Merge tag 'upstream-4.8-rc1' of git://git.infradead.org/linux-ubifs
[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/vmalloc.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
32                       *root, struct btrfs_key *ins_key,
33                       struct btrfs_path *path, int data_size, int extend);
34 static int push_node_left(struct btrfs_trans_handle *trans,
35                           struct btrfs_root *root, struct extent_buffer *dst,
36                           struct extent_buffer *src, int empty);
37 static int balance_node_right(struct btrfs_trans_handle *trans,
38                               struct btrfs_root *root,
39                               struct extent_buffer *dst_buf,
40                               struct extent_buffer *src_buf);
41 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
42                     int level, int slot);
43 static int tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
44                                  struct extent_buffer *eb);
45
46 struct btrfs_path *btrfs_alloc_path(void)
47 {
48         struct btrfs_path *path;
49         path = kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
50         return path;
51 }
52
53 /*
54  * set all locked nodes in the path to blocking locks.  This should
55  * be done before scheduling
56  */
57 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
58 {
59         int i;
60         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
61                 if (!p->nodes[i] || !p->locks[i])
62                         continue;
63                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
64                 if (p->locks[i] == BTRFS_READ_LOCK)
65                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
66                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
67                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
68         }
69 }
70
71 /*
72  * reset all the locked nodes in the patch to spinning locks.
73  *
74  * held is used to keep lockdep happy, when lockdep is enabled
75  * we set held to a blocking lock before we go around and
76  * retake all the spinlocks in the path.  You can safely use NULL
77  * for held
78  */
79 noinline void btrfs_clear_path_blocking(struct btrfs_path *p,
80                                         struct extent_buffer *held, int held_rw)
81 {
82         int i;
83
84         if (held) {
85                 btrfs_set_lock_blocking_rw(held, held_rw);
86                 if (held_rw == BTRFS_WRITE_LOCK)
87                         held_rw = BTRFS_WRITE_LOCK_BLOCKING;
88                 else if (held_rw == BTRFS_READ_LOCK)
89                         held_rw = BTRFS_READ_LOCK_BLOCKING;
90         }
91         btrfs_set_path_blocking(p);
92
93         for (i = BTRFS_MAX_LEVEL - 1; i >= 0; i--) {
94                 if (p->nodes[i] && p->locks[i]) {
95                         btrfs_clear_lock_blocking_rw(p->nodes[i], p->locks[i]);
96                         if (p->locks[i] == BTRFS_WRITE_LOCK_BLOCKING)
97                                 p->locks[i] = BTRFS_WRITE_LOCK;
98                         else if (p->locks[i] == BTRFS_READ_LOCK_BLOCKING)
99                                 p->locks[i] = BTRFS_READ_LOCK;
100                 }
101         }
102
103         if (held)
104                 btrfs_clear_lock_blocking_rw(held, held_rw);
105 }
106
107 /* this also releases the path */
108 void btrfs_free_path(struct btrfs_path *p)
109 {
110         if (!p)
111                 return;
112         btrfs_release_path(p);
113         kmem_cache_free(btrfs_path_cachep, p);
114 }
115
116 /*
117  * path release drops references on the extent buffers in the path
118  * and it drops any locks held by this path
119  *
120  * It is safe to call this on paths that no locks or extent buffers held.
121  */
122 noinline void btrfs_release_path(struct btrfs_path *p)
123 {
124         int i;
125
126         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
127                 p->slots[i] = 0;
128                 if (!p->nodes[i])
129                         continue;
130                 if (p->locks[i]) {
131                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
132                         p->locks[i] = 0;
133                 }
134                 free_extent_buffer(p->nodes[i]);
135                 p->nodes[i] = NULL;
136         }
137 }
138
139 /*
140  * safely gets a reference on the root node of a tree.  A lock
141  * is not taken, so a concurrent writer may put a different node
142  * at the root of the tree.  See btrfs_lock_root_node for the
143  * looping required.
144  *
145  * The extent buffer returned by this has a reference taken, so
146  * it won't disappear.  It may stop being the root of the tree
147  * at any time because there are no locks held.
148  */
149 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
150 {
151         struct extent_buffer *eb;
152
153         while (1) {
154                 rcu_read_lock();
155                 eb = rcu_dereference(root->node);
156
157                 /*
158                  * RCU really hurts here, we could free up the root node because
159                  * it was COWed but we may not get the new root node yet so do
160                  * the inc_not_zero dance and if it doesn't work then
161                  * synchronize_rcu and try again.
162                  */
163                 if (atomic_inc_not_zero(&eb->refs)) {
164                         rcu_read_unlock();
165                         break;
166                 }
167                 rcu_read_unlock();
168                 synchronize_rcu();
169         }
170         return eb;
171 }
172
173 /* loop around taking references on and locking the root node of the
174  * tree until you end up with a lock on the root.  A locked buffer
175  * is returned, with a reference held.
176  */
177 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
178 {
179         struct extent_buffer *eb;
180
181         while (1) {
182                 eb = btrfs_root_node(root);
183                 btrfs_tree_lock(eb);
184                 if (eb == root->node)
185                         break;
186                 btrfs_tree_unlock(eb);
187                 free_extent_buffer(eb);
188         }
189         return eb;
190 }
191
192 /* loop around taking references on and locking the root node of the
193  * tree until you end up with a lock on the root.  A locked buffer
194  * is returned, with a reference held.
195  */
196 static struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
197 {
198         struct extent_buffer *eb;
199
200         while (1) {
201                 eb = btrfs_root_node(root);
202                 btrfs_tree_read_lock(eb);
203                 if (eb == root->node)
204                         break;
205                 btrfs_tree_read_unlock(eb);
206                 free_extent_buffer(eb);
207         }
208         return eb;
209 }
210
211 /* cowonly root (everything not a reference counted cow subvolume), just get
212  * put onto a simple dirty list.  transaction.c walks this to make sure they
213  * get properly updated on disk.
214  */
215 static void add_root_to_dirty_list(struct btrfs_root *root)
216 {
217         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
218             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
219                 return;
220
221         spin_lock(&root->fs_info->trans_lock);
222         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
223                 /* Want the extent tree to be the last on the list */
224                 if (root->objectid == BTRFS_EXTENT_TREE_OBJECTID)
225                         list_move_tail(&root->dirty_list,
226                                        &root->fs_info->dirty_cowonly_roots);
227                 else
228                         list_move(&root->dirty_list,
229                                   &root->fs_info->dirty_cowonly_roots);
230         }
231         spin_unlock(&root->fs_info->trans_lock);
232 }
233
234 /*
235  * used by snapshot creation to make a copy of a root for a tree with
236  * a given objectid.  The buffer with the new root node is returned in
237  * cow_ret, and this func returns zero on success or a negative error code.
238  */
239 int btrfs_copy_root(struct btrfs_trans_handle *trans,
240                       struct btrfs_root *root,
241                       struct extent_buffer *buf,
242                       struct extent_buffer **cow_ret, u64 new_root_objectid)
243 {
244         struct extent_buffer *cow;
245         int ret = 0;
246         int level;
247         struct btrfs_disk_key disk_key;
248
249         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
250                 trans->transid != root->fs_info->running_transaction->transid);
251         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
252                 trans->transid != root->last_trans);
253
254         level = btrfs_header_level(buf);
255         if (level == 0)
256                 btrfs_item_key(buf, &disk_key, 0);
257         else
258                 btrfs_node_key(buf, &disk_key, 0);
259
260         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
261                         &disk_key, level, buf->start, 0);
262         if (IS_ERR(cow))
263                 return PTR_ERR(cow);
264
265         copy_extent_buffer(cow, buf, 0, 0, cow->len);
266         btrfs_set_header_bytenr(cow, cow->start);
267         btrfs_set_header_generation(cow, trans->transid);
268         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
269         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
270                                      BTRFS_HEADER_FLAG_RELOC);
271         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
272                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
273         else
274                 btrfs_set_header_owner(cow, new_root_objectid);
275
276         write_extent_buffer(cow, root->fs_info->fsid, btrfs_header_fsid(),
277                             BTRFS_FSID_SIZE);
278
279         WARN_ON(btrfs_header_generation(buf) > trans->transid);
280         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
281                 ret = btrfs_inc_ref(trans, root, cow, 1);
282         else
283                 ret = btrfs_inc_ref(trans, root, cow, 0);
284
285         if (ret)
286                 return ret;
287
288         btrfs_mark_buffer_dirty(cow);
289         *cow_ret = cow;
290         return 0;
291 }
292
293 enum mod_log_op {
294         MOD_LOG_KEY_REPLACE,
295         MOD_LOG_KEY_ADD,
296         MOD_LOG_KEY_REMOVE,
297         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
298         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
299         MOD_LOG_MOVE_KEYS,
300         MOD_LOG_ROOT_REPLACE,
301 };
302
303 struct tree_mod_move {
304         int dst_slot;
305         int nr_items;
306 };
307
308 struct tree_mod_root {
309         u64 logical;
310         u8 level;
311 };
312
313 struct tree_mod_elem {
314         struct rb_node node;
315         u64 logical;
316         u64 seq;
317         enum mod_log_op op;
318
319         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
320         int slot;
321
322         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
323         u64 generation;
324
325         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
326         struct btrfs_disk_key key;
327         u64 blockptr;
328
329         /* this is used for op == MOD_LOG_MOVE_KEYS */
330         struct tree_mod_move move;
331
332         /* this is used for op == MOD_LOG_ROOT_REPLACE */
333         struct tree_mod_root old_root;
334 };
335
336 static inline void tree_mod_log_read_lock(struct btrfs_fs_info *fs_info)
337 {
338         read_lock(&fs_info->tree_mod_log_lock);
339 }
340
341 static inline void tree_mod_log_read_unlock(struct btrfs_fs_info *fs_info)
342 {
343         read_unlock(&fs_info->tree_mod_log_lock);
344 }
345
346 static inline void tree_mod_log_write_lock(struct btrfs_fs_info *fs_info)
347 {
348         write_lock(&fs_info->tree_mod_log_lock);
349 }
350
351 static inline void tree_mod_log_write_unlock(struct btrfs_fs_info *fs_info)
352 {
353         write_unlock(&fs_info->tree_mod_log_lock);
354 }
355
356 /*
357  * Pull a new tree mod seq number for our operation.
358  */
359 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
360 {
361         return atomic64_inc_return(&fs_info->tree_mod_seq);
362 }
363
364 /*
365  * This adds a new blocker to the tree mod log's blocker list if the @elem
366  * passed does not already have a sequence number set. So when a caller expects
367  * to record tree modifications, it should ensure to set elem->seq to zero
368  * before calling btrfs_get_tree_mod_seq.
369  * Returns a fresh, unused tree log modification sequence number, even if no new
370  * blocker was added.
371  */
372 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
373                            struct seq_list *elem)
374 {
375         tree_mod_log_write_lock(fs_info);
376         spin_lock(&fs_info->tree_mod_seq_lock);
377         if (!elem->seq) {
378                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
379                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
380         }
381         spin_unlock(&fs_info->tree_mod_seq_lock);
382         tree_mod_log_write_unlock(fs_info);
383
384         return elem->seq;
385 }
386
387 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
388                             struct seq_list *elem)
389 {
390         struct rb_root *tm_root;
391         struct rb_node *node;
392         struct rb_node *next;
393         struct seq_list *cur_elem;
394         struct tree_mod_elem *tm;
395         u64 min_seq = (u64)-1;
396         u64 seq_putting = elem->seq;
397
398         if (!seq_putting)
399                 return;
400
401         spin_lock(&fs_info->tree_mod_seq_lock);
402         list_del(&elem->list);
403         elem->seq = 0;
404
405         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
406                 if (cur_elem->seq < min_seq) {
407                         if (seq_putting > cur_elem->seq) {
408                                 /*
409                                  * blocker with lower sequence number exists, we
410                                  * cannot remove anything from the log
411                                  */
412                                 spin_unlock(&fs_info->tree_mod_seq_lock);
413                                 return;
414                         }
415                         min_seq = cur_elem->seq;
416                 }
417         }
418         spin_unlock(&fs_info->tree_mod_seq_lock);
419
420         /*
421          * anything that's lower than the lowest existing (read: blocked)
422          * sequence number can be removed from the tree.
