Btrfs: send, fix race with transaction commits that create snapshots
[sfrench/cifs-2.6.git] / fs / btrfs / ctree.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007,2008 Oracle.  All rights reserved.
4  */
5
6 #include <linux/sched.h>
7 #include <linux/slab.h>
8 #include <linux/rbtree.h>
9 #include <linux/mm.h>
10 #include "ctree.h"
11 #include "disk-io.h"
12 #include "transaction.h"
13 #include "print-tree.h"
14 #include "locking.h"
15 #include "volumes.h"
16
17 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
18                       *root, struct btrfs_path *path, int level);
19 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root *root,
20                       const struct btrfs_key *ins_key, struct btrfs_path *path,
21                       int data_size, int extend);
22 static int push_node_left(struct btrfs_trans_handle *trans,
23                           struct btrfs_fs_info *fs_info,
24                           struct extent_buffer *dst,
25                           struct extent_buffer *src, int empty);
26 static int balance_node_right(struct btrfs_trans_handle *trans,
27                               struct btrfs_fs_info *fs_info,
28                               struct extent_buffer *dst_buf,
29                               struct extent_buffer *src_buf);
30 static void del_ptr(struct btrfs_root *root, struct btrfs_path *path,
31                     int level, int slot);
32
33 struct btrfs_path *btrfs_alloc_path(void)
34 {
35         return kmem_cache_zalloc(btrfs_path_cachep, GFP_NOFS);
36 }
37
38 /*
39  * set all locked nodes in the path to blocking locks.  This should
40  * be done before scheduling
41  */
42 noinline void btrfs_set_path_blocking(struct btrfs_path *p)
43 {
44         int i;
45         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
46                 if (!p->nodes[i] || !p->locks[i])
47                         continue;
48                 btrfs_set_lock_blocking_rw(p->nodes[i], p->locks[i]);
49                 if (p->locks[i] == BTRFS_READ_LOCK)
50                         p->locks[i] = BTRFS_READ_LOCK_BLOCKING;
51                 else if (p->locks[i] == BTRFS_WRITE_LOCK)
52                         p->locks[i] = BTRFS_WRITE_LOCK_BLOCKING;
53         }
54 }
55
56 /* this also releases the path */
57 void btrfs_free_path(struct btrfs_path *p)
58 {
59         if (!p)
60                 return;
61         btrfs_release_path(p);
62         kmem_cache_free(btrfs_path_cachep, p);
63 }
64
65 /*
66  * path release drops references on the extent buffers in the path
67  * and it drops any locks held by this path
68  *
69  * It is safe to call this on paths that no locks or extent buffers held.
70  */
71 noinline void btrfs_release_path(struct btrfs_path *p)
72 {
73         int i;
74
75         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
76                 p->slots[i] = 0;
77                 if (!p->nodes[i])
78                         continue;
79                 if (p->locks[i]) {
80                         btrfs_tree_unlock_rw(p->nodes[i], p->locks[i]);
81                         p->locks[i] = 0;
82                 }
83                 free_extent_buffer(p->nodes[i]);
84                 p->nodes[i] = NULL;
85         }
86 }
87
88 /*
89  * safely gets a reference on the root node of a tree.  A lock
90  * is not taken, so a concurrent writer may put a different node
91  * at the root of the tree.  See btrfs_lock_root_node for the
92  * looping required.
93  *
94  * The extent buffer returned by this has a reference taken, so
95  * it won't disappear.  It may stop being the root of the tree
96  * at any time because there are no locks held.
97  */
98 struct extent_buffer *btrfs_root_node(struct btrfs_root *root)
99 {
100         struct extent_buffer *eb;
101
102         while (1) {
103                 rcu_read_lock();
104                 eb = rcu_dereference(root->node);
105
106                 /*
107                  * RCU really hurts here, we could free up the root node because
108                  * it was COWed but we may not get the new root node yet so do
109                  * the inc_not_zero dance and if it doesn't work then
110                  * synchronize_rcu and try again.
111                  */
112                 if (atomic_inc_not_zero(&eb->refs)) {
113                         rcu_read_unlock();
114                         break;
115                 }
116                 rcu_read_unlock();
117                 synchronize_rcu();
118         }
119         return eb;
120 }
121
122 /* loop around taking references on and locking the root node of the
123  * tree until you end up with a lock on the root.  A locked buffer
124  * is returned, with a reference held.
125  */
126 struct extent_buffer *btrfs_lock_root_node(struct btrfs_root *root)
127 {
128         struct extent_buffer *eb;
129
130         while (1) {
131                 eb = btrfs_root_node(root);
132                 btrfs_tree_lock(eb);
133                 if (eb == root->node)
134                         break;
135                 btrfs_tree_unlock(eb);
136                 free_extent_buffer(eb);
137         }
138         return eb;
139 }
140
141 /* loop around taking references on and locking the root node of the
142  * tree until you end up with a lock on the root.  A locked buffer
143  * is returned, with a reference held.
144  */
145 struct extent_buffer *btrfs_read_lock_root_node(struct btrfs_root *root)
146 {
147         struct extent_buffer *eb;
148
149         while (1) {
150                 eb = btrfs_root_node(root);
151                 btrfs_tree_read_lock(eb);
152                 if (eb == root->node)
153                         break;
154                 btrfs_tree_read_unlock(eb);
155                 free_extent_buffer(eb);
156         }
157         return eb;
158 }
159
160 /* cowonly root (everything not a reference counted cow subvolume), just get
161  * put onto a simple dirty list.  transaction.c walks this to make sure they
162  * get properly updated on disk.
163  */
164 static void add_root_to_dirty_list(struct btrfs_root *root)
165 {
166         struct btrfs_fs_info *fs_info = root->fs_info;
167
168         if (test_bit(BTRFS_ROOT_DIRTY, &root->state) ||
169             !test_bit(BTRFS_ROOT_TRACK_DIRTY, &root->state))
170                 return;
171
172         spin_lock(&fs_info->trans_lock);
173         if (!test_and_set_bit(BTRFS_ROOT_DIRTY, &root->state)) {
174                 /* Want the extent tree to be the last on the list */
175                 if (root->root_key.objectid == BTRFS_EXTENT_TREE_OBJECTID)
176                         list_move_tail(&root->dirty_list,
177                                        &fs_info->dirty_cowonly_roots);
178                 else
179                         list_move(&root->dirty_list,
180                                   &fs_info->dirty_cowonly_roots);
181         }
182         spin_unlock(&fs_info->trans_lock);
183 }
184
185 /*
186  * used by snapshot creation to make a copy of a root for a tree with
187  * a given objectid.  The buffer with the new root node is returned in
188  * cow_ret, and this func returns zero on success or a negative error code.
189  */
190 int btrfs_copy_root(struct btrfs_trans_handle *trans,
191                       struct btrfs_root *root,
192                       struct extent_buffer *buf,
193                       struct extent_buffer **cow_ret, u64 new_root_objectid)
194 {
195         struct btrfs_fs_info *fs_info = root->fs_info;
196         struct extent_buffer *cow;
197         int ret = 0;
198         int level;
199         struct btrfs_disk_key disk_key;
200
201         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
202                 trans->transid != fs_info->running_transaction->transid);
203         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
204                 trans->transid != root->last_trans);
205
206         level = btrfs_header_level(buf);
207         if (level == 0)
208                 btrfs_item_key(buf, &disk_key, 0);
209         else
210                 btrfs_node_key(buf, &disk_key, 0);
211
212         cow = btrfs_alloc_tree_block(trans, root, 0, new_root_objectid,
213                         &disk_key, level, buf->start, 0);
214         if (IS_ERR(cow))
215                 return PTR_ERR(cow);
216
217         copy_extent_buffer_full(cow, buf);
218         btrfs_set_header_bytenr(cow, cow->start);
219         btrfs_set_header_generation(cow, trans->transid);
220         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
221         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
222                                      BTRFS_HEADER_FLAG_RELOC);
223         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
224                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
225         else
226                 btrfs_set_header_owner(cow, new_root_objectid);
227
228         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
229
230         WARN_ON(btrfs_header_generation(buf) > trans->transid);
231         if (new_root_objectid == BTRFS_TREE_RELOC_OBJECTID)
232                 ret = btrfs_inc_ref(trans, root, cow, 1);
233         else
234                 ret = btrfs_inc_ref(trans, root, cow, 0);
235
236         if (ret)
237                 return ret;
238
239         btrfs_mark_buffer_dirty(cow);
240         *cow_ret = cow;
241         return 0;
242 }
243
244 enum mod_log_op {
245         MOD_LOG_KEY_REPLACE,
246         MOD_LOG_KEY_ADD,
247         MOD_LOG_KEY_REMOVE,
248         MOD_LOG_KEY_REMOVE_WHILE_FREEING,
249         MOD_LOG_KEY_REMOVE_WHILE_MOVING,
250         MOD_LOG_MOVE_KEYS,
251         MOD_LOG_ROOT_REPLACE,
252 };
253
254 struct tree_mod_root {
255         u64 logical;
256         u8 level;
257 };
258
259 struct tree_mod_elem {
260         struct rb_node node;
261         u64 logical;
262         u64 seq;
263         enum mod_log_op op;
264
265         /* this is used for MOD_LOG_KEY_* and MOD_LOG_MOVE_KEYS operations */
266         int slot;
267
268         /* this is used for MOD_LOG_KEY* and MOD_LOG_ROOT_REPLACE */
269         u64 generation;
270
271         /* those are used for op == MOD_LOG_KEY_{REPLACE,REMOVE} */
272         struct btrfs_disk_key key;
273         u64 blockptr;
274
275         /* this is used for op == MOD_LOG_MOVE_KEYS */
276         struct {
277                 int dst_slot;
278                 int nr_items;
279         } move;
280
281         /* this is used for op == MOD_LOG_ROOT_REPLACE */
282         struct tree_mod_root old_root;
283 };
284
285 /*
286  * Pull a new tree mod seq number for our operation.
287  */
288 static inline u64 btrfs_inc_tree_mod_seq(struct btrfs_fs_info *fs_info)
289 {
290         return atomic64_inc_return(&fs_info->tree_mod_seq);
291 }
292
293 /*
294  * This adds a new blocker to the tree mod log's blocker list if the @elem
295  * passed does not already have a sequence number set. So when a caller expects
296  * to record tree modifications, it should ensure to set elem->seq to zero
297  * before calling btrfs_get_tree_mod_seq.
298  * Returns a fresh, unused tree log modification sequence number, even if no new
299  * blocker was added.
300  */
301 u64 btrfs_get_tree_mod_seq(struct btrfs_fs_info *fs_info,
302                            struct seq_list *elem)
303 {
304         write_lock(&fs_info->tree_mod_log_lock);
305         spin_lock(&fs_info->tree_mod_seq_lock);
306         if (!elem->seq) {
307                 elem->seq = btrfs_inc_tree_mod_seq(fs_info);
308                 list_add_tail(&elem->list, &fs_info->tree_mod_seq_list);
309         }
310         spin_unlock(&fs_info->tree_mod_seq_lock);
311         write_unlock(&fs_info->tree_mod_log_lock);
312
313         return elem->seq;
314 }
315
316 void btrfs_put_tree_mod_seq(struct btrfs_fs_info *fs_info,
317                             struct seq_list *elem)
318 {
319         struct rb_root *tm_root;
320         struct rb_node *node;
321         struct rb_node *next;
322         struct seq_list *cur_elem;
323         struct tree_mod_elem *tm;
324         u64 min_seq = (u64)-1;
325         u64 seq_putting = elem->seq;
326
327         if (!seq_putting)
328                 return;
329
330         spin_lock(&fs_info->tree_mod_seq_lock);
331         list_del(&elem->list);
332         elem->seq = 0;
333
334         list_for_each_entry(cur_elem, &fs_info->tree_mod_seq_list, list) {
335                 if (cur_elem->seq < min_seq) {
336                         if (seq_putting > cur_elem->seq) {
337                                 /*
338                                  * blocker with lower sequence number exists, we
339                                  * cannot remove anything from the log
340                                  */
341                                 spin_unlock(&fs_info->tree_mod_seq_lock);
342                                 return;
343                         }
344                         min_seq = cur_elem->seq;
345                 }
346         }
347         spin_unlock(&fs_info->tree_mod_seq_lock);
348
349         /*
350          * anything that's lower than the lowest existing (read: blocked)
351          * sequence number can be removed from the tree.
352          */
353         write_lock(&fs_info->tree_mod_log_lock);
354         tm_root = &fs_info->tree_mod_log;
355         for (node = rb_first(tm_root); node; node = next) {
356                 next = rb_next(node);
357                 tm = rb_entry(node, struct tree_mod_elem, node);
358                 if (tm->seq > min_seq)
359                         continue;
360                 rb_erase(node, tm_root);
361                 kfree(tm);
362         }
363         write_unlock(&fs_info->tree_mod_log_lock);
364 }
365
366 /*
367  * key order of the log:
368  *       node/leaf start address -> sequence
369  *
370  * The 'start address' is the logical address of the *new* root node
371  * for root replace operations, or the logical address of the affected
372  * block for all other operations.
373  *
374  * Note: must be called with write lock for fs_info::tree_mod_log_lock.
