Btrfs: make sure all pending extent operations are complete
[linux-2.6] / fs / btrfs / ctree.c
1 /*
2  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/sched.h>
20 #include "ctree.h"
21 #include "disk-io.h"
22 #include "transaction.h"
23 #include "print-tree.h"
24 #include "locking.h"
25
26 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
27                       *root, struct btrfs_path *path, int level);
28 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
29                       *root, struct btrfs_key *ins_key,
30                       struct btrfs_path *path, int data_size, int extend);
31 static int push_node_left(struct btrfs_trans_handle *trans,
32                           struct btrfs_root *root, struct extent_buffer *dst,
33                           struct extent_buffer *src, int empty);
34 static int balance_node_right(struct btrfs_trans_handle *trans,
35                               struct btrfs_root *root,
36                               struct extent_buffer *dst_buf,
37                               struct extent_buffer *src_buf);
38 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
39                    struct btrfs_path *path, int level, int slot);
40
41 inline void btrfs_init_path(struct btrfs_path *p)
42 {
43         memset(p, 0, sizeof(*p));
44 }
45
46 struct btrfs_path *btrfs_alloc_path(void)
47 {
48         struct btrfs_path *path;
49         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
50         if (path) {
51                 btrfs_init_path(path);
52                 path->reada = 1;
53         }
54         return path;
55 }
56
57 /*
58  * set all locked nodes in the path to blocking locks.  This should
59  * be done before scheduling
60  */
61 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
62 {
63         int i;
64         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
65                 if (p->nodes[i] && p->locks[i])
66                         btrfs_set_lock_blocking(p->nodes[i]);
67         }
68 }
69
70 /*
71  * reset all the locked nodes in the patch to spinning locks.
72  */
73 noinline void btrfs_clear_path_blocking(struct btrfs_path *p)
74 {
75         int i;
76         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
77                 if (p->nodes[i] && p->locks[i])
78                         btrfs_clear_lock_blocking(p->nodes[i]);
79         }
80 }
81
82 /* this also releases the path */
83 void btrfs_free_path(struct btrfs_path *p)
84 {
85         btrfs_release_path(NULL, p);
86         kmem_cache_free(btrfs_path_cachep, p);
87 }
88
89 /*
90  * path release drops references on the extent buffers in the path
91  * and it drops any locks held by this path
92  *
93  * It is safe to call this on paths that no locks or extent buffers held.
94  */
95 noinline void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
96 {
97         int i;
98
99         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
100                 p->slots[i] = 0;
101                 if (!p->nodes[i])
102                         continue;
103                 if (p->locks[i]) {
104                         btrfs_tree_unlock(p->nodes[i]);
105                         p->locks[i] = 0;
106                 }
107                 free_extent_buffer(p->nodes[i]);
108                 p->nodes[i] = NULL;
109         }
110 }
111
112 /*
113  * safely gets a reference on the root node of a tree.  A lock
114  * is not taken, so a concurrent writer may put a different node
115  * at the root of the tree.  See btrfs_lock_root_node for the
116  * looping required.
117  *
118  * The extent buffer returned by this has a reference taken, so
119  * it won't disappear.  It may stop being the root of the tree
120  * at any time because there are no locks held.
121  */
122 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
123 {
124         struct extent_buffer *eb;
125         spin_lock(&root->node_lock);
126         eb = root->node;
127         extent_buffer_get(eb);
128         spin_unlock(&root->node_lock);
129         return eb;
130 }
131
132 /* loop around taking references on and locking the root node of the
133  * tree until you end up with a lock on the root.  A locked buffer
134  * is returned, with a reference held.
135  */
136 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
137 {
138         struct extent_buffer *eb;
139
140         while (1) {
141                 eb = btrfs_root_node(root);
142                 btrfs_tree_lock(eb);
143
144                 spin_lock(&root->node_lock);
145                 if (eb == root->node) {
146                         spin_unlock(&root->node_lock);
147                         break;
148                 }
149                 spin_unlock(&root->node_lock);
150
151                 btrfs_tree_unlock(eb);
152                 free_extent_buffer(eb);
153         }
154         return eb;
155 }
156
157 /* cowonly root (everything not a reference counted cow subvolume), just get
158  * put onto a simple dirty list.  transaction.c walks this to make sure they
159  * get properly updated on disk.
160  */
161 static void add_root_to_dirty_list(struct btrfs_root *root)
162 {
163         if (root->track_dirty && list_empty(&root->dirty_list)) {
164                 list_add(&root->dirty_list,
165                          &root->fs_info->dirty_cowonly_roots);
166         }
167 }
168
169 /*
170  * used by snapshot creation to make a copy of a root for a tree with
171  * a given objectid.  The buffer with the new root node is returned in
172  * cow_ret, and this func returns zero on success or a negative error code.
173  */
174 int btrfs_copy_root(struct btrfs_trans_handle *trans,
175                       struct btrfs_root *root,
176                       struct extent_buffer *buf,
177                       struct extent_buffer **cow_ret, u64 new_root_objectid)
178 {
179         struct extent_buffer *cow;
180         u32 nritems;
181         int ret = 0;
182         int level;
183         struct btrfs_root *new_root;
184
185         new_root = kmalloc(sizeof(*new_root), GFP_NOFS);
186         if (!new_root)
187                 return -ENOMEM;
188
189         memcpy(new_root, root, sizeof(*new_root));
190         new_root->root_key.objectid = new_root_objectid;
191
192         WARN_ON(root->ref_cows && trans->transid !=
193                 root->fs_info->running_transaction->transid);
194         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
195
196         level = btrfs_header_level(buf);
197         nritems = btrfs_header_nritems(buf);
198
199         cow = btrfs_alloc_free_block(trans, new_root, buf->len, 0,
200                                      new_root_objectid, trans->transid,
201                                      level, buf->start, 0);
202         if (IS_ERR(cow)) {
203                 kfree(new_root);
204                 return PTR_ERR(cow);
205         }
206
207         copy_extent_buffer(cow, buf, 0, 0, cow->len);
208         btrfs_set_header_bytenr(cow, cow->start);
209         btrfs_set_header_generation(cow, trans->transid);
210         btrfs_set_header_owner(cow, new_root_objectid);
211         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
212
213         write_extent_buffer(cow, root->fs_info->fsid,
214                             (unsigned long)btrfs_header_fsid(cow),
215                             BTRFS_FSID_SIZE);
216
217         WARN_ON(btrfs_header_generation(buf) > trans->transid);
218         ret = btrfs_inc_ref(trans, new_root, buf, cow, NULL);
219         kfree(new_root);
220
221         if (ret)
222                 return ret;
223
224         btrfs_mark_buffer_dirty(cow);
225         *cow_ret = cow;
226         return 0;
227 }
228
229 /*
230  * does the dirty work in cow of a single block.  The parent block (if
231  * supplied) is updated to point to the new cow copy.  The new buffer is marked
232  * dirty and returned locked.  If you modify the block it needs to be marked
233  * dirty again.
234  *
235  * search_start -- an allocation hint for the new block
236  *
237  * empty_size -- a hint that you plan on doing more cow.  This is the size in
238  * bytes the allocator should try to find free next to the block it returns.
239  * This is just a hint and may be ignored by the allocator.
240  *
241  * prealloc_dest -- if you have already reserved a destination for the cow,
242  * this uses that block instead of allocating a new one.
243  * btrfs_alloc_reserved_extent is used to finish the allocation.
244  */
245 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
246                              struct btrfs_root *root,
247                              struct extent_buffer *buf,
248                              struct extent_buffer *parent, int parent_slot,
249                              struct extent_buffer **cow_ret,
250                              u64 search_start, u64 empty_size,
251                              u64 prealloc_dest)
252 {
253         u64 parent_start;
254         struct extent_buffer *cow;
255         u32 nritems;
256         int ret = 0;
257         int level;
258         int unlock_orig = 0;
259
260         if (*cow_ret == buf)
261                 unlock_orig = 1;
262
263         WARN_ON(!btrfs_tree_locked(buf));
264
265         if (parent)
266                 parent_start = parent->start;
267         else
268                 parent_start = 0;
269
270         WARN_ON(root->ref_cows && trans->transid !=
271                 root->fs_info->running_transaction->transid);
272         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
273
274         level = btrfs_header_level(buf);
275         nritems = btrfs_header_nritems(buf);
276
277         if (prealloc_dest) {
278                 struct btrfs_key ins;
279
280                 ins.objectid = prealloc_dest;
281                 ins.offset = buf->len;
282                 ins.type = BTRFS_EXTENT_ITEM_KEY;
283
284                 ret = btrfs_alloc_reserved_extent(trans, root, parent_start,
285                                                   root->root_key.objectid,
286                                                   trans->transid, level, &ins);
287                 BUG_ON(ret);
288                 cow = btrfs_init_new_buffer(trans, root, prealloc_dest,
289                                             buf->len);
290         } else {
291                 cow = btrfs_alloc_free_block(trans, root, buf->len,
292                                              parent_start,
293                                              root->root_key.objectid,
294                                              trans->transid, level,
295                                              search_start, empty_size);
296         }
297         if (IS_ERR(cow))
298                 return PTR_ERR(cow);
299
300         /* cow is set to blocking by btrfs_init_new_buffer */
301
302         copy_extent_buffer(cow, buf, 0, 0, cow->len);
303         btrfs_set_header_bytenr(cow, cow->start);
304         btrfs_set_header_generation(cow, trans->transid);
305         btrfs_set_header_owner(cow, root->root_key.objectid);
306         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN);
307
308         write_extent_buffer(cow, root->fs_info->fsid,
309                             (unsigned long)btrfs_header_fsid(cow),
310                             BTRFS_FSID_SIZE);
311
312         WARN_ON(btrfs_header_generation(buf) > trans->transid);
313         if (btrfs_header_generation(buf) != trans->transid) {
314                 u32 nr_extents;
315                 ret = btrfs_inc_ref(trans, root, buf, cow, &nr_extents);
316                 if (ret)
317                         return ret;
318
319                 ret = btrfs_cache_ref(trans, root, buf, nr_extents);
320                 WARN_ON(ret);
321         } else if (btrfs_header_owner(buf) == BTRFS_TREE_RELOC_OBJECTID) {
322                 /*
323                  * There are only two places that can drop reference to
324                  * tree blocks owned by living reloc trees, one is here,
325                  * the other place is btrfs_drop_subtree. In both places,
326                  * we check reference count while tree block is locked.
327                  * Furthermore, if reference count is one, it won't get
328                  * increased by someone else.
329                  */
330                 u32 refs;
331                 ret = btrfs_lookup_extent_ref(trans, root, buf->start,
332                                               buf->len, &refs);
333                 BUG_ON(ret);
334                 if (refs == 1) {
335                         ret = btrfs_update_ref(trans, root, buf, cow,
336                                                0, nritems);
337                         clean_tree_block(trans, root, buf);
338                 } else {
339                         ret = btrfs_inc_ref(trans, root, buf, cow, NULL);
340                 }
341                 BUG_ON(ret);
342         } else {
343                 ret = btrfs_update_ref(trans, root, buf, cow, 0, nritems);
344                 if (ret)
345                         return ret;
346                 clean_tree_block(trans, root, buf);
347         }
348
349         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
350                 ret = btrfs_reloc_tree_cache_ref(trans, root, cow, buf->start);
351                 WARN_ON(ret);
352         }
353
354         if (buf == root->node) {
355                 WARN_ON(parent && parent != buf);
356
357                 spin_lock(&root->node_lock);
358                 root->node = cow;
359                 extent_buffer_get(cow);
360                 spin_unlock(&root->node_lock);
361
362                 if (buf != root->commit_root) {
363                         btrfs_free_extent(trans, root, buf->start,
364                                           buf->len, buf->start,
365                                           root->root_key.objectid,
366                                           btrfs_header_generation(buf),
367                                           level, 1);
368                 }
369                 free_extent_buffer(buf);
370                 add_root_to_dirty_list(root);
371         } else {
372                 btrfs_set_node_blockptr(parent, parent_slot,
373                                         cow->start);
374                 WARN_ON(trans->transid == 0);
375                 btrfs_set_node_ptr_generation(parent, parent_slot,
376                                               trans->transid);
377                 btrfs_mark_buffer_dirty(parent);
378                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
379                 btrfs_free_extent(trans, root, buf->start, buf->len,
380                                   parent_start, btrfs_header_owner(parent),
381                                   btrfs_header_generation(parent), level, 1);
382         }
383         if (unlock_orig)
384                 btrfs_tree_unlock(buf);
385         free_extent_buffer(buf);
386         btrfs_mark_buffer_dirty(cow);
387         *cow_ret = cow;
388         return 0;
389 }
390
391 /*
392  * cows a single block, see __btrfs_cow_block for the real work.
393  * This version of it has extra checks so that a block isn't cow'd more than
394  * once per transaction, as long as it hasn't been written yet
395  */
396 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
397                     struct btrfs_root *root, struct extent_buffer *buf,
398                     struct extent_buffer *parent, int parent_slot,
399                     struct extent_buffer **cow_ret, u64 prealloc_dest)
400 {
401         u64 search_start;
402         int ret;
403
404         if (trans->transaction != root->fs_info->running_transaction) {
405                 printk(KERN_CRIT "trans %llu running %llu\n",
406                        (unsigned long long)trans->transid,
407                        (unsigned long long)
408                        root->fs_info->running_transaction->transid);
409                 WARN_ON(1);
410         }
411         if (trans->transid != root->fs_info->generation) {
412                 printk(KERN_CRIT "trans %llu running %llu\n",
413                        (unsigned long long)trans->transid,
414                        (unsigned long long)root->fs_info->generation);
415                 WARN_ON(1);
416         }
417
418         if (btrfs_header_generation(buf) == trans->transid &&
419             btrfs_header_owner(buf) == root->root_key.objectid &&
420             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
421                 *cow_ret = buf;
422                 WARN_ON(prealloc_dest);
423                 return 0;
424         }
425
426         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
427
428         if (parent)
429                 btrfs_set_lock_blocking(parent);
430         btrfs_set_lock_blocking(buf);
431
432         ret = __btrfs_cow_block(trans, root, buf, parent,
433                                  parent_slot, cow_ret, search_start, 0,
434                                  prealloc_dest);
435         return ret;
436 }
437
438 /*
439  * helper function for defrag to decide if two blocks pointed to by a
440  * node are actually close by
441  */
442 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
443 {
444         if (blocknr < other && other - (blocknr + blocksize) < 32768)
445                 return 1;
446         if (blocknr > other && blocknr - (other + blocksize) < 32768)
447                 return 1;
448         return 0;
449 }
450
451 /*
452  * compare two keys in a memcmp fashion
453  */
454 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
455 {
456         struct btrfs_key k1;
457
458         btrfs_disk_key_to_cpu(&k1, disk);
459
460         if (k1.objectid > k2->objectid)
461                 return 1;
462         if (k1.objectid < k2->objectid)
463                 return -1;
464         if (k1.type > k2->type)
465                 return 1;
466         if (k1.type < k2->type)
467                 return -1;
468         if (k1.offset > k2->offset)
469                 return 1;
470         if (k1.offset < k2->offset)
471                 return -1;
472         return 0;
473 }
474
475 /*
476  * same as comp_keys only with two btrfs_key's
477  */
478 static int comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
479 {
480         if (k1->objectid > k2->objectid)
481                 return 1;
482         if (k1->objectid < k2->objectid)
483                 return -1;
484         if (k1->type > k2->type)
485                 return 1;
486         if (k1->type < k2->type)
487                 return -1;
488         if (k1->offset > k2->offset)
489                 return 1;
490         if (k1->offset < k2->offset)
491                 return -1;
492         return 0;
493 }
494
495 /*
496  * this is used by the defrag code to go through all the
497  * leaves pointed to by a node and reallocate them so that
498  * disk order is close to key order
499  */
500 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
501                        struct btrfs_root *root, struct extent_buffer *parent,
502                        int start_slot, int cache_only, u64 *last_ret,
503                        struct btrfs_key *progress)
504 {
505         struct extent_buffer *cur;
506         u64 blocknr;
507         u64 gen;
508         u64 search_start = *last_ret;
509         u64 last_block = 0;
510         u64 other;
511         u32 parent_nritems;
512         int end_slot;
513         int i;
514         int err = 0;
515         int parent_level;
516         int uptodate;
517         u32 blocksize;
518         int progress_passed = 0;
519         struct btrfs_disk_key disk_key;
520
521         parent_level = btrfs_header_level(parent);
522         if (cache_only && parent_level != 1)
523                 return 0;
524
525         if (trans->transaction != root->fs_info->running_transaction)
526                 WARN_ON(1);
527         if (trans->transid != root->fs_info->generation)
528                 WARN_ON(1);
529
530         parent_nritems = btrfs_header_nritems(parent);
531         blocksize = btrfs_level_size(root, parent_level - 1);
532         end_slot = parent_nritems;
533
534         if (parent_nritems == 1)
535                 return 0;
536
537         btrfs_set_lock_blocking(parent);
538
539         for (i = start_slot; i < end_slot; i++) {
540                 int close = 1;
541
542                 if (!parent->map_token) {
543                         map_extent_buffer(parent,
544                                         btrfs_node_key_ptr_offset(i),
545                                         sizeof(struct btrfs_key_ptr),
546                                         &parent->map_token, &parent->kaddr,
547                                         &parent->map_start, &parent->map_len,
548                                         KM_USER1);
549                 }
550                 btrfs_node_key(parent, &disk_key, i);
551                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
552                         continue;
553
554                 progress_passed = 1;
555                 blocknr = btrfs_node_blockptr(parent, i);
556                 gen = btrfs_node_ptr_generation(parent, i);
557                 if (last_block == 0)
558                         last_block = blocknr;
559
560                 if (i > 0) {
561                         other = btrfs_node_blockptr(parent, i - 1);
562                         close = close_blocks(blocknr, other, blocksize);
563                 }
564                 if (!close && i < end_slot - 2) {
565                         other = btrfs_node_blockptr(parent, i + 1);
566                         close = close_blocks(blocknr, other, blocksize);
567                 }
568                 if (close) {
569                         last_block = blocknr;
570                         continue;
571                 }
572                 if (parent->map_token) {
573                         unmap_extent_buffer(parent, parent->map_token,
574                                             KM_USER1);
575                         parent->map_token = NULL;
576                 }
577
578                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
579                 if (cur)
580                         uptodate = btrfs_buffer_uptodate(cur, gen);
581                 else
582                         uptodate = 0;
583                 if (!cur || !uptodate) {
584                         if (cache_only) {
585                                 free_extent_buffer(cur);
586                                 continue;
587                         }
588                         if (!cur) {
589                                 cur = read_tree_block(root, blocknr,
590                                                          blocksize, gen);
591                         } else if (!uptodate) {
592                                 btrfs_read_buffer(cur, gen);
593                         }
594                 }
595                 if (search_start == 0)
596                         search_start = last_block;
597
598                 btrfs_tree_lock(cur);
599                 btrfs_set_lock_blocking(cur);
600                 err = __btrfs_cow_block(trans, root, cur, parent, i,
601                                         &cur, search_start,
602                                         min(16 * blocksize,
603                                             (end_slot - i) * blocksize), 0);
604                 if (err) {
605                         btrfs_tree_unlock(cur);
606                         free_extent_buffer(cur);
607                         break;
608                 }
609                 search_start = cur->start;
610                 last_block = cur->start;
611                 *last_ret = search_start;
612                 btrfs_tree_unlock(cur);
613                 free_extent_buffer(cur);
614         }
615         if (parent->map_token) {
616                 unmap_extent_buffer(parent, parent->map_token,
617                                     KM_USER1);
618                 parent->map_token = NULL;
619         }
620         return err;
621 }
622
623 /*
624  * The leaf data grows from end-to-front in the node.
