btrfs: Remove redundant comment from btrfs_search_forward
[sfrench/cifs-2.6.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 <linux/mm.h>
23 #include "ctree.h"
24 #include "disk-io.h"
25 #include "transaction.h"
26 #include "print-tree.h"
27 #include "locking.h"
28
29 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
30                       *root, struct btrfs_path *path, int level);
31 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
32                       const struct btrfs_key *ins_key, struct btrfs_path *path,
33                       int data_size, int extend);
34 static int push_node_left(struct btrfs_trans_handle *trans,
35                           struct btrfs_fs_info *fs_info,
36                           struct extent_buffer *dst,
37                           struct extent_buffer *src, int empty);
38 static int balance_node_right(struct btrfs_trans_handle *trans,
39                               struct btrfs_fs_info *fs_info,
40                               struct extent_buffer *dst_buf,
41                               struct extent_buffer *src_buf);
42 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
43                     int level, int slot);
44 static int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
45                                  struct extent_buffer *eb);
46
47 struct btrfs_path *btrfs_alloc_path(void)
48 {
49         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
50 }
51
52 /*
53  * set all locked nodes in the path to blocking locks.  This should
54  * be done before scheduling
55  */
56 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
57 {
58         int i;
59         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
60                 if (!p->nodes[i] || !p->locks[i])
61                         continue;
62                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
63                 if (p->locks[i] == BTRFS_READ_LOCK)
64                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
65                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
66                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
67         }
68 }
69
70 /*
71  * reset all the locked nodes in the patch to spinning locks.
72  *
73  * held is used to keep lockdep happy, when lockdep is enabled
74  * we set held to a blocking lock before we go around and
75  * retake all the spinlocks in the path.  You can safely use NULL
76  * for held
77  */
78 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
79                                         struct extent_buffer *held, int held_rw)
80 {
81         int i;
82
83         if (held) {
84                 btrfs_set_lock_blocking_rw(held, held_rw);
85                 if (held_rw == BTRFS_WRITE_LOCK)
86                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
87                 else if (held_rw == BTRFS_READ_LOCK)
88                         held_rw = BTRFS_READ_LOCK_BLOCKING;
89         }
90         btrfs_set_path_blocking(p);
91
92         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
93                 if (p->nodes[i] && p->locks[i]) {
94                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
95                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
96                                 p->locks[i] = BTRFS_WRITE_LOCK;
97                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
98                                 p->locks[i] = BTRFS_READ_LOCK;
99                 }
100         }
101
102         if (held)
103                 btrfs_clear_lock_blocking_rw(held, held_rw);
104 }
105
106 /* this also releases the path */
107 void btrfs_free_path(struct btrfs_path *p)
108 {
109         if (!p)
110                 return;
111         btrfs_release_path(p);
112         kmem_cache_free(btrfs_path_cachep, p);
113 }
114
115 /*
116  * path release drops references on the extent buffers in the path
117  * and it drops any locks held by this path
118  *
119  * It is safe to call this on paths that no locks or extent buffers held.
120  */
121 noinline void btrfs_release_path(struct btrfs_path *p)
122 {
123         int i;
124
125         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
126                 p->slots[i] = 0;
127                 if (!p->nodes[i])
128                         continue;
129                 if (p->locks[i]) {
130                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
131                         p->locks[i] = 0;
132                 }
133                 free_extent_buffer(p->nodes[i]);
134                 p->nodes[i] = NULL;
135         }
136 }
137
138 /*
139  * safely gets a reference on the root node of a tree.  A lock
140  * is not taken, so a concurrent writer may put a different node
141  * at the root of the tree.  See btrfs_lock_root_node for the
142  * looping required.
143  *
144  * The extent buffer returned by this has a reference taken, so
145  * it won't disappear.  It may stop being the root of the tree
146  * at any time because there are no locks held.
147  */
148 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
149 {
150         struct extent_buffer *eb;
151
152         while (1) {
153                 rcu_read_lock();
154                 eb = rcu_dereference(root->node);
155
156                 /*
157                  * RCU really hurts here, we could free up the root node because
158                  * it was COWed but we may not get the new root node yet so do
159                  * the inc_not_zero dance and if it doesn't work then
160                  * synchronize_rcu and try again.
161                  */
162                 if (atomic_inc_not_zero(&eb->refs)) {
163                         rcu_read_unlock();
164                         break;
165                 }
166                 rcu_read_unlock();
167                 synchronize_rcu();
168         }
169         return eb;
170 }
171
172 /* loop around taking references on and locking the root node of the
173  * tree until you end up with a lock on the root.  A locked buffer
174  * is returned, with a reference held.
175  */
176 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
177 {
178         struct extent_buffer *eb;
179
180         while (1) {
181                 eb = btrfs_root_node(root);
182                 btrfs_tree_lock(eb);
183                 if (eb == root->node)
184                         break;
185                 btrfs_tree_unlock(eb);
186                 free_extent_buffer(eb);
187         }
188         return eb;
189 }
190
191 /* loop around taking references on and locking the root node of the
192  * tree until you end up with a lock on the root.  A locked buffer
193  * is returned, with a reference held.
194  */
195 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
196 {
197         struct extent_buffer *eb;
198
199         while (1) {
200                 eb = btrfs_root_node(root);
201                 btrfs_tree_read_lock(eb);
202                 if (eb == root->node)
203                         break;
204                 btrfs_tree_read_unlock(eb);
205                 free_extent_buffer(eb);
206         }
207         return eb;
208 }
209
210 /* cowonly root (everything not a reference counted cow subvolume), just get
211  * put onto a simple dirty list.  transaction.c walks this to make sure they
212  * get properly updated on disk.
213  */
214 static void add_root_to_dirty_list(struct btrfs_root *root)
215 {
216         struct btrfs_fs_info *fs_info = root->fs_info;
217
218         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
219             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
220                 return;
221
222         spin_lock(&fs_info->trans_lock);
223         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
224                 /* Want the extent tree to be the last on the list */
225                 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
226                         list_move_tail(&root->dirty_list,
227                                        &fs_info->dirty_cowonly_roots);
228                 else
229                         list_move(&root->dirty_list,
230                                   &fs_info->dirty_cowonly_roots);
231         }
232         spin_unlock(&fs_info->trans_lock);
233 }
234
235 /*
236  * used by snapshot creation to make a copy of a root for a tree with
237  * a given objectid.  The buffer with the new root node is returned in
238  * cow_ret, and this func returns zero on success or a negative error code.
239  */
240 int btrfs_copy_root(struct btrfs_trans_handle *trans,
241                       struct btrfs_root *root,
242                       struct extent_buffer *buf,
243                       struct extent_buffer **cow_ret, u64 new_root_objectid)
244 {
245         struct btrfs_fs_info *fs_info = root->fs_info;
246         struct extent_buffer *cow;
247         int ret = 0;
248         int level;
249         struct btrfs_disk_key disk_key;
250
251         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
252                 trans->transid != fs_info->running_transaction->transid);
253         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
254                 trans->transid != root->last_trans);
255
256         level = btrfs_header_level(buf);
257         if (level == 0)
258                 btrfs_item_key(buf, &disk_key, 0);
259         else
260                 btrfs_node_key(buf, &disk_key, 0);
261
262         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
263                         &disk_key, level, buf->start, 0);
264         if (IS_ERR(cow))
265                 return PTR_ERR(cow);
266
267         copy_extent_buffer_full(cow, buf);
268         btrfs_set_header_bytenr(cow, cow->start);
269         btrfs_set_header_generation(cow, trans->transid);
270         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
271         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
272                                      BTRFS_HEADER_FLAG_RELOC);
273         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
274                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
275         else
276                 btrfs_set_header_owner(cow, new_root_objectid);
277
278         write_extent_buffer_fsid(cow, fs_info->fsid);
279
280         WARN_ON(btrfs_header_generation(buf) > trans->transid);
281         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
282                 ret = btrfs_inc_ref(trans, root, cow, 1);
283         else
284                 ret = btrfs_inc_ref(trans, root, cow, 0);
285
286         if (ret)
287                 return ret;
288
289         btrfs_mark_buffer_dirty(cow);
290         *cow_ret = cow;
291         return 0;
292 }
293
294 enum mod_log_op {
295         MOD_LOG_KEY_REPLACE,
296         MOD_LOG_KEY_ADD,
297         MOD_LOG_KEY_REMOVE,
298         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
299         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
300         MOD_LOG_MOVE_KEYS,
301         MOD_LOG_ROOT_REPLACE,
302 };
303
304 struct tree_mod_move {
305         int dst_slot;
306         int nr_items;
307 };
308
309 struct tree_mod_root {
310         u64 logical;
311         u8 level;
312 };
313
314 struct tree_mod_elem {
315         struct rb_node node;
316         u64 logical;
317         u64 seq;
318         enum mod_log_op op;
319
320         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
321         int slot;
322
323         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
324         u64 generation;
325
326         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
327         struct btrfs_disk_key key;
328         u64 blockptr;
329
330         /* this is used for op == MOD_LOG_MOVE_KEYS */
331         struct tree_mod_move move;
332
333         /* this is used for op == MOD_LOG_ROOT_REPLACE */
334         struct tree_mod_root old_root;
335 };
336
337 static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
338 {
339         read_lock(&fs_info->tree_mod_log_lock);
340 }
341
342 static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
343 {
344         read_unlock(&fs_info->tree_mod_log_lock);
345 }
346
347 static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
348 {
349         write_lock(&fs_info->tree_mod_log_lock);
350 }
351
352 static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
353 {
354         write_unlock(&fs_info->tree_mod_log_lock);
355 }
356
357 /*
358  * Pull a new tree mod seq number for our operation.
359  */
360 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
361 {
362         return atomic64_inc_return(&fs_info->tree_mod_seq);
363 }
364
365 /*
366  * This adds a new blocker to the tree mod log's blocker list if the @elem
367  * passed does not already have a sequence number set. So when a caller expects
368  * to record tree modifications, it should ensure to set elem->seq to zero
369  * before calling btrfs_get_tree_mod_seq.
370  * Returns a fresh, unused tree log modification sequence number, even if no new
371  * blocker was added.
372  */
373 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
374                            struct seq_list *elem)
375 {
376         tree_mod_log_write_lock(fs_info);
377         spin_lock(&fs_info->tree_mod_seq_lock);
378         if (!elem->seq) {
379                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
380                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
381         }
382         spin_unlock(&fs_info->tree_mod_seq_lock);
383         tree_mod_log_write_unlock(fs_info);
384
385         return elem->seq;
386 }
387
388 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
389                             struct seq_list *elem)
390 {
391         struct rb_root *tm_root;
392         struct rb_node *node;
393         struct rb_node *next;
394         struct seq_list *cur_elem;
395         struct tree_mod_elem *tm;
396         u64 min_seq = (u64)-1;
397         u64 seq_putting = elem->seq;
398
399         if (!seq_putting)
400                 return;
401
402         spin_lock(&fs_info->tree_mod_seq_lock);
403         list_del(&elem->list);
404         elem->seq = 0;
405
406         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
407                 if (cur_elem->seq < min_seq) {
408                         if (seq_putting > cur_elem->seq) {
409                                 /*
410                                  * blocker with lower sequence number exists, we
411                                  * cannot remove anything from the log
412                                  */
413                                 spin_unlock(&fs_info->tree_mod_seq_lock);
414                                 return;
415                         }
416                         min_seq = cur_elem->seq;
417                 }
418         }
419         spin_unlock(&fs_info->tree_mod_seq_lock);
420
421         /*
422          * anything that's lower than the lowest existing (read: blocked)
423          * sequence number can be removed from the tree.