375  */
376 static noinline int
377 __tree_mod_log_insert(struct btrfs_fs_info *fs_info, struct tree_mod_elem *tm)
378 {
379         struct rb_root *tm_root;
380         struct rb_node **new;
381         struct rb_node *parent = NULL;
382         struct tree_mod_elem *cur;
383
384         tm->seq = btrfs_inc_tree_mod_seq(fs_info);
385
386         tm_root = &fs_info->tree_mod_log;
387         new = &tm_root->rb_node;
388         while (*new) {
389                 cur = rb_entry(*new, struct tree_mod_elem, node);
390                 parent = *new;
391                 if (cur->logical < tm->logical)
392                         new = &((*new)->rb_left);
393                 else if (cur->logical > tm->logical)
394                         new = &((*new)->rb_right);
395                 else if (cur->seq < tm->seq)
396                         new = &((*new)->rb_left);
397                 else if (cur->seq > tm->seq)
398                         new = &((*new)->rb_right);
399                 else
400                         return -EEXIST;
401         }
402
403         rb_link_node(&tm->node, parent, new);
404         rb_insert_color(&tm->node, tm_root);
405         return 0;
406 }
407
408 /*
409  * Determines if logging can be omitted. Returns 1 if it can. Otherwise, it
410  * returns zero with the tree_mod_log_lock acquired. The caller must hold
411  * this until all tree mod log insertions are recorded in the rb tree and then
412  * write unlock fs_info::tree_mod_log_lock.
413  */
414 static inline int tree_mod_dont_log(struct btrfs_fs_info *fs_info,
415                                     struct extent_buffer *eb) {
416         smp_mb();
417         if (list_empty(&(fs_info)->tree_mod_seq_list))
418                 return 1;
419         if (eb && btrfs_header_level(eb) == 0)
420                 return 1;
421
422         write_lock(&fs_info->tree_mod_log_lock);
423         if (list_empty(&(fs_info)->tree_mod_seq_list)) {
424                 write_unlock(&fs_info->tree_mod_log_lock);
425                 return 1;
426         }
427
428         return 0;
429 }
430
431 /* Similar to tree_mod_dont_log, but doesn't acquire any locks. */
432 static inline int tree_mod_need_log(const struct btrfs_fs_info *fs_info,
433                                     struct extent_buffer *eb)
434 {
435         smp_mb();
436         if (list_empty(&(fs_info)->tree_mod_seq_list))
437                 return 0;
438         if (eb && btrfs_header_level(eb) == 0)
439                 return 0;
440
441         return 1;
442 }
443
444 static struct tree_mod_elem *
445 alloc_tree_mod_elem(struct extent_buffer *eb, int slot,
446                     enum mod_log_op op, gfp_t flags)
447 {
448         struct tree_mod_elem *tm;
449
450         tm = kzalloc(sizeof(*tm), flags);
451         if (!tm)
452                 return NULL;
453
454         tm->logical = eb->start;
455         if (op != MOD_LOG_KEY_ADD) {
456                 btrfs_node_key(eb, &tm->key, slot);
457                 tm->blockptr = btrfs_node_blockptr(eb, slot);
458         }
459         tm->op = op;
460         tm->slot = slot;
461         tm->generation = btrfs_node_ptr_generation(eb, slot);
462         RB_CLEAR_NODE(&tm->node);
463
464         return tm;
465 }
466
467 static noinline int tree_mod_log_insert_key(struct extent_buffer *eb, int slot,
468                 enum mod_log_op op, gfp_t flags)
469 {
470         struct tree_mod_elem *tm;
471         int ret;
472
473         if (!tree_mod_need_log(eb->fs_info, eb))
474                 return 0;
475
476         tm = alloc_tree_mod_elem(eb, slot, op, flags);
477         if (!tm)
478                 return -ENOMEM;
479
480         if (tree_mod_dont_log(eb->fs_info, eb)) {
481                 kfree(tm);
482                 return 0;
483         }
484
485         ret = __tree_mod_log_insert(eb->fs_info, tm);
486         write_unlock(&eb->fs_info->tree_mod_log_lock);
487         if (ret)
488                 kfree(tm);
489
490         return ret;
491 }
492
493 static noinline int tree_mod_log_insert_move(struct extent_buffer *eb,
494                 int dst_slot, int src_slot, int nr_items)
495 {
496         struct tree_mod_elem *tm = NULL;
497         struct tree_mod_elem **tm_list = NULL;
498         int ret = 0;
499         int i;
500         int locked = 0;
501
502         if (!tree_mod_need_log(eb->fs_info, eb))
503                 return 0;
504
505         tm_list = kcalloc(nr_items, sizeof(struct tree_mod_elem *), GFP_NOFS);
506         if (!tm_list)
507                 return -ENOMEM;
508
509         tm = kzalloc(sizeof(*tm), GFP_NOFS);
510         if (!tm) {
511                 ret = -ENOMEM;
512                 goto free_tms;
513         }
514
515         tm->logical = eb->start;
516         tm->slot = src_slot;
517         tm->move.dst_slot = dst_slot;
518         tm->move.nr_items = nr_items;
519         tm->op = MOD_LOG_MOVE_KEYS;
520
521         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
522                 tm_list[i] = alloc_tree_mod_elem(eb, i + dst_slot,
523                     MOD_LOG_KEY_REMOVE_WHILE_MOVING, GFP_NOFS);
524                 if (!tm_list[i]) {
525                         ret = -ENOMEM;
526                         goto free_tms;
527                 }
528         }
529
530         if (tree_mod_dont_log(eb->fs_info, eb))
531                 goto free_tms;
532         locked = 1;
533
534         /*
535          * When we override something during the move, we log these removals.
536          * This can only happen when we move towards the beginning of the
537          * buffer, i.e. dst_slot < src_slot.
538          */
539         for (i = 0; i + dst_slot < src_slot && i < nr_items; i++) {
540                 ret = __tree_mod_log_insert(eb->fs_info, tm_list[i]);
541                 if (ret)
542                         goto free_tms;
543         }
544
545         ret = __tree_mod_log_insert(eb->fs_info, tm);
546         if (ret)
547                 goto free_tms;
548         write_unlock(&eb->fs_info->tree_mod_log_lock);
549         kfree(tm_list);
550
551         return 0;
552 free_tms:
553         for (i = 0; i < nr_items; i++) {
554                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
555                         rb_erase(&tm_list[i]->node, &eb->fs_info->tree_mod_log);
556                 kfree(tm_list[i]);
557         }
558         if (locked)
559                 write_unlock(&eb->fs_info->tree_mod_log_lock);
560         kfree(tm_list);
561         kfree(tm);
562
563         return ret;
564 }
565
566 static inline int
567 __tree_mod_log_free_eb(struct btrfs_fs_info *fs_info,
568                        struct tree_mod_elem **tm_list,
569                        int nritems)
570 {
571         int i, j;
572         int ret;
573
574         for (i = nritems - 1; i >= 0; i--) {
575                 ret = __tree_mod_log_insert(fs_info, tm_list[i]);
576                 if (ret) {
577                         for (j = nritems - 1; j > i; j--)
578                                 rb_erase(&tm_list[j]->node,
579                                          &fs_info->tree_mod_log);
580                         return ret;
581                 }
582         }
583
584         return 0;
585 }
586
587 static noinline int tree_mod_log_insert_root(struct extent_buffer *old_root,
588                          struct extent_buffer *new_root, int log_removal)
589 {
590         struct btrfs_fs_info *fs_info = old_root->fs_info;
591         struct tree_mod_elem *tm = NULL;
592         struct tree_mod_elem **tm_list = NULL;
593         int nritems = 0;
594         int ret = 0;
595         int i;
596
597         if (!tree_mod_need_log(fs_info, NULL))
598                 return 0;
599
600         if (log_removal && btrfs_header_level(old_root) > 0) {
601                 nritems = btrfs_header_nritems(old_root);
602                 tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *),
603                                   GFP_NOFS);
604                 if (!tm_list) {
605                         ret = -ENOMEM;
606                         goto free_tms;
607                 }
608                 for (i = 0; i < nritems; i++) {
609                         tm_list[i] = alloc_tree_mod_elem(old_root, i,
610                             MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
611                         if (!tm_list[i]) {
612                                 ret = -ENOMEM;
613                                 goto free_tms;
614                         }
615                 }
616         }
617
618         tm = kzalloc(sizeof(*tm), GFP_NOFS);
619         if (!tm) {
620                 ret = -ENOMEM;
621                 goto free_tms;
622         }
623
624         tm->logical = new_root->start;
625         tm->old_root.logical = old_root->start;
626         tm->old_root.level = btrfs_header_level(old_root);
627         tm->generation = btrfs_header_generation(old_root);
628         tm->op = MOD_LOG_ROOT_REPLACE;
629
630         if (tree_mod_dont_log(fs_info, NULL))
631                 goto free_tms;
632
633         if (tm_list)
634                 ret = __tree_mod_log_free_eb(fs_info, tm_list, nritems);
635         if (!ret)
636                 ret = __tree_mod_log_insert(fs_info, tm);
637
638         write_unlock(&fs_info->tree_mod_log_lock);
639         if (ret)
640                 goto free_tms;
641         kfree(tm_list);
642
643         return ret;
644
645 free_tms:
646         if (tm_list) {
647                 for (i = 0; i < nritems; i++)
648                         kfree(tm_list[i]);
649                 kfree(tm_list);
650         }
651         kfree(tm);
652
653         return ret;
654 }
655
656 static struct tree_mod_elem *
657 __tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq,
658                       int smallest)
659 {
660         struct rb_root *tm_root;
661         struct rb_node *node;
662         struct tree_mod_elem *cur = NULL;
663         struct tree_mod_elem *found = NULL;
664
665         read_lock(&fs_info->tree_mod_log_lock);
666         tm_root = &fs_info->tree_mod_log;
667         node = tm_root->rb_node;
668         while (node) {
669                 cur = rb_entry(node, struct tree_mod_elem, node);
670                 if (cur->logical < start) {
671                         node = node->rb_left;
672                 } else if (cur->logical > start) {
673                         node = node->rb_right;
674                 } else if (cur->seq < min_seq) {
675                         node = node->rb_left;
676                 } else if (!smallest) {
677                         /* we want the node with the highest seq */
678                         if (found)
679                                 BUG_ON(found->seq > cur->seq);
680                         found = cur;
681                         node = node->rb_left;
682                 } else if (cur->seq > min_seq) {
683                         /* we want the node with the smallest seq */
684                         if (found)
685                                 BUG_ON(found->seq < cur->seq);
686                         found = cur;
687                         node = node->rb_right;
688                 } else {
689                         found = cur;
690                         break;
691                 }
692         }
693         read_unlock(&fs_info->tree_mod_log_lock);
694
695         return found;
696 }
697
698 /*
699  * this returns the element from the log with the smallest time sequence
700  * value that's in the log (the oldest log item). any element with a time
701  * sequence lower than min_seq will be ignored.
702  */
703 static struct tree_mod_elem *
704 tree_mod_log_search_oldest(struct btrfs_fs_info *fs_info, u64 start,
705                            u64 min_seq)
706 {
707         return __tree_mod_log_search(fs_info, start, min_seq, 1);
708 }
709
710 /*
711  * this returns the element from the log with the largest time sequence
712  * value that's in the log (the most recent log item). any element with
713  * a time sequence lower than min_seq will be ignored.
714  */
715 static struct tree_mod_elem *
716 tree_mod_log_search(struct btrfs_fs_info *fs_info, u64 start, u64 min_seq)
717 {
718         return __tree_mod_log_search(fs_info, start, min_seq, 0);
719 }
720
721 static noinline int
722 tree_mod_log_eb_copy(struct btrfs_fs_info *fs_info, struct extent_buffer *dst,
723                      struct extent_buffer *src, unsigned long dst_offset,
724                      unsigned long src_offset, int nr_items)
725 {
726         int ret = 0;
727         struct tree_mod_elem **tm_list = NULL;
728         struct tree_mod_elem **tm_list_add, **tm_list_rem;
729         int i;
730         int locked = 0;
731
732         if (!tree_mod_need_log(fs_info, NULL))
733                 return 0;
734
735         if (btrfs_header_level(dst) == 0 && btrfs_header_level(src) == 0)
736                 return 0;
737
738         tm_list = kcalloc(nr_items * 2, sizeof(struct tree_mod_elem *),
739                           GFP_NOFS);
740         if (!tm_list)
741                 return -ENOMEM;
742
743         tm_list_add = tm_list;
744         tm_list_rem = tm_list + nr_items;
745         for (i = 0; i < nr_items; i++) {
746                 tm_list_rem[i] = alloc_tree_mod_elem(src, i + src_offset,
747                     MOD_LOG_KEY_REMOVE, GFP_NOFS);
748                 if (!tm_list_rem[i]) {
749                         ret = -ENOMEM;
750                         goto free_tms;
751                 }
752
753                 tm_list_add[i] = alloc_tree_mod_elem(dst, i + dst_offset,
754                     MOD_LOG_KEY_ADD, GFP_NOFS);
755                 if (!tm_list_add[i]) {
756                         ret = -ENOMEM;
757                         goto free_tms;
758                 }
759         }
760
761         if (tree_mod_dont_log(fs_info, NULL))
762                 goto free_tms;
763         locked = 1;
764
765         for (i = 0; i < nr_items; i++) {
766                 ret = __tree_mod_log_insert(fs_info, tm_list_rem[i]);
767                 if (ret)
768                         goto free_tms;
769                 ret = __tree_mod_log_insert(fs_info, tm_list_add[i]);
770                 if (ret)
771                         goto free_tms;
772         }
773
774         write_unlock(&fs_info->tree_mod_log_lock);
775         kfree(tm_list);
776
777         return 0;
778
779 free_tms:
780         for (i = 0; i < nr_items * 2; i++) {
781                 if (tm_list[i] && !RB_EMPTY_NODE(&tm_list[i]->node))
782                         rb_erase(&tm_list[i]->node, &fs_info->tree_mod_log);
783                 kfree(tm_list[i]);
784         }
785         if (locked)
786                 write_unlock(&fs_info->tree_mod_log_lock);
787         kfree(tm_list);
788
789         return ret;
790 }
791
792 static noinline int tree_mod_log_free_eb(struct extent_buffer *eb)
793 {
794         struct tree_mod_elem **tm_list = NULL;
795         int nritems = 0;
796         int i;
797         int ret = 0;
798
799         if (btrfs_header_level(eb) == 0)
800                 return 0;
801
802         if (!tree_mod_need_log(eb->fs_info, NULL))
803                 return 0;
804
805         nritems = btrfs_header_nritems(eb);
806         tm_list = kcalloc(nritems, sizeof(struct tree_mod_elem *), GFP_NOFS);
807         if (!tm_list)
808                 return -ENOMEM;
809
810         for (i = 0; i < nritems; i++) {
811                 tm_list[i] = alloc_tree_mod_elem(eb, i,
812                     MOD_LOG_KEY_REMOVE_WHILE_FREEING, GFP_NOFS);
813                 if (!tm_list[i]) {
814                         ret = -ENOMEM;
815                         goto free_tms;
816                 }
817         }
818
819         if (tree_mod_dont_log(eb->fs_info, eb))
820                 goto free_tms;
821
822         ret = __tree_mod_log_free_eb(eb->fs_info, tm_list, nritems);
823         write_unlock(&eb->fs_info->tree_mod_log_lock);
824         if (ret)
825                 goto free_tms;
826         kfree(tm_list);
827
828         return 0;
829
830 free_tms:
831         for (i = 0; i < nritems; i++)
832                 kfree(tm_list[i]);
833         kfree(tm_list);
834
835         return ret;
836 }
837
838 /*
839  * check if the tree block can be shared by multiple trees
840  */
841 int btrfs_block_can_be_shared(struct btrfs_root *root,
842                               struct extent_buffer *buf)
843 {
844         /*
845          * Tree blocks not in reference counted trees and tree roots
846          * are never shared. If a block was allocated after the last
847          * snapshot and the block was not allocated by tree relocation,
848          * we know the block is not shared.
