Btrfs: mark delayed refs as for cow
[linux-3.10.git] / 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 <linux/slab.h>
21 #include "ctree.h"
22 #include "disk-io.h"
23 #include "transaction.h"
24 #include "print-tree.h"
25 #include "locking.h"
26
27 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
28                       *root, struct btrfs_path *path, int level);
29 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
30                       *root, struct btrfs_key *ins_key,
31                       struct btrfs_path *path, int data_size, int extend);
32 static int push_node_left(struct btrfs_trans_handle *trans,
33                           struct btrfs_root *root, struct extent_buffer *dst,
34                           struct extent_buffer *src, int empty);
35 static int balance_node_right(struct btrfs_trans_handle *trans,
36                               struct btrfs_root *root,
37                               struct extent_buffer *dst_buf,
38                               struct extent_buffer *src_buf);
39 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
40                    struct btrfs_path *path, int level, int slot);
41
42 struct btrfs_path *btrfs_alloc_path(void)
43 {
44         struct btrfs_path *path;
45         path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
46         return path;
47 }
48
49 /*
50  * set all locked nodes in the path to blocking locks.  This should
51  * be done before scheduling
52  */
53 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
54 {
55         int i;
56         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
57                 if (!p->nodes[i] || !p->locks[i])
58                         continue;
59                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
60                 if (p->locks[i] == BTRFS_READ_LOCK)
61                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
62                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
63                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
64         }
65 }
66
67 /*
68  * reset all the locked nodes in the patch to spinning locks.
69  *
70  * held is used to keep lockdep happy, when lockdep is enabled
71  * we set held to a blocking lock before we go around and
72  * retake all the spinlocks in the path.  You can safely use NULL
73  * for held
74  */
75 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
76                                         struct extent_buffer *held, int held_rw)
77 {
78         int i;
79
80 #ifdef CONFIG_DEBUG_LOCK_ALLOC
81         /* lockdep really cares that we take all of these spinlocks
82          * in the right order.  If any of the locks in the path are not
83          * currently blocking, it is going to complain.  So, make really
84          * really sure by forcing the path to blocking before we clear
85          * the path blocking.
86          */
87         if (held) {
88                 btrfs_set_lock_blocking_rw(held, held_rw);
89                 if (held_rw == BTRFS_WRITE_LOCK)
90                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
91                 else if (held_rw == BTRFS_READ_LOCK)
92                         held_rw = BTRFS_READ_LOCK_BLOCKING;
93         }
94         btrfs_set_path_blocking(p);
95 #endif
96
97         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
98                 if (p->nodes[i] && p->locks[i]) {
99                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
100                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
101                                 p->locks[i] = BTRFS_WRITE_LOCK;
102                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
103                                 p->locks[i] = BTRFS_READ_LOCK;
104                 }
105         }
106
107 #ifdef CONFIG_DEBUG_LOCK_ALLOC
108         if (held)
109                 btrfs_clear_lock_blocking_rw(held, held_rw);
110 #endif
111 }
112
113 /* this also releases the path */
114 void btrfs_free_path(struct btrfs_path *p)
115 {
116         if (!p)
117                 return;
118         btrfs_release_path(p);
119         kmem_cache_free(btrfs_path_cachep, p);
120 }
121
122 /*
123  * path release drops references on the extent buffers in the path
124  * and it drops any locks held by this path
125  *
126  * It is safe to call this on paths that no locks or extent buffers held.
127  */
128 noinline void btrfs_release_path(struct btrfs_path *p)
129 {
130         int i;
131
132         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
133                 p->slots[i] = 0;
134                 if (!p->nodes[i])
135                         continue;
136                 if (p->locks[i]) {
137                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
138                         p->locks[i] = 0;
139                 }
140                 free_extent_buffer(p->nodes[i]);
141                 p->nodes[i] = NULL;
142         }
143 }
144
145 /*
146  * safely gets a reference on the root node of a tree.  A lock
147  * is not taken, so a concurrent writer may put a different node
148  * at the root of the tree.  See btrfs_lock_root_node for the
149  * looping required.
150  *
151  * The extent buffer returned by this has a reference taken, so
152  * it won't disappear.  It may stop being the root of the tree
153  * at any time because there are no locks held.
154  */
155 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
156 {
157         struct extent_buffer *eb;
158
159         rcu_read_lock();
160         eb = rcu_dereference(root->node);
161         extent_buffer_get(eb);
162         rcu_read_unlock();
163         return eb;
164 }
165
166 /* loop around taking references on and locking the root node of the
167  * tree until you end up with a lock on the root.  A locked buffer
168  * is returned, with a reference held.
169  */
170 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
171 {
172         struct extent_buffer *eb;
173
174         while (1) {
175                 eb = btrfs_root_node(root);
176                 btrfs_tree_lock(eb);
177                 if (eb == root->node)
178                         break;
179                 btrfs_tree_unlock(eb);
180                 free_extent_buffer(eb);
181         }
182         return eb;
183 }
184
185 /* loop around taking references on and locking the root node of the
186  * tree until you end up with a lock on the root.  A locked buffer
187  * is returned, with a reference held.
188  */
189 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
190 {
191         struct extent_buffer *eb;
192
193         while (1) {
194                 eb = btrfs_root_node(root);
195                 btrfs_tree_read_lock(eb);
196                 if (eb == root->node)
197                         break;
198                 btrfs_tree_read_unlock(eb);
199                 free_extent_buffer(eb);
200         }
201         return eb;
202 }
203
204 /* cowonly root (everything not a reference counted cow subvolume), just get
205  * put onto a simple dirty list.  transaction.c walks this to make sure they
206  * get properly updated on disk.
207  */
208 static void add_root_to_dirty_list(struct btrfs_root *root)
209 {
210         if (root->track_dirty && list_empty(&root->dirty_list)) {
211                 list_add(&root->dirty_list,
212                          &root->fs_info->dirty_cowonly_roots);
213         }
214 }
215
216 /*
217  * used by snapshot creation to make a copy of a root for a tree with
218  * a given objectid.  The buffer with the new root node is returned in
219  * cow_ret, and this func returns zero on success or a negative error code.
220  */
221 int btrfs_copy_root(struct btrfs_trans_handle *trans,
222                       struct btrfs_root *root,
223                       struct extent_buffer *buf,
224                       struct extent_buffer **cow_ret, u64 new_root_objectid)
225 {
226         struct extent_buffer *cow;
227         int ret = 0;
228         int level;
229         struct btrfs_disk_key disk_key;
230
231         WARN_ON(root->ref_cows && trans->transid !=
232                 root->fs_info->running_transaction->transid);
233         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
234
235         level = btrfs_header_level(buf);
236         if (level == 0)
237                 btrfs_item_key(buf, &disk_key, 0);
238         else
239                 btrfs_node_key(buf, &disk_key, 0);
240
241         cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
242                                      new_root_objectid, &disk_key, level,
243                                      buf->start, 0, 1);
244         if (IS_ERR(cow))
245                 return PTR_ERR(cow);
246
247         copy_extent_buffer(cow, buf, 0, 0, cow->len);
248         btrfs_set_header_bytenr(cow, cow->start);
249         btrfs_set_header_generation(cow, trans->transid);
250         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
251         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
252                                      BTRFS_HEADER_FLAG_RELOC);
253         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
254                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
255         else
256                 btrfs_set_header_owner(cow, new_root_objectid);
257
258         write_extent_buffer(cow, root->fs_info->fsid,
259                             (unsigned long)btrfs_header_fsid(cow),
260                             BTRFS_FSID_SIZE);
261
262         WARN_ON(btrfs_header_generation(buf) > trans->transid);
263         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
264                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
265         else
266                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
267
268         if (ret)
269                 return ret;
270
271         btrfs_mark_buffer_dirty(cow);
272         *cow_ret = cow;
273         return 0;
274 }
275
276 /*
277  * check if the tree block can be shared by multiple trees
278  */
279 int btrfs_block_can_be_shared(struct btrfs_root *root,
280                               struct extent_buffer *buf)
281 {
282         /*
283          * Tree blocks not in refernece counted trees and tree roots
284          * are never shared. If a block was allocated after the last
285          * snapshot and the block was not allocated by tree relocation,
286          * we know the block is not shared.
287          */
288         if (root->ref_cows &&
289             buf != root->node && buf != root->commit_root &&
290             (btrfs_header_generation(buf) <=
291              btrfs_root_last_snapshot(&root->root_item) ||
292              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
293                 return 1;
294 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
295         if (root->ref_cows &&
296             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
297                 return 1;
298 #endif
299         return 0;
300 }
301
302 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
303                                        struct btrfs_root *root,
304                                        struct extent_buffer *buf,
305                                        struct extent_buffer *cow,
306                                        int *last_ref)
307 {
308         u64 refs;
309         u64 owner;
310         u64 flags;
311         u64 new_flags = 0;
312         int ret;
313
314         /*
315          * Backrefs update rules:
316          *
317          * Always use full backrefs for extent pointers in tree block
318          * allocated by tree relocation.
319          *
320          * If a shared tree block is no longer referenced by its owner
321          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
322          * use full backrefs for extent pointers in tree block.
323          *
324          * If a tree block is been relocating
325          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
326          * use full backrefs for extent pointers in tree block.
327          * The reason for this is some operations (such as drop tree)
328          * are only allowed for blocks use full backrefs.
329          */
330
331         if (btrfs_block_can_be_shared(root, buf)) {
332                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
333                                                buf->len, &refs, &flags);
334                 BUG_ON(ret);
335                 BUG_ON(refs == 0);
336         } else {
337                 refs = 1;
338                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
339                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
340                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
341                 else
342                         flags = 0;
343         }
344
345         owner = btrfs_header_owner(buf);
346         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
347                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
348
349         if (refs > 1) {
350                 if ((owner == root->root_key.objectid ||
351                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
352                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
353                         ret = btrfs_inc_ref(trans, root, buf, 1, 1);
354                         BUG_ON(ret);
355
356                         if (root->root_key.objectid ==
357                             BTRFS_TREE_RELOC_OBJECTID) {
358                                 ret = btrfs_dec_ref(trans, root, buf, 0, 1);
359                                 BUG_ON(ret);
360                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
361                                 BUG_ON(ret);
362                         }
363                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
364                 } else {
365
366                         if (root->root_key.objectid ==
367                             BTRFS_TREE_RELOC_OBJECTID)
368                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
369                         else
370                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
371                         BUG_ON(ret);
372                 }
373                 if (new_flags != 0) {
374                         ret = btrfs_set_disk_extent_flags(trans, root,
375                                                           buf->start,
376                                                           buf->len,
377                                                           new_flags, 0);
378                         BUG_ON(ret);
379                 }
380         } else {
381                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
382                         if (root->root_key.objectid ==
383                             BTRFS_TREE_RELOC_OBJECTID)
384                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
385                         else
386                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
387                         BUG_ON(ret);
388                         ret = btrfs_dec_ref(trans, root, buf, 1, 1);
389                         BUG_ON(ret);
390                 }
391                 clean_tree_block(trans, root, buf);
392                 *last_ref = 1;
393         }
394         return 0;
395 }
396
397 /*
398  * does the dirty work in cow of a single block.  The parent block (if
399  * supplied) is updated to point to the new cow copy.  The new buffer is marked
400  * dirty and returned locked.  If you modify the block it needs to be marked
401  * dirty again.
402  *
403  * search_start -- an allocation hint for the new block
404  *
405  * empty_size -- a hint that you plan on doing more cow.  This is the size in
406  * bytes the allocator should try to find free next to the block it returns.
407  * This is just a hint and may be ignored by the allocator.
408  */
409 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
410                              struct btrfs_root *root,
411                              struct extent_buffer *buf,
412                              struct extent_buffer *parent, int parent_slot,
413                              struct extent_buffer **cow_ret,
414                              u64 search_start, u64 empty_size)
415 {
416         struct btrfs_disk_key disk_key;
417         struct extent_buffer *cow;
418         int level;
419         int last_ref = 0;
420         int unlock_orig = 0;
421         u64 parent_start;
422
423         if (*cow_ret == buf)
424                 unlock_orig = 1;
425
426         btrfs_assert_tree_locked(buf);
427
428         WARN_ON(root->ref_cows && trans->transid !=
429                 root->fs_info->running_transaction->transid);
430         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
431
432         level = btrfs_header_level(buf);
433
434         if (level == 0)
435                 btrfs_item_key(buf, &disk_key, 0);
436         else
437                 btrfs_node_key(buf, &disk_key, 0);
438
439         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
440                 if (parent)
441                         parent_start = parent->start;
442                 else
443                         parent_start = 0;
444         } else
445                 parent_start = 0;
446
447         cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
448                                      root->root_key.objectid, &disk_key,
449                                      level, search_start, empty_size, 1);
450         if (IS_ERR(cow))
451                 return PTR_ERR(cow);
452
453         /* cow is set to blocking by btrfs_init_new_buffer */
454
455         copy_extent_buffer(cow, buf, 0, 0, cow->len);
456         btrfs_set_header_bytenr(cow, cow->start);
457         btrfs_set_header_generation(cow, trans->transid);
458         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
459         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
460                                      BTRFS_HEADER_FLAG_RELOC);
461         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
462                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
463         else
464                 btrfs_set_header_owner(cow, root->root_key.objectid);
465
466         write_extent_buffer(cow, root->fs_info->fsid,
467                             (unsigned long)btrfs_header_fsid(cow),
468                             BTRFS_FSID_SIZE);
469
470         update_ref_for_cow(trans, root, buf, cow, &last_ref);
471
472         if (root->ref_cows)
473                 btrfs_reloc_cow_block(trans, root, buf, cow);
474
475         if (buf == root->node) {
476                 WARN_ON(parent && parent != buf);
477                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
478                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
479                         parent_start = buf->start;
480                 else
481                         parent_start = 0;
482
483                 extent_buffer_get(cow);
484                 rcu_assign_pointer(root->node, cow);
485
486                 btrfs_free_tree_block(trans, root, buf, parent_start,
487                                       last_ref, 1);
488                 free_extent_buffer(buf);
489                 add_root_to_dirty_list(root);
490         } else {
491                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
492                         parent_start = parent->start;
493                 else
494                         parent_start = 0;
495
496                 WARN_ON(trans->transid != btrfs_header_generation(parent));
497                 btrfs_set_node_blockptr(parent, parent_slot,
498                                         cow->start);
499                 btrfs_set_node_ptr_generation(parent, parent_slot,
500                                               trans->transid);
501                 btrfs_mark_buffer_dirty(parent);
502                 btrfs_free_tree_block(trans, root, buf, parent_start,
503                                       last_ref, 1);
504         }
505         if (unlock_orig)
506                 btrfs_tree_unlock(buf);
507         free_extent_buffer(buf);
508         btrfs_mark_buffer_dirty(cow);
509         *cow_ret = cow;
510         return 0;
511 }
512
513 static inline int should_cow_block(struct btrfs_trans_handle *trans,
514                                    struct btrfs_root *root,
515                                    struct extent_buffer *buf)
516 {
517         /* ensure we can see the force_cow */
518         smp_rmb();
519
520         /*
521          * We do not need to cow a block if
522          * 1) this block is not created or changed in this transaction;
523          * 2) this block does not belong to TREE_RELOC tree;
524          * 3) the root is not forced COW.
525          *
526          * What is forced COW:
527          *    when we create snapshot during commiting the transaction,
528          *    after we've finished coping src root, we must COW the shared
529          *    block to ensure the metadata consistency.
530          */
531         if (btrfs_header_generation(buf) == trans->transid &&
532             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
533             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
534               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
535             !root->force_cow)
536                 return 0;
537         return 1;
538 }
539
540 /*
541  * cows a single block, see __btrfs_cow_block for the real work.
