283a3e967184243a2990601d0f7ce30dbebb01aa
[sfrench/cifs-2.6.git] / fs / btrfs / extent-tree.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/sched/signal.h>
8 #include <linux/pagemap.h>
9 #include <linux/writeback.h>
10 #include <linux/blkdev.h>
11 #include <linux/sort.h>
12 #include <linux/rcupdate.h>
13 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/ratelimit.h>
16 #include <linux/percpu_counter.h>
17 #include <linux/lockdep.h>
18 #include <linux/crc32c.h>
19 #include "tree-log.h"
20 #include "disk-io.h"
21 #include "print-tree.h"
22 #include "volumes.h"
23 #include "raid56.h"
24 #include "locking.h"
25 #include "free-space-cache.h"
26 #include "free-space-tree.h"
27 #include "math.h"
28 #include "sysfs.h"
29 #include "qgroup.h"
30 #include "ref-verify.h"
31
32 #undef SCRAMBLE_DELAYED_REFS
33
34 /*
35  * control flags for do_chunk_alloc's force field
36  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
37  * if we really need one.
38  *
39  * CHUNK_ALLOC_LIMITED means to only try and allocate one
40  * if we have very few chunks already allocated.  This is
41  * used as part of the clustering code to help make sure
42  * we have a good pool of storage to cluster in, without
43  * filling the FS with empty chunks
44  *
45  * CHUNK_ALLOC_FORCE means it must try to allocate one
46  *
47  */
48 enum {
49         CHUNK_ALLOC_NO_FORCE = 0,
50         CHUNK_ALLOC_LIMITED = 1,
51         CHUNK_ALLOC_FORCE = 2,
52 };
53
54 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
55                                struct btrfs_delayed_ref_node *node, u64 parent,
56                                u64 root_objectid, u64 owner_objectid,
57                                u64 owner_offset, int refs_to_drop,
58                                struct btrfs_delayed_extent_op *extra_op);
59 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
60                                     struct extent_buffer *leaf,
61                                     struct btrfs_extent_item *ei);
62 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
63                                       u64 parent, u64 root_objectid,
64                                       u64 flags, u64 owner, u64 offset,
65                                       struct btrfs_key *ins, int ref_mod);
66 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
67                                      struct btrfs_delayed_ref_node *node,
68                                      struct btrfs_delayed_extent_op *extent_op);
69 static int do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
70                           int force);
71 static int find_next_key(struct btrfs_path *path, int level,
72                          struct btrfs_key *key);
73 static void dump_space_info(struct btrfs_fs_info *fs_info,
74                             struct btrfs_space_info *info, u64 bytes,
75                             int dump_block_groups);
76 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
77                                u64 num_bytes);
78 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
79                                      struct btrfs_space_info *space_info,
80                                      u64 num_bytes);
81 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
82                                      struct btrfs_space_info *space_info,
83                                      u64 num_bytes);
84
85 static noinline int
86 block_group_cache_done(struct btrfs_block_group_cache *cache)
87 {
88         smp_mb();
89         return cache->cached == BTRFS_CACHE_FINISHED ||
90                 cache->cached == BTRFS_CACHE_ERROR;
91 }
92
93 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
94 {
95         return (cache->flags & bits) == bits;
96 }
97
98 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
99 {
100         atomic_inc(&cache->count);
101 }
102
103 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
104 {
105         if (atomic_dec_and_test(&cache->count)) {
106                 WARN_ON(cache->pinned > 0);
107                 WARN_ON(cache->reserved > 0);
108
109                 /*
110                  * If not empty, someone is still holding mutex of
111                  * full_stripe_lock, which can only be released by caller.
112                  * And it will definitely cause use-after-free when caller
113                  * tries to release full stripe lock.
114                  *
115                  * No better way to resolve, but only to warn.
116                  */
117                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
118                 kfree(cache->free_space_ctl);
119                 kfree(cache);
120         }
121 }
122
123 /*
124  * this adds the block group to the fs_info rb tree for the block group
125  * cache
126  */
127 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
128                                 struct btrfs_block_group_cache *block_group)
129 {
130         struct rb_node **p;
131         struct rb_node *parent = NULL;
132         struct btrfs_block_group_cache *cache;
133
134         spin_lock(&info->block_group_cache_lock);
135         p = &info->block_group_cache_tree.rb_node;
136
137         while (*p) {
138                 parent = *p;
139                 cache = rb_entry(parent, struct btrfs_block_group_cache,
140                                  cache_node);
141                 if (block_group->key.objectid < cache->key.objectid) {
142                         p = &(*p)->rb_left;
143                 } else if (block_group->key.objectid > cache->key.objectid) {
144                         p = &(*p)->rb_right;
145                 } else {
146                         spin_unlock(&info->block_group_cache_lock);
147                         return -EEXIST;
148                 }
149         }
150
151         rb_link_node(&block_group->cache_node, parent, p);
152         rb_insert_color(&block_group->cache_node,
153                         &info->block_group_cache_tree);
154
155         if (info->first_logical_byte > block_group->key.objectid)
156                 info->first_logical_byte = block_group->key.objectid;
157
158         spin_unlock(&info->block_group_cache_lock);
159
160         return 0;
161 }
162
163 /*
164  * This will return the block group at or after bytenr if contains is 0, else
165  * it will return the block group that contains the bytenr
166  */
167 static struct btrfs_block_group_cache *
168 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
169                               int contains)
170 {
171         struct btrfs_block_group_cache *cache, *ret = NULL;
172         struct rb_node *n;
173         u64 end, start;
174
175         spin_lock(&info->block_group_cache_lock);
176         n = info->block_group_cache_tree.rb_node;
177
178         while (n) {
179                 cache = rb_entry(n, struct btrfs_block_group_cache,
180                                  cache_node);
181                 end = cache->key.objectid + cache->key.offset - 1;
182                 start = cache->key.objectid;
183
184                 if (bytenr < start) {
185                         if (!contains && (!ret || start < ret->key.objectid))
186                                 ret = cache;
187                         n = n->rb_left;
188                 } else if (bytenr > start) {
189                         if (contains && bytenr <= end) {
190                                 ret = cache;
191                                 break;
192                         }
193                         n = n->rb_right;
194                 } else {
195                         ret = cache;
196                         break;
197                 }
198         }
199         if (ret) {
200                 btrfs_get_block_group(ret);
201                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
202                         info->first_logical_byte = ret->key.objectid;
203         }
204         spin_unlock(&info->block_group_cache_lock);
205
206         return ret;
207 }
208
209 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
210                                u64 start, u64 num_bytes)
211 {
212         u64 end = start + num_bytes - 1;
213         set_extent_bits(&fs_info->freed_extents[0],
214                         start, end, EXTENT_UPTODATE);
215         set_extent_bits(&fs_info->freed_extents[1],
216                         start, end, EXTENT_UPTODATE);
217         return 0;
218 }
219
220 static void free_excluded_extents(struct btrfs_block_group_cache *cache)
221 {
222         struct btrfs_fs_info *fs_info = cache->fs_info;
223         u64 start, end;
224
225         start = cache->key.objectid;
226         end = start + cache->key.offset - 1;
227
228         clear_extent_bits(&fs_info->freed_extents[0],
229                           start, end, EXTENT_UPTODATE);
230         clear_extent_bits(&fs_info->freed_extents[1],
231                           start, end, EXTENT_UPTODATE);
232 }
233
234 static int exclude_super_stripes(struct btrfs_block_group_cache *cache)
235 {
236         struct btrfs_fs_info *fs_info = cache->fs_info;
237         u64 bytenr;
238         u64 *logical;
239         int stripe_len;
240         int i, nr, ret;
241
242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
244                 cache->bytes_super += stripe_len;
245                 ret = add_excluded_extent(fs_info, cache->key.objectid,
246                                           stripe_len);
247                 if (ret)
248                         return ret;
249         }
250
251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
252                 bytenr = btrfs_sb_offset(i);
253                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
254                                        bytenr, &logical, &nr, &stripe_len);
255                 if (ret)
256                         return ret;
257
258                 while (nr--) {
259                         u64 start, len;
260
261                         if (logical[nr] > cache->key.objectid +
262                             cache->key.offset)
263                                 continue;
264
265                         if (logical[nr] + stripe_len <= cache->key.objectid)
266                                 continue;
267
268                         start = logical[nr];
269                         if (start < cache->key.objectid) {
270                                 start = cache->key.objectid;
271                                 len = (logical[nr] + stripe_len) - start;
272                         } else {
273                                 len = min_t(u64, stripe_len,
274                                             cache->key.objectid +
275                                             cache->key.offset - start);
276                         }
277
278                         cache->bytes_super += len;
279                         ret = add_excluded_extent(fs_info, start, len);
280                         if (ret) {
281                                 kfree(logical);
282                                 return ret;
283                         }
284                 }
285
286                 kfree(logical);
287         }
288         return 0;
289 }
290
291 static struct btrfs_caching_control *
292 get_caching_control(struct btrfs_block_group_cache *cache)
293 {
294         struct btrfs_caching_control *ctl;
295
296         spin_lock(&cache->lock);
297         if (!cache->caching_ctl) {
298                 spin_unlock(&cache->lock);
299                 return NULL;
300         }
301
302         ctl = cache->caching_ctl;
303         refcount_inc(&ctl->count);
304         spin_unlock(&cache->lock);
305         return ctl;
306 }
307
308 static void put_caching_control(struct btrfs_caching_control *ctl)
309 {
310         if (refcount_dec_and_test(&ctl->count))
311                 kfree(ctl);
312 }
313
314 #ifdef CONFIG_BTRFS_DEBUG
315 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
316 {
317         struct btrfs_fs_info *fs_info = block_group->fs_info;
318         u64 start = block_group->key.objectid;
319         u64 len = block_group->key.offset;
320         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
321                 fs_info->nodesize : fs_info->sectorsize;
322         u64 step = chunk << 1;
323
324         while (len > chunk) {
325                 btrfs_remove_free_space(block_group, start, chunk);
326                 start += step;
327                 if (len < step)
328                         len = 0;
329                 else
330                         len -= step;
331         }
332 }
333 #endif
334
335 /*
336  * this is only called by cache_block_group, since we could have freed extents
337  * we need to check the pinned_extents for any extents that can't be used yet
338  * since their free space will be released as soon as the transaction commits.
339  */
340 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
341                        u64 start, u64 end)
342 {
343         struct btrfs_fs_info *info = block_group->fs_info;
344         u64 extent_start, extent_end, size, total_added = 0;
345         int ret;
346
347         while (start < end) {
348                 ret = find_first_extent_bit(info->pinned_extents, start,
349                                             &extent_start, &extent_end,
350                                             EXTENT_DIRTY | EXTENT_UPTODATE,
351                                             NULL);
352                 if (ret)
353                         break;
354
355                 if (extent_start <= start) {
356                         start = extent_end + 1;
357                 } else if (extent_start > start && extent_start < end) {
358                         size = extent_start - start;
359                         total_added += size;
360                         ret = btrfs_add_free_space(block_group, start,
361                                                    size);
362                         BUG_ON(ret); /* -ENOMEM or logic error */
363                         start = extent_end + 1;
364                 } else {
365                         break;
366                 }
367         }
368
369         if (start < end) {
370                 size = end - start;
371                 total_added += size;
372                 ret = btrfs_add_free_space(block_group, start, size);
373                 BUG_ON(ret); /* -ENOMEM or logic error */
374         }
375
376         return total_added;
377 }
378
379 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
380 {
381         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
382         struct btrfs_fs_info *fs_info = block_group->fs_info;
383         struct btrfs_root *extent_root = fs_info->extent_root;
384         struct btrfs_path *path;
385         struct extent_buffer *leaf;
386         struct btrfs_key key;
387         u64 total_found = 0;
388         u64 last = 0;
389         u32 nritems;
390         int ret;
391         bool wakeup = true;
392
393         path = btrfs_alloc_path();
394         if (!path)
395                 return -ENOMEM;
396
397         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
398
399 #ifdef CONFIG_BTRFS_DEBUG
400         /*
401          * If we're fragmenting we don't want to make anybody think we can
402          * allocate from this block group until we've had a chance to fragment
403          * the free space.