423          */
424         tree_mod_log_write_lock(fs_info);
425         tm_root = &fs_info->tree_mod_log;
426         for (node = rb_first(tm_root); node; node = next) {
427                 next = rb_next(node);
428                 tm = container_of(node, struct tree_mod_elem, node);
429                 if (tm->seq > min_seq)
430                         continue;
431                 rb_erase(node, tm_root);
432                 kfree(tm);
433         }
434         tree_mod_log_write_unlock(fs_info);
435 }
436
437 /*
438  * key order of the log:
439  *       node/leaf start address -> sequence
440  *
441  * The 'start address' is the logical address of the *new* root node
442  * for root replace operations, or the logical address of the affected
443  * block for all other operations.
444  *
445  * Note: must be called with write lock (tree_mod_log_write_lock).
446  */
447 static noinline int
448 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
449 {
450         struct rb_root *tm_root;
451         struct rb_node **new;
452         struct rb_node *parent = NULL;
453         struct tree_mod_elem *cur;
454
455         BUG_ON(!tm);
456
457         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
458
459         tm_root = &fs_info->tree_mod_log;
460         new = &tm_root->rb_node;
461         while (*new) {
462                 cur = container_of(*new, struct tree_mod_elem, node);
463                 parent = *new;
464                 if (cur->logical < tm->logical)
465                         new = &((*new)->rb_left);
466                 else if (cur->logical > tm->logical)
467                         new = &((*new)->rb_right);
468                 else if (cur->seq < tm->seq)
469                         new = &((*new)->rb_left);
470                 else if (cur->seq > tm->seq)
471                         new = &((*new)->rb_right);
472                 else
473                         return -EEXIST;
474         }
475
476         rb_link_node(&tm->node, parent, new);
477         rb_insert_color(&tm->node, tm_root);
478         return 0;
479 }
480
481 /*
482  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
483  * returns zero with the tree_mod_log_lock acquired. The caller must hold
484  * this until all tree mod log insertions are recorded in the rb tree and then
485  * call tree_mod_log_write_unlock() to release.
486  */
487 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
488                                     struct extent_buffer *eb) {
489         smp_mb();
490         if (list_empty(&(fs_info)->tree_mod_seq_list))
491                 return 1;
492         if (eb && btrfs_header_level(eb) == 0)
493                 return 1;
494
495         tree_mod_log_write_lock(fs_info);
496         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
497                 tree_mod_log_write_unlock(fs_info);
498                 return 1;
499         }
500
501         return 0;
502 }
503
504 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
505 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
506                                     struct extent_buffer *eb)
507 {
508         smp_mb();
509         if (list_empty(&(fs_info)->tree_mod_seq_list))
510                 return 0;
511         if (eb && btrfs_header_level(eb) == 0)
512                 return 0;
513
514         return 1;
515 }
516
517 static struct tree_mod_elem *
518 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
519                     enum mod_log_op op, gfp_t flags)
520 {
521         struct tree_mod_elem *tm;
522
523         tm = kzalloc(sizeof(*tm), flags);
524         if (!tm)
525                 return NULL;
526
527         tm->logical = eb->start;
528         if (op != MOD_LOG_KEY_ADD) {
529                 btrfs_node_key(eb, &tm->key, slot);
530                 tm->blockptr = btrfs_node_blockptr(eb, slot);
531         }
532         tm->op = op;
533         tm->slot = slot;
534         tm->generation = btrfs_node_ptr_generation(eb, slot);
535         RB_CLEAR_NODE(&tm->node);
536
537         return tm;
538 }
539
540 static noinline int
541 tree_mod_log_insert_key(struct btrfs_fs_info *fs_info,
542                         struct extent_buffer *eb, int slot,
543                         enum mod_log_op op, gfp_t flags)
544 {
545         struct tree_mod_elem *tm;
546         int ret;
547
548         if (!tree_mod_need_log(fs_info, eb))
549                 return 0;
550
551         tm = alloc_tree_mod_elem(eb, slot, op, flags);
552         if (!tm)
553                 return -ENOMEM;
554
555         if (tree_mod_dont_log(fs_info, eb)) {
556                 kfree(tm);
557                 return 0;
558         }
559
560         ret = __tree_mod_log_insert(fs_info, tm);
561         tree_mod_log_write_unlock(fs_info);
562         if (ret)
563                 kfree(tm);
564
565         return ret;
566 }
567
568 static noinline int
569 tree_mod_log_insert_move(struct btrfs_fs_info *fs_info,
570                          struct extent_buffer *eb, int dst_slot, int src_slot,
571                          int nr_items, gfp_t flags)
572 {
573         struct tree_mod_elem *tm = NULL;
574         struct tree_mod_elem **tm_list = NULL;
575         int ret = 0;
576         int i;
577         int locked = 0;
578
579         if (!tree_mod_need_log(fs_info, eb))
580                 return 0;
581
582         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), flags);
583         if (!tm_list)
584                 return -ENOMEM;
585
586         tm = kzalloc(sizeof(*tm), flags);
587         if (!tm) {
588                 ret = -ENOMEM;
589                 goto free_tms;
590         }
591
592         tm->logical = eb->start;
593         tm->slot = src_slot;
594         tm->move.dst_slot = dst_slot;
595         tm->move.nr_items = nr_items;
596         tm->op = MOD_LOG_MOVE_KEYS;
597
598         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
599                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
600                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, flags);
601                 if (!tm_list[i]) {
602                         ret = -ENOMEM;
603                         goto free_tms;
604                 }
605         }
606
607         if (tree_mod_dont_log(fs_info, eb))
608                 goto free_tms;
609         locked = 1;
610
611         /*
612          * When we override something during the move, we log these removals.
613          * This can only happen when we move towards the beginning of the
614          * buffer, i.e. dst_slot < src_slot.
615          */
616         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
617                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
618                 if (ret)
619                         goto free_tms;
620         }
621
622         ret = __tree_mod_log_insert(fs_info, tm);
623         if (ret)
624                 goto free_tms;
625         tree_mod_log_write_unlock(fs_info);
626         kfree(tm_list);
627
628         return 0;
629 free_tms:
630         for (i = 0; i < nr_items; i++) {
631                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
632                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
633                 kfree(tm_list[i]);
634         }
635         if (locked)
636                 tree_mod_log_write_unlock(fs_info);
637         kfree(tm_list);
638         kfree(tm);
639
640         return ret;
641 }
642
643 static inline int
644 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
645                        struct tree_mod_elem **tm_list,
646                        int nritems)
647 {
648         int i, j;
649         int ret;
650
651         for (i = nritems - 1; i >= 0; i--) {
652                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
653                 if (ret) {
654                         for (j = nritems - 1; j > i; j--)
655                                 rb_erase(&tm_list[j]->node,
656                                          &fs_info->tree_mod_log);
657                         return ret;
658                 }
659         }
660
661         return 0;
662 }
663
664 static noinline int
665 tree_mod_log_insert_root(struct btrfs_fs_info *fs_info,
666                          struct extent_buffer *old_root,
667                          struct extent_buffer *new_root, gfp_t flags,
668                          int log_removal)
669 {
670         struct tree_mod_elem *tm = NULL;
671         struct tree_mod_elem **tm_list = NULL;
672         int nritems = 0;
673         int ret = 0;
674         int i;
675
676         if (!tree_mod_need_log(fs_info, NULL))
677                 return 0;
678
679         if (log_removal && btrfs_header_level(old_root) > 0) {
680                 nritems = btrfs_header_nritems(old_root);
681                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
682                                   flags);
683                 if (!tm_list) {
684                         ret = -ENOMEM;
685                         goto free_tms;
686                 }
687                 for (i = 0; i < nritems; i++) {
688                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
689                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, flags);
690                         if (!tm_list[i]) {
691                                 ret = -ENOMEM;
692                                 goto free_tms;
693                         }
694                 }
695         }
696
697         tm = kzalloc(sizeof(*tm), flags);
698         if (!tm) {
699                 ret = -ENOMEM;
700                 goto free_tms;
701         }
702
703         tm->logical = new_root->start;
704         tm->old_root.logical = old_root->start;
705         tm->old_root.level = btrfs_header_level(old_root);
706         tm->generation = btrfs_header_generation(old_root);
707         tm->op = MOD_LOG_ROOT_REPLACE;
708
709         if (tree_mod_dont_log(fs_info, NULL))
710                 goto free_tms;
711
712         if (tm_list)
713                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
714         if (!ret)
715                 ret = __tree_mod_log_insert(fs_info, tm);
716
717         tree_mod_log_write_unlock(fs_info);
718         if (ret)
719                 goto free_tms;
720         kfree(tm_list);
721
722         return ret;
723
724 free_tms:
725         if (tm_list) {
726                 for (i = 0; i < nritems; i++)
727                         kfree(tm_list[i]);
728                 kfree(tm_list);
729         }
730         kfree(tm);
731
732         return ret;
733 }
734
735 static struct tree_mod_elem *
736 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
737                       int smallest)
738 {
739         struct rb_root *tm_root;
740         struct rb_node *node;
741         struct tree_mod_elem *cur = NULL;
742         struct tree_mod_elem *found = NULL;
743
744         tree_mod_log_read_lock(fs_info);
745         tm_root = &fs_info->tree_mod_log;
746         node = tm_root->rb_node;
747         while (node) {
748                 cur = container_of(node, struct tree_mod_elem, node);
749                 if (cur->logical < start) {
750                         node = node->rb_left;
751                 } else if (cur->logical > start) {
752                         node = node->rb_right;
753                 } else if (cur->seq < min_seq) {
754                         node = node->rb_left;
755                 } else if (!smallest) {
756                         /* we want the node with the highest seq */
757                         if (found)
758                                 BUG_ON(found->seq > cur->seq);
759                         found = cur;
760                         node = node->rb_left;
761                 } else if (cur->seq > min_seq) {
762                         /* we want the node with the smallest seq */
763                         if (found)
764                                 BUG_ON(found->seq < cur->seq);
765                         found = cur;
766                         node = node->rb_right;
767                 } else {
768                         found = cur;
769                         break;
770                 }
771         }
772         tree_mod_log_read_unlock(fs_info);
773
774         return found;
775 }
776
777 /*
778  * this returns the element from the log with the smallest time sequence
779  * value that's in the log (the oldest log item). any element with a time
780  * sequence lower than min_seq will be ignored.
781  */
782 static struct tree_mod_elem *
783 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
784                            u64 min_seq)
785 {
786         return __tree_mod_log_search(fs_info, start, min_seq, 1);
787 }
788
789 /*
790  * this returns the element from the log with the largest time sequence
791  * value that's in the log (the most recent log item). any element with
792  * a time sequence lower than min_seq will be ignored.