849          */
850         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
851             buf != root->node && buf != root->commit_root &&
852             (btrfs_header_generation(buf) <=
853              btrfs_root_last_snapshot(&root->root_item) ||
854              btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)))
855                 return 1;
856
857         return 0;
858 }
859
860 static noinline int update_ref_for_cow(struct btrfs_trans_handle *trans,
861                                        struct btrfs_root *root,
862                                        struct extent_buffer *buf,
863                                        struct extent_buffer *cow,
864                                        int *last_ref)
865 {
866         struct btrfs_fs_info *fs_info = root->fs_info;
867         u64 refs;
868         u64 owner;
869         u64 flags;
870         u64 new_flags = 0;
871         int ret;
872
873         /*
874          * Backrefs update rules:
875          *
876          * Always use full backrefs for extent pointers in tree block
877          * allocated by tree relocation.
878          *
879          * If a shared tree block is no longer referenced by its owner
880          * tree (btrfs_header_owner(buf) == root->root_key.objectid),
881          * use full backrefs for extent pointers in tree block.
882          *
883          * If a tree block is been relocating
884          * (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID),
885          * use full backrefs for extent pointers in tree block.
886          * The reason for this is some operations (such as drop tree)
887          * are only allowed for blocks use full backrefs.
888          */
889
890         if (btrfs_block_can_be_shared(root, buf)) {
891                 ret = btrfs_lookup_extent_info(trans, fs_info, buf->start,
892                                                btrfs_header_level(buf), 1,
893                                                &refs, &flags);
894                 if (ret)
895                         return ret;
896                 if (refs == 0) {
897                         ret = -EROFS;
898                         btrfs_handle_fs_error(fs_info, ret, NULL);
899                         return ret;
900                 }
901         } else {
902                 refs = 1;
903                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
904                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
905                         flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
906                 else
907                         flags = 0;
908         }
909
910         owner = btrfs_header_owner(buf);
911         BUG_ON(owner == BTRFS_TREE_RELOC_OBJECTID &&
912                !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
913
914         if (refs > 1) {
915                 if ((owner == root->root_key.objectid ||
916                      root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) &&
917                     !(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) {
918                         ret = btrfs_inc_ref(trans, root, buf, 1);
919                         if (ret)
920                                 return ret;
921
922                         if (root->root_key.objectid ==
923                             BTRFS_TREE_RELOC_OBJECTID) {
924                                 ret = btrfs_dec_ref(trans, root, buf, 0);
925                                 if (ret)
926                                         return ret;
927                                 ret = btrfs_inc_ref(trans, root, cow, 1);
928                                 if (ret)
929                                         return ret;
930                         }
931                         new_flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
932                 } else {
933
934                         if (root->root_key.objectid ==
935                             BTRFS_TREE_RELOC_OBJECTID)
936                                 ret = btrfs_inc_ref(trans, root, cow, 1);
937                         else
938                                 ret = btrfs_inc_ref(trans, root, cow, 0);
939                         if (ret)
940                                 return ret;
941                 }
942                 if (new_flags != 0) {
943                         int level = btrfs_header_level(buf);
944
945                         ret = btrfs_set_disk_extent_flags(trans, fs_info,
946                                                           buf->start,
947                                                           buf->len,
948                                                           new_flags, level, 0);
949                         if (ret)
950                                 return ret;
951                 }
952         } else {
953                 if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
954                         if (root->root_key.objectid ==
955                             BTRFS_TREE_RELOC_OBJECTID)
956                                 ret = btrfs_inc_ref(trans, root, cow, 1);
957                         else
958                                 ret = btrfs_inc_ref(trans, root, cow, 0);
959                         if (ret)
960                                 return ret;
961                         ret = btrfs_dec_ref(trans, root, buf, 1);
962                         if (ret)
963                                 return ret;
964                 }
965                 clean_tree_block(fs_info, buf);
966                 *last_ref = 1;
967         }
968         return 0;
969 }
970
971 /*
972  * does the dirty work in cow of a single block.  The parent block (if
973  * supplied) is updated to point to the new cow copy.  The new buffer is marked
974  * dirty and returned locked.  If you modify the block it needs to be marked
975  * dirty again.
976  *
977  * search_start -- an allocation hint for the new block
978  *
979  * empty_size -- a hint that you plan on doing more cow.  This is the size in
980  * bytes the allocator should try to find free next to the block it returns.
981  * This is just a hint and may be ignored by the allocator.
982  */
983 static noinline int __btrfs_cow_block(struct btrfs_trans_handle *trans,
984                              struct btrfs_root *root,
985                              struct extent_buffer *buf,
986                              struct extent_buffer *parent, int parent_slot,
987                              struct extent_buffer **cow_ret,
988                              u64 search_start, u64 empty_size)
989 {
990         struct btrfs_fs_info *fs_info = root->fs_info;
991         struct btrfs_disk_key disk_key;
992         struct extent_buffer *cow;
993         int level, ret;
994         int last_ref = 0;
995         int unlock_orig = 0;
996         u64 parent_start = 0;
997
998         if (*cow_ret == buf)
999                 unlock_orig = 1;
1000
1001         btrfs_assert_tree_locked(buf);
1002
1003         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1004                 trans->transid != fs_info->running_transaction->transid);
1005         WARN_ON(test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
1006                 trans->transid != root->last_trans);
1007
1008         level = btrfs_header_level(buf);
1009
1010         if (level == 0)
1011                 btrfs_item_key(buf, &disk_key, 0);
1012         else
1013                 btrfs_node_key(buf, &disk_key, 0);
1014
1015         if ((root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) && parent)
1016                 parent_start = parent->start;
1017
1018         /*
1019          * If we are COWing a node/leaf from the extent, chunk or device trees,
1020          * make sure that we do not finish block group creation of pending block
1021          * groups. We do this to avoid a deadlock.
1022          * COWing can result in allocation of a new chunk, and flushing pending
1023          * block groups (btrfs_create_pending_block_groups()) can be triggered
1024          * when finishing allocation of a new chunk. Creation of a pending block
1025          * group modifies the extent, chunk and device trees, therefore we could
1026          * deadlock with ourselves since we are holding a lock on an extent
1027          * buffer that btrfs_create_pending_block_groups() may try to COW later.
1028          */
1029         if (root == fs_info->extent_root ||
1030             root == fs_info->chunk_root ||
1031             root == fs_info->dev_root)
1032                 trans->can_flush_pending_bgs = false;
1033
1034         cow = btrfs_alloc_tree_block(trans, root, parent_start,
1035                         root->root_key.objectid, &disk_key, level,
1036                         search_start, empty_size);
1037         trans->can_flush_pending_bgs = true;
1038         if (IS_ERR(cow))
1039                 return PTR_ERR(cow);
1040
1041         /* cow is set to blocking by btrfs_init_new_buffer */
1042
1043         copy_extent_buffer_full(cow, buf);
1044         btrfs_set_header_bytenr(cow, cow->start);
1045         btrfs_set_header_generation(cow, trans->transid);
1046         btrfs_set_header_backref_rev(cow, BTRFS_MIXED_BACKREF_REV);
1047         btrfs_clear_header_flag(cow, BTRFS_HEADER_FLAG_WRITTEN |
1048                                      BTRFS_HEADER_FLAG_RELOC);
1049         if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID)
1050                 btrfs_set_header_flag(cow, BTRFS_HEADER_FLAG_RELOC);
1051         else
1052                 btrfs_set_header_owner(cow, root->root_key.objectid);
1053
1054         write_extent_buffer_fsid(cow, fs_info->fs_devices->metadata_uuid);
1055
1056         ret = update_ref_for_cow(trans, root, buf, cow, &last_ref);
1057         if (ret) {
1058                 btrfs_abort_transaction(trans, ret);
1059                 return ret;
1060         }
1061
1062         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state)) {
1063                 ret = btrfs_reloc_cow_block(trans, root, buf, cow);
1064                 if (ret) {
1065                         btrfs_abort_transaction(trans, ret);
1066                         return ret;
1067                 }
1068         }
1069
1070         if (buf == root->node) {
1071                 WARN_ON(parent && parent != buf);
1072                 if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
1073                     btrfs_header_backref_rev(buf) < BTRFS_MIXED_BACKREF_REV)
1074                         parent_start = buf->start;
1075
1076                 extent_buffer_get(cow);
1077                 ret = tree_mod_log_insert_root(root->node, cow, 1);
1078                 BUG_ON(ret < 0);
1079                 rcu_assign_pointer(root->node, cow);
1080
1081                 btrfs_free_tree_block(trans, root, buf, parent_start,
1082                                       last_ref);
1083                 free_extent_buffer(buf);
1084                 add_root_to_dirty_list(root);
1085         } else {
1086                 WARN_ON(trans->transid != btrfs_header_generation(parent));
1087                 tree_mod_log_insert_key(parent, parent_slot,
1088                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1089                 btrfs_set_node_blockptr(parent, parent_slot,
1090                                         cow->start);
1091                 btrfs_set_node_ptr_generation(parent, parent_slot,
1092                                               trans->transid);
1093                 btrfs_mark_buffer_dirty(parent);
1094                 if (last_ref) {
1095                         ret = tree_mod_log_free_eb(buf);
1096                         if (ret) {
1097                                 btrfs_abort_transaction(trans, ret);
1098                                 return ret;
1099                         }
1100                 }
1101                 btrfs_free_tree_block(trans, root, buf, parent_start,
1102                                       last_ref);
1103         }
1104         if (unlock_orig)
1105                 btrfs_tree_unlock(buf);
1106         free_extent_buffer_stale(buf);
1107         btrfs_mark_buffer_dirty(cow);
1108         *cow_ret = cow;
1109         return 0;
1110 }
1111
1112 /*
1113  * returns the logical address of the oldest predecessor of the given root.
1114  * entries older than time_seq are ignored.
1115  */
1116 static struct tree_mod_elem *__tree_mod_log_oldest_root(
1117                 struct extent_buffer *eb_root, u64 time_seq)
1118 {
1119         struct tree_mod_elem *tm;
1120         struct tree_mod_elem *found = NULL;
1121         u64 root_logical = eb_root->start;
1122         int looped = 0;
1123
1124         if (!time_seq)
1125                 return NULL;
1126
1127         /*
1128          * the very last operation that's logged for a root is the
1129          * replacement operation (if it is replaced at all). this has
1130          * the logical address of the *new* root, making it the very
1131          * first operation that's logged for this root.
1132          */
1133         while (1) {
1134                 tm = tree_mod_log_search_oldest(eb_root->fs_info, root_logical,
1135                                                 time_seq);
1136                 if (!looped && !tm)
1137                         return NULL;
1138                 /*
1139                  * if there are no tree operation for the oldest root, we simply
1140                  * return it. this should only happen if that (old) root is at
1141                  * level 0.
1142                  */
1143                 if (!tm)
1144                         break;
1145
1146                 /*
1147                  * if there's an operation that's not a root replacement, we
1148                  * found the oldest version of our root. normally, we'll find a
1149                  * MOD_LOG_KEY_REMOVE_WHILE_FREEING operation here.
1150                  */
1151                 if (tm->op != MOD_LOG_ROOT_REPLACE)
1152                         break;
1153
1154                 found = tm;
1155                 root_logical = tm->old_root.logical;
1156                 looped = 1;
1157         }
1158
1159         /* if there's no old root to return, return what we found instead */
1160         if (!found)
1161                 found = tm;
1162
1163         return found;
1164 }
1165
1166 /*
1167  * tm is a pointer to the first operation to rewind within eb. then, all
1168  * previous operations will be rewound (until we reach something older than
1169  * time_seq).
1170  */
1171 static void
1172 __tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct extent_buffer *eb,
1173                       u64 time_seq, struct tree_mod_elem *first_tm)
1174 {
1175         u32 n;
1176         struct rb_node *next;
1177         struct tree_mod_elem *tm = first_tm;
1178         unsigned long o_dst;
1179         unsigned long o_src;
1180         unsigned long p_size = sizeof(struct btrfs_key_ptr);
1181
1182         n = btrfs_header_nritems(eb);
1183         read_lock(&fs_info->tree_mod_log_lock);
1184         while (tm && tm->seq >= time_seq) {
1185                 /*
1186                  * all the operations are recorded with the operator used for
1187                  * the modification. as we're going backwards, we do the
1188                  * opposite of each operation here.
