiate between types of back refs.
1018  * There are different meanings of the key offset for different types
1019  * of back refs.
1020  *
1021  * File extents can be referenced by:
1022  *
1023  * - multiple snapshots, subvolumes, or different generations in one subvol
1024  * - different files inside a single subvolume
1025  * - different offsets inside a file (bookend extents in file.c)
1026  *
1027  * The extent ref structure for the implicit back refs has fields for:
1028  *
1029  * - Objectid of the subvolume root
1030  * - objectid of the file holding the reference
1031  * - original offset in the file
1032  * - how many bookend extents
1033  *
1034  * The key offset for the implicit back refs is hash of the first
1035  * three fields.
1036  *
1037  * The extent ref structure for the full back refs has field for:
1038  *
1039  * - number of pointers in the tree leaf
1040  *
1041  * The key offset for the implicit back refs is the first byte of
1042  * the tree leaf
1043  *
1044  * When a file extent is allocated, The implicit back refs is used.
1045  * the fields are filled in:
1046  *
1047  *     (root_key.objectid, inode objectid, offset in file, 1)
1048  *
1049  * When a file extent is removed file truncation, we find the
1050  * corresponding implicit back refs and check the following fields:
1051  *
1052  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1053  *
1054  * Btree extents can be referenced by:
1055  *
1056  * - Different subvolumes
1057  *
1058  * Both the implicit back refs and the full back refs for tree blocks
1059  * only consist of key. The key offset for the implicit back refs is
1060  * objectid of block's owner tree. The key offset for the full back refs
1061  * is the first byte of parent block.
1062  *
1063  * When implicit back refs is used, information about the lowest key and
1064  * level of the tree block are required. These information are stored in
1065  * tree block info structure.
1066  */
1067
1068 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1069 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1070                                   struct btrfs_fs_info *fs_info,
1071                                   struct btrfs_path *path,
1072                                   u64 owner, u32 extra_size)
1073 {
1074         struct btrfs_root *root = fs_info->extent_root;
1075         struct btrfs_extent_item *item;
1076         struct btrfs_extent_item_v0 *ei0;
1077         struct btrfs_extent_ref_v0 *ref0;
1078         struct btrfs_tree_block_info *bi;
1079         struct extent_buffer *leaf;
1080         struct btrfs_key key;
1081         struct btrfs_key found_key;
1082         u32 new_size = sizeof(*item);
1083         u64 refs;
1084         int ret;
1085
1086         leaf = path->nodes[0];
1087         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1088
1089         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1090         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1091                              struct btrfs_extent_item_v0);
1092         refs = btrfs_extent_refs_v0(leaf, ei0);
1093
1094         if (owner == (u64)-1) {
1095                 while (1) {
1096                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1097                                 ret = btrfs_next_leaf(root, path);
1098                                 if (ret < 0)
1099                                         return ret;
1100                                 BUG_ON(ret > 0); /* Corruption */
1101                                 leaf = path->nodes[0];
1102                         }
1103                         btrfs_item_key_to_cpu(leaf, &found_key,
1104                                               path->slots[0]);
1105                         BUG_ON(key.objectid != found_key.objectid);
1106                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1107                                 path->slots[0]++;
1108                                 continue;
1109                         }
1110                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1111                                               struct btrfs_extent_ref_v0);
1112                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1113                         break;
1114                 }
1115         }
1116         btrfs_release_path(path);
1117
1118         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1119                 new_size += sizeof(*bi);
1120
1121         new_size -= sizeof(*ei0);
1122         ret = btrfs_search_slot(trans, root, &key, path,
1123                                 new_size + extra_size, 1);
1124         if (ret < 0)
1125                 return ret;
1126         BUG_ON(ret); /* Corruption */
1127
1128         btrfs_extend_item(fs_info, path, new_size);
1129
1130         leaf = path->nodes[0];
1131         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1132         btrfs_set_extent_refs(leaf, item, refs);
1133         /* FIXME: get real generation */
1134         btrfs_set_extent_generation(leaf, item, 0);
1135         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1136                 btrfs_set_extent_flags(leaf, item,
1137                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1138                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1139                 bi = (struct btrfs_tree_block_info *)(item + 1);
1140                 /* FIXME: get first key of the block */
1141                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1142                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1143         } else {
1144                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1145         }
1146         btrfs_mark_buffer_dirty(leaf);
1147         return 0;
1148 }
1149 #endif
1150
1151 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1152 {
1153         u32 high_crc = ~(u32)0;
1154         u32 low_crc = ~(u32)0;
1155         __le64 lenum;
1156
1157         lenum = cpu_to_le64(root_objectid);
1158         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1159         lenum = cpu_to_le64(owner);
1160         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1161         lenum = cpu_to_le64(offset);
1162         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1163
1164         return ((u64)high_crc << 31) ^ (u64)low_crc;
1165 }
1166
1167 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1168                                      struct btrfs_extent_data_ref *ref)
1169 {
1170         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1171                                     btrfs_extent_data_ref_objectid(leaf, ref),
1172                                     btrfs_extent_data_ref_offset(leaf, ref));
1173 }
1174
1175 static int match_extent_data_ref(struct extent_buffer *leaf,
1176                                  struct btrfs_extent_data_ref *ref,
1177                                  u64 root_objectid, u64 owner, u64 offset)
1178 {
1179         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1180             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1181             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1182                 return 0;
1183         return 1;
1184 }
1185
1186 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1187                                            struct btrfs_fs_info *fs_info,
1188                                            struct btrfs_path *path,
1189                                            u64 bytenr, u64 parent,
1190                                            u64 root_objectid,
1191                                            u64 owner, u64 offset)
1192 {
1193         struct btrfs_root *root = fs_info->extent_root;
1194         struct btrfs_key key;
1195         struct btrfs_extent_data_ref *ref;
1196         struct extent_buffer *leaf;
1197         u32 nritems;
1198         int ret;
1199         int recow;
1200         int err = -ENOENT;
1201
1202         key.objectid = bytenr;
1203         if (parent) {
1204                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1205                 key.offset = parent;
1206         } else {
1207                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1208                 key.offset = hash_extent_data_ref(root_objectid,
1209                                                   owner, offset);
1210         }
1211 again:
1212         recow = 0;
1213         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1214         if (ret < 0) {
1215                 err = ret;
1216                 goto fail;
1217         }
1218
1219         if (parent) {
1220                 if (!ret)
1221                         return 0;
1222 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1223                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1224                 btrfs_release_path(path);
1225                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1226                 if (ret < 0) {
1227                         err = ret;
1228                         goto fail;
1229                 }
1230                 if (!ret)
1231                         return 0;
1232 #endif
1233                 goto fail;
1234         }
1235
1236         leaf = path->nodes[0];
1237         nritems = btrfs_header_nritems(leaf);
1238         while (1) {
1239                 if (path->slots[0] >= nritems) {
1240                         ret = btrfs_next_leaf(root, path);
1241                         if (ret < 0)
1242                                 err = ret;
1243                         if (ret)
1244                                 goto fail;
1245
1246                         leaf = path->nodes[0];
1247                         nritems = btrfs_header_nritems(leaf);
1248                         recow = 1;
1249                 }
1250
1251                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252                 if (key.objectid != bytenr ||
1253                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1254                         goto fail;
1255
1256                 ref = btrfs_item_ptr(leaf, path->slots[0],
1257                                      struct btrfs_extent_data_ref);
1258
1259                 if (match_extent_data_ref(leaf, ref, root_objectid,
1260                                           owner, offset)) {
1261                         if (recow) {
1262                                 btrfs_release_path(path);
1263                                 goto again;
1264                         }
1265                         err = 0;
1266                         break;
1267                 }
1268                 path->slots[0]++;
1269         }
1270 fail:
1271         return err;
1272 }
1273
1274 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1275                                            struct btrfs_fs_info *fs_info,
1276                                            struct btrfs_path *path,
1277                                            u64 bytenr, u64 parent,
1278                                            u64 root_objectid, u64 owner,
1279                                            u64 offset, int refs_to_add)
1280 {
1281         struct btrfs_root *root = fs_info->extent_root;
1282         struct btrfs_key key;
1283         struct extent_buffer *leaf;
1284         u32 size;
1285         u32 num_refs;
1286         int ret;
1287
1288         key.objectid = bytenr;
1289         if (parent) {
1290                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1291                 key.offset = parent;
1292                 size = sizeof(struct btrfs_shared_data_ref);
1293         } else {