542  * This version of it has extra checks so that a block isn't cow'd more than
543  * once per transaction, as long as it hasn't been written yet
544  */
545 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
546                     struct btrfs_root *root, struct extent_buffer *buf,
547                     struct extent_buffer *parent, int parent_slot,
548                     struct extent_buffer **cow_ret)
549 {
550         u64 search_start;
551         int ret;
552
553         if (trans->transaction != root->fs_info->running_transaction) {
554                 printk(KERN_CRIT "trans %llu running %llu\n",
555                        (unsigned long long)trans->transid,
556                        (unsigned long long)
557                        root->fs_info->running_transaction->transid);
558                 WARN_ON(1);
559         }
560         if (trans->transid != root->fs_info->generation) {
561                 printk(KERN_CRIT "trans %llu running %llu\n",
562                        (unsigned long long)trans->transid,
563                        (unsigned long long)root->fs_info->generation);
564                 WARN_ON(1);
565         }
566
567         if (!should_cow_block(trans, root, buf)) {
568                 *cow_ret = buf;
569                 return 0;
570         }
571
572         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
573
574         if (parent)
575                 btrfs_set_lock_blocking(parent);
576         btrfs_set_lock_blocking(buf);
577
578         ret = __btrfs_cow_block(trans, root, buf, parent,
579                                  parent_slot, cow_ret, search_start, 0);
580
581         trace_btrfs_cow_block(root, buf, *cow_ret);
582
583         return ret;
584 }
585
586 /*
587  * helper function for defrag to decide if two blocks pointed to by a
588  * node are actually close by
589  */
590 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
591 {
592         if (blocknr < other && other - (blocknr + blocksize) < 32768)
593                 return 1;
594         if (blocknr > other && blocknr - (other + blocksize) < 32768)
595                 return 1;
596         return 0;
597 }
598
599 /*
600  * compare two keys in a memcmp fashion
601  */
602 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
603 {
604         struct btrfs_key k1;
605
606         btrfs_disk_key_to_cpu(&k1, disk);
607
608         return btrfs_comp_cpu_keys(&k1, k2);
609 }
610
611 /*
612  * same as comp_keys only with two btrfs_key's
613  */
614 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
615 {
616         if (k1->objectid > k2->objectid)
617                 return 1;
618         if (k1->objectid < k2->objectid)
619                 return -1;
620         if (k1->type > k2->type)
621                 return 1;
622         if (k1->type < k2->type)
623                 return -1;
624         if (k1->offset > k2->offset)
625                 return 1;
626         if (k1->offset < k2->offset)
627                 return -1;
628         return 0;
629 }
630
631 /*
632  * this is used by the defrag code to go through all the
633  * leaves pointed to by a node and reallocate them so that
634  * disk order is close to key order
635  */
636 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
637                        struct btrfs_root *root, struct extent_buffer *parent,
638                        int start_slot, int cache_only, u64 *last_ret,
639                        struct btrfs_key *progress)
640 {
641         struct extent_buffer *cur;
642         u64 blocknr;
643         u64 gen;
644         u64 search_start = *last_ret;
645         u64 last_block = 0;
646         u64 other;
647         u32 parent_nritems;
648         int end_slot;
649         int i;
650         int err = 0;
651         int parent_level;
652         int uptodate;
653         u32 blocksize;
654         int progress_passed = 0;
655         struct btrfs_disk_key disk_key;
656
657         parent_level = btrfs_header_level(parent);
658         if (cache_only && parent_level != 1)
659                 return 0;
660
661         if (trans->transaction != root->fs_info->running_transaction)
662                 WARN_ON(1);
663         if (trans->transid != root->fs_info->generation)
664                 WARN_ON(1);
665
666         parent_nritems = btrfs_header_nritems(parent);
667         blocksize = btrfs_level_size(root, parent_level - 1);
668         end_slot = parent_nritems;
669
670         if (parent_nritems == 1)
671                 return 0;
672
673         btrfs_set_lock_blocking(parent);
674
675         for (i = start_slot; i < end_slot; i++) {
676                 int close = 1;
677
678                 btrfs_node_key(parent, &disk_key, i);
679                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
680                         continue;
681
682                 progress_passed = 1;
683                 blocknr = btrfs_node_blockptr(parent, i);
684                 gen = btrfs_node_ptr_generation(parent, i);
685                 if (last_block == 0)
686                         last_block = blocknr;
687
688                 if (i > 0) {
689                         other = btrfs_node_blockptr(parent, i - 1);
690                         close = close_blocks(blocknr, other, blocksize);
691                 }
692                 if (!close && i < end_slot - 2) {
693                         other = btrfs_node_blockptr(parent, i + 1);
694                         close = close_blocks(blocknr, other, blocksize);
695                 }
696                 if (close) {
697                         last_block = blocknr;
698                         continue;
699                 }
700
701                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
702                 if (cur)
703                         uptodate = btrfs_buffer_uptodate(cur, gen);
704                 else
705                         uptodate = 0;
706                 if (!cur || !uptodate) {
707                         if (cache_only) {
708                                 free_extent_buffer(cur);
709                                 continue;
710                         }
711                         if (!cur) {
712                                 cur = read_tree_block(root, blocknr,
713                                                          blocksize, gen);
714                                 if (!cur)
715                                         return -EIO;
716                         } else if (!uptodate) {
717                                 btrfs_read_buffer(cur, gen);
718                         }
719                 }
720                 if (search_start == 0)
721                         search_start = last_block;
722
723                 btrfs_tree_lock(cur);
724                 btrfs_set_lock_blocking(cur);
725                 err = __btrfs_cow_block(trans, root, cur, parent, i,
726                                         &cur, search_start,
727                                         min(16 * blocksize,
728                                             (end_slot - i) * blocksize));
729                 if (err) {
730                         btrfs_tree_unlock(cur);
731                         free_extent_buffer(cur);
732                         break;
733                 }
734                 search_start = cur->start;
735                 last_block = cur->start;
736                 *last_ret = search_start;
737                 btrfs_tree_unlock(cur);
738                 free_extent_buffer(cur);
739         }
740         return err;
741 }
742
743 /*
744  * The leaf data grows from end-to-front in the node.
745  * this returns the address of the start of the last item,
746  * which is the stop of the leaf data stack
747  */
748 static inline unsigned int leaf_data_end(struct btrfs_root *root,
749                                          struct extent_buffer *leaf)
750 {
751         u32 nr = btrfs_header_nritems(leaf);
752         if (nr == 0)
753                 return BTRFS_LEAF_DATA_SIZE(root);
754         return btrfs_item_offset_nr(leaf, nr - 1);
755 }
756
757
758 /*
759  * search for key in the extent_buffer.  The items start at offset p,
760  * and they are item_size apart.  There are 'max' items in p.
761  *
762  * the slot in the array is returned via slot, and it points to
763  * the place where you would insert key if it is not found in
764  * the array.
765  *
766  * slot may point to max if the key is bigger than all of the keys
767  */
768 static noinline int generic_bin_search(struct extent_buffer *eb,
769                                        unsigned long p,
770                                        int item_size, struct btrfs_key *key,
771                                        int max, int *slot)
772 {
773         int low = 0;
774         int high = max;
775         int mid;
776         int ret;
777         struct btrfs_disk_key *tmp = NULL;
778         struct btrfs_disk_key unaligned;
779         unsigned long offset;
780         char *kaddr = NULL;
781         unsigned long map_start = 0;
782         unsigned long map_len = 0;
783         int err;
784
785         while (low < high) {
786                 mid = (low + high) / 2;
787                 offset = p + mid * item_size;
788
789                 if (!kaddr || offset < map_start ||
790                     (offset + sizeof(struct btrfs_disk_key)) >
791                     map_start + map_len) {
792
793                         err = map_private_extent_buffer(eb, offset,
794                                                 sizeof(struct btrfs_disk_key),
795                                                 &kaddr, &map_start, &map_len);
796
797                         if (!err) {
798                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
799                                                         map_start);
800                         } else {
801                                 read_extent_buffer(eb, &unaligned,
802                                                    offset, sizeof(unaligned));
803                                 tmp = &unaligned;
804                         }
805
806                 } else {
807                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
808                                                         map_start);
809                 }
810                 ret = comp_keys(tmp, key);
811
812                 if (ret < 0)
813                         low = mid + 1;
814                 else if (ret > 0)
815                         high = mid;
816                 else {
817                         *slot = mid;
818                         return 0;
819                 }
820         }
821         *slot = low;
822         return 1;
823 }
824
825 /*
826  * simple bin_search frontend that does the right thing for
827  * leaves vs nodes
828  */
829 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
830                       int level, int *slot)
831 {
832         if (level == 0) {
833                 return generic_bin_search(eb,
834                                           offsetof(struct btrfs_leaf, items),
835                                           sizeof(struct btrfs_item),
836                                           key, btrfs_header_nritems(eb),
837                                           slot);
838         } else {
839                 return generic_bin_search(eb,
840                                           offsetof(struct btrfs_node, ptrs),
841                                           sizeof(struct btrfs_key_ptr),
842                                           key, btrfs_header_nritems(eb),
843                                           slot);
844         }
845         return -1;
846 }
847
848 int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
849                      int level, int *slot)
850 {
851         return bin_search(eb, key, level, slot);
852 }
853
854 static void root_add_used(struct btrfs_root *root, u32 size)
855 {
856         spin_lock(&root->accounting_lock);
857         btrfs_set_root_used(&root->root_item,
858                             btrfs_root_used(&root->root_item) + size);
859         spin_unlock(&root->accounting_lock);
860 }
861
862 static void root_sub_used(struct btrfs_root *root, u32 size)
863 {
864         spin_lock(&root->accounting_lock);
865         btrfs_set_root_used(&root->root_item,
866                             btrfs_root_used(&root->root_item) - size);
867         spin_unlock(&root->accounting_lock);
868 }
869
870 /* given a node and slot number, this reads the blocks it points to.  The
871  * extent buffer is returned with a reference taken (but unlocked).
872  * NULL is returned on error.
873  */
874 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
875                                    struct extent_buffer *parent, int slot)
876 {
877         int level = btrfs_header_level(parent);
878         if (slot < 0)
879                 return NULL;
880         if (slot >= btrfs_header_nritems(parent))
881                 return NULL;
882
883         BUG_ON(level == 0);
884
885         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
886                        btrfs_level_size(root, level - 1),
887                        btrfs_node_ptr_generation(parent, slot));
888 }
889
890 /*
891  * node level balancing, used to make sure nodes are in proper order for
892  * item deletion.  We balance from the top down, so we have to make sure
893  * that a deletion won't leave an node completely empty later on.
894  */
895 static noinline int balance_level(struct btrfs_trans_handle *trans,
896                          struct btrfs_root *root,
897                          struct btrfs_path *path, int level)
898 {
899         struct extent_buffer *right = NULL;
900         struct extent_buffer *mid;
901         struct extent_buffer *left = NULL;
902         struct extent_buffer *parent = NULL;
903         int ret = 0;
904         int wret;
905         int pslot;
906         int orig_slot = path->slots[level];
907         u64 orig_ptr;
908
909         if (level == 0)
910                 return 0;
911
912         mid = path->nodes[level];
913
914         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
915                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
916         WARN_ON(btrfs_header_generation(mid) != trans->transid);
917
918         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
919
920         if (level < BTRFS_MAX_LEVEL - 1) {
921                 parent = path->nodes[level + 1];
922                 pslot = path->slots[level + 1];
923         }
924
925         /*
926          * deal with the case where there is only one pointer in the root
927          * by promoting the node below to a root
928          */
929         if (!parent) {
930                 struct extent_buffer *child;
931
932                 if (btrfs_header_nritems(mid) != 1)
933                         return 0;
934
935                 /* promote the child to a root */
936                 child = read_node_slot(root, mid, 0);
937                 BUG_ON(!child);
938                 btrfs_tree_lock(child);
939                 btrfs_set_lock_blocking(child);
940                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
941                 if (ret) {
942                         btrfs_tree_unlock(child);
943                         free_extent_buffer(child);
944                         goto enospc;
945                 }
946
947                 rcu_assign_pointer(root->node, child);
948
949                 add_root_to_dirty_list(root);
950                 btrfs_tree_unlock(child);
951
952                 path->locks[level] = 0;
953                 path->nodes[level] = NULL;
954                 clean_tree_block(trans, root, mid);
955                 btrfs_tree_unlock(mid);
956                 /* once for the path */
957                 free_extent_buffer(mid);
958
959                 root_sub_used(root, mid->len);
960                 btrfs_free_tree_block(trans, root, mid, 0, 1, 0);
961                 /* once for the root ptr */
962                 free_extent_buffer(mid);
963                 return 0;
964         }
965         if (btrfs_header_nritems(mid) >
966             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
967                 return 0;
968
969         btrfs_header_nritems(mid);
970
971         left = read_node_slot(root, parent, pslot - 1);
972         if (left) {
973                 btrfs_tree_lock(left);
974                 btrfs_set_lock_blocking(left);
975                 wret = btrfs_cow_block(trans, root, left,
976                                        parent, pslot - 1, &left);
977                 if (wret) {
978                         ret = wret;
979                         goto enospc;
980                 }
981         }
982         right = read_node_slot(root, parent, pslot + 1);
983         if (right) {
984                 btrfs_tree_lock(right);
985                 btrfs_set_lock_blocking(right);
986                 wret = btrfs_cow_block(trans, root, right,
987                                        parent, pslot + 1, &right);
988                 if (wret) {
989                         ret = wret;
990                         goto enospc;
991                 }
992         }
993
994         /* first, try to make some room in the middle buffer */
995         if (left) {
996                 orig_slot += btrfs_header_nritems(left);
997                 wret = push_node_left(trans, root, left, mid, 1);
998                 if (wret < 0)
999                         ret = wret;
1000                 btrfs_header_nritems(mid);
1001         }
1002
1003         /*
1004          * then try to empty the right most buffer into the middle
1005          */
1006         if (right) {
1007                 wret = push_node_left(trans, root, mid, right, 1);
1008                 if (wret < 0 && wret != -ENOSPC)
1009                         ret = wret;
1010                 if (btrfs_header_nritems(right) == 0) {
1011                         clean_tree_block(trans, root, right);
1012                         btrfs_tree_unlock(right);
1013                         wret = del_ptr(trans, root, path, level + 1, pslot +
1014                                        1);
1015                         if (wret)
1016                                 ret = wret;
1017                         root_sub_used(root, right->len);
1018                         btrfs_free_tree_block(trans, root, right, 0, 1, 0);
1019                         free_extent_buffer(right);
1020                         right = NULL;
1021                 } else {
1022                         struct btrfs_disk_key right_key;
1023                         btrfs_node_key(right, &right_key, 0);
1024                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1025                         btrfs_mark_buffer_dirty(parent);
1026                 }
1027         }
1028         if (btrfs_header_nritems(mid) == 1) {
1029                 /*
1030                  * we're not allowed to leave a node with one item in the
1031                  * tree during a delete.  A deletion from lower in the tree
1032                  * could try to delete the only pointer in this node.
1033                  * So, pull some keys from the left.
1034                  * There has to be a left pointer at this point because
1035                  * otherwise we would have pulled some pointers from the
1036                  * right
1037                  */
1038                 BUG_ON(!left);
1039                 wret = balance_node_right(trans, root, mid, left);
1040                 if (wret < 0) {
1041                         ret = wret;
1042                         goto enospc;
1043                 }
1044                 if (wret == 1) {
1045                         wret = push_node_left(trans, root, left, mid, 1);
1046                         if (wret < 0)
1047                                 ret = wret;
1048                 }
1049                 BUG_ON(wret == 1);
1050         }
1051         if (btrfs_header_nritems(mid) == 0) {
1052                 clean_tree_block(trans, root, mid);
1053                 btrfs_tree_unlock(mid);
1054                 wret = del_ptr(trans, root, path, level + 1, pslot);
1055                 if (wret)
1056                         ret = wret;
1057                 root_sub_used(root, mid->len);
1058                 btrfs_free_tree_block(trans, root, mid, 0, 1, 0);
1059                 free_extent_buffer(mid);
1060                 mid = NULL;
1061         } else {
1062                 /* update the parent key to reflect our changes */
1063                 struct btrfs_disk_key mid_key;
1064                 btrfs_node_key(mid, &mid_key, 0);
1065                 btrfs_set_node_key(parent, &mid_key, pslot);
1066                 btrfs_mark_buffer_dirty(parent);
1067         }
1068
1069         /* update the path */
1070         if (left) {
1071                 if (btrfs_header_nritems(left) > orig_slot) {
1072                         extent_buffer_get(left);
1073                         /* left was locked after cow */
1074                         path->nodes[level] = left;
1075                         path->slots[level + 1] -= 1;
1076                         path->slots[level] = orig_slot;
1077                         if (mid) {
1078                                 btrfs_tree_unlock(mid);
1079                                 free_extent_buffer(mid);
1080                         }
1081                 } else {
1082                         orig_slot -= btrfs_header_nritems(left);
1083                         path->slots[level] = orig_slot;
1084                 }
1085         }
1086         /* double check we haven't messed things up */
1087         if (orig_ptr !=
1088             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1089                 BUG();
1090 enospc:
1091         if (right) {
1092                 btrfs_tree_unlock(right);
1093                 free_extent_buffer(right);
1094         }
1095         if (left) {
1096                 if (path->nodes[level] != left)
1097                         btrfs_tree_unlock(left);
1098                 free_extent_buffer(left);
1099         }
1100         return ret;
1101 }
1102
1103 /* Node balancing for insertion.  Here we only split or push nodes around
1104  * when they are completely full.  This is also done top down, so we
1105  * have to be pessimistic.
1106  */
1107 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1108                                           struct btrfs_root *root,
1109                                           struct btrfs_path *path, int level)
1110 {
1111         struct extent_buffer *right = NULL;
1112         struct extent_buffer *mid;
1113         struct extent_buffer *left = NULL;
1114         struct extent_buffer *parent = NULL;
1115         int ret = 0;
1116         int wret;
1117         int pslot;
1118         int orig_slot = path->slots[level];
1119
1120         if (level == 0)
1121                 return 1;
1122
1123         mid = path->nodes[level];
1124         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1125
1126         if (level < BTRFS_MAX_LEVEL - 1) {
1127                 parent = path->nodes[level + 1];
1128                 pslot = path->slots[level + 1];
1129         }
1130
1131         if (!parent)
1132                 return 1;
1133
1134         left = read_node_slot(root, parent, pslot - 1);
1135
1136         /* first, try to make some room in the middle buffer */
1137         if (left) {
1138                 u32 left_nr;
1139
1140                 btrfs_tree_lock(left);
1141                 btrfs_set_lock_blocking(left);
1142
1143                 left_nr = btrfs_header_nritems(left);
1144                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1145                         wret = 1;
1146                 } else {
1147                         ret = btrfs_cow_block(trans, root, left, parent,
1148                                               pslot - 1, &left);
1149                         if (ret)
1150                                 wret = 1;
1151                         else {
1152                                 wret = push_node_left(trans, root,
1153                                                       left, mid, 0);
1154                         }
1155                 }
1156                 if (wret < 0)
1157                         ret = wret;
1158                 if (wret == 0) {
1159                         struct btrfs_disk_key disk_key;
1160                         orig_slot += left_nr;
1161                         btrfs_node_key(mid, &disk_key, 0);
1162                         btrfs_set_node_key(parent, &disk_key, pslot);
1163                         btrfs_mark_buffer_dirty(parent);
1164                         if (btrfs_header_nritems(left) > orig_slot) {
1165                                 path->nodes[level] = left;
1166                                 path->slots[level + 1] -= 1;
1167                                 path->slots[level] = orig_slot;
1168                                 btrfs_tree_unlock(mid);
1169                                 free_extent_buffer(mid);
1170                         } else {
1171                                 orig_slot -=
1172                                         btrfs_header_nritems(left);
1173                                 path->slots[level] = orig_slot;
1174                                 btrfs_tree_unlock(left);
1175                                 free_extent_buffer(left);
1176                         }
1177                         return 0;
1178                 }
1179                 btrfs_tree_unlock(left);
1180                 free_extent_buffer(left);
1181         }
1182         right = read_node_slot(root, parent, pslot + 1);
1183
1184         /*
1185          * then try to empty the right most buffer into the middle
1186          */
1187         if (right) {
1188                 u32 right_nr;
1189
1190                 btrfs_tree_lock(right);
1191                 btrfs_set_lock_blocking(right);
1192
1193                 right_nr = btrfs_header_nritems(right);
1194                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1195                         wret = 1;
1196                 } else {
1197                         ret = btrfs_cow_block(trans, root, right,
1198                                               parent, pslot + 1,
1199                                               &right);
1200                         if (ret)
1201                                 wret = 1;
1202                         else {
1203                                 wret = balance_node_right(trans, root,
1204                                                           right, mid);
1205                         }
1206                 }
1207                 if (wret < 0)
1208                         ret = wret;
1209                 if (wret == 0) {
1210                         struct btrfs_disk_key disk_key;
1211
1212                         btrfs_node_key(right, &disk_key, 0);
1213                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
1214                         btrfs_mark_buffer_dirty(parent);
1215
1216                         if (btrfs_header_nritems(mid) <= orig_slot) {
1217                                 path->nodes[level] = right;
1218                                 path->slots[level + 1] += 1;
1219                                 path->slots[level] = orig_slot -
1220                                         btrfs_header_nritems(mid);
1221                                 btrfs_tree_unlock(mid);
1222                                 free_extent_buffer(mid);
1223                         } else {
1224                                 btrfs_tree_unlock(right);
1225                                 free_extent_buffer(right);
1226                         }
1227                         return 0;
1228                 }
1229                 btrfs_tree_unlock(right);
1230                 free_extent_buffer(right);
1231         }
1232         return 1;
1233 }
1234
1235 /*
1236  * readahead one full node of leaves, finding things that are close
1237  * to the block in 'slot', and triggering ra on them.
