082e9f1c1f1a301fd2a0cc4b769aef73b0df854b
[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_fs_info *fs_info,
56                                 struct btrfs_delayed_ref_node *node, u64 parent,
57                                 u64 root_objectid, u64 owner_objectid,
58                                 u64 owner_offset, int refs_to_drop,
59                                 struct btrfs_delayed_extent_op *extra_op);
60 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
61                                     struct extent_buffer *leaf,
62                                     struct btrfs_extent_item *ei);
63 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
64                                       struct btrfs_fs_info *fs_info,
65                                       u64 parent, u64 root_objectid,
66                                       u64 flags, u64 owner, u64 offset,
67                                       struct btrfs_key *ins, int ref_mod);
68 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
69                                      struct btrfs_delayed_ref_node *node,
70                                      struct btrfs_delayed_extent_op *extent_op);
71 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
72                           struct btrfs_fs_info *fs_info, u64 flags,
73                           int force);
74 static int find_next_key(struct btrfs_path *path, int level,
75                          struct btrfs_key *key);
76 static void dump_space_info(struct btrfs_fs_info *fs_info,
77                             struct btrfs_space_info *info, u64 bytes,
78                             int dump_block_groups);
79 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
80                                u64 num_bytes);
81 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
82                                      struct btrfs_space_info *space_info,
83                                      u64 num_bytes);
84 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
85                                      struct btrfs_space_info *space_info,
86                                      u64 num_bytes);
87
88 static noinline int
89 block_group_cache_done(struct btrfs_block_group_cache *cache)
90 {
91         smp_mb();
92         return cache->cached == BTRFS_CACHE_FINISHED ||
93                 cache->cached == BTRFS_CACHE_ERROR;
94 }
95
96 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
97 {
98         return (cache->flags & bits) == bits;
99 }
100
101 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
102 {
103         atomic_inc(&cache->count);
104 }
105
106 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
107 {
108         if (atomic_dec_and_test(&cache->count)) {
109                 WARN_ON(cache->pinned > 0);
110                 WARN_ON(cache->reserved > 0);
111
112                 /*
113                  * If not empty, someone is still holding mutex of
114                  * full_stripe_lock, which can only be released by caller.
115                  * And it will definitely cause use-after-free when caller
116                  * tries to release full stripe lock.
117                  *
118                  * No better way to resolve, but only to warn.
119                  */
120                 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
121                 kfree(cache->free_space_ctl);
122                 kfree(cache);
123         }
124 }
125
126 /*
127  * this adds the block group to the fs_info rb tree for the block group
128  * cache
129  */
130 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
131                                 struct btrfs_block_group_cache *block_group)
132 {
133         struct rb_node **p;
134         struct rb_node *parent = NULL;
135         struct btrfs_block_group_cache *cache;
136
137         spin_lock(&info->block_group_cache_lock);
138         p = &info->block_group_cache_tree.rb_node;
139
140         while (*p) {
141                 parent = *p;
142                 cache = rb_entry(parent, struct btrfs_block_group_cache,
143                                  cache_node);
144                 if (block_group->key.objectid < cache->key.objectid) {
145                         p = &(*p)->rb_left;
146                 } else if (block_group->key.objectid > cache->key.objectid) {
147                         p = &(*p)->rb_right;
148                 } else {
149                         spin_unlock(&info->block_group_cache_lock);
150                         return -EEXIST;
151                 }
152         }
153
154         rb_link_node(&block_group->cache_node, parent, p);
155         rb_insert_color(&block_group->cache_node,
156                         &info->block_group_cache_tree);
157
158         if (info->first_logical_byte > block_group->key.objectid)
159                 info->first_logical_byte = block_group->key.objectid;
160
161         spin_unlock(&info->block_group_cache_lock);
162
163         return 0;
164 }
165
166 /*
167  * This will return the block group at or after bytenr if contains is 0, else
168  * it will return the block group that contains the bytenr
169  */
170 static struct btrfs_block_group_cache *
171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
172                               int contains)
173 {
174         struct btrfs_block_group_cache *cache, *ret = NULL;
175         struct rb_node *n;
176         u64 end, start;
177
178         spin_lock(&info->block_group_cache_lock);
179         n = info->block_group_cache_tree.rb_node;
180
181         while (n) {
182                 cache = rb_entry(n, struct btrfs_block_group_cache,
183                                  cache_node);
184                 end = cache->key.objectid + cache->key.offset - 1;
185                 start = cache->key.objectid;
186
187                 if (bytenr < start) {
188                         if (!contains && (!ret || start < ret->key.objectid))
189                                 ret = cache;
190                         n = n->rb_left;
191                 } else if (bytenr > start) {
192                         if (contains && bytenr <= end) {
193                                 ret = cache;
194                                 break;
195                         }
196                         n = n->rb_right;
197                 } else {
198                         ret = cache;
199                         break;
200                 }
201         }
202         if (ret) {
203                 btrfs_get_block_group(ret);
204                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
205                         info->first_logical_byte = ret->key.objectid;
206         }
207         spin_unlock(&info->block_group_cache_lock);
208
209         return ret;
210 }
211
212 static int add_excluded_extent(struct btrfs_fs_info *fs_info,
213                                u64 start, u64 num_bytes)
214 {
215         u64 end = start + num_bytes - 1;
216         set_extent_bits(&fs_info->freed_extents[0],
217                         start, end, EXTENT_UPTODATE);
218         set_extent_bits(&fs_info->freed_extents[1],
219                         start, end, EXTENT_UPTODATE);
220         return 0;
221 }
222
223 static void free_excluded_extents(struct btrfs_fs_info *fs_info,
224                                   struct btrfs_block_group_cache *cache)
225 {
226         u64 start, end;
227
228         start = cache->key.objectid;
229         end = start + cache->key.offset - 1;
230
231         clear_extent_bits(&fs_info->freed_extents[0],
232                           start, end, EXTENT_UPTODATE);
233         clear_extent_bits(&fs_info->freed_extents[1],
234                           start, end, EXTENT_UPTODATE);
235 }
236
237 static int exclude_super_stripes(struct btrfs_fs_info *fs_info,
238                                  struct btrfs_block_group_cache *cache)
239 {
240         u64 bytenr;
241         u64 *logical;
242         int stripe_len;
243         int i, nr, ret;
244
245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
247                 cache->bytes_super += stripe_len;
248                 ret = add_excluded_extent(fs_info, cache->key.objectid,
249                                           stripe_len);
250                 if (ret)
251                         return ret;
252         }
253
254         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
255                 bytenr = btrfs_sb_offset(i);
256                 ret = btrfs_rmap_block(fs_info, cache->key.objectid,
257                                        bytenr, &logical, &nr, &stripe_len);
258                 if (ret)
259                         return ret;
260
261                 while (nr--) {
262                         u64 start, len;
263
264                         if (logical[nr] > cache->key.objectid +
265                             cache->key.offset)
266                                 continue;
267
268                         if (logical[nr] + stripe_len <= cache->key.objectid)
269                                 continue;
270
271                         start = logical[nr];
272                         if (start < cache->key.objectid) {
273                                 start = cache->key.objectid;
274                                 len = (logical[nr] + stripe_len) - start;
275                         } else {
276                                 len = min_t(u64, stripe_len,
277                                             cache->key.objectid +
278                                             cache->key.offset - start);
279                         }
280
281                         cache->bytes_super += len;
282                         ret = add_excluded_extent(fs_info, start, len);
283                         if (ret) {
284                                 kfree(logical);
285                                 return ret;
286                         }
287                 }
288
289                 kfree(logical);
290         }
291         return 0;
292 }
293
294 static struct btrfs_caching_control *
295 get_caching_control(struct btrfs_block_group_cache *cache)
296 {
297         struct btrfs_caching_control *ctl;
298
299         spin_lock(&cache->lock);
300         if (!cache->caching_ctl) {
301                 spin_unlock(&cache->lock);
302                 return NULL;
303         }
304
305         ctl = cache->caching_ctl;
306         refcount_inc(&ctl->count);
307         spin_unlock(&cache->lock);
308         return ctl;
309 }
310
311 static void put_caching_control(struct btrfs_caching_control *ctl)
312 {
313         if (refcount_dec_and_test(&ctl->count))
314                 kfree(ctl);
315 }
316
317 #ifdef CONFIG_BTRFS_DEBUG
318 static void fragment_free_space(struct btrfs_block_group_cache *block_group)
319 {
320         struct btrfs_fs_info *fs_info = block_group->fs_info;
321         u64 start = block_group->key.objectid;
322         u64 len = block_group->key.offset;
323         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
324                 fs_info->nodesize : fs_info->sectorsize;
325         u64 step = chunk << 1;
326
327         while (len > chunk) {
328                 btrfs_remove_free_space(block_group, start, chunk);
329                 start += step;
330                 if (len < step)
331                         len = 0;
332                 else
333                         len -= step;
334         }
335 }
336 #endif
337
338 /*
339  * this is only called by cache_block_group, since we could have freed extents
340  * we need to check the pinned_extents for any extents that can't be used yet
341  * since their free space will be released as soon as the transaction commits.
342  */
343 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
344                        u64 start, u64 end)
345 {
346         struct btrfs_fs_info *info = block_group->fs_info;
347         u64 extent_start, extent_end, size, total_added = 0;
348         int ret;
349
350         while (start < end) {
351                 ret = find_first_extent_bit(info->pinned_extents, start,
352                                             &extent_start, &extent_end,
353                                             EXTENT_DIRTY | EXTENT_UPTODATE,
354                                             NULL);
355                 if (ret)
356                         break;
357
358                 if (extent_start <= start) {
359                         start = extent_end + 1;
360                 } else if (extent_start > start && extent_start < end) {
361                         size = extent_start - start;
362                         total_added += size;
363                         ret = btrfs_add_free_space(block_group, start,
364                                                    size);
365                         BUG_ON(ret); /* -ENOMEM or logic error */
366                         start = extent_end + 1;
367                 } else {
368                         break;
369                 }
370         }
371
372         if (start < end) {
373                 size = end - start;
374                 total_added += size;
375                 ret = btrfs_add_free_space(block_group, start, size);
376                 BUG_ON(ret); /* -ENOMEM or logic error */
377         }
378
379         return total_added;
380 }
381
382 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
383 {
384         struct btrfs_block_group_cache *block_group = caching_ctl->block_group;
385         struct btrfs_fs_info *fs_info = block_group->fs_info;
386         struct btrfs_root *extent_root = fs_info->extent_root;
387         struct btrfs_path *path;
388         struct extent_buffer *leaf;
389         struct btrfs_key key;
390         u64 total_found = 0;
391         u64 last = 0;
392         u32 nritems;
393         int ret;
394         bool wakeup = true;
395
396         path = btrfs_alloc_path();
397         if (!path)
398                 return -ENOMEM;
399
400         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
401
402 #ifdef CONFIG_BTRFS_DEBUG
403         /*
404          * If we're fragmenting we don't want to make anybody think we can
405          * allocate from this block group until we've had a chance to fragment
406          * the free space.
