]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - fs/btrfs/ctree.c
Btrfs: fix memory leak when cloning root's node
[~shefty/rdma-dev.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 <linux/rbtree.h>
22 #include "ctree.h"
23 #include "disk-io.h"
24 #include "transaction.h"
25 #include "print-tree.h"
26 #include "locking.h"
27
28 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
29                       *root, struct btrfs_path *path, int level);
30 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
31                       *root, struct btrfs_key *ins_key,
32                       struct btrfs_path *path, int data_size, int extend);
33 static int push_node_left(struct btrfs_trans_handle *trans,
34                           struct btrfs_root *root, struct extent_buffer *dst,
35                           struct extent_buffer *src, int empty);
36 static int balance_node_right(struct btrfs_trans_handle *trans,
37                               struct btrfs_root *root,
38                               struct extent_buffer *dst_buf,
39                               struct extent_buffer *src_buf);
40 static void del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
41                     struct btrfs_path *path, int level, int slot,
42                     int tree_mod_log);
43 static void tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
44                                  struct extent_buffer *eb);
45 struct extent_buffer *read_old_tree_block(struct btrfs_root *root, u64 bytenr,
46                                           u32 blocksize, u64 parent_transid,
47                                           u64 time_seq);
48 struct extent_buffer *btrfs_find_old_tree_block(struct btrfs_root *root,
49                                                 u64 bytenr, u32 blocksize,
50                                                 u64 time_seq);
51
52 struct btrfs_path *btrfs_alloc_path(void)
53 {
54         struct btrfs_path *path;
55         path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
56         return path;
57 }
58
59 /*
60  * set all locked nodes in the path to blocking locks.  This should
61  * be done before scheduling
62  */
63 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
64 {
65         int i;
66         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
67                 if (!p->nodes[i] || !p->locks[i])
68                         continue;
69                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
70                 if (p->locks[i] == BTRFS_READ_LOCK)
71                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
72                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
73                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
74         }
75 }
76
77 /*
78  * reset all the locked nodes in the patch to spinning locks.
79  *
80  * held is used to keep lockdep happy, when lockdep is enabled
81  * we set held to a blocking lock before we go around and
82  * retake all the spinlocks in the path.  You can safely use NULL
83  * for held
84  */
85 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
86                                         struct extent_buffer *held, int held_rw)
87 {
88         int i;
89
90 #ifdef CONFIG_DEBUG_LOCK_ALLOC
91         /* lockdep really cares that we take all of these spinlocks
92          * in the right order.  If any of the locks in the path are not
93          * currently blocking, it is going to complain.  So, make really
94          * really sure by forcing the path to blocking before we clear
95          * the path blocking.
96          */
97         if (held) {
98                 btrfs_set_lock_blocking_rw(held, held_rw);
99                 if (held_rw == BTRFS_WRITE_LOCK)
100                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
101                 else if (held_rw == BTRFS_READ_LOCK)
102                         held_rw = BTRFS_READ_LOCK_BLOCKING;
103         }
104         btrfs_set_path_blocking(p);
105 #endif
106
107         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
108                 if (p->nodes[i] && p->locks[i]) {
109                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
110                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
111                                 p->locks[i] = BTRFS_WRITE_LOCK;
112                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
113                                 p->locks[i] = BTRFS_READ_LOCK;
114                 }
115         }
116
117 #ifdef CONFIG_DEBUG_LOCK_ALLOC
118         if (held)
119                 btrfs_clear_lock_blocking_rw(held, held_rw);
120 #endif
121 }
122
123 /* this also releases the path */
124 void btrfs_free_path(struct btrfs_path *p)
125 {
126         if (!p)
127                 return;
128         btrfs_release_path(p);
129         kmem_cache_free(btrfs_path_cachep, p);
130 }
131
132 /*
133  * path release drops references on the extent buffers in the path
134  * and it drops any locks held by this path
135  *
136  * It is safe to call this on paths that no locks or extent buffers held.
137  */
138 noinline void btrfs_release_path(struct btrfs_path *p)
139 {
140         int i;
141
142         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
143                 p->slots[i] = 0;
144                 if (!p->nodes[i])
145                         continue;
146                 if (p->locks[i]) {
147                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
148                         p->locks[i] = 0;
149                 }
150                 free_extent_buffer(p->nodes[i]);
151                 p->nodes[i] = NULL;
152         }
153 }
154
155 /*
156  * safely gets a reference on the root node of a tree.  A lock
157  * is not taken, so a concurrent writer may put a different node
158  * at the root of the tree.  See btrfs_lock_root_node for the
159  * looping required.
160  *
161  * The extent buffer returned by this has a reference taken, so
162  * it won't disappear.  It may stop being the root of the tree
163  * at any time because there are no locks held.
164  */
165 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
166 {
167         struct extent_buffer *eb;
168
169         while (1) {
170                 rcu_read_lock();
171                 eb = rcu_dereference(root->node);
172
173                 /*
174                  * RCU really hurts here, we could free up the root node because
175                  * it was cow'ed but we may not get the new root node yet so do
176                  * the inc_not_zero dance and if it doesn't work then
177                  * synchronize_rcu and try again.
178                  */
179                 if (atomic_inc_not_zero(&eb->refs)) {
180                         rcu_read_unlock();
181                         break;
182                 }
183                 rcu_read_unlock();
184                 synchronize_rcu();
185         }
186         return eb;
187 }
188
189 /* loop around taking references on and locking the root node of the
190  * tree until you end up with a lock on the root.  A locked buffer
191  * is returned, with a reference held.
192  */
193 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
194 {
195         struct extent_buffer *eb;
196
197         while (1) {
198                 eb = btrfs_root_node(root);
199                 btrfs_tree_lock(eb);
200                 if (eb == root->node)
201                         break;
202                 btrfs_tree_unlock(eb);
203                 free_extent_buffer(eb);
204         }
205         return eb;
206 }
207
208 /* loop around taking references on and locking the root node of the
209  * tree until you end up with a lock on the root.  A locked buffer
210  * is returned, with a reference held.
211  */
212 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
213 {
214         struct extent_buffer *eb;
215
216         while (1) {
217                 eb = btrfs_root_node(root);
218                 btrfs_tree_read_lock(eb);
219                 if (eb == root->node)
220                         break;
221                 btrfs_tree_read_unlock(eb);
222                 free_extent_buffer(eb);
223         }
224         return eb;
225 }
226
227 /* cowonly root (everything not a reference counted cow subvolume), just get
228  * put onto a simple dirty list.  transaction.c walks this to make sure they
229  * get properly updated on disk.
230  */
231 static void add_root_to_dirty_list(struct btrfs_root *root)
232 {
233         spin_lock(&root->fs_info->trans_lock);
234         if (root->track_dirty && list_empty(&root->dirty_list)) {
235                 list_add(&root->dirty_list,
236                          &root->fs_info->dirty_cowonly_roots);
237         }
238         spin_unlock(&root->fs_info->trans_lock);
239 }
240
241 /*
242  * used by snapshot creation to make a copy of a root for a tree with
243  * a given objectid.  The buffer with the new root node is returned in
244  * cow_ret, and this func returns zero on success or a negative error code.
245  */
246 int btrfs_copy_root(struct btrfs_trans_handle *trans,
247                       struct btrfs_root *root,
248                       struct extent_buffer *buf,
249                       struct extent_buffer **cow_ret, u64 new_root_objectid)
250 {
251         struct extent_buffer *cow;
252         int ret = 0;
253         int level;
254         struct btrfs_disk_key disk_key;
255
256         WARN_ON(root->ref_cows && trans->transid !=
257                 root->fs_info->running_transaction->transid);
258         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
259
260         level = btrfs_header_level(buf);
261         if (level == 0)
262                 btrfs_item_key(buf, &disk_key, 0);
263         else
264                 btrfs_node_key(buf, &disk_key, 0);
265
266         cow = btrfs_alloc_free_block(trans, root, buf->len, 0,
267                                      new_root_objectid, &disk_key, level,
268                                      buf->start, 0);
269         if (IS_ERR(cow))
270                 return PTR_ERR(cow);
271
272         copy_extent_buffer(cow, buf, 0, 0, cow->len);
273         btrfs_set_header_bytenr(cow, cow->start);
274         btrfs_set_header_generation(cow, trans->transid);
275         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
276         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
277                                      BTRFS_HEADER_FLAG_RELOC);
278         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
279                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
280         else
281                 btrfs_set_header_owner(cow, new_root_objectid);
282
283         write_extent_buffer(cow, root->fs_info->fsid,
284                             (unsigned long)btrfs_header_fsid(cow),
285                             BTRFS_FSID_SIZE);
286
287         WARN_ON(btrfs_header_generation(buf) > trans->transid);
288         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
289                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
290         else
291                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
292
293         if (ret)
294                 return ret;
295
296         btrfs_mark_buffer_dirty(cow);
297         *cow_ret = cow;
298         return 0;
299 }
300
301 enum mod_log_op {
302         MOD_LOG_KEY_REPLACE,
303         MOD_LOG_KEY_ADD,
304         MOD_LOG_KEY_REMOVE,
305         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
306         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
307         MOD_LOG_MOVE_KEYS,
308         MOD_LOG_ROOT_REPLACE,
309 };
310
311 struct tree_mod_move {
312         int dst_slot;
313         int nr_items;
314 };
315
316 struct tree_mod_root {
317         u64 logical;
318         u8 level;
319 };
320
321 struct tree_mod_elem {
322         struct rb_node node;
323         u64 index;              /* shifted logical */
324         u64 seq;
325         enum mod_log_op op;
326
327         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
328         int slot;
329
330         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
331         u64 generation;
332
333         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
334         struct btrfs_disk_key key;
335         u64 blockptr;
336
337         /* this is used for op == MOD_LOG_MOVE_KEYS */
338         struct tree_mod_move move;
339
340         /* this is used for op == MOD_LOG_ROOT_REPLACE */
341         struct tree_mod_root old_root;
342 };
343
344 static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
345 {
346         read_lock(&fs_info->tree_mod_log_lock);
347 }
348
349 static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
350 {
351         read_unlock(&fs_info->tree_mod_log_lock);
352 }
353
354 static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
355 {
356         write_lock(&fs_info->tree_mod_log_lock);
357 }
358
359 static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
360 {
361         write_unlock(&fs_info->tree_mod_log_lock);
362 }
363
364 /*
365  * This adds a new blocker to the tree mod log's blocker list if the @elem
366  * passed does not already have a sequence number set. So when a caller expects
367  * to record tree modifications, it should ensure to set elem->seq to zero
368  * before calling btrfs_get_tree_mod_seq.
369  * Returns a fresh, unused tree log modification sequence number, even if no new
370  * blocker was added.
371  */
372 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
373                            struct seq_list *elem)
374 {
375         u64 seq;
376
377         tree_mod_log_write_lock(fs_info);
378         spin_lock(&fs_info->tree_mod_seq_lock);
379         if (!elem->seq) {
380                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
381                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
382         }
383         seq = btrfs_inc_tree_mod_seq(fs_info);
384         spin_unlock(&fs_info->tree_mod_seq_lock);
385         tree_mod_log_write_unlock(fs_info);
386
387         return seq;
388 }
389
390 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
391                             struct seq_list *elem)
392 {
393         struct rb_root *tm_root;
394         struct rb_node *node;
395         struct rb_node *next;
396         struct seq_list *cur_elem;
397         struct tree_mod_elem *tm;
398         u64 min_seq = (u64)-1;
399         u64 seq_putting = elem->seq;
400
401         if (!seq_putting)
402                 return;
403
404         spin_lock(&fs_info->tree_mod_seq_lock);
405         list_del(&elem->list);
406         elem->seq = 0;
407
408         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
409                 if (cur_elem->seq < min_seq) {
410                         if (seq_putting > cur_elem->seq) {
411                                 /*
412                                  * blocker with lower sequence number exists, we
413                                  * cannot remove anything from the log
414                                  */
415                                 spin_unlock(&fs_info->tree_mod_seq_lock);
416                                 return;
417                         }
418                         min_seq = cur_elem->seq;
419                 }
420         }
421         spin_unlock(&fs_info->tree_mod_seq_lock);
422
423         /*
424          * anything that's lower than the lowest existing (read: blocked)
425          * sequence number can be removed from the tree.
426          */
427         tree_mod_log_write_lock(fs_info);
428         tm_root = &fs_info->tree_mod_log;
429         for (node = rb_first(tm_root); node; node = next) {
430                 next = rb_next(node);
431                 tm = container_of(node, struct tree_mod_elem, node);
432                 if (tm->seq > min_seq)
433                         continue;
434                 rb_erase(node, tm_root);
435                 kfree(tm);
436         }
437         tree_mod_log_write_unlock(fs_info);
438 }
439
440 /*
441  * key order of the log:
442  *       index -> sequence
443  *
444  * the index is the shifted logical of the *new* root node for root replace
445  * operations, or the shifted logical of the affected block for all other
446  * operations.