424          */
425         tree_mod_log_write_lock(fs_info);
426         tm_root = &fs_info->tree_mod_log;
427         for (node = rb_first(tm_root); node; node = next) {
428                 next = rb_next(node);
429                 tm = rb_entry(node, struct tree_mod_elem, node);
430                 if (tm->seq > min_seq)
431                         continue;
432                 rb_erase(node, tm_root);
433                 kfree(tm);
434         }
435         tree_mod_log_write_unlock(fs_info);
436 }
437
438 /*
439  * key order of the log:
440  *       node/leaf start address -> sequence
441  *
442  * The 'start address' is the logical address of the *new* root node
443  * for root replace operations, or the logical address of the affected
444  * block for all other operations.
445  *
446  * Note: must be called with write lock (tree_mod_log_write_lock).
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         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
457
458         tm_root = &fs_info->tree_mod_log;
459         new = &tm_root->rb_node;
460         while (*new) {
461                 cur = rb_entry(*new, struct tree_mod_elem, node);
462                 parent = *new;
463                 if (cur->logical < tm->logical)
464                         new = &((*new)->rb_left);
465                 else if (cur->logical > tm->logical)
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                         return -EEXIST;
473         }
474
475         rb_link_node(&tm->node, parent, new);
476         rb_insert_color(&tm->node, tm_root);
477         return 0;
478 }
479
480 /*
481  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
482  * returns zero with the tree_mod_log_lock acquired. The caller must hold
483  * this until all tree mod log insertions are recorded in the rb tree and then
484  * call tree_mod_log_write_unlock() to release.
485  */
486 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
487                                     struct extent_buffer *eb) {
488         smp_mb();
489         if (list_empty(&(fs_info)->tree_mod_seq_list))
490                 return 1;
491         if (eb && btrfs_header_level(eb) == 0)
492                 return 1;
493
494         tree_mod_log_write_lock(fs_info);
495         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
496                 tree_mod_log_write_unlock(fs_info);
497                 return 1;
498         }
499
500         return 0;
501 }
502
503 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
504 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
505                                     struct extent_buffer *eb)
506 {
507         smp_mb();
508         if (list_empty(&(fs_info)->tree_mod_seq_list))
509                 return 0;
510         if (eb && btrfs_header_level(eb) == 0)
511                 return 0;
512
513         return 1;
514 }
515
516 static struct tree_mod_elem *
517 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
518                     enum mod_log_op op, gfp_t flags)
519 {
520         struct tree_mod_elem *tm;
521
522         tm = kzalloc(sizeof(*tm), flags);
523         if (!tm)
524                 return NULL;
525
526         tm->logical = eb->start;
527         if (op != MOD_LOG_KEY_ADD) {
528                 btrfs_node_key(eb, &tm->key, slot);
529                 tm->blockptr = btrfs_node_blockptr(eb, slot);
530         }
531         tm->op = op;
532         tm->slot = slot;
533         tm->generation = btrfs_node_ptr_generation(eb, slot);
534         RB_CLEAR_NODE(&tm->node);
535
536         return tm;
537 }
538
539 static noinline int
540 tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
541                         struct extent_buffer *eb, int slot,
542                         enum mod_log_op op, gfp_t flags)
543 {
544         struct tree_mod_elem *tm;
545         int ret;
546
547         if (!tree_mod_need_log(fs_info, eb))
548                 return 0;
549
550         tm = alloc_tree_mod_elem(eb, slot, op, flags);
551         if (!tm)
552                 return -ENOMEM;
553
554         if (tree_mod_dont_log(fs_info, eb)) {
555                 kfree(tm);
556                 return 0;
557         }
558
559         ret = __tree_mod_log_insert(fs_info, tm);
560         tree_mod_log_write_unlock(fs_info);
561         if (ret)
562                 kfree(tm);
563
564         return ret;
565 }
566
567 static noinline int
568 tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
569                          struct extent_buffer *eb, int dst_slot, int src_slot,
570                          int nr_items)
571 {
572         struct tree_mod_elem *tm = NULL;
573         struct tree_mod_elem **tm_list = NULL;
574         int ret = 0;
575         int i;
576         int locked = 0;
577
578         if (!tree_mod_need_log(fs_info, eb))
579                 return 0;
580
581         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
582         if (!tm_list)
583                 return -ENOMEM;
584
585         tm = kzalloc(sizeof(*tm), GFP_NOFS);
586         if (!tm) {
587                 ret = -ENOMEM;
588                 goto free_tms;
589         }
590
591         tm->logical = eb->start;
592         tm->slot = src_slot;
593         tm->move.dst_slot = dst_slot;
594         tm->move.nr_items = nr_items;
595         tm->op = MOD_LOG_MOVE_KEYS;
596
597         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
598                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
599                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
600                 if (!tm_list[i]) {
601                         ret = -ENOMEM;
602                         goto free_tms;
603                 }
604         }
605
606         if (tree_mod_dont_log(fs_info, eb))
607                 goto free_tms;
608         locked = 1;
609
610         /*
611          * When we override something during the move, we log these removals.
612          * This can only happen when we move towards the beginning of the
613          * buffer, i.e. dst_slot < src_slot.
614          */
615         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
616                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
617                 if (ret)
618                         goto free_tms;
619         }
620
621         ret = __tree_mod_log_insert(fs_info, tm);
622         if (ret)
623                 goto free_tms;
624         tree_mod_log_write_unlock(fs_info);
625         kfree(tm_list);
626
627         return 0;
628 free_tms:
629         for (i = 0; i < nr_items; i++) {
630                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
631                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
632                 kfree(tm_list[i]);
633         }
634         if (locked)
635                 tree_mod_log_write_unlock(fs_info);
636         kfree(tm_list);
637         kfree(tm);
638
639         return ret;
640 }
641
642 static inline int
643 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
644                        struct tree_mod_elem **tm_list,
645                        int nritems)
646 {
647         int i, j;
648         int ret;
649
650         for (i = nritems - 1; i >= 0; i--) {
651                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
652                 if (ret) {
653                         for (j = nritems - 1; j > i; j--)
654                                 rb_erase(&tm_list[j]->node,
655                                          &fs_info->tree_mod_log);
656                         return ret;
657                 }
658         }
659
660         return 0;
661 }
662
663 static noinline int
664 tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
665                          struct extent_buffer *old_root,
666                          struct extent_buffer *new_root,
667                          int log_removal)
668 {
669         struct tree_mod_elem *tm = NULL;
670         struct tree_mod_elem **tm_list = NULL;
671         int nritems = 0;
672         int ret = 0;
673         int i;
674
675         if (!tree_mod_need_log(fs_info, NULL))
676                 return 0;
677
678         if (log_removal && btrfs_header_level(old_root) > 0) {
679                 nritems = btrfs_header_nritems(old_root);
680                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
681                                   GFP_NOFS);
682                 if (!tm_list) {
683                         ret = -ENOMEM;
684                         goto free_tms;
685                 }
686                 for (i = 0; i < nritems; i++) {
687                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
688                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
689                         if (!tm_list[i]) {
690                                 ret = -ENOMEM;
691                                 goto free_tms;
692                         }
693                 }
694         }
695
696         tm = kzalloc(sizeof(*tm), GFP_NOFS);
697         if (!tm) {
698                 ret = -ENOMEM;
699                 goto free_tms;
700         }
701
702         tm->logical = new_root->start;
703         tm->old_root.logical = old_root->start;
704         tm->old_root.level = btrfs_header_level(old_root);
705         tm->generation = btrfs_header_generation(old_root);
706         tm->op = MOD_LOG_ROOT_REPLACE;
707
708         if (tree_mod_dont_log(fs_info, NULL))
709                 goto free_tms;
710
711         if (tm_list)
712                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
713         if (!ret)
714                 ret = __tree_mod_log_insert(fs_info, tm);
715
716         tree_mod_log_write_unlock(fs_info);
717         if (ret)
718                 goto free_tms;
719         kfree(tm_list);
720
721         return ret;
722
723 free_tms:
724         if (tm_list) {
725                 for (i = 0; i < nritems; i++)
726                         kfree(tm_list[i]);
727                 kfree(tm_list);
728         }
729         kfree(tm);
730
731         return ret;
732 }
733
734 static struct tree_mod_elem *
735 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
736                       int smallest)
737 {
738         struct rb_root *tm_root;
739         struct rb_node *node;
740         struct tree_mod_elem *cur = NULL;
741         struct tree_mod_elem *found = NULL;
742
743         tree_mod_log_read_lock(fs_info);
744         tm_root = &fs_info->tree_mod_log;
745         node = tm_root->rb_node;
746         while (node) {
747                 cur = rb_entry(node, struct tree_mod_elem, node);
748                 if (cur->logical < start) {
749                         node = node->rb_left;
750                 } else if (cur->logical > start) {
751                         node = node->rb_right;
752                 } else if (cur->seq < min_seq) {
753                         node = node->rb_left;
754                 } else if (!smallest) {
755                         /* we want the node with the highest seq */
756                         if (found)
757                                 BUG_ON(found->seq > cur->seq);
758                         found = cur;
759                         node = node->rb_left;
760                 } else if (cur->seq > min_seq) {
761                         /* we want the node with the smallest seq */
762                         if (found)
763                                 BUG_ON(found->seq < cur->seq);
764                         found = cur;
765                         node = node->rb_right;
766                 } else {
767                         found = cur;
768                         break;
769                 }
770         }
771         tree_mod_log_read_unlock(fs_info);
772
773         return found;
774 }
775
776 /*
777  * this returns the element from the log with the smallest time sequence
778  * value that's in the log (the oldest log item). any element with a time
779  * sequence lower than min_seq will be ignored.
780  */
781 static struct tree_mod_elem *
782 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
783                            u64 min_seq)
784 {
785         return __tree_mod_log_search(fs_info, start, min_seq, 1);
786 }
787
788 /*
789  * this returns the element from the log with the largest time sequence
790  * value that's in the log (the most recent log item). any element with
791  * a time sequence lower than min_seq will be ignored.
792  */
793 static struct tree_mod_elem *
794 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
795 {
796         return __tree_mod_log_search(fs_info, start, min_seq, 0);
797 }
798
799 static noinline int
800 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
801                      struct extent_buffer *src, unsigned long dst_offset,
802                      unsigned long src_offset, int nr_items)
803 {
804         int ret = 0;
805         struct tree_mod_elem **tm_list = NULL;
806         struct tree_mod_elem **tm_list_add, **tm_list_rem;
807         int i;
808         int locked = 0;
809
810         if (!tree_mod_need_log(fs_info, NULL))
811                 return 0;
812
813         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
814                 return 0;
815
816         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
817                           GFP_NOFS);
818         if (!tm_list)
819                 return -ENOMEM;
820
821         tm_list_add = tm_list;
822         tm_list_rem = tm_list + nr_items;
823         for (i = 0; i < nr_items; i++) {
824                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
825                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
826                 if (!tm_list_rem[i]) {
827                         ret = -ENOMEM;
828                         goto free_tms;
829                 }
830
831                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
832                     MOD_LOG_KEY_ADD, GFP_NOFS);
833                 if (!tm_list_add[i]) {
834                         ret = -ENOMEM;
835                         goto free_tms;
836                 }
837         }
838
839         if (tree_mod_dont_log(fs_info, NULL))
840                 goto free_tms;
841         locked = 1;
842
843         for (i = 0; i < nr_items; i++) {
844                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
845                 if (ret)
846                         goto free_tms;
847                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
848                 if (ret)
849                         goto free_tms;
850         }
851
852         tree_mod_log_write_unlock(fs_info);
853         kfree(tm_list);
854
855         return 0;
856
857 free_tms:
858         for (i = 0; i < nr_items * 2; i++) {
859                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
860                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
861                 kfree(tm_list[i]);
862         }
863         if (locked)
864                 tree_mod_log_write_unlock(fs_info);
865         kfree(tm_list);
866
867         return ret;
868 }
869
870 static inline void
871 tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
872                      int dst_offset, int src_offset, int nr_items)
873 {
874         int ret;
875         ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
876                                        nr_items);
877         BUG_ON(ret < 0);
878 }
879
880 static noinline void
881 tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
882                           struct extent_buffer *eb, int slot, int atomic)
883 {
884         int ret;
885
886         ret = tree_mod_log_insert_key(fs_info, eb, slot,
887                                         MOD_LOG_KEY_REPLACE,
888                                         atomic ? GFP_ATOMIC : GFP_NOFS);
889         BUG_ON(ret < 0);
890 }
891
892 static noinline int
893 tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
894 {
895         struct tree_mod_elem **tm_list = NULL;
896         int nritems = 0;
897         int i;
898         int ret = 0;
899
900         if (btrfs_header_level(eb) == 0)
901                 return 0;
902
903         if (!tree_mod_need_log(fs_info, NULL))
904                 return 0;
905
906         nritems = btrfs_header_nritems(eb);
907         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
908         if (!tm_list)
909                 return -ENOMEM;
910
911         for (i = 0; i < nritems; i++) {
912                 tm_list[i] = alloc_tree_mod_elem(eb, i,
913                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
914                 if (!tm_list[i]) {
915                         ret = -ENOMEM;
916                         goto free_tms;
917                 }
918         }
919
920         if (tree_mod_dont_log(fs_info, eb))
921                 goto free_tms;
922
923         ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
924         tree_mod_log_write_unlock(fs_info);
925         if (ret)
926                 goto free_tms;
927         kfree(tm_list);
928
929         return 0;
930
931 free_tms:
932         for (i = 0; i < nritems; i++)
933                 kfree(tm_list[i]);
934         kfree(tm_list);
935
936         return ret;
937 }
938
939 static noinline void
940 tree_mod_log_set_root_pointer(struct btrfs_root *root,
941                               struct extent_buffer *new_root_node,
942                               int log_removal)
943 {
944         int ret;
945         ret = tree_mod_log_insert_root(root->fs_info, root->node,
946                                        new_root_node, log_removal);
947         BUG_ON(ret < 0);
948 }
949
950 /*
951  * check if the tree block can be shared by multiple trees
952  */
953 int btrfs_block_can_be_shared(struct btrfs_root *root,
954                               struct extent_buffer *buf)
955 {
956         /*
957          * Tree blocks not in reference counted trees and tree roots
958          * are never shared. If a block was allocated after the last
959          * snapshot and the block was not allocated by tree relocation,
960          * we know the block is not shared.