404          */
405         if (btrfs_should_fragment_free_space(block_group))
406                 wakeup = false;
407 #endif
408         /*
409          * We don't want to deadlock with somebody trying to allocate a new
410          * extent for the extent root while also trying to search the extent
411          * root to add free space.  So we skip locking and search the commit
412          * root, since its read-only
413          */
414         path->skip_locking = 1;
415         path->search_commit_root = 1;
416         path->reada = READA_FORWARD;
417
418         key.objectid = last;
419         key.offset = 0;
420         key.type = BTRFS_EXTENT_ITEM_KEY;
421
422 next:
423         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
424         if (ret < 0)
425                 goto out;
426
427         leaf = path->nodes[0];
428         nritems = btrfs_header_nritems(leaf);
429
430         while (1) {
431                 if (btrfs_fs_closing(fs_info) > 1) {
432                         last = (u64)-1;
433                         break;
434                 }
435
436                 if (path->slots[0] < nritems) {
437                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
438                 } else {
439                         ret = find_next_key(path, 0, &key);
440                         if (ret)
441                                 break;
442
443                         if (need_resched() ||
444                             rwsem_is_contended(&fs_info->commit_root_sem)) {
445                                 if (wakeup)
446                                         caching_ctl->progress = last;
447                                 btrfs_release_path(path);
448                                 up_read(&fs_info->commit_root_sem);
449                                 mutex_unlock(&caching_ctl->mutex);
450                                 cond_resched();
451                                 mutex_lock(&caching_ctl->mutex);
452                                 down_read(&fs_info->commit_root_sem);
453                                 goto next;
454                         }
455
456                         ret = btrfs_next_leaf(extent_root, path);
457                         if (ret < 0)
458                                 goto out;
459                         if (ret)
460                                 break;
461                         leaf = path->nodes[0];
462                         nritems = btrfs_header_nritems(leaf);
463                         continue;
464                 }
465
466                 if (key.objectid < last) {
467                         key.objectid = last;
468                         key.offset = 0;
469                         key.type = BTRFS_EXTENT_ITEM_KEY;
470
471                         if (wakeup)
472                                 caching_ctl->progress = last;
473                         btrfs_release_path(path);
474                         goto next;
475                 }
476
477                 if (key.objectid < block_group->key.objectid) {
478                         path->slots[0]++;
479                         continue;
480                 }
481
482                 if (key.objectid >= block_group->key.objectid +
483                     block_group->key.offset)
484                         break;
485
486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
487                     key.type == BTRFS_METADATA_ITEM_KEY) {
488                         total_found += add_new_free_space(block_group, last,
489                                                           key.objectid);
490                         if (key.type == BTRFS_METADATA_ITEM_KEY)
491                                 last = key.objectid +
492                                         fs_info->nodesize;
493                         else
494                                 last = key.objectid + key.offset;
495
496                         if (total_found > CACHING_CTL_WAKE_UP) {
497                                 total_found = 0;
498                                 if (wakeup)
499                                         wake_up(&caching_ctl->wait);
500                         }
501                 }
502                 path->slots[0]++;
503         }
504         ret = 0;
505
506         total_found += add_new_free_space(block_group, last,
507                                           block_group->key.objectid +
508                                           block_group->key.offset);
509         caching_ctl->progress = (u64)-1;
510
511 out:
512         btrfs_free_path(path);
513         return ret;
514 }
515
516 static noinline void caching_thread(struct btrfs_work *work)
517 {
518         struct btrfs_block_group_cache *block_group;
519         struct btrfs_fs_info *fs_info;
520         struct btrfs_caching_control *caching_ctl;
521         int ret;
522
523         caching_ctl = container_of(work, struct btrfs_caching_control, work);
524         block_group = caching_ctl->block_group;
525         fs_info = block_group->fs_info;
526
527         mutex_lock(&caching_ctl->mutex);
528         down_read(&fs_info->commit_root_sem);
529
530         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
531                 ret = load_free_space_tree(caching_ctl);
532         else
533                 ret = load_extent_tree_free(caching_ctl);
534
535         spin_lock(&block_group->lock);
536         block_group->caching_ctl = NULL;
537         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
538         spin_unlock(&block_group->lock);
539
540 #ifdef CONFIG_BTRFS_DEBUG
541         if (btrfs_should_fragment_free_space(block_group)) {
542                 u64 bytes_used;
543
544                 spin_lock(&block_group->space_info->lock);
545                 spin_lock(&block_group->lock);
546                 bytes_used = block_group->key.offset -
547                         btrfs_block_group_used(&block_group->item);
548                 block_group->space_info->bytes_used += bytes_used >> 1;
549                 spin_unlock(&block_group->lock);
550                 spin_unlock(&block_group->space_info->lock);
551                 fragment_free_space(block_group);
552         }
553 #endif
554
555         caching_ctl->progress = (u64)-1;
556
557         up_read(&fs_info->commit_root_sem);
558         free_excluded_extents(block_group);
559         mutex_unlock(&caching_ctl->mutex);
560
561         wake_up(&caching_ctl->wait);
562
563         put_caching_control(caching_ctl);
564         btrfs_put_block_group(block_group);
565 }
566
567 static int cache_block_group(struct btrfs_block_group_cache *cache,
568                              int load_cache_only)
569 {
570         DEFINE_WAIT(wait);
571         struct btrfs_fs_info *fs_info = cache->fs_info;
572         struct btrfs_caching_control *caching_ctl;
573         int ret = 0;
574
575         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
576         if (!caching_ctl)
577                 return -ENOMEM;
578
579         INIT_LIST_HEAD(&caching_ctl->list);
580         mutex_init(&caching_ctl->mutex);
581         init_waitqueue_head(&caching_ctl->wait);
582         caching_ctl->block_group = cache;
583         caching_ctl->progress = cache->key.objectid;
584         refcount_set(&caching_ctl->count, 1);
585         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
586                         caching_thread, NULL, NULL);
587
588         spin_lock(&cache->lock);
589         /*
590          * This should be a rare occasion, but this could happen I think in the
591          * case where one thread starts to load the space cache info, and then
592          * some other thread starts a transaction commit which tries to do an
593          * allocation while the other thread is still loading the space cache
594          * info.  The previous loop should have kept us from choosing this block
595          * group, but if we've moved to the state where we will wait on caching
596          * block groups we need to first check if we're doing a fast load here,
597          * so we can wait for it to finish, otherwise we could end up allocating
598          * from a block group who's cache gets evicted for one reason or
599          * another.
600          */
601         while (cache->cached == BTRFS_CACHE_FAST) {
602                 struct btrfs_caching_control *ctl;
603
604                 ctl = cache->caching_ctl;
605                 refcount_inc(&ctl->count);
606                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
607                 spin_unlock(&cache->lock);
608
609                 schedule();
610
611                 finish_wait(&ctl->wait, &wait);
612                 put_caching_control(ctl);
613                 spin_lock(&cache->lock);
614         }
615
616         if (cache->cached != BTRFS_CACHE_NO) {
617                 spin_unlock(&cache->lock);
618                 kfree(caching_ctl);
619                 return 0;
620         }
621         WARN_ON(cache->caching_ctl);
622         cache->caching_ctl = caching_ctl;
623         cache->cached = BTRFS_CACHE_FAST;
624         spin_unlock(&cache->lock);
625
626         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
627                 mutex_lock(&caching_ctl->mutex);
628                 ret = load_free_space_cache(fs_info, cache);
629
630                 spin_lock(&cache->lock);
631                 if (ret == 1) {
632                         cache->caching_ctl = NULL;
633                         cache->cached = BTRFS_CACHE_FINISHED;
634                         cache->last_byte_to_unpin = (u64)-1;
635                         caching_ctl->progress = (u64)-1;
636                 } else {
637                         if (load_cache_only) {
638                                 cache->caching_ctl = NULL;
639                                 cache->cached = BTRFS_CACHE_NO;
640                         } else {
641                                 cache->cached = BTRFS_CACHE_STARTED;
642                                 cache->has_caching_ctl = 1;
643                         }
644                 }
645                 spin_unlock(&cache->lock);
646 #ifdef CONFIG_BTRFS_DEBUG
647                 if (ret == 1 &&
648                     btrfs_should_fragment_free_space(cache)) {
649                         u64 bytes_used;
650
651                         spin_lock(&cache->space_info->lock);
652                         spin_lock(&cache->lock);
653                         bytes_used = cache->key.offset -
654                                 btrfs_block_group_used(&cache->item);
655                         cache->space_info->bytes_used += bytes_used >> 1;
656                         spin_unlock(&cache->lock);
657                         spin_unlock(&cache->space_info->lock);
658                         fragment_free_space(cache);
659                 }
660 #endif
661                 mutex_unlock(&caching_ctl->mutex);
662
663                 wake_up(&caching_ctl->wait);
664                 if (ret == 1) {
665                         put_caching_control(caching_ctl);
666                         free_excluded_extents(cache);
667                         return 0;
668                 }
669         } else {
670                 /*
671                  * We're either using the free space tree or no caching at all.
672                  * Set cached to the appropriate value and wakeup any waiters.
673                  */
674                 spin_lock(&cache->lock);
675                 if (load_cache_only) {
676                         cache->caching_ctl = NULL;
677                         cache->cached = BTRFS_CACHE_NO;
678                 } else {
679                         cache->cached = BTRFS_CACHE_STARTED;
680                         cache->has_caching_ctl = 1;
681                 }
682                 spin_unlock(&cache->lock);
683                 wake_up(&caching_ctl->wait);
684         }
685
686         if (load_cache_only) {
687                 put_caching_control(caching_ctl);
688                 return 0;
689         }
690
691         down_write(&fs_info->commit_root_sem);
692         refcount_inc(&caching_ctl->count);
693         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
694         up_write(&fs_info->commit_root_sem);
695
696         btrfs_get_block_group(cache);
697
698         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
699
700         return ret;
701 }
702
703 /*
704  * return the block group that starts at or after bytenr
705  */
706 static struct btrfs_block_group_cache *
707 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
708 {
709         return block_group_cache_tree_search(info, bytenr, 0);
710 }
711
712 /*
713  * return the block group that contains the given bytenr
714  */
715 struct btrfs_block_group_cache *btrfs_lookup_block_group(
716                                                  struct btrfs_fs_info *info,
717                                                  u64 bytenr)
718 {
719         return block_group_cache_tree_search(info, bytenr, 1);
720 }
721
722 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
723                                                   u64 flags)
724 {
725         struct list_head *head = &info->space_info;
726         struct btrfs_space_info *found;
727
728         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
729
730         rcu_read_lock();
731         list_for_each_entry_rcu(found, head, list) {
732                 if (found->flags & flags) {
733                         rcu_read_unlock();
734                         return found;
735                 }
736         }
737         rcu_read_unlock();
738         return NULL;
739 }
740
741 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
742                              bool metadata, u64 root_objectid)
743 {
744         struct btrfs_space_info *space_info;
745         u64 flags;
746
747         if (metadata) {
748                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
749                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
750                 else
751                         flags = BTRFS_BLOCK_GROUP_METADATA;
752         } else {
753                 flags = BTRFS_BLOCK_GROUP_DATA;
754         }
755
756         space_info = __find_space_info(fs_info, flags);
757         ASSERT(space_info);
758         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
759 }
760
761 /*
762  * after adding space to the filesystem, we need to clear the full flags
763  * on all the space infos.
764  */
765 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
766 {
767         struct list_head *head = &info->space_info;
768         struct btrfs_space_info *found;
769
770         rcu_read_lock();
771         list_for_each_entry_rcu(found, head, list)
772                 found->full = 0;
773         rcu_read_unlock();
774 }
775
776 /* simple helper to search for an existing data extent at a given offset */
777 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
778 {
779         int ret;
780         struct btrfs_key key;
781         struct btrfs_path *path;
782
783         path = btrfs_alloc_path();
784         if (!path)
785                 return -ENOMEM;
786
787         key.objectid = start;
788         key.offset = len;
789         key.type = BTRFS_EXTENT_ITEM_KEY;
790         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
791         btrfs_free_path(path);
792         return ret;
793 }
794
795 /*
796  * helper function to lookup reference count and flags of a tree block.
797  *
798  * the head node for delayed ref is used to store the sum of all the
799  * reference count modifications queued up in the rbtree. the head
800  * node may also store the extent flags to set. This way you can check
801  * to see what the reference count and extent flags would be if all of
802  * the delayed refs are not processed.
803  */
804 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
805                              struct btrfs_fs_info *fs_info, u64 bytenr,
806                              u64 offset, int metadata, u64 *refs, u64 *flags)
807 {
808         struct btrfs_delayed_ref_head *head;
809         struct btrfs_delayed_ref_root *delayed_refs;
810         struct btrfs_path *path;
811         struct btrfs_extent_item *ei;
812         struct extent_buffer *leaf;
813         struct btrfs_key key;
814         u32 item_size;
815         u64 num_refs;
816         u64 extent_flags;
817         int ret;
818
819         /*
820          * If we don't have skinny metadata, don't bother doing anything
821          * different
822          */
823         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
824                 offset = fs_info->nodesize;
825                 metadata = 0;
826         }
827
828         path = btrfs_alloc_path();
829         if (!path)
830                 return -ENOMEM;
831
832         if (!trans) {
833                 path->skip_locking = 1;
834                 path->search_commit_root = 1;
835         }
836
837 search_again:
838         key.objectid = bytenr;
839         key.offset = offset;
840         if (metadata)
841                 key.type = BTRFS_METADATA_ITEM_KEY;
842         else
843                 key.type = BTRFS_EXTENT_ITEM_KEY;
844
845         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
846         if (ret < 0)
847                 goto out_free;
848
849         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
850                 if (path->slots[0]) {
851                         path->slots[0]--;
852                         btrfs_item_key_to_cpu(path->nodes[0], &key,
853                                               path->slots[0]);
854                         if (key.objectid == bytenr &&
855                             key.type == BTRFS_EXTENT_ITEM_KEY &&
856                             key.offset == fs_info->nodesize)
857                                 ret = 0;
858                 }
859         }
860
861         if (ret == 0) {
862                 leaf = path->nodes[0];
863                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
864                 if (item_size >= sizeof(*ei)) {
865                         ei = btrfs_item_ptr(leaf, path->slots[0],
866                                             struct btrfs_extent_item);
867                         num_refs = btrfs_extent_refs(leaf, ei);
868                         extent_flags = btrfs_extent_flags(leaf, ei);
869                 } else {
870 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
871                         struct btrfs_extent_item_v0 *ei0;
872                         BUG_ON(item_size != sizeof(*ei0));
873                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
874                                              struct btrfs_extent_item_v0);
875                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
876                         /* FIXME: this isn't correct for data */
877                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
878 #else
879                         BUG();
880 #endif
881                 }
882                 BUG_ON(num_refs == 0);
883         } else {
884                 num_refs = 0;
885                 extent_flags = 0;
886                 ret = 0;
887         }
888
889         if (!trans)
890                 goto out;
891
892         delayed_refs = &trans->transaction->delayed_refs;
893         spin_lock(&delayed_refs->lock);
894         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
895         if (head) {
896                 if (!mutex_trylock(&head->mutex)) {
897                         refcount_inc(&head->refs);
898                         spin_unlock(&delayed_refs->lock);
899
900                         btrfs_release_path(path);
901
902                         /*
903                          * Mutex was contended, block until it's released and try
904                          * again
905                          */
906                         mutex_lock(&head->mutex);
907                         mutex_unlock(&head->mutex);
908                         btrfs_put_delayed_ref_head(head);
909                         goto search_again;
910                 }
911                 spin_lock(&head->lock);
912                 if (head->extent_op && head->extent_op->update_flags)
913                         extent_flags |= head->extent_op->flags_to_set;
914                 else
915                         BUG_ON(num_refs == 0);
916
917                 num_refs += head->ref_mod;
918                 spin_unlock(&head->lock);
919                 mutex_unlock(&head->mutex);
920         }
921         spin_unlock(&delayed_refs->lock);
922 out:
923         WARN_ON(num_refs == 0);
924         if (refs)
925                 *refs = num_refs;
926         if (flags)
927                 *flags = extent_flags;
928 out_free:
929         btrfs_free_path(path);
930         return ret;
931 }
932
933 /*
934  * Back reference rules.  Back refs have three main goals:
935  *
936  * 1) differentiate between all holders of references to an extent so that
937  *    when a reference is dropped we can make sure it was a valid reference
938  *    before freeing the extent.