1189                  */
1190                 switch (tm->op) {
1191                 case MOD_LOG_KEY_REMOVE_WHILE_FREEING:
1192                         BUG_ON(tm->slot < n);
1193                         /* Fallthrough */
1194                 case MOD_LOG_KEY_REMOVE_WHILE_MOVING:
1195                 case MOD_LOG_KEY_REMOVE:
1196                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1197                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1198                         btrfs_set_node_ptr_generation(eb, tm->slot,
1199                                                       tm->generation);
1200                         n++;
1201                         break;
1202                 case MOD_LOG_KEY_REPLACE:
1203                         BUG_ON(tm->slot >= n);
1204                         btrfs_set_node_key(eb, &tm->key, tm->slot);
1205                         btrfs_set_node_blockptr(eb, tm->slot, tm->blockptr);
1206                         btrfs_set_node_ptr_generation(eb, tm->slot,
1207                                                       tm->generation);
1208                         break;
1209                 case MOD_LOG_KEY_ADD:
1210                         /* if a move operation is needed it's in the log */
1211                         n--;
1212                         break;
1213                 case MOD_LOG_MOVE_KEYS:
1214                         o_dst = btrfs_node_key_ptr_offset(tm->slot);
1215                         o_src = btrfs_node_key_ptr_offset(tm->move.dst_slot);
1216                         memmove_extent_buffer(eb, o_dst, o_src,
1217                                               tm->move.nr_items * p_size);
1218                         break;
1219                 case MOD_LOG_ROOT_REPLACE:
1220                         /*
1221                          * this operation is special. for roots, this must be
1222                          * handled explicitly before rewinding.
1223                          * for non-roots, this operation may exist if the node
1224                          * was a root: root A -> child B; then A gets empty and
1225                          * B is promoted to the new root. in the mod log, we'll
1226                          * have a root-replace operation for B, a tree block
1227                          * that is no root. we simply ignore that operation.
1228                          */
1229                         break;
1230                 }
1231                 next = rb_next(&tm->node);
1232                 if (!next)
1233                         break;
1234                 tm = rb_entry(next, struct tree_mod_elem, node);
1235                 if (tm->logical != first_tm->logical)
1236                         break;
1237         }
1238         read_unlock(&fs_info->tree_mod_log_lock);
1239         btrfs_set_header_nritems(eb, n);
1240 }
1241
1242 /*
1243  * Called with eb read locked. If the buffer cannot be rewound, the same buffer
1244  * is returned. If rewind operations happen, a fresh buffer is returned. The
1245  * returned buffer is always read-locked. If the returned buffer is not the
1246  * input buffer, the lock on the input buffer is released and the input buffer
1247  * is freed (its refcount is decremented).
1248  */
1249 static struct extent_buffer *
1250 tree_mod_log_rewind(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
1251                     struct extent_buffer *eb, u64 time_seq)
1252 {
1253         struct extent_buffer *eb_rewin;
1254         struct tree_mod_elem *tm;
1255
1256         if (!time_seq)
1257                 return eb;
1258
1259         if (btrfs_header_level(eb) == 0)
1260                 return eb;
1261
1262         tm = tree_mod_log_search(fs_info, eb->start, time_seq);
1263         if (!tm)
1264                 return eb;
1265
1266         btrfs_set_path_blocking(path);
1267         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
1268
1269         if (tm->op == MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1270                 BUG_ON(tm->slot != 0);
1271                 eb_rewin = alloc_dummy_extent_buffer(fs_info, eb->start);
1272                 if (!eb_rewin) {
1273                         btrfs_tree_read_unlock_blocking(eb);
1274                         free_extent_buffer(eb);
1275                         return NULL;
1276                 }
1277                 btrfs_set_header_bytenr(eb_rewin, eb->start);
1278                 btrfs_set_header_backref_rev(eb_rewin,
1279                                              btrfs_header_backref_rev(eb));
1280                 btrfs_set_header_owner(eb_rewin, btrfs_header_owner(eb));
1281                 btrfs_set_header_level(eb_rewin, btrfs_header_level(eb));
1282         } else {
1283                 eb_rewin = btrfs_clone_extent_buffer(eb);
1284                 if (!eb_rewin) {
1285                         btrfs_tree_read_unlock_blocking(eb);
1286                         free_extent_buffer(eb);
1287                         return NULL;
1288                 }
1289         }
1290
1291         btrfs_tree_read_unlock_blocking(eb);
1292         free_extent_buffer(eb);
1293
1294         btrfs_tree_read_lock(eb_rewin);
1295         __tree_mod_log_rewind(fs_info, eb_rewin, time_seq, tm);
1296         WARN_ON(btrfs_header_nritems(eb_rewin) >
1297                 BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1298
1299         return eb_rewin;
1300 }
1301
1302 /*
1303  * get_old_root() rewinds the state of @root's root node to the given @time_seq
1304  * value. If there are no changes, the current root->root_node is returned. If
1305  * anything changed in between, there's a fresh buffer allocated on which the
1306  * rewind operations are done. In any case, the returned buffer is read locked.
1307  * Returns NULL on error (with no locks held).
1308  */
1309 static inline struct extent_buffer *
1310 get_old_root(struct btrfs_root *root, u64 time_seq)
1311 {
1312         struct btrfs_fs_info *fs_info = root->fs_info;
1313         struct tree_mod_elem *tm;
1314         struct extent_buffer *eb = NULL;
1315         struct extent_buffer *eb_root;
1316         struct extent_buffer *old;
1317         struct tree_mod_root *old_root = NULL;
1318         u64 old_generation = 0;
1319         u64 logical;
1320         int level;
1321
1322         eb_root = btrfs_read_lock_root_node(root);
1323         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1324         if (!tm)
1325                 return eb_root;
1326
1327         if (tm->op == MOD_LOG_ROOT_REPLACE) {
1328                 old_root = &tm->old_root;
1329                 old_generation = tm->generation;
1330                 logical = old_root->logical;
1331                 level = old_root->level;
1332         } else {
1333                 logical = eb_root->start;
1334                 level = btrfs_header_level(eb_root);
1335         }
1336
1337         tm = tree_mod_log_search(fs_info, logical, time_seq);
1338         if (old_root && tm && tm->op != MOD_LOG_KEY_REMOVE_WHILE_FREEING) {
1339                 btrfs_tree_read_unlock(eb_root);
1340                 free_extent_buffer(eb_root);
1341                 old = read_tree_block(fs_info, logical, 0, level, NULL);
1342                 if (WARN_ON(IS_ERR(old) || !extent_buffer_uptodate(old))) {
1343                         if (!IS_ERR(old))
1344                                 free_extent_buffer(old);
1345                         btrfs_warn(fs_info,
1346                                    "failed to read tree block %llu from get_old_root",
1347                                    logical);
1348                 } else {
1349                         eb = btrfs_clone_extent_buffer(old);
1350                         free_extent_buffer(old);
1351                 }
1352         } else if (old_root) {
1353                 btrfs_tree_read_unlock(eb_root);
1354                 free_extent_buffer(eb_root);
1355                 eb = alloc_dummy_extent_buffer(fs_info, logical);
1356         } else {
1357                 btrfs_set_lock_blocking_rw(eb_root, BTRFS_READ_LOCK);
1358                 eb = btrfs_clone_extent_buffer(eb_root);
1359                 btrfs_tree_read_unlock_blocking(eb_root);
1360                 free_extent_buffer(eb_root);
1361         }
1362
1363         if (!eb)
1364                 return NULL;
1365         btrfs_tree_read_lock(eb);
1366         if (old_root) {
1367                 btrfs_set_header_bytenr(eb, eb->start);
1368                 btrfs_set_header_backref_rev(eb, BTRFS_MIXED_BACKREF_REV);
1369                 btrfs_set_header_owner(eb, btrfs_header_owner(eb_root));
1370                 btrfs_set_header_level(eb, old_root->level);
1371                 btrfs_set_header_generation(eb, old_generation);
1372         }
1373         if (tm)
1374                 __tree_mod_log_rewind(fs_info, eb, time_seq, tm);
1375         else
1376                 WARN_ON(btrfs_header_level(eb) != 0);
1377         WARN_ON(btrfs_header_nritems(eb) > BTRFS_NODEPTRS_PER_BLOCK(fs_info));
1378
1379         return eb;
1380 }
1381
1382 int btrfs_old_root_level(struct btrfs_root *root, u64 time_seq)
1383 {
1384         struct tree_mod_elem *tm;
1385         int level;
1386         struct extent_buffer *eb_root = btrfs_root_node(root);
1387
1388         tm = __tree_mod_log_oldest_root(eb_root, time_seq);
1389         if (tm && tm->op == MOD_LOG_ROOT_REPLACE) {
1390                 level = tm->old_root.level;
1391         } else {
1392                 level = btrfs_header_level(eb_root);
1393         }
1394         free_extent_buffer(eb_root);
1395
1396         return level;
1397 }
1398
1399 static inline int should_cow_block(struct btrfs_trans_handle *trans,
1400                                    struct btrfs_root *root,
1401                                    struct extent_buffer *buf)
1402 {
1403         if (btrfs_is_testing(root->fs_info))
1404                 return 0;
1405
1406         /* Ensure we can see the FORCE_COW bit */
1407         smp_mb__before_atomic();
1408
1409         /*
1410          * We do not need to cow a block if
1411          * 1) this block is not created or changed in this transaction;
1412          * 2) this block does not belong to TREE_RELOC tree;
1413          * 3) the root is not forced COW.
1414          *
1415          * What is forced COW:
1416          *    when we create snapshot during committing the transaction,
1417          *    after we've finished coping src root, we must COW the shared
1418          *    block to ensure the metadata consistency.
1419          */
1420         if (btrfs_header_generation(buf) == trans->transid &&
1421             !btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN) &&
1422             !(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID &&
1423               btrfs_header_flag(buf, BTRFS_HEADER_FLAG_RELOC)) &&
1424             !test_bit(BTRFS_ROOT_FORCE_COW, &root->state))
1425                 return 0;
1426         return 1;
1427 }
1428
1429 /*
1430  * cows a single block, see __btrfs_cow_block for the real work.
1431  * This version of it has extra checks so that a block isn't COWed more than
1432  * once per transaction, as long as it hasn't been written yet
1433  */
1434 noinline int btrfs_cow_block(struct btrfs_trans_handle *trans,
1435                     struct btrfs_root *root, struct extent_buffer *buf,
1436                     struct extent_buffer *parent, int parent_slot,
1437                     struct extent_buffer **cow_ret)
1438 {
1439         struct btrfs_fs_info *fs_info = root->fs_info;
1440         u64 search_start;
1441         int ret;
1442
1443         if (test_bit(BTRFS_ROOT_DELETING, &root->state))
1444                 btrfs_err(fs_info,
1445                         "COW'ing blocks on a fs root that's being dropped");
1446
1447         if (trans->transaction != fs_info->running_transaction)
1448                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1449                        trans->transid,
1450                        fs_info->running_transaction->transid);
1451
1452         if (trans->transid != fs_info->generation)
1453                 WARN(1, KERN_CRIT "trans %llu running %llu\n",
1454                        trans->transid, fs_info->generation);
1455
1456         if (!should_cow_block(trans, root, buf)) {
1457                 trans->dirty = true;
1458                 *cow_ret = buf;
1459                 return 0;
1460         }
1461
1462         search_start = buf->start & ~((u64)SZ_1G - 1);
1463
1464         if (parent)
1465                 btrfs_set_lock_blocking(parent);
1466         btrfs_set_lock_blocking(buf);
1467
1468         ret = __btrfs_cow_block(trans, root, buf, parent,
1469                                  parent_slot, cow_ret, search_start, 0);
1470
1471         trace_btrfs_cow_block(root, buf, *cow_ret);
1472
1473         return ret;
1474 }
1475
1476 /*
1477  * helper function for defrag to decide if two blocks pointed to by a
1478  * node are actually close by
1479  */
1480 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
1481 {
1482         if (blocknr < other && other - (blocknr + blocksize) < 32768)
1483                 return 1;
1484         if (blocknr > other && blocknr - (other + blocksize) < 32768)
1485                 return 1;
1486         return 0;
1487 }
1488
1489 /*
1490  * compare two keys in a memcmp fashion
1491  */
1492 static int comp_keys(const struct btrfs_disk_key *disk,
1493                      const struct btrfs_key *k2)
1494 {
1495         struct btrfs_key k1;
1496
1497         btrfs_disk_key_to_cpu(&k1, disk);
1498
1499         return btrfs_comp_cpu_keys(&k1, k2);
1500 }
1501
1502 /*
1503  * same as comp_keys only with two btrfs_key's
1504  */
1505 int btrfs_comp_cpu_keys(const struct btrfs_key *k1, const struct btrfs_key *k2)
1506 {
1507         if (k1->objectid > k2->objectid)
1508                 return 1;
1509         if (k1->objectid < k2->objectid)
1510                 return -1;
1511         if (k1->type > k2->type)
1512                 return 1;
1513         if (k1->type < k2->type)
1514                 return -1;
1515         if (k1->offset > k2->offset)
1516                 return 1;
1517         if (k1->offset < k2->offset)
1518                 return -1;
1519         return 0;
1520 }
1521
1522 /*
1523  * this is used by the defrag code to go through all the
1524  * leaves pointed to by a node and reallocate them so that
1525  * disk order is close to key order
1526  */
1527 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
1528                        struct btrfs_root *root, struct extent_buffer *parent,
1529                        int start_slot, u64 *last_ret,
1530                        struct btrfs_key *progress)
1531 {
1532         struct btrfs_fs_info *fs_info = root->fs_info;
1533         struct extent_buffer *cur;
1534         u64 blocknr;
1535         u64 gen;
1536         u64 search_start = *last_ret;
1537         u64 last_block = 0;
1538         u64 other;
1539         u32 parent_nritems;
1540         int end_slot;
1541         int i;
1542         int err = 0;
1543         int parent_level;
1544         int uptodate;
1545         u32 blocksize;
1546         int progress_passed = 0;
1547         struct btrfs_disk_key disk_key;
1548
1549         parent_level = btrfs_header_level(parent);
1550
1551         WARN_ON(trans->transaction != fs_info->running_transaction);
1552         WARN_ON(trans->transid != fs_info->generation);
1553
1554         parent_nritems = btrfs_header_nritems(parent);
1555         blocksize = fs_info->nodesize;
1556         end_slot = parent_nritems - 1;
1557
1558         if (parent_nritems <= 1)
1559                 return 0;
1560
1561         btrfs_set_lock_blocking(parent);
1562
1563         for (i = start_slot; i <= end_slot; i++) {
1564                 struct btrfs_key first_key;
1565                 int close = 1;
1566
1567                 btrfs_node_key(parent, &disk_key, i);
1568                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
1569                         continue;
1570
1571                 progress_passed = 1;
1572                 blocknr = btrfs_node_blockptr(parent, i);
1573                 gen = btrfs_node_ptr_generation(parent, i);
1574                 btrfs_node_key_to_cpu(parent, &first_key, i);
1575                 if (last_block == 0)
1576                         last_block = blocknr;
1577
1578                 if (i > 0) {
1579                         other = btrfs_node_blockptr(parent, i - 1);
1580                         close = close_blocks(blocknr, other, blocksize);
1581                 }
1582                 if (!close && i < end_slot) {
1583                         other = btrfs_node_blockptr(parent, i + 1);
1584                         close = close_blocks(blocknr, other, blocksize);
1585                 }
1586                 if (close) {
1587                         last_block = blocknr;
1588                         continue;
1589                 }
1590
1591                 cur = find_extent_buffer(fs_info, blocknr);
1592                 if (cur)
1593                         uptodate = btrfs_buffer_uptodate(cur, gen, 0);
1594                 else
1595                         uptodate = 0;
1596                 if (!cur || !uptodate) {
1597                         if (!cur) {
1598                                 cur = read_tree_block(fs_info, blocknr, gen,
1599                                                       parent_level - 1,
1600                                                       &first_key);
1601                                 if (IS_ERR(cur)) {
1602                                         return PTR_ERR(cur);
1603                                 } else if (!extent_buffer_uptodate(cur)) {
1604                                         free_extent_buffer(cur);
1605                                         return -EIO;
1606                                 }
1607                         } else if (!uptodate) {
1608                                 err = btrfs_read_buffer(cur, gen,
1609                                                 parent_level - 1,&first_key);
1610                                 if (err) {
1611                                         free_extent_buffer(cur);
1612                                         return err;
1613                                 }
1614                         }
1615                 }
1616                 if (search_start == 0)
1617                         search_start = last_block;
1618
1619                 btrfs_tree_lock(cur);
1620                 btrfs_set_lock_blocking(cur);
1621                 err = __btrfs_cow_block(trans, root, cur, parent, i,
1622                                         &cur, search_start,
1623                                         min(16 * blocksize,
1624                                             (end_slot - i) * blocksize));
1625                 if (err) {
1626                         btrfs_tree_unlock(cur);
1627                         free_extent_buffer(cur);
1628                         break;
1629                 }
1630                 search_start = cur->start;
1631                 last_block = cur->start;
1632                 *last_ret = search_start;
1633                 btrfs_tree_unlock(cur);
1634                 free_extent_buffer(cur);
1635         }
1636         return err;
1637 }
1638
1639 /*
1640  * search for key in the extent_buffer.  The items start at offset p,
1641  * and they are item_size apart.  There are 'max' items in p.