961          */
962         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
963             buf != root->node && buf != root->commit_root &&
964             (btrfs_header_generation(buf) <=
965              btrfs_root_last_snapshot(&root->root_item) ||
966              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
967                 return 1;
968 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
969         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
970             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
971                 return 1;
972 #endif
973         return 0;
974 }
975
976 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
977                                        struct btrfs_root *root,
978                                        struct extent_buffer *buf,
979                                        struct extent_buffer *cow,
980                                        int *last_ref)
981 {
982         struct btrfs_fs_info *fs_info = root->fs_info;
983         u64 refs;
984         u64 owner;
985         u64 flags;
986         u64 new_flags = 0;
987         int ret;
988
989         /*
990          * Backrefs update rules:
991          *
992          * Always use full backrefs for extent pointers in tree block
993          * allocated by tree relocation.
994          *
995          * If a shared tree block is no longer referenced by its owner
996          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
997          * use full backrefs for extent pointers in tree block.
998          *
999          * If a tree block is been relocating
1000          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
1001          * use full backrefs for extent pointers in tree block.
1002          * The reason for this is some operations (such as drop tree)
1003          * are only allowed for blocks use full backrefs.
1004          */
1005
1006         if (btrfs_block_can_be_shared(root, buf)) {
1007                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
1008                                                btrfs_header_level(buf), 1,
1009                                                &refs, &flags);
1010                 if (ret)
1011                         return ret;
1012                 if (refs == 0) {
1013                         ret = -EROFS;
1014                         btrfs_handle_fs_error(fs_info, ret, NULL);
1015                         return ret;
1016                 }
1017         } else {
1018                 refs = 1;
1019                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1020                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1021                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
1022                 else
1023                         flags = 0;
1024         }
1025
1026         owner = btrfs_header_owner(buf);
1027         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
1028                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
1029
1030         if (refs > 1) {
1031                 if ((owner == root->root_key.objectid ||
1032                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
1033                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
1034                         ret = btrfs_inc_ref(trans, root, buf, 1);
1035                         if (ret)
1036                                 return ret;
1037
1038                         if (root->root_key.objectid ==
1039                             BTRFS_TREE_RELOC_OBJECTID) {
1040                                 ret = btrfs_dec_ref(trans, root, buf, 0);
1041                                 if (ret)
1042                                         return ret;
1043                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1044                                 if (ret)
1045                                         return ret;
1046                         }
1047                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
1048                 } else {
1049
1050                         if (root->root_key.objectid ==
1051                             BTRFS_TREE_RELOC_OBJECTID)
1052                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1053                         else
1054                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1055                         if (ret)
1056                                 return ret;
1057                 }
1058                 if (new_flags != 0) {
1059                         int level = btrfs_header_level(buf);
1060
1061                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
1062                                                           buf->start,
1063                                                           buf->len,
1064                                                           new_flags, level, 0);
1065                         if (ret)
1066                                 return ret;
1067                 }
1068         } else {
1069                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
1070                         if (root->root_key.objectid ==
1071                             BTRFS_TREE_RELOC_OBJECTID)
1072                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1073                         else
1074                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1075                         if (ret)
1076                                 return ret;
1077                         ret = btrfs_dec_ref(trans, root, buf, 1);
1078                         if (ret)
1079                                 return ret;
1080                 }
1081                 clean_tree_block(fs_info, buf);
1082                 *last_ref = 1;
1083         }
1084         return 0;
1085 }
1086
1087 /*
1088  * does the dirty work in cow of a single block.  The parent block (if
1089  * supplied) is updated to point to the new cow copy.  The new buffer is marked
1090  * dirty and returned locked.  If you modify the block it needs to be marked
1091  * dirty again.
1092  *
1093  * search_start -- an allocation hint for the new block
1094  *
1095  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1096  * bytes the allocator should try to find free next to the block it returns.
1097  * This is just a hint and may be ignored by the allocator.
1098  */
1099 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1100                              struct btrfs_root *root,
1101                              struct extent_buffer *buf,
1102                              struct extent_buffer *parent, int parent_slot,
1103                              struct extent_buffer **cow_ret,
1104                              u64 search_start, u64 empty_size)
1105 {
1106         struct btrfs_fs_info *fs_info = root->fs_info;
1107         struct btrfs_disk_key disk_key;
1108         struct extent_buffer *cow;
1109         int level, ret;
1110         int last_ref = 0;
1111         int unlock_orig = 0;
1112         u64 parent_start = 0;
1113
1114         if (*cow_ret == buf)
1115                 unlock_orig = 1;
1116
1117         btrfs_assert_tree_locked(buf);
1118
1119         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1120                 trans->transid != fs_info->running_transaction->transid);
1121         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1122                 trans->transid != root->last_trans);
1123
1124         level = btrfs_header_level(buf);
1125
1126         if (level == 0)
1127                 btrfs_item_key(buf, &disk_key, 0);
1128         else
1129                 btrfs_node_key(buf, &disk_key, 0);
1130
1131         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1132                 parent_start = parent->start;
1133
1134         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1135                         root->root_key.objectid, &disk_key, level,
1136                         search_start, empty_size);
1137         if (IS_ERR(cow))
1138                 return PTR_ERR(cow);
1139
1140         /* cow is set to blocking by btrfs_init_new_buffer */
1141
1142         copy_extent_buffer_full(cow, buf);
1143         btrfs_set_header_bytenr(cow, cow->start);
1144         btrfs_set_header_generation(cow, trans->transid);
1145         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1146         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1147                                      BTRFS_HEADER_FLAG_RELOC);
1148         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1149                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1150         else
1151                 btrfs_set_header_owner(cow, root->root_key.objectid);
1152
1153         write_extent_buffer_fsid(cow, fs_info->fsid);
1154
1155         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1156         if (ret) {
1157                 btrfs_abort_transaction(trans, ret);
1158                 return ret;
1159         }
1160
1161         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1162                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1163                 if (ret) {
1164                         btrfs_abort_transaction(trans, ret);
1165                         return ret;
1166                 }
1167         }
1168
1169         if (buf == root->node) {
1170                 WARN_ON(parent && parent != buf);
1171                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1172                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1173                         parent_start = buf->start;
1174
1175                 extent_buffer_get(cow);
1176                 tree_mod_log_set_root_pointer(root, cow, 1);
1177                 rcu_assign_pointer(root->node, cow);
1178
1179                 btrfs_free_tree_block(trans, root, buf, parent_start,
1180                                       last_ref);
1181                 free_extent_buffer(buf);
1182                 add_root_to_dirty_list(root);
1183         } else {
1184                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1185                 tree_mod_log_insert_key(fs_info, parent, parent_slot,
1186                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1187                 btrfs_set_node_blockptr(parent, parent_slot,
1188                                         cow->start);
1189                 btrfs_set_node_ptr_generation(parent, parent_slot,
1190                                               trans->transid);
1191                 btrfs_mark_buffer_dirty(parent);
1192                 if (last_ref) {
1193                         ret = tree_mod_log_free_eb(fs_info, buf);
1194                         if (ret) {
1195                                 btrfs_abort_transaction(trans, ret);
1196                                 return ret;
1197                         }
1198                 }
1199                 btrfs_free_tree_block(trans, root, buf, parent_start,
1200                                       last_ref);
1201         }
1202         if (unlock_orig)
1203                 btrfs_tree_unlock(buf);
1204         free_extent_buffer_stale(buf);
1205         btrfs_mark_buffer_dirty(cow);
1206         *cow_ret = cow;
1207         return 0;
1208 }
1209
1210 /*
1211  * returns the logical address of the oldest predecessor of the given root.
1212  * entries older than time_seq are ignored.
1213  */
1214 static struct tree_mod_elem *
1215 __tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1216                            struct extent_buffer *eb_root, u64 time_seq)
1217 {
1218         struct tree_mod_elem *tm;
1219         struct tree_mod_elem *found = NULL;
1220         u64 root_logical = eb_root->start;
1221         int looped = 0;
1222
1223         if (!time_seq)
1224                 return NULL;
1225
1226         /*
1227          * the very last operation that's logged for a root is the
1228          * replacement operation (if it is replaced at all). this has
1229          * the logical address of the *new* root, making it the very
1230          * first operation that's logged for this root.
1231          */
1232         while (1) {
1233                 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1234                                                 time_seq);
1235                 if (!looped && !tm)
1236                         return NULL;
1237                 /*
1238                  * if there are no tree operation for the oldest root, we simply
1239                  * return it. this should only happen if that (old) root is at
1240                  * level 0.
1241                  */
1242                 if (!tm)
1243                         break;
1244
1245                 /*
1246                  * if there's an operation that's not a root replacement, we
1247                  * found the oldest version of our root. normally, we'll find a
1248                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1249                  */
1250                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1251                         break;
1252
1253                 found = tm;
1254                 root_logical = tm->old_root.logical;
1255                 looped = 1;
1256         }
1257
1258         /* if there's no old root to return, return what we found instead */
1259         if (!found)
1260                 found = tm;
1261
1262         return found;
1263 }
1264
1265 /*
1266  * tm is a pointer to the first operation to rewind within eb. then, all
1267  * previous operations will be rewound (until we reach something older than
1268  * time_seq).
1269  */
1270 static void
1271 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1272                       u64 time_seq, struct tree_mod_elem *first_tm)
1273 {
1274         u32 n;
1275         struct rb_node *next;
1276         struct tree_mod_elem *tm = first_tm;
1277         unsigned long o_dst;
1278         unsigned long o_src;
1279         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1280
1281         n = btrfs_header_nritems(eb);
1282         tree_mod_log_read_lock(fs_info);
1283         while (tm && tm->seq >= time_seq) {
1284                 /*
1285                  * all the operations are recorded with the operator used for
1286                  * the modification. as we're going backwards, we do the
1287                  * opposite of each operation here.