939  *
940  * 2) Provide enough information to quickly find the holders of an extent
941  *    if we notice a given block is corrupted or bad.
942  *
943  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
944  *    maintenance.  This is actually the same as #2, but with a slightly
945  *    different use case.
946  *
947  * There are two kinds of back refs. The implicit back refs is optimized
948  * for pointers in non-shared tree blocks. For a given pointer in a block,
949  * back refs of this kind provide information about the block's owner tree
950  * and the pointer's key. These information allow us to find the block by
951  * b-tree searching. The full back refs is for pointers in tree blocks not
952  * referenced by their owner trees. The location of tree block is recorded
953  * in the back refs. Actually the full back refs is generic, and can be
954  * used in all cases the implicit back refs is used. The major shortcoming
955  * of the full back refs is its overhead. Every time a tree block gets
956  * COWed, we have to update back refs entry for all pointers in it.
957  *
958  * For a newly allocated tree block, we use implicit back refs for
959  * pointers in it. This means most tree related operations only involve
960  * implicit back refs. For a tree block created in old transaction, the
961  * only way to drop a reference to it is COW it. So we can detect the
962  * event that tree block loses its owner tree's reference and do the
963  * back refs conversion.
964  *
965  * When a tree block is COWed through a tree, there are four cases:
966  *
967  * The reference count of the block is one and the tree is the block's
968  * owner tree. Nothing to do in this case.
969  *
970  * The reference count of the block is one and the tree is not the
971  * block's owner tree. In this case, full back refs is used for pointers
972  * in the block. Remove these full back refs, add implicit back refs for
973  * every pointers in the new block.
974  *
975  * The reference count of the block is greater than one and the tree is
976  * the block's owner tree. In this case, implicit back refs is used for
977  * pointers in the block. Add full back refs for every pointers in the
978  * block, increase lower level extents' reference counts. The original
979  * implicit back refs are entailed to the new block.
980  *
981  * The reference count of the block is greater than one and the tree is
982  * not the block's owner tree. Add implicit back refs for every pointer in
983  * the new block, increase lower level extents' reference count.
984  *
985  * Back Reference Key composing:
986  *
987  * The key objectid corresponds to the first byte in the extent,
988  * The key type is used to differentiate between types of back refs.
989  * There are different meanings of the key offset for different types
990  * of back refs.
991  *
992  * File extents can be referenced by:
993  *
994  * - multiple snapshots, subvolumes, or different generations in one subvol
995  * - different files inside a single subvolume
996  * - different offsets inside a file (bookend extents in file.c)
997  *
998  * The extent ref structure for the implicit back refs has fields for:
999  *
1000  * - Objectid of the subvolume root
1001  * - objectid of the file holding the reference
1002  * - original offset in the file
1003  * - how many bookend extents
1004  *
1005  * The key offset for the implicit back refs is hash of the first
1006  * three fields.
1007  *
1008  * The extent ref structure for the full back refs has field for:
1009  *
1010  * - number of pointers in the tree leaf
1011  *
1012  * The key offset for the implicit back refs is the first byte of
1013  * the tree leaf
1014  *
1015  * When a file extent is allocated, The implicit back refs is used.
1016  * the fields are filled in:
1017  *
1018  *     (root_key.objectid, inode objectid, offset in file, 1)
1019  *
1020  * When a file extent is removed file truncation, we find the
1021  * corresponding implicit back refs and check the following fields:
1022  *
1023  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1024  *
1025  * Btree extents can be referenced by:
1026  *
1027  * - Different subvolumes
1028  *
1029  * Both the implicit back refs and the full back refs for tree blocks
1030  * only consist of key. The key offset for the implicit back refs is
1031  * objectid of block's owner tree. The key offset for the full back refs
1032  * is the first byte of parent block.
1033  *
1034  * When implicit back refs is used, information about the lowest key and
1035  * level of the tree block are required. These information are stored in
1036  * tree block info structure.
1037  */
1038
1039 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1040 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1041                                   struct btrfs_fs_info *fs_info,
1042                                   struct btrfs_path *path,
1043                                   u64 owner, u32 extra_size)
1044 {
1045         struct btrfs_root *root = fs_info->extent_root;
1046         struct btrfs_extent_item *item;
1047         struct btrfs_extent_item_v0 *ei0;
1048         struct btrfs_extent_ref_v0 *ref0;
1049         struct btrfs_tree_block_info *bi;
1050         struct extent_buffer *leaf;
1051         struct btrfs_key key;
1052         struct btrfs_key found_key;
1053         u32 new_size = sizeof(*item);
1054         u64 refs;
1055         int ret;
1056
1057         leaf = path->nodes[0];
1058         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1059
1060         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1061         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1062                              struct btrfs_extent_item_v0);
1063         refs = btrfs_extent_refs_v0(leaf, ei0);
1064
1065         if (owner == (u64)-1) {
1066                 while (1) {
1067                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1068                                 ret = btrfs_next_leaf(root, path);
1069                                 if (ret < 0)
1070                                         return ret;
1071                                 BUG_ON(ret > 0); /* Corruption */
1072                                 leaf = path->nodes[0];
1073                         }
1074                         btrfs_item_key_to_cpu(leaf, &found_key,
1075                                               path->slots[0]);
1076                         BUG_ON(key.objectid != found_key.objectid);
1077                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1078                                 path->slots[0]++;
1079                                 continue;
1080                         }
1081                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1082                                               struct btrfs_extent_ref_v0);
1083                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1084                         break;
1085                 }
1086         }
1087         btrfs_release_path(path);
1088
1089         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1090                 new_size += sizeof(*bi);
1091
1092         new_size -= sizeof(*ei0);
1093         ret = btrfs_search_slot(trans, root, &key, path,
1094                                 new_size + extra_size, 1);
1095         if (ret < 0)
1096                 return ret;
1097         BUG_ON(ret); /* Corruption */
1098
1099         btrfs_extend_item(fs_info, path, new_size);
1100
1101         leaf = path->nodes[0];
1102         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1103         btrfs_set_extent_refs(leaf, item, refs);
1104         /* FIXME: get real generation */
1105         btrfs_set_extent_generation(leaf, item, 0);
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1107                 btrfs_set_extent_flags(leaf, item,
1108                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1109                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1110                 bi = (struct btrfs_tree_block_info *)(item + 1);
1111                 /* FIXME: get first key of the block */
1112                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1113                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1114         } else {
1115                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1116         }
1117         btrfs_mark_buffer_dirty(leaf);
1118         return 0;
1119 }
1120 #endif
1121
1122 /*
1123  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1124  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1125  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1126  */
1127 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1128                                      struct btrfs_extent_inline_ref *iref,
1129                                      enum btrfs_inline_ref_type is_data)
1130 {
1131         int type = btrfs_extent_inline_ref_type(eb, iref);
1132         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1133
1134         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1135             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1136             type == BTRFS_SHARED_DATA_REF_KEY ||
1137             type == BTRFS_EXTENT_DATA_REF_KEY) {
1138                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1139                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1140                                 return type;
1141                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1142                                 ASSERT(eb->fs_info);
1143                                 /*
1144                                  * Every shared one has parent tree
1145                                  * block, which must be aligned to
1146                                  * nodesize.
1147                                  */
1148                                 if (offset &&
1149                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1150                                         return type;
1151                         }
1152                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1153                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1154                                 return type;
1155                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1156                                 ASSERT(eb->fs_info);
1157                                 /*
1158                                  * Every shared one has parent tree
1159                                  * block, which must be aligned to
1160                                  * nodesize.
1161                                  */
1162                                 if (offset &&
1163                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1164                                         return type;
1165                         }
1166                 } else {
1167                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1168                         return type;
1169                 }
1170         }
1171
1172         btrfs_print_leaf((struct extent_buffer *)eb);
1173         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1174                   eb->start, type);
1175         WARN_ON(1);
1176
1177         return BTRFS_REF_TYPE_INVALID;
1178 }
1179
1180 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1181 {
1182         u32 high_crc = ~(u32)0;
1183         u32 low_crc = ~(u32)0;
1184         __le64 lenum;
1185
1186         lenum = cpu_to_le64(root_objectid);
1187         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1188         lenum = cpu_to_le64(owner);
1189         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1190         lenum = cpu_to_le64(offset);
1191         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1192
1193         return ((u64)high_crc << 31) ^ (u64)low_crc;
1194 }
1195
1196 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1197                                      struct btrfs_extent_data_ref *ref)
1198 {
1199         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1200                                     btrfs_extent_data_ref_objectid(leaf, ref),
1201                                     btrfs_extent_data_ref_offset(leaf, ref));
1202 }
1203
1204 static int match_extent_data_ref(struct extent_buffer *leaf,
1205                                  struct btrfs_extent_data_ref *ref,
1206                                  u64 root_objectid, u64 owner, u64 offset)
1207 {
1208         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1209             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1210             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1211                 return 0;
1212         return 1;
1213 }
1214
1215 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1216                                            struct btrfs_path *path,
1217                                            u64 bytenr, u64 parent,
1218                                            u64 root_objectid,
1219                                            u64 owner, u64 offset)
1220 {
1221         struct btrfs_root *root = trans->fs_info->extent_root;
1222         struct btrfs_key key;
1223         struct btrfs_extent_data_ref *ref;
1224         struct extent_buffer *leaf;
1225         u32 nritems;
1226         int ret;
1227         int recow;
1228         int err = -ENOENT;
1229
1230         key.objectid = bytenr;
1231         if (parent) {
1232                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1233                 key.offset = parent;
1234         } else {
1235                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1236                 key.offset = hash_extent_data_ref(root_objectid,
1237                                                   owner, offset);
1238         }
1239 again:
1240         recow = 0;
1241         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1242         if (ret < 0) {
1243                 err = ret;
1244                 goto fail;
1245         }
1246
1247         if (parent) {
1248                 if (!ret)
1249                         return 0;
1250 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1251                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1252                 btrfs_release_path(path);
1253                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1254                 if (ret < 0) {
1255                         err = ret;
1256                         goto fail;
1257                 }
1258                 if (!ret)
1259                         return 0;
1260 #endif
1261                 goto fail;
1262         }
1263
1264         leaf = path->nodes[0];
1265         nritems = btrfs_header_nritems(leaf);
1266         while (1) {
1267                 if (path->slots[0] >= nritems) {
1268                         ret = btrfs_next_leaf(root, path);
1269                         if (ret < 0)
1270                                 err = ret;
1271                         if (ret)
1272                                 goto fail;
1273
1274                         leaf = path->nodes[0];
1275                         nritems = btrfs_header_nritems(leaf);
1276                         recow = 1;
1277                 }
1278
1279                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1280                 if (key.objectid != bytenr ||
1281                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1282                         goto fail;
1283
1284                 ref = btrfs_item_ptr(leaf, path->slots[0],
1285                                      struct btrfs_extent_data_ref);
1286
1287                 if (match_extent_data_ref(leaf, ref, root_objectid,
1288                                           owner, offset)) {
1289                         if (recow) {
1290                                 btrfs_release_path(path);
1291                                 goto again;
1292                         }
1293                         err = 0;
1294                         break;
1295                 }
1296                 path->slots[0]++;
1297         }
1298 fail:
1299         return err;
1300 }
1301
1302 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1303                                            struct btrfs_path *path,
1304                                            u64 bytenr, u64 parent,
1305                                            u64 root_objectid, u64 owner,
1306                                            u64 offset, int refs_to_add)
1307 {
1308         struct btrfs_root *root = trans->fs_info->extent_root;
1309         struct btrfs_key key;
1310         struct extent_buffer *leaf;
1311         u32 size;
1312         u32 num_refs;
1313         int ret;
1314
1315         key.objectid = bytenr;
1316         if (parent) {
1317                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1318                 key.offset = parent;
1319                 size = sizeof(struct btrfs_shared_data_ref);
1320         } else {
1321                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1322                 key.offset = hash_extent_data_ref(root_objectid,
1323                                                   owner, offset);
1324                 size = sizeof(struct btrfs_extent_data_ref);
1325         }
1326
1327         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1328         if (ret && ret != -EEXIST)
1329                 goto fail;
1330
1331         leaf = path->nodes[0];
1332         if (parent) {
1333                 struct btrfs_shared_data_ref *ref;
1334                 ref = btrfs_item_ptr(leaf, path->slots[0],
1335                                      struct btrfs_shared_data_ref);
1336                 if (ret == 0) {
1337                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1338                 } else {
1339                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1340                         num_refs += refs_to_add;
1341                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1342                 }
1343         } else {
1344                 struct btrfs_extent_data_ref *ref;
1345                 while (ret == -EEXIST) {
1346                         ref = btrfs_item_ptr(leaf, path->slots[0],
1347                                              struct btrfs_extent_data_ref);
1348                         if (match_extent_data_ref(leaf, ref, root_objectid,
1349                                                   owner, offset))
1350                                 break;
1351                         btrfs_release_path(path);
1352                         key.offset++;
1353                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1354                                                       size);
1355                         if (ret && ret != -EEXIST)
1356                                 goto fail;
1357
1358                         leaf = path->nodes[0];
1359                 }
1360                 ref = btrfs_item_ptr(leaf, path->slots[0],
1361                                      struct btrfs_extent_data_ref);
1362                 if (ret == 0) {
1363                         btrfs_set_extent_data_ref_root(leaf, ref,
1364                                                        root_objectid);
1365                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1366                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1367                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1368                 } else {
1369                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1370                         num_refs += refs_to_add;
1371                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1372                 }
1373         }
1374         btrfs_mark_buffer_dirty(leaf);
1375         ret = 0;
1376 fail:
1377         btrfs_release_path(path);
1378         return ret;
1379 }
1380
1381 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1382                                            struct btrfs_path *path,
1383                                            int refs_to_drop, int *last_ref)
1384 {
1385         struct btrfs_key key;
1386         struct btrfs_extent_data_ref *ref1 = NULL;
1387         struct btrfs_shared_data_ref *ref2 = NULL;
1388         struct extent_buffer *leaf;
1389         u32 num_refs = 0;
1390         int ret = 0;
1391
1392         leaf = path->nodes[0];
1393         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1394
1395         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1396                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1397                                       struct btrfs_extent_data_ref);
1398                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1399         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1400                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1401                                       struct btrfs_shared_data_ref);
1402                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1403 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1404         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1405                 struct btrfs_extent_ref_v0 *ref0;
1406                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1407                                       struct btrfs_extent_ref_v0);
1408                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1409 #endif
1410         } else {
1411                 BUG();
1412         }
1413
1414         BUG_ON(num_refs < refs_to_drop);
1415         num_refs -= refs_to_drop;
1416
1417         if (num_refs == 0) {
1418                 ret = btrfs_del_item(trans, trans->fs_info->extent_root, path);
1419                 *last_ref = 1;
1420         } else {
1421                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1422                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1423                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1424                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1425 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1426                 else {
1427                         struct btrfs_extent_ref_v0 *ref0;
1428                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1429                                         struct btrfs_extent_ref_v0);
1430                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1431                 }
1432 #endif
1433                 btrfs_mark_buffer_dirty(leaf);
1434         }
1435         return ret;
1436 }
1437
1438 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1439                                           struct btrfs_extent_inline_ref *iref)
1440 {
1441         struct btrfs_key key;
1442         struct extent_buffer *leaf;
1443         struct btrfs_extent_data_ref *ref1;
1444         struct btrfs_shared_data_ref *ref2;
1445         u32 num_refs = 0;
1446         int type;
1447
1448         leaf = path->nodes[0];
1449         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1450         if (iref) {
1451                 /*
1452                  * If type is invalid, we should have bailed out earlier than
1453                  * this call.