407          */
408         if (btrfs_should_fragment_free_space(block_group))
409                 wakeup = false;
410 #endif
411         /*
412          * We don't want to deadlock with somebody trying to allocate a new
413          * extent for the extent root while also trying to search the extent
414          * root to add free space.  So we skip locking and search the commit
415          * root, since its read-only
416          */
417         path->skip_locking = 1;
418         path->search_commit_root = 1;
419         path->reada = READA_FORWARD;
420
421         key.objectid = last;
422         key.offset = 0;
423         key.type = BTRFS_EXTENT_ITEM_KEY;
424
425 next:
426         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
427         if (ret < 0)
428                 goto out;
429
430         leaf = path->nodes[0];
431         nritems = btrfs_header_nritems(leaf);
432
433         while (1) {
434                 if (btrfs_fs_closing(fs_info) > 1) {
435                         last = (u64)-1;
436                         break;
437                 }
438
439                 if (path->slots[0] < nritems) {
440                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
441                 } else {
442                         ret = find_next_key(path, 0, &key);
443                         if (ret)
444                                 break;
445
446                         if (need_resched() ||
447                             rwsem_is_contended(&fs_info->commit_root_sem)) {
448                                 if (wakeup)
449                                         caching_ctl->progress = last;
450                                 btrfs_release_path(path);
451                                 up_read(&fs_info->commit_root_sem);
452                                 mutex_unlock(&caching_ctl->mutex);
453                                 cond_resched();
454                                 mutex_lock(&caching_ctl->mutex);
455                                 down_read(&fs_info->commit_root_sem);
456                                 goto next;
457                         }
458
459                         ret = btrfs_next_leaf(extent_root, path);
460                         if (ret < 0)
461                                 goto out;
462                         if (ret)
463                                 break;
464                         leaf = path->nodes[0];
465                         nritems = btrfs_header_nritems(leaf);
466                         continue;
467                 }
468
469                 if (key.objectid < last) {
470                         key.objectid = last;
471                         key.offset = 0;
472                         key.type = BTRFS_EXTENT_ITEM_KEY;
473
474                         if (wakeup)
475                                 caching_ctl->progress = last;
476                         btrfs_release_path(path);
477                         goto next;
478                 }
479
480                 if (key.objectid < block_group->key.objectid) {
481                         path->slots[0]++;
482                         continue;
483                 }
484
485                 if (key.objectid >= block_group->key.objectid +
486                     block_group->key.offset)
487                         break;
488
489                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
490                     key.type == BTRFS_METADATA_ITEM_KEY) {
491                         total_found += add_new_free_space(block_group, last,
492                                                           key.objectid);
493                         if (key.type == BTRFS_METADATA_ITEM_KEY)
494                                 last = key.objectid +
495                                         fs_info->nodesize;
496                         else
497                                 last = key.objectid + key.offset;
498
499                         if (total_found > CACHING_CTL_WAKE_UP) {
500                                 total_found = 0;
501                                 if (wakeup)
502                                         wake_up(&caching_ctl->wait);
503                         }
504                 }
505                 path->slots[0]++;
506         }
507         ret = 0;
508
509         total_found += add_new_free_space(block_group, last,
510                                           block_group->key.objectid +
511                                           block_group->key.offset);
512         caching_ctl->progress = (u64)-1;
513
514 out:
515         btrfs_free_path(path);
516         return ret;
517 }
518
519 static noinline void caching_thread(struct btrfs_work *work)
520 {
521         struct btrfs_block_group_cache *block_group;
522         struct btrfs_fs_info *fs_info;
523         struct btrfs_caching_control *caching_ctl;
524         int ret;
525
526         caching_ctl = container_of(work, struct btrfs_caching_control, work);
527         block_group = caching_ctl->block_group;
528         fs_info = block_group->fs_info;
529
530         mutex_lock(&caching_ctl->mutex);
531         down_read(&fs_info->commit_root_sem);
532
533         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
534                 ret = load_free_space_tree(caching_ctl);
535         else
536                 ret = load_extent_tree_free(caching_ctl);
537
538         spin_lock(&block_group->lock);
539         block_group->caching_ctl = NULL;
540         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
541         spin_unlock(&block_group->lock);
542
543 #ifdef CONFIG_BTRFS_DEBUG
544         if (btrfs_should_fragment_free_space(block_group)) {
545                 u64 bytes_used;
546
547                 spin_lock(&block_group->space_info->lock);
548                 spin_lock(&block_group->lock);
549                 bytes_used = block_group->key.offset -
550                         btrfs_block_group_used(&block_group->item);
551                 block_group->space_info->bytes_used += bytes_used >> 1;
552                 spin_unlock(&block_group->lock);
553                 spin_unlock(&block_group->space_info->lock);
554                 fragment_free_space(block_group);
555         }
556 #endif
557
558         caching_ctl->progress = (u64)-1;
559
560         up_read(&fs_info->commit_root_sem);
561         free_excluded_extents(fs_info, block_group);
562         mutex_unlock(&caching_ctl->mutex);
563
564         wake_up(&caching_ctl->wait);
565
566         put_caching_control(caching_ctl);
567         btrfs_put_block_group(block_group);
568 }
569
570 static int cache_block_group(struct btrfs_block_group_cache *cache,
571                              int load_cache_only)
572 {
573         DEFINE_WAIT(wait);
574         struct btrfs_fs_info *fs_info = cache->fs_info;
575         struct btrfs_caching_control *caching_ctl;
576         int ret = 0;
577
578         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
579         if (!caching_ctl)
580                 return -ENOMEM;
581
582         INIT_LIST_HEAD(&caching_ctl->list);
583         mutex_init(&caching_ctl->mutex);
584         init_waitqueue_head(&caching_ctl->wait);
585         caching_ctl->block_group = cache;
586         caching_ctl->progress = cache->key.objectid;
587         refcount_set(&caching_ctl->count, 1);
588         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
589                         caching_thread, NULL, NULL);
590
591         spin_lock(&cache->lock);
592         /*
593          * This should be a rare occasion, but this could happen I think in the
594          * case where one thread starts to load the space cache info, and then
595          * some other thread starts a transaction commit which tries to do an
596          * allocation while the other thread is still loading the space cache
597          * info.  The previous loop should have kept us from choosing this block
598          * group, but if we've moved to the state where we will wait on caching
599          * block groups we need to first check if we're doing a fast load here,
600          * so we can wait for it to finish, otherwise we could end up allocating
601          * from a block group who's cache gets evicted for one reason or
602          * another.
603          */
604         while (cache->cached == BTRFS_CACHE_FAST) {
605                 struct btrfs_caching_control *ctl;
606
607                 ctl = cache->caching_ctl;
608                 refcount_inc(&ctl->count);
609                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
610                 spin_unlock(&cache->lock);
611
612                 schedule();
613
614                 finish_wait(&ctl->wait, &wait);
615                 put_caching_control(ctl);
616                 spin_lock(&cache->lock);
617         }
618
619         if (cache->cached != BTRFS_CACHE_NO) {
620                 spin_unlock(&cache->lock);
621                 kfree(caching_ctl);
622                 return 0;
623         }
624         WARN_ON(cache->caching_ctl);
625         cache->caching_ctl = caching_ctl;
626         cache->cached = BTRFS_CACHE_FAST;
627         spin_unlock(&cache->lock);
628
629         if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
630                 mutex_lock(&caching_ctl->mutex);
631                 ret = load_free_space_cache(fs_info, cache);
632
633                 spin_lock(&cache->lock);
634                 if (ret == 1) {
635                         cache->caching_ctl = NULL;
636                         cache->cached = BTRFS_CACHE_FINISHED;
637                         cache->last_byte_to_unpin = (u64)-1;
638                         caching_ctl->progress = (u64)-1;
639                 } else {
640                         if (load_cache_only) {
641                                 cache->caching_ctl = NULL;
642                                 cache->cached = BTRFS_CACHE_NO;
643                         } else {
644                                 cache->cached = BTRFS_CACHE_STARTED;
645                                 cache->has_caching_ctl = 1;
646                         }
647                 }
648                 spin_unlock(&cache->lock);
649 #ifdef CONFIG_BTRFS_DEBUG
650                 if (ret == 1 &&
651                     btrfs_should_fragment_free_space(cache)) {
652                         u64 bytes_used;
653
654                         spin_lock(&cache->space_info->lock);
655                         spin_lock(&cache->lock);
656                         bytes_used = cache->key.offset -
657                                 btrfs_block_group_used(&cache->item);
658                         cache->space_info->bytes_used += bytes_used >> 1;
659                         spin_unlock(&cache->lock);
660                         spin_unlock(&cache->space_info->lock);
661                         fragment_free_space(cache);
662                 }
663 #endif
664                 mutex_unlock(&caching_ctl->mutex);
665
666                 wake_up(&caching_ctl->wait);
667                 if (ret == 1) {
668                         put_caching_control(caching_ctl);
669                         free_excluded_extents(fs_info, cache);
670                         return 0;
671                 }
672         } else {
673                 /*
674                  * We're either using the free space tree or no caching at all.
675                  * Set cached to the appropriate value and wakeup any waiters.
676                  */
677                 spin_lock(&cache->lock);
678                 if (load_cache_only) {
679                         cache->caching_ctl = NULL;
680                         cache->cached = BTRFS_CACHE_NO;
681                 } else {
682                         cache->cached = BTRFS_CACHE_STARTED;
683                         cache->has_caching_ctl = 1;
684                 }
685                 spin_unlock(&cache->lock);
686                 wake_up(&caching_ctl->wait);
687         }
688
689         if (load_cache_only) {
690                 put_caching_control(caching_ctl);
691                 return 0;
692         }
693
694         down_write(&fs_info->commit_root_sem);
695         refcount_inc(&caching_ctl->count);
696         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
697         up_write(&fs_info->commit_root_sem);
698
699         btrfs_get_block_group(cache);
700
701         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
702
703         return ret;
704 }
705
706 /*
707  * return the block group that starts at or after bytenr
708  */
709 static struct btrfs_block_group_cache *
710 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
711 {
712         return block_group_cache_tree_search(info, bytenr, 0);
713 }
714
715 /*
716  * return the block group that contains the given bytenr
717  */
718 struct btrfs_block_group_cache *btrfs_lookup_block_group(
719                                                  struct btrfs_fs_info *info,
720                                                  u64 bytenr)
721 {
722         return block_group_cache_tree_search(info, bytenr, 1);
723 }
724
725 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
726                                                   u64 flags)
727 {
728         struct list_head *head = &info->space_info;
729         struct btrfs_space_info *found;
730
731         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
732
733         rcu_read_lock();
734         list_for_each_entry_rcu(found, head, list) {
735                 if (found->flags & flags) {
736                         rcu_read_unlock();
737                         return found;
738                 }
739         }
740         rcu_read_unlock();
741         return NULL;
742 }
743
744 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, s64 num_bytes,
745                              bool metadata, u64 root_objectid)
746 {
747         struct btrfs_space_info *space_info;
748         u64 flags;
749
750         if (metadata) {
751                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
752                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
753                 else
754                         flags = BTRFS_BLOCK_GROUP_METADATA;
755         } else {
756                 flags = BTRFS_BLOCK_GROUP_DATA;
757         }
758
759         space_info = __find_space_info(fs_info, flags);
760         ASSERT(space_info);
761         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
762 }
763
764 /*
765  * after adding space to the filesystem, we need to clear the full flags
766  * on all the space infos.
767  */
768 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
769 {
770         struct list_head *head = &info->space_info;
771         struct btrfs_space_info *found;
772
773         rcu_read_lock();
774         list_for_each_entry_rcu(found, head, list)
775                 found->full = 0;
776         rcu_read_unlock();
777 }
778
779 /* simple helper to search for an existing data extent at a given offset */
780 int btrfs_lookup_data_extent(struct btrfs_fs_info *fs_info, u64 start, u64 len)
781 {
782         int ret;
783         struct btrfs_key key;
784         struct btrfs_path *path;
785
786         path = btrfs_alloc_path();
787         if (!path)
788                 return -ENOMEM;
789
790         key.objectid = start;
791         key.offset = len;
792         key.type = BTRFS_EXTENT_ITEM_KEY;
793         ret = btrfs_search_slot(NULL, fs_info->extent_root, &key, path, 0, 0);
794         btrfs_free_path(path);
795         return ret;
796 }
797
798 /*
799  * helper function to lookup reference count and flags of a tree block.
800  *
801  * the head node for delayed ref is used to store the sum of all the
802  * reference count modifications queued up in the rbtree. the head
803  * node may also store the extent flags to set. This way you can check
804  * to see what the reference count and extent flags would be if all of
805  * the delayed refs are not processed.