1238  */
1239 static void reada_for_search(struct btrfs_root *root,
1240                              struct btrfs_path *path,
1241                              int level, int slot, u64 objectid)
1242 {
1243         struct extent_buffer *node;
1244         struct btrfs_disk_key disk_key;
1245         u32 nritems;
1246         u64 search;
1247         u64 target;
1248         u64 nread = 0;
1249         u64 gen;
1250         int direction = path->reada;
1251         struct extent_buffer *eb;
1252         u32 nr;
1253         u32 blocksize;
1254         u32 nscan = 0;
1255
1256         if (level != 1)
1257                 return;
1258
1259         if (!path->nodes[level])
1260                 return;
1261
1262         node = path->nodes[level];
1263
1264         search = btrfs_node_blockptr(node, slot);
1265         blocksize = btrfs_level_size(root, level - 1);
1266         eb = btrfs_find_tree_block(root, search, blocksize);
1267         if (eb) {
1268                 free_extent_buffer(eb);
1269                 return;
1270         }
1271
1272         target = search;
1273
1274         nritems = btrfs_header_nritems(node);
1275         nr = slot;
1276
1277         while (1) {
1278                 if (direction < 0) {
1279                         if (nr == 0)
1280                                 break;
1281                         nr--;
1282                 } else if (direction > 0) {
1283                         nr++;
1284                         if (nr >= nritems)
1285                                 break;
1286                 }
1287                 if (path->reada < 0 && objectid) {
1288                         btrfs_node_key(node, &disk_key, nr);
1289                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
1290                                 break;
1291                 }
1292                 search = btrfs_node_blockptr(node, nr);
1293                 if ((search <= target && target - search <= 65536) ||
1294                     (search > target && search - target <= 65536)) {
1295                         gen = btrfs_node_ptr_generation(node, nr);
1296                         readahead_tree_block(root, search, blocksize, gen);
1297                         nread += blocksize;
1298                 }
1299                 nscan++;
1300                 if ((nread > 65536 || nscan > 32))
1301                         break;
1302         }
1303 }
1304
1305 /*
1306  * returns -EAGAIN if it had to drop the path, or zero if everything was in
1307  * cache
1308  */
1309 static noinline int reada_for_balance(struct btrfs_root *root,
1310                                       struct btrfs_path *path, int level)
1311 {
1312         int slot;
1313         int nritems;
1314         struct extent_buffer *parent;
1315         struct extent_buffer *eb;
1316         u64 gen;
1317         u64 block1 = 0;
1318         u64 block2 = 0;
1319         int ret = 0;
1320         int blocksize;
1321
1322         parent = path->nodes[level + 1];
1323         if (!parent)
1324                 return 0;
1325
1326         nritems = btrfs_header_nritems(parent);
1327         slot = path->slots[level + 1];
1328         blocksize = btrfs_level_size(root, level);
1329
1330         if (slot > 0) {
1331                 block1 = btrfs_node_blockptr(parent, slot - 1);
1332                 gen = btrfs_node_ptr_generation(parent, slot - 1);
1333                 eb = btrfs_find_tree_block(root, block1, blocksize);
1334                 if (eb && btrfs_buffer_uptodate(eb, gen))
1335                         block1 = 0;
1336                 free_extent_buffer(eb);
1337         }
1338         if (slot + 1 < nritems) {
1339                 block2 = btrfs_node_blockptr(parent, slot + 1);
1340                 gen = btrfs_node_ptr_generation(parent, slot + 1);
1341                 eb = btrfs_find_tree_block(root, block2, blocksize);
1342                 if (eb && btrfs_buffer_uptodate(eb, gen))
1343                         block2 = 0;
1344                 free_extent_buffer(eb);
1345         }
1346         if (block1 || block2) {
1347                 ret = -EAGAIN;
1348
1349                 /* release the whole path */
1350                 btrfs_release_path(path);
1351
1352                 /* read the blocks */
1353                 if (block1)
1354                         readahead_tree_block(root, block1, blocksize, 0);
1355                 if (block2)
1356                         readahead_tree_block(root, block2, blocksize, 0);
1357
1358                 if (block1) {
1359                         eb = read_tree_block(root, block1, blocksize, 0);
1360                         free_extent_buffer(eb);
1361                 }
1362                 if (block2) {
1363                         eb = read_tree_block(root, block2, blocksize, 0);
1364                         free_extent_buffer(eb);
1365                 }
1366         }
1367         return ret;
1368 }
1369
1370
1371 /*
1372  * when we walk down the tree, it is usually safe to unlock the higher layers
1373  * in the tree.  The exceptions are when our path goes through slot 0, because
1374  * operations on the tree might require changing key pointers higher up in the
1375  * tree.
1376  *
1377  * callers might also have set path->keep_locks, which tells this code to keep
1378  * the lock if the path points to the last slot in the block.  This is part of
1379  * walking through the tree, and selecting the next slot in the higher block.
1380  *
1381  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
1382  * if lowest_unlock is 1, level 0 won't be unlocked
1383  */
1384 static noinline void unlock_up(struct btrfs_path *path, int level,
1385                                int lowest_unlock)
1386 {
1387         int i;
1388         int skip_level = level;
1389         int no_skips = 0;
1390         struct extent_buffer *t;
1391
1392         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1393                 if (!path->nodes[i])
1394                         break;
1395                 if (!path->locks[i])
1396                         break;
1397                 if (!no_skips && path->slots[i] == 0) {
1398                         skip_level = i + 1;
1399                         continue;
1400                 }
1401                 if (!no_skips && path->keep_locks) {
1402                         u32 nritems;
1403                         t = path->nodes[i];
1404                         nritems = btrfs_header_nritems(t);
1405                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
1406                                 skip_level = i + 1;
1407                                 continue;
1408                         }
1409                 }
1410                 if (skip_level < i && i >= lowest_unlock)
1411                         no_skips = 1;
1412
1413                 t = path->nodes[i];
1414                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
1415                         btrfs_tree_unlock_rw(t, path->locks[i]);
1416                         path->locks[i] = 0;
1417                 }
1418         }
1419 }
1420
1421 /*
1422  * This releases any locks held in the path starting at level and
1423  * going all the way up to the root.
1424  *
1425  * btrfs_search_slot will keep the lock held on higher nodes in a few
1426  * corner cases, such as COW of the block at slot zero in the node.  This
1427  * ignores those rules, and it should only be called when there are no
1428  * more updates to be done higher up in the tree.
1429  */
1430 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
1431 {
1432         int i;
1433
1434         if (path->keep_locks)
1435                 return;
1436
1437         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1438                 if (!path->nodes[i])
1439                         continue;
1440                 if (!path->locks[i])
1441                         continue;
1442                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
1443                 path->locks[i] = 0;
1444         }
1445 }
1446
1447 /*
1448  * helper function for btrfs_search_slot.  The goal is to find a block
1449  * in cache without setting the path to blocking.  If we find the block
1450  * we return zero and the path is unchanged.
1451  *
1452  * If we can't find the block, we set the path blocking and do some
1453  * reada.  -EAGAIN is returned and the search must be repeated.
1454  */
1455 static int
1456 read_block_for_search(struct btrfs_trans_handle *trans,
1457                        struct btrfs_root *root, struct btrfs_path *p,
1458                        struct extent_buffer **eb_ret, int level, int slot,
1459                        struct btrfs_key *key)
1460 {
1461         u64 blocknr;
1462         u64 gen;
1463         u32 blocksize;
1464         struct extent_buffer *b = *eb_ret;
1465         struct extent_buffer *tmp;
1466         int ret;
1467
1468         blocknr = btrfs_node_blockptr(b, slot);
1469         gen = btrfs_node_ptr_generation(b, slot);
1470         blocksize = btrfs_level_size(root, level - 1);
1471
1472         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
1473         if (tmp) {
1474                 if (btrfs_buffer_uptodate(tmp, 0)) {
1475                         if (btrfs_buffer_uptodate(tmp, gen)) {
1476                                 /*
1477                                  * we found an up to date block without
1478                                  * sleeping, return
1479                                  * right away
1480                                  */
1481                                 *eb_ret = tmp;
1482                                 return 0;
1483                         }
1484                         /* the pages were up to date, but we failed
1485                          * the generation number check.  Do a full
1486                          * read for the generation number that is correct.
1487                          * We must do this without dropping locks so
1488                          * we can trust our generation number
1489                          */
1490                         free_extent_buffer(tmp);
1491                         btrfs_set_path_blocking(p);
1492
1493                         tmp = read_tree_block(root, blocknr, blocksize, gen);
1494                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
1495                                 *eb_ret = tmp;
1496                                 return 0;
1497                         }
1498                         free_extent_buffer(tmp);
1499                         btrfs_release_path(p);
1500                         return -EIO;
1501                 }
1502         }
1503
1504         /*
1505          * reduce lock contention at high levels
1506          * of the btree by dropping locks before
1507          * we read.  Don't release the lock on the current
1508          * level because we need to walk this node to figure
1509          * out which blocks to read.
1510          */
1511         btrfs_unlock_up_safe(p, level + 1);
1512         btrfs_set_path_blocking(p);
1513
1514         free_extent_buffer(tmp);
1515         if (p->reada)
1516                 reada_for_search(root, p, level, slot, key->objectid);
1517
1518         btrfs_release_path(p);
1519
1520         ret = -EAGAIN;
1521         tmp = read_tree_block(root, blocknr, blocksize, 0);
1522         if (tmp) {
1523                 /*
1524                  * If the read above didn't mark this buffer up to date,
1525                  * it will never end up being up to date.  Set ret to EIO now
1526                  * and give up so that our caller doesn't loop forever
1527                  * on our EAGAINs.
1528                  */
1529                 if (!btrfs_buffer_uptodate(tmp, 0))
1530                         ret = -EIO;
1531                 free_extent_buffer(tmp);
1532         }
1533         return ret;
1534 }
1535
1536 /*
1537  * helper function for btrfs_search_slot.  This does all of the checks
1538  * for node-level blocks and does any balancing required based on
1539  * the ins_len.
1540  *
1541  * If no extra work was required, zero is returned.  If we had to
1542  * drop the path, -EAGAIN is returned and btrfs_search_slot must
1543  * start over
1544  */
1545 static int
1546 setup_nodes_for_search(struct btrfs_trans_handle *trans,
1547                        struct btrfs_root *root, struct btrfs_path *p,
1548                        struct extent_buffer *b, int level, int ins_len,
1549                        int *write_lock_level)
1550 {
1551         int ret;
1552         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
1553             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
1554                 int sret;
1555
1556                 if (*write_lock_level < level + 1) {
1557                         *write_lock_level = level + 1;
1558                         btrfs_release_path(p);
1559                         goto again;
1560                 }
1561
1562                 sret = reada_for_balance(root, p, level);
1563                 if (sret)
1564                         goto again;
1565
1566                 btrfs_set_path_blocking(p);
1567                 sret = split_node(trans, root, p, level);
1568                 btrfs_clear_path_blocking(p, NULL, 0);
1569
1570                 BUG_ON(sret > 0);
1571                 if (sret) {
1572                         ret = sret;
1573                         goto done;
1574                 }
1575                 b = p->nodes[level];
1576         } else if (ins_len < 0 && btrfs_header_nritems(b) <
1577                    BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
1578                 int sret;
1579
1580                 if (*write_lock_level < level + 1) {
1581                         *write_lock_level = level + 1;
1582                         btrfs_release_path(p);
1583                         goto again;
1584                 }
1585
1586                 sret = reada_for_balance(root, p, level);
1587                 if (sret)
1588                         goto again;
1589
1590                 btrfs_set_path_blocking(p);
1591                 sret = balance_level(trans, root, p, level);
1592                 btrfs_clear_path_blocking(p, NULL, 0);
1593
1594                 if (sret) {
1595                         ret = sret;
1596                         goto done;
1597                 }
1598                 b = p->nodes[level];
1599                 if (!b) {
1600                         btrfs_release_path(p);
1601                         goto again;
1602                 }
1603                 BUG_ON(btrfs_header_nritems(b) == 1);
1604         }
1605         return 0;
1606
1607 again:
1608         ret = -EAGAIN;
1609 done:
1610         return ret;
1611 }
1612
1613 /*
1614  * look for key in the tree.  path is filled in with nodes along the way
1615  * if key is found, we return zero and you can find the item in the leaf
1616  * level of the path (level 0)
1617  *
1618  * If the key isn't found, the path points to the slot where it should
1619  * be inserted, and 1 is returned.  If there are other errors during the
1620  * search a negative error number is returned.
1621  *
1622  * if ins_len > 0, nodes and leaves will be split as we walk down the
1623  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1624  * possible)
1625  */
1626 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1627                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1628                       ins_len, int cow)
1629 {
1630         struct extent_buffer *b;
1631         int slot;
1632         int ret;
1633         int err;
1634         int level;
1635         int lowest_unlock = 1;
1636         int root_lock;
1637         /* everything at write_lock_level or lower must be write locked */
1638         int write_lock_level = 0;
1639         u8 lowest_level = 0;
1640
1641         lowest_level = p->lowest_level;
1642         WARN_ON(lowest_level && ins_len > 0);
1643         WARN_ON(p->nodes[0] != NULL);
1644
1645         if (ins_len < 0) {
1646                 lowest_unlock = 2;
1647
1648                 /* when we are removing items, we might have to go up to level
1649                  * two as we update tree pointers  Make sure we keep write
1650                  * for those levels as well
1651                  */
1652                 write_lock_level = 2;
1653         } else if (ins_len > 0) {
1654                 /*
1655                  * for inserting items, make sure we have a write lock on
1656                  * level 1 so we can update keys
1657                  */
1658                 write_lock_level = 1;
1659         }
1660
1661         if (!cow)
1662                 write_lock_level = -1;
1663
1664         if (cow && (p->keep_locks || p->lowest_level))
1665                 write_lock_level = BTRFS_MAX_LEVEL;
1666
1667 again:
1668         /*
1669          * we try very hard to do read locks on the root
1670          */
1671         root_lock = BTRFS_READ_LOCK;
1672         level = 0;
1673         if (p->search_commit_root) {
1674                 /*
1675                  * the commit roots are read only
1676                  * so we always do read locks
1677                  */
1678                 b = root->commit_root;
1679                 extent_buffer_get(b);
1680                 level = btrfs_header_level(b);
1681                 if (!p->skip_locking)
1682                         btrfs_tree_read_lock(b);
1683         } else {
1684                 if (p->skip_locking) {
1685                         b = btrfs_root_node(root);
1686                         level = btrfs_header_level(b);
1687                 } else {
1688                         /* we don't know the level of the root node
1689                          * until we actually have it read locked
1690                          */
1691                         b = btrfs_read_lock_root_node(root);
1692                         level = btrfs_header_level(b);
1693                         if (level <= write_lock_level) {
1694                                 /* whoops, must trade for write lock */
1695                                 btrfs_tree_read_unlock(b);
1696                                 free_extent_buffer(b);
1697                                 b = btrfs_lock_root_node(root);
1698                                 root_lock = BTRFS_WRITE_LOCK;
1699
1700                                 /* the level might have changed, check again */
1701                                 level = btrfs_header_level(b);
1702                         }
1703                 }
1704         }
1705         p->nodes[level] = b;
1706         if (!p->skip_locking)
1707                 p->locks[level] = root_lock;
1708
1709         while (b) {
1710                 level = btrfs_header_level(b);
1711
1712                 /*
1713                  * setup the path here so we can release it under lock
1714                  * contention with the cow code
1715                  */
1716                 if (cow) {
1717                         /*
1718                          * if we don't really need to cow this block
1719                          * then we don't want to set the path blocking,
1720                          * so we test it here
1721                          */
1722                         if (!should_cow_block(trans, root, b))
1723                                 goto cow_done;
1724
1725                         btrfs_set_path_blocking(p);
1726
1727                         /*
1728                          * must have write locks on this node and the
1729                          * parent
1730                          */
1731                         if (level + 1 > write_lock_level) {
1732                                 write_lock_level = level + 1;
1733                                 btrfs_release_path(p);
1734                                 goto again;
1735                         }
1736
1737                         err = btrfs_cow_block(trans, root, b,
1738                                               p->nodes[level + 1],
1739                                               p->slots[level + 1], &b);
1740                         if (err) {
1741                                 ret = err;
1742                                 goto done;
1743                         }
1744                 }
1745 cow_done:
1746                 BUG_ON(!cow && ins_len);
1747
1748                 p->nodes[level] = b;
1749                 btrfs_clear_path_blocking(p, NULL, 0);
1750
1751                 /*
1752                  * we have a lock on b and as long as we aren't changing
1753                  * the tree, there is no way to for the items in b to change.