625  * this returns the address of the start of the last item,
626  * which is the stop of the leaf data stack
627  */
628 static inline unsigned int leaf_data_end(struct btrfs_root *root,
629                                          struct extent_buffer *leaf)
630 {
631         u32 nr = btrfs_header_nritems(leaf);
632         if (nr == 0)
633                 return BTRFS_LEAF_DATA_SIZE(root);
634         return btrfs_item_offset_nr(leaf, nr - 1);
635 }
636
637 /*
638  * extra debugging checks to make sure all the items in a key are
639  * well formed and in the proper order
640  */
641 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
642                       int level)
643 {
644         struct extent_buffer *parent = NULL;
645         struct extent_buffer *node = path->nodes[level];
646         struct btrfs_disk_key parent_key;
647         struct btrfs_disk_key node_key;
648         int parent_slot;
649         int slot;
650         struct btrfs_key cpukey;
651         u32 nritems = btrfs_header_nritems(node);
652
653         if (path->nodes[level + 1])
654                 parent = path->nodes[level + 1];
655
656         slot = path->slots[level];
657         BUG_ON(nritems == 0);
658         if (parent) {
659                 parent_slot = path->slots[level + 1];
660                 btrfs_node_key(parent, &parent_key, parent_slot);
661                 btrfs_node_key(node, &node_key, 0);
662                 BUG_ON(memcmp(&parent_key, &node_key,
663                               sizeof(struct btrfs_disk_key)));
664                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
665                        btrfs_header_bytenr(node));
666         }
667         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
668         if (slot != 0) {
669                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
670                 btrfs_node_key(node, &node_key, slot);
671                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
672         }
673         if (slot < nritems - 1) {
674                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
675                 btrfs_node_key(node, &node_key, slot);
676                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
677         }
678         return 0;
679 }
680
681 /*
682  * extra checking to make sure all the items in a leaf are
683  * well formed and in the proper order
684  */
685 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
686                       int level)
687 {
688         struct extent_buffer *leaf = path->nodes[level];
689         struct extent_buffer *parent = NULL;
690         int parent_slot;
691         struct btrfs_key cpukey;
692         struct btrfs_disk_key parent_key;
693         struct btrfs_disk_key leaf_key;
694         int slot = path->slots[0];
695
696         u32 nritems = btrfs_header_nritems(leaf);
697
698         if (path->nodes[level + 1])
699                 parent = path->nodes[level + 1];
700
701         if (nritems == 0)
702                 return 0;
703
704         if (parent) {
705                 parent_slot = path->slots[level + 1];
706                 btrfs_node_key(parent, &parent_key, parent_slot);
707                 btrfs_item_key(leaf, &leaf_key, 0);
708
709                 BUG_ON(memcmp(&parent_key, &leaf_key,
710                        sizeof(struct btrfs_disk_key)));
711                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
712                        btrfs_header_bytenr(leaf));
713         }
714         if (slot != 0 && slot < nritems - 1) {
715                 btrfs_item_key(leaf, &leaf_key, slot);
716                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
717                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
718                         btrfs_print_leaf(root, leaf);
719                         printk(KERN_CRIT "slot %d offset bad key\n", slot);
720                         BUG_ON(1);
721                 }
722                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
723                        btrfs_item_end_nr(leaf, slot)) {
724                         btrfs_print_leaf(root, leaf);
725                         printk(KERN_CRIT "slot %d offset bad\n", slot);
726                         BUG_ON(1);
727                 }
728         }
729         if (slot < nritems - 1) {
730                 btrfs_item_key(leaf, &leaf_key, slot);
731                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
732                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
733                 if (btrfs_item_offset_nr(leaf, slot) !=
734                         btrfs_item_end_nr(leaf, slot + 1)) {
735                         btrfs_print_leaf(root, leaf);
736                         printk(KERN_CRIT "slot %d offset bad\n", slot);
737                         BUG_ON(1);
738                 }
739         }
740         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
741                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
742         return 0;
743 }
744
745 static noinline int check_block(struct btrfs_root *root,
746                                 struct btrfs_path *path, int level)
747 {
748         return 0;
749         if (level == 0)
750                 return check_leaf(root, path, level);
751         return check_node(root, path, level);
752 }
753
754 /*
755  * search for key in the extent_buffer.  The items start at offset p,
756  * and they are item_size apart.  There are 'max' items in p.
757  *
758  * the slot in the array is returned via slot, and it points to
759  * the place where you would insert key if it is not found in
760  * the array.
761  *
762  * slot may point to max if the key is bigger than all of the keys
763  */
764 static noinline int generic_bin_search(struct extent_buffer *eb,
765                                        unsigned long p,
766                                        int item_size, struct btrfs_key *key,
767                                        int max, int *slot)
768 {
769         int low = 0;
770         int high = max;
771         int mid;
772         int ret;
773         struct btrfs_disk_key *tmp = NULL;
774         struct btrfs_disk_key unaligned;
775         unsigned long offset;
776         char *map_token = NULL;
777         char *kaddr = NULL;
778         unsigned long map_start = 0;
779         unsigned long map_len = 0;
780         int err;
781
782         while (low < high) {
783                 mid = (low + high) / 2;
784                 offset = p + mid * item_size;
785
786                 if (!map_token || offset < map_start ||
787                     (offset + sizeof(struct btrfs_disk_key)) >
788                     map_start + map_len) {
789                         if (map_token) {
790                                 unmap_extent_buffer(eb, map_token, KM_USER0);
791                                 map_token = NULL;
792                         }
793
794                         err = map_private_extent_buffer(eb, offset,
795                                                 sizeof(struct btrfs_disk_key),
796                                                 &map_token, &kaddr,
797                                                 &map_start, &map_len, KM_USER0);
798
799                         if (!err) {
800                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
801                                                         map_start);
802                         } else {
803                                 read_extent_buffer(eb, &unaligned,
804                                                    offset, sizeof(unaligned));
805                                 tmp = &unaligned;
806                         }
807
808                 } else {
809                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
810                                                         map_start);
811                 }
812                 ret = comp_keys(tmp, key);
813
814                 if (ret < 0)
815                         low = mid + 1;
816                 else if (ret > 0)
817                         high = mid;
818                 else {
819                         *slot = mid;
820                         if (map_token)
821                                 unmap_extent_buffer(eb, map_token, KM_USER0);
822                         return 0;
823                 }
824         }
825         *slot = low;
826         if (map_token)
827                 unmap_extent_buffer(eb, map_token, KM_USER0);
828         return 1;
829 }
830
831 /*
832  * simple bin_search frontend that does the right thing for
833  * leaves vs nodes
834  */
835 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
836                       int level, int *slot)
837 {
838         if (level == 0) {
839                 return generic_bin_search(eb,
840                                           offsetof(struct btrfs_leaf, items),
841                                           sizeof(struct btrfs_item),
842                                           key, btrfs_header_nritems(eb),
843                                           slot);
844         } else {
845                 return generic_bin_search(eb,
846                                           offsetof(struct btrfs_node, ptrs),
847                                           sizeof(struct btrfs_key_ptr),
848                                           key, btrfs_header_nritems(eb),
849                                           slot);
850         }
851         return -1;
852 }
853
854 /* given a node and slot number, this reads the blocks it points to.  The
855  * extent buffer is returned with a reference taken (but unlocked).
856  * NULL is returned on error.
857  */
858 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
859                                    struct extent_buffer *parent, int slot)
860 {
861         int level = btrfs_header_level(parent);
862         if (slot < 0)
863                 return NULL;
864         if (slot >= btrfs_header_nritems(parent))
865                 return NULL;
866
867         BUG_ON(level == 0);
868
869         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
870                        btrfs_level_size(root, level - 1),
871                        btrfs_node_ptr_generation(parent, slot));
872 }
873
874 /*
875  * node level balancing, used to make sure nodes are in proper order for
876  * item deletion.  We balance from the top down, so we have to make sure
877  * that a deletion won't leave an node completely empty later on.
878  */
879 static noinline int balance_level(struct btrfs_trans_handle *trans,
880                          struct btrfs_root *root,
881                          struct btrfs_path *path, int level)
882 {
883         struct extent_buffer *right = NULL;
884         struct extent_buffer *mid;
885         struct extent_buffer *left = NULL;
886         struct extent_buffer *parent = NULL;
887         int ret = 0;
888         int wret;
889         int pslot;
890         int orig_slot = path->slots[level];
891         int err_on_enospc = 0;
892         u64 orig_ptr;
893
894         if (level == 0)
895                 return 0;
896
897         mid = path->nodes[level];
898
899         WARN_ON(!path->locks[level]);
900         WARN_ON(btrfs_header_generation(mid) != trans->transid);
901
902         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
903
904         if (level < BTRFS_MAX_LEVEL - 1)
905                 parent = path->nodes[level + 1];
906         pslot = path->slots[level + 1];
907
908         /*
909          * deal with the case where there is only one pointer in the root
910          * by promoting the node below to a root
911          */
912         if (!parent) {
913                 struct extent_buffer *child;
914
915                 if (btrfs_header_nritems(mid) != 1)
916                         return 0;
917
918                 /* promote the child to a root */
919                 child = read_node_slot(root, mid, 0);
920                 btrfs_tree_lock(child);
921                 btrfs_set_lock_blocking(child);
922                 BUG_ON(!child);
923                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child, 0);
924                 BUG_ON(ret);
925
926                 spin_lock(&root->node_lock);
927                 root->node = child;
928                 spin_unlock(&root->node_lock);
929
930                 ret = btrfs_update_extent_ref(trans, root, child->start,
931                                               mid->start, child->start,
932                                               root->root_key.objectid,
933                                               trans->transid, level - 1);
934                 BUG_ON(ret);
935
936                 add_root_to_dirty_list(root);
937                 btrfs_tree_unlock(child);
938
939                 path->locks[level] = 0;
940                 path->nodes[level] = NULL;
941                 clean_tree_block(trans, root, mid);
942                 btrfs_tree_unlock(mid);
943                 /* once for the path */
944                 free_extent_buffer(mid);
945                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
946                                         mid->start, root->root_key.objectid,
947                                         btrfs_header_generation(mid),
948                                         level, 1);
949                 /* once for the root ptr */
950                 free_extent_buffer(mid);
951                 return ret;
952         }
953         if (btrfs_header_nritems(mid) >
954             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
955                 return 0;
956
957         if (btrfs_header_nritems(mid) < 2)
958                 err_on_enospc = 1;
959
960         left = read_node_slot(root, parent, pslot - 1);
961         if (left) {
962                 btrfs_tree_lock(left);
963                 btrfs_set_lock_blocking(left);
964                 wret = btrfs_cow_block(trans, root, left,
965                                        parent, pslot - 1, &left, 0);
966                 if (wret) {
967                         ret = wret;
968                         goto enospc;
969                 }
970         }
971         right = read_node_slot(root, parent, pslot + 1);
972         if (right) {
973                 btrfs_tree_lock(right);
974                 btrfs_set_lock_blocking(right);
975                 wret = btrfs_cow_block(trans, root, right,
976                                        parent, pslot + 1, &right, 0);
977                 if (wret) {
978                         ret = wret;
979                         goto enospc;
980                 }
981         }
982
983         /* first, try to make some room in the middle buffer */
984         if (left) {
985                 orig_slot += btrfs_header_nritems(left);
986                 wret = push_node_left(trans, root, left, mid, 1);
987                 if (wret < 0)
988                         ret = wret;
989                 if (btrfs_header_nritems(mid) < 2)
990                         err_on_enospc = 1;
991         }
992
993         /*
994          * then try to empty the right most buffer into the middle
995          */
996         if (right) {
997                 wret = push_node_left(trans, root, mid, right, 1);
998                 if (wret < 0 && wret != -ENOSPC)
999                         ret = wret;
1000                 if (btrfs_header_nritems(right) == 0) {
1001                         u64 bytenr = right->start;
1002                         u64 generation = btrfs_header_generation(parent);
1003                         u32 blocksize = right->len;
1004
1005                         clean_tree_block(trans, root, right);
1006                         btrfs_tree_unlock(right);
1007                         free_extent_buffer(right);
1008                         right = NULL;
1009                         wret = del_ptr(trans, root, path, level + 1, pslot +
1010                                        1);
1011                         if (wret)
1012                                 ret = wret;
1013                         wret = btrfs_free_extent(trans, root, bytenr,
1014                                                  blocksize, parent->start,
1015                                                  btrfs_header_owner(parent),
1016                                                  generation, level, 1);
1017                         if (wret)
1018                                 ret = wret;
1019                 } else {
1020                         struct btrfs_disk_key right_key;
1021                         btrfs_node_key(right, &right_key, 0);
1022                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1023                         btrfs_mark_buffer_dirty(parent);
1024                 }
1025         }
1026         if (btrfs_header_nritems(mid) == 1) {
1027                 /*
1028                  * we're not allowed to leave a node with one item in the
1029                  * tree during a delete.  A deletion from lower in the tree
1030                  * could try to delete the only pointer in this node.
1031                  * So, pull some keys from the left.
1032                  * There has to be a left pointer at this point because
1033                  * otherwise we would have pulled some pointers from the
1034                  * right
1035                  */
1036                 BUG_ON(!left);
1037                 wret = balance_node_right(trans, root, mid, left);
1038                 if (wret < 0) {
1039                         ret = wret;
1040                         goto enospc;
1041                 }
1042                 if (wret == 1) {
1043                         wret = push_node_left(trans, root, left, mid, 1);
1044                         if (wret < 0)
1045                                 ret = wret;
1046                 }
1047                 BUG_ON(wret == 1);
1048         }
1049         if (btrfs_header_nritems(mid) == 0) {
1050                 /* we've managed to empty the middle node, drop it */
1051                 u64 root_gen = btrfs_header_generation(parent);
1052                 u64 bytenr = mid->start;
1053                 u32 blocksize = mid->len;
1054
1055                 clean_tree_block(trans, root, mid);
1056                 btrfs_tree_unlock(mid);
1057                 free_extent_buffer(mid);
1058                 mid = NULL;
1059                 wret = del_ptr(trans, root, path, level + 1, pslot);
1060                 if (wret)
1061                         ret = wret;
1062                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
1063                                          parent->start,
1064                                          btrfs_header_owner(parent),
1065                                          root_gen, level, 1);
1066                 if (wret)
1067                         ret = wret;
1068         } else {
1069                 /* update the parent key to reflect our changes */
1070                 struct btrfs_disk_key mid_key;
1071                 btrfs_node_key(mid, &mid_key, 0);
1072                 btrfs_set_node_key(parent, &mid_key, pslot);
1073                 btrfs_mark_buffer_dirty(parent);
1074         }
1075
1076         /* update the path */
1077         if (left) {
1078                 if (btrfs_header_nritems(left) > orig_slot) {
1079                         extent_buffer_get(left);
1080                         /* left was locked after cow */
1081                         path->nodes[level] = left;
1082                         path->slots[level + 1] -= 1;
1083                         path->slots[level] = orig_slot;
1084                         if (mid) {
1085                                 btrfs_tree_unlock(mid);
1086                                 free_extent_buffer(mid);
1087                         }
1088                 } else {
1089                         orig_slot -= btrfs_header_nritems(left);
1090                         path->slots[level] = orig_slot;
1091                 }
1092         }
1093         /* double check we haven't messed things up */
1094         check_block(root, path, level);
1095         if (orig_ptr !=
1096             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1097                 BUG();
1098 enospc:
1099         if (right) {
1100                 btrfs_tree_unlock(right);
1101                 free_extent_buffer(right);
1102         }
1103         if (left) {
1104                 if (path->nodes[level] != left)
1105                         btrfs_tree_unlock(left);
1106                 free_extent_buffer(left);
1107         }
1108         return ret;
1109 }
1110
1111 /* Node balancing for insertion.  Here we only split or push nodes around
1112  * when they are completely full.  This is also done top down, so we
1113  * have to be pessimistic.