793  */
794 static struct tree_mod_elem *
795 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
796 {
797         return __tree_mod_log_search(fs_info, start, min_seq, 0);
798 }
799
800 static noinline int
801 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
802                      struct extent_buffer *src, unsigned long dst_offset,
803                      unsigned long src_offset, int nr_items)
804 {
805         int ret = 0;
806         struct tree_mod_elem **tm_list = NULL;
807         struct tree_mod_elem **tm_list_add, **tm_list_rem;
808         int i;
809         int locked = 0;
810
811         if (!tree_mod_need_log(fs_info, NULL))
812                 return 0;
813
814         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
815                 return 0;
816
817         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
818                           GFP_NOFS);
819         if (!tm_list)
820                 return -ENOMEM;
821
822         tm_list_add = tm_list;
823         tm_list_rem = tm_list + nr_items;
824         for (i = 0; i < nr_items; i++) {
825                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
826                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
827                 if (!tm_list_rem[i]) {
828                         ret = -ENOMEM;
829                         goto free_tms;
830                 }
831
832                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
833                     MOD_LOG_KEY_ADD, GFP_NOFS);
834                 if (!tm_list_add[i]) {
835                         ret = -ENOMEM;
836                         goto free_tms;
837                 }
838         }
839
840         if (tree_mod_dont_log(fs_info, NULL))
841                 goto free_tms;
842         locked = 1;
843
844         for (i = 0; i < nr_items; i++) {
845                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
846                 if (ret)
847                         goto free_tms;
848                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
849                 if (ret)
850                         goto free_tms;
851         }
852
853         tree_mod_log_write_unlock(fs_info);
854         kfree(tm_list);
855
856         return 0;
857
858 free_tms:
859         for (i = 0; i < nr_items * 2; i++) {
860                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
861                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
862                 kfree(tm_list[i]);
863         }
864         if (locked)
865                 tree_mod_log_write_unlock(fs_info);
866         kfree(tm_list);
867
868         return ret;
869 }
870
871 static inline void
872 tree_mod_log_eb_move(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
873                      int dst_offset, int src_offset, int nr_items)
874 {
875         int ret;
876         ret = tree_mod_log_insert_move(fs_info, dst, dst_offset, src_offset,
877                                        nr_items, GFP_NOFS);
878         BUG_ON(ret < 0);
879 }
880
881 static noinline void
882 tree_mod_log_set_node_key(struct btrfs_fs_info *fs_info,
883                           struct extent_buffer *eb, int slot, int atomic)
884 {
885         int ret;
886
887         ret = tree_mod_log_insert_key(fs_info, eb, slot,
888                                         MOD_LOG_KEY_REPLACE,
889                                         atomic ? GFP_ATOMIC : GFP_NOFS);
890         BUG_ON(ret < 0);
891 }
892
893 static noinline int
894 tree_mod_log_free_eb(struct btrfs_fs_info *fs_info, struct extent_buffer *eb)
895 {
896         struct tree_mod_elem **tm_list = NULL;
897         int nritems = 0;
898         int i;
899         int ret = 0;
900
901         if (btrfs_header_level(eb) == 0)
902                 return 0;
903
904         if (!tree_mod_need_log(fs_info, NULL))
905                 return 0;
906
907         nritems = btrfs_header_nritems(eb);
908         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
909         if (!tm_list)
910                 return -ENOMEM;
911
912         for (i = 0; i < nritems; i++) {
913                 tm_list[i] = alloc_tree_mod_elem(eb, i,
914                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
915                 if (!tm_list[i]) {
916                         ret = -ENOMEM;
917                         goto free_tms;
918                 }
919         }
920
921         if (tree_mod_dont_log(fs_info, eb))
922                 goto free_tms;
923
924         ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
925         tree_mod_log_write_unlock(fs_info);
926         if (ret)
927                 goto free_tms;
928         kfree(tm_list);
929
930         return 0;
931
932 free_tms:
933         for (i = 0; i < nritems; i++)
934                 kfree(tm_list[i]);
935         kfree(tm_list);
936
937         return ret;
938 }
939
940 static noinline void
941 tree_mod_log_set_root_pointer(struct btrfs_root *root,
942                               struct extent_buffer *new_root_node,
943                               int log_removal)
944 {
945         int ret;
946         ret = tree_mod_log_insert_root(root->fs_info, root->node,
947                                        new_root_node, GFP_NOFS, log_removal);
948         BUG_ON(ret < 0);
949 }
950
951 /*
952  * check if the tree block can be shared by multiple trees
953  */
954 int btrfs_block_can_be_shared(struct btrfs_root *root,
955                               struct extent_buffer *buf)
956 {
957         /*
958          * Tree blocks not in reference counted trees and tree roots
959          * are never shared. If a block was allocated after the last
960          * snapshot and the block was not allocated by tree relocation,
961          * we know the block is not shared.
962          */
963         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
964             buf != root->node && buf != root->commit_root &&
965             (btrfs_header_generation(buf) <=
966              btrfs_root_last_snapshot(&root->root_item) ||
967              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
968                 return 1;
969 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
970         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
971             btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
972                 return 1;
973 #endif
974         return 0;
975 }
976
977 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
978                                        struct btrfs_root *root,
979                                        struct extent_buffer *buf,
980                                        struct extent_buffer *cow,
981                                        int *last_ref)
982 {
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, root, 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(root->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                         BUG_ON(ret); /* -ENOMEM */
1036
1037                         if (root->root_key.objectid ==
1038                             BTRFS_TREE_RELOC_OBJECTID) {
1039                                 ret = btrfs_dec_ref(trans, root, buf, 0);
1040                                 BUG_ON(ret); /* -ENOMEM */
1041                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1042                                 BUG_ON(ret); /* -ENOMEM */
1043                         }
1044                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
1045                 } else {
1046
1047                         if (root->root_key.objectid ==
1048                             BTRFS_TREE_RELOC_OBJECTID)
1049                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1050                         else
1051                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1052                         BUG_ON(ret); /* -ENOMEM */
1053                 }
1054                 if (new_flags != 0) {
1055                         int level = btrfs_header_level(buf);
1056
1057                         ret = btrfs_set_disk_extent_flags(trans, root,
1058                                                           buf->start,
1059                                                           buf->len,
1060                                                           new_flags, level, 0);
1061                         if (ret)
1062                                 return ret;
1063                 }
1064         } else {
1065                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
1066                         if (root->root_key.objectid ==
1067                             BTRFS_TREE_RELOC_OBJECTID)
1068                                 ret = btrfs_inc_ref(trans, root, cow, 1);
1069                         else
1070                                 ret = btrfs_inc_ref(trans, root, cow, 0);
1071                         BUG_ON(ret); /* -ENOMEM */
1072                         ret = btrfs_dec_ref(trans, root, buf, 1);
1073                         BUG_ON(ret); /* -ENOMEM */
1074                 }
1075                 clean_tree_block(trans, root->fs_info, buf);
1076                 *last_ref = 1;
1077         }
1078         return 0;
1079 }
1080
1081 /*
1082  * does the dirty work in cow of a single block.  The parent block (if
1083  * supplied) is updated to point to the new cow copy.  The new buffer is marked
1084  * dirty and returned locked.  If you modify the block it needs to be marked
1085  * dirty again.
1086  *
1087  * search_start -- an allocation hint for the new block
1088  *
1089  * empty_size -- a hint that you plan on doing more cow.  This is the size in
1090  * bytes the allocator should try to find free next to the block it returns.
1091  * This is just a hint and may be ignored by the allocator.
1092  */
1093 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
1094                              struct btrfs_root *root,
1095                              struct extent_buffer *buf,
1096                              struct extent_buffer *parent, int parent_slot,
1097                              struct extent_buffer **cow_ret,
1098                              u64 search_start, u64 empty_size)
1099 {
1100         struct btrfs_disk_key disk_key;
1101         struct extent_buffer *cow;
1102         int level, ret;
1103         int last_ref = 0;
1104         int unlock_orig = 0;
1105         u64 parent_start;
1106
1107         if (*cow_ret == buf)
1108                 unlock_orig = 1;
1109
1110         btrfs_assert_tree_locked(buf);
1111
1112         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1113                 trans->transid != root->fs_info->running_transaction->transid);
1114         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1115                 trans->transid != root->last_trans);
1116
1117         level = btrfs_header_level(buf);
1118
1119         if (level == 0)
1120                 btrfs_item_key(buf, &disk_key, 0);
1121         else
1122                 btrfs_node_key(buf, &disk_key, 0);
1123
1124         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) {
1125                 if (parent)
1126                         parent_start = parent->start;
1127                 else
1128                         parent_start = 0;
1129         } else
1130                 parent_start = 0;
1131
1132         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1133                         root->root_key.objectid, &disk_key, level,
1134                         search_start, empty_size);
1135         if (IS_ERR(cow))
1136                 return PTR_ERR(cow);
1137
1138         /* cow is set to blocking by btrfs_init_new_buffer */
1139
1140         copy_extent_buffer(cow, buf, 0, 0, cow->len);
1141         btrfs_set_header_bytenr(cow, cow->start);
1142         btrfs_set_header_generation(cow, trans->transid);
1143         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1144         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1145                                      BTRFS_HEADER_FLAG_RELOC);
1146         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1147                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1148         else
1149                 btrfs_set_header_owner(cow, root->root_key.objectid);
1150
1151         write_extent_buffer(cow, root->fs_info->fsid, btrfs_header_fsid(),
1152                             BTRFS_FSID_SIZE);
1153
1154         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1155         if (ret) {
1156                 btrfs_abort_transaction(trans, root, ret);
1157                 return ret;
1158         }
1159
1160         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1161                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1162                 if (ret) {
1163                         btrfs_abort_transaction(trans, root, ret);
1164                         return ret;
1165                 }
1166         }
1167
1168         if (buf == root->node) {
1169                 WARN_ON(parent && parent != buf);
1170                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1171                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1172                         parent_start = buf->start;
1173                 else
1174                         parent_start = 0;
1175
1176                 extent_buffer_get(cow);
1177                 tree_mod_log_set_root_pointer(root, cow, 1);
1178                 rcu_assign_pointer(root->node, cow);
1179
1180                 btrfs_free_tree_block(trans, root, buf, parent_start,
1181                                       last_ref);
1182                 free_extent_buffer(buf);
1183                 add_root_to_dirty_list(root);
1184         } else {
1185                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1186                         parent_start = parent->start;
1187                 else
1188                         parent_start = 0;
1189
1190                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1191                 tree_mod_log_insert_key(root->fs_info, parent, parent_slot,
1192                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1193                 btrfs_set_node_blockptr(parent, parent_slot,
1194                                         cow->start);
1195                 btrfs_set_node_ptr_generation(parent, parent_slot,
1196                                               trans->transid);
1197                 btrfs_mark_buffer_dirty(parent);
1198                 if (last_ref) {
1199                         ret = tree_mod_log_free_eb(root->fs_info, buf);
1200                         if (ret) {
1201                                 btrfs_abort_transaction(trans, root, ret);
1202                                 return ret;
1203                         }
1204                 }
1205                 btrfs_free_tree_block(trans, root, buf, parent_start,
1206                                       last_ref);
1207         }
1208         if (unlock_orig)
1209                 btrfs_tree_unlock(buf);
1210         free_extent_buffer_stale(buf);
1211         btrfs_mark_buffer_dirty(cow);
1212         *cow_ret = cow;
1213         return 0;
1214 }
1215
1216 /*
1217  * returns the logical address of the oldest predecessor of the given root.
1218  * entries older than time_seq are ignored.
1219  */
1220 static struct tree_mod_elem *
1221 __tree_mod_log_oldest_root(struct btrfs_fs_info *fs_info,
1222                            struct extent_buffer *eb_root, u64 time_seq)
1223 {
1224         struct tree_mod_elem *tm;
1225         struct tree_mod_elem *found = NULL;
1226         u64 root_logical = eb_root->start;
1227         int looped = 0;
1228
1229         if (!time_seq)
1230                 return NULL;
1231
1232         /*
1233          * the very last operation that's logged for a root is the
1234          * replacement operation (if it is replaced at all). this has
1235          * the logical address of the *new* root, making it the very
1236          * first operation that's logged for this root.
1237          */
1238         while (1) {
1239                 tm = tree_mod_log_search_oldest(fs_info, root_logical,
1240                                                 time_seq);
1241                 if (!looped && !tm)
1242                         return NULL;
1243                 /*
1244                  * if there are no tree operation for the oldest root, we simply
1245                  * return it. this should only happen if that (old) root is at
1246                  * level 0.
1247                  */
1248                 if (!tm)
1249                         break;
1250
1251                 /*
1252                  * if there's an operation that's not a root replacement, we
1253                  * found the oldest version of our root. normally, we'll find a
1254                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1255                  */
1256                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1257                         break;
1258
1259                 found = tm;
1260                 root_logical = tm->old_root.logical;
1261                 looped = 1;
1262         }
1263
1264         /* if there's no old root to return, return what we found instead */
1265         if (!found)
1266                 found = tm;
1267
1268         return found;
1269 }
1270
1271 /*
1272  * tm is a pointer to the first operation to rewind within eb. then, all
1273  * previous operations will be rewound (until we reach something older than
1274  * time_seq).
1275  */
1276 static void
1277 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1278                       u64 time_seq, struct tree_mod_elem *first_tm)
1279 {
1280         u32 n;
1281         struct rb_node *next;
1282         struct tree_mod_elem *tm = first_tm;
1283         unsigned long o_dst;
1284         unsigned long o_src;
1285         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1286
1287         n = btrfs_header_nritems(eb);
1288         tree_mod_log_read_lock(fs_info);
1289         while (tm && tm->seq >= time_seq) {
1290                 /*
1291                  * all the operations are recorded with the operator used for
1292                  * the modification. as we're going backwards, we do the
1293                  * opposite of each operation here.