447  */
448 static noinline int
449 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
450 {
451         struct rb_root *tm_root;
452         struct rb_node **new;
453         struct rb_node *parent = NULL;
454         struct tree_mod_elem *cur;
455
456         BUG_ON(!tm || !tm->seq);
457
458         tm_root = &fs_info->tree_mod_log;
459         new = &tm_root->rb_node;
460         while (*new) {
461                 cur = container_of(*new, struct tree_mod_elem, node);
462                 parent = *new;
463                 if (cur->index < tm->index)
464                         new = &((*new)->rb_left);
465                 else if (cur->index > tm->index)
466                         new = &((*new)->rb_right);
467                 else if (cur->seq < tm->seq)
468                         new = &((*new)->rb_left);
469                 else if (cur->seq > tm->seq)
470                         new = &((*new)->rb_right);
471                 else {
472                         kfree(tm);
473                         return -EEXIST;
474                 }
475         }
476
477         rb_link_node(&tm->node, parent, new);
478         rb_insert_color(&tm->node, tm_root);
479         return 0;
480 }
481
482 /*
483  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
484  * returns zero with the tree_mod_log_lock acquired. The caller must hold
485  * this until all tree mod log insertions are recorded in the rb tree and then
486  * call tree_mod_log_write_unlock() to release.
487  */
488 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
489                                     struct extent_buffer *eb) {
490         smp_mb();
491         if (list_empty(&(fs_info)->tree_mod_seq_list))
492                 return 1;
493         if (eb && btrfs_header_level(eb) == 0)
494                 return 1;
495
496         tree_mod_log_write_lock(fs_info);
497         if (list_empty(&fs_info->tree_mod_seq_list)) {
498                 /*
499                  * someone emptied the list while we were waiting for the lock.
500                  * we must not add to the list when no blocker exists.
501                  */
502                 tree_mod_log_write_unlock(fs_info);
503                 return 1;
504         }
505
506         return 0;
507 }
508
509 /*
510  * This allocates memory and gets a tree modification sequence number.
511  *
512  * Returns <0 on error.
513  * Returns >0 (the added sequence number) on success.
514  */
515 static inline int tree_mod_alloc(struct btrfs_fs_info *fs_info, gfp_t flags,
516                                  struct tree_mod_elem **tm_ret)
517 {
518         struct tree_mod_elem *tm;
519
520         /*
521          * once we switch from spin locks to something different, we should
522          * honor the flags parameter here.
523          */
524         tm = *tm_ret = kzalloc(sizeof(*tm), GFP_ATOMIC);
525         if (!tm)
526                 return -ENOMEM;
527
528         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
529         return tm->seq;
530 }
531
532 static inline int
533 __tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
534                           struct extent_buffer *eb, int slot,
535                           enum mod_log_op op, gfp_t flags)
536 {
537         int ret;
538         struct tree_mod_elem *tm;
539
540         ret = tree_mod_alloc(fs_info, flags, &tm);
541         if (ret < 0)
542                 return ret;
543
544         tm->index = eb->start >> PAGE_CACHE_SHIFT;
545         if (op != MOD_LOG_KEY_ADD) {
546                 btrfs_node_key(eb, &tm->key, slot);
547                 tm->blockptr = btrfs_node_blockptr(eb, slot);
548         }
549         tm->op = op;
550         tm->slot = slot;
551         tm->generation = btrfs_node_ptr_generation(eb, slot);
552
553         return __tree_mod_log_insert(fs_info, tm);
554 }
555
556 static noinline int
557 tree_mod_log_insert_key_mask(struct btrfs_fs_info *fs_info,
558                              struct extent_buffer *eb, int slot,
559                              enum mod_log_op op, gfp_t flags)
560 {
561         int ret;
562
563         if (tree_mod_dont_log(fs_info, eb))
564                 return 0;
565
566         ret = __tree_mod_log_insert_key(fs_info, eb, slot, op, flags);
567
568         tree_mod_log_write_unlock(fs_info);
569         return ret;
570 }
571
572 static noinline int
573 tree_mod_log_insert_key(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
574                         int slot, enum mod_log_op op)
575 {
576         return tree_mod_log_insert_key_mask(fs_info, eb, slot, op, GFP_NOFS);
577 }
578
579 static noinline int
580 tree_mod_log_insert_key_locked(struct btrfs_fs_info *fs_info,
581                              struct extent_buffer *eb, int slot,
582                              enum mod_log_op op)
583 {
584         return __tree_mod_log_insert_key(fs_info, eb, slot, op, GFP_NOFS);
585 }
586
587 static noinline int
588 tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
589                          struct extent_buffer *eb, int dst_slot, int src_slot,
590                          int nr_items, gfp_t flags)
591 {
592         struct tree_mod_elem *tm;
593         int ret;
594         int i;
595
596         if (tree_mod_dont_log(fs_info, eb))
597                 return 0;
598
599         /*
600          * When we override something during the move, we log these removals.
601          * This can only happen when we move towards the beginning of the
602          * buffer, i.e. dst_slot < src_slot.
603          */
604         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
605                 ret = tree_mod_log_insert_key_locked(fs_info, eb, i + dst_slot,
606                                               MOD_LOG_KEY_REMOVE_WHILE_MOVING);
607                 BUG_ON(ret < 0);
608         }
609
610         ret = tree_mod_alloc(fs_info, flags, &tm);
611         if (ret < 0)
612                 goto out;
613
614         tm->index = eb->start >> PAGE_CACHE_SHIFT;
615         tm->slot = src_slot;
616         tm->move.dst_slot = dst_slot;
617         tm->move.nr_items = nr_items;
618         tm->op = MOD_LOG_MOVE_KEYS;
619
620         ret = __tree_mod_log_insert(fs_info, tm);
621 out:
622         tree_mod_log_write_unlock(fs_info);
623         return ret;
624 }
625
626 static inline void
627 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
628 {
629         int i;
630         u32 nritems;
631         int ret;
632
633         if (btrfs_header_level(eb) == 0)
634                 return;
635
636         nritems = btrfs_header_nritems(eb);
637         for (i = nritems - 1; i >= 0; i--) {
638                 ret = tree_mod_log_insert_key_locked(fs_info, eb, i,
639                                               MOD_LOG_KEY_REMOVE_WHILE_FREEING);
640                 BUG_ON(ret < 0);
641         }
642 }
643
644 static noinline int
645 tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
646                          struct extent_buffer *old_root,
647                          struct extent_buffer *new_root, gfp_t flags)
648 {
649         struct tree_mod_elem *tm;
650         int ret;
651
652         if (tree_mod_dont_log(fs_info, NULL))
653                 return 0;
654
655         ret = tree_mod_alloc(fs_info, flags, &tm);
656         if (ret < 0)
657                 goto out;
658
659         tm->index = new_root->start >> PAGE_CACHE_SHIFT;
660         tm->old_root.logical = old_root->start;
661         tm->old_root.level = btrfs_header_level(old_root);
662         tm->generation = btrfs_header_generation(old_root);
663         tm->op = MOD_LOG_ROOT_REPLACE;
664
665         ret = __tree_mod_log_insert(fs_info, tm);
666 out:
667         tree_mod_log_write_unlock(fs_info);
668         return ret;
669 }
670
671 static struct tree_mod_elem *
672 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
673                       int smallest)
674 {
675         struct rb_root *tm_root;
676         struct rb_node *node;
677         struct tree_mod_elem *cur = NULL;
678         struct tree_mod_elem *found = NULL;
679         u64 index = start >> PAGE_CACHE_SHIFT;
680
681         tree_mod_log_read_lock(fs_info);
682         tm_root = &fs_info->tree_mod_log;
683         node = tm_root->rb_node;
684         while (node) {
685                 cur = container_of(node, struct tree_mod_elem, node);
686                 if (cur->index < index) {
687                         node = node->rb_left;
688                 } else if (cur->index > index) {
689                         node = node->rb_right;
690                 } else if (cur->seq < min_seq) {
691                         node = node->rb_left;
692                 } else if (!smallest) {
693                         /* we want the node with the highest seq */
694                         if (found)
695                                 BUG_ON(found->seq > cur->seq);
696                         found = cur;
697                         node = node->rb_left;
698                 } else if (cur->seq > min_seq) {
699                         /* we want the node with the smallest seq */
700                         if (found)
701                                 BUG_ON(found->seq < cur->seq);
702                         found = cur;
703                         node = node->rb_right;
704                 } else {
705                         found = cur;
706                         break;
707                 }
708         }
709         tree_mod_log_read_unlock(fs_info);
710
711         return found;
712 }
713
714 /*
715  * this returns the element from the log with the smallest time sequence
716  * value that's in the log (the oldest log item). any element with a time
717  * sequence lower than min_seq will be ignored.
718  */
719 static struct tree_mod_elem *
720 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
721                            u64 min_seq)
722 {
723         return __tree_mod_log_search(fs_info, start, min_seq, 1);
724 }
725
726 /*
727  * this returns the element from the log with the largest time sequence
728  * value that's in the log (the most recent log item). any element with
729  * a time sequence lower than min_seq will be ignored.
730  */
731 static struct tree_mod_elem *
732 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
733 {
734         return __tree_mod_log_search(fs_info, start, min_seq, 0);
735 }
736
737 static noinline void
738 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
739                      struct extent_buffer *src, unsigned long dst_offset,
740                      unsigned long src_offset, int nr_items)
741 {
742         int ret;
743         int i;
744
745         if (tree_mod_dont_log(fs_info, NULL))
746                 return;
747
748         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0) {
749                 tree_mod_log_write_unlock(fs_info);
750                 return;
751         }
752
753         for (i = 0; i < nr_items; i++) {
754                 ret = tree_mod_log_insert_key_locked(fs_info, src,
755                                                      i + src_offset,
756                                                      MOD_LOG_KEY_REMOVE);
757                 BUG_ON(ret < 0);
758                 ret = tree_mod_log_insert_key_locked(fs_info, dst,
759                                                      i + dst_offset,
760                                                      MOD_LOG_KEY_ADD);
761                 BUG_ON(ret < 0);
762         }
763
764         tree_mod_log_write_unlock(fs_info);
765 }
766
767 static inline void
768 tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
769                      int dst_offset, int src_offset, int nr_items)
770 {
771         int ret;
772         ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
773                                        nr_items, GFP_NOFS);
774         BUG_ON(ret < 0);
775 }
776
777 static noinline void
778 tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
779                           struct extent_buffer *eb,
780                           struct btrfs_disk_key *disk_key, int slot, int atomic)
781 {
782         int ret;
783
784         ret = tree_mod_log_insert_key_mask(fs_info, eb, slot,
785                                            MOD_LOG_KEY_REPLACE,
786                                            atomic ? GFP_ATOMIC : GFP_NOFS);
787         BUG_ON(ret < 0);
788 }
789
790 static noinline void
791 tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
792 {
793         if (tree_mod_dont_log(fs_info, eb))
794                 return;
795
796         __tree_mod_log_free_eb(fs_info, eb);
797
798         tree_mod_log_write_unlock(fs_info);
799 }
800
801 static noinline void
802 tree_mod_log_set_root_pointer(struct btrfs_root *root,
803                               struct extent_buffer *new_root_node)
804 {
805         int ret;
806         ret = tree_mod_log_insert_root(root->fs_info, root->node,
807                                        new_root_node, GFP_NOFS);
808         BUG_ON(ret < 0);
809 }
810
811 /*
812  * check if the tree block can be shared by multiple trees
813  */
814 int btrfs_block_can_be_shared(struct btrfs_root *root,
815                               struct extent_buffer *buf)
816 {
817         /*
818          * Tree blocks not in refernece counted trees and tree roots
819          * are never shared. If a block was allocated after the last
820          * snapshot and the block was not allocated by tree relocation,
821          * we know the block is not shared.
822          */
823         if (root->ref_cows &&
824             buf != root->node && buf != root->commit_root &&
825             (btrfs_header_generation(buf) <=
826              btrfs_root_last_snapshot(&root->root_item) ||
827              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
828                 return 1;
829 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
830         if (root->ref_cows &&
831             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
832                 return 1;
833 #endif
834         return 0;
835 }
836
837 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
838                                        struct btrfs_root *root,
839                                        struct extent_buffer *buf,
840                                        struct extent_buffer *cow,
841                                        int *last_ref)
842 {
843         u64 refs;
844         u64 owner;
845         u64 flags;
846         u64 new_flags = 0;
847         int ret;
848
849         /*
850          * Backrefs update rules:
851          *
852          * Always use full backrefs for extent pointers in tree block
853          * allocated by tree relocation.
854          *
855          * If a shared tree block is no longer referenced by its owner
856          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
857          * use full backrefs for extent pointers in tree block.
858          *
859          * If a tree block is been relocating
860          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
861          * use full backrefs for extent pointers in tree block.
862          * The reason for this is some operations (such as drop tree)
863          * are only allowed for blocks use full backrefs.