1754                  * It is safe to drop the lock on our parent before we
1755                  * go through the expensive btree search on b.
1756                  *
1757                  * If cow is true, then we might be changing slot zero,
1758                  * which may require changing the parent.  So, we can't
1759                  * drop the lock until after we know which slot we're
1760                  * operating on.
1761                  */
1762                 if (!cow)
1763                         btrfs_unlock_up_safe(p, level + 1);
1764
1765                 ret = bin_search(b, key, level, &slot);
1766
1767                 if (level != 0) {
1768                         int dec = 0;
1769                         if (ret && slot > 0) {
1770                                 dec = 1;
1771                                 slot -= 1;
1772                         }
1773                         p->slots[level] = slot;
1774                         err = setup_nodes_for_search(trans, root, p, b, level,
1775                                              ins_len, &write_lock_level);
1776                         if (err == -EAGAIN)
1777                                 goto again;
1778                         if (err) {
1779                                 ret = err;
1780                                 goto done;
1781                         }
1782                         b = p->nodes[level];
1783                         slot = p->slots[level];
1784
1785                         /*
1786                          * slot 0 is special, if we change the key
1787                          * we have to update the parent pointer
1788                          * which means we must have a write lock
1789                          * on the parent
1790                          */
1791                         if (slot == 0 && cow &&
1792                             write_lock_level < level + 1) {
1793                                 write_lock_level = level + 1;
1794                                 btrfs_release_path(p);
1795                                 goto again;
1796                         }
1797
1798                         unlock_up(p, level, lowest_unlock);
1799
1800                         if (level == lowest_level) {
1801                                 if (dec)
1802                                         p->slots[level]++;
1803                                 goto done;
1804                         }
1805
1806                         err = read_block_for_search(trans, root, p,
1807                                                     &b, level, slot, key);
1808                         if (err == -EAGAIN)
1809                                 goto again;
1810                         if (err) {
1811                                 ret = err;
1812                                 goto done;
1813                         }
1814
1815                         if (!p->skip_locking) {
1816                                 level = btrfs_header_level(b);
1817                                 if (level <= write_lock_level) {
1818                                         err = btrfs_try_tree_write_lock(b);
1819                                         if (!err) {
1820                                                 btrfs_set_path_blocking(p);
1821                                                 btrfs_tree_lock(b);
1822                                                 btrfs_clear_path_blocking(p, b,
1823                                                                   BTRFS_WRITE_LOCK);
1824                                         }
1825                                         p->locks[level] = BTRFS_WRITE_LOCK;
1826                                 } else {
1827                                         err = btrfs_try_tree_read_lock(b);
1828                                         if (!err) {
1829                                                 btrfs_set_path_blocking(p);
1830                                                 btrfs_tree_read_lock(b);
1831                                                 btrfs_clear_path_blocking(p, b,
1832                                                                   BTRFS_READ_LOCK);
1833                                         }
1834                                         p->locks[level] = BTRFS_READ_LOCK;
1835                                 }
1836                                 p->nodes[level] = b;
1837                         }
1838                 } else {
1839                         p->slots[level] = slot;
1840                         if (ins_len > 0 &&
1841                             btrfs_leaf_free_space(root, b) < ins_len) {
1842                                 if (write_lock_level < 1) {
1843                                         write_lock_level = 1;
1844                                         btrfs_release_path(p);
1845                                         goto again;
1846                                 }
1847
1848                                 btrfs_set_path_blocking(p);
1849                                 err = split_leaf(trans, root, key,
1850                                                  p, ins_len, ret == 0);
1851                                 btrfs_clear_path_blocking(p, NULL, 0);
1852
1853                                 BUG_ON(err > 0);
1854                                 if (err) {
1855                                         ret = err;
1856                                         goto done;
1857                                 }
1858                         }
1859                         if (!p->search_for_split)
1860                                 unlock_up(p, level, lowest_unlock);
1861                         goto done;
1862                 }
1863         }
1864         ret = 1;
1865 done:
1866         /*
1867          * we don't really know what they plan on doing with the path
1868          * from here on, so for now just mark it as blocking
1869          */
1870         if (!p->leave_spinning)
1871                 btrfs_set_path_blocking(p);
1872         if (ret < 0)
1873                 btrfs_release_path(p);
1874         return ret;
1875 }
1876
1877 /*
1878  * adjust the pointers going up the tree, starting at level
1879  * making sure the right key of each node is points to 'key'.
1880  * This is used after shifting pointers to the left, so it stops
1881  * fixing up pointers when a given leaf/node is not in slot 0 of the
1882  * higher levels
1883  *
1884  * If this fails to write a tree block, it returns -1, but continues
1885  * fixing up the blocks in ram so the tree is consistent.
1886  */
1887 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1888                           struct btrfs_root *root, struct btrfs_path *path,
1889                           struct btrfs_disk_key *key, int level)
1890 {
1891         int i;
1892         int ret = 0;
1893         struct extent_buffer *t;
1894
1895         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1896                 int tslot = path->slots[i];
1897                 if (!path->nodes[i])
1898                         break;
1899                 t = path->nodes[i];
1900                 btrfs_set_node_key(t, key, tslot);
1901                 btrfs_mark_buffer_dirty(path->nodes[i]);
1902                 if (tslot != 0)
1903                         break;
1904         }
1905         return ret;
1906 }
1907
1908 /*
1909  * update item key.
1910  *
1911  * This function isn't completely safe. It's the caller's responsibility
1912  * that the new key won't break the order
1913  */
1914 int btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
1915                             struct btrfs_root *root, struct btrfs_path *path,
1916                             struct btrfs_key *new_key)
1917 {
1918         struct btrfs_disk_key disk_key;
1919         struct extent_buffer *eb;
1920         int slot;
1921
1922         eb = path->nodes[0];
1923         slot = path->slots[0];
1924         if (slot > 0) {
1925                 btrfs_item_key(eb, &disk_key, slot - 1);
1926                 if (comp_keys(&disk_key, new_key) >= 0)
1927                         return -1;
1928         }
1929         if (slot < btrfs_header_nritems(eb) - 1) {
1930                 btrfs_item_key(eb, &disk_key, slot + 1);
1931                 if (comp_keys(&disk_key, new_key) <= 0)
1932                         return -1;
1933         }
1934
1935         btrfs_cpu_key_to_disk(&disk_key, new_key);
1936         btrfs_set_item_key(eb, &disk_key, slot);
1937         btrfs_mark_buffer_dirty(eb);
1938         if (slot == 0)
1939                 fixup_low_keys(trans, root, path, &disk_key, 1);
1940         return 0;
1941 }
1942
1943 /*
1944  * try to push data from one node into the next node left in the
1945  * tree.
1946  *
1947  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1948  * error, and > 0 if there was no room in the left hand block.
1949  */
1950 static int push_node_left(struct btrfs_trans_handle *trans,
1951                           struct btrfs_root *root, struct extent_buffer *dst,
1952                           struct extent_buffer *src, int empty)
1953 {
1954         int push_items = 0;
1955         int src_nritems;
1956         int dst_nritems;
1957         int ret = 0;
1958
1959         src_nritems = btrfs_header_nritems(src);
1960         dst_nritems = btrfs_header_nritems(dst);
1961         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1962         WARN_ON(btrfs_header_generation(src) != trans->transid);
1963         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1964
1965         if (!empty && src_nritems <= 8)
1966                 return 1;
1967
1968         if (push_items <= 0)
1969                 return 1;
1970
1971         if (empty) {
1972                 push_items = min(src_nritems, push_items);
1973                 if (push_items < src_nritems) {
1974                         /* leave at least 8 pointers in the node if
1975                          * we aren't going to empty it
1976                          */
1977                         if (src_nritems - push_items < 8) {
1978                                 if (push_items <= 8)
1979                                         return 1;
1980                                 push_items -= 8;
1981                         }
1982                 }
1983         } else
1984                 push_items = min(src_nritems - 8, push_items);
1985
1986         copy_extent_buffer(dst, src,
1987                            btrfs_node_key_ptr_offset(dst_nritems),
1988                            btrfs_node_key_ptr_offset(0),
1989                            push_items * sizeof(struct btrfs_key_ptr));
1990
1991         if (push_items < src_nritems) {
1992                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1993                                       btrfs_node_key_ptr_offset(push_items),
1994                                       (src_nritems - push_items) *
1995                                       sizeof(struct btrfs_key_ptr));
1996         }
1997         btrfs_set_header_nritems(src, src_nritems - push_items);
1998         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1999         btrfs_mark_buffer_dirty(src);
2000         btrfs_mark_buffer_dirty(dst);
2001
2002         return ret;
2003 }
2004
2005 /*
2006  * try to push data from one node into the next node right in the
2007  * tree.
2008  *
2009  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
2010  * error, and > 0 if there was no room in the right hand block.
2011  *
2012  * this will  only push up to 1/2 the contents of the left node over
2013  */
2014 static int balance_node_right(struct btrfs_trans_handle *trans,
2015                               struct btrfs_root *root,
2016                               struct extent_buffer *dst,
2017                               struct extent_buffer *src)
2018 {
2019         int push_items = 0;
2020         int max_push;
2021         int src_nritems;
2022         int dst_nritems;
2023         int ret = 0;
2024
2025         WARN_ON(btrfs_header_generation(src) != trans->transid);
2026         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2027
2028         src_nritems = btrfs_header_nritems(src);
2029         dst_nritems = btrfs_header_nritems(dst);
2030         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2031         if (push_items <= 0)
2032                 return 1;
2033
2034         if (src_nritems < 4)
2035                 return 1;
2036
2037         max_push = src_nritems / 2 + 1;
2038         /* don't try to empty the node */
2039         if (max_push >= src_nritems)
2040                 return 1;
2041
2042         if (max_push < push_items)
2043                 push_items = max_push;
2044
2045         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
2046                                       btrfs_node_key_ptr_offset(0),
2047                                       (dst_nritems) *
2048                                       sizeof(struct btrfs_key_ptr));
2049
2050         copy_extent_buffer(dst, src,
2051                            btrfs_node_key_ptr_offset(0),
2052                            btrfs_node_key_ptr_offset(src_nritems - push_items),
2053                            push_items * sizeof(struct btrfs_key_ptr));
2054
2055         btrfs_set_header_nritems(src, src_nritems - push_items);
2056         btrfs_set_header_nritems(dst, dst_nritems + push_items);
2057
2058         btrfs_mark_buffer_dirty(src);
2059         btrfs_mark_buffer_dirty(dst);
2060
2061         return ret;
2062 }
2063
2064 /*
2065  * helper function to insert a new root level in the tree.
2066  * A new node is allocated, and a single item is inserted to
2067  * point to the existing root
2068  *
2069  * returns zero on success or < 0 on failure.
2070  */
2071 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
2072                            struct btrfs_root *root,
2073                            struct btrfs_path *path, int level)
2074 {
2075         u64 lower_gen;
2076         struct extent_buffer *lower;
2077         struct extent_buffer *c;
2078         struct extent_buffer *old;
2079         struct btrfs_disk_key lower_key;
2080
2081         BUG_ON(path->nodes[level]);
2082         BUG_ON(path->nodes[level-1] != root->node);
2083
2084         lower = path->nodes[level-1];
2085         if (level == 1)
2086                 btrfs_item_key(lower, &lower_key, 0);
2087         else
2088                 btrfs_node_key(lower, &lower_key, 0);
2089
2090         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2091                                    root->root_key.objectid, &lower_key,
2092                                    level, root->node->start, 0, 0);
2093         if (IS_ERR(c))
2094                 return PTR_ERR(c);
2095
2096         root_add_used(root, root->nodesize);
2097
2098         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
2099         btrfs_set_header_nritems(c, 1);
2100         btrfs_set_header_level(c, level);
2101         btrfs_set_header_bytenr(c, c->start);
2102         btrfs_set_header_generation(c, trans->transid);
2103         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
2104         btrfs_set_header_owner(c, root->root_key.objectid);
2105
2106         write_extent_buffer(c, root->fs_info->fsid,
2107                             (unsigned long)btrfs_header_fsid(c),
2108                             BTRFS_FSID_SIZE);
2109
2110         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
2111                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
2112                             BTRFS_UUID_SIZE);
2113
2114         btrfs_set_node_key(c, &lower_key, 0);
2115         btrfs_set_node_blockptr(c, 0, lower->start);
2116         lower_gen = btrfs_header_generation(lower);
2117         WARN_ON(lower_gen != trans->transid);
2118
2119         btrfs_set_node_ptr_generation(c, 0, lower_gen);
2120
2121         btrfs_mark_buffer_dirty(c);
2122
2123         old = root->node;
2124         rcu_assign_pointer(root->node, c);
2125
2126         /* the super has an extra ref to root->node */
2127         free_extent_buffer(old);
2128
2129         add_root_to_dirty_list(root);
2130         extent_buffer_get(c);
2131         path->nodes[level] = c;
2132         path->locks[level] = BTRFS_WRITE_LOCK;
2133         path->slots[level] = 0;
2134         return 0;
2135 }
2136
2137 /*
2138  * worker function to insert a single pointer in a node.
2139  * the node should have enough room for the pointer already
2140  *
2141  * slot and level indicate where you want the key to go, and
2142  * blocknr is the block the key points to.
2143  *
2144  * returns zero on success and < 0 on any error
2145  */
2146 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
2147                       *root, struct btrfs_path *path, struct btrfs_disk_key
2148                       *key, u64 bytenr, int slot, int level)
2149 {
2150         struct extent_buffer *lower;
2151         int nritems;
2152
2153         BUG_ON(!path->nodes[level]);
2154         btrfs_assert_tree_locked(path->nodes[level]);
2155         lower = path->nodes[level];
2156         nritems = btrfs_header_nritems(lower);
2157         BUG_ON(slot > nritems);
2158         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
2159                 BUG();
2160         if (slot != nritems) {
2161                 memmove_extent_buffer(lower,
2162                               btrfs_node_key_ptr_offset(slot + 1),
2163                               btrfs_node_key_ptr_offset(slot),
2164                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
2165         }
2166         btrfs_set_node_key(lower, key, slot);
2167         btrfs_set_node_blockptr(lower, slot, bytenr);
2168         WARN_ON(trans->transid == 0);
2169         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
2170         btrfs_set_header_nritems(lower, nritems + 1);
2171         btrfs_mark_buffer_dirty(lower);
2172         return 0;
2173 }
2174
2175 /*
2176  * split the node at the specified level in path in two.
2177  * The path is corrected to point to the appropriate node after the split
2178  *
2179  * Before splitting this tries to make some room in the node by pushing
2180  * left and right, if either one works, it returns right away.