1294                  */
1295                 switch (tm->op) {
1296                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1297                         BUG_ON(tm->slot < n);
1298                         /* Fallthrough */
1299                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1300                 case MOD_LOG_KEY_REMOVE:
1301                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1302                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1303                         btrfs_set_node_ptr_generation(eb, tm->slot,
1304                                                       tm->generation);
1305                         n++;
1306                         break;
1307                 case MOD_LOG_KEY_REPLACE:
1308                         BUG_ON(tm->slot >= n);
1309                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1310                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1311                         btrfs_set_node_ptr_generation(eb, tm->slot,
1312                                                       tm->generation);
1313                         break;
1314                 case MOD_LOG_KEY_ADD:
1315                         /* if a move operation is needed it's in the log */
1316                         n--;
1317                         break;
1318                 case MOD_LOG_MOVE_KEYS:
1319                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1320                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1321                         memmove_extent_buffer(eb, o_dst, o_src,
1322                                               tm->move.nr_items * p_size);
1323                         break;
1324                 case MOD_LOG_ROOT_REPLACE:
1325                         /*
1326                          * this operation is special. for roots, this must be
1327                          * handled explicitly before rewinding.
1328                          * for non-roots, this operation may exist if the node
1329                          * was a root: root A -> child B; then A gets empty and
1330                          * B is promoted to the new root. in the mod log, we'll
1331                          * have a root-replace operation for B, a tree block
1332                          * that is no root. we simply ignore that operation.
1333                          */
1334                         break;
1335                 }
1336                 next = rb_next(&tm->node);
1337                 if (!next)
1338                         break;
1339                 tm = container_of(next, struct tree_mod_elem, node);
1340                 if (tm->logical != first_tm->logical)
1341                         break;
1342         }
1343         tree_mod_log_read_unlock(fs_info);
1344         btrfs_set_header_nritems(eb, n);
1345 }
1346
1347 /*
1348  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1349  * is returned. If rewind operations happen, a fresh buffer is returned. The
1350  * returned buffer is always read-locked. If the returned buffer is not the
1351  * input buffer, the lock on the input buffer is released and the input buffer
1352  * is freed (its refcount is decremented).
1353  */
1354 static struct extent_buffer *
1355 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1356                     struct extent_buffer *eb, u64 time_seq)
1357 {
1358         struct extent_buffer *eb_rewin;
1359         struct tree_mod_elem *tm;
1360
1361         if (!time_seq)
1362                 return eb;
1363
1364         if (btrfs_header_level(eb) == 0)
1365                 return eb;
1366
1367         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1368         if (!tm)
1369                 return eb;
1370
1371         btrfs_set_path_blocking(path);
1372         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1373
1374         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1375                 BUG_ON(tm->slot != 0);
1376                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start,
1377                                                 eb->len);
1378                 if (!eb_rewin) {
1379                         btrfs_tree_read_unlock_blocking(eb);
1380                         free_extent_buffer(eb);
1381                         return NULL;
1382                 }
1383                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1384                 btrfs_set_header_backref_rev(eb_rewin,
1385                                              btrfs_header_backref_rev(eb));
1386                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1387                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1388         } else {
1389                 eb_rewin = btrfs_clone_extent_buffer(eb);
1390                 if (!eb_rewin) {
1391                         btrfs_tree_read_unlock_blocking(eb);
1392                         free_extent_buffer(eb);
1393                         return NULL;
1394                 }
1395         }
1396
1397         btrfs_clear_path_blocking(path, NULL, BTRFS_READ_LOCK);
1398         btrfs_tree_read_unlock_blocking(eb);
1399         free_extent_buffer(eb);
1400
1401         extent_buffer_get(eb_rewin);
1402         btrfs_tree_read_lock(eb_rewin);
1403         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1404         WARN_ON(btrfs_header_nritems(eb_rewin) >
1405                 BTRFS_NODEPTRS_PER_BLOCK(fs_info->tree_root));
1406
1407         return eb_rewin;
1408 }
1409
1410 /*
1411  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1412  * value. If there are no changes, the current root->root_node is returned. If
1413  * anything changed in between, there's a fresh buffer allocated on which the
1414  * rewind operations are done. In any case, the returned buffer is read locked.
1415  * Returns NULL on error (with no locks held).
1416  */
1417 static inline struct extent_buffer *
1418 get_old_root(struct btrfs_root *root, u64 time_seq)
1419 {
1420         struct tree_mod_elem *tm;
1421         struct extent_buffer *eb = NULL;
1422         struct extent_buffer *eb_root;
1423         struct extent_buffer *old;
1424         struct tree_mod_root *old_root = NULL;
1425         u64 old_generation = 0;
1426         u64 logical;
1427
1428         eb_root = btrfs_read_lock_root_node(root);
1429         tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1430         if (!tm)
1431                 return eb_root;
1432
1433         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1434                 old_root = &tm->old_root;
1435                 old_generation = tm->generation;
1436                 logical = old_root->logical;
1437         } else {
1438                 logical = eb_root->start;
1439         }
1440
1441         tm = tree_mod_log_search(root->fs_info, logical, time_seq);
1442         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1443                 btrfs_tree_read_unlock(eb_root);
1444                 free_extent_buffer(eb_root);
1445                 old = read_tree_block(root, logical, 0);
1446                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1447                         if (!IS_ERR(old))
1448                                 free_extent_buffer(old);
1449                         btrfs_warn(root->fs_info,
1450                                 "failed to read tree block %llu from get_old_root", logical);
1451                 } else {
1452                         eb = btrfs_clone_extent_buffer(old);
1453                         free_extent_buffer(old);
1454                 }
1455         } else if (old_root) {
1456                 btrfs_tree_read_unlock(eb_root);
1457                 free_extent_buffer(eb_root);
1458                 eb = alloc_dummy_extent_buffer(root->fs_info, logical,
1459                                         root->nodesize);
1460         } else {
1461                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1462                 eb = btrfs_clone_extent_buffer(eb_root);
1463                 btrfs_tree_read_unlock_blocking(eb_root);
1464                 free_extent_buffer(eb_root);
1465         }
1466
1467         if (!eb)
1468                 return NULL;
1469         extent_buffer_get(eb);
1470         btrfs_tree_read_lock(eb);
1471         if (old_root) {
1472                 btrfs_set_header_bytenr(eb, eb->start);
1473                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1474                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1475                 btrfs_set_header_level(eb, old_root->level);
1476                 btrfs_set_header_generation(eb, old_generation);
1477         }
1478         if (tm)
1479                 __tree_mod_log_rewind(root->fs_info, eb, time_seq, tm);
1480         else
1481                 WARN_ON(btrfs_header_level(eb) != 0);
1482         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(root));
1483
1484         return eb;
1485 }
1486
1487 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1488 {
1489         struct tree_mod_elem *tm;
1490         int level;
1491         struct extent_buffer *eb_root = btrfs_root_node(root);
1492
1493         tm = __tree_mod_log_oldest_root(root->fs_info, eb_root, time_seq);
1494         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1495                 level = tm->old_root.level;
1496         } else {
1497                 level = btrfs_header_level(eb_root);
1498         }
1499         free_extent_buffer(eb_root);
1500
1501         return level;
1502 }
1503
1504 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1505                                    struct btrfs_root *root,
1506                                    struct extent_buffer *buf)
1507 {
1508         if (btrfs_test_is_dummy_root(root))
1509                 return 0;
1510
1511         /* ensure we can see the force_cow */
1512         smp_rmb();
1513
1514         /*
1515          * We do not need to cow a block if
1516          * 1) this block is not created or changed in this transaction;
1517          * 2) this block does not belong to TREE_RELOC tree;
1518          * 3) the root is not forced COW.
1519          *
1520          * What is forced COW:
1521          *    when we create snapshot during committing the transaction,
1522          *    after we've finished coping src root, we must COW the shared
1523          *    block to ensure the metadata consistency.
1524          */
1525         if (btrfs_header_generation(buf) == trans->transid &&
1526             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1527             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1528               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1529             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1530                 return 0;
1531         return 1;
1532 }
1533
1534 /*
1535  * cows a single block, see __btrfs_cow_block for the real work.
1536  * This version of it has extra checks so that a block isn't COWed more than
1537  * once per transaction, as long as it hasn't been written yet
1538  */
1539 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1540                     struct btrfs_root *root, struct extent_buffer *buf,
1541                     struct extent_buffer *parent, int parent_slot,
1542                     struct extent_buffer **cow_ret)
1543 {
1544         u64 search_start;
1545         int ret;
1546
1547         if (trans->transaction != root->fs_info->running_transaction)
1548                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1549                        trans->transid,
1550                        root->fs_info->running_transaction->transid);
1551
1552         if (trans->transid != root->fs_info->generation)
1553                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1554                        trans->transid, root->fs_info->generation);
1555
1556         if (!should_cow_block(trans, root, buf)) {
1557                 trans->dirty = true;
1558                 *cow_ret = buf;
1559                 return 0;
1560         }
1561
1562         search_start = buf->start & ~((u64)SZ_1G - 1);
1563
1564         if (parent)
1565                 btrfs_set_lock_blocking(parent);
1566         btrfs_set_lock_blocking(buf);
1567
1568         ret = __btrfs_cow_block(trans, root, buf, parent,
1569                                  parent_slot, cow_ret, search_start, 0);
1570
1571         trace_btrfs_cow_block(root, buf, *cow_ret);
1572
1573         return ret;
1574 }
1575
1576 /*
1577  * helper function for defrag to decide if two blocks pointed to by a
1578  * node are actually close by
1579  */
1580 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1581 {
1582         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1583                 return 1;
1584         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1585                 return 1;
1586         return 0;
1587 }
1588
1589 /*
1590  * compare two keys in a memcmp fashion
1591  */
1592 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
1593 {
1594         struct btrfs_key k1;
1595
1596         btrfs_disk_key_to_cpu(&k1, disk);
1597
1598         return btrfs_comp_cpu_keys(&k1, k2);
1599 }
1600
1601 /*
1602  * same as comp_keys only with two btrfs_key's
1603  */
1604 int btrfs_comp_cpu_keys(struct btrfs_key *k1, struct btrfs_key *k2)
1605 {
1606         if (k1->objectid > k2->objectid)
1607                 return 1;
1608         if (k1->objectid < k2->objectid)
1609                 return -1;
1610         if (k1->type > k2->type)
1611                 return 1;
1612         if (k1->type < k2->type)
1613                 return -1;
1614         if (k1->offset > k2->offset)
1615                 return 1;
1616         if (k1->offset < k2->offset)
1617                 return -1;
1618         return 0;
1619 }
1620
1621 /*
1622  * this is used by the defrag code to go through all the
1623  * leaves pointed to by a node and reallocate them so that
1624  * disk order is close to key order
1625  */
1626 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1627                        struct btrfs_root *root, struct extent_buffer *parent,
1628                        int start_slot, u64 *last_ret,
1629                        struct btrfs_key *progress)
1630 {
1631         struct extent_buffer *cur;
1632         u64 blocknr;
1633         u64 gen;
1634         u64 search_start = *last_ret;
1635         u64 last_block = 0;
1636         u64 other;
1637         u32 parent_nritems;
1638         int end_slot;
1639         int i;
1640         int err = 0;
1641         int parent_level;
1642         int uptodate;
1643         u32 blocksize;
1644         int progress_passed = 0;
1645         struct btrfs_disk_key disk_key;
1646
1647         parent_level = btrfs_header_level(parent);
1648
1649         WARN_ON(trans->transaction != root->fs_info->running_transaction);
1650         WARN_ON(trans->transid != root->fs_info->generation);
1651
1652         parent_nritems = btrfs_header_nritems(parent);
1653         blocksize = root->nodesize;
1654         end_slot = parent_nritems - 1;
1655
1656         if (parent_nritems <= 1)
1657                 return 0;
1658
1659         btrfs_set_lock_blocking(parent);
1660
1661         for (i = start_slot; i <= end_slot; i++) {
1662                 int close = 1;
1663
1664                 btrfs_node_key(parent, &disk_key, i);
1665                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1666                         continue;
1667
1668                 progress_passed = 1;
1669                 blocknr = btrfs_node_blockptr(parent, i);
1670                 gen = btrfs_node_ptr_generation(parent, i);
1671                 if (last_block == 0)
1672                         last_block = blocknr;
1673
1674                 if (i > 0) {
1675                         other = btrfs_node_blockptr(parent, i - 1);
1676                         close = close_blocks(blocknr, other, blocksize);
1677                 }
1678                 if (!close && i < end_slot) {
1679                         other = btrfs_node_blockptr(parent, i + 1);
1680                         close = close_blocks(blocknr, other, blocksize);
1681                 }
1682                 if (close) {
1683                         last_block = blocknr;
1684                         continue;
1685                 }
1686
1687                 cur = btrfs_find_tree_block(root->fs_info, blocknr);
1688                 if (cur)
1689                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1690                 else
1691                         uptodate = 0;
1692                 if (!cur || !uptodate) {
1693                         if (!cur) {
1694                                 cur = read_tree_block(root, blocknr, gen);
1695                                 if (IS_ERR(cur)) {
1696                                         return PTR_ERR(cur);
1697                                 } else if (!extent_buffer_uptodate(cur)) {
1698                                         free_extent_buffer(cur);
1699                                         return -EIO;
1700                                 }
1701                         } else if (!uptodate) {
1702                                 err = btrfs_read_buffer(cur, gen);
1703                                 if (err) {
1704                                         free_extent_buffer(cur);
1705                                         return err;
1706                                 }
1707                         }
1708                 }
1709                 if (search_start == 0)
1710                         search_start = last_block;
1711
1712                 btrfs_tree_lock(cur);
1713                 btrfs_set_lock_blocking(cur);
1714                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1715                                         &cur, search_start,
1716                                         min(16 * blocksize,
1717                                             (end_slot - i) * blocksize));
1718                 if (err) {
1719                         btrfs_tree_unlock(cur);
1720                         free_extent_buffer(cur);
1721                         break;
1722                 }
1723                 search_start = cur->start;
1724                 last_block = cur->start;
1725                 *last_ret = search_start;
1726                 btrfs_tree_unlock(cur);
1727                 free_extent_buffer(cur);
1728         }
1729         return err;
1730 }
1731
1732 /*
1733  * The leaf data grows from end-to-front in the node.