1454                  */
1455                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1456                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1457                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1458                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1459                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1460                 } else {
1461                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1462                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1463                 }
1464         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1465                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1466                                       struct btrfs_extent_data_ref);
1467                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1468         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1469                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1470                                       struct btrfs_shared_data_ref);
1471                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1472 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1473         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1474                 struct btrfs_extent_ref_v0 *ref0;
1475                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1476                                       struct btrfs_extent_ref_v0);
1477                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1478 #endif
1479         } else {
1480                 WARN_ON(1);
1481         }
1482         return num_refs;
1483 }
1484
1485 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1486                                           struct btrfs_path *path,
1487                                           u64 bytenr, u64 parent,
1488                                           u64 root_objectid)
1489 {
1490         struct btrfs_root *root = trans->fs_info->extent_root;
1491         struct btrfs_key key;
1492         int ret;
1493
1494         key.objectid = bytenr;
1495         if (parent) {
1496                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1497                 key.offset = parent;
1498         } else {
1499                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1500                 key.offset = root_objectid;
1501         }
1502
1503         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1504         if (ret > 0)
1505                 ret = -ENOENT;
1506 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1507         if (ret == -ENOENT && parent) {
1508                 btrfs_release_path(path);
1509                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1510                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1511                 if (ret > 0)
1512                         ret = -ENOENT;
1513         }
1514 #endif
1515         return ret;
1516 }
1517
1518 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1519                                           struct btrfs_path *path,
1520                                           u64 bytenr, u64 parent,
1521                                           u64 root_objectid)
1522 {
1523         struct btrfs_key key;
1524         int ret;
1525
1526         key.objectid = bytenr;
1527         if (parent) {
1528                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1529                 key.offset = parent;
1530         } else {
1531                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1532                 key.offset = root_objectid;
1533         }
1534
1535         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1536                                       path, &key, 0);
1537         btrfs_release_path(path);
1538         return ret;
1539 }
1540
1541 static inline int extent_ref_type(u64 parent, u64 owner)
1542 {
1543         int type;
1544         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1545                 if (parent > 0)
1546                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1547                 else
1548                         type = BTRFS_TREE_BLOCK_REF_KEY;
1549         } else {
1550                 if (parent > 0)
1551                         type = BTRFS_SHARED_DATA_REF_KEY;
1552                 else
1553                         type = BTRFS_EXTENT_DATA_REF_KEY;
1554         }
1555         return type;
1556 }
1557
1558 static int find_next_key(struct btrfs_path *path, int level,
1559                          struct btrfs_key *key)
1560
1561 {
1562         for (; level < BTRFS_MAX_LEVEL; level++) {
1563                 if (!path->nodes[level])
1564                         break;
1565                 if (path->slots[level] + 1 >=
1566                     btrfs_header_nritems(path->nodes[level]))
1567                         continue;
1568                 if (level == 0)
1569                         btrfs_item_key_to_cpu(path->nodes[level], key,
1570                                               path->slots[level] + 1);
1571                 else
1572                         btrfs_node_key_to_cpu(path->nodes[level], key,
1573                                               path->slots[level] + 1);
1574                 return 0;
1575         }
1576         return 1;
1577 }
1578
1579 /*
1580  * look for inline back ref. if back ref is found, *ref_ret is set
1581  * to the address of inline back ref, and 0 is returned.
1582  *
1583  * if back ref isn't found, *ref_ret is set to the address where it
1584  * should be inserted, and -ENOENT is returned.
1585  *
1586  * if insert is true and there are too many inline back refs, the path
1587  * points to the extent item, and -EAGAIN is returned.
1588  *
1589  * NOTE: inline back refs are ordered in the same way that back ref
1590  *       items in the tree are ordered.
1591  */
1592 static noinline_for_stack
1593 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1594                                  struct btrfs_path *path,
1595                                  struct btrfs_extent_inline_ref **ref_ret,
1596                                  u64 bytenr, u64 num_bytes,
1597                                  u64 parent, u64 root_objectid,
1598                                  u64 owner, u64 offset, int insert)
1599 {
1600         struct btrfs_fs_info *fs_info = trans->fs_info;
1601         struct btrfs_root *root = fs_info->extent_root;
1602         struct btrfs_key key;
1603         struct extent_buffer *leaf;
1604         struct btrfs_extent_item *ei;
1605         struct btrfs_extent_inline_ref *iref;
1606         u64 flags;
1607         u64 item_size;
1608         unsigned long ptr;
1609         unsigned long end;
1610         int extra_size;
1611         int type;
1612         int want;
1613         int ret;
1614         int err = 0;
1615         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1616         int needed;
1617
1618         key.objectid = bytenr;
1619         key.type = BTRFS_EXTENT_ITEM_KEY;
1620         key.offset = num_bytes;
1621
1622         want = extent_ref_type(parent, owner);
1623         if (insert) {
1624                 extra_size = btrfs_extent_inline_ref_size(want);
1625                 path->keep_locks = 1;
1626         } else
1627                 extra_size = -1;
1628
1629         /*
1630          * Owner is our level, so we can just add one to get the level for the
1631          * block we are interested in.
1632          */
1633         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1634                 key.type = BTRFS_METADATA_ITEM_KEY;
1635                 key.offset = owner;
1636         }
1637
1638 again:
1639         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1640         if (ret < 0) {
1641                 err = ret;
1642                 goto out;
1643         }
1644
1645         /*
1646          * We may be a newly converted file system which still has the old fat
1647          * extent entries for metadata, so try and see if we have one of those.
1648          */
1649         if (ret > 0 && skinny_metadata) {
1650                 skinny_metadata = false;
1651                 if (path->slots[0]) {
1652                         path->slots[0]--;
1653                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1654                                               path->slots[0]);
1655                         if (key.objectid == bytenr &&
1656                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1657                             key.offset == num_bytes)
1658                                 ret = 0;
1659                 }
1660                 if (ret) {
1661                         key.objectid = bytenr;
1662                         key.type = BTRFS_EXTENT_ITEM_KEY;
1663                         key.offset = num_bytes;
1664                         btrfs_release_path(path);
1665                         goto again;
1666                 }
1667         }
1668
1669         if (ret && !insert) {
1670                 err = -ENOENT;
1671                 goto out;
1672         } else if (WARN_ON(ret)) {
1673                 err = -EIO;
1674                 goto out;
1675         }
1676
1677         leaf = path->nodes[0];
1678         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1679 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1680         if (item_size < sizeof(*ei)) {
1681                 if (!insert) {
1682                         err = -ENOENT;
1683                         goto out;
1684                 }
1685                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1686                                              extra_size);
1687                 if (ret < 0) {
1688                         err = ret;
1689                         goto out;
1690                 }
1691                 leaf = path->nodes[0];
1692                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1693         }
1694 #endif
1695         BUG_ON(item_size < sizeof(*ei));
1696
1697         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1698         flags = btrfs_extent_flags(leaf, ei);
1699
1700         ptr = (unsigned long)(ei + 1);
1701         end = (unsigned long)ei + item_size;
1702
1703         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1704                 ptr += sizeof(struct btrfs_tree_block_info);
1705                 BUG_ON(ptr > end);
1706         }
1707
1708         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1709                 needed = BTRFS_REF_TYPE_DATA;
1710         else
1711                 needed = BTRFS_REF_TYPE_BLOCK;
1712
1713         err = -ENOENT;
1714         while (1) {
1715                 if (ptr >= end) {
1716                         WARN_ON(ptr > end);
1717                         break;
1718                 }
1719                 iref = (struct btrfs_extent_inline_ref *)ptr;
1720                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1721                 if (type == BTRFS_REF_TYPE_INVALID) {
1722                         err = -EINVAL;
1723                         goto out;
1724                 }
1725
1726                 if (want < type)
1727                         break;
1728                 if (want > type) {
1729                         ptr += btrfs_extent_inline_ref_size(type);
1730                         continue;
1731                 }
1732
1733                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1734                         struct btrfs_extent_data_ref *dref;
1735                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1736                         if (match_extent_data_ref(leaf, dref, root_objectid,
1737                                                   owner, offset)) {
1738                                 err = 0;
1739                                 break;
1740                         }
1741                         if (hash_extent_data_ref_item(leaf, dref) <
1742                             hash_extent_data_ref(root_objectid, owner, offset))
1743                                 break;
1744                 } else {
1745                         u64 ref_offset;
1746                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1747                         if (parent > 0) {
1748                                 if (parent == ref_offset) {
1749                                         err = 0;
1750                                         break;
1751                                 }
1752                                 if (ref_offset < parent)
1753                                         break;
1754                         } else {
1755                                 if (root_objectid == ref_offset) {
1756                                         err = 0;
1757                                         break;
1758                                 }
1759                                 if (ref_offset < root_objectid)
1760                                         break;
1761                         }
1762                 }
1763                 ptr += btrfs_extent_inline_ref_size(type);
1764         }
1765         if (err == -ENOENT && insert) {
1766                 if (item_size + extra_size >=
1767                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1768                         err = -EAGAIN;
1769                         goto out;
1770                 }
1771                 /*
1772                  * To add new inline back ref, we have to make sure
1773                  * there is no corresponding back ref item.