864          */
865
866         if (btrfs_block_can_be_shared(root, buf)) {
867                 ret = btrfs_lookup_extent_info(trans, root, buf->start,
868                                                buf->len, &refs, &flags);
869                 if (ret)
870                         return ret;
871                 if (refs == 0) {
872                         ret = -EROFS;
873                         btrfs_std_error(root->fs_info, ret);
874                         return ret;
875                 }
876         } else {
877                 refs = 1;
878                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
879                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
880                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
881                 else
882                         flags = 0;
883         }
884
885         owner = btrfs_header_owner(buf);
886         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
887                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
888
889         if (refs > 1) {
890                 if ((owner == root->root_key.objectid ||
891                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
892                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
893                         ret = btrfs_inc_ref(trans, root, buf, 1, 1);
894                         BUG_ON(ret); /* -ENOMEM */
895
896                         if (root->root_key.objectid ==
897                             BTRFS_TREE_RELOC_OBJECTID) {
898                                 ret = btrfs_dec_ref(trans, root, buf, 0, 1);
899                                 BUG_ON(ret); /* -ENOMEM */
900                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
901                                 BUG_ON(ret); /* -ENOMEM */
902                         }
903                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
904                 } else {
905
906                         if (root->root_key.objectid ==
907                             BTRFS_TREE_RELOC_OBJECTID)
908                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
909                         else
910                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
911                         BUG_ON(ret); /* -ENOMEM */
912                 }
913                 if (new_flags != 0) {
914                         ret = btrfs_set_disk_extent_flags(trans, root,
915                                                           buf->start,
916                                                           buf->len,
917                                                           new_flags, 0);
918                         if (ret)
919                                 return ret;
920                 }
921         } else {
922                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
923                         if (root->root_key.objectid ==
924                             BTRFS_TREE_RELOC_OBJECTID)
925                                 ret = btrfs_inc_ref(trans, root, cow, 1, 1);
926                         else
927                                 ret = btrfs_inc_ref(trans, root, cow, 0, 1);
928                         BUG_ON(ret); /* -ENOMEM */
929                         ret = btrfs_dec_ref(trans, root, buf, 1, 1);
930                         BUG_ON(ret); /* -ENOMEM */
931                 }
932                 tree_mod_log_free_eb(root->fs_info, buf);
933                 clean_tree_block(trans, root, buf);
934                 *last_ref = 1;
935         }
936         return 0;
937 }
938
939 /*
940  * does the dirty work in cow of a single block.  The parent block (if
941  * supplied) is updated to point to the new cow copy.  The new buffer is marked
942  * dirty and returned locked.  If you modify the block it needs to be marked
943  * dirty again.
944  *
945  * search_start -- an allocation hint for the new block
946  *
947  * empty_size -- a hint that you plan on doing more cow.  This is the size in
948  * bytes the allocator should try to find free next to the block it returns.
949  * This is just a hint and may be ignored by the allocator.
950  */
951 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
952                              struct btrfs_root *root,
953                              struct extent_buffer *buf,
954                              struct extent_buffer *parent, int parent_slot,
955                              struct extent_buffer **cow_ret,
956                              u64 search_start, u64 empty_size)
957 {
958         struct btrfs_disk_key disk_key;
959         struct extent_buffer *cow;
960         int level, ret;
961         int last_ref = 0;
962         int unlock_orig = 0;
963         u64 parent_start;
964
965         if (*cow_ret == buf)
966                 unlock_orig = 1;
967
968         btrfs_assert_tree_locked(buf);
969
970         WARN_ON(root->ref_cows && trans->transid !=
971                 root->fs_info->running_transaction->transid);
972         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
973
974         level = btrfs_header_level(buf);
975
976         if (level == 0)
977                 btrfs_item_key(buf, &disk_key, 0);
978         else
979                 btrfs_node_key(buf, &disk_key, 0);
980
981         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
982                 if (parent)
983                         parent_start = parent->start;
984                 else
985                         parent_start = 0;
986         } else
987                 parent_start = 0;
988
989         cow = btrfs_alloc_free_block(trans, root, buf->len, parent_start,
990                                      root->root_key.objectid, &disk_key,
991                                      level, search_start, empty_size);
992         if (IS_ERR(cow))
993                 return PTR_ERR(cow);
994
995         /* cow is set to blocking by btrfs_init_new_buffer */
996
997         copy_extent_buffer(cow, buf, 0, 0, cow->len);
998         btrfs_set_header_bytenr(cow, cow->start);
999         btrfs_set_header_generation(cow, trans->transid);
1000         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1001         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1002                                      BTRFS_HEADER_FLAG_RELOC);
1003         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1004                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1005         else
1006                 btrfs_set_header_owner(cow, root->root_key.objectid);
1007
1008         write_extent_buffer(cow, root->fs_info->fsid,
1009                             (unsigned long)btrfs_header_fsid(cow),
1010                             BTRFS_FSID_SIZE);
1011
1012         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1013         if (ret) {
1014                 btrfs_abort_transaction(trans, root, ret);
1015                 return ret;
1016         }
1017
1018         if (root->ref_cows)
1019                 btrfs_reloc_cow_block(trans, root, buf, cow);
1020
1021         if (buf == root->node) {
1022                 WARN_ON(parent && parent != buf);
1023                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1024                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1025                         parent_start = buf->start;
1026                 else
1027                         parent_start = 0;
1028
1029                 extent_buffer_get(cow);
1030                 tree_mod_log_set_root_pointer(root, cow);
1031                 rcu_assign_pointer(root->node, cow);
1032
1033                 btrfs_free_tree_block(trans, root, buf, parent_start,
1034                                       last_ref);
1035                 free_extent_buffer(buf);
1036                 add_root_to_dirty_list(root);
1037         } else {
1038                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1039                         parent_start = parent->start;
1040                 else
1041                         parent_start = 0;
1042
1043                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1044                 tree_mod_log_insert_key(root->fs_info, parent, parent_slot,
1045                                         MOD_LOG_KEY_REPLACE);
1046                 btrfs_set_node_blockptr(parent, parent_slot,
1047                                         cow->start);
1048                 btrfs_set_node_ptr_generation(parent, parent_slot,
1049                                               trans->transid);
1050                 btrfs_mark_buffer_dirty(parent);
1051                 btrfs_free_tree_block(trans, root, buf, parent_start,
1052                                       last_ref);
1053         }
1054         if (unlock_orig)
1055                 btrfs_tree_unlock(buf);
1056         free_extent_buffer_stale(buf);
1057         btrfs_mark_buffer_dirty(cow);
1058         *cow_ret = cow;
1059         return 0;
1060 }
1061
1062 /*
1063  * returns the logical address of the oldest predecessor of the given root.
1064  * entries older than time_seq are ignored.
1065  */
1066 static struct tree_mod_elem *
1067 __tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1068                            struct btrfs_root *root, u64 time_seq)
1069 {
1070         struct tree_mod_elem *tm;
1071         struct tree_mod_elem *found = NULL;
1072         u64 root_logical = root->node->start;
1073         int looped = 0;
1074
1075         if (!time_seq)
1076                 return 0;
1077
1078         /*
1079          * the very last operation that's logged for a root is the replacement
1080          * operation (if it is replaced at all). this has the index of the *new*
1081          * root, making it the very first operation that's logged for this root.
1082          */
1083         while (1) {
1084                 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1085                                                 time_seq);
1086                 if (!looped && !tm)
1087                         return 0;
1088                 /*
1089                  * if there are no tree operation for the oldest root, we simply
1090                  * return it. this should only happen if that (old) root is at
1091                  * level 0.
1092                  */
1093                 if (!tm)
1094                         break;
1095
1096                 /*
1097                  * if there's an operation that's not a root replacement, we
1098                  * found the oldest version of our root. normally, we'll find a
1099                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1100                  */
1101                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1102                         break;
1103
1104                 found = tm;
1105                 root_logical = tm->old_root.logical;
1106                 BUG_ON(root_logical == root->node->start);
1107                 looped = 1;
1108         }
1109
1110         /* if there's no old root to return, return what we found instead */
1111         if (!found)
1112                 found = tm;
1113
1114         return found;
1115 }
1116
1117 /*
1118  * tm is a pointer to the first operation to rewind within eb. then, all
1119  * previous operations will be rewinded (until we reach something older than
1120  * time_seq).
1121  */
1122 static void
1123 __tree_mod_log_rewind(struct extent_buffer *eb, u64 time_seq,
1124                       struct tree_mod_elem *first_tm)
1125 {
1126         u32 n;
1127         struct rb_node *next;
1128         struct tree_mod_elem *tm = first_tm;
1129         unsigned long o_dst;
1130         unsigned long o_src;
1131         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1132
1133         n = btrfs_header_nritems(eb);
1134         while (tm && tm->seq >= time_seq) {
1135                 /*
1136                  * all the operations are recorded with the operator used for
1137                  * the modification. as we're going backwards, we do the
1138                  * opposite of each operation here.
1139                  */
1140                 switch (tm->op) {
1141                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1142                         BUG_ON(tm->slot < n);
1143                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1144                 case MOD_LOG_KEY_REMOVE:
1145                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1146                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1147                         btrfs_set_node_ptr_generation(eb, tm->slot,
1148                                                       tm->generation);
1149                         n++;
1150                         break;
1151                 case MOD_LOG_KEY_REPLACE:
1152                         BUG_ON(tm->slot >= n);
1153                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1154                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1155                         btrfs_set_node_ptr_generation(eb, tm->slot,
1156                                                       tm->generation);
1157                         break;
1158                 case MOD_LOG_KEY_ADD:
1159                         /* if a move operation is needed it's in the log */
1160                         n--;
1161                         break;
1162                 case MOD_LOG_MOVE_KEYS:
1163                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1164                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1165                         memmove_extent_buffer(eb, o_dst, o_src,
1166                                               tm->move.nr_items * p_size);
1167                         break;
1168                 case MOD_LOG_ROOT_REPLACE:
1169                         /*
1170                          * this operation is special. for roots, this must be
1171                          * handled explicitly before rewinding.
1172                          * for non-roots, this operation may exist if the node
1173                          * was a root: root A -> child B; then A gets empty and
1174                          * B is promoted to the new root. in the mod log, we'll
1175                          * have a root-replace operation for B, a tree block
1176                          * that is no root. we simply ignore that operation.
1177                          */
1178                         break;
1179                 }
1180                 next = rb_next(&tm->node);
1181                 if (!next)
1182                         break;
1183                 tm = container_of(next, struct tree_mod_elem, node);
1184                 if (tm->index != first_tm->index)
1185                         break;
1186         }
1187         btrfs_set_header_nritems(eb, n);
1188 }
1189
1190 static struct extent_buffer *
1191 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1192                     u64 time_seq)
1193 {
1194         struct extent_buffer *eb_rewin;
1195         struct tree_mod_elem *tm;
1196
1197         if (!time_seq)
1198                 return eb;
1199
1200         if (btrfs_header_level(eb) == 0)
1201                 return eb;
1202
1203         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1204         if (!tm)
1205                 return eb;
1206
1207         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1208                 BUG_ON(tm->slot != 0);
1209                 eb_rewin = alloc_dummy_extent_buffer(eb->start,
1210                                                 fs_info->tree_root->nodesize);
1211                 BUG_ON(!eb_rewin);
1212                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1213                 btrfs_set_header_backref_rev(eb_rewin,
1214                                              btrfs_header_backref_rev(eb));
1215                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1216                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1217         } else {
1218                 eb_rewin = btrfs_clone_extent_buffer(eb);
1219                 BUG_ON(!eb_rewin);
1220         }
1221
1222         extent_buffer_get(eb_rewin);
1223         free_extent_buffer(eb);
1224
1225         __tree_mod_log_rewind(eb_rewin, time_seq, tm);
1226         WARN_ON(btrfs_header_nritems(eb_rewin) >
1227                 BTRFS_NODEPTRS_PER_BLOCK(fs_info->fs_root));
1228
1229         return eb_rewin;
1230 }
1231
1232 /*
1233  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1234  * value. If there are no changes, the current root->root_node is returned. If
1235  * anything changed in between, there's a fresh buffer allocated on which the
1236  * rewind operations are done. In any case, the returned buffer is read locked.
1237  * Returns NULL on error (with no locks held).