1734  * this returns the address of the start of the last item,
1735  * which is the stop of the leaf data stack
1736  */
1737 static inline unsigned int leaf_data_end(struct btrfs_root *root,
1738                                          struct extent_buffer *leaf)
1739 {
1740         u32 nr = btrfs_header_nritems(leaf);
1741         if (nr == 0)
1742                 return BTRFS_LEAF_DATA_SIZE(root);
1743         return btrfs_item_offset_nr(leaf, nr - 1);
1744 }
1745
1746
1747 /*
1748  * search for key in the extent_buffer.  The items start at offset p,
1749  * and they are item_size apart.  There are 'max' items in p.
1750  *
1751  * the slot in the array is returned via slot, and it points to
1752  * the place where you would insert key if it is not found in
1753  * the array.
1754  *
1755  * slot may point to max if the key is bigger than all of the keys
1756  */
1757 static noinline int generic_bin_search(struct extent_buffer *eb,
1758                                        unsigned long p,
1759                                        int item_size, struct btrfs_key *key,
1760                                        int max, int *slot)
1761 {
1762         int low = 0;
1763         int high = max;
1764         int mid;
1765         int ret;
1766         struct btrfs_disk_key *tmp = NULL;
1767         struct btrfs_disk_key unaligned;
1768         unsigned long offset;
1769         char *kaddr = NULL;
1770         unsigned long map_start = 0;
1771         unsigned long map_len = 0;
1772         int err;
1773
1774         while (low < high) {
1775                 mid = (low + high) / 2;
1776                 offset = p + mid * item_size;
1777
1778                 if (!kaddr || offset < map_start ||
1779                     (offset + sizeof(struct btrfs_disk_key)) >
1780                     map_start + map_len) {
1781
1782                         err = map_private_extent_buffer(eb, offset,
1783                                                 sizeof(struct btrfs_disk_key),
1784                                                 &kaddr, &map_start, &map_len);
1785
1786                         if (!err) {
1787                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1788                                                         map_start);
1789                         } else if (err == 1) {
1790                                 read_extent_buffer(eb, &unaligned,
1791                                                    offset, sizeof(unaligned));
1792                                 tmp = &unaligned;
1793                         } else {
1794                                 return err;
1795                         }
1796
1797                 } else {
1798                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1799                                                         map_start);
1800                 }
1801                 ret = comp_keys(tmp, key);
1802
1803                 if (ret < 0)
1804                         low = mid + 1;
1805                 else if (ret > 0)
1806                         high = mid;
1807                 else {
1808                         *slot = mid;
1809                         return 0;
1810                 }
1811         }
1812         *slot = low;
1813         return 1;
1814 }
1815
1816 /*
1817  * simple bin_search frontend that does the right thing for
1818  * leaves vs nodes
1819  */
1820 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1821                       int level, int *slot)
1822 {
1823         if (level == 0)
1824                 return generic_bin_search(eb,
1825                                           offsetof(struct btrfs_leaf, items),
1826                                           sizeof(struct btrfs_item),
1827                                           key, btrfs_header_nritems(eb),
1828                                           slot);
1829         else
1830                 return generic_bin_search(eb,
1831                                           offsetof(struct btrfs_node, ptrs),
1832                                           sizeof(struct btrfs_key_ptr),
1833                                           key, btrfs_header_nritems(eb),
1834                                           slot);
1835 }
1836
1837 int btrfs_bin_search(struct extent_buffer *eb, struct btrfs_key *key,
1838                      int level, int *slot)
1839 {
1840         return bin_search(eb, key, level, slot);
1841 }
1842
1843 static void root_add_used(struct btrfs_root *root, u32 size)
1844 {
1845         spin_lock(&root->accounting_lock);
1846         btrfs_set_root_used(&root->root_item,
1847                             btrfs_root_used(&root->root_item) + size);
1848         spin_unlock(&root->accounting_lock);
1849 }
1850
1851 static void root_sub_used(struct btrfs_root *root, u32 size)
1852 {
1853         spin_lock(&root->accounting_lock);
1854         btrfs_set_root_used(&root->root_item,
1855                             btrfs_root_used(&root->root_item) - size);
1856         spin_unlock(&root->accounting_lock);
1857 }
1858
1859 /* given a node and slot number, this reads the blocks it points to.  The
1860  * extent buffer is returned with a reference taken (but unlocked).
1861  * NULL is returned on error.
1862  */
1863 static noinline struct extent_buffer *read_node_slot(struct btrfs_root *root,
1864                                    struct extent_buffer *parent, int slot)
1865 {
1866         int level = btrfs_header_level(parent);
1867         struct extent_buffer *eb;
1868
1869         if (slot < 0)
1870                 return NULL;
1871         if (slot >= btrfs_header_nritems(parent))
1872                 return NULL;
1873
1874         BUG_ON(level == 0);
1875
1876         eb = read_tree_block(root, btrfs_node_blockptr(parent, slot),
1877                              btrfs_node_ptr_generation(parent, slot));
1878         if (IS_ERR(eb) || !extent_buffer_uptodate(eb)) {
1879                 if (!IS_ERR(eb))
1880                         free_extent_buffer(eb);
1881                 eb = NULL;
1882         }
1883
1884         return eb;
1885 }
1886
1887 /*
1888  * node level balancing, used to make sure nodes are in proper order for
1889  * item deletion.  We balance from the top down, so we have to make sure
1890  * that a deletion won't leave an node completely empty later on.
1891  */
1892 static noinline int balance_level(struct btrfs_trans_handle *trans,
1893                          struct btrfs_root *root,
1894                          struct btrfs_path *path, int level)
1895 {
1896         struct extent_buffer *right = NULL;
1897         struct extent_buffer *mid;
1898         struct extent_buffer *left = NULL;
1899         struct extent_buffer *parent = NULL;
1900         int ret = 0;
1901         int wret;
1902         int pslot;
1903         int orig_slot = path->slots[level];
1904         u64 orig_ptr;
1905
1906         if (level == 0)
1907                 return 0;
1908
1909         mid = path->nodes[level];
1910
1911         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1912                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1913         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1914
1915         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1916
1917         if (level < BTRFS_MAX_LEVEL - 1) {
1918                 parent = path->nodes[level + 1];
1919                 pslot = path->slots[level + 1];
1920         }
1921
1922         /*
1923          * deal with the case where there is only one pointer in the root
1924          * by promoting the node below to a root
1925          */
1926         if (!parent) {
1927                 struct extent_buffer *child;
1928
1929                 if (btrfs_header_nritems(mid) != 1)
1930                         return 0;
1931
1932                 /* promote the child to a root */
1933                 child = read_node_slot(root, mid, 0);
1934                 if (!child) {
1935                         ret = -EROFS;
1936                         btrfs_handle_fs_error(root->fs_info, ret, NULL);
1937                         goto enospc;
1938                 }
1939
1940                 btrfs_tree_lock(child);
1941                 btrfs_set_lock_blocking(child);
1942                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1943                 if (ret) {
1944                         btrfs_tree_unlock(child);
1945                         free_extent_buffer(child);
1946                         goto enospc;
1947                 }
1948
1949                 tree_mod_log_set_root_pointer(root, child, 1);
1950                 rcu_assign_pointer(root->node, child);
1951
1952                 add_root_to_dirty_list(root);
1953                 btrfs_tree_unlock(child);
1954
1955                 path->locks[level] = 0;
1956                 path->nodes[level] = NULL;
1957                 clean_tree_block(trans, root->fs_info, mid);
1958                 btrfs_tree_unlock(mid);
1959                 /* once for the path */
1960                 free_extent_buffer(mid);
1961
1962                 root_sub_used(root, mid->len);
1963                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1964                 /* once for the root ptr */
1965                 free_extent_buffer_stale(mid);
1966                 return 0;
1967         }
1968         if (btrfs_header_nritems(mid) >
1969             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
1970                 return 0;
1971
1972         left = read_node_slot(root, parent, pslot - 1);
1973         if (left) {
1974                 btrfs_tree_lock(left);
1975                 btrfs_set_lock_blocking(left);
1976                 wret = btrfs_cow_block(trans, root, left,
1977                                        parent, pslot - 1, &left);
1978                 if (wret) {
1979                         ret = wret;
1980                         goto enospc;
1981                 }
1982         }
1983         right = read_node_slot(root, parent, pslot + 1);
1984         if (right) {
1985                 btrfs_tree_lock(right);
1986                 btrfs_set_lock_blocking(right);
1987                 wret = btrfs_cow_block(trans, root, right,
1988                                        parent, pslot + 1, &right);
1989                 if (wret) {
1990                         ret = wret;
1991                         goto enospc;
1992                 }
1993         }
1994
1995         /* first, try to make some room in the middle buffer */
1996         if (left) {
1997                 orig_slot += btrfs_header_nritems(left);
1998                 wret = push_node_left(trans, root, left, mid, 1);
1999                 if (wret < 0)
2000                         ret = wret;
2001         }
2002
2003         /*
2004          * then try to empty the right most buffer into the middle
2005          */
2006         if (right) {
2007                 wret = push_node_left(trans, root, mid, right, 1);
2008                 if (wret < 0 && wret != -ENOSPC)
2009                         ret = wret;
2010                 if (btrfs_header_nritems(right) == 0) {
2011                         clean_tree_block(trans, root->fs_info, right);
2012                         btrfs_tree_unlock(right);
2013                         del_ptr(root, path, level + 1, pslot + 1);
2014                         root_sub_used(root, right->len);
2015                         btrfs_free_tree_block(trans, root, right, 0, 1);
2016                         free_extent_buffer_stale(right);
2017                         right = NULL;
2018                 } else {
2019                         struct btrfs_disk_key right_key;
2020                         btrfs_node_key(right, &right_key, 0);
2021                         tree_mod_log_set_node_key(root->fs_info, parent,
2022                                                   pslot + 1, 0);
2023                         btrfs_set_node_key(parent, &right_key, pslot + 1);
2024                         btrfs_mark_buffer_dirty(parent);
2025                 }
2026         }
2027         if (btrfs_header_nritems(mid) == 1) {
2028                 /*
2029                  * we're not allowed to leave a node with one item in the
2030                  * tree during a delete.  A deletion from lower in the tree
2031                  * could try to delete the only pointer in this node.