2181  *
2182  * returns 0 on success and < 0 on failure
2183  */
2184 static noinline int split_node(struct btrfs_trans_handle *trans,
2185                                struct btrfs_root *root,
2186                                struct btrfs_path *path, int level)
2187 {
2188         struct extent_buffer *c;
2189         struct extent_buffer *split;
2190         struct btrfs_disk_key disk_key;
2191         int mid;
2192         int ret;
2193         int wret;
2194         u32 c_nritems;
2195
2196         c = path->nodes[level];
2197         WARN_ON(btrfs_header_generation(c) != trans->transid);
2198         if (c == root->node) {
2199                 /* trying to split the root, lets make a new one */
2200                 ret = insert_new_root(trans, root, path, level + 1);
2201                 if (ret)
2202                         return ret;
2203         } else {
2204                 ret = push_nodes_for_insert(trans, root, path, level);
2205                 c = path->nodes[level];
2206                 if (!ret && btrfs_header_nritems(c) <
2207                     BTRFS_NODEPTRS_PER_BLOCK(root) - 3)
2208                         return 0;
2209                 if (ret < 0)
2210                         return ret;
2211         }
2212
2213         c_nritems = btrfs_header_nritems(c);
2214         mid = (c_nritems + 1) / 2;
2215         btrfs_node_key(c, &disk_key, mid);
2216
2217         split = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
2218                                         root->root_key.objectid,
2219                                         &disk_key, level, c->start, 0, 0);
2220         if (IS_ERR(split))
2221                 return PTR_ERR(split);
2222
2223         root_add_used(root, root->nodesize);
2224
2225         memset_extent_buffer(split, 0, 0, sizeof(struct btrfs_header));
2226         btrfs_set_header_level(split, btrfs_header_level(c));
2227         btrfs_set_header_bytenr(split, split->start);
2228         btrfs_set_header_generation(split, trans->transid);
2229         btrfs_set_header_backref_rev(split, BTRFS_MIXED_BACKREF_REV);
2230         btrfs_set_header_owner(split, root->root_key.objectid);
2231         write_extent_buffer(split, root->fs_info->fsid,
2232                             (unsigned long)btrfs_header_fsid(split),
2233                             BTRFS_FSID_SIZE);
2234         write_extent_buffer(split, root->fs_info->chunk_tree_uuid,
2235                             (unsigned long)btrfs_header_chunk_tree_uuid(split),
2236                             BTRFS_UUID_SIZE);
2237
2238
2239         copy_extent_buffer(split, c,
2240                            btrfs_node_key_ptr_offset(0),
2241                            btrfs_node_key_ptr_offset(mid),
2242                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
2243         btrfs_set_header_nritems(split, c_nritems - mid);
2244         btrfs_set_header_nritems(c, mid);
2245         ret = 0;
2246
2247         btrfs_mark_buffer_dirty(c);
2248         btrfs_mark_buffer_dirty(split);
2249
2250         wret = insert_ptr(trans, root, path, &disk_key, split->start,
2251                           path->slots[level + 1] + 1,
2252                           level + 1);
2253         if (wret)
2254                 ret = wret;
2255
2256         if (path->slots[level] >= mid) {
2257                 path->slots[level] -= mid;
2258                 btrfs_tree_unlock(c);
2259                 free_extent_buffer(c);
2260                 path->nodes[level] = split;
2261                 path->slots[level + 1] += 1;
2262         } else {
2263                 btrfs_tree_unlock(split);
2264                 free_extent_buffer(split);
2265         }
2266         return ret;
2267 }
2268
2269 /*
2270  * how many bytes are required to store the items in a leaf.  start
2271  * and nr indicate which items in the leaf to check.  This totals up the
2272  * space used both by the item structs and the item data
2273  */
2274 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
2275 {
2276         int data_len;
2277         int nritems = btrfs_header_nritems(l);
2278         int end = min(nritems, start + nr) - 1;
2279
2280         if (!nr)
2281                 return 0;
2282         data_len = btrfs_item_end_nr(l, start);
2283         data_len = data_len - btrfs_item_offset_nr(l, end);
2284         data_len += sizeof(struct btrfs_item) * nr;
2285         WARN_ON(data_len < 0);
2286         return data_len;
2287 }
2288
2289 /*
2290  * The space between the end of the leaf items and
2291  * the start of the leaf data.  IOW, how much room
2292  * the leaf has left for both items and data
2293  */
2294 noinline int btrfs_leaf_free_space(struct btrfs_root *root,
2295                                    struct extent_buffer *leaf)
2296 {
2297         int nritems = btrfs_header_nritems(leaf);
2298         int ret;
2299         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
2300         if (ret < 0) {
2301                 printk(KERN_CRIT "leaf free space ret %d, leaf data size %lu, "
2302                        "used %d nritems %d\n",
2303                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
2304                        leaf_space_used(leaf, 0, nritems), nritems);
2305         }
2306         return ret;
2307 }
2308
2309 /*
2310  * min slot controls the lowest index we're willing to push to the
2311  * right.  We'll push up to and including min_slot, but no lower
2312  */
2313 static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,
2314                                       struct btrfs_root *root,
2315                                       struct btrfs_path *path,
2316                                       int data_size, int empty,
2317                                       struct extent_buffer *right,
2318                                       int free_space, u32 left_nritems,
2319                                       u32 min_slot)
2320 {
2321         struct extent_buffer *left = path->nodes[0];
2322         struct extent_buffer *upper = path->nodes[1];
2323         struct btrfs_disk_key disk_key;
2324         int slot;
2325         u32 i;
2326         int push_space = 0;
2327         int push_items = 0;
2328         struct btrfs_item *item;
2329         u32 nr;
2330         u32 right_nritems;
2331         u32 data_end;
2332         u32 this_item_size;
2333
2334         if (empty)
2335                 nr = 0;
2336         else
2337                 nr = max_t(u32, 1, min_slot);
2338
2339         if (path->slots[0] >= left_nritems)
2340                 push_space += data_size;
2341
2342         slot = path->slots[1];
2343         i = left_nritems - 1;
2344         while (i >= nr) {
2345                 item = btrfs_item_nr(left, i);
2346
2347                 if (!empty && push_items > 0) {
2348                         if (path->slots[0] > i)
2349                                 break;
2350                         if (path->slots[0] == i) {
2351                                 int space = btrfs_leaf_free_space(root, left);
2352                                 if (space + push_space * 2 > free_space)
2353                                         break;
2354                         }
2355                 }
2356
2357                 if (path->slots[0] == i)
2358                         push_space += data_size;
2359
2360                 this_item_size = btrfs_item_size(left, item);
2361                 if (this_item_size + sizeof(*item) + push_space > free_space)
2362                         break;
2363
2364                 push_items++;
2365                 push_space += this_item_size + sizeof(*item);
2366                 if (i == 0)
2367                         break;
2368                 i--;
2369         }
2370
2371         if (push_items == 0)
2372                 goto out_unlock;
2373
2374         if (!empty && push_items == left_nritems)
2375                 WARN_ON(1);
2376
2377         /* push left to right */
2378         right_nritems = btrfs_header_nritems(right);
2379
2380         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
2381         push_space -= leaf_data_end(root, left);
2382
2383         /* make room in the right data area */
2384         data_end = leaf_data_end(root, right);
2385         memmove_extent_buffer(right,
2386                               btrfs_leaf_data(right) + data_end - push_space,
2387                               btrfs_leaf_data(right) + data_end,
2388                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
2389
2390         /* copy from the left data area */
2391         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
2392                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
2393                      btrfs_leaf_data(left) + leaf_data_end(root, left),
2394                      push_space);
2395
2396         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
2397                               btrfs_item_nr_offset(0),
2398                               right_nritems * sizeof(struct btrfs_item));
2399
2400         /* copy the items from left to right */
2401         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
2402                    btrfs_item_nr_offset(left_nritems - push_items),
2403                    push_items * sizeof(struct btrfs_item));
2404
2405         /* update the item pointers */
2406         right_nritems += push_items;
2407         btrfs_set_header_nritems(right, right_nritems);
2408         push_space = BTRFS_LEAF_DATA_SIZE(root);
2409         for (i = 0; i < right_nritems; i++) {
2410                 item = btrfs_item_nr(right, i);
2411                 push_space -= btrfs_item_size(right, item);
2412                 btrfs_set_item_offset(right, item, push_space);
2413         }
2414
2415         left_nritems -= push_items;
2416         btrfs_set_header_nritems(left, left_nritems);
2417
2418         if (left_nritems)
2419                 btrfs_mark_buffer_dirty(left);
2420         else
2421                 clean_tree_block(trans, root, left);
2422
2423         btrfs_mark_buffer_dirty(right);
2424
2425         btrfs_item_key(right, &disk_key, 0);
2426         btrfs_set_node_key(upper, &disk_key, slot + 1);
2427         btrfs_mark_buffer_dirty(upper);
2428
2429         /* then fixup the leaf pointer in the path */
2430         if (path->slots[0] >= left_nritems) {
2431                 path->slots[0] -= left_nritems;
2432                 if (btrfs_header_nritems(path->nodes[0]) == 0)
2433                         clean_tree_block(trans, root, path->nodes[0]);
2434                 btrfs_tree_unlock(path->nodes[0]);
2435                 free_extent_buffer(path->nodes[0]);
2436                 path->nodes[0] = right;
2437                 path->slots[1] += 1;
2438         } else {
2439                 btrfs_tree_unlock(right);
2440                 free_extent_buffer(right);
2441         }
2442         return 0;
2443
2444 out_unlock:
2445         btrfs_tree_unlock(right);
2446         free_extent_buffer(right);
2447         return 1;
2448 }
2449
2450 /*
2451  * push some data in the path leaf to the right, trying to free up at
2452  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2453  *
2454  * returns 1 if the push failed because the other node didn't have enough
2455  * room, 0 if everything worked out and < 0 if there were major errors.
2456  *
2457  * this will push starting from min_slot to the end of the leaf.  It won't
2458  * push any slot lower than min_slot
2459  */
2460 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
2461                            *root, struct btrfs_path *path,
2462                            int min_data_size, int data_size,
2463                            int empty, u32 min_slot)
2464 {
2465         struct extent_buffer *left = path->nodes[0];
2466         struct extent_buffer *right;
2467         struct extent_buffer *upper;
2468         int slot;
2469         int free_space;
2470         u32 left_nritems;
2471         int ret;
2472
2473         if (!path->nodes[1])
2474                 return 1;
2475
2476         slot = path->slots[1];
2477         upper = path->nodes[1];
2478         if (slot >= btrfs_header_nritems(upper) - 1)
2479                 return 1;
2480
2481         btrfs_assert_tree_locked(path->nodes[1]);
2482
2483         right = read_node_slot(root, upper, slot + 1);
2484         if (right == NULL)
2485                 return 1;
2486
2487         btrfs_tree_lock(right);
2488         btrfs_set_lock_blocking(right);
2489
2490         free_space = btrfs_leaf_free_space(root, right);
2491         if (free_space < data_size)
2492                 goto out_unlock;
2493
2494         /* cow and double check */
2495         ret = btrfs_cow_block(trans, root, right, upper,
2496                               slot + 1, &right);
2497         if (ret)
2498                 goto out_unlock;
2499
2500         free_space = btrfs_leaf_free_space(root, right);
2501         if (free_space < data_size)
2502                 goto out_unlock;
2503
2504         left_nritems = btrfs_header_nritems(left);
2505         if (left_nritems == 0)
2506                 goto out_unlock;
2507
2508         return __push_leaf_right(trans, root, path, min_data_size, empty,
2509                                 right, free_space, left_nritems, min_slot);
2510 out_unlock:
2511         btrfs_tree_unlock(right);
2512         free_extent_buffer(right);
2513         return 1;
2514 }
2515
2516 /*
2517  * push some data in the path leaf to the left, trying to free up at
2518  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2519  *
2520  * max_slot can put a limit on how far into the leaf we'll push items.  The
2521  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us do all the
2522  * items
2523  */
2524 static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,
2525                                      struct btrfs_root *root,
2526                                      struct btrfs_path *path, int data_size,
2527                                      int empty, struct extent_buffer *left,
2528                                      int free_space, u32 right_nritems,
2529                                      u32 max_slot)
2530 {
2531         struct btrfs_disk_key disk_key;
2532         struct extent_buffer *right = path->nodes[0];
2533         int i;
2534         int push_space = 0;
2535         int push_items = 0;
2536         struct btrfs_item *item;
2537         u32 old_left_nritems;
2538         u32 nr;
2539         int ret = 0;
2540         int wret;
2541         u32 this_item_size;
2542         u32 old_left_item_size;
2543
2544         if (empty)
2545                 nr = min(right_nritems, max_slot);
2546         else
2547                 nr = min(right_nritems - 1, max_slot);
2548
2549         for (i = 0; i < nr; i++) {
2550                 item = btrfs_item_nr(right, i);
2551
2552                 if (!empty && push_items > 0) {
2553                         if (path->slots[0] < i)
2554                                 break;
2555                         if (path->slots[0] == i) {
2556                                 int space = btrfs_leaf_free_space(root, right);
2557                                 if (space + push_space * 2 > free_space)
2558                                         break;
2559                         }
2560                 }
2561
2562                 if (path->slots[0] == i)
2563                         push_space += data_size;
2564
2565                 this_item_size = btrfs_item_size(right, item);
2566                 if (this_item_size + sizeof(*item) + push_space > free_space)
2567                         break;
2568
2569                 push_items++;
2570                 push_space += this_item_size + sizeof(*item);
2571         }
2572
2573         if (push_items == 0) {
2574                 ret = 1;
2575                 goto out;
2576         }
2577         if (!empty && push_items == btrfs_header_nritems(right))
2578                 WARN_ON(1);
2579
2580         /* push data from right to left */
2581         copy_extent_buffer(left, right,
2582                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
2583                            btrfs_item_nr_offset(0),
2584                            push_items * sizeof(struct btrfs_item));
2585
2586         push_space = BTRFS_LEAF_DATA_SIZE(root) -
2587                      btrfs_item_offset_nr(right, push_items - 1);
2588
2589         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
2590                      leaf_data_end(root, left) - push_space,
2591                      btrfs_leaf_data(right) +
2592                      btrfs_item_offset_nr(right, push_items - 1),
2593                      push_space);
2594         old_left_nritems = btrfs_header_nritems(left);
2595         BUG_ON(old_left_nritems <= 0);
2596
2597         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
2598         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
2599                 u32 ioff;
2600
2601                 item = btrfs_item_nr(left, i);
2602
2603                 ioff = btrfs_item_offset(left, item);
2604                 btrfs_set_item_offset(left, item,
2605                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
2606         }
2607         btrfs_set_header_nritems(left, old_left_nritems + push_items);
2608
2609         /* fixup right node */
2610         if (push_items > right_nritems) {
2611                 printk(KERN_CRIT "push items %d nr %u\n", push_items,
2612                        right_nritems);
2613                 WARN_ON(1);
2614         }
2615
2616         if (push_items < right_nritems) {
2617                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
2618                                                   leaf_data_end(root, right);
2619                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
2620                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
2621                                       btrfs_leaf_data(right) +
2622                                       leaf_data_end(root, right), push_space);
2623
2624                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
2625                               btrfs_item_nr_offset(push_items),
2626                              (btrfs_header_nritems(right) - push_items) *
2627                              sizeof(struct btrfs_item));
2628         }
2629         right_nritems -= push_items;
2630         btrfs_set_header_nritems(right, right_nritems);
2631         push_space = BTRFS_LEAF_DATA_SIZE(root);
2632         for (i = 0; i < right_nritems; i++) {
2633                 item = btrfs_item_nr(right, i);
2634
2635                 push_space = push_space - btrfs_item_size(right, item);
2636                 btrfs_set_item_offset(right, item, push_space);
2637         }
2638
2639         btrfs_mark_buffer_dirty(left);
2640         if (right_nritems)
2641                 btrfs_mark_buffer_dirty(right);
2642         else
2643                 clean_tree_block(trans, root, right);
2644
2645         btrfs_item_key(right, &disk_key, 0);
2646         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
2647         if (wret)
2648                 ret = wret;
2649
2650         /* then fixup the leaf pointer in the path */
2651         if (path->slots[0] < push_items) {
2652                 path->slots[0] += old_left_nritems;
2653                 btrfs_tree_unlock(path->nodes[0]);
2654                 free_extent_buffer(path->nodes[0]);
2655                 path->nodes[0] = left;
2656                 path->slots[1] -= 1;
2657         } else {
2658                 btrfs_tree_unlock(left);
2659                 free_extent_buffer(left);
2660                 path->slots[0] -= push_items;
2661         }
2662         BUG_ON(path->slots[0] < 0);
2663         return ret;
2664 out:
2665         btrfs_tree_unlock(left);
2666         free_extent_buffer(left);
2667         return ret;
2668 }
2669
2670 /*
2671  * push some data in the path leaf to the left, trying to free up at
2672  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
2673  *
2674  * max_slot can put a limit on how far into the leaf we'll push items.  The
2675  * item at 'max_slot' won't be touched.  Use (u32)-1 to make us push all the
2676  * items
2677  */
2678 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
2679                           *root, struct btrfs_path *path, int min_data_size,
2680                           int data_size, int empty, u32 max_slot)
2681 {
2682         struct extent_buffer *right = path->nodes[0];
2683         struct extent_buffer *left;
2684         int slot;
2685         int free_space;
2686         u32 right_nritems;
2687         int ret = 0;
2688
2689         slot = path->slots[1];
2690         if (slot == 0)
2691                 return 1;
2692         if (!path->nodes[1])
2693                 return 1;
2694
2695         right_nritems = btrfs_header_nritems(right);
2696         if (right_nritems == 0)
2697                 return 1;
2698
2699         btrfs_assert_tree_locked(path->nodes[1]);
2700
2701         left = read_node_slot(root, path->nodes[1], slot - 1);
2702         if (left == NULL)
2703                 return 1;
2704
2705         btrfs_tree_lock(left);
2706         btrfs_set_lock_blocking(left);
2707
2708         free_space = btrfs_leaf_free_space(root, left);
2709         if (free_space < data_size) {
2710                 ret = 1;
2711                 goto out;
2712         }
2713
2714         /* cow and double check */
2715         ret = btrfs_cow_block(trans, root, left,
2716                               path->nodes[1], slot - 1, &left);
2717         if (ret) {
2718                 /* we hit -ENOSPC, but it isn't fatal here */
2719                 ret = 1;
2720                 goto out;
2721         }
2722
2723         free_space = btrfs_leaf_free_space(root, left);
2724         if (free_space < data_size) {
2725                 ret = 1;
2726                 goto out;
2727         }
2728
2729         return __push_leaf_left(trans, root, path, min_data_size,
2730                                empty, left, free_space, right_nritems,
2731                                max_slot);
2732 out:
2733         btrfs_tree_unlock(left);
2734         free_extent_buffer(left);
2735         return ret;
2736 }
2737
2738 /*
2739  * split the path's leaf in two, making sure there is at least data_size
2740  * available for the resulting leaf level of the path.
2741  *
2742  * returns 0 if all went well and < 0 on failure.
2743  */
2744 static noinline int copy_for_split(struct btrfs_trans_handle *trans,
2745                                struct btrfs_root *root,
2746                                struct btrfs_path *path,
2747                                struct extent_buffer *l,
2748                                struct extent_buffer *right,
2749                                int slot, int mid, int nritems)
2750 {
2751         int data_copy_size;
2752         int rt_data_off;
2753         int i;
2754         int ret = 0;
2755         int wret;
2756         struct btrfs_disk_key disk_key;
2757
2758         nritems = nritems - mid;
2759         btrfs_set_header_nritems(right, nritems);
2760         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2761
2762         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2763                            btrfs_item_nr_offset(mid),
2764                            nritems * sizeof(struct btrfs_item));
2765
2766         copy_extent_buffer(right, l,
2767                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2768                      data_copy_size, btrfs_leaf_data(l) +
2769                      leaf_data_end(root, l), data_copy_size);
2770
2771         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2772                       btrfs_item_end_nr(l, mid);
2773
2774         for (i = 0; i < nritems; i++) {
2775                 struct btrfs_item *item = btrfs_item_nr(right, i);
2776                 u32 ioff;
2777
2778                 ioff = btrfs_item_offset(right, item);
2779                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2780         }
2781
2782         btrfs_set_header_nritems(l, mid);
2783         ret = 0;
2784         btrfs_item_key(right, &disk_key, 0);
2785         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2786                           path->slots[1] + 1, 1);
2787         if (wret)
2788                 ret = wret;
2789
2790         btrfs_mark_buffer_dirty(right);
2791         btrfs_mark_buffer_dirty(l);
2792         BUG_ON(path->slots[0] != slot);
2793
2794         if (mid <= slot) {
2795                 btrfs_tree_unlock(path->nodes[0]);
2796                 free_extent_buffer(path->nodes[0]);
2797                 path->nodes[0] = right;
2798                 path->slots[0] -= mid;
2799                 path->slots[1] += 1;
2800         } else {
2801                 btrfs_tree_unlock(right);
2802                 free_extent_buffer(right);
2803         }
2804
2805         BUG_ON(path->slots[0] < 0);
2806
2807         return ret;
2808 }
2809
2810 /*
2811  * double splits happen when we need to insert a big item in the middle
2812  * of a leaf.  A double split can leave us with 3 mostly empty leaves:
2813  * leaf: [ slots 0 - N] [ our target ] [ N + 1 - total in leaf ]
2814  *          A                 B                 C
2815  *
2816  * We avoid this by trying to push the items on either side of our target
2817  * into the adjacent leaves.  If all goes well we can avoid the double split
2818  * completely.