1288                  */
1289                 switch (tm->op) {
1290                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1291                         BUG_ON(tm->slot < n);
1292                         /* Fallthrough */
1293                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1294                 case MOD_LOG_KEY_REMOVE:
1295                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1296                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1297                         btrfs_set_node_ptr_generation(eb, tm->slot,
1298                                                       tm->generation);
1299                         n++;
1300                         break;
1301                 case MOD_LOG_KEY_REPLACE:
1302                         BUG_ON(tm->slot >= n);
1303                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1304                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1305                         btrfs_set_node_ptr_generation(eb, tm->slot,
1306                                                       tm->generation);
1307                         break;
1308                 case MOD_LOG_KEY_ADD:
1309                         /* if a move operation is needed it's in the log */
1310                         n--;
1311                         break;
1312                 case MOD_LOG_MOVE_KEYS:
1313                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1314                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1315                         memmove_extent_buffer(eb, o_dst, o_src,
1316                                               tm->move.nr_items * p_size);
1317                         break;
1318                 case MOD_LOG_ROOT_REPLACE:
1319                         /*
1320                          * this operation is special. for roots, this must be
1321                          * handled explicitly before rewinding.
1322                          * for non-roots, this operation may exist if the node
1323                          * was a root: root A -> child B; then A gets empty and
1324                          * B is promoted to the new root. in the mod log, we'll
1325                          * have a root-replace operation for B, a tree block
1326                          * that is no root. we simply ignore that operation.
1327                          */
1328                         break;
1329                 }
1330                 next = rb_next(&tm->node);
1331                 if (!next)
1332                         break;
1333                 tm = rb_entry(next, struct tree_mod_elem, node);
1334                 if (tm->logical != first_tm->logical)
1335                         break;
1336         }
1337         tree_mod_log_read_unlock(fs_info);
1338         btrfs_set_header_nritems(eb, n);
1339 }
1340
1341 /*
1342  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1343  * is returned. If rewind operations happen, a fresh buffer is returned. The
1344  * returned buffer is always read-locked. If the returned buffer is not the
1345  * input buffer, the lock on the input buffer is released and the input buffer
1346  * is freed (its refcount is decremented).
1347  */
1348 static struct extent_buffer *
1349 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1350                     struct extent_buffer *eb, u64 time_seq)
1351 {
1352         struct extent_buffer *eb_rewin;
1353         struct tree_mod_elem *tm;
1354
1355         if (!time_seq)
1356                 return eb;
1357
1358         if (btrfs_header_level(eb) == 0)
1359                 return eb;
1360
1361         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1362         if (!tm)
1363                 return eb;
1364
1365         btrfs_set_path_blocking(path);
1366         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1367
1368         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1369                 BUG_ON(tm->slot != 0);
1370                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1371                 if (!eb_rewin) {
1372                         btrfs_tree_read_unlock_blocking(eb);
1373                         free_extent_buffer(eb);
1374                         return NULL;
1375                 }
1376                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1377                 btrfs_set_header_backref_rev(eb_rewin,
1378                                              btrfs_header_backref_rev(eb));
1379                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1380                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1381         } else {
1382                 eb_rewin = btrfs_clone_extent_buffer(eb);
1383                 if (!eb_rewin) {
1384                         btrfs_tree_read_unlock_blocking(eb);
1385                         free_extent_buffer(eb);
1386                         return NULL;
1387                 }
1388         }
1389
1390         btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1391         btrfs_tree_read_unlock_blocking(eb);
1392         free_extent_buffer(eb);
1393
1394         extent_buffer_get(eb_rewin);
1395         btrfs_tree_read_lock(eb_rewin);
1396         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1397         WARN_ON(btrfs_header_nritems(eb_rewin) >
1398                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1399
1400         return eb_rewin;
1401 }
1402
1403 /*
1404  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1405  * value. If there are no changes, the current root->root_node is returned. If
1406  * anything changed in between, there's a fresh buffer allocated on which the
1407  * rewind operations are done. In any case, the returned buffer is read locked.
1408  * Returns NULL on error (with no locks held).
1409  */
1410 static inline struct extent_buffer *
1411 get_old_root(struct btrfs_root *root, u64 time_seq)
1412 {
1413         struct btrfs_fs_info *fs_info = root->fs_info;
1414         struct tree_mod_elem *tm;
1415         struct extent_buffer *eb = NULL;
1416         struct extent_buffer *eb_root;
1417         struct extent_buffer *old;
1418         struct tree_mod_root *old_root = NULL;
1419         u64 old_generation = 0;
1420         u64 logical;
1421
1422         eb_root = btrfs_read_lock_root_node(root);
1423         tm = __tree_mod_log_oldest_root(fs_info, eb_root, time_seq);
1424         if (!tm)
1425                 return eb_root;
1426
1427         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1428                 old_root = &tm->old_root;
1429                 old_generation = tm->generation;
1430                 logical = old_root->logical;
1431         } else {
1432                 logical = eb_root->start;
1433         }
1434
1435         tm = tree_mod_log_search(fs_info, logical, time_seq);
1436         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1437                 btrfs_tree_read_unlock(eb_root);
1438                 free_extent_buffer(eb_root);
1439                 old = read_tree_block(fs_info, logical, 0);
1440                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1441                         if (!IS_ERR(old))
1442                                 free_extent_buffer(old);
1443                         btrfs_warn(fs_info,
1444                                    "failed to read tree block %llu from get_old_root",
1445                                    logical);
1446                 } else {
1447                         eb = btrfs_clone_extent_buffer(old);
1448                         free_extent_buffer(old);
1449                 }
1450         } else if (old_root) {
1451                 btrfs_tree_read_unlock(eb_root);
1452                 free_extent_buffer(eb_root);
1453                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1454         } else {
1455                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1456                 eb = btrfs_clone_extent_buffer(eb_root);
1457                 btrfs_tree_read_unlock_blocking(eb_root);
1458                 free_extent_buffer(eb_root);
1459         }
1460
1461         if (!eb)
1462                 return NULL;
1463         extent_buffer_get(eb);
1464         btrfs_tree_read_lock(eb);
1465         if (old_root) {
1466                 btrfs_set_header_bytenr(eb, eb->start);
1467                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1468                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1469                 btrfs_set_header_level(eb, old_root->level);
1470                 btrfs_set_header_generation(eb, old_generation);
1471         }
1472         if (tm)
1473                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1474         else
1475                 WARN_ON(btrfs_header_level(eb) != 0);
1476         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1477
1478         return eb;
1479 }
1480
1481 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1482 {
1483         struct tree_mod_elem *tm;
1484         int level;
1485         struct extent_buffer *eb_root = btrfs_root_node(root);
1486
1487         tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1488         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1489                 level = tm->old_root.level;
1490         } else {
1491                 level = btrfs_header_level(eb_root);
1492         }
1493         free_extent_buffer(eb_root);
1494
1495         return level;
1496 }
1497
1498 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1499                                    struct btrfs_root *root,
1500                                    struct extent_buffer *buf)
1501 {
1502         if (btrfs_is_testing(root->fs_info))
1503                 return 0;
1504
1505         /* ensure we can see the force_cow */
1506         smp_rmb();
1507
1508         /*
1509          * We do not need to cow a block if
1510          * 1) this block is not created or changed in this transaction;
1511          * 2) this block does not belong to TREE_RELOC tree;
1512          * 3) the root is not forced COW.
1513          *
1514          * What is forced COW:
1515          *    when we create snapshot during committing the transaction,
1516          *    after we've finished coping src root, we must COW the shared
1517          *    block to ensure the metadata consistency.
1518          */
1519         if (btrfs_header_generation(buf) == trans->transid &&
1520             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1521             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1522               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1523             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1524                 return 0;
1525         return 1;
1526 }
1527
1528 /*
1529  * cows a single block, see __btrfs_cow_block for the real work.
1530  * This version of it has extra checks so that a block isn't COWed more than
1531  * once per transaction, as long as it hasn't been written yet
1532  */
1533 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1534                     struct btrfs_root *root, struct extent_buffer *buf,
1535                     struct extent_buffer *parent, int parent_slot,
1536                     struct extent_buffer **cow_ret)
1537 {
1538         struct btrfs_fs_info *fs_info = root->fs_info;
1539         u64 search_start;
1540         int ret;
1541
1542         if (trans->transaction != fs_info->running_transaction)
1543                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1544                        trans->transid,
1545                        fs_info->running_transaction->transid);
1546
1547         if (trans->transid != fs_info->generation)
1548                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1549                        trans->transid, fs_info->generation);
1550
1551         if (!should_cow_block(trans, root, buf)) {
1552                 trans->dirty = true;
1553                 *cow_ret = buf;
1554                 return 0;
1555         }
1556
1557         search_start = buf->start & ~((u64)SZ_1G - 1);
1558
1559         if (parent)
1560                 btrfs_set_lock_blocking(parent);
1561         btrfs_set_lock_blocking(buf);
1562
1563         ret = __btrfs_cow_block(trans, root, buf, parent,
1564                                  parent_slot, cow_ret, search_start, 0);
1565
1566         trace_btrfs_cow_block(root, buf, *cow_ret);
1567
1568         return ret;
1569 }
1570
1571 /*
1572  * helper function for defrag to decide if two blocks pointed to by a
1573  * node are actually close by
1574  */
1575 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1576 {
1577         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1578                 return 1;
1579         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1580                 return 1;
1581         return 0;
1582 }
1583
1584 /*
1585  * compare two keys in a memcmp fashion
1586  */
1587 static int comp_keys(const struct btrfs_disk_key *disk,
1588                      const struct btrfs_key *k2)
1589 {
1590         struct btrfs_key k1;
1591
1592         btrfs_disk_key_to_cpu(&k1, disk);
1593
1594         return btrfs_comp_cpu_keys(&k1, k2);
1595 }
1596
1597 /*
1598  * same as comp_keys only with two btrfs_key's
1599  */
1600 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1601 {
1602         if (k1->objectid > k2->objectid)
1603                 return 1;
1604         if (k1->objectid < k2->objectid)
1605                 return -1;
1606         if (k1->type > k2->type)
1607                 return 1;
1608         if (k1->type < k2->type)
1609                 return -1;
1610         if (k1->offset > k2->offset)
1611                 return 1;
1612         if (k1->offset < k2->offset)
1613                 return -1;
1614         return 0;
1615 }
1616
1617 /*
1618  * this is used by the defrag code to go through all the
1619  * leaves pointed to by a node and reallocate them so that
1620  * disk order is close to key order
1621  */
1622 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1623                        struct btrfs_root *root, struct extent_buffer *parent,
1624                        int start_slot, u64 *last_ret,
1625                        struct btrfs_key *progress)
1626 {
1627         struct btrfs_fs_info *fs_info = root->fs_info;
1628         struct extent_buffer *cur;
1629         u64 blocknr;
1630         u64 gen;
1631         u64 search_start = *last_ret;
1632         u64 last_block = 0;
1633         u64 other;
1634         u32 parent_nritems;
1635         int end_slot;
1636         int i;
1637         int err = 0;
1638         int parent_level;
1639         int uptodate;
1640         u32 blocksize;
1641         int progress_passed = 0;
1642         struct btrfs_disk_key disk_key;
1643
1644         parent_level = btrfs_header_level(parent);
1645
1646         WARN_ON(trans->transaction != fs_info->running_transaction);
1647         WARN_ON(trans->transid != fs_info->generation);
1648
1649         parent_nritems = btrfs_header_nritems(parent);
1650         blocksize = fs_info->nodesize;
1651         end_slot = parent_nritems - 1;
1652
1653         if (parent_nritems <= 1)
1654                 return 0;
1655
1656         btrfs_set_lock_blocking(parent);
1657
1658         for (i = start_slot; i <= end_slot; i++) {
1659                 int close = 1;
1660
1661                 btrfs_node_key(parent, &disk_key, i);
1662                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1663                         continue;
1664
1665                 progress_passed = 1;
1666                 blocknr = btrfs_node_blockptr(parent, i);
1667                 gen = btrfs_node_ptr_generation(parent, i);
1668                 if (last_block == 0)
1669                         last_block = blocknr;
1670
1671                 if (i > 0) {
1672                         other = btrfs_node_blockptr(parent, i - 1);
1673                         close = close_blocks(blocknr, other, blocksize);
1674                 }
1675                 if (!close && i < end_slot) {
1676                         other = btrfs_node_blockptr(parent, i + 1);
1677                         close = close_blocks(blocknr, other, blocksize);
1678                 }
1679                 if (close) {
1680                         last_block = blocknr;
1681                         continue;
1682                 }
1683
1684                 cur = find_extent_buffer(fs_info, blocknr);
1685                 if (cur)
1686                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1687                 else
1688                         uptodate = 0;
1689                 if (!cur || !uptodate) {
1690                         if (!cur) {
1691                                 cur = read_tree_block(fs_info, blocknr, gen);
1692                                 if (IS_ERR(cur)) {
1693                                         return PTR_ERR(cur);
1694                                 } else if (!extent_buffer_uptodate(cur)) {
1695                                         free_extent_buffer(cur);
1696                                         return -EIO;
1697                                 }
1698                         } else if (!uptodate) {
1699                                 err = btrfs_read_buffer(cur, gen);
1700                                 if (err) {
1701                                         free_extent_buffer(cur);
1702                                         return err;
1703                                 }
1704                         }
1705                 }
1706                 if (search_start == 0)
1707                         search_start = last_block;
1708
1709                 btrfs_tree_lock(cur);
1710                 btrfs_set_lock_blocking(cur);
1711                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1712                                         &cur, search_start,
1713                                         min(16 * blocksize,
1714                                             (end_slot - i) * blocksize));
1715                 if (err) {
1716                         btrfs_tree_unlock(cur);
1717                         free_extent_buffer(cur);
1718                         break;
1719                 }
1720                 search_start = cur->start;
1721                 last_block = cur->start;
1722                 *last_ret = search_start;
1723                 btrfs_tree_unlock(cur);
1724                 free_extent_buffer(cur);
1725         }
1726         return err;
1727 }
1728
1729 /*
1730  * search for key in the extent_buffer.  The items start at offset p,
1731  * and they are item_size apart.  There are 'max' items in p.