2032                  * So, pull some keys from the left.
2033                  * There has to be a left pointer at this point because
2034                  * otherwise we would have pulled some pointers from the
2035                  * right
2036                  */
2037                 if (!left) {
2038                         ret = -EROFS;
2039                         btrfs_handle_fs_error(root->fs_info, ret, NULL);
2040                         goto enospc;
2041                 }
2042                 wret = balance_node_right(trans, root, mid, left);
2043                 if (wret < 0) {
2044                         ret = wret;
2045                         goto enospc;
2046                 }
2047                 if (wret == 1) {
2048                         wret = push_node_left(trans, root, left, mid, 1);
2049                         if (wret < 0)
2050                                 ret = wret;
2051                 }
2052                 BUG_ON(wret == 1);
2053         }
2054         if (btrfs_header_nritems(mid) == 0) {
2055                 clean_tree_block(trans, root->fs_info, mid);
2056                 btrfs_tree_unlock(mid);
2057                 del_ptr(root, path, level + 1, pslot);
2058                 root_sub_used(root, mid->len);
2059                 btrfs_free_tree_block(trans, root, mid, 0, 1);
2060                 free_extent_buffer_stale(mid);
2061                 mid = NULL;
2062         } else {
2063                 /* update the parent key to reflect our changes */
2064                 struct btrfs_disk_key mid_key;
2065                 btrfs_node_key(mid, &mid_key, 0);
2066                 tree_mod_log_set_node_key(root->fs_info, parent,
2067                                           pslot, 0);
2068                 btrfs_set_node_key(parent, &mid_key, pslot);
2069                 btrfs_mark_buffer_dirty(parent);
2070         }
2071
2072         /* update the path */
2073         if (left) {
2074                 if (btrfs_header_nritems(left) > orig_slot) {
2075                         extent_buffer_get(left);
2076                         /* left was locked after cow */
2077                         path->nodes[level] = left;
2078                         path->slots[level + 1] -= 1;
2079                         path->slots[level] = orig_slot;
2080                         if (mid) {
2081                                 btrfs_tree_unlock(mid);
2082                                 free_extent_buffer(mid);
2083                         }
2084                 } else {
2085                         orig_slot -= btrfs_header_nritems(left);
2086                         path->slots[level] = orig_slot;
2087                 }
2088         }
2089         /* double check we haven't messed things up */
2090         if (orig_ptr !=
2091             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
2092                 BUG();
2093 enospc:
2094         if (right) {
2095                 btrfs_tree_unlock(right);
2096                 free_extent_buffer(right);
2097         }
2098         if (left) {
2099                 if (path->nodes[level] != left)
2100                         btrfs_tree_unlock(left);
2101                 free_extent_buffer(left);
2102         }
2103         return ret;
2104 }
2105
2106 /* Node balancing for insertion.  Here we only split or push nodes around
2107  * when they are completely full.  This is also done top down, so we
2108  * have to be pessimistic.
2109  */
2110 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2111                                           struct btrfs_root *root,
2112                                           struct btrfs_path *path, int level)
2113 {
2114         struct extent_buffer *right = NULL;
2115         struct extent_buffer *mid;
2116         struct extent_buffer *left = NULL;
2117         struct extent_buffer *parent = NULL;
2118         int ret = 0;
2119         int wret;
2120         int pslot;
2121         int orig_slot = path->slots[level];
2122
2123         if (level == 0)
2124                 return 1;
2125
2126         mid = path->nodes[level];
2127         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2128
2129         if (level < BTRFS_MAX_LEVEL - 1) {
2130                 parent = path->nodes[level + 1];
2131                 pslot = path->slots[level + 1];
2132         }
2133
2134         if (!parent)
2135                 return 1;
2136
2137         left = read_node_slot(root, parent, pslot - 1);
2138
2139         /* first, try to make some room in the middle buffer */
2140         if (left) {
2141                 u32 left_nr;
2142
2143                 btrfs_tree_lock(left);
2144                 btrfs_set_lock_blocking(left);
2145
2146                 left_nr = btrfs_header_nritems(left);
2147                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
2148                         wret = 1;
2149                 } else {
2150                         ret = btrfs_cow_block(trans, root, left, parent,
2151                                               pslot - 1, &left);
2152                         if (ret)
2153                                 wret = 1;
2154                         else {
2155                                 wret = push_node_left(trans, root,
2156                                                       left, mid, 0);
2157                         }
2158                 }
2159                 if (wret < 0)
2160                         ret = wret;
2161                 if (wret == 0) {
2162                         struct btrfs_disk_key disk_key;
2163                         orig_slot += left_nr;
2164                         btrfs_node_key(mid, &disk_key, 0);
2165                         tree_mod_log_set_node_key(root->fs_info, parent,
2166                                                   pslot, 0);
2167                         btrfs_set_node_key(parent, &disk_key, pslot);
2168                         btrfs_mark_buffer_dirty(parent);
2169                         if (btrfs_header_nritems(left) > orig_slot) {
2170                                 path->nodes[level] = left;
2171                                 path->slots[level + 1] -= 1;
2172                                 path->slots[level] = orig_slot;
2173                                 btrfs_tree_unlock(mid);
2174                                 free_extent_buffer(mid);
2175                         } else {
2176                                 orig_slot -=
2177                                         btrfs_header_nritems(left);
2178                                 path->slots[level] = orig_slot;
2179                                 btrfs_tree_unlock(left);
2180                                 free_extent_buffer(left);
2181                         }
2182                         return 0;
2183                 }
2184                 btrfs_tree_unlock(left);
2185                 free_extent_buffer(left);
2186         }
2187         right = read_node_slot(root, parent, pslot + 1);
2188
2189         /*
2190          * then try to empty the right most buffer into the middle
2191          */
2192         if (right) {
2193                 u32 right_nr;
2194
2195                 btrfs_tree_lock(right);
2196                 btrfs_set_lock_blocking(right);
2197
2198                 right_nr = btrfs_header_nritems(right);
2199                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
2200                         wret = 1;
2201                 } else {
2202                         ret = btrfs_cow_block(trans, root, right,
2203                                               parent, pslot + 1,
2204                                               &right);
2205                         if (ret)
2206                                 wret = 1;
2207                         else {
2208                                 wret = balance_node_right(trans, root,
2209                                                           right, mid);
2210                         }
2211                 }
2212                 if (wret < 0)
2213                         ret = wret;
2214                 if (wret == 0) {
2215                         struct btrfs_disk_key disk_key;
2216
2217                         btrfs_node_key(right, &disk_key, 0);
2218                         tree_mod_log_set_node_key(root->fs_info, parent,
2219                                                   pslot + 1, 0);
2220                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2221                         btrfs_mark_buffer_dirty(parent);
2222
2223                         if (btrfs_header_nritems(mid) <= orig_slot) {
2224                                 path->nodes[level] = right;
2225                                 path->slots[level + 1] += 1;
2226                                 path->slots[level] = orig_slot -
2227                                         btrfs_header_nritems(mid);
2228                                 btrfs_tree_unlock(mid);
2229                                 free_extent_buffer(mid);
2230                         } else {
2231                                 btrfs_tree_unlock(right);
2232                                 free_extent_buffer(right);
2233                         }
2234                         return 0;
2235                 }
2236                 btrfs_tree_unlock(right);
2237                 free_extent_buffer(right);
2238         }
2239         return 1;
2240 }
2241
2242 /*
2243  * readahead one full node of leaves, finding things that are close
2244  * to the block in 'slot', and triggering ra on them.
2245  */
2246 static void reada_for_search(struct btrfs_root *root,
2247                              struct btrfs_path *path,
2248                              int level, int slot, u64 objectid)
2249 {
2250         struct extent_buffer *node;
2251         struct btrfs_disk_key disk_key;
2252         u32 nritems;
2253         u64 search;
2254         u64 target;
2255         u64 nread = 0;
2256         u64 gen;
2257         struct extent_buffer *eb;
2258         u32 nr;
2259         u32 blocksize;
2260         u32 nscan = 0;
2261
2262         if (level != 1)
2263                 return;
2264
2265         if (!path->nodes[level])
2266                 return;
2267
2268         node = path->nodes[level];
2269
2270         search = btrfs_node_blockptr(node, slot);
2271         blocksize = root->nodesize;
2272         eb = btrfs_find_tree_block(root->fs_info, search);
2273         if (eb) {
2274                 free_extent_buffer(eb);
2275                 return;
2276         }
2277
2278         target = search;
2279
2280         nritems = btrfs_header_nritems(node);
2281         nr = slot;
2282
2283         while (1) {
2284                 if (path->reada == READA_BACK) {
2285                         if (nr == 0)
2286                                 break;
2287                         nr--;
2288                 } else if (path->reada == READA_FORWARD) {
2289                         nr++;
2290                         if (nr >= nritems)
2291                                 break;
2292                 }
2293                 if (path->reada == READA_BACK && objectid) {
2294                         btrfs_node_key(node, &disk_key, nr);
2295                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2296                                 break;
2297                 }
2298                 search = btrfs_node_blockptr(node, nr);
2299                 if ((search <= target && target - search <= 65536) ||
2300                     (search > target && search - target <= 65536)) {
2301                         gen = btrfs_node_ptr_generation(node, nr);
2302                         readahead_tree_block(root, search);
2303                         nread += blocksize;
2304                 }
2305                 nscan++;
2306                 if ((nread > 65536 || nscan > 32))
2307                         break;
2308         }
2309 }
2310
2311 static noinline void reada_for_balance(struct btrfs_root *root,
2312                                        struct btrfs_path *path, int level)
2313 {
2314         int slot;
2315         int nritems;
2316         struct extent_buffer *parent;
2317         struct extent_buffer *eb;
2318         u64 gen;
2319         u64 block1 = 0;
2320         u64 block2 = 0;
2321
2322         parent = path->nodes[level + 1];
2323         if (!parent)
2324                 return;
2325
2326         nritems = btrfs_header_nritems(parent);
2327         slot = path->slots[level + 1];
2328
2329         if (slot > 0) {
2330                 block1 = btrfs_node_blockptr(parent, slot - 1);
2331                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2332                 eb = btrfs_find_tree_block(root->fs_info, block1);
2333                 /*
2334                  * if we get -eagain from btrfs_buffer_uptodate, we
2335                  * don't want to return eagain here.  That will loop
2336                  * forever
2337                  */
2338                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2339                         block1 = 0;
2340                 free_extent_buffer(eb);
2341         }
2342         if (slot + 1 < nritems) {
2343                 block2 = btrfs_node_blockptr(parent, slot + 1);
2344                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2345                 eb = btrfs_find_tree_block(root->fs_info, block2);
2346                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2347                         block2 = 0;
2348                 free_extent_buffer(eb);
2349         }
2350
2351         if (block1)
2352                 readahead_tree_block(root, block1);
2353         if (block2)
2354                 readahead_tree_block(root, block2);
2355 }
2356
2357
2358 /*
2359  * when we walk down the tree, it is usually safe to unlock the higher layers
2360  * in the tree.  The exceptions are when our path goes through slot 0, because
2361  * operations on the tree might require changing key pointers higher up in the
2362  * tree.
2363  *
2364  * callers might also have set path->keep_locks, which tells this code to keep
2365  * the lock if the path points to the last slot in the block.  This is part of
2366  * walking through the tree, and selecting the next slot in the higher block.