2819  */
2820 static noinline int push_for_double_split(struct btrfs_trans_handle *trans,
2821                                           struct btrfs_root *root,
2822                                           struct btrfs_path *path,
2823                                           int data_size)
2824 {
2825         int ret;
2826         int progress = 0;
2827         int slot;
2828         u32 nritems;
2829
2830         slot = path->slots[0];
2831
2832         /*
2833          * try to push all the items after our slot into the
2834          * right leaf
2835          */
2836         ret = push_leaf_right(trans, root, path, 1, data_size, 0, slot);
2837         if (ret < 0)
2838                 return ret;
2839
2840         if (ret == 0)
2841                 progress++;
2842
2843         nritems = btrfs_header_nritems(path->nodes[0]);
2844         /*
2845          * our goal is to get our slot at the start or end of a leaf.  If
2846          * we've done so we're done
2847          */
2848         if (path->slots[0] == 0 || path->slots[0] == nritems)
2849                 return 0;
2850
2851         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
2852                 return 0;
2853
2854         /* try to push all the items before our slot into the next leaf */
2855         slot = path->slots[0];
2856         ret = push_leaf_left(trans, root, path, 1, data_size, 0, slot);
2857         if (ret < 0)
2858                 return ret;
2859
2860         if (ret == 0)
2861                 progress++;
2862
2863         if (progress)
2864                 return 0;
2865         return 1;
2866 }
2867
2868 /*
2869  * split the path's leaf in two, making sure there is at least data_size
2870  * available for the resulting leaf level of the path.
2871  *
2872  * returns 0 if all went well and < 0 on failure.
2873  */
2874 static noinline int split_leaf(struct btrfs_trans_handle *trans,
2875                                struct btrfs_root *root,
2876                                struct btrfs_key *ins_key,
2877                                struct btrfs_path *path, int data_size,
2878                                int extend)
2879 {
2880         struct btrfs_disk_key disk_key;
2881         struct extent_buffer *l;
2882         u32 nritems;
2883         int mid;
2884         int slot;
2885         struct extent_buffer *right;
2886         int ret = 0;
2887         int wret;
2888         int split;
2889         int num_doubles = 0;
2890         int tried_avoid_double = 0;
2891
2892         l = path->nodes[0];
2893         slot = path->slots[0];
2894         if (extend && data_size + btrfs_item_size_nr(l, slot) +
2895             sizeof(struct btrfs_item) > BTRFS_LEAF_DATA_SIZE(root))
2896                 return -EOVERFLOW;
2897
2898         /* first try to make some room by pushing left and right */
2899         if (data_size) {
2900                 wret = push_leaf_right(trans, root, path, data_size,
2901                                        data_size, 0, 0);
2902                 if (wret < 0)
2903                         return wret;
2904                 if (wret) {
2905                         wret = push_leaf_left(trans, root, path, data_size,
2906                                               data_size, 0, (u32)-1);
2907                         if (wret < 0)
2908                                 return wret;
2909                 }
2910                 l = path->nodes[0];
2911
2912                 /* did the pushes work? */
2913                 if (btrfs_leaf_free_space(root, l) >= data_size)
2914                         return 0;
2915         }
2916
2917         if (!path->nodes[1]) {
2918                 ret = insert_new_root(trans, root, path, 1);
2919                 if (ret)
2920                         return ret;
2921         }
2922 again:
2923         split = 1;
2924         l = path->nodes[0];
2925         slot = path->slots[0];
2926         nritems = btrfs_header_nritems(l);
2927         mid = (nritems + 1) / 2;
2928
2929         if (mid <= slot) {
2930                 if (nritems == 1 ||
2931                     leaf_space_used(l, mid, nritems - mid) + data_size >
2932                         BTRFS_LEAF_DATA_SIZE(root)) {
2933                         if (slot >= nritems) {
2934                                 split = 0;
2935                         } else {
2936                                 mid = slot;
2937                                 if (mid != nritems &&
2938                                     leaf_space_used(l, mid, nritems - mid) +
2939                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2940                                         if (data_size && !tried_avoid_double)
2941                                                 goto push_for_double;
2942                                         split = 2;
2943                                 }
2944                         }
2945                 }
2946         } else {
2947                 if (leaf_space_used(l, 0, mid) + data_size >
2948                         BTRFS_LEAF_DATA_SIZE(root)) {
2949                         if (!extend && data_size && slot == 0) {
2950                                 split = 0;
2951                         } else if ((extend || !data_size) && slot == 0) {
2952                                 mid = 1;
2953                         } else {
2954                                 mid = slot;
2955                                 if (mid != nritems &&
2956                                     leaf_space_used(l, mid, nritems - mid) +
2957                                     data_size > BTRFS_LEAF_DATA_SIZE(root)) {
2958                                         if (data_size && !tried_avoid_double)
2959                                                 goto push_for_double;
2960                                         split = 2 ;
2961                                 }
2962                         }
2963                 }
2964         }
2965
2966         if (split == 0)
2967                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
2968         else
2969                 btrfs_item_key(l, &disk_key, mid);
2970
2971         right = btrfs_alloc_free_block(trans, root, root->leafsize, 0,
2972                                         root->root_key.objectid,
2973                                         &disk_key, 0, l->start, 0, 0);
2974         if (IS_ERR(right))
2975                 return PTR_ERR(right);
2976
2977         root_add_used(root, root->leafsize);
2978
2979         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
2980         btrfs_set_header_bytenr(right, right->start);
2981         btrfs_set_header_generation(right, trans->transid);
2982         btrfs_set_header_backref_rev(right, BTRFS_MIXED_BACKREF_REV);
2983         btrfs_set_header_owner(right, root->root_key.objectid);
2984         btrfs_set_header_level(right, 0);
2985         write_extent_buffer(right, root->fs_info->fsid,
2986                             (unsigned long)btrfs_header_fsid(right),
2987                             BTRFS_FSID_SIZE);
2988
2989         write_extent_buffer(right, root->fs_info->chunk_tree_uuid,
2990                             (unsigned long)btrfs_header_chunk_tree_uuid(right),
2991                             BTRFS_UUID_SIZE);
2992
2993         if (split == 0) {
2994                 if (mid <= slot) {
2995                         btrfs_set_header_nritems(right, 0);
2996                         wret = insert_ptr(trans, root, path,
2997                                           &disk_key, right->start,
2998                                           path->slots[1] + 1, 1);
2999                         if (wret)
3000                                 ret = wret;
3001
3002                         btrfs_tree_unlock(path->nodes[0]);
3003                         free_extent_buffer(path->nodes[0]);
3004                         path->nodes[0] = right;
3005                         path->slots[0] = 0;
3006                         path->slots[1] += 1;
3007                 } else {
3008                         btrfs_set_header_nritems(right, 0);
3009                         wret = insert_ptr(trans, root, path,
3010                                           &disk_key,
3011                                           right->start,
3012                                           path->slots[1], 1);
3013                         if (wret)
3014                                 ret = wret;
3015                         btrfs_tree_unlock(path->nodes[0]);
3016                         free_extent_buffer(path->nodes[0]);
3017                         path->nodes[0] = right;
3018                         path->slots[0] = 0;
3019                         if (path->slots[1] == 0) {
3020                                 wret = fixup_low_keys(trans, root,
3021                                                 path, &disk_key, 1);
3022                                 if (wret)
3023                                         ret = wret;
3024                         }
3025                 }
3026                 btrfs_mark_buffer_dirty(right);
3027                 return ret;
3028         }
3029
3030         ret = copy_for_split(trans, root, path, l, right, slot, mid, nritems);
3031         BUG_ON(ret);
3032
3033         if (split == 2) {
3034                 BUG_ON(num_doubles != 0);
3035                 num_doubles++;
3036                 goto again;
3037         }
3038
3039         return ret;
3040
3041 push_for_double:
3042         push_for_double_split(trans, root, path, data_size);
3043         tried_avoid_double = 1;
3044         if (btrfs_leaf_free_space(root, path->nodes[0]) >= data_size)
3045                 return 0;
3046         goto again;
3047 }
3048
3049 static noinline int setup_leaf_for_split(struct btrfs_trans_handle *trans,
3050                                          struct btrfs_root *root,
3051                                          struct btrfs_path *path, int ins_len)
3052 {
3053         struct btrfs_key key;
3054         struct extent_buffer *leaf;
3055         struct btrfs_file_extent_item *fi;
3056         u64 extent_len = 0;
3057         u32 item_size;
3058         int ret;
3059
3060         leaf = path->nodes[0];
3061         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3062
3063         BUG_ON(key.type != BTRFS_EXTENT_DATA_KEY &&
3064                key.type != BTRFS_EXTENT_CSUM_KEY);
3065
3066         if (btrfs_leaf_free_space(root, leaf) >= ins_len)
3067                 return 0;
3068
3069         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3070         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3071                 fi = btrfs_item_ptr(leaf, path->slots[0],
3072                                     struct btrfs_file_extent_item);
3073                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
3074         }
3075         btrfs_release_path(path);
3076
3077         path->keep_locks = 1;
3078         path->search_for_split = 1;
3079         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3080         path->search_for_split = 0;
3081         if (ret < 0)
3082                 goto err;
3083
3084         ret = -EAGAIN;
3085         leaf = path->nodes[0];
3086         /* if our item isn't there or got smaller, return now */
3087         if (ret > 0 || item_size != btrfs_item_size_nr(leaf, path->slots[0]))
3088                 goto err;
3089
3090         /* the leaf has  changed, it now has room.  return now */
3091         if (btrfs_leaf_free_space(root, path->nodes[0]) >= ins_len)
3092                 goto err;
3093
3094         if (key.type == BTRFS_EXTENT_DATA_KEY) {
3095                 fi = btrfs_item_ptr(leaf, path->slots[0],
3096                                     struct btrfs_file_extent_item);
3097                 if (extent_len != btrfs_file_extent_num_bytes(leaf, fi))
3098                         goto err;
3099         }
3100
3101         btrfs_set_path_blocking(path);
3102         ret = split_leaf(trans, root, &key, path, ins_len, 1);
3103         if (ret)
3104                 goto err;
3105
3106         path->keep_locks = 0;
3107         btrfs_unlock_up_safe(path, 1);
3108         return 0;
3109 err:
3110         path->keep_locks = 0;
3111         return ret;
3112 }
3113
3114 static noinline int split_item(struct btrfs_trans_handle *trans,
3115                                struct btrfs_root *root,
3116                                struct btrfs_path *path,
3117                                struct btrfs_key *new_key,
3118                                unsigned long split_offset)
3119 {
3120         struct extent_buffer *leaf;
3121         struct btrfs_item *item;
3122         struct btrfs_item *new_item;
3123         int slot;
3124         char *buf;
3125         u32 nritems;
3126         u32 item_size;
3127         u32 orig_offset;
3128         struct btrfs_disk_key disk_key;
3129
3130         leaf = path->nodes[0];
3131         BUG_ON(btrfs_leaf_free_space(root, leaf) < sizeof(struct btrfs_item));
3132
3133         btrfs_set_path_blocking(path);
3134
3135         item = btrfs_item_nr(leaf, path->slots[0]);
3136         orig_offset = btrfs_item_offset(leaf, item);
3137         item_size = btrfs_item_size(leaf, item);
3138
3139         buf = kmalloc(item_size, GFP_NOFS);
3140         if (!buf)
3141                 return -ENOMEM;
3142
3143         read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf,
3144                             path->slots[0]), item_size);
3145
3146         slot = path->slots[0] + 1;
3147         nritems = btrfs_header_nritems(leaf);
3148         if (slot != nritems) {
3149                 /* shift the items */
3150                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
3151                                 btrfs_item_nr_offset(slot),
3152                                 (nritems - slot) * sizeof(struct btrfs_item));
3153         }
3154
3155         btrfs_cpu_key_to_disk(&disk_key, new_key);
3156         btrfs_set_item_key(leaf, &disk_key, slot);
3157
3158         new_item = btrfs_item_nr(leaf, slot);
3159
3160         btrfs_set_item_offset(leaf, new_item, orig_offset);
3161         btrfs_set_item_size(leaf, new_item, item_size - split_offset);
3162
3163         btrfs_set_item_offset(leaf, item,
3164                               orig_offset + item_size - split_offset);
3165         btrfs_set_item_size(leaf, item, split_offset);
3166
3167         btrfs_set_header_nritems(leaf, nritems + 1);
3168
3169         /* write the data for the start of the original item */
3170         write_extent_buffer(leaf, buf,
3171                             btrfs_item_ptr_offset(leaf, path->slots[0]),
3172                             split_offset);
3173
3174         /* write the data for the new item */
3175         write_extent_buffer(leaf, buf + split_offset,
3176                             btrfs_item_ptr_offset(leaf, slot),
3177                             item_size - split_offset);
3178         btrfs_mark_buffer_dirty(leaf);
3179
3180         BUG_ON(btrfs_leaf_free_space(root, leaf) < 0);
3181         kfree(buf);
3182         return 0;
3183 }
3184
3185 /*
3186  * This function splits a single item into two items,
3187  * giving 'new_key' to the new item and splitting the
3188  * old one at split_offset (from the start of the item).
3189  *
3190  * The path may be released by this operation.  After
3191  * the split, the path is pointing to the old item.  The
3192  * new item is going to be in the same node as the old one.
3193  *
3194  * Note, the item being split must be smaller enough to live alone on
3195  * a tree block with room for one extra struct btrfs_item
3196  *
3197  * This allows us to split the item in place, keeping a lock on the
3198  * leaf the entire time.
3199  */
3200 int btrfs_split_item(struct btrfs_trans_handle *trans,
3201                      struct btrfs_root *root,
3202                      struct btrfs_path *path,
3203                      struct btrfs_key *new_key,
3204                      unsigned long split_offset)
3205 {
3206         int ret;
3207         ret = setup_leaf_for_split(trans, root, path,
3208                                    sizeof(struct btrfs_item));
3209         if (ret)
3210                 return ret;
3211
3212         ret = split_item(trans, root, path, new_key, split_offset);
3213         return ret;
3214 }
3215
3216 /*
3217  * This function duplicate a item, giving 'new_key' to the new item.
3218  * It guarantees both items live in the same tree leaf and the new item
3219  * is contiguous with the original item.
3220  *
3221  * This allows us to split file extent in place, keeping a lock on the
3222  * leaf the entire time.
3223  */
3224 int btrfs_duplicate_item(struct btrfs_trans_handle *trans,
3225                          struct btrfs_root *root,
3226                          struct btrfs_path *path,
3227                          struct btrfs_key *new_key)
3228 {
3229         struct extent_buffer *leaf;
3230         int ret;
3231         u32 item_size;
3232
3233         leaf = path->nodes[0];
3234         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3235         ret = setup_leaf_for_split(trans, root, path,
3236                                    item_size + sizeof(struct btrfs_item));
3237         if (ret)
3238                 return ret;
3239
3240         path->slots[0]++;
3241         ret = setup_items_for_insert(trans, root, path, new_key, &item_size,
3242                                      item_size, item_size +
3243                                      sizeof(struct btrfs_item), 1);
3244         BUG_ON(ret);
3245
3246         leaf = path->nodes[0];
3247         memcpy_extent_buffer(leaf,
3248                              btrfs_item_ptr_offset(leaf, path->slots[0]),
3249                              btrfs_item_ptr_offset(leaf, path->slots[0] - 1),
3250                              item_size);
3251         return 0;
3252 }
3253
3254 /*
3255  * make the item pointed to by the path smaller.  new_size indicates
3256  * how small to make it, and from_end tells us if we just chop bytes
3257  * off the end of the item or if we shift the item to chop bytes off
3258  * the front.