1642  *
1643  * the slot in the array is returned via slot, and it points to
1644  * the place where you would insert key if it is not found in
1645  * the array.
1646  *
1647  * slot may point to max if the key is bigger than all of the keys
1648  */
1649 static noinline int generic_bin_search(struct extent_buffer *eb,
1650                                        unsigned long p, int item_size,
1651                                        const struct btrfs_key *key,
1652                                        int max, int *slot)
1653 {
1654         int low = 0;
1655         int high = max;
1656         int mid;
1657         int ret;
1658         struct btrfs_disk_key *tmp = NULL;
1659         struct btrfs_disk_key unaligned;
1660         unsigned long offset;
1661         char *kaddr = NULL;
1662         unsigned long map_start = 0;
1663         unsigned long map_len = 0;
1664         int err;
1665
1666         if (low > high) {
1667                 btrfs_err(eb->fs_info,
1668                  "%s: low (%d) > high (%d) eb %llu owner %llu level %d",
1669                           __func__, low, high, eb->start,
1670                           btrfs_header_owner(eb), btrfs_header_level(eb));
1671                 return -EINVAL;
1672         }
1673
1674         while (low < high) {
1675                 mid = (low + high) / 2;
1676                 offset = p + mid * item_size;
1677
1678                 if (!kaddr || offset < map_start ||
1679                     (offset + sizeof(struct btrfs_disk_key)) >
1680                     map_start + map_len) {
1681
1682                         err = map_private_extent_buffer(eb, offset,
1683                                                 sizeof(struct btrfs_disk_key),
1684                                                 &kaddr, &map_start, &map_len);
1685
1686                         if (!err) {
1687                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
1688                                                         map_start);
1689                         } else if (err == 1) {
1690                                 read_extent_buffer(eb, &unaligned,
1691                                                    offset, sizeof(unaligned));
1692                                 tmp = &unaligned;
1693                         } else {
1694                                 return err;
1695                         }
1696
1697                 } else {
1698                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
1699                                                         map_start);
1700                 }
1701                 ret = comp_keys(tmp, key);
1702
1703                 if (ret < 0)
1704                         low = mid + 1;
1705                 else if (ret > 0)
1706                         high = mid;
1707                 else {
1708                         *slot = mid;
1709                         return 0;
1710                 }
1711         }
1712         *slot = low;
1713         return 1;
1714 }
1715
1716 /*
1717  * simple bin_search frontend that does the right thing for
1718  * leaves vs nodes
1719  */
1720 int btrfs_bin_search(struct extent_buffer *eb, const struct btrfs_key *key,
1721                      int level, int *slot)
1722 {
1723         if (level == 0)
1724                 return generic_bin_search(eb,
1725                                           offsetof(struct btrfs_leaf, items),
1726                                           sizeof(struct btrfs_item),
1727                                           key, btrfs_header_nritems(eb),
1728                                           slot);
1729         else
1730                 return generic_bin_search(eb,
1731                                           offsetof(struct btrfs_node, ptrs),
1732                                           sizeof(struct btrfs_key_ptr),
1733                                           key, btrfs_header_nritems(eb),
1734                                           slot);
1735 }
1736
1737 static void root_add_used(struct btrfs_root *root, u32 size)
1738 {
1739         spin_lock(&root->accounting_lock);
1740         btrfs_set_root_used(&root->root_item,
1741                             btrfs_root_used(&root->root_item) + size);
1742         spin_unlock(&root->accounting_lock);
1743 }
1744
1745 static void root_sub_used(struct btrfs_root *root, u32 size)
1746 {
1747         spin_lock(&root->accounting_lock);
1748         btrfs_set_root_used(&root->root_item,
1749                             btrfs_root_used(&root->root_item) - size);
1750         spin_unlock(&root->accounting_lock);
1751 }
1752
1753 /* given a node and slot number, this reads the blocks it points to.  The
1754  * extent buffer is returned with a reference taken (but unlocked).
1755  */
1756 static noinline struct extent_buffer *
1757 read_node_slot(struct btrfs_fs_info *fs_info, struct extent_buffer *parent,
1758                int slot)
1759 {
1760         int level = btrfs_header_level(parent);
1761         struct extent_buffer *eb;
1762         struct btrfs_key first_key;
1763
1764         if (slot < 0 || slot >= btrfs_header_nritems(parent))
1765                 return ERR_PTR(-ENOENT);
1766
1767         BUG_ON(level == 0);
1768
1769         btrfs_node_key_to_cpu(parent, &first_key, slot);
1770         eb = read_tree_block(fs_info, btrfs_node_blockptr(parent, slot),
1771                              btrfs_node_ptr_generation(parent, slot),
1772                              level - 1, &first_key);
1773         if (!IS_ERR(eb) && !extent_buffer_uptodate(eb)) {
1774                 free_extent_buffer(eb);
1775                 eb = ERR_PTR(-EIO);
1776         }
1777
1778         return eb;
1779 }
1780
1781 /*
1782  * node level balancing, used to make sure nodes are in proper order for
1783  * item deletion.  We balance from the top down, so we have to make sure
1784  * that a deletion won't leave an node completely empty later on.
1785  */
1786 static noinline int balance_level(struct btrfs_trans_handle *trans,
1787                          struct btrfs_root *root,
1788                          struct btrfs_path *path, int level)
1789 {
1790         struct btrfs_fs_info *fs_info = root->fs_info;
1791         struct extent_buffer *right = NULL;
1792         struct extent_buffer *mid;
1793         struct extent_buffer *left = NULL;
1794         struct extent_buffer *parent = NULL;
1795         int ret = 0;
1796         int wret;
1797         int pslot;
1798         int orig_slot = path->slots[level];
1799         u64 orig_ptr;
1800
1801         ASSERT(level > 0);
1802
1803         mid = path->nodes[level];
1804
1805         WARN_ON(path->locks[level] != BTRFS_WRITE_LOCK &&
1806                 path->locks[level] != BTRFS_WRITE_LOCK_BLOCKING);
1807         WARN_ON(btrfs_header_generation(mid) != trans->transid);
1808
1809         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
1810
1811         if (level < BTRFS_MAX_LEVEL - 1) {
1812                 parent = path->nodes[level + 1];
1813                 pslot = path->slots[level + 1];
1814         }
1815
1816         /*
1817          * deal with the case where there is only one pointer in the root
1818          * by promoting the node below to a root
1819          */
1820         if (!parent) {
1821                 struct extent_buffer *child;
1822
1823                 if (btrfs_header_nritems(mid) != 1)
1824                         return 0;
1825
1826                 /* promote the child to a root */
1827                 child = read_node_slot(fs_info, mid, 0);
1828                 if (IS_ERR(child)) {
1829                         ret = PTR_ERR(child);
1830                         btrfs_handle_fs_error(fs_info, ret, NULL);
1831                         goto enospc;
1832                 }
1833
1834                 btrfs_tree_lock(child);
1835                 btrfs_set_lock_blocking(child);
1836                 ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
1837                 if (ret) {
1838                         btrfs_tree_unlock(child);
1839                         free_extent_buffer(child);
1840                         goto enospc;
1841                 }
1842
1843                 ret = tree_mod_log_insert_root(root->node, child, 1);
1844                 BUG_ON(ret < 0);
1845                 rcu_assign_pointer(root->node, child);
1846
1847                 add_root_to_dirty_list(root);
1848                 btrfs_tree_unlock(child);
1849
1850                 path->locks[level] = 0;
1851                 path->nodes[level] = NULL;
1852                 clean_tree_block(fs_info, mid);
1853                 btrfs_tree_unlock(mid);
1854                 /* once for the path */
1855                 free_extent_buffer(mid);
1856
1857                 root_sub_used(root, mid->len);
1858                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1859                 /* once for the root ptr */
1860                 free_extent_buffer_stale(mid);
1861                 return 0;
1862         }
1863         if (btrfs_header_nritems(mid) >
1864             BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 4)
1865                 return 0;
1866
1867         left = read_node_slot(fs_info, parent, pslot - 1);
1868         if (IS_ERR(left))
1869                 left = NULL;
1870
1871         if (left) {
1872                 btrfs_tree_lock(left);
1873                 btrfs_set_lock_blocking(left);
1874                 wret = btrfs_cow_block(trans, root, left,
1875                                        parent, pslot - 1, &left);
1876                 if (wret) {
1877                         ret = wret;
1878                         goto enospc;
1879                 }
1880         }
1881
1882         right = read_node_slot(fs_info, parent, pslot + 1);
1883         if (IS_ERR(right))
1884                 right = NULL;
1885
1886         if (right) {
1887                 btrfs_tree_lock(right);
1888                 btrfs_set_lock_blocking(right);
1889                 wret = btrfs_cow_block(trans, root, right,
1890                                        parent, pslot + 1, &right);
1891                 if (wret) {
1892                         ret = wret;
1893                         goto enospc;
1894                 }
1895         }
1896
1897         /* first, try to make some room in the middle buffer */
1898         if (left) {
1899                 orig_slot += btrfs_header_nritems(left);
1900                 wret = push_node_left(trans, fs_info, left, mid, 1);
1901                 if (wret < 0)
1902                         ret = wret;
1903         }
1904
1905         /*
1906          * then try to empty the right most buffer into the middle
1907          */
1908         if (right) {
1909                 wret = push_node_left(trans, fs_info, mid, right, 1);
1910                 if (wret < 0 && wret != -ENOSPC)
1911                         ret = wret;
1912                 if (btrfs_header_nritems(right) == 0) {
1913                         clean_tree_block(fs_info, right);
1914                         btrfs_tree_unlock(right);
1915                         del_ptr(root, path, level + 1, pslot + 1);
1916                         root_sub_used(root, right->len);
1917                         btrfs_free_tree_block(trans, root, right, 0, 1);
1918                         free_extent_buffer_stale(right);
1919                         right = NULL;
1920                 } else {
1921                         struct btrfs_disk_key right_key;
1922                         btrfs_node_key(right, &right_key, 0);
1923                         ret = tree_mod_log_insert_key(parent, pslot + 1,
1924                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
1925                         BUG_ON(ret < 0);
1926                         btrfs_set_node_key(parent, &right_key, pslot + 1);
1927                         btrfs_mark_buffer_dirty(parent);
1928                 }
1929         }
1930         if (btrfs_header_nritems(mid) == 1) {
1931                 /*
1932                  * we're not allowed to leave a node with one item in the
1933                  * tree during a delete.  A deletion from lower in the tree
1934                  * could try to delete the only pointer in this node.