1114  */
1115 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1116                                           struct btrfs_root *root,
1117                                           struct btrfs_path *path, int level)
1118 {
1119         struct extent_buffer *right = NULL;
1120         struct extent_buffer *mid;
1121         struct extent_buffer *left = NULL;
1122         struct extent_buffer *parent = NULL;
1123         int ret = 0;
1124         int wret;
1125         int pslot;
1126         int orig_slot = path->slots[level];
1127         u64 orig_ptr;
1128
1129         if (level == 0)
1130                 return 1;
1131
1132         mid = path->nodes[level];
1133         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1134         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1135
1136         if (level < BTRFS_MAX_LEVEL - 1)
1137                 parent = path->nodes[level + 1];
1138         pslot = path->slots[level + 1];
1139
1140         if (!parent)
1141                 return 1;
1142
1143         left = read_node_slot(root, parent, pslot - 1);
1144
1145         /* first, try to make some room in the middle buffer */
1146         if (left) {
1147                 u32 left_nr;
1148
1149                 btrfs_tree_lock(left);
1150                 btrfs_set_lock_blocking(left);
1151
1152                 left_nr = btrfs_header_nritems(left);
1153                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1154                         wret = 1;
1155                 } else {
1156                         ret = btrfs_cow_block(trans, root, left, parent,
1157                                               pslot - 1, &left, 0);
1158                         if (ret)
1159                                 wret = 1;
1160                         else {
1161                                 wret = push_node_left(trans, root,
1162                                                       left, mid, 0);
1163                         }
1164                 }
1165                 if (wret < 0)
1166                         ret = wret;
1167                 if (wret == 0) {
1168                         struct btrfs_disk_key disk_key;
1169                         orig_slot += left_nr;
1170                         btrfs_node_key(mid, &disk_key, 0);
1171                         btrfs_set_node_key(parent, &disk_key, pslot);
1172                         btrfs_mark_buffer_dirty(parent);
1173                         if (btrfs_header_nritems(left) > orig_slot) {
1174                                 path->nodes[level] = left;
1175                                 path->slots[level + 1] -= 1;
1176                                 path->slots[level] = orig_slot;
1177                                 btrfs_tree_unlock(mid);
1178                                 free_extent_buffer(mid);
1179                         } else {
1180                                 orig_slot -=
1181                                         btrfs_header_nritems(left);
1182                                 path->slots[level] = orig_slot;
1183                                 btrfs_tree_unlock(left);
1184                                 free_extent_buffer(left);
1185                         }
1186                         return 0;
1187                 }
1188                 btrfs_tree_unlock(left);
1189                 free_extent_buffer(left);
1190         }
1191         right = read_node_slot(root, parent, pslot + 1);
1192
1193         /*
1194          * then try to empty the right most buffer into the middle
1195          */
1196         if (right) {
1197                 u32 right_nr;
1198
1199                 btrfs_tree_lock(right);
1200                 btrfs_set_lock_blocking(right);
1201
1202                 right_nr = btrfs_header_nritems(right);
1203                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1204                         wret = 1;
1205                 } else {
1206                         ret = btrfs_cow_block(trans, root, right,
1207                                               parent, pslot + 1,
1208                                               &right, 0);
1209                         if (ret)
1210                                 wret = 1;
1211                         else {
1212                                 wret = balance_node_right(trans, root,
1213                                                           right, mid);
1214                         }
1215                 }
1216                 if (wret < 0)
1217                         ret = wret;
1218                 if (wret == 0) {
1219                         struct btrfs_disk_key disk_key;
1220
1221                         btrfs_node_key(right, &disk_key, 0);
1222                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1223                         btrfs_mark_buffer_dirty(parent);
1224
1225                         if (btrfs_header_nritems(mid) <= orig_slot) {
1226                                 path->nodes[level] = right;
1227                                 path->slots[level + 1] += 1;
1228                                 path->slots[level] = orig_slot -
1229                                         btrfs_header_nritems(mid);
1230                                 btrfs_tree_unlock(mid);
1231                                 free_extent_buffer(mid);
1232                         } else {
1233                                 btrfs_tree_unlock(right);
1234                                 free_extent_buffer(right);
1235                         }
1236                         return 0;
1237                 }
1238                 btrfs_tree_unlock(right);
1239                 free_extent_buffer(right);
1240         }
1241         return 1;
1242 }
1243
1244 /*
1245  * readahead one full node of leaves, finding things that are close
1246  * to the block in 'slot', and triggering ra on them.
1247  */
1248 static noinline void reada_for_search(struct btrfs_root *root,
1249                                       struct btrfs_path *path,
1250                                       int level, int slot, u64 objectid)
1251 {
1252         struct extent_buffer *node;
1253         struct btrfs_disk_key disk_key;
1254         u32 nritems;
1255         u64 search;
1256         u64 target;
1257         u64 nread = 0;
1258         int direction = path->reada;
1259         struct extent_buffer *eb;
1260         u32 nr;
1261         u32 blocksize;
1262         u32 nscan = 0;
1263
1264         if (level != 1)
1265                 return;
1266
1267         if (!path->nodes[level])
1268                 return;
1269
1270         node = path->nodes[level];
1271
1272         search = btrfs_node_blockptr(node, slot);
1273         blocksize = btrfs_level_size(root, level - 1);
1274         eb = btrfs_find_tree_block(root, search, blocksize);
1275         if (eb) {
1276                 free_extent_buffer(eb);
1277                 return;
1278         }
1279
1280         target = search;
1281
1282         nritems = btrfs_header_nritems(node);
1283         nr = slot;
1284         while (1) {
1285                 if (direction < 0) {
1286                         if (nr == 0)
1287                                 break;
1288                         nr--;
1289                 } else if (direction > 0) {
1290                         nr++;
1291                         if (nr >= nritems)
1292                                 break;
1293                 }
1294                 if (path->reada < 0 && objectid) {
1295                         btrfs_node_key(node, &disk_key, nr);
1296                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1297                                 break;
1298                 }
1299                 search = btrfs_node_blockptr(node, nr);
1300                 if ((search <= target && target - search <= 65536) ||
1301                     (search > target && search - target <= 65536)) {
1302                         readahead_tree_block(root, search, blocksize,
1303                                      btrfs_node_ptr_generation(node, nr));
1304                         nread += blocksize;
1305                 }
1306                 nscan++;
1307                 if ((nread > 65536 || nscan > 32))
1308                         break;
1309         }
1310 }
1311
1312 /*
1313  * returns -EAGAIN if it had to drop the path, or zero if everything was in
1314  * cache
1315  */
1316 static noinline int reada_for_balance(struct btrfs_root *root,
1317                                       struct btrfs_path *path, int level)
1318 {
1319         int slot;
1320         int nritems;
1321         struct extent_buffer *parent;
1322         struct extent_buffer *eb;
1323         u64 gen;
1324         u64 block1 = 0;
1325         u64 block2 = 0;
1326         int ret = 0;
1327         int blocksize;
1328
1329         parent = path->nodes[level - 1];
1330         if (!parent)
1331                 return 0;
1332
1333         nritems = btrfs_header_nritems(parent);
1334         slot = path->slots[level];
1335         blocksize = btrfs_level_size(root, level);
1336
1337         if (slot > 0) {
1338                 block1 = btrfs_node_blockptr(parent, slot - 1);
1339                 gen = btrfs_node_ptr_generation(parent, slot - 1);
1340                 eb = btrfs_find_tree_block(root, block1, blocksize);
1341                 if (eb && btrfs_buffer_uptodate(eb, gen))
1342                         block1 = 0;
1343                 free_extent_buffer(eb);
1344         }
1345         if (slot < nritems) {
1346                 block2 = btrfs_node_blockptr(parent, slot + 1);
1347                 gen = btrfs_node_ptr_generation(parent, slot + 1);
1348                 eb = btrfs_find_tree_block(root, block2, blocksize);
1349                 if (eb && btrfs_buffer_uptodate(eb, gen))
1350                         block2 = 0;
1351                 free_extent_buffer(eb);
1352         }
1353         if (block1 || block2) {
1354                 ret = -EAGAIN;
1355                 btrfs_release_path(root, path);
1356                 if (block1)
1357                         readahead_tree_block(root, block1, blocksize, 0);
1358                 if (block2)
1359                         readahead_tree_block(root, block2, blocksize, 0);
1360
1361                 if (block1) {
1362                         eb = read_tree_block(root, block1, blocksize, 0);
1363                         free_extent_buffer(eb);
1364                 }
1365                 if (block1) {
1366                         eb = read_tree_block(root, block2, blocksize, 0);
1367                         free_extent_buffer(eb);
1368                 }
1369         }
1370         return ret;
1371 }
1372
1373
1374 /*
1375  * when we walk down the tree, it is usually safe to unlock the higher layers
1376  * in the tree.  The exceptions are when our path goes through slot 0, because
1377  * operations on the tree might require changing key pointers higher up in the
1378  * tree.
1379  *
1380  * callers might also have set path->keep_locks, which tells this code to keep
1381  * the lock if the path points to the last slot in the block.  This is part of
1382  * walking through the tree, and selecting the next slot in the higher block.
1383  *
1384  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1385  * if lowest_unlock is 1, level 0 won't be unlocked
1386  */
1387 static noinline void unlock_up(struct btrfs_path *path, int level,
1388                                int lowest_unlock)
1389 {
1390         int i;
1391         int skip_level = level;
1392         int no_skips = 0;
1393         struct extent_buffer *t;
1394
1395         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1396                 if (!path->nodes[i])
1397                         break;
1398                 if (!path->locks[i])
1399                         break;
1400                 if (!no_skips && path->slots[i] == 0) {
1401                         skip_level = i + 1;
1402                         continue;
1403                 }
1404                 if (!no_skips && path->keep_locks) {
1405                         u32 nritems;
1406                         t = path->nodes[i];
1407                         nritems = btrfs_header_nritems(t);
1408                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1409                                 skip_level = i + 1;
1410                                 continue;
1411                         }
1412                 }
1413                 if (skip_level < i && i >= lowest_unlock)
1414                         no_skips = 1;
1415
1416                 t = path->nodes[i];
1417                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1418                         btrfs_tree_unlock(t);
1419                         path->locks[i] = 0;
1420                 }
1421         }
1422 }
1423
1424 /*
1425  * This releases any locks held in the path starting at level and
1426  * going all the way up to the root.
1427  *
1428  * btrfs_search_slot will keep the lock held on higher nodes in a few
1429  * corner cases, such as COW of the block at slot zero in the node.  This
1430  * ignores those rules, and it should only be called when there are no
1431  * more updates to be done higher up in the tree.
1432  */
1433 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
1434 {
1435         int i;
1436
1437         if (path->keep_locks || path->lowest_level)
1438                 return;
1439
1440         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1441                 if (!path->nodes[i])
1442                         continue;
1443                 if (!path->locks[i])
1444                         continue;
1445                 btrfs_tree_unlock(path->nodes[i]);
1446                 path->locks[i] = 0;
1447         }
1448 }
1449
1450 /*
1451  * look for key in the tree.  path is filled in with nodes along the way
1452  * if key is found, we return zero and you can find the item in the leaf
1453  * level of the path (level 0)
1454  *
1455  * If the key isn't found, the path points to the slot where it should
1456  * be inserted, and 1 is returned.  If there are other errors during the
1457  * search a negative error number is returned.
1458  *
1459  * if ins_len > 0, nodes and leaves will be split as we walk down the
1460  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1461  * possible)
1462  */
1463 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1464                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1465                       ins_len, int cow)
1466 {
1467         struct extent_buffer *b;
1468         struct extent_buffer *tmp;
1469         int slot;
1470         int ret;
1471         int level;
1472         int should_reada = p->reada;
1473         int lowest_unlock = 1;
1474         int blocksize;
1475         u8 lowest_level = 0;
1476         u64 blocknr;
1477         u64 gen;
1478         struct btrfs_key prealloc_block;
1479
1480         lowest_level = p->lowest_level;
1481         WARN_ON(lowest_level && ins_len > 0);
1482         WARN_ON(p->nodes[0] != NULL);
1483
1484         if (ins_len < 0)
1485                 lowest_unlock = 2;
1486
1487         prealloc_block.objectid = 0;
1488
1489 again:
1490         if (p->skip_locking)
1491                 b = btrfs_root_node(root);
1492         else
1493                 b = btrfs_lock_root_node(root);
1494
1495         while (b) {
1496                 level = btrfs_header_level(b);
1497
1498                 /*
1499                  * setup the path here so we can release it under lock
1500                  * contention with the cow code
1501                  */
1502                 p->nodes[level] = b;
1503                 if (!p->skip_locking)
1504                         p->locks[level] = 1;
1505
1506                 if (cow) {
1507                         int wret;
1508
1509                         /* is a cow on this block not required */
1510                         if (btrfs_header_generation(b) == trans->transid &&
1511                             btrfs_header_owner(b) == root->root_key.objectid &&
1512                             !btrfs_header_flag(b, BTRFS_HEADER_FLAG_WRITTEN)) {
1513                                 goto cow_done;
1514                         }
1515
1516                         /* ok, we have to cow, is our old prealloc the right
1517                          * size?
1518                          */
1519                         if (prealloc_block.objectid &&
1520                             prealloc_block.offset != b->len) {
1521                                 btrfs_release_path(root, p);
1522                                 btrfs_free_reserved_extent(root,
1523                                            prealloc_block.objectid,
1524                                            prealloc_block.offset);
1525                                 prealloc_block.objectid = 0;
1526                                 goto again;
1527                         }
1528
1529                         /*
1530                          * for higher level blocks, try not to allocate blocks
1531                          * with the block and the parent locks held.
1532                          */
1533                         if (level > 0 && !prealloc_block.objectid) {
1534                                 u32 size = b->len;
1535                                 u64 hint = b->start;
1536
1537                                 btrfs_release_path(root, p);
1538                                 ret = btrfs_reserve_extent(trans, root,
1539                                                            size, size, 0,
1540                                                            hint, (u64)-1,
1541                                                            &prealloc_block, 0);
1542                                 BUG_ON(ret);
1543                                 goto again;
1544                         }
1545
1546                         btrfs_set_path_blocking(p);
1547
1548                         wret = btrfs_cow_block(trans, root, b,
1549                                                p->nodes[level + 1],
1550                                                p->slots[level + 1],
1551                                                &b, prealloc_block.objectid);
1552                         prealloc_block.objectid = 0;
1553                         if (wret) {
1554                                 free_extent_buffer(b);
1555                                 ret = wret;
1556                                 goto done;
1557                         }
1558                 }
1559 cow_done:
1560                 BUG_ON(!cow && ins_len);
1561                 if (level != btrfs_header_level(b))
1562                         WARN_ON(1);
1563                 level = btrfs_header_level(b);
1564
1565                 p->nodes[level] = b;
1566                 if (!p->skip_locking)
1567                         p->locks[level] = 1;
1568
1569                 btrfs_clear_path_blocking(p);
1570
1571                 /*
1572                  * we have a lock on b and as long as we aren't changing
1573                  * the tree, there is no way to for the items in b to change.
1574                  * It is safe to drop the lock on our parent before we
1575                  * go through the expensive btree search on b.
1576                  *
1577                  * If cow is true, then we might be changing slot zero,
1578                  * which may require changing the parent.  So, we can't
1579                  * drop the lock until after we know which slot we're
1580                  * operating on.