1238  */
1239 static inline struct extent_buffer *
1240 get_old_root(struct btrfs_root *root, u64 time_seq)
1241 {
1242         struct tree_mod_elem *tm;
1243         struct extent_buffer *eb;
1244         struct extent_buffer *old;
1245         struct tree_mod_root *old_root = NULL;
1246         u64 old_generation = 0;
1247         u64 logical;
1248         u32 blocksize;
1249
1250         eb = btrfs_read_lock_root_node(root);
1251         tm = __tree_mod_log_oldest_root(root->fs_info, root, time_seq);
1252         if (!tm)
1253                 return root->node;
1254
1255         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1256                 old_root = &tm->old_root;
1257                 old_generation = tm->generation;
1258                 logical = old_root->logical;
1259         } else {
1260                 logical = root->node->start;
1261         }
1262
1263         tm = tree_mod_log_search(root->fs_info, logical, time_seq);
1264         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1265                 btrfs_tree_read_unlock(root->node);
1266                 free_extent_buffer(root->node);
1267                 blocksize = btrfs_level_size(root, old_root->level);
1268                 old = read_tree_block(root, logical, blocksize, 0);
1269                 if (!old) {
1270                         pr_warn("btrfs: failed to read tree block %llu from get_old_root\n",
1271                                 logical);
1272                         WARN_ON(1);
1273                 } else {
1274                         eb = btrfs_clone_extent_buffer(old);
1275                         free_extent_buffer(old);
1276                 }
1277         } else if (old_root) {
1278                 btrfs_tree_read_unlock(root->node);
1279                 free_extent_buffer(root->node);
1280                 eb = alloc_dummy_extent_buffer(logical, root->nodesize);
1281         } else {
1282                 eb = btrfs_clone_extent_buffer(root->node);
1283                 btrfs_tree_read_unlock(root->node);
1284                 free_extent_buffer(root->node);
1285         }
1286
1287         if (!eb)
1288                 return NULL;
1289         extent_buffer_get(eb);
1290         btrfs_tree_read_lock(eb);
1291         if (old_root) {
1292                 btrfs_set_header_bytenr(eb, eb->start);
1293                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1294                 btrfs_set_header_owner(eb, root->root_key.objectid);
1295                 btrfs_set_header_level(eb, old_root->level);
1296                 btrfs_set_header_generation(eb, old_generation);
1297         }
1298         if (tm)
1299                 __tree_mod_log_rewind(eb, time_seq, tm);
1300         else
1301                 WARN_ON(btrfs_header_level(eb) != 0);
1302         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(root));
1303
1304         return eb;
1305 }
1306
1307 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1308 {
1309         struct tree_mod_elem *tm;
1310         int level;
1311
1312         tm = __tree_mod_log_oldest_root(root->fs_info, root, time_seq);
1313         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1314                 level = tm->old_root.level;
1315         } else {
1316                 rcu_read_lock();
1317                 level = btrfs_header_level(root->node);
1318                 rcu_read_unlock();
1319         }
1320
1321         return level;
1322 }
1323
1324 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1325                                    struct btrfs_root *root,
1326                                    struct extent_buffer *buf)
1327 {
1328         /* ensure we can see the force_cow */
1329         smp_rmb();
1330
1331         /*
1332          * We do not need to cow a block if
1333          * 1) this block is not created or changed in this transaction;
1334          * 2) this block does not belong to TREE_RELOC tree;
1335          * 3) the root is not forced COW.
1336          *
1337          * What is forced COW:
1338          *    when we create snapshot during commiting the transaction,
1339          *    after we've finished coping src root, we must COW the shared
1340          *    block to ensure the metadata consistency.
1341          */
1342         if (btrfs_header_generation(buf) == trans->transid &&
1343             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1344             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1345               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1346             !root->force_cow)
1347                 return 0;
1348         return 1;
1349 }
1350
1351 /*
1352  * cows a single block, see __btrfs_cow_block for the real work.
1353  * This version of it has extra checks so that a block isn't cow'd more than
1354  * once per transaction, as long as it hasn't been written yet
1355  */
1356 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1357                     struct btrfs_root *root, struct extent_buffer *buf,
1358                     struct extent_buffer *parent, int parent_slot,
1359                     struct extent_buffer **cow_ret)
1360 {
1361         u64 search_start;
1362         int ret;
1363
1364         if (trans->transaction != root->fs_info->running_transaction) {
1365                 printk(KERN_CRIT "trans %llu running %llu\n",
1366                        (unsigned long long)trans->transid,
1367                        (unsigned long long)
1368                        root->fs_info->running_transaction->transid);
1369                 WARN_ON(1);
1370         }
1371         if (trans->transid != root->fs_info->generation) {
1372                 printk(KERN_CRIT "trans %llu running %llu\n",
1373                        (unsigned long long)trans->transid,
1374                        (unsigned long long)root->fs_info->generation);
1375                 WARN_ON(1);
1376         }
1377
1378         if (!should_cow_block(trans, root, buf)) {
1379                 *cow_ret = buf;
1380                 return 0;
1381         }
1382
1383         search_start = buf->start & ~((u64)(1024 * 1024 * 1024) - 1);
1384
1385         if (parent)
1386                 btrfs_set_lock_blocking(parent);
1387         btrfs_set_lock_blocking(buf);
1388
1389         ret = __btrfs_cow_block(trans, root, buf, parent,
1390                                  parent_slot, cow_ret, search_start, 0);
1391
1392         trace_btrfs_cow_block(root, buf, *cow_ret);
1393
1394         return ret;
1395 }
1396
1397 /*
1398  * helper function for defrag to decide if two blocks pointed to by a
1399  * node are actually close by
1400  */
1401 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1402 {
1403         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1404                 return 1;
1405         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1406                 return 1;
1407         return 0;
1408 }
1409
1410 /*
1411  * compare two keys in a memcmp fashion
1412  */
1413 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
1414 {
1415         struct btrfs_key k1;
1416
1417         btrfs_disk_key_to_cpu(&k1, disk);
1418
1419         return btrfs_comp_cpu_keys(&k1, k2);
1420 }
1421
1422 /*
1423  * same as comp_keys only with two btrfs_key's
1424  */
1425 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
1426 {
1427         if (k1->objectid > k2->objectid)
1428                 return 1;
1429         if (k1->objectid < k2->objectid)
1430                 return -1;
1431         if (k1->type > k2->type)
1432                 return 1;
1433         if (k1->type < k2->type)
1434                 return -1;
1435         if (k1->offset > k2->offset)
1436                 return 1;
1437         if (k1->offset < k2->offset)
1438                 return -1;
1439         return 0;
1440 }
1441
1442 /*
1443  * this is used by the defrag code to go through all the
1444  * leaves pointed to by a node and reallocate them so that
1445  * disk order is close to key order
1446  */
1447 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1448                        struct btrfs_root *root, struct extent_buffer *parent,
1449                        int start_slot, int cache_only, u64 *last_ret,
1450                        struct btrfs_key *progress)
1451 {
1452         struct extent_buffer *cur;
1453         u64 blocknr;
1454         u64 gen;
1455         u64 search_start = *last_ret;
1456         u64 last_block = 0;
1457         u64 other;
1458         u32 parent_nritems;
1459         int end_slot;
1460         int i;
1461         int err = 0;
1462         int parent_level;
1463         int uptodate;
1464         u32 blocksize;
1465         int progress_passed = 0;
1466         struct btrfs_disk_key disk_key;
1467
1468         parent_level = btrfs_header_level(parent);
1469         if (cache_only && parent_level != 1)
1470                 return 0;
1471
1472         if (trans->transaction != root->fs_info->running_transaction)
1473                 WARN_ON(1);
1474         if (trans->transid != root->fs_info->generation)
1475                 WARN_ON(1);
1476
1477         parent_nritems = btrfs_header_nritems(parent);
1478         blocksize = btrfs_level_size(root, parent_level - 1);
1479         end_slot = parent_nritems;
1480
1481         if (parent_nritems == 1)
1482                 return 0;
1483
1484         btrfs_set_lock_blocking(parent);
1485
1486         for (i = start_slot; i < end_slot; i++) {
1487                 int close = 1;
1488
1489                 btrfs_node_key(parent, &disk_key, i);
1490                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1491                         continue;
1492
1493                 progress_passed = 1;
1494                 blocknr = btrfs_node_blockptr(parent, i);
1495                 gen = btrfs_node_ptr_generation(parent, i);
1496                 if (last_block == 0)
1497                         last_block = blocknr;
1498
1499                 if (i > 0) {
1500                         other = btrfs_node_blockptr(parent, i - 1);
1501                         close = close_blocks(blocknr, other, blocksize);
1502                 }
1503                 if (!close && i < end_slot - 2) {
1504                         other = btrfs_node_blockptr(parent, i + 1);
1505                         close = close_blocks(blocknr, other, blocksize);
1506                 }
1507                 if (close) {
1508                         last_block = blocknr;
1509                         continue;
1510                 }
1511
1512                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
1513                 if (cur)
1514                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1515                 else
1516                         uptodate = 0;
1517                 if (!cur || !uptodate) {
1518                         if (cache_only) {
1519                                 free_extent_buffer(cur);
1520                                 continue;
1521                         }
1522                         if (!cur) {
1523                                 cur = read_tree_block(root, blocknr,
1524                                                          blocksize, gen);
1525                                 if (!cur)
1526                                         return -EIO;
1527                         } else if (!uptodate) {
1528                                 err = btrfs_read_buffer(cur, gen);
1529                                 if (err) {
1530                                         free_extent_buffer(cur);
1531                                         return err;
1532                                 }
1533                         }
1534                 }
1535                 if (search_start == 0)
1536                         search_start = last_block;
1537
1538                 btrfs_tree_lock(cur);
1539                 btrfs_set_lock_blocking(cur);
1540                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1541                                         &cur, search_start,
1542                                         min(16 * blocksize,
1543                                             (end_slot - i) * blocksize));
1544                 if (err) {
1545                         btrfs_tree_unlock(cur);
1546                         free_extent_buffer(cur);
1547                         break;
1548                 }
1549                 search_start = cur->start;
1550                 last_block = cur->start;
1551                 *last_ret = search_start;
1552                 btrfs_tree_unlock(cur);
1553                 free_extent_buffer(cur);
1554         }
1555         return err;
1556 }
1557
1558 /*
1559  * The leaf data grows from end-to-front in the node.
1560  * this returns the address of the start of the last item,
1561  * which is the stop of the leaf data stack
1562  */
1563 static inline unsigned int leaf_data_end(struct btrfs_root *root,
1564                                          struct extent_buffer *leaf)
1565 {
1566         u32 nr = btrfs_header_nritems(leaf);
1567         if (nr == 0)
1568                 return BTRFS_LEAF_DATA_SIZE(root);
1569         return btrfs_item_offset_nr(leaf, nr - 1);
1570 }
1571
1572
1573 /*
1574  * search for key in the extent_buffer.  The items start at offset p,
1575  * and they are item_size apart.  There are 'max' items in p.
1576  *
1577  * the slot in the array is returned via slot, and it points to
1578  * the place where you would insert key if it is not found in
1579  * the array.
1580  *
1581  * slot may point to max if the key is bigger than all of the keys
1582  */
1583 static noinline int generic_bin_search(struct extent_buffer *eb,
1584                                        unsigned long p,
1585                                        int item_size, struct btrfs_key *key,
1586                                        int max, int *slot)
1587 {
1588         int low = 0;
1589         int high = max;
1590         int mid;
1591         int ret;
1592         struct btrfs_disk_key *tmp = NULL;
1593         struct btrfs_disk_key unaligned;
1594         unsigned long offset;
1595         char *kaddr = NULL;
1596         unsigned long map_start = 0;
1597         unsigned long map_len = 0;
1598         int err;
1599
1600         while (low < high) {
1601                 mid = (low + high) / 2;
1602                 offset = p + mid * item_size;
1603
1604                 if (!kaddr || offset < map_start ||
1605                     (offset + sizeof(struct btrfs_disk_key)) >
1606                     map_start + map_len) {
1607
1608                         err = map_private_extent_buffer(eb, offset,
1609                                                 sizeof(struct btrfs_disk_key),
1610                                                 &kaddr, &map_start, &map_len);
1611
1612                         if (!err) {
1613                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1614                                                         map_start);
1615                         } else {
1616                                 read_extent_buffer(eb, &unaligned,
1617                                                    offset, sizeof(unaligned));
1618                                 tmp = &unaligned;
1619                         }
1620
1621                 } else {
1622                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1623                                                         map_start);
1624                 }
1625                 ret = comp_keys(tmp, key);
1626
1627                 if (ret < 0)
1628                         low = mid + 1;
1629                 else if (ret > 0)
1630                         high = mid;
1631                 else {
1632                         *slot = mid;
1633                         return 0;
1634                 }
1635         }
1636         *slot = low;
1637         return 1;
1638 }
1639
1640 /*
1641  * simple bin_search frontend that does the right thing for
1642  * leaves vs nodes
1643  */
1644 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1645                       int level, int *slot)
1646 {
1647         if (level == 0)
1648                 return generic_bin_search(eb,
1649                                           offsetof(struct btrfs_leaf, items),
1650                                           sizeof(struct btrfs_item),
1651                                           key, btrfs_header_nritems(eb),
1652                                           slot);
1653         else
1654                 return generic_bin_search(eb,
1655                                           offsetof(struct btrfs_node, ptrs),
1656                                           sizeof(struct btrfs_key_ptr),
1657                                           key, btrfs_header_nritems(eb),
1658                                           slot);
1659 }
1660
1661 int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1662                      int level, int *slot)
1663 {
1664         return bin_search(eb, key, level, slot);
1665 }
1666
1667 static void root_add_used(struct btrfs_root *root, u32 size)
1668 {
1669         spin_lock(&root->accounting_lock);
1670         btrfs_set_root_used(&root->root_item,
1671                             btrfs_root_used(&root->root_item) + size);
1672         spin_unlock(&root->accounting_lock);
1673 }
1674
1675 static void root_sub_used(struct btrfs_root *root, u32 size)
1676 {
1677         spin_lock(&root->accounting_lock);
1678         btrfs_set_root_used(&root->root_item,
1679                             btrfs_root_used(&root->root_item) - size);
1680         spin_unlock(&root->accounting_lock);
1681 }
1682
1683 /* given a node and slot number, this reads the blocks it points to.  The
1684  * extent buffer is returned with a reference taken (but unlocked).
1685  * NULL is returned on error.