806  */
807 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
808                              struct btrfs_fs_info *fs_info, u64 bytenr,
809                              u64 offset, int metadata, u64 *refs, u64 *flags)
810 {
811         struct btrfs_delayed_ref_head *head;
812         struct btrfs_delayed_ref_root *delayed_refs;
813         struct btrfs_path *path;
814         struct btrfs_extent_item *ei;
815         struct extent_buffer *leaf;
816         struct btrfs_key key;
817         u32 item_size;
818         u64 num_refs;
819         u64 extent_flags;
820         int ret;
821
822         /*
823          * If we don't have skinny metadata, don't bother doing anything
824          * different
825          */
826         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA)) {
827                 offset = fs_info->nodesize;
828                 metadata = 0;
829         }
830
831         path = btrfs_alloc_path();
832         if (!path)
833                 return -ENOMEM;
834
835         if (!trans) {
836                 path->skip_locking = 1;
837                 path->search_commit_root = 1;
838         }
839
840 search_again:
841         key.objectid = bytenr;
842         key.offset = offset;
843         if (metadata)
844                 key.type = BTRFS_METADATA_ITEM_KEY;
845         else
846                 key.type = BTRFS_EXTENT_ITEM_KEY;
847
848         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 0);
849         if (ret < 0)
850                 goto out_free;
851
852         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
853                 if (path->slots[0]) {
854                         path->slots[0]--;
855                         btrfs_item_key_to_cpu(path->nodes[0], &key,
856                                               path->slots[0]);
857                         if (key.objectid == bytenr &&
858                             key.type == BTRFS_EXTENT_ITEM_KEY &&
859                             key.offset == fs_info->nodesize)
860                                 ret = 0;
861                 }
862         }
863
864         if (ret == 0) {
865                 leaf = path->nodes[0];
866                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
867                 if (item_size >= sizeof(*ei)) {
868                         ei = btrfs_item_ptr(leaf, path->slots[0],
869                                             struct btrfs_extent_item);
870                         num_refs = btrfs_extent_refs(leaf, ei);
871                         extent_flags = btrfs_extent_flags(leaf, ei);
872                 } else {
873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
874                         struct btrfs_extent_item_v0 *ei0;
875                         BUG_ON(item_size != sizeof(*ei0));
876                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
877                                              struct btrfs_extent_item_v0);
878                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
879                         /* FIXME: this isn't correct for data */
880                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
881 #else
882                         BUG();
883 #endif
884                 }
885                 BUG_ON(num_refs == 0);
886         } else {
887                 num_refs = 0;
888                 extent_flags = 0;
889                 ret = 0;
890         }
891
892         if (!trans)
893                 goto out;
894
895         delayed_refs = &trans->transaction->delayed_refs;
896         spin_lock(&delayed_refs->lock);
897         head = btrfs_find_delayed_ref_head(delayed_refs, bytenr);
898         if (head) {
899                 if (!mutex_trylock(&head->mutex)) {
900                         refcount_inc(&head->refs);
901                         spin_unlock(&delayed_refs->lock);
902
903                         btrfs_release_path(path);
904
905                         /*
906                          * Mutex was contended, block until it's released and try
907                          * again
908                          */
909                         mutex_lock(&head->mutex);
910                         mutex_unlock(&head->mutex);
911                         btrfs_put_delayed_ref_head(head);
912                         goto search_again;
913                 }
914                 spin_lock(&head->lock);
915                 if (head->extent_op && head->extent_op->update_flags)
916                         extent_flags |= head->extent_op->flags_to_set;
917                 else
918                         BUG_ON(num_refs == 0);
919
920                 num_refs += head->ref_mod;
921                 spin_unlock(&head->lock);
922                 mutex_unlock(&head->mutex);
923         }
924         spin_unlock(&delayed_refs->lock);
925 out:
926         WARN_ON(num_refs == 0);
927         if (refs)
928                 *refs = num_refs;
929         if (flags)
930                 *flags = extent_flags;
931 out_free:
932         btrfs_free_path(path);
933         return ret;
934 }
935
936 /*
937  * Back reference rules.  Back refs have three main goals:
938  *
939  * 1) differentiate between all holders of references to an extent so that
940  *    when a reference is dropped we can make sure it was a valid reference
941  *    before freeing the extent.
942  *
943  * 2) Provide enough information to quickly find the holders of an extent
944  *    if we notice a given block is corrupted or bad.
945  *
946  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
947  *    maintenance.  This is actually the same as #2, but with a slightly
948  *    different use case.
949  *
950  * There are two kinds of back refs. The implicit back refs is optimized
951  * for pointers in non-shared tree blocks. For a given pointer in a block,
952  * back refs of this kind provide information about the block's owner tree
953  * and the pointer's key. These information allow us to find the block by
954  * b-tree searching. The full back refs is for pointers in tree blocks not
955  * referenced by their owner trees. The location of tree block is recorded
956  * in the back refs. Actually the full back refs is generic, and can be
957  * used in all cases the implicit back refs is used. The major shortcoming
958  * of the full back refs is its overhead. Every time a tree block gets
959  * COWed, we have to update back refs entry for all pointers in it.
960  *
961  * For a newly allocated tree block, we use implicit back refs for
962  * pointers in it. This means most tree related operations only involve
963  * implicit back refs. For a tree block created in old transaction, the
964  * only way to drop a reference to it is COW it. So we can detect the
965  * event that tree block loses its owner tree's reference and do the
966  * back refs conversion.
967  *
968  * When a tree block is COWed through a tree, there are four cases:
969  *
970  * The reference count of the block is one and the tree is the block's
971  * owner tree. Nothing to do in this case.
972  *
973  * The reference count of the block is one and the tree is not the
974  * block's owner tree. In this case, full back refs is used for pointers
975  * in the block. Remove these full back refs, add implicit back refs for
976  * every pointers in the new block.
977  *
978  * The reference count of the block is greater than one and the tree is
979  * the block's owner tree. In this case, implicit back refs is used for
980  * pointers in the block. Add full back refs for every pointers in the
981  * block, increase lower level extents' reference counts. The original
982  * implicit back refs are entailed to the new block.
983  *
984  * The reference count of the block is greater than one and the tree is
985  * not the block's owner tree. Add implicit back refs for every pointer in
986  * the new block, increase lower level extents' reference count.
987  *
988  * Back Reference Key composing:
989  *
990  * The key objectid corresponds to the first byte in the extent,
991  * The key type is used to differentiate between types of back refs.
992  * There are different meanings of the key offset for different types
993  * of back refs.
994  *
995  * File extents can be referenced by:
996  *
997  * - multiple snapshots, subvolumes, or different generations in one subvol
998  * - different files inside a single subvolume
999  * - different offsets inside a file (bookend extents in file.c)
1000  *
1001  * The extent ref structure for the implicit back refs has fields for:
1002  *
1003  * - Objectid of the subvolume root
1004  * - objectid of the file holding the reference
1005  * - original offset in the file
1006  * - how many bookend extents
1007  *
1008  * The key offset for the implicit back refs is hash of the first
1009  * three fields.
1010  *
1011  * The extent ref structure for the full back refs has field for:
1012  *
1013  * - number of pointers in the tree leaf
1014  *
1015  * The key offset for the implicit back refs is the first byte of
1016  * the tree leaf
1017  *
1018  * When a file extent is allocated, The implicit back refs is used.
1019  * the fields are filled in:
1020  *
1021  *     (root_key.objectid, inode objectid, offset in file, 1)
1022  *
1023  * When a file extent is removed file truncation, we find the
1024  * corresponding implicit back refs and check the following fields:
1025  *
1026  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1027  *
1028  * Btree extents can be referenced by:
1029  *
1030  * - Different subvolumes
1031  *
1032  * Both the implicit back refs and the full back refs for tree blocks
1033  * only consist of key. The key offset for the implicit back refs is
1034  * objectid of block's owner tree. The key offset for the full back refs
1035  * is the first byte of parent block.
1036  *
1037  * When implicit back refs is used, information about the lowest key and
1038  * level of the tree block are required. These information are stored in
1039  * tree block info structure.
1040  */
1041
1042 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1043 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1044                                   struct btrfs_fs_info *fs_info,
1045                                   struct btrfs_path *path,
1046                                   u64 owner, u32 extra_size)
1047 {
1048         struct btrfs_root *root = fs_info->extent_root;
1049         struct btrfs_extent_item *item;
1050         struct btrfs_extent_item_v0 *ei0;
1051         struct btrfs_extent_ref_v0 *ref0;
1052         struct btrfs_tree_block_info *bi;
1053         struct extent_buffer *leaf;
1054         struct btrfs_key key;
1055         struct btrfs_key found_key;
1056         u32 new_size = sizeof(*item);
1057         u64 refs;
1058         int ret;
1059
1060         leaf = path->nodes[0];
1061         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1062
1063         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1064         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1065                              struct btrfs_extent_item_v0);
1066         refs = btrfs_extent_refs_v0(leaf, ei0);
1067
1068         if (owner == (u64)-1) {
1069                 while (1) {
1070                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1071                                 ret = btrfs_next_leaf(root, path);
1072                                 if (ret < 0)
1073                                         return ret;
1074                                 BUG_ON(ret > 0); /* Corruption */
1075                                 leaf = path->nodes[0];
1076                         }
1077                         btrfs_item_key_to_cpu(leaf, &found_key,
1078                                               path->slots[0]);
1079                         BUG_ON(key.objectid != found_key.objectid);
1080                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1081                                 path->slots[0]++;
1082                                 continue;
1083                         }
1084                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1085                                               struct btrfs_extent_ref_v0);
1086                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1087                         break;
1088                 }
1089         }
1090         btrfs_release_path(path);
1091
1092         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1093                 new_size += sizeof(*bi);
1094
1095         new_size -= sizeof(*ei0);
1096         ret = btrfs_search_slot(trans, root, &key, path,
1097                                 new_size + extra_size, 1);
1098         if (ret < 0)
1099                 return ret;
1100         BUG_ON(ret); /* Corruption */
1101
1102         btrfs_extend_item(fs_info, path, new_size);
1103
1104         leaf = path->nodes[0];
1105         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1106         btrfs_set_extent_refs(leaf, item, refs);
1107         /* FIXME: get real generation */
1108         btrfs_set_extent_generation(leaf, item, 0);
1109         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1110                 btrfs_set_extent_flags(leaf, item,
1111                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1112                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1113                 bi = (struct btrfs_tree_block_info *)(item + 1);
1114                 /* FIXME: get first key of the block */
1115                 memzero_extent_buffer(leaf, (unsigned long)bi, sizeof(*bi));
1116                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1117         } else {
1118                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1119         }
1120         btrfs_mark_buffer_dirty(leaf);
1121         return 0;
1122 }
1123 #endif
1124
1125 /*
1126  * is_data == BTRFS_REF_TYPE_BLOCK, tree block type is required,
1127  * is_data == BTRFS_REF_TYPE_DATA, data type is requried,
1128  * is_data == BTRFS_REF_TYPE_ANY, either type is OK.
1129  */
1130 int btrfs_get_extent_inline_ref_type(const struct extent_buffer *eb,
1131                                      struct btrfs_extent_inline_ref *iref,
1132                                      enum btrfs_inline_ref_type is_data)
1133 {
1134         int type = btrfs_extent_inline_ref_type(eb, iref);
1135         u64 offset = btrfs_extent_inline_ref_offset(eb, iref);
1136
1137         if (type == BTRFS_TREE_BLOCK_REF_KEY ||
1138             type == BTRFS_SHARED_BLOCK_REF_KEY ||
1139             type == BTRFS_SHARED_DATA_REF_KEY ||
1140             type == BTRFS_EXTENT_DATA_REF_KEY) {
1141                 if (is_data == BTRFS_REF_TYPE_BLOCK) {
1142                         if (type == BTRFS_TREE_BLOCK_REF_KEY)
1143                                 return type;
1144                         if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1145                                 ASSERT(eb->fs_info);
1146                                 /*
1147                                  * Every shared one has parent tree
1148                                  * block, which must be aligned to
1149                                  * nodesize.
1150                                  */
1151                                 if (offset &&
1152                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1153                                         return type;
1154                         }
1155                 } else if (is_data == BTRFS_REF_TYPE_DATA) {
1156                         if (type == BTRFS_EXTENT_DATA_REF_KEY)
1157                                 return type;
1158                         if (type == BTRFS_SHARED_DATA_REF_KEY) {
1159                                 ASSERT(eb->fs_info);
1160                                 /*
1161                                  * Every shared one has parent tree
1162                                  * block, which must be aligned to
1163                                  * nodesize.