1581                  */
1582                 if (!cow)
1583                         btrfs_unlock_up_safe(p, level + 1);
1584
1585                 ret = check_block(root, p, level);
1586                 if (ret) {
1587                         ret = -1;
1588                         goto done;
1589                 }
1590
1591                 ret = bin_search(b, key, level, &slot);
1592
1593                 if (level != 0) {
1594                         if (ret && slot > 0)
1595                                 slot -= 1;
1596                         p->slots[level] = slot;
1597                         if ((p->search_for_split || ins_len > 0) &&
1598                             btrfs_header_nritems(b) >=
1599                             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1600                                 int sret;
1601
1602                                 sret = reada_for_balance(root, p, level);
1603                                 if (sret)
1604                                         goto again;
1605
1606                                 btrfs_set_path_blocking(p);
1607                                 sret = split_node(trans, root, p, level);
1608                                 btrfs_clear_path_blocking(p);
1609
1610                                 BUG_ON(sret > 0);
1611                                 if (sret) {
1612                                         ret = sret;
1613                                         goto done;
1614                                 }
1615                                 b = p->nodes[level];
1616                                 slot = p->slots[level];
1617                         } else if (ins_len < 0 &&
1618                                    btrfs_header_nritems(b) <
1619                                    BTRFS_NODEPTRS_PER_BLOCK(root) / 4) {
1620                                 int sret;
1621
1622                                 sret = reada_for_balance(root, p, level);
1623                                 if (sret)
1624                                         goto again;
1625
1626                                 btrfs_set_path_blocking(p);
1627                                 sret = balance_level(trans, root, p, level);
1628                                 btrfs_clear_path_blocking(p);
1629
1630                                 if (sret) {
1631                                         ret = sret;
1632                                         goto done;
1633                                 }
1634                                 b = p->nodes[level];
1635                                 if (!b) {
1636                                         btrfs_release_path(NULL, p);
1637                                         goto again;
1638                                 }
1639                                 slot = p->slots[level];
1640                                 BUG_ON(btrfs_header_nritems(b) == 1);
1641                         }
1642                         unlock_up(p, level, lowest_unlock);
1643
1644                         /* this is only true while dropping a snapshot */
1645                         if (level == lowest_level) {
1646                                 ret = 0;
1647                                 goto done;
1648                         }
1649
1650                         blocknr = btrfs_node_blockptr(b, slot);
1651                         gen = btrfs_node_ptr_generation(b, slot);
1652                         blocksize = btrfs_level_size(root, level - 1);
1653
1654                         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1655                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1656                                 b = tmp;
1657                         } else {
1658                                 /*
1659                                  * reduce lock contention at high levels
1660                                  * of the btree by dropping locks before
1661                                  * we read.
1662                                  */
1663                                 if (level > 0) {
1664                                         btrfs_release_path(NULL, p);
1665                                         if (tmp)
1666                                                 free_extent_buffer(tmp);
1667                                         if (should_reada)
1668                                                 reada_for_search(root, p,
1669                                                                  level, slot,
1670                                                                  key->objectid);
1671
1672                                         tmp = read_tree_block(root, blocknr,
1673                                                          blocksize, gen);
1674                                         if (tmp)
1675                                                 free_extent_buffer(tmp);
1676                                         goto again;
1677                                 } else {
1678                                         btrfs_set_path_blocking(p);
1679                                         if (tmp)
1680                                                 free_extent_buffer(tmp);
1681                                         if (should_reada)
1682                                                 reada_for_search(root, p,
1683                                                                  level, slot,
1684                                                                  key->objectid);
1685                                         b = read_node_slot(root, b, slot);
1686                                 }
1687                         }
1688                         if (!p->skip_locking) {
1689                                 int lret;
1690
1691                                 btrfs_clear_path_blocking(p);
1692                                 lret = btrfs_try_spin_lock(b);
1693
1694                                 if (!lret) {
1695                                         btrfs_set_path_blocking(p);
1696                                         btrfs_tree_lock(b);
1697                                         btrfs_clear_path_blocking(p);
1698                                 }
1699                         }
1700                 } else {
1701                         p->slots[level] = slot;
1702                         if (ins_len > 0 &&
1703                             btrfs_leaf_free_space(root, b) < ins_len) {
1704                                 int sret;
1705
1706                                 btrfs_set_path_blocking(p);
1707                                 sret = split_leaf(trans, root, key,
1708                                                       p, ins_len, ret == 0);
1709                                 btrfs_clear_path_blocking(p);
1710
1711                                 BUG_ON(sret > 0);
1712                                 if (sret) {
1713                                         ret = sret;
1714                                         goto done;
1715                                 }
1716                         }
1717                         if (!p->search_for_split)
1718                                 unlock_up(p, level, lowest_unlock);
1719                         goto done;
1720                 }
1721         }
1722         ret = 1;
1723 done:
1724         /*
1725          * we don't really know what they plan on doing with the path
1726          * from here on, so for now just mark it as blocking
1727          */
1728         btrfs_set_path_blocking(p);
1729         if (prealloc_block.objectid) {
1730                 btrfs_free_reserved_extent(root,
1731                            prealloc_block.objectid,
1732                            prealloc_block.offset);
1733         }
1734         return ret;
1735 }
1736
1737 int btrfs_merge_path(struct btrfs_trans_handle *trans,
1738                      struct btrfs_root *root,
1739                      struct btrfs_key *node_keys,
1740                      u64 *nodes, int lowest_level)
1741 {
1742         struct extent_buffer *eb;
1743         struct extent_buffer *parent;
1744         struct btrfs_key key;
1745         u64 bytenr;
1746         u64 generation;
1747         u32 blocksize;
1748         int level;
1749         int slot;
1750         int key_match;
1751         int ret;
1752
1753         eb = btrfs_lock_root_node(root);
1754         ret = btrfs_cow_block(trans, root, eb, NULL, 0, &eb, 0);
1755         BUG_ON(ret);
1756
1757         btrfs_set_lock_blocking(eb);
1758
1759         parent = eb;
1760         while (1) {
1761                 level = btrfs_header_level(parent);
1762                 if (level == 0 || level <= lowest_level)
1763                         break;
1764
1765                 ret = bin_search(parent, &node_keys[lowest_level], level,
1766                                  &slot);
1767                 if (ret && slot > 0)
1768                         slot--;
1769
1770                 bytenr = btrfs_node_blockptr(parent, slot);
1771                 if (nodes[level - 1] == bytenr)
1772                         break;
1773
1774                 blocksize = btrfs_level_size(root, level - 1);
1775                 generation = btrfs_node_ptr_generation(parent, slot);
1776                 btrfs_node_key_to_cpu(eb, &key, slot);
1777                 key_match = !memcmp(&key, &node_keys[level - 1], sizeof(key));
1778
1779                 if (generation == trans->transid) {
1780                         eb = read_tree_block(root, bytenr, blocksize,
1781                                              generation);
1782                         btrfs_tree_lock(eb);
1783                         btrfs_set_lock_blocking(eb);
1784                 }
1785
1786                 /*
1787                  * if node keys match and node pointer hasn't been modified
1788                  * in the running transaction, we can merge the path. for
1789                  * blocks owened by reloc trees, the node pointer check is
1790                  * skipped, this is because these blocks are fully controlled
1791                  * by the space balance code, no one else can modify them.
1792                  */
1793                 if (!nodes[level - 1] || !key_match ||
1794                     (generation == trans->transid &&
1795                      btrfs_header_owner(eb) != BTRFS_TREE_RELOC_OBJECTID)) {
1796                         if (level == 1 || level == lowest_level + 1) {
1797                                 if (generation == trans->transid) {
1798                                         btrfs_tree_unlock(eb);
1799                                         free_extent_buffer(eb);
1800                                 }
1801                                 break;
1802                         }
1803
1804                         if (generation != trans->transid) {
1805                                 eb = read_tree_block(root, bytenr, blocksize,
1806                                                 generation);
1807                                 btrfs_tree_lock(eb);
1808                                 btrfs_set_lock_blocking(eb);
1809                         }
1810
1811                         ret = btrfs_cow_block(trans, root, eb, parent, slot,
1812                                               &eb, 0);
1813                         BUG_ON(ret);
1814
1815                         if (root->root_key.objectid ==
1816                             BTRFS_TREE_RELOC_OBJECTID) {
1817                                 if (!nodes[level - 1]) {
1818                                         nodes[level - 1] = eb->start;
1819                                         memcpy(&node_keys[level - 1], &key,
1820                                                sizeof(node_keys[0]));
1821                                 } else {
1822                                         WARN_ON(1);
1823                                 }
1824                         }
1825
1826                         btrfs_tree_unlock(parent);
1827                         free_extent_buffer(parent);
1828                         parent = eb;
1829                         continue;
1830                 }
1831
1832                 btrfs_set_node_blockptr(parent, slot, nodes[level - 1]);
1833                 btrfs_set_node_ptr_generation(parent, slot, trans->transid);
1834                 btrfs_mark_buffer_dirty(parent);
1835
1836                 ret = btrfs_inc_extent_ref(trans, root,
1837                                         nodes[level - 1],
1838                                         blocksize, parent->start,
1839                                         btrfs_header_owner(parent),
1840                                         btrfs_header_generation(parent),
1841                                         level - 1);
1842                 BUG_ON(ret);
1843
1844                 /*
1845                  * If the block was created in the running transaction,
1846                  * it's possible this is the last reference to it, so we
1847                  * should drop the subtree.
1848                  */
1849                 if (generation == trans->transid) {
1850                         ret = btrfs_drop_subtree(trans, root, eb, parent);
1851                         BUG_ON(ret);
1852                         btrfs_tree_unlock(eb);
1853                         free_extent_buffer(eb);
1854                 } else {
1855                         ret = btrfs_free_extent(trans, root, bytenr,
1856                                         blocksize, parent->start,
1857                                         btrfs_header_owner(parent),
1858                                         btrfs_header_generation(parent),
1859                                         level - 1, 1);
1860                         BUG_ON(ret);
1861                 }
1862                 break;
1863         }
1864         btrfs_tree_unlock(parent);
1865         free_extent_buffer(parent);
1866         return 0;
1867 }
1868
1869 /*
1870  * adjust the pointers going up the tree, starting at level
1871  * making sure the right key of each node is points to 'key'.
1872  * This is used after shifting pointers to the left, so it stops
1873  * fixing up pointers when a given leaf/node is not in slot 0 of the
1874  * higher levels
1875  *
1876  * If this fails to write a tree block, it returns -1, but continues
1877  * fixing up the blocks in ram so the tree is consistent.
1878  */
1879 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1880                           struct btrfs_root *root, struct btrfs_path *path,
1881                           struct btrfs_disk_key *key, int level)
1882 {
1883         int i;
1884         int ret = 0;
1885         struct extent_buffer *t;
1886
1887         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1888                 int tslot = path->slots[i];
1889                 if (!path->nodes[i])
1890                         break;
1891                 t = path->nodes[i];
1892                 btrfs_set_node_key(t, key, tslot);
1893                 btrfs_mark_buffer_dirty(path->nodes[i]);
1894                 if (tslot != 0)
1895                         break;
1896         }
1897         return ret;
1898 }
1899
1900 /*
1901  * update item key.
1902  *
1903  * This function isn't completely safe. It's the caller's responsibility
1904  * that the new key won't break the order
1905  */
1906 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1907                             struct btrfs_root *root, struct btrfs_path *path,
1908                             struct btrfs_key *new_key)
1909 {
1910         struct btrfs_disk_key disk_key;
1911         struct extent_buffer *eb;
1912         int slot;
1913
1914         eb = path->nodes[0];
1915         slot = path->slots[0];
1916         if (slot > 0) {
1917                 btrfs_item_key(eb, &disk_key, slot - 1);
1918                 if (comp_keys(&disk_key, new_key) >= 0)
1919                         return -1;
1920         }
1921         if (slot < btrfs_header_nritems(eb) - 1) {
1922                 btrfs_item_key(eb, &disk_key, slot + 1);
1923                 if (comp_keys(&disk_key, new_key) <= 0)
1924                         return -1;
1925         }
1926
1927         btrfs_cpu_key_to_disk(&disk_key, new_key);
1928         btrfs_set_item_key(eb, &disk_key, slot);
1929         btrfs_mark_buffer_dirty(eb);
1930         if (slot == 0)
1931                 fixup_low_keys(trans, root, path, &disk_key, 1);
1932         return 0;
1933 }
1934
1935 /*
1936  * try to push data from one node into the next node left in the
1937  * tree.
1938  *
1939  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1940  * error, and > 0 if there was no room in the left hand block.
1941  */
1942 static int push_node_left(struct btrfs_trans_handle *trans,
1943                           struct btrfs_root *root, struct extent_buffer *dst,
1944                           struct extent_buffer *src, int empty)
1945 {
1946         int push_items = 0;
1947         int src_nritems;
1948         int dst_nritems;
1949         int ret = 0;
1950
1951         src_nritems = btrfs_header_nritems(src);
1952         dst_nritems = btrfs_header_nritems(dst);
1953         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1954         WARN_ON(btrfs_header_generation(src) != trans->transid);
1955         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1956
1957         if (!empty && src_nritems <= 8)
1958                 return 1;
1959
1960         if (push_items <= 0)
1961                 return 1;
1962
1963         if (empty) {
1964                 push_items = min(src_nritems, push_items);
1965                 if (push_items < src_nritems) {
1966                         /* leave at least 8 pointers in the node if
1967                          * we aren't going to empty it
1968                          */
1969                         if (src_nritems - push_items < 8) {
1970                                 if (push_items <= 8)
1971                                         return 1;
1972                                 push_items -= 8;
1973                         }
1974                 }
1975         } else
1976                 push_items = min(src_nritems - 8, push_items);
1977
1978         copy_extent_buffer(dst, src,
1979                            btrfs_node_key_ptr_offset(dst_nritems),
1980                            btrfs_node_key_ptr_offset(0),
1981                            push_items * sizeof(struct btrfs_key_ptr));
1982
1983         if (push_items < src_nritems) {
1984                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1985                                       btrfs_node_key_ptr_offset(push_items),
1986                                       (src_nritems - push_items) *
1987                                       sizeof(struct btrfs_key_ptr));
1988         }
1989         btrfs_set_header_nritems(src, src_nritems - push_items);
1990         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1991         btrfs_mark_buffer_dirty(src);
1992         btrfs_mark_buffer_dirty(dst);
1993
1994         ret = btrfs_update_ref(trans, root, src, dst, dst_nritems, push_items);
1995         BUG_ON(ret);
1996
1997         return ret;
1998 }
1999
2000 /*
2001  * try to push data from one node into the next node right in the
2002  * tree.
2003  *
2004  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2005  * error, and > 0 if there was no room in the right hand block.
2006  *
2007  * this will  only push up to 1/2 the contents of the left node over
2008  */
2009 static int balance_node_right(struct btrfs_trans_handle *trans,
2010                               struct btrfs_root *root,
2011                               struct extent_buffer *dst,
2012                               struct extent_buffer *src)
2013 {
2014         int push_items = 0;
2015         int max_push;
2016         int src_nritems;
2017         int dst_nritems;
2018         int ret = 0;
2019
2020         WARN_ON(btrfs_header_generation(src) != trans->transid);
2021         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2022
2023         src_nritems = btrfs_header_nritems(src);
2024         dst_nritems = btrfs_header_nritems(dst);
2025         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2026         if (push_items <= 0)
2027                 return 1;
2028
2029         if (src_nritems < 4)
2030                 return 1;
2031
2032         max_push = src_nritems / 2 + 1;
2033         /* don't try to empty the node */
2034         if (max_push >= src_nritems)
2035                 return 1;
2036
2037         if (max_push < push_items)
2038                 push_items = max_push;
2039
2040         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2041                                       btrfs_node_key_ptr_offset(0),
2042                                       (dst_nritems) *
2043                                       sizeof(struct btrfs_key_ptr));
2044
2045         copy_extent_buffer(dst, src,
2046                            btrfs_node_key_ptr_offset(0),
2047                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2048                            push_items * sizeof(struct btrfs_key_ptr));
2049
2050         btrfs_set_header_nritems(src, src_nritems - push_items);
2051         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2052
2053         btrfs_mark_buffer_dirty(src);
2054         btrfs_mark_buffer_dirty(dst);
2055
2056         ret = btrfs_update_ref(trans, root, src, dst, 0, push_items);
2057         BUG_ON(ret);
2058
2059         return ret;
2060 }
2061
2062 /*
2063  * helper function to insert a new root level in the tree.
2064  * A new node is allocated, and a single item is inserted to
2065  * point to the existing root
2066  *
2067  * returns zero on success or < 0 on failure.