1294                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1295                 key.offset = hash_extent_data_ref(root_objectid,
1296                                                   owner, offset);
1297                 size = sizeof(struct btrfs_extent_data_ref);
1298         }
1299
1300         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1301         if (ret && ret != -EEXIST)
1302                 goto fail;
1303
1304         leaf = path->nodes[0];
1305         if (parent) {
1306                 struct btrfs_shared_data_ref *ref;
1307                 ref = btrfs_item_ptr(leaf, path->slots[0],
1308                                      struct btrfs_shared_data_ref);
1309                 if (ret == 0) {
1310                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1311                 } else {
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1313                         num_refs += refs_to_add;
1314                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1315                 }
1316         } else {
1317                 struct btrfs_extent_data_ref *ref;
1318                 while (ret == -EEXIST) {
1319                         ref = btrfs_item_ptr(leaf, path->slots[0],
1320                                              struct btrfs_extent_data_ref);
1321                         if (match_extent_data_ref(leaf, ref, root_objectid,
1322                                                   owner, offset))
1323                                 break;
1324                         btrfs_release_path(path);
1325                         key.offset++;
1326                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1327                                                       size);
1328                         if (ret && ret != -EEXIST)
1329                                 goto fail;
1330
1331                         leaf = path->nodes[0];
1332                 }
1333                 ref = btrfs_item_ptr(leaf, path->slots[0],
1334                                      struct btrfs_extent_data_ref);
1335                 if (ret == 0) {
1336                         btrfs_set_extent_data_ref_root(leaf, ref,
1337                                                        root_objectid);
1338                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1339                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1340                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1341                 } else {
1342                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1343                         num_refs += refs_to_add;
1344                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1345                 }
1346         }
1347         btrfs_mark_buffer_dirty(leaf);
1348         ret = 0;
1349 fail:
1350         btrfs_release_path(path);
1351         return ret;
1352 }
1353
1354 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1355                                            struct btrfs_fs_info *fs_info,
1356                                            struct btrfs_path *path,
1357                                            int refs_to_drop, int *last_ref)
1358 {
1359         struct btrfs_key key;
1360         struct btrfs_extent_data_ref *ref1 = NULL;
1361         struct btrfs_shared_data_ref *ref2 = NULL;
1362         struct extent_buffer *leaf;
1363         u32 num_refs = 0;
1364         int ret = 0;
1365
1366         leaf = path->nodes[0];
1367         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1368
1369         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1370                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1371                                       struct btrfs_extent_data_ref);
1372                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1373         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1374                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1375                                       struct btrfs_shared_data_ref);
1376                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1377 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1378         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1379                 struct btrfs_extent_ref_v0 *ref0;
1380                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1381                                       struct btrfs_extent_ref_v0);
1382                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1383 #endif
1384         } else {
1385                 BUG();
1386         }
1387
1388         BUG_ON(num_refs < refs_to_drop);
1389         num_refs -= refs_to_drop;
1390
1391         if (num_refs == 0) {
1392                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1393                 *last_ref = 1;
1394         } else {
1395                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1396                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1397                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1398                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1399 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1400                 else {
1401                         struct btrfs_extent_ref_v0 *ref0;
1402                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1403                                         struct btrfs_extent_ref_v0);
1404                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1405                 }
1406 #endif
1407                 btrfs_mark_buffer_dirty(leaf);
1408         }
1409         return ret;
1410 }
1411
1412 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1413                                           struct btrfs_extent_inline_ref *iref)
1414 {
1415         struct btrfs_key key;
1416         struct extent_buffer *leaf;
1417         struct btrfs_extent_data_ref *ref1;
1418         struct btrfs_shared_data_ref *ref2;
1419         u32 num_refs = 0;
1420
1421         leaf = path->nodes[0];
1422         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1423         if (iref) {
1424                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1425                     BTRFS_EXTENT_DATA_REF_KEY) {
1426                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1427                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1428                 } else {
1429                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1430                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1431                 }
1432         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1433                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1434                                       struct btrfs_extent_data_ref);
1435                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1436         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1437                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1438                                       struct btrfs_shared_data_ref);
1439                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1440 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1441         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1442                 struct btrfs_extent_ref_v0 *ref0;
1443                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1444                                       struct btrfs_extent_ref_v0);
1445                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1446 #endif
1447         } else {
1448                 WARN_ON(1);
1449         }
1450         return num_refs;
1451 }
1452
1453 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1454                                           struct btrfs_fs_info *fs_info,
1455                                           struct btrfs_path *path,
1456                                           u64 bytenr, u64 parent,
1457                                           u64 root_objectid)
1458 {
1459         struct btrfs_root *root = fs_info->extent_root;
1460         struct btrfs_key key;
1461         int ret;
1462
1463         key.objectid = bytenr;
1464         if (parent) {
1465                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1466                 key.offset = parent;
1467         } else {
1468                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1469                 key.offset = root_objectid;
1470         }
1471
1472         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1473         if (ret > 0)
1474                 ret = -ENOENT;
1475 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1476         if (ret == -ENOENT && parent) {
1477                 btrfs_release_path(path);
1478                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1479                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1480                 if (ret > 0)
1481                         ret = -ENOENT;
1482         }
1483 #endif
1484         return ret;
1485 }
1486
1487 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1488                                           struct btrfs_fs_info *fs_info,
1489                                           struct btrfs_path *path,
1490                                           u64 bytenr, u64 parent,
1491                                           u64 root_objectid)
1492 {
1493         struct btrfs_key key;
1494         int ret;
1495
1496         key.objectid = bytenr;
1497         if (parent) {
1498                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1499                 key.offset = parent;
1500         } else {
1501                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1502                 key.offset = root_objectid;
1503         }
1504
1505         ret = btrfs_insert_empty_item(trans, fs_info->extent_root,
1506                                       path, &key, 0);
1507         btrfs_release_path(path);
1508         return ret;
1509 }
1510
1511 static inline int extent_ref_type(u64 parent, u64 owner)
1512 {
1513         int type;
1514         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1515                 if (parent > 0)
1516                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1517                 else
1518                         type = BTRFS_TREE_BLOCK_REF_KEY;
1519         } else {
1520                 if (parent > 0)
1521                         type = BTRFS_SHARED_DATA_REF_KEY;
1522                 else
1523                         type = BTRFS_EXTENT_DATA_REF_KEY;
1524         }
1525         return type;
1526 }
1527
1528 static int find_next_key(struct btrfs_path *path, int level,
1529                          struct btrfs_key *key)
1530
1531 {
1532         for (; level < BTRFS_MAX_LEVEL; level++) {
1533                 if (!path->nodes[level])
1534                         break;
1535                 if (path->slots[level] + 1 >=
1536                     btrfs_header_nritems(path->nodes[level]))
1537                         continue;
1538                 if (level == 0)
1539                         btrfs_item_key_to_cpu(path->nodes[level], key,
1540                                               path->slots[level] + 1);
1541                 else
1542                         btrfs_node_key_to_cpu(path->nodes[level], key,
1543                                               path->slots[level] + 1);
1544                 return 0;
1545         }
1546         return 1;
1547 }
1548
1549 /*
1550  * look for inline back ref. if back ref is found, *ref_ret is set
1551  * to the address of inline back ref, and 0 is returned.