1164                                  */
1165                                 if (offset &&
1166                                     IS_ALIGNED(offset, eb->fs_info->nodesize))
1167                                         return type;
1168                         }
1169                 } else {
1170                         ASSERT(is_data == BTRFS_REF_TYPE_ANY);
1171                         return type;
1172                 }
1173         }
1174
1175         btrfs_print_leaf((struct extent_buffer *)eb);
1176         btrfs_err(eb->fs_info, "eb %llu invalid extent inline ref type %d",
1177                   eb->start, type);
1178         WARN_ON(1);
1179
1180         return BTRFS_REF_TYPE_INVALID;
1181 }
1182
1183 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1184 {
1185         u32 high_crc = ~(u32)0;
1186         u32 low_crc = ~(u32)0;
1187         __le64 lenum;
1188
1189         lenum = cpu_to_le64(root_objectid);
1190         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1191         lenum = cpu_to_le64(owner);
1192         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1193         lenum = cpu_to_le64(offset);
1194         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1195
1196         return ((u64)high_crc << 31) ^ (u64)low_crc;
1197 }
1198
1199 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1200                                      struct btrfs_extent_data_ref *ref)
1201 {
1202         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1203                                     btrfs_extent_data_ref_objectid(leaf, ref),
1204                                     btrfs_extent_data_ref_offset(leaf, ref));
1205 }
1206
1207 static int match_extent_data_ref(struct extent_buffer *leaf,
1208                                  struct btrfs_extent_data_ref *ref,
1209                                  u64 root_objectid, u64 owner, u64 offset)
1210 {
1211         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1212             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1213             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1214                 return 0;
1215         return 1;
1216 }
1217
1218 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1219                                            struct btrfs_fs_info *fs_info,
1220                                            struct btrfs_path *path,
1221                                            u64 bytenr, u64 parent,
1222                                            u64 root_objectid,
1223                                            u64 owner, u64 offset)
1224 {
1225         struct btrfs_root *root = fs_info->extent_root;
1226         struct btrfs_key key;
1227         struct btrfs_extent_data_ref *ref;
1228         struct extent_buffer *leaf;
1229         u32 nritems;
1230         int ret;
1231         int recow;
1232         int err = -ENOENT;
1233
1234         key.objectid = bytenr;
1235         if (parent) {
1236                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1237                 key.offset = parent;
1238         } else {
1239                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1240                 key.offset = hash_extent_data_ref(root_objectid,
1241                                                   owner, offset);
1242         }
1243 again:
1244         recow = 0;
1245         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1246         if (ret < 0) {
1247                 err = ret;
1248                 goto fail;
1249         }
1250
1251         if (parent) {
1252                 if (!ret)
1253                         return 0;
1254 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1255                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1256                 btrfs_release_path(path);
1257                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1258                 if (ret < 0) {
1259                         err = ret;
1260                         goto fail;
1261                 }
1262                 if (!ret)
1263                         return 0;
1264 #endif
1265                 goto fail;
1266         }
1267
1268         leaf = path->nodes[0];
1269         nritems = btrfs_header_nritems(leaf);
1270         while (1) {
1271                 if (path->slots[0] >= nritems) {
1272                         ret = btrfs_next_leaf(root, path);
1273                         if (ret < 0)
1274                                 err = ret;
1275                         if (ret)
1276                                 goto fail;
1277
1278                         leaf = path->nodes[0];
1279                         nritems = btrfs_header_nritems(leaf);
1280                         recow = 1;
1281                 }
1282
1283                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1284                 if (key.objectid != bytenr ||
1285                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1286                         goto fail;
1287
1288                 ref = btrfs_item_ptr(leaf, path->slots[0],
1289                                      struct btrfs_extent_data_ref);
1290
1291                 if (match_extent_data_ref(leaf, ref, root_objectid,
1292                                           owner, offset)) {
1293                         if (recow) {
1294                                 btrfs_release_path(path);
1295                                 goto again;
1296                         }
1297                         err = 0;
1298                         break;
1299                 }
1300                 path->slots[0]++;
1301         }
1302 fail:
1303         return err;
1304 }
1305
1306 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1307                                            struct btrfs_fs_info *fs_info,
1308                                            struct btrfs_path *path,
1309                                            u64 bytenr, u64 parent,
1310                                            u64 root_objectid, u64 owner,
1311                                            u64 offset, int refs_to_add)
1312 {
1313         struct btrfs_root *root = fs_info->extent_root;
1314         struct btrfs_key key;
1315         struct extent_buffer *leaf;
1316         u32 size;
1317         u32 num_refs;
1318         int ret;
1319
1320         key.objectid = bytenr;
1321         if (parent) {
1322                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1323                 key.offset = parent;
1324                 size = sizeof(struct btrfs_shared_data_ref);
1325         } else {
1326                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1327                 key.offset = hash_extent_data_ref(root_objectid,
1328                                                   owner, offset);
1329                 size = sizeof(struct btrfs_extent_data_ref);
1330         }
1331
1332         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1333         if (ret && ret != -EEXIST)
1334                 goto fail;
1335
1336         leaf = path->nodes[0];
1337         if (parent) {
1338                 struct btrfs_shared_data_ref *ref;
1339                 ref = btrfs_item_ptr(leaf, path->slots[0],
1340                                      struct btrfs_shared_data_ref);
1341                 if (ret == 0) {
1342                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1343                 } else {
1344                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1345                         num_refs += refs_to_add;
1346                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1347                 }
1348         } else {
1349                 struct btrfs_extent_data_ref *ref;
1350                 while (ret == -EEXIST) {
1351                         ref = btrfs_item_ptr(leaf, path->slots[0],
1352                                              struct btrfs_extent_data_ref);
1353                         if (match_extent_data_ref(leaf, ref, root_objectid,
1354                                                   owner, offset))
1355                                 break;
1356                         btrfs_release_path(path);
1357                         key.offset++;
1358                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1359                                                       size);
1360                         if (ret && ret != -EEXIST)
1361                                 goto fail;
1362
1363                         leaf = path->nodes[0];
1364                 }
1365                 ref = btrfs_item_ptr(leaf, path->slots[0],
1366                                      struct btrfs_extent_data_ref);
1367                 if (ret == 0) {
1368                         btrfs_set_extent_data_ref_root(leaf, ref,
1369                                                        root_objectid);
1370                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1371                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1372                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1373                 } else {
1374                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1375                         num_refs += refs_to_add;
1376                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1377                 }
1378         }
1379         btrfs_mark_buffer_dirty(leaf);
1380         ret = 0;
1381 fail:
1382         btrfs_release_path(path);
1383         return ret;
1384 }
1385
1386 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1387                                            struct btrfs_fs_info *fs_info,
1388                                            struct btrfs_path *path,
1389                                            int refs_to_drop, int *last_ref)
1390 {
1391         struct btrfs_key key;
1392         struct btrfs_extent_data_ref *ref1 = NULL;
1393         struct btrfs_shared_data_ref *ref2 = NULL;
1394         struct extent_buffer *leaf;
1395         u32 num_refs = 0;
1396         int ret = 0;
1397
1398         leaf = path->nodes[0];
1399         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1400
1401         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1402                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1403                                       struct btrfs_extent_data_ref);
1404                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1405         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1406                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1407                                       struct btrfs_shared_data_ref);
1408                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1409 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1410         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1411                 struct btrfs_extent_ref_v0 *ref0;
1412                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1413                                       struct btrfs_extent_ref_v0);
1414                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1415 #endif
1416         } else {
1417                 BUG();
1418         }
1419
1420         BUG_ON(num_refs < refs_to_drop);
1421         num_refs -= refs_to_drop;
1422
1423         if (num_refs == 0) {
1424                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
1425                 *last_ref = 1;
1426         } else {
1427                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1428                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1429                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1430                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1431 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1432                 else {
1433                         struct btrfs_extent_ref_v0 *ref0;
1434                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1435                                         struct btrfs_extent_ref_v0);
1436                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1437                 }
1438 #endif
1439                 btrfs_mark_buffer_dirty(leaf);
1440         }
1441         return ret;
1442 }
1443
1444 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1445                                           struct btrfs_extent_inline_ref *iref)
1446 {
1447         struct btrfs_key key;
1448         struct extent_buffer *leaf;
1449         struct btrfs_extent_data_ref *ref1;
1450         struct btrfs_shared_data_ref *ref2;
1451         u32 num_refs = 0;
1452         int type;
1453
1454         leaf = path->nodes[0];
1455         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1456         if (iref) {
1457                 /*
1458                  * If type is invalid, we should have bailed out earlier than
1459                  * this call.
1460                  */
1461                 type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_DATA);
1462                 ASSERT(type != BTRFS_REF_TYPE_INVALID);
1463                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1464                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1465                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1466                 } else {
1467                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1468                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1469                 }
1470         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1471                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1472                                       struct btrfs_extent_data_ref);
1473                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1474         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1475                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1476                                       struct btrfs_shared_data_ref);
1477                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1479         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1480                 struct btrfs_extent_ref_v0 *ref0;
1481                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1482                                       struct btrfs_extent_ref_v0);
1483                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1484 #endif
1485         } else {
1486                 WARN_ON(1);
1487         }
1488         return num_refs;
1489 }
1490
1491 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1492                                           struct btrfs_fs_info *fs_info,
1493                                           struct btrfs_path *path,
1494                                           u64 bytenr, u64 parent,
1495                                           u64 root_objectid)
1496 {
1497         struct btrfs_root *root = fs_info->extent_root;
1498         struct btrfs_key key;
1499         int ret;
1500
1501         key.objectid = bytenr;
1502         if (parent) {
1503                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1504                 key.offset = parent;
1505         } else {
1506                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1507                 key.offset = root_objectid;
1508         }
1509
1510         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1511         if (ret > 0)
1512                 ret = -ENOENT;
1513 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1514         if (ret == -ENOENT && parent) {
1515                 btrfs_release_path(path);
1516                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1517                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1518                 if (ret > 0)
1519                         ret = -ENOENT;
1520         }
1521 #endif
1522         return ret;
1523 }
1524
1525 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1526                                           struct btrfs_path *path,
1527                                           u64 bytenr, u64 parent,
1528                                           u64 root_objectid)
1529 {
1530         struct btrfs_key key;
1531         int ret;
1532
1533         key.objectid = bytenr;
1534         if (parent) {
1535                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1536                 key.offset = parent;
1537         } else {
1538                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1539                 key.offset = root_objectid;
1540         }
1541
1542         ret = btrfs_insert_empty_item(trans, trans->fs_info->extent_root,
1543                                       path, &key, 0);
1544         btrfs_release_path(path);
1545         return ret;
1546 }
1547
1548 static inline int extent_ref_type(u64 parent, u64 owner)
1549 {
1550         int type;
1551         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1552                 if (parent > 0)
1553                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1554                 else
1555                         type = BTRFS_TREE_BLOCK_REF_KEY;
1556         } else {
1557                 if (parent > 0)
1558                         type = BTRFS_SHARED_DATA_REF_KEY;
1559                 else
1560                         type = BTRFS_EXTENT_DATA_REF_KEY;
1561         }
1562         return type;
1563 }
1564
1565 static int find_next_key(struct btrfs_path *path, int level,
1566                          struct btrfs_key *key)
1567
1568 {
1569         for (; level < BTRFS_MAX_LEVEL; level++) {
1570                 if (!path->nodes[level])
1571                         break;
1572                 if (path->slots[level] + 1 >=
1573                     btrfs_header_nritems(path->nodes[level]))
1574                         continue;
1575                 if (level == 0)
1576                         btrfs_item_key_to_cpu(path->nodes[level], key,
1577                                               path->slots[level] + 1);
1578                 else
1579                         btrfs_node_key_to_cpu(path->nodes[level], key,
1580                                               path->slots[level] + 1);
1581                 return 0;
1582         }
1583         return 1;
1584 }
1585
1586 /*
1587  * look for inline back ref. if back ref is found, *ref_ret is set
1588  * to the address of inline back ref, and 0 is returned.
1589  *
1590  * if back ref isn't found, *ref_ret is set to the address where it
1591  * should be inserted, and -ENOENT is returned.
1592  *
1593  * if insert is true and there are too many inline back refs, the path
1594  * points to the extent item, and -EAGAIN is returned.
1595  *
1596  * NOTE: inline back refs are ordered in the same way that back ref
1597  *       items in the tree are ordered.