2068  */
2069 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2070                            struct btrfs_root *root,
2071                            struct btrfs_path *path, int level)
2072 {
2073         u64 lower_gen;
2074         struct extent_buffer *lower;
2075         struct extent_buffer *c;
2076         struct extent_buffer *old;
2077         struct btrfs_disk_key lower_key;
2078         int ret;
2079
2080         BUG_ON(path->nodes[level]);
2081         BUG_ON(path->nodes[level-1] != root->node);
2082
2083         lower = path->nodes[level-1];
2084         if (level == 1)
2085                 btrfs_item_key(lower, &lower_key, 0);
2086         else
2087                 btrfs_node_key(lower, &lower_key, 0);
2088
2089         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2090                                    root->root_key.objectid, trans->transid,
2091                                    level, root->node->start, 0);
2092         if (IS_ERR(c))
2093                 return PTR_ERR(c);
2094
2095         memset_extent_buffer(c, 0, 0, root->nodesize);
2096         btrfs_set_header_nritems(c, 1);
2097         btrfs_set_header_level(c, level);
2098         btrfs_set_header_bytenr(c, c->start);
2099         btrfs_set_header_generation(c, trans->transid);
2100         btrfs_set_header_owner(c, root->root_key.objectid);
2101
2102         write_extent_buffer(c, root->fs_info->fsid,
2103                             (unsigned long)btrfs_header_fsid(c),
2104                             BTRFS_FSID_SIZE);
2105
2106         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2107                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
2108                             BTRFS_UUID_SIZE);
2109
2110         btrfs_set_node_key(c, &lower_key, 0);
2111         btrfs_set_node_blockptr(c, 0, lower->start);
2112         lower_gen = btrfs_header_generation(lower);
2113         WARN_ON(lower_gen != trans->transid);
2114
2115         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2116
2117         btrfs_mark_buffer_dirty(c);
2118
2119         spin_lock(&root->node_lock);
2120         old = root->node;
2121         root->node = c;
2122         spin_unlock(&root->node_lock);
2123
2124         ret = btrfs_update_extent_ref(trans, root, lower->start,
2125                                       lower->start, c->start,
2126                                       root->root_key.objectid,
2127                                       trans->transid, level - 1);
2128         BUG_ON(ret);
2129
2130         /* the super has an extra ref to root->node */
2131         free_extent_buffer(old);
2132
2133         add_root_to_dirty_list(root);
2134         extent_buffer_get(c);
2135         path->nodes[level] = c;
2136         path->locks[level] = 1;
2137         path->slots[level] = 0;
2138         return 0;
2139 }
2140
2141 /*
2142  * worker function to insert a single pointer in a node.
2143  * the node should have enough room for the pointer already
2144  *
2145  * slot and level indicate where you want the key to go, and
2146  * blocknr is the block the key points to.
2147  *
2148  * returns zero on success and < 0 on any error
2149  */
2150 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
2151                       *root, struct btrfs_path *path, struct btrfs_disk_key
2152                       *key, u64 bytenr, int slot, int level)
2153 {
2154         struct extent_buffer *lower;
2155         int nritems;
2156
2157         BUG_ON(!path->nodes[level]);
2158         lower = path->nodes[level];
2159         nritems = btrfs_header_nritems(lower);
2160         if (slot > nritems)
2161                 BUG();
2162         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
2163                 BUG();
2164         if (slot != nritems) {
2165                 memmove_extent_buffer(lower,
2166                               btrfs_node_key_ptr_offset(slot + 1),
2167                               btrfs_node_key_ptr_offset(slot),
2168                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2169         }
2170         btrfs_set_node_key(lower, key, slot);
2171         btrfs_set_node_blockptr(lower, slot, bytenr);
2172         WARN_ON(trans->transid == 0);
2173         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2174         btrfs_set_header_nritems(lower, nritems + 1);
2175         btrfs_mark_buffer_dirty(lower);
2176         return 0;
2177 }
2178
2179 /*
2180  * split the node at the specified level in path in two.
2181  * The path is corrected to point to the appropriate node after the split
2182  *
2183  * Before splitting this tries to make some room in the node by pushing
2184  * left and right, if either one works, it returns right away.
2185  *
2186  * returns 0 on success and < 0 on failure
2187  */
2188 static noinline int split_node(struct btrfs_trans_handle *trans,
2189                                struct btrfs_root *root,
2190                                struct btrfs_path *path, int level)
2191 {
2192         struct extent_buffer *c;
2193         struct extent_buffer *split;
2194         struct btrfs_disk_key disk_key;
2195         int mid;
2196         int ret;
2197         int wret;
2198         u32 c_nritems;
2199
2200         c = path->nodes[level];
2201         WARN_ON(btrfs_header_generation(c) != trans->transid);
2202         if (c == root->node) {
2203                 /* trying to split the root, lets make a new one */
2204                 ret = insert_new_root(trans, root, path, level + 1);
2205                 if (ret)
2206                         return ret;
2207         } else {
2208                 ret = push_nodes_for_insert(trans, root, path, level);
2209                 c = path->nodes[level];
2210                 if (!ret && btrfs_header_nritems(c) <
2211                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
2212                         return 0;
2213                 if (ret < 0)
2214                         return ret;
2215         }
2216
2217         c_nritems = btrfs_header_nritems(c);
2218
2219         split = btrfs_alloc_free_block(trans, root, root->nodesize,
2220                                         path->nodes[level + 1]->start,
2221                                         root->root_key.objectid,
2222                                         trans->transid, level, c->start, 0);
2223         if (IS_ERR(split))
2224                 return PTR_ERR(split);
2225
2226         btrfs_set_header_flags(split, btrfs_header_flags(c));
2227         btrfs_set_header_level(split, btrfs_header_level(c));
2228         btrfs_set_header_bytenr(split, split->start);
2229         btrfs_set_header_generation(split, trans->transid);
2230         btrfs_set_header_owner(split, root->root_key.objectid);
2231         btrfs_set_header_flags(split, 0);
2232         write_extent_buffer(split, root->fs_info->fsid,
2233                             (unsigned long)btrfs_header_fsid(split),
2234                             BTRFS_FSID_SIZE);
2235         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
2236                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
2237                             BTRFS_UUID_SIZE);
2238
2239         mid = (c_nritems + 1) / 2;
2240
2241         copy_extent_buffer(split, c,
2242                            btrfs_node_key_ptr_offset(0),
2243                            btrfs_node_key_ptr_offset(mid),
2244                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2245         btrfs_set_header_nritems(split, c_nritems - mid);
2246         btrfs_set_header_nritems(c, mid);
2247         ret = 0;
2248
2249         btrfs_mark_buffer_dirty(c);
2250         btrfs_mark_buffer_dirty(split);
2251
2252         btrfs_node_key(split, &disk_key, 0);
2253         wret = insert_ptr(trans, root, path, &disk_key, split->start,
2254                           path->slots[level + 1] + 1,
2255                           level + 1);
2256         if (wret)
2257                 ret = wret;
2258
2259         ret = btrfs_update_ref(trans, root, c, split, 0, c_nritems - mid);
2260         BUG_ON(ret);
2261
2262         if (path->slots[level] >= mid) {
2263                 path->slots[level] -= mid;
2264                 btrfs_tree_unlock(c);
2265                 free_extent_buffer(c);
2266                 path->nodes[level] = split;
2267                 path->slots[level + 1] += 1;
2268         } else {
2269                 btrfs_tree_unlock(split);
2270                 free_extent_buffer(split);
2271         }
2272         return ret;
2273 }
2274
2275 /*
2276  * how many bytes are required to store the items in a leaf.  start
2277  * and nr indicate which items in the leaf to check.  This totals up the
2278  * space used both by the item structs and the item data
2279  */
2280 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2281 {
2282         int data_len;
2283         int nritems = btrfs_header_nritems(l);
2284         int end = min(nritems, start + nr) - 1;
2285
2286         if (!nr)
2287                 return 0;
2288         data_len = btrfs_item_end_nr(l, start);
2289         data_len = data_len - btrfs_item_offset_nr(l, end);
2290         data_len += sizeof(struct btrfs_item) * nr;
2291         WARN_ON(data_len < 0);
2292         return data_len;
2293 }
2294
2295 /*
2296  * The space between the end of the leaf items and
2297  * the start of the leaf data.  IOW, how much room
2298  * the leaf has left for both items and data
2299  */
2300 noinline int btrfs_leaf_free_space(struct btrfs_root *root,
2301                                    struct extent_buffer *leaf)
2302 {
2303         int nritems = btrfs_header_nritems(leaf);
2304         int ret;
2305         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
2306         if (ret < 0) {
2307                 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
2308                        "used %d nritems %d\n",
2309                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
2310                        leaf_space_used(leaf, 0, nritems), nritems);
2311         }
2312         return ret;
2313 }
2314
2315 /*
2316  * push some data in the path leaf to the right, trying to free up at
2317  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2318  *
2319  * returns 1 if the push failed because the other node didn't have enough
2320  * room, 0 if everything worked out and < 0 if there were major errors.
2321  */
2322 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2323                            *root, struct btrfs_path *path, int data_size,
2324                            int empty)
2325 {
2326         struct extent_buffer *left = path->nodes[0];
2327         struct extent_buffer *right;
2328         struct extent_buffer *upper;
2329         struct btrfs_disk_key disk_key;
2330         int slot;
2331         u32 i;
2332         int free_space;
2333         int push_space = 0;
2334         int push_items = 0;
2335         struct btrfs_item *item;
2336         u32 left_nritems;
2337         u32 nr;
2338         u32 right_nritems;
2339         u32 data_end;
2340         u32 this_item_size;
2341         int ret;
2342
2343         slot = path->slots[1];
2344         if (!path->nodes[1])
2345                 return 1;
2346
2347         upper = path->nodes[1];
2348         if (slot >= btrfs_header_nritems(upper) - 1)
2349                 return 1;
2350
2351         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
2352
2353         right = read_node_slot(root, upper, slot + 1);
2354         btrfs_tree_lock(right);
2355         btrfs_set_lock_blocking(right);
2356
2357         free_space = btrfs_leaf_free_space(root, right);
2358         if (free_space < data_size)
2359                 goto out_unlock;
2360
2361         /* cow and double check */
2362         ret = btrfs_cow_block(trans, root, right, upper,
2363                               slot + 1, &right, 0);
2364         if (ret)
2365                 goto out_unlock;
2366
2367         free_space = btrfs_leaf_free_space(root, right);
2368         if (free_space < data_size)
2369                 goto out_unlock;
2370
2371         left_nritems = btrfs_header_nritems(left);
2372         if (left_nritems == 0)
2373                 goto out_unlock;
2374
2375         if (empty)
2376                 nr = 0;
2377         else
2378                 nr = 1;
2379
2380         if (path->slots[0] >= left_nritems)
2381                 push_space += data_size;
2382
2383         i = left_nritems - 1;
2384         while (i >= nr) {
2385                 item = btrfs_item_nr(left, i);
2386
2387                 if (!empty && push_items > 0) {
2388                         if (path->slots[0] > i)
2389                                 break;
2390                         if (path->slots[0] == i) {
2391                                 int space = btrfs_leaf_free_space(root, left);
2392                                 if (space + push_space * 2 > free_space)
2393                                         break;
2394                         }
2395                 }
2396
2397                 if (path->slots[0] == i)
2398                         push_space += data_size;
2399
2400                 if (!left->map_token) {
2401                         map_extent_buffer(left, (unsigned long)item,
2402                                         sizeof(struct btrfs_item),
2403                                         &left->map_token, &left->kaddr,
2404                                         &left->map_start, &left->map_len,
2405                                         KM_USER1);
2406                 }
2407
2408                 this_item_size = btrfs_item_size(left, item);
2409                 if (this_item_size + sizeof(*item) + push_space > free_space)
2410                         break;
2411
2412                 push_items++;
2413                 push_space += this_item_size + sizeof(*item);
2414                 if (i == 0)
2415                         break;
2416                 i--;
2417         }
2418         if (left->map_token) {
2419                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2420                 left->map_token = NULL;
2421         }
2422
2423         if (push_items == 0)
2424                 goto out_unlock;
2425
2426         if (!empty && push_items == left_nritems)
2427                 WARN_ON(1);
2428
2429         /* push left to right */
2430         right_nritems = btrfs_header_nritems(right);
2431
2432         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2433         push_space -= leaf_data_end(root, left);
2434
2435         /* make room in the right data area */
2436         data_end = leaf_data_end(root, right);
2437         memmove_extent_buffer(right,
2438                               btrfs_leaf_data(right) + data_end - push_space,
2439                               btrfs_leaf_data(right) + data_end,
2440                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
2441
2442         /* copy from the left data area */
2443         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2444                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
2445                      btrfs_leaf_data(left) + leaf_data_end(root, left),
2446                      push_space);
2447
2448         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2449                               btrfs_item_nr_offset(0),
2450                               right_nritems * sizeof(struct btrfs_item));
2451
2452         /* copy the items from left to right */
2453         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2454                    btrfs_item_nr_offset(left_nritems - push_items),
2455                    push_items * sizeof(struct btrfs_item));
2456
2457         /* update the item pointers */
2458         right_nritems += push_items;
2459         btrfs_set_header_nritems(right, right_nritems);
2460         push_space = BTRFS_LEAF_DATA_SIZE(root);
2461         for (i = 0; i < right_nritems; i++) {
2462                 item = btrfs_item_nr(right, i);
2463                 if (!right->map_token) {
2464                         map_extent_buffer(right, (unsigned long)item,
2465                                         sizeof(struct btrfs_item),
2466                                         &right->map_token, &right->kaddr,
2467                                         &right->map_start, &right->map_len,
2468                                         KM_USER1);
2469                 }
2470                 push_space -= btrfs_item_size(right, item);
2471                 btrfs_set_item_offset(right, item, push_space);
2472         }
2473
2474         if (right->map_token) {
2475                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2476                 right->map_token = NULL;
2477         }
2478         left_nritems -= push_items;
2479         btrfs_set_header_nritems(left, left_nritems);
2480
2481         if (left_nritems)
2482                 btrfs_mark_buffer_dirty(left);
2483         btrfs_mark_buffer_dirty(right);
2484
2485         ret = btrfs_update_ref(trans, root, left, right, 0, push_items);
2486         BUG_ON(ret);
2487
2488         btrfs_item_key(right, &disk_key, 0);
2489         btrfs_set_node_key(upper, &disk_key, slot + 1);
2490         btrfs_mark_buffer_dirty(upper);
2491
2492         /* then fixup the leaf pointer in the path */
2493         if (path->slots[0] >= left_nritems) {
2494                 path->slots[0] -= left_nritems;
2495                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2496                         clean_tree_block(trans, root, path->nodes[0]);
2497                 btrfs_tree_unlock(path->nodes[0]);
2498                 free_extent_buffer(path->nodes[0]);
2499                 path->nodes[0] = right;
2500                 path->slots[1] += 1;
2501         } else {
2502                 btrfs_tree_unlock(right);
2503                 free_extent_buffer(right);
2504         }
2505         return 0;
2506
2507 out_unlock:
2508         btrfs_tree_unlock(right);
2509         free_extent_buffer(right);
2510         return 1;
2511 }
2512
2513 /*
2514  * push some data in the path leaf to the left, trying to free up at
2515  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2516  */
2517 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2518                           *root, struct btrfs_path *path, int data_size,
2519                           int empty)
2520 {
2521         struct btrfs_disk_key disk_key;
2522         struct extent_buffer *right = path->nodes[0];
2523         struct extent_buffer *left;
2524         int slot;
2525         int i;
2526         int free_space;
2527         int push_space = 0;
2528         int push_items = 0;
2529         struct btrfs_item *item;
2530         u32 old_left_nritems;
2531         u32 right_nritems;
2532         u32 nr;
2533         int ret = 0;
2534         int wret;
2535         u32 this_item_size;
2536         u32 old_left_item_size;
2537
2538         slot = path->slots[1];
2539         if (slot == 0)
2540                 return 1;
2541         if (!path->nodes[1])
2542                 return 1;
2543
2544         right_nritems = btrfs_header_nritems(right);
2545         if (right_nritems == 0)
2546                 return 1;
2547
2548         WARN_ON(!btrfs_tree_locked(path->nodes[1]));
2549
2550         left = read_node_slot(root, path->nodes[1], slot - 1);
2551         btrfs_tree_lock(left);
2552         btrfs_set_lock_blocking(left);
2553
2554         free_space = btrfs_leaf_free_space(root, left);
2555         if (free_space < data_size) {
2556                 ret = 1;
2557                 goto out;
2558         }
2559
2560         /* cow and double check */
2561         ret = btrfs_cow_block(trans, root, left,
2562                               path->nodes[1], slot - 1, &left, 0);
2563         if (ret) {
2564                 /* we hit -ENOSPC, but it isn't fatal here */
2565                 ret = 1;
2566                 goto out;
2567         }
2568
2569         free_space = btrfs_leaf_free_space(root, left);
2570         if (free_space < data_size) {
2571                 ret = 1;
2572                 goto out;
2573         }
2574
2575         if (empty)
2576                 nr = right_nritems;
2577         else
2578                 nr = right_nritems - 1;
2579
2580         for (i = 0; i < nr; i++) {
2581                 item = btrfs_item_nr(right, i);
2582                 if (!right->map_token) {
2583                         map_extent_buffer(right, (unsigned long)item,
2584                                         sizeof(struct btrfs_item),
2585                                         &right->map_token, &right->kaddr,
2586                                         &right->map_start, &right->map_len,
2587                                         KM_USER1);
2588                 }
2589
2590                 if (!empty && push_items > 0) {
2591                         if (path->slots[0] < i)
2592                                 break;
2593                         if (path->slots[0] == i) {
2594                                 int space = btrfs_leaf_free_space(root, right);
2595                                 if (space + push_space * 2 > free_space)
2596                                         break;
2597                         }
2598                 }
2599
2600                 if (path->slots[0] == i)
2601                         push_space += data_size;
2602
2603                 this_item_size = btrfs_item_size(right, item);
2604                 if (this_item_size + sizeof(*item) + push_space > free_space)
2605                         break;
2606
2607                 push_items++;
2608                 push_space += this_item_size + sizeof(*item);
2609         }
2610
2611         if (right->map_token) {
2612                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2613                 right->map_token = NULL;
2614         }
2615
2616         if (push_items == 0) {
2617                 ret = 1;
2618                 goto out;
2619         }
2620         if (!empty && push_items == btrfs_header_nritems(right))
2621                 WARN_ON(1);
2622
2623         /* push data from right to left */
2624         copy_extent_buffer(left, right,
2625                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2626                            btrfs_item_nr_offset(0),
2627                            push_items * sizeof(struct btrfs_item));
2628
2629         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2630                      btrfs_item_offset_nr(right, push_items - 1);
2631
2632         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2633                      leaf_data_end(root, left) - push_space,
2634                      btrfs_leaf_data(right) +
2635                      btrfs_item_offset_nr(right, push_items - 1),
2636                      push_space);
2637         old_left_nritems = btrfs_header_nritems(left);
2638         BUG_ON(old_left_nritems <= 0);
2639
2640         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2641         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2642                 u32 ioff;
2643
2644                 item = btrfs_item_nr(left, i);
2645                 if (!left->map_token) {
2646                         map_extent_buffer(left, (unsigned long)item,
2647                                         sizeof(struct btrfs_item),
2648                                         &left->map_token, &left->kaddr,
2649                                         &left->map_start, &left->map_len,
2650                                         KM_USER1);
2651                 }
2652
2653                 ioff = btrfs_item_offset(left, item);
2654                 btrfs_set_item_offset(left, item,
2655                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2656         }
2657         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2658         if (left->map_token) {
2659                 unmap_extent_buffer(left, left->map_token, KM_USER1);
2660                 left->map_token = NULL;
2661         }
2662
2663         /* fixup right node */
2664         if (push_items > right_nritems) {
2665                 printk(KERN_CRIT "push items %d nr %u\n", push_items,
2666                        right_nritems);
2667                 WARN_ON(1);
2668         }
2669
2670         if (push_items < right_nritems) {
2671                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2672                                                   leaf_data_end(root, right);
2673                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2674                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2675                                       btrfs_leaf_data(right) +
2676                                       leaf_data_end(root, right), push_space);
2677
2678                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2679                               btrfs_item_nr_offset(push_items),
2680                              (btrfs_header_nritems(right) - push_items) *
2681                              sizeof(struct btrfs_item));
2682         }
2683         right_nritems -= push_items;
2684         btrfs_set_header_nritems(right, right_nritems);
2685         push_space = BTRFS_LEAF_DATA_SIZE(root);
2686         for (i = 0; i < right_nritems; i++) {
2687                 item = btrfs_item_nr(right, i);
2688
2689                 if (!right->map_token) {
2690                         map_extent_buffer(right, (unsigned long)item,
2691                                         sizeof(struct btrfs_item),
2692                                         &right->map_token, &right->kaddr,
2693                                         &right->map_start, &right->map_len,
2694                                         KM_USER1);
2695                 }
2696
2697                 push_space = push_space - btrfs_item_size(right, item);
2698                 btrfs_set_item_offset(right, item, push_space);
2699         }
2700         if (right->map_token) {
2701                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2702                 right->map_token = NULL;
2703         }
2704
2705         btrfs_mark_buffer_dirty(left);
2706         if (right_nritems)
2707                 btrfs_mark_buffer_dirty(right);
2708
2709         ret = btrfs_update_ref(trans, root, right, left,
2710                                old_left_nritems, push_items);
2711         BUG_ON(ret);
2712
2713         btrfs_item_key(right, &disk_key, 0);
2714         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2715         if (wret)
2716                 ret = wret;
2717
2718         /* then fixup the leaf pointer in the path */
2719         if (path->slots[0] < push_items) {
2720                 path->slots[0] += old_left_nritems;
2721                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2722                         clean_tree_block(trans, root, path->nodes[0]);
2723                 btrfs_tree_unlock(path->nodes[0]);
2724                 free_extent_buffer(path->nodes[0]);
2725                 path->nodes[0] = left;
2726                 path->slots[1] -= 1;
2727         } else {
2728                 btrfs_tree_unlock(left);
2729                 free_extent_buffer(left);
2730                 path->slots[0] -= push_items;
2731         }
2732         BUG_ON(path->slots[0] < 0);
2733         return ret;
2734 out:
2735         btrfs_tree_unlock(left);
2736         free_extent_buffer(left);
2737         return ret;
2738 }
2739
2740 /*
2741  * split the path's leaf in two, making sure there is at least data_size
2742  * available for the resulting leaf level of the path.