1774                  * For simplicity, we just do not add new inline back
1775                  * ref if there is any kind of item for this block
1776                  */
1777                 if (find_next_key(path, 0, &key) == 0 &&
1778                     key.objectid == bytenr &&
1779                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1780                         err = -EAGAIN;
1781                         goto out;
1782                 }
1783         }
1784         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1785 out:
1786         if (insert) {
1787                 path->keep_locks = 0;
1788                 btrfs_unlock_up_safe(path, 1);
1789         }
1790         return err;
1791 }
1792
1793 /*
1794  * helper to add new inline back ref
1795  */
1796 static noinline_for_stack
1797 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1798                                  struct btrfs_path *path,
1799                                  struct btrfs_extent_inline_ref *iref,
1800                                  u64 parent, u64 root_objectid,
1801                                  u64 owner, u64 offset, int refs_to_add,
1802                                  struct btrfs_delayed_extent_op *extent_op)
1803 {
1804         struct extent_buffer *leaf;
1805         struct btrfs_extent_item *ei;
1806         unsigned long ptr;
1807         unsigned long end;
1808         unsigned long item_offset;
1809         u64 refs;
1810         int size;
1811         int type;
1812
1813         leaf = path->nodes[0];
1814         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1815         item_offset = (unsigned long)iref - (unsigned long)ei;
1816
1817         type = extent_ref_type(parent, owner);
1818         size = btrfs_extent_inline_ref_size(type);
1819
1820         btrfs_extend_item(fs_info, path, size);
1821
1822         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1823         refs = btrfs_extent_refs(leaf, ei);
1824         refs += refs_to_add;
1825         btrfs_set_extent_refs(leaf, ei, refs);
1826         if (extent_op)
1827                 __run_delayed_extent_op(extent_op, leaf, ei);
1828
1829         ptr = (unsigned long)ei + item_offset;
1830         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1831         if (ptr < end - size)
1832                 memmove_extent_buffer(leaf, ptr + size, ptr,
1833                                       end - size - ptr);
1834
1835         iref = (struct btrfs_extent_inline_ref *)ptr;
1836         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1837         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1838                 struct btrfs_extent_data_ref *dref;
1839                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1840                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1841                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1842                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1843                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1844         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1845                 struct btrfs_shared_data_ref *sref;
1846                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1847                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1848                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1849         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1850                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1851         } else {
1852                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1853         }
1854         btrfs_mark_buffer_dirty(leaf);
1855 }
1856
1857 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1858                                  struct btrfs_path *path,
1859                                  struct btrfs_extent_inline_ref **ref_ret,
1860                                  u64 bytenr, u64 num_bytes, u64 parent,
1861                                  u64 root_objectid, u64 owner, u64 offset)
1862 {
1863         int ret;
1864
1865         ret = lookup_inline_extent_backref(trans, path, ref_ret, bytenr,
1866                                            num_bytes, parent, root_objectid,
1867                                            owner, offset, 0);
1868         if (ret != -ENOENT)
1869                 return ret;
1870
1871         btrfs_release_path(path);
1872         *ref_ret = NULL;
1873
1874         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1875                 ret = lookup_tree_block_ref(trans, path, bytenr, parent,
1876                                             root_objectid);
1877         } else {
1878                 ret = lookup_extent_data_ref(trans, path, bytenr, parent,
1879                                              root_objectid, owner, offset);
1880         }
1881         return ret;
1882 }
1883
1884 /*
1885  * helper to update/remove inline back ref
1886  */
1887 static noinline_for_stack
1888 void update_inline_extent_backref(struct btrfs_path *path,
1889                                   struct btrfs_extent_inline_ref *iref,
1890                                   int refs_to_mod,
1891                                   struct btrfs_delayed_extent_op *extent_op,
1892                                   int *last_ref)
1893 {
1894         struct extent_buffer *leaf = path->nodes[0];
1895         struct btrfs_fs_info *fs_info = leaf->fs_info;
1896         struct btrfs_extent_item *ei;
1897         struct btrfs_extent_data_ref *dref = NULL;
1898         struct btrfs_shared_data_ref *sref = NULL;
1899         unsigned long ptr;
1900         unsigned long end;
1901         u32 item_size;
1902         int size;
1903         int type;
1904         u64 refs;
1905
1906         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1907         refs = btrfs_extent_refs(leaf, ei);
1908         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1909         refs += refs_to_mod;
1910         btrfs_set_extent_refs(leaf, ei, refs);
1911         if (extent_op)
1912                 __run_delayed_extent_op(extent_op, leaf, ei);
1913
1914         /*
1915          * If type is invalid, we should have bailed out after
1916          * lookup_inline_extent_backref().
1917          */
1918         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1919         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1920
1921         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1922                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1923                 refs = btrfs_extent_data_ref_count(leaf, dref);
1924         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1925                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1926                 refs = btrfs_shared_data_ref_count(leaf, sref);
1927         } else {
1928                 refs = 1;
1929                 BUG_ON(refs_to_mod != -1);
1930         }
1931
1932         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1933         refs += refs_to_mod;
1934
1935         if (refs > 0) {
1936                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1937                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1938                 else
1939                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1940         } else {
1941                 *last_ref = 1;
1942                 size =  btrfs_extent_inline_ref_size(type);
1943                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1944                 ptr = (unsigned long)iref;
1945                 end = (unsigned long)ei + item_size;
1946                 if (ptr + size < end)
1947                         memmove_extent_buffer(leaf, ptr, ptr + size,
1948                                               end - ptr - size);
1949                 item_size -= size;
1950                 btrfs_truncate_item(fs_info, path, item_size, 1);
1951         }
1952         btrfs_mark_buffer_dirty(leaf);
1953 }
1954
1955 static noinline_for_stack
1956 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1957                                  struct btrfs_path *path,
1958                                  u64 bytenr, u64 num_bytes, u64 parent,
1959                                  u64 root_objectid, u64 owner,
1960                                  u64 offset, int refs_to_add,
1961                                  struct btrfs_delayed_extent_op *extent_op)
1962 {
1963         struct btrfs_extent_inline_ref *iref;
1964         int ret;
1965
1966         ret = lookup_inline_extent_backref(trans, path, &iref, bytenr,
1967                                            num_bytes, parent, root_objectid,
1968                                            owner, offset, 1);
1969         if (ret == 0) {
1970                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1971                 update_inline_extent_backref(path, iref, refs_to_add,
1972                                              extent_op, NULL);
1973         } else if (ret == -ENOENT) {
1974                 setup_inline_extent_backref(trans->fs_info, path, iref, parent,
1975                                             root_objectid, owner, offset,
1976                                             refs_to_add, extent_op);
1977                 ret = 0;
1978         }
1979         return ret;
1980 }
1981
1982 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1983                                  struct btrfs_path *path,
1984                                  u64 bytenr, u64 parent, u64 root_objectid,
1985                                  u64 owner, u64 offset, int refs_to_add)
1986 {
1987         int ret;
1988         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1989                 BUG_ON(refs_to_add != 1);
1990                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
1991                                             root_objectid);
1992         } else {
1993                 ret = insert_extent_data_ref(trans, path, bytenr, parent,
1994                                              root_objectid, owner, offset,
1995                                              refs_to_add);
1996         }
1997         return ret;
1998 }
1999
2000 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2001                                  struct btrfs_fs_info *fs_info,
2002                                  struct btrfs_path *path,
2003                                  struct btrfs_extent_inline_ref *iref,
2004                                  int refs_to_drop, int is_data, int *last_ref)
2005 {
2006         int ret = 0;
2007
2008         BUG_ON(!is_data && refs_to_drop != 1);
2009         if (iref) {
2010                 update_inline_extent_backref(path, iref, -refs_to_drop, NULL,
2011                                              last_ref);
2012         } else if (is_data) {
2013                 ret = remove_extent_data_ref(trans, path, refs_to_drop,
2014                                              last_ref);
2015         } else {
2016                 *last_ref = 1;
2017                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2018         }
2019         return ret;
2020 }
2021
2022 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2023 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2024                                u64 *discarded_bytes)
2025 {
2026         int j, ret = 0;
2027         u64 bytes_left, end;
2028         u64 aligned_start = ALIGN(start, 1 << 9);
2029
2030         if (WARN_ON(start != aligned_start)) {
2031                 len -= aligned_start - start;
2032                 len = round_down(len, 1 << 9);
2033                 start = aligned_start;
2034         }
2035
2036         *discarded_bytes = 0;
2037
2038         if (!len)
2039                 return 0;
2040
2041         end = start + len;
2042         bytes_left = len;
2043
2044         /* Skip any superblocks on this device. */
2045         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2046                 u64 sb_start = btrfs_sb_offset(j);
2047                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2048                 u64 size = sb_start - start;
2049
2050                 if (!in_range(sb_start, start, bytes_left) &&
2051                     !in_range(sb_end, start, bytes_left) &&
2052                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2053                         continue;
2054
2055                 /*
2056                  * Superblock spans beginning of range.  Adjust start and
2057                  * try again.
2058                  */
2059                 if (sb_start <= start) {
2060                         start += sb_end - start;
2061                         if (start > end) {
2062                                 bytes_left = 0;
2063                                 break;
2064                         }
2065                         bytes_left = end - start;
2066                         continue;
2067                 }
2068
2069                 if (size) {
2070                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2071                                                    GFP_NOFS, 0);
2072                         if (!ret)
2073                                 *discarded_bytes += size;
2074                         else if (ret != -EOPNOTSUPP)
2075                                 return ret;
2076                 }
2077
2078                 start = sb_end;
2079                 if (start > end) {
2080                         bytes_left = 0;
2081                         break;
2082                 }
2083                 bytes_left = end - start;
2084         }
2085
2086         if (bytes_left) {
2087                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2088                                            GFP_NOFS, 0);
2089                 if (!ret)
2090                         *discarded_bytes += bytes_left;
2091         }
2092         return ret;
2093 }
2094
2095 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2096                          u64 num_bytes, u64 *actual_bytes)
2097 {
2098         int ret;
2099         u64 discarded_bytes = 0;
2100         struct btrfs_bio *bbio = NULL;
2101
2102
2103         /*
2104          * Avoid races with device replace and make sure our bbio has devices
2105          * associated to its stripes that don't go away while we are discarding.
2106          */
2107         btrfs_bio_counter_inc_blocked(fs_info);
2108         /* Tell the block device(s) that the sectors can be discarded */
2109         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2110                               &bbio, 0);
2111         /* Error condition is -ENOMEM */
2112         if (!ret) {
2113                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2114                 int i;
2115
2116
2117                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2118                         u64 bytes;
2119                         struct request_queue *req_q;
2120
2121                         if (!stripe->dev->bdev) {
2122                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2123                                 continue;
2124                         }
2125                         req_q = bdev_get_queue(stripe->dev->bdev);
2126                         if (!blk_queue_discard(req_q))
2127                                 continue;
2128
2129                         ret = btrfs_issue_discard(stripe->dev->bdev,
2130                                                   stripe->physical,
2131                                                   stripe->length,
2132                                                   &bytes);
2133                         if (!ret)
2134                                 discarded_bytes += bytes;
2135                         else if (ret != -EOPNOTSUPP)
2136                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2137
2138                         /*
2139                          * Just in case we get back EOPNOTSUPP for some reason,
2140                          * just ignore the return value so we don't screw up
2141                          * people calling discard_extent.
2142                          */
2143                         ret = 0;
2144                 }
2145                 btrfs_put_bbio(bbio);
2146         }
2147         btrfs_bio_counter_dec(fs_info);
2148
2149         if (actual_bytes)
2150                 *actual_bytes = discarded_bytes;
2151
2152
2153         if (ret == -EOPNOTSUPP)
2154                 ret = 0;
2155         return ret;
2156 }
2157
2158 /* Can return -ENOMEM */
2159 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2160                          struct btrfs_root *root,
2161                          u64 bytenr, u64 num_bytes, u64 parent,
2162                          u64 root_objectid, u64 owner, u64 offset)
2163 {
2164         struct btrfs_fs_info *fs_info = root->fs_info;
2165         int old_ref_mod, new_ref_mod;
2166         int ret;
2167
2168         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2169                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2170
2171         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2172                            owner, offset, BTRFS_ADD_DELAYED_REF);
2173
2174         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2175                 ret = btrfs_add_delayed_tree_ref(trans, bytenr,
2176                                                  num_bytes, parent,
2177                                                  root_objectid, (int)owner,
2178                                                  BTRFS_ADD_DELAYED_REF, NULL,
2179                                                  &old_ref_mod, &new_ref_mod);
2180         } else {
2181                 ret = btrfs_add_delayed_data_ref(trans, bytenr,
2182                                                  num_bytes, parent,
2183                                                  root_objectid, owner, offset,
2184                                                  0, BTRFS_ADD_DELAYED_REF,
2185                                                  &old_ref_mod, &new_ref_mod);
2186         }
2187
2188         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2189                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2190
2191                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2192         }
2193
2194         return ret;
2195 }
2196
2197 /*
2198  * __btrfs_inc_extent_ref - insert backreference for a given extent
2199  *
2200  * @trans:          Handle of transaction
2201  *
2202  * @node:           The delayed ref node used to get the bytenr/length for
2203  *                  extent whose references are incremented.
2204  *
2205  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2206  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2207  *                  bytenr of the parent block. Since new extents are always
2208  *                  created with indirect references, this will only be the case
2209  *                  when relocating a shared extent. In that case, root_objectid
2210  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2211  *                  be 0
2212  *
2213  * @root_objectid:  The id of the root where this modification has originated,
2214  *                  this can be either one of the well-known metadata trees or
2215  *                  the subvolume id which references this extent.
2216  *
2217  * @owner:          For data extents it is the inode number of the owning file.
2218  *                  For metadata extents this parameter holds the level in the
2219  *                  tree of the extent.
2220  *
2221  * @offset:         For metadata extents the offset is ignored and is currently
2222  *                  always passed as 0. For data extents it is the fileoffset
2223  *                  this extent belongs to.
2224  *
2225  * @refs_to_add     Number of references to add
2226  *
2227  * @extent_op       Pointer to a structure, holding information necessary when
2228  *                  updating a tree block's flags
2229  *
2230  */
2231 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2232                                   struct btrfs_delayed_ref_node *node,
2233                                   u64 parent, u64 root_objectid,
2234                                   u64 owner, u64 offset, int refs_to_add,
2235                                   struct btrfs_delayed_extent_op *extent_op)
2236 {
2237         struct btrfs_path *path;
2238         struct extent_buffer *leaf;
2239         struct btrfs_extent_item *item;
2240         struct btrfs_key key;
2241         u64 bytenr = node->bytenr;
2242         u64 num_bytes = node->num_bytes;
2243         u64 refs;
2244         int ret;
2245
2246         path = btrfs_alloc_path();
2247         if (!path)
2248                 return -ENOMEM;
2249
2250         path->reada = READA_FORWARD;
2251         path->leave_spinning = 1;
2252         /* this will setup the path even if it fails to insert the back ref */
2253         ret = insert_inline_extent_backref(trans, path, bytenr, num_bytes,
2254                                            parent, root_objectid, owner,
2255                                            offset, refs_to_add, extent_op);
2256         if ((ret < 0 && ret != -EAGAIN) || !ret)
2257                 goto out;
2258
2259         /*
2260          * Ok we had -EAGAIN which means we didn't have space to insert and
2261          * inline extent ref, so just update the reference count and add a
2262          * normal backref.