3259  */
3260 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
3261                         struct btrfs_root *root,
3262                         struct btrfs_path *path,
3263                         u32 new_size, int from_end)
3264 {
3265         int slot;
3266         struct extent_buffer *leaf;
3267         struct btrfs_item *item;
3268         u32 nritems;
3269         unsigned int data_end;
3270         unsigned int old_data_start;
3271         unsigned int old_size;
3272         unsigned int size_diff;
3273         int i;
3274
3275         leaf = path->nodes[0];
3276         slot = path->slots[0];
3277
3278         old_size = btrfs_item_size_nr(leaf, slot);
3279         if (old_size == new_size)
3280                 return 0;
3281
3282         nritems = btrfs_header_nritems(leaf);
3283         data_end = leaf_data_end(root, leaf);
3284
3285         old_data_start = btrfs_item_offset_nr(leaf, slot);
3286
3287         size_diff = old_size - new_size;
3288
3289         BUG_ON(slot < 0);
3290         BUG_ON(slot >= nritems);
3291
3292         /*
3293          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3294          */
3295         /* first correct the data pointers */
3296         for (i = slot; i < nritems; i++) {
3297                 u32 ioff;
3298                 item = btrfs_item_nr(leaf, i);
3299
3300                 ioff = btrfs_item_offset(leaf, item);
3301                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
3302         }
3303
3304         /* shift the data */
3305         if (from_end) {
3306                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3307                               data_end + size_diff, btrfs_leaf_data(leaf) +
3308                               data_end, old_data_start + new_size - data_end);
3309         } else {
3310                 struct btrfs_disk_key disk_key;
3311                 u64 offset;
3312
3313                 btrfs_item_key(leaf, &disk_key, slot);
3314
3315                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
3316                         unsigned long ptr;
3317                         struct btrfs_file_extent_item *fi;
3318
3319                         fi = btrfs_item_ptr(leaf, slot,
3320                                             struct btrfs_file_extent_item);
3321                         fi = (struct btrfs_file_extent_item *)(
3322                              (unsigned long)fi - size_diff);
3323
3324                         if (btrfs_file_extent_type(leaf, fi) ==
3325                             BTRFS_FILE_EXTENT_INLINE) {
3326                                 ptr = btrfs_item_ptr_offset(leaf, slot);
3327                                 memmove_extent_buffer(leaf, ptr,
3328                                       (unsigned long)fi,
3329                                       offsetof(struct btrfs_file_extent_item,
3330                                                  disk_bytenr));
3331                         }
3332                 }
3333
3334                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3335                               data_end + size_diff, btrfs_leaf_data(leaf) +
3336                               data_end, old_data_start - data_end);
3337
3338                 offset = btrfs_disk_key_offset(&disk_key);
3339                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
3340                 btrfs_set_item_key(leaf, &disk_key, slot);
3341                 if (slot == 0)
3342                         fixup_low_keys(trans, root, path, &disk_key, 1);
3343         }
3344
3345         item = btrfs_item_nr(leaf, slot);
3346         btrfs_set_item_size(leaf, item, new_size);
3347         btrfs_mark_buffer_dirty(leaf);
3348
3349         if (btrfs_leaf_free_space(root, leaf) < 0) {
3350                 btrfs_print_leaf(root, leaf);
3351                 BUG();
3352         }
3353         return 0;
3354 }
3355
3356 /*
3357  * make the item pointed to by the path bigger, data_size is the new size.
3358  */
3359 int btrfs_extend_item(struct btrfs_trans_handle *trans,
3360                       struct btrfs_root *root, struct btrfs_path *path,
3361                       u32 data_size)
3362 {
3363         int slot;
3364         struct extent_buffer *leaf;
3365         struct btrfs_item *item;
3366         u32 nritems;
3367         unsigned int data_end;
3368         unsigned int old_data;
3369         unsigned int old_size;
3370         int i;
3371
3372         leaf = path->nodes[0];
3373
3374         nritems = btrfs_header_nritems(leaf);
3375         data_end = leaf_data_end(root, leaf);
3376
3377         if (btrfs_leaf_free_space(root, leaf) < data_size) {
3378                 btrfs_print_leaf(root, leaf);
3379                 BUG();
3380         }
3381         slot = path->slots[0];
3382         old_data = btrfs_item_end_nr(leaf, slot);
3383
3384         BUG_ON(slot < 0);
3385         if (slot >= nritems) {
3386                 btrfs_print_leaf(root, leaf);
3387                 printk(KERN_CRIT "slot %d too large, nritems %d\n",
3388                        slot, nritems);
3389                 BUG_ON(1);
3390         }
3391
3392         /*
3393          * item0..itemN ... dataN.offset..dataN.size .. data0.size
3394          */
3395         /* first correct the data pointers */
3396         for (i = slot; i < nritems; i++) {
3397                 u32 ioff;
3398                 item = btrfs_item_nr(leaf, i);
3399
3400                 ioff = btrfs_item_offset(leaf, item);
3401                 btrfs_set_item_offset(leaf, item, ioff - data_size);
3402         }
3403
3404         /* shift the data */
3405         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3406                       data_end - data_size, btrfs_leaf_data(leaf) +
3407                       data_end, old_data - data_end);
3408
3409         data_end = old_data;
3410         old_size = btrfs_item_size_nr(leaf, slot);
3411         item = btrfs_item_nr(leaf, slot);
3412         btrfs_set_item_size(leaf, item, old_size + data_size);
3413         btrfs_mark_buffer_dirty(leaf);
3414
3415         if (btrfs_leaf_free_space(root, leaf) < 0) {
3416                 btrfs_print_leaf(root, leaf);
3417                 BUG();
3418         }
3419         return 0;
3420 }
3421
3422 /*
3423  * Given a key and some data, insert items into the tree.
3424  * This does all the path init required, making room in the tree if needed.
3425  * Returns the number of keys that were inserted.
3426  */
3427 int btrfs_insert_some_items(struct btrfs_trans_handle *trans,
3428                             struct btrfs_root *root,
3429                             struct btrfs_path *path,
3430                             struct btrfs_key *cpu_key, u32 *data_size,
3431                             int nr)
3432 {
3433         struct extent_buffer *leaf;
3434         struct btrfs_item *item;
3435         int ret = 0;
3436         int slot;
3437         int i;
3438         u32 nritems;
3439         u32 total_data = 0;
3440         u32 total_size = 0;
3441         unsigned int data_end;
3442         struct btrfs_disk_key disk_key;
3443         struct btrfs_key found_key;
3444
3445         for (i = 0; i < nr; i++) {
3446                 if (total_size + data_size[i] + sizeof(struct btrfs_item) >
3447                     BTRFS_LEAF_DATA_SIZE(root)) {
3448                         break;
3449                         nr = i;
3450                 }
3451                 total_data += data_size[i];
3452                 total_size += data_size[i] + sizeof(struct btrfs_item);
3453         }
3454         BUG_ON(nr == 0);
3455
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         leaf = path->nodes[0];
3463
3464         nritems = btrfs_header_nritems(leaf);
3465         data_end = leaf_data_end(root, leaf);
3466
3467         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3468                 for (i = nr; i >= 0; i--) {
3469                         total_data -= data_size[i];
3470                         total_size -= data_size[i] + sizeof(struct btrfs_item);
3471                         if (total_size < btrfs_leaf_free_space(root, leaf))
3472                                 break;
3473                 }
3474                 nr = i;
3475         }
3476
3477         slot = path->slots[0];
3478         BUG_ON(slot < 0);
3479
3480         if (slot != nritems) {
3481                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3482
3483                 item = btrfs_item_nr(leaf, slot);
3484                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3485
3486                 /* figure out how many keys we can insert in here */
3487                 total_data = data_size[0];
3488                 for (i = 1; i < nr; i++) {
3489                         if (btrfs_comp_cpu_keys(&found_key, cpu_key + i) <= 0)
3490                                 break;
3491                         total_data += data_size[i];
3492                 }
3493                 nr = i;
3494
3495                 if (old_data < data_end) {
3496                         btrfs_print_leaf(root, leaf);
3497                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3498                                slot, old_data, data_end);
3499                         BUG_ON(1);
3500                 }
3501                 /*
3502                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3503                  */
3504                 /* first correct the data pointers */
3505                 for (i = slot; i < nritems; i++) {
3506                         u32 ioff;
3507
3508                         item = btrfs_item_nr(leaf, i);
3509                         ioff = btrfs_item_offset(leaf, item);
3510                         btrfs_set_item_offset(leaf, item, ioff - total_data);
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         } else {
3523                 /*
3524                  * this sucks but it has to be done, if we are inserting at
3525                  * the end of the leaf only insert 1 of the items, since we
3526                  * have no way of knowing whats on the next leaf and we'd have
3527                  * to drop our current locks to figure it out
3528                  */
3529                 nr = 1;
3530         }
3531
3532         /* setup the item for the new data */
3533         for (i = 0; i < nr; i++) {
3534                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3535                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3536                 item = btrfs_item_nr(leaf, slot + i);
3537                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3538                 data_end -= data_size[i];
3539                 btrfs_set_item_size(leaf, item, data_size[i]);
3540         }
3541         btrfs_set_header_nritems(leaf, nritems + nr);
3542         btrfs_mark_buffer_dirty(leaf);
3543
3544         ret = 0;
3545         if (slot == 0) {
3546                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3547                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3548         }
3549
3550         if (btrfs_leaf_free_space(root, leaf) < 0) {
3551                 btrfs_print_leaf(root, leaf);
3552                 BUG();
3553         }
3554 out:
3555         if (!ret)
3556                 ret = nr;
3557         return ret;
3558 }
3559
3560 /*
3561  * this is a helper for btrfs_insert_empty_items, the main goal here is
3562  * to save stack depth by doing the bulk of the work in a function
3563  * that doesn't call btrfs_search_slot
3564  */
3565 int setup_items_for_insert(struct btrfs_trans_handle *trans,
3566                            struct btrfs_root *root, struct btrfs_path *path,
3567                            struct btrfs_key *cpu_key, u32 *data_size,
3568                            u32 total_data, u32 total_size, int nr)
3569 {
3570         struct btrfs_item *item;
3571         int i;
3572         u32 nritems;
3573         unsigned int data_end;
3574         struct btrfs_disk_key disk_key;
3575         int ret;
3576         struct extent_buffer *leaf;
3577         int slot;
3578
3579         leaf = path->nodes[0];
3580         slot = path->slots[0];
3581
3582         nritems = btrfs_header_nritems(leaf);
3583         data_end = leaf_data_end(root, leaf);
3584
3585         if (btrfs_leaf_free_space(root, leaf) < total_size) {
3586                 btrfs_print_leaf(root, leaf);
3587                 printk(KERN_CRIT "not enough freespace need %u have %d\n",
3588                        total_size, btrfs_leaf_free_space(root, leaf));
3589                 BUG();
3590         }
3591
3592         if (slot != nritems) {
3593                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
3594
3595                 if (old_data < data_end) {
3596                         btrfs_print_leaf(root, leaf);
3597                         printk(KERN_CRIT "slot %d old_data %d data_end %d\n",
3598                                slot, old_data, data_end);
3599                         BUG_ON(1);
3600                 }
3601                 /*
3602                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
3603                  */
3604                 /* first correct the data pointers */
3605                 for (i = slot; i < nritems; i++) {
3606                         u32 ioff;
3607
3608                         item = btrfs_item_nr(leaf, i);
3609                         ioff = btrfs_item_offset(leaf, item);
3610                         btrfs_set_item_offset(leaf, item, ioff - total_data);
3611                 }
3612                 /* shift the items */
3613                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + nr),
3614                               btrfs_item_nr_offset(slot),
3615                               (nritems - slot) * sizeof(struct btrfs_item));
3616
3617                 /* shift the data */
3618                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3619                               data_end - total_data, btrfs_leaf_data(leaf) +
3620                               data_end, old_data - data_end);
3621                 data_end = old_data;
3622         }
3623
3624         /* setup the item for the new data */
3625         for (i = 0; i < nr; i++) {
3626                 btrfs_cpu_key_to_disk(&disk_key, cpu_key + i);
3627                 btrfs_set_item_key(leaf, &disk_key, slot + i);
3628                 item = btrfs_item_nr(leaf, slot + i);
3629                 btrfs_set_item_offset(leaf, item, data_end - data_size[i]);
3630                 data_end -= data_size[i];
3631                 btrfs_set_item_size(leaf, item, data_size[i]);
3632         }
3633
3634         btrfs_set_header_nritems(leaf, nritems + nr);
3635
3636         ret = 0;
3637         if (slot == 0) {
3638                 btrfs_cpu_key_to_disk(&disk_key, cpu_key);
3639                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
3640         }
3641         btrfs_unlock_up_safe(path, 1);
3642         btrfs_mark_buffer_dirty(leaf);
3643
3644         if (btrfs_leaf_free_space(root, leaf) < 0) {
3645                 btrfs_print_leaf(root, leaf);
3646                 BUG();
3647         }
3648         return ret;
3649 }
3650
3651 /*
3652  * Given a key and some data, insert items into the tree.
3653  * This does all the path init required, making room in the tree if needed.
3654  */
3655 int btrfs_insert_empty_items(struct btrfs_trans_handle *trans,
3656                             struct btrfs_root *root,
3657                             struct btrfs_path *path,
3658                             struct btrfs_key *cpu_key, u32 *data_size,
3659                             int nr)
3660 {
3661         int ret = 0;
3662         int slot;
3663         int i;
3664         u32 total_size = 0;
3665         u32 total_data = 0;
3666
3667         for (i = 0; i < nr; i++)
3668                 total_data += data_size[i];
3669
3670         total_size = total_data + (nr * sizeof(struct btrfs_item));
3671         ret = btrfs_search_slot(trans, root, cpu_key, path, total_size, 1);
3672         if (ret == 0)
3673                 return -EEXIST;
3674         if (ret < 0)
3675                 goto out;
3676
3677         slot = path->slots[0];
3678         BUG_ON(slot < 0);
3679
3680         ret = setup_items_for_insert(trans, root, path, cpu_key, data_size,
3681                                total_data, total_size, nr);
3682
3683 out:
3684         return ret;
3685 }
3686
3687 /*
3688  * Given a key and some data, insert an item into the tree.
3689  * This does all the path init required, making room in the tree if needed.
3690  */
3691 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
3692                       *root, struct btrfs_key *cpu_key, void *data, u32
3693                       data_size)
3694 {
3695         int ret = 0;
3696         struct btrfs_path *path;
3697         struct extent_buffer *leaf;
3698         unsigned long ptr;
3699
3700         path = btrfs_alloc_path();
3701         if (!path)
3702                 return -ENOMEM;
3703         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
3704         if (!ret) {
3705                 leaf = path->nodes[0];
3706                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3707                 write_extent_buffer(leaf, data, ptr, data_size);
3708                 btrfs_mark_buffer_dirty(leaf);
3709         }
3710         btrfs_free_path(path);
3711         return ret;
3712 }
3713
3714 /*
3715  * delete the pointer from a given node.
3716  *
3717  * the tree should have been previously balanced so the deletion does not
3718  * empty a node.
3719  */
3720 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3721                    struct btrfs_path *path, int level, int slot)
3722 {
3723         struct extent_buffer *parent = path->nodes[level];
3724         u32 nritems;
3725         int ret = 0;
3726         int wret;
3727
3728         nritems = btrfs_header_nritems(parent);
3729         if (slot != nritems - 1) {
3730                 memmove_extent_buffer(parent,
3731                               btrfs_node_key_ptr_offset(slot),
3732                               btrfs_node_key_ptr_offset(slot + 1),
3733                               sizeof(struct btrfs_key_ptr) *
3734                               (nritems - slot - 1));
3735         }
3736         nritems--;
3737         btrfs_set_header_nritems(parent, nritems);
3738         if (nritems == 0 && parent == root->node) {
3739                 BUG_ON(btrfs_header_level(root->node) != 1);
3740                 /* just turn the root into a leaf and break */
3741                 btrfs_set_header_level(root->node, 0);
3742         } else if (slot == 0) {
3743                 struct btrfs_disk_key disk_key;
3744
3745                 btrfs_node_key(parent, &disk_key, 0);
3746                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
3747                 if (wret)
3748                         ret = wret;
3749         }
3750         btrfs_mark_buffer_dirty(parent);
3751         return ret;
3752 }
3753
3754 /*
3755  * a helper function to delete the leaf pointed to by path->slots[1] and
3756  * path->nodes[1].
3757  *
3758  * This deletes the pointer in path->nodes[1] and frees the leaf
3759  * block extent.  zero is returned if it all worked out, < 0 otherwise.
3760  *
3761  * The path must have already been setup for deleting the leaf, including
3762  * all the proper balancing.  path->nodes[1] must be locked.
3763  */
3764 static noinline int btrfs_del_leaf(struct btrfs_trans_handle *trans,
3765                                    struct btrfs_root *root,
3766                                    struct btrfs_path *path,
3767                                    struct extent_buffer *leaf)
3768 {
3769         int ret;
3770
3771         WARN_ON(btrfs_header_generation(leaf) != trans->transid);
3772         ret = del_ptr(trans, root, path, 1, path->slots[1]);
3773         if (ret)
3774                 return ret;
3775
3776         /*
3777          * btrfs_free_extent is expensive, we want to make sure we
3778          * aren't holding any locks when we call it
3779          */
3780         btrfs_unlock_up_safe(path, 0);
3781
3782         root_sub_used(root, leaf->len);
3783
3784         btrfs_free_tree_block(trans, root, leaf, 0, 1, 0);
3785         return 0;
3786 }
3787 /*
3788  * delete the item at the leaf level in path.  If that empties
3789  * the leaf, remove it from the tree
3790  */
3791 int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3792                     struct btrfs_path *path, int slot, int nr)
3793 {
3794         struct extent_buffer *leaf;
3795         struct btrfs_item *item;
3796         int last_off;
3797         int dsize = 0;
3798         int ret = 0;
3799         int wret;
3800         int i;
3801         u32 nritems;
3802
3803         leaf = path->nodes[0];
3804         last_off = btrfs_item_offset_nr(leaf, slot + nr - 1);
3805
3806         for (i = 0; i < nr; i++)
3807                 dsize += btrfs_item_size_nr(leaf, slot + i);
3808
3809         nritems = btrfs_header_nritems(leaf);
3810
3811         if (slot + nr != nritems) {
3812                 int data_end = leaf_data_end(root, leaf);
3813
3814                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
3815                               data_end + dsize,
3816                               btrfs_leaf_data(leaf) + data_end,
3817                               last_off - data_end);
3818
3819                 for (i = slot + nr; i < nritems; i++) {
3820                         u32 ioff;
3821
3822                         item = btrfs_item_nr(leaf, i);
3823                         ioff = btrfs_item_offset(leaf, item);
3824                         btrfs_set_item_offset(leaf, item, ioff + dsize);
3825                 }
3826
3827                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
3828                               btrfs_item_nr_offset(slot + nr),
3829                               sizeof(struct btrfs_item) *
3830                               (nritems - slot - nr));
3831         }
3832         btrfs_set_header_nritems(leaf, nritems - nr);
3833         nritems -= nr;
3834
3835         /* delete the leaf if we've emptied it */
3836         if (nritems == 0) {
3837                 if (leaf == root->node) {
3838                         btrfs_set_header_level(leaf, 0);
3839                 } else {
3840                         btrfs_set_path_blocking(path);
3841                         clean_tree_block(trans, root, leaf);
3842                         ret = btrfs_del_leaf(trans, root, path, leaf);
3843                         BUG_ON(ret);
3844                 }
3845         } else {
3846                 int used = leaf_space_used(leaf, 0, nritems);
3847                 if (slot == 0) {
3848                         struct btrfs_disk_key disk_key;
3849
3850                         btrfs_item_key(leaf, &disk_key, 0);
3851                         wret = fixup_low_keys(trans, root, path,
3852                                               &disk_key, 1);
3853                         if (wret)
3854                                 ret = wret;
3855                 }
3856
3857                 /* delete the leaf if it is mostly empty */
3858                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
3859                         /* push_leaf_left fixes the path.