2743  *
2744  * returns 0 if all went well and < 0 on failure.
2745  */
2746 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2747                                struct btrfs_root *root,
2748                                struct btrfs_key *ins_key,
2749                                struct btrfs_path *path, int data_size,
2750                                int extend)
2751 {
2752         struct extent_buffer *l;
2753         u32 nritems;
2754         int mid;
2755         int slot;
2756         struct extent_buffer *right;
2757         int data_copy_size;
2758         int rt_data_off;
2759         int i;
2760         int ret = 0;
2761         int wret;
2762         int double_split;
2763         int num_doubles = 0;
2764         struct btrfs_disk_key disk_key;
2765
2766         /* first try to make some room by pushing left and right */
2767         if (data_size && ins_key->type != BTRFS_DIR_ITEM_KEY) {
2768                 wret = push_leaf_right(trans, root, path, data_size, 0);
2769                 if (wret < 0)
2770                         return wret;
2771                 if (wret) {
2772                         wret = push_leaf_left(trans, root, path, data_size, 0);
2773                         if (wret < 0)
2774                                 return wret;
2775                 }
2776                 l = path->nodes[0];
2777
2778                 /* did the pushes work? */
2779                 if (btrfs_leaf_free_space(root, l) >= data_size)
2780                         return 0;
2781         }
2782
2783         if (!path->nodes[1]) {
2784                 ret = insert_new_root(trans, root, path, 1);
2785                 if (ret)
2786                         return ret;
2787         }
2788 again:
2789         double_split = 0;
2790         l = path->nodes[0];
2791         slot = path->slots[0];
2792         nritems = btrfs_header_nritems(l);
2793         mid = (nritems + 1) / 2;
2794
2795         right = btrfs_alloc_free_block(trans, root, root->leafsize,
2796                                         path->nodes[1]->start,
2797                                         root->root_key.objectid,
2798                                         trans->transid, 0, l->start, 0);
2799         if (IS_ERR(right)) {
2800                 BUG_ON(1);
2801                 return PTR_ERR(right);
2802         }
2803
2804         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2805         btrfs_set_header_bytenr(right, right->start);
2806         btrfs_set_header_generation(right, trans->transid);
2807         btrfs_set_header_owner(right, root->root_key.objectid);
2808         btrfs_set_header_level(right, 0);
2809         write_extent_buffer(right, root->fs_info->fsid,
2810                             (unsigned long)btrfs_header_fsid(right),
2811                             BTRFS_FSID_SIZE);
2812
2813         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2814                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2815                             BTRFS_UUID_SIZE);
2816         if (mid <= slot) {
2817                 if (nritems == 1 ||
2818                     leaf_space_used(l, mid, nritems - mid) + data_size >
2819                         BTRFS_LEAF_DATA_SIZE(root)) {
2820                         if (slot >= nritems) {
2821                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2822                                 btrfs_set_header_nritems(right, 0);
2823                                 wret = insert_ptr(trans, root, path,
2824                                                   &disk_key, right->start,
2825                                                   path->slots[1] + 1, 1);
2826                                 if (wret)
2827                                         ret = wret;
2828
2829                                 btrfs_tree_unlock(path->nodes[0]);
2830                                 free_extent_buffer(path->nodes[0]);
2831                                 path->nodes[0] = right;
2832                                 path->slots[0] = 0;
2833                                 path->slots[1] += 1;
2834                                 btrfs_mark_buffer_dirty(right);
2835                                 return ret;
2836                         }
2837                         mid = slot;
2838                         if (mid != nritems &&
2839                             leaf_space_used(l, mid, nritems - mid) +
2840                             data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2841                                 double_split = 1;
2842                         }
2843                 }
2844         } else {
2845                 if (leaf_space_used(l, 0, mid) + data_size >
2846                         BTRFS_LEAF_DATA_SIZE(root)) {
2847                         if (!extend && data_size && slot == 0) {
2848                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2849                                 btrfs_set_header_nritems(right, 0);
2850                                 wret = insert_ptr(trans, root, path,
2851                                                   &disk_key,
2852                                                   right->start,
2853                                                   path->slots[1], 1);
2854                                 if (wret)
2855                                         ret = wret;
2856                                 btrfs_tree_unlock(path->nodes[0]);
2857                                 free_extent_buffer(path->nodes[0]);
2858                                 path->nodes[0] = right;
2859                                 path->slots[0] = 0;
2860                                 if (path->slots[1] == 0) {
2861                                         wret = fixup_low_keys(trans, root,
2862                                                       path, &disk_key, 1);
2863                                         if (wret)
2864                                                 ret = wret;
2865                                 }
2866                                 btrfs_mark_buffer_dirty(right);
2867                                 return ret;
2868                         } else if ((extend || !data_size) && slot == 0) {
2869                                 mid = 1;
2870                         } else {
2871                                 mid = slot;
2872                                 if (mid != nritems &&
2873                                     leaf_space_used(l, mid, nritems - mid) +
2874                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2875                                         double_split = 1;
2876                                 }
2877                         }
2878                 }
2879         }
2880         nritems = nritems - mid;
2881         btrfs_set_header_nritems(right, nritems);
2882         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2883
2884         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2885                            btrfs_item_nr_offset(mid),
2886                            nritems * sizeof(struct btrfs_item));
2887
2888         copy_extent_buffer(right, l,
2889                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2890                      data_copy_size, btrfs_leaf_data(l) +
2891                      leaf_data_end(root, l), data_copy_size);
2892
2893         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2894                       btrfs_item_end_nr(l, mid);
2895
2896         for (i = 0; i < nritems; i++) {
2897                 struct btrfs_item *item = btrfs_item_nr(right, i);
2898                 u32 ioff;
2899
2900                 if (!right->map_token) {
2901                         map_extent_buffer(right, (unsigned long)item,
2902                                         sizeof(struct btrfs_item),
2903                                         &right->map_token, &right->kaddr,
2904                                         &right->map_start, &right->map_len,
2905                                         KM_USER1);
2906                 }
2907
2908                 ioff = btrfs_item_offset(right, item);
2909                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2910         }
2911
2912         if (right->map_token) {
2913                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2914                 right->map_token = NULL;
2915         }
2916
2917         btrfs_set_header_nritems(l, mid);
2918         ret = 0;
2919         btrfs_item_key(right, &disk_key, 0);
2920         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2921                           path->slots[1] + 1, 1);
2922         if (wret)
2923                 ret = wret;
2924
2925         btrfs_mark_buffer_dirty(right);
2926         btrfs_mark_buffer_dirty(l);
2927         BUG_ON(path->slots[0] != slot);
2928
2929         ret = btrfs_update_ref(trans, root, l, right, 0, nritems);
2930         BUG_ON(ret);
2931
2932         if (mid <= slot) {
2933                 btrfs_tree_unlock(path->nodes[0]);
2934                 free_extent_buffer(path->nodes[0]);
2935                 path->nodes[0] = right;
2936                 path->slots[0] -= mid;
2937                 path->slots[1] += 1;
2938         } else {
2939                 btrfs_tree_unlock(right);
2940                 free_extent_buffer(right);
2941         }
2942
2943         BUG_ON(path->slots[0] < 0);
2944
2945         if (double_split) {
2946                 BUG_ON(num_doubles != 0);
2947                 num_doubles++;
2948                 goto again;
2949         }
2950         return ret;
2951 }
2952
2953 /*
2954  * This function splits a single item into two items,
2955  * giving 'new_key' to the new item and splitting the
2956  * old one at split_offset (from the start of the item).
2957  *
2958  * The path may be released by this operation.  After
2959  * the split, the path is pointing to the old item.  The
2960  * new item is going to be in the same node as the old one.
2961  *
2962  * Note, the item being split must be smaller enough to live alone on
2963  * a tree block with room for one extra struct btrfs_item
2964  *
2965  * This allows us to split the item in place, keeping a lock on the
2966  * leaf the entire time.
2967  */
2968 int btrfs_split_item(struct btrfs_trans_handle *trans,
2969                      struct btrfs_root *root,
2970                      struct btrfs_path *path,
2971                      struct btrfs_key *new_key,
2972                      unsigned long split_offset)
2973 {
2974         u32 item_size;
2975         struct extent_buffer *leaf;
2976         struct btrfs_key orig_key;
2977         struct btrfs_item *item;
2978         struct btrfs_item *new_item;
2979         int ret = 0;
2980         int slot;
2981         u32 nritems;
2982         u32 orig_offset;
2983         struct btrfs_disk_key disk_key;
2984         char *buf;
2985
2986         leaf = path->nodes[0];
2987         btrfs_item_key_to_cpu(leaf, &orig_key, path->slots[0]);
2988         if (btrfs_leaf_free_space(root, leaf) >= sizeof(struct btrfs_item))
2989                 goto split;
2990
2991         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2992         btrfs_release_path(root, path);
2993
2994         path->search_for_split = 1;
2995         path->keep_locks = 1;
2996
2997         ret = btrfs_search_slot(trans, root, &orig_key, path, 0, 1);
2998         path->search_for_split = 0;
2999
3000         /* if our item isn't there or got smaller, return now */
3001         if (ret != 0 || item_size != btrfs_item_size_nr(path->nodes[0],
3002                                                         path->slots[0])) {
3003                 path->keep_locks = 0;
3004                 return -EAGAIN;
3005         }
3006
3007         ret = split_leaf(trans, root, &orig_key, path,
3008                          sizeof(struct btrfs_item), 1);
3009         path->keep_locks = 0;
3010         BUG_ON(ret);
3011
3012         /*
3013          * make sure any changes to the path from split_leaf leave it
3014          * in a blocking state
3015          */
3016         btrfs_set_path_blocking(path);
3017
3018         leaf = path->nodes[0];
3019         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3020
3021 split:
3022         item = btrfs_item_nr(leaf, path->slots[0]);
3023         orig_offset = btrfs_item_offset(leaf, item);
3024         item_size = btrfs_item_size(leaf, item);
3025
3026
3027         buf = kmalloc(item_size, GFP_NOFS);
3028         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3029                             path->slots[0]), item_size);
3030         slot = path->slots[0] + 1;
3031         leaf = path->nodes[0];
3032
3033         nritems = btrfs_header_nritems(leaf);
3034
3035         if (slot != nritems) {
3036                 /* shift the items */
3037                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3038                               btrfs_item_nr_offset(slot),
3039                               (nritems - slot) * sizeof(struct btrfs_item));
3040
3041         }
3042
3043         btrfs_cpu_key_to_disk(&disk_key, new_key);
3044         btrfs_set_item_key(leaf, &disk_key, slot);
3045
3046         new_item = btrfs_item_nr(leaf, slot);
3047
3048         btrfs_set_item_offset(leaf, new_item, orig_offset);
3049         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3050
3051         btrfs_set_item_offset(leaf, item,
3052                               orig_offset + item_size - split_offset);
3053         btrfs_set_item_size(leaf, item, split_offset);
3054
3055         btrfs_set_header_nritems(leaf, nritems + 1);
3056
3057         /* write the data for the start of the original item */
3058         write_extent_buffer(leaf, buf,
3059                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3060                             split_offset);
3061
3062         /* write the data for the new item */
3063         write_extent_buffer(leaf, buf + split_offset,
3064                             btrfs_item_ptr_offset(leaf, slot),
3065                             item_size - split_offset);
3066         btrfs_mark_buffer_dirty(leaf);
3067
3068         ret = 0;
3069         if (btrfs_leaf_free_space(root, leaf) < 0) {
3070                 btrfs_print_leaf(root, leaf);
3071                 BUG();
3072         }
3073         kfree(buf);
3074         return ret;
3075 }
3076
3077 /*
3078  * make the item pointed to by the path smaller.  new_size indicates
3079  * how small to make it, and from_end tells us if we just chop bytes
3080  * off the end of the item or if we shift the item to chop bytes off
3081  * the front.