1598  */
1599 static noinline_for_stack
1600 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_fs_info *fs_info,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref **ref_ret,
1604                                  u64 bytenr, u64 num_bytes,
1605                                  u64 parent, u64 root_objectid,
1606                                  u64 owner, u64 offset, int insert)
1607 {
1608         struct btrfs_root *root = fs_info->extent_root;
1609         struct btrfs_key key;
1610         struct extent_buffer *leaf;
1611         struct btrfs_extent_item *ei;
1612         struct btrfs_extent_inline_ref *iref;
1613         u64 flags;
1614         u64 item_size;
1615         unsigned long ptr;
1616         unsigned long end;
1617         int extra_size;
1618         int type;
1619         int want;
1620         int ret;
1621         int err = 0;
1622         bool skinny_metadata = btrfs_fs_incompat(fs_info, SKINNY_METADATA);
1623         int needed;
1624
1625         key.objectid = bytenr;
1626         key.type = BTRFS_EXTENT_ITEM_KEY;
1627         key.offset = num_bytes;
1628
1629         want = extent_ref_type(parent, owner);
1630         if (insert) {
1631                 extra_size = btrfs_extent_inline_ref_size(want);
1632                 path->keep_locks = 1;
1633         } else
1634                 extra_size = -1;
1635
1636         /*
1637          * Owner is our level, so we can just add one to get the level for the
1638          * block we are interested in.
1639          */
1640         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1641                 key.type = BTRFS_METADATA_ITEM_KEY;
1642                 key.offset = owner;
1643         }
1644
1645 again:
1646         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1647         if (ret < 0) {
1648                 err = ret;
1649                 goto out;
1650         }
1651
1652         /*
1653          * We may be a newly converted file system which still has the old fat
1654          * extent entries for metadata, so try and see if we have one of those.
1655          */
1656         if (ret > 0 && skinny_metadata) {
1657                 skinny_metadata = false;
1658                 if (path->slots[0]) {
1659                         path->slots[0]--;
1660                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1661                                               path->slots[0]);
1662                         if (key.objectid == bytenr &&
1663                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1664                             key.offset == num_bytes)
1665                                 ret = 0;
1666                 }
1667                 if (ret) {
1668                         key.objectid = bytenr;
1669                         key.type = BTRFS_EXTENT_ITEM_KEY;
1670                         key.offset = num_bytes;
1671                         btrfs_release_path(path);
1672                         goto again;
1673                 }
1674         }
1675
1676         if (ret && !insert) {
1677                 err = -ENOENT;
1678                 goto out;
1679         } else if (WARN_ON(ret)) {
1680                 err = -EIO;
1681                 goto out;
1682         }
1683
1684         leaf = path->nodes[0];
1685         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1686 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1687         if (item_size < sizeof(*ei)) {
1688                 if (!insert) {
1689                         err = -ENOENT;
1690                         goto out;
1691                 }
1692                 ret = convert_extent_item_v0(trans, fs_info, path, owner,
1693                                              extra_size);
1694                 if (ret < 0) {
1695                         err = ret;
1696                         goto out;
1697                 }
1698                 leaf = path->nodes[0];
1699                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1700         }
1701 #endif
1702         BUG_ON(item_size < sizeof(*ei));
1703
1704         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1705         flags = btrfs_extent_flags(leaf, ei);
1706
1707         ptr = (unsigned long)(ei + 1);
1708         end = (unsigned long)ei + item_size;
1709
1710         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1711                 ptr += sizeof(struct btrfs_tree_block_info);
1712                 BUG_ON(ptr > end);
1713         }
1714
1715         if (owner >= BTRFS_FIRST_FREE_OBJECTID)
1716                 needed = BTRFS_REF_TYPE_DATA;
1717         else
1718                 needed = BTRFS_REF_TYPE_BLOCK;
1719
1720         err = -ENOENT;
1721         while (1) {
1722                 if (ptr >= end) {
1723                         WARN_ON(ptr > end);
1724                         break;
1725                 }
1726                 iref = (struct btrfs_extent_inline_ref *)ptr;
1727                 type = btrfs_get_extent_inline_ref_type(leaf, iref, needed);
1728                 if (type == BTRFS_REF_TYPE_INVALID) {
1729                         err = -EINVAL;
1730                         goto out;
1731                 }
1732
1733                 if (want < type)
1734                         break;
1735                 if (want > type) {
1736                         ptr += btrfs_extent_inline_ref_size(type);
1737                         continue;
1738                 }
1739
1740                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1741                         struct btrfs_extent_data_ref *dref;
1742                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1743                         if (match_extent_data_ref(leaf, dref, root_objectid,
1744                                                   owner, offset)) {
1745                                 err = 0;
1746                                 break;
1747                         }
1748                         if (hash_extent_data_ref_item(leaf, dref) <
1749                             hash_extent_data_ref(root_objectid, owner, offset))
1750                                 break;
1751                 } else {
1752                         u64 ref_offset;
1753                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1754                         if (parent > 0) {
1755                                 if (parent == ref_offset) {
1756                                         err = 0;
1757                                         break;
1758                                 }
1759                                 if (ref_offset < parent)
1760                                         break;
1761                         } else {
1762                                 if (root_objectid == ref_offset) {
1763                                         err = 0;
1764                                         break;
1765                                 }
1766                                 if (ref_offset < root_objectid)
1767                                         break;
1768                         }
1769                 }
1770                 ptr += btrfs_extent_inline_ref_size(type);
1771         }
1772         if (err == -ENOENT && insert) {
1773                 if (item_size + extra_size >=
1774                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1775                         err = -EAGAIN;
1776                         goto out;
1777                 }
1778                 /*
1779                  * To add new inline back ref, we have to make sure
1780                  * there is no corresponding back ref item.
1781                  * For simplicity, we just do not add new inline back
1782                  * ref if there is any kind of item for this block
1783                  */
1784                 if (find_next_key(path, 0, &key) == 0 &&
1785                     key.objectid == bytenr &&
1786                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1787                         err = -EAGAIN;
1788                         goto out;
1789                 }
1790         }
1791         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1792 out:
1793         if (insert) {
1794                 path->keep_locks = 0;
1795                 btrfs_unlock_up_safe(path, 1);
1796         }
1797         return err;
1798 }
1799
1800 /*
1801  * helper to add new inline back ref
1802  */
1803 static noinline_for_stack
1804 void setup_inline_extent_backref(struct btrfs_fs_info *fs_info,
1805                                  struct btrfs_path *path,
1806                                  struct btrfs_extent_inline_ref *iref,
1807                                  u64 parent, u64 root_objectid,
1808                                  u64 owner, u64 offset, int refs_to_add,
1809                                  struct btrfs_delayed_extent_op *extent_op)
1810 {
1811         struct extent_buffer *leaf;
1812         struct btrfs_extent_item *ei;
1813         unsigned long ptr;
1814         unsigned long end;
1815         unsigned long item_offset;
1816         u64 refs;
1817         int size;
1818         int type;
1819
1820         leaf = path->nodes[0];
1821         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1822         item_offset = (unsigned long)iref - (unsigned long)ei;
1823
1824         type = extent_ref_type(parent, owner);
1825         size = btrfs_extent_inline_ref_size(type);
1826
1827         btrfs_extend_item(fs_info, path, size);
1828
1829         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1830         refs = btrfs_extent_refs(leaf, ei);
1831         refs += refs_to_add;
1832         btrfs_set_extent_refs(leaf, ei, refs);
1833         if (extent_op)
1834                 __run_delayed_extent_op(extent_op, leaf, ei);
1835
1836         ptr = (unsigned long)ei + item_offset;
1837         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1838         if (ptr < end - size)
1839                 memmove_extent_buffer(leaf, ptr + size, ptr,
1840                                       end - size - ptr);
1841
1842         iref = (struct btrfs_extent_inline_ref *)ptr;
1843         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1844         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1845                 struct btrfs_extent_data_ref *dref;
1846                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1847                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1848                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1849                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1850                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1851         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1852                 struct btrfs_shared_data_ref *sref;
1853                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1854                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1855                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1856         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1857                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1858         } else {
1859                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1860         }
1861         btrfs_mark_buffer_dirty(leaf);
1862 }
1863
1864 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1865                                  struct btrfs_fs_info *fs_info,
1866                                  struct btrfs_path *path,
1867                                  struct btrfs_extent_inline_ref **ref_ret,
1868                                  u64 bytenr, u64 num_bytes, u64 parent,
1869                                  u64 root_objectid, u64 owner, u64 offset)
1870 {
1871         int ret;
1872
1873         ret = lookup_inline_extent_backref(trans, fs_info, path, ref_ret,
1874                                            bytenr, num_bytes, parent,
1875                                            root_objectid, owner, offset, 0);
1876         if (ret != -ENOENT)
1877                 return ret;
1878
1879         btrfs_release_path(path);
1880         *ref_ret = NULL;
1881
1882         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1883                 ret = lookup_tree_block_ref(trans, fs_info, path, bytenr,
1884                                             parent, root_objectid);
1885         } else {
1886                 ret = lookup_extent_data_ref(trans, fs_info, path, bytenr,
1887                                              parent, root_objectid, owner,
1888                                              offset);
1889         }
1890         return ret;
1891 }
1892
1893 /*
1894  * helper to update/remove inline back ref
1895  */
1896 static noinline_for_stack
1897 void update_inline_extent_backref(struct btrfs_fs_info *fs_info,
1898                                   struct btrfs_path *path,
1899                                   struct btrfs_extent_inline_ref *iref,
1900                                   int refs_to_mod,
1901                                   struct btrfs_delayed_extent_op *extent_op,
1902                                   int *last_ref)
1903 {
1904         struct extent_buffer *leaf;
1905         struct btrfs_extent_item *ei;
1906         struct btrfs_extent_data_ref *dref = NULL;
1907         struct btrfs_shared_data_ref *sref = NULL;
1908         unsigned long ptr;
1909         unsigned long end;
1910         u32 item_size;
1911         int size;
1912         int type;
1913         u64 refs;
1914
1915         leaf = path->nodes[0];
1916         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1917         refs = btrfs_extent_refs(leaf, ei);
1918         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1919         refs += refs_to_mod;
1920         btrfs_set_extent_refs(leaf, ei, refs);
1921         if (extent_op)
1922                 __run_delayed_extent_op(extent_op, leaf, ei);
1923
1924         /*
1925          * If type is invalid, we should have bailed out after
1926          * lookup_inline_extent_backref().