1732  *
1733  * the slot in the array is returned via slot, and it points to
1734  * the place where you would insert key if it is not found in
1735  * the array.
1736  *
1737  * slot may point to max if the key is bigger than all of the keys
1738  */
1739 static noinline int generic_bin_search(struct extent_buffer *eb,
1740                                        unsigned long p, int item_size,
1741                                        const struct btrfs_key *key,
1742                                        int max, int *slot)
1743 {
1744         int low = 0;
1745         int high = max;
1746         int mid;
1747         int ret;
1748         struct btrfs_disk_key *tmp = NULL;
1749         struct btrfs_disk_key unaligned;
1750         unsigned long offset;
1751         char *kaddr = NULL;
1752         unsigned long map_start = 0;
1753         unsigned long map_len = 0;
1754         int err;
1755
1756         if (low > high) {
1757                 btrfs_err(eb->fs_info,
1758                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1759                           __func__, low, high, eb->start,
1760                           btrfs_header_owner(eb), btrfs_header_level(eb));
1761                 return -EINVAL;
1762         }
1763
1764         while (low < high) {
1765                 mid = (low + high) / 2;
1766                 offset = p + mid * item_size;
1767
1768                 if (!kaddr || offset < map_start ||
1769                     (offset + sizeof(struct btrfs_disk_key)) >
1770                     map_start + map_len) {
1771
1772                         err = map_private_extent_buffer(eb, offset,
1773                                                 sizeof(struct btrfs_disk_key),
1774                                                 &kaddr, &map_start, &map_len);
1775
1776                         if (!err) {
1777                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1778                                                         map_start);
1779                         } else if (err == 1) {
1780                                 read_extent_buffer(eb, &unaligned,
1781                                                    offset, sizeof(unaligned));
1782                                 tmp = &unaligned;
1783                         } else {
1784                                 return err;
1785                         }
1786
1787                 } else {
1788                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1789                                                         map_start);
1790                 }
1791                 ret = comp_keys(tmp, key);
1792
1793                 if (ret < 0)
1794                         low = mid + 1;
1795                 else if (ret > 0)
1796                         high = mid;
1797                 else {
1798                         *slot = mid;
1799                         return 0;
1800                 }
1801         }
1802         *slot = low;
1803         return 1;
1804 }
1805
1806 /*
1807  * simple bin_search frontend that does the right thing for
1808  * leaves vs nodes
1809  */
1810 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1811                      int level, int *slot)
1812 {
1813         if (level == 0)
1814                 return generic_bin_search(eb,
1815                                           offsetof(struct btrfs_leaf, items),
1816                                           sizeof(struct btrfs_item),
1817                                           key, btrfs_header_nritems(eb),
1818                                           slot);
1819         else
1820                 return generic_bin_search(eb,
1821                                           offsetof(struct btrfs_node, ptrs),
1822                                           sizeof(struct btrfs_key_ptr),
1823                                           key, btrfs_header_nritems(eb),
1824                                           slot);
1825 }
1826
1827 static void root_add_used(struct btrfs_root *root, u32 size)
1828 {
1829         spin_lock(&root->accounting_lock);
1830         btrfs_set_root_used(&root->root_item,
1831                             btrfs_root_used(&root->root_item) + size);
1832         spin_unlock(&root->accounting_lock);
1833 }
1834
1835 static void root_sub_used(struct btrfs_root *root, u32 size)
1836 {
1837         spin_lock(&root->accounting_lock);
1838         btrfs_set_root_used(&root->root_item,
1839                             btrfs_root_used(&root->root_item) - size);
1840         spin_unlock(&root->accounting_lock);
1841 }
1842
1843 /* given a node and slot number, this reads the blocks it points to.  The
1844  * extent buffer is returned with a reference taken (but unlocked).
1845  */
1846 static noinline struct extent_buffer *
1847 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1848                int slot)
1849 {
1850         int level = btrfs_header_level(parent);
1851         struct extent_buffer *eb;
1852
1853         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1854                 return ERR_PTR(-ENOENT);
1855
1856         BUG_ON(level == 0);
1857
1858         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1859                              btrfs_node_ptr_generation(parent, slot));
1860         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1861                 free_extent_buffer(eb);
1862                 eb = ERR_PTR(-EIO);
1863         }
1864
1865         return eb;
1866 }
1867
1868 /*
1869  * node level balancing, used to make sure nodes are in proper order for
1870  * item deletion.  We balance from the top down, so we have to make sure
1871  * that a deletion won't leave an node completely empty later on.
1872  */
1873 static noinline int balance_level(struct btrfs_trans_handle *trans,
1874                          struct btrfs_root *root,
1875                          struct btrfs_path *path, int level)
1876 {
1877         struct btrfs_fs_info *fs_info = root->fs_info;
1878         struct extent_buffer *right = NULL;
1879         struct extent_buffer *mid;
1880         struct extent_buffer *left = NULL;
1881         struct extent_buffer *parent = NULL;
1882         int ret = 0;
1883         int wret;
1884         int pslot;
1885         int orig_slot = path->slots[level];
1886         u64 orig_ptr;
1887
1888         if (level == 0)
1889                 return 0;
1890
1891         mid = path->nodes[level];
1892
1893         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1894                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1895         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1896
1897         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1898
1899         if (level < BTRFS_MAX_LEVEL - 1) {
1900                 parent = path->nodes[level + 1];
1901                 pslot = path->slots[level + 1];
1902         }
1903
1904         /*
1905          * deal with the case where there is only one pointer in the root
1906          * by promoting the node below to a root
1907          */
1908         if (!parent) {
1909                 struct extent_buffer *child;
1910
1911                 if (btrfs_header_nritems(mid) != 1)
1912                         return 0;
1913
1914                 /* promote the child to a root */
1915                 child = read_node_slot(fs_info, mid, 0);
1916                 if (IS_ERR(child)) {
1917                         ret = PTR_ERR(child);
1918                         btrfs_handle_fs_error(fs_info, ret, NULL);
1919                         goto enospc;
1920                 }
1921
1922                 btrfs_tree_lock(child);
1923                 btrfs_set_lock_blocking(child);
1924                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1925                 if (ret) {
1926                         btrfs_tree_unlock(child);
1927                         free_extent_buffer(child);
1928                         goto enospc;
1929                 }
1930
1931                 tree_mod_log_set_root_pointer(root, child, 1);
1932                 rcu_assign_pointer(root->node, child);
1933
1934                 add_root_to_dirty_list(root);
1935                 btrfs_tree_unlock(child);
1936
1937                 path->locks[level] = 0;
1938                 path->nodes[level] = NULL;
1939                 clean_tree_block(fs_info, mid);
1940                 btrfs_tree_unlock(mid);
1941                 /* once for the path */
1942                 free_extent_buffer(mid);
1943
1944                 root_sub_used(root, mid->len);
1945                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1946                 /* once for the root ptr */
1947                 free_extent_buffer_stale(mid);
1948                 return 0;
1949         }
1950         if (btrfs_header_nritems(mid) >
1951             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1952                 return 0;
1953
1954         left = read_node_slot(fs_info, parent, pslot - 1);
1955         if (IS_ERR(left))
1956                 left = NULL;
1957
1958         if (left) {
1959                 btrfs_tree_lock(left);
1960                 btrfs_set_lock_blocking(left);
1961                 wret = btrfs_cow_block(trans, root, left,
1962                                        parent, pslot - 1, &left);
1963                 if (wret) {
1964                         ret = wret;
1965                         goto enospc;
1966                 }
1967         }
1968
1969         right = read_node_slot(fs_info, parent, pslot + 1);
1970         if (IS_ERR(right))
1971                 right = NULL;
1972
1973         if (right) {
1974                 btrfs_tree_lock(right);
1975                 btrfs_set_lock_blocking(right);
1976                 wret = btrfs_cow_block(trans, root, right,
1977                                        parent, pslot + 1, &right);
1978                 if (wret) {
1979                         ret = wret;
1980                         goto enospc;
1981                 }
1982         }
1983
1984         /* first, try to make some room in the middle buffer */
1985         if (left) {
1986                 orig_slot += btrfs_header_nritems(left);
1987                 wret = push_node_left(trans, fs_info, left, mid, 1);
1988                 if (wret < 0)
1989                         ret = wret;
1990         }
1991
1992         /*
1993          * then try to empty the right most buffer into the middle
1994          */
1995         if (right) {
1996                 wret = push_node_left(trans, fs_info, mid, right, 1);
1997                 if (wret < 0 && wret != -ENOSPC)
1998                         ret = wret;
1999                 if (btrfs_header_nritems(right) == 0) {
2000                         clean_tree_block(fs_info, right);
2001                         btrfs_tree_unlock(right);
2002                         del_ptr(root, path, level + 1, pslot + 1);
2003                         root_sub_used(root, right->len);
2004                         btrfs_free_tree_block(trans, root, right, 0, 1);
2005                         free_extent_buffer_stale(right);
2006                         right = NULL;
2007                 } else {
2008                         struct btrfs_disk_key right_key;
2009                         btrfs_node_key(right, &right_key, 0);
2010                         tree_mod_log_set_node_key(fs_info, parent,
2011                                                   pslot + 1, 0);
2012                         btrfs_set_node_key(parent, &right_key, pslot + 1);
2013                         btrfs_mark_buffer_dirty(parent);
2014                 }
2015         }
2016         if (btrfs_header_nritems(mid) == 1) {
2017                 /*
2018                  * we're not allowed to leave a node with one item in the
2019                  * tree during a delete.  A deletion from lower in the tree
2020                  * could try to delete the only pointer in this node.
2021                  * So, pull some keys from the left.