1935                  * So, pull some keys from the left.
1936                  * There has to be a left pointer at this point because
1937                  * otherwise we would have pulled some pointers from the
1938                  * right
1939                  */
1940                 if (!left) {
1941                         ret = -EROFS;
1942                         btrfs_handle_fs_error(fs_info, ret, NULL);
1943                         goto enospc;
1944                 }
1945                 wret = balance_node_right(trans, fs_info, mid, left);
1946                 if (wret < 0) {
1947                         ret = wret;
1948                         goto enospc;
1949                 }
1950                 if (wret == 1) {
1951                         wret = push_node_left(trans, fs_info, left, mid, 1);
1952                         if (wret < 0)
1953                                 ret = wret;
1954                 }
1955                 BUG_ON(wret == 1);
1956         }
1957         if (btrfs_header_nritems(mid) == 0) {
1958                 clean_tree_block(fs_info, mid);
1959                 btrfs_tree_unlock(mid);
1960                 del_ptr(root, path, level + 1, pslot);
1961                 root_sub_used(root, mid->len);
1962                 btrfs_free_tree_block(trans, root, mid, 0, 1);
1963                 free_extent_buffer_stale(mid);
1964                 mid = NULL;
1965         } else {
1966                 /* update the parent key to reflect our changes */
1967                 struct btrfs_disk_key mid_key;
1968                 btrfs_node_key(mid, &mid_key, 0);
1969                 ret = tree_mod_log_insert_key(parent, pslot,
1970                                 MOD_LOG_KEY_REPLACE, GFP_NOFS);
1971                 BUG_ON(ret < 0);
1972                 btrfs_set_node_key(parent, &mid_key, pslot);
1973                 btrfs_mark_buffer_dirty(parent);
1974         }
1975
1976         /* update the path */
1977         if (left) {
1978                 if (btrfs_header_nritems(left) > orig_slot) {
1979                         extent_buffer_get(left);
1980                         /* left was locked after cow */
1981                         path->nodes[level] = left;
1982                         path->slots[level + 1] -= 1;
1983                         path->slots[level] = orig_slot;
1984                         if (mid) {
1985                                 btrfs_tree_unlock(mid);
1986                                 free_extent_buffer(mid);
1987                         }
1988                 } else {
1989                         orig_slot -= btrfs_header_nritems(left);
1990                         path->slots[level] = orig_slot;
1991                 }
1992         }
1993         /* double check we haven't messed things up */
1994         if (orig_ptr !=
1995             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
1996                 BUG();
1997 enospc:
1998         if (right) {
1999                 btrfs_tree_unlock(right);
2000                 free_extent_buffer(right);
2001         }
2002         if (left) {
2003                 if (path->nodes[level] != left)
2004                         btrfs_tree_unlock(left);
2005                 free_extent_buffer(left);
2006         }
2007         return ret;
2008 }
2009
2010 /* Node balancing for insertion.  Here we only split or push nodes around
2011  * when they are completely full.  This is also done top down, so we
2012  * have to be pessimistic.
2013  */
2014 static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
2015                                           struct btrfs_root *root,
2016                                           struct btrfs_path *path, int level)
2017 {
2018         struct btrfs_fs_info *fs_info = root->fs_info;
2019         struct extent_buffer *right = NULL;
2020         struct extent_buffer *mid;
2021         struct extent_buffer *left = NULL;
2022         struct extent_buffer *parent = NULL;
2023         int ret = 0;
2024         int wret;
2025         int pslot;
2026         int orig_slot = path->slots[level];
2027
2028         if (level == 0)
2029                 return 1;
2030
2031         mid = path->nodes[level];
2032         WARN_ON(btrfs_header_generation(mid) != trans->transid);
2033
2034         if (level < BTRFS_MAX_LEVEL - 1) {
2035                 parent = path->nodes[level + 1];
2036                 pslot = path->slots[level + 1];
2037         }
2038
2039         if (!parent)
2040                 return 1;
2041
2042         left = read_node_slot(fs_info, parent, pslot - 1);
2043         if (IS_ERR(left))
2044                 left = NULL;
2045
2046         /* first, try to make some room in the middle buffer */
2047         if (left) {
2048                 u32 left_nr;
2049
2050                 btrfs_tree_lock(left);
2051                 btrfs_set_lock_blocking(left);
2052
2053                 left_nr = btrfs_header_nritems(left);
2054                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2055                         wret = 1;
2056                 } else {
2057                         ret = btrfs_cow_block(trans, root, left, parent,
2058                                               pslot - 1, &left);
2059                         if (ret)
2060                                 wret = 1;
2061                         else {
2062                                 wret = push_node_left(trans, fs_info,
2063                                                       left, mid, 0);
2064                         }
2065                 }
2066                 if (wret < 0)
2067                         ret = wret;
2068                 if (wret == 0) {
2069                         struct btrfs_disk_key disk_key;
2070                         orig_slot += left_nr;
2071                         btrfs_node_key(mid, &disk_key, 0);
2072                         ret = tree_mod_log_insert_key(parent, pslot,
2073                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2074                         BUG_ON(ret < 0);
2075                         btrfs_set_node_key(parent, &disk_key, pslot);
2076                         btrfs_mark_buffer_dirty(parent);
2077                         if (btrfs_header_nritems(left) > orig_slot) {
2078                                 path->nodes[level] = left;
2079                                 path->slots[level + 1] -= 1;
2080                                 path->slots[level] = orig_slot;
2081                                 btrfs_tree_unlock(mid);
2082                                 free_extent_buffer(mid);
2083                         } else {
2084                                 orig_slot -=
2085                                         btrfs_header_nritems(left);
2086                                 path->slots[level] = orig_slot;
2087                                 btrfs_tree_unlock(left);
2088                                 free_extent_buffer(left);
2089                         }
2090                         return 0;
2091                 }
2092                 btrfs_tree_unlock(left);
2093                 free_extent_buffer(left);
2094         }
2095         right = read_node_slot(fs_info, parent, pslot + 1);
2096         if (IS_ERR(right))
2097                 right = NULL;
2098
2099         /*
2100          * then try to empty the right most buffer into the middle
2101          */
2102         if (right) {
2103                 u32 right_nr;
2104
2105                 btrfs_tree_lock(right);
2106                 btrfs_set_lock_blocking(right);
2107
2108                 right_nr = btrfs_header_nritems(right);
2109                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
2110                         wret = 1;
2111                 } else {
2112                         ret = btrfs_cow_block(trans, root, right,
2113                                               parent, pslot + 1,
2114                                               &right);
2115                         if (ret)
2116                                 wret = 1;
2117                         else {
2118                                 wret = balance_node_right(trans, fs_info,
2119                                                           right, mid);
2120                         }
2121                 }
2122                 if (wret < 0)
2123                         ret = wret;
2124                 if (wret == 0) {
2125                         struct btrfs_disk_key disk_key;
2126
2127                         btrfs_node_key(right, &disk_key, 0);
2128                         ret = tree_mod_log_insert_key(parent, pslot + 1,
2129                                         MOD_LOG_KEY_REPLACE, GFP_NOFS);
2130                         BUG_ON(ret < 0);
2131                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
2132                         btrfs_mark_buffer_dirty(parent);
2133
2134                         if (btrfs_header_nritems(mid) <= orig_slot) {
2135                                 path->nodes[level] = right;
2136                                 path->slots[level + 1] += 1;
2137                                 path->slots[level] = orig_slot -
2138                                         btrfs_header_nritems(mid);
2139                                 btrfs_tree_unlock(mid);
2140                                 free_extent_buffer(mid);
2141                         } else {
2142                                 btrfs_tree_unlock(right);
2143                                 free_extent_buffer(right);
2144                         }
2145                         return 0;
2146                 }
2147                 btrfs_tree_unlock(right);
2148                 free_extent_buffer(right);
2149         }
2150         return 1;
2151 }
2152
2153 /*
2154  * readahead one full node of leaves, finding things that are close
2155  * to the block in 'slot', and triggering ra on them.
2156  */
2157 static void reada_for_search(struct btrfs_fs_info *fs_info,
2158                              struct btrfs_path *path,
2159                              int level, int slot, u64 objectid)
2160 {
2161         struct extent_buffer *node;
2162         struct btrfs_disk_key disk_key;
2163         u32 nritems;
2164         u64 search;
2165         u64 target;
2166         u64 nread = 0;
2167         struct extent_buffer *eb;
2168         u32 nr;
2169         u32 blocksize;
2170         u32 nscan = 0;
2171
2172         if (level != 1)
2173                 return;
2174
2175         if (!path->nodes[level])
2176                 return;
2177
2178         node = path->nodes[level];
2179
2180         search = btrfs_node_blockptr(node, slot);
2181         blocksize = fs_info->nodesize;
2182         eb = find_extent_buffer(fs_info, search);
2183         if (eb) {
2184                 free_extent_buffer(eb);
2185                 return;
2186         }
2187
2188         target = search;
2189
2190         nritems = btrfs_header_nritems(node);
2191         nr = slot;
2192
2193         while (1) {
2194                 if (path->reada == READA_BACK) {
2195                         if (nr == 0)
2196                                 break;
2197                         nr--;
2198                 } else if (path->reada == READA_FORWARD) {
2199                         nr++;
2200                         if (nr >= nritems)
2201                                 break;
2202                 }
2203                 if (path->reada == READA_BACK && objectid) {
2204                         btrfs_node_key(node, &disk_key, nr);
2205                         if (btrfs_disk_key_objectid(&disk_key) != objectid)
2206                                 break;
2207                 }
2208                 search = btrfs_node_blockptr(node, nr);
2209                 if ((search <= target && target - search <= 65536) ||
2210                     (search > target && search - target <= 65536)) {
2211                         readahead_tree_block(fs_info, search);
2212                         nread += blocksize;
2213                 }
2214                 nscan++;
2215                 if ((nread > 65536 || nscan > 32))
2216                         break;
2217         }
2218 }
2219
2220 static noinline void reada_for_balance(struct btrfs_fs_info *fs_info,
2221                                        struct btrfs_path *path, int level)
2222 {
2223         int slot;
2224         int nritems;
2225         struct extent_buffer *parent;
2226         struct extent_buffer *eb;
2227         u64 gen;
2228         u64 block1 = 0;
2229         u64 block2 = 0;
2230
2231         parent = path->nodes[level + 1];
2232         if (!parent)
2233                 return;
2234
2235         nritems = btrfs_header_nritems(parent);
2236         slot = path->slots[level + 1];
2237
2238         if (slot > 0) {
2239                 block1 = btrfs_node_blockptr(parent, slot - 1);
2240                 gen = btrfs_node_ptr_generation(parent, slot - 1);
2241                 eb = find_extent_buffer(fs_info, block1);
2242                 /*
2243                  * if we get -eagain from btrfs_buffer_uptodate, we
2244                  * don't want to return eagain here.  That will loop
2245                  * forever
2246                  */
2247                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2248                         block1 = 0;
2249                 free_extent_buffer(eb);
2250         }
2251         if (slot + 1 < nritems) {
2252                 block2 = btrfs_node_blockptr(parent, slot + 1);
2253                 gen = btrfs_node_ptr_generation(parent, slot + 1);
2254                 eb = find_extent_buffer(fs_info, block2);
2255                 if (eb && btrfs_buffer_uptodate(eb, gen, 1) != 0)
2256                         block2 = 0;
2257                 free_extent_buffer(eb);
2258         }
2259
2260         if (block1)
2261                 readahead_tree_block(fs_info, block1);
2262         if (block2)
2263                 readahead_tree_block(fs_info, block2);
2264 }
2265
2266
2267 /*
2268  * when we walk down the tree, it is usually safe to unlock the higher layers
2269  * in the tree.  The exceptions are when our path goes through slot 0, because
2270  * operations on the tree might require changing key pointers higher up in the
2271  * tree.
2272  *
2273  * callers might also have set path->keep_locks, which tells this code to keep
2274  * the lock if the path points to the last slot in the block.  This is part of
2275  * walking through the tree, and selecting the next slot in the higher block.
2276  *
2277  * lowest_unlock sets the lowest level in the tree we're allowed to unlock.  so
2278  * if lowest_unlock is 1, level 0 won't be unlocked
2279  */
2280 static noinline void unlock_up(struct btrfs_path *path, int level,
2281                                int lowest_unlock, int min_write_lock_level,
2282                                int *write_lock_level)
2283 {
2284         int i;
2285         int skip_level = level;
2286         int no_skips = 0;
2287         struct extent_buffer *t;
2288
2289         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2290                 if (!path->nodes[i])
2291                         break;
2292                 if (!path->locks[i])
2293                         break;
2294                 if (!no_skips && path->slots[i] == 0) {
2295                         skip_level = i + 1;
2296                         continue;
2297                 }
2298                 if (!no_skips && path->keep_locks) {
2299                         u32 nritems;
2300                         t = path->nodes[i];
2301                         nritems = btrfs_header_nritems(t);
2302                         if (nritems < 1 || path->slots[i] >= nritems - 1) {
2303                                 skip_level = i + 1;
2304                                 continue;
2305                         }
2306                 }
2307                 if (skip_level < i && i >= lowest_unlock)
2308                         no_skips = 1;
2309
2310                 t = path->nodes[i];
2311                 if (i >= lowest_unlock && i > skip_level) {
2312                         btrfs_tree_unlock_rw(t, path->locks[i]);
2313                         path->locks[i] = 0;
2314                         if (write_lock_level &&
2315                             i > min_write_lock_level &&
2316                             i <= *write_lock_level) {
2317                                 *write_lock_level = i - 1;
2318                         }
2319                 }
2320         }
2321 }
2322
2323 /*
2324  * This releases any locks held in the path starting at level and
2325  * going all the way up to the root.