1927          */
1928         type = btrfs_get_extent_inline_ref_type(leaf, iref, BTRFS_REF_TYPE_ANY);
1929         ASSERT(type != BTRFS_REF_TYPE_INVALID);
1930
1931         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1932                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1933                 refs = btrfs_extent_data_ref_count(leaf, dref);
1934         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1935                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1936                 refs = btrfs_shared_data_ref_count(leaf, sref);
1937         } else {
1938                 refs = 1;
1939                 BUG_ON(refs_to_mod != -1);
1940         }
1941
1942         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1943         refs += refs_to_mod;
1944
1945         if (refs > 0) {
1946                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1947                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1948                 else
1949                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1950         } else {
1951                 *last_ref = 1;
1952                 size =  btrfs_extent_inline_ref_size(type);
1953                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1954                 ptr = (unsigned long)iref;
1955                 end = (unsigned long)ei + item_size;
1956                 if (ptr + size < end)
1957                         memmove_extent_buffer(leaf, ptr, ptr + size,
1958                                               end - ptr - size);
1959                 item_size -= size;
1960                 btrfs_truncate_item(fs_info, path, item_size, 1);
1961         }
1962         btrfs_mark_buffer_dirty(leaf);
1963 }
1964
1965 static noinline_for_stack
1966 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1967                                  struct btrfs_fs_info *fs_info,
1968                                  struct btrfs_path *path,
1969                                  u64 bytenr, u64 num_bytes, u64 parent,
1970                                  u64 root_objectid, u64 owner,
1971                                  u64 offset, int refs_to_add,
1972                                  struct btrfs_delayed_extent_op *extent_op)
1973 {
1974         struct btrfs_extent_inline_ref *iref;
1975         int ret;
1976
1977         ret = lookup_inline_extent_backref(trans, fs_info, path, &iref,
1978                                            bytenr, num_bytes, parent,
1979                                            root_objectid, owner, offset, 1);
1980         if (ret == 0) {
1981                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1982                 update_inline_extent_backref(fs_info, path, iref,
1983                                              refs_to_add, extent_op, NULL);
1984         } else if (ret == -ENOENT) {
1985                 setup_inline_extent_backref(fs_info, path, iref, parent,
1986                                             root_objectid, owner, offset,
1987                                             refs_to_add, extent_op);
1988                 ret = 0;
1989         }
1990         return ret;
1991 }
1992
1993 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1994                                  struct btrfs_fs_info *fs_info,
1995                                  struct btrfs_path *path,
1996                                  u64 bytenr, u64 parent, u64 root_objectid,
1997                                  u64 owner, u64 offset, int refs_to_add)
1998 {
1999         int ret;
2000         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2001                 BUG_ON(refs_to_add != 1);
2002                 ret = insert_tree_block_ref(trans, path, bytenr, parent,
2003                                             root_objectid);
2004         } else {
2005                 ret = insert_extent_data_ref(trans, fs_info, path, bytenr,
2006                                              parent, root_objectid,
2007                                              owner, offset, refs_to_add);
2008         }
2009         return ret;
2010 }
2011
2012 static int remove_extent_backref(struct btrfs_trans_handle *trans,
2013                                  struct btrfs_fs_info *fs_info,
2014                                  struct btrfs_path *path,
2015                                  struct btrfs_extent_inline_ref *iref,
2016                                  int refs_to_drop, int is_data, int *last_ref)
2017 {
2018         int ret = 0;
2019
2020         BUG_ON(!is_data && refs_to_drop != 1);
2021         if (iref) {
2022                 update_inline_extent_backref(fs_info, path, iref,
2023                                              -refs_to_drop, NULL, last_ref);
2024         } else if (is_data) {
2025                 ret = remove_extent_data_ref(trans, fs_info, path, refs_to_drop,
2026                                              last_ref);
2027         } else {
2028                 *last_ref = 1;
2029                 ret = btrfs_del_item(trans, fs_info->extent_root, path);
2030         }
2031         return ret;
2032 }
2033
2034 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
2035 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
2036                                u64 *discarded_bytes)
2037 {
2038         int j, ret = 0;
2039         u64 bytes_left, end;
2040         u64 aligned_start = ALIGN(start, 1 << 9);
2041
2042         if (WARN_ON(start != aligned_start)) {
2043                 len -= aligned_start - start;
2044                 len = round_down(len, 1 << 9);
2045                 start = aligned_start;
2046         }
2047
2048         *discarded_bytes = 0;
2049
2050         if (!len)
2051                 return 0;
2052
2053         end = start + len;
2054         bytes_left = len;
2055
2056         /* Skip any superblocks on this device. */
2057         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
2058                 u64 sb_start = btrfs_sb_offset(j);
2059                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2060                 u64 size = sb_start - start;
2061
2062                 if (!in_range(sb_start, start, bytes_left) &&
2063                     !in_range(sb_end, start, bytes_left) &&
2064                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2065                         continue;
2066
2067                 /*
2068                  * Superblock spans beginning of range.  Adjust start and
2069                  * try again.
2070                  */
2071                 if (sb_start <= start) {
2072                         start += sb_end - start;
2073                         if (start > end) {
2074                                 bytes_left = 0;
2075                                 break;
2076                         }
2077                         bytes_left = end - start;
2078                         continue;
2079                 }
2080
2081                 if (size) {
2082                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2083                                                    GFP_NOFS, 0);
2084                         if (!ret)
2085                                 *discarded_bytes += size;
2086                         else if (ret != -EOPNOTSUPP)
2087                                 return ret;
2088                 }
2089
2090                 start = sb_end;
2091                 if (start > end) {
2092                         bytes_left = 0;
2093                         break;
2094                 }
2095                 bytes_left = end - start;
2096         }
2097
2098         if (bytes_left) {
2099                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2100                                            GFP_NOFS, 0);
2101                 if (!ret)
2102                         *discarded_bytes += bytes_left;
2103         }
2104         return ret;
2105 }
2106
2107 int btrfs_discard_extent(struct btrfs_fs_info *fs_info, u64 bytenr,
2108                          u64 num_bytes, u64 *actual_bytes)
2109 {
2110         int ret;
2111         u64 discarded_bytes = 0;
2112         struct btrfs_bio *bbio = NULL;
2113
2114
2115         /*
2116          * Avoid races with device replace and make sure our bbio has devices
2117          * associated to its stripes that don't go away while we are discarding.
2118          */
2119         btrfs_bio_counter_inc_blocked(fs_info);
2120         /* Tell the block device(s) that the sectors can be discarded */
2121         ret = btrfs_map_block(fs_info, BTRFS_MAP_DISCARD, bytenr, &num_bytes,
2122                               &bbio, 0);
2123         /* Error condition is -ENOMEM */
2124         if (!ret) {
2125                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2126                 int i;
2127
2128
2129                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2130                         u64 bytes;
2131                         struct request_queue *req_q;
2132
2133                         if (!stripe->dev->bdev) {
2134                                 ASSERT(btrfs_test_opt(fs_info, DEGRADED));
2135                                 continue;
2136                         }
2137                         req_q = bdev_get_queue(stripe->dev->bdev);
2138                         if (!blk_queue_discard(req_q))
2139                                 continue;
2140
2141                         ret = btrfs_issue_discard(stripe->dev->bdev,
2142                                                   stripe->physical,
2143                                                   stripe->length,
2144                                                   &bytes);
2145                         if (!ret)
2146                                 discarded_bytes += bytes;
2147                         else if (ret != -EOPNOTSUPP)
2148                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2149
2150                         /*
2151                          * Just in case we get back EOPNOTSUPP for some reason,
2152                          * just ignore the return value so we don't screw up
2153                          * people calling discard_extent.
2154                          */
2155                         ret = 0;
2156                 }
2157                 btrfs_put_bbio(bbio);
2158         }
2159         btrfs_bio_counter_dec(fs_info);
2160
2161         if (actual_bytes)
2162                 *actual_bytes = discarded_bytes;
2163
2164
2165         if (ret == -EOPNOTSUPP)
2166                 ret = 0;
2167         return ret;
2168 }
2169
2170 /* Can return -ENOMEM */
2171 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2172                          struct btrfs_root *root,
2173                          u64 bytenr, u64 num_bytes, u64 parent,
2174                          u64 root_objectid, u64 owner, u64 offset)
2175 {
2176         struct btrfs_fs_info *fs_info = root->fs_info;
2177         int old_ref_mod, new_ref_mod;
2178         int ret;
2179
2180         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2181                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2182
2183         btrfs_ref_tree_mod(root, bytenr, num_bytes, parent, root_objectid,
2184                            owner, offset, BTRFS_ADD_DELAYED_REF);
2185
2186         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2187                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2188                                                  num_bytes, parent,
2189                                                  root_objectid, (int)owner,
2190                                                  BTRFS_ADD_DELAYED_REF, NULL,
2191                                                  &old_ref_mod, &new_ref_mod);
2192         } else {
2193                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2194                                                  num_bytes, parent,
2195                                                  root_objectid, owner, offset,
2196                                                  0, BTRFS_ADD_DELAYED_REF,
2197                                                  &old_ref_mod, &new_ref_mod);
2198         }
2199
2200         if (ret == 0 && old_ref_mod < 0 && new_ref_mod >= 0) {
2201                 bool metadata = owner < BTRFS_FIRST_FREE_OBJECTID;
2202
2203                 add_pinned_bytes(fs_info, -num_bytes, metadata, root_objectid);
2204         }
2205
2206         return ret;
2207 }
2208
2209 /*
2210  * __btrfs_inc_extent_ref - insert backreference for a given extent
2211  *
2212  * @trans:          Handle of transaction
2213  *
2214  * @node:           The delayed ref node used to get the bytenr/length for
2215  *                  extent whose references are incremented.
2216  *
2217  * @parent:         If this is a shared extent (BTRFS_SHARED_DATA_REF_KEY/
2218  *                  BTRFS_SHARED_BLOCK_REF_KEY) then it holds the logical
2219  *                  bytenr of the parent block. Since new extents are always
2220  *                  created with indirect references, this will only be the case
2221  *                  when relocating a shared extent. In that case, root_objectid
2222  *                  will be BTRFS_TREE_RELOC_OBJECTID. Otheriwse, parent must
2223  *                  be 0
2224  *
2225  * @root_objectid:  The id of the root where this modification has originated,
2226  *                  this can be either one of the well-known metadata trees or
2227  *                  the subvolume id which references this extent.
2228  *
2229  * @owner:          For data extents it is the inode number of the owning file.
2230  *                  For metadata extents this parameter holds the level in the
2231  *                  tree of the extent.
2232  *
2233  * @offset:         For metadata extents the offset is ignored and is currently
2234  *                  always passed as 0. For data extents it is the fileoffset
2235  *                  this extent belongs to.
2236  *
2237  * @refs_to_add     Number of references to add
2238  *
2239  * @extent_op       Pointer to a structure, holding information necessary when
2240  *                  updating a tree block's flags
2241  *
2242  */
2243 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2244                                   struct btrfs_fs_info *fs_info,
2245                                   struct btrfs_delayed_ref_node *node,
2246                                   u64 parent, u64 root_objectid,
2247                                   u64 owner, u64 offset, int refs_to_add,
2248                                   struct btrfs_delayed_extent_op *extent_op)
2249 {
2250         struct btrfs_path *path;
2251         struct extent_buffer *leaf;
2252         struct btrfs_extent_item *item;
2253         struct btrfs_key key;
2254         u64 bytenr = node->bytenr;
2255         u64 num_bytes = node->num_bytes;
2256         u64 refs;
2257         int ret;
2258
2259         path = btrfs_alloc_path();
2260         if (!path)
2261                 return -ENOMEM;
2262
2263         path->reada = READA_FORWARD;
2264         path->leave_spinning = 1;
2265         /* this will setup the path even if it fails to insert the back ref */
2266         ret = insert_inline_extent_backref(trans, fs_info, path, bytenr,
2267                                            num_bytes, parent, root_objectid,
2268                                            owner, offset,
2269                                            refs_to_add, extent_op);
2270         if ((ret < 0 && ret != -EAGAIN) || !ret)
2271                 goto out;
2272
2273         /*
2274          * Ok we had -EAGAIN which means we didn't have space to insert and
2275          * inline extent ref, so just update the reference count and add a
2276          * normal backref.