1686  */
1687 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
1688                                    struct extent_buffer *parent, int slot)
1689 {
1690         int level = btrfs_header_level(parent);
1691         if (slot < 0)
1692                 return NULL;
1693         if (slot >= btrfs_header_nritems(parent))
1694                 return NULL;
1695
1696         BUG_ON(level == 0);
1697
1698         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
1699                        btrfs_level_size(root, level - 1),
1700                        btrfs_node_ptr_generation(parent, slot));
1701 }
1702
1703 /*
1704  * node level balancing, used to make sure nodes are in proper order for
1705  * item deletion.  We balance from the top down, so we have to make sure
1706  * that a deletion won't leave an node completely empty later on.
1707  */
1708 static noinline int balance_level(struct btrfs_trans_handle *trans,
1709                          struct btrfs_root *root,
1710                          struct btrfs_path *path, int level)
1711 {
1712         struct extent_buffer *right = NULL;
1713         struct extent_buffer *mid;
1714         struct extent_buffer *left = NULL;
1715         struct extent_buffer *parent = NULL;
1716         int ret = 0;
1717         int wret;
1718         int pslot;
1719         int orig_slot = path->slots[level];
1720         u64 orig_ptr;
1721
1722         if (level == 0)
1723                 return 0;
1724
1725         mid = path->nodes[level];
1726
1727         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1728                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1729         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1730
1731         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1732
1733         if (level < BTRFS_MAX_LEVEL - 1) {
1734                 parent = path->nodes[level + 1];
1735                 pslot = path->slots[level + 1];
1736         }
1737
1738         /*
1739          * deal with the case where there is only one pointer in the root
1740          * by promoting the node below to a root
1741          */
1742         if (!parent) {
1743                 struct extent_buffer *child;
1744
1745                 if (btrfs_header_nritems(mid) != 1)
1746                         return 0;
1747
1748                 /* promote the child to a root */
1749                 child = read_node_slot(root, mid, 0);
1750                 if (!child) {
1751                         ret = -EROFS;
1752                         btrfs_std_error(root->fs_info, ret);
1753                         goto enospc;
1754                 }
1755
1756                 btrfs_tree_lock(child);
1757                 btrfs_set_lock_blocking(child);
1758                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1759                 if (ret) {
1760                         btrfs_tree_unlock(child);
1761                         free_extent_buffer(child);
1762                         goto enospc;
1763                 }
1764
1765                 tree_mod_log_free_eb(root->fs_info, root->node);
1766                 tree_mod_log_set_root_pointer(root, child);
1767                 rcu_assign_pointer(root->node, child);
1768
1769                 add_root_to_dirty_list(root);
1770                 btrfs_tree_unlock(child);
1771
1772                 path->locks[level] = 0;
1773                 path->nodes[level] = NULL;
1774                 clean_tree_block(trans, root, mid);
1775                 btrfs_tree_unlock(mid);
1776                 /* once for the path */
1777                 free_extent_buffer(mid);
1778
1779                 root_sub_used(root, mid->len);
1780                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1781                 /* once for the root ptr */
1782                 free_extent_buffer_stale(mid);
1783                 return 0;
1784         }
1785         if (btrfs_header_nritems(mid) >
1786             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
1787                 return 0;
1788
1789         left = read_node_slot(root, parent, pslot - 1);
1790         if (left) {
1791                 btrfs_tree_lock(left);
1792                 btrfs_set_lock_blocking(left);
1793                 wret = btrfs_cow_block(trans, root, left,
1794                                        parent, pslot - 1, &left);
1795                 if (wret) {
1796                         ret = wret;
1797                         goto enospc;
1798                 }
1799         }
1800         right = read_node_slot(root, parent, pslot + 1);
1801         if (right) {
1802                 btrfs_tree_lock(right);
1803                 btrfs_set_lock_blocking(right);
1804                 wret = btrfs_cow_block(trans, root, right,
1805                                        parent, pslot + 1, &right);
1806                 if (wret) {
1807                         ret = wret;
1808                         goto enospc;
1809                 }
1810         }
1811
1812         /* first, try to make some room in the middle buffer */
1813         if (left) {
1814                 orig_slot += btrfs_header_nritems(left);
1815                 wret = push_node_left(trans, root, left, mid, 1);
1816                 if (wret < 0)
1817                         ret = wret;
1818         }
1819
1820         /*
1821          * then try to empty the right most buffer into the middle
1822          */
1823         if (right) {
1824                 wret = push_node_left(trans, root, mid, right, 1);
1825                 if (wret < 0 && wret != -ENOSPC)
1826                         ret = wret;
1827                 if (btrfs_header_nritems(right) == 0) {
1828                         clean_tree_block(trans, root, right);
1829                         btrfs_tree_unlock(right);
1830                         del_ptr(trans, root, path, level + 1, pslot + 1, 1);
1831                         root_sub_used(root, right->len);
1832                         btrfs_free_tree_block(trans, root, right, 0, 1);
1833                         free_extent_buffer_stale(right);
1834                         right = NULL;
1835                 } else {
1836                         struct btrfs_disk_key right_key;
1837                         btrfs_node_key(right, &right_key, 0);
1838                         tree_mod_log_set_node_key(root->fs_info, parent,
1839                                                   &right_key, pslot + 1, 0);
1840                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1841                         btrfs_mark_buffer_dirty(parent);
1842                 }
1843         }
1844         if (btrfs_header_nritems(mid) == 1) {
1845                 /*
1846                  * we're not allowed to leave a node with one item in the
1847                  * tree during a delete.  A deletion from lower in the tree
1848                  * could try to delete the only pointer in this node.
1849                  * So, pull some keys from the left.
1850                  * There has to be a left pointer at this point because
1851                  * otherwise we would have pulled some pointers from the
1852                  * right
1853                  */
1854                 if (!left) {
1855                         ret = -EROFS;
1856                         btrfs_std_error(root->fs_info, ret);
1857                         goto enospc;
1858                 }
1859                 wret = balance_node_right(trans, root, mid, left);
1860                 if (wret < 0) {
1861                         ret = wret;
1862                         goto enospc;
1863                 }
1864                 if (wret == 1) {
1865                         wret = push_node_left(trans, root, left, mid, 1);
1866                         if (wret < 0)
1867                                 ret = wret;
1868                 }
1869                 BUG_ON(wret == 1);
1870         }
1871         if (btrfs_header_nritems(mid) == 0) {
1872                 clean_tree_block(trans, root, mid);
1873                 btrfs_tree_unlock(mid);
1874                 del_ptr(trans, root, path, level + 1, pslot, 1);
1875                 root_sub_used(root, mid->len);
1876                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1877                 free_extent_buffer_stale(mid);
1878                 mid = NULL;
1879         } else {
1880                 /* update the parent key to reflect our changes */
1881                 struct btrfs_disk_key mid_key;
1882                 btrfs_node_key(mid, &mid_key, 0);
1883                 tree_mod_log_set_node_key(root->fs_info, parent, &mid_key,
1884                                           pslot, 0);
1885                 btrfs_set_node_key(parent, &mid_key, pslot);
1886                 btrfs_mark_buffer_dirty(parent);
1887         }
1888
1889         /* update the path */
1890         if (left) {
1891                 if (btrfs_header_nritems(left) > orig_slot) {
1892                         extent_buffer_get(left);
1893                         /* left was locked after cow */
1894                         path->nodes[level] = left;
1895                         path->slots[level + 1] -= 1;
1896                         path->slots[level] = orig_slot;
1897                         if (mid) {
1898                                 btrfs_tree_unlock(mid);
1899                                 free_extent_buffer(mid);
1900                         }
1901                 } else {
1902                         orig_slot -= btrfs_header_nritems(left);
1903                         path->slots[level] = orig_slot;
1904                 }
1905         }
1906         /* double check we haven't messed things up */
1907         if (orig_ptr !=
1908             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1909                 BUG();
1910 enospc:
1911         if (right) {
1912                 btrfs_tree_unlock(right);
1913                 free_extent_buffer(right);
1914         }
1915         if (left) {
1916                 if (path->nodes[level] != left)
1917                         btrfs_tree_unlock(left);
1918                 free_extent_buffer(left);
1919         }
1920         return ret;
1921 }
1922
1923 /* Node balancing for insertion.  Here we only split or push nodes around
1924  * when they are completely full.  This is also done top down, so we
1925  * have to be pessimistic.
1926  */
1927 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
1928                                           struct btrfs_root *root,
1929                                           struct btrfs_path *path, int level)
1930 {
1931         struct extent_buffer *right = NULL;
1932         struct extent_buffer *mid;
1933         struct extent_buffer *left = NULL;
1934         struct extent_buffer *parent = NULL;
1935         int ret = 0;
1936         int wret;
1937         int pslot;
1938         int orig_slot = path->slots[level];
1939
1940         if (level == 0)
1941                 return 1;
1942
1943         mid = path->nodes[level];
1944         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1945
1946         if (level < BTRFS_MAX_LEVEL - 1) {
1947                 parent = path->nodes[level + 1];
1948                 pslot = path->slots[level + 1];
1949         }
1950
1951         if (!parent)
1952                 return 1;
1953
1954         left = read_node_slot(root, parent, pslot - 1);
1955
1956         /* first, try to make some room in the middle buffer */
1957         if (left) {
1958                 u32 left_nr;
1959
1960                 btrfs_tree_lock(left);
1961                 btrfs_set_lock_blocking(left);
1962
1963                 left_nr = btrfs_header_nritems(left);
1964                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1965                         wret = 1;
1966                 } else {
1967                         ret = btrfs_cow_block(trans, root, left, parent,
1968                                               pslot - 1, &left);
1969                         if (ret)
1970                                 wret = 1;
1971                         else {
1972                                 wret = push_node_left(trans, root,
1973                                                       left, mid, 0);
1974                         }
1975                 }
1976                 if (wret < 0)
1977                         ret = wret;
1978                 if (wret == 0) {
1979                         struct btrfs_disk_key disk_key;
1980                         orig_slot += left_nr;
1981                         btrfs_node_key(mid, &disk_key, 0);
1982                         tree_mod_log_set_node_key(root->fs_info, parent,
1983                                                   &disk_key, pslot, 0);
1984                         btrfs_set_node_key(parent, &disk_key, pslot);
1985                         btrfs_mark_buffer_dirty(parent);
1986                         if (btrfs_header_nritems(left) > orig_slot) {
1987                                 path->nodes[level] = left;
1988                                 path->slots[level + 1] -= 1;
1989                                 path->slots[level] = orig_slot;
1990                                 btrfs_tree_unlock(mid);
1991                                 free_extent_buffer(mid);
1992                         } else {
1993                                 orig_slot -=
1994                                         btrfs_header_nritems(left);
1995                                 path->slots[level] = orig_slot;
1996                                 btrfs_tree_unlock(left);
1997                                 free_extent_buffer(left);
1998                         }
1999                         return 0;
2000                 }
2001                 btrfs_tree_unlock(left);
2002                 free_extent_buffer(left);
2003         }
2004         right = read_node_slot(root, parent, pslot + 1);
2005
2006         /*
2007          * then try to empty the right most buffer into the middle
2008          */
2009         if (right) {
2010                 u32 right_nr;
2011
2012                 btrfs_tree_lock(right);
2013                 btrfs_set_lock_blocking(right);
2014
2015                 right_nr = btrfs_header_nritems(right);
2016                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
2017                         wret = 1;
2018                 } else {
2019                         ret = btrfs_cow_block(trans, root, right,
2020                                               parent, pslot + 1,
2021                                               &right);
2022                         if (ret)
2023                                 wret = 1;
2024                         else {
2025                                 wret = balance_node_right(trans, root,
2026                                                           right, mid);
2027                         }
2028                 }
2029                 if (wret < 0)
2030                         ret = wret;
2031                 if (wret == 0) {
2032                         struct btrfs_disk_key disk_key;
2033
2034                         btrfs_node_key(right, &disk_key, 0);
2035                         tree_mod_log_set_node_key(root->fs_info, parent,
2036                                                   &disk_key, pslot + 1, 0);
2037                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2038                         btrfs_mark_buffer_dirty(parent);
2039
2040                         if (btrfs_header_nritems(mid) <= orig_slot) {
2041                                 path->nodes[level] = right;
2042                                 path->slots[level + 1] += 1;
2043                                 path->slots[level] = orig_slot -
2044                                         btrfs_header_nritems(mid);
2045                                 btrfs_tree_unlock(mid);
2046                                 free_extent_buffer(mid);
2047                         } else {
2048                                 btrfs_tree_unlock(right);
2049                                 free_extent_buffer(right);
2050                         }
2051                         return 0;
2052                 }
2053                 btrfs_tree_unlock(right);
2054                 free_extent_buffer(right);
2055         }
2056         return 1;
2057 }
2058
2059 /*
2060  * readahead one full node of leaves, finding things that are close
2061  * to the block in 'slot', and triggering ra on them.