2326  *
2327  * btrfs_search_slot will keep the lock held on higher nodes in a few
2328  * corner cases, such as COW of the block at slot zero in the node.  This
2329  * ignores those rules, and it should only be called when there are no
2330  * more updates to be done higher up in the tree.
2331  */
2332 noinline void btrfs_unlock_up_safe(struct btrfs_path *path, int level)
2333 {
2334         int i;
2335
2336         if (path->keep_locks)
2337                 return;
2338
2339         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
2340                 if (!path->nodes[i])
2341                         continue;
2342                 if (!path->locks[i])
2343                         continue;
2344                 btrfs_tree_unlock_rw(path->nodes[i], path->locks[i]);
2345                 path->locks[i] = 0;
2346         }
2347 }
2348
2349 /*
2350  * helper function for btrfs_search_slot.  The goal is to find a block
2351  * in cache without setting the path to blocking.  If we find the block
2352  * we return zero and the path is unchanged.
2353  *
2354  * If we can't find the block, we set the path blocking and do some
2355  * reada.  -EAGAIN is returned and the search must be repeated.
2356  */
2357 static int
2358 read_block_for_search(struct btrfs_root *root, struct btrfs_path *p,
2359                       struct extent_buffer **eb_ret, int level, int slot,
2360                       const struct btrfs_key *key)
2361 {
2362         struct btrfs_fs_info *fs_info = root->fs_info;
2363         u64 blocknr;
2364         u64 gen;
2365         struct extent_buffer *b = *eb_ret;
2366         struct extent_buffer *tmp;
2367         struct btrfs_key first_key;
2368         int ret;
2369         int parent_level;
2370
2371         blocknr = btrfs_node_blockptr(b, slot);
2372         gen = btrfs_node_ptr_generation(b, slot);
2373         parent_level = btrfs_header_level(b);
2374         btrfs_node_key_to_cpu(b, &first_key, slot);
2375
2376         tmp = find_extent_buffer(fs_info, blocknr);
2377         if (tmp) {
2378                 /* first we do an atomic uptodate check */
2379                 if (btrfs_buffer_uptodate(tmp, gen, 1) > 0) {
2380                         *eb_ret = tmp;
2381                         return 0;
2382                 }
2383
2384                 /* the pages were up to date, but we failed
2385                  * the generation number check.  Do a full
2386                  * read for the generation number that is correct.
2387                  * We must do this without dropping locks so
2388                  * we can trust our generation number
2389                  */
2390                 btrfs_set_path_blocking(p);
2391
2392                 /* now we're allowed to do a blocking uptodate check */
2393                 ret = btrfs_read_buffer(tmp, gen, parent_level - 1, &first_key);
2394                 if (!ret) {
2395                         *eb_ret = tmp;
2396                         return 0;
2397                 }
2398                 free_extent_buffer(tmp);
2399                 btrfs_release_path(p);
2400                 return -EIO;
2401         }
2402
2403         /*
2404          * reduce lock contention at high levels
2405          * of the btree by dropping locks before
2406          * we read.  Don't release the lock on the current
2407          * level because we need to walk this node to figure
2408          * out which blocks to read.
2409          */
2410         btrfs_unlock_up_safe(p, level + 1);
2411         btrfs_set_path_blocking(p);
2412
2413         if (p->reada != READA_NONE)
2414                 reada_for_search(fs_info, p, level, slot, key->objectid);
2415
2416         ret = -EAGAIN;
2417         tmp = read_tree_block(fs_info, blocknr, gen, parent_level - 1,
2418                               &first_key);
2419         if (!IS_ERR(tmp)) {
2420                 /*
2421                  * If the read above didn't mark this buffer up to date,
2422                  * it will never end up being up to date.  Set ret to EIO now
2423                  * and give up so that our caller doesn't loop forever
2424                  * on our EAGAINs.
2425                  */
2426                 if (!extent_buffer_uptodate(tmp))
2427                         ret = -EIO;
2428                 free_extent_buffer(tmp);
2429         } else {
2430                 ret = PTR_ERR(tmp);
2431         }
2432
2433         btrfs_release_path(p);
2434         return ret;
2435 }
2436
2437 /*
2438  * helper function for btrfs_search_slot.  This does all of the checks
2439  * for node-level blocks and does any balancing required based on
2440  * the ins_len.
2441  *
2442  * If no extra work was required, zero is returned.  If we had to
2443  * drop the path, -EAGAIN is returned and btrfs_search_slot must
2444  * start over
2445  */
2446 static int
2447 setup_nodes_for_search(struct btrfs_trans_handle *trans,
2448                        struct btrfs_root *root, struct btrfs_path *p,
2449                        struct extent_buffer *b, int level, int ins_len,
2450                        int *write_lock_level)
2451 {
2452         struct btrfs_fs_info *fs_info = root->fs_info;
2453         int ret;
2454
2455         if ((p->search_for_split || ins_len > 0) && btrfs_header_nritems(b) >=
2456             BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 3) {
2457                 int sret;
2458
2459                 if (*write_lock_level < level + 1) {
2460                         *write_lock_level = level + 1;
2461                         btrfs_release_path(p);
2462                         goto again;
2463                 }
2464
2465                 btrfs_set_path_blocking(p);
2466                 reada_for_balance(fs_info, p, level);
2467                 sret = split_node(trans, root, p, level);
2468
2469                 BUG_ON(sret > 0);
2470                 if (sret) {
2471                         ret = sret;
2472                         goto done;
2473                 }
2474                 b = p->nodes[level];
2475         } else if (ins_len < 0 && btrfs_header_nritems(b) <
2476                    BTRFS_NODEPTRS_PER_BLOCK(fs_info) / 2) {
2477                 int sret;
2478
2479                 if (*write_lock_level < level + 1) {
2480                         *write_lock_level = level + 1;
2481                         btrfs_release_path(p);
2482                         goto again;
2483                 }
2484
2485                 btrfs_set_path_blocking(p);
2486                 reada_for_balance(fs_info, p, level);
2487                 sret = balance_level(trans, root, p, level);
2488
2489                 if (sret) {
2490                         ret = sret;
2491                         goto done;
2492                 }
2493                 b = p->nodes[level];
2494                 if (!b) {
2495                         btrfs_release_path(p);
2496                         goto again;
2497                 }
2498                 BUG_ON(btrfs_header_nritems(b) == 1);
2499         }
2500         return 0;
2501
2502 again:
2503         ret = -EAGAIN;
2504 done:
2505         return ret;
2506 }
2507
2508 static void key_search_validate(struct extent_buffer *b,
2509                                 const struct btrfs_key *key,
2510                                 int level)
2511 {
2512 #ifdef CONFIG_BTRFS_ASSERT
2513         struct btrfs_disk_key disk_key;
2514
2515         btrfs_cpu_key_to_disk(&disk_key, key);
2516
2517         if (level == 0)
2518                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2519                     offsetof(struct btrfs_leaf, items[0].key),
2520                     sizeof(disk_key)));
2521         else
2522                 ASSERT(!memcmp_extent_buffer(b, &disk_key,
2523                     offsetof(struct btrfs_node, ptrs[0].key),
2524                     sizeof(disk_key)));
2525 #endif
2526 }
2527
2528 static int key_search(struct extent_buffer *b, const struct btrfs_key *key,
2529                       int level, int *prev_cmp, int *slot)
2530 {
2531         if (*prev_cmp != 0) {
2532                 *prev_cmp = btrfs_bin_search(b, key, level, slot);
2533                 return *prev_cmp;
2534         }
2535
2536         key_search_validate(b, key, level);
2537         *slot = 0;
2538
2539         return 0;
2540 }
2541
2542 int btrfs_find_item(struct btrfs_root *fs_root, struct btrfs_path *path,
2543                 u64 iobjectid, u64 ioff, u8 key_type,
2544                 struct btrfs_key *found_key)
2545 {
2546         int ret;
2547         struct btrfs_key key;
2548         struct extent_buffer *eb;
2549
2550         ASSERT(path);
2551         ASSERT(found_key);
2552
2553         key.type = key_type;
2554         key.objectid = iobjectid;
2555         key.offset = ioff;
2556
2557         ret = btrfs_search_slot(NULL, fs_root, &key, path, 0, 0);
2558         if (ret < 0)
2559                 return ret;
2560
2561         eb = path->nodes[0];
2562         if (ret && path->slots[0] >= btrfs_header_nritems(eb)) {
2563                 ret = btrfs_next_leaf(fs_root, path);
2564                 if (ret)
2565                         return ret;
2566                 eb = path->nodes[0];
2567         }
2568
2569         btrfs_item_key_to_cpu(eb, found_key, path->slots[0]);
2570         if (found_key->type != key.type ||
2571                         found_key->objectid != key.objectid)
2572                 return 1;
2573
2574         return 0;
2575 }
2576
2577 static struct extent_buffer *btrfs_search_slot_get_root(struct btrfs_root *root,
2578                                                         struct btrfs_path *p,
2579                                                         int write_lock_level)
2580 {
2581         struct btrfs_fs_info *fs_info = root->fs_info;
2582         struct extent_buffer *b;
2583         int root_lock;
2584         int level = 0;
2585
2586         /* We try very hard to do read locks on the root */
2587         root_lock = BTRFS_READ_LOCK;
2588
2589         if (p->search_commit_root) {
2590                 /*
2591                  * The commit roots are read only so we always do read locks,
2592                  * and we always must hold the commit_root_sem when doing
2593                  * searches on them, the only exception is send where we don't
2594                  * want to block transaction commits for a long time, so
2595                  * we need to clone the commit root in order to avoid races
2596                  * with transaction commits that create a snapshot of one of
2597                  * the roots used by a send operation.
2598                  */
2599                 if (p->need_commit_sem) {
2600                         down_read(&fs_info->commit_root_sem);
2601                         b = btrfs_clone_extent_buffer(root->commit_root);
2602                         up_read(&fs_info->commit_root_sem);
2603                         if (!b)
2604                                 return ERR_PTR(-ENOMEM);
2605
2606                 } else {
2607                         b = root->commit_root;
2608                         extent_buffer_get(b);
2609                 }
2610                 level = btrfs_header_level(b);
2611                 /*
2612                  * Ensure that all callers have set skip_locking when
2613                  * p->search_commit_root = 1.
2614                  */
2615                 ASSERT(p->skip_locking == 1);
2616
2617                 goto out;
2618         }
2619
2620         if (p->skip_locking) {
2621                 b = btrfs_root_node(root);
2622                 level = btrfs_header_level(b);
2623                 goto out;
2624         }
2625
2626         /*
2627          * If the level is set to maximum, we can skip trying to get the read
2628          * lock.
2629          */
2630         if (write_lock_level < BTRFS_MAX_LEVEL) {
2631                 /*
2632                  * We don't know the level of the root node until we actually
2633                  * have it read locked
2634                  */
2635                 b = btrfs_read_lock_root_node(root);
2636                 level = btrfs_header_level(b);
2637                 if (level > write_lock_level)
2638                         goto out;
2639
2640                 /* Whoops, must trade for write lock */
2641                 btrfs_tree_read_unlock(b);
2642                 free_extent_buffer(b);
2643         }
2644
2645         b = btrfs_lock_root_node(root);
2646         root_lock = BTRFS_WRITE_LOCK;
2647
2648         /* The level might have changed, check again */
2649         level = btrfs_header_level(b);
2650
2651 out:
2652         p->nodes[level] = b;
2653         if (!p->skip_locking)
2654                 p->locks[level] = root_lock;
2655         /*
2656          * Callers are responsible for dropping b's references.
2657          */
2658         return b;
2659 }
2660
2661
2662 /*
2663  * btrfs_search_slot - look for a key in a tree and perform necessary
2664  * modifications to preserve tree invariants.
2665  *
2666  * @trans:      Handle of transaction, used when modifying the tree
2667  * @p:          Holds all btree nodes along the search path
2668  * @root:       The root node of the tree
2669  * @key:        The key we are looking for
2670  * @ins_len:    Indicates purpose of search, for inserts it is 1, for
2671  *              deletions it's -1. 0 for plain searches
2672  * @cow:        boolean should CoW operations be performed. Must always be 1
2673  *              when modifying the tree.
2674  *
2675  * If @ins_len > 0, nodes and leaves will be split as we walk down the tree.