3082  */
3083 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
3084                         struct btrfs_root *root,
3085                         struct btrfs_path *path,
3086                         u32 new_size, int from_end)
3087 {
3088         int ret = 0;
3089         int slot;
3090         int slot_orig;
3091         struct extent_buffer *leaf;
3092         struct btrfs_item *item;
3093         u32 nritems;
3094         unsigned int data_end;
3095         unsigned int old_data_start;
3096         unsigned int old_size;
3097         unsigned int size_diff;
3098         int i;
3099
3100         slot_orig = path->slots[0];
3101         leaf = path->nodes[0];
3102         slot = path->slots[0];
3103
3104         old_size = btrfs_item_size_nr(leaf, slot);
3105         if (old_size == new_size)
3106                 return 0;
3107
3108         nritems = btrfs_header_nritems(leaf);
3109         data_end = leaf_data_end(root, leaf);
3110
3111         old_data_start = btrfs_item_offset_nr(leaf, slot);
3112
3113         size_diff = old_size - new_size;
3114
3115         BUG_ON(slot < 0);
3116         BUG_ON(slot >= nritems);
3117
3118         /*
3119          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3120          */
3121         /* first correct the data pointers */
3122         for (i = slot; i < nritems; i++) {
3123                 u32 ioff;
3124                 item = btrfs_item_nr(leaf, i);
3125
3126                 if (!leaf->map_token) {
3127                         map_extent_buffer(leaf, (unsigned long)item,
3128                                         sizeof(struct btrfs_item),
3129                                         &leaf->map_token, &leaf->kaddr,
3130                                         &leaf->map_start, &leaf->map_len,
3131                                         KM_USER1);
3132                 }
3133
3134                 ioff = btrfs_item_offset(leaf, item);
3135                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
3136         }
3137
3138         if (leaf->map_token) {
3139                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3140                 leaf->map_token = NULL;
3141         }
3142
3143         /* shift the data */
3144         if (from_end) {
3145                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3146                               data_end + size_diff, btrfs_leaf_data(leaf) +
3147                               data_end, old_data_start + new_size - data_end);
3148         } else {
3149                 struct btrfs_disk_key disk_key;
3150                 u64 offset;
3151
3152                 btrfs_item_key(leaf, &disk_key, slot);
3153
3154                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3155                         unsigned long ptr;
3156                         struct btrfs_file_extent_item *fi;
3157
3158                         fi = btrfs_item_ptr(leaf, slot,
3159                                             struct btrfs_file_extent_item);
3160                         fi = (struct btrfs_file_extent_item *)(
3161                              (unsigned long)fi - size_diff);
3162
3163                         if (btrfs_file_extent_type(leaf, fi) ==
3164                             BTRFS_FILE_EXTENT_INLINE) {
3165                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3166                                 memmove_extent_buffer(leaf, ptr,
3167                                       (unsigned long)fi,
3168                                       offsetof(struct btrfs_file_extent_item,
3169                                                  disk_bytenr));
3170                         }
3171                 }
3172
3173                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3174                               data_end + size_diff, btrfs_leaf_data(leaf) +
3175                               data_end, old_data_start - data_end);
3176
3177                 offset = btrfs_disk_key_offset(&disk_key);
3178                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3179                 btrfs_set_item_key(leaf, &disk_key, slot);
3180                 if (slot == 0)
3181                         fixup_low_keys(trans, root, path, &disk_key, 1);
3182         }
3183
3184         item = btrfs_item_nr(leaf, slot);
3185         btrfs_set_item_size(leaf, item, new_size);
3186         btrfs_mark_buffer_dirty(leaf);
3187
3188         ret = 0;
3189         if (btrfs_leaf_free_space(root, leaf) < 0) {
3190                 btrfs_print_leaf(root, leaf);
3191                 BUG();
3192         }
3193         return ret;
3194 }
3195
3196 /*
3197  * make the item pointed to by the path bigger, data_size is the new size.
3198  */
3199 int btrfs_extend_item(struct btrfs_trans_handle *trans,
3200                       struct btrfs_root *root, struct btrfs_path *path,
3201                       u32 data_size)
3202 {
3203         int ret = 0;
3204         int slot;
3205         int slot_orig;
3206         struct extent_buffer *leaf;
3207         struct btrfs_item *item;
3208         u32 nritems;
3209         unsigned int data_end;
3210         unsigned int old_data;
3211         unsigned int old_size;
3212         int i;
3213
3214         slot_orig = path->slots[0];
3215         leaf = path->nodes[0];
3216
3217         nritems = btrfs_header_nritems(leaf);
3218         data_end = leaf_data_end(root, leaf);
3219
3220         if (btrfs_leaf_free_space(root, leaf) < data_size) {
3221                 btrfs_print_leaf(root, leaf);
3222                 BUG();
3223         }
3224         slot = path->slots[0];
3225         old_data = btrfs_item_end_nr(leaf, slot);
3226
3227         BUG_ON(slot < 0);
3228         if (slot >= nritems) {
3229                 btrfs_print_leaf(root, leaf);
3230                 printk(KERN_CRIT "slot %d too large, nritems %d\n",
3231                        slot, nritems);
3232                 BUG_ON(1);
3233         }
3234
3235         /*
3236          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3237          */
3238         /* first correct the data pointers */
3239         for (i = slot; i < nritems; i++) {
3240                 u32 ioff;
3241                 item = btrfs_item_nr(leaf, i);
3242
3243                 if (!leaf->map_token) {
3244                         map_extent_buffer(leaf, (unsigned long)item,
3245                                         sizeof(struct btrfs_item),
3246                                         &leaf->map_token, &leaf->kaddr,
3247                                         &leaf->map_start, &leaf->map_len,
3248                                         KM_USER1);
3249                 }
3250                 ioff = btrfs_item_offset(leaf, item);
3251                 btrfs_set_item_offset(leaf, item, ioff - data_size);
3252         }
3253
3254         if (leaf->map_token) {
3255                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3256                 leaf->map_token = NULL;
3257         }
3258
3259         /* shift the data */
3260         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3261                       data_end - data_size, btrfs_leaf_data(leaf) +
3262                       data_end, old_data - data_end);
3263
3264         data_end = old_data;
3265         old_size = btrfs_item_size_nr(leaf, slot);
3266         item = btrfs_item_nr(leaf, slot);
3267         btrfs_set_item_size(leaf, item, old_size + data_size);
3268         btrfs_mark_buffer_dirty(leaf);
3269
3270         ret = 0;
3271         if (btrfs_leaf_free_space(root, leaf) < 0) {
3272                 btrfs_print_leaf(root, leaf);
3273                 BUG();
3274         }
3275         return ret;
3276 }
3277
3278 /*
3279  * Given a key and some data, insert items into the tree.
3280  * This does all the path init required, making room in the tree if needed.
3281  * Returns the number of keys that were inserted.
3282  */
3283 int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
3284                             struct btrfs_root *root,
3285                             struct btrfs_path *path,
3286                             struct btrfs_key *cpu_key, u32 *data_size,
3287                             int nr)
3288 {
3289         struct extent_buffer *leaf;
3290         struct btrfs_item *item;
3291         int ret = 0;
3292         int slot;
3293         int i;
3294         u32 nritems;
3295         u32 total_data = 0;
3296         u32 total_size = 0;
3297         unsigned int data_end;
3298         struct btrfs_disk_key disk_key;
3299         struct btrfs_key found_key;
3300
3301         for (i = 0; i < nr; i++) {
3302                 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
3303                     BTRFS_LEAF_DATA_SIZE(root)) {
3304                         break;
3305                         nr = i;
3306                 }
3307                 total_data += data_size[i];
3308                 total_size += data_size[i] + sizeof(struct btrfs_item);
3309         }
3310         BUG_ON(nr == 0);
3311
3312         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3313         if (ret == 0)
3314                 return -EEXIST;
3315         if (ret < 0)
3316                 goto out;
3317
3318         leaf = path->nodes[0];
3319
3320         nritems = btrfs_header_nritems(leaf);
3321         data_end = leaf_data_end(root, leaf);
3322
3323         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3324                 for (i = nr; i >= 0; i--) {
3325                         total_data -= data_size[i];
3326                         total_size -= data_size[i] + sizeof(struct btrfs_item);
3327                         if (total_size < btrfs_leaf_free_space(root, leaf))
3328                                 break;
3329                 }
3330                 nr = i;
3331         }
3332
3333         slot = path->slots[0];
3334         BUG_ON(slot < 0);
3335
3336         if (slot != nritems) {
3337                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3338
3339                 item = btrfs_item_nr(leaf, slot);
3340                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3341
3342                 /* figure out how many keys we can insert in here */
3343                 total_data = data_size[0];
3344                 for (i = 1; i < nr; i++) {
3345                         if (comp_cpu_keys(&found_key, cpu_key + i) <= 0)
3346                                 break;
3347                         total_data += data_size[i];
3348                 }
3349                 nr = i;
3350
3351                 if (old_data < data_end) {
3352                         btrfs_print_leaf(root, leaf);
3353                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3354                                slot, old_data, data_end);
3355                         BUG_ON(1);
3356                 }
3357                 /*
3358                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3359                  */
3360                 /* first correct the data pointers */
3361                 WARN_ON(leaf->map_token);
3362                 for (i = slot; i < nritems; i++) {
3363                         u32 ioff;
3364
3365                         item = btrfs_item_nr(leaf, i);
3366                         if (!leaf->map_token) {
3367                                 map_extent_buffer(leaf, (unsigned long)item,
3368                                         sizeof(struct btrfs_item),
3369                                         &leaf->map_token, &leaf->kaddr,
3370                                         &leaf->map_start, &leaf->map_len,
3371                                         KM_USER1);
3372                         }
3373
3374                         ioff = btrfs_item_offset(leaf, item);
3375                         btrfs_set_item_offset(leaf, item, ioff - total_data);
3376                 }
3377                 if (leaf->map_token) {
3378                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3379                         leaf->map_token = NULL;
3380                 }
3381
3382                 /* shift the items */
3383                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3384                               btrfs_item_nr_offset(slot),
3385                               (nritems - slot) * sizeof(struct btrfs_item));
3386
3387                 /* shift the data */
3388                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3389                               data_end - total_data, btrfs_leaf_data(leaf) +
3390                               data_end, old_data - data_end);
3391                 data_end = old_data;
3392         } else {
3393                 /*
3394                  * this sucks but it has to be done, if we are inserting at
3395                  * the end of the leaf only insert 1 of the items, since we
3396                  * have no way of knowing whats on the next leaf and we'd have
3397                  * to drop our current locks to figure it out
3398                  */
3399                 nr = 1;
3400         }
3401
3402         /* setup the item for the new data */
3403         for (i = 0; i < nr; i++) {
3404                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3405                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3406                 item = btrfs_item_nr(leaf, slot + i);
3407                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3408                 data_end -= data_size[i];
3409                 btrfs_set_item_size(leaf, item, data_size[i]);
3410         }
3411         btrfs_set_header_nritems(leaf, nritems + nr);
3412         btrfs_mark_buffer_dirty(leaf);
3413
3414         ret = 0;
3415         if (slot == 0) {
3416                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3417                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3418         }
3419
3420         if (btrfs_leaf_free_space(root, leaf) < 0) {
3421                 btrfs_print_leaf(root, leaf);
3422                 BUG();
3423         }
3424 out:
3425         if (!ret)
3426                 ret = nr;
3427         return ret;
3428 }
3429
3430 /*
3431  * Given a key and some data, insert items into the tree.
3432  * This does all the path init required, making room in the tree if needed.
3433  */
3434 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3435                             struct btrfs_root *root,
3436                             struct btrfs_path *path,
3437                             struct btrfs_key *cpu_key, u32 *data_size,
3438                             int nr)
3439 {
3440         struct extent_buffer *leaf;
3441         struct btrfs_item *item;
3442         int ret = 0;
3443         int slot;
3444         int slot_orig;
3445         int i;
3446         u32 nritems;
3447         u32 total_size = 0;
3448         u32 total_data = 0;
3449         unsigned int data_end;
3450         struct btrfs_disk_key disk_key;
3451
3452         for (i = 0; i < nr; i++)
3453                 total_data += data_size[i];
3454
3455         total_size = total_data + (nr * sizeof(struct btrfs_item));
3456         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3457         if (ret == 0)
3458                 return -EEXIST;
3459         if (ret < 0)
3460                 goto out;
3461
3462         slot_orig = path->slots[0];
3463         leaf = path->nodes[0];
3464
3465         nritems = btrfs_header_nritems(leaf);
3466         data_end = leaf_data_end(root, leaf);
3467
3468         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3469                 btrfs_print_leaf(root, leaf);
3470                 printk(KERN_CRIT "not enough freespace need %u have %d\n",
3471                        total_size, btrfs_leaf_free_space(root, leaf));
3472                 BUG();
3473         }
3474
3475         slot = path->slots[0];
3476         BUG_ON(slot < 0);
3477
3478         if (slot != nritems) {
3479                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3480
3481                 if (old_data < data_end) {
3482                         btrfs_print_leaf(root, leaf);
3483                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3484                                slot, old_data, data_end);
3485                         BUG_ON(1);
3486                 }
3487                 /*
3488                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3489                  */
3490                 /* first correct the data pointers */
3491                 WARN_ON(leaf->map_token);
3492                 for (i = slot; i < nritems; i++) {
3493                         u32 ioff;
3494
3495                         item = btrfs_item_nr(leaf, i);
3496                         if (!leaf->map_token) {
3497                                 map_extent_buffer(leaf, (unsigned long)item,
3498                                         sizeof(struct btrfs_item),
3499                                         &leaf->map_token, &leaf->kaddr,
3500                                         &leaf->map_start, &leaf->map_len,
3501                                         KM_USER1);
3502                         }
3503
3504                         ioff = btrfs_item_offset(leaf, item);
3505                         btrfs_set_item_offset(leaf, item, ioff - total_data);
3506                 }
3507                 if (leaf->map_token) {
3508                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3509                         leaf->map_token = NULL;
3510                 }
3511
3512                 /* shift the items */
3513                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3514                               btrfs_item_nr_offset(slot),
3515                               (nritems - slot) * sizeof(struct btrfs_item));
3516
3517                 /* shift the data */
3518                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3519                               data_end - total_data, btrfs_leaf_data(leaf) +
3520                               data_end, old_data - data_end);
3521                 data_end = old_data;
3522         }
3523
3524         /* setup the item for the new data */
3525         for (i = 0; i < nr; i++) {
3526                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3527                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3528                 item = btrfs_item_nr(leaf, slot + i);
3529                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3530                 data_end -= data_size[i];
3531                 btrfs_set_item_size(leaf, item, data_size[i]);
3532         }
3533         btrfs_set_header_nritems(leaf, nritems + nr);
3534         btrfs_mark_buffer_dirty(leaf);
3535
3536         ret = 0;
3537         if (slot == 0) {
3538                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3539                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3540         }
3541
3542         if (btrfs_leaf_free_space(root, leaf) < 0) {
3543                 btrfs_print_leaf(root, leaf);
3544                 BUG();
3545         }
3546 out:
3547         btrfs_unlock_up_safe(path, 1);
3548         return ret;
3549 }
3550
3551 /*
3552  * Given a key and some data, insert an item into the tree.
3553  * This does all the path init required, making room in the tree if needed.
3554  */
3555 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
3556                       *root, struct btrfs_key *cpu_key, void *data, u32
3557                       data_size)
3558 {
3559         int ret = 0;
3560         struct btrfs_path *path;
3561         struct extent_buffer *leaf;
3562         unsigned long ptr;
3563
3564         path = btrfs_alloc_path();
3565         BUG_ON(!path);
3566         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3567         if (!ret) {
3568                 leaf = path->nodes[0];
3569                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3570                 write_extent_buffer(leaf, data, ptr, data_size);
3571                 btrfs_mark_buffer_dirty(leaf);
3572         }
3573         btrfs_free_path(path);
3574         return ret;
3575 }
3576
3577 /*
3578  * delete the pointer from a given node.
3579  *
3580  * the tree should have been previously balanced so the deletion does not
3581  * empty a node.
3582  */
3583 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3584                    struct btrfs_path *path, int level, int slot)
3585 {
3586         struct extent_buffer *parent = path->nodes[level];
3587         u32 nritems;
3588         int ret = 0;
3589         int wret;
3590
3591         nritems = btrfs_header_nritems(parent);
3592         if (slot != nritems - 1) {
3593                 memmove_extent_buffer(parent,
3594                               btrfs_node_key_ptr_offset(slot),
3595                               btrfs_node_key_ptr_offset(slot + 1),
3596                               sizeof(struct btrfs_key_ptr) *
3597                               (nritems - slot - 1));
3598         }
3599         nritems--;
3600         btrfs_set_header_nritems(parent, nritems);
3601         if (nritems == 0 && parent == root->node) {
3602                 BUG_ON(btrfs_header_level(root->node) != 1);
3603                 /* just turn the root into a leaf and break */
3604                 btrfs_set_header_level(root->node, 0);
3605         } else if (slot == 0) {
3606                 struct btrfs_disk_key disk_key;
3607
3608                 btrfs_node_key(parent, &disk_key, 0);
3609                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
3610                 if (wret)
3611                         ret = wret;
3612         }
3613         btrfs_mark_buffer_dirty(parent);
3614         return ret;
3615 }
3616
3617 /*
3618  * a helper function to delete the leaf pointed to by path->slots[1] and
3619  * path->nodes[1].  bytenr is the node block pointer, but since the callers
3620  * already know it, it is faster to have them pass it down than to
3621  * read it out of the node again.
3622  *
3623  * This deletes the pointer in path->nodes[1] and frees the leaf
3624  * block extent.  zero is returned if it all worked out, < 0 otherwise.
3625  *
3626  * The path must have already been setup for deleting the leaf, including
3627  * all the proper balancing.  path->nodes[1] must be locked.