2062  */
2063 static void reada_for_search(struct btrfs_root *root,
2064                              struct btrfs_path *path,
2065                              int level, int slot, u64 objectid)
2066 {
2067         struct extent_buffer *node;
2068         struct btrfs_disk_key disk_key;
2069         u32 nritems;
2070         u64 search;
2071         u64 target;
2072         u64 nread = 0;
2073         u64 gen;
2074         int direction = path->reada;
2075         struct extent_buffer *eb;
2076         u32 nr;
2077         u32 blocksize;
2078         u32 nscan = 0;
2079
2080         if (level != 1)
2081                 return;
2082
2083         if (!path->nodes[level])
2084                 return;
2085
2086         node = path->nodes[level];
2087
2088         search = btrfs_node_blockptr(node, slot);
2089         blocksize = btrfs_level_size(root, level - 1);
2090         eb = btrfs_find_tree_block(root, search, blocksize);
2091         if (eb) {
2092                 free_extent_buffer(eb);
2093                 return;
2094         }
2095
2096         target = search;
2097
2098         nritems = btrfs_header_nritems(node);
2099         nr = slot;
2100
2101         while (1) {
2102                 if (direction < 0) {
2103                         if (nr == 0)
2104                                 break;
2105                         nr--;
2106                 } else if (direction > 0) {
2107                         nr++;
2108                         if (nr >= nritems)
2109                                 break;
2110                 }
2111                 if (path->reada < 0 && objectid) {
2112                         btrfs_node_key(node, &disk_key, nr);
2113                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2114                                 break;
2115                 }
2116                 search = btrfs_node_blockptr(node, nr);
2117                 if ((search <= target && target - search <= 65536) ||
2118                     (search > target && search - target <= 65536)) {
2119                         gen = btrfs_node_ptr_generation(node, nr);
2120                         readahead_tree_block(root, search, blocksize, gen);
2121                         nread += blocksize;
2122                 }
2123                 nscan++;
2124                 if ((nread > 65536 || nscan > 32))
2125                         break;
2126         }
2127 }
2128
2129 /*
2130  * returns -EAGAIN if it had to drop the path, or zero if everything was in
2131  * cache
2132  */
2133 static noinline int reada_for_balance(struct btrfs_root *root,
2134                                       struct btrfs_path *path, int level)
2135 {
2136         int slot;
2137         int nritems;
2138         struct extent_buffer *parent;
2139         struct extent_buffer *eb;
2140         u64 gen;
2141         u64 block1 = 0;
2142         u64 block2 = 0;
2143         int ret = 0;
2144         int blocksize;
2145
2146         parent = path->nodes[level + 1];
2147         if (!parent)
2148                 return 0;
2149
2150         nritems = btrfs_header_nritems(parent);
2151         slot = path->slots[level + 1];
2152         blocksize = btrfs_level_size(root, level);
2153
2154         if (slot > 0) {
2155                 block1 = btrfs_node_blockptr(parent, slot - 1);
2156                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2157                 eb = btrfs_find_tree_block(root, block1, blocksize);
2158                 /*
2159                  * if we get -eagain from btrfs_buffer_uptodate, we
2160                  * don't want to return eagain here.  That will loop
2161                  * forever
2162                  */
2163                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2164                         block1 = 0;
2165                 free_extent_buffer(eb);
2166         }
2167         if (slot + 1 < nritems) {
2168                 block2 = btrfs_node_blockptr(parent, slot + 1);
2169                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2170                 eb = btrfs_find_tree_block(root, block2, blocksize);
2171                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2172                         block2 = 0;
2173                 free_extent_buffer(eb);
2174         }
2175         if (block1 || block2) {
2176                 ret = -EAGAIN;
2177
2178                 /* release the whole path */
2179                 btrfs_release_path(path);
2180
2181                 /* read the blocks */
2182                 if (block1)
2183                         readahead_tree_block(root, block1, blocksize, 0);
2184                 if (block2)
2185                         readahead_tree_block(root, block2, blocksize, 0);
2186
2187                 if (block1) {
2188                         eb = read_tree_block(root, block1, blocksize, 0);
2189                         free_extent_buffer(eb);
2190                 }
2191                 if (block2) {
2192                         eb = read_tree_block(root, block2, blocksize, 0);
2193                         free_extent_buffer(eb);
2194                 }
2195         }
2196         return ret;
2197 }
2198
2199
2200 /*
2201  * when we walk down the tree, it is usually safe to unlock the higher layers
2202  * in the tree.  The exceptions are when our path goes through slot 0, because
2203  * operations on the tree might require changing key pointers higher up in the
2204  * tree.
2205  *
2206  * callers might also have set path->keep_locks, which tells this code to keep
2207  * the lock if the path points to the last slot in the block.  This is part of
2208  * walking through the tree, and selecting the next slot in the higher block.
2209  *
2210  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2211  * if lowest_unlock is 1, level 0 won't be unlocked
2212  */
2213 static noinline void unlock_up(struct btrfs_path *path, int level,
2214                                int lowest_unlock, int min_write_lock_level,
2215                                int *write_lock_level)
2216 {
2217         int i;
2218         int skip_level = level;
2219         int no_skips = 0;
2220         struct extent_buffer *t;
2221
2222         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2223                 if (!path->nodes[i])
2224                         break;
2225                 if (!path->locks[i])
2226                         break;
2227                 if (!no_skips && path->slots[i] == 0) {
2228                         skip_level = i + 1;
2229                         continue;
2230                 }
2231                 if (!no_skips && path->keep_locks) {
2232                         u32 nritems;
2233                         t = path->nodes[i];
2234                         nritems = btrfs_header_nritems(t);
2235                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2236                                 skip_level = i + 1;
2237                                 continue;
2238                         }
2239                 }
2240                 if (skip_level < i && i >= lowest_unlock)
2241                         no_skips = 1;
2242
2243                 t = path->nodes[i];
2244                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2245                         btrfs_tree_unlock_rw(t, path->locks[i]);
2246                         path->locks[i] = 0;
2247                         if (write_lock_level &&
2248                             i > min_write_lock_level &&
2249                             i <= *write_lock_level) {
2250                                 *write_lock_level = i - 1;
2251                         }
2252                 }
2253         }
2254 }
2255
2256 /*
2257  * This releases any locks held in the path starting at level and
2258  * going all the way up to the root.
2259  *
2260  * btrfs_search_slot will keep the lock held on higher nodes in a few
2261  * corner cases, such as COW of the block at slot zero in the node.  This
2262  * ignores those rules, and it should only be called when there are no
2263  * more updates to be done higher up in the tree.
2264  */
2265 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2266 {
2267         int i;
2268
2269         if (path->keep_locks)
2270                 return;
2271
2272         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2273                 if (!path->nodes[i])
2274                         continue;
2275                 if (!path->locks[i])
2276                         continue;
2277                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2278                 path->locks[i] = 0;
2279         }
2280 }
2281
2282 /*
2283  * helper function for btrfs_search_slot.  The goal is to find a block
2284  * in cache without setting the path to blocking.  If we find the block
2285  * we return zero and the path is unchanged.
2286  *
2287  * If we can't find the block, we set the path blocking and do some
2288  * reada.  -EAGAIN is returned and the search must be repeated.
2289  */
2290 static int
2291 read_block_for_search(struct btrfs_trans_handle *trans,
2292                        struct btrfs_root *root, struct btrfs_path *p,
2293                        struct extent_buffer **eb_ret, int level, int slot,
2294                        struct btrfs_key *key, u64 time_seq)
2295 {
2296         u64 blocknr;
2297         u64 gen;
2298         u32 blocksize;
2299         struct extent_buffer *b = *eb_ret;
2300         struct extent_buffer *tmp;
2301         int ret;
2302
2303         blocknr = btrfs_node_blockptr(b, slot);
2304         gen = btrfs_node_ptr_generation(b, slot);
2305         blocksize = btrfs_level_size(root, level - 1);
2306
2307         tmp = btrfs_find_tree_block(root, blocknr, blocksize);
2308         if (tmp) {
2309                 /* first we do an atomic uptodate check */
2310                 if (btrfs_buffer_uptodate(tmp, 0, 1) > 0) {
2311                         if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2312                                 /*
2313                                  * we found an up to date block without
2314                                  * sleeping, return
2315                                  * right away
2316                                  */
2317                                 *eb_ret = tmp;
2318                                 return 0;
2319                         }
2320                         /* the pages were up to date, but we failed
2321                          * the generation number check.  Do a full
2322                          * read for the generation number that is correct.
2323                          * We must do this without dropping locks so
2324                          * we can trust our generation number
2325                          */
2326                         free_extent_buffer(tmp);
2327                         btrfs_set_path_blocking(p);
2328
2329                         /* now we're allowed to do a blocking uptodate check */
2330                         tmp = read_tree_block(root, blocknr, blocksize, gen);
2331                         if (tmp && btrfs_buffer_uptodate(tmp, gen, 0) > 0) {
2332                                 *eb_ret = tmp;
2333                                 return 0;
2334                         }
2335                         free_extent_buffer(tmp);
2336                         btrfs_release_path(p);
2337                         return -EIO;
2338                 }
2339         }
2340
2341         /*
2342          * reduce lock contention at high levels
2343          * of the btree by dropping locks before
2344          * we read.  Don't release the lock on the current
2345          * level because we need to walk this node to figure
2346          * out which blocks to read.
2347          */
2348         btrfs_unlock_up_safe(p, level + 1);
2349         btrfs_set_path_blocking(p);
2350
2351         free_extent_buffer(tmp);
2352         if (p->reada)
2353                 reada_for_search(root, p, level, slot, key->objectid);
2354
2355         btrfs_release_path(p);
2356
2357         ret = -EAGAIN;
2358         tmp = read_tree_block(root, blocknr, blocksize, 0);
2359         if (tmp) {
2360                 /*
2361                  * If the read above didn't mark this buffer up to date,
2362                  * it will never end up being up to date.  Set ret to EIO now
2363                  * and give up so that our caller doesn't loop forever
2364                  * on our EAGAINs.
2365                  */
2366                 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2367                         ret = -EIO;
2368                 free_extent_buffer(tmp);
2369         }
2370         return ret;
2371 }
2372
2373 /*
2374  * helper function for btrfs_search_slot.  This does all of the checks
2375  * for node-level blocks and does any balancing required based on
2376  * the ins_len.
2377  *
2378  * If no extra work was required, zero is returned.  If we had to
2379  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2380  * start over
2381  */
2382 static int
2383 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2384                        struct btrfs_root *root, struct btrfs_path *p,
2385                        struct extent_buffer *b, int level, int ins_len,
2386                        int *write_lock_level)
2387 {
2388         int ret;
2389         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2390             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
2391                 int sret;
2392
2393                 if (*write_lock_level < level + 1) {
2394                         *write_lock_level = level + 1;
2395                         btrfs_release_path(p);
2396                         goto again;
2397                 }
2398
2399                 sret = reada_for_balance(root, p, level);
2400                 if (sret)
2401                         goto again;
2402
2403                 btrfs_set_path_blocking(p);
2404                 sret = split_node(trans, root, p, level);
2405                 btrfs_clear_path_blocking(p, NULL, 0);
2406
2407                 BUG_ON(sret > 0);
2408                 if (sret) {
2409                         ret = sret;
2410                         goto done;
2411                 }
2412                 b = p->nodes[level];
2413         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2414                    BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
2415                 int sret;
2416
2417                 if (*write_lock_level < level + 1) {
2418                         *write_lock_level = level + 1;
2419                         btrfs_release_path(p);
2420                         goto again;
2421                 }
2422
2423                 sret = reada_for_balance(root, p, level);
2424                 if (sret)
2425                         goto again;
2426
2427                 btrfs_set_path_blocking(p);
2428                 sret = balance_level(trans, root, p, level);
2429                 btrfs_clear_path_blocking(p, NULL, 0);
2430
2431                 if (sret) {
2432                         ret = sret;
2433                         goto done;
2434                 }
2435                 b = p->nodes[level];
2436                 if (!b) {
2437                         btrfs_release_path(p);
2438                         goto again;
2439                 }
2440                 BUG_ON(btrfs_header_nritems(b) == 1);
2441         }
2442         return 0;
2443
2444 again:
2445         ret = -EAGAIN;
2446 done:
2447         return ret;
2448 }
2449
2450 /*
2451  * look for key in the tree.  path is filled in with nodes along the way
2452  * if key is found, we return zero and you can find the item in the leaf
2453  * level of the path (level 0)
2454  *
2455  * If the key isn't found, the path points to the slot where it should
2456  * be inserted, and 1 is returned.  If there are other errors during the
2457  * search a negative error number is returned.