2676  * If @ins_len < 0, nodes will be merged as we walk down the tree (if possible)
2677  *
2678  * If @key is found, 0 is returned and you can find the item in the leaf level
2679  * of the path (level 0)
2680  *
2681  * If @key isn't found, 1 is returned and the leaf level of the path (level 0)
2682  * points to the slot where it should be inserted
2683  *
2684  * If an error is encountered while searching the tree a negative error number
2685  * is returned
2686  */
2687 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2688                       const struct btrfs_key *key, struct btrfs_path *p,
2689                       int ins_len, int cow)
2690 {
2691         struct btrfs_fs_info *fs_info = root->fs_info;
2692         struct extent_buffer *b;
2693         int slot;
2694         int ret;
2695         int err;
2696         int level;
2697         int lowest_unlock = 1;
2698         /* everything at write_lock_level or lower must be write locked */
2699         int write_lock_level = 0;
2700         u8 lowest_level = 0;
2701         int min_write_lock_level;
2702         int prev_cmp;
2703
2704         lowest_level = p->lowest_level;
2705         WARN_ON(lowest_level && ins_len > 0);
2706         WARN_ON(p->nodes[0] != NULL);
2707         BUG_ON(!cow && ins_len);
2708
2709         if (ins_len < 0) {
2710                 lowest_unlock = 2;
2711
2712                 /* when we are removing items, we might have to go up to level
2713                  * two as we update tree pointers  Make sure we keep write
2714                  * for those levels as well
2715                  */
2716                 write_lock_level = 2;
2717         } else if (ins_len > 0) {
2718                 /*
2719                  * for inserting items, make sure we have a write lock on
2720                  * level 1 so we can update keys
2721                  */
2722                 write_lock_level = 1;
2723         }
2724
2725         if (!cow)
2726                 write_lock_level = -1;
2727
2728         if (cow && (p->keep_locks || p->lowest_level))
2729                 write_lock_level = BTRFS_MAX_LEVEL;
2730
2731         min_write_lock_level = write_lock_level;
2732
2733 again:
2734         prev_cmp = -1;
2735         b = btrfs_search_slot_get_root(root, p, write_lock_level);
2736         if (IS_ERR(b)) {
2737                 ret = PTR_ERR(b);
2738                 goto done;
2739         }
2740
2741         while (b) {
2742                 level = btrfs_header_level(b);
2743
2744                 /*
2745                  * setup the path here so we can release it under lock
2746                  * contention with the cow code
2747                  */
2748                 if (cow) {
2749                         bool last_level = (level == (BTRFS_MAX_LEVEL - 1));
2750
2751                         /*
2752                          * if we don't really need to cow this block
2753                          * then we don't want to set the path blocking,
2754                          * so we test it here
2755                          */
2756                         if (!should_cow_block(trans, root, b)) {
2757                                 trans->dirty = true;
2758                                 goto cow_done;
2759                         }
2760
2761                         /*
2762                          * must have write locks on this node and the
2763                          * parent
2764                          */
2765                         if (level > write_lock_level ||
2766                             (level + 1 > write_lock_level &&
2767                             level + 1 < BTRFS_MAX_LEVEL &&
2768                             p->nodes[level + 1])) {
2769                                 write_lock_level = level + 1;
2770                                 btrfs_release_path(p);
2771                                 goto again;
2772                         }
2773
2774                         btrfs_set_path_blocking(p);
2775                         if (last_level)
2776                                 err = btrfs_cow_block(trans, root, b, NULL, 0,
2777                                                       &b);
2778                         else
2779                                 err = btrfs_cow_block(trans, root, b,
2780                                                       p->nodes[level + 1],
2781                                                       p->slots[level + 1], &b);
2782                         if (err) {
2783                                 ret = err;
2784                                 goto done;
2785                         }
2786                 }
2787 cow_done:
2788                 p->nodes[level] = b;
2789                 /*
2790                  * Leave path with blocking locks to avoid massive
2791                  * lock context switch, this is made on purpose.
2792                  */
2793
2794                 /*
2795                  * we have a lock on b and as long as we aren't changing
2796                  * the tree, there is no way to for the items in b to change.
2797                  * It is safe to drop the lock on our parent before we
2798                  * go through the expensive btree search on b.
2799                  *
2800                  * If we're inserting or deleting (ins_len != 0), then we might
2801                  * be changing slot zero, which may require changing the parent.
2802                  * So, we can't drop the lock until after we know which slot
2803                  * we're operating on.
2804                  */
2805                 if (!ins_len && !p->keep_locks) {
2806                         int u = level + 1;
2807
2808                         if (u < BTRFS_MAX_LEVEL && p->locks[u]) {
2809                                 btrfs_tree_unlock_rw(p->nodes[u], p->locks[u]);
2810                                 p->locks[u] = 0;
2811                         }
2812                 }
2813
2814                 ret = key_search(b, key, level, &prev_cmp, &slot);
2815                 if (ret < 0)
2816                         goto done;
2817
2818                 if (level != 0) {
2819                         int dec = 0;
2820                         if (ret && slot > 0) {
2821                                 dec = 1;
2822                                 slot -= 1;
2823                         }
2824                         p->slots[level] = slot;
2825                         err = setup_nodes_for_search(trans, root, p, b, level,
2826                                              ins_len, &write_lock_level);
2827                         if (err == -EAGAIN)
2828                                 goto again;
2829                         if (err) {
2830                                 ret = err;
2831                                 goto done;
2832                         }
2833                         b = p->nodes[level];
2834                         slot = p->slots[level];
2835
2836                         /*
2837                          * slot 0 is special, if we change the key
2838                          * we have to update the parent pointer
2839                          * which means we must have a write lock
2840                          * on the parent
2841                          */
2842                         if (slot == 0 && ins_len &&
2843                             write_lock_level < level + 1) {
2844                                 write_lock_level = level + 1;
2845                                 btrfs_release_path(p);
2846                                 goto again;
2847                         }
2848
2849                         unlock_up(p, level, lowest_unlock,
2850                                   min_write_lock_level, &write_lock_level);
2851
2852                         if (level == lowest_level) {
2853                                 if (dec)
2854                                         p->slots[level]++;
2855                                 goto done;
2856                         }
2857
2858                         err = read_block_for_search(root, p, &b, level,
2859                                                     slot, key);
2860                         if (err == -EAGAIN)
2861                                 goto again;
2862                         if (err) {
2863                                 ret = err;
2864                                 goto done;
2865                         }
2866
2867                         if (!p->skip_locking) {
2868                                 level = btrfs_header_level(b);
2869                                 if (level <= write_lock_level) {
2870                                         err = btrfs_try_tree_write_lock(b);
2871                                         if (!err) {
2872                                                 btrfs_set_path_blocking(p);
2873                                                 btrfs_tree_lock(b);
2874                                         }
2875                                         p->locks[level] = BTRFS_WRITE_LOCK;
2876                                 } else {
2877                                         err = btrfs_tree_read_lock_atomic(b);
2878                                         if (!err) {
2879                                                 btrfs_set_path_blocking(p);
2880                                                 btrfs_tree_read_lock(b);
2881                                         }
2882                                         p->locks[level] = BTRFS_READ_LOCK;
2883                                 }
2884                                 p->nodes[level] = b;
2885                         }
2886                 } else {
2887                         p->slots[level] = slot;
2888                         if (ins_len > 0 &&
2889                             btrfs_leaf_free_space(fs_info, b) < ins_len) {
2890                                 if (write_lock_level < 1) {
2891                                         write_lock_level = 1;
2892                                         btrfs_release_path(p);
2893                                         goto again;
2894                                 }
2895
2896                                 btrfs_set_path_blocking(p);
2897                                 err = split_leaf(trans, root, key,
2898                                                  p, ins_len, ret == 0);
2899
2900                                 BUG_ON(err > 0);
2901                                 if (err) {
2902                                         ret = err;
2903                                         goto done;
2904                                 }
2905                         }
2906                         if (!p->search_for_split)
2907                                 unlock_up(p, level, lowest_unlock,
2908                                           min_write_lock_level, NULL);
2909                         goto done;
2910                 }
2911         }
2912         ret = 1;
2913 done:
2914         /*
2915          * we don't really know what they plan on doing with the path
2916          * from here on, so for now just mark it as blocking
2917          */
2918         if (!p->leave_spinning)
2919                 btrfs_set_path_blocking(p);
2920         if (ret < 0 && !p->skip_release_on_error)
2921                 btrfs_release_path(p);
2922         return ret;
2923 }
2924
2925 /*
2926  * Like btrfs_search_slot, this looks for a key in the given tree. It uses the
2927  * current state of the tree together with the operations recorded in the tree
2928  * modification log to search for the key in a previous version of this tree, as
2929  * denoted by the time_seq parameter.
2930  *
2931  * Naturally, there is no support for insert, delete or cow operations.
2932  *
2933  * The resulting path and return value will be set up as if we called
2934  * btrfs_search_slot at that point in time with ins_len and cow both set to 0.
2935  */
2936 int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
2937                           struct btrfs_path *p, u64 time_seq)
2938 {
2939         struct btrfs_fs_info *fs_info = root->fs_info;
2940         struct extent_buffer *b;
2941         int slot;
2942         int ret;
2943         int err;
2944         int level;
2945         int lowest_unlock = 1;
2946         u8 lowest_level = 0;
2947         int prev_cmp = -1;
2948
2949         lowest_level = p->lowest_level;
2950         WARN_ON(p->nodes[0] != NULL);
2951
2952         if (p->search_commit_root) {
2953                 BUG_ON(time_seq);
2954                 return btrfs_search_slot(NULL, root, key, p, 0, 0);
2955         }
2956
2957 again:
2958         b = get_old_root(root, time_seq);
2959         if (!b) {
2960                 ret = -EIO;
2961                 goto done;
2962         }
2963         level = btrfs_header_level(b);
2964         p->locks[level] = BTRFS_READ_LOCK;
2965
2966         while (b) {
2967                 level = btrfs_header_level(b);
2968                 p->nodes[level] = b;
2969
2970                 /*
2971                  * we have a lock on b and as long as we aren't changing
2972                  * the tree, there is no way to for the items in b to change.
2973                  * It is safe to drop the lock on our parent before we
2974                  * go through the expensive btree search on b.
2975                  */
2976                 btrfs_unlock_up_safe(p, level + 1);
2977
2978                 /*
2979                  * Since we can unwind ebs we want to do a real search every
2980                  * time.
2981                  */
2982                 prev_cmp = -1;
2983                 ret = key_search(b, key, level, &prev_cmp, &slot);
2984
2985                 if (level != 0) {
2986                         int dec = 0;
2987                         if (ret && slot > 0) {
2988                                 dec = 1;
2989                                 slot -= 1;
2990                         }
2991                         p->slots[level] = slot;
2992                         unlock_up(p, level, lowest_unlock, 0, NULL);
2993
2994                         if (level == lowest_level) {
2995                                 if (dec)
2996                                         p->slots[level]++;
2997                                 goto done;
2998                         }
2999
3000                         err = read_block_for_search(root, p, &b, level,
3001                                                     slot, key);
3002                         if (err == -EAGAIN)
3003                                 goto again;
3004                         if (err) {
3005                                 ret = err;
3006                                 goto done;
3007                         }
3008
3009                         level = btrfs_header_level(b);
3010                         err = btrfs_tree_read_lock_atomic(b);
3011                         if (!err) {
3012                                 btrfs_set_path_blocking(p);
3013                                 btrfs_tree_read_lock(b);
3014                         }
3015                         b = tree_mod_log_rewind(fs_info, p, b, time_seq);
3016                         if (!b) {
3017                                 ret = -ENOMEM;
3018                                 goto done;
3019                         }
3020                         p->locks[level] = BTRFS_READ_LOCK;
3021                         p->nodes[level] = b;
3022                 } else {
3023                         p->slots[level] = slot;
3024                         unlock_up(p, level, lowest_unlock, 0, NULL);
3025                         goto done;
3026                 }
3027         }
3028         ret = 1;
3029 done:
3030         if (!p->leave_spinning)
3031                 btrfs_set_path_blocking(p);
3032         if (ret < 0)
3033                 btrfs_release_path(p);
3034
3035         return ret;
3036 }
3037
3038 /*
3039  * helper to use instead of search slot if no exact match is needed but
3040  * instead the next or previous item should be returned.
3041  * When find_higher is true, the next higher item is returned, the next lower
3042  * otherwise.
3043  * When return_any and find_higher are both true, and no higher item is found,
3044  * return the next lower instead.
3045  * When return_any is true and find_higher is false, and no lower item is found,
3046  * return the next higher instead.
3047  * It returns 0 if any item is found, 1 if none is found (tree empty), and
3048  * < 0 on error
3049  */
3050 int btrfs_search_slot_for_read(struct btrfs_root *root,
3051                                const struct btrfs_key *key,
3052                                struct btrfs_path *p, int find_higher,
3053                                int return_any)
3054 {
3055         int ret;
3056         struct extent_buffer *leaf;
3057
3058 again:
3059         ret = btrfs_search_slot(NULL, root, key, p, 0, 0);
3060         if (ret <= 0)
3061                 return ret;
3062         /*
3063          * a return value of 1 means the path is at the position where the
3064          * item should be inserted. Normally this is the next bigger item,
3065          * but in case the previous item is the last in a leaf, path points
3066          * to the first free slot in the previous leaf, i.e. at an invalid
3067          * item.
3068          */
3069         leaf = p->nodes[0];
3070
3071         if (find_higher) {
3072                 if (p->slots[0] >= btrfs_header_nritems(leaf)) {
3073                         ret = btrfs_next_leaf(root, p);
3074                         if (ret <= 0)
3075                                 return ret;
3076                         if (!return_any)
3077                                 return 1;
3078                         /*
3079                          * no higher item found, return the next
3080                          * lower instead
3081                          */
3082                         return_any = 0;
3083                         find_higher = 0;
3084                         btrfs_release_path(p);
3085                         goto again;
3086                 }
3087         } else {
3088                 if (p->slots[0] == 0) {
3089                         ret = btrfs_prev_leaf(root, p);
3090                         if (ret < 0)
3091                                 return ret;
3092                         if (!ret) {
3093                                 leaf = p->nodes[0];
3094                                 if (p->slots[0] == btrfs_header_nritems(leaf))
3095                                         p->slots[0]--;
3096                                 return 0;
3097                         }
3098                         if (!return_any)
3099                                 return 1;
3100                         /*
3101                          * no lower item found, return the next
3102                          * higher instead
3103                          */
3104                         return_any = 0;
3105                         find_higher = 1;
3106                         btrfs_release_path(p);
3107                         goto again;
3108                 } else {
3109                         --p->slots[0];
3110                 }
3111         }
3112         return 0;
3113 }
3114
3115 /*
3116  * adjust the pointers going up the tree, starting at level
3117  * making sure the right key of each node is points to 'key'.
3118  * This is used after shifting pointers to the left, so it stops
3119  * fixing up pointers when a given leaf/node is not in slot 0 of the
3120  * higher levels
3121  *
3122  */
3123 static void fixup_low_keys(struct btrfs_path *path,
3124                            struct btrfs_disk_key *key, int level)
3125 {
3126         int i;
3127         struct extent_buffer *t;
3128         int ret;
3129
3130         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
3131                 int tslot = path->slots[i];
3132
3133                 if (!path->nodes[i])
3134                         break;
3135                 t&n