1552  *
1553  * if back ref isn't found, *ref_ret is set to the address where it
1554  * should be inserted, and -ENOENT is returned.
1555  *
1556  * if insert is true and there are too many inline back refs, the path
1557  * points to the extent item, and -EAGAIN is returned.
1558  *
1559  * NOTE: inline back refs are ordered in the same way that back ref
1560  *       items in the tree are ordered.
1561  */
1562 static noinline_for_stack
1563 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1564                                  struct btrfs_fs_info *fs_info,
1565                                  struct btrfs_path *path,
1566                                  struct btrfs_extent_inline_ref **ref_ret,
1567                                  u64 bytenr, u64 num_bytes,
1568                                  u64 parent, u64 root_objectid,
1569                                  u64 owner, u64 offset, int insert)
1570 {
1571         struct btrfs_root *root = fs_info->extent_root;
1572         struct btrfs_key key;
1573         struct extent_buffer *leaf;
1574         struct btrfs_extent_item *ei;
1575         struct btrfs_extent_inline_ref *iref;
1576         u64 flags;
1577         u64 item_size;
1578         unsigned long ptr;
1579         unsigned long end;
1580         int extra_size;
1581         int type;
1582         int want;
1583         int ret;
1584         int err = 0;
1585         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1586
1587         key.objectid = bytenr;
1588         key.type = BTRFS_EXTENT_ITEM_KEY;
1589         key.offset = num_bytes;
1590
1591         want = extent_ref_type(parent, owner);
1592         if (insert) {
1593                 extra_size = btrfs_extent_inline_ref_size(want);
1594                 path->keep_locks = 1;
1595         } else
1596                 extra_size = -1;
1597
1598         /*
1599          * Owner is our parent level, so we can just add one to get the level
1600          * for the block we are interested in.
1601          */
1602         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1603                 key.type = BTRFS_METADATA_ITEM_KEY;
1604                 key.offset = owner;
1605         }
1606
1607 again:
1608         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1609         if (ret < 0) {
1610                 err = ret;
1611                 goto out;
1612         }
1613
1614         /*
1615          * We may be a newly converted file system which still has the old fat
1616          * extent entries for metadata, so try and see if we have one of those.
1617          */
1618         if (ret > 0 && skinny_metadata) {
1619                 skinny_metadata = false;
1620                 if (path->slots[0]) {
1621                         path->slots[0]--;
1622                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1623                                               path->slots[0]);
1624                         if (key.objectid == bytenr &&
1625                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1626                             key.offset == num_bytes)
1627                                 ret = 0;
1628                 }
1629                 if (ret) {
1630                         key.objectid = bytenr;
1631                         key.type = BTRFS_EXTENT_ITEM_KEY;
1632                         key.offset = num_bytes;
1633                         btrfs_release_path(path);
1634                         goto again;
1635                 }
1636         }
1637
1638         if (ret && !insert) {
1639                 err = -ENOENT;
1640                 goto out;
1641         } else if (WARN_ON(ret)) {
1642                 err = -EIO;
1643                 goto out;
1644         }
1645
1646         leaf = path->nodes[0];
1647         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1648 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1649         if (item_size < sizeof(*ei)) {
1650                 if (!insert) {
1651                         err = -ENOENT;
1652                         goto out;
1653                 }
1654                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1655                                              extra_size);
1656                 if (ret < 0) {
1657                         err = ret;
1658                         goto out;
1659                 }
1660                 leaf = path->nodes[0];
1661                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1662         }
1663 #endif
1664         BUG_ON(item_size < sizeof(*ei));
1665
1666         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1667         flags = btrfs_extent_flags(leaf, ei);
1668
1669         ptr = (unsigned long)(ei + 1);
1670         end = (unsigned long)ei + item_size;
1671
1672         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1673                 ptr += sizeof(struct btrfs_tree_block_info);
1674                 BUG_ON(ptr > end);
1675         }
1676
1677         err = -ENOENT;
1678         while (1) {
1679                 if (ptr >= end) {
1680                         WARN_ON(ptr > end);
1681                         break;
1682                 }
1683                 iref = (struct btrfs_extent_inline_ref *)ptr;
1684                 type = btrfs_extent_inline_ref_type(leaf, iref);
1685                 if (want < type)
1686                         break;
1687                 if (want > type) {
1688                         ptr += btrfs_extent_inline_ref_size(type);
1689                         continue;
1690                 }
1691
1692                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1693                         struct btrfs_extent_data_ref *dref;
1694                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1695                         if (match_extent_data_ref(leaf, dref, root_objectid,
1696                                                   owner, offset)) {
1697                                 err = 0;
1698                                 break;
1699                         }
1700                         if (hash_extent_data_ref_item(leaf, dref) <
1701                             hash_extent_data_ref(root_objectid, owner, offset))
1702                                 break;
1703                 } else {
1704                         u64 ref_offset;
1705                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1706                         if (parent > 0) {
1707                                 if (parent == ref_offset) {
1708                                         err = 0;
1709                                         break;
1710                                 }
1711                                 if (ref_offset < parent)
1712                                         break;
1713                         } else {
1714                                 if (root_objectid == ref_offset) {
1715                                         err = 0;
1716                                         break;
1717                                 }
1718                                 if (ref_offset < root_objectid)
1719                                         break;
1720                         }
1721                 }
1722                 ptr += btrfs_extent_inline_ref_size(type);
1723         }
1724         if (err == -ENOENT && insert) {
1725                 if (item_size + extra_size >=
1726                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1727                         err = -EAGAIN;
1728                         goto out;
1729                 }
1730                 /*
1731                  * To add new inline back ref, we have to make sure
1732                  * there is no corresponding back ref item.