2263          */
2264         leaf = path->nodes[0];
2265         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2266         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2267         refs = btrfs_extent_refs(leaf, item);
2268         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2269         if (extent_op)
2270                 __run_delayed_extent_op(extent_op, leaf, item);
2271
2272         btrfs_mark_buffer_dirty(leaf);
2273         btrfs_release_path(path);
2274
2275         path->reada = READA_FORWARD;
2276         path->leave_spinning = 1;
2277         /* now insert the actual backref */
2278         ret = insert_extent_backref(trans, path, bytenr, parent, root_objectid,
2279                                     owner, offset, refs_to_add);
2280         if (ret)
2281                 btrfs_abort_transaction(trans, ret);
2282 out:
2283         btrfs_free_path(path);
2284         return ret;
2285 }
2286
2287 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2288                                 struct btrfs_delayed_ref_node *node,
2289                                 struct btrfs_delayed_extent_op *extent_op,
2290                                 int insert_reserved)
2291 {
2292         int ret = 0;
2293         struct btrfs_delayed_data_ref *ref;
2294         struct btrfs_key ins;
2295         u64 parent = 0;
2296         u64 ref_root = 0;
2297         u64 flags = 0;
2298
2299         ins.objectid = node->bytenr;
2300         ins.offset = node->num_bytes;
2301         ins.type = BTRFS_EXTENT_ITEM_KEY;
2302
2303         ref = btrfs_delayed_node_to_data_ref(node);
2304         trace_run_delayed_data_ref(trans->fs_info, node, ref, node->action);
2305
2306         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2307                 parent = ref->parent;
2308         ref_root = ref->root;
2309
2310         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2311                 if (extent_op)
2312                         flags |= extent_op->flags_to_set;
2313                 ret = alloc_reserved_file_extent(trans, parent, ref_root,
2314                                                  flags, ref->objectid,
2315                                                  ref->offset, &ins,
2316                                                  node->ref_mod);
2317         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2318                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2319                                              ref->objectid, ref->offset,
2320                                              node->ref_mod, extent_op);
2321         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2322                 ret = __btrfs_free_extent(trans, node, parent,
2323                                           ref_root, ref->objectid,
2324                                           ref->offset, node->ref_mod,
2325                                           extent_op);
2326         } else {
2327                 BUG();
2328         }
2329         return ret;
2330 }
2331
2332 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2333                                     struct extent_buffer *leaf,
2334                                     struct btrfs_extent_item *ei)
2335 {
2336         u64 flags = btrfs_extent_flags(leaf, ei);
2337         if (extent_op->update_flags) {
2338                 flags |= extent_op->flags_to_set;
2339                 btrfs_set_extent_flags(leaf, ei, flags);
2340         }
2341
2342         if (extent_op->update_key) {
2343                 struct btrfs_tree_block_info *bi;
2344                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2345                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2346                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2347         }
2348 }
2349
2350 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2351                                  struct btrfs_delayed_ref_head *head,
2352                                  struct btrfs_delayed_extent_op *extent_op)
2353 {
2354         struct btrfs_fs_info *fs_info = trans->fs_info;
2355         struct btrfs_key key;
2356         struct btrfs_path *path;
2357         struct btrfs_extent_item *ei;
2358         struct extent_buffer *leaf;
2359         u32 item_size;
2360         int ret;
2361         int err = 0;
2362         int metadata = !extent_op->is_data;
2363
2364         if (trans->aborted)
2365                 return 0;
2366
2367         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2368                 metadata = 0;
2369
2370         path = btrfs_alloc_path();
2371         if (!path)
2372                 return -ENOMEM;
2373
2374         key.objectid = head->bytenr;
2375
2376         if (metadata) {
2377                 key.type = BTRFS_METADATA_ITEM_KEY;
2378                 key.offset = extent_op->level;
2379         } else {
2380                 key.type = BTRFS_EXTENT_ITEM_KEY;
2381                 key.offset = head->num_bytes;
2382         }
2383
2384 again:
2385         path->reada = READA_FORWARD;
2386         path->leave_spinning = 1;
2387         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2388         if (ret < 0) {
2389                 err = ret;
2390                 goto out;
2391         }
2392         if (ret > 0) {
2393                 if (metadata) {
2394                         if (path->slots[0] > 0) {
2395                                 path->slots[0]--;
2396                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2397                                                       path->slots[0]);
2398                                 if (key.objectid == head->bytenr &&
2399                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2400                                     key.offset == head->num_bytes)
2401                                         ret = 0;
2402                         }
2403                         if (ret > 0) {
2404                                 btrfs_release_path(path);
2405                                 metadata = 0;
2406
2407                                 key.objectid = head->bytenr;
2408                                 key.offset = head->num_bytes;
2409                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2410                                 goto again;
2411                         }
2412                 } else {
2413                         err = -EIO;
2414                         goto out;
2415                 }
2416         }
2417
2418         leaf = path->nodes[0];
2419         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2420 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2421         if (item_size < sizeof(*ei)) {
2422                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2423                 if (ret < 0) {
2424                         err = ret;
2425                         goto out;
2426                 }
2427                 leaf = path->nodes[0];
2428                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2429         }
2430 #endif
2431         BUG_ON(item_size < sizeof(*ei));
2432         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2433         __run_delayed_extent_op(extent_op, leaf, ei);
2434
2435         btrfs_mark_buffer_dirty(leaf);
2436 out:
2437         btrfs_free_path(path);
2438         return err;
2439 }
2440
2441 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2442                                 struct btrfs_delayed_ref_node *node,
2443                                 struct btrfs_delayed_extent_op *extent_op,
2444                                 int insert_reserved)
2445 {
2446         int ret = 0;
2447         struct btrfs_delayed_tree_ref *ref;
2448         u64 parent = 0;
2449         u64 ref_root = 0;
2450
2451         ref = btrfs_delayed_node_to_tree_ref(node);
2452         trace_run_delayed_tree_ref(trans->fs_info, node, ref, node->action);
2453
2454         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2455                 parent = ref->parent;
2456         ref_root = ref->root;
2457
2458         if (node->ref_mod != 1) {
2459                 btrfs_err(trans->fs_info,
2460         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2461                           node->bytenr, node->ref_mod, node->action, ref_root,
2462                           parent);
2463                 return -EIO;
2464         }
2465         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2466                 BUG_ON(!extent_op || !extent_op->update_flags);
2467                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2468         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2469                 ret = __btrfs_inc_extent_ref(trans, node, parent, ref_root,
2470                                              ref->level, 0, 1, extent_op);
2471         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2472                 ret = __btrfs_free_extent(trans, node, parent, ref_root,
2473                                           ref->level, 0, 1, extent_op);
2474         } else {
2475                 BUG();
2476         }
2477         return ret;
2478 }
2479
2480 /* helper function to actually process a single delayed ref entry */
2481 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2482                                struct btrfs_fs_info *fs_info,
2483                                struct btrfs_delayed_ref_node *node,
2484                                struct btrfs_delayed_extent_op *extent_op,
2485                                int insert_reserved)
2486 {
2487         int ret = 0;
2488
2489         if (trans->aborted) {
2490                 if (insert_reserved)
2491                         btrfs_pin_extent(fs_info, node->bytenr,
2492                                          node->num_bytes, 1);
2493                 return 0;
2494         }
2495
2496         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2497             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2498                 ret = run_delayed_tree_ref(trans, node, extent_op,
2499                                            insert_reserved);
2500         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2501                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2502                 ret = run_delayed_data_ref(trans, node, extent_op,
2503                                            insert_reserved);
2504         else
2505                 BUG();
2506         return ret;
2507 }
2508
2509 static inline struct btrfs_delayed_ref_node *
2510 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2511 {
2512         struct btrfs_delayed_ref_node *ref;
2513
2514         if (RB_EMPTY_ROOT(&head->ref_tree))
2515                 return NULL;
2516
2517         /*
2518          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2519          * This is to prevent a ref count from going down to zero, which deletes
2520          * the extent item from the extent tree, when there still are references
2521          * to add, which would fail because they would not find the extent item.
2522          */
2523         if (!list_empty(&head->ref_add_list))
2524                 return list_first_entry(&head->ref_add_list,
2525                                 struct btrfs_delayed_ref_node, add_list);
2526
2527         ref = rb_entry(rb_first(&head->ref_tree),
2528                        struct btrfs_delayed_ref_node, ref_node);
2529         ASSERT(list_empty(&ref->add_list));
2530         return ref;
2531 }
2532
2533 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2534                                       struct btrfs_delayed_ref_head *head)
2535 {
2536         spin_lock(&delayed_refs->lock);
2537         head->processing = 0;
2538         delayed_refs->num_heads_ready++;
2539         spin_unlock(&delayed_refs->lock);
2540         btrfs_delayed_ref_unlock(head);
2541 }
2542
2543 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2544                              struct btrfs_delayed_ref_head *head)
2545 {
2546         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2547         int ret;
2548
2549         if (!extent_op)
2550                 return 0;
2551         head->extent_op = NULL;
2552         if (head->must_insert_reserved) {
2553                 btrfs_free_delayed_extent_op(extent_op);
2554                 return 0;
2555         }
2556         spin_unlock(&head->lock);
2557         ret = run_delayed_extent_op(trans, head, extent_op);
2558         btrfs_free_delayed_extent_op(extent_op);
2559         return ret ? ret : 1;
2560 }
2561
2562 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2563                             struct btrfs_delayed_ref_head *head)
2564 {
2565
2566         struct btrfs_fs_info *fs_info = trans->fs_info;
2567         struct btrfs_delayed_ref_root *delayed_refs;
2568         int ret;
2569
2570         delayed_refs = &trans->transaction->delayed_refs;
2571
2572         ret = cleanup_extent_op(trans, head);
2573         if (ret < 0) {
2574                 unselect_delayed_ref_head(delayed_refs, head);
2575                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2576                 return ret;
2577         } else if (ret) {
2578                 return ret;
2579         }
2580
2581         /*
2582          * Need to drop our head ref lock and re-acquire the delayed ref lock
2583          * and then re-check to make sure nobody got added.
2584          */
2585         spin_unlock(&head->lock);
2586         spin_lock(&delayed_refs->lock);
2587         spin_lock(&head->lock);
2588         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2589                 spin_unlock(&head->lock);
2590                 spin_unlock(&delayed_refs->lock);
2591                 return 1;
2592         }
2593         delayed_refs->num_heads--;
2594         rb_erase(&head->href_node, &delayed_refs->href_root);
2595         RB_CLEAR_NODE(&head->href_node);
2596         spin_unlock(&head->lock);
2597         spin_unlock(&delayed_refs->lock);
2598         atomic_dec(&delayed_refs->num_entries);
2599
2600         trace_run_delayed_ref_head(fs_info, head, 0);
2601
2602         if (head->total_ref_mod < 0) {
2603                 struct btrfs_space_info *space_info;
2604                 u64 flags;
2605
2606                 if (head->is_data)
2607                         flags = BTRFS_BLOCK_GROUP_DATA;
2608                 else if (head->is_system)
2609                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2610                 else
2611                         flags = BTRFS_BLOCK_GROUP_METADATA;
2612                 space_info = __find_space_info(fs_info, flags);
2613                 ASSERT(space_info);
2614                 percpu_counter_add(&space_info->total_bytes_pinned,
2615                                    -head->num_bytes);
2616
2617                 if (head->is_data) {
2618                         spin_lock(&delayed_refs->lock);
2619                         delayed_refs->pending_csums -= head->num_bytes;
2620                         spin_unlock(&delayed_refs->lock);
2621                 }
2622         }
2623
2624         if (head->must_insert_reserved) {
2625                 btrfs_pin_extent(fs_info, head->bytenr,
2626                                  head->num_bytes, 1);
2627                 if (head->is_data) {
2628                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2629                                               head->num_bytes);
2630                 }
2631         }
2632
2633         /* Also free its reserved qgroup space */
2634         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2635                                       head->qgroup_reserved);
2636         btrfs_delayed_ref_unlock(head);
2637         btrfs_put_delayed_ref_head(head);
2638         return 0;
2639 }
2640
2641 /*
2642  * Returns 0 on success or if called with an already aborted transaction.
2643  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2644  */
2645 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2646                                              unsigned long nr)
2647 {
2648         struct btrfs_fs_info *fs_info = trans->fs_info;
2649         struct btrfs_delayed_ref_root *delayed_refs;
2650         struct btrfs_delayed_ref_node *ref;
2651         struct btrfs_delayed_ref_head *locked_ref = NULL;
2652         struct btrfs_delayed_extent_op *extent_op;
2653         ktime_t start = ktime_get();
2654         int ret;
2655         unsigned long count = 0;
2656         unsigned long actual_count = 0;
2657         int must_insert_reserved = 0;
2658
2659         delayed_refs = &trans->transaction->delayed_refs;
2660         while (1) {
2661                 if (!locked_ref) {
2662                         if (count >= nr)
2663                                 break;
2664
2665                         spin_lock(&delayed_refs->lock);
2666                         locked_ref = btrfs_select_ref_head(trans);
2667                         if (!locked_ref) {
2668                                 spin_unlock(&delayed_refs->lock);
2669                                 break;
2670                         }
2671
2672                         /* grab the lock that says we are going to process
2673                          * all the refs for this head */
2674                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2675                         spin_unlock(&delayed_refs->lock);
2676                         /*
2677                          * we may have dropped the spin lock to get the head
2678                          * mutex lock, and that might have given someone else
2679                          * time to free the head.  If that's true, it has been
2680                          * removed from our list and we can move on.