2458  *
2459  * if ins_len > 0, nodes and leaves will be split as we walk down the
2460  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
2461  * possible)
2462  */
2463 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
2464                       *root, struct btrfs_key *key, struct btrfs_path *p, int
2465                       ins_len, int cow)
2466 {
2467         struct extent_buffer *b;
2468         int slot;
2469         int ret;
2470         int err;
2471         int level;
2472         int lowest_unlock = 1;
2473         int root_lock;
2474         /* everything at write_lock_level or lower must be write locked */
2475         int write_lock_level = 0;
2476         u8 lowest_level = 0;
2477         int min_write_lock_level;
2478
2479         lowest_level = p->lowest_level;
2480         WARN_ON(lowest_level && ins_len > 0);
2481         WARN_ON(p->nodes[0] != NULL);
2482
2483         if (ins_len < 0) {
2484                 lowest_unlock = 2;
2485
2486                 /* when we are removing items, we might have to go up to level
2487                  * two as we update tree pointers  Make sure we keep write
2488                  * for those levels as well
2489                  */
2490                 write_lock_level = 2;
2491         } else if (ins_len > 0) {
2492                 /*
2493                  * for inserting items, make sure we have a write lock on
2494                  * level 1 so we can update keys
2495                  */
2496                 write_lock_level = 1;
2497         }
2498
2499         if (!cow)
2500                 write_lock_level = -1;
2501
2502         if (cow && (p->keep_locks || p->lowest_level))
2503                 write_lock_level = BTRFS_MAX_LEVEL;
2504
2505         min_write_lock_level = write_lock_level;
2506
2507 again:
2508         /*
2509          * we try very hard to do read locks on the root
2510          */
2511         root_lock = BTRFS_READ_LOCK;
2512         level = 0;
2513         if (p->search_commit_root) {
2514                 /*
2515                  * the commit roots are read only
2516                  * so we always do read locks
2517                  */
2518                 b = root->commit_root;
2519                 extent_buffer_get(b);
2520                 level = btrfs_header_level(b);
2521                 if (!p->skip_locking)
2522                         btrfs_tree_read_lock(b);
2523         } else {
2524                 if (p->skip_locking) {
2525                         b = btrfs_root_node(root);
2526                         level = btrfs_header_level(b);
2527                 } else {
2528                         /* we don't know the level of the root node
2529                          * until we actually have it read locked
2530                          */
2531                         b = btrfs_read_lock_root_node(root);
2532                         level = btrfs_header_level(b);
2533                         if (level <= write_lock_level) {
2534                                 /* whoops, must trade for write lock */
2535                                 btrfs_tree_read_unlock(b);
2536                                 free_extent_buffer(b);
2537                                 b = btrfs_lock_root_node(root);
2538                                 root_lock = BTRFS_WRITE_LOCK;
2539
2540                                 /* the level might have changed, check again */
2541                                 level = btrfs_header_level(b);
2542                         }
2543                 }
2544         }
2545         p->nodes[level] = b;
2546         if (!p->skip_locking)
2547                 p->locks[level] = root_lock;
2548
2549         while (b) {
2550                 level = btrfs_header_level(b);
2551
2552                 /*
2553                  * setup the path here so we can release it under lock
2554                  * contention with the cow code
2555                  */
2556                 if (cow) {
2557                         /*
2558                          * if we don't really need to cow this block
2559                          * then we don't want to set the path blocking,
2560                          * so we test it here
2561                          */
2562                         if (!should_cow_block(trans, root, b))
2563                                 goto cow_done;
2564
2565                         btrfs_set_path_blocking(p);
2566
2567                         /*
2568                          * must have write locks on this node and the
2569                          * parent
2570                          */
2571                         if (level + 1 > write_lock_level) {
2572                                 write_lock_level = level + 1;
2573                                 btrfs_release_path(p);
2574                                 goto again;
2575                         }
2576
2577                         err = btrfs_cow_block(trans, root, b,
2578                                               p->nodes[level + 1],
2579                                               p->slots[level + 1], &b);
2580                         if (err) {
2581                                 ret = err;
2582                                 goto done;
2583                         }
2584                 }
2585 cow_done:
2586                 BUG_ON(!cow && ins_len);
2587
2588                 p->nodes[level] = b;
2589                 btrfs_clear_path_blocking(p, NULL, 0);
2590
2591                 /*
2592                  * we have a lock on b and as long as we aren't changing
2593                  * the tree, there is no way to for the items in b to change.
2594                  * It is safe to drop the lock on our parent before we
2595                  * go through the expensive btree search on b.
2596                  *
2597                  * If cow is true, then we might be changing slot zero,
2598                  * which may require changing the parent.  So, we can't
2599                  * drop the lock until after we know which slot we're
2600                  * operating on.
2601                  */
2602                 if (!cow)
2603                         btrfs_unlock_up_safe(p, level + 1);
2604
2605                 ret = bin_search(b, key, level, &slot);
2606
2607                 if (level != 0) {
2608                         int dec = 0;
2609                         if (ret && slot > 0) {
2610                                 dec = 1;
2611                                 slot -= 1;
2612                         }
2613                         p->slots[level] = slot;
2614                         err = setup_nodes_for_search(trans, root, p, b, level,
2615                                              ins_len, &write_lock_level);
2616                         if (err == -EAGAIN)
2617                                 goto again;
2618                         if (err) {
2619                                 ret = err;
2620                                 goto done;
2621                         }
2622                         b = p->nodes[level];
2623                         slot = p->slots[level];
2624
2625                         /*
2626                          * slot 0 is special, if we change the key
2627                          * we have to update the parent pointer
2628                          * which means we must have a write lock
2629                          * on the parent
2630                          */
2631                         if (slot == 0 && cow &&
2632                             write_lock_level < level + 1) {
2633                                 write_lock_level = level + 1;
2634                                 btrfs_release_path(p);
2635                                 goto again;
2636                         }
2637
2638                         unlock_up(p, level, lowest_unlock,
2639                                   min_write_lock_level, &write_lock_level);
2640
2641                         if (level == lowest_level) {
2642                                 if (dec)
2643                                         p->slots[level]++;
2644                                 goto done;
2645                         }
2646
2647                         err = read_block_for_search(trans, root, p,
2648                                                     &b, level, slot, key, 0);
2649                         if (err == -EAGAIN)
2650                                 goto again;
2651                         if (err) {
2652                                 ret = err;
2653                                 goto done;
2654                         }
2655
2656                         if (!p->skip_locking) {
2657                                 level = btrfs_header_level(b);
2658                                 if (level <= write_lock_level) {
2659                                         err = btrfs_try_tree_write_lock(b);
2660                                         if (!err) {
2661                                                 btrfs_set_path_blocking(p);
2662                                                 btrfs_tree_lock(b);
2663                                                 btrfs_clear_path_blocking(p, b,
2664                                                                   BTRFS_WRITE_LOCK);
2665                                         }
2666                                         p->locks[level] = BTRFS_WRITE_LOCK;
2667                                 } else {
2668                                         err = btrfs_try_tree_read_lock(b);
2669                                         if (!err) {
2670                                                 btrfs_set_path_blocking(p);
2671                                                 btrfs_tree_read_lock(b);
2672                                                 btrfs_clear_path_blocking(p, b,
2673                                                                   BTRFS_READ_LOCK);
2674                                         }
2675                                         p->locks[level] = BTRFS_READ_LOCK;
2676                                 }
2677                                 p->nodes[level] = b;
2678                         }
2679                 } else {
2680                         p->slots[level] = slot;
2681                         if (ins_len > 0 &&
2682                             btrfs_leaf_free_space(root, b) < ins_len) {
2683                                 if (write_lock_level < 1) {
2684                                         write_lock_level = 1;
2685                                         btrfs_release_path(p);
2686                                         goto again;
2687                                 }
2688
2689                                 btrfs_set_path_blocking(p);
2690                                 err = split_leaf(trans, root, key,
2691                                                  p, ins_len, ret == 0);
2692                                 btrfs_clear_path_blocking(p, NULL, 0);
2693
2694                                 BUG_ON(err > 0);
2695                                 if (err) {
2696                                         ret = err;
2697                                         goto done;
2698                                 }
2699                         }
2700                         if (!p->search_for_split)
2701                                 unlock_up(p, level, lowest_unlock,
2702                                           min_write_lock_level, &write_lock_level);
2703                         goto done;
2704                 }
2705         }
2706         ret = 1;
2707 done:
2708         /*
2709          * we don't really know what they plan on doing with the path
2710          * from here on, so for now just mark it as blocking
2711          */
2712         if (!p->leave_spinning)
2713                 btrfs_set_path_blocking(p);
2714         if (ret < 0)
2715                 btrfs_release_path(p);
2716         return ret;
2717 }
2718
2719 /*
2720  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2721  * current state of the tree together with the operations recorded in the tree
2722  * modification log to search for the key in a previous version of this tree, as
2723  * denoted by the time_seq parameter.
2724  *
2725  * Naturally, there is no support for insert, delete or cow operations.
2726  *
2727  * The resulting path and return value will be set up as if we called
2728  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2729  */
2730 int btrfs_search_old_slot(struct btrfs_root *root, struct btrfs_key *key,
2731                           struct btrfs_path *p, u64 time_seq)
2732 {
2733         struct extent_buffer *b;
2734         int slot;
2735         int ret;
2736         int err;
2737         int level;
2738         int lowest_unlock = 1;
2739         u8 lowest_level = 0;
2740
2741         lowest_level = p->lowest_level;
2742         WARN_ON(p->nodes[0] != NULL);
2743
2744         if (p->search_commit_root) {
2745                 BUG_ON(time_seq);
2746                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2747         }
2748
2749 again:
2750         b = get_old_root(root, time_seq);
2751         level = btrfs_header_level(b);
2752         p->locks[level] = BTRFS_READ_LOCK;
2753
2754         while (b) {
2755                 level = btrfs_header_level(b);
2756                 p->nodes[level] = b;
2757                 btrfs_clear_path_blocking(p, NULL, 0);
2758
2759                 /*
2760                  * we have a lock on b and as long as we aren't changing
2761                  * the tree, there is no way to for the items in b to change.
2762                  * It is safe to drop the lock on our parent before we
2763                  * go through the expensive btree search on b.
2764                  */
2765                 btrfs_unlock_up_safe(p, level + 1);
2766
2767                 ret = bin_search(b, key, level, &slot);
2768
2769                 if (level != 0) {
2770                         int dec = 0;
2771                         if (ret && slot > 0) {
2772                                 dec = 1;
2773                                 slot -= 1;
2774                         }
2775                         p->slots[level] = slot;
2776                         unlock_up(p, level, lowest_unlock, 0, NULL);
2777
2778                         if (level == lowest_level) {
2779                                 if (dec)
2780                                         p->slots[level]++;
2781                                 goto done;
2782                         }
2783
2784                         err = read_block_for_search(NULL, root, p, &b, level,
2785                                                     slot, key, time_seq);
2786                         if (err == -EAGAIN)
2787                                 goto again;
2788                         if (err) {
2789                                 ret = err;
2790                                 goto done;
2791                         }
2792
2793                         level = btrfs_header_level(b);
2794                         err = btrfs_try_tree_read_lock(b);
2795                         if (!err) {
2796                                 btrfs_set_path_blocking(p);
2797                                 btrfs_tree_read_lock(b);
2798                                 btrfs_clear_path_blocking(p, b,
2799                                                           BTRFS_READ_LOCK);
2800                         }
2801                         p->locks[level] = BTRFS_READ_LOCK;
2802                         p->nodes[level] = b;
2803                         b = tree_mod_log_rewind(root->fs_info, b, time_seq);
2804                         if (b != p->nodes[level]) {
2805                                 btrfs_tree_unlock_rw(p->nodes[level],
2806                                                      p->locks[level]);
2807                                 p->locks[level] = 0;
2808                                 p->nodes[level] = b;
2809                         }
2810                 } else {
2811                         p->slots[level] = slot;
2812                         unlock_up(p, level, lowest_unlock, 0, NULL);
2813                         goto done;
2814                 }
2815         }
2816         ret = 1;
2817 done:
2818         if (!p->leave_spinning)
2819                 btrfs_set_path_blocking(p);
2820         if (ret < 0)
2821                 btrfs_release_path(p);
2822
2823         return ret;
2824 }
2825
2826 /*
2827  * helper to use instead of search slot if no exact match is needed but
2828  * instead the next or previous item should be returned.
2829  * When find_higher is true, the next higher item is returned, the next lower
2830  * otherwise.
2831  * When return_any and find_higher are both true, and no higher item is found,
2832  * return the next lower instead.
2833  * When return_any is true and find_higher is false, and no lower item is found,
2834  * return the next higher instead.
2835  * It returns 0 if any item is found, 1 if none is found (tree empty), and
2836  * < 0 on error
2837  */
2838 int btrfs_search_slot_for_read(struct btrfs_root *root,
2839                                struct btrfs_key *key, struct btrfs_path *p,
2840                                int find_higher, int return_any)
2841 {
2842         int ret;
2843         struct extent_buffer *leaf;
2844
2845 again:
2846         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
2847         if (ret <= 0)
2848                 return ret;
2849         /*
2850          * a return value of 1 means the path is at the position where the
2851          * item should be inserted. Normally this is the next bigger item,
2852          * but in case the previous item is the last in a leaf, path points
2853          * to the first free slot in the previous leaf, i.e. at an invalid
2854          * item.
2855          */
2856         leaf = p->nodes[0];
2857
2858         if (find_higher) {
2859                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
2860                         ret = btrfs_next_leaf(root, p);
2861                         if (ret <= 0)
2862                                 return ret;
2863                         if (!return_any)
2864                                 return 1;
2865                         /*
2866                          * no higher item found, return the next
2867                          * lower instead
2868                          */
2869                         return_any = 0;
2870                         find_higher = 0;
2871                         btrfs_release_path(p);
2872                         goto again;
2873                 }
2874         } else {
2875                 if (p->slots[0] == 0) {
2876                         ret = btrfs_prev_leaf(root, p);
2877                         if (ret < 0)
2878                                 return ret;
2879                         if (!ret) {
2880                                 p->slots[0] = btrfs_header_nritems(leaf) - 1;
2881                                 return 0;
2882                         }
2883                         if (!return_any)
2884                                 return 1;
2885                         /*
2886                          * no lower item found, return the next
2887                          * higher instead
2888                          */
2889                         return_any = 0;
2890                         find_higher = 1;
2891                         btrfs_release_path(p);
2892                         goto again;
2893                 } else {
2894                         --p->slots[0];
2895                 }
2896         }
2897         return 0;
2898 }
2899
2900 /*
2901  * adjust the pointers going up the tree, starting at level
2902  * making sure the right key of each node is points to 'key'.