1733                  * For simplicity, we just do not add new inline back
1734                  * ref if there is any kind of item for this block
1735                  */
1736                 if (find_next_key(path, 0, &key) == 0 &&
1737                     key.objectid == bytenr &&
1738                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1739                         err = -EAGAIN;
1740                         goto out;
1741                 }
1742         }
1743         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1744 out:
1745         if (insert) {
1746                 path->keep_locks = 0;
1747                 btrfs_unlock_up_safe(path, 1);
1748         }
1749         return err;
1750 }
1751
1752 /*
1753  * helper to add new inline back ref
1754  */
1755 static noinline_for_stack
1756 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1757                                  struct btrfs_path *path,
1758                                  struct btrfs_extent_inline_ref *iref,
1759                                  u64 parent, u64 root_objectid,
1760                                  u64 owner, u64 offset, int refs_to_add,
1761                                  struct btrfs_delayed_extent_op *extent_op)
1762 {
1763         struct extent_buffer *leaf;
1764         struct btrfs_extent_item *ei;
1765         unsigned long ptr;
1766         unsigned long end;
1767         unsigned long item_offset;
1768         u64 refs;
1769         int size;
1770         int type;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         item_offset = (unsigned long)iref - (unsigned long)ei;
1775
1776         type = extent_ref_type(parent, owner);
1777         size = btrfs_extent_inline_ref_size(type);
1778
1779         btrfs_extend_item(fs_info, path, size);
1780
1781         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1782         refs = btrfs_extent_refs(leaf, ei);
1783         refs += refs_to_add;
1784         btrfs_set_extent_refs(leaf, ei, refs);
1785         if (extent_op)
1786                 __run_delayed_extent_op(extent_op, leaf, ei);
1787
1788         ptr = (unsigned long)ei + item_offset;
1789         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1790         if (ptr < end - size)
1791                 memmove_extent_buffer(leaf, ptr + size, ptr,
1792                                       end - size - ptr);
1793
1794         iref = (struct btrfs_extent_inline_ref *)ptr;
1795         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1796         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1797                 struct btrfs_extent_data_ref *dref;
1798                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1799                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1800                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1801                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1802                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1803         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1804                 struct btrfs_shared_data_ref *sref;
1805                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1806                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1807                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1808         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1809                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1810         } else {
1811                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1812         }
1813         btrfs_mark_buffer_dirty(leaf);
1814 }
1815
1816 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1817                                  struct btrfs_fs_info *fs_info,
1818                                  struct btrfs_path *path,
1819                                  struct btrfs_extent_inline_ref **ref_ret,
1820                                  u64 bytenr, u64 num_bytes, u64 parent,
1821                                  u64 root_objectid, u64 owner, u64 offset)
1822 {
1823         int ret;
1824
1825         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1826                                            bytenr, num_bytes, parent,
1827                                            root_objectid, owner, offset, 0);
1828         if (ret != -ENOENT)
1829                 return ret;
1830
1831         btrfs_release_path(path);
1832         *ref_ret = NULL;
1833
1834         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1835                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1836                                             parent, root_objectid);
1837         } else {
1838                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1839                                              parent, root_objectid, owner,
1840                                              offset);
1841         }
1842         return ret;
1843 }
1844
1845 /*
1846  * helper to update/remove inline back ref
1847  */
1848 static noinline_for_stack
1849 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1850                                   struct btrfs_path *path,
1851                                   struct btrfs_extent_inline_ref *iref,
1852                                   int refs_to_mod,
1853                                   struct btrfs_delayed_extent_op *extent_op,
1854                                   int *last_ref)
1855 {
1856         struct extent_buffer *leaf;
1857         struct btrfs_extent_item *ei;
1858         struct btrfs_extent_data_ref *dref = NULL;
1859         struct btrfs_shared_data_ref *sref = NULL;
1860         unsigned long ptr;
1861         unsigned long end;
1862         u32 item_size;
1863         int size;
1864         int type;
1865         u64 refs;
1866
1867         leaf = path->nodes[0];
1868         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1869         refs = btrfs_extent_refs(leaf, ei);
1870         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1871         refs += refs_to_mod;
1872         btrfs_set_extent_refs(leaf, ei, refs);
1873         if (extent_op)
1874                 __run_delayed_extent_op(extent_op, leaf, ei);
1875
1876         type = btrfs_extent_inline_ref_type(leaf, iref);
1877
1878         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1879                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1880                 refs = btrfs_extent_data_ref_count(leaf, dref);
1881         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1882                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1883                 refs = btrfs_shared_data_ref_count(leaf, sref);
1884         } else {
1885                 refs = 1;
1886                 BUG_ON(refs_to_mod != -1);
1887         }
1888
1889         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1890         refs += refs_to_mod;
1891
1892         if (refs > 0) {
1893                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1894                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1895                 else
1896                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1897         } else {
1898                 *last_ref = 1;
1899                 size =  btrfs_extent_inline_ref_size(type);
1900                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1901                 ptr = (unsigned long)iref;
1902                 end = (unsigned long)ei + item_size;
1903                 if (ptr + size < end)
1904                         memmove_extent_buffer(leaf, ptr, ptr + size,
1905                                               end - ptr - size);
1906                 item_size -= size;
1907                 btrfs_truncate_item(fs_info, path, item_size, 1);
1908         }
1909         btrfs_mark_buffer_dirty(leaf);
1910 }
1911
1912 static noinline_for_stack
1913 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1914                                  struct btrfs_fs_info *fs_info,
1915                                  struct btrfs_path *path,
1916                                  u64 bytenr, u64 num_bytes, u64 parent,
1917                                  u64 root_objectid, u64 owner,
1918                                  u64 offset, int refs_to_add,
1919                                  struct btrfs_delayed_extent_op *extent_op)
1920 {
1921         struct btrfs_extent_inline_ref *iref;
1922         int ret;
1923
1924         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1925                                            bytenr, num_bytes, parent,
1926                                            root_objectid, owner, offset, 1);
1927         if (ret == 0) {
1928                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1929                 update_inline_extent_backref(fs_info, path, iref,
1930                                              refs_to_add, extent_op, NULL);
1931         } else if (ret == -ENOENT) {
1932                 setup_inline_extent_backref(fs_info, path, iref, parent,
1933                                             root_objectid, owner, offset,
1934                                             refs_to_add, extent_op);
1935                 ret = 0;
1936         }
1937         return ret;
1938 }
1939
1940 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1941                                  struct btrfs_fs_info *fs_info,
1942                                  struct btrfs_path *path,
1943                                  u64 bytenr, u64 parent, u64 root_objectid,
1944                                  u64 owner, u64 offset, int refs_to_add)
1945 {
1946         int ret;
1947         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1948                 BUG_ON(refs_to_add != 1);
1949                 ret = insert_tree_block_ref(trans, fs_info, path, bytenr,
1950                                             parent, root_objectid);
1951         } else {
1952                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
1953                                              parent, root_objectid,
1954                                              owner, offset, refs_to_add);
1955         }
1956         return ret;
1957 }
1958
1959 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1960                                  struct btrfs_fs_info *fs_info,
1961                                  struct btrfs_path *path,
1962                                  struct btrfs_extent_inline_ref *iref,
1963                                  int refs_to_drop, int is_data, int *last_ref)
1964 {
1965         int ret = 0;
1966
1967         BUG_ON(!is_data && refs_to_drop != 1);
1968         if (iref) {
1969                 update_inline_extent_backref(fs_info, path, iref,
1970                                              -refs_to_drop, NULL, last_ref);
1971         } else if (is_data) {
1972                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
1973                                              last_ref);
1974         } else {
1975                 *last_ref = 1;
1976                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1977         }
1978         return ret;
1979 }
1980
1981 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1982 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1983                                u64 *discarded_bytes)
1984 {
1985         int j, ret = 0;
1986         u64 bytes_left, end;
1987         u64 aligned_start = ALIGN(start, 1 << 9);
1988
1989         if (WARN_ON(start != aligned_start)) {
1990                 len -= aligned_start - start;
1991                 len = round_down(len, 1 << 9);
1992                 start = aligned_start;
1993         }
1994
1995         *discarded_bytes = 0;
1996
1997         if (!len)
1998                 return 0;
1999
2000         end = start + len;
2001         bytes_left = len;
2002
2003         /* Skip any superblocks on this device. */
2004         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2005                 u64 sb_start = btrfs_sb_offset(j);
2006                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2007                 u64 size = sb_start - start;
2008
2009                 if (!in_range(sb_start, start, bytes_left) &&
2010                     !in_range(sb_end, start, bytes_left) &&
2011                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2012                         continue;
2013
2014                 /*
2015                  * Superblock spans beginning of range.  Adjust start and
2016                  * try again.
2017                  */
2018                 if (sb_start <= start) {
2019                         start += sb_end - start;
2020                         if (start > end) {
2021                                 bytes_left = 0;
2022                                 break;
2023                         }
2024                         bytes_left = end - start;
2025                         continue;
2026                 }
2027
2028                 if (size) {
2029                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2030                                                    GFP_NOFS, 0);
2031                         if (!ret)
2032                                 *discarded_bytes += size;
2033                         else if (ret != -EOPNOTSUPP)
2034                                 return ret;
2035                 }
2036
2037                 start = sb_end;
2038                 if (start > end) {
2039                         bytes_left = 0;
2040                         break;
2041                 }
2042                 bytes_left = end - start;
2043         }
2044
2045         if (bytes_left) {
2046                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2047                                            GFP_NOFS, 0);
2048                 if (!ret)
2049                         *discarded_bytes += bytes_left;
2050         }
2051         return ret;
2052 }
2053
2054 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2055                          u64 num_bytes, u64 *actual_bytes)
2056 {
2057         int ret;
2058         u64 discarded_bytes = 0;
2059         struct btrfs_bio *bbio = NULL;
2060
2061
2062         /*
2063          * Avoid races with device replace and make sure our bbio has devices
2064          * associated to its stripes that don't go away while we are discarding.