3860                          * make sure the path still points to our leaf
3861                          * for possible call to del_ptr below
3862                          */
3863                         slot = path->slots[1];
3864                         extent_buffer_get(leaf);
3865
3866                         btrfs_set_path_blocking(path);
3867                         wret = push_leaf_left(trans, root, path, 1, 1,
3868                                               1, (u32)-1);
3869                         if (wret < 0 && wret != -ENOSPC)
3870                                 ret = wret;
3871
3872                         if (path->nodes[0] == leaf &&
3873                             btrfs_header_nritems(leaf)) {
3874                                 wret = push_leaf_right(trans, root, path, 1,
3875                                                        1, 1, 0);
3876                                 if (wret < 0 && wret != -ENOSPC)
3877                                         ret = wret;
3878                         }
3879
3880                         if (btrfs_header_nritems(leaf) == 0) {
3881                                 path->slots[1] = slot;
3882                                 ret = btrfs_del_leaf(trans, root, path, leaf);
3883                                 BUG_ON(ret);
3884                                 free_extent_buffer(leaf);
3885                         } else {
3886                                 /* if we're still in the path, make sure
3887                                  * we're dirty.  Otherwise, one of the
3888                                  * push_leaf functions must have already
3889                                  * dirtied this buffer
3890                                  */
3891                                 if (path->nodes[0] == leaf)
3892                                         btrfs_mark_buffer_dirty(leaf);
3893                                 free_extent_buffer(leaf);
3894                         }
3895                 } else {
3896                         btrfs_mark_buffer_dirty(leaf);
3897                 }
3898         }
3899         return ret;
3900 }
3901
3902 /*
3903  * search the tree again to find a leaf with lesser keys
3904  * returns 0 if it found something or 1 if there are no lesser leaves.
3905  * returns < 0 on io errors.
3906  *
3907  * This may release the path, and so you may lose any locks held at the
3908  * time you call it.
3909  */
3910 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
3911 {
3912         struct btrfs_key key;
3913         struct btrfs_disk_key found_key;
3914         int ret;
3915
3916         btrfs_item_key_to_cpu(path->nodes[0], &key, 0);
3917
3918         if (key.offset > 0)
3919                 key.offset--;
3920         else if (key.type > 0)
3921                 key.type--;
3922         else if (key.objectid > 0)
3923                 key.objectid--;
3924         else
3925                 return 1;
3926
3927         btrfs_release_path(path);
3928         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3929         if (ret < 0)
3930                 return ret;
3931         btrfs_item_key(path->nodes[0], &found_key, 0);
3932         ret = comp_keys(&found_key, &key);
3933         if (ret < 0)
3934                 return 0;
3935         return 1;
3936 }
3937
3938 /*
3939  * A helper function to walk down the tree starting at min_key, and looking
3940  * for nodes or leaves that are either in cache or have a minimum
3941  * transaction id.  This is used by the btree defrag code, and tree logging
3942  *
3943  * This does not cow, but it does stuff the starting key it finds back
3944  * into min_key, so you can call btrfs_search_slot with cow=1 on the
3945  * key and get a writable path.
3946  *
3947  * This does lock as it descends, and path->keep_locks should be set
3948  * to 1 by the caller.
3949  *
3950  * This honors path->lowest_level to prevent descent past a given level
3951  * of the tree.
3952  *
3953  * min_trans indicates the oldest transaction that you are interested
3954  * in walking through.  Any nodes or leaves older than min_trans are
3955  * skipped over (without reading them).
3956  *
3957  * returns zero if something useful was found, < 0 on error and 1 if there
3958  * was nothing in the tree that matched the search criteria.
3959  */
3960 int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
3961                          struct btrfs_key *max_key,
3962                          struct btrfs_path *path, int cache_only,
3963                          u64 min_trans)
3964 {
3965         struct extent_buffer *cur;
3966         struct btrfs_key found_key;
3967         int slot;
3968         int sret;
3969         u32 nritems;
3970         int level;
3971         int ret = 1;
3972
3973         WARN_ON(!path->keep_locks);
3974 again:
3975         cur = btrfs_read_lock_root_node(root);
3976         level = btrfs_header_level(cur);
3977         WARN_ON(path->nodes[level]);
3978         path->nodes[level] = cur;
3979         path->locks[level] = BTRFS_READ_LOCK;
3980
3981         if (btrfs_header_generation(cur) < min_trans) {
3982                 ret = 1;
3983                 goto out;
3984         }
3985         while (1) {
3986                 nritems = btrfs_header_nritems(cur);
3987                 level = btrfs_header_level(cur);
3988                 sret = bin_search(cur, min_key, level, &slot);
3989
3990                 /* at the lowest level, we're done, setup the path and exit */
3991                 if (level == path->lowest_level) {
3992                         if (slot >= nritems)
3993                                 goto find_next_key;
3994                         ret = 0;
3995                         path->slots[level] = slot;
3996                         btrfs_item_key_to_cpu(cur, &found_key, slot);
3997                         goto out;
3998                 }
3999                 if (sret && slot > 0)
4000                         slot--;
4001                 /*
4002                  * check this node pointer against the cache_only and
4003                  * min_trans parameters.  If it isn't in cache or is too
4004                  * old, skip to the next one.
4005                  */
4006                 while (slot < nritems) {
4007                         u64 blockptr;
4008                         u64 gen;
4009                         struct extent_buffer *tmp;
4010                         struct btrfs_disk_key disk_key;
4011
4012                         blockptr = btrfs_node_blockptr(cur, slot);
4013                         gen = btrfs_node_ptr_generation(cur, slot);
4014                         if (gen < min_trans) {
4015                                 slot++;
4016                                 continue;
4017                         }
4018                         if (!cache_only)
4019                                 break;
4020
4021                         if (max_key) {
4022                                 btrfs_node_key(cur, &disk_key, slot);
4023                                 if (comp_keys(&disk_key, max_key) >= 0) {
4024                                         ret = 1;
4025                                         goto out;
4026                                 }
4027                         }
4028
4029                         tmp = btrfs_find_tree_block(root, blockptr,
4030                                             btrfs_level_size(root, level - 1));
4031
4032                         if (tmp && btrfs_buffer_uptodate(tmp, gen)) {
4033                                 free_extent_buffer(tmp);
4034                                 break;
4035                         }
4036                         if (tmp)
4037                                 free_extent_buffer(tmp);
4038                         slot++;
4039                 }
4040 find_next_key:
4041                 /*
4042                  * we didn't find a candidate key in this node, walk forward
4043                  * and find another one
4044                  */
4045                 if (slot >= nritems) {
4046                         path->slots[level] = slot;
4047                         btrfs_set_path_blocking(path);
4048                         sret = btrfs_find_next_key(root, path, min_key, level,
4049                                                   cache_only, min_trans);
4050                         if (sret == 0) {
4051                                 btrfs_release_path(path);
4052                                 goto again;
4053                         } else {
4054                                 goto out;
4055                         }
4056                 }
4057                 /* save our key for returning back */
4058                 btrfs_node_key_to_cpu(cur, &found_key, slot);
4059                 path->slots[level] = slot;
4060                 if (level == path->lowest_level) {
4061                         ret = 0;
4062                         unlock_up(path, level, 1);
4063                         goto out;
4064                 }
4065                 btrfs_set_path_blocking(path);
4066                 cur = read_node_slot(root, cur, slot);
4067                 BUG_ON(!cur);
4068
4069                 btrfs_tree_read_lock(cur);
4070
4071                 path->locks[level - 1] = BTRFS_READ_LOCK;
4072                 path->nodes[level - 1] = cur;
4073                 unlock_up(path, level, 1);
4074                 btrfs_clear_path_blocking(path, NULL, 0);
4075         }
4076 out:
4077         if (ret == 0)
4078                 memcpy(min_key, &found_key, sizeof(found_key));
4079         btrfs_set_path_blocking(path);
4080         return ret;
4081 }
4082
4083 /*
4084  * this is similar to btrfs_next_leaf, but does not try to preserve
4085  * and fixup the path.  It looks for and returns the next key in the
4086  * tree based on the current path and the cache_only and min_trans
4087  * parameters.
4088  *
4089  * 0 is returned if another key is found, < 0 if there are any errors
4090  * and 1 is returned if there are no higher keys in the tree
4091  *
4092  * path->keep_locks should be set to 1 on the search made before
4093  * calling this function.
4094  */
4095 int btrfs_find_next_key(struct btrfs_root *root, struct btrfs_path *path,
4096                         struct btrfs_key *key, int level,
4097                         int cache_only, u64 min_trans)
4098 {
4099         int slot;
4100         struct extent_buffer *c;
4101
4102         WARN_ON(!path->keep_locks);
4103         while (level < BTRFS_MAX_LEVEL) {
4104                 if (!path->nodes[level])
4105                         return 1;
4106
4107                 slot = path->slots[level] + 1;
4108                 c = path->nodes[level];
4109 next:
4110                 if (slot >= btrfs_header_nritems(c)) {
4111                         int ret;
4112                         int orig_lowest;
4113                         struct btrfs_key cur_key;
4114                         if (level + 1 >= BTRFS_MAX_LEVEL ||
4115                             !path->nodes[level + 1])
4116                                 return 1;
4117
4118                         if (path->locks[level + 1]) {
4119                                 level++;
4120                                 continue;
4121                         }
4122
4123                         slot = btrfs_header_nritems(c) - 1;
4124                         if (level == 0)
4125                                 btrfs_item_key_to_cpu(c, &cur_key, slot);
4126                         else
4127                                 btrfs_node_key_to_cpu(c, &cur_key, slot);
4128
4129                         orig_lowest = path->lowest_level;
4130                         btrfs_release_path(path);
4131                         path->lowest_level = level;
4132                         ret = btrfs_search_slot(NULL, root, &cur_key, path,
4133                                                 0, 0);
4134                         path->lowest_level = orig_lowest;
4135                         if (ret < 0)
4136                                 return ret;
4137
4138                         c = path->nodes[level];
4139                         slot = path->slots[level];
4140                         if (ret == 0)
4141                                 slot++;
4142                         goto next;
4143                 }
4144
4145                 if (level == 0)
4146                         btrfs_item_key_to_cpu(c, key, slot);
4147                 else {
4148                         u64 blockptr = btrfs_node_blockptr(c, slot);
4149                         u64 gen = btrfs_node_ptr_generation(c, slot);
4150
4151                         if (cache_only) {
4152                                 struct extent_buffer *cur;
4153                                 cur = btrfs_find_tree_block(root, blockptr,
4154                                             btrfs_level_size(root, level - 1));
4155                                 if (!cur || !btrfs_buffer_uptodate(cur, gen)) {
4156                                         slot++;
4157                                         if (cur)
4158                                                 free_extent_buffer(cur);
4159                                         goto next;
4160                                 }
4161                                 free_extent_buffer(cur);
4162                         }
4163                         if (gen < min_trans) {
4164                                 slot++;
4165                                 goto next;
4166                         }
4167                         btrfs_node_key_to_cpu(c, key, slot);
4168                 }
4169                 return 0;
4170         }
4171         return 1;
4172 }
4173
4174 /*
4175  * search the tree again to find a leaf with greater keys
4176  * returns 0 if it found something or 1 if there are no greater leaves.
4177  * returns < 0 on io errors.
4178  */
4179 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
4180 {
4181         int slot;
4182         int level;
4183         struct extent_buffer *c;
4184         struct extent_buffer *next;
4185         struct btrfs_key key;
4186         u32 nritems;
4187         int ret;
4188         int old_spinning = path->leave_spinning;
4189         int next_rw_lock = 0;
4190
4191         nritems = btrfs_header_nritems(path->nodes[0]);
4192         if (nritems == 0)
4193                 return 1;
4194
4195         btrfs_item_key_to_cpu(path->nodes[0], &key, nritems - 1);
4196 again:
4197         level = 1;
4198         next = NULL;
4199         next_rw_lock = 0;
4200         btrfs_release_path(path);
4201
4202         path->keep_locks = 1;
4203         path->leave_spinning = 1;
4204
4205         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
4206         path->keep_locks = 0;
4207
4208         if (ret < 0)
4209                 return ret;
4210
4211         nritems = btrfs_header_nritems(path->nodes[0]);
4212         /*
4213          * by releasing the path above we dropped all our locks.  A balance
4214          * could have added more items next to the key that used to be
4215          * at the very end of the block.  So, check again here and
4216          * advance the path if there are now more items available.
4217          */
4218         if (nritems > 0 && path->slots[0] < nritems - 1) {
4219                 if (ret == 0)
4220                         path->slots[0]++;
4221                 ret = 0;
4222                 goto done;
4223         }
4224
4225         while (level < BTRFS_MAX_LEVEL) {
4226                 if (!path->nodes[level]) {
4227                         ret = 1;
4228                         goto done;
4229                 }
4230
4231                 slot = path->slots[level] + 1;
4232                 c = path->nodes[level];
4233                 if (slot >= btrfs_header_nritems(c)) {
4234                         level++;
4235                         if (level == BTRFS_MAX_LEVEL) {
4236                                 ret = 1;
4237                                 goto done;
4238                         }
4239                         continue;
4240                 }
4241
4242                 if (next) {
4243                         btrfs_tree_unlock_rw(next, next_rw_lock);
4244                         free_extent_buffer(next);
4245                 }
4246
4247                 next = c;
4248                 next_rw_lock = path->locks[level];
4249                 ret = read_block_for_search(NULL, root, path, &next, level,
4250                                             slot, &key);
4251                 if (ret == -EAGAIN)
4252                         goto again;
4253
4254                 if (ret < 0) {
4255                         btrfs_release_path(path);
4256                         goto done;
4257                 }
4258
4259                 if (!path->skip_locking) {
4260                         ret = btrfs_try_tree_read_lock(next);
4261                         if (!ret) {
4262                                 btrfs_set_path_blocking(path);
4263                                 btrfs_tree_read_lock(next);
4264                                 btrfs_clear_path_blocking(path, next,
4265                                                           BTRFS_READ_LOCK);
4266                         }
4267                         next_rw_lock = BTRFS_READ_LOCK;
4268                 }
4269                 break;
4270         }
4271         path->slots[level] = slot;
4272         while (1) {
4273                 level--;
4274                 c = path->nodes[level];
4275                 if (path->locks[level])
4276                         btrfs_tree_unlock_rw(c, path->locks[level]);
4277
4278                 free_extent_buffer(c);
4279                 path->nodes[level] = next;
4280                 path->slots[level] = 0;
4281                 if (!path->skip_locking)
4282                         path->locks[level] = next_rw_lock;
4283                 if (!level)
4284                         break;
4285
4286                 ret = read_block_for_search(NULL, root, path, &next, level,
4287                                             0, &key);
4288                 if (ret == -EAGAIN)
4289                         goto again;
4290
4291                 if (ret < 0) {
4292                         btrfs_release_path(path);
4293                         goto done;
4294                 }
4295
4296                 if (!path->skip_locking) {
4297                         ret = btrfs_try_tree_read_lock(next);
4298                         if (!ret) {
4299                                 btrfs_set_path_blocking(path);
4300                                 btrfs_tree_read_lock(next);
4301                                 btrfs_clear_path_blocking(path, next,
4302                                                           BTRFS_READ_LOCK);
4303                         }
4304                         next_rw_lock = BTRFS_READ_LOCK;
4305                 }
4306         }
4307         ret = 0;
4308 done:
4309         unlock_up(path, 0, 1);
4310         path->leave_spinning = old_spinning;
4311         if (!old_spinning)
4312                 btrfs_set_path_blocking(path);
4313
4314         return ret;
4315 }
4316
4317 /*
4318  * this uses btrfs_prev_leaf to walk backwards in the tree, and keeps
4319  * searching until it gets past min_objectid or finds an item of 'type'
4320  *
4321  * returns 0 if something is found, 1 if nothing was found and < 0 on error
4322  */
4323 int btrfs_previous_item(struct btrfs_root *root,
4324                         struct btrfs_path *path, u64 min_objectid,
4325                         int type)
4326 {
4327         struct btrfs_key found_key;
4328         struct extent_buffer *leaf;
4329         u32 nritems;
4330         int ret;
4331
4332         while (1) {
4333                 if (path->slots[0] == 0) {
4334                         btrfs_set_path_blocking(path);
4335                         ret = btrfs_prev_leaf(root, path);
4336                         if (ret != 0)
4337                                 return ret;
4338                 } else {
4339                         path->slots[0]--;
4340                 }
4341                 leaf = path->nodes[0];
4342                 nritems = btrfs_header_nritems(leaf);
4343                 if (nritems == 0)
4344                         return 1;
4345                 if (path->slots[0] == nritems)
4346                         path->slots[0]--;
4347
4348                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4349                 if (found_key.objectid < min_objectid)
4350                         break;
4351                 if (found_key.type == type)
4352                         return 0;
4353                 if (found_key.objectid == min_objectid &&
4354                     found_key.type < type)
4355                         break;
4356         }
4357         return 1;
4358 }