3628  */
3629 noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
3630                             struct btrfs_root *root,
3631                             struct btrfs_path *path, u64 bytenr)
3632 {
3633         int ret;
3634         u64 root_gen = btrfs_header_generation(path->nodes[1]);
3635         u64 parent_start = path->nodes[1]->start;
3636         u64 parent_owner = btrfs_header_owner(path->nodes[1]);
3637
3638         ret = del_ptr(trans, root, path, 1, path->slots[1]);
3639         if (ret)
3640                 return ret;
3641
3642         /*
3643          * btrfs_free_extent is expensive, we want to make sure we
3644          * aren't holding any locks when we call it
3645          */
3646         btrfs_unlock_up_safe(path, 0);
3647
3648         ret = btrfs_free_extent(trans, root, bytenr,
3649                                 btrfs_level_size(root, 0),
3650                                 parent_start, parent_owner,
3651                                 root_gen, 0, 1);
3652         return ret;
3653 }
3654 /*
3655  * delete the item at the leaf level in path.  If that empties
3656  * the leaf, remove it from the tree
3657  */
3658 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3659                     struct btrfs_path *path, int slot, int nr)
3660 {
3661         struct extent_buffer *leaf;
3662         struct btrfs_item *item;
3663         int last_off;
3664         int dsize = 0;
3665         int ret = 0;
3666         int wret;
3667         int i;
3668         u32 nritems;
3669
3670         leaf = path->nodes[0];
3671         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
3672
3673         for (i = 0; i < nr; i++)
3674                 dsize += btrfs_item_size_nr(leaf, slot + i);
3675
3676         nritems = btrfs_header_nritems(leaf);
3677
3678         if (slot + nr != nritems) {
3679                 int data_end = leaf_data_end(root, leaf);
3680
3681                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3682                               data_end + dsize,
3683                               btrfs_leaf_data(leaf) + data_end,
3684                               last_off - data_end);
3685
3686                 for (i = slot + nr; i < nritems; i++) {
3687                         u32 ioff;
3688
3689                         item = btrfs_item_nr(leaf, i);
3690                         if (!leaf->map_token) {
3691                                 map_extent_buffer(leaf, (unsigned long)item,
3692                                         sizeof(struct btrfs_item),
3693                                         &leaf->map_token, &leaf->kaddr,
3694                                         &leaf->map_start, &leaf->map_len,
3695                                         KM_USER1);
3696                         }
3697                         ioff = btrfs_item_offset(leaf, item);
3698                         btrfs_set_item_offset(leaf, item, ioff + dsize);
3699                 }
3700
3701                 if (leaf->map_token) {
3702                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
3703                         leaf->map_token = NULL;
3704                 }
3705
3706                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
3707                               btrfs_item_nr_offset(slot + nr),
3708                               sizeof(struct btrfs_item) *
3709                               (nritems - slot - nr));
3710         }
3711         btrfs_set_header_nritems(leaf, nritems - nr);
3712         nritems -= nr;
3713
3714         /* delete the leaf if we've emptied it */
3715         if (nritems == 0) {
3716                 if (leaf == root->node) {
3717                         btrfs_set_header_level(leaf, 0);
3718                 } else {
3719                         ret = btrfs_del_leaf(trans, root, path, leaf->start);
3720                         BUG_ON(ret);
3721                 }
3722         } else {
3723                 int used = leaf_space_used(leaf, 0, nritems);
3724                 if (slot == 0) {
3725                         struct btrfs_disk_key disk_key;
3726
3727                         btrfs_item_key(leaf, &disk_key, 0);
3728                         wret = fixup_low_keys(trans, root, path,
3729                                               &disk_key, 1);
3730                         if (wret)
3731                                 ret = wret;
3732                 }
3733
3734                 /* delete the leaf if it is mostly empty */
3735                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 4) {
3736                         /* push_leaf_left fixes the path.
3737                          * make sure the path still points to our leaf
3738                          * for possible call to del_ptr below
3739                          */
3740                         slot = path->slots[1];
3741                         extent_buffer_get(leaf);
3742
3743                         wret = push_leaf_left(trans, root, path, 1, 1);
3744                         if (wret < 0 && wret != -ENOSPC)
3745                                 ret = wret;
3746
3747                         if (path->nodes[0] == leaf &&
3748                             btrfs_header_nritems(leaf)) {
3749                                 wret = push_leaf_right(trans, root, path, 1, 1);
3750                                 if (wret < 0 && wret != -ENOSPC)
3751                                         ret = wret;
3752                         }
3753
3754                         if (btrfs_header_nritems(leaf) == 0) {
3755                                 path->slots[1] = slot;
3756                                 ret = btrfs_del_leaf(trans, root, path,
3757                                                      leaf->start);
3758                                 BUG_ON(ret);
3759                                 free_extent_buffer(leaf);
3760                         } else {
3761                                 /* if we're still in the path, make sure
3762                                  * we're dirty.  Otherwise, one of the
3763                                  * push_leaf functions must have already
3764                                  * dirtied this buffer
3765                                  */
3766                                 if (path->nodes[0] == leaf)
3767                                         btrfs_mark_buffer_dirty(leaf);
3768                                 free_extent_buffer(leaf);
3769                         }
3770                 } else {
3771                         btrfs_mark_buffer_dirty(leaf);
3772                 }
3773         }
3774         return ret;
3775 }
3776
3777 /*
3778  * search the tree again to find a leaf with lesser keys
3779  * returns 0 if it found something or 1 if there are no lesser leaves.
3780  * returns < 0 on io errors.
3781  *
3782  * This may release the path, and so you may lose any locks held at the
3783  * time you call it.
3784  */
3785 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
3786 {
3787         struct btrfs_key key;
3788         struct btrfs_disk_key found_key;
3789         int ret;
3790
3791         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
3792
3793         if (key.offset > 0)
3794                 key.offset--;
3795         else if (key.type > 0)
3796                 key.type--;
3797         else if (key.objectid > 0)
3798                 key.objectid--;
3799         else
3800                 return 1;
3801
3802         btrfs_release_path(root, path);
3803         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3804         if (ret < 0)
3805                 return ret;
3806         btrfs_item_key(path->nodes[0], &found_key, 0);
3807         ret = comp_keys(&found_key, &key);
3808         if (ret < 0)
3809                 return 0;
3810         return 1;
3811 }
3812
3813 /*
3814  * A helper function to walk down the tree starting at min_key, and looking
3815  * for nodes or leaves that are either in cache or have a minimum
3816  * transaction id.  This is used by the btree defrag code, and tree logging
3817  *
3818  * This does not cow, but it does stuff the starting key it finds back
3819  * into min_key, so you can call btrfs_search_slot with cow=1 on the
3820  * key and get a writable path.
3821  *
3822  * This does lock as it descends, and path->keep_locks should be set
3823  * to 1 by the caller.
3824  *
3825  * This honors path->lowest_level to prevent descent past a given level
3826  * of the tree.
3827  *
3828  * min_trans indicates the oldest transaction that you are interested
3829  * in walking through.  Any nodes or leaves older than min_trans are
3830  * skipped over (without reading them).
3831  *
3832  * returns zero if something useful was found, < 0 on error and 1 if there
3833  * was nothing in the tree that matched the search criteria.
3834  */
3835 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
3836                          struct btrfs_key *max_key,
3837                          struct btrfs_path *path, int cache_only,
3838                          u64 min_trans)
3839 {
3840         struct extent_buffer *cur;
3841         struct btrfs_key found_key;
3842         int slot;
3843         int sret;
3844         u32 nritems;
3845         int level;
3846         int ret = 1;
3847
3848         WARN_ON(!path->keep_locks);
3849 again:
3850         cur = btrfs_lock_root_node(root);
3851         level = btrfs_header_level(cur);
3852         WARN_ON(path->nodes[level]);
3853         path->nodes[level] = cur;
3854         path->locks[level] = 1;
3855
3856         if (btrfs_header_generation(cur) < min_trans) {
3857                 ret = 1;
3858                 goto out;
3859         }
3860         while (1) {
3861                 nritems = btrfs_header_nritems(cur);
3862                 level = btrfs_header_level(cur);
3863                 sret = bin_search(cur, min_key, level, &slot);
3864
3865                 /* at the lowest level, we're done, setup the path and exit */
3866                 if (level == path->lowest_level) {
3867                         if (slot >= nritems)
3868                                 goto find_next_key;
3869                         ret = 0;
3870                         path->slots[level] = slot;
3871                         btrfs_item_key_to_cpu(cur, &found_key, slot);
3872                         goto out;
3873                 }
3874                 if (sret && slot > 0)
3875                         slot--;
3876                 /*
3877                  * check this node pointer against the cache_only and
3878                  * min_trans parameters.  If it isn't in cache or is too
3879                  * old, skip to the next one.
3880                  */
3881                 while (slot < nritems) {
3882                         u64 blockptr;
3883                         u64 gen;
3884                         struct extent_buffer *tmp;
3885                         struct btrfs_disk_key disk_key;
3886
3887                         blockptr = btrfs_node_blockptr(cur, slot);
3888                         gen = btrfs_node_ptr_generation(cur, slot);
3889                         if (gen < min_trans) {
3890                                 slot++;
3891                                 continue;
3892                         }
3893                         if (!cache_only)
3894                                 break;
3895
3896                         if (max_key) {
3897                                 btrfs_node_key(cur, &disk_key, slot);
3898                                 if (comp_keys(&disk_key, max_key) >= 0) {
3899                                         ret = 1;
3900                                         goto out;
3901                                 }
3902                         }
3903
3904                         tmp = btrfs_find_tree_block(root, blockptr,
3905                                             btrfs_level_size(root, level - 1));
3906
3907                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
3908                                 free_extent_buffer(tmp);
3909                                 break;
3910                         }
3911                         if (tmp)
3912                                 free_extent_buffer(tmp);
3913                         slot++;
3914                 }
3915 find_next_key:
3916                 /*
3917                  * we didn't find a candidate key in this node, walk forward
3918                  * and find another one
3919                  */
3920                 if (slot >= nritems) {
3921                         path->slots[level] = slot;
3922                         btrfs_set_path_blocking(path);
3923                         sret = btrfs_find_next_key(root, path, min_key, level,
3924                                                   cache_only, min_trans);
3925                         if (sret == 0) {
3926                                 btrfs_release_path(root, path);
3927                                 goto again;
3928                         } else {
3929                                 btrfs_clear_path_blocking(path);
3930                                 goto out;
3931                         }
3932                 }
3933                 /* save our key for returning back */
3934                 btrfs_node_key_to_cpu(cur, &found_key, slot);
3935                 path->slots[level] = slot;
3936                 if (level == path->lowest_level) {
3937                         ret = 0;
3938                         unlock_up(path, level, 1);
3939                         goto out;
3940                 }
3941                 btrfs_set_path_blocking(path);
3942                 cur = read_node_slot(root, cur, slot);
3943
3944                 btrfs_tree_lock(cur);
3945
3946                 path->locks[level - 1] = 1;
3947                 path->nodes[level - 1] = cur;
3948                 unlock_up(path, level, 1);
3949                 btrfs_clear_path_blocking(path);
3950         }
3951 out:
3952         if (ret == 0)
3953                 memcpy(min_key, &found_key, sizeof(found_key));
3954         btrfs_set_path_blocking(path);
3955         return ret;
3956 }
3957
3958 /*
3959  * this is similar to btrfs_next_leaf, but does not try to preserve
3960  * and fixup the path.  It looks for and returns the next key in the
3961  * tree based on the current path and the cache_only and min_trans
3962  * parameters.
3963  *
3964  * 0 is returned if another key is found, < 0 if there are any errors
3965  * and 1 is returned if there are no higher keys in the tree
3966  *
3967  * path->keep_locks should be set to 1 on the search made before
3968  * calling this function.
3969  */
3970 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
3971                         struct btrfs_key *key, int lowest_level,
3972                         int cache_only, u64 min_trans)
3973 {
3974         int level = lowest_level;
3975         int slot;
3976         struct extent_buffer *c;
3977
3978         WARN_ON(!path->keep_locks);
3979         while (level < BTRFS_MAX_LEVEL) {
3980                 if (!path->nodes[level])
3981                         return 1;
3982
3983                 slot = path->slots[level] + 1;
3984                 c = path->nodes[level];
3985 next:
3986                 if (slot >= btrfs_header_nritems(c)) {
3987                         level++;
3988                         if (level == BTRFS_MAX_LEVEL)
3989                                 return 1;
3990                         continue;
3991                 }
3992                 if (level == 0)
3993                         btrfs_item_key_to_cpu(c, key, slot);
3994                 else {
3995                         u64 blockptr = btrfs_node_blockptr(c, slot);
3996                         u64 gen = btrfs_node_ptr_generation(c, slot);
3997
3998                         if (cache_only) {
3999                                 struct extent_buffer *cur;
4000                                 cur = btrfs_find_tree_block(root, blockptr,
4001                                             btrfs_level_size(root, level - 1));
4002                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
4003                                         slot++;
4004                                         if (cur)
4005                                                 free_extent_buffer(cur);
4006                                         goto next;
4007                                 }
4008                                 free_extent_buffer(cur);
4009                         }
4010                         if (gen < min_trans) {
4011                                 slot++;
4012                                 goto next;
4013                         }
4014                         btrfs_node_key_to_cpu(c, key, slot);
4015                 }
4016                 return 0;
4017         }
4018         return 1;
4019 }
4020
4021 /*
4022  * search the tree again to find a leaf with greater keys
4023  * returns 0 if it found something or 1 if there are no greater leaves.
4024  * returns < 0 on io errors.
4025  */
4026 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4027 {
4028         int slot;
4029         int level = 1;
4030         struct extent_buffer *c;
4031         struct extent_buffer *next = NULL;
4032         struct btrfs_key key;
4033         u32 nritems;
4034         int ret;
4035
4036         nritems = btrfs_header_nritems(path->nodes[0]);
4037         if (nritems == 0)
4038                 return 1;
4039
4040         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4041
4042         btrfs_release_path(root, path);
4043         path->keep_locks = 1;
4044         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4045         path->keep_locks = 0;
4046
4047         if (ret < 0)
4048                 return ret;
4049
4050         btrfs_set_path_blocking(path);
4051         nritems = btrfs_header_nritems(path->nodes[0]);
4052         /*
4053          * by releasing the path above we dropped all our locks.  A balance
4054          * could have added more items next to the key that used to be
4055          * at the very end of the block.  So, check again here and
4056          * advance the path if there are now more items available.
4057          */
4058         if (nritems > 0 && path->slots[0] < nritems - 1) {
4059                 path->slots[0]++;
4060                 goto done;
4061         }
4062
4063         while (level < BTRFS_MAX_LEVEL) {
4064                 if (!path->nodes[level])
4065                         return 1;
4066
4067                 slot = path->slots[level] + 1;
4068                 c = path->nodes[level];
4069                 if (slot >= btrfs_header_nritems(c)) {
4070                         level++;
4071                         if (level == BTRFS_MAX_LEVEL)
4072                                 return 1;
4073                         continue;
4074                 }
4075
4076                 if (next) {
4077                         btrfs_tree_unlock(next);
4078                         free_extent_buffer(next);
4079                 }
4080
4081                 /* the path was set to blocking above */
4082                 if (level == 1 && (path->locks[1] || path->skip_locking) &&
4083                     path->reada)
4084                         reada_for_search(root, path, level, slot, 0);
4085
4086                 next = read_node_slot(root, c, slot);
4087                 if (!path->skip_locking) {
4088                         WARN_ON(!btrfs_tree_locked(c));
4089                         btrfs_tree_lock(next);
4090                         btrfs_set_lock_blocking(next);
4091                 }
4092                 break;
4093         }
4094         path->slots[level] = slot;
4095         while (1) {
4096                 level--;
4097                 c = path->nodes[level];
4098                 if (path->locks[level])
4099                         btrfs_tree_unlock(c);
4100                 free_extent_buffer(c);
4101                 path->nodes[level] = next;
4102                 path->slots[level] = 0;
4103                 if (!path->skip_locking)
4104                         path->locks[level] = 1;
4105                 if (!level)
4106                         break;
4107
4108                 btrfs_set_path_blocking(path);
4109                 if (level == 1 && path->locks[1] && path->reada)
4110                         reada_for_search(root, path, level, slot, 0);
4111                 next = read_node_slot(root, next, 0);
4112                 if (!path->skip_locking) {
4113                         WARN_ON(!btrfs_tree_locked(path->nodes[level]));
4114                         btrfs_tree_lock(next);
4115                         btrfs_set_lock_blocking(next);
4116                 }
4117         }
4118 done:
4119         unlock_up(path, 0, 1);
4120         return 0;
4121 }
4122
4123 /*
4124  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4125  * searching until it gets past min_objectid or finds an item of 'type'
4126  *
4127  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4128  */
4129 int btrfs_previous_item(struct btrfs_root *root,
4130                         struct btrfs_path *path, u64 min_objectid,
4131                         int type)
4132 {
4133         struct btrfs_key found_key;
4134         struct extent_buffer *leaf;
4135         u32 nritems;
4136         int ret;
4137
4138         while (1) {
4139                 if (path->slots[0] == 0) {
4140                         btrfs_set_path_blocking(path);
4141                         ret = btrfs_prev_leaf(root, path);
4142                         if (ret != 0)
4143                                 return ret;
4144                 } else {
4145                         path->slots[0]--;
4146                 }
4147                 leaf = path->nodes[0];
4148                 nritems = btrfs_header_nritems(leaf);
4149                 if (nritems == 0)
4150                         return 1;
4151                 if (path->slots[0] == nritems)
4152                         path->slots[0]--;
4153
4154                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4155                 if (found_key.type == type)
4156                         return 0;
4157                 if (found_key.objectid < min_objectid)
4158                         break;
4159                 if (found_key.objectid == min_objectid &&
4160                     found_key.type < type)
4161                         break;
4162         }
4163         return 1;
4164 }