2065          */
2066         btrfs_bio_counter_inc_blocked(fs_info);
2067         /* Tell the block device(s) that the sectors can be discarded */
2068         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2069                               &bbio, 0);
2070         /* Error condition is -ENOMEM */
2071         if (!ret) {
2072                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2073                 int i;
2074
2075
2076                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2077                         u64 bytes;
2078                         if (!stripe->dev->can_discard)
2079                                 continue;
2080
2081                         ret = btrfs_issue_discard(stripe->dev->bdev,
2082                                                   stripe->physical,
2083                                                   stripe->length,
2084                                                   &bytes);
2085                         if (!ret)
2086                                 discarded_bytes += bytes;
2087                         else if (ret != -EOPNOTSUPP)
2088                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2089
2090                         /*
2091                          * Just in case we get back EOPNOTSUPP for some reason,
2092                          * just ignore the return value so we don't screw up
2093                          * people calling discard_extent.
2094                          */
2095                         ret = 0;
2096                 }
2097                 btrfs_put_bbio(bbio);
2098         }
2099         btrfs_bio_counter_dec(fs_info);
2100
2101         if (actual_bytes)
2102                 *actual_bytes = discarded_bytes;
2103
2104
2105         if (ret == -EOPNOTSUPP)
2106                 ret = 0;
2107         return ret;
2108 }
2109
2110 /* Can return -ENOMEM */
2111 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2112                          struct btrfs_fs_info *fs_info,
2113                          u64 bytenr, u64 num_bytes, u64 parent,
2114                          u64 root_objectid, u64 owner, u64 offset)
2115 {
2116         int old_ref_mod, new_ref_mod;
2117         int ret;
2118
2119         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2120                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2121
2122         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2123                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2124                                                  num_bytes, parent,
2125                                                  root_objectid, (int)owner,
2126                                                  BTRFS_ADD_DELAYED_REF, NULL,
2127                                                  &old_ref_mod, &new_ref_mod);
2128         } else {
2129                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2130                                                  num_bytes, parent,
2131                                                  root_objectid, owner, offset,
2132                                                  0, BTRFS_ADD_DELAYED_REF,
2133                                                  &old_ref_mod, &new_ref_mod);
2134         }
2135
2136         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0)
2137                 add_pinned_bytes(fs_info, -num_bytes, owner, root_objectid);
2138
2139         return ret;
2140 }
2141
2142 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2143                                   struct btrfs_fs_info *fs_info,
2144                                   struct btrfs_delayed_ref_node *node,
2145                                   u64 parent, u64 root_objectid,
2146                                   u64 owner, u64 offset, int refs_to_add,
2147                                   struct btrfs_delayed_extent_op *extent_op)
2148 {
2149         struct btrfs_path *path;
2150         struct extent_buffer *leaf;
2151         struct btrfs_extent_item *item;
2152         struct btrfs_key key;
2153         u64 bytenr = node->bytenr;
2154         u64 num_bytes = node->num_bytes;
2155         u64 refs;
2156         int ret;
2157
2158         path = btrfs_alloc_path();
2159         if (!path)
2160                 return -ENOMEM;
2161
2162         path->reada = READA_FORWARD;
2163         path->leave_spinning = 1;
2164         /* this will setup the path even if it fails to insert the back ref */
2165         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2166                                            num_bytes, parent, root_objectid,
2167                                            owner, offset,
2168                                            refs_to_add, extent_op);
2169         if ((ret < 0 && ret != -EAGAIN) || !ret)
2170                 goto out;
2171
2172         /*
2173          * Ok we had -EAGAIN which means we didn't have space to insert and
2174          * inline extent ref, so just update the reference count and add a
2175          * normal backref.
2176          */
2177         leaf = path->nodes[0];
2178         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2179         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2180         refs = btrfs_extent_refs(leaf, item);
2181         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2182         if (extent_op)
2183                 __run_delayed_extent_op(extent_op, leaf, item);
2184
2185         btrfs_mark_buffer_dirty(leaf);
2186         btrfs_release_path(path);
2187
2188         path->reada = READA_FORWARD;
2189         path->leave_spinning = 1;
2190         /* now insert the actual backref */
2191         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2192                                     root_objectid, owner, offset, refs_to_add);
2193         if (ret)
2194                 btrfs_abort_transaction(trans, ret);
2195 out:
2196         btrfs_free_path(path);
2197         return ret;
2198 }
2199
2200 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2201                                 struct btrfs_fs_info *fs_info,
2202                                 struct btrfs_delayed_ref_node *node,
2203                                 struct btrfs_delayed_extent_op *extent_op,
2204                                 int insert_reserved)
2205 {
2206         int ret = 0;
2207         struct btrfs_delayed_data_ref *ref;
2208         struct btrfs_key ins;
2209         u64 parent = 0;
2210         u64 ref_root = 0;
2211         u64 flags = 0;
2212
2213         ins.objectid = node->bytenr;
2214         ins.offset = node->num_bytes;
2215         ins.type = BTRFS_EXTENT_ITEM_KEY;
2216
2217         ref = btrfs_delayed_node_to_data_ref(node);
2218         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2219
2220         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2221                 parent = ref->parent;
2222         ref_root = ref->root;
2223
2224         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2225                 if (extent_op)
2226                         flags |= extent_op->flags_to_set;
2227                 ret = alloc_reserved_file_extent(trans, fs_info,
2228                                                  parent, ref_root, flags,
2229                                                  ref->objectid, ref->offset,
2230                                                  &ins, node->ref_mod);
2231         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2232                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2233                                              ref_root, ref->objectid,
2234                                              ref->offset, node->ref_mod,
2235                                              extent_op);
2236         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2237                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2238                                           ref_root, ref->objectid,
2239                                           ref->offset, node->ref_mod,
2240                                           extent_op);
2241         } else {
2242                 BUG();
2243         }
2244         return ret;
2245 }
2246
2247 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2248                                     struct extent_buffer *leaf,
2249                                     struct btrfs_extent_item *ei)
2250 {
2251         u64 flags = btrfs_extent_flags(leaf, ei);
2252         if (extent_op->update_flags) {
2253                 flags |= extent_op->flags_to_set;
2254                 btrfs_set_extent_flags(leaf, ei, flags);
2255         }
2256
2257         if (extent_op->update_key) {
2258                 struct btrfs_tree_block_info *bi;
2259                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2260                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2261                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2262         }
2263 }
2264
2265 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2266                                  struct btrfs_fs_info *fs_info,
2267                                  struct btrfs_delayed_ref_node *node,
2268                                  struct btrfs_delayed_extent_op *extent_op)
2269 {
2270         struct btrfs_key key;
2271         struct btrfs_path *path;
2272         struct btrfs_extent_item *ei;
2273         struct extent_buffer *leaf;
2274         u32 item_size;
2275         int ret;
2276         int err = 0;
2277         int metadata = !extent_op->is_data;
2278
2279         if (trans->aborted)
2280                 return 0;
2281
2282         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2283                 metadata = 0;
2284
2285         path = btrfs_alloc_path();
2286         if (!path)
2287                 return -ENOMEM;
2288
2289         key.objectid = node->bytenr;
2290
2291         if (metadata) {
2292                 key.type = BTRFS_METADATA_ITEM_KEY;
2293                 key.offset = extent_op->level;
2294         } else {
2295                 key.type = BTRFS_EXTENT_ITEM_KEY;
2296                 key.offset = node->num_bytes;
2297         }
2298
2299 again:
2300         path->reada = READA_FORWARD;
2301         path->leave_spinning = 1;
2302         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2303         if (ret < 0) {
2304                 err = ret;
2305                 goto out;
2306         }
2307         if (ret > 0) {
2308                 if (metadata) {
2309                         if (path->slots[0] > 0) {
2310                                 path->slots[0]--;
2311                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2312                                                       path->slots[0]);
2313                                 if (key.objectid == node->bytenr &&
2314                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2315                                     key.offset == node->num_bytes)
2316                                         ret = 0;
2317                         }
2318                         if (ret > 0) {