2367  *
2368  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2369  * if lowest_unlock is 1, level 0 won't be unlocked
2370  */
2371 static noinline void unlock_up(struct btrfs_path *path, int level,
2372                                int lowest_unlock, int min_write_lock_level,
2373                                int *write_lock_level)
2374 {
2375         int i;
2376         int skip_level = level;
2377         int no_skips = 0;
2378         struct extent_buffer *t;
2379
2380         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2381                 if (!path->nodes[i])
2382                         break;
2383                 if (!path->locks[i])
2384                         break;
2385                 if (!no_skips && path->slots[i] == 0) {
2386                         skip_level = i + 1;
2387                         continue;
2388                 }
2389                 if (!no_skips && path->keep_locks) {
2390                         u32 nritems;
2391                         t = path->nodes[i];
2392                         nritems = btrfs_header_nritems(t);
2393                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2394                                 skip_level = i + 1;
2395                                 continue;
2396                         }
2397                 }
2398                 if (skip_level < i && i >= lowest_unlock)
2399                         no_skips = 1;
2400
2401                 t = path->nodes[i];
2402                 if (i >= lowest_unlock && i > skip_level && path->locks[i]) {
2403                         btrfs_tree_unlock_rw(t, path->locks[i]);
2404                         path->locks[i] = 0;
2405                         if (write_lock_level &&
2406                             i > min_write_lock_level &&
2407                             i <= *write_lock_level) {
2408                                 *write_lock_level = i - 1;
2409                         }
2410                 }
2411         }
2412 }
2413
2414 /*
2415  * This releases any locks held in the path starting at level and
2416  * going all the way up to the root.
2417  *
2418  * btrfs_search_slot will keep the lock held on higher nodes in a few
2419  * corner cases, such as COW of the block at slot zero in the node.  This
2420  * ignores those rules, and it should only be called when there are no
2421  * more updates to be done higher up in the tree.
2422  */
2423 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2424 {
2425         int i;
2426
2427         if (path->keep_locks)
2428                 return;
2429
2430         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2431                 if (!path->nodes[i])
2432                         continue;
2433                 if (!path->locks[i])
2434                         continue;
2435                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2436                 path->locks[i] = 0;
2437         }
2438 }
2439
2440 /*
2441  * helper function for btrfs_search_slot.  The goal is to find a block
2442  * in cache without setting the path to blocking.  If we find the block
2443  * we return zero and the path is unchanged.
2444  *
2445  * If we can't find the block, we set the path blocking and do some
2446  * reada.  -EAGAIN is returned and the search must be repeated.
2447  */
2448 static int
2449 read_block_for_search(struct btrfs_trans_handle *trans,
2450                        struct btrfs_root *root, struct btrfs_path *p,
2451                        struct extent_buffer **eb_ret, int level, int slot,
2452                        struct btrfs_key *key, u64 time_seq)
2453 {
2454         u64 blocknr;
2455         u64 gen;
2456         struct extent_buffer *b = *eb_ret;
2457         struct extent_buffer *tmp;
2458         int ret;
2459
2460         blocknr = btrfs_node_blockptr(b, slot);
2461         gen = btrfs_node_ptr_generation(b, slot);
2462
2463         tmp = btrfs_find_tree_block(root->fs_info, blocknr);
2464         if (tmp) {
2465                 /* first we do an atomic uptodate check */
2466                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2467                         *eb_ret = tmp;
2468                         return 0;
2469                 }
2470
2471                 /* the pages were up to date, but we failed
2472                  * the generation number check.  Do a full
2473                  * read for the generation number that is correct.
2474                  * We must do this without dropping locks so
2475                  * we can trust our generation number
2476                  */
2477                 btrfs_set_path_blocking(p);
2478
2479                 /* now we're allowed to do a blocking uptodate check */
2480                 ret = btrfs_read_buffer(tmp, gen);
2481                 if (!ret) {
2482                         *eb_ret = tmp;
2483                         return 0;
2484                 }
2485                 free_extent_buffer(tmp);
2486                 btrfs_release_path(p);
2487                 return -EIO;
2488         }
2489
2490         /*
2491          * reduce lock contention at high levels
2492          * of the btree by dropping locks before
2493          * we read.  Don't release the lock on the current
2494          * level because we need to walk this node to figure
2495          * out which blocks to read.
2496          */
2497         btrfs_unlock_up_safe(p, level + 1);
2498         btrfs_set_path_blocking(p);
2499
2500         free_extent_buffer(tmp);
2501         if (p->reada != READA_NONE)
2502                 reada_for_search(root, p, level, slot, key->objectid);
2503
2504         btrfs_release_path(p);
2505
2506         ret = -EAGAIN;
2507         tmp = read_tree_block(root, blocknr, 0);
2508         if (!IS_ERR(tmp)) {
2509                 /*
2510                  * If the read above didn't mark this buffer up to date,
2511                  * it will never end up being up to date.  Set ret to EIO now
2512                  * and give up so that our caller doesn't loop forever
2513                  * on our EAGAINs.
2514                  */
2515                 if (!btrfs_buffer_uptodate(tmp, 0, 0))
2516                         ret = -EIO;
2517                 free_extent_buffer(tmp);
2518         } else {
2519                 ret = PTR_ERR(tmp);
2520         }
2521         return ret;
2522 }
2523
2524 /*
2525  * helper function for btrfs_search_slot.  This does all of the checks
2526  * for node-level blocks and does any balancing required based on
2527  * the ins_len.
2528  *
2529  * If no extra work was required, zero is returned.  If we had to
2530  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2531  * start over
2532  */
2533 static int
2534 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2535                        struct btrfs_root *root, struct btrfs_path *p,
2536                        struct extent_buffer *b, int level, int ins_len,
2537                        int *write_lock_level)
2538 {
2539         int ret;
2540         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2541             BTRFS_NODEPTRS_PER_BLOCK(root) - 3) {
2542                 int sret;
2543
2544                 if (*write_lock_level < level + 1) {
2545                         *write_lock_level = level + 1;
2546                         btrfs_release_path(p);
2547                         goto again;
2548                 }
2549
2550                 btrfs_set_path_blocking(p);
2551                 reada_for_balance(root, p, level);
2552                 sret = split_node(trans, root, p, level);
2553                 btrfs_clear_path_blocking(p, NULL, 0);
2554
2555                 BUG_ON(sret > 0);
2556                 if (sret) {
2557                         ret = sret;
2558                         goto done;
2559                 }
2560                 b = p->nodes[level];
2561         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2562                    BTRFS_NODEPTRS_PER_BLOCK(root) / 2) {
2563                 int sret;
2564
2565                 if (*write_lock_level < level + 1) {
2566                         *write_lock_level = level + 1;
2567                         btrfs_release_path(p);
2568                         goto again;
2569                 }
2570
2571                 btrfs_set_path_blocking(p);
2572                 reada_for_balance(root, p, level);
2573                 sret = balance_level(trans, root, p, level);
2574                 btrfs_clear_path_blocking(p, NULL, 0);
2575
2576                 if (sret) {
2577                         ret = sret;
2578                         goto done;
2579                 }
2580                 b = p->nodes[level];
2581                 if (!b) {
2582                         btrfs_release_path(p);
2583                         goto again;
2584                 }
2585                 BUG_ON(btrfs_header_nritems(b) == 1);
2586         }
2587         return 0;
2588
2589 again:
2590         ret = -EAGAIN;
2591 done:
2592         return ret;
2593 }
2594
2595 static void key_search_validate(struct extent_buffer *b,
2596                                 struct btrfs_key *key,
2597                                 int level)
2598 {
2599 #ifdef CONFIG_BTRFS_ASSERT
2600         struct btrfs_disk_key disk_key;
2601
2602         btrfs_cpu_key_to_disk(&disk_key, key);
2603
2604         if (level == 0)
2605                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2606                     offsetof(struct btrfs_leaf, items[0].key),
2607                     sizeof(disk_key)));
2608         else
2609                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2610                     offsetof(struct btrfs_node, ptrs[0].key),
2611                     sizeof(disk_key)));
2612 #endif
2613 }
2614
2615 static int key_search(struct extent_buffer *b, struct btrfs_key *key,
2616                       int level, int *prev_cmp, int *slot)
2617 {
2618         if (*prev_cmp != 0) {
2619                 *prev_cmp = bin_search(b, key, level, slot);
2620                 return *prev_cmp;
2621         }
2622
2623         key_search_validate(b, key, level);
2624         *slot = 0;
2625
2626         return 0;
2627 }
2628
2629 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2630                 u64 iobjectid, u64 ioff, u8 key_type,
2631                 struct btrfs_key *found_key)
2632 {
2633         int ret;
2634         struct btrfs_key key;
2635         struct extent_buffer *eb;
2636
2637         ASSERT(path);
2638         ASSERT(found_key);
2639
2640         key.type = key_type;
2641         key.objectid = iobjectid;
2642         key.offset = ioff;
2643
2644         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2645         if (ret < 0)
2646                 return ret;
2647
2648         eb = path->nodes[0];
2649         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2650                 ret = btrfs_next_leaf(fs_root, path);
2651                 if (ret)
2652                         return ret;
2653                 eb = path->nodes[0];
2654         }
2655
2656         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2657         if (found_key->type != key.type ||
2658                         found_key->objectid != key.objectid)
2659                 return 1;
2660
2661         return 0;
2662 }
2663
2664 /*
2665  * look for key in the tree.  path is filled in with nodes along the way
2666  * if key is found, we return zero and you can find the item in the leaf
2667  * level of the path (level 0)
2668  *
2669  * If the key isn't found, the path points to the slot where it should
2670  * be inserted, and 1 is returned.  If there are other errors during the
2671  * search a negative error number is returned.
2672  *
2673  * if ins_len > 0, nodes and leaves will be split as we walk down the
2674  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
2675  * possible)
2676  */
2677 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
2678                       *root, struct btrfs_key *key, struct btrfs_path *p, int
2679                       ins_len, int cow)
2680 {
2681         struct extent_buffer *b;
2682         int slot;
2683         int ret;
2684         int err;
2685         int level;
2686         int lowest_unlock = 1;
2687         int root_lock;
2688         /* everything at write_lock_level or lower must be write locked */
2689         int write_lock_level = 0;
2690         u8 lowest_level = 0;
2691         int min_write_lock_level;
2692         int prev_cmp;
2693
2694         lowest_level = p->lowest_level;
2695         WARN_ON(lowest_level && ins_len > 0);
2696         WARN_ON(p->nodes[0] != NULL);
2697         BUG_ON(!cow && ins_len);
2698
2699         if (ins_len < 0) {
2700                 lowest_unlock = 2;
2701
2702                 /* when we are removing items, we might have to go up to level
2703                  * two as we update tree pointers  Make sure we keep write
2704                  * for those levels as well
2705                  */
2706                 write_lock_level = 2;
2707         } else if (ins_len > 0) {
2708                 /*
2709                  * for inserting items, make sure we have a write lock on
2710                  * level 1 so we can update keys
2711                  */
2712                 write_lock_level = 1;
2713         }
2714
2715         if (!cow)
2716                 write_lock_level = -1;
2717
2718         if (cow && (p->keep_locks || p->lowest_level))
2719                 write_lock_level = BTRFS_MAX_LEVEL;
2720
2721         min_write_lock_level = write_lock_level;
2722
2723 again:
2724         prev_cmp = -1;
2725         /*
2726          * we try very hard to do read locks on the root
2727          */
2728         root_lock = BTRFS_READ_LOCK;
2729         level = 0;
2730         if (p->search_commit_root) {
2731                 /*
2732                  * the commit roots are read only
2733                  * so we always do read locks
2734                  */
2735                 if (p->need_commit_sem)
2736                         down_read(&root->fs_info->commit_root_sem);
2737                 b = root->commit_root;
2738                 extent_buffer_get(b);
2739                 level = btrfs_header_level(b);
2740                 if (p->need_commit_sem)
2741                         up_read(&root->fs_info->commit_root_sem);
2742                 if (!p->skip_locking)
2743                         btrfs_tree_read_lock(b);
2744         } else {
2745                 if (p->skip_locking) {
2746                         b = btrfs_root_node(root);
2747                         level = btrfs_header_level(b);
2748                 } else {
2749                         /* we don't know the level of the root node
2750                          * until we actually have it read locked
2751                          */
2752                         b = btrfs_read_lock_root_node(root);
2753                         level = btrfs_header_level(b);
2754                         if (level <= write_lock_level) {
2755                                 /* whoops, must trade for write lock */
2756                                 btrfs_tree_read_unlock(b);
2757                                 free_extent_buffer(b);
2758                                 b = btrfs_lock_root_node(root);
2759                                 root_lock = BTRFS_WRITE_LOCK;
2760
2761                                 /* the level might have changed, check again */
2762                                 level = btrfs_header_level(b);
2763                         }
2764                 }
2765         }
2766         p->nodes[level] = b;
2767         if (!p->skip_locking)
2768                 p->locks[level] = root_lock;
2769
2770         while (b) {
2771                 level = btrfs_header_level(b);
2772
2773                 /*
2774                  * setup the path here so we can release it under lock
2775                  * contention with the cow code
2776                  */
2777                 if (cow) {
2778                         /*
2779                          * if we don't really need to cow this block
2780                          * then we don't want to set the path blocking,
2781                          * so we test it here
2782                          */
2783                         if (!should_cow_block(trans, root, b)) {
2784                                 trans->dirty = true;
2785                                 goto cow_done;
2786                         }
2787
2788                         /*
2789                          * must have write locks on this node and the
2790                          * parent
2791                          */
2792                         if (level > write_lock_level ||
2793                             (level + 1 > write_lock_level &&
2794                             level + 1 < BTRFS_MAX_LEVEL &&
2795                             p->nodes[level + 1])) {
2796                                 write_lock_level = level + 1;
2797                                 btrfs_release_path(p);
2798                                 goto again;
2799                         }
2800
2801                         btrfs_set_path_blocking(p);
2802                         err = btrfs_cow_block(trans, root, b,
2803                                               p->nodes[level + 1],
2804                                               p->slots[level + 1], &b);
2805                         if (err) {
2806                                 ret = err;
2807                                 goto done;
2808                         }
2809                 }
2810 cow_done:
2811                 p->nodes[level] = b;
2812                 btrfs_clear_path_blocking(p, NULL, 0);
2813
2814                 /*
2815                  * we have a lock on b and as long as we aren't changing
2816                  * the tree, there is no way to for the items in b to change.