2681                          */
2682                         if (ret == -EAGAIN) {
2683                                 locked_ref = NULL;
2684                                 count++;
2685                                 continue;
2686                         }
2687                 }
2688
2689                 /*
2690                  * We need to try and merge add/drops of the same ref since we
2691                  * can run into issues with relocate dropping the implicit ref
2692                  * and then it being added back again before the drop can
2693                  * finish.  If we merged anything we need to re-loop so we can
2694                  * get a good ref.
2695                  * Or we can get node references of the same type that weren't
2696                  * merged when created due to bumps in the tree mod seq, and
2697                  * we need to merge them to prevent adding an inline extent
2698                  * backref before dropping it (triggering a BUG_ON at
2699                  * insert_inline_extent_backref()).
2700                  */
2701                 spin_lock(&locked_ref->lock);
2702                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2703
2704                 ref = select_delayed_ref(locked_ref);
2705
2706                 if (ref && ref->seq &&
2707                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2708                         spin_unlock(&locked_ref->lock);
2709                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2710                         locked_ref = NULL;
2711                         cond_resched();
2712                         count++;
2713                         continue;
2714                 }
2715
2716                 /*
2717                  * We're done processing refs in this ref_head, clean everything
2718                  * up and move on to the next ref_head.
2719                  */
2720                 if (!ref) {
2721                         ret = cleanup_ref_head(trans, locked_ref);
2722                         if (ret > 0 ) {
2723                                 /* We dropped our lock, we need to loop. */
2724                                 ret = 0;
2725                                 continue;
2726                         } else if (ret) {
2727                                 return ret;
2728                         }
2729                         locked_ref = NULL;
2730                         count++;
2731                         continue;
2732                 }
2733
2734                 actual_count++;
2735                 ref->in_tree = 0;
2736                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2737                 RB_CLEAR_NODE(&ref->ref_node);
2738                 if (!list_empty(&ref->add_list))
2739                         list_del(&ref->add_list);
2740                 /*
2741                  * When we play the delayed ref, also correct the ref_mod on
2742                  * head
2743                  */
2744                 switch (ref->action) {
2745                 case BTRFS_ADD_DELAYED_REF:
2746                 case BTRFS_ADD_DELAYED_EXTENT:
2747                         locked_ref->ref_mod -= ref->ref_mod;
2748                         break;
2749                 case BTRFS_DROP_DELAYED_REF:
2750                         locked_ref->ref_mod += ref->ref_mod;
2751                         break;
2752                 default:
2753                         WARN_ON(1);
2754                 }
2755                 atomic_dec(&delayed_refs->num_entries);
2756
2757                 /*
2758                  * Record the must-insert_reserved flag before we drop the spin
2759                  * lock.
2760                  */
2761                 must_insert_reserved = locked_ref->must_insert_reserved;
2762                 locked_ref->must_insert_reserved = 0;
2763
2764                 extent_op = locked_ref->extent_op;
2765                 locked_ref->extent_op = NULL;
2766                 spin_unlock(&locked_ref->lock);
2767
2768                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2769                                           must_insert_reserved);
2770
2771                 btrfs_free_delayed_extent_op(extent_op);
2772                 if (ret) {
2773                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2774                         btrfs_put_delayed_ref(ref);
2775                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2776                                     ret);
2777                         return ret;
2778                 }
2779
2780                 btrfs_put_delayed_ref(ref);
2781                 count++;
2782                 cond_resched();
2783         }
2784
2785         /*
2786          * We don't want to include ref heads since we can have empty ref heads
2787          * and those will drastically skew our runtime down since we just do
2788          * accounting, no actual extent tree updates.
2789          */
2790         if (actual_count > 0) {
2791                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2792                 u64 avg;
2793
2794                 /*
2795                  * We weigh the current average higher than our current runtime
2796                  * to avoid large swings in the average.
2797                  */
2798                 spin_lock(&delayed_refs->lock);
2799                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2800                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2801                 spin_unlock(&delayed_refs->lock);
2802         }
2803         return 0;
2804 }
2805
2806 #ifdef SCRAMBLE_DELAYED_REFS
2807 /*
2808  * Normally delayed refs get processed in ascending bytenr order. This
2809  * correlates in most cases to the order added. To expose dependencies on this
2810  * order, we start to process the tree in the middle instead of the beginning
2811  */
2812 static u64 find_middle(struct rb_root *root)
2813 {
2814         struct rb_node *n = root->rb_node;
2815         struct btrfs_delayed_ref_node *entry;
2816         int alt = 1;
2817         u64 middle;
2818         u64 first = 0, last = 0;
2819
2820         n = rb_first(root);
2821         if (n) {
2822                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2823                 first = entry->bytenr;
2824         }
2825         n = rb_last(root);
2826         if (n) {
2827                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2828                 last = entry->bytenr;
2829         }
2830         n = root->rb_node;
2831
2832         while (n) {
2833                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2834                 WARN_ON(!entry->in_tree);
2835
2836                 middle = entry->bytenr;
2837
2838                 if (alt)
2839                         n = n->rb_left;
2840                 else
2841                         n = n->rb_right;
2842
2843                 alt = 1 - alt;
2844         }
2845         return middle;
2846 }
2847 #endif
2848
2849 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2850 {
2851         u64 num_bytes;
2852
2853         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2854                              sizeof(struct btrfs_extent_inline_ref));
2855         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2856                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2857
2858         /*
2859          * We don't ever fill up leaves all the way so multiply by 2 just to be
2860          * closer to what we're really going to want to use.
2861          */
2862         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2863 }
2864
2865 /*
2866  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2867  * would require to store the csums for that many bytes.
2868  */
2869 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2870 {
2871         u64 csum_size;
2872         u64 num_csums_per_leaf;
2873         u64 num_csums;
2874
2875         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2876         num_csums_per_leaf = div64_u64(csum_size,
2877                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2878         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2879         num_csums += num_csums_per_leaf - 1;
2880         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2881         return num_csums;
2882 }
2883
2884 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2885                                        struct btrfs_fs_info *fs_info)
2886 {
2887         struct btrfs_block_rsv *global_rsv;
2888         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2889         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2890         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2891         u64 num_bytes, num_dirty_bgs_bytes;
2892         int ret = 0;
2893
2894         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2895         num_heads = heads_to_leaves(fs_info, num_heads);
2896         if (num_heads > 1)
2897                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2898         num_bytes <<= 1;
2899         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2900                                                         fs_info->nodesize;
2901         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2902                                                              num_dirty_bgs);
2903         global_rsv = &fs_info->global_block_rsv;
2904
2905         /*
2906          * If we can't allocate any more chunks lets make sure we have _lots_ of
2907          * wiggle room since running delayed refs can create more delayed refs.
2908          */
2909         if (global_rsv->space_info->full) {
2910                 num_dirty_bgs_bytes <<= 1;
2911                 num_bytes <<= 1;
2912         }
2913
2914         spin_lock(&global_rsv->lock);
2915         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2916                 ret = 1;
2917         spin_unlock(&global_rsv->lock);
2918         return ret;
2919 }
2920
2921 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2922                                        struct btrfs_fs_info *fs_info)
2923 {
2924         u64 num_entries =
2925                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2926         u64 avg_runtime;
2927         u64 val;
2928
2929         smp_mb();
2930         avg_runtime = fs_info->avg_delayed_ref_runtime;
2931         val = num_entries * avg_runtime;
2932         if (val >= NSEC_PER_SEC)
2933                 return 1;
2934         if (val >= NSEC_PER_SEC / 2)
2935                 return 2;
2936
2937         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2938 }
2939
2940 struct async_delayed_refs {
2941         struct btrfs_root *root;
2942         u64 transid;
2943         int count;
2944         int error;
2945         int sync;
2946         struct completion wait;
2947         struct btrfs_work work;
2948 };
2949
2950 static inline struct async_delayed_refs *
2951 to_async_delayed_refs(struct btrfs_work *work)
2952 {
2953         return container_of(work, struct async_delayed_refs, work);
2954 }
2955
2956 static void delayed_ref_async_start(struct btrfs_work *work)
2957 {
2958         struct async_delayed_refs *async = to_async_delayed_refs(work);
2959         struct btrfs_trans_handle *trans;
2960         struct btrfs_fs_info *fs_info = async->root->fs_info;
2961         int ret;
2962
2963         /* if the commit is already started, we don't need to wait here */
2964         if (btrfs_transaction_blocked(fs_info))
2965                 goto done;
2966
2967         trans = btrfs_join_transaction(async->root);
2968         if (IS_ERR(trans)) {
2969                 async->error = PTR_ERR(trans);
2970                 goto done;
2971         }
2972
2973         /*
2974          * trans->sync means that when we call end_transaction, we won't
2975          * wait on delayed refs
2976          */
2977         trans->sync = true;
2978
2979         /* Don't bother flushing if we got into a different transaction */
2980         if (trans->transid > async->transid)
2981                 goto end;
2982
2983         ret = btrfs_run_delayed_refs(trans, async->count);
2984         if (ret)
2985                 async->error = ret;
2986 end:
2987         ret = btrfs_end_transaction(trans);
2988         if (ret && !async->error)
2989                 async->error = ret;
2990 done:
2991         if (async->sync)
2992                 complete(&async->wait);
2993         else
2994                 kfree(async);
2995 }
2996
2997 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
2998                                  unsigned long count, u64 transid, int wait)
2999 {
3000         struct async_delayed_refs *async;
3001         int ret;
3002
3003         async = kmalloc(sizeof(*async), GFP_NOFS);
3004         if (!async)
3005                 return -ENOMEM;
3006
3007         async->root = fs_info->tree_root;
3008         async->count = count;
3009         async->error = 0;
3010         async->transid = transid;
3011         if (wait)
3012                 async->sync = 1;
3013         else
3014                 async->sync = 0;
3015         init_completion(&async->wait);
3016
3017         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3018                         delayed_ref_async_start, NULL, NULL);
3019
3020         btrfs_queue_work(fs_info->extent_workers, &async->work);
3021
3022         if (wait) {
3023                 wait_for_completion(&async->wait);
3024                 ret = async->error;
3025                 kfree(async);
3026                 return ret;
3027         }
3028         return 0;
3029 }
3030
3031 /*
3032  * this starts processing the delayed reference count updates and
3033  * extent insertions we have queued up so far.  count can be
3034  * 0, which means to process everything in the tree at the start
3035  * of the run (but not newly added entries), or it can be some target
3036  * number you'd like to process.
3037  *
3038  * Returns 0 on success or if called with an aborted transaction
3039  * Returns <0 on error and aborts the transaction
3040  */
3041 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3042                            unsigned long count)
3043 {
3044         struct btrfs_fs_info *fs_info = trans->fs_info;
3045         struct rb_node *node;
3046         struct btrfs_delayed_ref_root *delayed_refs;
3047         struct btrfs_delayed_ref_head *head;
3048         int ret;
3049         int run_all = count == (unsigned long)-1;
3050         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
3051
3052         /* We'll clean this up in btrfs_cleanup_transaction */
3053         if (trans->aborted)
3054                 return 0;
3055
3056         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags))
3057                 return 0;
3058
3059         delayed_refs = &trans->transaction->delayed_refs;
3060         if (count == 0)
3061                 count = atomic_read(&delayed_refs->num_entries) * 2;
3062
3063 again:
3064 #ifdef SCRAMBLE_DELAYED_REFS
3065         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
3066 #endif
3067         trans->can_flush_pending_bgs = false;
3068         ret = __btrfs_run_delayed_refs(trans, count);
3069         if (ret < 0) {
3070                 btrfs_abort_transaction(trans, ret);
3071                 return ret;
3072         }
3073
3074         if (run_all) {
3075                 if (!list_empty(&trans->new_bgs))
3076                         btrfs_create_pending_block_groups(trans);
3077
3078                 spin_lock(&delayed_refs->lock);
3079                 node = rb_first(&delayed_refs->href_root);
3080                 if (!node) {
3081                         spin_unlock(&delayed_refs->lock);
3082                         goto out;
3083                 }
3084                 head = rb_entry(node, struct btrfs_delayed_ref_head,
3085                                 href_node);
3086                 refcount_inc(&head->refs);
3087                 spin_unlock(&delayed_refs->lock);
3088
3089                 /* Mutex was contended, block until it's released and retry. */
3090                 mutex_lock(&head->mutex);
3091                 mutex_unlock(&head->mutex);
3092
3093                 btrfs_put_delayed_ref_head(head);
3094                 cond_resched();
3095                 goto again;
3096         }
3097 out:
3098         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3099         return 0;
3100 }
3101
3102 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3103                                 struct btrfs_fs_info *fs_info,
3104                                 u64 bytenr, u64 num_bytes, u64 flags,
3105                                 int level, int is_data)
3106 {
3107         struct btrfs_delayed_extent_op *extent_op;
3108         int ret;
3109
3110         extent_op = btrfs_alloc_delayed_extent_op();
3111         if (!extent_op)
3112                 return -ENOMEM;
3113
3114         extent_op->flags_to_set = flags;
3115         extent_op->update_flags = true;
3116         extent_op->update_key = false;
3117         extent_op->is_data = is_data ? true : false;
3118         extent_op->level = level;
3119
3120         ret = btrfs_add_delayed_extent_op(fs_info, trans, bytenr,
3121                                           num_bytes, extent_op);
3122         if (ret)
3123                 btrfs_free_delayed_extent_op(extent_op);
3124         return ret;
3125 }
3126
3127 static noinline int check_delayed_ref(struct btrfs_root *root,
3128                                       struct btrfs_path *path,
3129                                       u64 objectid, u64 offset, u64 bytenr)
3130 {
3131         struct btrfs_delayed_ref_head *head;
3132         struct btrfs_delayed_ref_node *ref;
3133         struct btrfs_delayed_data_ref *data_ref;
3134         struct btrfs_delayed_ref_root *delayed_refs;
3135         struct btrfs_transaction *cur_trans;
3136         struct rb_node *node;
3137         int ret = 0;
3138
3139         spin_lock(&root->fs_info->trans_lock);
3140         cur_trans = root->fs_info->running_transaction;
3141         if (cur_trans)
3142                 refcount_inc(&cur_trans->use_count);
3143         spin_unlock(&root->fs_info->trans_lock);
3144         if (!cur_trans)
3145                 return 0;
3146
3147         delayed_refs = &cur_trans->delayed_refs;
3148         spin_lock(&delayed_refs->lock);
3149         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
3150         if (!head) {
3151                 spin_unlock(&delayed_refs->lock);
3152                 btrfs_put_transaction(cur_trans);
3153                 return 0;
3154         }
3155
3156         if (!mutex_trylock(&head->mutex)) {
3157                 refcount_inc(&head->refs);
3158                 spin_unlock(&delayed_refs->lock);
3159
3160                 btrfs_release_path(path);
3161
3162                 /*
3163                  * Mutex was contended, block until it's released and let
3164                  * caller try again
3165                  */
3166                 mutex_lock(&head->mutex);
3167                 mutex_unlock(&head->mutex);
3168                 btrfs_put_delayed_ref_head(head);
3169                 btrfs_put_transaction(cur_trans);
3170                 return -EAGAIN;
3171         }
3172         spin_unlock(&delayed_refs->lock);
3173
3174         spin_lock(&head->lock);
3175         /*
3176          * XXX: We should replace this with a proper search function in the
3177          * future.