2319                                 btrfs_release_path(path);
2320                                 metadata = 0;
2321
2322                                 key.objectid = node->bytenr;
2323                                 key.offset = node->num_bytes;
2324                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2325                                 goto again;
2326                         }
2327                 } else {
2328                         err = -EIO;
2329                         goto out;
2330                 }
2331         }
2332
2333         leaf = path->nodes[0];
2334         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2335 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2336         if (item_size < sizeof(*ei)) {
2337                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2338                 if (ret < 0) {
2339                         err = ret;
2340                         goto out;
2341                 }
2342                 leaf = path->nodes[0];
2343                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2344         }
2345 #endif
2346         BUG_ON(item_size < sizeof(*ei));
2347         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2348         __run_delayed_extent_op(extent_op, leaf, ei);
2349
2350         btrfs_mark_buffer_dirty(leaf);
2351 out:
2352         btrfs_free_path(path);
2353         return err;
2354 }
2355
2356 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2357                                 struct btrfs_fs_info *fs_info,
2358                                 struct btrfs_delayed_ref_node *node,
2359                                 struct btrfs_delayed_extent_op *extent_op,
2360                                 int insert_reserved)
2361 {
2362         int ret = 0;
2363         struct btrfs_delayed_tree_ref *ref;
2364         struct btrfs_key ins;
2365         u64 parent = 0;
2366         u64 ref_root = 0;
2367         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
2368
2369         ref = btrfs_delayed_node_to_tree_ref(node);
2370         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2371
2372         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2373                 parent = ref->parent;
2374         ref_root = ref->root;
2375
2376         ins.objectid = node->bytenr;
2377         if (skinny_metadata) {
2378                 ins.offset = ref->level;
2379                 ins.type = BTRFS_METADATA_ITEM_KEY;
2380         } else {
2381                 ins.offset = node->num_bytes;
2382                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2383         }
2384
2385         if (node->ref_mod != 1) {
2386                 btrfs_err(fs_info,
2387         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2388                           node->bytenr, node->ref_mod, node->action, ref_root,
2389                           parent);
2390                 return -EIO;
2391         }
2392         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2393                 BUG_ON(!extent_op || !extent_op->update_flags);
2394                 ret = alloc_reserved_tree_block(trans, fs_info,
2395                                                 parent, ref_root,
2396                                                 extent_op->flags_to_set,
2397                                                 &extent_op->key,
2398                                                 ref->level, &ins);
2399         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2400                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2401                                              parent, ref_root,
2402                                              ref->level, 0, 1,
2403                                              extent_op);
2404         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2405                 ret = __btrfs_free_extent(trans, fs_info, node,
2406                                           parent, ref_root,
2407                                           ref->level, 0, 1, extent_op);
2408         } else {
2409                 BUG();
2410         }
2411         return ret;
2412 }
2413
2414 /* helper function to actually process a single delayed ref entry */
2415 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2416                                struct btrfs_fs_info *fs_info,
2417                                struct btrfs_delayed_ref_node *node,
2418                                struct btrfs_delayed_extent_op *extent_op,
2419                                int insert_reserved)
2420 {
2421         int ret = 0;
2422
2423         if (trans->aborted) {
2424                 if (insert_reserved)
2425                         btrfs_pin_extent(fs_info, node->bytenr,
2426                                          node->num_bytes, 1);
2427                 return 0;
2428         }
2429
2430         if (btrfs_delayed_ref_is_head(node)) {
2431                 struct btrfs_delayed_ref_head *head;
2432                 /*
2433                  * we've hit the end of the chain and we were supposed
2434                  * to insert this extent into the tree.  But, it got
2435                  * deleted before we ever needed to insert it, so all
2436                  * we have to do is clean up the accounting
2437                  */
2438                 BUG_ON(extent_op);
2439                 head = btrfs_delayed_node_to_head(node);
2440                 trace_run_delayed_ref_head(fs_info, node, head, node->action);
2441
2442                 if (head->total_ref_mod < 0) {
2443                         struct btrfs_block_group_cache *cache;
2444
2445                         cache = btrfs_lookup_block_group(fs_info, node->bytenr);
2446                         ASSERT(cache);
2447                         percpu_counter_add(&cache->space_info->total_bytes_pinned,
2448                                            -node->num_bytes);
2449                         btrfs_put_block_group(cache);
2450                 }
2451
2452                 if (insert_reserved) {
2453                         btrfs_pin_extent(fs_info, node->bytenr,
2454                                          node->num_bytes, 1);
2455                         if (head->is_data) {
2456                                 ret = btrfs_del_csums(trans, fs_info,
2457                                                       node->bytenr,
2458                                                       node->num_bytes);
2459                         }
2460                 }
2461
2462                 /* Also free its reserved qgroup space */
2463                 btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2464                                               head->qgroup_reserved);
2465                 return ret;
2466         }
2467
2468         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2469             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2470                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2471                                            insert_reserved);
2472         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2473                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2474                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2475                                            insert_reserved);
2476         else
2477                 BUG();
2478         return ret;
2479 }
2480
2481 static inline struct btrfs_delayed_ref_node *
2482 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2483 {
2484         struct btrfs_delayed_ref_node *ref;
2485
2486         if (list_empty(&head->ref_list))
2487                 return NULL;
2488
2489         /*
2490          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2491          * This is to prevent a ref count from going down to zero, which deletes
2492          * the extent item from the extent tree, when there still are references
2493          * to add, which would fail because they would not find the extent item.
2494          */
2495         if (!list_empty(&head->ref_add_list))
2496                 return list_first_entry(&head->ref_add_list,
2497                                 struct btrfs_delayed_ref_node, add_list);
2498
2499         ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
2500                                list);
2501         ASSERT(list_empty(&ref->add_list));
2502         return ref;
2503 }
2504
2505 /*
2506  * Returns 0 on success or if called with an already aborted transaction.
2507  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2508  */
2509 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2510                                              struct btrfs_fs_info *fs_info,
2511                                              unsigned long nr)
2512 {
2513         struct btrfs_delayed_ref_root *delayed_refs;
2514         struct btrfs_delayed_ref_node *ref;
2515         struct btrfs_delayed_ref_head *locked_ref = NULL;
2516         struct btrfs_delayed_extent_op *extent_op;
2517         ktime_t start = ktime_get();
2518         int ret;
2519         unsigned long count = 0;
2520         unsigned long actual_count = 0;
2521         int must_insert_reserved = 0;
2522
2523         delayed_refs = &trans->transaction->delayed_refs;
2524         while (1) {
2525                 if (!locked_ref) {
2526                         if (count >= nr)
2527                                 break;
2528
2529                         spin_lock(&delayed_refs->lock);
2530                         locked_ref = btrfs_select_ref_head(trans);
2531                         if (!locked_ref) {
2532                                 spin_unlock(&delayed_refs->lock);
2533                                 break;
2534                         }
2535
2536                         /* grab the lock that says we are going to process
2537                          * all the refs for this head */
2538                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2539                         spin_unlock(&delayed_refs->lock);
2540                         /*
2541                          * we may have dropped the spin lock to get the head
2542                          * mutex lock, and that might have given someone else
2543                          * time to free the head.  If that's true, it has been
2544                          * removed from our list and we can move on.
2545                          */
2546                         if (ret == -EAGAIN) {
2547                                 locked_ref = NULL;
2548                                 count++;
2549                                 continue;
2550                         }
2551                 }
2552
2553                 /*
2554                  * We need to try and merge add/drops of the same ref since we
2555                  * can run into issues with relocate dropping the implicit ref
2556                  * and then it being added back again before the drop can
2557                  * finish.  If we merged anything we need to re-loop so we can
2558                  * get a good ref.