2277          */
2278         leaf = path->nodes[0];
2279         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2280         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2281         refs = btrfs_extent_refs(leaf, item);
2282         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2283         if (extent_op)
2284                 __run_delayed_extent_op(extent_op, leaf, item);
2285
2286         btrfs_mark_buffer_dirty(leaf);
2287         btrfs_release_path(path);
2288
2289         path->reada = READA_FORWARD;
2290         path->leave_spinning = 1;
2291         /* now insert the actual backref */
2292         ret = insert_extent_backref(trans, fs_info, path, bytenr, parent,
2293                                     root_objectid, owner, offset, refs_to_add);
2294         if (ret)
2295                 btrfs_abort_transaction(trans, ret);
2296 out:
2297         btrfs_free_path(path);
2298         return ret;
2299 }
2300
2301 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2302                                 struct btrfs_fs_info *fs_info,
2303                                 struct btrfs_delayed_ref_node *node,
2304                                 struct btrfs_delayed_extent_op *extent_op,
2305                                 int insert_reserved)
2306 {
2307         int ret = 0;
2308         struct btrfs_delayed_data_ref *ref;
2309         struct btrfs_key ins;
2310         u64 parent = 0;
2311         u64 ref_root = 0;
2312         u64 flags = 0;
2313
2314         ins.objectid = node->bytenr;
2315         ins.offset = node->num_bytes;
2316         ins.type = BTRFS_EXTENT_ITEM_KEY;
2317
2318         ref = btrfs_delayed_node_to_data_ref(node);
2319         trace_run_delayed_data_ref(fs_info, node, ref, node->action);
2320
2321         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2322                 parent = ref->parent;
2323         ref_root = ref->root;
2324
2325         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2326                 if (extent_op)
2327                         flags |= extent_op->flags_to_set;
2328                 ret = alloc_reserved_file_extent(trans, fs_info,
2329                                                  parent, ref_root, flags,
2330                                                  ref->objectid, ref->offset,
2331                                                  &ins, node->ref_mod);
2332         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2333                 ret = __btrfs_inc_extent_ref(trans, fs_info, node, parent,
2334                                              ref_root, ref->objectid,
2335                                              ref->offset, node->ref_mod,
2336                                              extent_op);
2337         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2338                 ret = __btrfs_free_extent(trans, fs_info, node, parent,
2339                                           ref_root, ref->objectid,
2340                                           ref->offset, node->ref_mod,
2341                                           extent_op);
2342         } else {
2343                 BUG();
2344         }
2345         return ret;
2346 }
2347
2348 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2349                                     struct extent_buffer *leaf,
2350                                     struct btrfs_extent_item *ei)
2351 {
2352         u64 flags = btrfs_extent_flags(leaf, ei);
2353         if (extent_op->update_flags) {
2354                 flags |= extent_op->flags_to_set;
2355                 btrfs_set_extent_flags(leaf, ei, flags);
2356         }
2357
2358         if (extent_op->update_key) {
2359                 struct btrfs_tree_block_info *bi;
2360                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2361                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2362                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2363         }
2364 }
2365
2366 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2367                                  struct btrfs_fs_info *fs_info,
2368                                  struct btrfs_delayed_ref_head *head,
2369                                  struct btrfs_delayed_extent_op *extent_op)
2370 {
2371         struct btrfs_key key;
2372         struct btrfs_path *path;
2373         struct btrfs_extent_item *ei;
2374         struct extent_buffer *leaf;
2375         u32 item_size;
2376         int ret;
2377         int err = 0;
2378         int metadata = !extent_op->is_data;
2379
2380         if (trans->aborted)
2381                 return 0;
2382
2383         if (metadata && !btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2384                 metadata = 0;
2385
2386         path = btrfs_alloc_path();
2387         if (!path)
2388                 return -ENOMEM;
2389
2390         key.objectid = head->bytenr;
2391
2392         if (metadata) {
2393                 key.type = BTRFS_METADATA_ITEM_KEY;
2394                 key.offset = extent_op->level;
2395         } else {
2396                 key.type = BTRFS_EXTENT_ITEM_KEY;
2397                 key.offset = head->num_bytes;
2398         }
2399
2400 again:
2401         path->reada = READA_FORWARD;
2402         path->leave_spinning = 1;
2403         ret = btrfs_search_slot(trans, fs_info->extent_root, &key, path, 0, 1);
2404         if (ret < 0) {
2405                 err = ret;
2406                 goto out;
2407         }
2408         if (ret > 0) {
2409                 if (metadata) {
2410                         if (path->slots[0] > 0) {
2411                                 path->slots[0]--;
2412                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2413                                                       path->slots[0]);
2414                                 if (key.objectid == head->bytenr &&
2415                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2416                                     key.offset == head->num_bytes)
2417                                         ret = 0;
2418                         }
2419                         if (ret > 0) {
2420                                 btrfs_release_path(path);
2421                                 metadata = 0;
2422
2423                                 key.objectid = head->bytenr;
2424                                 key.offset = head->num_bytes;
2425                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2426                                 goto again;
2427                         }
2428                 } else {
2429                         err = -EIO;
2430                         goto out;
2431                 }
2432         }
2433
2434         leaf = path->nodes[0];
2435         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2436 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2437         if (item_size < sizeof(*ei)) {
2438                 ret = convert_extent_item_v0(trans, fs_info, path, (u64)-1, 0);
2439                 if (ret < 0) {
2440                         err = ret;
2441                         goto out;
2442                 }
2443                 leaf = path->nodes[0];
2444                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2445         }
2446 #endif
2447         BUG_ON(item_size < sizeof(*ei));
2448         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2449         __run_delayed_extent_op(extent_op, leaf, ei);
2450
2451         btrfs_mark_buffer_dirty(leaf);
2452 out:
2453         btrfs_free_path(path);
2454         return err;
2455 }
2456
2457 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2458                                 struct btrfs_fs_info *fs_info,
2459                                 struct btrfs_delayed_ref_node *node,
2460                                 struct btrfs_delayed_extent_op *extent_op,
2461                                 int insert_reserved)
2462 {
2463         int ret = 0;
2464         struct btrfs_delayed_tree_ref *ref;
2465         u64 parent = 0;
2466         u64 ref_root = 0;
2467
2468         ref = btrfs_delayed_node_to_tree_ref(node);
2469         trace_run_delayed_tree_ref(fs_info, node, ref, node->action);
2470
2471         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2472                 parent = ref->parent;
2473         ref_root = ref->root;
2474
2475         if (node->ref_mod != 1) {
2476                 btrfs_err(fs_info,
2477         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2478                           node->bytenr, node->ref_mod, node->action, ref_root,
2479                           parent);
2480                 return -EIO;
2481         }
2482         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2483                 BUG_ON(!extent_op || !extent_op->update_flags);
2484                 ret = alloc_reserved_tree_block(trans, node, extent_op);
2485         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2486                 ret = __btrfs_inc_extent_ref(trans, fs_info, node,
2487                                              parent, ref_root,
2488                                              ref->level, 0, 1,
2489                                              extent_op);
2490         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2491                 ret = __btrfs_free_extent(trans, fs_info, node,
2492                                           parent, ref_root,
2493                                           ref->level, 0, 1, extent_op);
2494         } else {
2495                 BUG();
2496         }
2497         return ret;
2498 }
2499
2500 /* helper function to actually process a single delayed ref entry */
2501 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2502                                struct btrfs_fs_info *fs_info,
2503                                struct btrfs_delayed_ref_node *node,
2504                                struct btrfs_delayed_extent_op *extent_op,
2505                                int insert_reserved)
2506 {
2507         int ret = 0;
2508
2509         if (trans->aborted) {
2510                 if (insert_reserved)
2511                         btrfs_pin_extent(fs_info, node->bytenr,
2512                                          node->num_bytes, 1);
2513                 return 0;
2514         }
2515
2516         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2517             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2518                 ret = run_delayed_tree_ref(trans, fs_info, node, extent_op,
2519                                            insert_reserved);
2520         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2521                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2522                 ret = run_delayed_data_ref(trans, fs_info, node, extent_op,
2523                                            insert_reserved);
2524         else
2525                 BUG();
2526         return ret;
2527 }
2528
2529 static inline struct btrfs_delayed_ref_node *
2530 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2531 {
2532         struct btrfs_delayed_ref_node *ref;
2533
2534         if (RB_EMPTY_ROOT(&head->ref_tree))
2535                 return NULL;
2536
2537         /*
2538          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2539          * This is to prevent a ref count from going down to zero, which deletes
2540          * the extent item from the extent tree, when there still are references
2541          * to add, which would fail because they would not find the extent item.
2542          */
2543         if (!list_empty(&head->ref_add_list))
2544                 return list_first_entry(&head->ref_add_list,
2545                                 struct btrfs_delayed_ref_node, add_list);
2546
2547         ref = rb_entry(rb_first(&head->ref_tree),
2548                        struct btrfs_delayed_ref_node, ref_node);
2549         ASSERT(list_empty(&ref->add_list));
2550         return ref;
2551 }
2552
2553 static void unselect_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
2554                                       struct btrfs_delayed_ref_head *head)
2555 {
2556         spin_lock(&delayed_refs->lock);
2557         head->processing = 0;
2558         delayed_refs->num_heads_ready++;
2559         spin_unlock(&delayed_refs->lock);
2560         btrfs_delayed_ref_unlock(head);
2561 }
2562
2563 static int cleanup_extent_op(struct btrfs_trans_handle *trans,
2564                              struct btrfs_fs_info *fs_info,
2565                              struct btrfs_delayed_ref_head *head)
2566 {
2567         struct btrfs_delayed_extent_op *extent_op = head->extent_op;
2568         int ret;
2569
2570         if (!extent_op)
2571                 return 0;
2572         head->extent_op = NULL;
2573         if (head->must_insert_reserved) {
2574                 btrfs_free_delayed_extent_op(extent_op);
2575                 return 0;
2576         }
2577         spin_unlock(&head->lock);
2578         ret = run_delayed_extent_op(trans, fs_info, head, extent_op);
2579         btrfs_free_delayed_extent_op(extent_op);
2580         return ret ? ret : 1;
2581 }
2582
2583 static int cleanup_ref_head(struct btrfs_trans_handle *trans,
2584                             struct btrfs_fs_info *fs_info,
2585                             struct btrfs_delayed_ref_head *head)
2586 {
2587         struct btrfs_delayed_ref_root *delayed_refs;
2588         int ret;
2589
2590         delayed_refs = &trans->transaction->delayed_refs;
2591
2592         ret = cleanup_extent_op(trans, fs_info, head);
2593         if (ret < 0) {
2594                 unselect_delayed_ref_head(delayed_refs, head);
2595                 btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2596                 return ret;
2597         } else if (ret) {
2598                 return ret;
2599         }
2600
2601         /*
2602          * Need to drop our head ref lock and re-acquire the delayed ref lock
2603          * and then re-check to make sure nobody got added.
2604          */
2605         spin_unlock(&head->lock);
2606         spin_lock(&delayed_refs->lock);
2607         spin_lock(&head->lock);
2608         if (!RB_EMPTY_ROOT(&head->ref_tree) || head->extent_op) {
2609                 spin_unlock(&head->lock);
2610                 spin_unlock(&delayed_refs->lock);
2611                 return 1;
2612         }
2613         delayed_refs->num_heads--;
2614         rb_erase(&head->href_node, &delayed_refs->href_root);
2615         RB_CLEAR_NODE(&head->href_node);
2616         spin_unlock(&head->lock);
2617         spin_unlock(&delayed_refs->lock);
2618         atomic_dec(&delayed_refs->num_entries);
2619
2620         trace_run_delayed_ref_head(fs_info, head, 0);
2621
2622         if (head->total_ref_mod < 0) {
2623                 struct btrfs_space_info *space_info;
2624                 u64 flags;
2625
2626                 if (head->is_data)
2627                         flags = BTRFS_BLOCK_GROUP_DATA;
2628                 else if (head->is_system)
2629                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
2630                 else
2631                         flags = BTRFS_BLOCK_GROUP_METADATA;
2632                 space_info = __find_space_info(fs_info, flags);
2633                 ASSERT(space_info);
2634                 percpu_counter_add(&space_info->total_bytes_pinned,
2635                                    -head->num_bytes);
2636
2637                 if (head->is_data) {
2638                         spin_lock(&delayed_refs->lock);
2639                         delayed_refs->pending_csums -= head->num_bytes;
2640                         spin_unlock(&delayed_refs->lock);
2641                 }
2642         }
2643
2644         if (head->must_insert_reserved) {
2645                 btrfs_pin_extent(fs_info, head->bytenr,
2646                                  head->num_bytes, 1);
2647                 if (head->is_data) {
2648                         ret = btrfs_del_csums(trans, fs_info, head->bytenr,
2649                                               head->num_bytes);
2650                 }
2651         }
2652
2653         /* Also free its reserved qgroup space */
2654         btrfs_qgroup_free_delayed_ref(fs_info, head->qgroup_ref_root,
2655                                       head->qgroup_reserved);
2656         btrfs_delayed_ref_unlock(head);
2657         btrfs_put_delayed_ref_head(head);
2658         return 0;
2659 }
2660
2661 /*
2662  * Returns 0 on success or if called with an already aborted transaction.
2663  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2664  */
2665 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2666                                              unsigned long nr)
2667 {
2668         struct btrfs_fs_info *fs_info = trans->fs_info;
2669         struct btrfs_delayed_ref_root *delayed_refs;
2670         struct btrfs_delayed_ref_node *ref;
2671         struct btrfs_delayed_ref_head *locked_ref = NULL;
2672         struct btrfs_delayed_extent_op *extent_op;
2673         ktime_t start = ktime_get();
2674         int ret;
2675         unsigned long count = 0;
2676         unsigned long actual_count = 0;
2677         int must_insert_reserved = 0;
2678
2679         delayed_refs = &trans->transaction->delayed_refs;
2680         while (1) {
2681                 if (!locked_ref) {
2682                         if (count >= nr)
2683                                 break;
2684
2685                         spin_lock(&delayed_refs->lock);
2686                         locked_ref = btrfs_select_ref_head(trans);
2687                         if (!locked_ref) {
2688                                 spin_unlock(&delayed_refs->lock);
2689                                 break;
2690                         }
2691
2692                         /* grab the lock that says we are going to process
2693                          * all the refs for this head */
2694                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2695                         spin_unlock(&delayed_refs->lock);
2696                         /*
2697                          * we may have dropped the spin lock to get the head
2698                          * mutex lock, and that might have given someone else
2699                          * time to free the head.  If that's true, it has been
2700                          * removed from our list and we can move on.