2817                  * It is safe to drop the lock on our parent before we
2818                  * go through the expensive btree search on b.
2819                  *
2820                  * If we're inserting or deleting (ins_len != 0), then we might
2821                  * be changing slot zero, which may require changing the parent.
2822                  * So, we can't drop the lock until after we know which slot
2823                  * we're operating on.
2824                  */
2825                 if (!ins_len && !p->keep_locks) {
2826                         int u = level + 1;
2827
2828                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2829                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2830                                 p->locks[u] = 0;
2831                         }
2832                 }
2833
2834                 ret = key_search(b, key, level, &prev_cmp, &slot);
2835                 if (ret < 0)
2836                         goto done;
2837
2838                 if (level != 0) {
2839                         int dec = 0;
2840                         if (ret && slot > 0) {
2841                                 dec = 1;
2842                                 slot -= 1;
2843                         }
2844                         p->slots[level] = slot;
2845                         err = setup_nodes_for_search(trans, root, p, b, level,
2846                                              ins_len, &write_lock_level);
2847                         if (err == -EAGAIN)
2848                                 goto again;
2849                         if (err) {
2850                                 ret = err;
2851                                 goto done;
2852                         }
2853                         b = p->nodes[level];
2854                         slot = p->slots[level];
2855
2856                         /*
2857                          * slot 0 is special, if we change the key
2858                          * we have to update the parent pointer
2859                          * which means we must have a write lock
2860                          * on the parent
2861                          */
2862                         if (slot == 0 && ins_len &&
2863                             write_lock_level < level + 1) {
2864                                 write_lock_level = level + 1;
2865                                 btrfs_release_path(p);
2866                                 goto again;
2867                         }
2868
2869                         unlock_up(p, level, lowest_unlock,
2870                                   min_write_lock_level, &write_lock_level);
2871
2872                         if (level == lowest_level) {
2873                                 if (dec)
2874                                         p->slots[level]++;
2875                                 goto done;
2876                         }
2877
2878                         err = read_block_for_search(trans, root, p,
2879                                                     &b, level, slot, key, 0);
2880                         if (err == -EAGAIN)
2881                                 goto again;
2882                         if (err) {
2883                                 ret = err;
2884                                 goto done;
2885                         }
2886
2887                         if (!p->skip_locking) {
2888                                 level = btrfs_header_level(b);
2889                                 if (level <= write_lock_level) {
2890                                         err = btrfs_try_tree_write_lock(b);
2891                                         if (!err) {
2892                                                 btrfs_set_path_blocking(p);
2893                                                 btrfs_tree_lock(b);
2894                                                 btrfs_clear_path_blocking(p, b,
2895                                                                   BTRFS_WRITE_LOCK);
2896                                         }
2897                                         p->locks[level] = BTRFS_WRITE_LOCK;
2898                                 } else {
2899                                         err = btrfs_tree_read_lock_atomic(b);
2900                                         if (!err) {
2901                                                 btrfs_set_path_blocking(p);
2902                                                 btrfs_tree_read_lock(b);
2903                                                 btrfs_clear_path_blocking(p, b,
2904                                                                   BTRFS_READ_LOCK);
2905                                         }
2906                                         p->locks[level] = BTRFS_READ_LOCK;
2907                                 }
2908                                 p->nodes[level] = b;
2909                         }
2910                 } else {
2911                         p->slots[level] = slot;
2912                         if (ins_len > 0 &&
2913                             btrfs_leaf_free_space(root, b) < ins_len) {
2914                                 if (write_lock_level < 1) {
2915                                         write_lock_level = 1;
2916                                         btrfs_release_path(p);
2917                                         goto again;
2918                                 }
2919
2920                                 btrfs_set_path_blocking(p);
2921                                 err = split_leaf(trans, root, key,
2922                                                  p, ins_len, ret == 0);
2923                                 btrfs_clear_path_blocking(p, NULL, 0);
2924
2925                                 BUG_ON(err > 0);
2926                                 if (err) {
2927                                         ret = err;
2928                                         goto done;
2929                                 }
2930                         }
2931                         if (!p->search_for_split)
2932                                 unlock_up(p, level, lowest_unlock,
2933                                           min_write_lock_level, &write_lock_level);
2934                         goto done;
2935                 }
2936         }
2937         ret = 1;
2938 done:
2939         /*
2940          * we don't really know what they plan on doing with the path
2941          * from here on, so for now just mark it as blocking
2942          */
2943         if (!p->leave_spinning)
2944                 btrfs_set_path_blocking(p);
2945         if (ret < 0 && !p->skip_release_on_error)
2946                 btrfs_release_path(p);
2947         return ret;
2948 }
2949
2950 /*
2951  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2952  * current state of the tree together with the operations recorded in the tree
2953  * modification log to search for the key in a previous version of this tree, as
2954  * denoted by the time_seq parameter.
2955  *
2956  * Naturally, there is no support for insert, delete or cow operations.
2957  *
2958  * The resulting path and return value will be set up as if we called
2959  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2960  */
2961 int btrfs_search_old_slot(struct btrfs_root *root, struct btrfs_key *key,
2962                           struct btrfs_path *p, u64 time_seq)
2963 {
2964         struct extent_buffer *b;
2965         int slot;
2966         int ret;
2967         int err;
2968         int level;
2969         int lowest_unlock = 1;
2970         u8 lowest_level = 0;
2971         int prev_cmp = -1;
2972
2973         lowest_level = p->lowest_level;
2974         WARN_ON(p->nodes[0] != NULL);
2975
2976         if (p->search_commit_root) {
2977                 BUG_ON(time_seq);
2978                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2979         }
2980
2981 again:
2982         b = get_old_root(root, time_seq);
2983         level = btrfs_header_level(b);
2984         p->locks[level] = BTRFS_READ_LOCK;
2985
2986         while (b) {
2987                 level = btrfs_header_level(b);
2988                 p->nodes[level] = b;
2989                 btrfs_clear_path_blocking(p, NULL, 0);
2990
2991                 /*
2992                  * we have a lock on b and as long as we aren't changing
2993                  * the tree, there is no way to for the items in b to change.
2994                  * It is safe to drop the lock on our parent before we
2995                  * go through the expensive btree search on b.
2996                  */
2997                 btrfs_unlock_up_safe(p, level + 1);
2998
2999                 /*
3000                  * Since we can unwind ebs we want to do a real search every
3001                  * time.
3002                  */
3003                 prev_cmp = -1;
3004                 ret = key_search(b, key, level, &prev_cmp, &slot);
3005
3006                 if (level != 0) {
3007                         int dec = 0;
3008                         if (ret && slot > 0) {
3009                                 dec = 1;
3010                                 slot -= 1;
3011                         }
3012                         p->slots[level] = slot;
3013                         unlock_up(p, level, lowest_unlock, 0, NULL);
3014
3015                         if (level == lowest_level) {
3016                                 if (dec)
3017                                         p->slots[level]++;
3018                                 goto done;
3019                         }
3020
3021                         err = read_block_for_search(NULL, root, p, &b, level,
3022                                                     slot, key, time_seq);
3023                         if (err == -EAGAIN)
3024                                 goto again;
3025                         if (err) {
3026                                 ret = err;
3027                                 goto done;
3028                         }
3029
3030                         level = btrfs_header_level(b);
3031                         err = btrfs_tree_read_lock_atomic(b);
3032                         if (!err) {
3033                                 btrfs_set_path_blocking(p);
3034                                 btrfs_tree_read_lock(b);
3035                                 btrfs_clear_path_blocking(p, b,
3036                                                           BTRFS_READ_LOCK);
3037                         }
3038                         b = tree_mod_log_rewind(root->fs_info, p, b, time_seq);
3039                         if (!b) {
3040                                 ret = -ENOMEM;
3041                                 goto done;
3042                         }
3043                         p->locks[level] = BTRFS_READ_LOCK;
3044                         p->nodes[level] = b;
3045                 } else {
3046                         p->slots[level] = slot;
3047                         unlock_up(p, level, lowest_unlock, 0, NULL);
3048                         goto done;
3049                 }
3050         }
3051         ret = 1;
3052 done:
3053         if (!p->leave_spinning)
3054                 btrfs_set_path_blocking(p);
3055         if (ret < 0)
3056                 btrfs_release_path(p);
3057
3058         return ret;
3059 }
3060
3061 /*
3062  * helper to use instead of search slot if no exact match is needed but
3063  * instead the next or previous item should be returned.
3064  * When find_higher is true, the next higher item is returned, the next lower
3065  * otherwise.
3066  * When return_any and find_higher are both true, and no higher item is found,
3067  * return the next lower instead.
3068  * When return_any is true and find_higher is false, and no lower item is found,
3069  * return the next higher instead.
3070  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3071  * < 0 on error
3072  */
3073 int btrfs_search_slot_for_read(struct btrfs_root *root,
3074                                struct btrfs_key *key, struct btrfs_path *p,
3075                                int find_higher, int return_any)
3076 {
3077         int ret;
3078         struct extent_buffer *leaf;
3079
3080 again:
3081         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3082         if (ret <= 0)
3083                 return ret;
3084         /*
3085          * a return value of 1 means the path is at the position where the
3086          * item should be inserted. Normally this is the next bigger item,
3087          * but in case the previous item is the last in a leaf, path points
3088          * to the first free slot in the previous leaf, i.e. at an invalid
3089          * item.
3090          */
3091         leaf = p->nodes[0];
3092
3093         if (find_higher) {
3094                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3095                         ret = btrfs_next_leaf(root, p);
3096                         if (ret <= 0)
3097                                 return ret;
3098                         if (!return_any)
3099                                 return 1;
3100                         /*
3101                          * no higher item found, return the next
3102                          * lower instead
3103                          */
3104                         return_any = 0;
3105                         find_higher = 0;
3106                         btrfs_release_path(p);
3107                         goto again;
3108                 }
3109         } else {
3110                 if (p->slots[0] == 0) {
3111                         ret = btrfs_prev_leaf(root, p);
3112                         if (ret < 0)
3113                                 return ret;
3114                         if (!ret) {
3115                                 leaf = p->nodes[0];
3116                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3117                                         p->slots[0]--;
3118                                 return 0;
3119                         }
3120                         if (!return_any)
3121                                 return 1;
3122                         /*
3123                          * no lower item found, return the next
3124                          * higher instead
3125                          */
3126                         return_any = 0;
3127                         find_higher = 1;
3128                         btrfs_release_path(p);
3129                         goto again;
3130                 } else {
3131                         --p->slots[0];
3132                 }
3133         }
3134         return 0;
3135 }
3136
3137 /*
3138  * adjust the pointers going up the tree, starting at level
3139  * making sure the right key of each node is points to 'key'.