2559                  * Or we can get node references of the same type that weren't
2560                  * merged when created due to bumps in the tree mod seq, and
2561                  * we need to merge them to prevent adding an inline extent
2562                  * backref before dropping it (triggering a BUG_ON at
2563                  * insert_inline_extent_backref()).
2564                  */
2565                 spin_lock(&locked_ref->lock);
2566                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2567                                          locked_ref);
2568
2569                 /*
2570                  * locked_ref is the head node, so we have to go one
2571                  * node back for any delayed ref updates
2572                  */
2573                 ref = select_delayed_ref(locked_ref);
2574
2575                 if (ref && ref->seq &&
2576                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2577                         spin_unlock(&locked_ref->lock);
2578                         spin_lock(&delayed_refs->lock);
2579                         locked_ref->processing = 0;
2580                         delayed_refs->num_heads_ready++;
2581                         spin_unlock(&delayed_refs->lock);
2582                         btrfs_delayed_ref_unlock(locked_ref);
2583                         locked_ref = NULL;
2584                         cond_resched();
2585                         count++;
2586                         continue;
2587                 }
2588
2589                 /*
2590                  * record the must insert reserved flag before we
2591                  * drop the spin lock.
2592                  */
2593                 must_insert_reserved = locked_ref->must_insert_reserved;
2594                 locked_ref->must_insert_reserved = 0;
2595
2596                 extent_op = locked_ref->extent_op;
2597                 locked_ref->extent_op = NULL;
2598
2599                 if (!ref) {
2600
2601
2602                         /* All delayed refs have been processed, Go ahead
2603                          * and send the head node to run_one_delayed_ref,
2604                          * so that any accounting fixes can happen
2605                          */
2606                         ref = &locked_ref->node;
2607
2608                         if (extent_op && must_insert_reserved) {
2609                                 btrfs_free_delayed_extent_op(extent_op);
2610                                 extent_op = NULL;
2611                         }
2612
2613                         if (extent_op) {
2614                                 spin_unlock(&locked_ref->lock);
2615                                 ret = run_delayed_extent_op(trans, fs_info,
2616                                                             ref, extent_op);
2617                                 btrfs_free_delayed_extent_op(extent_op);
2618
2619                                 if (ret) {
2620                                         /*
2621                                          * Need to reset must_insert_reserved if
2622                                          * there was an error so the abort stuff
2623                                          * can cleanup the reserved space
2624                                          * properly.
2625                                          */
2626                                         if (must_insert_reserved)
2627                                                 locked_ref->must_insert_reserved = 1;
2628                                         spin_lock(&delayed_refs->lock);
2629                                         locked_ref->processing = 0;
2630                                         delayed_refs->num_heads_ready++;
2631                                         spin_unlock(&delayed_refs->lock);
2632                                         btrfs_debug(fs_info,
2633                                                     "run_delayed_extent_op returned %d",
2634                                                     ret);
2635                                         btrfs_delayed_ref_unlock(locked_ref);
2636                                         return ret;
2637                                 }
2638                                 continue;
2639                         }
2640
2641                         /*
2642                          * Need to drop our head ref lock and re-acquire the
2643                          * delayed ref lock and then re-check to make sure
2644                          * nobody got added.
2645                          */
2646                         spin_unlock(&locked_ref->lock);
2647                         spin_lock(&delayed_refs->lock);
2648                         spin_lock(&locked_ref->lock);
2649                         if (!list_empty(&locked_ref->ref_list) ||
2650                             locked_ref->extent_op) {
2651                                 spin_unlock(&locked_ref->lock);
2652                                 spin_unlock(&delayed_refs->lock);
2653                                 continue;
2654                         }
2655                         ref->in_tree = 0;
2656                         delayed_refs->num_heads--;
2657                         rb_erase(&locked_ref->href_node,
2658                                  &delayed_refs->href_root);
2659                         spin_unlock(&delayed_refs->lock);
2660                 } else {
2661                         actual_count++;
2662                         ref->in_tree = 0;
2663                         list_del(&ref->list);
2664                         if (!list_empty(&ref->add_list))
2665                                 list_del(&ref->add_list);
2666                 }
2667                 atomic_dec(&delayed_refs->num_entries);
2668
2669                 if (!btrfs_delayed_ref_is_head(ref)) {
2670                         /*
2671                          * when we play the delayed ref, also correct the
2672                          * ref_mod on head
2673                          */
2674                         switch (ref->action) {
2675                         case BTRFS_ADD_DELAYED_REF:
2676                         case BTRFS_ADD_DELAYED_EXTENT:
2677                                 locked_ref->node.ref_mod -= ref->ref_mod;
2678                                 break;
2679                         case BTRFS_DROP_DELAYED_REF:
2680                                 locked_ref->node.ref_mod += ref->ref_mod;
2681                                 break;
2682                         default:
2683                                 WARN_ON(1);
2684                         }
2685                 }
2686                 spin_unlock(&locked_ref->lock);
2687
2688                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2689                                           must_insert_reserved);
2690
2691                 btrfs_free_delayed_extent_op(extent_op);
2692                 if (ret) {
2693                         spin_lock(&delayed_refs->lock);
2694                         locked_ref->processing = 0;
2695                         delayed_refs->num_heads_ready++;
2696                         spin_unlock(&delayed_refs->lock);
2697                         btrfs_delayed_ref_unlock(locked_ref);
2698                         btrfs_put_delayed_ref(ref);
2699                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2700                                     ret);
2701                         return ret;
2702                 }
2703
2704                 /*
2705                  * If this node is a head, that means all the refs in this head
2706                  * have been dealt with, and we will pick the next head to deal
2707                  * with, so we must unlock the head and drop it from the cluster
2708                  * list before we release it.
2709                  */
2710                 if (btrfs_delayed_ref_is_head(ref)) {
2711                         if (locked_ref->is_data &&
2712                             locked_ref->total_ref_mod < 0) {
2713                                 spin_lock(&delayed_refs->lock);
2714                                 delayed_refs->pending_csums -= ref->num_bytes;
2715                                 spin_unlock(&delayed_refs->lock);
2716                         }
2717                         btrfs_delayed_ref_unlock(locked_ref);
2718                         locked_ref = NULL;
2719                 }
2720                 btrfs_put_delayed_ref(ref);
2721                 count++;
2722                 cond_resched();
2723         }
2724
2725         /*
2726          * We don't want to include ref heads since we can have empty ref heads
2727          * and those will drastically skew our runtime down since we just do
2728          * accounting, no actual extent tree updates.
2729          */
2730         if (actual_count > 0) {
2731                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2732                 u64 avg;
2733
2734                 /*
2735                  * We weigh the current average higher than our current runtime
2736                  * to avoid large swings in the average.
2737                  */
2738                 spin_lock(&delayed_refs->lock);
2739                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2740                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2741                 spin_unlock(&delayed_refs->lock);
2742         }
2743         return 0;
2744 }
2745
2746 #ifdef SCRAMBLE_DELAYED_REFS
2747 /*
2748  * Normally delayed refs get processed in ascending bytenr order. This
2749  * correlates in most cases to the order added. To expose dependencies on this
2750  * order, we start to process the tree in the middle instead of the beginning
2751  */
2752 static u64 find_middle(struct rb_root *root)
2753 {
2754         struct rb_node *n = root->rb_node;
2755         struct btrfs_delayed_ref_node *entry;
2756         int alt = 1;
2757         u64 middle;
2758         u64 first = 0, last = 0;
2759
2760         n = rb_first(root);
2761         if (n) {
2762                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2763                 first = entry->bytenr;
2764         }
2765         n = rb_last(root);
2766         if (n) {
2767                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2768                 last = entry->bytenr;
2769         }
2770         n = root->rb_node;
2771
2772         while (n) {
2773                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2774                 WARN_ON(!entry->in_tree);
2775
2776                 middle = entry->bytenr;
2777
2778                 if (alt)
2779                         n = n->rb_left;
2780                 else
2781                         n = n->rb_right;
2782
2783                 alt = 1 - alt;
2784         }
2785         return middle;
2786 }
2787 #endif
2788
2789 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2790 {
2791         u64 num_bytes;
2792
2793         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2794                              sizeof(struct btrfs_extent_inline_ref));
2795         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2796                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2797
2798         /*
2799          * We don't ever fill up leaves all the way so multiply by 2 just to be
2800          * closer to what we're really going to want to use.