2022                  * There has to be a left pointer at this point because
2023                  * otherwise we would have pulled some pointers from the
2024                  * right
2025                  */
2026                 if (!left) {
2027                         ret = -EROFS;
2028                         btrfs_handle_fs_error(fs_info, ret, NULL);
2029                         goto enospc;
2030                 }
2031                 wret = balance_node_right(trans, fs_info, mid, left);
2032                 if (wret < 0) {
2033                         ret = wret;
2034                         goto enospc;
2035                 }
2036                 if (wret == 1) {
2037                         wret = push_node_left(trans, fs_info, left, mid, 1);
2038                         if (wret < 0)
2039                                 ret = wret;
2040                 }
2041                 BUG_ON(wret == 1);
2042         }
2043         if (btrfs_header_nritems(mid) == 0) {
2044                 clean_tree_block(fs_info, mid);
2045                 btrfs_tree_unlock(mid);
2046                 del_ptr(root, path, level + 1, pslot);
2047                 root_sub_used(root, mid->len);
2048                 btrfs_free_tree_block(trans, root, mid, 0, 1);
2049                 free_extent_buffer_stale(mid);
2050                 mid = NULL;
2051         } else {
2052                 /* update the parent key to reflect our changes */
2053                 struct btrfs_disk_key mid_key;
2054                 btrfs_node_key(mid, &mid_key, 0);
2055                 tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2056                 btrfs_set_node_key(parent, &mid_key, pslot);
2057                 btrfs_mark_buffer_dirty(parent);
2058         }
2059
2060         /* update the path */
2061         if (left) {
2062                 if (btrfs_header_nritems(left) > orig_slot) {
2063                         extent_buffer_get(left);
2064                         /* left was locked after cow */
2065                         path->nodes[level] = left;
2066                         path->slots[level + 1] -= 1;
2067                         path->slots[level] = orig_slot;
2068                         if (mid) {
2069                                 btrfs_tree_unlock(mid);
2070                                 free_extent_buffer(mid);
2071                         }
2072                 } else {
2073                         orig_slot -= btrfs_header_nritems(left);
2074                         path->slots[level] = orig_slot;
2075                 }
2076         }
2077         /* double check we haven't messed things up */
2078         if (orig_ptr !=
2079             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2080                 BUG();
2081 enospc:
2082         if (right) {
2083                 btrfs_tree_unlock(right);
2084                 free_extent_buffer(right);
2085         }
2086         if (left) {
2087                 if (path->nodes[level] != left)
2088                         btrfs_tree_unlock(left);
2089                 free_extent_buffer(left);
2090         }
2091         return ret;
2092 }
2093
2094 /* Node balancing for insertion.  Here we only split or push nodes around
2095  * when they are completely full.  This is also done top down, so we
2096  * have to be pessimistic.
2097  */
2098 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2099                                           struct btrfs_root *root,
2100                                           struct btrfs_path *path, int level)
2101 {
2102         struct btrfs_fs_info *fs_info = root->fs_info;
2103         struct extent_buffer *right = NULL;
2104         struct extent_buffer *mid;
2105         struct extent_buffer *left = NULL;
2106         struct extent_buffer *parent = NULL;
2107         int ret = 0;
2108         int wret;
2109         int pslot;
2110         int orig_slot = path->slots[level];
2111
2112         if (level == 0)
2113                 return 1;
2114
2115         mid = path->nodes[level];
2116         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2117
2118         if (level < BTRFS_MAX_LEVEL - 1) {
2119                 parent = path->nodes[level + 1];
2120                 pslot = path->slots[level + 1];
2121         }
2122
2123         if (!parent)
2124                 return 1;
2125
2126         left = read_node_slot(fs_info, parent, pslot - 1);
2127         if (IS_ERR(left))
2128                 left = NULL;
2129
2130         /* first, try to make some room in the middle buffer */
2131         if (left) {
2132                 u32 left_nr;
2133
2134                 btrfs_tree_lock(left);
2135                 btrfs_set_lock_blocking(left);
2136
2137                 left_nr = btrfs_header_nritems(left);
2138                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2139                         wret = 1;
2140                 } else {
2141                         ret = btrfs_cow_block(trans, root, left, parent,
2142                                               pslot - 1, &left);
2143                         if (ret)
2144                                 wret = 1;
2145                         else {
2146                                 wret = push_node_left(trans, fs_info,
2147                                                       left, mid, 0);
2148                         }
2149                 }
2150                 if (wret < 0)
2151                         ret = wret;
2152                 if (wret == 0) {
2153                         struct btrfs_disk_key disk_key;
2154                         orig_slot += left_nr;
2155                         btrfs_node_key(mid, &disk_key, 0);
2156                         tree_mod_log_set_node_key(fs_info, parent, pslot, 0);
2157                         btrfs_set_node_key(parent, &disk_key, pslot);
2158                         btrfs_mark_buffer_dirty(parent);
2159                         if (btrfs_header_nritems(left) > orig_slot) {
2160                                 path->nodes[level] = left;
2161                                 path->slots[level + 1] -= 1;
2162                                 path->slots[level] = orig_slot;
2163                                 btrfs_tree_unlock(mid);
2164                                 free_extent_buffer(mid);
2165                         } else {
2166                                 orig_slot -=
2167                                         btrfs_header_nritems(left);
2168                                 path->slots[level] = orig_slot;
2169                                 btrfs_tree_unlock(left);
2170                                 free_extent_buffer(left);
2171                         }
2172                         return 0;
2173                 }
2174                 btrfs_tree_unlock(left);
2175                 free_extent_buffer(left);
2176         }
2177         right = read_node_slot(fs_info, parent, pslot + 1);
2178         if (IS_ERR(right))
2179                 right = NULL;
2180
2181         /*
2182          * then try to empty the right most buffer into the middle
2183          */
2184         if (right) {
2185                 u32 right_nr;
2186
2187                 btrfs_tree_lock(right);
2188                 btrfs_set_lock_blocking(right);
2189
2190                 right_nr = btrfs_header_nritems(right);
2191                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2192                         wret = 1;
2193                 } else {
2194                         ret = btrfs_cow_block(trans, root, right,
2195                                               parent, pslot + 1,
2196                                               &right);
2197                         if (ret)
2198                                 wret = 1;
2199                         else {
2200                                 wret = balance_node_right(trans, fs_info,
2201                                                           right, mid);
2202                         }
2203                 }
2204                 if (wret < 0)
2205                         ret = wret;
2206                 if (wret == 0) {
2207                         struct btrfs_disk_key disk_key;
2208
2209                         btrfs_node_key(right, &disk_key, 0);
2210                         tree_mod_log_set_node_key(fs_info, parent,
2211                                                   pslot + 1, 0);
2212                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2213                         btrfs_mark_buffer_dirty(parent);
2214
2215                         if (btrfs_header_nritems(mid) <= orig_slot) {
2216                                 path->nodes[level] = right;
2217                                 path->slots[level + 1] += 1;
2218                                 path->slots[level] = orig_slot -
2219                                         btrfs_header_nritems(mid);
2220                                 btrfs_tree_unlock(mid);
2221                                 free_extent_buffer(mid);
2222                         } else {
2223                                 btrfs_tree_unlock(right);
2224                                 free_extent_buffer(right);
2225                         }
2226                         return 0;
2227                 }
2228                 btrfs_tree_unlock(right);
2229                 free_extent_buffer(right);
2230         }
2231         return 1;
2232 }
2233
2234 /*
2235  * readahead one full node of leaves, finding things that are close
2236  * to the block in 'slot', and triggering ra on them.
2237  */
2238 static void reada_for_search(struct btrfs_fs_info *fs_info,
2239                              struct btrfs_path *path,
2240                              int level, int slot, u64 objectid)
2241 {
2242         struct extent_buffer *node;
2243         struct btrfs_disk_key disk_key;
2244         u32 nritems;
2245         u64 search;
2246         u64 target;
2247         u64 nread = 0;
2248         struct extent_buffer *eb;
2249         u32 nr;
2250         u32 blocksize;
2251         u32 nscan = 0;
2252
2253         if (level != 1)
2254                 return;
2255
2256         if (!path->nodes[level])
2257                 return;
2258
2259         node = path->nodes[level];
2260
2261         search = btrfs_node_blockptr(node, slot);
2262         blocksize = fs_info->nodesize;
2263         eb = find_extent_buffer(fs_info, search);
2264         if (eb) {
2265                 free_extent_buffer(eb);
2266                 return;
2267         }
2268
2269         target = search;
2270
2271         nritems = btrfs_header_nritems(node);
2272         nr = slot;
2273
2274         while (1) {
2275                 if (path->reada == READA_BACK) {
2276                         if (nr == 0)
2277                                 break;
2278                         nr--;
2279                 } else if (path->reada == READA_FORWARD) {
2280                         nr++;
2281                         if (nr >= nritems)
2282                                 break;
2283                 }
2284                 if (path->reada == READA_BACK && objectid) {
2285                         btrfs_node_key(node, &disk_key, nr);
2286                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2287                                 break;
2288                 }
2289                 search = btrfs_node_blockptr(node, nr);
2290                 if ((search <= target && target - search <= 65536) ||
2291                     (search > target && search - target <= 65536)) {
2292                         readahead_tree_block(fs_info, search);
2293                         nread += blocksize;
2294                 }
2295                 nscan++;
2296                 if ((nread > 65536 || nscan > 32))
2297                         break;
2298         }
2299 }
2300
2301 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2302                                        struct btrfs_path *path, int level)
2303 {
2304         int slot;
2305         int nritems;
2306         struct extent_buffer *parent;
2307         struct extent_buffer *eb;
2308         u64 gen;
2309         u64 block1 = 0;
2310         u64 block2 = 0;
2311
2312         parent = path->nodes[level + 1];
2313         if (!parent)
2314                 return;
2315
2316         nritems = btrfs_header_nritems(parent);
2317         slot = path->slots[level + 1];
2318
2319         if (slot > 0) {
2320                 block1 = btrfs_node_blockptr(parent, slot - 1);
2321                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2322                 eb = find_extent_buffer(fs_info, block1);
2323                 /*
2324                  * if we get -eagain from btrfs_buffer_uptodate, we
2325                  * don't want to return eagain here.  That will loop
2326                  * forever
2327                  */
2328                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2329                         block1 = 0;
2330                 free_extent_buffer(eb);
2331         }
2332         if (slot + 1 < nritems) {
2333                 block2 = btrfs_node_blockptr(parent, slot + 1);
2334                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2335                 eb = find_extent_buffer(fs_info, block2);
2336                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2337                         block2 = 0;
2338                 free_extent_buffer(eb);
2339         }
2340
2341         if (block1)
2342                 readahead_tree_block(fs_info, block1);
2343         if (block2)
2344                 readahead_tree_block(fs_info, block2);
2345 }
2346
2347
2348 /*
2349  * when we walk down the tree, it is usually safe to unlock the higher layers
2350  * in the tree.  The exceptions are when our path goes through slot 0, because
2351  * operations on the tree might require changing key pointers higher up in the
2352  * tree.
2353  *
2354  * callers might also have set path->keep_locks, which tells this code to keep
2355  * the lock if the path points to the last slot in the block.  This is part of
2356  * walking through the tree, and selecting the next slot in the higher block.
2357  *
2358  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2359  * if lowest_unlock is 1, level 0 won't be unlocked
2360  */
2361 static noinline void unlock_up(struct btrfs_path *path, int level,
2362                                int lowest_unlock, int min_write_lock_level,
2363                                int *write_lock_level)
2364 {
2365         int i;
2366         int skip_level = level;
2367         int no_skips = 0;
2368         struct extent_buffer *t;
2369
2370         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2371                 if (!path->nodes[i])
2372                         break;
2373                 if (!path->locks[i])
2374                         break;
2375                 if (!no_skips && path->slots[i] == 0) {
2376                         skip_level = i + 1;
2377                         continue;
2378                 }
2379                 if (!no_skips && path->keep_locks) {
2380                         u32 nritems;
2381                         t = path->nodes[i];
2382                         nritems = btrfs_header_nritems(t);
2383                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2384                                 skip_level = i + 1;
2385                                 continue;
2386                         }
2387                 }
2388                 if (skip_level < i && i >= lowest_unlock)
2389                         no_skips = 1;
2390
2391                 t = path->nodes[i];
2392                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2393                         btrfs_tree_unlock_rw(t, path->locks[i]);
2394                         path->locks[i] = 0;
2395                         if (write_lock_level &&
2396                             i > min_write_lock_level &&
2397                             i <= *write_lock_level) {
2398                                 *write_lock_level = i - 1;
2399                         }
2400                 }
2401         }
2402 }
2403
2404 /*
2405  * This releases any locks held in the path starting at level and
2406  * going all the way up to the root.