2903  * This is used after shifting pointers to the left, so it stops
2904  * fixing up pointers when a given leaf/node is not in slot 0 of the
2905  * higher levels
2906  *
2907  */
2908 static void fixup_low_keys(struct btrfs_trans_handle *trans,
2909                            struct btrfs_root *root, struct btrfs_path *path,
2910                            struct btrfs_disk_key *key, int level)
2911 {
2912         int i;
2913         struct extent_buffer *t;
2914
2915         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2916                 int tslot = path->slots[i];
2917                 if (!path->nodes[i])
2918                         break;
2919                 t = path->nodes[i];
2920                 tree_mod_log_set_node_key(root->fs_info, t, key, tslot, 1);
2921                 btrfs_set_node_key(t, key, tslot);
2922                 btrfs_mark_buffer_dirty(path->nodes[i]);
2923                 if (tslot != 0)
2924                         break;
2925         }
2926 }
2927
2928 /*
2929  * update item key.
2930  *
2931  * This function isn't completely safe. It's the caller's responsibility
2932  * that the new key won't break the order
2933  */
2934 void btrfs_set_item_key_safe(struct btrfs_trans_handle *trans,
2935                              struct btrfs_root *root, struct btrfs_path *path,
2936                              struct btrfs_key *new_key)
2937 {
2938         struct btrfs_disk_key disk_key;
2939         struct extent_buffer *eb;
2940         int slot;
2941
2942         eb = path->nodes[0];
2943         slot = path->slots[0];
2944         if (slot > 0) {
2945                 btrfs_item_key(eb, &disk_key, slot - 1);
2946                 BUG_ON(comp_keys(&disk_key, new_key) >= 0);
2947         }
2948         if (slot < btrfs_header_nritems(eb) - 1) {
2949                 btrfs_item_key(eb, &disk_key, slot + 1);
2950                 BUG_ON(comp_keys(&disk_key, new_key) <= 0);
2951         }
2952
2953         btrfs_cpu_key_to_disk(&disk_key, new_key);
2954         btrfs_set_item_key(eb, &disk_key, slot);
2955         btrfs_mark_buffer_dirty(eb);
2956         if (slot == 0)
2957                 fixup_low_keys(trans, root, path, &disk_key, 1);
2958 }
2959
2960 /*
2961  * try to push data from one node into the next node left in the
2962  * tree.
2963  *
2964  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
2965  * error, and > 0 if there was no room in the left hand block.
2966  */
2967 static int push_node_left(struct btrfs_trans_handle *trans,
2968                           struct btrfs_root *root, struct extent_buffer *dst,
2969                           struct extent_buffer *src, int empty)
2970 {
2971         int push_items = 0;
2972         int src_nritems;
2973         int dst_nritems;
2974         int ret = 0;
2975
2976         src_nritems = btrfs_header_nritems(src);
2977         dst_nritems = btrfs_header_nritems(dst);
2978         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
2979         WARN_ON(btrfs_header_generation(src) != trans->transid);
2980         WARN_ON(btrfs_header_generation(dst) != trans->transid);
2981
2982         if (!empty && src_nritems <= 8)
2983                 return 1;
2984
2985         if (push_items <= 0)
2986                 return 1;
2987
2988         if (empty) {
2989                 push_items = min(src_nritems, push_items);
2990                 if (push_items < src_nritems) {
2991                         /* leave at least 8 pointers in the node if
2992                          * we aren't going to empty it
2993                          */
2994                         if (src_nritems - push_items < 8) {
2995                                 if (push_items <= 8)
2996                                         return 1;
2997                                 push_items -= 8;
2998                         }
2999                 }
3000         } else
3001                 push_items = min(src_nritems - 8, push_items);
3002
3003         tree_mod_log_eb_copy(root->fs_info, dst, src, dst_nritems, 0,
3004                              push_items);
3005         copy_extent_buffer(dst, src,
3006                            btrfs_node_key_ptr_offset(dst_nritems),
3007                            btrfs_node_key_ptr_offset(0),
3008                            push_items * sizeof(struct btrfs_key_ptr));
3009
3010         if (push_items < src_nritems) {
3011                 /*
3012                  * don't call tree_mod_log_eb_move here, key removal was already
3013                  * fully logged by tree_mod_log_eb_copy above.
3014                  */
3015                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
3016                                       btrfs_node_key_ptr_offset(push_items),
3017                                       (src_nritems - push_items) *
3018                                       sizeof(struct btrfs_key_ptr));
3019         }
3020         btrfs_set_header_nritems(src, src_nritems - push_items);
3021         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3022         btrfs_mark_buffer_dirty(src);
3023         btrfs_mark_buffer_dirty(dst);
3024
3025         return ret;
3026 }
3027
3028 /*
3029  * try to push data from one node into the next node right in the
3030  * tree.
3031  *
3032  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
3033  * error, and > 0 if there was no room in the right hand block.
3034  *
3035  * this will  only push up to 1/2 the contents of the left node over
3036  */
3037 static int balance_node_right(struct btrfs_trans_handle *trans,
3038                               struct btrfs_root *root,
3039                               struct extent_buffer *dst,
3040                               struct extent_buffer *src)
3041 {
3042         int push_items = 0;
3043         int max_push;
3044         int src_nritems;
3045         int dst_nritems;
3046         int ret = 0;
3047
3048         WARN_ON(btrfs_header_generation(src) != trans->transid);
3049         WARN_ON(btrfs_header_generation(dst) != trans->transid);
3050
3051         src_nritems = btrfs_header_nritems(src);
3052         dst_nritems = btrfs_header_nritems(dst);
3053         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
3054         if (push_items <= 0)
3055                 return 1;
3056
3057         if (src_nritems < 4)
3058                 return 1;
3059
3060         max_push = src_nritems / 2 + 1;
3061         /* don't try to empty the node */
3062         if (max_push >= src_nritems)
3063                 return 1;
3064
3065         if (max_push < push_items)
3066                 push_items = max_push;
3067
3068         tree_mod_log_eb_move(root->fs_info, dst, push_items, 0, dst_nritems);
3069         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
3070                                       btrfs_node_key_ptr_offset(0),
3071                                       (dst_nritems) *
3072                                       sizeof(struct btrfs_key_ptr));
3073
3074         tree_mod_log_eb_copy(root->fs_info, dst, src, 0,
3075                              src_nritems - push_items, push_items);
3076         copy_extent_buffer(dst, src,
3077                            btrfs_node_key_ptr_offset(0),
3078                            btrfs_node_key_ptr_offset(src_nritems - push_items),
3079                            push_items * sizeof(struct btrfs_key_ptr));
3080
3081         btrfs_set_header_nritems(src, src_nritems - push_items);
3082         btrfs_set_header_nritems(dst, dst_nritems + push_items);
3083
3084         btrfs_mark_buffer_dirty(src);
3085         btrfs_mark_buffer_dirty(dst);
3086
3087         return ret;
3088 }
3089
3090 /*
3091  * helper function to insert a new root level in the tree.
3092  * A new node is allocated, and a single item is inserted to
3093  * point to the existing root
3094  *
3095  * returns zero on success or < 0 on failure.
3096  */
3097 static noinline int insert_new_root(struct btrfs_trans_handle *trans,
3098                            struct btrfs_root *root,
3099                            struct btrfs_path *path, int level)
3100 {
3101         u64 lower_gen;
3102         struct extent_buffer *lower;
3103         struct extent_buffer *c;
3104         struct extent_buffer *old;
3105         struct btrfs_disk_key lower_key;
3106
3107         BUG_ON(path->nodes[level]);
3108         BUG_ON(path->nodes[level-1] != root->node);
3109
3110         lower = path->nodes[level-1];
3111         if (level == 1)
3112                 btrfs_item_key(lower, &lower_key, 0);
3113         else
3114                 btrfs_node_key(lower, &lower_key, 0);
3115
3116         c = btrfs_alloc_free_block(trans, root, root->nodesize, 0,
3117                                    root->root_key.objectid, &lower_key,
3118                                    level, root->node->start, 0);
3119         if (IS_ERR(c))
3120                 return PTR_ERR(c);
3121
3122         root_add_used(root, root->nodesize);
3123
3124         memset_extent_buffer(c, 0, 0, sizeof(struct btrfs_header));
3125         btrfs_set_header_nritems(c, 1);
3126         btrfs_set_header_level(c, level);
3127         btrfs_set_header_bytenr(c, c->start);
3128         btrfs_set_header_generation(c, trans->transid);
3129         btrfs_set_header_backref_rev(c, BTRFS_MIXED_BACKREF_REV);
3130         btrfs_set_header_owner(c, root->root_key.objectid);
3131
3132         write_extent_buffer(c, root->fs_info->fsid,
3133                             (unsigned long)btrfs_header_fsid(c),
3134                             BTRFS_FSID_SIZE);
3135
3136         write_extent_buffer(c, root->fs_info->chunk_tree_uuid,
3137                             (unsigned long)btrfs_header_chunk_tree_uuid(c),
3138                             BTRFS_UUID_SIZE);
3139
3140         btrfs_set_node_key(c, &lower_key, 0);
3141         btrfs_set_node_blockptr(c, 0, lower->start);
3142         lower_gen = btrfs_header_generation(lower);
3143         WARN_ON(lower_gen != trans->transid);
3144
3145         btrfs_set_node_ptr_generation(c, 0, lower_gen);
3146
3147         btrfs_mark_buffer_dirty(c);
3148
3149         old = root->node;
3150         tree_mod_log_set_root_pointer(root, c);
3151         rcu_assign_pointer(root->node, c);
3152
3153         /* the super has an extra ref to root->node */
3154         free_extent_buffer(old);
3155
3156         add_root_to_dirty_list(root);
3157         extent_buffer_get(c);
3158         path->nodes[level] = c;
3159         path->locks[level] = BTRFS_WRITE_LOCK;
3160         path->slots[level] = 0;
3161         return 0;
3162 }
3163
3164 /*
3165  * worker function to insert a single pointer in a node.
3166  * the node should have enough room for the pointer already
3167  *
3168  * slot and level indicate where you want the key to go, and
3169  * blocknr is the block the key points to.
3170  */
3171 static void insert_ptr(struct btrfs_trans_handle *trans,
3172                        struct btrfs_root *root, struct btrfs_path *path,
3173                        struct btrfs_disk_key *key, u64 bytenr,
3174                        int slot, int level)
3175 {
3176         struct extent_buffer *lower;
3177         int nritems;
3178         int ret;
3179
3180         BUG_ON(!path->nodes[level]);
3181         btrfs_assert_tree_locked(path->nodes[level]);
3182         lower = path->nodes[level];
3183         nritems = btrfs_header_nritems(lower);
3184         BUG_ON(slot > nritems);
3185         BUG_ON(nritems == BTRFS_NODEPTRS_PER_BLOCK(root));
3186         if (slot != nritems) {
3187                 if (level)
3188                         tree_mod_log_eb_move(root->fs_info, lower, slot + 1,
3189                                              slot, nritems - slot);
3190                 memmove_extent_buffer(lower,
3191                               btrfs_node_key_ptr_offset(slot + 1),
3192                               btrfs_node_key_ptr_offset(slot),
3193                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
3194         }
3195         if (level) {
3196                 ret = tree_mod_log_insert_key(root->fs_info, lower, slot,
3197                                               MOD_LOG_KEY_ADD);
3198                 BUG_ON(ret < 0);
3199         }
3200         btrfs_set_node_key(lower, key, slot);
3201         btrfs_set_node_blockptr(lower, slot, bytenr);
3202         WARN_ON(trans->transid == 0);
3203         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
3204         btrfs_set_header_nritems(lower, nritems + 1);
3205         btrfs_mark_buffer_dirty(lower);
3206 }
3207
3208 /*
3209  * split the node at the specified level in path in two.
3210  * The path is corrected to point to the appropriate node after the split
3211  *
3212  * Before splitting this tries to make some room in the node by pushing
3213  * left and right, if either one works, it returns right away.
3214  *
3215  * returns 0 on success and < 0 on failure
3216  */
3217 static noinline int split_node(struct btrfs_trans_handle *trans,
3218                                struct btrfs_root *root,
3219                                struct btrfs_path *path, int level)
3220 {
3221         struct extent_buffer *c;
3222         struct extent_buffer *split;
3223         struct btrfs_disk_key disk_key;
3224         int mid;
3225         int ret;
3226         u32 c_nritems;
3227
3228         c = path->nodes[level];
3229         WARN_ON(btrfs_header_generation(c) != trans->transid);
3230         if (c == root->node) {
3231                 /* trying to split the root, lets make a new one */