2801          */
2802         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2803 }
2804
2805 /*
2806  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2807  * would require to store the csums for that many bytes.
2808  */
2809 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2810 {
2811         u64 csum_size;
2812         u64 num_csums_per_leaf;
2813         u64 num_csums;
2814
2815         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2816         num_csums_per_leaf = div64_u64(csum_size,
2817                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2818         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2819         num_csums += num_csums_per_leaf - 1;
2820         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2821         return num_csums;
2822 }
2823
2824 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2825                                        struct btrfs_fs_info *fs_info)
2826 {
2827         struct btrfs_block_rsv *global_rsv;
2828         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2829         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2830         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2831         u64 num_bytes, num_dirty_bgs_bytes;
2832         int ret = 0;
2833
2834         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2835         num_heads = heads_to_leaves(fs_info, num_heads);
2836         if (num_heads > 1)
2837                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2838         num_bytes <<= 1;
2839         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2840                                                         fs_info->nodesize;
2841         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2842                                                              num_dirty_bgs);
2843         global_rsv = &fs_info->global_block_rsv;
2844
2845         /*
2846          * If we can't allocate any more chunks lets make sure we have _lots_ of
2847          * wiggle room since running delayed refs can create more delayed refs.
2848          */
2849         if (global_rsv->space_info->full) {
2850                 num_dirty_bgs_bytes <<= 1;
2851                 num_bytes <<= 1;
2852         }
2853
2854         spin_lock(&global_rsv->lock);
2855         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2856                 ret = 1;
2857         spin_unlock(&global_rsv->lock);
2858         return ret;
2859 }
2860
2861 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2862                                        struct btrfs_fs_info *fs_info)
2863 {
2864         u64 num_entries =
2865                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2866         u64 avg_runtime;
2867         u64 val;
2868
2869         smp_mb();
2870         avg_runtime = fs_info->avg_delayed_ref_runtime;
2871         val = num_entries * avg_runtime;
2872         if (val >= NSEC_PER_SEC)
2873                 return 1;
2874         if (val >= NSEC_PER_SEC / 2)
2875                 return 2;
2876
2877         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2878 }
2879
2880 struct async_delayed_refs {
2881         struct btrfs_root *root;
2882         u64 transid;
2883         int count;
2884         int error;
2885         int sync;
2886         struct completion wait;
2887         struct btrfs_work work;
2888 };
2889
2890 static inline struct async_delayed_refs *
2891 to_async_delayed_refs(struct btrfs_work *work)
2892 {
2893         return container_of(work, struct async_delayed_refs, work);
2894 }
2895
2896 static void delayed_ref_async_start(struct btrfs_work *work)
2897 {
2898         struct async_delayed_refs *async = to_async_delayed_refs(work);
2899         struct btrfs_trans_handle *trans;
2900         struct btrfs_fs_info *fs_info = async->root->fs_info;
2901         int ret;
2902
2903         /* if the commit is already started, we don't need to wait here */
2904         if (btrfs_transaction_blocked(fs_info))
2905                 goto done;
2906
2907         trans = btrfs_join_transaction(async->root);
2908         if (IS_ERR(trans)) {
2909                 async->error = PTR_ERR(trans);
2910                 goto done;
2911         }
2912
2913         /*
2914          * trans->sync means that when we call end_transaction, we won't
2915          * wait on delayed refs
2916          */
2917         trans->sync = true;
2918
2919         /* Don't bother flushing if we got into a different transaction */
2920         if (trans->transid > async->transid)
2921                 goto end;
2922
2923         ret = btrfs_run_delayed_refs(trans, fs_info, async->count);
2924         if (ret)
2925                 async->error = ret;
2926 end:
2927         ret = btrfs_end_transaction(trans);
2928         if (ret && !async->error)
2929                 async->error = ret;
2930 done:
2931         if (async->sync)
2932                 complete(&async->wait);
2933         else
2934                 kfree(async);
2935 }
2936
2937 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2938                                  unsigned long count, u64 transid, int wait)
2939 {
2940         struct async_delayed_refs *async;
2941         int ret;
2942
2943         async = kmalloc(sizeof(*async), GFP_NOFS);
2944         if (!async)
2945                 return -ENOMEM;
2946
2947         async->root = fs_info->tree_root;
2948         async->count = count;
2949         async->error = 0;
2950         async->transid = transid;
2951         if (wait)
2952                 async->sync = 1;
2953         else
2954                 async->sync = 0;
2955         init_completion(&async->wait);
2956
2957         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2958                         delayed_ref_async_start, NULL, NULL);
2959
2960         btrfs_queue_work(fs_info->extent_workers, &async->work);
2961
2962         if (wait) {
2963                 wait_for_completion(&async->wait);
2964                 ret = async->error;
2965                 kfree(async);
2966                 return ret;
2967         }
2968         return 0;
2969 }
2970
2971 /*
2972  * this starts processing the delayed reference count updates and
2973  * extent insertions we have queued up so far.  count can be
2974  * 0, which means to process everything in the tree at the start
2975  * of the run (but not newly added entries), or it can be some target
2976  * number you'd like to process.
2977  *
2978  * Returns 0 on success or if called with an aborted transaction
2979  * Returns <0 on error and aborts the transaction
2980  */
2981 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2982                            struct btrfs_fs_info *fs_info, unsigned long count)
2983 {
2984         struct rb_node *node;
2985         struct btrfs_delayed_ref_root *delayed_refs;
2986         struct btrfs_delayed_ref_head *head;
2987         int ret;
2988         int run_all = count == (unsigned long)-1;
2989         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2990
2991         /* We'll clean this up in btrfs_cleanup_transaction */
2992         if (trans->aborted)
2993                 return 0;
2994
2995         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
2996                 return 0;
2997
2998         delayed_refs = &trans->transaction->delayed_refs;
2999         if (count == 0)
3000                 count = atomic_read(&delayed_refs->num_entries) * 2;
3001
3002 again:
3003 #ifdef SCRAMBLE_DELAYED_REFS
3004         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3005 #endif
3006         trans->can_flush_pending_bgs = false;
3007         ret = __btrfs_run_delayed_refs(trans, fs_info, count);
3008         if (ret < 0) {
3009                 btrfs_abort_transaction(trans, ret);
3010                 return ret;
3011         }
3012
3013         if (run_all) {
3014                 if (!list_empty(&trans->new_bgs))
3015                         btrfs_create_pending_block_groups(trans, fs_info);
3016
3017                 spin_lock(&delayed_refs->lock);
3018                 node = rb_first(&delayed_refs->href_root);
3019                 if (!node) {
3020                         spin_unlock(&delayed_refs->lock);
3021                         goto out;
3022                 }
3023
3024                 while (node) {
3025                         head = rb_entry(node, struct btrfs_delayed_ref_head,
3026                                         href_node);
3027                         if (btrfs_delayed_ref_is_head(&head->node)) {
3028                                 struct btrfs_delayed_ref_node *ref;
3029
3030                                 ref = &head->node;
3031                                 refcount_inc(&ref->refs);
3032