2407  *
2408  * btrfs_search_slot will keep the lock held on higher nodes in a few
2409  * corner cases, such as COW of the block at slot zero in the node.  This
2410  * ignores those rules, and it should only be called when there are no
2411  * more updates to be done higher up in the tree.
2412  */
2413 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2414 {
2415         int i;
2416
2417         if (path->keep_locks)
2418                 return;
2419
2420         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2421                 if (!path->nodes[i])
2422                         continue;
2423                 if (!path->locks[i])
2424                         continue;
2425                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2426                 path->locks[i] = 0;
2427         }
2428 }
2429
2430 /*
2431  * helper function for btrfs_search_slot.  The goal is to find a block
2432  * in cache without setting the path to blocking.  If we find the block
2433  * we return zero and the path is unchanged.
2434  *
2435  * If we can't find the block, we set the path blocking and do some
2436  * reada.  -EAGAIN is returned and the search must be repeated.
2437  */
2438 static int
2439 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2440                       struct extent_buffer **eb_ret, int level, int slot,
2441                       const struct btrfs_key *key)
2442 {
2443         struct btrfs_fs_info *fs_info = root->fs_info;
2444         u64 blocknr;
2445         u64 gen;
2446         struct extent_buffer *b = *eb_ret;
2447         struct extent_buffer *tmp;
2448         int ret;
2449
2450         blocknr = btrfs_node_blockptr(b, slot);
2451         gen = btrfs_node_ptr_generation(b, slot);
2452
2453         tmp = find_extent_buffer(fs_info, blocknr);
2454         if (tmp) {
2455                 /* first we do an atomic uptodate check */
2456                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2457                         *eb_ret = tmp;
2458                         return 0;
2459                 }
2460
2461                 /* the pages were up to date, but we failed
2462                  * the generation number check.  Do a full
2463                  * read for the generation number that is correct.
2464                  * We must do this without dropping locks so
2465                  * we can trust our generation number
2466                  */
2467                 btrfs_set_path_blocking(p);
2468
2469                 /* now we're allowed to do a blocking uptodate check */
2470                 ret = btrfs_read_buffer(tmp, gen);
2471                 if (!ret) {
2472                         *eb_ret = tmp;
2473                         return 0;
2474                 }
2475                 free_extent_buffer(tmp);
2476                 btrfs_release_path(p);
2477                 return -EIO;
2478         }
2479
2480         /*
2481          * reduce lock contention at high levels
2482          * of the btree by dropping locks before
2483          * we read.  Don't release the lock on the current
2484          * level because we need to walk this node to figure
2485          * out which blocks to read.
2486          */
2487         btrfs_unlock_up_safe(p, level + 1);
2488         btrfs_set_path_blocking(p);
2489
2490         free_extent_buffer(tmp);
2491         if (p->reada != READA_NONE)
2492                 reada_for_search(fs_info, p, level, slot, key->objectid);
2493
2494         btrfs_release_path(p);
2495
2496         ret = -EAGAIN;
2497         tmp = read_tree_block(fs_info, blocknr, 0);
2498         if (!IS_ERR(tmp)) {
2499                 /*
2500                  * If the read above didn't mark this buffer up to date,
2501                  * it will never end up being up to date.  Set ret to EIO now
2502                  * and give up so that our caller doesn't loop forever
2503                  * on our EAGAINs.
2504                  */
2505                 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2506                         ret = -EIO;
2507                 free_extent_buffer(tmp);
2508         } else {
2509                 ret = PTR_ERR(tmp);
2510         }
2511         return ret;
2512 }
2513
2514 /*
2515  * helper function for btrfs_search_slot.  This does all of the checks
2516  * for node-level blocks and does any balancing required based on
2517  * the ins_len.
2518  *
2519  * If no extra work was required, zero is returned.  If we had to
2520  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2521  * start over
2522  */
2523 static int
2524 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2525                        struct btrfs_root *root, struct btrfs_path *p,
2526                        struct extent_buffer *b, int level, int ins_len,
2527                        int *write_lock_level)
2528 {
2529         struct btrfs_fs_info *fs_info = root->fs_info;
2530         int ret;
2531
2532         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2533             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2534                 int sret;
2535
2536                 if (*write_lock_level < level + 1) {
2537                         *write_lock_level = level + 1;
2538                         btrfs_release_path(p);
2539                         goto again;
2540                 }
2541
2542                 btrfs_set_path_blocking(p);
2543                 reada_for_balance(fs_info, p, level);
2544                 sret = split_node(trans, root, p, level);
2545                 btrfs_clear_path_blocking(p, NULL, 0);
2546
2547                 BUG_ON(sret > 0);
2548                 if (sret) {
2549                         ret = sret;
2550                         goto done;
2551                 }
2552                 b = p->nodes[level];
2553         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2554                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2555                 int sret;
2556
2557                 if (*write_lock_level < level + 1) {
2558                         *write_lock_level = level + 1;
2559                         btrfs_release_path(p);
2560                         goto again;
2561                 }
2562
2563                 btrfs_set_path_blocking(p);
2564                 reada_for_balance(fs_info, p, level);
2565                 sret = balance_level(trans, root, p, level);
2566                 btrfs_clear_path_blocking(p, NULL, 0);
2567
2568                 if (sret) {
2569                         ret = sret;
2570                         goto done;
2571                 }
2572                 b = p->nodes[level];
2573                 if (!b) {
2574                         btrfs_release_path(p);
2575                         goto again;
2576                 }
2577                 BUG_ON(btrfs_header_nritems(b) == 1);
2578         }
2579         return 0;
2580
2581 again:
2582         ret = -EAGAIN;
2583 done:
2584         return ret;
2585 }
2586
2587 static void key_search_validate(struct extent_buffer *b,
2588                                 const struct btrfs_key *key,
2589                                 int level)
2590 {
2591 #ifdef CONFIG_BTRFS_ASSERT
2592         struct btrfs_disk_key disk_key;
2593
2594         btrfs_cpu_key_to_disk(&disk_key, key);
2595
2596         if (level == 0)
2597                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2598                     offsetof(struct btrfs_leaf, items[0].key),
2599                     sizeof(disk_key)));
2600         else
2601                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2602                     offsetof(struct btrfs_node, ptrs[0].key),
2603                     sizeof(disk_key)));
2604 #endif
2605 }
2606
2607 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2608                       int level, int *prev_cmp, int *slot)
2609 {
2610         if (*prev_cmp != 0) {
2611                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2612                 return *prev_cmp;
2613         }
2614
2615         key_search_validate(b, key, level);
2616         *slot = 0;
2617
2618         return 0;
2619 }
2620
2621 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2622                 u64 iobjectid, u64 ioff, u8 key_type,
2623                 struct btrfs_key *found_key)
2624 {
2625         int ret;
2626         struct btrfs_key key;
2627         struct extent_buffer *eb;
2628
2629         ASSERT(path);
2630         ASSERT(found_key);
2631
2632         key.type = key_type;
2633         key.objectid = iobjectid;
2634         key.offset = ioff;
2635
2636         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2637         if (ret < 0)
2638                 return ret;
2639
2640         eb = path->nodes[0];
2641         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2642                 ret = btrfs_next_leaf(fs_root, path);
2643                 if (ret)
2644                         return ret;
2645                 eb = path->nodes[0];
2646         }
2647
2648         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2649         if (found_key->type != key.type ||
2650                         found_key->objectid != key.objectid)
2651                 return 1;
2652
2653         return 0;
2654 }
2655
2656 /*
2657  * btrfs_search_slot - look for a key in a tree and perform necessary
2658  * modifications to preserve tree invariants.
2659  *
2660  * @trans:      Handle of transaction, used when modifying the tree
2661  * @p:          Holds all btree nodes along the search path
2662  * @root:       The root node of the tree
2663  * @key:        The key we are looking for
2664  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2665  *              deletions it's -1. 0 for plain searches
2666  * @cow:        boolean should CoW operations be performed. Must always be 1
2667  *              when modifying the tree.
2668  *
2669  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2670  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2671  *
2672  * If @key is found, 0 is returned and you can find the item in the leaf level
2673  * of the path (level 0)
2674  *
2675  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2676  * points to the slot where it should be inserted
2677  *
2678  * If an error is encountered while searching the tree a negative error number
2679  * is returned
2680  */
2681 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2682                       const struct btrfs_key *key, struct btrfs_path *p,
2683                       int ins_len, int cow)
2684 {
2685         struct btrfs_fs_info *fs_info = root->fs_info;
2686         struct extent_buffer *b;
2687         int slot;
2688         int ret;
2689         int err;
2690         int level;
2691         int lowest_unlock = 1;
2692         int root_lock;
2693         /* everything at write_lock_level or lower must be write locked */
2694         int write_lock_level = 0;
2695         u8 lowest_level = 0;
2696         int min_write_lock_level;
2697         int prev_cmp;
2698
2699         lowest_level = p->lowest_level;
2700         WARN_ON(lowest_level && ins_len > 0);
2701         WARN_ON(p->nodes[0] != NULL);
2702         BUG_ON(!cow && ins_len);
2703
2704         if (ins_len < 0) {
2705                 lowest_unlock = 2;
2706
2707                 /* when we are removing items, we might have to go up to level
2708                  * two as we update tree pointers  Make sure we keep write
2709                  * for those levels as well
2710                  */
2711                 write_lock_level = 2;
2712         } else if (ins_len > 0) {
2713                 /*
2714                  * for inserting items, make sure we have a write lock on
2715                  * level 1 so we can update keys
2716                  */
2717                 write_lock_level = 1;
2718         }
2719
2720         if (!cow)
2721                 write_lock_level = -1;
2722
2723         if (cow && (p->keep_locks || p->lowest_level))
2724                 write_lock_level = BTRFS_MAX_LEVEL;
2725
2726         min_write_lock_level = write_lock_level;
2727
2728 again:
2729         prev_cmp = -1;
2730         /*
2731          * we try very hard to do read locks on the root
2732          */
2733         root_lock = BTRFS_READ_LOCK;
2734         level = 0;
2735         if (p->search_commit_root) {
2736                 /*
2737                  * the commit roots are read only
2738                  * so we always do read locks
2739                  */
2740                 if (p->need_commit_sem)
2741                         down_read(&fs_info->commit_root_sem);
2742                 b = root->commit_root;
2743                 extent_buffer_get(b);
2744                 level = btrfs_header_level(b);
2745                 if (p->need_commit_sem)
2746                         up_read(&fs_info->commit_root_sem);
2747                 if (!p->skip_locking)
2748                         btrfs_tree_read_lock(b);
2749         } else {
2750                 if (p->skip_locking) {
2751                         b = btrfs_root_node(root);
2752                         level = btrfs_header_level(b);
2753                 } else {
2754                         /* we don't know the level of the root node
2755                          * until we actually have it read locked
2756                          */
2757                         b = btrfs_read_lock_root_node(root);
2758                         level = btrfs_header_level(b);
2759                         if (level <= write_lock_level) {
2760                                 /* whoops, must trade for write lock */
2761                                 btrfs_tree_read_unlock(b);
2762                                 free_extent_buffer(b);
2763                                 b = btrfs_lock_root_node(root);
2764                                 root_lock = BTRFS_WRITE_LOCK;
2765
2766                                 /* the level might have changed, check again */
2767                                 level = btrfs_header_level(b);
2768                         }
2769                 }
2770         }
2771         p->nodes[level] = b;
2772         if (!p->skip_locking)
2773                 p->locks[level] = root_lock;
2774
2775         while (b) {
2776                 level = btrfs_header_level(b);
2777
2778                 /*
2779                  * setup the path here so we can release it under lock
2780                  * contention with the cow code
2781                  */
2782                 if (cow) {
2783                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2784
2785                         /*
2786                          * if we don't really need to cow this block
2787                          * then we don't want to set the path blocking,
2788                          * so we test it here
2789                          */
2790                         if (!should_cow_block(trans, root, b)) {
2791                                 trans->dirty = true;
2792                                 goto cow_done;
2793                         }
2794
2795                         /*
2796                          * must have write locks on this node and the
2797                          * parent
2798                          */
2799                         if (level > write_lock_level ||
2800                             (level + 1 > write_lock_level &&
2801                             level + 1 < BTRFS_MAX_LEVEL &&
2802                             p->nodes[level + 1])) {
2803                                 write_lock_level = level + 1;
2804                                 btrfs_release_path(p);
2805                                 goto again;
2806                         }
2807
2808                         btrfs_set_path_blocking(p);
2809                         if (last_level)
2810                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2811                                                       &b);
2812                         else
2813                                 err = btrfs_cow_block(trans, root, b,
2814                                                       p->nodes[level + 1],
2815                                                       p->slots[level + 1], &b);
2816                         if (err) {
2817                                 ret = err;
2818                                 goto done;
2819                         }
2820                 }
2821 cow_done:
2822                 p->nodes[level] = b;
2823                 btrfs_clear_path_blocking(p, NULL, 0);
2824
2825                 /*
2826                  * we have a lock on b and as long as we aren't changing
2827                  * the tree, there is no way to for the items in b to change.