3178          */
3179         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
3180                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
3181                 /* If it's a shared ref we know a cross reference exists */
3182                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3183                         ret = 1;
3184                         break;
3185                 }
3186
3187                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3188
3189                 /*
3190                  * If our ref doesn't match the one we're currently looking at
3191                  * then we have a cross reference.
3192                  */
3193                 if (data_ref->root != root->root_key.objectid ||
3194                     data_ref->objectid != objectid ||
3195                     data_ref->offset != offset) {
3196                         ret = 1;
3197                         break;
3198                 }
3199         }
3200         spin_unlock(&head->lock);
3201         mutex_unlock(&head->mutex);
3202         btrfs_put_transaction(cur_trans);
3203         return ret;
3204 }
3205
3206 static noinline int check_committed_ref(struct btrfs_root *root,
3207                                         struct btrfs_path *path,
3208                                         u64 objectid, u64 offset, u64 bytenr)
3209 {
3210         struct btrfs_fs_info *fs_info = root->fs_info;
3211         struct btrfs_root *extent_root = fs_info->extent_root;
3212         struct extent_buffer *leaf;
3213         struct btrfs_extent_data_ref *ref;
3214         struct btrfs_extent_inline_ref *iref;
3215         struct btrfs_extent_item *ei;
3216         struct btrfs_key key;
3217         u32 item_size;
3218         int type;
3219         int ret;
3220
3221         key.objectid = bytenr;
3222         key.offset = (u64)-1;
3223         key.type = BTRFS_EXTENT_ITEM_KEY;
3224
3225         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3226         if (ret < 0)
3227                 goto out;
3228         BUG_ON(ret == 0); /* Corruption */
3229
3230         ret = -ENOENT;
3231         if (path->slots[0] == 0)
3232                 goto out;
3233
3234         path->slots[0]--;
3235         leaf = path->nodes[0];
3236         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3237
3238         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3239                 goto out;
3240
3241         ret = 1;
3242         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3243 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3244         if (item_size < sizeof(*ei)) {
3245                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3246                 goto out;
3247         }
3248 #endif
3249         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3250
3251         if (item_size != sizeof(*ei) +
3252             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3253                 goto out;
3254
3255         if (btrfs_extent_generation(leaf, ei) <=
3256             btrfs_root_last_snapshot(&root->root_item))
3257                 goto out;
3258
3259         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3260
3261         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
3262         if (type != BTRFS_EXTENT_DATA_REF_KEY)
3263                 goto out;
3264
3265         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3266         if (btrfs_extent_refs(leaf, ei) !=
3267             btrfs_extent_data_ref_count(leaf, ref) ||
3268             btrfs_extent_data_ref_root(leaf, ref) !=
3269             root->root_key.objectid ||
3270             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3271             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3272                 goto out;
3273
3274         ret = 0;
3275 out:
3276         return ret;
3277 }
3278
3279 int btrfs_cross_ref_exist(struct btrfs_root *root, u64 objectid, u64 offset,
3280                           u64 bytenr)
3281 {
3282         struct btrfs_path *path;
3283         int ret;
3284         int ret2;
3285
3286         path = btrfs_alloc_path();
3287         if (!path)
3288                 return -ENOMEM;
3289
3290         do {
3291                 ret = check_committed_ref(root, path, objectid,
3292                                           offset, bytenr);
3293                 if (ret && ret != -ENOENT)
3294                         goto out;
3295
3296                 ret2 = check_delayed_ref(root, path, objectid,
3297                                          offset, bytenr);
3298         } while (ret2 == -EAGAIN);
3299
3300         if (ret2 && ret2 != -ENOENT) {
3301                 ret = ret2;
3302                 goto out;
3303         }
3304
3305         if (ret != -ENOENT || ret2 != -ENOENT)
3306                 ret = 0;
3307 out:
3308         btrfs_free_path(path);
3309         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3310                 WARN_ON(ret > 0);
3311         return ret;
3312 }
3313
3314 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3315                            struct btrfs_root *root,
3316                            struct extent_buffer *buf,
3317                            int full_backref, int inc)
3318 {
3319         struct btrfs_fs_info *fs_info = root->fs_info;
3320         u64 bytenr;
3321         u64 num_bytes;
3322         u64 parent;
3323         u64 ref_root;
3324         u32 nritems;
3325         struct btrfs_key key;
3326         struct btrfs_file_extent_item *fi;
3327         int i;
3328         int level;
3329         int ret = 0;
3330         int (*process_func)(struct btrfs_trans_handle *,
3331                             struct btrfs_root *,
3332                             u64, u64, u64, u64, u64, u64);
3333
3334
3335         if (btrfs_is_testing(fs_info))
3336                 return 0;
3337
3338         ref_root = btrfs_header_owner(buf);
3339         nritems = btrfs_header_nritems(buf);
3340         level = btrfs_header_level(buf);
3341
3342         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3343                 return 0;
3344
3345         if (inc)
3346                 process_func = btrfs_inc_extent_ref;
3347         else
3348                 process_func = btrfs_free_extent;
3349
3350         if (full_backref)
3351                 parent = buf->start;
3352         else
3353                 parent = 0;
3354
3355         for (i = 0; i < nritems; i++) {
3356                 if (level == 0) {
3357                         btrfs_item_key_to_cpu(buf, &key, i);
3358                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3359                                 continue;
3360                         fi = btrfs_item_ptr(buf, i,
3361                                             struct btrfs_file_extent_item);
3362                         if (btrfs_file_extent_type(buf, fi) ==
3363                             BTRFS_FILE_EXTENT_INLINE)
3364                                 continue;
3365                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3366                         if (bytenr == 0)
3367                                 continue;
3368
3369                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3370                         key.offset -= btrfs_file_extent_offset(buf, fi);
3371                         ret = process_func(trans, root, bytenr, num_bytes,
3372                                            parent, ref_root, key.objectid,
3373                                            key.offset);
3374                         if (ret)
3375                                 goto fail;
3376                 } else {
3377                         bytenr = btrfs_node_blockptr(buf, i);
3378                         num_bytes = fs_info->nodesize;
3379                         ret = process_func(trans, root, bytenr, num_bytes,
3380                                            parent, ref_root, level - 1, 0);
3381                         if (ret)
3382                                 goto fail;
3383                 }
3384         }
3385         return 0;
3386 fail:
3387         return ret;
3388 }
3389
3390 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3391                   struct extent_buffer *buf, int full_backref)
3392 {
3393         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3394 }
3395
3396 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3397                   struct extent_buffer *buf, int full_backref)
3398 {
3399         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3400 }
3401
3402 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3403                                  struct btrfs_fs_info *fs_info,
3404                                  struct btrfs_path *path,
3405                                  struct btrfs_block_group_cache *cache)
3406 {
3407         int ret;
3408         struct btrfs_root *extent_root = fs_info->extent_root;
3409         unsigned long bi;
3410         struct extent_buffer *leaf;
3411
3412         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3413         if (ret) {
3414                 if (ret > 0)
3415                         ret = -ENOENT;
3416                 goto fail;
3417         }
3418
3419         leaf = path->nodes[0];
3420         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3421         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3422         btrfs_mark_buffer_dirty(leaf);
3423 fail:
3424         btrfs_release_path(path);
3425         return ret;
3426
3427 }
3428
3429 static struct btrfs_block_group_cache *
3430 next_block_group(struct btrfs_fs_info *fs_info,
3431                  struct btrfs_block_group_cache *cache)
3432 {
3433         struct rb_node *node;
3434
3435         spin_lock(&fs_info->block_group_cache_lock);
3436
3437         /* If our block group was removed, we need a full search. */
3438         if (RB_EMPTY_NODE(&cache->cache_node)) {
3439                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3440
3441                 spin_unlock(&fs_info->block_group_cache_lock);
3442                 btrfs_put_block_group(cache);
3443                 cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
3444         }
3445         node = rb_next(&cache->cache_node);
3446         btrfs_put_block_group(cache);
3447         if (node) {
3448                 cache = rb_entry(node, struct btrfs_block_group_cache,
3449                                  cache_node);
3450                 btrfs_get_block_group(cache);
3451         } else
3452                 cache = NULL;
3453         spin_unlock(&fs_info->block_group_cache_lock);
3454         return cache;
3455 }
3456
3457 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3458                             struct btrfs_trans_handle *trans,
3459                             struct btrfs_path *path)
3460 {
3461         struct btrfs_fs_info *fs_info = block_group->fs_info;
3462         struct btrfs_root *root = fs_info->tree_root;
3463         struct inode *inode = NULL;
3464         struct extent_changeset *data_reserved = NULL;
3465         u64 alloc_hint = 0;
3466         int dcs = BTRFS_DC_ERROR;
3467         u64 num_pages = 0;
3468         int retries = 0;
3469         int ret = 0;
3470
3471         /*
3472          * If this block group is smaller than 100 megs don't bother caching the
3473          * block group.
3474          */
3475         if (block_group->key.offset < (100 * SZ_1M)) {
3476                 spin_lock(&block_group->lock);
3477                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3478                 spin_unlock(&block_group->lock);
3479                 return 0;
3480         }
3481
3482         if (trans->aborted)
3483                 return 0;
3484 again:
3485         inode = lookup_free_space_inode(fs_info, block_group, path);
3486         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3487                 ret = PTR_ERR(inode);
3488                 btrfs_release_path(path);
3489                 goto out;
3490         }
3491
3492         if (IS_ERR(inode)) {
3493                 BUG_ON(retries);
3494                 retries++;
3495
3496                 if (block_group->ro)
3497                         goto out_free;
3498
3499                 ret = create_free_space_inode(fs_info, trans, block_group,
3500                                               path);
3501                 if (ret)
3502                         goto out_free;
3503                 goto again;
3504         }
3505
3506         /*
3507          * We want to set the generation to 0, that way if anything goes wrong
3508          * from here on out we know not to trust this cache when we load up next
3509          * time.
3510          */
3511         BTRFS_I(inode)->generation = 0;
3512         ret = btrfs_update_inode(trans, root, inode);
3513         if (ret) {
3514                 /*
3515                  * So theoretically we could recover from this, simply set the
3516                  * super cache generation to 0 so we know to invalidate the
3517                  * cache, but then we'd have to keep track of the block groups
3518                  * that fail this way so we know we _have_ to reset this cache
3519                  * before the next commit or risk reading stale cache.  So to
3520                  * limit our exposure to horrible edge cases lets just abort the
3521                  * transaction, this only happens in really bad situations
3522                  * anyway.
3523                  */
3524                 btrfs_abort_transaction(trans, ret);
3525                 goto out_put;
3526         }
3527         WARN_ON(ret);
3528
3529         /* We've already setup this transaction, go ahead and exit */
3530         if (block_group->cache_generation == trans->transid &&
3531             i_size_read(inode)) {
3532                 dcs = BTRFS_DC_SETUP;
3533                 goto out_put;
3534         }
3535
3536         if (i_size_read(inode) > 0) {
3537                 ret = btrfs_check_trunc_cache_free_space(fs_info,
3538                                         &fs_info->global_block_rsv);
3539                 if (ret)
3540                         goto out_put;
3541
3542                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3543                 if (ret)
3544                         goto out_put;
3545         }
3546
3547         spin_lock(&block_group->lock);
3548         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3549             !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3550                 /*
3551                  * don't bother trying to write stuff out _if_
3552                  * a) we're not cached,
3553                  * b) we're with nospace_cache mount option,
3554