2701                          */
2702                         if (ret == -EAGAIN) {
2703                                 locked_ref = NULL;
2704                                 count++;
2705                                 continue;
2706                         }
2707                 }
2708
2709                 /*
2710                  * We need to try and merge add/drops of the same ref since we
2711                  * can run into issues with relocate dropping the implicit ref
2712                  * and then it being added back again before the drop can
2713                  * finish.  If we merged anything we need to re-loop so we can
2714                  * get a good ref.
2715                  * Or we can get node references of the same type that weren't
2716                  * merged when created due to bumps in the tree mod seq, and
2717                  * we need to merge them to prevent adding an inline extent
2718                  * backref before dropping it (triggering a BUG_ON at
2719                  * insert_inline_extent_backref()).
2720                  */
2721                 spin_lock(&locked_ref->lock);
2722                 btrfs_merge_delayed_refs(trans, delayed_refs, locked_ref);
2723
2724                 ref = select_delayed_ref(locked_ref);
2725
2726                 if (ref && ref->seq &&
2727                     btrfs_check_delayed_seq(fs_info, ref->seq)) {
2728                         spin_unlock(&locked_ref->lock);
2729                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2730                         locked_ref = NULL;
2731                         cond_resched();
2732                         count++;
2733                         continue;
2734                 }
2735
2736                 /*
2737                  * We're done processing refs in this ref_head, clean everything
2738                  * up and move on to the next ref_head.
2739                  */
2740                 if (!ref) {
2741                         ret = cleanup_ref_head(trans, fs_info, locked_ref);
2742                         if (ret > 0 ) {
2743                                 /* We dropped our lock, we need to loop. */
2744                                 ret = 0;
2745                                 continue;
2746                         } else if (ret) {
2747                                 return ret;
2748                         }
2749                         locked_ref = NULL;
2750                         count++;
2751                         continue;
2752                 }
2753
2754                 actual_count++;
2755                 ref->in_tree = 0;
2756                 rb_erase(&ref->ref_node, &locked_ref->ref_tree);
2757                 RB_CLEAR_NODE(&ref->ref_node);
2758                 if (!list_empty(&ref->add_list))
2759                         list_del(&ref->add_list);
2760                 /*
2761                  * When we play the delayed ref, also correct the ref_mod on
2762                  * head
2763                  */
2764                 switch (ref->action) {
2765                 case BTRFS_ADD_DELAYED_REF:
2766                 case BTRFS_ADD_DELAYED_EXTENT:
2767                         locked_ref->ref_mod -= ref->ref_mod;
2768                         break;
2769                 case BTRFS_DROP_DELAYED_REF:
2770                         locked_ref->ref_mod += ref->ref_mod;
2771                         break;
2772                 default:
2773                         WARN_ON(1);
2774                 }
2775                 atomic_dec(&delayed_refs->num_entries);
2776
2777                 /*
2778                  * Record the must-insert_reserved flag before we drop the spin
2779                  * lock.
2780                  */
2781                 must_insert_reserved = locked_ref->must_insert_reserved;
2782                 locked_ref->must_insert_reserved = 0;
2783
2784                 extent_op = locked_ref->extent_op;
2785                 locked_ref->extent_op = NULL;
2786                 spin_unlock(&locked_ref->lock);
2787
2788                 ret = run_one_delayed_ref(trans, fs_info, ref, extent_op,
2789                                           must_insert_reserved);
2790
2791                 btrfs_free_delayed_extent_op(extent_op);
2792                 if (ret) {
2793                         unselect_delayed_ref_head(delayed_refs, locked_ref);
2794                         btrfs_put_delayed_ref(ref);
2795                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d",
2796                                     ret);
2797                         return ret;
2798                 }
2799
2800                 btrfs_put_delayed_ref(ref);
2801                 count++;
2802                 cond_resched();
2803         }
2804
2805         /*
2806          * We don't want to include ref heads since we can have empty ref heads
2807          * and those will drastically skew our runtime down since we just do
2808          * accounting, no actual extent tree updates.
2809          */
2810         if (actual_count > 0) {
2811                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2812                 u64 avg;
2813
2814                 /*
2815                  * We weigh the current average higher than our current runtime
2816                  * to avoid large swings in the average.
2817                  */
2818                 spin_lock(&delayed_refs->lock);
2819                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2820                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2821                 spin_unlock(&delayed_refs->lock);
2822         }
2823         return 0;
2824 }
2825
2826 #ifdef SCRAMBLE_DELAYED_REFS
2827 /*
2828  * Normally delayed refs get processed in ascending bytenr order. This
2829  * correlates in most cases to the order added. To expose dependencies on this
2830  * order, we start to process the tree in the middle instead of the beginning
2831  */
2832 static u64 find_middle(struct rb_root *root)
2833 {
2834         struct rb_node *n = root->rb_node;
2835         struct btrfs_delayed_ref_node *entry;
2836         int alt = 1;
2837         u64 middle;
2838         u64 first = 0, last = 0;
2839
2840         n = rb_first(root);
2841         if (n) {
2842                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2843                 first = entry->bytenr;
2844         }
2845         n = rb_last(root);
2846         if (n) {
2847                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2848                 last = entry->bytenr;
2849         }
2850         n = root->rb_node;
2851
2852         while (n) {
2853                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2854                 WARN_ON(!entry->in_tree);
2855
2856                 middle = entry->bytenr;
2857
2858                 if (alt)
2859                         n = n->rb_left;
2860                 else
2861                         n = n->rb_right;
2862
2863                 alt = 1 - alt;
2864         }
2865         return middle;
2866 }
2867 #endif
2868
2869 static inline u64 heads_to_leaves(struct btrfs_fs_info *fs_info, u64 heads)
2870 {
2871         u64 num_bytes;
2872
2873         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2874                              sizeof(struct btrfs_extent_inline_ref));
2875         if (!btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2876                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2877
2878         /*
2879          * We don't ever fill up leaves all the way so multiply by 2 just to be
2880          * closer to what we're really going to want to use.
2881          */
2882         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(fs_info));
2883 }
2884
2885 /*
2886  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2887  * would require to store the csums for that many bytes.
2888  */
2889 u64 btrfs_csum_bytes_to_leaves(struct btrfs_fs_info *fs_info, u64 csum_bytes)
2890 {
2891         u64 csum_size;
2892         u64 num_csums_per_leaf;
2893         u64 num_csums;
2894
2895         csum_size = BTRFS_MAX_ITEM_SIZE(fs_info);
2896         num_csums_per_leaf = div64_u64(csum_size,
2897                         (u64)btrfs_super_csum_size(fs_info->super_copy));
2898         num_csums = div64_u64(csum_bytes, fs_info->sectorsize);
2899         num_csums += num_csums_per_leaf - 1;
2900         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2901         return num_csums;
2902 }
2903
2904 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2905                                        struct btrfs_fs_info *fs_info)
2906 {
2907         struct btrfs_block_rsv *global_rsv;
2908         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2909         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2910         unsigned int num_dirty_bgs = trans->transaction->num_dirty_bgs;
2911         u64 num_bytes, num_dirty_bgs_bytes;
2912         int ret = 0;
2913
2914         num_bytes = btrfs_calc_trans_metadata_size(fs_info, 1);
2915         num_heads = heads_to_leaves(fs_info, num_heads);
2916         if (num_heads > 1)
2917                 num_bytes += (num_heads - 1) * fs_info->nodesize;
2918         num_bytes <<= 1;
2919         num_bytes += btrfs_csum_bytes_to_leaves(fs_info, csum_bytes) *
2920                                                         fs_info->nodesize;
2921         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(fs_info,
2922                                                              num_dirty_bgs);
2923         global_rsv = &fs_info->global_block_rsv;
2924
2925         /*
2926          * If we can't allocate any more chunks lets make sure we have _lots_ of
2927          * wiggle room since running delayed refs can create more delayed refs.
2928          */
2929         if (global_rsv->space_info->full) {
2930                 num_dirty_bgs_bytes <<= 1;
2931                 num_bytes <<= 1;
2932         }
2933
2934         spin_lock(&global_rsv->lock);
2935         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2936                 ret = 1;
2937         spin_unlock(&global_rsv->lock);
2938         return ret;
2939 }
2940
2941 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2942                                        struct btrfs_fs_info *fs_info)
2943 {
2944         u64 num_entries =
2945                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2946         u64 avg_runtime;
2947         u64 val;
2948
2949         smp_mb();
2950         avg_runtime = fs_info->avg_delayed_ref_runtime;
2951         val = num_entries * avg_runtime;
2952         if (val >= NSEC_PER_SEC)
2953                 return 1;
2954         if (val >= NSEC_PER_SEC / 2)
2955                 return 2;
2956
2957         return btrfs_check_space_for_delayed_refs(trans, fs_info);
2958 }
2959
2960 struct async_delayed_refs {
2961         struct btrfs_root *root;
2962         u64 transid;
2963         int count;
2964         int error;
2965         int sync;
2966         struct completion wait;
2967         struct btrfs_work work;
2968 };
2969
2970 static inline struct async_delayed_refs *
2971 to_async_delayed_refs(struct btrfs_work *work)
2972 {
2973         return container_of(work, struct async_delayed_refs, work);
2974 }
2975
2976 static void delayed_ref_async_start(struct btrfs_work *work)
2977 {
2978         struct async_delayed_refs *async = to_async_delayed_refs(work);
2979         struct btrfs_trans_handle *trans;
2980         struct btrfs_fs_info *fs_info = async->root->fs_info;
2981         int ret;
2982
2983         /* if the commit is already started, we don't need to wait here */
2984         if (btrfs_transaction_blocked(fs_info))
2985                 goto done;
2986
2987         trans = btrfs_join_transaction(async->root);
2988         if (IS_ERR(trans)) {
2989                 async->error = PTR_ERR(trans);
2990                 goto done;
2991         }
2992
2993         /*
2994          * trans->sync means that when we call end_transaction, we won't
2995          * wait on delayed refs
2996          */
2997         trans->sync = true;
2998
2999         /* Don't bother flushing if we got into a different transaction */
3000         if (trans->transid > async->transid)
3001                 goto end;
3002
3003         ret = btrfs_run_delayed_refs(trans, async->count);
3004         if (ret)
3005                 async->error = ret;
3006 end:
3007         ret = btrfs_end_transaction(trans);
3008         if (ret && !async->error)
3009                 async->error = ret;
3010 done:
3011         if (async->sync)
3012                 complete(&async->wait);
3013         else
3014                 kfree(async);
3015 }
3016
3017 int btrfs_async_run_delayed_refs(struct btrfs_fs_info *fs_info,
3018                                  unsigned long count, u64 transid, int wait)
3019 {
3020         struct async_delayed_refs *async;
3021         int ret;
3022
3023         async = kmalloc(sizeof(*async), GFP_NOFS);
3024         if (!async)
3025                 return -ENOMEM;
3026
3027         async->root = fs_info->tree_root;
3028         async->count = count;
3029         async->error = 0;
3030         async->transid = transid;
3031         if (wait)
3032                 async->sync = 1;
3033         else
3034                 async->sync = 0;
3035         init_completion(&async->wait);
3036
3037         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
3038                         delayed_ref_async_start, NULL, NULL);
3039
3040         btrfs_queue_work(fs_info->extent_workers, &async->work);
3041
3042         if (wait) {
3043                 wait_for_completion(&async->wait);
3044                 ret = async->error;
3045                 kfree(async);
3046                 return ret;
3047         }
3048         return 0;
3049 }
3050
3051 /*
3052  * this starts processing the delayed reference count updates and
3053  * extent insertions we have queued up so far.  count can be
3054  * 0, which means to process everything in the tree at the start
3055  * of the run (but not newly added entries), or it can be some target
3056  * number you'd like to process.
3057  *
3058  * Returns 0 on success or if called with an aborted transaction
3059  * Returns <0 on error and aborts the transaction
3060  */
3061 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
3062                            unsigned long count)
3063 {
3064         struct btrfs_fs_info *fs_info = trans->fs_info;
3065         struct rb_node *node;