2828                  * It is safe to drop the lock on our parent before we
2829                  * go through the expensive btree search on b.
2830                  *
2831                  * If we're inserting or deleting (ins_len != 0), then we might
2832                  * be changing slot zero, which may require changing the parent.
2833                  * So, we can't drop the lock until after we know which slot
2834                  * we're operating on.
2835                  */
2836                 if (!ins_len && !p->keep_locks) {
2837                         int u = level + 1;
2838
2839                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2840                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2841                                 p->locks[u] = 0;
2842                         }
2843                 }
2844
2845                 ret = key_search(b, key, level, &prev_cmp, &slot);
2846                 if (ret < 0)
2847                         goto done;
2848
2849                 if (level != 0) {
2850                         int dec = 0;
2851                         if (ret && slot > 0) {
2852                                 dec = 1;
2853                                 slot -= 1;
2854                         }
2855                         p->slots[level] = slot;
2856                         err = setup_nodes_for_search(trans, root, p, b, level,
2857                                              ins_len, &write_lock_level);
2858                         if (err == -EAGAIN)
2859                                 goto again;
2860                         if (err) {
2861                                 ret = err;
2862                                 goto done;
2863                         }
2864                         b = p->nodes[level];
2865                         slot = p->slots[level];
2866
2867                         /*
2868                          * slot 0 is special, if we change the key
2869                          * we have to update the parent pointer
2870                          * which means we must have a write lock
2871                          * on the parent
2872                          */
2873                         if (slot == 0 && ins_len &&
2874                             write_lock_level < level + 1) {
2875                                 write_lock_level = level + 1;
2876                                 btrfs_release_path(p);
2877                                 goto again;
2878                         }
2879
2880                         unlock_up(p, level, lowest_unlock,
2881                                   min_write_lock_level, &write_lock_level);
2882
2883                         if (level == lowest_level) {
2884                                 if (dec)
2885                                         p->slots[level]++;
2886                                 goto done;
2887                         }
2888
2889                         err = read_block_for_search(root, p, &b, level,
2890                                                     slot, key);
2891                         if (err == -EAGAIN)
2892                                 goto again;
2893                         if (err) {
2894                                 ret = err;
2895                                 goto done;
2896                         }
2897
2898                         if (!p->skip_locking) {
2899                                 level = btrfs_header_level(b);
2900                                 if (level <= write_lock_level) {
2901                                         err = btrfs_try_tree_write_lock(b);
2902                                         if (!err) {
2903                                                 btrfs_set_path_blocking(p);
2904                                                 btrfs_tree_lock(b);
2905                                                 btrfs_clear_path_blocking(p, b,
2906                                                                   BTRFS_WRITE_LOCK);
2907                                         }
2908                                         p->locks[level] = BTRFS_WRITE_LOCK;
2909                                 } else {
2910                                         err = btrfs_tree_read_lock_atomic(b);
2911                                         if (!err) {
2912                                                 btrfs_set_path_blocking(p);
2913                                                 btrfs_tree_read_lock(b);
2914                                                 btrfs_clear_path_blocking(p, b,
2915                                                                   BTRFS_READ_LOCK);
2916                                         }
2917                                         p->locks[level] = BTRFS_READ_LOCK;
2918                                 }
2919                                 p->nodes[level] = b;
2920                         }
2921                 } else {
2922                         p->slots[level] = slot;
2923                         if (ins_len > 0 &&
2924                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2925                                 if (write_lock_level < 1) {
2926                                         write_lock_level = 1;
2927                                         btrfs_release_path(p);
2928                                         goto again;
2929                                 }
2930
2931                                 btrfs_set_path_blocking(p);
2932                                 err = split_leaf(trans, root, key,
2933                                                  p, ins_len, ret == 0);
2934                                 btrfs_clear_path_blocking(p, NULL, 0);
2935
2936                                 BUG_ON(err > 0);
2937                                 if (err) {
2938                                         ret = err;
2939                                         goto done;
2940                                 }
2941                         }
2942                         if (!p->search_for_split)
2943                                 unlock_up(p, level, lowest_unlock,
2944                                           min_write_lock_level, &write_lock_level);
2945                         goto done;
2946                 }
2947         }
2948         ret = 1;
2949 done:
2950         /*
2951          * we don't really know what they plan on doing with the path
2952          * from here on, so for now just mark it as blocking
2953          */
2954         if (!p->leave_spinning)
2955                 btrfs_set_path_blocking(p);
2956         if (ret < 0 && !p->skip_release_on_error)
2957                 btrfs_release_path(p);
2958         return ret;
2959 }
2960
2961 /*
2962  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2963  * current state of the tree together with the operations recorded in the tree
2964  * modification log to search for the key in a previous version of this tree, as
2965  * denoted by the time_seq parameter.
2966  *
2967  * Naturally, there is no support for insert, delete or cow operations.
2968  *
2969  * The resulting path and return value will be set up as if we called
2970  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2971  */
2972 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2973                           struct btrfs_path *p, u64 time_seq)
2974 {
2975         struct btrfs_fs_info *fs_info = root->fs_info;
2976         struct extent_buffer *b;
2977         int slot;
2978         int ret;
2979         int err;
2980         int level;
2981         int lowest_unlock = 1;
2982         u8 lowest_level = 0;
2983         int prev_cmp = -1;
2984
2985         lowest_level = p->lowest_level;
2986         WARN_ON(p->nodes[0] != NULL);
2987
2988         if (p->search_commit_root) {
2989                 BUG_ON(time_seq);
2990                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2991         }
2992
2993 again:
2994         b = get_old_root(root, time_seq);
2995         level = btrfs_header_level(b);
2996         p->locks[level] = BTRFS_READ_LOCK;
2997
2998         while (b) {
2999                 level = btrfs_header_level(b);
3000                 p->nodes[level] = b;
3001                 btrfs_clear_path_blocking(p, NULL, 0);
3002
3003                 /*
3004                  * we have a lock on b and as long as we aren't changing
3005                  * the tree, there is no way to for the items in b to change.
3006                  * It is safe to drop the lock on our parent before we
3007                  * go through the expensive btree search on b.
3008                  */
3009                 btrfs_unlock_up_safe(p, level + 1);
3010
3011                 /*
3012                  * Since we can unwind ebs we want to do a real search every
3013                  * time.
3014                  */
3015                 prev_cmp = -1;
3016                 ret = key_search(b, key, level, &prev_cmp, &slot);
3017
3018                 if (level != 0) {
3019                         int dec = 0;
3020                         if (ret && slot > 0) {
3021                                 dec = 1;
3022                                 slot -= 1;
3023                         }
3024                         p->slots[level] = slot;
3025                         unlock_up(p, level, lowest_unlock, 0, NULL);
3026
3027                         if (level == lowest_level) {
3028                                 if (dec)
3029                                         p->slots[level]++;
3030                                 goto done;
3031                         }
3032
3033                         err = read_block_for_search(root, p, &b, level,
3034                                                     slot, key);
3035                         if (err == -EAGAIN)
3036                                 goto again;
3037                         if (err) {
3038                                 ret = err;
3039                                 goto done;
3040                         }
3041
3042                         level = btrfs_header_level(b);
3043                         err = btrfs_tree_read_lock_atomic(b);
3044                         if (!err) {
3045                                 btrfs_set_path_blocking(p);
3046                                 btrfs_tree_read_lock(b);
3047                                 btrfs_clear_path_blocking(p, b,
3048                                                           BTRFS_READ_LOCK);
3049                         }
3050                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3051                         if (!b) {
3052                                 ret = -ENOMEM;
3053                                 goto done;
3054                         }
3055                         p->locks[level] = BTRFS_READ_LOCK;
3056                         p->nodes[level] = b;
3057                 } else {
3058                         p->slots[level] = slot;
3059                         unlock_up(p, level, lowest_unlock, 0, NULL);
3060                         goto done;
3061                 }
3062         }
3063         ret = 1;
3064 done:
3065         if (!p->leave_spinning)
3066                 btrfs_set_path_blocking(p);
3067         if (ret < 0)
3068                 btrfs_release_path(p);
3069
3070         return ret;
3071 }
3072
3073 /*
3074  * helper to use instead of search slot if no exact match is needed but
3075  * instead the next or previous item should be returned.
3076  * When find_higher is true, the next higher item is returned, the next lower
3077  * otherwise.
3078  * When return_any and find_higher are both true, and no higher item is found,
3079  * return the next lower instead.
3080  * When return_any is true and find_higher is false, and no lower item is found,
3081  * return the next higher instead.
3082  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3083  * < 0 on error
3084  */
3085 int btrfs_search_slot_for_read(struct btrfs_root *root,
3086                                const struct btrfs_key *key,
3087                                struct btrfs_path *p, int find_higher,
3088                                int return_any)
3089 {
3090         int ret;
3091         struct extent_buffer *leaf;
3092
3093 again:
3094         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3095         if (ret <= 0)
3096                 return ret;
3097         /*
3098          * a return value of 1 means the path is at the position where the
3099          * item should be inserted. Normally this is the next bigger item,
3100          * but in case the previous item is the last in a leaf, path points
3101          * to the first free slot in the previous leaf, i.e. at an invalid
3102          * item.
3103          */
3104         leaf = p->nodes[0];
3105
3106         if (find_higher) {
3107                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3108                         ret = btrfs_next_leaf(root, p);
3109                         if (ret <= 0)
3110                                 return ret;
3111                         if (!return_any)
3112                                 return 1;
3113                         /*
3114                          * no higher item found, return the next
3115                          * lower instead
3116                          */
3117                         return_any = 0;
3118                         find_higher = 0;
3119                         btrfs_release_path(p);
3120                         goto again;
3121                 }
3122         } else {
3123                 if (p->slots[0] == 0) {
3124                         ret = btrfs_prev_leaf(root, p);
3125                         if (ret < 0)
3126                                 return ret;
3127                         if (!ret) {
3128                                 leaf = p->nodes[0];
3129                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3130                                         p->slots[0]--;
3131                                 return 0;
3132                         }
3133                         if (!return_any)
3134                                 return 1;
3135                         /*
3136                          * no lower item found, return the next
3137                          * higher instead
3138                          */